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 if (!Matcher::has_match_rule(Op_PopCountI)) return NULL;
342 break;
343
344 case vmIntrinsics::_bitCount_l:
345 if (!Matcher::has_match_rule(Op_PopCountL)) return NULL;
346 break;
347
348 case vmIntrinsics::_numberOfLeadingZeros_i:
349 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
350 break;
351
352 case vmIntrinsics::_numberOfLeadingZeros_l:
353 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
354 break;
355
356 case vmIntrinsics::_numberOfTrailingZeros_i:
357 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
358 break;
359
360 case vmIntrinsics::_numberOfTrailingZeros_l:
361 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
362 break;
363
364 case vmIntrinsics::_Reference_get:
365 // It is only when G1 is enabled that we absolutely
366 // need to use the intrinsic version of Reference.get()
367 // so that the value in the referent field, if necessary,
368 // can be registered by the pre-barrier code.
369 if (!UseG1GC) return NULL;
370 break;
371
372 default:
373 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
374 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
375 break;
376 }
377
378 // -XX:-InlineClassNatives disables natives from the Class class.
379 // The flag applies to all reflective calls, notably Array.newArray
380 // (visible to Java programmers as Array.newInstance).
381 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
382 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
383 if (!InlineClassNatives) return NULL;
384 }
385
386 // -XX:-InlineThreadNatives disables natives from the Thread class.
387 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
388 if (!InlineThreadNatives) return NULL;
389 }
390
391 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
392 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
393 m->holder()->name() == ciSymbol::java_lang_Float() ||
394 m->holder()->name() == ciSymbol::java_lang_Double()) {
395 if (!InlineMathNatives) return NULL;
396 }
397
398 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
399 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
400 if (!InlineUnsafeOps) return NULL;
401 }
402
403 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
404 }
405
406 //----------------------register_library_intrinsics-----------------------
407 // Initialize this file's data structures, for each Compile instance.
408 void Compile::register_library_intrinsics() {
409 // Nothing to do here.
410 }
411
412 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
413 LibraryCallKit kit(jvms, this);
414 Compile* C = kit.C;
415 int nodes = C->unique();
416 #ifndef PRODUCT
417 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
418 char buf[1000];
419 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
420 tty->print_cr("Intrinsic %s", str);
421 }
422 #endif
423
424 if (kit.try_to_inline()) {
425 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
426 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
427 }
428 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
429 if (C->log()) {
430 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
431 vmIntrinsics::name_at(intrinsic_id()),
432 (is_virtual() ? " virtual='1'" : ""),
433 C->unique() - nodes);
434 }
435 return kit.transfer_exceptions_into_jvms();
436 }
437
438 // The intrinsic bailed out
439 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
440 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
441 if (jvms->has_method()) {
442 // Not a root compile.
443 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg);
444 } else {
445 // Root compile
446 tty->print("Did not generate intrinsic %s%s at bci:%d in",
447 vmIntrinsics::name_at(intrinsic_id()),
448 (is_virtual() ? " (virtual)" : ""), kit.bci());
449 }
450 }
451 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
452 return NULL;
453 }
454
455 bool LibraryCallKit::try_to_inline() {
456 // Handle symbolic names for otherwise undistinguished boolean switches:
457 const bool is_store = true;
458 const bool is_native_ptr = true;
459 const bool is_static = true;
460
461 if (!jvms()->has_method()) {
462 // Root JVMState has a null method.
463 assert(map()->memory()->Opcode() == Op_Parm, "");
464 // Insert the memory aliasing node
465 set_all_memory(reset_memory());
466 }
467 assert(merged_memory(), "");
468
469 switch (intrinsic_id()) {
470 case vmIntrinsics::_hashCode:
471 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
472 case vmIntrinsics::_identityHashCode:
473 return inline_native_hashcode(/*!virtual*/ false, is_static);
474 case vmIntrinsics::_getClass:
475 return inline_native_getClass();
476
477 case vmIntrinsics::_dsin:
478 case vmIntrinsics::_dcos:
479 case vmIntrinsics::_dtan:
480 case vmIntrinsics::_dabs:
481 case vmIntrinsics::_datan2:
482 case vmIntrinsics::_dsqrt:
483 case vmIntrinsics::_dexp:
484 case vmIntrinsics::_dlog:
485 case vmIntrinsics::_dlog10:
486 case vmIntrinsics::_dpow:
487 return inline_math_native(intrinsic_id());
488
489 case vmIntrinsics::_min:
490 case vmIntrinsics::_max:
491 return inline_min_max(intrinsic_id());
492
493 case vmIntrinsics::_arraycopy:
494 return inline_arraycopy();
495
496 case vmIntrinsics::_compareTo:
497 return inline_string_compareTo();
498 case vmIntrinsics::_indexOf:
499 return inline_string_indexOf();
500 case vmIntrinsics::_equals:
501 return inline_string_equals();
502
503 case vmIntrinsics::_getObject:
504 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
505 case vmIntrinsics::_getBoolean:
506 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
507 case vmIntrinsics::_getByte:
508 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
509 case vmIntrinsics::_getShort:
510 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
511 case vmIntrinsics::_getChar:
512 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
513 case vmIntrinsics::_getInt:
514 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
515 case vmIntrinsics::_getLong:
516 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
517 case vmIntrinsics::_getFloat:
518 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
519 case vmIntrinsics::_getDouble:
520 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
521
522 case vmIntrinsics::_putObject:
523 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
524 case vmIntrinsics::_putBoolean:
525 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
526 case vmIntrinsics::_putByte:
527 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
528 case vmIntrinsics::_putShort:
529 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
530 case vmIntrinsics::_putChar:
531 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
532 case vmIntrinsics::_putInt:
533 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
534 case vmIntrinsics::_putLong:
535 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
536 case vmIntrinsics::_putFloat:
537 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
538 case vmIntrinsics::_putDouble:
539 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
540
541 case vmIntrinsics::_getByte_raw:
542 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
543 case vmIntrinsics::_getShort_raw:
544 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
545 case vmIntrinsics::_getChar_raw:
546 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
547 case vmIntrinsics::_getInt_raw:
548 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
549 case vmIntrinsics::_getLong_raw:
550 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
551 case vmIntrinsics::_getFloat_raw:
552 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
553 case vmIntrinsics::_getDouble_raw:
554 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
555 case vmIntrinsics::_getAddress_raw:
556 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
557
558 case vmIntrinsics::_putByte_raw:
559 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
560 case vmIntrinsics::_putShort_raw:
561 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
562 case vmIntrinsics::_putChar_raw:
563 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
564 case vmIntrinsics::_putInt_raw:
565 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
566 case vmIntrinsics::_putLong_raw:
567 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
568 case vmIntrinsics::_putFloat_raw:
569 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
570 case vmIntrinsics::_putDouble_raw:
571 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
572 case vmIntrinsics::_putAddress_raw:
573 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
574
575 case vmIntrinsics::_getObjectVolatile:
576 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
577 case vmIntrinsics::_getBooleanVolatile:
578 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
579 case vmIntrinsics::_getByteVolatile:
580 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
581 case vmIntrinsics::_getShortVolatile:
582 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
583 case vmIntrinsics::_getCharVolatile:
584 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
585 case vmIntrinsics::_getIntVolatile:
586 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
587 case vmIntrinsics::_getLongVolatile:
588 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
589 case vmIntrinsics::_getFloatVolatile:
590 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
591 case vmIntrinsics::_getDoubleVolatile:
592 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
593
594 case vmIntrinsics::_putObjectVolatile:
595 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
596 case vmIntrinsics::_putBooleanVolatile:
597 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
598 case vmIntrinsics::_putByteVolatile:
599 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
600 case vmIntrinsics::_putShortVolatile:
601 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
602 case vmIntrinsics::_putCharVolatile:
603 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
604 case vmIntrinsics::_putIntVolatile:
605 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
606 case vmIntrinsics::_putLongVolatile:
607 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
608 case vmIntrinsics::_putFloatVolatile:
609 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
610 case vmIntrinsics::_putDoubleVolatile:
611 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
612
613 case vmIntrinsics::_prefetchRead:
614 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
615 case vmIntrinsics::_prefetchWrite:
616 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
617 case vmIntrinsics::_prefetchReadStatic:
618 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
619 case vmIntrinsics::_prefetchWriteStatic:
620 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
621
622 case vmIntrinsics::_compareAndSwapObject:
623 return inline_unsafe_CAS(T_OBJECT);
624 case vmIntrinsics::_compareAndSwapInt:
625 return inline_unsafe_CAS(T_INT);
626 case vmIntrinsics::_compareAndSwapLong:
627 return inline_unsafe_CAS(T_LONG);
628
629 case vmIntrinsics::_putOrderedObject:
630 return inline_unsafe_ordered_store(T_OBJECT);
631 case vmIntrinsics::_putOrderedInt:
632 return inline_unsafe_ordered_store(T_INT);
633 case vmIntrinsics::_putOrderedLong:
634 return inline_unsafe_ordered_store(T_LONG);
635
636 case vmIntrinsics::_currentThread:
637 return inline_native_currentThread();
638 case vmIntrinsics::_isInterrupted:
639 return inline_native_isInterrupted();
640
641 case vmIntrinsics::_currentTimeMillis:
642 return inline_native_time_funcs(false);
643 case vmIntrinsics::_nanoTime:
644 return inline_native_time_funcs(true);
645 case vmIntrinsics::_allocateInstance:
646 return inline_unsafe_allocate();
647 case vmIntrinsics::_copyMemory:
648 return inline_unsafe_copyMemory();
649 case vmIntrinsics::_newArray:
650 return inline_native_newArray();
651 case vmIntrinsics::_getLength:
652 return inline_native_getLength();
653 case vmIntrinsics::_copyOf:
654 return inline_array_copyOf(false);
655 case vmIntrinsics::_copyOfRange:
656 return inline_array_copyOf(true);
657 case vmIntrinsics::_equalsC:
658 return inline_array_equals();
659 case vmIntrinsics::_clone:
660 return inline_native_clone(intrinsic()->is_virtual());
661
662 case vmIntrinsics::_isAssignableFrom:
663 return inline_native_subtype_check();
664
665 case vmIntrinsics::_isInstance:
666 case vmIntrinsics::_getModifiers:
667 case vmIntrinsics::_isInterface:
668 case vmIntrinsics::_isArray:
669 case vmIntrinsics::_isPrimitive:
670 case vmIntrinsics::_getSuperclass:
671 case vmIntrinsics::_getComponentType:
672 case vmIntrinsics::_getClassAccessFlags:
673 return inline_native_Class_query(intrinsic_id());
674
675 case vmIntrinsics::_floatToRawIntBits:
676 case vmIntrinsics::_floatToIntBits:
677 case vmIntrinsics::_intBitsToFloat:
678 case vmIntrinsics::_doubleToRawLongBits:
679 case vmIntrinsics::_doubleToLongBits:
680 case vmIntrinsics::_longBitsToDouble:
681 return inline_fp_conversions(intrinsic_id());
682
683 case vmIntrinsics::_numberOfLeadingZeros_i:
684 case vmIntrinsics::_numberOfLeadingZeros_l:
685 return inline_numberOfLeadingZeros(intrinsic_id());
686
687 case vmIntrinsics::_numberOfTrailingZeros_i:
688 case vmIntrinsics::_numberOfTrailingZeros_l:
689 return inline_numberOfTrailingZeros(intrinsic_id());
690
691 case vmIntrinsics::_bitCount_i:
692 case vmIntrinsics::_bitCount_l:
693 return inline_bitCount(intrinsic_id());
694
695 case vmIntrinsics::_reverseBytes_i:
696 case vmIntrinsics::_reverseBytes_l:
697 case vmIntrinsics::_reverseBytes_s:
698 case vmIntrinsics::_reverseBytes_c:
699 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
700
701 case vmIntrinsics::_get_AtomicLong:
702 return inline_native_AtomicLong_get();
703 case vmIntrinsics::_attemptUpdate:
704 return inline_native_AtomicLong_attemptUpdate();
705
706 case vmIntrinsics::_getCallerClass:
707 return inline_native_Reflection_getCallerClass();
708
709 case vmIntrinsics::_Reference_get:
710 return inline_reference_get();
711
712 default:
713 // If you get here, it may be that someone has added a new intrinsic
714 // to the list in vmSymbols.hpp without implementing it here.
715 #ifndef PRODUCT
716 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
717 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
718 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
719 }
720 #endif
721 return false;
722 }
723 }
724
725 //------------------------------push_result------------------------------
726 // Helper function for finishing intrinsics.
727 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
728 record_for_igvn(region);
729 set_control(_gvn.transform(region));
730 BasicType value_type = value->type()->basic_type();
731 push_node(value_type, _gvn.transform(value));
732 }
733
734 //------------------------------generate_guard---------------------------
735 // Helper function for generating guarded fast-slow graph structures.
736 // The given 'test', if true, guards a slow path. If the test fails
737 // then a fast path can be taken. (We generally hope it fails.)
738 // In all cases, GraphKit::control() is updated to the fast path.
739 // The returned value represents the control for the slow path.
740 // The return value is never 'top'; it is either a valid control
741 // or NULL if it is obvious that the slow path can never be taken.
742 // Also, if region and the slow control are not NULL, the slow edge
743 // is appended to the region.
744 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
745 if (stopped()) {
746 // Already short circuited.
747 return NULL;
748 }
749
750 // Build an if node and its projections.
751 // If test is true we take the slow path, which we assume is uncommon.
752 if (_gvn.type(test) == TypeInt::ZERO) {
753 // The slow branch is never taken. No need to build this guard.
754 return NULL;
755 }
756
757 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
758
759 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
760 if (if_slow == top()) {
761 // The slow branch is never taken. No need to build this guard.
762 return NULL;
763 }
764
765 if (region != NULL)
766 region->add_req(if_slow);
767
768 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
769 set_control(if_fast);
770
771 return if_slow;
772 }
773
774 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
775 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
776 }
777 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
778 return generate_guard(test, region, PROB_FAIR);
779 }
780
781 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
782 Node* *pos_index) {
783 if (stopped())
784 return NULL; // already stopped
785 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
786 return NULL; // index is already adequately typed
787 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
788 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
789 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
790 if (is_neg != NULL && pos_index != NULL) {
791 // Emulate effect of Parse::adjust_map_after_if.
792 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
793 ccast->set_req(0, control());
794 (*pos_index) = _gvn.transform(ccast);
795 }
796 return is_neg;
797 }
798
799 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
800 Node* *pos_index) {
801 if (stopped())
802 return NULL; // already stopped
803 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
804 return NULL; // index is already adequately typed
805 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
806 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
807 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
808 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
809 if (is_notp != NULL && pos_index != NULL) {
810 // Emulate effect of Parse::adjust_map_after_if.
811 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
812 ccast->set_req(0, control());
813 (*pos_index) = _gvn.transform(ccast);
814 }
815 return is_notp;
816 }
817
818 // Make sure that 'position' is a valid limit index, in [0..length].
819 // There are two equivalent plans for checking this:
820 // A. (offset + copyLength) unsigned<= arrayLength
821 // B. offset <= (arrayLength - copyLength)
822 // We require that all of the values above, except for the sum and
823 // difference, are already known to be non-negative.
824 // Plan A is robust in the face of overflow, if offset and copyLength
825 // are both hugely positive.
826 //
827 // Plan B is less direct and intuitive, but it does not overflow at
828 // all, since the difference of two non-negatives is always
829 // representable. Whenever Java methods must perform the equivalent
830 // check they generally use Plan B instead of Plan A.
831 // For the moment we use Plan A.
832 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
833 Node* subseq_length,
834 Node* array_length,
835 RegionNode* region) {
836 if (stopped())
837 return NULL; // already stopped
838 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
839 if (zero_offset && subseq_length->eqv_uncast(array_length))
840 return NULL; // common case of whole-array copy
841 Node* last = subseq_length;
842 if (!zero_offset) // last += offset
843 last = _gvn.transform( new (C, 3) AddINode(last, offset));
844 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
845 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
846 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
847 return is_over;
848 }
849
850
851 //--------------------------generate_current_thread--------------------
852 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
853 ciKlass* thread_klass = env()->Thread_klass();
854 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
855 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
856 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
857 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
858 tls_output = thread;
859 return threadObj;
860 }
861
862
863 //------------------------------make_string_method_node------------------------
864 // Helper method for String intrinsic finctions.
865 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
866 const int value_offset = java_lang_String::value_offset_in_bytes();
867 const int count_offset = java_lang_String::count_offset_in_bytes();
868 const int offset_offset = java_lang_String::offset_offset_in_bytes();
869
870 Node* no_ctrl = NULL;
871
872 ciInstanceKlass* klass = env()->String_klass();
873 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
874
875 const TypeAryPtr* value_type =
876 TypeAryPtr::make(TypePtr::NotNull,
877 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
878 ciTypeArrayKlass::make(T_CHAR), true, 0);
879
880 // Get start addr of string and substring
881 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
882 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
883 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
884 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
885 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
886
887 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
888 Node* str2_value = make_load(no_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
889 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
890 Node* str2_offset = make_load(no_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
891 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
892
893 Node* result = NULL;
894 switch (opcode) {
895 case Op_StrIndexOf:
896 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
897 str1_start, cnt1, str2_start, cnt2);
898 break;
899 case Op_StrComp:
900 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
901 str1_start, cnt1, str2_start, cnt2);
902 break;
903 case Op_StrEquals:
904 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
905 str1_start, str2_start, cnt1);
906 break;
907 default:
908 ShouldNotReachHere();
909 return NULL;
910 }
911
912 // All these intrinsics have checks.
913 C->set_has_split_ifs(true); // Has chance for split-if optimization
914
915 return _gvn.transform(result);
916 }
917
918 //------------------------------inline_string_compareTo------------------------
919 bool LibraryCallKit::inline_string_compareTo() {
920
921 if (!Matcher::has_match_rule(Op_StrComp)) return false;
922
923 const int value_offset = java_lang_String::value_offset_in_bytes();
924 const int count_offset = java_lang_String::count_offset_in_bytes();
925 const int offset_offset = java_lang_String::offset_offset_in_bytes();
926
927 _sp += 2;
928 Node *argument = pop(); // pop non-receiver first: it was pushed second
929 Node *receiver = pop();
930
931 // Null check on self without removing any arguments. The argument
932 // null check technically happens in the wrong place, which can lead to
933 // invalid stack traces when string compare is inlined into a method
934 // which handles NullPointerExceptions.
935 _sp += 2;
936 receiver = do_null_check(receiver, T_OBJECT);
937 argument = do_null_check(argument, T_OBJECT);
938 _sp -= 2;
939 if (stopped()) {
940 return true;
941 }
942
943 ciInstanceKlass* klass = env()->String_klass();
944 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
945 Node* no_ctrl = NULL;
946
947 // Get counts for string and argument
948 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
949 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
950
951 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
952 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
953
954 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
955 push(compare);
956 return true;
957 }
958
959 //------------------------------inline_string_equals------------------------
960 bool LibraryCallKit::inline_string_equals() {
961
962 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
963
964 const int value_offset = java_lang_String::value_offset_in_bytes();
965 const int count_offset = java_lang_String::count_offset_in_bytes();
966 const int offset_offset = java_lang_String::offset_offset_in_bytes();
967
968 int nargs = 2;
969 _sp += nargs;
970 Node* argument = pop(); // pop non-receiver first: it was pushed second
971 Node* receiver = pop();
972
973 // Null check on self without removing any arguments. The argument
974 // null check technically happens in the wrong place, which can lead to
975 // invalid stack traces when string compare is inlined into a method
976 // which handles NullPointerExceptions.
977 _sp += nargs;
978 receiver = do_null_check(receiver, T_OBJECT);
979 //should not do null check for argument for String.equals(), because spec
980 //allows to specify NULL as argument.
981 _sp -= nargs;
982
983 if (stopped()) {
984 return true;
985 }
986
987 // paths (plus control) merge
988 RegionNode* region = new (C, 5) RegionNode(5);
989 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
990
991 // does source == target string?
992 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
993 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
994
995 Node* if_eq = generate_slow_guard(bol, NULL);
996 if (if_eq != NULL) {
997 // receiver == argument
998 phi->init_req(2, intcon(1));
999 region->init_req(2, if_eq);
1000 }
1001
1002 // get String klass for instanceOf
1003 ciInstanceKlass* klass = env()->String_klass();
1004
1005 if (!stopped()) {
1006 _sp += nargs; // gen_instanceof might do an uncommon trap
1007 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1008 _sp -= nargs;
1009 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
1010 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
1011
1012 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1013 //instanceOf == true, fallthrough
1014
1015 if (inst_false != NULL) {
1016 phi->init_req(3, intcon(0));
1017 region->init_req(3, inst_false);
1018 }
1019 }
1020
1021 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1022
1023 Node* no_ctrl = NULL;
1024 Node* receiver_cnt;
1025 Node* argument_cnt;
1026
1027 if (!stopped()) {
1028 // Properly cast the argument to String
1029 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
1030 // This path is taken only when argument's type is String:NotNull.
1031 argument = cast_not_null(argument, false);
1032
1033 // Get counts for string and argument
1034 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
1035 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1036
1037 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
1038 argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1039
1040 // Check for receiver count != argument count
1041 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
1042 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
1043 Node* if_ne = generate_slow_guard(bol, NULL);
1044 if (if_ne != NULL) {
1045 phi->init_req(4, intcon(0));
1046 region->init_req(4, if_ne);
1047 }
1048 }
1049
1050 // Check for count == 0 is done by mach node StrEquals.
1051
1052 if (!stopped()) {
1053 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1054 phi->init_req(1, equals);
1055 region->init_req(1, control());
1056 }
1057
1058 // post merge
1059 set_control(_gvn.transform(region));
1060 record_for_igvn(region);
1061
1062 push(_gvn.transform(phi));
1063
1064 return true;
1065 }
1066
1067 //------------------------------inline_array_equals----------------------------
1068 bool LibraryCallKit::inline_array_equals() {
1069
1070 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1071
1072 _sp += 2;
1073 Node *argument2 = pop();
1074 Node *argument1 = pop();
1075
1076 Node* equals =
1077 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1078 argument1, argument2) );
1079 push(equals);
1080 return true;
1081 }
1082
1083 // Java version of String.indexOf(constant string)
1084 // class StringDecl {
1085 // StringDecl(char[] ca) {
1086 // offset = 0;
1087 // count = ca.length;
1088 // value = ca;
1089 // }
1090 // int offset;
1091 // int count;
1092 // char[] value;
1093 // }
1094 //
1095 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1096 // int targetOffset, int cache_i, int md2) {
1097 // int cache = cache_i;
1098 // int sourceOffset = string_object.offset;
1099 // int sourceCount = string_object.count;
1100 // int targetCount = target_object.length;
1101 //
1102 // int targetCountLess1 = targetCount - 1;
1103 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1104 //
1105 // char[] source = string_object.value;
1106 // char[] target = target_object;
1107 // int lastChar = target[targetCountLess1];
1108 //
1109 // outer_loop:
1110 // for (int i = sourceOffset; i < sourceEnd; ) {
1111 // int src = source[i + targetCountLess1];
1112 // if (src == lastChar) {
1113 // // With random strings and a 4-character alphabet,
1114 // // reverse matching at this point sets up 0.8% fewer
1115 // // frames, but (paradoxically) makes 0.3% more probes.
1116 // // Since those probes are nearer the lastChar probe,
1117 // // there is may be a net D$ win with reverse matching.
1118 // // But, reversing loop inhibits unroll of inner loop
1119 // // for unknown reason. So, does running outer loop from
1120 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1121 // for (int j = 0; j < targetCountLess1; j++) {
1122 // if (target[targetOffset + j] != source[i+j]) {
1123 // if ((cache & (1 << source[i+j])) == 0) {
1124 // if (md2 < j+1) {
1125 // i += j+1;
1126 // continue outer_loop;
1127 // }
1128 // }
1129 // i += md2;
1130 // continue outer_loop;
1131 // }
1132 // }
1133 // return i - sourceOffset;
1134 // }
1135 // if ((cache & (1 << src)) == 0) {
1136 // i += targetCountLess1;
1137 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1138 // i++;
1139 // }
1140 // return -1;
1141 // }
1142
1143 //------------------------------string_indexOf------------------------
1144 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1145 jint cache_i, jint md2_i) {
1146
1147 Node* no_ctrl = NULL;
1148 float likely = PROB_LIKELY(0.9);
1149 float unlikely = PROB_UNLIKELY(0.9);
1150
1151 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1152
1153 const int value_offset = java_lang_String::value_offset_in_bytes();
1154 const int count_offset = java_lang_String::count_offset_in_bytes();
1155 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1156
1157 ciInstanceKlass* klass = env()->String_klass();
1158 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1159 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1160
1161 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1162 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1163 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1164 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1165 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1166 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1167
1168 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1169 jint target_length = target_array->length();
1170 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1171 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1172
1173 IdealKit kit(this, false, true);
1174 #define __ kit.
1175 Node* zero = __ ConI(0);
1176 Node* one = __ ConI(1);
1177 Node* cache = __ ConI(cache_i);
1178 Node* md2 = __ ConI(md2_i);
1179 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1180 Node* targetCount = __ ConI(target_length);
1181 Node* targetCountLess1 = __ ConI(target_length - 1);
1182 Node* targetOffset = __ ConI(targetOffset_i);
1183 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1184
1185 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1186 Node* outer_loop = __ make_label(2 /* goto */);
1187 Node* return_ = __ make_label(1);
1188
1189 __ set(rtn,__ ConI(-1));
1190 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1191 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1192 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1193 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1194 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1195 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1196 Node* tpj = __ AddI(targetOffset, __ value(j));
1197 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1198 Node* ipj = __ AddI(__ value(i), __ value(j));
1199 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1200 __ if_then(targ, BoolTest::ne, src2); {
1201 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1202 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1203 __ increment(i, __ AddI(__ value(j), one));
1204 __ goto_(outer_loop);
1205 } __ end_if(); __ dead(j);
1206 }__ end_if(); __ dead(j);
1207 __ increment(i, md2);
1208 __ goto_(outer_loop);
1209 }__ end_if();
1210 __ increment(j, one);
1211 }__ end_loop(); __ dead(j);
1212 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1213 __ goto_(return_);
1214 }__ end_if();
1215 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1216 __ increment(i, targetCountLess1);
1217 }__ end_if();
1218 __ increment(i, one);
1219 __ bind(outer_loop);
1220 }__ end_loop(); __ dead(i);
1221 __ bind(return_);
1222
1223 // Final sync IdealKit and GraphKit.
1224 final_sync(kit);
1225 Node* result = __ value(rtn);
1226 #undef __
1227 C->set_has_loops(true);
1228 return result;
1229 }
1230
1231 //------------------------------inline_string_indexOf------------------------
1232 bool LibraryCallKit::inline_string_indexOf() {
1233
1234 const int value_offset = java_lang_String::value_offset_in_bytes();
1235 const int count_offset = java_lang_String::count_offset_in_bytes();
1236 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1237
1238 _sp += 2;
1239 Node *argument = pop(); // pop non-receiver first: it was pushed second
1240 Node *receiver = pop();
1241
1242 Node* result;
1243 // Disable the use of pcmpestri until it can be guaranteed that
1244 // the load doesn't cross into the uncommited space.
1245 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1246 UseSSE42Intrinsics) {
1247 // Generate SSE4.2 version of indexOf
1248 // We currently only have match rules that use SSE4.2
1249
1250 // Null check on self without removing any arguments. The argument
1251 // null check technically happens in the wrong place, which can lead to
1252 // invalid stack traces when string compare is inlined into a method
1253 // which handles NullPointerExceptions.
1254 _sp += 2;
1255 receiver = do_null_check(receiver, T_OBJECT);
1256 argument = do_null_check(argument, T_OBJECT);
1257 _sp -= 2;
1258
1259 if (stopped()) {
1260 return true;
1261 }
1262
1263 ciInstanceKlass* str_klass = env()->String_klass();
1264 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1265
1266 // Make the merge point
1267 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1268 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1269 Node* no_ctrl = NULL;
1270
1271 // Get counts for string and substr
1272 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1273 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1274
1275 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1276 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1277
1278 // Check for substr count > string count
1279 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1280 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1281 Node* if_gt = generate_slow_guard(bol, NULL);
1282 if (if_gt != NULL) {
1283 result_phi->init_req(2, intcon(-1));
1284 result_rgn->init_req(2, if_gt);
1285 }
1286
1287 if (!stopped()) {
1288 // Check for substr count == 0
1289 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1290 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1291 Node* if_zero = generate_slow_guard(bol, NULL);
1292 if (if_zero != NULL) {
1293 result_phi->init_req(3, intcon(0));
1294 result_rgn->init_req(3, if_zero);
1295 }
1296 }
1297
1298 if (!stopped()) {
1299 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1300 result_phi->init_req(1, result);
1301 result_rgn->init_req(1, control());
1302 }
1303 set_control(_gvn.transform(result_rgn));
1304 record_for_igvn(result_rgn);
1305 result = _gvn.transform(result_phi);
1306
1307 } else { // Use LibraryCallKit::string_indexOf
1308 // don't intrinsify if argument isn't a constant string.
1309 if (!argument->is_Con()) {
1310 return false;
1311 }
1312 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1313 if (str_type == NULL) {
1314 return false;
1315 }
1316 ciInstanceKlass* klass = env()->String_klass();
1317 ciObject* str_const = str_type->const_oop();
1318 if (str_const == NULL || str_const->klass() != klass) {
1319 return false;
1320 }
1321 ciInstance* str = str_const->as_instance();
1322 assert(str != NULL, "must be instance");
1323
1324 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1325 int o = str->field_value_by_offset(offset_offset).as_int();
1326 int c = str->field_value_by_offset(count_offset).as_int();
1327 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1328
1329 // constant strings have no offset and count == length which
1330 // simplifies the resulting code somewhat so lets optimize for that.
1331 if (o != 0 || c != pat->length()) {
1332 return false;
1333 }
1334
1335 // Null check on self without removing any arguments. The argument
1336 // null check technically happens in the wrong place, which can lead to
1337 // invalid stack traces when string compare is inlined into a method
1338 // which handles NullPointerExceptions.
1339 _sp += 2;
1340 receiver = do_null_check(receiver, T_OBJECT);
1341 // No null check on the argument is needed since it's a constant String oop.
1342 _sp -= 2;
1343 if (stopped()) {
1344 return true;
1345 }
1346
1347 // The null string as a pattern always returns 0 (match at beginning of string)
1348 if (c == 0) {
1349 push(intcon(0));
1350 return true;
1351 }
1352
1353 // Generate default indexOf
1354 jchar lastChar = pat->char_at(o + (c - 1));
1355 int cache = 0;
1356 int i;
1357 for (i = 0; i < c - 1; i++) {
1358 assert(i < pat->length(), "out of range");
1359 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1360 }
1361
1362 int md2 = c;
1363 for (i = 0; i < c - 1; i++) {
1364 assert(i < pat->length(), "out of range");
1365 if (pat->char_at(o + i) == lastChar) {
1366 md2 = (c - 1) - i;
1367 }
1368 }
1369
1370 result = string_indexOf(receiver, pat, o, cache, md2);
1371 }
1372
1373 push(result);
1374 return true;
1375 }
1376
1377 //--------------------------pop_math_arg--------------------------------
1378 // Pop a double argument to a math function from the stack
1379 // rounding it if necessary.
1380 Node * LibraryCallKit::pop_math_arg() {
1381 Node *arg = pop_pair();
1382 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1383 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1384 return arg;
1385 }
1386
1387 //------------------------------inline_trig----------------------------------
1388 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1389 // argument reduction which will turn into a fast/slow diamond.
1390 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1391 _sp += arg_size(); // restore stack pointer
1392 Node* arg = pop_math_arg();
1393 Node* trig = NULL;
1394
1395 switch (id) {
1396 case vmIntrinsics::_dsin:
1397 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1398 break;
1399 case vmIntrinsics::_dcos:
1400 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1401 break;
1402 case vmIntrinsics::_dtan:
1403 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1404 break;
1405 default:
1406 assert(false, "bad intrinsic was passed in");
1407 return false;
1408 }
1409
1410 // Rounding required? Check for argument reduction!
1411 if( Matcher::strict_fp_requires_explicit_rounding ) {
1412
1413 static const double pi_4 = 0.7853981633974483;
1414 static const double neg_pi_4 = -0.7853981633974483;
1415 // pi/2 in 80-bit extended precision
1416 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1417 // -pi/2 in 80-bit extended precision
1418 // 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};
1419 // Cutoff value for using this argument reduction technique
1420 //static const double pi_2_minus_epsilon = 1.564660403643354;
1421 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1422
1423 // Pseudocode for sin:
1424 // if (x <= Math.PI / 4.0) {
1425 // if (x >= -Math.PI / 4.0) return fsin(x);
1426 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1427 // } else {
1428 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1429 // }
1430 // return StrictMath.sin(x);
1431
1432 // Pseudocode for cos:
1433 // if (x <= Math.PI / 4.0) {
1434 // if (x >= -Math.PI / 4.0) return fcos(x);
1435 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1436 // } else {
1437 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1438 // }
1439 // return StrictMath.cos(x);
1440
1441 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1442 // requires a special machine instruction to load it. Instead we'll try
1443 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1444 // probably do the math inside the SIN encoding.
1445
1446 // Make the merge point
1447 RegionNode *r = new (C, 3) RegionNode(3);
1448 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1449
1450 // Flatten arg so we need only 1 test
1451 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1452 // Node for PI/4 constant
1453 Node *pi4 = makecon(TypeD::make(pi_4));
1454 // Check PI/4 : abs(arg)
1455 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1456 // Check: If PI/4 < abs(arg) then go slow
1457 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1458 // Branch either way
1459 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1460 set_control(opt_iff(r,iff));
1461
1462 // Set fast path result
1463 phi->init_req(2,trig);
1464
1465 // Slow path - non-blocking leaf call
1466 Node* call = NULL;
1467 switch (id) {
1468 case vmIntrinsics::_dsin:
1469 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1470 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1471 "Sin", NULL, arg, top());
1472 break;
1473 case vmIntrinsics::_dcos:
1474 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1475 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1476 "Cos", NULL, arg, top());
1477 break;
1478 case vmIntrinsics::_dtan:
1479 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1480 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1481 "Tan", NULL, arg, top());
1482 break;
1483 }
1484 assert(control()->in(0) == call, "");
1485 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1486 r->init_req(1,control());
1487 phi->init_req(1,slow_result);
1488
1489 // Post-merge
1490 set_control(_gvn.transform(r));
1491 record_for_igvn(r);
1492 trig = _gvn.transform(phi);
1493
1494 C->set_has_split_ifs(true); // Has chance for split-if optimization
1495 }
1496 // Push result back on JVM stack
1497 push_pair(trig);
1498 return true;
1499 }
1500
1501 //------------------------------inline_sqrt-------------------------------------
1502 // Inline square root instruction, if possible.
1503 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1504 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1505 _sp += arg_size(); // restore stack pointer
1506 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1507 return true;
1508 }
1509
1510 //------------------------------inline_abs-------------------------------------
1511 // Inline absolute value instruction, if possible.
1512 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1513 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1514 _sp += arg_size(); // restore stack pointer
1515 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1516 return true;
1517 }
1518
1519 //------------------------------inline_exp-------------------------------------
1520 // Inline exp instructions, if possible. The Intel hardware only misses
1521 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1522 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1523 assert(id == vmIntrinsics::_dexp, "Not exp");
1524
1525 // If this inlining ever returned NaN in the past, we do not intrinsify it
1526 // every again. NaN results requires StrictMath.exp handling.
1527 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1528
1529 // Do not intrinsify on older platforms which lack cmove.
1530 if (ConditionalMoveLimit == 0) return false;
1531
1532 _sp += arg_size(); // restore stack pointer
1533 Node *x = pop_math_arg();
1534 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1535
1536 //-------------------
1537 //result=(result.isNaN())? StrictMath::exp():result;
1538 // Check: If isNaN() by checking result!=result? then go to Strict Math
1539 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1540 // Build the boolean node
1541 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1542
1543 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1544 // End the current control-flow path
1545 push_pair(x);
1546 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1547 // to handle. Recompile without intrinsifying Math.exp
1548 uncommon_trap(Deoptimization::Reason_intrinsic,
1549 Deoptimization::Action_make_not_entrant);
1550 }
1551
1552 C->set_has_split_ifs(true); // Has chance for split-if optimization
1553
1554 push_pair(result);
1555
1556 return true;
1557 }
1558
1559 //------------------------------inline_pow-------------------------------------
1560 // Inline power instructions, if possible.
1561 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1562 assert(id == vmIntrinsics::_dpow, "Not pow");
1563
1564 // If this inlining ever returned NaN in the past, we do not intrinsify it
1565 // every again. NaN results requires StrictMath.pow handling.
1566 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1567
1568 // Do not intrinsify on older platforms which lack cmove.
1569 if (ConditionalMoveLimit == 0) return false;
1570
1571 // Pseudocode for pow
1572 // if (x <= 0.0) {
1573 // if ((double)((int)y)==y) { // if y is int
1574 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1575 // } else {
1576 // result = NaN;
1577 // }
1578 // } else {
1579 // result = DPow(x,y);
1580 // }
1581 // if (result != result)? {
1582 // uncommon_trap();
1583 // }
1584 // return result;
1585
1586 _sp += arg_size(); // restore stack pointer
1587 Node* y = pop_math_arg();
1588 Node* x = pop_math_arg();
1589
1590 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1591
1592 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1593 // inside of something) then skip the fancy tests and just check for
1594 // NaN result.
1595 Node *result = NULL;
1596 if( jvms()->depth() >= 1 ) {
1597 result = fast_result;
1598 } else {
1599
1600 // Set the merge point for If node with condition of (x <= 0.0)
1601 // There are four possible paths to region node and phi node
1602 RegionNode *r = new (C, 4) RegionNode(4);
1603 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1604
1605 // Build the first if node: if (x <= 0.0)
1606 // Node for 0 constant
1607 Node *zeronode = makecon(TypeD::ZERO);
1608 // Check x:0
1609 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1610 // Check: If (x<=0) then go complex path
1611 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1612 // Branch either way
1613 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1614 Node *opt_test = _gvn.transform(if1);
1615 //assert( opt_test->is_If(), "Expect an IfNode");
1616 IfNode *opt_if1 = (IfNode*)opt_test;
1617 // Fast path taken; set region slot 3
1618 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1619 r->init_req(3,fast_taken); // Capture fast-control
1620
1621 // Fast path not-taken, i.e. slow path
1622 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1623
1624 // Set fast path result
1625 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1626 phi->init_req(3, fast_result);
1627
1628 // Complex path
1629 // Build the second if node (if y is int)
1630 // Node for (int)y
1631 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1632 // Node for (double)((int) y)
1633 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1634 // Check (double)((int) y) : y
1635 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1636 // Check if (y isn't int) then go to slow path
1637
1638 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1639 // Branch either way
1640 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1641 Node *slow_path = opt_iff(r,if2); // Set region path 2
1642
1643 // Calculate DPow(abs(x), y)*(1 & (int)y)
1644 // Node for constant 1
1645 Node *conone = intcon(1);
1646 // 1& (int)y
1647 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1648 // zero node
1649 Node *conzero = intcon(0);
1650 // Check (1&(int)y)==0?
1651 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1652 // Check if (1&(int)y)!=0?, if so the result is negative
1653 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1654 // abs(x)
1655 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1656 // abs(x)^y
1657 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1658 // -abs(x)^y
1659 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1660 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1661 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1662 // Set complex path fast result
1663 phi->init_req(2, signresult);
1664
1665 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1666 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1667 r->init_req(1,slow_path);
1668 phi->init_req(1,slow_result);
1669
1670 // Post merge
1671 set_control(_gvn.transform(r));
1672 record_for_igvn(r);
1673 result=_gvn.transform(phi);
1674 }
1675
1676 //-------------------
1677 //result=(result.isNaN())? uncommon_trap():result;
1678 // Check: If isNaN() by checking result!=result? then go to Strict Math
1679 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1680 // Build the boolean node
1681 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1682
1683 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1684 // End the current control-flow path
1685 push_pair(x);
1686 push_pair(y);
1687 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1688 // to handle. Recompile without intrinsifying Math.pow.
1689 uncommon_trap(Deoptimization::Reason_intrinsic,
1690 Deoptimization::Action_make_not_entrant);
1691 }
1692
1693 C->set_has_split_ifs(true); // Has chance for split-if optimization
1694
1695 push_pair(result);
1696
1697 return true;
1698 }
1699
1700 //------------------------------inline_trans-------------------------------------
1701 // Inline transcendental instructions, if possible. The Intel hardware gets
1702 // these right, no funny corner cases missed.
1703 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1704 _sp += arg_size(); // restore stack pointer
1705 Node* arg = pop_math_arg();
1706 Node* trans = NULL;
1707
1708 switch (id) {
1709 case vmIntrinsics::_dlog:
1710 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1711 break;
1712 case vmIntrinsics::_dlog10:
1713 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1714 break;
1715 default:
1716 assert(false, "bad intrinsic was passed in");
1717 return false;
1718 }
1719
1720 // Push result back on JVM stack
1721 push_pair(trans);
1722 return true;
1723 }
1724
1725 //------------------------------runtime_math-----------------------------
1726 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1727 Node* a = NULL;
1728 Node* b = NULL;
1729
1730 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1731 "must be (DD)D or (D)D type");
1732
1733 // Inputs
1734 _sp += arg_size(); // restore stack pointer
1735 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1736 b = pop_math_arg();
1737 }
1738 a = pop_math_arg();
1739
1740 const TypePtr* no_memory_effects = NULL;
1741 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1742 no_memory_effects,
1743 a, top(), b, b ? top() : NULL);
1744 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1745 #ifdef ASSERT
1746 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1747 assert(value_top == top(), "second value must be top");
1748 #endif
1749
1750 push_pair(value);
1751 return true;
1752 }
1753
1754 //------------------------------inline_math_native-----------------------------
1755 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1756 switch (id) {
1757 // These intrinsics are not properly supported on all hardware
1758 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1759 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1760 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1761 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1762 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1763 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1764
1765 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1766 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1767 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1768 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1769
1770 // These intrinsics are supported on all hardware
1771 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1772 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1773
1774 // These intrinsics don't work on X86. The ad implementation doesn't
1775 // handle NaN's properly. Instead of returning infinity, the ad
1776 // implementation returns a NaN on overflow. See bug: 6304089
1777 // Once the ad implementations are fixed, change the code below
1778 // to match the intrinsics above
1779
1780 case vmIntrinsics::_dexp: return
1781 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1782 case vmIntrinsics::_dpow: return
1783 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1784
1785 // These intrinsics are not yet correctly implemented
1786 case vmIntrinsics::_datan2:
1787 return false;
1788
1789 default:
1790 ShouldNotReachHere();
1791 return false;
1792 }
1793 }
1794
1795 static bool is_simple_name(Node* n) {
1796 return (n->req() == 1 // constant
1797 || (n->is_Type() && n->as_Type()->type()->singleton())
1798 || n->is_Proj() // parameter or return value
1799 || n->is_Phi() // local of some sort
1800 );
1801 }
1802
1803 //----------------------------inline_min_max-----------------------------------
1804 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1805 push(generate_min_max(id, argument(0), argument(1)));
1806
1807 return true;
1808 }
1809
1810 Node*
1811 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1812 // These are the candidate return value:
1813 Node* xvalue = x0;
1814 Node* yvalue = y0;
1815
1816 if (xvalue == yvalue) {
1817 return xvalue;
1818 }
1819
1820 bool want_max = (id == vmIntrinsics::_max);
1821
1822 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1823 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1824 if (txvalue == NULL || tyvalue == NULL) return top();
1825 // This is not really necessary, but it is consistent with a
1826 // hypothetical MaxINode::Value method:
1827 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1828
1829 // %%% This folding logic should (ideally) be in a different place.
1830 // Some should be inside IfNode, and there to be a more reliable
1831 // transformation of ?: style patterns into cmoves. We also want
1832 // more powerful optimizations around cmove and min/max.
1833
1834 // Try to find a dominating comparison of these guys.
1835 // It can simplify the index computation for Arrays.copyOf
1836 // and similar uses of System.arraycopy.
1837 // First, compute the normalized version of CmpI(x, y).
1838 int cmp_op = Op_CmpI;
1839 Node* xkey = xvalue;
1840 Node* ykey = yvalue;
1841 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1842 if (ideal_cmpxy->is_Cmp()) {
1843 // E.g., if we have CmpI(length - offset, count),
1844 // it might idealize to CmpI(length, count + offset)
1845 cmp_op = ideal_cmpxy->Opcode();
1846 xkey = ideal_cmpxy->in(1);
1847 ykey = ideal_cmpxy->in(2);
1848 }
1849
1850 // Start by locating any relevant comparisons.
1851 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1852 Node* cmpxy = NULL;
1853 Node* cmpyx = NULL;
1854 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1855 Node* cmp = start_from->fast_out(k);
1856 if (cmp->outcnt() > 0 && // must have prior uses
1857 cmp->in(0) == NULL && // must be context-independent
1858 cmp->Opcode() == cmp_op) { // right kind of compare
1859 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1860 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1861 }
1862 }
1863
1864 const int NCMPS = 2;
1865 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1866 int cmpn;
1867 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1868 if (cmps[cmpn] != NULL) break; // find a result
1869 }
1870 if (cmpn < NCMPS) {
1871 // Look for a dominating test that tells us the min and max.
1872 int depth = 0; // Limit search depth for speed
1873 Node* dom = control();
1874 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1875 if (++depth >= 100) break;
1876 Node* ifproj = dom;
1877 if (!ifproj->is_Proj()) continue;
1878 Node* iff = ifproj->in(0);
1879 if (!iff->is_If()) continue;
1880 Node* bol = iff->in(1);
1881 if (!bol->is_Bool()) continue;
1882 Node* cmp = bol->in(1);
1883 if (cmp == NULL) continue;
1884 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1885 if (cmps[cmpn] == cmp) break;
1886 if (cmpn == NCMPS) continue;
1887 BoolTest::mask btest = bol->as_Bool()->_test._test;
1888 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1889 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1890 // At this point, we know that 'x btest y' is true.
1891 switch (btest) {
1892 case BoolTest::eq:
1893 // They are proven equal, so we can collapse the min/max.
1894 // Either value is the answer. Choose the simpler.
1895 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1896 return yvalue;
1897 return xvalue;
1898 case BoolTest::lt: // x < y
1899 case BoolTest::le: // x <= y
1900 return (want_max ? yvalue : xvalue);
1901 case BoolTest::gt: // x > y
1902 case BoolTest::ge: // x >= y
1903 return (want_max ? xvalue : yvalue);
1904 }
1905 }
1906 }
1907
1908 // We failed to find a dominating test.
1909 // Let's pick a test that might GVN with prior tests.
1910 Node* best_bol = NULL;
1911 BoolTest::mask best_btest = BoolTest::illegal;
1912 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1913 Node* cmp = cmps[cmpn];
1914 if (cmp == NULL) continue;
1915 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1916 Node* bol = cmp->fast_out(j);
1917 if (!bol->is_Bool()) continue;
1918 BoolTest::mask btest = bol->as_Bool()->_test._test;
1919 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1920 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1921 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1922 best_bol = bol->as_Bool();
1923 best_btest = btest;
1924 }
1925 }
1926 }
1927
1928 Node* answer_if_true = NULL;
1929 Node* answer_if_false = NULL;
1930 switch (best_btest) {
1931 default:
1932 if (cmpxy == NULL)
1933 cmpxy = ideal_cmpxy;
1934 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1935 // and fall through:
1936 case BoolTest::lt: // x < y
1937 case BoolTest::le: // x <= y
1938 answer_if_true = (want_max ? yvalue : xvalue);
1939 answer_if_false = (want_max ? xvalue : yvalue);
1940 break;
1941 case BoolTest::gt: // x > y
1942 case BoolTest::ge: // x >= y
1943 answer_if_true = (want_max ? xvalue : yvalue);
1944 answer_if_false = (want_max ? yvalue : xvalue);
1945 break;
1946 }
1947
1948 jint hi, lo;
1949 if (want_max) {
1950 // We can sharpen the minimum.
1951 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1952 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1953 } else {
1954 // We can sharpen the maximum.
1955 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1956 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1957 }
1958
1959 // Use a flow-free graph structure, to avoid creating excess control edges
1960 // which could hinder other optimizations.
1961 // Since Math.min/max is often used with arraycopy, we want
1962 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1963 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1964 answer_if_false, answer_if_true,
1965 TypeInt::make(lo, hi, widen));
1966
1967 return _gvn.transform(cmov);
1968
1969 /*
1970 // This is not as desirable as it may seem, since Min and Max
1971 // nodes do not have a full set of optimizations.
1972 // And they would interfere, anyway, with 'if' optimizations
1973 // and with CMoveI canonical forms.
1974 switch (id) {
1975 case vmIntrinsics::_min:
1976 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1977 case vmIntrinsics::_max:
1978 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1979 default:
1980 ShouldNotReachHere();
1981 }
1982 */
1983 }
1984
1985 inline int
1986 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1987 const TypePtr* base_type = TypePtr::NULL_PTR;
1988 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1989 if (base_type == NULL) {
1990 // Unknown type.
1991 return Type::AnyPtr;
1992 } else if (base_type == TypePtr::NULL_PTR) {
1993 // Since this is a NULL+long form, we have to switch to a rawptr.
1994 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1995 offset = MakeConX(0);
1996 return Type::RawPtr;
1997 } else if (base_type->base() == Type::RawPtr) {
1998 return Type::RawPtr;
1999 } else if (base_type->isa_oopptr()) {
2000 // Base is never null => always a heap address.
2001 if (base_type->ptr() == TypePtr::NotNull) {
2002 return Type::OopPtr;
2003 }
2004 // Offset is small => always a heap address.
2005 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2006 if (offset_type != NULL &&
2007 base_type->offset() == 0 && // (should always be?)
2008 offset_type->_lo >= 0 &&
2009 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2010 return Type::OopPtr;
2011 }
2012 // Otherwise, it might either be oop+off or NULL+addr.
2013 return Type::AnyPtr;
2014 } else {
2015 // No information:
2016 return Type::AnyPtr;
2017 }
2018 }
2019
2020 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2021 int kind = classify_unsafe_addr(base, offset);
2022 if (kind == Type::RawPtr) {
2023 return basic_plus_adr(top(), base, offset);
2024 } else {
2025 return basic_plus_adr(base, offset);
2026 }
2027 }
2028
2029 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2030 // inline int Integer.numberOfLeadingZeros(int)
2031 // inline int Long.numberOfLeadingZeros(long)
2032 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2033 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2034 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2035 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2036 _sp += arg_size(); // restore stack pointer
2037 switch (id) {
2038 case vmIntrinsics::_numberOfLeadingZeros_i:
2039 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
2040 break;
2041 case vmIntrinsics::_numberOfLeadingZeros_l:
2042 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
2043 break;
2044 default:
2045 ShouldNotReachHere();
2046 }
2047 return true;
2048 }
2049
2050 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2051 // inline int Integer.numberOfTrailingZeros(int)
2052 // inline int Long.numberOfTrailingZeros(long)
2053 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2054 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2055 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2056 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2057 _sp += arg_size(); // restore stack pointer
2058 switch (id) {
2059 case vmIntrinsics::_numberOfTrailingZeros_i:
2060 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2061 break;
2062 case vmIntrinsics::_numberOfTrailingZeros_l:
2063 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2064 break;
2065 default:
2066 ShouldNotReachHere();
2067 }
2068 return true;
2069 }
2070
2071 //----------------------------inline_bitCount_int/long-----------------------
2072 // inline int Integer.bitCount(int)
2073 // inline int Long.bitCount(long)
2074 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2075 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2076 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2077 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2078 _sp += arg_size(); // restore stack pointer
2079 switch (id) {
2080 case vmIntrinsics::_bitCount_i:
2081 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2082 break;
2083 case vmIntrinsics::_bitCount_l:
2084 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2085 break;
2086 default:
2087 ShouldNotReachHere();
2088 }
2089 return true;
2090 }
2091
2092 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2093 // inline Integer.reverseBytes(int)
2094 // inline Long.reverseBytes(long)
2095 // inline Character.reverseBytes(char)
2096 // inline Short.reverseBytes(short)
2097 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2098 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2099 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2100 "not reverse Bytes");
2101 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2102 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2103 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2104 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2105 _sp += arg_size(); // restore stack pointer
2106 switch (id) {
2107 case vmIntrinsics::_reverseBytes_i:
2108 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2109 break;
2110 case vmIntrinsics::_reverseBytes_l:
2111 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2112 break;
2113 case vmIntrinsics::_reverseBytes_c:
2114 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2115 break;
2116 case vmIntrinsics::_reverseBytes_s:
2117 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2118 break;
2119 default:
2120 ;
2121 }
2122 return true;
2123 }
2124
2125 //----------------------------inline_unsafe_access----------------------------
2126
2127 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2128
2129 // Helper that guards and inserts a G1 pre-barrier.
2130 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) {
2131 assert(UseG1GC, "should not call this otherwise");
2132
2133 // We could be accessing the referent field of a reference object. If so, when G1
2134 // is enabled, we need to log the value in the referent field in an SATB buffer.
2135 // This routine performs some compile time filters and generates suitable
2136 // runtime filters that guard the pre-barrier code.
2137
2138 // Some compile time checks.
2139
2140 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2141 const TypeX* otype = offset->find_intptr_t_type();
2142 if (otype != NULL && otype->is_con() &&
2143 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2144 // Constant offset but not the reference_offset so just return
2145 return;
2146 }
2147
2148 // We only need to generate the runtime guards for instances.
2149 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2150 if (btype != NULL) {
2151 if (btype->isa_aryptr()) {
2152 // Array type so nothing to do
2153 return;
2154 }
2155
2156 const TypeInstPtr* itype = btype->isa_instptr();
2157 if (itype != NULL) {
2158 // Can the klass of base_oop be statically determined
2159 // to be _not_ a sub-class of Reference?
2160 ciKlass* klass = itype->klass();
2161 if (klass->is_subtype_of(env()->Reference_klass()) &&
2162 !env()->Reference_klass()->is_subtype_of(klass)) {
2163 return;
2164 }
2165 }
2166 }
2167
2168 // The compile time filters did not reject base_oop/offset so
2169 // we need to generate the following runtime filters
2170 //
2171 // if (offset == java_lang_ref_Reference::_reference_offset) {
2172 // if (base != null) {
2173 // if (instance_of(base, java.lang.ref.Reference)) {
2174 // pre_barrier(_, pre_val, ...);
2175 // }
2176 // }
2177 // }
2178
2179 float likely = PROB_LIKELY(0.999);
2180 float unlikely = PROB_UNLIKELY(0.999);
2181
2182 IdealKit ideal(this);
2183 #define __ ideal.
2184
2185 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2186
2187 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2188 __ if_then(base_oop, BoolTest::ne, null(), likely); {
2189
2190 // Update graphKit memory and control from IdealKit.
2191 sync_kit(ideal);
2192
2193 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2194 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2195
2196 // Update IdealKit memory and control from graphKit.
2197 __ sync_kit(this);
2198
2199 Node* one = __ ConI(1);
2200
2201 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2202
2203 // Update graphKit from IdeakKit.
2204 sync_kit(ideal);
2205
2206 // Use the pre-barrier to record the value in the referent field
2207 pre_barrier(false /* do_load */,
2208 __ ctrl(),
2209 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2210 pre_val /* pre_val */,
2211 T_OBJECT);
2212
2213 // Update IdealKit from graphKit.
2214 __ sync_kit(this);
2215
2216 } __ end_if(); // _ref_type != ref_none
2217 } __ end_if(); // base != NULL
2218 } __ end_if(); // offset == referent_offset
2219
2220 // Final sync IdealKit and GraphKit.
2221 final_sync(ideal);
2222 #undef __
2223 }
2224
2225
2226 // Interpret Unsafe.fieldOffset cookies correctly:
2227 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2228
2229 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2230 if (callee()->is_static()) return false; // caller must have the capability!
2231
2232 #ifndef PRODUCT
2233 {
2234 ResourceMark rm;
2235 // Check the signatures.
2236 ciSignature* sig = signature();
2237 #ifdef ASSERT
2238 if (!is_store) {
2239 // Object getObject(Object base, int/long offset), etc.
2240 BasicType rtype = sig->return_type()->basic_type();
2241 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2242 rtype = T_ADDRESS; // it is really a C void*
2243 assert(rtype == type, "getter must return the expected value");
2244 if (!is_native_ptr) {
2245 assert(sig->count() == 2, "oop getter has 2 arguments");
2246 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2247 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2248 } else {
2249 assert(sig->count() == 1, "native getter has 1 argument");
2250 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2251 }
2252 } else {
2253 // void putObject(Object base, int/long offset, Object x), etc.
2254 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2255 if (!is_native_ptr) {
2256 assert(sig->count() == 3, "oop putter has 3 arguments");
2257 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2258 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2259 } else {
2260 assert(sig->count() == 2, "native putter has 2 arguments");
2261 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2262 }
2263 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2264 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2265 vtype = T_ADDRESS; // it is really a C void*
2266 assert(vtype == type, "putter must accept the expected value");
2267 }
2268 #endif // ASSERT
2269 }
2270 #endif //PRODUCT
2271
2272 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2273
2274 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2275
2276 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2277 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2278
2279 debug_only(int saved_sp = _sp);
2280 _sp += nargs;
2281
2282 Node* val;
2283 debug_only(val = (Node*)(uintptr_t)-1);
2284
2285
2286 if (is_store) {
2287 // Get the value being stored. (Pop it first; it was pushed last.)
2288 switch (type) {
2289 case T_DOUBLE:
2290 case T_LONG:
2291 case T_ADDRESS:
2292 val = pop_pair();
2293 break;
2294 default:
2295 val = pop();
2296 }
2297 }
2298
2299 // Build address expression. See the code in inline_unsafe_prefetch.
2300 Node *adr;
2301 Node *heap_base_oop = top();
2302 Node* offset = top();
2303
2304 if (!is_native_ptr) {
2305 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2306 offset = pop_pair();
2307 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2308 Node* base = pop();
2309 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2310 // to be plain byte offsets, which are also the same as those accepted
2311 // by oopDesc::field_base.
2312 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2313 "fieldOffset must be byte-scaled");
2314 // 32-bit machines ignore the high half!
2315 offset = ConvL2X(offset);
2316 adr = make_unsafe_address(base, offset);
2317 heap_base_oop = base;
2318 } else {
2319 Node* ptr = pop_pair();
2320 // Adjust Java long to machine word:
2321 ptr = ConvL2X(ptr);
2322 adr = make_unsafe_address(NULL, ptr);
2323 }
2324
2325 // Pop receiver last: it was pushed first.
2326 Node *receiver = pop();
2327
2328 assert(saved_sp == _sp, "must have correct argument count");
2329
2330 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2331
2332 // First guess at the value type.
2333 const Type *value_type = Type::get_const_basic_type(type);
2334
2335 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2336 // there was not enough information to nail it down.
2337 Compile::AliasType* alias_type = C->alias_type(adr_type);
2338 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2339
2340 // We will need memory barriers unless we can determine a unique
2341 // alias category for this reference. (Note: If for some reason
2342 // the barriers get omitted and the unsafe reference begins to "pollute"
2343 // the alias analysis of the rest of the graph, either Compile::can_alias
2344 // or Compile::must_alias will throw a diagnostic assert.)
2345 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2346
2347 // If we are reading the value of the referent field of a Reference
2348 // object (either by using Unsafe directly or through reflection)
2349 // then, if G1 is enabled, we need to record the referent in an
2350 // SATB log buffer using the pre-barrier mechanism.
2351 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store &&
2352 offset != top() && heap_base_oop != top();
2353
2354 if (!is_store && type == T_OBJECT) {
2355 // Attempt to infer a sharper value type from the offset and base type.
2356 ciKlass* sharpened_klass = NULL;
2357
2358 // See if it is an instance field, with an object type.
2359 if (alias_type->field() != NULL) {
2360 assert(!is_native_ptr, "native pointer op cannot use a java address");
2361 if (alias_type->field()->type()->is_klass()) {
2362 sharpened_klass = alias_type->field()->type()->as_klass();
2363 }
2364 }
2365
2366 // See if it is a narrow oop array.
2367 if (adr_type->isa_aryptr()) {
2368 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2369 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2370 if (elem_type != NULL) {
2371 sharpened_klass = elem_type->klass();
2372 }
2373 }
2374 }
2375
2376 if (sharpened_klass != NULL) {
2377 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2378
2379 // Sharpen the value type.
2380 value_type = tjp;
2381
2382 #ifndef PRODUCT
2383 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2384 tty->print(" from base type: "); adr_type->dump();
2385 tty->print(" sharpened value: "); value_type->dump();
2386 }
2387 #endif
2388 }
2389 }
2390
2391 // Null check on self without removing any arguments. The argument
2392 // null check technically happens in the wrong place, which can lead to
2393 // invalid stack traces when the primitive is inlined into a method
2394 // which handles NullPointerExceptions.
2395 _sp += nargs;
2396 do_null_check(receiver, T_OBJECT);
2397 _sp -= nargs;
2398 if (stopped()) {
2399 return true;
2400 }
2401 // Heap pointers get a null-check from the interpreter,
2402 // as a courtesy. However, this is not guaranteed by Unsafe,
2403 // and it is not possible to fully distinguish unintended nulls
2404 // from intended ones in this API.
2405
2406 if (is_volatile) {
2407 // We need to emit leading and trailing CPU membars (see below) in
2408 // addition to memory membars when is_volatile. This is a little
2409 // too strong, but avoids the need to insert per-alias-type
2410 // volatile membars (for stores; compare Parse::do_put_xxx), which
2411 // we cannot do effectively here because we probably only have a
2412 // rough approximation of type.
2413 need_mem_bar = true;
2414 // For Stores, place a memory ordering barrier now.
2415 if (is_store)
2416 insert_mem_bar(Op_MemBarRelease);
2417 }
2418
2419 // Memory barrier to prevent normal and 'unsafe' accesses from
2420 // bypassing each other. Happens after null checks, so the
2421 // exception paths do not take memory state from the memory barrier,
2422 // so there's no problems making a strong assert about mixing users
2423 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2424 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2425 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2426
2427 if (!is_store) {
2428 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2429 // load value and push onto stack
2430 switch (type) {
2431 case T_BOOLEAN:
2432 case T_CHAR:
2433 case T_BYTE:
2434 case T_SHORT:
2435 case T_INT:
2436 case T_FLOAT:
2437 push(p);
2438 break;
2439 case T_OBJECT:
2440 if (need_read_barrier) {
2441 insert_g1_pre_barrier(heap_base_oop, offset, p);
2442 }
2443 push(p);
2444 break;
2445 case T_ADDRESS:
2446 // Cast to an int type.
2447 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2448 p = ConvX2L(p);
2449 push_pair(p);
2450 break;
2451 case T_DOUBLE:
2452 case T_LONG:
2453 push_pair( p );
2454 break;
2455 default: ShouldNotReachHere();
2456 }
2457 } else {
2458 // place effect of store into memory
2459 switch (type) {
2460 case T_DOUBLE:
2461 val = dstore_rounding(val);
2462 break;
2463 case T_ADDRESS:
2464 // Repackage the long as a pointer.
2465 val = ConvL2X(val);
2466 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2467 break;
2468 }
2469
2470 if (type != T_OBJECT ) {
2471 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2472 } else {
2473 // Possibly an oop being stored to Java heap or native memory
2474 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2475 // oop to Java heap.
2476 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2477 } else {
2478 // We can't tell at compile time if we are storing in the Java heap or outside
2479 // of it. So we need to emit code to conditionally do the proper type of
2480 // store.
2481
2482 IdealKit ideal(this);
2483 #define __ ideal.
2484 // QQQ who knows what probability is here??
2485 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2486 // Sync IdealKit and graphKit.
2487 sync_kit(ideal);
2488 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2489 // Update IdealKit memory.
2490 __ sync_kit(this);
2491 } __ else_(); {
2492 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2493 } __ end_if();
2494 // Final sync IdealKit and GraphKit.
2495 final_sync(ideal);
2496 #undef __
2497 }
2498 }
2499 }
2500
2501 if (is_volatile) {
2502 if (!is_store)
2503 insert_mem_bar(Op_MemBarAcquire);
2504 else
2505 insert_mem_bar(Op_MemBarVolatile);
2506 }
2507
2508 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2509
2510 return true;
2511 }
2512
2513 //----------------------------inline_unsafe_prefetch----------------------------
2514
2515 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2516 #ifndef PRODUCT
2517 {
2518 ResourceMark rm;
2519 // Check the signatures.
2520 ciSignature* sig = signature();
2521 #ifdef ASSERT
2522 // Object getObject(Object base, int/long offset), etc.
2523 BasicType rtype = sig->return_type()->basic_type();
2524 if (!is_native_ptr) {
2525 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2526 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2527 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2528 } else {
2529 assert(sig->count() == 1, "native prefetch has 1 argument");
2530 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2531 }
2532 #endif // ASSERT
2533 }
2534 #endif // !PRODUCT
2535
2536 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2537
2538 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2539 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2540
2541 debug_only(int saved_sp = _sp);
2542 _sp += nargs;
2543
2544 // Build address expression. See the code in inline_unsafe_access.
2545 Node *adr;
2546 if (!is_native_ptr) {
2547 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2548 Node* offset = pop_pair();
2549 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2550 Node* base = pop();
2551 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2552 // to be plain byte offsets, which are also the same as those accepted
2553 // by oopDesc::field_base.
2554 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2555 "fieldOffset must be byte-scaled");
2556 // 32-bit machines ignore the high half!
2557 offset = ConvL2X(offset);
2558 adr = make_unsafe_address(base, offset);
2559 } else {
2560 Node* ptr = pop_pair();
2561 // Adjust Java long to machine word:
2562 ptr = ConvL2X(ptr);
2563 adr = make_unsafe_address(NULL, ptr);
2564 }
2565
2566 if (is_static) {
2567 assert(saved_sp == _sp, "must have correct argument count");
2568 } else {
2569 // Pop receiver last: it was pushed first.
2570 Node *receiver = pop();
2571 assert(saved_sp == _sp, "must have correct argument count");
2572
2573 // Null check on self without removing any arguments. The argument
2574 // null check technically happens in the wrong place, which can lead to
2575 // invalid stack traces when the primitive is inlined into a method
2576 // which handles NullPointerExceptions.
2577 _sp += nargs;
2578 do_null_check(receiver, T_OBJECT);
2579 _sp -= nargs;
2580 if (stopped()) {
2581 return true;
2582 }
2583 }
2584
2585 // Generate the read or write prefetch
2586 Node *prefetch;
2587 if (is_store) {
2588 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2589 } else {
2590 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2591 }
2592 prefetch->init_req(0, control());
2593 set_i_o(_gvn.transform(prefetch));
2594
2595 return true;
2596 }
2597
2598 //----------------------------inline_unsafe_CAS----------------------------
2599
2600 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2601 // This basic scheme here is the same as inline_unsafe_access, but
2602 // differs in enough details that combining them would make the code
2603 // overly confusing. (This is a true fact! I originally combined
2604 // them, but even I was confused by it!) As much code/comments as
2605 // possible are retained from inline_unsafe_access though to make
2606 // the correspondences clearer. - dl
2607
2608 if (callee()->is_static()) return false; // caller must have the capability!
2609
2610 #ifndef PRODUCT
2611 {
2612 ResourceMark rm;
2613 // Check the signatures.
2614 ciSignature* sig = signature();
2615 #ifdef ASSERT
2616 BasicType rtype = sig->return_type()->basic_type();
2617 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2618 assert(sig->count() == 4, "CAS has 4 arguments");
2619 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2620 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2621 #endif // ASSERT
2622 }
2623 #endif //PRODUCT
2624
2625 // number of stack slots per value argument (1 or 2)
2626 int type_words = type2size[type];
2627
2628 // Cannot inline wide CAS on machines that don't support it natively
2629 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2630 return false;
2631
2632 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2633
2634 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2635 int nargs = 1 + 1 + 2 + type_words + type_words;
2636
2637 // pop arguments: newval, oldval, offset, base, and receiver
2638 debug_only(int saved_sp = _sp);
2639 _sp += nargs;
2640 Node* newval = (type_words == 1) ? pop() : pop_pair();
2641 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2642 Node *offset = pop_pair();
2643 Node *base = pop();
2644 Node *receiver = pop();
2645 assert(saved_sp == _sp, "must have correct argument count");
2646
2647 // Null check receiver.
2648 _sp += nargs;
2649 do_null_check(receiver, T_OBJECT);
2650 _sp -= nargs;
2651 if (stopped()) {
2652 return true;
2653 }
2654
2655 // Build field offset expression.
2656 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2657 // to be plain byte offsets, which are also the same as those accepted
2658 // by oopDesc::field_base.
2659 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2660 // 32-bit machines ignore the high half of long offsets
2661 offset = ConvL2X(offset);
2662 Node* adr = make_unsafe_address(base, offset);
2663 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2664
2665 // (Unlike inline_unsafe_access, there seems no point in trying
2666 // to refine types. Just use the coarse types here.
2667 const Type *value_type = Type::get_const_basic_type(type);
2668 Compile::AliasType* alias_type = C->alias_type(adr_type);
2669 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2670 int alias_idx = C->get_alias_index(adr_type);
2671
2672 // Memory-model-wise, a CAS acts like a little synchronized block,
2673 // so needs barriers on each side. These don't translate into
2674 // actual barriers on most machines, but we still need rest of
2675 // compiler to respect ordering.
2676
2677 insert_mem_bar(Op_MemBarRelease);
2678 insert_mem_bar(Op_MemBarCPUOrder);
2679
2680 // 4984716: MemBars must be inserted before this
2681 // memory node in order to avoid a false
2682 // dependency which will confuse the scheduler.
2683 Node *mem = memory(alias_idx);
2684
2685 // For now, we handle only those cases that actually exist: ints,
2686 // longs, and Object. Adding others should be straightforward.
2687 Node* cas;
2688 switch(type) {
2689 case T_INT:
2690 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2691 break;
2692 case T_LONG:
2693 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2694 break;
2695 case T_OBJECT:
2696 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2697 // could be delayed during Parse (for example, in adjust_map_after_if()).
2698 // Execute transformation here to avoid barrier generation in such case.
2699 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2700 newval = _gvn.makecon(TypePtr::NULL_PTR);
2701
2702 // Reference stores need a store barrier.
2703 // (They don't if CAS fails, but it isn't worth checking.)
2704 pre_barrier(true /* do_load*/,
2705 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2706 NULL /* pre_val*/,
2707 T_OBJECT);
2708 #ifdef _LP64
2709 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2710 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2711 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2712 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2713 newval_enc, oldval_enc));
2714 } else
2715 #endif
2716 {
2717 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2718 }
2719 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2720 break;
2721 default:
2722 ShouldNotReachHere();
2723 break;
2724 }
2725
2726 // SCMemProjNodes represent the memory state of CAS. Their main
2727 // role is to prevent CAS nodes from being optimized away when their
2728 // results aren't used.
2729 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2730 set_memory(proj, alias_idx);
2731
2732 // Add the trailing membar surrounding the access
2733 insert_mem_bar(Op_MemBarCPUOrder);
2734 insert_mem_bar(Op_MemBarAcquire);
2735
2736 push(cas);
2737 return true;
2738 }
2739
2740 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2741 // This is another variant of inline_unsafe_access, differing in
2742 // that it always issues store-store ("release") barrier and ensures
2743 // store-atomicity (which only matters for "long").
2744
2745 if (callee()->is_static()) return false; // caller must have the capability!
2746
2747 #ifndef PRODUCT
2748 {
2749 ResourceMark rm;
2750 // Check the signatures.
2751 ciSignature* sig = signature();
2752 #ifdef ASSERT
2753 BasicType rtype = sig->return_type()->basic_type();
2754 assert(rtype == T_VOID, "must return void");
2755 assert(sig->count() == 3, "has 3 arguments");
2756 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2757 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2758 #endif // ASSERT
2759 }
2760 #endif //PRODUCT
2761
2762 // number of stack slots per value argument (1 or 2)
2763 int type_words = type2size[type];
2764
2765 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2766
2767 // Argument words: "this" plus oop plus offset plus value;
2768 int nargs = 1 + 1 + 2 + type_words;
2769
2770 // pop arguments: val, offset, base, and receiver
2771 debug_only(int saved_sp = _sp);
2772 _sp += nargs;
2773 Node* val = (type_words == 1) ? pop() : pop_pair();
2774 Node *offset = pop_pair();
2775 Node *base = pop();
2776 Node *receiver = pop();
2777 assert(saved_sp == _sp, "must have correct argument count");
2778
2779 // Null check receiver.
2780 _sp += nargs;
2781 do_null_check(receiver, T_OBJECT);
2782 _sp -= nargs;
2783 if (stopped()) {
2784 return true;
2785 }
2786
2787 // Build field offset expression.
2788 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2789 // 32-bit machines ignore the high half of long offsets
2790 offset = ConvL2X(offset);
2791 Node* adr = make_unsafe_address(base, offset);
2792 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2793 const Type *value_type = Type::get_const_basic_type(type);
2794 Compile::AliasType* alias_type = C->alias_type(adr_type);
2795
2796 insert_mem_bar(Op_MemBarRelease);
2797 insert_mem_bar(Op_MemBarCPUOrder);
2798 // Ensure that the store is atomic for longs:
2799 bool require_atomic_access = true;
2800 Node* store;
2801 if (type == T_OBJECT) // reference stores need a store barrier.
2802 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2803 else {
2804 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2805 }
2806 insert_mem_bar(Op_MemBarCPUOrder);
2807 return true;
2808 }
2809
2810 bool LibraryCallKit::inline_unsafe_allocate() {
2811 if (callee()->is_static()) return false; // caller must have the capability!
2812 int nargs = 1 + 1;
2813 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2814 null_check_receiver(callee()); // check then ignore argument(0)
2815 _sp += nargs; // set original stack for use by uncommon_trap
2816 Node* cls = do_null_check(argument(1), T_OBJECT);
2817 _sp -= nargs;
2818 if (stopped()) return true;
2819
2820 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2821 _sp += nargs; // set original stack for use by uncommon_trap
2822 kls = do_null_check(kls, T_OBJECT);
2823 _sp -= nargs;
2824 if (stopped()) return true; // argument was like int.class
2825
2826 // Note: The argument might still be an illegal value like
2827 // Serializable.class or Object[].class. The runtime will handle it.
2828 // But we must make an explicit check for initialization.
2829 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset()));
2830 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler
2831 // can generate code to load it as unsigned byte.
2832 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2833 Node* bits = intcon(instanceKlass::fully_initialized);
2834 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2835 // The 'test' is non-zero if we need to take a slow path.
2836
2837 Node* obj = new_instance(kls, test);
2838 push(obj);
2839
2840 return true;
2841 }
2842
2843 //------------------------inline_native_time_funcs--------------
2844 // inline code for System.currentTimeMillis() and System.nanoTime()
2845 // these have the same type and signature
2846 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2847 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2848 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2849 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2850 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2851 const TypePtr* no_memory_effects = NULL;
2852 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2853 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2854 #ifdef ASSERT
2855 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2856 assert(value_top == top(), "second value must be top");
2857 #endif
2858 push_pair(value);
2859 return true;
2860 }
2861
2862 //------------------------inline_native_currentThread------------------
2863 bool LibraryCallKit::inline_native_currentThread() {
2864 Node* junk = NULL;
2865 push(generate_current_thread(junk));
2866 return true;
2867 }
2868
2869 //------------------------inline_native_isInterrupted------------------
2870 bool LibraryCallKit::inline_native_isInterrupted() {
2871 const int nargs = 1+1; // receiver + boolean
2872 assert(nargs == arg_size(), "sanity");
2873 // Add a fast path to t.isInterrupted(clear_int):
2874 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2875 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2876 // So, in the common case that the interrupt bit is false,
2877 // we avoid making a call into the VM. Even if the interrupt bit
2878 // is true, if the clear_int argument is false, we avoid the VM call.
2879 // However, if the receiver is not currentThread, we must call the VM,
2880 // because there must be some locking done around the operation.
2881
2882 // We only go to the fast case code if we pass two guards.
2883 // Paths which do not pass are accumulated in the slow_region.
2884 RegionNode* slow_region = new (C, 1) RegionNode(1);
2885 record_for_igvn(slow_region);
2886 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2887 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2888 enum { no_int_result_path = 1,
2889 no_clear_result_path = 2,
2890 slow_result_path = 3
2891 };
2892
2893 // (a) Receiving thread must be the current thread.
2894 Node* rec_thr = argument(0);
2895 Node* tls_ptr = NULL;
2896 Node* cur_thr = generate_current_thread(tls_ptr);
2897 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2898 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2899
2900 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2901 if (!known_current_thread)
2902 generate_slow_guard(bol_thr, slow_region);
2903
2904 // (b) Interrupt bit on TLS must be false.
2905 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2906 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2907 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2908 // Set the control input on the field _interrupted read to prevent it floating up.
2909 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2910 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2911 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2912
2913 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2914
2915 // First fast path: if (!TLS._interrupted) return false;
2916 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2917 result_rgn->init_req(no_int_result_path, false_bit);
2918 result_val->init_req(no_int_result_path, intcon(0));
2919
2920 // drop through to next case
2921 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2922
2923 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2924 Node* clr_arg = argument(1);
2925 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2926 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2927 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2928
2929 // Second fast path: ... else if (!clear_int) return true;
2930 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2931 result_rgn->init_req(no_clear_result_path, false_arg);
2932 result_val->init_req(no_clear_result_path, intcon(1));
2933
2934 // drop through to next case
2935 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2936
2937 // (d) Otherwise, go to the slow path.
2938 slow_region->add_req(control());
2939 set_control( _gvn.transform(slow_region) );
2940
2941 if (stopped()) {
2942 // There is no slow path.
2943 result_rgn->init_req(slow_result_path, top());
2944 result_val->init_req(slow_result_path, top());
2945 } else {
2946 // non-virtual because it is a private non-static
2947 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2948
2949 Node* slow_val = set_results_for_java_call(slow_call);
2950 // this->control() comes from set_results_for_java_call
2951
2952 // If we know that the result of the slow call will be true, tell the optimizer!
2953 if (known_current_thread) slow_val = intcon(1);
2954
2955 Node* fast_io = slow_call->in(TypeFunc::I_O);
2956 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2957 // These two phis are pre-filled with copies of of the fast IO and Memory
2958 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2959 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2960
2961 result_rgn->init_req(slow_result_path, control());
2962 io_phi ->init_req(slow_result_path, i_o());
2963 mem_phi ->init_req(slow_result_path, reset_memory());
2964 result_val->init_req(slow_result_path, slow_val);
2965
2966 set_all_memory( _gvn.transform(mem_phi) );
2967 set_i_o( _gvn.transform(io_phi) );
2968 }
2969
2970 push_result(result_rgn, result_val);
2971 C->set_has_split_ifs(true); // Has chance for split-if optimization
2972
2973 return true;
2974 }
2975
2976 //---------------------------load_mirror_from_klass----------------------------
2977 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2978 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2979 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
2980 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2981 }
2982
2983 //-----------------------load_klass_from_mirror_common-------------------------
2984 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2985 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2986 // and branch to the given path on the region.
2987 // If never_see_null, take an uncommon trap on null, so we can optimistically
2988 // compile for the non-null case.
2989 // If the region is NULL, force never_see_null = true.
2990 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2991 bool never_see_null,
2992 int nargs,
2993 RegionNode* region,
2994 int null_path,
2995 int offset) {
2996 if (region == NULL) never_see_null = true;
2997 Node* p = basic_plus_adr(mirror, offset);
2998 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2999 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3000 _sp += nargs; // any deopt will start just before call to enclosing method
3001 Node* null_ctl = top();
3002 kls = null_check_oop(kls, &null_ctl, never_see_null);
3003 if (region != NULL) {
3004 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3005 region->init_req(null_path, null_ctl);
3006 } else {
3007 assert(null_ctl == top(), "no loose ends");
3008 }
3009 _sp -= nargs;
3010 return kls;
3011 }
3012
3013 //--------------------(inline_native_Class_query helpers)---------------------
3014 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3015 // Fall through if (mods & mask) == bits, take the guard otherwise.
3016 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3017 // Branch around if the given klass has the given modifier bit set.
3018 // Like generate_guard, adds a new path onto the region.
3019 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3020 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3021 Node* mask = intcon(modifier_mask);
3022 Node* bits = intcon(modifier_bits);
3023 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
3024 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
3025 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
3026 return generate_fair_guard(bol, region);
3027 }
3028 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3029 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3030 }
3031
3032 //-------------------------inline_native_Class_query-------------------
3033 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3034 int nargs = 1+0; // just the Class mirror, in most cases
3035 const Type* return_type = TypeInt::BOOL;
3036 Node* prim_return_value = top(); // what happens if it's a primitive class?
3037 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3038 bool expect_prim = false; // most of these guys expect to work on refs
3039
3040 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3041
3042 switch (id) {
3043 case vmIntrinsics::_isInstance:
3044 nargs = 1+1; // the Class mirror, plus the object getting queried about
3045 // nothing is an instance of a primitive type
3046 prim_return_value = intcon(0);
3047 break;
3048 case vmIntrinsics::_getModifiers:
3049 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3050 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3051 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3052 break;
3053 case vmIntrinsics::_isInterface:
3054 prim_return_value = intcon(0);
3055 break;
3056 case vmIntrinsics::_isArray:
3057 prim_return_value = intcon(0);
3058 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3059 break;
3060 case vmIntrinsics::_isPrimitive:
3061 prim_return_value = intcon(1);
3062 expect_prim = true; // obviously
3063 break;
3064 case vmIntrinsics::_getSuperclass:
3065 prim_return_value = null();
3066 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3067 break;
3068 case vmIntrinsics::_getComponentType:
3069 prim_return_value = null();
3070 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3071 break;
3072 case vmIntrinsics::_getClassAccessFlags:
3073 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3074 return_type = TypeInt::INT; // not bool! 6297094
3075 break;
3076 default:
3077 ShouldNotReachHere();
3078 }
3079
3080 Node* mirror = argument(0);
3081 Node* obj = (nargs <= 1)? top(): argument(1);
3082
3083 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3084 if (mirror_con == NULL) return false; // cannot happen?
3085
3086 #ifndef PRODUCT
3087 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3088 ciType* k = mirror_con->java_mirror_type();
3089 if (k) {
3090 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3091 k->print_name();
3092 tty->cr();
3093 }
3094 }
3095 #endif
3096
3097 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3098 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3099 record_for_igvn(region);
3100 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
3101
3102 // The mirror will never be null of Reflection.getClassAccessFlags, however
3103 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3104 // if it is. See bug 4774291.
3105
3106 // For Reflection.getClassAccessFlags(), the null check occurs in
3107 // the wrong place; see inline_unsafe_access(), above, for a similar
3108 // situation.
3109 _sp += nargs; // set original stack for use by uncommon_trap
3110 mirror = do_null_check(mirror, T_OBJECT);
3111 _sp -= nargs;
3112 // If mirror or obj is dead, only null-path is taken.
3113 if (stopped()) return true;
3114
3115 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3116
3117 // Now load the mirror's klass metaobject, and null-check it.
3118 // Side-effects region with the control path if the klass is null.
3119 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3120 region, _prim_path);
3121 // If kls is null, we have a primitive mirror.
3122 phi->init_req(_prim_path, prim_return_value);
3123 if (stopped()) { push_result(region, phi); return true; }
3124
3125 Node* p; // handy temp
3126 Node* null_ctl;
3127
3128 // Now that we have the non-null klass, we can perform the real query.
3129 // For constant classes, the query will constant-fold in LoadNode::Value.
3130 Node* query_value = top();
3131 switch (id) {
3132 case vmIntrinsics::_isInstance:
3133 // nothing is an instance of a primitive type
3134 _sp += nargs; // gen_instanceof might do an uncommon trap
3135 query_value = gen_instanceof(obj, kls);
3136 _sp -= nargs;
3137 break;
3138
3139 case vmIntrinsics::_getModifiers:
3140 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3141 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3142 break;
3143
3144 case vmIntrinsics::_isInterface:
3145 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3146 if (generate_interface_guard(kls, region) != NULL)
3147 // A guard was added. If the guard is taken, it was an interface.
3148 phi->add_req(intcon(1));
3149 // If we fall through, it's a plain class.
3150 query_value = intcon(0);
3151 break;
3152
3153 case vmIntrinsics::_isArray:
3154 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3155 if (generate_array_guard(kls, region) != NULL)
3156 // A guard was added. If the guard is taken, it was an array.
3157 phi->add_req(intcon(1));
3158 // If we fall through, it's a plain class.
3159 query_value = intcon(0);
3160 break;
3161
3162 case vmIntrinsics::_isPrimitive:
3163 query_value = intcon(0); // "normal" path produces false
3164 break;
3165
3166 case vmIntrinsics::_getSuperclass:
3167 // The rules here are somewhat unfortunate, but we can still do better
3168 // with random logic than with a JNI call.
3169 // Interfaces store null or Object as _super, but must report null.
3170 // Arrays store an intermediate super as _super, but must report Object.
3171 // Other types can report the actual _super.
3172 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3173 if (generate_interface_guard(kls, region) != NULL)
3174 // A guard was added. If the guard is taken, it was an interface.
3175 phi->add_req(null());
3176 if (generate_array_guard(kls, region) != NULL)
3177 // A guard was added. If the guard is taken, it was an array.
3178 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3179 // If we fall through, it's a plain class. Get its _super.
3180 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3181 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3182 null_ctl = top();
3183 kls = null_check_oop(kls, &null_ctl);
3184 if (null_ctl != top()) {
3185 // If the guard is taken, Object.superClass is null (both klass and mirror).
3186 region->add_req(null_ctl);
3187 phi ->add_req(null());
3188 }
3189 if (!stopped()) {
3190 query_value = load_mirror_from_klass(kls);
3191 }
3192 break;
3193
3194 case vmIntrinsics::_getComponentType:
3195 if (generate_array_guard(kls, region) != NULL) {
3196 // Be sure to pin the oop load to the guard edge just created:
3197 Node* is_array_ctrl = region->in(region->req()-1);
3198 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3199 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3200 phi->add_req(cmo);
3201 }
3202 query_value = null(); // non-array case is null
3203 break;
3204
3205 case vmIntrinsics::_getClassAccessFlags:
3206 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3207 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3208 break;
3209
3210 default:
3211 ShouldNotReachHere();
3212 }
3213
3214 // Fall-through is the normal case of a query to a real class.
3215 phi->init_req(1, query_value);
3216 region->init_req(1, control());
3217
3218 push_result(region, phi);
3219 C->set_has_split_ifs(true); // Has chance for split-if optimization
3220
3221 return true;
3222 }
3223
3224 //--------------------------inline_native_subtype_check------------------------
3225 // This intrinsic takes the JNI calls out of the heart of
3226 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3227 bool LibraryCallKit::inline_native_subtype_check() {
3228 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3229
3230 // Pull both arguments off the stack.
3231 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3232 args[0] = argument(0);
3233 args[1] = argument(1);
3234 Node* klasses[2]; // corresponding Klasses: superk, subk
3235 klasses[0] = klasses[1] = top();
3236
3237 enum {
3238 // A full decision tree on {superc is prim, subc is prim}:
3239 _prim_0_path = 1, // {P,N} => false
3240 // {P,P} & superc!=subc => false
3241 _prim_same_path, // {P,P} & superc==subc => true
3242 _prim_1_path, // {N,P} => false
3243 _ref_subtype_path, // {N,N} & subtype check wins => true
3244 _both_ref_path, // {N,N} & subtype check loses => false
3245 PATH_LIMIT
3246 };
3247
3248 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3249 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3250 record_for_igvn(region);
3251
3252 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3253 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3254 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3255
3256 // First null-check both mirrors and load each mirror's klass metaobject.
3257 int which_arg;
3258 for (which_arg = 0; which_arg <= 1; which_arg++) {
3259 Node* arg = args[which_arg];
3260 _sp += nargs; // set original stack for use by uncommon_trap
3261 arg = do_null_check(arg, T_OBJECT);
3262 _sp -= nargs;
3263 if (stopped()) break;
3264 args[which_arg] = _gvn.transform(arg);
3265
3266 Node* p = basic_plus_adr(arg, class_klass_offset);
3267 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3268 klasses[which_arg] = _gvn.transform(kls);
3269 }
3270
3271 // Having loaded both klasses, test each for null.
3272 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3273 for (which_arg = 0; which_arg <= 1; which_arg++) {
3274 Node* kls = klasses[which_arg];
3275 Node* null_ctl = top();
3276 _sp += nargs; // set original stack for use by uncommon_trap
3277 kls = null_check_oop(kls, &null_ctl, never_see_null);
3278 _sp -= nargs;
3279 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3280 region->init_req(prim_path, null_ctl);
3281 if (stopped()) break;
3282 klasses[which_arg] = kls;
3283 }
3284
3285 if (!stopped()) {
3286 // now we have two reference types, in klasses[0..1]
3287 Node* subk = klasses[1]; // the argument to isAssignableFrom
3288 Node* superk = klasses[0]; // the receiver
3289 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3290 // now we have a successful reference subtype check
3291 region->set_req(_ref_subtype_path, control());
3292 }
3293
3294 // If both operands are primitive (both klasses null), then
3295 // we must return true when they are identical primitives.
3296 // It is convenient to test this after the first null klass check.
3297 set_control(region->in(_prim_0_path)); // go back to first null check
3298 if (!stopped()) {
3299 // Since superc is primitive, make a guard for the superc==subc case.
3300 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3301 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3302 generate_guard(bol_eq, region, PROB_FAIR);
3303 if (region->req() == PATH_LIMIT+1) {
3304 // A guard was added. If the added guard is taken, superc==subc.
3305 region->swap_edges(PATH_LIMIT, _prim_same_path);
3306 region->del_req(PATH_LIMIT);
3307 }
3308 region->set_req(_prim_0_path, control()); // Not equal after all.
3309 }
3310
3311 // these are the only paths that produce 'true':
3312 phi->set_req(_prim_same_path, intcon(1));
3313 phi->set_req(_ref_subtype_path, intcon(1));
3314
3315 // pull together the cases:
3316 assert(region->req() == PATH_LIMIT, "sane region");
3317 for (uint i = 1; i < region->req(); i++) {
3318 Node* ctl = region->in(i);
3319 if (ctl == NULL || ctl == top()) {
3320 region->set_req(i, top());
3321 phi ->set_req(i, top());
3322 } else if (phi->in(i) == NULL) {
3323 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3324 }
3325 }
3326
3327 set_control(_gvn.transform(region));
3328 push(_gvn.transform(phi));
3329
3330 return true;
3331 }
3332
3333 //---------------------generate_array_guard_common------------------------
3334 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3335 bool obj_array, bool not_array) {
3336 // If obj_array/non_array==false/false:
3337 // Branch around if the given klass is in fact an array (either obj or prim).
3338 // If obj_array/non_array==false/true:
3339 // Branch around if the given klass is not an array klass of any kind.
3340 // If obj_array/non_array==true/true:
3341 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3342 // If obj_array/non_array==true/false:
3343 // Branch around if the kls is an oop array (Object[] or subtype)
3344 //
3345 // Like generate_guard, adds a new path onto the region.
3346 jint layout_con = 0;
3347 Node* layout_val = get_layout_helper(kls, layout_con);
3348 if (layout_val == NULL) {
3349 bool query = (obj_array
3350 ? Klass::layout_helper_is_objArray(layout_con)
3351 : Klass::layout_helper_is_javaArray(layout_con));
3352 if (query == not_array) {
3353 return NULL; // never a branch
3354 } else { // always a branch
3355 Node* always_branch = control();
3356 if (region != NULL)
3357 region->add_req(always_branch);
3358 set_control(top());
3359 return always_branch;
3360 }
3361 }
3362 // Now test the correct condition.
3363 jint nval = (obj_array
3364 ? ((jint)Klass::_lh_array_tag_type_value
3365 << Klass::_lh_array_tag_shift)
3366 : Klass::_lh_neutral_value);
3367 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3368 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3369 // invert the test if we are looking for a non-array
3370 if (not_array) btest = BoolTest(btest).negate();
3371 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3372 return generate_fair_guard(bol, region);
3373 }
3374
3375
3376 //-----------------------inline_native_newArray--------------------------
3377 bool LibraryCallKit::inline_native_newArray() {
3378 int nargs = 2;
3379 Node* mirror = argument(0);
3380 Node* count_val = argument(1);
3381
3382 _sp += nargs; // set original stack for use by uncommon_trap
3383 mirror = do_null_check(mirror, T_OBJECT);
3384 _sp -= nargs;
3385 // If mirror or obj is dead, only null-path is taken.
3386 if (stopped()) return true;
3387
3388 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3389 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3390 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3391 TypeInstPtr::NOTNULL);
3392 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3393 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3394 TypePtr::BOTTOM);
3395
3396 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3397 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3398 nargs,
3399 result_reg, _slow_path);
3400 Node* normal_ctl = control();
3401 Node* no_array_ctl = result_reg->in(_slow_path);
3402
3403 // Generate code for the slow case. We make a call to newArray().
3404 set_control(no_array_ctl);
3405 if (!stopped()) {
3406 // Either the input type is void.class, or else the
3407 // array klass has not yet been cached. Either the
3408 // ensuing call will throw an exception, or else it
3409 // will cache the array klass for next time.
3410 PreserveJVMState pjvms(this);
3411 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3412 Node* slow_result = set_results_for_java_call(slow_call);
3413 // this->control() comes from set_results_for_java_call
3414 result_reg->set_req(_slow_path, control());
3415 result_val->set_req(_slow_path, slow_result);
3416 result_io ->set_req(_slow_path, i_o());
3417 result_mem->set_req(_slow_path, reset_memory());
3418 }
3419
3420 set_control(normal_ctl);
3421 if (!stopped()) {
3422 // Normal case: The array type has been cached in the java.lang.Class.
3423 // The following call works fine even if the array type is polymorphic.
3424 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3425 Node* obj = new_array(klass_node, count_val, nargs);
3426 result_reg->init_req(_normal_path, control());
3427 result_val->init_req(_normal_path, obj);
3428 result_io ->init_req(_normal_path, i_o());
3429 result_mem->init_req(_normal_path, reset_memory());
3430 }
3431
3432 // Return the combined state.
3433 set_i_o( _gvn.transform(result_io) );
3434 set_all_memory( _gvn.transform(result_mem) );
3435 push_result(result_reg, result_val);
3436 C->set_has_split_ifs(true); // Has chance for split-if optimization
3437
3438 return true;
3439 }
3440
3441 //----------------------inline_native_getLength--------------------------
3442 bool LibraryCallKit::inline_native_getLength() {
3443 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3444
3445 int nargs = 1;
3446 Node* array = argument(0);
3447
3448 _sp += nargs; // set original stack for use by uncommon_trap
3449 array = do_null_check(array, T_OBJECT);
3450 _sp -= nargs;
3451
3452 // If array is dead, only null-path is taken.
3453 if (stopped()) return true;
3454
3455 // Deoptimize if it is a non-array.
3456 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3457
3458 if (non_array != NULL) {
3459 PreserveJVMState pjvms(this);
3460 set_control(non_array);
3461 _sp += nargs; // push the arguments back on the stack
3462 uncommon_trap(Deoptimization::Reason_intrinsic,
3463 Deoptimization::Action_maybe_recompile);
3464 }
3465
3466 // If control is dead, only non-array-path is taken.
3467 if (stopped()) return true;
3468
3469 // The works fine even if the array type is polymorphic.
3470 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3471 push( load_array_length(array) );
3472
3473 C->set_has_split_ifs(true); // Has chance for split-if optimization
3474
3475 return true;
3476 }
3477
3478 //------------------------inline_array_copyOf----------------------------
3479 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3480 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3481
3482 // Restore the stack and pop off the arguments.
3483 int nargs = 3 + (is_copyOfRange? 1: 0);
3484 Node* original = argument(0);
3485 Node* start = is_copyOfRange? argument(1): intcon(0);
3486 Node* end = is_copyOfRange? argument(2): argument(1);
3487 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3488
3489 Node* newcopy;
3490
3491 //set the original stack and the reexecute bit for the interpreter to reexecute
3492 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3493 { PreserveReexecuteState preexecs(this);
3494 _sp += nargs;
3495 jvms()->set_should_reexecute(true);
3496
3497 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3498 original = do_null_check(original, T_OBJECT);
3499
3500 // Check if a null path was taken unconditionally.
3501 if (stopped()) return true;
3502
3503 Node* orig_length = load_array_length(original);
3504
3505 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3506 NULL, 0);
3507 klass_node = do_null_check(klass_node, T_OBJECT);
3508
3509 RegionNode* bailout = new (C, 1) RegionNode(1);
3510 record_for_igvn(bailout);
3511
3512 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3513 // Bail out if that is so.
3514 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3515 if (not_objArray != NULL) {
3516 // Improve the klass node's type from the new optimistic assumption:
3517 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3518 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3519 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3520 cast->init_req(0, control());
3521 klass_node = _gvn.transform(cast);
3522 }
3523
3524 // Bail out if either start or end is negative.
3525 generate_negative_guard(start, bailout, &start);
3526 generate_negative_guard(end, bailout, &end);
3527
3528 Node* length = end;
3529 if (_gvn.type(start) != TypeInt::ZERO) {
3530 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3531 }
3532
3533 // Bail out if length is negative.
3534 // ...Not needed, since the new_array will throw the right exception.
3535 //generate_negative_guard(length, bailout, &length);
3536
3537 if (bailout->req() > 1) {
3538 PreserveJVMState pjvms(this);
3539 set_control( _gvn.transform(bailout) );
3540 uncommon_trap(Deoptimization::Reason_intrinsic,
3541 Deoptimization::Action_maybe_recompile);
3542 }
3543
3544 if (!stopped()) {
3545
3546 // How many elements will we copy from the original?
3547 // The answer is MinI(orig_length - start, length).
3548 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3549 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3550
3551 newcopy = new_array(klass_node, length, 0);
3552
3553 // Generate a direct call to the right arraycopy function(s).
3554 // We know the copy is disjoint but we might not know if the
3555 // oop stores need checking.
3556 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3557 // This will fail a store-check if x contains any non-nulls.
3558 bool disjoint_bases = true;
3559 bool length_never_negative = true;
3560 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3561 original, start, newcopy, intcon(0), moved,
3562 disjoint_bases, length_never_negative);
3563 }
3564 } //original reexecute and sp are set back here
3565
3566 if(!stopped()) {
3567 push(newcopy);
3568 }
3569
3570 C->set_has_split_ifs(true); // Has chance for split-if optimization
3571
3572 return true;
3573 }
3574
3575
3576 //----------------------generate_virtual_guard---------------------------
3577 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3578 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3579 RegionNode* slow_region) {
3580 ciMethod* method = callee();
3581 int vtable_index = method->vtable_index();
3582 // Get the methodOop out of the appropriate vtable entry.
3583 int entry_offset = (instanceKlass::vtable_start_offset() +
3584 vtable_index*vtableEntry::size()) * wordSize +
3585 vtableEntry::method_offset_in_bytes();
3586 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3587 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3588
3589 // Compare the target method with the expected method (e.g., Object.hashCode).
3590 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3591
3592 Node* native_call = makecon(native_call_addr);
3593 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3594 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3595
3596 return generate_slow_guard(test_native, slow_region);
3597 }
3598
3599 //-----------------------generate_method_call----------------------------
3600 // Use generate_method_call to make a slow-call to the real
3601 // method if the fast path fails. An alternative would be to
3602 // use a stub like OptoRuntime::slow_arraycopy_Java.
3603 // This only works for expanding the current library call,
3604 // not another intrinsic. (E.g., don't use this for making an
3605 // arraycopy call inside of the copyOf intrinsic.)
3606 CallJavaNode*
3607 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3608 // When compiling the intrinsic method itself, do not use this technique.
3609 guarantee(callee() != C->method(), "cannot make slow-call to self");
3610
3611 ciMethod* method = callee();
3612 // ensure the JVMS we have will be correct for this call
3613 guarantee(method_id == method->intrinsic_id(), "must match");
3614
3615 const TypeFunc* tf = TypeFunc::make(method);
3616 int tfdc = tf->domain()->cnt();
3617 CallJavaNode* slow_call;
3618 if (is_static) {
3619 assert(!is_virtual, "");
3620 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3621 SharedRuntime::get_resolve_static_call_stub(),
3622 method, bci());
3623 } else if (is_virtual) {
3624 null_check_receiver(method);
3625 int vtable_index = methodOopDesc::invalid_vtable_index;
3626 if (UseInlineCaches) {
3627 // Suppress the vtable call
3628 } else {
3629 // hashCode and clone are not a miranda methods,
3630 // so the vtable index is fixed.
3631 // No need to use the linkResolver to get it.
3632 vtable_index = method->vtable_index();
3633 }
3634 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3635 SharedRuntime::get_resolve_virtual_call_stub(),
3636 method, vtable_index, bci());
3637 } else { // neither virtual nor static: opt_virtual
3638 null_check_receiver(method);
3639 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3640 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3641 method, bci());
3642 slow_call->set_optimized_virtual(true);
3643 }
3644 set_arguments_for_java_call(slow_call);
3645 set_edges_for_java_call(slow_call);
3646 return slow_call;
3647 }
3648
3649
3650 //------------------------------inline_native_hashcode--------------------
3651 // Build special case code for calls to hashCode on an object.
3652 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3653 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3654 assert(!(is_virtual && is_static), "either virtual, special, or static");
3655
3656 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3657
3658 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3659 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3660 TypeInt::INT);
3661 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3662 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3663 TypePtr::BOTTOM);
3664 Node* obj = NULL;
3665 if (!is_static) {
3666 // Check for hashing null object
3667 obj = null_check_receiver(callee());
3668 if (stopped()) return true; // unconditionally null
3669 result_reg->init_req(_null_path, top());
3670 result_val->init_req(_null_path, top());
3671 } else {
3672 // Do a null check, and return zero if null.
3673 // System.identityHashCode(null) == 0
3674 obj = argument(0);
3675 Node* null_ctl = top();
3676 obj = null_check_oop(obj, &null_ctl);
3677 result_reg->init_req(_null_path, null_ctl);
3678 result_val->init_req(_null_path, _gvn.intcon(0));
3679 }
3680
3681 // Unconditionally null? Then return right away.
3682 if (stopped()) {
3683 set_control( result_reg->in(_null_path) );
3684 if (!stopped())
3685 push( result_val ->in(_null_path) );
3686 return true;
3687 }
3688
3689 // After null check, get the object's klass.
3690 Node* obj_klass = load_object_klass(obj);
3691
3692 // This call may be virtual (invokevirtual) or bound (invokespecial).
3693 // For each case we generate slightly different code.
3694
3695 // We only go to the fast case code if we pass a number of guards. The
3696 // paths which do not pass are accumulated in the slow_region.
3697 RegionNode* slow_region = new (C, 1) RegionNode(1);
3698 record_for_igvn(slow_region);
3699
3700 // If this is a virtual call, we generate a funny guard. We pull out
3701 // the vtable entry corresponding to hashCode() from the target object.
3702 // If the target method which we are calling happens to be the native
3703 // Object hashCode() method, we pass the guard. We do not need this
3704 // guard for non-virtual calls -- the caller is known to be the native
3705 // Object hashCode().
3706 if (is_virtual) {
3707 generate_virtual_guard(obj_klass, slow_region);
3708 }
3709
3710 // Get the header out of the object, use LoadMarkNode when available
3711 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3712 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3713
3714 // Test the header to see if it is unlocked.
3715 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3716 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3717 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3718 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3719 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3720
3721 generate_slow_guard(test_unlocked, slow_region);
3722
3723 // Get the hash value and check to see that it has been properly assigned.
3724 // We depend on hash_mask being at most 32 bits and avoid the use of
3725 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3726 // vm: see markOop.hpp.
3727 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3728 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3729 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3730 // This hack lets the hash bits live anywhere in the mark object now, as long
3731 // as the shift drops the relevant bits into the low 32 bits. Note that
3732 // Java spec says that HashCode is an int so there's no point in capturing
3733 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3734 hshifted_header = ConvX2I(hshifted_header);
3735 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3736
3737 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3738 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3739 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3740
3741 generate_slow_guard(test_assigned, slow_region);
3742
3743 Node* init_mem = reset_memory();
3744 // fill in the rest of the null path:
3745 result_io ->init_req(_null_path, i_o());
3746 result_mem->init_req(_null_path, init_mem);
3747
3748 result_val->init_req(_fast_path, hash_val);
3749 result_reg->init_req(_fast_path, control());
3750 result_io ->init_req(_fast_path, i_o());
3751 result_mem->init_req(_fast_path, init_mem);
3752
3753 // Generate code for the slow case. We make a call to hashCode().
3754 set_control(_gvn.transform(slow_region));
3755 if (!stopped()) {
3756 // No need for PreserveJVMState, because we're using up the present state.
3757 set_all_memory(init_mem);
3758 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3759 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3760 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3761 Node* slow_result = set_results_for_java_call(slow_call);
3762 // this->control() comes from set_results_for_java_call
3763 result_reg->init_req(_slow_path, control());
3764 result_val->init_req(_slow_path, slow_result);
3765 result_io ->set_req(_slow_path, i_o());
3766 result_mem ->set_req(_slow_path, reset_memory());
3767 }
3768
3769 // Return the combined state.
3770 set_i_o( _gvn.transform(result_io) );
3771 set_all_memory( _gvn.transform(result_mem) );
3772 push_result(result_reg, result_val);
3773
3774 return true;
3775 }
3776
3777 //---------------------------inline_native_getClass----------------------------
3778 // Build special case code for calls to getClass on an object.
3779 bool LibraryCallKit::inline_native_getClass() {
3780 Node* obj = null_check_receiver(callee());
3781 if (stopped()) return true;
3782 push( load_mirror_from_klass(load_object_klass(obj)) );
3783 return true;
3784 }
3785
3786 //-----------------inline_native_Reflection_getCallerClass---------------------
3787 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3788 //
3789 // NOTE that this code must perform the same logic as
3790 // vframeStream::security_get_caller_frame in that it must skip
3791 // Method.invoke() and auxiliary frames.
3792
3793
3794
3795
3796 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3797 ciMethod* method = callee();
3798
3799 #ifndef PRODUCT
3800 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3801 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3802 }
3803 #endif
3804
3805 debug_only(int saved_sp = _sp);
3806
3807 // Argument words: (int depth)
3808 int nargs = 1;
3809
3810 _sp += nargs;
3811 Node* caller_depth_node = pop();
3812
3813 assert(saved_sp == _sp, "must have correct argument count");
3814
3815 // The depth value must be a constant in order for the runtime call
3816 // to be eliminated.
3817 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3818 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3819 #ifndef PRODUCT
3820 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3821 tty->print_cr(" Bailing out because caller depth was not a constant");
3822 }
3823 #endif
3824 return false;
3825 }
3826 // Note that the JVM state at this point does not include the
3827 // getCallerClass() frame which we are trying to inline. The
3828 // semantics of getCallerClass(), however, are that the "first"
3829 // frame is the getCallerClass() frame, so we subtract one from the
3830 // requested depth before continuing. We don't inline requests of
3831 // getCallerClass(0).
3832 int caller_depth = caller_depth_type->get_con() - 1;
3833 if (caller_depth < 0) {
3834 #ifndef PRODUCT
3835 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3836 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3837 }
3838 #endif
3839 return false;
3840 }
3841
3842 if (!jvms()->has_method()) {
3843 #ifndef PRODUCT
3844 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3845 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3846 }
3847 #endif
3848 return false;
3849 }
3850 int _depth = jvms()->depth(); // cache call chain depth
3851
3852 // Walk back up the JVM state to find the caller at the required
3853 // depth. NOTE that this code must perform the same logic as
3854 // vframeStream::security_get_caller_frame in that it must skip
3855 // Method.invoke() and auxiliary frames. Note also that depth is
3856 // 1-based (1 is the bottom of the inlining).
3857 int inlining_depth = _depth;
3858 JVMState* caller_jvms = NULL;
3859
3860 if (inlining_depth > 0) {
3861 caller_jvms = jvms();
3862 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3863 do {
3864 // The following if-tests should be performed in this order
3865 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3866 // Skip a Method.invoke() or auxiliary frame
3867 } else if (caller_depth > 0) {
3868 // Skip real frame
3869 --caller_depth;
3870 } else {
3871 // We're done: reached desired caller after skipping.
3872 break;
3873 }
3874 caller_jvms = caller_jvms->caller();
3875 --inlining_depth;
3876 } while (inlining_depth > 0);
3877 }
3878
3879 if (inlining_depth == 0) {
3880 #ifndef PRODUCT
3881 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3882 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3883 tty->print_cr(" JVM state at this point:");
3884 for (int i = _depth; i >= 1; i--) {
3885 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3886 }
3887 }
3888 #endif
3889 return false; // Reached end of inlining
3890 }
3891
3892 // Acquire method holder as java.lang.Class
3893 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3894 ciInstance* caller_mirror = caller_klass->java_mirror();
3895 // Push this as a constant
3896 push(makecon(TypeInstPtr::make(caller_mirror)));
3897 #ifndef PRODUCT
3898 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3899 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);
3900 tty->print_cr(" JVM state at this point:");
3901 for (int i = _depth; i >= 1; i--) {
3902 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3903 }
3904 }
3905 #endif
3906 return true;
3907 }
3908
3909 // Helper routine for above
3910 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3911 ciMethod* method = jvms->method();
3912
3913 // Is this the Method.invoke method itself?
3914 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3915 return true;
3916
3917 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3918 ciKlass* k = method->holder();
3919 if (k->is_instance_klass()) {
3920 ciInstanceKlass* ik = k->as_instance_klass();
3921 for (; ik != NULL; ik = ik->super()) {
3922 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3923 ik == env()->find_system_klass(ik->name())) {
3924 return true;
3925 }
3926 }
3927 }
3928 else if (method->is_method_handle_adapter()) {
3929 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3930 return true;
3931 }
3932
3933 return false;
3934 }
3935
3936 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3937 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3938 // computing it since there is no lookup field by name function in the
3939 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3940 // Using a static variable here is safe even if we have multiple compilation
3941 // threads because the offset is constant. At worst the same offset will be
3942 // computed and stored multiple
3943
3944 bool LibraryCallKit::inline_native_AtomicLong_get() {
3945 // Restore the stack and pop off the argument
3946 _sp+=1;
3947 Node *obj = pop();
3948
3949 // get the offset of the "value" field. Since the CI interfaces
3950 // does not provide a way to look up a field by name, we scan the bytecodes
3951 // to get the field index. We expect the first 2 instructions of the method
3952 // to be:
3953 // 0 aload_0
3954 // 1 getfield "value"
3955 ciMethod* method = callee();
3956 if (value_field_offset == -1)
3957 {
3958 ciField* value_field;
3959 ciBytecodeStream iter(method);
3960 Bytecodes::Code bc = iter.next();
3961
3962 if ((bc != Bytecodes::_aload_0) &&
3963 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3964 return false;
3965 bc = iter.next();
3966 if (bc != Bytecodes::_getfield)
3967 return false;
3968 bool ignore;
3969 value_field = iter.get_field(ignore);
3970 value_field_offset = value_field->offset_in_bytes();
3971 }
3972
3973 // Null check without removing any arguments.
3974 _sp++;
3975 obj = do_null_check(obj, T_OBJECT);
3976 _sp--;
3977 // Check for locking null object
3978 if (stopped()) return true;
3979
3980 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3981 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3982 int alias_idx = C->get_alias_index(adr_type);
3983
3984 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3985
3986 push_pair(result);
3987
3988 return true;
3989 }
3990
3991 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3992 // Restore the stack and pop off the arguments
3993 _sp+=5;
3994 Node *newVal = pop_pair();
3995 Node *oldVal = pop_pair();
3996 Node *obj = pop();
3997
3998 // we need the offset of the "value" field which was computed when
3999 // inlining the get() method. Give up if we don't have it.
4000 if (value_field_offset == -1)
4001 return false;
4002
4003 // Null check without removing any arguments.
4004 _sp+=5;
4005 obj = do_null_check(obj, T_OBJECT);
4006 _sp-=5;
4007 // Check for locking null object
4008 if (stopped()) return true;
4009
4010 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
4011 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
4012 int alias_idx = C->get_alias_index(adr_type);
4013
4014 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
4015 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
4016 set_memory(store_proj, alias_idx);
4017 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
4018
4019 Node *result;
4020 // CMove node is not used to be able fold a possible check code
4021 // after attemptUpdate() call. This code could be transformed
4022 // into CMove node by loop optimizations.
4023 {
4024 RegionNode *r = new (C, 3) RegionNode(3);
4025 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
4026
4027 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4028 Node *iftrue = opt_iff(r, iff);
4029 r->init_req(1, iftrue);
4030 result->init_req(1, intcon(1));
4031 result->init_req(2, intcon(0));
4032
4033 set_control(_gvn.transform(r));
4034 record_for_igvn(r);
4035
4036 C->set_has_split_ifs(true); // Has chance for split-if optimization
4037 }
4038
4039 push(_gvn.transform(result));
4040 return true;
4041 }
4042
4043 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4044 // restore the arguments
4045 _sp += arg_size();
4046
4047 switch (id) {
4048 case vmIntrinsics::_floatToRawIntBits:
4049 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
4050 break;
4051
4052 case vmIntrinsics::_intBitsToFloat:
4053 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
4054 break;
4055
4056 case vmIntrinsics::_doubleToRawLongBits:
4057 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
4058 break;
4059
4060 case vmIntrinsics::_longBitsToDouble:
4061 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
4062 break;
4063
4064 case vmIntrinsics::_doubleToLongBits: {
4065 Node* value = pop_pair();
4066
4067 // two paths (plus control) merge in a wood
4068 RegionNode *r = new (C, 3) RegionNode(3);
4069 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
4070
4071 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
4072 // Build the boolean node
4073 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4074
4075 // Branch either way.
4076 // NaN case is less traveled, which makes all the difference.
4077 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4078 Node *opt_isnan = _gvn.transform(ifisnan);
4079 assert( opt_isnan->is_If(), "Expect an IfNode");
4080 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4081 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4082
4083 set_control(iftrue);
4084
4085 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4086 Node *slow_result = longcon(nan_bits); // return NaN
4087 phi->init_req(1, _gvn.transform( slow_result ));
4088 r->init_req(1, iftrue);
4089
4090 // Else fall through
4091 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4092 set_control(iffalse);
4093
4094 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
4095 r->init_req(2, iffalse);
4096
4097 // Post merge
4098 set_control(_gvn.transform(r));
4099 record_for_igvn(r);
4100
4101 Node* result = _gvn.transform(phi);
4102 assert(result->bottom_type()->isa_long(), "must be");
4103 push_pair(result);
4104
4105 C->set_has_split_ifs(true); // Has chance for split-if optimization
4106
4107 break;
4108 }
4109
4110 case vmIntrinsics::_floatToIntBits: {
4111 Node* value = pop();
4112
4113 // two paths (plus control) merge in a wood
4114 RegionNode *r = new (C, 3) RegionNode(3);
4115 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
4116
4117 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
4118 // Build the boolean node
4119 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4120
4121 // Branch either way.
4122 // NaN case is less traveled, which makes all the difference.
4123 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4124 Node *opt_isnan = _gvn.transform(ifisnan);
4125 assert( opt_isnan->is_If(), "Expect an IfNode");
4126 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4127 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4128
4129 set_control(iftrue);
4130
4131 static const jint nan_bits = 0x7fc00000;
4132 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4133 phi->init_req(1, _gvn.transform( slow_result ));
4134 r->init_req(1, iftrue);
4135
4136 // Else fall through
4137 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4138 set_control(iffalse);
4139
4140 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
4141 r->init_req(2, iffalse);
4142
4143 // Post merge
4144 set_control(_gvn.transform(r));
4145 record_for_igvn(r);
4146
4147 Node* result = _gvn.transform(phi);
4148 assert(result->bottom_type()->isa_int(), "must be");
4149 push(result);
4150
4151 C->set_has_split_ifs(true); // Has chance for split-if optimization
4152
4153 break;
4154 }
4155
4156 default:
4157 ShouldNotReachHere();
4158 }
4159
4160 return true;
4161 }
4162
4163 #ifdef _LP64
4164 #define XTOP ,top() /*additional argument*/
4165 #else //_LP64
4166 #define XTOP /*no additional argument*/
4167 #endif //_LP64
4168
4169 //----------------------inline_unsafe_copyMemory-------------------------
4170 bool LibraryCallKit::inline_unsafe_copyMemory() {
4171 if (callee()->is_static()) return false; // caller must have the capability!
4172 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4173 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4174 null_check_receiver(callee()); // check then ignore argument(0)
4175 if (stopped()) return true;
4176
4177 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4178
4179 Node* src_ptr = argument(1);
4180 Node* src_off = ConvL2X(argument(2));
4181 assert(argument(3)->is_top(), "2nd half of long");
4182 Node* dst_ptr = argument(4);
4183 Node* dst_off = ConvL2X(argument(5));
4184 assert(argument(6)->is_top(), "2nd half of long");
4185 Node* size = ConvL2X(argument(7));
4186 assert(argument(8)->is_top(), "2nd half of long");
4187
4188 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4189 "fieldOffset must be byte-scaled");
4190
4191 Node* src = make_unsafe_address(src_ptr, src_off);
4192 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4193
4194 // Conservatively insert a memory barrier on all memory slices.
4195 // Do not let writes of the copy source or destination float below the copy.
4196 insert_mem_bar(Op_MemBarCPUOrder);
4197
4198 // Call it. Note that the length argument is not scaled.
4199 make_runtime_call(RC_LEAF|RC_NO_FP,
4200 OptoRuntime::fast_arraycopy_Type(),
4201 StubRoutines::unsafe_arraycopy(),
4202 "unsafe_arraycopy",
4203 TypeRawPtr::BOTTOM,
4204 src, dst, size XTOP);
4205
4206 // Do not let reads of the copy destination float above the copy.
4207 insert_mem_bar(Op_MemBarCPUOrder);
4208
4209 return true;
4210 }
4211
4212 //------------------------clone_coping-----------------------------------
4213 // Helper function for inline_native_clone.
4214 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4215 assert(obj_size != NULL, "");
4216 Node* raw_obj = alloc_obj->in(1);
4217 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4218
4219 AllocateNode* alloc = NULL;
4220 if (ReduceBulkZeroing) {
4221 // We will be completely responsible for initializing this object -
4222 // mark Initialize node as complete.
4223 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4224 // The object was just allocated - there should be no any stores!
4225 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4226 // Mark as complete_with_arraycopy so that on AllocateNode
4227 // expansion, we know this AllocateNode is initialized by an array
4228 // copy and a StoreStore barrier exists after the array copy.
4229 alloc->initialization()->set_complete_with_arraycopy();
4230 }
4231
4232 // Copy the fastest available way.
4233 // TODO: generate fields copies for small objects instead.
4234 Node* src = obj;
4235 Node* dest = alloc_obj;
4236 Node* size = _gvn.transform(obj_size);
4237
4238 // Exclude the header but include array length to copy by 8 bytes words.
4239 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4240 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4241 instanceOopDesc::base_offset_in_bytes();
4242 // base_off:
4243 // 8 - 32-bit VM
4244 // 12 - 64-bit VM, compressed oops
4245 // 16 - 64-bit VM, normal oops
4246 if (base_off % BytesPerLong != 0) {
4247 assert(UseCompressedOops, "");
4248 if (is_array) {
4249 // Exclude length to copy by 8 bytes words.
4250 base_off += sizeof(int);
4251 } else {
4252 // Include klass to copy by 8 bytes words.
4253 base_off = instanceOopDesc::klass_offset_in_bytes();
4254 }
4255 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4256 }
4257 src = basic_plus_adr(src, base_off);
4258 dest = basic_plus_adr(dest, base_off);
4259
4260 // Compute the length also, if needed:
4261 Node* countx = size;
4262 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4263 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4264
4265 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4266 bool disjoint_bases = true;
4267 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4268 src, NULL, dest, NULL, countx,
4269 /*dest_uninitialized*/true);
4270
4271 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4272 if (card_mark) {
4273 assert(!is_array, "");
4274 // Put in store barrier for any and all oops we are sticking
4275 // into this object. (We could avoid this if we could prove
4276 // that the object type contains no oop fields at all.)
4277 Node* no_particular_value = NULL;
4278 Node* no_particular_field = NULL;
4279 int raw_adr_idx = Compile::AliasIdxRaw;
4280 post_barrier(control(),
4281 memory(raw_adr_type),
4282 alloc_obj,
4283 no_particular_field,
4284 raw_adr_idx,
4285 no_particular_value,
4286 T_OBJECT,
4287 false);
4288 }
4289
4290 // Do not let reads from the cloned object float above the arraycopy.
4291 if (alloc != NULL) {
4292 // Do not let stores that initialize this object be reordered with
4293 // a subsequent store that would make this object accessible by
4294 // other threads.
4295 // Record what AllocateNode this StoreStore protects so that
4296 // escape analysis can go from the MemBarStoreStoreNode to the
4297 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4298 // based on the escape status of the AllocateNode.
4299 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4300 } else {
4301 insert_mem_bar(Op_MemBarCPUOrder);
4302 }
4303 }
4304
4305 //------------------------inline_native_clone----------------------------
4306 // Here are the simple edge cases:
4307 // null receiver => normal trap
4308 // virtual and clone was overridden => slow path to out-of-line clone
4309 // not cloneable or finalizer => slow path to out-of-line Object.clone
4310 //
4311 // The general case has two steps, allocation and copying.
4312 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4313 //
4314 // Copying also has two cases, oop arrays and everything else.
4315 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4316 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4317 //
4318 // These steps fold up nicely if and when the cloned object's klass
4319 // can be sharply typed as an object array, a type array, or an instance.
4320 //
4321 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4322 int nargs = 1;
4323 PhiNode* result_val;
4324
4325 //set the original stack and the reexecute bit for the interpreter to reexecute
4326 //the bytecode that invokes Object.clone if deoptimization happens
4327 { PreserveReexecuteState preexecs(this);
4328 jvms()->set_should_reexecute(true);
4329
4330 //null_check_receiver will adjust _sp (push and pop)
4331 Node* obj = null_check_receiver(callee());
4332 if (stopped()) return true;
4333
4334 _sp += nargs;
4335
4336 Node* obj_klass = load_object_klass(obj);
4337 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4338 const TypeOopPtr* toop = ((tklass != NULL)
4339 ? tklass->as_instance_type()
4340 : TypeInstPtr::NOTNULL);
4341
4342 // Conservatively insert a memory barrier on all memory slices.
4343 // Do not let writes into the original float below the clone.
4344 insert_mem_bar(Op_MemBarCPUOrder);
4345
4346 // paths into result_reg:
4347 enum {
4348 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4349 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4350 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4351 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4352 PATH_LIMIT
4353 };
4354 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4355 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4356 TypeInstPtr::NOTNULL);
4357 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4358 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4359 TypePtr::BOTTOM);
4360 record_for_igvn(result_reg);
4361
4362 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4363 int raw_adr_idx = Compile::AliasIdxRaw;
4364
4365 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4366 if (array_ctl != NULL) {
4367 // It's an array.
4368 PreserveJVMState pjvms(this);
4369 set_control(array_ctl);
4370 Node* obj_length = load_array_length(obj);
4371 Node* obj_size = NULL;
4372 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);
4373
4374 if (!use_ReduceInitialCardMarks()) {
4375 // If it is an oop array, it requires very special treatment,
4376 // because card marking is required on each card of the array.
4377 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4378 if (is_obja != NULL) {
4379 PreserveJVMState pjvms2(this);
4380 set_control(is_obja);
4381 // Generate a direct call to the right arraycopy function(s).
4382 bool disjoint_bases = true;
4383 bool length_never_negative = true;
4384 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4385 obj, intcon(0), alloc_obj, intcon(0),
4386 obj_length,
4387 disjoint_bases, length_never_negative);
4388 result_reg->init_req(_objArray_path, control());
4389 result_val->init_req(_objArray_path, alloc_obj);
4390 result_i_o ->set_req(_objArray_path, i_o());
4391 result_mem ->set_req(_objArray_path, reset_memory());
4392 }
4393 }
4394 // Otherwise, there are no card marks to worry about.
4395 // (We can dispense with card marks if we know the allocation
4396 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4397 // causes the non-eden paths to take compensating steps to
4398 // simulate a fresh allocation, so that no further
4399 // card marks are required in compiled code to initialize
4400 // the object.)
4401
4402 if (!stopped()) {
4403 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4404
4405 // Present the results of the copy.
4406 result_reg->init_req(_array_path, control());
4407 result_val->init_req(_array_path, alloc_obj);
4408 result_i_o ->set_req(_array_path, i_o());
4409 result_mem ->set_req(_array_path, reset_memory());
4410 }
4411 }
4412
4413 // We only go to the instance fast case code if we pass a number of guards.
4414 // The paths which do not pass are accumulated in the slow_region.
4415 RegionNode* slow_region = new (C, 1) RegionNode(1);
4416 record_for_igvn(slow_region);
4417 if (!stopped()) {
4418 // It's an instance (we did array above). Make the slow-path tests.
4419 // If this is a virtual call, we generate a funny guard. We grab
4420 // the vtable entry corresponding to clone() from the target object.
4421 // If the target method which we are calling happens to be the
4422 // Object clone() method, we pass the guard. We do not need this
4423 // guard for non-virtual calls; the caller is known to be the native
4424 // Object clone().
4425 if (is_virtual) {
4426 generate_virtual_guard(obj_klass, slow_region);
4427 }
4428
4429 // The object must be cloneable and must not have a finalizer.
4430 // Both of these conditions may be checked in a single test.
4431 // We could optimize the cloneable test further, but we don't care.
4432 generate_access_flags_guard(obj_klass,
4433 // Test both conditions:
4434 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4435 // Must be cloneable but not finalizer:
4436 JVM_ACC_IS_CLONEABLE,
4437 slow_region);
4438 }
4439
4440 if (!stopped()) {
4441 // It's an instance, and it passed the slow-path tests.
4442 PreserveJVMState pjvms(this);
4443 Node* obj_size = NULL;
4444 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4445
4446 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4447
4448 // Present the results of the slow call.
4449 result_reg->init_req(_instance_path, control());
4450 result_val->init_req(_instance_path, alloc_obj);
4451 result_i_o ->set_req(_instance_path, i_o());
4452 result_mem ->set_req(_instance_path, reset_memory());
4453 }
4454
4455 // Generate code for the slow case. We make a call to clone().
4456 set_control(_gvn.transform(slow_region));
4457 if (!stopped()) {
4458 PreserveJVMState pjvms(this);
4459 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4460 Node* slow_result = set_results_for_java_call(slow_call);
4461 // this->control() comes from set_results_for_java_call
4462 result_reg->init_req(_slow_path, control());
4463 result_val->init_req(_slow_path, slow_result);
4464 result_i_o ->set_req(_slow_path, i_o());
4465 result_mem ->set_req(_slow_path, reset_memory());
4466 }
4467
4468 // Return the combined state.
4469 set_control( _gvn.transform(result_reg) );
4470 set_i_o( _gvn.transform(result_i_o) );
4471 set_all_memory( _gvn.transform(result_mem) );
4472 } //original reexecute and sp are set back here
4473
4474 push(_gvn.transform(result_val));
4475
4476 return true;
4477 }
4478
4479 //------------------------------basictype2arraycopy----------------------------
4480 address LibraryCallKit::basictype2arraycopy(BasicType t,
4481 Node* src_offset,
4482 Node* dest_offset,
4483 bool disjoint_bases,
4484 const char* &name,
4485 bool dest_uninitialized) {
4486 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4487 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4488
4489 bool aligned = false;
4490 bool disjoint = disjoint_bases;
4491
4492 // if the offsets are the same, we can treat the memory regions as
4493 // disjoint, because either the memory regions are in different arrays,
4494 // or they are identical (which we can treat as disjoint.) We can also
4495 // treat a copy with a destination index less that the source index
4496 // as disjoint since a low->high copy will work correctly in this case.
4497 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4498 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4499 // both indices are constants
4500 int s_offs = src_offset_inttype->get_con();
4501 int d_offs = dest_offset_inttype->get_con();
4502 int element_size = type2aelembytes(t);
4503 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4504 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4505 if (s_offs >= d_offs) disjoint = true;
4506 } else if (src_offset == dest_offset && src_offset != NULL) {
4507 // This can occur if the offsets are identical non-constants.
4508 disjoint = true;
4509 }
4510
4511 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4512 }
4513
4514
4515 //------------------------------inline_arraycopy-----------------------
4516 bool LibraryCallKit::inline_arraycopy() {
4517 // Restore the stack and pop off the arguments.
4518 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4519 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4520
4521 Node *src = argument(0);
4522 Node *src_offset = argument(1);
4523 Node *dest = argument(2);
4524 Node *dest_offset = argument(3);
4525 Node *length = argument(4);
4526
4527 // Compile time checks. If any of these checks cannot be verified at compile time,
4528 // we do not make a fast path for this call. Instead, we let the call remain as it
4529 // is. The checks we choose to mandate at compile time are:
4530 //
4531 // (1) src and dest are arrays.
4532 const Type* src_type = src->Value(&_gvn);
4533 const Type* dest_type = dest->Value(&_gvn);
4534 const TypeAryPtr* top_src = src_type->isa_aryptr();
4535 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4536 if (top_src == NULL || top_src->klass() == NULL ||
4537 top_dest == NULL || top_dest->klass() == NULL) {
4538 // Conservatively insert a memory barrier on all memory slices.
4539 // Do not let writes into the source float below the arraycopy.
4540 insert_mem_bar(Op_MemBarCPUOrder);
4541
4542 // Call StubRoutines::generic_arraycopy stub.
4543 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4544 src, src_offset, dest, dest_offset, length);
4545
4546 // Do not let reads from the destination float above the arraycopy.
4547 // Since we cannot type the arrays, we don't know which slices
4548 // might be affected. We could restrict this barrier only to those
4549 // memory slices which pertain to array elements--but don't bother.
4550 if (!InsertMemBarAfterArraycopy)
4551 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4552 insert_mem_bar(Op_MemBarCPUOrder);
4553 return true;
4554 }
4555
4556 // (2) src and dest arrays must have elements of the same BasicType
4557 // Figure out the size and type of the elements we will be copying.
4558 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4559 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4560 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4561 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4562
4563 if (src_elem != dest_elem || dest_elem == T_VOID) {
4564 // The component types are not the same or are not recognized. Punt.
4565 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4566 generate_slow_arraycopy(TypePtr::BOTTOM,
4567 src, src_offset, dest, dest_offset, length,
4568 /*dest_uninitialized*/false);
4569 return true;
4570 }
4571
4572 //---------------------------------------------------------------------------
4573 // We will make a fast path for this call to arraycopy.
4574
4575 // We have the following tests left to perform:
4576 //
4577 // (3) src and dest must not be null.
4578 // (4) src_offset must not be negative.
4579 // (5) dest_offset must not be negative.
4580 // (6) length must not be negative.
4581 // (7) src_offset + length must not exceed length of src.
4582 // (8) dest_offset + length must not exceed length of dest.
4583 // (9) each element of an oop array must be assignable
4584
4585 RegionNode* slow_region = new (C, 1) RegionNode(1);
4586 record_for_igvn(slow_region);
4587
4588 // (3) operands must not be null
4589 // We currently perform our null checks with the do_null_check routine.
4590 // This means that the null exceptions will be reported in the caller
4591 // rather than (correctly) reported inside of the native arraycopy call.
4592 // This should be corrected, given time. We do our null check with the
4593 // stack pointer restored.
4594 _sp += nargs;
4595 src = do_null_check(src, T_ARRAY);
4596 dest = do_null_check(dest, T_ARRAY);
4597 _sp -= nargs;
4598
4599 // (4) src_offset must not be negative.
4600 generate_negative_guard(src_offset, slow_region);
4601
4602 // (5) dest_offset must not be negative.
4603 generate_negative_guard(dest_offset, slow_region);
4604
4605 // (6) length must not be negative (moved to generate_arraycopy()).
4606 // generate_negative_guard(length, slow_region);
4607
4608 // (7) src_offset + length must not exceed length of src.
4609 generate_limit_guard(src_offset, length,
4610 load_array_length(src),
4611 slow_region);
4612
4613 // (8) dest_offset + length must not exceed length of dest.
4614 generate_limit_guard(dest_offset, length,
4615 load_array_length(dest),
4616 slow_region);
4617
4618 // (9) each element of an oop array must be assignable
4619 // The generate_arraycopy subroutine checks this.
4620
4621 // This is where the memory effects are placed:
4622 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4623 generate_arraycopy(adr_type, dest_elem,
4624 src, src_offset, dest, dest_offset, length,
4625 false, false, slow_region);
4626
4627 return true;
4628 }
4629
4630 //-----------------------------generate_arraycopy----------------------
4631 // Generate an optimized call to arraycopy.
4632 // Caller must guard against non-arrays.
4633 // Caller must determine a common array basic-type for both arrays.
4634 // Caller must validate offsets against array bounds.
4635 // The slow_region has already collected guard failure paths
4636 // (such as out of bounds length or non-conformable array types).
4637 // The generated code has this shape, in general:
4638 //
4639 // if (length == 0) return // via zero_path
4640 // slowval = -1
4641 // if (types unknown) {
4642 // slowval = call generic copy loop
4643 // if (slowval == 0) return // via checked_path
4644 // } else if (indexes in bounds) {
4645 // if ((is object array) && !(array type check)) {
4646 // slowval = call checked copy loop
4647 // if (slowval == 0) return // via checked_path
4648 // } else {
4649 // call bulk copy loop
4650 // return // via fast_path
4651 // }
4652 // }
4653 // // adjust params for remaining work:
4654 // if (slowval != -1) {
4655 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4656 // }
4657 // slow_region:
4658 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4659 // return // via slow_call_path
4660 //
4661 // This routine is used from several intrinsics: System.arraycopy,
4662 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4663 //
4664 void
4665 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4666 BasicType basic_elem_type,
4667 Node* src, Node* src_offset,
4668 Node* dest, Node* dest_offset,
4669 Node* copy_length,
4670 bool disjoint_bases,
4671 bool length_never_negative,
4672 RegionNode* slow_region) {
4673
4674 if (slow_region == NULL) {
4675 slow_region = new(C,1) RegionNode(1);
4676 record_for_igvn(slow_region);
4677 }
4678
4679 Node* original_dest = dest;
4680 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4681 bool dest_uninitialized = false;
4682
4683 // See if this is the initialization of a newly-allocated array.
4684 // If so, we will take responsibility here for initializing it to zero.
4685 // (Note: Because tightly_coupled_allocation performs checks on the
4686 // out-edges of the dest, we need to avoid making derived pointers
4687 // from it until we have checked its uses.)
4688 if (ReduceBulkZeroing
4689 && !ZeroTLAB // pointless if already zeroed
4690 && basic_elem_type != T_CONFLICT // avoid corner case
4691 && !src->eqv_uncast(dest)
4692 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4693 != NULL)
4694 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4695 && alloc->maybe_set_complete(&_gvn)) {
4696 // "You break it, you buy it."
4697 InitializeNode* init = alloc->initialization();
4698 assert(init->is_complete(), "we just did this");
4699 init->set_complete_with_arraycopy();
4700 assert(dest->is_CheckCastPP(), "sanity");
4701 assert(dest->in(0)->in(0) == init, "dest pinned");
4702 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4703 // From this point on, every exit path is responsible for
4704 // initializing any non-copied parts of the object to zero.
4705 // Also, if this flag is set we make sure that arraycopy interacts properly
4706 // with G1, eliding pre-barriers. See CR 6627983.
4707 dest_uninitialized = true;
4708 } else {
4709 // No zeroing elimination here.
4710 alloc = NULL;
4711 //original_dest = dest;
4712 //dest_uninitialized = false;
4713 }
4714
4715 // Results are placed here:
4716 enum { fast_path = 1, // normal void-returning assembly stub
4717 checked_path = 2, // special assembly stub with cleanup
4718 slow_call_path = 3, // something went wrong; call the VM
4719 zero_path = 4, // bypass when length of copy is zero
4720 bcopy_path = 5, // copy primitive array by 64-bit blocks
4721 PATH_LIMIT = 6
4722 };
4723 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4724 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4725 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4726 record_for_igvn(result_region);
4727 _gvn.set_type_bottom(result_i_o);
4728 _gvn.set_type_bottom(result_memory);
4729 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4730
4731 // The slow_control path:
4732 Node* slow_control;
4733 Node* slow_i_o = i_o();
4734 Node* slow_mem = memory(adr_type);
4735 debug_only(slow_control = (Node*) badAddress);
4736
4737 // Checked control path:
4738 Node* checked_control = top();
4739 Node* checked_mem = NULL;
4740 Node* checked_i_o = NULL;
4741 Node* checked_value = NULL;
4742
4743 if (basic_elem_type == T_CONFLICT) {
4744 assert(!dest_uninitialized, "");
4745 Node* cv = generate_generic_arraycopy(adr_type,
4746 src, src_offset, dest, dest_offset,
4747 copy_length, dest_uninitialized);
4748 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4749 checked_control = control();
4750 checked_i_o = i_o();
4751 checked_mem = memory(adr_type);
4752 checked_value = cv;
4753 set_control(top()); // no fast path
4754 }
4755
4756 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4757 if (not_pos != NULL) {
4758 PreserveJVMState pjvms(this);
4759 set_control(not_pos);
4760
4761 // (6) length must not be negative.
4762 if (!length_never_negative) {
4763 generate_negative_guard(copy_length, slow_region);
4764 }
4765
4766 // copy_length is 0.
4767 if (!stopped() && dest_uninitialized) {
4768 Node* dest_length = alloc->in(AllocateNode::ALength);
4769 if (copy_length->eqv_uncast(dest_length)
4770 || _gvn.find_int_con(dest_length, 1) <= 0) {
4771 // There is no zeroing to do. No need for a secondary raw memory barrier.
4772 } else {
4773 // Clear the whole thing since there are no source elements to copy.
4774 generate_clear_array(adr_type, dest, basic_elem_type,
4775 intcon(0), NULL,
4776 alloc->in(AllocateNode::AllocSize));
4777 // Use a secondary InitializeNode as raw memory barrier.
4778 // Currently it is needed only on this path since other
4779 // paths have stub or runtime calls as raw memory barriers.
4780 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4781 Compile::AliasIdxRaw,
4782 top())->as_Initialize();
4783 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4784 }
4785 }
4786
4787 // Present the results of the fast call.
4788 result_region->init_req(zero_path, control());
4789 result_i_o ->init_req(zero_path, i_o());
4790 result_memory->init_req(zero_path, memory(adr_type));
4791 }
4792
4793 if (!stopped() && dest_uninitialized) {
4794 // We have to initialize the *uncopied* part of the array to zero.
4795 // The copy destination is the slice dest[off..off+len]. The other slices
4796 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4797 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4798 Node* dest_length = alloc->in(AllocateNode::ALength);
4799 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4800 copy_length) );
4801
4802 // If there is a head section that needs zeroing, do it now.
4803 if (find_int_con(dest_offset, -1) != 0) {
4804 generate_clear_array(adr_type, dest, basic_elem_type,
4805 intcon(0), dest_offset,
4806 NULL);
4807 }
4808
4809 // Next, perform a dynamic check on the tail length.
4810 // It is often zero, and we can win big if we prove this.
4811 // There are two wins: Avoid generating the ClearArray
4812 // with its attendant messy index arithmetic, and upgrade
4813 // the copy to a more hardware-friendly word size of 64 bits.
4814 Node* tail_ctl = NULL;
4815 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4816 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4817 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4818 tail_ctl = generate_slow_guard(bol_lt, NULL);
4819 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4820 }
4821
4822 // At this point, let's assume there is no tail.
4823 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4824 // There is no tail. Try an upgrade to a 64-bit copy.
4825 bool didit = false;
4826 { PreserveJVMState pjvms(this);
4827 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4828 src, src_offset, dest, dest_offset,
4829 dest_size, dest_uninitialized);
4830 if (didit) {
4831 // Present the results of the block-copying fast call.
4832 result_region->init_req(bcopy_path, control());
4833 result_i_o ->init_req(bcopy_path, i_o());
4834 result_memory->init_req(bcopy_path, memory(adr_type));
4835 }
4836 }
4837 if (didit)
4838 set_control(top()); // no regular fast path
4839 }
4840
4841 // Clear the tail, if any.
4842 if (tail_ctl != NULL) {
4843 Node* notail_ctl = stopped() ? NULL : control();
4844 set_control(tail_ctl);
4845 if (notail_ctl == NULL) {
4846 generate_clear_array(adr_type, dest, basic_elem_type,
4847 dest_tail, NULL,
4848 dest_size);
4849 } else {
4850 // Make a local merge.
4851 Node* done_ctl = new(C,3) RegionNode(3);
4852 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4853 done_ctl->init_req(1, notail_ctl);
4854 done_mem->init_req(1, memory(adr_type));
4855 generate_clear_array(adr_type, dest, basic_elem_type,
4856 dest_tail, NULL,
4857 dest_size);
4858 done_ctl->init_req(2, control());
4859 done_mem->init_req(2, memory(adr_type));
4860 set_control( _gvn.transform(done_ctl) );
4861 set_memory( _gvn.transform(done_mem), adr_type );
4862 }
4863 }
4864 }
4865
4866 BasicType copy_type = basic_elem_type;
4867 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4868 if (!stopped() && copy_type == T_OBJECT) {
4869 // If src and dest have compatible element types, we can copy bits.
4870 // Types S[] and D[] are compatible if D is a supertype of S.
4871 //
4872 // If they are not, we will use checked_oop_disjoint_arraycopy,
4873 // which performs a fast optimistic per-oop check, and backs off
4874 // further to JVM_ArrayCopy on the first per-oop check that fails.
4875 // (Actually, we don't move raw bits only; the GC requires card marks.)
4876
4877 // Get the klassOop for both src and dest
4878 Node* src_klass = load_object_klass(src);
4879 Node* dest_klass = load_object_klass(dest);
4880
4881 // Generate the subtype check.
4882 // This might fold up statically, or then again it might not.
4883 //
4884 // Non-static example: Copying List<String>.elements to a new String[].
4885 // The backing store for a List<String> is always an Object[],
4886 // but its elements are always type String, if the generic types
4887 // are correct at the source level.
4888 //
4889 // Test S[] against D[], not S against D, because (probably)
4890 // the secondary supertype cache is less busy for S[] than S.
4891 // This usually only matters when D is an interface.
4892 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4893 // Plug failing path into checked_oop_disjoint_arraycopy
4894 if (not_subtype_ctrl != top()) {
4895 PreserveJVMState pjvms(this);
4896 set_control(not_subtype_ctrl);
4897 // (At this point we can assume disjoint_bases, since types differ.)
4898 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
4899 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4900 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4901 Node* dest_elem_klass = _gvn.transform(n1);
4902 Node* cv = generate_checkcast_arraycopy(adr_type,
4903 dest_elem_klass,
4904 src, src_offset, dest, dest_offset,
4905 ConvI2X(copy_length), dest_uninitialized);
4906 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4907 checked_control = control();
4908 checked_i_o = i_o();
4909 checked_mem = memory(adr_type);
4910 checked_value = cv;
4911 }
4912 // At this point we know we do not need type checks on oop stores.
4913
4914 // Let's see if we need card marks:
4915 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4916 // If we do not need card marks, copy using the jint or jlong stub.
4917 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4918 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4919 "sizes agree");
4920 }
4921 }
4922
4923 if (!stopped()) {
4924 // Generate the fast path, if possible.
4925 PreserveJVMState pjvms(this);
4926 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4927 src, src_offset, dest, dest_offset,
4928 ConvI2X(copy_length), dest_uninitialized);
4929
4930 // Present the results of the fast call.
4931 result_region->init_req(fast_path, control());
4932 result_i_o ->init_req(fast_path, i_o());
4933 result_memory->init_req(fast_path, memory(adr_type));
4934 }
4935
4936 // Here are all the slow paths up to this point, in one bundle:
4937 slow_control = top();
4938 if (slow_region != NULL)
4939 slow_control = _gvn.transform(slow_region);
4940 debug_only(slow_region = (RegionNode*)badAddress);
4941
4942 set_control(checked_control);
4943 if (!stopped()) {
4944 // Clean up after the checked call.
4945 // The returned value is either 0 or -1^K,
4946 // where K = number of partially transferred array elements.
4947 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4948 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4949 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4950
4951 // If it is 0, we are done, so transfer to the end.
4952 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4953 result_region->init_req(checked_path, checks_done);
4954 result_i_o ->init_req(checked_path, checked_i_o);
4955 result_memory->init_req(checked_path, checked_mem);
4956
4957 // If it is not zero, merge into the slow call.
4958 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4959 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4960 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4961 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4962 record_for_igvn(slow_reg2);
4963 slow_reg2 ->init_req(1, slow_control);
4964 slow_i_o2 ->init_req(1, slow_i_o);
4965 slow_mem2 ->init_req(1, slow_mem);
4966 slow_reg2 ->init_req(2, control());
4967 slow_i_o2 ->init_req(2, checked_i_o);
4968 slow_mem2 ->init_req(2, checked_mem);
4969
4970 slow_control = _gvn.transform(slow_reg2);
4971 slow_i_o = _gvn.transform(slow_i_o2);
4972 slow_mem = _gvn.transform(slow_mem2);
4973
4974 if (alloc != NULL) {
4975 // We'll restart from the very beginning, after zeroing the whole thing.
4976 // This can cause double writes, but that's OK since dest is brand new.
4977 // So we ignore the low 31 bits of the value returned from the stub.
4978 } else {
4979 // We must continue the copy exactly where it failed, or else
4980 // another thread might see the wrong number of writes to dest.
4981 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4982 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4983 slow_offset->init_req(1, intcon(0));
4984 slow_offset->init_req(2, checked_offset);
4985 slow_offset = _gvn.transform(slow_offset);
4986
4987 // Adjust the arguments by the conditionally incoming offset.
4988 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4989 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4990 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4991
4992 // Tweak the node variables to adjust the code produced below:
4993 src_offset = src_off_plus;
4994 dest_offset = dest_off_plus;
4995 copy_length = length_minus;
4996 }
4997 }
4998
4999 set_control(slow_control);
5000 if (!stopped()) {
5001 // Generate the slow path, if needed.
5002 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5003
5004 set_memory(slow_mem, adr_type);
5005 set_i_o(slow_i_o);
5006
5007 if (dest_uninitialized) {
5008 generate_clear_array(adr_type, dest, basic_elem_type,
5009 intcon(0), NULL,
5010 alloc->in(AllocateNode::AllocSize));
5011 }
5012
5013 generate_slow_arraycopy(adr_type,
5014 src, src_offset, dest, dest_offset,
5015 copy_length, /*dest_uninitialized*/false);
5016
5017 result_region->init_req(slow_call_path, control());
5018 result_i_o ->init_req(slow_call_path, i_o());
5019 result_memory->init_req(slow_call_path, memory(adr_type));
5020 }
5021
5022 // Remove unused edges.
5023 for (uint i = 1; i < result_region->req(); i++) {
5024 if (result_region->in(i) == NULL)
5025 result_region->init_req(i, top());
5026 }
5027
5028 // Finished; return the combined state.
5029 set_control( _gvn.transform(result_region) );
5030 set_i_o( _gvn.transform(result_i_o) );
5031 set_memory( _gvn.transform(result_memory), adr_type );
5032
5033 // The memory edges above are precise in order to model effects around
5034 // array copies accurately to allow value numbering of field loads around
5035 // arraycopy. Such field loads, both before and after, are common in Java
5036 // collections and similar classes involving header/array data structures.
5037 //
5038 // But with low number of register or when some registers are used or killed
5039 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5040 // The next memory barrier is added to avoid it. If the arraycopy can be
5041 // optimized away (which it can, sometimes) then we can manually remove
5042 // the membar also.
5043 //
5044 // Do not let reads from the cloned object float above the arraycopy.
5045 if (alloc != NULL) {
5046 // Do not let stores that initialize this object be reordered with
5047 // a subsequent store that would make this object accessible by
5048 // other threads.
5049 // Record what AllocateNode this StoreStore protects so that
5050 // escape analysis can go from the MemBarStoreStoreNode to the
5051 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5052 // based on the escape status of the AllocateNode.
5053 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5054 } else if (InsertMemBarAfterArraycopy)
5055 insert_mem_bar(Op_MemBarCPUOrder);
5056 }
5057
5058
5059 // Helper function which determines if an arraycopy immediately follows
5060 // an allocation, with no intervening tests or other escapes for the object.
5061 AllocateArrayNode*
5062 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5063 RegionNode* slow_region) {
5064 if (stopped()) return NULL; // no fast path
5065 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5066
5067 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5068 if (alloc == NULL) return NULL;
5069
5070 Node* rawmem = memory(Compile::AliasIdxRaw);
5071 // Is the allocation's memory state untouched?
5072 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5073 // Bail out if there have been raw-memory effects since the allocation.
5074 // (Example: There might have been a call or safepoint.)
5075 return NULL;
5076 }
5077 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5078 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5079 return NULL;
5080 }
5081
5082 // There must be no unexpected observers of this allocation.
5083 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5084 Node* obs = ptr->fast_out(i);
5085 if (obs != this->map()) {
5086 return NULL;
5087 }
5088 }
5089
5090 // This arraycopy must unconditionally follow the allocation of the ptr.
5091 Node* alloc_ctl = ptr->in(0);
5092 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5093
5094 Node* ctl = control();
5095 while (ctl != alloc_ctl) {
5096 // There may be guards which feed into the slow_region.
5097 // Any other control flow means that we might not get a chance
5098 // to finish initializing the allocated object.
5099 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5100 IfNode* iff = ctl->in(0)->as_If();
5101 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5102 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5103 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5104 ctl = iff->in(0); // This test feeds the known slow_region.
5105 continue;
5106 }
5107 // One more try: Various low-level checks bottom out in
5108 // uncommon traps. If the debug-info of the trap omits
5109 // any reference to the allocation, as we've already
5110 // observed, then there can be no objection to the trap.
5111 bool found_trap = false;
5112 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5113 Node* obs = not_ctl->fast_out(j);
5114 if (obs->in(0) == not_ctl && obs->is_Call() &&
5115 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5116 found_trap = true; break;
5117 }
5118 }
5119 if (found_trap) {
5120 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5121 continue;
5122 }
5123 }
5124 return NULL;
5125 }
5126
5127 // If we get this far, we have an allocation which immediately
5128 // precedes the arraycopy, and we can take over zeroing the new object.
5129 // The arraycopy will finish the initialization, and provide
5130 // a new control state to which we will anchor the destination pointer.
5131
5132 return alloc;
5133 }
5134
5135 // Helper for initialization of arrays, creating a ClearArray.
5136 // It writes zero bits in [start..end), within the body of an array object.
5137 // The memory effects are all chained onto the 'adr_type' alias category.
5138 //
5139 // Since the object is otherwise uninitialized, we are free
5140 // to put a little "slop" around the edges of the cleared area,
5141 // as long as it does not go back into the array's header,
5142 // or beyond the array end within the heap.
5143 //
5144 // The lower edge can be rounded down to the nearest jint and the
5145 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5146 //
5147 // Arguments:
5148 // adr_type memory slice where writes are generated
5149 // dest oop of the destination array
5150 // basic_elem_type element type of the destination
5151 // slice_idx array index of first element to store
5152 // slice_len number of elements to store (or NULL)
5153 // dest_size total size in bytes of the array object
5154 //
5155 // Exactly one of slice_len or dest_size must be non-NULL.
5156 // If dest_size is non-NULL, zeroing extends to the end of the object.
5157 // If slice_len is non-NULL, the slice_idx value must be a constant.
5158 void
5159 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5160 Node* dest,
5161 BasicType basic_elem_type,
5162 Node* slice_idx,
5163 Node* slice_len,
5164 Node* dest_size) {
5165 // one or the other but not both of slice_len and dest_size:
5166 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5167 if (slice_len == NULL) slice_len = top();
5168 if (dest_size == NULL) dest_size = top();
5169
5170 // operate on this memory slice:
5171 Node* mem = memory(adr_type); // memory slice to operate on
5172
5173 // scaling and rounding of indexes:
5174 int scale = exact_log2(type2aelembytes(basic_elem_type));
5175 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5176 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5177 int bump_bit = (-1 << scale) & BytesPerInt;
5178
5179 // determine constant starts and ends
5180 const intptr_t BIG_NEG = -128;
5181 assert(BIG_NEG + 2*abase < 0, "neg enough");
5182 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5183 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5184 if (slice_len_con == 0) {
5185 return; // nothing to do here
5186 }
5187 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5188 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5189 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5190 assert(end_con < 0, "not two cons");
5191 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5192 BytesPerLong);
5193 }
5194
5195 if (start_con >= 0 && end_con >= 0) {
5196 // Constant start and end. Simple.
5197 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5198 start_con, end_con, &_gvn);
5199 } else if (start_con >= 0 && dest_size != top()) {
5200 // Constant start, pre-rounded end after the tail of the array.
5201 Node* end = dest_size;
5202 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5203 start_con, end, &_gvn);
5204 } else if (start_con >= 0 && slice_len != top()) {
5205 // Constant start, non-constant end. End needs rounding up.
5206 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5207 intptr_t end_base = abase + (slice_idx_con << scale);
5208 int end_round = (-1 << scale) & (BytesPerLong - 1);
5209 Node* end = ConvI2X(slice_len);
5210 if (scale != 0)
5211 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5212 end_base += end_round;
5213 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5214 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5215 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5216 start_con, end, &_gvn);
5217 } else if (start_con < 0 && dest_size != top()) {
5218 // Non-constant start, pre-rounded end after the tail of the array.
5219 // This is almost certainly a "round-to-end" operation.
5220 Node* start = slice_idx;
5221 start = ConvI2X(start);
5222 if (scale != 0)
5223 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5224 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5225 if ((bump_bit | clear_low) != 0) {
5226 int to_clear = (bump_bit | clear_low);
5227 // Align up mod 8, then store a jint zero unconditionally
5228 // just before the mod-8 boundary.
5229 if (((abase + bump_bit) & ~to_clear) - bump_bit
5230 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5231 bump_bit = 0;
5232 assert((abase & to_clear) == 0, "array base must be long-aligned");
5233 } else {
5234 // Bump 'start' up to (or past) the next jint boundary:
5235 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5236 assert((abase & clear_low) == 0, "array base must be int-aligned");
5237 }
5238 // Round bumped 'start' down to jlong boundary in body of array.
5239 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5240 if (bump_bit != 0) {
5241 // Store a zero to the immediately preceding jint:
5242 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5243 Node* p1 = basic_plus_adr(dest, x1);
5244 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5245 mem = _gvn.transform(mem);
5246 }
5247 }
5248 Node* end = dest_size; // pre-rounded
5249 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5250 start, end, &_gvn);
5251 } else {
5252 // Non-constant start, unrounded non-constant end.
5253 // (Nobody zeroes a random midsection of an array using this routine.)
5254 ShouldNotReachHere(); // fix caller
5255 }
5256
5257 // Done.
5258 set_memory(mem, adr_type);
5259 }
5260
5261
5262 bool
5263 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5264 BasicType basic_elem_type,
5265 AllocateNode* alloc,
5266 Node* src, Node* src_offset,
5267 Node* dest, Node* dest_offset,
5268 Node* dest_size, bool dest_uninitialized) {
5269 // See if there is an advantage from block transfer.
5270 int scale = exact_log2(type2aelembytes(basic_elem_type));
5271 if (scale >= LogBytesPerLong)
5272 return false; // it is already a block transfer
5273
5274 // Look at the alignment of the starting offsets.
5275 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5276
5277 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5278 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5279 if (src_off_con < 0 || dest_off_con < 0)
5280 // At present, we can only understand constants.
5281 return false;
5282
5283 intptr_t src_off = abase + (src_off_con << scale);
5284 intptr_t dest_off = abase + (dest_off_con << scale);
5285
5286 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5287 // Non-aligned; too bad.
5288 // One more chance: Pick off an initial 32-bit word.
5289 // This is a common case, since abase can be odd mod 8.
5290 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5291 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5292 Node* sptr = basic_plus_adr(src, src_off);
5293 Node* dptr = basic_plus_adr(dest, dest_off);
5294 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5295 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5296 src_off += BytesPerInt;
5297 dest_off += BytesPerInt;
5298 } else {
5299 return false;
5300 }
5301 }
5302 assert(src_off % BytesPerLong == 0, "");
5303 assert(dest_off % BytesPerLong == 0, "");
5304
5305 // Do this copy by giant steps.
5306 Node* sptr = basic_plus_adr(src, src_off);
5307 Node* dptr = basic_plus_adr(dest, dest_off);
5308 Node* countx = dest_size;
5309 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5310 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5311
5312 bool disjoint_bases = true; // since alloc != NULL
5313 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5314 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5315
5316 return true;
5317 }
5318
5319
5320 // Helper function; generates code for the slow case.
5321 // We make a call to a runtime method which emulates the native method,
5322 // but without the native wrapper overhead.
5323 void
5324 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5325 Node* src, Node* src_offset,
5326 Node* dest, Node* dest_offset,
5327 Node* copy_length, bool dest_uninitialized) {
5328 assert(!dest_uninitialized, "Invariant");
5329 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5330 OptoRuntime::slow_arraycopy_Type(),
5331 OptoRuntime::slow_arraycopy_Java(),
5332 "slow_arraycopy", adr_type,
5333 src, src_offset, dest, dest_offset,
5334 copy_length);
5335
5336 // Handle exceptions thrown by this fellow:
5337 make_slow_call_ex(call, env()->Throwable_klass(), false);
5338 }
5339
5340 // Helper function; generates code for cases requiring runtime checks.
5341 Node*
5342 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5343 Node* dest_elem_klass,
5344 Node* src, Node* src_offset,
5345 Node* dest, Node* dest_offset,
5346 Node* copy_length, bool dest_uninitialized) {
5347 if (stopped()) return NULL;
5348
5349 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5350 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5351 return NULL;
5352 }
5353
5354 // Pick out the parameters required to perform a store-check
5355 // for the target array. This is an optimistic check. It will
5356 // look in each non-null element's class, at the desired klass's
5357 // super_check_offset, for the desired klass.
5358 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5359 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5360 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5361 Node* check_offset = ConvI2X(_gvn.transform(n3));
5362 Node* check_value = dest_elem_klass;
5363
5364 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5365 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5366
5367 // (We know the arrays are never conjoint, because their types differ.)
5368 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5369 OptoRuntime::checkcast_arraycopy_Type(),
5370 copyfunc_addr, "checkcast_arraycopy", adr_type,
5371 // five arguments, of which two are
5372 // intptr_t (jlong in LP64)
5373 src_start, dest_start,
5374 copy_length XTOP,
5375 check_offset XTOP,
5376 check_value);
5377
5378 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5379 }
5380
5381
5382 // Helper function; generates code for cases requiring runtime checks.
5383 Node*
5384 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5385 Node* src, Node* src_offset,
5386 Node* dest, Node* dest_offset,
5387 Node* copy_length, bool dest_uninitialized) {
5388 assert(!dest_uninitialized, "Invariant");
5389 if (stopped()) return NULL;
5390 address copyfunc_addr = StubRoutines::generic_arraycopy();
5391 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5392 return NULL;
5393 }
5394
5395 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5396 OptoRuntime::generic_arraycopy_Type(),
5397 copyfunc_addr, "generic_arraycopy", adr_type,
5398 src, src_offset, dest, dest_offset, copy_length);
5399
5400 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5401 }
5402
5403 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5404 void
5405 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5406 BasicType basic_elem_type,
5407 bool disjoint_bases,
5408 Node* src, Node* src_offset,
5409 Node* dest, Node* dest_offset,
5410 Node* copy_length, bool dest_uninitialized) {
5411 if (stopped()) return; // nothing to do
5412
5413 Node* src_start = src;
5414 Node* dest_start = dest;
5415 if (src_offset != NULL || dest_offset != NULL) {
5416 assert(src_offset != NULL && dest_offset != NULL, "");
5417 src_start = array_element_address(src, src_offset, basic_elem_type);
5418 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5419 }
5420
5421 // Figure out which arraycopy runtime method to call.
5422 const char* copyfunc_name = "arraycopy";
5423 address copyfunc_addr =
5424 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5425 disjoint_bases, copyfunc_name, dest_uninitialized);
5426
5427 // Call it. Note that the count_ix value is not scaled to a byte-size.
5428 make_runtime_call(RC_LEAF|RC_NO_FP,
5429 OptoRuntime::fast_arraycopy_Type(),
5430 copyfunc_addr, copyfunc_name, adr_type,
5431 src_start, dest_start, copy_length XTOP);
5432 }
5433
5434 //----------------------------inline_reference_get----------------------------
5435
5436 bool LibraryCallKit::inline_reference_get() {
5437 const int nargs = 1; // self
5438
5439 guarantee(java_lang_ref_Reference::referent_offset > 0,
5440 "should have already been set");
5441
5442 int referent_offset = java_lang_ref_Reference::referent_offset;
5443
5444 // Restore the stack and pop off the argument
5445 _sp += nargs;
5446 Node *reference_obj = pop();
5447
5448 // Null check on self without removing any arguments.
5449 _sp += nargs;
5450 reference_obj = do_null_check(reference_obj, T_OBJECT);
5451 _sp -= nargs;;
5452
5453 if (stopped()) return true;
5454
5455 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5456
5457 ciInstanceKlass* klass = env()->Object_klass();
5458 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5459
5460 Node* no_ctrl = NULL;
5461 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5462
5463 // Use the pre-barrier to record the value in the referent field
5464 pre_barrier(false /* do_load */,
5465 control(),
5466 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5467 result /* pre_val */,
5468 T_OBJECT);
5469
5470 push(result);
5471 return true;
5472 }
5473