1 /*
2 * Copyright (c) 1999, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
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 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
69
70 public:
71 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
72 : GraphKit(caller),
73 _intrinsic(intrinsic)
74 {
75 }
76
77 ciMethod* caller() const { return jvms()->method(); }
78 int bci() const { return jvms()->bci(); }
79 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
81 ciMethod* callee() const { return _intrinsic->method(); }
82 ciSignature* signature() const { return callee()->signature(); }
83 int arg_size() const { return callee()->arg_size(); }
84
85 bool try_to_inline();
86
87 // Helper functions to inline natives
88 void push_result(RegionNode* region, PhiNode* value);
89 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
90 Node* generate_slow_guard(Node* test, RegionNode* region);
91 Node* generate_fair_guard(Node* test, RegionNode* region);
92 Node* generate_negative_guard(Node* index, RegionNode* region,
93 // resulting CastII of index:
94 Node* *pos_index = NULL);
95 Node* generate_nonpositive_guard(Node* index, bool never_negative,
96 // resulting CastII of index:
97 Node* *pos_index = NULL);
98 Node* generate_limit_guard(Node* offset, Node* subseq_length,
99 Node* array_length,
100 RegionNode* region);
101 Node* generate_current_thread(Node* &tls_output);
102 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
103 bool disjoint_bases, const char* &name, bool dest_uninitialized);
104 Node* load_mirror_from_klass(Node* klass);
105 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
106 int nargs,
107 RegionNode* region, int null_path,
108 int offset);
109 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
110 RegionNode* region, int null_path) {
111 int offset = java_lang_Class::klass_offset_in_bytes();
112 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
113 region, null_path,
114 offset);
115 }
116 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
117 int nargs,
118 RegionNode* region, int null_path) {
119 int offset = java_lang_Class::array_klass_offset_in_bytes();
120 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
121 region, null_path,
122 offset);
123 }
124 Node* generate_access_flags_guard(Node* kls,
125 int modifier_mask, int modifier_bits,
126 RegionNode* region);
127 Node* generate_interface_guard(Node* kls, RegionNode* region);
128 Node* generate_array_guard(Node* kls, RegionNode* region) {
129 return generate_array_guard_common(kls, region, false, false);
130 }
131 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
132 return generate_array_guard_common(kls, region, false, true);
133 }
134 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
135 return generate_array_guard_common(kls, region, true, false);
136 }
137 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
138 return generate_array_guard_common(kls, region, true, true);
139 }
140 Node* generate_array_guard_common(Node* kls, RegionNode* region,
141 bool obj_array, bool not_array);
142 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
143 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
144 bool is_virtual = false, bool is_static = false);
145 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
146 return generate_method_call(method_id, false, true);
147 }
148 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
149 return generate_method_call(method_id, true, false);
150 }
151
152 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
153 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
154 bool inline_string_compareTo();
155 bool inline_string_indexOf();
156 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
157 bool inline_string_equals();
158 Node* pop_math_arg();
159 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
160 bool inline_math_native(vmIntrinsics::ID id);
161 bool inline_trig(vmIntrinsics::ID id);
162 bool inline_trans(vmIntrinsics::ID id);
163 bool inline_abs(vmIntrinsics::ID id);
164 bool inline_sqrt(vmIntrinsics::ID id);
165 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
166 bool inline_pow(vmIntrinsics::ID id);
167 bool inline_exp(vmIntrinsics::ID id);
168 bool inline_min_max(vmIntrinsics::ID id);
169 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
170 // This returns Type::AnyPtr, RawPtr, or OopPtr.
171 int classify_unsafe_addr(Node* &base, Node* &offset);
172 Node* make_unsafe_address(Node* base, Node* offset);
173 // Helper for inline_unsafe_access.
174 // Generates the guards that check whether the result of
175 // Unsafe.getObject should be recorded in an SATB log buffer.
176 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, int nargs, bool need_mem_bar);
177 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
178 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
179 bool inline_unsafe_allocate();
180 bool inline_unsafe_copyMemory();
181 bool inline_native_currentThread();
182 #ifdef TRACE_HAVE_INTRINSICS
183 bool inline_native_classID();
184 bool inline_native_threadID();
185 #endif
186 bool inline_native_time_funcs(address method, const char* funcName);
187 bool inline_native_isInterrupted();
188 bool inline_native_Class_query(vmIntrinsics::ID id);
189 bool inline_native_subtype_check();
190
191 bool inline_native_newArray();
192 bool inline_native_getLength();
193 bool inline_array_copyOf(bool is_copyOfRange);
194 bool inline_array_equals();
195 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
196 bool inline_native_clone(bool is_virtual);
197 bool inline_native_Reflection_getCallerClass();
198 bool is_method_invoke_or_aux_frame(JVMState* jvms);
199 // Helper function for inlining native object hash method
200 bool inline_native_hashcode(bool is_virtual, bool is_static);
201 bool inline_native_getClass();
202
203 // Helper functions for inlining arraycopy
204 bool inline_arraycopy();
205 void generate_arraycopy(const TypePtr* adr_type,
206 BasicType basic_elem_type,
207 Node* src, Node* src_offset,
208 Node* dest, Node* dest_offset,
209 Node* copy_length,
210 bool disjoint_bases = false,
211 bool length_never_negative = false,
212 RegionNode* slow_region = NULL);
213 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
214 RegionNode* slow_region);
215 void generate_clear_array(const TypePtr* adr_type,
216 Node* dest,
217 BasicType basic_elem_type,
218 Node* slice_off,
219 Node* slice_len,
220 Node* slice_end);
221 bool generate_block_arraycopy(const TypePtr* adr_type,
222 BasicType basic_elem_type,
223 AllocateNode* alloc,
224 Node* src, Node* src_offset,
225 Node* dest, Node* dest_offset,
226 Node* dest_size, bool dest_uninitialized);
227 void generate_slow_arraycopy(const TypePtr* adr_type,
228 Node* src, Node* src_offset,
229 Node* dest, Node* dest_offset,
230 Node* copy_length, bool dest_uninitialized);
231 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
232 Node* dest_elem_klass,
233 Node* src, Node* src_offset,
234 Node* dest, Node* dest_offset,
235 Node* copy_length, bool dest_uninitialized);
236 Node* generate_generic_arraycopy(const TypePtr* adr_type,
237 Node* src, Node* src_offset,
238 Node* dest, Node* dest_offset,
239 Node* copy_length, bool dest_uninitialized);
240 void generate_unchecked_arraycopy(const TypePtr* adr_type,
241 BasicType basic_elem_type,
242 bool disjoint_bases,
243 Node* src, Node* src_offset,
244 Node* dest, Node* dest_offset,
245 Node* copy_length, bool dest_uninitialized);
246 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
247 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
248 bool inline_unsafe_ordered_store(BasicType type);
249 bool inline_fp_conversions(vmIntrinsics::ID id);
250 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
251 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
252 bool inline_bitCount(vmIntrinsics::ID id);
253 bool inline_reverseBytes(vmIntrinsics::ID id);
254
255 bool inline_reference_get();
256 };
257
258
259 //---------------------------make_vm_intrinsic----------------------------
260 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
261 vmIntrinsics::ID id = m->intrinsic_id();
262 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
263
264 if (DisableIntrinsic[0] != '\0'
265 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
266 // disabled by a user request on the command line:
267 // example: -XX:DisableIntrinsic=_hashCode,_getClass
268 return NULL;
269 }
270
271 if (!m->is_loaded()) {
272 // do not attempt to inline unloaded methods
273 return NULL;
274 }
275
276 // Only a few intrinsics implement a virtual dispatch.
277 // They are expensive calls which are also frequently overridden.
278 if (is_virtual) {
279 switch (id) {
280 case vmIntrinsics::_hashCode:
281 case vmIntrinsics::_clone:
282 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
283 break;
284 default:
285 return NULL;
286 }
287 }
288
289 // -XX:-InlineNatives disables nearly all intrinsics:
290 if (!InlineNatives) {
291 switch (id) {
292 case vmIntrinsics::_indexOf:
293 case vmIntrinsics::_compareTo:
294 case vmIntrinsics::_equals:
295 case vmIntrinsics::_equalsC:
296 case vmIntrinsics::_getAndAddInt:
297 case vmIntrinsics::_getAndAddLong:
298 case vmIntrinsics::_getAndSetInt:
299 case vmIntrinsics::_getAndSetLong:
300 case vmIntrinsics::_getAndSetObject:
301 break; // InlineNatives does not control String.compareTo
302 case vmIntrinsics::_Reference_get:
303 break; // InlineNatives does not control Reference.get
304 default:
305 return NULL;
306 }
307 }
308
309 switch (id) {
310 case vmIntrinsics::_compareTo:
311 if (!SpecialStringCompareTo) return NULL;
312 break;
313 case vmIntrinsics::_indexOf:
314 if (!SpecialStringIndexOf) return NULL;
315 break;
316 case vmIntrinsics::_equals:
317 if (!SpecialStringEquals) return NULL;
318 break;
319 case vmIntrinsics::_equalsC:
320 if (!SpecialArraysEquals) return NULL;
321 break;
322 case vmIntrinsics::_arraycopy:
323 if (!InlineArrayCopy) return NULL;
324 break;
325 case vmIntrinsics::_copyMemory:
326 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
327 if (!InlineArrayCopy) return NULL;
328 break;
329 case vmIntrinsics::_hashCode:
330 if (!InlineObjectHash) return NULL;
331 break;
332 case vmIntrinsics::_clone:
333 case vmIntrinsics::_copyOf:
334 case vmIntrinsics::_copyOfRange:
335 if (!InlineObjectCopy) return NULL;
336 // These also use the arraycopy intrinsic mechanism:
337 if (!InlineArrayCopy) return NULL;
338 break;
339 case vmIntrinsics::_checkIndex:
340 // We do not intrinsify this. The optimizer does fine with it.
341 return NULL;
342
343 case vmIntrinsics::_getCallerClass:
344 if (!UseNewReflection) return NULL;
345 if (!InlineReflectionGetCallerClass) return NULL;
346 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
347 break;
348
349 case vmIntrinsics::_bitCount_i:
350 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
351 break;
352
353 case vmIntrinsics::_bitCount_l:
354 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
355 break;
356
357 case vmIntrinsics::_numberOfLeadingZeros_i:
358 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
359 break;
360
361 case vmIntrinsics::_numberOfLeadingZeros_l:
362 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
363 break;
364
365 case vmIntrinsics::_numberOfTrailingZeros_i:
366 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
367 break;
368
369 case vmIntrinsics::_numberOfTrailingZeros_l:
370 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
371 break;
372
373 case vmIntrinsics::_Reference_get:
374 // Use the intrinsic version of Reference.get() so that the value in
375 // the referent field can be registered by the G1 pre-barrier code.
376 // Also add memory barrier to prevent commoning reads from this field
377 // across safepoint since GC can change it value.
378 break;
379
380 case vmIntrinsics::_compareAndSwapObject:
381 #ifdef _LP64
382 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
383 #endif
384 break;
385
386 case vmIntrinsics::_compareAndSwapLong:
387 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
388 break;
389
390 case vmIntrinsics::_getAndAddInt:
391 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
392 break;
393
394 case vmIntrinsics::_getAndAddLong:
395 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
396 break;
397
398 case vmIntrinsics::_getAndSetInt:
399 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
400 break;
401
402 case vmIntrinsics::_getAndSetLong:
403 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
404 break;
405
406 case vmIntrinsics::_getAndSetObject:
407 #ifdef _LP64
408 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
409 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
410 break;
411 #else
412 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
413 break;
414 #endif
415
416 default:
417 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
418 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
419 break;
420 }
421
422 // -XX:-InlineClassNatives disables natives from the Class class.
423 // The flag applies to all reflective calls, notably Array.newArray
424 // (visible to Java programmers as Array.newInstance).
425 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
426 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
427 if (!InlineClassNatives) return NULL;
428 }
429
430 // -XX:-InlineThreadNatives disables natives from the Thread class.
431 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
432 if (!InlineThreadNatives) return NULL;
433 }
434
435 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
436 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
437 m->holder()->name() == ciSymbol::java_lang_Float() ||
438 m->holder()->name() == ciSymbol::java_lang_Double()) {
439 if (!InlineMathNatives) return NULL;
440 }
441
442 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
443 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
444 if (!InlineUnsafeOps) return NULL;
445 }
446
447 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
448 }
449
450 //----------------------register_library_intrinsics-----------------------
451 // Initialize this file's data structures, for each Compile instance.
452 void Compile::register_library_intrinsics() {
453 // Nothing to do here.
454 }
455
456 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
457 LibraryCallKit kit(jvms, this);
458 Compile* C = kit.C;
459 int nodes = C->unique();
460 #ifndef PRODUCT
461 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
462 char buf[1000];
463 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
464 tty->print_cr("Intrinsic %s", str);
465 }
466 #endif
467
468 if (kit.try_to_inline()) {
469 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
470 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
471 }
472 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
473 if (C->log()) {
474 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
475 vmIntrinsics::name_at(intrinsic_id()),
476 (is_virtual() ? " virtual='1'" : ""),
477 C->unique() - nodes);
478 }
479 return kit.transfer_exceptions_into_jvms();
480 }
481
482 // The intrinsic bailed out
483 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
484 if (jvms->has_method()) {
485 // Not a root compile.
486 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
487 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg);
488 } else {
489 // Root compile
490 tty->print("Did not generate intrinsic %s%s at bci:%d in",
491 vmIntrinsics::name_at(intrinsic_id()),
492 (is_virtual() ? " (virtual)" : ""), kit.bci());
493 }
494 }
495 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
496 return NULL;
497 }
498
499 bool LibraryCallKit::try_to_inline() {
500 // Handle symbolic names for otherwise undistinguished boolean switches:
501 const bool is_store = true;
502 const bool is_native_ptr = true;
503 const bool is_static = true;
504
505 if (!jvms()->has_method()) {
506 // Root JVMState has a null method.
507 assert(map()->memory()->Opcode() == Op_Parm, "");
508 // Insert the memory aliasing node
509 set_all_memory(reset_memory());
510 }
511 assert(merged_memory(), "");
512
513 switch (intrinsic_id()) {
514 case vmIntrinsics::_hashCode:
515 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
516 case vmIntrinsics::_identityHashCode:
517 return inline_native_hashcode(/*!virtual*/ false, is_static);
518 case vmIntrinsics::_getClass:
519 return inline_native_getClass();
520
521 case vmIntrinsics::_dsin:
522 case vmIntrinsics::_dcos:
523 case vmIntrinsics::_dtan:
524 case vmIntrinsics::_dabs:
525 case vmIntrinsics::_datan2:
526 case vmIntrinsics::_dsqrt:
527 case vmIntrinsics::_dexp:
528 case vmIntrinsics::_dlog:
529 case vmIntrinsics::_dlog10:
530 case vmIntrinsics::_dpow:
531 return inline_math_native(intrinsic_id());
532
533 case vmIntrinsics::_min:
534 case vmIntrinsics::_max:
535 return inline_min_max(intrinsic_id());
536
537 case vmIntrinsics::_arraycopy:
538 return inline_arraycopy();
539
540 case vmIntrinsics::_compareTo:
541 return inline_string_compareTo();
542 case vmIntrinsics::_indexOf:
543 return inline_string_indexOf();
544 case vmIntrinsics::_equals:
545 return inline_string_equals();
546
547 case vmIntrinsics::_getObject:
548 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
549 case vmIntrinsics::_getBoolean:
550 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
551 case vmIntrinsics::_getByte:
552 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
553 case vmIntrinsics::_getShort:
554 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
555 case vmIntrinsics::_getChar:
556 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
557 case vmIntrinsics::_getInt:
558 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
559 case vmIntrinsics::_getLong:
560 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
561 case vmIntrinsics::_getFloat:
562 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
563 case vmIntrinsics::_getDouble:
564 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
565
566 case vmIntrinsics::_putObject:
567 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
568 case vmIntrinsics::_putBoolean:
569 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
570 case vmIntrinsics::_putByte:
571 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
572 case vmIntrinsics::_putShort:
573 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
574 case vmIntrinsics::_putChar:
575 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
576 case vmIntrinsics::_putInt:
577 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
578 case vmIntrinsics::_putLong:
579 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
580 case vmIntrinsics::_putFloat:
581 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
582 case vmIntrinsics::_putDouble:
583 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
584
585 case vmIntrinsics::_getByte_raw:
586 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
587 case vmIntrinsics::_getShort_raw:
588 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
589 case vmIntrinsics::_getChar_raw:
590 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
591 case vmIntrinsics::_getInt_raw:
592 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
593 case vmIntrinsics::_getLong_raw:
594 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
595 case vmIntrinsics::_getFloat_raw:
596 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
597 case vmIntrinsics::_getDouble_raw:
598 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
599 case vmIntrinsics::_getAddress_raw:
600 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
601
602 case vmIntrinsics::_putByte_raw:
603 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
604 case vmIntrinsics::_putShort_raw:
605 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
606 case vmIntrinsics::_putChar_raw:
607 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
608 case vmIntrinsics::_putInt_raw:
609 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
610 case vmIntrinsics::_putLong_raw:
611 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
612 case vmIntrinsics::_putFloat_raw:
613 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
614 case vmIntrinsics::_putDouble_raw:
615 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
616 case vmIntrinsics::_putAddress_raw:
617 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
618
619 case vmIntrinsics::_getObjectVolatile:
620 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
621 case vmIntrinsics::_getBooleanVolatile:
622 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
623 case vmIntrinsics::_getByteVolatile:
624 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
625 case vmIntrinsics::_getShortVolatile:
626 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
627 case vmIntrinsics::_getCharVolatile:
628 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
629 case vmIntrinsics::_getIntVolatile:
630 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
631 case vmIntrinsics::_getLongVolatile:
632 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
633 case vmIntrinsics::_getFloatVolatile:
634 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
635 case vmIntrinsics::_getDoubleVolatile:
636 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
637
638 case vmIntrinsics::_putObjectVolatile:
639 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
640 case vmIntrinsics::_putBooleanVolatile:
641 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
642 case vmIntrinsics::_putByteVolatile:
643 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
644 case vmIntrinsics::_putShortVolatile:
645 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
646 case vmIntrinsics::_putCharVolatile:
647 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
648 case vmIntrinsics::_putIntVolatile:
649 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
650 case vmIntrinsics::_putLongVolatile:
651 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
652 case vmIntrinsics::_putFloatVolatile:
653 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
654 case vmIntrinsics::_putDoubleVolatile:
655 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
656
657 case vmIntrinsics::_prefetchRead:
658 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
659 case vmIntrinsics::_prefetchWrite:
660 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
661 case vmIntrinsics::_prefetchReadStatic:
662 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
663 case vmIntrinsics::_prefetchWriteStatic:
664 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
665
666 case vmIntrinsics::_compareAndSwapObject:
667 return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
668 case vmIntrinsics::_compareAndSwapInt:
669 return inline_unsafe_load_store(T_INT, LS_cmpxchg);
670 case vmIntrinsics::_compareAndSwapLong:
671 return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
672
673 case vmIntrinsics::_putOrderedObject:
674 return inline_unsafe_ordered_store(T_OBJECT);
675 case vmIntrinsics::_putOrderedInt:
676 return inline_unsafe_ordered_store(T_INT);
677 case vmIntrinsics::_putOrderedLong:
678 return inline_unsafe_ordered_store(T_LONG);
679
680 case vmIntrinsics::_getAndAddInt:
681 return inline_unsafe_load_store(T_INT, LS_xadd);
682 case vmIntrinsics::_getAndAddLong:
683 return inline_unsafe_load_store(T_LONG, LS_xadd);
684 case vmIntrinsics::_getAndSetInt:
685 return inline_unsafe_load_store(T_INT, LS_xchg);
686 case vmIntrinsics::_getAndSetLong:
687 return inline_unsafe_load_store(T_LONG, LS_xchg);
688 case vmIntrinsics::_getAndSetObject:
689 return inline_unsafe_load_store(T_OBJECT, LS_xchg);
690
691 case vmIntrinsics::_currentThread:
692 return inline_native_currentThread();
693 case vmIntrinsics::_isInterrupted:
694 return inline_native_isInterrupted();
695
696 #ifdef TRACE_HAVE_INTRINSICS
697 case vmIntrinsics::_classID:
698 return inline_native_classID();
699 case vmIntrinsics::_threadID:
700 return inline_native_threadID();
701 case vmIntrinsics::_counterTime:
702 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
703 #endif
704 case vmIntrinsics::_currentTimeMillis:
705 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
706 case vmIntrinsics::_nanoTime:
707 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
708 case vmIntrinsics::_allocateInstance:
709 return inline_unsafe_allocate();
710 case vmIntrinsics::_copyMemory:
711 return inline_unsafe_copyMemory();
712 case vmIntrinsics::_newArray:
713 return inline_native_newArray();
714 case vmIntrinsics::_getLength:
715 return inline_native_getLength();
716 case vmIntrinsics::_copyOf:
717 return inline_array_copyOf(false);
718 case vmIntrinsics::_copyOfRange:
719 return inline_array_copyOf(true);
720 case vmIntrinsics::_equalsC:
721 return inline_array_equals();
722 case vmIntrinsics::_clone:
723 return inline_native_clone(intrinsic()->is_virtual());
724
725 case vmIntrinsics::_isAssignableFrom:
726 return inline_native_subtype_check();
727
728 case vmIntrinsics::_isInstance:
729 case vmIntrinsics::_getModifiers:
730 case vmIntrinsics::_isInterface:
731 case vmIntrinsics::_isArray:
732 case vmIntrinsics::_isPrimitive:
733 case vmIntrinsics::_getSuperclass:
734 case vmIntrinsics::_getComponentType:
735 case vmIntrinsics::_getClassAccessFlags:
736 return inline_native_Class_query(intrinsic_id());
737
738 case vmIntrinsics::_floatToRawIntBits:
739 case vmIntrinsics::_floatToIntBits:
740 case vmIntrinsics::_intBitsToFloat:
741 case vmIntrinsics::_doubleToRawLongBits:
742 case vmIntrinsics::_doubleToLongBits:
743 case vmIntrinsics::_longBitsToDouble:
744 return inline_fp_conversions(intrinsic_id());
745
746 case vmIntrinsics::_numberOfLeadingZeros_i:
747 case vmIntrinsics::_numberOfLeadingZeros_l:
748 return inline_numberOfLeadingZeros(intrinsic_id());
749
750 case vmIntrinsics::_numberOfTrailingZeros_i:
751 case vmIntrinsics::_numberOfTrailingZeros_l:
752 return inline_numberOfTrailingZeros(intrinsic_id());
753
754 case vmIntrinsics::_bitCount_i:
755 case vmIntrinsics::_bitCount_l:
756 return inline_bitCount(intrinsic_id());
757
758 case vmIntrinsics::_reverseBytes_i:
759 case vmIntrinsics::_reverseBytes_l:
760 case vmIntrinsics::_reverseBytes_s:
761 case vmIntrinsics::_reverseBytes_c:
762 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
763
764 case vmIntrinsics::_getCallerClass:
765 return inline_native_Reflection_getCallerClass();
766
767 case vmIntrinsics::_Reference_get:
768 return inline_reference_get();
769
770 default:
771 // If you get here, it may be that someone has added a new intrinsic
772 // to the list in vmSymbols.hpp without implementing it here.
773 #ifndef PRODUCT
774 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
775 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
776 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
777 }
778 #endif
779 return false;
780 }
781 }
782
783 //------------------------------push_result------------------------------
784 // Helper function for finishing intrinsics.
785 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
786 record_for_igvn(region);
787 set_control(_gvn.transform(region));
788 BasicType value_type = value->type()->basic_type();
789 push_node(value_type, _gvn.transform(value));
790 }
791
792 //------------------------------generate_guard---------------------------
793 // Helper function for generating guarded fast-slow graph structures.
794 // The given 'test', if true, guards a slow path. If the test fails
795 // then a fast path can be taken. (We generally hope it fails.)
796 // In all cases, GraphKit::control() is updated to the fast path.
797 // The returned value represents the control for the slow path.
798 // The return value is never 'top'; it is either a valid control
799 // or NULL if it is obvious that the slow path can never be taken.
800 // Also, if region and the slow control are not NULL, the slow edge
801 // is appended to the region.
802 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
803 if (stopped()) {
804 // Already short circuited.
805 return NULL;
806 }
807
808 // Build an if node and its projections.
809 // If test is true we take the slow path, which we assume is uncommon.
810 if (_gvn.type(test) == TypeInt::ZERO) {
811 // The slow branch is never taken. No need to build this guard.
812 return NULL;
813 }
814
815 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
816
817 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
818 if (if_slow == top()) {
819 // The slow branch is never taken. No need to build this guard.
820 return NULL;
821 }
822
823 if (region != NULL)
824 region->add_req(if_slow);
825
826 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
827 set_control(if_fast);
828
829 return if_slow;
830 }
831
832 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
833 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
834 }
835 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
836 return generate_guard(test, region, PROB_FAIR);
837 }
838
839 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
840 Node* *pos_index) {
841 if (stopped())
842 return NULL; // already stopped
843 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
844 return NULL; // index is already adequately typed
845 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
846 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
847 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
848 if (is_neg != NULL && pos_index != NULL) {
849 // Emulate effect of Parse::adjust_map_after_if.
850 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
851 ccast->set_req(0, control());
852 (*pos_index) = _gvn.transform(ccast);
853 }
854 return is_neg;
855 }
856
857 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
858 Node* *pos_index) {
859 if (stopped())
860 return NULL; // already stopped
861 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
862 return NULL; // index is already adequately typed
863 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
864 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
865 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
866 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
867 if (is_notp != NULL && pos_index != NULL) {
868 // Emulate effect of Parse::adjust_map_after_if.
869 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
870 ccast->set_req(0, control());
871 (*pos_index) = _gvn.transform(ccast);
872 }
873 return is_notp;
874 }
875
876 // Make sure that 'position' is a valid limit index, in [0..length].
877 // There are two equivalent plans for checking this:
878 // A. (offset + copyLength) unsigned<= arrayLength
879 // B. offset <= (arrayLength - copyLength)
880 // We require that all of the values above, except for the sum and
881 // difference, are already known to be non-negative.
882 // Plan A is robust in the face of overflow, if offset and copyLength
883 // are both hugely positive.
884 //
885 // Plan B is less direct and intuitive, but it does not overflow at
886 // all, since the difference of two non-negatives is always
887 // representable. Whenever Java methods must perform the equivalent
888 // check they generally use Plan B instead of Plan A.
889 // For the moment we use Plan A.
890 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
891 Node* subseq_length,
892 Node* array_length,
893 RegionNode* region) {
894 if (stopped())
895 return NULL; // already stopped
896 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
897 if (zero_offset && subseq_length->eqv_uncast(array_length))
898 return NULL; // common case of whole-array copy
899 Node* last = subseq_length;
900 if (!zero_offset) // last += offset
901 last = _gvn.transform( new (C) AddINode(last, offset));
902 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
903 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
904 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
905 return is_over;
906 }
907
908
909 //--------------------------generate_current_thread--------------------
910 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
911 ciKlass* thread_klass = env()->Thread_klass();
912 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
913 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
914 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
915 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
916 tls_output = thread;
917 return threadObj;
918 }
919
920
921 //------------------------------make_string_method_node------------------------
922 // Helper method for String intrinsic functions. This version is called
923 // with str1 and str2 pointing to String object nodes.
924 //
925 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
926 Node* no_ctrl = NULL;
927
928 // Get start addr of string
929 Node* str1_value = load_String_value(no_ctrl, str1);
930 Node* str1_offset = load_String_offset(no_ctrl, str1);
931 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
932
933 // Get length of string 1
934 Node* str1_len = load_String_length(no_ctrl, str1);
935
936 Node* str2_value = load_String_value(no_ctrl, str2);
937 Node* str2_offset = load_String_offset(no_ctrl, str2);
938 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
939
940 Node* str2_len = NULL;
941 Node* result = NULL;
942
943 switch (opcode) {
944 case Op_StrIndexOf:
945 // Get length of string 2
946 str2_len = load_String_length(no_ctrl, str2);
947
948 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
949 str1_start, str1_len, str2_start, str2_len);
950 break;
951 case Op_StrComp:
952 // Get length of string 2
953 str2_len = load_String_length(no_ctrl, str2);
954
955 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
956 str1_start, str1_len, str2_start, str2_len);
957 break;
958 case Op_StrEquals:
959 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
960 str1_start, str2_start, str1_len);
961 break;
962 default:
963 ShouldNotReachHere();
964 return NULL;
965 }
966
967 // All these intrinsics have checks.
968 C->set_has_split_ifs(true); // Has chance for split-if optimization
969
970 return _gvn.transform(result);
971 }
972
973 // Helper method for String intrinsic functions. This version is called
974 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
975 // to Int nodes containing the lenghts of str1 and str2.
976 //
977 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
978
979 Node* result = NULL;
980 switch (opcode) {
981 case Op_StrIndexOf:
982 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
983 str1_start, cnt1, str2_start, cnt2);
984 break;
985 case Op_StrComp:
986 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
987 str1_start, cnt1, str2_start, cnt2);
988 break;
989 case Op_StrEquals:
990 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
991 str1_start, str2_start, cnt1);
992 break;
993 default:
994 ShouldNotReachHere();
995 return NULL;
996 }
997
998 // All these intrinsics have checks.
999 C->set_has_split_ifs(true); // Has chance for split-if optimization
1000
1001 return _gvn.transform(result);
1002 }
1003
1004 //------------------------------inline_string_compareTo------------------------
1005 bool LibraryCallKit::inline_string_compareTo() {
1006
1007 if (!Matcher::has_match_rule(Op_StrComp)) return false;
1008
1009 _sp += 2;
1010 Node *argument = pop(); // pop non-receiver first: it was pushed second
1011 Node *receiver = pop();
1012
1013 // Null check on self without removing any arguments. The argument
1014 // null check technically happens in the wrong place, which can lead to
1015 // invalid stack traces when string compare is inlined into a method
1016 // which handles NullPointerExceptions.
1017 _sp += 2;
1018 receiver = do_null_check(receiver, T_OBJECT);
1019 argument = do_null_check(argument, T_OBJECT);
1020 _sp -= 2;
1021 if (stopped()) {
1022 return true;
1023 }
1024
1025 Node* compare = make_string_method_node(Op_StrComp, receiver, argument);
1026 push(compare);
1027 return true;
1028 }
1029
1030 //------------------------------inline_string_equals------------------------
1031 bool LibraryCallKit::inline_string_equals() {
1032
1033 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
1034
1035 int nargs = 2;
1036 _sp += nargs;
1037 Node* argument = pop(); // pop non-receiver first: it was pushed second
1038 Node* receiver = pop();
1039
1040 // Null check on self without removing any arguments. The argument
1041 // null check technically happens in the wrong place, which can lead to
1042 // invalid stack traces when string compare is inlined into a method
1043 // which handles NullPointerExceptions.
1044 _sp += nargs;
1045 receiver = do_null_check(receiver, T_OBJECT);
1046 //should not do null check for argument for String.equals(), because spec
1047 //allows to specify NULL as argument.
1048 _sp -= nargs;
1049
1050 if (stopped()) {
1051 return true;
1052 }
1053
1054 // paths (plus control) merge
1055 RegionNode* region = new (C) RegionNode(5);
1056 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1057
1058 // does source == target string?
1059 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1060 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1061
1062 Node* if_eq = generate_slow_guard(bol, NULL);
1063 if (if_eq != NULL) {
1064 // receiver == argument
1065 phi->init_req(2, intcon(1));
1066 region->init_req(2, if_eq);
1067 }
1068
1069 // get String klass for instanceOf
1070 ciInstanceKlass* klass = env()->String_klass();
1071
1072 if (!stopped()) {
1073 _sp += nargs; // gen_instanceof might do an uncommon trap
1074 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1075 _sp -= nargs;
1076 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1077 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1078
1079 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1080 //instanceOf == true, fallthrough
1081
1082 if (inst_false != NULL) {
1083 phi->init_req(3, intcon(0));
1084 region->init_req(3, inst_false);
1085 }
1086 }
1087
1088 if (!stopped()) {
1089 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1090
1091 // Properly cast the argument to String
1092 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1093 // This path is taken only when argument's type is String:NotNull.
1094 argument = cast_not_null(argument, false);
1095
1096 Node* no_ctrl = NULL;
1097
1098 // Get start addr of receiver
1099 Node* receiver_val = load_String_value(no_ctrl, receiver);
1100 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1101 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1102
1103 // Get length of receiver
1104 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1105
1106 // Get start addr of argument
1107 Node* argument_val = load_String_value(no_ctrl, argument);
1108 Node* argument_offset = load_String_offset(no_ctrl, argument);
1109 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1110
1111 // Get length of argument
1112 Node* argument_cnt = load_String_length(no_ctrl, argument);
1113
1114 // Check for receiver count != argument count
1115 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1116 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1117 Node* if_ne = generate_slow_guard(bol, NULL);
1118 if (if_ne != NULL) {
1119 phi->init_req(4, intcon(0));
1120 region->init_req(4, if_ne);
1121 }
1122
1123 // Check for count == 0 is done by assembler code for StrEquals.
1124
1125 if (!stopped()) {
1126 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1127 phi->init_req(1, equals);
1128 region->init_req(1, control());
1129 }
1130 }
1131
1132 // post merge
1133 set_control(_gvn.transform(region));
1134 record_for_igvn(region);
1135
1136 push(_gvn.transform(phi));
1137
1138 return true;
1139 }
1140
1141 //------------------------------inline_array_equals----------------------------
1142 bool LibraryCallKit::inline_array_equals() {
1143
1144 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1145
1146 _sp += 2;
1147 Node *argument2 = pop();
1148 Node *argument1 = pop();
1149
1150 Node* equals =
1151 _gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1152 argument1, argument2) );
1153 push(equals);
1154 return true;
1155 }
1156
1157 // Java version of String.indexOf(constant string)
1158 // class StringDecl {
1159 // StringDecl(char[] ca) {
1160 // offset = 0;
1161 // count = ca.length;
1162 // value = ca;
1163 // }
1164 // int offset;
1165 // int count;
1166 // char[] value;
1167 // }
1168 //
1169 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1170 // int targetOffset, int cache_i, int md2) {
1171 // int cache = cache_i;
1172 // int sourceOffset = string_object.offset;
1173 // int sourceCount = string_object.count;
1174 // int targetCount = target_object.length;
1175 //
1176 // int targetCountLess1 = targetCount - 1;
1177 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1178 //
1179 // char[] source = string_object.value;
1180 // char[] target = target_object;
1181 // int lastChar = target[targetCountLess1];
1182 //
1183 // outer_loop:
1184 // for (int i = sourceOffset; i < sourceEnd; ) {
1185 // int src = source[i + targetCountLess1];
1186 // if (src == lastChar) {
1187 // // With random strings and a 4-character alphabet,
1188 // // reverse matching at this point sets up 0.8% fewer
1189 // // frames, but (paradoxically) makes 0.3% more probes.
1190 // // Since those probes are nearer the lastChar probe,
1191 // // there is may be a net D$ win with reverse matching.
1192 // // But, reversing loop inhibits unroll of inner loop
1193 // // for unknown reason. So, does running outer loop from
1194 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1195 // for (int j = 0; j < targetCountLess1; j++) {
1196 // if (target[targetOffset + j] != source[i+j]) {
1197 // if ((cache & (1 << source[i+j])) == 0) {
1198 // if (md2 < j+1) {
1199 // i += j+1;
1200 // continue outer_loop;
1201 // }
1202 // }
1203 // i += md2;
1204 // continue outer_loop;
1205 // }
1206 // }
1207 // return i - sourceOffset;
1208 // }
1209 // if ((cache & (1 << src)) == 0) {
1210 // i += targetCountLess1;
1211 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1212 // i++;
1213 // }
1214 // return -1;
1215 // }
1216
1217 //------------------------------string_indexOf------------------------
1218 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1219 jint cache_i, jint md2_i) {
1220
1221 Node* no_ctrl = NULL;
1222 float likely = PROB_LIKELY(0.9);
1223 float unlikely = PROB_UNLIKELY(0.9);
1224
1225 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1226
1227 Node* source = load_String_value(no_ctrl, string_object);
1228 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1229 Node* sourceCount = load_String_length(no_ctrl, string_object);
1230
1231 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1232 jint target_length = target_array->length();
1233 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1234 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1235
1236 IdealKit kit(this, false, true);
1237 #define __ kit.
1238 Node* zero = __ ConI(0);
1239 Node* one = __ ConI(1);
1240 Node* cache = __ ConI(cache_i);
1241 Node* md2 = __ ConI(md2_i);
1242 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1243 Node* targetCount = __ ConI(target_length);
1244 Node* targetCountLess1 = __ ConI(target_length - 1);
1245 Node* targetOffset = __ ConI(targetOffset_i);
1246 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1247
1248 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1249 Node* outer_loop = __ make_label(2 /* goto */);
1250 Node* return_ = __ make_label(1);
1251
1252 __ set(rtn,__ ConI(-1));
1253 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1254 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1255 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1256 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1257 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1258 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1259 Node* tpj = __ AddI(targetOffset, __ value(j));
1260 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1261 Node* ipj = __ AddI(__ value(i), __ value(j));
1262 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1263 __ if_then(targ, BoolTest::ne, src2); {
1264 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1265 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1266 __ increment(i, __ AddI(__ value(j), one));
1267 __ goto_(outer_loop);
1268 } __ end_if(); __ dead(j);
1269 }__ end_if(); __ dead(j);
1270 __ increment(i, md2);
1271 __ goto_(outer_loop);
1272 }__ end_if();
1273 __ increment(j, one);
1274 }__ end_loop(); __ dead(j);
1275 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1276 __ goto_(return_);
1277 }__ end_if();
1278 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1279 __ increment(i, targetCountLess1);
1280 }__ end_if();
1281 __ increment(i, one);
1282 __ bind(outer_loop);
1283 }__ end_loop(); __ dead(i);
1284 __ bind(return_);
1285
1286 // Final sync IdealKit and GraphKit.
1287 final_sync(kit);
1288 Node* result = __ value(rtn);
1289 #undef __
1290 C->set_has_loops(true);
1291 return result;
1292 }
1293
1294 //------------------------------inline_string_indexOf------------------------
1295 bool LibraryCallKit::inline_string_indexOf() {
1296
1297 _sp += 2;
1298 Node *argument = pop(); // pop non-receiver first: it was pushed second
1299 Node *receiver = pop();
1300
1301 Node* result;
1302 // Disable the use of pcmpestri until it can be guaranteed that
1303 // the load doesn't cross into the uncommited space.
1304 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1305 UseSSE42Intrinsics) {
1306 // Generate SSE4.2 version of indexOf
1307 // We currently only have match rules that use SSE4.2
1308
1309 // Null check on self without removing any arguments. The argument
1310 // null check technically happens in the wrong place, which can lead to
1311 // invalid stack traces when string compare is inlined into a method
1312 // which handles NullPointerExceptions.
1313 _sp += 2;
1314 receiver = do_null_check(receiver, T_OBJECT);
1315 argument = do_null_check(argument, T_OBJECT);
1316 _sp -= 2;
1317
1318 if (stopped()) {
1319 return true;
1320 }
1321
1322 ciInstanceKlass* str_klass = env()->String_klass();
1323 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1324
1325 // Make the merge point
1326 RegionNode* result_rgn = new (C) RegionNode(4);
1327 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1328 Node* no_ctrl = NULL;
1329
1330 // Get start addr of source string
1331 Node* source = load_String_value(no_ctrl, receiver);
1332 Node* source_offset = load_String_offset(no_ctrl, receiver);
1333 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1334
1335 // Get length of source string
1336 Node* source_cnt = load_String_length(no_ctrl, receiver);
1337
1338 // Get start addr of substring
1339 Node* substr = load_String_value(no_ctrl, argument);
1340 Node* substr_offset = load_String_offset(no_ctrl, argument);
1341 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1342
1343 // Get length of source string
1344 Node* substr_cnt = load_String_length(no_ctrl, argument);
1345
1346 // Check for substr count > string count
1347 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1348 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1349 Node* if_gt = generate_slow_guard(bol, NULL);
1350 if (if_gt != NULL) {
1351 result_phi->init_req(2, intcon(-1));
1352 result_rgn->init_req(2, if_gt);
1353 }
1354
1355 if (!stopped()) {
1356 // Check for substr count == 0
1357 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1358 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1359 Node* if_zero = generate_slow_guard(bol, NULL);
1360 if (if_zero != NULL) {
1361 result_phi->init_req(3, intcon(0));
1362 result_rgn->init_req(3, if_zero);
1363 }
1364 }
1365
1366 if (!stopped()) {
1367 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1368 result_phi->init_req(1, result);
1369 result_rgn->init_req(1, control());
1370 }
1371 set_control(_gvn.transform(result_rgn));
1372 record_for_igvn(result_rgn);
1373 result = _gvn.transform(result_phi);
1374
1375 } else { // Use LibraryCallKit::string_indexOf
1376 // don't intrinsify if argument isn't a constant string.
1377 if (!argument->is_Con()) {
1378 return false;
1379 }
1380 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1381 if (str_type == NULL) {
1382 return false;
1383 }
1384 ciInstanceKlass* klass = env()->String_klass();
1385 ciObject* str_const = str_type->const_oop();
1386 if (str_const == NULL || str_const->klass() != klass) {
1387 return false;
1388 }
1389 ciInstance* str = str_const->as_instance();
1390 assert(str != NULL, "must be instance");
1391
1392 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1393 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1394
1395 int o;
1396 int c;
1397 if (java_lang_String::has_offset_field()) {
1398 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1399 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1400 } else {
1401 o = 0;
1402 c = pat->length();
1403 }
1404
1405 // constant strings have no offset and count == length which
1406 // simplifies the resulting code somewhat so lets optimize for that.
1407 if (o != 0 || c != pat->length()) {
1408 return false;
1409 }
1410
1411 // Null check on self without removing any arguments. The argument
1412 // null check technically happens in the wrong place, which can lead to
1413 // invalid stack traces when string compare is inlined into a method
1414 // which handles NullPointerExceptions.
1415 _sp += 2;
1416 receiver = do_null_check(receiver, T_OBJECT);
1417 // No null check on the argument is needed since it's a constant String oop.
1418 _sp -= 2;
1419 if (stopped()) {
1420 return true;
1421 }
1422
1423 // The null string as a pattern always returns 0 (match at beginning of string)
1424 if (c == 0) {
1425 push(intcon(0));
1426 return true;
1427 }
1428
1429 // Generate default indexOf
1430 jchar lastChar = pat->char_at(o + (c - 1));
1431 int cache = 0;
1432 int i;
1433 for (i = 0; i < c - 1; i++) {
1434 assert(i < pat->length(), "out of range");
1435 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1436 }
1437
1438 int md2 = c;
1439 for (i = 0; i < c - 1; i++) {
1440 assert(i < pat->length(), "out of range");
1441 if (pat->char_at(o + i) == lastChar) {
1442 md2 = (c - 1) - i;
1443 }
1444 }
1445
1446 result = string_indexOf(receiver, pat, o, cache, md2);
1447 }
1448
1449 push(result);
1450 return true;
1451 }
1452
1453 //--------------------------pop_math_arg--------------------------------
1454 // Pop a double argument to a math function from the stack
1455 // rounding it if necessary.
1456 Node * LibraryCallKit::pop_math_arg() {
1457 Node *arg = pop_pair();
1458 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1459 arg = _gvn.transform( new (C) RoundDoubleNode(0, arg) );
1460 return arg;
1461 }
1462
1463 //------------------------------inline_trig----------------------------------
1464 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1465 // argument reduction which will turn into a fast/slow diamond.
1466 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1467 _sp += arg_size(); // restore stack pointer
1468 Node* arg = pop_math_arg();
1469 Node* trig = NULL;
1470
1471 switch (id) {
1472 case vmIntrinsics::_dsin:
1473 trig = _gvn.transform((Node*)new (C) SinDNode(arg));
1474 break;
1475 case vmIntrinsics::_dcos:
1476 trig = _gvn.transform((Node*)new (C) CosDNode(arg));
1477 break;
1478 case vmIntrinsics::_dtan:
1479 trig = _gvn.transform((Node*)new (C) TanDNode(arg));
1480 break;
1481 default:
1482 assert(false, "bad intrinsic was passed in");
1483 return false;
1484 }
1485
1486 // Rounding required? Check for argument reduction!
1487 if( Matcher::strict_fp_requires_explicit_rounding ) {
1488
1489 static const double pi_4 = 0.7853981633974483;
1490 static const double neg_pi_4 = -0.7853981633974483;
1491 // pi/2 in 80-bit extended precision
1492 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1493 // -pi/2 in 80-bit extended precision
1494 // 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};
1495 // Cutoff value for using this argument reduction technique
1496 //static const double pi_2_minus_epsilon = 1.564660403643354;
1497 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1498
1499 // Pseudocode for sin:
1500 // if (x <= Math.PI / 4.0) {
1501 // if (x >= -Math.PI / 4.0) return fsin(x);
1502 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1503 // } else {
1504 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1505 // }
1506 // return StrictMath.sin(x);
1507
1508 // Pseudocode for cos:
1509 // if (x <= Math.PI / 4.0) {
1510 // if (x >= -Math.PI / 4.0) return fcos(x);
1511 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1512 // } else {
1513 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1514 // }
1515 // return StrictMath.cos(x);
1516
1517 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1518 // requires a special machine instruction to load it. Instead we'll try
1519 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1520 // probably do the math inside the SIN encoding.
1521
1522 // Make the merge point
1523 RegionNode *r = new (C) RegionNode(3);
1524 Node *phi = new (C) PhiNode(r,Type::DOUBLE);
1525
1526 // Flatten arg so we need only 1 test
1527 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1528 // Node for PI/4 constant
1529 Node *pi4 = makecon(TypeD::make(pi_4));
1530 // Check PI/4 : abs(arg)
1531 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1532 // Check: If PI/4 < abs(arg) then go slow
1533 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1534 // Branch either way
1535 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1536 set_control(opt_iff(r,iff));
1537
1538 // Set fast path result
1539 phi->init_req(2,trig);
1540
1541 // Slow path - non-blocking leaf call
1542 Node* call = NULL;
1543 switch (id) {
1544 case vmIntrinsics::_dsin:
1545 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1546 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1547 "Sin", NULL, arg, top());
1548 break;
1549 case vmIntrinsics::_dcos:
1550 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1551 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1552 "Cos", NULL, arg, top());
1553 break;
1554 case vmIntrinsics::_dtan:
1555 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1556 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1557 "Tan", NULL, arg, top());
1558 break;
1559 }
1560 assert(control()->in(0) == call, "");
1561 Node* slow_result = _gvn.transform(new (C) ProjNode(call,TypeFunc::Parms));
1562 r->init_req(1,control());
1563 phi->init_req(1,slow_result);
1564
1565 // Post-merge
1566 set_control(_gvn.transform(r));
1567 record_for_igvn(r);
1568 trig = _gvn.transform(phi);
1569
1570 C->set_has_split_ifs(true); // Has chance for split-if optimization
1571 }
1572 // Push result back on JVM stack
1573 push_pair(trig);
1574 return true;
1575 }
1576
1577 //------------------------------inline_sqrt-------------------------------------
1578 // Inline square root instruction, if possible.
1579 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1580 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1581 _sp += arg_size(); // restore stack pointer
1582 push_pair(_gvn.transform(new (C) SqrtDNode(0, pop_math_arg())));
1583 return true;
1584 }
1585
1586 //------------------------------inline_abs-------------------------------------
1587 // Inline absolute value instruction, if possible.
1588 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1589 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1590 _sp += arg_size(); // restore stack pointer
1591 push_pair(_gvn.transform(new (C) AbsDNode(pop_math_arg())));
1592 return true;
1593 }
1594
1595 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1596 //-------------------
1597 //result=(result.isNaN())? funcAddr():result;
1598 // Check: If isNaN() by checking result!=result? then either trap
1599 // or go to runtime
1600 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result,result));
1601 // Build the boolean node
1602 Node* bolisnum = _gvn.transform( new (C) BoolNode(cmpisnan, BoolTest::eq) );
1603
1604 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1605 {
1606 BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1607 // End the current control-flow path
1608 push_pair(x);
1609 if (y != NULL) {
1610 push_pair(y);
1611 }
1612 // The pow or exp intrinsic returned a NaN, which requires a call
1613 // to the runtime. Recompile with the runtime call.
1614 uncommon_trap(Deoptimization::Reason_intrinsic,
1615 Deoptimization::Action_make_not_entrant);
1616 }
1617 push_pair(result);
1618 } else {
1619 // If this inlining ever returned NaN in the past, we compile a call
1620 // to the runtime to properly handle corner cases
1621
1622 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1623 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1624 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1625
1626 if (!if_slow->is_top()) {
1627 RegionNode* result_region = new(C) RegionNode(3);
1628 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1629
1630 result_region->init_req(1, if_fast);
1631 result_val->init_req(1, result);
1632
1633 set_control(if_slow);
1634
1635 const TypePtr* no_memory_effects = NULL;
1636 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1637 no_memory_effects,
1638 x, top(), y, y ? top() : NULL);
1639 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1640 #ifdef ASSERT
1641 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1642 assert(value_top == top(), "second value must be top");
1643 #endif
1644
1645 result_region->init_req(2, control());
1646 result_val->init_req(2, value);
1647 push_result(result_region, result_val);
1648 } else {
1649 push_pair(result);
1650 }
1651 }
1652 }
1653
1654 //------------------------------inline_exp-------------------------------------
1655 // Inline exp instructions, if possible. The Intel hardware only misses
1656 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1657 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1658 assert(id == vmIntrinsics::_dexp, "Not exp");
1659
1660 _sp += arg_size(); // restore stack pointer
1661 Node *x = pop_math_arg();
1662 Node *result = _gvn.transform(new (C) ExpDNode(0,x));
1663
1664 finish_pow_exp(result, x, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1665
1666 C->set_has_split_ifs(true); // Has chance for split-if optimization
1667
1668 return true;
1669 }
1670
1671 //------------------------------inline_pow-------------------------------------
1672 // Inline power instructions, if possible.
1673 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1674 assert(id == vmIntrinsics::_dpow, "Not pow");
1675
1676 // Pseudocode for pow
1677 // if (x <= 0.0) {
1678 // long longy = (long)y;
1679 // if ((double)longy == y) { // if y is long
1680 // if (y + 1 == y) longy = 0; // huge number: even
1681 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1682 // } else {
1683 // result = NaN;
1684 // }
1685 // } else {
1686 // result = DPow(x,y);
1687 // }
1688 // if (result != result)? {
1689 // result = uncommon_trap() or runtime_call();
1690 // }
1691 // return result;
1692
1693 _sp += arg_size(); // restore stack pointer
1694 Node* y = pop_math_arg();
1695 Node* x = pop_math_arg();
1696
1697 Node* result = NULL;
1698
1699 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1700 // Short form: skip the fancy tests and just check for NaN result.
1701 result = _gvn.transform( new (C) PowDNode(0, x, y) );
1702 } else {
1703 // If this inlining ever returned NaN in the past, include all
1704 // checks + call to the runtime.
1705
1706 // Set the merge point for If node with condition of (x <= 0.0)
1707 // There are four possible paths to region node and phi node
1708 RegionNode *r = new (C) RegionNode(4);
1709 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1710
1711 // Build the first if node: if (x <= 0.0)
1712 // Node for 0 constant
1713 Node *zeronode = makecon(TypeD::ZERO);
1714 // Check x:0
1715 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1716 // Check: If (x<=0) then go complex path
1717 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1718 // Branch either way
1719 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1720 // Fast path taken; set region slot 3
1721 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1722 r->init_req(3,fast_taken); // Capture fast-control
1723
1724 // Fast path not-taken, i.e. slow path
1725 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1726
1727 // Set fast path result
1728 Node *fast_result = _gvn.transform( new (C) PowDNode(0, x, y) );
1729 phi->init_req(3, fast_result);
1730
1731 // Complex path
1732 // Build the second if node (if y is long)
1733 // Node for (long)y
1734 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1735 // Node for (double)((long) y)
1736 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1737 // Check (double)((long) y) : y
1738 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1739 // Check if (y isn't long) then go to slow path
1740
1741 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1742 // Branch either way
1743 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1744 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1745
1746 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1747
1748 // Calculate DPow(abs(x), y)*(1 & (long)y)
1749 // Node for constant 1
1750 Node *conone = longcon(1);
1751 // 1& (long)y
1752 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1753
1754 // A huge number is always even. Detect a huge number by checking
1755 // if y + 1 == y and set integer to be tested for parity to 0.
1756 // Required for corner case:
1757 // (long)9.223372036854776E18 = max_jlong
1758 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1759 // max_jlong is odd but 9.223372036854776E18 is even
1760 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1761 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1762 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1763 Node* correctedsign = NULL;
1764 if (ConditionalMoveLimit != 0) {
1765 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1766 } else {
1767 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1768 RegionNode *r = new (C) RegionNode(3);
1769 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1770 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1771 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1772 phi->init_req(1, signnode);
1773 phi->init_req(2, longcon(0));
1774 correctedsign = _gvn.transform(phi);
1775 ylong_path = _gvn.transform(r);
1776 record_for_igvn(r);
1777 }
1778
1779 // zero node
1780 Node *conzero = longcon(0);
1781 // Check (1&(long)y)==0?
1782 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1783 // Check if (1&(long)y)!=0?, if so the result is negative
1784 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1785 // abs(x)
1786 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1787 // abs(x)^y
1788 Node *absxpowy = _gvn.transform( new (C) PowDNode(0, absx, y) );
1789 // -abs(x)^y
1790 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1791 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1792 Node *signresult = NULL;
1793 if (ConditionalMoveLimit != 0) {
1794 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1795 } else {
1796 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1797 RegionNode *r = new (C) RegionNode(3);
1798 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1799 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1800 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1801 phi->init_req(1, absxpowy);
1802 phi->init_req(2, negabsxpowy);
1803 signresult = _gvn.transform(phi);
1804 ylong_path = _gvn.transform(r);
1805 record_for_igvn(r);
1806 }
1807 // Set complex path fast result
1808 r->init_req(2, ylong_path);
1809 phi->init_req(2, signresult);
1810
1811 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1812 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1813 r->init_req(1,slow_path);
1814 phi->init_req(1,slow_result);
1815
1816 // Post merge
1817 set_control(_gvn.transform(r));
1818 record_for_igvn(r);
1819 result=_gvn.transform(phi);
1820 }
1821
1822 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1823
1824 C->set_has_split_ifs(true); // Has chance for split-if optimization
1825
1826 return true;
1827 }
1828
1829 //------------------------------inline_trans-------------------------------------
1830 // Inline transcendental instructions, if possible. The Intel hardware gets
1831 // these right, no funny corner cases missed.
1832 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1833 _sp += arg_size(); // restore stack pointer
1834 Node* arg = pop_math_arg();
1835 Node* trans = NULL;
1836
1837 switch (id) {
1838 case vmIntrinsics::_dlog:
1839 trans = _gvn.transform((Node*)new (C) LogDNode(arg));
1840 break;
1841 case vmIntrinsics::_dlog10:
1842 trans = _gvn.transform((Node*)new (C) Log10DNode(arg));
1843 break;
1844 default:
1845 assert(false, "bad intrinsic was passed in");
1846 return false;
1847 }
1848
1849 // Push result back on JVM stack
1850 push_pair(trans);
1851 return true;
1852 }
1853
1854 //------------------------------runtime_math-----------------------------
1855 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1856 Node* a = NULL;
1857 Node* b = NULL;
1858
1859 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1860 "must be (DD)D or (D)D type");
1861
1862 // Inputs
1863 _sp += arg_size(); // restore stack pointer
1864 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1865 b = pop_math_arg();
1866 }
1867 a = pop_math_arg();
1868
1869 const TypePtr* no_memory_effects = NULL;
1870 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1871 no_memory_effects,
1872 a, top(), b, b ? top() : NULL);
1873 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1874 #ifdef ASSERT
1875 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1876 assert(value_top == top(), "second value must be top");
1877 #endif
1878
1879 push_pair(value);
1880 return true;
1881 }
1882
1883 //------------------------------inline_math_native-----------------------------
1884 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1885 switch (id) {
1886 // These intrinsics are not properly supported on all hardware
1887 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1888 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1889 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1890 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1891 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1892 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1893
1894 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1895 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1896 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1897 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1898
1899 // These intrinsics are supported on all hardware
1900 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1901 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1902
1903 case vmIntrinsics::_dexp: return
1904 Matcher::has_match_rule(Op_ExpD) ? inline_exp(id) :
1905 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1906 case vmIntrinsics::_dpow: return
1907 Matcher::has_match_rule(Op_PowD) ? inline_pow(id) :
1908 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1909
1910 // These intrinsics are not yet correctly implemented
1911 case vmIntrinsics::_datan2:
1912 return false;
1913
1914 default:
1915 ShouldNotReachHere();
1916 return false;
1917 }
1918 }
1919
1920 static bool is_simple_name(Node* n) {
1921 return (n->req() == 1 // constant
1922 || (n->is_Type() && n->as_Type()->type()->singleton())
1923 || n->is_Proj() // parameter or return value
1924 || n->is_Phi() // local of some sort
1925 );
1926 }
1927
1928 //----------------------------inline_min_max-----------------------------------
1929 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1930 push(generate_min_max(id, argument(0), argument(1)));
1931
1932 return true;
1933 }
1934
1935 Node*
1936 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1937 // These are the candidate return value:
1938 Node* xvalue = x0;
1939 Node* yvalue = y0;
1940
1941 if (xvalue == yvalue) {
1942 return xvalue;
1943 }
1944
1945 bool want_max = (id == vmIntrinsics::_max);
1946
1947 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1948 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1949 if (txvalue == NULL || tyvalue == NULL) return top();
1950 // This is not really necessary, but it is consistent with a
1951 // hypothetical MaxINode::Value method:
1952 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1953
1954 // %%% This folding logic should (ideally) be in a different place.
1955 // Some should be inside IfNode, and there to be a more reliable
1956 // transformation of ?: style patterns into cmoves. We also want
1957 // more powerful optimizations around cmove and min/max.
1958
1959 // Try to find a dominating comparison of these guys.
1960 // It can simplify the index computation for Arrays.copyOf
1961 // and similar uses of System.arraycopy.
1962 // First, compute the normalized version of CmpI(x, y).
1963 int cmp_op = Op_CmpI;
1964 Node* xkey = xvalue;
1965 Node* ykey = yvalue;
1966 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
1967 if (ideal_cmpxy->is_Cmp()) {
1968 // E.g., if we have CmpI(length - offset, count),
1969 // it might idealize to CmpI(length, count + offset)
1970 cmp_op = ideal_cmpxy->Opcode();
1971 xkey = ideal_cmpxy->in(1);
1972 ykey = ideal_cmpxy->in(2);
1973 }
1974
1975 // Start by locating any relevant comparisons.
1976 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1977 Node* cmpxy = NULL;
1978 Node* cmpyx = NULL;
1979 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1980 Node* cmp = start_from->fast_out(k);
1981 if (cmp->outcnt() > 0 && // must have prior uses
1982 cmp->in(0) == NULL && // must be context-independent
1983 cmp->Opcode() == cmp_op) { // right kind of compare
1984 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1985 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1986 }
1987 }
1988
1989 const int NCMPS = 2;
1990 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1991 int cmpn;
1992 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1993 if (cmps[cmpn] != NULL) break; // find a result
1994 }
1995 if (cmpn < NCMPS) {
1996 // Look for a dominating test that tells us the min and max.
1997 int depth = 0; // Limit search depth for speed
1998 Node* dom = control();
1999 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2000 if (++depth >= 100) break;
2001 Node* ifproj = dom;
2002 if (!ifproj->is_Proj()) continue;
2003 Node* iff = ifproj->in(0);
2004 if (!iff->is_If()) continue;
2005 Node* bol = iff->in(1);
2006 if (!bol->is_Bool()) continue;
2007 Node* cmp = bol->in(1);
2008 if (cmp == NULL) continue;
2009 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2010 if (cmps[cmpn] == cmp) break;
2011 if (cmpn == NCMPS) continue;
2012 BoolTest::mask btest = bol->as_Bool()->_test._test;
2013 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2014 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2015 // At this point, we know that 'x btest y' is true.
2016 switch (btest) {
2017 case BoolTest::eq:
2018 // They are proven equal, so we can collapse the min/max.
2019 // Either value is the answer. Choose the simpler.
2020 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2021 return yvalue;
2022 return xvalue;
2023 case BoolTest::lt: // x < y
2024 case BoolTest::le: // x <= y
2025 return (want_max ? yvalue : xvalue);
2026 case BoolTest::gt: // x > y
2027 case BoolTest::ge: // x >= y
2028 return (want_max ? xvalue : yvalue);
2029 }
2030 }
2031 }
2032
2033 // We failed to find a dominating test.
2034 // Let's pick a test that might GVN with prior tests.
2035 Node* best_bol = NULL;
2036 BoolTest::mask best_btest = BoolTest::illegal;
2037 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2038 Node* cmp = cmps[cmpn];
2039 if (cmp == NULL) continue;
2040 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2041 Node* bol = cmp->fast_out(j);
2042 if (!bol->is_Bool()) continue;
2043 BoolTest::mask btest = bol->as_Bool()->_test._test;
2044 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2045 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2046 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2047 best_bol = bol->as_Bool();
2048 best_btest = btest;
2049 }
2050 }
2051 }
2052
2053 Node* answer_if_true = NULL;
2054 Node* answer_if_false = NULL;
2055 switch (best_btest) {
2056 default:
2057 if (cmpxy == NULL)
2058 cmpxy = ideal_cmpxy;
2059 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2060 // and fall through:
2061 case BoolTest::lt: // x < y
2062 case BoolTest::le: // x <= y
2063 answer_if_true = (want_max ? yvalue : xvalue);
2064 answer_if_false = (want_max ? xvalue : yvalue);
2065 break;
2066 case BoolTest::gt: // x > y
2067 case BoolTest::ge: // x >= y
2068 answer_if_true = (want_max ? xvalue : yvalue);
2069 answer_if_false = (want_max ? yvalue : xvalue);
2070 break;
2071 }
2072
2073 jint hi, lo;
2074 if (want_max) {
2075 // We can sharpen the minimum.
2076 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2077 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2078 } else {
2079 // We can sharpen the maximum.
2080 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2081 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2082 }
2083
2084 // Use a flow-free graph structure, to avoid creating excess control edges
2085 // which could hinder other optimizations.
2086 // Since Math.min/max is often used with arraycopy, we want
2087 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2088 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2089 answer_if_false, answer_if_true,
2090 TypeInt::make(lo, hi, widen));
2091
2092 return _gvn.transform(cmov);
2093
2094 /*
2095 // This is not as desirable as it may seem, since Min and Max
2096 // nodes do not have a full set of optimizations.
2097 // And they would interfere, anyway, with 'if' optimizations
2098 // and with CMoveI canonical forms.
2099 switch (id) {
2100 case vmIntrinsics::_min:
2101 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2102 case vmIntrinsics::_max:
2103 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2104 default:
2105 ShouldNotReachHere();
2106 }
2107 */
2108 }
2109
2110 inline int
2111 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2112 const TypePtr* base_type = TypePtr::NULL_PTR;
2113 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2114 if (base_type == NULL) {
2115 // Unknown type.
2116 return Type::AnyPtr;
2117 } else if (base_type == TypePtr::NULL_PTR) {
2118 // Since this is a NULL+long form, we have to switch to a rawptr.
2119 base = _gvn.transform( new (C) CastX2PNode(offset) );
2120 offset = MakeConX(0);
2121 return Type::RawPtr;
2122 } else if (base_type->base() == Type::RawPtr) {
2123 return Type::RawPtr;
2124 } else if (base_type->isa_oopptr()) {
2125 // Base is never null => always a heap address.
2126 if (base_type->ptr() == TypePtr::NotNull) {
2127 return Type::OopPtr;
2128 }
2129 // Offset is small => always a heap address.
2130 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2131 if (offset_type != NULL &&
2132 base_type->offset() == 0 && // (should always be?)
2133 offset_type->_lo >= 0 &&
2134 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2135 return Type::OopPtr;
2136 }
2137 // Otherwise, it might either be oop+off or NULL+addr.
2138 return Type::AnyPtr;
2139 } else {
2140 // No information:
2141 return Type::AnyPtr;
2142 }
2143 }
2144
2145 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2146 int kind = classify_unsafe_addr(base, offset);
2147 if (kind == Type::RawPtr) {
2148 return basic_plus_adr(top(), base, offset);
2149 } else {
2150 return basic_plus_adr(base, offset);
2151 }
2152 }
2153
2154 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2155 // inline int Integer.numberOfLeadingZeros(int)
2156 // inline int Long.numberOfLeadingZeros(long)
2157 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2158 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2159 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2160 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2161 _sp += arg_size(); // restore stack pointer
2162 switch (id) {
2163 case vmIntrinsics::_numberOfLeadingZeros_i:
2164 push(_gvn.transform(new (C) CountLeadingZerosINode(pop())));
2165 break;
2166 case vmIntrinsics::_numberOfLeadingZeros_l:
2167 push(_gvn.transform(new (C) CountLeadingZerosLNode(pop_pair())));
2168 break;
2169 default:
2170 ShouldNotReachHere();
2171 }
2172 return true;
2173 }
2174
2175 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2176 // inline int Integer.numberOfTrailingZeros(int)
2177 // inline int Long.numberOfTrailingZeros(long)
2178 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2179 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2180 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2181 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2182 _sp += arg_size(); // restore stack pointer
2183 switch (id) {
2184 case vmIntrinsics::_numberOfTrailingZeros_i:
2185 push(_gvn.transform(new (C) CountTrailingZerosINode(pop())));
2186 break;
2187 case vmIntrinsics::_numberOfTrailingZeros_l:
2188 push(_gvn.transform(new (C) CountTrailingZerosLNode(pop_pair())));
2189 break;
2190 default:
2191 ShouldNotReachHere();
2192 }
2193 return true;
2194 }
2195
2196 //----------------------------inline_bitCount_int/long-----------------------
2197 // inline int Integer.bitCount(int)
2198 // inline int Long.bitCount(long)
2199 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2200 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2201 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2202 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2203 _sp += arg_size(); // restore stack pointer
2204 switch (id) {
2205 case vmIntrinsics::_bitCount_i:
2206 push(_gvn.transform(new (C) PopCountINode(pop())));
2207 break;
2208 case vmIntrinsics::_bitCount_l:
2209 push(_gvn.transform(new (C) PopCountLNode(pop_pair())));
2210 break;
2211 default:
2212 ShouldNotReachHere();
2213 }
2214 return true;
2215 }
2216
2217 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2218 // inline Integer.reverseBytes(int)
2219 // inline Long.reverseBytes(long)
2220 // inline Character.reverseBytes(char)
2221 // inline Short.reverseBytes(short)
2222 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2223 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2224 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2225 "not reverse Bytes");
2226 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2227 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2228 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2229 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2230 _sp += arg_size(); // restore stack pointer
2231 switch (id) {
2232 case vmIntrinsics::_reverseBytes_i:
2233 push(_gvn.transform(new (C) ReverseBytesINode(0, pop())));
2234 break;
2235 case vmIntrinsics::_reverseBytes_l:
2236 push_pair(_gvn.transform(new (C) ReverseBytesLNode(0, pop_pair())));
2237 break;
2238 case vmIntrinsics::_reverseBytes_c:
2239 push(_gvn.transform(new (C) ReverseBytesUSNode(0, pop())));
2240 break;
2241 case vmIntrinsics::_reverseBytes_s:
2242 push(_gvn.transform(new (C) ReverseBytesSNode(0, pop())));
2243 break;
2244 default:
2245 ;
2246 }
2247 return true;
2248 }
2249
2250 //----------------------------inline_unsafe_access----------------------------
2251
2252 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2253
2254 // Helper that guards and inserts a pre-barrier.
2255 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2256 Node* pre_val, int nargs, bool need_mem_bar) {
2257 // We could be accessing the referent field of a reference object. If so, when G1
2258 // is enabled, we need to log the value in the referent field in an SATB buffer.
2259 // This routine performs some compile time filters and generates suitable
2260 // runtime filters that guard the pre-barrier code.
2261 // Also add memory barrier for non volatile load from the referent field
2262 // to prevent commoning of loads across safepoint.
2263 if (!UseG1GC && !need_mem_bar)
2264 return;
2265
2266 // Some compile time checks.
2267
2268 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2269 const TypeX* otype = offset->find_intptr_t_type();
2270 if (otype != NULL && otype->is_con() &&
2271 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2272 // Constant offset but not the reference_offset so just return
2273 return;
2274 }
2275
2276 // We only need to generate the runtime guards for instances.
2277 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2278 if (btype != NULL) {
2279 if (btype->isa_aryptr()) {
2280 // Array type so nothing to do
2281 return;
2282 }
2283
2284 const TypeInstPtr* itype = btype->isa_instptr();
2285 if (itype != NULL) {
2286 // Can the klass of base_oop be statically determined to be
2287 // _not_ a sub-class of Reference and _not_ Object?
2288 ciKlass* klass = itype->klass();
2289 if ( klass->is_loaded() &&
2290 !klass->is_subtype_of(env()->Reference_klass()) &&
2291 !env()->Object_klass()->is_subtype_of(klass)) {
2292 return;
2293 }
2294 }
2295 }
2296
2297 // The compile time filters did not reject base_oop/offset so
2298 // we need to generate the following runtime filters
2299 //
2300 // if (offset == java_lang_ref_Reference::_reference_offset) {
2301 // if (instance_of(base, java.lang.ref.Reference)) {
2302 // pre_barrier(_, pre_val, ...);
2303 // }
2304 // }
2305
2306 float likely = PROB_LIKELY(0.999);
2307 float unlikely = PROB_UNLIKELY(0.999);
2308
2309 IdealKit ideal(this);
2310 #define __ ideal.
2311
2312 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2313
2314 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2315 // Update graphKit memory and control from IdealKit.
2316 sync_kit(ideal);
2317
2318 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2319 _sp += nargs; // gen_instanceof might do an uncommon trap
2320 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2321 _sp -= nargs;
2322
2323 // Update IdealKit memory and control from graphKit.
2324 __ sync_kit(this);
2325
2326 Node* one = __ ConI(1);
2327 // is_instof == 0 if base_oop == NULL
2328 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2329
2330 // Update graphKit from IdeakKit.
2331 sync_kit(ideal);
2332
2333 // Use the pre-barrier to record the value in the referent field
2334 pre_barrier(false /* do_load */,
2335 __ ctrl(),
2336 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2337 pre_val /* pre_val */,
2338 T_OBJECT);
2339 if (need_mem_bar) {
2340 // Add memory barrier to prevent commoning reads from this field
2341 // across safepoint since GC can change its value.
2342 insert_mem_bar(Op_MemBarCPUOrder);
2343 }
2344 // Update IdealKit from graphKit.
2345 __ sync_kit(this);
2346
2347 } __ end_if(); // _ref_type != ref_none
2348 } __ end_if(); // offset == referent_offset
2349
2350 // Final sync IdealKit and GraphKit.
2351 final_sync(ideal);
2352 #undef __
2353 }
2354
2355
2356 // Interpret Unsafe.fieldOffset cookies correctly:
2357 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2358
2359 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2360 // Attempt to infer a sharper value type from the offset and base type.
2361 ciKlass* sharpened_klass = NULL;
2362
2363 // See if it is an instance field, with an object type.
2364 if (alias_type->field() != NULL) {
2365 assert(!is_native_ptr, "native pointer op cannot use a java address");
2366 if (alias_type->field()->type()->is_klass()) {
2367 sharpened_klass = alias_type->field()->type()->as_klass();
2368 }
2369 }
2370
2371 // See if it is a narrow oop array.
2372 if (adr_type->isa_aryptr()) {
2373 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2374 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2375 if (elem_type != NULL) {
2376 sharpened_klass = elem_type->klass();
2377 }
2378 }
2379 }
2380
2381 if (sharpened_klass != NULL) {
2382 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2383
2384 #ifndef PRODUCT
2385 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2386 tty->print(" from base type: "); adr_type->dump();
2387 tty->print(" sharpened value: "); tjp->dump();
2388 }
2389 #endif
2390 // Sharpen the value type.
2391 return tjp;
2392 }
2393 return NULL;
2394 }
2395
2396 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2397 if (callee()->is_static()) return false; // caller must have the capability!
2398
2399 #ifndef PRODUCT
2400 {
2401 ResourceMark rm;
2402 // Check the signatures.
2403 ciSignature* sig = signature();
2404 #ifdef ASSERT
2405 if (!is_store) {
2406 // Object getObject(Object base, int/long offset), etc.
2407 BasicType rtype = sig->return_type()->basic_type();
2408 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2409 rtype = T_ADDRESS; // it is really a C void*
2410 assert(rtype == type, "getter must return the expected value");
2411 if (!is_native_ptr) {
2412 assert(sig->count() == 2, "oop getter has 2 arguments");
2413 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2414 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2415 } else {
2416 assert(sig->count() == 1, "native getter has 1 argument");
2417 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2418 }
2419 } else {
2420 // void putObject(Object base, int/long offset, Object x), etc.
2421 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2422 if (!is_native_ptr) {
2423 assert(sig->count() == 3, "oop putter has 3 arguments");
2424 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2425 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2426 } else {
2427 assert(sig->count() == 2, "native putter has 2 arguments");
2428 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2429 }
2430 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2431 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2432 vtype = T_ADDRESS; // it is really a C void*
2433 assert(vtype == type, "putter must accept the expected value");
2434 }
2435 #endif // ASSERT
2436 }
2437 #endif //PRODUCT
2438
2439 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2440
2441 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2442
2443 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2444 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2445 assert(callee()->arg_size() == nargs, "must be");
2446
2447 debug_only(int saved_sp = _sp);
2448 _sp += nargs;
2449
2450 Node* val;
2451 debug_only(val = (Node*)(uintptr_t)-1);
2452
2453
2454 if (is_store) {
2455 // Get the value being stored. (Pop it first; it was pushed last.)
2456 switch (type) {
2457 case T_DOUBLE:
2458 case T_LONG:
2459 case T_ADDRESS:
2460 val = pop_pair();
2461 break;
2462 default:
2463 val = pop();
2464 }
2465 }
2466
2467 // Build address expression. See the code in inline_unsafe_prefetch.
2468 Node *adr;
2469 Node *heap_base_oop = top();
2470 Node* offset = top();
2471
2472 if (!is_native_ptr) {
2473 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2474 offset = pop_pair();
2475 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2476 Node* base = pop();
2477 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2478 // to be plain byte offsets, which are also the same as those accepted
2479 // by oopDesc::field_base.
2480 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2481 "fieldOffset must be byte-scaled");
2482 // 32-bit machines ignore the high half!
2483 offset = ConvL2X(offset);
2484 adr = make_unsafe_address(base, offset);
2485 heap_base_oop = base;
2486 } else {
2487 Node* ptr = pop_pair();
2488 // Adjust Java long to machine word:
2489 ptr = ConvL2X(ptr);
2490 adr = make_unsafe_address(NULL, ptr);
2491 }
2492
2493 // Pop receiver last: it was pushed first.
2494 Node *receiver = pop();
2495
2496 assert(saved_sp == _sp, "must have correct argument count");
2497
2498 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2499
2500 // First guess at the value type.
2501 const Type *value_type = Type::get_const_basic_type(type);
2502
2503 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2504 // there was not enough information to nail it down.
2505 Compile::AliasType* alias_type = C->alias_type(adr_type);
2506 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2507
2508 // We will need memory barriers unless we can determine a unique
2509 // alias category for this reference. (Note: If for some reason
2510 // the barriers get omitted and the unsafe reference begins to "pollute"
2511 // the alias analysis of the rest of the graph, either Compile::can_alias
2512 // or Compile::must_alias will throw a diagnostic assert.)
2513 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2514
2515 // If we are reading the value of the referent field of a Reference
2516 // object (either by using Unsafe directly or through reflection)
2517 // then, if G1 is enabled, we need to record the referent in an
2518 // SATB log buffer using the pre-barrier mechanism.
2519 // Also we need to add memory barrier to prevent commoning reads
2520 // from this field across safepoint since GC can change its value.
2521 bool need_read_barrier = !is_native_ptr && !is_store &&
2522 offset != top() && heap_base_oop != top();
2523
2524 if (!is_store && type == T_OBJECT) {
2525 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2526 if (tjp != NULL) {
2527 value_type = tjp;
2528 }
2529 }
2530
2531 // Null check on self without removing any arguments. The argument
2532 // null check technically happens in the wrong place, which can lead to
2533 // invalid stack traces when the primitive is inlined into a method
2534 // which handles NullPointerExceptions.
2535 _sp += nargs;
2536 do_null_check(receiver, T_OBJECT);
2537 _sp -= nargs;
2538 if (stopped()) {
2539 return true;
2540 }
2541 // Heap pointers get a null-check from the interpreter,
2542 // as a courtesy. However, this is not guaranteed by Unsafe,
2543 // and it is not possible to fully distinguish unintended nulls
2544 // from intended ones in this API.
2545
2546 if (is_volatile) {
2547 // We need to emit leading and trailing CPU membars (see below) in
2548 // addition to memory membars when is_volatile. This is a little
2549 // too strong, but avoids the need to insert per-alias-type
2550 // volatile membars (for stores; compare Parse::do_put_xxx), which
2551 // we cannot do effectively here because we probably only have a
2552 // rough approximation of type.
2553 need_mem_bar = true;
2554 // For Stores, place a memory ordering barrier now.
2555 if (is_store)
2556 insert_mem_bar(Op_MemBarRelease);
2557 }
2558
2559 // Memory barrier to prevent normal and 'unsafe' accesses from
2560 // bypassing each other. Happens after null checks, so the
2561 // exception paths do not take memory state from the memory barrier,
2562 // so there's no problems making a strong assert about mixing users
2563 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2564 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2565 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2566
2567 if (!is_store) {
2568 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2569 // load value and push onto stack
2570 switch (type) {
2571 case T_BOOLEAN:
2572 case T_CHAR:
2573 case T_BYTE:
2574 case T_SHORT:
2575 case T_INT:
2576 case T_FLOAT:
2577 push(p);
2578 break;
2579 case T_OBJECT:
2580 if (need_read_barrier) {
2581 insert_pre_barrier(heap_base_oop, offset, p, nargs, !(is_volatile || need_mem_bar));
2582 }
2583 push(p);
2584 break;
2585 case T_ADDRESS:
2586 // Cast to an int type.
2587 p = _gvn.transform( new (C) CastP2XNode(NULL,p) );
2588 p = ConvX2L(p);
2589 push_pair(p);
2590 break;
2591 case T_DOUBLE:
2592 case T_LONG:
2593 push_pair( p );
2594 break;
2595 default: ShouldNotReachHere();
2596 }
2597 } else {
2598 // place effect of store into memory
2599 switch (type) {
2600 case T_DOUBLE:
2601 val = dstore_rounding(val);
2602 break;
2603 case T_ADDRESS:
2604 // Repackage the long as a pointer.
2605 val = ConvL2X(val);
2606 val = _gvn.transform( new (C) CastX2PNode(val) );
2607 break;
2608 }
2609
2610 if (type != T_OBJECT ) {
2611 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2612 } else {
2613 // Possibly an oop being stored to Java heap or native memory
2614 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2615 // oop to Java heap.
2616 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2617 } else {
2618 // We can't tell at compile time if we are storing in the Java heap or outside
2619 // of it. So we need to emit code to conditionally do the proper type of
2620 // store.
2621
2622 IdealKit ideal(this);
2623 #define __ ideal.
2624 // QQQ who knows what probability is here??
2625 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2626 // Sync IdealKit and graphKit.
2627 sync_kit(ideal);
2628 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2629 // Update IdealKit memory.
2630 __ sync_kit(this);
2631 } __ else_(); {
2632 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2633 } __ end_if();
2634 // Final sync IdealKit and GraphKit.
2635 final_sync(ideal);
2636 #undef __
2637 }
2638 }
2639 }
2640
2641 if (is_volatile) {
2642 if (!is_store)
2643 insert_mem_bar(Op_MemBarAcquire);
2644 else
2645 insert_mem_bar(Op_MemBarVolatile);
2646 }
2647
2648 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2649
2650 return true;
2651 }
2652
2653 //----------------------------inline_unsafe_prefetch----------------------------
2654
2655 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2656 #ifndef PRODUCT
2657 {
2658 ResourceMark rm;
2659 // Check the signatures.
2660 ciSignature* sig = signature();
2661 #ifdef ASSERT
2662 // Object getObject(Object base, int/long offset), etc.
2663 BasicType rtype = sig->return_type()->basic_type();
2664 if (!is_native_ptr) {
2665 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2666 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2667 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2668 } else {
2669 assert(sig->count() == 1, "native prefetch has 1 argument");
2670 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2671 }
2672 #endif // ASSERT
2673 }
2674 #endif // !PRODUCT
2675
2676 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2677
2678 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2679 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2680
2681 debug_only(int saved_sp = _sp);
2682 _sp += nargs;
2683
2684 // Build address expression. See the code in inline_unsafe_access.
2685 Node *adr;
2686 if (!is_native_ptr) {
2687 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2688 Node* offset = pop_pair();
2689 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2690 Node* base = pop();
2691 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2692 // to be plain byte offsets, which are also the same as those accepted
2693 // by oopDesc::field_base.
2694 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2695 "fieldOffset must be byte-scaled");
2696 // 32-bit machines ignore the high half!
2697 offset = ConvL2X(offset);
2698 adr = make_unsafe_address(base, offset);
2699 } else {
2700 Node* ptr = pop_pair();
2701 // Adjust Java long to machine word:
2702 ptr = ConvL2X(ptr);
2703 adr = make_unsafe_address(NULL, ptr);
2704 }
2705
2706 if (is_static) {
2707 assert(saved_sp == _sp, "must have correct argument count");
2708 } else {
2709 // Pop receiver last: it was pushed first.
2710 Node *receiver = pop();
2711 assert(saved_sp == _sp, "must have correct argument count");
2712
2713 // Null check on self without removing any arguments. The argument
2714 // null check technically happens in the wrong place, which can lead to
2715 // invalid stack traces when the primitive is inlined into a method
2716 // which handles NullPointerExceptions.
2717 _sp += nargs;
2718 do_null_check(receiver, T_OBJECT);
2719 _sp -= nargs;
2720 if (stopped()) {
2721 return true;
2722 }
2723 }
2724
2725 // Generate the read or write prefetch
2726 Node *prefetch;
2727 if (is_store) {
2728 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2729 } else {
2730 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2731 }
2732 prefetch->init_req(0, control());
2733 set_i_o(_gvn.transform(prefetch));
2734
2735 return true;
2736 }
2737
2738 //----------------------------inline_unsafe_load_store----------------------------
2739
2740 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2741 // This basic scheme here is the same as inline_unsafe_access, but
2742 // differs in enough details that combining them would make the code
2743 // overly confusing. (This is a true fact! I originally combined
2744 // them, but even I was confused by it!) As much code/comments as
2745 // possible are retained from inline_unsafe_access though to make
2746 // the correspondences clearer. - dl
2747
2748 if (callee()->is_static()) return false; // caller must have the capability!
2749
2750 #ifndef PRODUCT
2751 BasicType rtype;
2752 {
2753 ResourceMark rm;
2754 ciSignature* sig = signature();
2755 rtype = sig->return_type()->basic_type();
2756 if (kind == LS_xadd || kind == LS_xchg) {
2757 // Check the signatures.
2758 #ifdef ASSERT
2759 assert(rtype == type, "get and set must return the expected type");
2760 assert(sig->count() == 3, "get and set has 3 arguments");
2761 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2762 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2763 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2764 #endif // ASSERT
2765 } else if (kind == LS_cmpxchg) {
2766 // Check the signatures.
2767 #ifdef ASSERT
2768 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2769 assert(sig->count() == 4, "CAS has 4 arguments");
2770 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2771 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2772 #endif // ASSERT
2773 } else {
2774 ShouldNotReachHere();
2775 }
2776 }
2777 #endif //PRODUCT
2778
2779 // number of stack slots per value argument (1 or 2)
2780 int type_words = type2size[type];
2781
2782 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2783
2784 // Argument words: "this" plus oop plus offset (plus oldvalue) plus newvalue/delta;
2785 int nargs = 1 + 1 + 2 + ((kind == LS_cmpxchg) ? type_words : 0) + type_words;
2786
2787 // pop arguments: newval, offset, base, and receiver
2788 debug_only(int saved_sp = _sp);
2789 _sp += nargs;
2790 Node* newval = (type_words == 1) ? pop() : pop_pair();
2791 Node* oldval = (kind == LS_cmpxchg) ? ((type_words == 1) ? pop() : pop_pair()) : NULL;
2792 Node *offset = pop_pair();
2793 Node *base = pop();
2794 Node *receiver = pop();
2795 assert(saved_sp == _sp, "must have correct argument count");
2796
2797 // Null check receiver.
2798 _sp += nargs;
2799 do_null_check(receiver, T_OBJECT);
2800 _sp -= nargs;
2801 if (stopped()) {
2802 return true;
2803 }
2804
2805 // Build field offset expression.
2806 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2807 // to be plain byte offsets, which are also the same as those accepted
2808 // by oopDesc::field_base.
2809 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2810 // 32-bit machines ignore the high half of long offsets
2811 offset = ConvL2X(offset);
2812 Node* adr = make_unsafe_address(base, offset);
2813 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2814
2815 // For CAS, unlike inline_unsafe_access, there seems no point in
2816 // trying to refine types. Just use the coarse types here.
2817 const Type *value_type = Type::get_const_basic_type(type);
2818 Compile::AliasType* alias_type = C->alias_type(adr_type);
2819 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2820
2821 if (kind == LS_xchg && type == T_OBJECT) {
2822 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2823 if (tjp != NULL) {
2824 value_type = tjp;
2825 }
2826 }
2827
2828 int alias_idx = C->get_alias_index(adr_type);
2829
2830 // Memory-model-wise, a LoadStore acts like a little synchronized
2831 // block, so needs barriers on each side. These don't translate
2832 // into actual barriers on most machines, but we still need rest of
2833 // compiler to respect ordering.
2834
2835 insert_mem_bar(Op_MemBarRelease);
2836 insert_mem_bar(Op_MemBarCPUOrder);
2837
2838 // 4984716: MemBars must be inserted before this
2839 // memory node in order to avoid a false
2840 // dependency which will confuse the scheduler.
2841 Node *mem = memory(alias_idx);
2842
2843 // For now, we handle only those cases that actually exist: ints,
2844 // longs, and Object. Adding others should be straightforward.
2845 Node* load_store;
2846 switch(type) {
2847 case T_INT:
2848 if (kind == LS_xadd) {
2849 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2850 } else if (kind == LS_xchg) {
2851 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2852 } else if (kind == LS_cmpxchg) {
2853 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2854 } else {
2855 ShouldNotReachHere();
2856 }
2857 break;
2858 case T_LONG:
2859 if (kind == LS_xadd) {
2860 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2861 } else if (kind == LS_xchg) {
2862 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2863 } else if (kind == LS_cmpxchg) {
2864 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2865 } else {
2866 ShouldNotReachHere();
2867 }
2868 break;
2869 case T_OBJECT:
2870 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2871 // could be delayed during Parse (for example, in adjust_map_after_if()).
2872 // Execute transformation here to avoid barrier generation in such case.
2873 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2874 newval = _gvn.makecon(TypePtr::NULL_PTR);
2875
2876 // Reference stores need a store barrier.
2877 pre_barrier(true /* do_load*/,
2878 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2879 NULL /* pre_val*/,
2880 T_OBJECT);
2881 #ifdef _LP64
2882 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2883 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2884 if (kind == LS_xchg) {
2885 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2886 newval_enc, adr_type, value_type->make_narrowoop()));
2887 } else {
2888 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2889 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2890 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2891 newval_enc, oldval_enc));
2892 }
2893 } else
2894 #endif
2895 {
2896 if (kind == LS_xchg) {
2897 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2898 } else {
2899 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2900 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2901 }
2902 }
2903 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2904 break;
2905 default:
2906 ShouldNotReachHere();
2907 break;
2908 }
2909
2910 // SCMemProjNodes represent the memory state of a LoadStore. Their
2911 // main role is to prevent LoadStore nodes from being optimized away
2912 // when their results aren't used.
2913 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
2914 set_memory(proj, alias_idx);
2915
2916 // Add the trailing membar surrounding the access
2917 insert_mem_bar(Op_MemBarCPUOrder);
2918 insert_mem_bar(Op_MemBarAcquire);
2919
2920 #ifdef _LP64
2921 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
2922 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->bottom_type()->make_ptr()));
2923 }
2924 #endif
2925
2926 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2927 push_node(load_store->bottom_type()->basic_type(), load_store);
2928 return true;
2929 }
2930
2931 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2932 // This is another variant of inline_unsafe_access, differing in
2933 // that it always issues store-store ("release") barrier and ensures
2934 // store-atomicity (which only matters for "long").
2935
2936 if (callee()->is_static()) return false; // caller must have the capability!
2937
2938 #ifndef PRODUCT
2939 {
2940 ResourceMark rm;
2941 // Check the signatures.
2942 ciSignature* sig = signature();
2943 #ifdef ASSERT
2944 BasicType rtype = sig->return_type()->basic_type();
2945 assert(rtype == T_VOID, "must return void");
2946 assert(sig->count() == 3, "has 3 arguments");
2947 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2948 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2949 #endif // ASSERT
2950 }
2951 #endif //PRODUCT
2952
2953 // number of stack slots per value argument (1 or 2)
2954 int type_words = type2size[type];
2955
2956 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2957
2958 // Argument words: "this" plus oop plus offset plus value;
2959 int nargs = 1 + 1 + 2 + type_words;
2960
2961 // pop arguments: val, offset, base, and receiver
2962 debug_only(int saved_sp = _sp);
2963 _sp += nargs;
2964 Node* val = (type_words == 1) ? pop() : pop_pair();
2965 Node *offset = pop_pair();
2966 Node *base = pop();
2967 Node *receiver = pop();
2968 assert(saved_sp == _sp, "must have correct argument count");
2969
2970 // Null check receiver.
2971 _sp += nargs;
2972 do_null_check(receiver, T_OBJECT);
2973 _sp -= nargs;
2974 if (stopped()) {
2975 return true;
2976 }
2977
2978 // Build field offset expression.
2979 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2980 // 32-bit machines ignore the high half of long offsets
2981 offset = ConvL2X(offset);
2982 Node* adr = make_unsafe_address(base, offset);
2983 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2984 const Type *value_type = Type::get_const_basic_type(type);
2985 Compile::AliasType* alias_type = C->alias_type(adr_type);
2986
2987 insert_mem_bar(Op_MemBarRelease);
2988 insert_mem_bar(Op_MemBarCPUOrder);
2989 // Ensure that the store is atomic for longs:
2990 bool require_atomic_access = true;
2991 Node* store;
2992 if (type == T_OBJECT) // reference stores need a store barrier.
2993 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2994 else {
2995 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2996 }
2997 insert_mem_bar(Op_MemBarCPUOrder);
2998 return true;
2999 }
3000
3001 bool LibraryCallKit::inline_unsafe_allocate() {
3002 if (callee()->is_static()) return false; // caller must have the capability!
3003 int nargs = 1 + 1;
3004 assert(signature()->size() == nargs-1, "alloc has 1 argument");
3005 null_check_receiver(callee()); // check then ignore argument(0)
3006 _sp += nargs; // set original stack for use by uncommon_trap
3007 Node* cls = do_null_check(argument(1), T_OBJECT);
3008 _sp -= nargs;
3009 if (stopped()) return true;
3010
3011 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
3012 _sp += nargs; // set original stack for use by uncommon_trap
3013 kls = do_null_check(kls, T_OBJECT);
3014 _sp -= nargs;
3015 if (stopped()) return true; // argument was like int.class
3016
3017 // Note: The argument might still be an illegal value like
3018 // Serializable.class or Object[].class. The runtime will handle it.
3019 // But we must make an explicit check for initialization.
3020 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3021 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3022 // can generate code to load it as unsigned byte.
3023 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
3024 Node* bits = intcon(InstanceKlass::fully_initialized);
3025 Node* test = _gvn.transform( new (C) SubINode(inst, bits) );
3026 // The 'test' is non-zero if we need to take a slow path.
3027
3028 Node* obj = new_instance(kls, test);
3029 push(obj);
3030
3031 return true;
3032 }
3033
3034 #ifdef TRACE_HAVE_INTRINSICS
3035 /*
3036 * oop -> myklass
3037 * myklass->trace_id |= USED
3038 * return myklass->trace_id & ~0x3
3039 */
3040 bool LibraryCallKit::inline_native_classID() {
3041 int nargs = 1 + 1;
3042 null_check_receiver(callee()); // check then ignore argument(0)
3043 _sp += nargs;
3044 Node* cls = do_null_check(argument(1), T_OBJECT);
3045 _sp -= nargs;
3046 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
3047 _sp += nargs;
3048 kls = do_null_check(kls, T_OBJECT);
3049 _sp -= nargs;
3050 ByteSize offset = TRACE_ID_OFFSET;
3051 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3052 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
3053 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3054 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3055 Node* clsused = longcon(0x01l); // set the class bit
3056 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3057
3058 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3059 store_to_memory(control(), insp, orl, T_LONG, adr_type);
3060 push_pair(andl);
3061 return true;
3062 }
3063
3064 bool LibraryCallKit::inline_native_threadID() {
3065 Node* tls_ptr = NULL;
3066 Node* cur_thr = generate_current_thread(tls_ptr);
3067 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3068 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3069 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3070
3071 Node* threadid = NULL;
3072 size_t thread_id_size = OSThread::thread_id_size();
3073 if (thread_id_size == (size_t) BytesPerLong) {
3074 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
3075 push(threadid);
3076 } else if (thread_id_size == (size_t) BytesPerInt) {
3077 threadid = make_load(control(), p, TypeInt::INT, T_INT);
3078 push(threadid);
3079 } else {
3080 ShouldNotReachHere();
3081 }
3082 return true;
3083 }
3084 #endif
3085
3086 //------------------------inline_native_time_funcs--------------
3087 // inline code for System.currentTimeMillis() and System.nanoTime()
3088 // these have the same type and signature
3089 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3090 const TypeFunc *tf = OptoRuntime::void_long_Type();
3091 const TypePtr* no_memory_effects = NULL;
3092 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3093 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3094 #ifdef ASSERT
3095 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms + 1));
3096 assert(value_top == top(), "second value must be top");
3097 #endif
3098 push_pair(value);
3099 return true;
3100 }
3101
3102 //------------------------inline_native_currentThread------------------
3103 bool LibraryCallKit::inline_native_currentThread() {
3104 Node* junk = NULL;
3105 push(generate_current_thread(junk));
3106 return true;
3107 }
3108
3109 //------------------------inline_native_isInterrupted------------------
3110 bool LibraryCallKit::inline_native_isInterrupted() {
3111 const int nargs = 1+1; // receiver + boolean
3112 assert(nargs == arg_size(), "sanity");
3113 // Add a fast path to t.isInterrupted(clear_int):
3114 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3115 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3116 // So, in the common case that the interrupt bit is false,
3117 // we avoid making a call into the VM. Even if the interrupt bit
3118 // is true, if the clear_int argument is false, we avoid the VM call.
3119 // However, if the receiver is not currentThread, we must call the VM,
3120 // because there must be some locking done around the operation.
3121
3122 // We only go to the fast case code if we pass two guards.
3123 // Paths which do not pass are accumulated in the slow_region.
3124 RegionNode* slow_region = new (C) RegionNode(1);
3125 record_for_igvn(slow_region);
3126 RegionNode* result_rgn = new (C) RegionNode(1+3); // fast1, fast2, slow
3127 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3128 enum { no_int_result_path = 1,
3129 no_clear_result_path = 2,
3130 slow_result_path = 3
3131 };
3132
3133 // (a) Receiving thread must be the current thread.
3134 Node* rec_thr = argument(0);
3135 Node* tls_ptr = NULL;
3136 Node* cur_thr = generate_current_thread(tls_ptr);
3137 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
3138 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
3139
3140 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
3141 if (!known_current_thread)
3142 generate_slow_guard(bol_thr, slow_region);
3143
3144 // (b) Interrupt bit on TLS must be false.
3145 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3146 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3147 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3148 // Set the control input on the field _interrupted read to prevent it floating up.
3149 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3150 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
3151 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
3152
3153 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3154
3155 // First fast path: if (!TLS._interrupted) return false;
3156 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
3157 result_rgn->init_req(no_int_result_path, false_bit);
3158 result_val->init_req(no_int_result_path, intcon(0));
3159
3160 // drop through to next case
3161 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
3162
3163 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3164 Node* clr_arg = argument(1);
3165 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
3166 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
3167 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3168
3169 // Second fast path: ... else if (!clear_int) return true;
3170 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3171 result_rgn->init_req(no_clear_result_path, false_arg);
3172 result_val->init_req(no_clear_result_path, intcon(1));
3173
3174 // drop through to next case
3175 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3176
3177 // (d) Otherwise, go to the slow path.
3178 slow_region->add_req(control());
3179 set_control( _gvn.transform(slow_region) );
3180
3181 if (stopped()) {
3182 // There is no slow path.
3183 result_rgn->init_req(slow_result_path, top());
3184 result_val->init_req(slow_result_path, top());
3185 } else {
3186 // non-virtual because it is a private non-static
3187 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3188
3189 Node* slow_val = set_results_for_java_call(slow_call);
3190 // this->control() comes from set_results_for_java_call
3191
3192 // If we know that the result of the slow call will be true, tell the optimizer!
3193 if (known_current_thread) slow_val = intcon(1);
3194
3195 Node* fast_io = slow_call->in(TypeFunc::I_O);
3196 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3197 // These two phis are pre-filled with copies of of the fast IO and Memory
3198 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3199 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3200
3201 result_rgn->init_req(slow_result_path, control());
3202 io_phi ->init_req(slow_result_path, i_o());
3203 mem_phi ->init_req(slow_result_path, reset_memory());
3204 result_val->init_req(slow_result_path, slow_val);
3205
3206 set_all_memory( _gvn.transform(mem_phi) );
3207 set_i_o( _gvn.transform(io_phi) );
3208 }
3209
3210 push_result(result_rgn, result_val);
3211 C->set_has_split_ifs(true); // Has chance for split-if optimization
3212
3213 return true;
3214 }
3215
3216 //---------------------------load_mirror_from_klass----------------------------
3217 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3218 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3219 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3220 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3221 }
3222
3223 //-----------------------load_klass_from_mirror_common-------------------------
3224 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3225 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3226 // and branch to the given path on the region.
3227 // If never_see_null, take an uncommon trap on null, so we can optimistically
3228 // compile for the non-null case.
3229 // If the region is NULL, force never_see_null = true.
3230 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3231 bool never_see_null,
3232 int nargs,
3233 RegionNode* region,
3234 int null_path,
3235 int offset) {
3236 if (region == NULL) never_see_null = true;
3237 Node* p = basic_plus_adr(mirror, offset);
3238 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3239 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3240 _sp += nargs; // any deopt will start just before call to enclosing method
3241 Node* null_ctl = top();
3242 kls = null_check_oop(kls, &null_ctl, never_see_null);
3243 if (region != NULL) {
3244 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3245 region->init_req(null_path, null_ctl);
3246 } else {
3247 assert(null_ctl == top(), "no loose ends");
3248 }
3249 _sp -= nargs;
3250 return kls;
3251 }
3252
3253 //--------------------(inline_native_Class_query helpers)---------------------
3254 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3255 // Fall through if (mods & mask) == bits, take the guard otherwise.
3256 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3257 // Branch around if the given klass has the given modifier bit set.
3258 // Like generate_guard, adds a new path onto the region.
3259 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3260 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3261 Node* mask = intcon(modifier_mask);
3262 Node* bits = intcon(modifier_bits);
3263 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3264 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3265 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3266 return generate_fair_guard(bol, region);
3267 }
3268 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3269 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3270 }
3271
3272 //-------------------------inline_native_Class_query-------------------
3273 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3274 int nargs = 1+0; // just the Class mirror, in most cases
3275 const Type* return_type = TypeInt::BOOL;
3276 Node* prim_return_value = top(); // what happens if it's a primitive class?
3277 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3278 bool expect_prim = false; // most of these guys expect to work on refs
3279
3280 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3281
3282 switch (id) {
3283 case vmIntrinsics::_isInstance:
3284 nargs = 1+1; // the Class mirror, plus the object getting queried about
3285 // nothing is an instance of a primitive type
3286 prim_return_value = intcon(0);
3287 break;
3288 case vmIntrinsics::_getModifiers:
3289 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3290 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3291 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3292 break;
3293 case vmIntrinsics::_isInterface:
3294 prim_return_value = intcon(0);
3295 break;
3296 case vmIntrinsics::_isArray:
3297 prim_return_value = intcon(0);
3298 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3299 break;
3300 case vmIntrinsics::_isPrimitive:
3301 prim_return_value = intcon(1);
3302 expect_prim = true; // obviously
3303 break;
3304 case vmIntrinsics::_getSuperclass:
3305 prim_return_value = null();
3306 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3307 break;
3308 case vmIntrinsics::_getComponentType:
3309 prim_return_value = null();
3310 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3311 break;
3312 case vmIntrinsics::_getClassAccessFlags:
3313 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3314 return_type = TypeInt::INT; // not bool! 6297094
3315 break;
3316 default:
3317 ShouldNotReachHere();
3318 }
3319
3320 Node* mirror = argument(0);
3321 Node* obj = (nargs <= 1)? top(): argument(1);
3322
3323 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3324 if (mirror_con == NULL) return false; // cannot happen?
3325
3326 #ifndef PRODUCT
3327 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3328 ciType* k = mirror_con->java_mirror_type();
3329 if (k) {
3330 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3331 k->print_name();
3332 tty->cr();
3333 }
3334 }
3335 #endif
3336
3337 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3338 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3339 record_for_igvn(region);
3340 PhiNode* phi = new (C) PhiNode(region, return_type);
3341
3342 // The mirror will never be null of Reflection.getClassAccessFlags, however
3343 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3344 // if it is. See bug 4774291.
3345
3346 // For Reflection.getClassAccessFlags(), the null check occurs in
3347 // the wrong place; see inline_unsafe_access(), above, for a similar
3348 // situation.
3349 _sp += nargs; // set original stack for use by uncommon_trap
3350 mirror = do_null_check(mirror, T_OBJECT);
3351 _sp -= nargs;
3352 // If mirror or obj is dead, only null-path is taken.
3353 if (stopped()) return true;
3354
3355 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3356
3357 // Now load the mirror's klass metaobject, and null-check it.
3358 // Side-effects region with the control path if the klass is null.
3359 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3360 region, _prim_path);
3361 // If kls is null, we have a primitive mirror.
3362 phi->init_req(_prim_path, prim_return_value);
3363 if (stopped()) { push_result(region, phi); return true; }
3364
3365 Node* p; // handy temp
3366 Node* null_ctl;
3367
3368 // Now that we have the non-null klass, we can perform the real query.
3369 // For constant classes, the query will constant-fold in LoadNode::Value.
3370 Node* query_value = top();
3371 switch (id) {
3372 case vmIntrinsics::_isInstance:
3373 // nothing is an instance of a primitive type
3374 _sp += nargs; // gen_instanceof might do an uncommon trap
3375 query_value = gen_instanceof(obj, kls);
3376 _sp -= nargs;
3377 break;
3378
3379 case vmIntrinsics::_getModifiers:
3380 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3381 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3382 break;
3383
3384 case vmIntrinsics::_isInterface:
3385 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3386 if (generate_interface_guard(kls, region) != NULL)
3387 // A guard was added. If the guard is taken, it was an interface.
3388 phi->add_req(intcon(1));
3389 // If we fall through, it's a plain class.
3390 query_value = intcon(0);
3391 break;
3392
3393 case vmIntrinsics::_isArray:
3394 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3395 if (generate_array_guard(kls, region) != NULL)
3396 // A guard was added. If the guard is taken, it was an array.
3397 phi->add_req(intcon(1));
3398 // If we fall through, it's a plain class.
3399 query_value = intcon(0);
3400 break;
3401
3402 case vmIntrinsics::_isPrimitive:
3403 query_value = intcon(0); // "normal" path produces false
3404 break;
3405
3406 case vmIntrinsics::_getSuperclass:
3407 // The rules here are somewhat unfortunate, but we can still do better
3408 // with random logic than with a JNI call.
3409 // Interfaces store null or Object as _super, but must report null.
3410 // Arrays store an intermediate super as _super, but must report Object.
3411 // Other types can report the actual _super.
3412 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3413 if (generate_interface_guard(kls, region) != NULL)
3414 // A guard was added. If the guard is taken, it was an interface.
3415 phi->add_req(null());
3416 if (generate_array_guard(kls, region) != NULL)
3417 // A guard was added. If the guard is taken, it was an array.
3418 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3419 // If we fall through, it's a plain class. Get its _super.
3420 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3421 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3422 null_ctl = top();
3423 kls = null_check_oop(kls, &null_ctl);
3424 if (null_ctl != top()) {
3425 // If the guard is taken, Object.superClass is null (both klass and mirror).
3426 region->add_req(null_ctl);
3427 phi ->add_req(null());
3428 }
3429 if (!stopped()) {
3430 query_value = load_mirror_from_klass(kls);
3431 }
3432 break;
3433
3434 case vmIntrinsics::_getComponentType:
3435 if (generate_array_guard(kls, region) != NULL) {
3436 // Be sure to pin the oop load to the guard edge just created:
3437 Node* is_array_ctrl = region->in(region->req()-1);
3438 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3439 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3440 phi->add_req(cmo);
3441 }
3442 query_value = null(); // non-array case is null
3443 break;
3444
3445 case vmIntrinsics::_getClassAccessFlags:
3446 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3447 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3448 break;
3449
3450 default:
3451 ShouldNotReachHere();
3452 }
3453
3454 // Fall-through is the normal case of a query to a real class.
3455 phi->init_req(1, query_value);
3456 region->init_req(1, control());
3457
3458 push_result(region, phi);
3459 C->set_has_split_ifs(true); // Has chance for split-if optimization
3460
3461 return true;
3462 }
3463
3464 //--------------------------inline_native_subtype_check------------------------
3465 // This intrinsic takes the JNI calls out of the heart of
3466 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3467 bool LibraryCallKit::inline_native_subtype_check() {
3468 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3469
3470 // Pull both arguments off the stack.
3471 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3472 args[0] = argument(0);
3473 args[1] = argument(1);
3474 Node* klasses[2]; // corresponding Klasses: superk, subk
3475 klasses[0] = klasses[1] = top();
3476
3477 enum {
3478 // A full decision tree on {superc is prim, subc is prim}:
3479 _prim_0_path = 1, // {P,N} => false
3480 // {P,P} & superc!=subc => false
3481 _prim_same_path, // {P,P} & superc==subc => true
3482 _prim_1_path, // {N,P} => false
3483 _ref_subtype_path, // {N,N} & subtype check wins => true
3484 _both_ref_path, // {N,N} & subtype check loses => false
3485 PATH_LIMIT
3486 };
3487
3488 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3489 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3490 record_for_igvn(region);
3491
3492 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3493 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3494 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3495
3496 // First null-check both mirrors and load each mirror's klass metaobject.
3497 int which_arg;
3498 for (which_arg = 0; which_arg <= 1; which_arg++) {
3499 Node* arg = args[which_arg];
3500 _sp += nargs; // set original stack for use by uncommon_trap
3501 arg = do_null_check(arg, T_OBJECT);
3502 _sp -= nargs;
3503 if (stopped()) break;
3504 args[which_arg] = _gvn.transform(arg);
3505
3506 Node* p = basic_plus_adr(arg, class_klass_offset);
3507 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3508 klasses[which_arg] = _gvn.transform(kls);
3509 }
3510
3511 // Having loaded both klasses, test each for null.
3512 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3513 for (which_arg = 0; which_arg <= 1; which_arg++) {
3514 Node* kls = klasses[which_arg];
3515 Node* null_ctl = top();
3516 _sp += nargs; // set original stack for use by uncommon_trap
3517 kls = null_check_oop(kls, &null_ctl, never_see_null);
3518 _sp -= nargs;
3519 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3520 region->init_req(prim_path, null_ctl);
3521 if (stopped()) break;
3522 klasses[which_arg] = kls;
3523 }
3524
3525 if (!stopped()) {
3526 // now we have two reference types, in klasses[0..1]
3527 Node* subk = klasses[1]; // the argument to isAssignableFrom
3528 Node* superk = klasses[0]; // the receiver
3529 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3530 // now we have a successful reference subtype check
3531 region->set_req(_ref_subtype_path, control());
3532 }
3533
3534 // If both operands are primitive (both klasses null), then
3535 // we must return true when they are identical primitives.
3536 // It is convenient to test this after the first null klass check.
3537 set_control(region->in(_prim_0_path)); // go back to first null check
3538 if (!stopped()) {
3539 // Since superc is primitive, make a guard for the superc==subc case.
3540 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3541 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3542 generate_guard(bol_eq, region, PROB_FAIR);
3543 if (region->req() == PATH_LIMIT+1) {
3544 // A guard was added. If the added guard is taken, superc==subc.
3545 region->swap_edges(PATH_LIMIT, _prim_same_path);
3546 region->del_req(PATH_LIMIT);
3547 }
3548 region->set_req(_prim_0_path, control()); // Not equal after all.
3549 }
3550
3551 // these are the only paths that produce 'true':
3552 phi->set_req(_prim_same_path, intcon(1));
3553 phi->set_req(_ref_subtype_path, intcon(1));
3554
3555 // pull together the cases:
3556 assert(region->req() == PATH_LIMIT, "sane region");
3557 for (uint i = 1; i < region->req(); i++) {
3558 Node* ctl = region->in(i);
3559 if (ctl == NULL || ctl == top()) {
3560 region->set_req(i, top());
3561 phi ->set_req(i, top());
3562 } else if (phi->in(i) == NULL) {
3563 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3564 }
3565 }
3566
3567 set_control(_gvn.transform(region));
3568 push(_gvn.transform(phi));
3569
3570 return true;
3571 }
3572
3573 //---------------------generate_array_guard_common------------------------
3574 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3575 bool obj_array, bool not_array) {
3576 // If obj_array/non_array==false/false:
3577 // Branch around if the given klass is in fact an array (either obj or prim).
3578 // If obj_array/non_array==false/true:
3579 // Branch around if the given klass is not an array klass of any kind.
3580 // If obj_array/non_array==true/true:
3581 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3582 // If obj_array/non_array==true/false:
3583 // Branch around if the kls is an oop array (Object[] or subtype)
3584 //
3585 // Like generate_guard, adds a new path onto the region.
3586 jint layout_con = 0;
3587 Node* layout_val = get_layout_helper(kls, layout_con);
3588 if (layout_val == NULL) {
3589 bool query = (obj_array
3590 ? Klass::layout_helper_is_objArray(layout_con)
3591 : Klass::layout_helper_is_array(layout_con));
3592 if (query == not_array) {
3593 return NULL; // never a branch
3594 } else { // always a branch
3595 Node* always_branch = control();
3596 if (region != NULL)
3597 region->add_req(always_branch);
3598 set_control(top());
3599 return always_branch;
3600 }
3601 }
3602 // Now test the correct condition.
3603 jint nval = (obj_array
3604 ? ((jint)Klass::_lh_array_tag_type_value
3605 << Klass::_lh_array_tag_shift)
3606 : Klass::_lh_neutral_value);
3607 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3608 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3609 // invert the test if we are looking for a non-array
3610 if (not_array) btest = BoolTest(btest).negate();
3611 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3612 return generate_fair_guard(bol, region);
3613 }
3614
3615
3616 //-----------------------inline_native_newArray--------------------------
3617 bool LibraryCallKit::inline_native_newArray() {
3618 int nargs = 2;
3619 Node* mirror = argument(0);
3620 Node* count_val = argument(1);
3621
3622 _sp += nargs; // set original stack for use by uncommon_trap
3623 mirror = do_null_check(mirror, T_OBJECT);
3624 _sp -= nargs;
3625 // If mirror or obj is dead, only null-path is taken.
3626 if (stopped()) return true;
3627
3628 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3629 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3630 PhiNode* result_val = new(C) PhiNode(result_reg,
3631 TypeInstPtr::NOTNULL);
3632 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3633 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3634 TypePtr::BOTTOM);
3635
3636 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3637 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3638 nargs,
3639 result_reg, _slow_path);
3640 Node* normal_ctl = control();
3641 Node* no_array_ctl = result_reg->in(_slow_path);
3642
3643 // Generate code for the slow case. We make a call to newArray().
3644 set_control(no_array_ctl);
3645 if (!stopped()) {
3646 // Either the input type is void.class, or else the
3647 // array klass has not yet been cached. Either the
3648 // ensuing call will throw an exception, or else it
3649 // will cache the array klass for next time.
3650 PreserveJVMState pjvms(this);
3651 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3652 Node* slow_result = set_results_for_java_call(slow_call);
3653 // this->control() comes from set_results_for_java_call
3654 result_reg->set_req(_slow_path, control());
3655 result_val->set_req(_slow_path, slow_result);
3656 result_io ->set_req(_slow_path, i_o());
3657 result_mem->set_req(_slow_path, reset_memory());
3658 }
3659
3660 set_control(normal_ctl);
3661 if (!stopped()) {
3662 // Normal case: The array type has been cached in the java.lang.Class.
3663 // The following call works fine even if the array type is polymorphic.
3664 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3665 Node* obj = new_array(klass_node, count_val, nargs);
3666 result_reg->init_req(_normal_path, control());
3667 result_val->init_req(_normal_path, obj);
3668 result_io ->init_req(_normal_path, i_o());
3669 result_mem->init_req(_normal_path, reset_memory());
3670 }
3671
3672 // Return the combined state.
3673 set_i_o( _gvn.transform(result_io) );
3674 set_all_memory( _gvn.transform(result_mem) );
3675 push_result(result_reg, result_val);
3676 C->set_has_split_ifs(true); // Has chance for split-if optimization
3677
3678 return true;
3679 }
3680
3681 //----------------------inline_native_getLength--------------------------
3682 bool LibraryCallKit::inline_native_getLength() {
3683 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3684
3685 int nargs = 1;
3686 Node* array = argument(0);
3687
3688 _sp += nargs; // set original stack for use by uncommon_trap
3689 array = do_null_check(array, T_OBJECT);
3690 _sp -= nargs;
3691
3692 // If array is dead, only null-path is taken.
3693 if (stopped()) return true;
3694
3695 // Deoptimize if it is a non-array.
3696 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3697
3698 if (non_array != NULL) {
3699 PreserveJVMState pjvms(this);
3700 set_control(non_array);
3701 _sp += nargs; // push the arguments back on the stack
3702 uncommon_trap(Deoptimization::Reason_intrinsic,
3703 Deoptimization::Action_maybe_recompile);
3704 }
3705
3706 // If control is dead, only non-array-path is taken.
3707 if (stopped()) return true;
3708
3709 // The works fine even if the array type is polymorphic.
3710 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3711 push( load_array_length(array) );
3712
3713 C->set_has_split_ifs(true); // Has chance for split-if optimization
3714
3715 return true;
3716 }
3717
3718 //------------------------inline_array_copyOf----------------------------
3719 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3720 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3721
3722 // Restore the stack and pop off the arguments.
3723 int nargs = 3 + (is_copyOfRange? 1: 0);
3724 Node* original = argument(0);
3725 Node* start = is_copyOfRange? argument(1): intcon(0);
3726 Node* end = is_copyOfRange? argument(2): argument(1);
3727 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3728
3729 Node* newcopy;
3730
3731 //set the original stack and the reexecute bit for the interpreter to reexecute
3732 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3733 { PreserveReexecuteState preexecs(this);
3734 _sp += nargs;
3735 jvms()->set_should_reexecute(true);
3736
3737 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3738 original = do_null_check(original, T_OBJECT);
3739
3740 // Check if a null path was taken unconditionally.
3741 if (stopped()) return true;
3742
3743 Node* orig_length = load_array_length(original);
3744
3745 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3746 NULL, 0);
3747 klass_node = do_null_check(klass_node, T_OBJECT);
3748
3749 RegionNode* bailout = new (C) RegionNode(1);
3750 record_for_igvn(bailout);
3751
3752 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3753 // Bail out if that is so.
3754 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3755 if (not_objArray != NULL) {
3756 // Improve the klass node's type from the new optimistic assumption:
3757 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3758 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3759 Node* cast = new (C) CastPPNode(klass_node, akls);
3760 cast->init_req(0, control());
3761 klass_node = _gvn.transform(cast);
3762 }
3763
3764 // Bail out if either start or end is negative.
3765 generate_negative_guard(start, bailout, &start);
3766 generate_negative_guard(end, bailout, &end);
3767
3768 Node* length = end;
3769 if (_gvn.type(start) != TypeInt::ZERO) {
3770 length = _gvn.transform( new (C) SubINode(end, start) );
3771 }
3772
3773 // Bail out if length is negative.
3774 // Without this the new_array would throw
3775 // NegativeArraySizeException but IllegalArgumentException is what
3776 // should be thrown
3777 generate_negative_guard(length, bailout, &length);
3778
3779 if (bailout->req() > 1) {
3780 PreserveJVMState pjvms(this);
3781 set_control( _gvn.transform(bailout) );
3782 uncommon_trap(Deoptimization::Reason_intrinsic,
3783 Deoptimization::Action_maybe_recompile);
3784 }
3785
3786 if (!stopped()) {
3787
3788 // How many elements will we copy from the original?
3789 // The answer is MinI(orig_length - start, length).
3790 Node* orig_tail = _gvn.transform( new(C) SubINode(orig_length, start) );
3791 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3792
3793 newcopy = new_array(klass_node, length, 0);
3794
3795 // Generate a direct call to the right arraycopy function(s).
3796 // We know the copy is disjoint but we might not know if the
3797 // oop stores need checking.
3798 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3799 // This will fail a store-check if x contains any non-nulls.
3800 bool disjoint_bases = true;
3801 // if start > orig_length then the length of the copy may be
3802 // negative.
3803 bool length_never_negative = !is_copyOfRange;
3804 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3805 original, start, newcopy, intcon(0), moved,
3806 disjoint_bases, length_never_negative);
3807 }
3808 } //original reexecute and sp are set back here
3809
3810 if(!stopped()) {
3811 push(newcopy);
3812 }
3813
3814 C->set_has_split_ifs(true); // Has chance for split-if optimization
3815
3816 return true;
3817 }
3818
3819
3820 //----------------------generate_virtual_guard---------------------------
3821 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3822 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3823 RegionNode* slow_region) {
3824 ciMethod* method = callee();
3825 int vtable_index = method->vtable_index();
3826 // Get the Method* out of the appropriate vtable entry.
3827 int entry_offset = (InstanceKlass::vtable_start_offset() +
3828 vtable_index*vtableEntry::size()) * wordSize +
3829 vtableEntry::method_offset_in_bytes();
3830 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3831 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3832
3833 // Compare the target method with the expected method (e.g., Object.hashCode).
3834 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3835
3836 Node* native_call = makecon(native_call_addr);
3837 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3838 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3839
3840 return generate_slow_guard(test_native, slow_region);
3841 }
3842
3843 //-----------------------generate_method_call----------------------------
3844 // Use generate_method_call to make a slow-call to the real
3845 // method if the fast path fails. An alternative would be to
3846 // use a stub like OptoRuntime::slow_arraycopy_Java.
3847 // This only works for expanding the current library call,
3848 // not another intrinsic. (E.g., don't use this for making an
3849 // arraycopy call inside of the copyOf intrinsic.)
3850 CallJavaNode*
3851 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3852 // When compiling the intrinsic method itself, do not use this technique.
3853 guarantee(callee() != C->method(), "cannot make slow-call to self");
3854
3855 ciMethod* method = callee();
3856 // ensure the JVMS we have will be correct for this call
3857 guarantee(method_id == method->intrinsic_id(), "must match");
3858
3859 const TypeFunc* tf = TypeFunc::make(method);
3860 CallJavaNode* slow_call;
3861 if (is_static) {
3862 assert(!is_virtual, "");
3863 slow_call = new(C) CallStaticJavaNode(tf,
3864 SharedRuntime::get_resolve_static_call_stub(),
3865 method, bci());
3866 } else if (is_virtual) {
3867 null_check_receiver(method);
3868 int vtable_index = Method::invalid_vtable_index;
3869 if (UseInlineCaches) {
3870 // Suppress the vtable call
3871 } else {
3872 // hashCode and clone are not a miranda methods,
3873 // so the vtable index is fixed.
3874 // No need to use the linkResolver to get it.
3875 vtable_index = method->vtable_index();
3876 }
3877 slow_call = new(C) CallDynamicJavaNode(tf,
3878 SharedRuntime::get_resolve_virtual_call_stub(),
3879 method, vtable_index, bci());
3880 } else { // neither virtual nor static: opt_virtual
3881 null_check_receiver(method);
3882 slow_call = new(C) CallStaticJavaNode(tf,
3883 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3884 method, bci());
3885 slow_call->set_optimized_virtual(true);
3886 }
3887 set_arguments_for_java_call(slow_call);
3888 set_edges_for_java_call(slow_call);
3889 return slow_call;
3890 }
3891
3892
3893 //------------------------------inline_native_hashcode--------------------
3894 // Build special case code for calls to hashCode on an object.
3895 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3896 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3897 assert(!(is_virtual && is_static), "either virtual, special, or static");
3898
3899 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3900
3901 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3902 PhiNode* result_val = new(C) PhiNode(result_reg,
3903 TypeInt::INT);
3904 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3905 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3906 TypePtr::BOTTOM);
3907 Node* obj = NULL;
3908 if (!is_static) {
3909 // Check for hashing null object
3910 obj = null_check_receiver(callee());
3911 if (stopped()) return true; // unconditionally null
3912 result_reg->init_req(_null_path, top());
3913 result_val->init_req(_null_path, top());
3914 } else {
3915 // Do a null check, and return zero if null.
3916 // System.identityHashCode(null) == 0
3917 obj = argument(0);
3918 Node* null_ctl = top();
3919 obj = null_check_oop(obj, &null_ctl);
3920 result_reg->init_req(_null_path, null_ctl);
3921 result_val->init_req(_null_path, _gvn.intcon(0));
3922 }
3923
3924 // Unconditionally null? Then return right away.
3925 if (stopped()) {
3926 set_control( result_reg->in(_null_path) );
3927 if (!stopped())
3928 push( result_val ->in(_null_path) );
3929 return true;
3930 }
3931
3932 // After null check, get the object's klass.
3933 Node* obj_klass = load_object_klass(obj);
3934
3935 // This call may be virtual (invokevirtual) or bound (invokespecial).
3936 // For each case we generate slightly different code.
3937
3938 // We only go to the fast case code if we pass a number of guards. The
3939 // paths which do not pass are accumulated in the slow_region.
3940 RegionNode* slow_region = new (C) RegionNode(1);
3941 record_for_igvn(slow_region);
3942
3943 // If this is a virtual call, we generate a funny guard. We pull out
3944 // the vtable entry corresponding to hashCode() from the target object.
3945 // If the target method which we are calling happens to be the native
3946 // Object hashCode() method, we pass the guard. We do not need this
3947 // guard for non-virtual calls -- the caller is known to be the native
3948 // Object hashCode().
3949 if (is_virtual) {
3950 generate_virtual_guard(obj_klass, slow_region);
3951 }
3952
3953 // Get the header out of the object, use LoadMarkNode when available
3954 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3955 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3956
3957 // Test the header to see if it is unlocked.
3958 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3959 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) );
3960 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3961 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val));
3962 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) );
3963
3964 generate_slow_guard(test_unlocked, slow_region);
3965
3966 // Get the hash value and check to see that it has been properly assigned.
3967 // We depend on hash_mask being at most 32 bits and avoid the use of
3968 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3969 // vm: see markOop.hpp.
3970 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3971 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3972 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) );
3973 // This hack lets the hash bits live anywhere in the mark object now, as long
3974 // as the shift drops the relevant bits into the low 32 bits. Note that
3975 // Java spec says that HashCode is an int so there's no point in capturing
3976 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3977 hshifted_header = ConvX2I(hshifted_header);
3978 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) );
3979
3980 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3981 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val));
3982 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) );
3983
3984 generate_slow_guard(test_assigned, slow_region);
3985
3986 Node* init_mem = reset_memory();
3987 // fill in the rest of the null path:
3988 result_io ->init_req(_null_path, i_o());
3989 result_mem->init_req(_null_path, init_mem);
3990
3991 result_val->init_req(_fast_path, hash_val);
3992 result_reg->init_req(_fast_path, control());
3993 result_io ->init_req(_fast_path, i_o());
3994 result_mem->init_req(_fast_path, init_mem);
3995
3996 // Generate code for the slow case. We make a call to hashCode().
3997 set_control(_gvn.transform(slow_region));
3998 if (!stopped()) {
3999 // No need for PreserveJVMState, because we're using up the present state.
4000 set_all_memory(init_mem);
4001 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
4002 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
4003 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4004 Node* slow_result = set_results_for_java_call(slow_call);
4005 // this->control() comes from set_results_for_java_call
4006 result_reg->init_req(_slow_path, control());
4007 result_val->init_req(_slow_path, slow_result);
4008 result_io ->set_req(_slow_path, i_o());
4009 result_mem ->set_req(_slow_path, reset_memory());
4010 }
4011
4012 // Return the combined state.
4013 set_i_o( _gvn.transform(result_io) );
4014 set_all_memory( _gvn.transform(result_mem) );
4015 push_result(result_reg, result_val);
4016
4017 return true;
4018 }
4019
4020 //---------------------------inline_native_getClass----------------------------
4021 // Build special case code for calls to getClass on an object.
4022 bool LibraryCallKit::inline_native_getClass() {
4023 Node* obj = null_check_receiver(callee());
4024 if (stopped()) return true;
4025 push( load_mirror_from_klass(load_object_klass(obj)) );
4026 return true;
4027 }
4028
4029 //-----------------inline_native_Reflection_getCallerClass---------------------
4030 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4031 //
4032 // NOTE that this code must perform the same logic as
4033 // vframeStream::security_get_caller_frame in that it must skip
4034 // Method.invoke() and auxiliary frames.
4035
4036
4037
4038
4039 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4040 ciMethod* method = callee();
4041
4042 #ifndef PRODUCT
4043 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4044 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4045 }
4046 #endif
4047
4048 debug_only(int saved_sp = _sp);
4049
4050 // Argument words: (int depth)
4051 int nargs = 1;
4052
4053 _sp += nargs;
4054 Node* caller_depth_node = pop();
4055
4056 assert(saved_sp == _sp, "must have correct argument count");
4057
4058 // The depth value must be a constant in order for the runtime call
4059 // to be eliminated.
4060 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
4061 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
4062 #ifndef PRODUCT
4063 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4064 tty->print_cr(" Bailing out because caller depth was not a constant");
4065 }
4066 #endif
4067 return false;
4068 }
4069 // Note that the JVM state at this point does not include the
4070 // getCallerClass() frame which we are trying to inline. The
4071 // semantics of getCallerClass(), however, are that the "first"
4072 // frame is the getCallerClass() frame, so we subtract one from the
4073 // requested depth before continuing. We don't inline requests of
4074 // getCallerClass(0).
4075 int caller_depth = caller_depth_type->get_con() - 1;
4076 if (caller_depth < 0) {
4077 #ifndef PRODUCT
4078 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4079 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
4080 }
4081 #endif
4082 return false;
4083 }
4084
4085 if (!jvms()->has_method()) {
4086 #ifndef PRODUCT
4087 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4088 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4089 }
4090 #endif
4091 return false;
4092 }
4093 int _depth = jvms()->depth(); // cache call chain depth
4094
4095 // Walk back up the JVM state to find the caller at the required
4096 // depth. NOTE that this code must perform the same logic as
4097 // vframeStream::security_get_caller_frame in that it must skip
4098 // Method.invoke() and auxiliary frames. Note also that depth is
4099 // 1-based (1 is the bottom of the inlining).
4100 int inlining_depth = _depth;
4101 JVMState* caller_jvms = NULL;
4102
4103 if (inlining_depth > 0) {
4104 caller_jvms = jvms();
4105 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
4106 do {
4107 // The following if-tests should be performed in this order
4108 if (is_method_invoke_or_aux_frame(caller_jvms)) {
4109 // Skip a Method.invoke() or auxiliary frame
4110 } else if (caller_depth > 0) {
4111 // Skip real frame
4112 --caller_depth;
4113 } else {
4114 // We're done: reached desired caller after skipping.
4115 break;
4116 }
4117 caller_jvms = caller_jvms->caller();
4118 --inlining_depth;
4119 } while (inlining_depth > 0);
4120 }
4121
4122 if (inlining_depth == 0) {
4123 #ifndef PRODUCT
4124 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4125 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
4126 tty->print_cr(" JVM state at this point:");
4127 for (int i = _depth; i >= 1; i--) {
4128 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
4129 }
4130 }
4131 #endif
4132 return false; // Reached end of inlining
4133 }
4134
4135 // Acquire method holder as java.lang.Class
4136 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4137 ciInstance* caller_mirror = caller_klass->java_mirror();
4138 // Push this as a constant
4139 push(makecon(TypeInstPtr::make(caller_mirror)));
4140 #ifndef PRODUCT
4141 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4142 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);
4143 tty->print_cr(" JVM state at this point:");
4144 for (int i = _depth; i >= 1; i--) {
4145 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
4146 }
4147 }
4148 #endif
4149 return true;
4150 }
4151
4152 // Helper routine for above
4153 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
4154 ciMethod* method = jvms->method();
4155
4156 // Is this the Method.invoke method itself?
4157 if (method->intrinsic_id() == vmIntrinsics::_invoke)
4158 return true;
4159
4160 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
4161 ciKlass* k = method->holder();
4162 if (k->is_instance_klass()) {
4163 ciInstanceKlass* ik = k->as_instance_klass();
4164 for (; ik != NULL; ik = ik->super()) {
4165 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
4166 ik == env()->find_system_klass(ik->name())) {
4167 return true;
4168 }
4169 }
4170 }
4171 else if (method->is_method_handle_intrinsic() ||
4172 method->is_compiled_lambda_form()) {
4173 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
4174 return true;
4175 }
4176
4177 return false;
4178 }
4179
4180 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4181 // restore the arguments
4182 _sp += arg_size();
4183
4184 switch (id) {
4185 case vmIntrinsics::_floatToRawIntBits:
4186 push(_gvn.transform( new (C) MoveF2INode(pop())));
4187 break;
4188
4189 case vmIntrinsics::_intBitsToFloat:
4190 push(_gvn.transform( new (C) MoveI2FNode(pop())));
4191 break;
4192
4193 case vmIntrinsics::_doubleToRawLongBits:
4194 push_pair(_gvn.transform( new (C) MoveD2LNode(pop_pair())));
4195 break;
4196
4197 case vmIntrinsics::_longBitsToDouble:
4198 push_pair(_gvn.transform( new (C) MoveL2DNode(pop_pair())));
4199 break;
4200
4201 case vmIntrinsics::_doubleToLongBits: {
4202 Node* value = pop_pair();
4203
4204 // two paths (plus control) merge in a wood
4205 RegionNode *r = new (C) RegionNode(3);
4206 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4207
4208 Node *cmpisnan = _gvn.transform( new (C) CmpDNode(value, value));
4209 // Build the boolean node
4210 Node *bolisnan = _gvn.transform( new (C) BoolNode( cmpisnan, BoolTest::ne ) );
4211
4212 // Branch either way.
4213 // NaN case is less traveled, which makes all the difference.
4214 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4215 Node *opt_isnan = _gvn.transform(ifisnan);
4216 assert( opt_isnan->is_If(), "Expect an IfNode");
4217 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4218 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4219
4220 set_control(iftrue);
4221
4222 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4223 Node *slow_result = longcon(nan_bits); // return NaN
4224 phi->init_req(1, _gvn.transform( slow_result ));
4225 r->init_req(1, iftrue);
4226
4227 // Else fall through
4228 Node *iffalse = _gvn.transform( new (C) IfFalseNode(opt_ifisnan) );
4229 set_control(iffalse);
4230
4231 phi->init_req(2, _gvn.transform( new (C) MoveD2LNode(value)));
4232 r->init_req(2, iffalse);
4233
4234 // Post merge
4235 set_control(_gvn.transform(r));
4236 record_for_igvn(r);
4237
4238 Node* result = _gvn.transform(phi);
4239 assert(result->bottom_type()->isa_long(), "must be");
4240 push_pair(result);
4241
4242 C->set_has_split_ifs(true); // Has chance for split-if optimization
4243
4244 break;
4245 }
4246
4247 case vmIntrinsics::_floatToIntBits: {
4248 Node* value = pop();
4249
4250 // two paths (plus control) merge in a wood
4251 RegionNode *r = new (C) RegionNode(3);
4252 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4253
4254 Node *cmpisnan = _gvn.transform( new (C) CmpFNode(value, value));
4255 // Build the boolean node
4256 Node *bolisnan = _gvn.transform( new (C) BoolNode( cmpisnan, BoolTest::ne ) );
4257
4258 // Branch either way.
4259 // NaN case is less traveled, which makes all the difference.
4260 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4261 Node *opt_isnan = _gvn.transform(ifisnan);
4262 assert( opt_isnan->is_If(), "Expect an IfNode");
4263 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4264 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4265
4266 set_control(iftrue);
4267
4268 static const jint nan_bits = 0x7fc00000;
4269 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4270 phi->init_req(1, _gvn.transform( slow_result ));
4271 r->init_req(1, iftrue);
4272
4273 // Else fall through
4274 Node *iffalse = _gvn.transform( new (C) IfFalseNode(opt_ifisnan) );
4275 set_control(iffalse);
4276
4277 phi->init_req(2, _gvn.transform( new (C) MoveF2INode(value)));
4278 r->init_req(2, iffalse);
4279
4280 // Post merge
4281 set_control(_gvn.transform(r));
4282 record_for_igvn(r);
4283
4284 Node* result = _gvn.transform(phi);
4285 assert(result->bottom_type()->isa_int(), "must be");
4286 push(result);
4287
4288 C->set_has_split_ifs(true); // Has chance for split-if optimization
4289
4290 break;
4291 }
4292
4293 default:
4294 ShouldNotReachHere();
4295 }
4296
4297 return true;
4298 }
4299
4300 #ifdef _LP64
4301 #define XTOP ,top() /*additional argument*/
4302 #else //_LP64
4303 #define XTOP /*no additional argument*/
4304 #endif //_LP64
4305
4306 //----------------------inline_unsafe_copyMemory-------------------------
4307 bool LibraryCallKit::inline_unsafe_copyMemory() {
4308 if (callee()->is_static()) return false; // caller must have the capability!
4309 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4310 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4311 null_check_receiver(callee()); // check then ignore argument(0)
4312 if (stopped()) return true;
4313
4314 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4315
4316 Node* src_ptr = argument(1);
4317 Node* src_off = ConvL2X(argument(2));
4318 assert(argument(3)->is_top(), "2nd half of long");
4319 Node* dst_ptr = argument(4);
4320 Node* dst_off = ConvL2X(argument(5));
4321 assert(argument(6)->is_top(), "2nd half of long");
4322 Node* size = ConvL2X(argument(7));
4323 assert(argument(8)->is_top(), "2nd half of long");
4324
4325 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4326 "fieldOffset must be byte-scaled");
4327
4328 Node* src = make_unsafe_address(src_ptr, src_off);
4329 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4330
4331 // Conservatively insert a memory barrier on all memory slices.
4332 // Do not let writes of the copy source or destination float below the copy.
4333 insert_mem_bar(Op_MemBarCPUOrder);
4334
4335 // Call it. Note that the length argument is not scaled.
4336 make_runtime_call(RC_LEAF|RC_NO_FP,
4337 OptoRuntime::fast_arraycopy_Type(),
4338 StubRoutines::unsafe_arraycopy(),
4339 "unsafe_arraycopy",
4340 TypeRawPtr::BOTTOM,
4341 src, dst, size XTOP);
4342
4343 // Do not let reads of the copy destination float above the copy.
4344 insert_mem_bar(Op_MemBarCPUOrder);
4345
4346 return true;
4347 }
4348
4349 //------------------------clone_coping-----------------------------------
4350 // Helper function for inline_native_clone.
4351 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4352 assert(obj_size != NULL, "");
4353 Node* raw_obj = alloc_obj->in(1);
4354 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4355
4356 AllocateNode* alloc = NULL;
4357 if (ReduceBulkZeroing) {
4358 // We will be completely responsible for initializing this object -
4359 // mark Initialize node as complete.
4360 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4361 // The object was just allocated - there should be no any stores!
4362 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4363 // Mark as complete_with_arraycopy so that on AllocateNode
4364 // expansion, we know this AllocateNode is initialized by an array
4365 // copy and a StoreStore barrier exists after the array copy.
4366 alloc->initialization()->set_complete_with_arraycopy();
4367 }
4368
4369 // Copy the fastest available way.
4370 // TODO: generate fields copies for small objects instead.
4371 Node* src = obj;
4372 Node* dest = alloc_obj;
4373 Node* size = _gvn.transform(obj_size);
4374
4375 // Exclude the header but include array length to copy by 8 bytes words.
4376 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4377 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4378 instanceOopDesc::base_offset_in_bytes();
4379 // base_off:
4380 // 8 - 32-bit VM
4381 // 12 - 64-bit VM, compressed klass
4382 // 16 - 64-bit VM, normal klass
4383 if (base_off % BytesPerLong != 0) {
4384 assert(UseCompressedOops, "");
4385 if (is_array) {
4386 // Exclude length to copy by 8 bytes words.
4387 base_off += sizeof(int);
4388 } else {
4389 // Include klass to copy by 8 bytes words.
4390 base_off = instanceOopDesc::klass_offset_in_bytes();
4391 }
4392 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4393 }
4394 src = basic_plus_adr(src, base_off);
4395 dest = basic_plus_adr(dest, base_off);
4396
4397 // Compute the length also, if needed:
4398 Node* countx = size;
4399 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4400 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4401
4402 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4403 bool disjoint_bases = true;
4404 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4405 src, NULL, dest, NULL, countx,
4406 /*dest_uninitialized*/true);
4407
4408 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4409 if (card_mark) {
4410 assert(!is_array, "");
4411 // Put in store barrier for any and all oops we are sticking
4412 // into this object. (We could avoid this if we could prove
4413 // that the object type contains no oop fields at all.)
4414 Node* no_particular_value = NULL;
4415 Node* no_particular_field = NULL;
4416 int raw_adr_idx = Compile::AliasIdxRaw;
4417 post_barrier(control(),
4418 memory(raw_adr_type),
4419 alloc_obj,
4420 no_particular_field,
4421 raw_adr_idx,
4422 no_particular_value,
4423 T_OBJECT,
4424 false);
4425 }
4426
4427 // Do not let reads from the cloned object float above the arraycopy.
4428 if (alloc != NULL) {
4429 // Do not let stores that initialize this object be reordered with
4430 // a subsequent store that would make this object accessible by
4431 // other threads.
4432 // Record what AllocateNode this StoreStore protects so that
4433 // escape analysis can go from the MemBarStoreStoreNode to the
4434 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4435 // based on the escape status of the AllocateNode.
4436 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4437 } else {
4438 insert_mem_bar(Op_MemBarCPUOrder);
4439 }
4440 }
4441
4442 //------------------------inline_native_clone----------------------------
4443 // Here are the simple edge cases:
4444 // null receiver => normal trap
4445 // virtual and clone was overridden => slow path to out-of-line clone
4446 // not cloneable or finalizer => slow path to out-of-line Object.clone
4447 //
4448 // The general case has two steps, allocation and copying.
4449 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4450 //
4451 // Copying also has two cases, oop arrays and everything else.
4452 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4453 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4454 //
4455 // These steps fold up nicely if and when the cloned object's klass
4456 // can be sharply typed as an object array, a type array, or an instance.
4457 //
4458 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4459 int nargs = 1;
4460 PhiNode* result_val;
4461
4462 //set the original stack and the reexecute bit for the interpreter to reexecute
4463 //the bytecode that invokes Object.clone if deoptimization happens
4464 { PreserveReexecuteState preexecs(this);
4465 jvms()->set_should_reexecute(true);
4466
4467 //null_check_receiver will adjust _sp (push and pop)
4468 Node* obj = null_check_receiver(callee());
4469 if (stopped()) return true;
4470
4471 _sp += nargs;
4472
4473 Node* obj_klass = load_object_klass(obj);
4474 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4475 const TypeOopPtr* toop = ((tklass != NULL)
4476 ? tklass->as_instance_type()
4477 : TypeInstPtr::NOTNULL);
4478
4479 // Conservatively insert a memory barrier on all memory slices.
4480 // Do not let writes into the original float below the clone.
4481 insert_mem_bar(Op_MemBarCPUOrder);
4482
4483 // paths into result_reg:
4484 enum {
4485 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4486 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4487 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4488 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4489 PATH_LIMIT
4490 };
4491 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4492 result_val = new(C) PhiNode(result_reg,
4493 TypeInstPtr::NOTNULL);
4494 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4495 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4496 TypePtr::BOTTOM);
4497 record_for_igvn(result_reg);
4498
4499 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4500 int raw_adr_idx = Compile::AliasIdxRaw;
4501
4502 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4503 if (array_ctl != NULL) {
4504 // It's an array.
4505 PreserveJVMState pjvms(this);
4506 set_control(array_ctl);
4507 Node* obj_length = load_array_length(obj);
4508 Node* obj_size = NULL;
4509 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);
4510
4511 if (!use_ReduceInitialCardMarks()) {
4512 // If it is an oop array, it requires very special treatment,
4513 // because card marking is required on each card of the array.
4514 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4515 if (is_obja != NULL) {
4516 PreserveJVMState pjvms2(this);
4517 set_control(is_obja);
4518 // Generate a direct call to the right arraycopy function(s).
4519 bool disjoint_bases = true;
4520 bool length_never_negative = true;
4521 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4522 obj, intcon(0), alloc_obj, intcon(0),
4523 obj_length,
4524 disjoint_bases, length_never_negative);
4525 result_reg->init_req(_objArray_path, control());
4526 result_val->init_req(_objArray_path, alloc_obj);
4527 result_i_o ->set_req(_objArray_path, i_o());
4528 result_mem ->set_req(_objArray_path, reset_memory());
4529 }
4530 }
4531 // Otherwise, there are no card marks to worry about.
4532 // (We can dispense with card marks if we know the allocation
4533 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4534 // causes the non-eden paths to take compensating steps to
4535 // simulate a fresh allocation, so that no further
4536 // card marks are required in compiled code to initialize
4537 // the object.)
4538
4539 if (!stopped()) {
4540 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4541
4542 // Present the results of the copy.
4543 result_reg->init_req(_array_path, control());
4544 result_val->init_req(_array_path, alloc_obj);
4545 result_i_o ->set_req(_array_path, i_o());
4546 result_mem ->set_req(_array_path, reset_memory());
4547 }
4548 }
4549
4550 // We only go to the instance fast case code if we pass a number of guards.
4551 // The paths which do not pass are accumulated in the slow_region.
4552 RegionNode* slow_region = new (C) RegionNode(1);
4553 record_for_igvn(slow_region);
4554 if (!stopped()) {
4555 // It's an instance (we did array above). Make the slow-path tests.
4556 // If this is a virtual call, we generate a funny guard. We grab
4557 // the vtable entry corresponding to clone() from the target object.
4558 // If the target method which we are calling happens to be the
4559 // Object clone() method, we pass the guard. We do not need this
4560 // guard for non-virtual calls; the caller is known to be the native
4561 // Object clone().
4562 if (is_virtual) {
4563 generate_virtual_guard(obj_klass, slow_region);
4564 }
4565
4566 // The object must be cloneable and must not have a finalizer.
4567 // Both of these conditions may be checked in a single test.
4568 // We could optimize the cloneable test further, but we don't care.
4569 generate_access_flags_guard(obj_klass,
4570 // Test both conditions:
4571 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4572 // Must be cloneable but not finalizer:
4573 JVM_ACC_IS_CLONEABLE,
4574 slow_region);
4575 }
4576
4577 if (!stopped()) {
4578 // It's an instance, and it passed the slow-path tests.
4579 PreserveJVMState pjvms(this);
4580 Node* obj_size = NULL;
4581 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4582
4583 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4584
4585 // Present the results of the slow call.
4586 result_reg->init_req(_instance_path, control());
4587 result_val->init_req(_instance_path, alloc_obj);
4588 result_i_o ->set_req(_instance_path, i_o());
4589 result_mem ->set_req(_instance_path, reset_memory());
4590 }
4591
4592 // Generate code for the slow case. We make a call to clone().
4593 set_control(_gvn.transform(slow_region));
4594 if (!stopped()) {
4595 PreserveJVMState pjvms(this);
4596 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4597 Node* slow_result = set_results_for_java_call(slow_call);
4598 // this->control() comes from set_results_for_java_call
4599 result_reg->init_req(_slow_path, control());
4600 result_val->init_req(_slow_path, slow_result);
4601 result_i_o ->set_req(_slow_path, i_o());
4602 result_mem ->set_req(_slow_path, reset_memory());
4603 }
4604
4605 // Return the combined state.
4606 set_control( _gvn.transform(result_reg) );
4607 set_i_o( _gvn.transform(result_i_o) );
4608 set_all_memory( _gvn.transform(result_mem) );
4609 } //original reexecute and sp are set back here
4610
4611 push(_gvn.transform(result_val));
4612
4613 return true;
4614 }
4615
4616 //------------------------------basictype2arraycopy----------------------------
4617 address LibraryCallKit::basictype2arraycopy(BasicType t,
4618 Node* src_offset,
4619 Node* dest_offset,
4620 bool disjoint_bases,
4621 const char* &name,
4622 bool dest_uninitialized) {
4623 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4624 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4625
4626 bool aligned = false;
4627 bool disjoint = disjoint_bases;
4628
4629 // if the offsets are the same, we can treat the memory regions as
4630 // disjoint, because either the memory regions are in different arrays,
4631 // or they are identical (which we can treat as disjoint.) We can also
4632 // treat a copy with a destination index less that the source index
4633 // as disjoint since a low->high copy will work correctly in this case.
4634 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4635 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4636 // both indices are constants
4637 int s_offs = src_offset_inttype->get_con();
4638 int d_offs = dest_offset_inttype->get_con();
4639 int element_size = type2aelembytes(t);
4640 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4641 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4642 if (s_offs >= d_offs) disjoint = true;
4643 } else if (src_offset == dest_offset && src_offset != NULL) {
4644 // This can occur if the offsets are identical non-constants.
4645 disjoint = true;
4646 }
4647
4648 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4649 }
4650
4651
4652 //------------------------------inline_arraycopy-----------------------
4653 bool LibraryCallKit::inline_arraycopy() {
4654 // Restore the stack and pop off the arguments.
4655 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4656 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4657
4658 Node *src = argument(0);
4659 Node *src_offset = argument(1);
4660 Node *dest = argument(2);
4661 Node *dest_offset = argument(3);
4662 Node *length = argument(4);
4663
4664 // Compile time checks. If any of these checks cannot be verified at compile time,
4665 // we do not make a fast path for this call. Instead, we let the call remain as it
4666 // is. The checks we choose to mandate at compile time are:
4667 //
4668 // (1) src and dest are arrays.
4669 const Type* src_type = src->Value(&_gvn);
4670 const Type* dest_type = dest->Value(&_gvn);
4671 const TypeAryPtr* top_src = src_type->isa_aryptr();
4672 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4673 if (top_src == NULL || top_src->klass() == NULL ||
4674 top_dest == NULL || top_dest->klass() == NULL) {
4675 // Conservatively insert a memory barrier on all memory slices.
4676 // Do not let writes into the source float below the arraycopy.
4677 insert_mem_bar(Op_MemBarCPUOrder);
4678
4679 // Call StubRoutines::generic_arraycopy stub.
4680 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4681 src, src_offset, dest, dest_offset, length);
4682
4683 // Do not let reads from the destination float above the arraycopy.
4684 // Since we cannot type the arrays, we don't know which slices
4685 // might be affected. We could restrict this barrier only to those
4686 // memory slices which pertain to array elements--but don't bother.
4687 if (!InsertMemBarAfterArraycopy)
4688 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4689 insert_mem_bar(Op_MemBarCPUOrder);
4690 return true;
4691 }
4692
4693 // (2) src and dest arrays must have elements of the same BasicType
4694 // Figure out the size and type of the elements we will be copying.
4695 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4696 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4697 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4698 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4699
4700 if (src_elem != dest_elem || dest_elem == T_VOID) {
4701 // The component types are not the same or are not recognized. Punt.
4702 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4703 generate_slow_arraycopy(TypePtr::BOTTOM,
4704 src, src_offset, dest, dest_offset, length,
4705 /*dest_uninitialized*/false);
4706 return true;
4707 }
4708
4709 //---------------------------------------------------------------------------
4710 // We will make a fast path for this call to arraycopy.
4711
4712 // We have the following tests left to perform:
4713 //
4714 // (3) src and dest must not be null.
4715 // (4) src_offset must not be negative.
4716 // (5) dest_offset must not be negative.
4717 // (6) length must not be negative.
4718 // (7) src_offset + length must not exceed length of src.
4719 // (8) dest_offset + length must not exceed length of dest.
4720 // (9) each element of an oop array must be assignable
4721
4722 RegionNode* slow_region = new (C) RegionNode(1);
4723 record_for_igvn(slow_region);
4724
4725 // (3) operands must not be null
4726 // We currently perform our null checks with the do_null_check routine.
4727 // This means that the null exceptions will be reported in the caller
4728 // rather than (correctly) reported inside of the native arraycopy call.
4729 // This should be corrected, given time. We do our null check with the
4730 // stack pointer restored.
4731 _sp += nargs;
4732 src = do_null_check(src, T_ARRAY);
4733 dest = do_null_check(dest, T_ARRAY);
4734 _sp -= nargs;
4735
4736 // (4) src_offset must not be negative.
4737 generate_negative_guard(src_offset, slow_region);
4738
4739 // (5) dest_offset must not be negative.
4740 generate_negative_guard(dest_offset, slow_region);
4741
4742 // (6) length must not be negative (moved to generate_arraycopy()).
4743 // generate_negative_guard(length, slow_region);
4744
4745 // (7) src_offset + length must not exceed length of src.
4746 generate_limit_guard(src_offset, length,
4747 load_array_length(src),
4748 slow_region);
4749
4750 // (8) dest_offset + length must not exceed length of dest.
4751 generate_limit_guard(dest_offset, length,
4752 load_array_length(dest),
4753 slow_region);
4754
4755 // (9) each element of an oop array must be assignable
4756 // The generate_arraycopy subroutine checks this.
4757
4758 // This is where the memory effects are placed:
4759 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4760 generate_arraycopy(adr_type, dest_elem,
4761 src, src_offset, dest, dest_offset, length,
4762 false, false, slow_region);
4763
4764 return true;
4765 }
4766
4767 //-----------------------------generate_arraycopy----------------------
4768 // Generate an optimized call to arraycopy.
4769 // Caller must guard against non-arrays.
4770 // Caller must determine a common array basic-type for both arrays.
4771 // Caller must validate offsets against array bounds.
4772 // The slow_region has already collected guard failure paths
4773 // (such as out of bounds length or non-conformable array types).
4774 // The generated code has this shape, in general:
4775 //
4776 // if (length == 0) return // via zero_path
4777 // slowval = -1
4778 // if (types unknown) {
4779 // slowval = call generic copy loop
4780 // if (slowval == 0) return // via checked_path
4781 // } else if (indexes in bounds) {
4782 // if ((is object array) && !(array type check)) {
4783 // slowval = call checked copy loop
4784 // if (slowval == 0) return // via checked_path
4785 // } else {
4786 // call bulk copy loop
4787 // return // via fast_path
4788 // }
4789 // }
4790 // // adjust params for remaining work:
4791 // if (slowval != -1) {
4792 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4793 // }
4794 // slow_region:
4795 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4796 // return // via slow_call_path
4797 //
4798 // This routine is used from several intrinsics: System.arraycopy,
4799 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4800 //
4801 void
4802 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4803 BasicType basic_elem_type,
4804 Node* src, Node* src_offset,
4805 Node* dest, Node* dest_offset,
4806 Node* copy_length,
4807 bool disjoint_bases,
4808 bool length_never_negative,
4809 RegionNode* slow_region) {
4810
4811 if (slow_region == NULL) {
4812 slow_region = new(C) RegionNode(1);
4813 record_for_igvn(slow_region);
4814 }
4815
4816 Node* original_dest = dest;
4817 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4818 bool dest_uninitialized = false;
4819
4820 // See if this is the initialization of a newly-allocated array.
4821 // If so, we will take responsibility here for initializing it to zero.
4822 // (Note: Because tightly_coupled_allocation performs checks on the
4823 // out-edges of the dest, we need to avoid making derived pointers
4824 // from it until we have checked its uses.)
4825 if (ReduceBulkZeroing
4826 && !ZeroTLAB // pointless if already zeroed
4827 && basic_elem_type != T_CONFLICT // avoid corner case
4828 && !src->eqv_uncast(dest)
4829 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4830 != NULL)
4831 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4832 && alloc->maybe_set_complete(&_gvn)) {
4833 // "You break it, you buy it."
4834 InitializeNode* init = alloc->initialization();
4835 assert(init->is_complete(), "we just did this");
4836 init->set_complete_with_arraycopy();
4837 assert(dest->is_CheckCastPP(), "sanity");
4838 assert(dest->in(0)->in(0) == init, "dest pinned");
4839 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4840 // From this point on, every exit path is responsible for
4841 // initializing any non-copied parts of the object to zero.
4842 // Also, if this flag is set we make sure that arraycopy interacts properly
4843 // with G1, eliding pre-barriers. See CR 6627983.
4844 dest_uninitialized = true;
4845 } else {
4846 // No zeroing elimination here.
4847 alloc = NULL;
4848 //original_dest = dest;
4849 //dest_uninitialized = false;
4850 }
4851
4852 // Results are placed here:
4853 enum { fast_path = 1, // normal void-returning assembly stub
4854 checked_path = 2, // special assembly stub with cleanup
4855 slow_call_path = 3, // something went wrong; call the VM
4856 zero_path = 4, // bypass when length of copy is zero
4857 bcopy_path = 5, // copy primitive array by 64-bit blocks
4858 PATH_LIMIT = 6
4859 };
4860 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4861 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4862 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4863 record_for_igvn(result_region);
4864 _gvn.set_type_bottom(result_i_o);
4865 _gvn.set_type_bottom(result_memory);
4866 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4867
4868 // The slow_control path:
4869 Node* slow_control;
4870 Node* slow_i_o = i_o();
4871 Node* slow_mem = memory(adr_type);
4872 debug_only(slow_control = (Node*) badAddress);
4873
4874 // Checked control path:
4875 Node* checked_control = top();
4876 Node* checked_mem = NULL;
4877 Node* checked_i_o = NULL;
4878 Node* checked_value = NULL;
4879
4880 if (basic_elem_type == T_CONFLICT) {
4881 assert(!dest_uninitialized, "");
4882 Node* cv = generate_generic_arraycopy(adr_type,
4883 src, src_offset, dest, dest_offset,
4884 copy_length, dest_uninitialized);
4885 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4886 checked_control = control();
4887 checked_i_o = i_o();
4888 checked_mem = memory(adr_type);
4889 checked_value = cv;
4890 set_control(top()); // no fast path
4891 }
4892
4893 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4894 if (not_pos != NULL) {
4895 PreserveJVMState pjvms(this);
4896 set_control(not_pos);
4897
4898 // (6) length must not be negative.
4899 if (!length_never_negative) {
4900 generate_negative_guard(copy_length, slow_region);
4901 }
4902
4903 // copy_length is 0.
4904 if (!stopped() && dest_uninitialized) {
4905 Node* dest_length = alloc->in(AllocateNode::ALength);
4906 if (copy_length->eqv_uncast(dest_length)
4907 || _gvn.find_int_con(dest_length, 1) <= 0) {
4908 // There is no zeroing to do. No need for a secondary raw memory barrier.
4909 } else {
4910 // Clear the whole thing since there are no source elements to copy.
4911 generate_clear_array(adr_type, dest, basic_elem_type,
4912 intcon(0), NULL,
4913 alloc->in(AllocateNode::AllocSize));
4914 // Use a secondary InitializeNode as raw memory barrier.
4915 // Currently it is needed only on this path since other
4916 // paths have stub or runtime calls as raw memory barriers.
4917 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4918 Compile::AliasIdxRaw,
4919 top())->as_Initialize();
4920 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4921 }
4922 }
4923
4924 // Present the results of the fast call.
4925 result_region->init_req(zero_path, control());
4926 result_i_o ->init_req(zero_path, i_o());
4927 result_memory->init_req(zero_path, memory(adr_type));
4928 }
4929
4930 if (!stopped() && dest_uninitialized) {
4931 // We have to initialize the *uncopied* part of the array to zero.
4932 // The copy destination is the slice dest[off..off+len]. The other slices
4933 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4934 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4935 Node* dest_length = alloc->in(AllocateNode::ALength);
4936 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
4937 copy_length) );
4938
4939 // If there is a head section that needs zeroing, do it now.
4940 if (find_int_con(dest_offset, -1) != 0) {
4941 generate_clear_array(adr_type, dest, basic_elem_type,
4942 intcon(0), dest_offset,
4943 NULL);
4944 }
4945
4946 // Next, perform a dynamic check on the tail length.
4947 // It is often zero, and we can win big if we prove this.
4948 // There are two wins: Avoid generating the ClearArray
4949 // with its attendant messy index arithmetic, and upgrade
4950 // the copy to a more hardware-friendly word size of 64 bits.
4951 Node* tail_ctl = NULL;
4952 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4953 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
4954 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
4955 tail_ctl = generate_slow_guard(bol_lt, NULL);
4956 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4957 }
4958
4959 // At this point, let's assume there is no tail.
4960 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4961 // There is no tail. Try an upgrade to a 64-bit copy.
4962 bool didit = false;
4963 { PreserveJVMState pjvms(this);
4964 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4965 src, src_offset, dest, dest_offset,
4966 dest_size, dest_uninitialized);
4967 if (didit) {
4968 // Present the results of the block-copying fast call.
4969 result_region->init_req(bcopy_path, control());
4970 result_i_o ->init_req(bcopy_path, i_o());
4971 result_memory->init_req(bcopy_path, memory(adr_type));
4972 }
4973 }
4974 if (didit)
4975 set_control(top()); // no regular fast path
4976 }
4977
4978 // Clear the tail, if any.
4979 if (tail_ctl != NULL) {
4980 Node* notail_ctl = stopped() ? NULL : control();
4981 set_control(tail_ctl);
4982 if (notail_ctl == NULL) {
4983 generate_clear_array(adr_type, dest, basic_elem_type,
4984 dest_tail, NULL,
4985 dest_size);
4986 } else {
4987 // Make a local merge.
4988 Node* done_ctl = new(C) RegionNode(3);
4989 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4990 done_ctl->init_req(1, notail_ctl);
4991 done_mem->init_req(1, memory(adr_type));
4992 generate_clear_array(adr_type, dest, basic_elem_type,
4993 dest_tail, NULL,
4994 dest_size);
4995 done_ctl->init_req(2, control());
4996 done_mem->init_req(2, memory(adr_type));
4997 set_control( _gvn.transform(done_ctl) );
4998 set_memory( _gvn.transform(done_mem), adr_type );
4999 }
5000 }
5001 }
5002
5003 BasicType copy_type = basic_elem_type;
5004 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5005 if (!stopped() && copy_type == T_OBJECT) {
5006 // If src and dest have compatible element types, we can copy bits.
5007 // Types S[] and D[] are compatible if D is a supertype of S.
5008 //
5009 // If they are not, we will use checked_oop_disjoint_arraycopy,
5010 // which performs a fast optimistic per-oop check, and backs off
5011 // further to JVM_ArrayCopy on the first per-oop check that fails.
5012 // (Actually, we don't move raw bits only; the GC requires card marks.)
5013
5014 // Get the Klass* for both src and dest
5015 Node* src_klass = load_object_klass(src);
5016 Node* dest_klass = load_object_klass(dest);
5017
5018 // Generate the subtype check.
5019 // This might fold up statically, or then again it might not.
5020 //
5021 // Non-static example: Copying List<String>.elements to a new String[].
5022 // The backing store for a List<String> is always an Object[],
5023 // but its elements are always type String, if the generic types
5024 // are correct at the source level.
5025 //
5026 // Test S[] against D[], not S against D, because (probably)
5027 // the secondary supertype cache is less busy for S[] than S.
5028 // This usually only matters when D is an interface.
5029 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5030 // Plug failing path into checked_oop_disjoint_arraycopy
5031 if (not_subtype_ctrl != top()) {
5032 PreserveJVMState pjvms(this);
5033 set_control(not_subtype_ctrl);
5034 // (At this point we can assume disjoint_bases, since types differ.)
5035 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
5036 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5037 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5038 Node* dest_elem_klass = _gvn.transform(n1);
5039 Node* cv = generate_checkcast_arraycopy(adr_type,
5040 dest_elem_klass,
5041 src, src_offset, dest, dest_offset,
5042 ConvI2X(copy_length), dest_uninitialized);
5043 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5044 checked_control = control();
5045 checked_i_o = i_o();
5046 checked_mem = memory(adr_type);
5047 checked_value = cv;
5048 }
5049 // At this point we know we do not need type checks on oop stores.
5050
5051 // Let's see if we need card marks:
5052 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5053 // If we do not need card marks, copy using the jint or jlong stub.
5054 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5055 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5056 "sizes agree");
5057 }
5058 }
5059
5060 if (!stopped()) {
5061 // Generate the fast path, if possible.
5062 PreserveJVMState pjvms(this);
5063 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5064 src, src_offset, dest, dest_offset,
5065 ConvI2X(copy_length), dest_uninitialized);
5066
5067 // Present the results of the fast call.
5068 result_region->init_req(fast_path, control());
5069 result_i_o ->init_req(fast_path, i_o());
5070 result_memory->init_req(fast_path, memory(adr_type));
5071 }
5072
5073 // Here are all the slow paths up to this point, in one bundle:
5074 slow_control = top();
5075 if (slow_region != NULL)
5076 slow_control = _gvn.transform(slow_region);
5077 debug_only(slow_region = (RegionNode*)badAddress);
5078
5079 set_control(checked_control);
5080 if (!stopped()) {
5081 // Clean up after the checked call.
5082 // The returned value is either 0 or -1^K,
5083 // where K = number of partially transferred array elements.
5084 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
5085 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
5086 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5087
5088 // If it is 0, we are done, so transfer to the end.
5089 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
5090 result_region->init_req(checked_path, checks_done);
5091 result_i_o ->init_req(checked_path, checked_i_o);
5092 result_memory->init_req(checked_path, checked_mem);
5093
5094 // If it is not zero, merge into the slow call.
5095 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
5096 RegionNode* slow_reg2 = new(C) RegionNode(3);
5097 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5098 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5099 record_for_igvn(slow_reg2);
5100 slow_reg2 ->init_req(1, slow_control);
5101 slow_i_o2 ->init_req(1, slow_i_o);
5102 slow_mem2 ->init_req(1, slow_mem);
5103 slow_reg2 ->init_req(2, control());
5104 slow_i_o2 ->init_req(2, checked_i_o);
5105 slow_mem2 ->init_req(2, checked_mem);
5106
5107 slow_control = _gvn.transform(slow_reg2);
5108 slow_i_o = _gvn.transform(slow_i_o2);
5109 slow_mem = _gvn.transform(slow_mem2);
5110
5111 if (alloc != NULL) {
5112 // We'll restart from the very beginning, after zeroing the whole thing.
5113 // This can cause double writes, but that's OK since dest is brand new.
5114 // So we ignore the low 31 bits of the value returned from the stub.
5115 } else {
5116 // We must continue the copy exactly where it failed, or else
5117 // another thread might see the wrong number of writes to dest.
5118 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
5119 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5120 slow_offset->init_req(1, intcon(0));
5121 slow_offset->init_req(2, checked_offset);
5122 slow_offset = _gvn.transform(slow_offset);
5123
5124 // Adjust the arguments by the conditionally incoming offset.
5125 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
5126 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
5127 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
5128
5129 // Tweak the node variables to adjust the code produced below:
5130 src_offset = src_off_plus;
5131 dest_offset = dest_off_plus;
5132 copy_length = length_minus;
5133 }
5134 }
5135
5136 set_control(slow_control);
5137 if (!stopped()) {
5138 // Generate the slow path, if needed.
5139 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5140
5141 set_memory(slow_mem, adr_type);
5142 set_i_o(slow_i_o);
5143
5144 if (dest_uninitialized) {
5145 generate_clear_array(adr_type, dest, basic_elem_type,
5146 intcon(0), NULL,
5147 alloc->in(AllocateNode::AllocSize));
5148 }
5149
5150 generate_slow_arraycopy(adr_type,
5151 src, src_offset, dest, dest_offset,
5152 copy_length, /*dest_uninitialized*/false);
5153
5154 result_region->init_req(slow_call_path, control());
5155 result_i_o ->init_req(slow_call_path, i_o());
5156 result_memory->init_req(slow_call_path, memory(adr_type));
5157 }
5158
5159 // Remove unused edges.
5160 for (uint i = 1; i < result_region->req(); i++) {
5161 if (result_region->in(i) == NULL)
5162 result_region->init_req(i, top());
5163 }
5164
5165 // Finished; return the combined state.
5166 set_control( _gvn.transform(result_region) );
5167 set_i_o( _gvn.transform(result_i_o) );
5168 set_memory( _gvn.transform(result_memory), adr_type );
5169
5170 // The memory edges above are precise in order to model effects around
5171 // array copies accurately to allow value numbering of field loads around
5172 // arraycopy. Such field loads, both before and after, are common in Java
5173 // collections and similar classes involving header/array data structures.
5174 //
5175 // But with low number of register or when some registers are used or killed
5176 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5177 // The next memory barrier is added to avoid it. If the arraycopy can be
5178 // optimized away (which it can, sometimes) then we can manually remove
5179 // the membar also.
5180 //
5181 // Do not let reads from the cloned object float above the arraycopy.
5182 if (alloc != NULL) {
5183 // Do not let stores that initialize this object be reordered with
5184 // a subsequent store that would make this object accessible by
5185 // other threads.
5186 // Record what AllocateNode this StoreStore protects so that
5187 // escape analysis can go from the MemBarStoreStoreNode to the
5188 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5189 // based on the escape status of the AllocateNode.
5190 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5191 } else if (InsertMemBarAfterArraycopy)
5192 insert_mem_bar(Op_MemBarCPUOrder);
5193 }
5194
5195
5196 // Helper function which determines if an arraycopy immediately follows
5197 // an allocation, with no intervening tests or other escapes for the object.
5198 AllocateArrayNode*
5199 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5200 RegionNode* slow_region) {
5201 if (stopped()) return NULL; // no fast path
5202 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5203
5204 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5205 if (alloc == NULL) return NULL;
5206
5207 Node* rawmem = memory(Compile::AliasIdxRaw);
5208 // Is the allocation's memory state untouched?
5209 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5210 // Bail out if there have been raw-memory effects since the allocation.
5211 // (Example: There might have been a call or safepoint.)
5212 return NULL;
5213 }
5214 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5215 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5216 return NULL;
5217 }
5218
5219 // There must be no unexpected observers of this allocation.
5220 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5221 Node* obs = ptr->fast_out(i);
5222 if (obs != this->map()) {
5223 return NULL;
5224 }
5225 }
5226
5227 // This arraycopy must unconditionally follow the allocation of the ptr.
5228 Node* alloc_ctl = ptr->in(0);
5229 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5230
5231 Node* ctl = control();
5232 while (ctl != alloc_ctl) {
5233 // There may be guards which feed into the slow_region.
5234 // Any other control flow means that we might not get a chance
5235 // to finish initializing the allocated object.
5236 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5237 IfNode* iff = ctl->in(0)->as_If();
5238 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5239 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5240 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5241 ctl = iff->in(0); // This test feeds the known slow_region.
5242 continue;
5243 }
5244 // One more try: Various low-level checks bottom out in
5245 // uncommon traps. If the debug-info of the trap omits
5246 // any reference to the allocation, as we've already
5247 // observed, then there can be no objection to the trap.
5248 bool found_trap = false;
5249 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5250 Node* obs = not_ctl->fast_out(j);
5251 if (obs->in(0) == not_ctl && obs->is_Call() &&
5252 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5253 found_trap = true; break;
5254 }
5255 }
5256 if (found_trap) {
5257 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5258 continue;
5259 }
5260 }
5261 return NULL;
5262 }
5263
5264 // If we get this far, we have an allocation which immediately
5265 // precedes the arraycopy, and we can take over zeroing the new object.
5266 // The arraycopy will finish the initialization, and provide
5267 // a new control state to which we will anchor the destination pointer.
5268
5269 return alloc;
5270 }
5271
5272 // Helper for initialization of arrays, creating a ClearArray.
5273 // It writes zero bits in [start..end), within the body of an array object.
5274 // The memory effects are all chained onto the 'adr_type' alias category.
5275 //
5276 // Since the object is otherwise uninitialized, we are free
5277 // to put a little "slop" around the edges of the cleared area,
5278 // as long as it does not go back into the array's header,
5279 // or beyond the array end within the heap.
5280 //
5281 // The lower edge can be rounded down to the nearest jint and the
5282 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5283 //
5284 // Arguments:
5285 // adr_type memory slice where writes are generated
5286 // dest oop of the destination array
5287 // basic_elem_type element type of the destination
5288 // slice_idx array index of first element to store
5289 // slice_len number of elements to store (or NULL)
5290 // dest_size total size in bytes of the array object
5291 //
5292 // Exactly one of slice_len or dest_size must be non-NULL.
5293 // If dest_size is non-NULL, zeroing extends to the end of the object.
5294 // If slice_len is non-NULL, the slice_idx value must be a constant.
5295 void
5296 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5297 Node* dest,
5298 BasicType basic_elem_type,
5299 Node* slice_idx,
5300 Node* slice_len,
5301 Node* dest_size) {
5302 // one or the other but not both of slice_len and dest_size:
5303 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5304 if (slice_len == NULL) slice_len = top();
5305 if (dest_size == NULL) dest_size = top();
5306
5307 // operate on this memory slice:
5308 Node* mem = memory(adr_type); // memory slice to operate on
5309
5310 // scaling and rounding of indexes:
5311 int scale = exact_log2(type2aelembytes(basic_elem_type));
5312 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5313 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5314 int bump_bit = (-1 << scale) & BytesPerInt;
5315
5316 // determine constant starts and ends
5317 const intptr_t BIG_NEG = -128;
5318 assert(BIG_NEG + 2*abase < 0, "neg enough");
5319 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5320 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5321 if (slice_len_con == 0) {
5322 return; // nothing to do here
5323 }
5324 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5325 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5326 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5327 assert(end_con < 0, "not two cons");
5328 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5329 BytesPerLong);
5330 }
5331
5332 if (start_con >= 0 && end_con >= 0) {
5333 // Constant start and end. Simple.
5334 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5335 start_con, end_con, &_gvn);
5336 } else if (start_con >= 0 && dest_size != top()) {
5337 // Constant start, pre-rounded end after the tail of the array.
5338 Node* end = dest_size;
5339 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5340 start_con, end, &_gvn);
5341 } else if (start_con >= 0 && slice_len != top()) {
5342 // Constant start, non-constant end. End needs rounding up.
5343 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5344 intptr_t end_base = abase + (slice_idx_con << scale);
5345 int end_round = (-1 << scale) & (BytesPerLong - 1);
5346 Node* end = ConvI2X(slice_len);
5347 if (scale != 0)
5348 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5349 end_base += end_round;
5350 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5351 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5352 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5353 start_con, end, &_gvn);
5354 } else if (start_con < 0 && dest_size != top()) {
5355 // Non-constant start, pre-rounded end after the tail of the array.
5356 // This is almost certainly a "round-to-end" operation.
5357 Node* start = slice_idx;
5358 start = ConvI2X(start);
5359 if (scale != 0)
5360 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5361 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5362 if ((bump_bit | clear_low) != 0) {
5363 int to_clear = (bump_bit | clear_low);
5364 // Align up mod 8, then store a jint zero unconditionally
5365 // just before the mod-8 boundary.
5366 if (((abase + bump_bit) & ~to_clear) - bump_bit
5367 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5368 bump_bit = 0;
5369 assert((abase & to_clear) == 0, "array base must be long-aligned");
5370 } else {
5371 // Bump 'start' up to (or past) the next jint boundary:
5372 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5373 assert((abase & clear_low) == 0, "array base must be int-aligned");
5374 }
5375 // Round bumped 'start' down to jlong boundary in body of array.
5376 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5377 if (bump_bit != 0) {
5378 // Store a zero to the immediately preceding jint:
5379 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5380 Node* p1 = basic_plus_adr(dest, x1);
5381 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5382 mem = _gvn.transform(mem);
5383 }
5384 }
5385 Node* end = dest_size; // pre-rounded
5386 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5387 start, end, &_gvn);
5388 } else {
5389 // Non-constant start, unrounded non-constant end.
5390 // (Nobody zeroes a random midsection of an array using this routine.)
5391 ShouldNotReachHere(); // fix caller
5392 }
5393
5394 // Done.
5395 set_memory(mem, adr_type);
5396 }
5397
5398
5399 bool
5400 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5401 BasicType basic_elem_type,
5402 AllocateNode* alloc,
5403 Node* src, Node* src_offset,
5404 Node* dest, Node* dest_offset,
5405 Node* dest_size, bool dest_uninitialized) {
5406 // See if there is an advantage from block transfer.
5407 int scale = exact_log2(type2aelembytes(basic_elem_type));
5408 if (scale >= LogBytesPerLong)
5409 return false; // it is already a block transfer
5410
5411 // Look at the alignment of the starting offsets.
5412 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5413
5414 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5415 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5416 if (src_off_con < 0 || dest_off_con < 0)
5417 // At present, we can only understand constants.
5418 return false;
5419
5420 intptr_t src_off = abase + (src_off_con << scale);
5421 intptr_t dest_off = abase + (dest_off_con << scale);
5422
5423 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5424 // Non-aligned; too bad.
5425 // One more chance: Pick off an initial 32-bit word.
5426 // This is a common case, since abase can be odd mod 8.
5427 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5428 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5429 Node* sptr = basic_plus_adr(src, src_off);
5430 Node* dptr = basic_plus_adr(dest, dest_off);
5431 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5432 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5433 src_off += BytesPerInt;
5434 dest_off += BytesPerInt;
5435 } else {
5436 return false;
5437 }
5438 }
5439 assert(src_off % BytesPerLong == 0, "");
5440 assert(dest_off % BytesPerLong == 0, "");
5441
5442 // Do this copy by giant steps.
5443 Node* sptr = basic_plus_adr(src, src_off);
5444 Node* dptr = basic_plus_adr(dest, dest_off);
5445 Node* countx = dest_size;
5446 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5447 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5448
5449 bool disjoint_bases = true; // since alloc != NULL
5450 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5451 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5452
5453 return true;
5454 }
5455
5456
5457 // Helper function; generates code for the slow case.
5458 // We make a call to a runtime method which emulates the native method,
5459 // but without the native wrapper overhead.
5460 void
5461 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5462 Node* src, Node* src_offset,
5463 Node* dest, Node* dest_offset,
5464 Node* copy_length, bool dest_uninitialized) {
5465 assert(!dest_uninitialized, "Invariant");
5466 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5467 OptoRuntime::slow_arraycopy_Type(),
5468 OptoRuntime::slow_arraycopy_Java(),
5469 "slow_arraycopy", adr_type,
5470 src, src_offset, dest, dest_offset,
5471 copy_length);
5472
5473 // Handle exceptions thrown by this fellow:
5474 make_slow_call_ex(call, env()->Throwable_klass(), false);
5475 }
5476
5477 // Helper function; generates code for cases requiring runtime checks.
5478 Node*
5479 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5480 Node* dest_elem_klass,
5481 Node* src, Node* src_offset,
5482 Node* dest, Node* dest_offset,
5483 Node* copy_length, bool dest_uninitialized) {
5484 if (stopped()) return NULL;
5485
5486 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5487 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5488 return NULL;
5489 }
5490
5491 // Pick out the parameters required to perform a store-check
5492 // for the target array. This is an optimistic check. It will
5493 // look in each non-null element's class, at the desired klass's
5494 // super_check_offset, for the desired klass.
5495 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5496 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5497 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5498 Node* check_offset = ConvI2X(_gvn.transform(n3));
5499 Node* check_value = dest_elem_klass;
5500
5501 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5502 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5503
5504 // (We know the arrays are never conjoint, because their types differ.)
5505 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5506 OptoRuntime::checkcast_arraycopy_Type(),
5507 copyfunc_addr, "checkcast_arraycopy", adr_type,
5508 // five arguments, of which two are
5509 // intptr_t (jlong in LP64)
5510 src_start, dest_start,
5511 copy_length XTOP,
5512 check_offset XTOP,
5513 check_value);
5514
5515 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5516 }
5517
5518
5519 // Helper function; generates code for cases requiring runtime checks.
5520 Node*
5521 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5522 Node* src, Node* src_offset,
5523 Node* dest, Node* dest_offset,
5524 Node* copy_length, bool dest_uninitialized) {
5525 assert(!dest_uninitialized, "Invariant");
5526 if (stopped()) return NULL;
5527 address copyfunc_addr = StubRoutines::generic_arraycopy();
5528 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5529 return NULL;
5530 }
5531
5532 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5533 OptoRuntime::generic_arraycopy_Type(),
5534 copyfunc_addr, "generic_arraycopy", adr_type,
5535 src, src_offset, dest, dest_offset, copy_length);
5536
5537 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5538 }
5539
5540 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5541 void
5542 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5543 BasicType basic_elem_type,
5544 bool disjoint_bases,
5545 Node* src, Node* src_offset,
5546 Node* dest, Node* dest_offset,
5547 Node* copy_length, bool dest_uninitialized) {
5548 if (stopped()) return; // nothing to do
5549
5550 Node* src_start = src;
5551 Node* dest_start = dest;
5552 if (src_offset != NULL || dest_offset != NULL) {
5553 assert(src_offset != NULL && dest_offset != NULL, "");
5554 src_start = array_element_address(src, src_offset, basic_elem_type);
5555 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5556 }
5557
5558 // Figure out which arraycopy runtime method to call.
5559 const char* copyfunc_name = "arraycopy";
5560 address copyfunc_addr =
5561 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5562 disjoint_bases, copyfunc_name, dest_uninitialized);
5563
5564 // Call it. Note that the count_ix value is not scaled to a byte-size.
5565 make_runtime_call(RC_LEAF|RC_NO_FP,
5566 OptoRuntime::fast_arraycopy_Type(),
5567 copyfunc_addr, copyfunc_name, adr_type,
5568 src_start, dest_start, copy_length XTOP);
5569 }
5570
5571 //----------------------------inline_reference_get----------------------------
5572
5573 bool LibraryCallKit::inline_reference_get() {
5574 const int nargs = 1; // self
5575
5576 guarantee(java_lang_ref_Reference::referent_offset > 0,
5577 "should have already been set");
5578
5579 int referent_offset = java_lang_ref_Reference::referent_offset;
5580
5581 // Restore the stack and pop off the argument
5582 _sp += nargs;
5583 Node *reference_obj = pop();
5584
5585 // Null check on self without removing any arguments.
5586 _sp += nargs;
5587 reference_obj = do_null_check(reference_obj, T_OBJECT);
5588 _sp -= nargs;;
5589
5590 if (stopped()) return true;
5591
5592 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5593
5594 ciInstanceKlass* klass = env()->Object_klass();
5595 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5596
5597 Node* no_ctrl = NULL;
5598 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5599
5600 // Use the pre-barrier to record the value in the referent field
5601 pre_barrier(false /* do_load */,
5602 control(),
5603 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5604 result /* pre_val */,
5605 T_OBJECT);
5606
5607 // Add memory barrier to prevent commoning reads from this field
5608 // across safepoint since GC can change its value.
5609 insert_mem_bar(Op_MemBarCPUOrder);
5610
5611 push(result);
5612 return true;
5613 }