1 /*
2 * Copyright 1999-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_library_call.cpp.incl"
27
28 class LibraryIntrinsic : public InlineCallGenerator {
29 // Extend the set of intrinsics known to the runtime:
30 public:
31 private:
32 bool _is_virtual;
33 vmIntrinsics::ID _intrinsic_id;
34
35 public:
36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
37 : InlineCallGenerator(m),
38 _is_virtual(is_virtual),
39 _intrinsic_id(id)
40 {
41 }
42 virtual bool is_intrinsic() const { return true; }
43 virtual bool is_virtual() const { return _is_virtual; }
44 virtual JVMState* generate(JVMState* jvms);
45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
46 };
47
48
49 // Local helper class for LibraryIntrinsic:
50 class LibraryCallKit : public GraphKit {
51 private:
52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
53
54 public:
55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
56 : GraphKit(caller),
57 _intrinsic(intrinsic)
58 {
59 }
60
61 ciMethod* caller() const { return jvms()->method(); }
62 int bci() const { return jvms()->bci(); }
63 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
65 ciMethod* callee() const { return _intrinsic->method(); }
66 ciSignature* signature() const { return callee()->signature(); }
67 int arg_size() const { return callee()->arg_size(); }
68
69 bool try_to_inline();
70
71 // Helper functions to inline natives
72 void push_result(RegionNode* region, PhiNode* value);
73 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
74 Node* generate_slow_guard(Node* test, RegionNode* region);
75 Node* generate_fair_guard(Node* test, RegionNode* region);
76 Node* generate_negative_guard(Node* index, RegionNode* region,
77 // resulting CastII of index:
78 Node* *pos_index = NULL);
79 Node* generate_nonpositive_guard(Node* index, bool never_negative,
80 // resulting CastII of index:
81 Node* *pos_index = NULL);
82 Node* generate_limit_guard(Node* offset, Node* subseq_length,
83 Node* array_length,
84 RegionNode* region);
85 Node* generate_current_thread(Node* &tls_output);
86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
87 bool disjoint_bases, const char* &name);
88 Node* load_mirror_from_klass(Node* klass);
89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
90 int nargs,
91 RegionNode* region, int null_path,
92 int offset);
93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
94 RegionNode* region, int null_path) {
95 int offset = java_lang_Class::klass_offset_in_bytes();
96 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
97 region, null_path,
98 offset);
99 }
100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
101 int nargs,
102 RegionNode* region, int null_path) {
103 int offset = java_lang_Class::array_klass_offset_in_bytes();
104 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
105 region, null_path,
106 offset);
107 }
108 Node* generate_access_flags_guard(Node* kls,
109 int modifier_mask, int modifier_bits,
110 RegionNode* region);
111 Node* generate_interface_guard(Node* kls, RegionNode* region);
112 Node* generate_array_guard(Node* kls, RegionNode* region) {
113 return generate_array_guard_common(kls, region, false, false);
114 }
115 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
116 return generate_array_guard_common(kls, region, false, true);
117 }
118 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
119 return generate_array_guard_common(kls, region, true, false);
120 }
121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
122 return generate_array_guard_common(kls, region, true, true);
123 }
124 Node* generate_array_guard_common(Node* kls, RegionNode* region,
125 bool obj_array, bool not_array);
126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
128 bool is_virtual = false, bool is_static = false);
129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
130 return generate_method_call(method_id, false, true);
131 }
132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
133 return generate_method_call(method_id, true, false);
134 }
135
136 bool inline_string_compareTo();
137 bool inline_string_indexOf();
138 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
139 bool inline_string_equals();
140 Node* pop_math_arg();
141 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
142 bool inline_math_native(vmIntrinsics::ID id);
143 bool inline_trig(vmIntrinsics::ID id);
144 bool inline_trans(vmIntrinsics::ID id);
145 bool inline_abs(vmIntrinsics::ID id);
146 bool inline_sqrt(vmIntrinsics::ID id);
147 bool inline_pow(vmIntrinsics::ID id);
148 bool inline_exp(vmIntrinsics::ID id);
149 bool inline_min_max(vmIntrinsics::ID id);
150 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
151 // This returns Type::AnyPtr, RawPtr, or OopPtr.
152 int classify_unsafe_addr(Node* &base, Node* &offset);
153 Node* make_unsafe_address(Node* base, Node* offset);
154 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
155 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
156 bool inline_unsafe_allocate();
157 bool inline_unsafe_copyMemory();
158 bool inline_native_currentThread();
159 bool inline_native_time_funcs(bool isNano);
160 bool inline_native_isInterrupted();
161 bool inline_native_Class_query(vmIntrinsics::ID id);
162 bool inline_native_subtype_check();
163
164 bool inline_native_newArray();
165 bool inline_native_getLength();
166 bool inline_array_copyOf(bool is_copyOfRange);
167 bool inline_array_equals();
168 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
169 bool inline_native_clone(bool is_virtual);
170 bool inline_native_Reflection_getCallerClass();
171 bool inline_native_AtomicLong_get();
172 bool inline_native_AtomicLong_attemptUpdate();
173 bool is_method_invoke_or_aux_frame(JVMState* jvms);
174 // Helper function for inlining native object hash method
175 bool inline_native_hashcode(bool is_virtual, bool is_static);
176 bool inline_native_getClass();
177
178 // Helper functions for inlining arraycopy
179 bool inline_arraycopy();
180 void generate_arraycopy(const TypePtr* adr_type,
181 BasicType basic_elem_type,
182 Node* src, Node* src_offset,
183 Node* dest, Node* dest_offset,
184 Node* copy_length,
185 bool disjoint_bases = false,
186 bool length_never_negative = false,
187 RegionNode* slow_region = NULL);
188 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
189 RegionNode* slow_region);
190 void generate_clear_array(const TypePtr* adr_type,
191 Node* dest,
192 BasicType basic_elem_type,
193 Node* slice_off,
194 Node* slice_len,
195 Node* slice_end);
196 bool generate_block_arraycopy(const TypePtr* adr_type,
197 BasicType basic_elem_type,
198 AllocateNode* alloc,
199 Node* src, Node* src_offset,
200 Node* dest, Node* dest_offset,
201 Node* dest_size);
202 void generate_slow_arraycopy(const TypePtr* adr_type,
203 Node* src, Node* src_offset,
204 Node* dest, Node* dest_offset,
205 Node* copy_length);
206 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
207 Node* dest_elem_klass,
208 Node* src, Node* src_offset,
209 Node* dest, Node* dest_offset,
210 Node* copy_length);
211 Node* generate_generic_arraycopy(const TypePtr* adr_type,
212 Node* src, Node* src_offset,
213 Node* dest, Node* dest_offset,
214 Node* copy_length);
215 void generate_unchecked_arraycopy(const TypePtr* adr_type,
216 BasicType basic_elem_type,
217 bool disjoint_bases,
218 Node* src, Node* src_offset,
219 Node* dest, Node* dest_offset,
220 Node* copy_length);
221 bool inline_unsafe_CAS(BasicType type);
222 bool inline_unsafe_ordered_store(BasicType type);
223 bool inline_fp_conversions(vmIntrinsics::ID id);
224 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
225 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
226 bool inline_bitCount(vmIntrinsics::ID id);
227 bool inline_reverseBytes(vmIntrinsics::ID id);
228 };
229
230
231 //---------------------------make_vm_intrinsic----------------------------
232 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
233 vmIntrinsics::ID id = m->intrinsic_id();
234 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
235
236 if (DisableIntrinsic[0] != '\0'
237 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
238 // disabled by a user request on the command line:
239 // example: -XX:DisableIntrinsic=_hashCode,_getClass
240 return NULL;
241 }
242
243 if (!m->is_loaded()) {
244 // do not attempt to inline unloaded methods
245 return NULL;
246 }
247
248 // Only a few intrinsics implement a virtual dispatch.
249 // They are expensive calls which are also frequently overridden.
250 if (is_virtual) {
251 switch (id) {
252 case vmIntrinsics::_hashCode:
253 case vmIntrinsics::_clone:
254 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
255 break;
256 default:
257 return NULL;
258 }
259 }
260
261 // -XX:-InlineNatives disables nearly all intrinsics:
262 if (!InlineNatives) {
263 switch (id) {
264 case vmIntrinsics::_indexOf:
265 case vmIntrinsics::_compareTo:
266 case vmIntrinsics::_equals:
267 case vmIntrinsics::_equalsC:
268 break; // InlineNatives does not control String.compareTo
269 default:
270 return NULL;
271 }
272 }
273
274 switch (id) {
275 case vmIntrinsics::_compareTo:
276 if (!SpecialStringCompareTo) return NULL;
277 break;
278 case vmIntrinsics::_indexOf:
279 if (!SpecialStringIndexOf) return NULL;
280 break;
281 case vmIntrinsics::_equals:
282 if (!SpecialStringEquals) return NULL;
283 break;
284 case vmIntrinsics::_equalsC:
285 if (!SpecialArraysEquals) return NULL;
286 break;
287 case vmIntrinsics::_arraycopy:
288 if (!InlineArrayCopy) return NULL;
289 break;
290 case vmIntrinsics::_copyMemory:
291 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
292 if (!InlineArrayCopy) return NULL;
293 break;
294 case vmIntrinsics::_hashCode:
295 if (!InlineObjectHash) return NULL;
296 break;
297 case vmIntrinsics::_clone:
298 case vmIntrinsics::_copyOf:
299 case vmIntrinsics::_copyOfRange:
300 if (!InlineObjectCopy) return NULL;
301 // These also use the arraycopy intrinsic mechanism:
302 if (!InlineArrayCopy) return NULL;
303 break;
304 case vmIntrinsics::_checkIndex:
305 // We do not intrinsify this. The optimizer does fine with it.
306 return NULL;
307
308 case vmIntrinsics::_get_AtomicLong:
309 case vmIntrinsics::_attemptUpdate:
310 if (!InlineAtomicLong) return NULL;
311 break;
312
313 case vmIntrinsics::_getCallerClass:
314 if (!UseNewReflection) return NULL;
315 if (!InlineReflectionGetCallerClass) return NULL;
316 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
317 break;
318
319 case vmIntrinsics::_bitCount_i:
320 case vmIntrinsics::_bitCount_l:
321 if (!UsePopCountInstruction) return NULL;
322 break;
323
324 default:
325 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
326 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
327 break;
328 }
329
330 // -XX:-InlineClassNatives disables natives from the Class class.
331 // The flag applies to all reflective calls, notably Array.newArray
332 // (visible to Java programmers as Array.newInstance).
333 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
334 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
335 if (!InlineClassNatives) return NULL;
336 }
337
338 // -XX:-InlineThreadNatives disables natives from the Thread class.
339 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
340 if (!InlineThreadNatives) return NULL;
341 }
342
343 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
344 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
345 m->holder()->name() == ciSymbol::java_lang_Float() ||
346 m->holder()->name() == ciSymbol::java_lang_Double()) {
347 if (!InlineMathNatives) return NULL;
348 }
349
350 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
351 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
352 if (!InlineUnsafeOps) return NULL;
353 }
354
355 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
356 }
357
358 //----------------------register_library_intrinsics-----------------------
359 // Initialize this file's data structures, for each Compile instance.
360 void Compile::register_library_intrinsics() {
361 // Nothing to do here.
362 }
363
364 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
365 LibraryCallKit kit(jvms, this);
366 Compile* C = kit.C;
367 int nodes = C->unique();
368 #ifndef PRODUCT
369 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
370 char buf[1000];
371 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
372 tty->print_cr("Intrinsic %s", str);
373 }
374 #endif
375 if (kit.try_to_inline()) {
376 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
377 tty->print("Inlining intrinsic %s%s at bci:%d in",
378 vmIntrinsics::name_at(intrinsic_id()),
379 (is_virtual() ? " (virtual)" : ""), kit.bci());
380 kit.caller()->print_short_name(tty);
381 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
382 }
383 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
384 if (C->log()) {
385 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
386 vmIntrinsics::name_at(intrinsic_id()),
387 (is_virtual() ? " virtual='1'" : ""),
388 C->unique() - nodes);
389 }
390 return kit.transfer_exceptions_into_jvms();
391 }
392
393 if (PrintIntrinsics) {
394 tty->print("Did not inline intrinsic %s%s at bci:%d in",
395 vmIntrinsics::name_at(intrinsic_id()),
396 (is_virtual() ? " (virtual)" : ""), kit.bci());
397 kit.caller()->print_short_name(tty);
398 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
399 }
400 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
401 return NULL;
402 }
403
404 bool LibraryCallKit::try_to_inline() {
405 // Handle symbolic names for otherwise undistinguished boolean switches:
406 const bool is_store = true;
407 const bool is_native_ptr = true;
408 const bool is_static = true;
409
410 switch (intrinsic_id()) {
411 case vmIntrinsics::_hashCode:
412 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
413 case vmIntrinsics::_identityHashCode:
414 return inline_native_hashcode(/*!virtual*/ false, is_static);
415 case vmIntrinsics::_getClass:
416 return inline_native_getClass();
417
418 case vmIntrinsics::_dsin:
419 case vmIntrinsics::_dcos:
420 case vmIntrinsics::_dtan:
421 case vmIntrinsics::_dabs:
422 case vmIntrinsics::_datan2:
423 case vmIntrinsics::_dsqrt:
424 case vmIntrinsics::_dexp:
425 case vmIntrinsics::_dlog:
426 case vmIntrinsics::_dlog10:
427 case vmIntrinsics::_dpow:
428 return inline_math_native(intrinsic_id());
429
430 case vmIntrinsics::_min:
431 case vmIntrinsics::_max:
432 return inline_min_max(intrinsic_id());
433
434 case vmIntrinsics::_arraycopy:
435 return inline_arraycopy();
436
437 case vmIntrinsics::_compareTo:
438 return inline_string_compareTo();
439 case vmIntrinsics::_indexOf:
440 return inline_string_indexOf();
441 case vmIntrinsics::_equals:
442 return inline_string_equals();
443
444 case vmIntrinsics::_getObject:
445 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
446 case vmIntrinsics::_getBoolean:
447 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
448 case vmIntrinsics::_getByte:
449 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
450 case vmIntrinsics::_getShort:
451 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
452 case vmIntrinsics::_getChar:
453 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
454 case vmIntrinsics::_getInt:
455 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
456 case vmIntrinsics::_getLong:
457 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
458 case vmIntrinsics::_getFloat:
459 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
460 case vmIntrinsics::_getDouble:
461 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
462
463 case vmIntrinsics::_putObject:
464 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
465 case vmIntrinsics::_putBoolean:
466 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
467 case vmIntrinsics::_putByte:
468 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
469 case vmIntrinsics::_putShort:
470 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
471 case vmIntrinsics::_putChar:
472 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
473 case vmIntrinsics::_putInt:
474 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
475 case vmIntrinsics::_putLong:
476 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
477 case vmIntrinsics::_putFloat:
478 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
479 case vmIntrinsics::_putDouble:
480 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
481
482 case vmIntrinsics::_getByte_raw:
483 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
484 case vmIntrinsics::_getShort_raw:
485 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
486 case vmIntrinsics::_getChar_raw:
487 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
488 case vmIntrinsics::_getInt_raw:
489 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
490 case vmIntrinsics::_getLong_raw:
491 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
492 case vmIntrinsics::_getFloat_raw:
493 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
494 case vmIntrinsics::_getDouble_raw:
495 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
496 case vmIntrinsics::_getAddress_raw:
497 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
498
499 case vmIntrinsics::_putByte_raw:
500 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
501 case vmIntrinsics::_putShort_raw:
502 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
503 case vmIntrinsics::_putChar_raw:
504 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
505 case vmIntrinsics::_putInt_raw:
506 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
507 case vmIntrinsics::_putLong_raw:
508 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
509 case vmIntrinsics::_putFloat_raw:
510 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
511 case vmIntrinsics::_putDouble_raw:
512 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
513 case vmIntrinsics::_putAddress_raw:
514 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
515
516 case vmIntrinsics::_getObjectVolatile:
517 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
518 case vmIntrinsics::_getBooleanVolatile:
519 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
520 case vmIntrinsics::_getByteVolatile:
521 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
522 case vmIntrinsics::_getShortVolatile:
523 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
524 case vmIntrinsics::_getCharVolatile:
525 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
526 case vmIntrinsics::_getIntVolatile:
527 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
528 case vmIntrinsics::_getLongVolatile:
529 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
530 case vmIntrinsics::_getFloatVolatile:
531 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
532 case vmIntrinsics::_getDoubleVolatile:
533 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
534
535 case vmIntrinsics::_putObjectVolatile:
536 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
537 case vmIntrinsics::_putBooleanVolatile:
538 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
539 case vmIntrinsics::_putByteVolatile:
540 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
541 case vmIntrinsics::_putShortVolatile:
542 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
543 case vmIntrinsics::_putCharVolatile:
544 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
545 case vmIntrinsics::_putIntVolatile:
546 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
547 case vmIntrinsics::_putLongVolatile:
548 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
549 case vmIntrinsics::_putFloatVolatile:
550 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
551 case vmIntrinsics::_putDoubleVolatile:
552 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
553
554 case vmIntrinsics::_prefetchRead:
555 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
556 case vmIntrinsics::_prefetchWrite:
557 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
558 case vmIntrinsics::_prefetchReadStatic:
559 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
560 case vmIntrinsics::_prefetchWriteStatic:
561 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
562
563 case vmIntrinsics::_compareAndSwapObject:
564 return inline_unsafe_CAS(T_OBJECT);
565 case vmIntrinsics::_compareAndSwapInt:
566 return inline_unsafe_CAS(T_INT);
567 case vmIntrinsics::_compareAndSwapLong:
568 return inline_unsafe_CAS(T_LONG);
569
570 case vmIntrinsics::_putOrderedObject:
571 return inline_unsafe_ordered_store(T_OBJECT);
572 case vmIntrinsics::_putOrderedInt:
573 return inline_unsafe_ordered_store(T_INT);
574 case vmIntrinsics::_putOrderedLong:
575 return inline_unsafe_ordered_store(T_LONG);
576
577 case vmIntrinsics::_currentThread:
578 return inline_native_currentThread();
579 case vmIntrinsics::_isInterrupted:
580 return inline_native_isInterrupted();
581
582 case vmIntrinsics::_currentTimeMillis:
583 return inline_native_time_funcs(false);
584 case vmIntrinsics::_nanoTime:
585 return inline_native_time_funcs(true);
586 case vmIntrinsics::_allocateInstance:
587 return inline_unsafe_allocate();
588 case vmIntrinsics::_copyMemory:
589 return inline_unsafe_copyMemory();
590 case vmIntrinsics::_newArray:
591 return inline_native_newArray();
592 case vmIntrinsics::_getLength:
593 return inline_native_getLength();
594 case vmIntrinsics::_copyOf:
595 return inline_array_copyOf(false);
596 case vmIntrinsics::_copyOfRange:
597 return inline_array_copyOf(true);
598 case vmIntrinsics::_equalsC:
599 return inline_array_equals();
600 case vmIntrinsics::_clone:
601 return inline_native_clone(intrinsic()->is_virtual());
602
603 case vmIntrinsics::_isAssignableFrom:
604 return inline_native_subtype_check();
605
606 case vmIntrinsics::_isInstance:
607 case vmIntrinsics::_getModifiers:
608 case vmIntrinsics::_isInterface:
609 case vmIntrinsics::_isArray:
610 case vmIntrinsics::_isPrimitive:
611 case vmIntrinsics::_getSuperclass:
612 case vmIntrinsics::_getComponentType:
613 case vmIntrinsics::_getClassAccessFlags:
614 return inline_native_Class_query(intrinsic_id());
615
616 case vmIntrinsics::_floatToRawIntBits:
617 case vmIntrinsics::_floatToIntBits:
618 case vmIntrinsics::_intBitsToFloat:
619 case vmIntrinsics::_doubleToRawLongBits:
620 case vmIntrinsics::_doubleToLongBits:
621 case vmIntrinsics::_longBitsToDouble:
622 return inline_fp_conversions(intrinsic_id());
623
624 case vmIntrinsics::_numberOfLeadingZeros_i:
625 case vmIntrinsics::_numberOfLeadingZeros_l:
626 return inline_numberOfLeadingZeros(intrinsic_id());
627
628 case vmIntrinsics::_numberOfTrailingZeros_i:
629 case vmIntrinsics::_numberOfTrailingZeros_l:
630 return inline_numberOfTrailingZeros(intrinsic_id());
631
632 case vmIntrinsics::_bitCount_i:
633 case vmIntrinsics::_bitCount_l:
634 return inline_bitCount(intrinsic_id());
635
636 case vmIntrinsics::_reverseBytes_i:
637 case vmIntrinsics::_reverseBytes_l:
638 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
639
640 case vmIntrinsics::_get_AtomicLong:
641 return inline_native_AtomicLong_get();
642 case vmIntrinsics::_attemptUpdate:
643 return inline_native_AtomicLong_attemptUpdate();
644
645 case vmIntrinsics::_getCallerClass:
646 return inline_native_Reflection_getCallerClass();
647
648 default:
649 // If you get here, it may be that someone has added a new intrinsic
650 // to the list in vmSymbols.hpp without implementing it here.
651 #ifndef PRODUCT
652 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
653 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
654 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
655 }
656 #endif
657 return false;
658 }
659 }
660
661 //------------------------------push_result------------------------------
662 // Helper function for finishing intrinsics.
663 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
664 record_for_igvn(region);
665 set_control(_gvn.transform(region));
666 BasicType value_type = value->type()->basic_type();
667 push_node(value_type, _gvn.transform(value));
668 }
669
670 //------------------------------generate_guard---------------------------
671 // Helper function for generating guarded fast-slow graph structures.
672 // The given 'test', if true, guards a slow path. If the test fails
673 // then a fast path can be taken. (We generally hope it fails.)
674 // In all cases, GraphKit::control() is updated to the fast path.
675 // The returned value represents the control for the slow path.
676 // The return value is never 'top'; it is either a valid control
677 // or NULL if it is obvious that the slow path can never be taken.
678 // Also, if region and the slow control are not NULL, the slow edge
679 // is appended to the region.
680 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
681 if (stopped()) {
682 // Already short circuited.
683 return NULL;
684 }
685
686 // Build an if node and its projections.
687 // If test is true we take the slow path, which we assume is uncommon.
688 if (_gvn.type(test) == TypeInt::ZERO) {
689 // The slow branch is never taken. No need to build this guard.
690 return NULL;
691 }
692
693 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
694
695 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
696 if (if_slow == top()) {
697 // The slow branch is never taken. No need to build this guard.
698 return NULL;
699 }
700
701 if (region != NULL)
702 region->add_req(if_slow);
703
704 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
705 set_control(if_fast);
706
707 return if_slow;
708 }
709
710 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
711 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
712 }
713 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
714 return generate_guard(test, region, PROB_FAIR);
715 }
716
717 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
718 Node* *pos_index) {
719 if (stopped())
720 return NULL; // already stopped
721 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
722 return NULL; // index is already adequately typed
723 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
724 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
725 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
726 if (is_neg != NULL && pos_index != NULL) {
727 // Emulate effect of Parse::adjust_map_after_if.
728 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
729 ccast->set_req(0, control());
730 (*pos_index) = _gvn.transform(ccast);
731 }
732 return is_neg;
733 }
734
735 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
736 Node* *pos_index) {
737 if (stopped())
738 return NULL; // already stopped
739 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
740 return NULL; // index is already adequately typed
741 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
742 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
743 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
744 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
745 if (is_notp != NULL && pos_index != NULL) {
746 // Emulate effect of Parse::adjust_map_after_if.
747 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
748 ccast->set_req(0, control());
749 (*pos_index) = _gvn.transform(ccast);
750 }
751 return is_notp;
752 }
753
754 // Make sure that 'position' is a valid limit index, in [0..length].
755 // There are two equivalent plans for checking this:
756 // A. (offset + copyLength) unsigned<= arrayLength
757 // B. offset <= (arrayLength - copyLength)
758 // We require that all of the values above, except for the sum and
759 // difference, are already known to be non-negative.
760 // Plan A is robust in the face of overflow, if offset and copyLength
761 // are both hugely positive.
762 //
763 // Plan B is less direct and intuitive, but it does not overflow at
764 // all, since the difference of two non-negatives is always
765 // representable. Whenever Java methods must perform the equivalent
766 // check they generally use Plan B instead of Plan A.
767 // For the moment we use Plan A.
768 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
769 Node* subseq_length,
770 Node* array_length,
771 RegionNode* region) {
772 if (stopped())
773 return NULL; // already stopped
774 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
775 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
776 return NULL; // common case of whole-array copy
777 Node* last = subseq_length;
778 if (!zero_offset) // last += offset
779 last = _gvn.transform( new (C, 3) AddINode(last, offset));
780 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
781 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
782 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
783 return is_over;
784 }
785
786
787 //--------------------------generate_current_thread--------------------
788 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
789 ciKlass* thread_klass = env()->Thread_klass();
790 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
791 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
792 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
793 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
794 tls_output = thread;
795 return threadObj;
796 }
797
798
799 //------------------------------inline_string_compareTo------------------------
800 bool LibraryCallKit::inline_string_compareTo() {
801
802 if (!Matcher::has_match_rule(Op_StrComp)) return false;
803
804 const int value_offset = java_lang_String::value_offset_in_bytes();
805 const int count_offset = java_lang_String::count_offset_in_bytes();
806 const int offset_offset = java_lang_String::offset_offset_in_bytes();
807
808 _sp += 2;
809 Node *argument = pop(); // pop non-receiver first: it was pushed second
810 Node *receiver = pop();
811
812 // Null check on self without removing any arguments. The argument
813 // null check technically happens in the wrong place, which can lead to
814 // invalid stack traces when string compare is inlined into a method
815 // which handles NullPointerExceptions.
816 _sp += 2;
817 receiver = do_null_check(receiver, T_OBJECT);
818 argument = do_null_check(argument, T_OBJECT);
819 _sp -= 2;
820 if (stopped()) {
821 return true;
822 }
823
824 ciInstanceKlass* klass = env()->String_klass();
825 const TypeInstPtr* string_type =
826 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
827
828 Node* compare =
829 _gvn.transform(new (C, 7) StrCompNode(
830 control(),
831 memory(TypeAryPtr::CHARS),
832 memory(string_type->add_offset(value_offset)),
833 memory(string_type->add_offset(count_offset)),
834 memory(string_type->add_offset(offset_offset)),
835 receiver,
836 argument));
837 push(compare);
838 return true;
839 }
840
841 //------------------------------inline_string_equals------------------------
842 bool LibraryCallKit::inline_string_equals() {
843
844 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
845
846 const int value_offset = java_lang_String::value_offset_in_bytes();
847 const int count_offset = java_lang_String::count_offset_in_bytes();
848 const int offset_offset = java_lang_String::offset_offset_in_bytes();
849
850 _sp += 2;
851 Node* argument = pop(); // pop non-receiver first: it was pushed second
852 Node* receiver = pop();
853
854 // Null check on self without removing any arguments. The argument
855 // null check technically happens in the wrong place, which can lead to
856 // invalid stack traces when string compare is inlined into a method
857 // which handles NullPointerExceptions.
858 _sp += 2;
859 receiver = do_null_check(receiver, T_OBJECT);
860 //should not do null check for argument for String.equals(), because spec
861 //allows to specify NULL as argument.
862 _sp -= 2;
863
864 if (stopped()) {
865 return true;
866 }
867
868 // get String klass for instanceOf
869 ciInstanceKlass* klass = env()->String_klass();
870
871 // two paths (plus control) merge
872 RegionNode* region = new (C, 3) RegionNode(3);
873 Node* phi = new (C, 3) PhiNode(region, TypeInt::BOOL);
874
875 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
876 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
877 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
878
879 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
880
881 Node* if_true = _gvn.transform(new (C, 1) IfTrueNode(iff));
882 set_control(if_true);
883
884 const TypeInstPtr* string_type =
885 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
886
887 // instanceOf == true
888 Node* equals =
889 _gvn.transform(new (C, 7) StrEqualsNode(
890 control(),
891 memory(TypeAryPtr::CHARS),
892 memory(string_type->add_offset(value_offset)),
893 memory(string_type->add_offset(count_offset)),
894 memory(string_type->add_offset(offset_offset)),
895 receiver,
896 argument));
897
898 phi->init_req(1, _gvn.transform(equals));
899 region->init_req(1, if_true);
900
901 //instanceOf == false, fallthrough
902 Node* if_false = _gvn.transform(new (C, 1) IfFalseNode(iff));
903 set_control(if_false);
904
905 phi->init_req(2, _gvn.transform(intcon(0)));
906 region->init_req(2, if_false);
907
908 // post merge
909 set_control(_gvn.transform(region));
910 record_for_igvn(region);
911
912 push(_gvn.transform(phi));
913
914 return true;
915 }
916
917 //------------------------------inline_array_equals----------------------------
918 bool LibraryCallKit::inline_array_equals() {
919
920 if (!Matcher::has_match_rule(Op_AryEq)) return false;
921
922 _sp += 2;
923 Node *argument2 = pop();
924 Node *argument1 = pop();
925
926 Node* equals =
927 _gvn.transform(new (C, 3) AryEqNode(control(),
928 argument1,
929 argument2)
930 );
931 push(equals);
932 return true;
933 }
934
935 // Java version of String.indexOf(constant string)
936 // class StringDecl {
937 // StringDecl(char[] ca) {
938 // offset = 0;
939 // count = ca.length;
940 // value = ca;
941 // }
942 // int offset;
943 // int count;
944 // char[] value;
945 // }
946 //
947 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
948 // int targetOffset, int cache_i, int md2) {
949 // int cache = cache_i;
950 // int sourceOffset = string_object.offset;
951 // int sourceCount = string_object.count;
952 // int targetCount = target_object.length;
953 //
954 // int targetCountLess1 = targetCount - 1;
955 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
956 //
957 // char[] source = string_object.value;
958 // char[] target = target_object;
959 // int lastChar = target[targetCountLess1];
960 //
961 // outer_loop:
962 // for (int i = sourceOffset; i < sourceEnd; ) {
963 // int src = source[i + targetCountLess1];
964 // if (src == lastChar) {
965 // // With random strings and a 4-character alphabet,
966 // // reverse matching at this point sets up 0.8% fewer
967 // // frames, but (paradoxically) makes 0.3% more probes.
968 // // Since those probes are nearer the lastChar probe,
969 // // there is may be a net D$ win with reverse matching.
970 // // But, reversing loop inhibits unroll of inner loop
971 // // for unknown reason. So, does running outer loop from
972 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
973 // for (int j = 0; j < targetCountLess1; j++) {
974 // if (target[targetOffset + j] != source[i+j]) {
975 // if ((cache & (1 << source[i+j])) == 0) {
976 // if (md2 < j+1) {
977 // i += j+1;
978 // continue outer_loop;
979 // }
980 // }
981 // i += md2;
982 // continue outer_loop;
983 // }
984 // }
985 // return i - sourceOffset;
986 // }
987 // if ((cache & (1 << src)) == 0) {
988 // i += targetCountLess1;
989 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
990 // i++;
991 // }
992 // return -1;
993 // }
994
995 //------------------------------string_indexOf------------------------
996 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
997 jint cache_i, jint md2_i) {
998
999 Node* no_ctrl = NULL;
1000 float likely = PROB_LIKELY(0.9);
1001 float unlikely = PROB_UNLIKELY(0.9);
1002
1003 const int value_offset = java_lang_String::value_offset_in_bytes();
1004 const int count_offset = java_lang_String::count_offset_in_bytes();
1005 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1006
1007 ciInstanceKlass* klass = env()->String_klass();
1008 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1009 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1010
1011 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1012 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1013 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1014 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1015 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1016 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1017
1018 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
1019 jint target_length = target_array->length();
1020 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1021 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1022
1023 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1024 #define __ kit.
1025 Node* zero = __ ConI(0);
1026 Node* one = __ ConI(1);
1027 Node* cache = __ ConI(cache_i);
1028 Node* md2 = __ ConI(md2_i);
1029 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1030 Node* targetCount = __ ConI(target_length);
1031 Node* targetCountLess1 = __ ConI(target_length - 1);
1032 Node* targetOffset = __ ConI(targetOffset_i);
1033 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1034
1035 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1036 Node* outer_loop = __ make_label(2 /* goto */);
1037 Node* return_ = __ make_label(1);
1038
1039 __ set(rtn,__ ConI(-1));
1040 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
1041 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1042 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1043 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1044 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1045 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
1046 Node* tpj = __ AddI(targetOffset, __ value(j));
1047 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1048 Node* ipj = __ AddI(__ value(i), __ value(j));
1049 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1050 __ if_then(targ, BoolTest::ne, src2); {
1051 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1052 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1053 __ increment(i, __ AddI(__ value(j), one));
1054 __ goto_(outer_loop);
1055 } __ end_if(); __ dead(j);
1056 }__ end_if(); __ dead(j);
1057 __ increment(i, md2);
1058 __ goto_(outer_loop);
1059 }__ end_if();
1060 __ increment(j, one);
1061 }__ end_loop(); __ dead(j);
1062 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1063 __ goto_(return_);
1064 }__ end_if();
1065 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1066 __ increment(i, targetCountLess1);
1067 }__ end_if();
1068 __ increment(i, one);
1069 __ bind(outer_loop);
1070 }__ end_loop(); __ dead(i);
1071 __ bind(return_);
1072
1073 // Final sync IdealKit and GraphKit.
1074 sync_kit(kit);
1075 Node* result = __ value(rtn);
1076 #undef __
1077 C->set_has_loops(true);
1078 return result;
1079 }
1080
1081 //------------------------------inline_string_indexOf------------------------
1082 bool LibraryCallKit::inline_string_indexOf() {
1083
1084 const int value_offset = java_lang_String::value_offset_in_bytes();
1085 const int count_offset = java_lang_String::count_offset_in_bytes();
1086 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1087
1088 _sp += 2;
1089 Node *argument = pop(); // pop non-receiver first: it was pushed second
1090 Node *receiver = pop();
1091
1092 Node* alloc = NULL;
1093 if (DoEscapeAnalysis && argument->is_Con())
1094 alloc = AllocateNode::Ideal_allocation(receiver, &_gvn);
1095
1096 Node* result;
1097 if (Matcher::has_match_rule(Op_StrIndexOf) && alloc == NULL &&
1098 UseSSE42Intrinsics) {
1099 // Generate SSE4.2 version of indexOf
1100 // We currently only have match rules that use SSE4.2
1101
1102 // Null check on self without removing any arguments. The argument
1103 // null check technically happens in the wrong place, which can lead to
1104 // invalid stack traces when string compare is inlined into a method
1105 // which handles NullPointerExceptions.
1106 _sp += 2;
1107 receiver = do_null_check(receiver, T_OBJECT);
1108 argument = do_null_check(argument, T_OBJECT);
1109 _sp -= 2;
1110
1111 if (stopped()) {
1112 return true;
1113 }
1114
1115 ciInstanceKlass* klass = env()->String_klass();
1116 const TypeInstPtr* string_type =
1117 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1118
1119 result =
1120 _gvn.transform(new (C, 7)
1121 StrIndexOfNode(control(),
1122 memory(TypeAryPtr::CHARS),
1123 memory(string_type->add_offset(value_offset)),
1124 memory(string_type->add_offset(count_offset)),
1125 memory(string_type->add_offset(offset_offset)),
1126 receiver,
1127 argument));
1128 } else { //Use LibraryCallKit::string_indexOf
1129 // don't intrinsify is argument isn't a constant string.
1130 if (!argument->is_Con()) {
1131 return false;
1132 }
1133 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1134 if (str_type == NULL) {
1135 return false;
1136 }
1137 ciInstanceKlass* klass = env()->String_klass();
1138 ciObject* str_const = str_type->const_oop();
1139 if (str_const == NULL || str_const->klass() != klass) {
1140 return false;
1141 }
1142 ciInstance* str = str_const->as_instance();
1143 assert(str != NULL, "must be instance");
1144
1145 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1146 int o = str->field_value_by_offset(offset_offset).as_int();
1147 int c = str->field_value_by_offset(count_offset).as_int();
1148 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1149
1150 // constant strings have no offset and count == length which
1151 // simplifies the resulting code somewhat so lets optimize for that.
1152 if (o != 0 || c != pat->length()) {
1153 return false;
1154 }
1155
1156 // Null check on self without removing any arguments. The argument
1157 // null check technically happens in the wrong place, which can lead to
1158 // invalid stack traces when string compare is inlined into a method
1159 // which handles NullPointerExceptions.
1160 _sp += 2;
1161 receiver = do_null_check(receiver, T_OBJECT);
1162 // No null check on the argument is needed since it's a constant String oop.
1163 _sp -= 2;
1164 if (stopped()) {
1165 return true;
1166 }
1167
1168 // The null string as a pattern always returns 0 (match at beginning of string)
1169 if (c == 0) {
1170 push(intcon(0));
1171 return true;
1172 }
1173
1174 // Generate default indexOf
1175 jchar lastChar = pat->char_at(o + (c - 1));
1176 int cache = 0;
1177 int i;
1178 for (i = 0; i < c - 1; i++) {
1179 assert(i < pat->length(), "out of range");
1180 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1181 }
1182
1183 int md2 = c;
1184 for (i = 0; i < c - 1; i++) {
1185 assert(i < pat->length(), "out of range");
1186 if (pat->char_at(o + i) == lastChar) {
1187 md2 = (c - 1) - i;
1188 }
1189 }
1190
1191 result = string_indexOf(receiver, pat, o, cache, md2);
1192 }
1193
1194 push(result);
1195 return true;
1196 }
1197
1198 //--------------------------pop_math_arg--------------------------------
1199 // Pop a double argument to a math function from the stack
1200 // rounding it if necessary.
1201 Node * LibraryCallKit::pop_math_arg() {
1202 Node *arg = pop_pair();
1203 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1204 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1205 return arg;
1206 }
1207
1208 //------------------------------inline_trig----------------------------------
1209 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1210 // argument reduction which will turn into a fast/slow diamond.
1211 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1212 _sp += arg_size(); // restore stack pointer
1213 Node* arg = pop_math_arg();
1214 Node* trig = NULL;
1215
1216 switch (id) {
1217 case vmIntrinsics::_dsin:
1218 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1219 break;
1220 case vmIntrinsics::_dcos:
1221 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1222 break;
1223 case vmIntrinsics::_dtan:
1224 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1225 break;
1226 default:
1227 assert(false, "bad intrinsic was passed in");
1228 return false;
1229 }
1230
1231 // Rounding required? Check for argument reduction!
1232 if( Matcher::strict_fp_requires_explicit_rounding ) {
1233
1234 static const double pi_4 = 0.7853981633974483;
1235 static const double neg_pi_4 = -0.7853981633974483;
1236 // pi/2 in 80-bit extended precision
1237 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1238 // -pi/2 in 80-bit extended precision
1239 // 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};
1240 // Cutoff value for using this argument reduction technique
1241 //static const double pi_2_minus_epsilon = 1.564660403643354;
1242 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1243
1244 // Pseudocode for sin:
1245 // if (x <= Math.PI / 4.0) {
1246 // if (x >= -Math.PI / 4.0) return fsin(x);
1247 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1248 // } else {
1249 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1250 // }
1251 // return StrictMath.sin(x);
1252
1253 // Pseudocode for cos:
1254 // if (x <= Math.PI / 4.0) {
1255 // if (x >= -Math.PI / 4.0) return fcos(x);
1256 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1257 // } else {
1258 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1259 // }
1260 // return StrictMath.cos(x);
1261
1262 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1263 // requires a special machine instruction to load it. Instead we'll try
1264 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1265 // probably do the math inside the SIN encoding.
1266
1267 // Make the merge point
1268 RegionNode *r = new (C, 3) RegionNode(3);
1269 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1270
1271 // Flatten arg so we need only 1 test
1272 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1273 // Node for PI/4 constant
1274 Node *pi4 = makecon(TypeD::make(pi_4));
1275 // Check PI/4 : abs(arg)
1276 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1277 // Check: If PI/4 < abs(arg) then go slow
1278 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1279 // Branch either way
1280 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1281 set_control(opt_iff(r,iff));
1282
1283 // Set fast path result
1284 phi->init_req(2,trig);
1285
1286 // Slow path - non-blocking leaf call
1287 Node* call = NULL;
1288 switch (id) {
1289 case vmIntrinsics::_dsin:
1290 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1291 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1292 "Sin", NULL, arg, top());
1293 break;
1294 case vmIntrinsics::_dcos:
1295 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1296 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1297 "Cos", NULL, arg, top());
1298 break;
1299 case vmIntrinsics::_dtan:
1300 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1301 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1302 "Tan", NULL, arg, top());
1303 break;
1304 }
1305 assert(control()->in(0) == call, "");
1306 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1307 r->init_req(1,control());
1308 phi->init_req(1,slow_result);
1309
1310 // Post-merge
1311 set_control(_gvn.transform(r));
1312 record_for_igvn(r);
1313 trig = _gvn.transform(phi);
1314
1315 C->set_has_split_ifs(true); // Has chance for split-if optimization
1316 }
1317 // Push result back on JVM stack
1318 push_pair(trig);
1319 return true;
1320 }
1321
1322 //------------------------------inline_sqrt-------------------------------------
1323 // Inline square root instruction, if possible.
1324 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1325 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1326 _sp += arg_size(); // restore stack pointer
1327 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1328 return true;
1329 }
1330
1331 //------------------------------inline_abs-------------------------------------
1332 // Inline absolute value instruction, if possible.
1333 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1334 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1335 _sp += arg_size(); // restore stack pointer
1336 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1337 return true;
1338 }
1339
1340 //------------------------------inline_exp-------------------------------------
1341 // Inline exp instructions, if possible. The Intel hardware only misses
1342 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1343 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1344 assert(id == vmIntrinsics::_dexp, "Not exp");
1345
1346 // If this inlining ever returned NaN in the past, we do not intrinsify it
1347 // every again. NaN results requires StrictMath.exp handling.
1348 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1349
1350 // Do not intrinsify on older platforms which lack cmove.
1351 if (ConditionalMoveLimit == 0) return false;
1352
1353 _sp += arg_size(); // restore stack pointer
1354 Node *x = pop_math_arg();
1355 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1356
1357 //-------------------
1358 //result=(result.isNaN())? StrictMath::exp():result;
1359 // Check: If isNaN() by checking result!=result? then go to Strict Math
1360 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1361 // Build the boolean node
1362 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1363
1364 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1365 // End the current control-flow path
1366 push_pair(x);
1367 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1368 // to handle. Recompile without intrinsifying Math.exp
1369 uncommon_trap(Deoptimization::Reason_intrinsic,
1370 Deoptimization::Action_make_not_entrant);
1371 }
1372
1373 C->set_has_split_ifs(true); // Has chance for split-if optimization
1374
1375 push_pair(result);
1376
1377 return true;
1378 }
1379
1380 //------------------------------inline_pow-------------------------------------
1381 // Inline power instructions, if possible.
1382 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1383 assert(id == vmIntrinsics::_dpow, "Not pow");
1384
1385 // If this inlining ever returned NaN in the past, we do not intrinsify it
1386 // every again. NaN results requires StrictMath.pow handling.
1387 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1388
1389 // Do not intrinsify on older platforms which lack cmove.
1390 if (ConditionalMoveLimit == 0) return false;
1391
1392 // Pseudocode for pow
1393 // if (x <= 0.0) {
1394 // if ((double)((int)y)==y) { // if y is int
1395 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1396 // } else {
1397 // result = NaN;
1398 // }
1399 // } else {
1400 // result = DPow(x,y);
1401 // }
1402 // if (result != result)? {
1403 // uncommon_trap();
1404 // }
1405 // return result;
1406
1407 _sp += arg_size(); // restore stack pointer
1408 Node* y = pop_math_arg();
1409 Node* x = pop_math_arg();
1410
1411 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1412
1413 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1414 // inside of something) then skip the fancy tests and just check for
1415 // NaN result.
1416 Node *result = NULL;
1417 if( jvms()->depth() >= 1 ) {
1418 result = fast_result;
1419 } else {
1420
1421 // Set the merge point for If node with condition of (x <= 0.0)
1422 // There are four possible paths to region node and phi node
1423 RegionNode *r = new (C, 4) RegionNode(4);
1424 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1425
1426 // Build the first if node: if (x <= 0.0)
1427 // Node for 0 constant
1428 Node *zeronode = makecon(TypeD::ZERO);
1429 // Check x:0
1430 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1431 // Check: If (x<=0) then go complex path
1432 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1433 // Branch either way
1434 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1435 Node *opt_test = _gvn.transform(if1);
1436 //assert( opt_test->is_If(), "Expect an IfNode");
1437 IfNode *opt_if1 = (IfNode*)opt_test;
1438 // Fast path taken; set region slot 3
1439 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1440 r->init_req(3,fast_taken); // Capture fast-control
1441
1442 // Fast path not-taken, i.e. slow path
1443 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1444
1445 // Set fast path result
1446 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1447 phi->init_req(3, fast_result);
1448
1449 // Complex path
1450 // Build the second if node (if y is int)
1451 // Node for (int)y
1452 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1453 // Node for (double)((int) y)
1454 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1455 // Check (double)((int) y) : y
1456 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1457 // Check if (y isn't int) then go to slow path
1458
1459 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1460 // Branch either way
1461 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1462 Node *slow_path = opt_iff(r,if2); // Set region path 2
1463
1464 // Calculate DPow(abs(x), y)*(1 & (int)y)
1465 // Node for constant 1
1466 Node *conone = intcon(1);
1467 // 1& (int)y
1468 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1469 // zero node
1470 Node *conzero = intcon(0);
1471 // Check (1&(int)y)==0?
1472 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1473 // Check if (1&(int)y)!=0?, if so the result is negative
1474 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1475 // abs(x)
1476 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1477 // abs(x)^y
1478 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1479 // -abs(x)^y
1480 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1481 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1482 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1483 // Set complex path fast result
1484 phi->init_req(2, signresult);
1485
1486 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1487 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1488 r->init_req(1,slow_path);
1489 phi->init_req(1,slow_result);
1490
1491 // Post merge
1492 set_control(_gvn.transform(r));
1493 record_for_igvn(r);
1494 result=_gvn.transform(phi);
1495 }
1496
1497 //-------------------
1498 //result=(result.isNaN())? uncommon_trap():result;
1499 // Check: If isNaN() by checking result!=result? then go to Strict Math
1500 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1501 // Build the boolean node
1502 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1503
1504 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1505 // End the current control-flow path
1506 push_pair(x);
1507 push_pair(y);
1508 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1509 // to handle. Recompile without intrinsifying Math.pow.
1510 uncommon_trap(Deoptimization::Reason_intrinsic,
1511 Deoptimization::Action_make_not_entrant);
1512 }
1513
1514 C->set_has_split_ifs(true); // Has chance for split-if optimization
1515
1516 push_pair(result);
1517
1518 return true;
1519 }
1520
1521 //------------------------------inline_trans-------------------------------------
1522 // Inline transcendental instructions, if possible. The Intel hardware gets
1523 // these right, no funny corner cases missed.
1524 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1525 _sp += arg_size(); // restore stack pointer
1526 Node* arg = pop_math_arg();
1527 Node* trans = NULL;
1528
1529 switch (id) {
1530 case vmIntrinsics::_dlog:
1531 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1532 break;
1533 case vmIntrinsics::_dlog10:
1534 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1535 break;
1536 default:
1537 assert(false, "bad intrinsic was passed in");
1538 return false;
1539 }
1540
1541 // Push result back on JVM stack
1542 push_pair(trans);
1543 return true;
1544 }
1545
1546 //------------------------------runtime_math-----------------------------
1547 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1548 Node* a = NULL;
1549 Node* b = NULL;
1550
1551 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1552 "must be (DD)D or (D)D type");
1553
1554 // Inputs
1555 _sp += arg_size(); // restore stack pointer
1556 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1557 b = pop_math_arg();
1558 }
1559 a = pop_math_arg();
1560
1561 const TypePtr* no_memory_effects = NULL;
1562 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1563 no_memory_effects,
1564 a, top(), b, b ? top() : NULL);
1565 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1566 #ifdef ASSERT
1567 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1568 assert(value_top == top(), "second value must be top");
1569 #endif
1570
1571 push_pair(value);
1572 return true;
1573 }
1574
1575 //------------------------------inline_math_native-----------------------------
1576 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1577 switch (id) {
1578 // These intrinsics are not properly supported on all hardware
1579 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1580 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1581 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1582 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1583 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1584 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1585
1586 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1587 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1588 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1589 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1590
1591 // These intrinsics are supported on all hardware
1592 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1593 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1594
1595 // These intrinsics don't work on X86. The ad implementation doesn't
1596 // handle NaN's properly. Instead of returning infinity, the ad
1597 // implementation returns a NaN on overflow. See bug: 6304089
1598 // Once the ad implementations are fixed, change the code below
1599 // to match the intrinsics above
1600
1601 case vmIntrinsics::_dexp: return
1602 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1603 case vmIntrinsics::_dpow: return
1604 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1605
1606 // These intrinsics are not yet correctly implemented
1607 case vmIntrinsics::_datan2:
1608 return false;
1609
1610 default:
1611 ShouldNotReachHere();
1612 return false;
1613 }
1614 }
1615
1616 static bool is_simple_name(Node* n) {
1617 return (n->req() == 1 // constant
1618 || (n->is_Type() && n->as_Type()->type()->singleton())
1619 || n->is_Proj() // parameter or return value
1620 || n->is_Phi() // local of some sort
1621 );
1622 }
1623
1624 //----------------------------inline_min_max-----------------------------------
1625 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1626 push(generate_min_max(id, argument(0), argument(1)));
1627
1628 return true;
1629 }
1630
1631 Node*
1632 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1633 // These are the candidate return value:
1634 Node* xvalue = x0;
1635 Node* yvalue = y0;
1636
1637 if (xvalue == yvalue) {
1638 return xvalue;
1639 }
1640
1641 bool want_max = (id == vmIntrinsics::_max);
1642
1643 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1644 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1645 if (txvalue == NULL || tyvalue == NULL) return top();
1646 // This is not really necessary, but it is consistent with a
1647 // hypothetical MaxINode::Value method:
1648 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1649
1650 // %%% This folding logic should (ideally) be in a different place.
1651 // Some should be inside IfNode, and there to be a more reliable
1652 // transformation of ?: style patterns into cmoves. We also want
1653 // more powerful optimizations around cmove and min/max.
1654
1655 // Try to find a dominating comparison of these guys.
1656 // It can simplify the index computation for Arrays.copyOf
1657 // and similar uses of System.arraycopy.
1658 // First, compute the normalized version of CmpI(x, y).
1659 int cmp_op = Op_CmpI;
1660 Node* xkey = xvalue;
1661 Node* ykey = yvalue;
1662 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1663 if (ideal_cmpxy->is_Cmp()) {
1664 // E.g., if we have CmpI(length - offset, count),
1665 // it might idealize to CmpI(length, count + offset)
1666 cmp_op = ideal_cmpxy->Opcode();
1667 xkey = ideal_cmpxy->in(1);
1668 ykey = ideal_cmpxy->in(2);
1669 }
1670
1671 // Start by locating any relevant comparisons.
1672 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1673 Node* cmpxy = NULL;
1674 Node* cmpyx = NULL;
1675 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1676 Node* cmp = start_from->fast_out(k);
1677 if (cmp->outcnt() > 0 && // must have prior uses
1678 cmp->in(0) == NULL && // must be context-independent
1679 cmp->Opcode() == cmp_op) { // right kind of compare
1680 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1681 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1682 }
1683 }
1684
1685 const int NCMPS = 2;
1686 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1687 int cmpn;
1688 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1689 if (cmps[cmpn] != NULL) break; // find a result
1690 }
1691 if (cmpn < NCMPS) {
1692 // Look for a dominating test that tells us the min and max.
1693 int depth = 0; // Limit search depth for speed
1694 Node* dom = control();
1695 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1696 if (++depth >= 100) break;
1697 Node* ifproj = dom;
1698 if (!ifproj->is_Proj()) continue;
1699 Node* iff = ifproj->in(0);
1700 if (!iff->is_If()) continue;
1701 Node* bol = iff->in(1);
1702 if (!bol->is_Bool()) continue;
1703 Node* cmp = bol->in(1);
1704 if (cmp == NULL) continue;
1705 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1706 if (cmps[cmpn] == cmp) break;
1707 if (cmpn == NCMPS) continue;
1708 BoolTest::mask btest = bol->as_Bool()->_test._test;
1709 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1710 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1711 // At this point, we know that 'x btest y' is true.
1712 switch (btest) {
1713 case BoolTest::eq:
1714 // They are proven equal, so we can collapse the min/max.
1715 // Either value is the answer. Choose the simpler.
1716 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1717 return yvalue;
1718 return xvalue;
1719 case BoolTest::lt: // x < y
1720 case BoolTest::le: // x <= y
1721 return (want_max ? yvalue : xvalue);
1722 case BoolTest::gt: // x > y
1723 case BoolTest::ge: // x >= y
1724 return (want_max ? xvalue : yvalue);
1725 }
1726 }
1727 }
1728
1729 // We failed to find a dominating test.
1730 // Let's pick a test that might GVN with prior tests.
1731 Node* best_bol = NULL;
1732 BoolTest::mask best_btest = BoolTest::illegal;
1733 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1734 Node* cmp = cmps[cmpn];
1735 if (cmp == NULL) continue;
1736 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1737 Node* bol = cmp->fast_out(j);
1738 if (!bol->is_Bool()) continue;
1739 BoolTest::mask btest = bol->as_Bool()->_test._test;
1740 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1741 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1742 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1743 best_bol = bol->as_Bool();
1744 best_btest = btest;
1745 }
1746 }
1747 }
1748
1749 Node* answer_if_true = NULL;
1750 Node* answer_if_false = NULL;
1751 switch (best_btest) {
1752 default:
1753 if (cmpxy == NULL)
1754 cmpxy = ideal_cmpxy;
1755 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1756 // and fall through:
1757 case BoolTest::lt: // x < y
1758 case BoolTest::le: // x <= y
1759 answer_if_true = (want_max ? yvalue : xvalue);
1760 answer_if_false = (want_max ? xvalue : yvalue);
1761 break;
1762 case BoolTest::gt: // x > y
1763 case BoolTest::ge: // x >= y
1764 answer_if_true = (want_max ? xvalue : yvalue);
1765 answer_if_false = (want_max ? yvalue : xvalue);
1766 break;
1767 }
1768
1769 jint hi, lo;
1770 if (want_max) {
1771 // We can sharpen the minimum.
1772 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1773 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1774 } else {
1775 // We can sharpen the maximum.
1776 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1777 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1778 }
1779
1780 // Use a flow-free graph structure, to avoid creating excess control edges
1781 // which could hinder other optimizations.
1782 // Since Math.min/max is often used with arraycopy, we want
1783 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1784 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1785 answer_if_false, answer_if_true,
1786 TypeInt::make(lo, hi, widen));
1787
1788 return _gvn.transform(cmov);
1789
1790 /*
1791 // This is not as desirable as it may seem, since Min and Max
1792 // nodes do not have a full set of optimizations.
1793 // And they would interfere, anyway, with 'if' optimizations
1794 // and with CMoveI canonical forms.
1795 switch (id) {
1796 case vmIntrinsics::_min:
1797 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1798 case vmIntrinsics::_max:
1799 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1800 default:
1801 ShouldNotReachHere();
1802 }
1803 */
1804 }
1805
1806 inline int
1807 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1808 const TypePtr* base_type = TypePtr::NULL_PTR;
1809 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1810 if (base_type == NULL) {
1811 // Unknown type.
1812 return Type::AnyPtr;
1813 } else if (base_type == TypePtr::NULL_PTR) {
1814 // Since this is a NULL+long form, we have to switch to a rawptr.
1815 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1816 offset = MakeConX(0);
1817 return Type::RawPtr;
1818 } else if (base_type->base() == Type::RawPtr) {
1819 return Type::RawPtr;
1820 } else if (base_type->isa_oopptr()) {
1821 // Base is never null => always a heap address.
1822 if (base_type->ptr() == TypePtr::NotNull) {
1823 return Type::OopPtr;
1824 }
1825 // Offset is small => always a heap address.
1826 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1827 if (offset_type != NULL &&
1828 base_type->offset() == 0 && // (should always be?)
1829 offset_type->_lo >= 0 &&
1830 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1831 return Type::OopPtr;
1832 }
1833 // Otherwise, it might either be oop+off or NULL+addr.
1834 return Type::AnyPtr;
1835 } else {
1836 // No information:
1837 return Type::AnyPtr;
1838 }
1839 }
1840
1841 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1842 int kind = classify_unsafe_addr(base, offset);
1843 if (kind == Type::RawPtr) {
1844 return basic_plus_adr(top(), base, offset);
1845 } else {
1846 return basic_plus_adr(base, offset);
1847 }
1848 }
1849
1850 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1851 // inline int Integer.numberOfLeadingZeros(int)
1852 // inline int Long.numberOfLeadingZeros(long)
1853 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1854 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1855 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1856 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1857 _sp += arg_size(); // restore stack pointer
1858 switch (id) {
1859 case vmIntrinsics::_numberOfLeadingZeros_i:
1860 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1861 break;
1862 case vmIntrinsics::_numberOfLeadingZeros_l:
1863 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1864 break;
1865 default:
1866 ShouldNotReachHere();
1867 }
1868 return true;
1869 }
1870
1871 //-------------------inline_numberOfTrailingZeros_int/long----------------------
1872 // inline int Integer.numberOfTrailingZeros(int)
1873 // inline int Long.numberOfTrailingZeros(long)
1874 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
1875 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
1876 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
1877 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
1878 _sp += arg_size(); // restore stack pointer
1879 switch (id) {
1880 case vmIntrinsics::_numberOfTrailingZeros_i:
1881 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
1882 break;
1883 case vmIntrinsics::_numberOfTrailingZeros_l:
1884 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
1885 break;
1886 default:
1887 ShouldNotReachHere();
1888 }
1889 return true;
1890 }
1891
1892 //----------------------------inline_bitCount_int/long-----------------------
1893 // inline int Integer.bitCount(int)
1894 // inline int Long.bitCount(long)
1895 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
1896 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
1897 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
1898 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
1899 _sp += arg_size(); // restore stack pointer
1900 switch (id) {
1901 case vmIntrinsics::_bitCount_i:
1902 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
1903 break;
1904 case vmIntrinsics::_bitCount_l:
1905 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
1906 break;
1907 default:
1908 ShouldNotReachHere();
1909 }
1910 return true;
1911 }
1912
1913 //----------------------------inline_reverseBytes_int/long-------------------
1914 // inline Integer.reverseBytes(int)
1915 // inline Long.reverseBytes(long)
1916 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
1917 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
1918 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
1919 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
1920 _sp += arg_size(); // restore stack pointer
1921 switch (id) {
1922 case vmIntrinsics::_reverseBytes_i:
1923 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
1924 break;
1925 case vmIntrinsics::_reverseBytes_l:
1926 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
1927 break;
1928 default:
1929 ;
1930 }
1931 return true;
1932 }
1933
1934 //----------------------------inline_unsafe_access----------------------------
1935
1936 const static BasicType T_ADDRESS_HOLDER = T_LONG;
1937
1938 // Interpret Unsafe.fieldOffset cookies correctly:
1939 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
1940
1941 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
1942 if (callee()->is_static()) return false; // caller must have the capability!
1943
1944 #ifndef PRODUCT
1945 {
1946 ResourceMark rm;
1947 // Check the signatures.
1948 ciSignature* sig = signature();
1949 #ifdef ASSERT
1950 if (!is_store) {
1951 // Object getObject(Object base, int/long offset), etc.
1952 BasicType rtype = sig->return_type()->basic_type();
1953 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
1954 rtype = T_ADDRESS; // it is really a C void*
1955 assert(rtype == type, "getter must return the expected value");
1956 if (!is_native_ptr) {
1957 assert(sig->count() == 2, "oop getter has 2 arguments");
1958 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
1959 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
1960 } else {
1961 assert(sig->count() == 1, "native getter has 1 argument");
1962 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
1963 }
1964 } else {
1965 // void putObject(Object base, int/long offset, Object x), etc.
1966 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
1967 if (!is_native_ptr) {
1968 assert(sig->count() == 3, "oop putter has 3 arguments");
1969 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
1970 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
1971 } else {
1972 assert(sig->count() == 2, "native putter has 2 arguments");
1973 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
1974 }
1975 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
1976 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
1977 vtype = T_ADDRESS; // it is really a C void*
1978 assert(vtype == type, "putter must accept the expected value");
1979 }
1980 #endif // ASSERT
1981 }
1982 #endif //PRODUCT
1983
1984 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1985
1986 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
1987
1988 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
1989 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
1990
1991 debug_only(int saved_sp = _sp);
1992 _sp += nargs;
1993
1994 Node* val;
1995 debug_only(val = (Node*)(uintptr_t)-1);
1996
1997
1998 if (is_store) {
1999 // Get the value being stored. (Pop it first; it was pushed last.)
2000 switch (type) {
2001 case T_DOUBLE:
2002 case T_LONG:
2003 case T_ADDRESS:
2004 val = pop_pair();
2005 break;
2006 default:
2007 val = pop();
2008 }
2009 }
2010
2011 // Build address expression. See the code in inline_unsafe_prefetch.
2012 Node *adr;
2013 Node *heap_base_oop = top();
2014 if (!is_native_ptr) {
2015 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2016 Node* offset = pop_pair();
2017 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2018 Node* base = pop();
2019 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2020 // to be plain byte offsets, which are also the same as those accepted
2021 // by oopDesc::field_base.
2022 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2023 "fieldOffset must be byte-scaled");
2024 // 32-bit machines ignore the high half!
2025 offset = ConvL2X(offset);
2026 adr = make_unsafe_address(base, offset);
2027 heap_base_oop = base;
2028 } else {
2029 Node* ptr = pop_pair();
2030 // Adjust Java long to machine word:
2031 ptr = ConvL2X(ptr);
2032 adr = make_unsafe_address(NULL, ptr);
2033 }
2034
2035 // Pop receiver last: it was pushed first.
2036 Node *receiver = pop();
2037
2038 assert(saved_sp == _sp, "must have correct argument count");
2039
2040 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2041
2042 // First guess at the value type.
2043 const Type *value_type = Type::get_const_basic_type(type);
2044
2045 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2046 // there was not enough information to nail it down.
2047 Compile::AliasType* alias_type = C->alias_type(adr_type);
2048 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2049
2050 // We will need memory barriers unless we can determine a unique
2051 // alias category for this reference. (Note: If for some reason
2052 // the barriers get omitted and the unsafe reference begins to "pollute"
2053 // the alias analysis of the rest of the graph, either Compile::can_alias
2054 // or Compile::must_alias will throw a diagnostic assert.)
2055 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2056
2057 if (!is_store && type == T_OBJECT) {
2058 // Attempt to infer a sharper value type from the offset and base type.
2059 ciKlass* sharpened_klass = NULL;
2060
2061 // See if it is an instance field, with an object type.
2062 if (alias_type->field() != NULL) {
2063 assert(!is_native_ptr, "native pointer op cannot use a java address");
2064 if (alias_type->field()->type()->is_klass()) {
2065 sharpened_klass = alias_type->field()->type()->as_klass();
2066 }
2067 }
2068
2069 // See if it is a narrow oop array.
2070 if (adr_type->isa_aryptr()) {
2071 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2072 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2073 if (elem_type != NULL) {
2074 sharpened_klass = elem_type->klass();
2075 }
2076 }
2077 }
2078
2079 if (sharpened_klass != NULL) {
2080 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2081
2082 // Sharpen the value type.
2083 value_type = tjp;
2084
2085 #ifndef PRODUCT
2086 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2087 tty->print(" from base type: "); adr_type->dump();
2088 tty->print(" sharpened value: "); value_type->dump();
2089 }
2090 #endif
2091 }
2092 }
2093
2094 // Null check on self without removing any arguments. The argument
2095 // null check technically happens in the wrong place, which can lead to
2096 // invalid stack traces when the primitive is inlined into a method
2097 // which handles NullPointerExceptions.
2098 _sp += nargs;
2099 do_null_check(receiver, T_OBJECT);
2100 _sp -= nargs;
2101 if (stopped()) {
2102 return true;
2103 }
2104 // Heap pointers get a null-check from the interpreter,
2105 // as a courtesy. However, this is not guaranteed by Unsafe,
2106 // and it is not possible to fully distinguish unintended nulls
2107 // from intended ones in this API.
2108
2109 if (is_volatile) {
2110 // We need to emit leading and trailing CPU membars (see below) in
2111 // addition to memory membars when is_volatile. This is a little
2112 // too strong, but avoids the need to insert per-alias-type
2113 // volatile membars (for stores; compare Parse::do_put_xxx), which
2114 // we cannot do effectively here because we probably only have a
2115 // rough approximation of type.
2116 need_mem_bar = true;
2117 // For Stores, place a memory ordering barrier now.
2118 if (is_store)
2119 insert_mem_bar(Op_MemBarRelease);
2120 }
2121
2122 // Memory barrier to prevent normal and 'unsafe' accesses from
2123 // bypassing each other. Happens after null checks, so the
2124 // exception paths do not take memory state from the memory barrier,
2125 // so there's no problems making a strong assert about mixing users
2126 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2127 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2128 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2129
2130 if (!is_store) {
2131 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2132 // load value and push onto stack
2133 switch (type) {
2134 case T_BOOLEAN:
2135 case T_CHAR:
2136 case T_BYTE:
2137 case T_SHORT:
2138 case T_INT:
2139 case T_FLOAT:
2140 case T_OBJECT:
2141 push( p );
2142 break;
2143 case T_ADDRESS:
2144 // Cast to an int type.
2145 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2146 p = ConvX2L(p);
2147 push_pair(p);
2148 break;
2149 case T_DOUBLE:
2150 case T_LONG:
2151 push_pair( p );
2152 break;
2153 default: ShouldNotReachHere();
2154 }
2155 } else {
2156 // place effect of store into memory
2157 switch (type) {
2158 case T_DOUBLE:
2159 val = dstore_rounding(val);
2160 break;
2161 case T_ADDRESS:
2162 // Repackage the long as a pointer.
2163 val = ConvL2X(val);
2164 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2165 break;
2166 }
2167
2168 if (type != T_OBJECT ) {
2169 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2170 } else {
2171 // Possibly an oop being stored to Java heap or native memory
2172 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2173 // oop to Java heap.
2174 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2175 } else {
2176 // We can't tell at compile time if we are storing in the Java heap or outside
2177 // of it. So we need to emit code to conditionally do the proper type of
2178 // store.
2179
2180 IdealKit ideal(gvn(), control(), merged_memory());
2181 #define __ ideal.
2182 // QQQ who knows what probability is here??
2183 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2184 // Sync IdealKit and graphKit.
2185 set_all_memory( __ merged_memory());
2186 set_control(__ ctrl());
2187 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2188 // Update IdealKit memory.
2189 __ set_all_memory(merged_memory());
2190 __ set_ctrl(control());
2191 } __ else_(); {
2192 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2193 } __ end_if();
2194 // Final sync IdealKit and GraphKit.
2195 sync_kit(ideal);
2196 #undef __
2197 }
2198 }
2199 }
2200
2201 if (is_volatile) {
2202 if (!is_store)
2203 insert_mem_bar(Op_MemBarAcquire);
2204 else
2205 insert_mem_bar(Op_MemBarVolatile);
2206 }
2207
2208 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2209
2210 return true;
2211 }
2212
2213 //----------------------------inline_unsafe_prefetch----------------------------
2214
2215 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2216 #ifndef PRODUCT
2217 {
2218 ResourceMark rm;
2219 // Check the signatures.
2220 ciSignature* sig = signature();
2221 #ifdef ASSERT
2222 // Object getObject(Object base, int/long offset), etc.
2223 BasicType rtype = sig->return_type()->basic_type();
2224 if (!is_native_ptr) {
2225 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2226 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2227 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2228 } else {
2229 assert(sig->count() == 1, "native prefetch has 1 argument");
2230 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2231 }
2232 #endif // ASSERT
2233 }
2234 #endif // !PRODUCT
2235
2236 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2237
2238 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2239 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2240
2241 debug_only(int saved_sp = _sp);
2242 _sp += nargs;
2243
2244 // Build address expression. See the code in inline_unsafe_access.
2245 Node *adr;
2246 if (!is_native_ptr) {
2247 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2248 Node* offset = pop_pair();
2249 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2250 Node* base = pop();
2251 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2252 // to be plain byte offsets, which are also the same as those accepted
2253 // by oopDesc::field_base.
2254 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2255 "fieldOffset must be byte-scaled");
2256 // 32-bit machines ignore the high half!
2257 offset = ConvL2X(offset);
2258 adr = make_unsafe_address(base, offset);
2259 } else {
2260 Node* ptr = pop_pair();
2261 // Adjust Java long to machine word:
2262 ptr = ConvL2X(ptr);
2263 adr = make_unsafe_address(NULL, ptr);
2264 }
2265
2266 if (is_static) {
2267 assert(saved_sp == _sp, "must have correct argument count");
2268 } else {
2269 // Pop receiver last: it was pushed first.
2270 Node *receiver = pop();
2271 assert(saved_sp == _sp, "must have correct argument count");
2272
2273 // Null check on self without removing any arguments. The argument
2274 // null check technically happens in the wrong place, which can lead to
2275 // invalid stack traces when the primitive is inlined into a method
2276 // which handles NullPointerExceptions.
2277 _sp += nargs;
2278 do_null_check(receiver, T_OBJECT);
2279 _sp -= nargs;
2280 if (stopped()) {
2281 return true;
2282 }
2283 }
2284
2285 // Generate the read or write prefetch
2286 Node *prefetch;
2287 if (is_store) {
2288 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2289 } else {
2290 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2291 }
2292 prefetch->init_req(0, control());
2293 set_i_o(_gvn.transform(prefetch));
2294
2295 return true;
2296 }
2297
2298 //----------------------------inline_unsafe_CAS----------------------------
2299
2300 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2301 // This basic scheme here is the same as inline_unsafe_access, but
2302 // differs in enough details that combining them would make the code
2303 // overly confusing. (This is a true fact! I originally combined
2304 // them, but even I was confused by it!) As much code/comments as
2305 // possible are retained from inline_unsafe_access though to make
2306 // the correspondences clearer. - dl
2307
2308 if (callee()->is_static()) return false; // caller must have the capability!
2309
2310 #ifndef PRODUCT
2311 {
2312 ResourceMark rm;
2313 // Check the signatures.
2314 ciSignature* sig = signature();
2315 #ifdef ASSERT
2316 BasicType rtype = sig->return_type()->basic_type();
2317 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2318 assert(sig->count() == 4, "CAS has 4 arguments");
2319 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2320 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2321 #endif // ASSERT
2322 }
2323 #endif //PRODUCT
2324
2325 // number of stack slots per value argument (1 or 2)
2326 int type_words = type2size[type];
2327
2328 // Cannot inline wide CAS on machines that don't support it natively
2329 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2330 return false;
2331
2332 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2333
2334 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2335 int nargs = 1 + 1 + 2 + type_words + type_words;
2336
2337 // pop arguments: newval, oldval, offset, base, and receiver
2338 debug_only(int saved_sp = _sp);
2339 _sp += nargs;
2340 Node* newval = (type_words == 1) ? pop() : pop_pair();
2341 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2342 Node *offset = pop_pair();
2343 Node *base = pop();
2344 Node *receiver = pop();
2345 assert(saved_sp == _sp, "must have correct argument count");
2346
2347 // Null check receiver.
2348 _sp += nargs;
2349 do_null_check(receiver, T_OBJECT);
2350 _sp -= nargs;
2351 if (stopped()) {
2352 return true;
2353 }
2354
2355 // Build field offset expression.
2356 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2357 // to be plain byte offsets, which are also the same as those accepted
2358 // by oopDesc::field_base.
2359 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2360 // 32-bit machines ignore the high half of long offsets
2361 offset = ConvL2X(offset);
2362 Node* adr = make_unsafe_address(base, offset);
2363 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2364
2365 // (Unlike inline_unsafe_access, there seems no point in trying
2366 // to refine types. Just use the coarse types here.
2367 const Type *value_type = Type::get_const_basic_type(type);
2368 Compile::AliasType* alias_type = C->alias_type(adr_type);
2369 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2370 int alias_idx = C->get_alias_index(adr_type);
2371
2372 // Memory-model-wise, a CAS acts like a little synchronized block,
2373 // so needs barriers on each side. These don't translate into
2374 // actual barriers on most machines, but we still need rest of
2375 // compiler to respect ordering.
2376
2377 insert_mem_bar(Op_MemBarRelease);
2378 insert_mem_bar(Op_MemBarCPUOrder);
2379
2380 // 4984716: MemBars must be inserted before this
2381 // memory node in order to avoid a false
2382 // dependency which will confuse the scheduler.
2383 Node *mem = memory(alias_idx);
2384
2385 // For now, we handle only those cases that actually exist: ints,
2386 // longs, and Object. Adding others should be straightforward.
2387 Node* cas;
2388 switch(type) {
2389 case T_INT:
2390 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2391 break;
2392 case T_LONG:
2393 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2394 break;
2395 case T_OBJECT:
2396 // reference stores need a store barrier.
2397 // (They don't if CAS fails, but it isn't worth checking.)
2398 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2399 #ifdef _LP64
2400 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2401 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2402 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2403 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2404 newval_enc, oldval_enc));
2405 } else
2406 #endif
2407 {
2408 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2409 }
2410 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2411 break;
2412 default:
2413 ShouldNotReachHere();
2414 break;
2415 }
2416
2417 // SCMemProjNodes represent the memory state of CAS. Their main
2418 // role is to prevent CAS nodes from being optimized away when their
2419 // results aren't used.
2420 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2421 set_memory(proj, alias_idx);
2422
2423 // Add the trailing membar surrounding the access
2424 insert_mem_bar(Op_MemBarCPUOrder);
2425 insert_mem_bar(Op_MemBarAcquire);
2426
2427 push(cas);
2428 return true;
2429 }
2430
2431 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2432 // This is another variant of inline_unsafe_access, differing in
2433 // that it always issues store-store ("release") barrier and ensures
2434 // store-atomicity (which only matters for "long").
2435
2436 if (callee()->is_static()) return false; // caller must have the capability!
2437
2438 #ifndef PRODUCT
2439 {
2440 ResourceMark rm;
2441 // Check the signatures.
2442 ciSignature* sig = signature();
2443 #ifdef ASSERT
2444 BasicType rtype = sig->return_type()->basic_type();
2445 assert(rtype == T_VOID, "must return void");
2446 assert(sig->count() == 3, "has 3 arguments");
2447 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2448 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2449 #endif // ASSERT
2450 }
2451 #endif //PRODUCT
2452
2453 // number of stack slots per value argument (1 or 2)
2454 int type_words = type2size[type];
2455
2456 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2457
2458 // Argument words: "this" plus oop plus offset plus value;
2459 int nargs = 1 + 1 + 2 + type_words;
2460
2461 // pop arguments: val, offset, base, and receiver
2462 debug_only(int saved_sp = _sp);
2463 _sp += nargs;
2464 Node* val = (type_words == 1) ? pop() : pop_pair();
2465 Node *offset = pop_pair();
2466 Node *base = pop();
2467 Node *receiver = pop();
2468 assert(saved_sp == _sp, "must have correct argument count");
2469
2470 // Null check receiver.
2471 _sp += nargs;
2472 do_null_check(receiver, T_OBJECT);
2473 _sp -= nargs;
2474 if (stopped()) {
2475 return true;
2476 }
2477
2478 // Build field offset expression.
2479 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2480 // 32-bit machines ignore the high half of long offsets
2481 offset = ConvL2X(offset);
2482 Node* adr = make_unsafe_address(base, offset);
2483 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2484 const Type *value_type = Type::get_const_basic_type(type);
2485 Compile::AliasType* alias_type = C->alias_type(adr_type);
2486
2487 insert_mem_bar(Op_MemBarRelease);
2488 insert_mem_bar(Op_MemBarCPUOrder);
2489 // Ensure that the store is atomic for longs:
2490 bool require_atomic_access = true;
2491 Node* store;
2492 if (type == T_OBJECT) // reference stores need a store barrier.
2493 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2494 else {
2495 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2496 }
2497 insert_mem_bar(Op_MemBarCPUOrder);
2498 return true;
2499 }
2500
2501 bool LibraryCallKit::inline_unsafe_allocate() {
2502 if (callee()->is_static()) return false; // caller must have the capability!
2503 int nargs = 1 + 1;
2504 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2505 null_check_receiver(callee()); // check then ignore argument(0)
2506 _sp += nargs; // set original stack for use by uncommon_trap
2507 Node* cls = do_null_check(argument(1), T_OBJECT);
2508 _sp -= nargs;
2509 if (stopped()) return true;
2510
2511 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2512 _sp += nargs; // set original stack for use by uncommon_trap
2513 kls = do_null_check(kls, T_OBJECT);
2514 _sp -= nargs;
2515 if (stopped()) return true; // argument was like int.class
2516
2517 // Note: The argument might still be an illegal value like
2518 // Serializable.class or Object[].class. The runtime will handle it.
2519 // But we must make an explicit check for initialization.
2520 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2521 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2522 Node* bits = intcon(instanceKlass::fully_initialized);
2523 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2524 // The 'test' is non-zero if we need to take a slow path.
2525
2526 Node* obj = new_instance(kls, test);
2527 push(obj);
2528
2529 return true;
2530 }
2531
2532 //------------------------inline_native_time_funcs--------------
2533 // inline code for System.currentTimeMillis() and System.nanoTime()
2534 // these have the same type and signature
2535 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2536 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2537 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2538 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2539 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2540 const TypePtr* no_memory_effects = NULL;
2541 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2542 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2543 #ifdef ASSERT
2544 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2545 assert(value_top == top(), "second value must be top");
2546 #endif
2547 push_pair(value);
2548 return true;
2549 }
2550
2551 //------------------------inline_native_currentThread------------------
2552 bool LibraryCallKit::inline_native_currentThread() {
2553 Node* junk = NULL;
2554 push(generate_current_thread(junk));
2555 return true;
2556 }
2557
2558 //------------------------inline_native_isInterrupted------------------
2559 bool LibraryCallKit::inline_native_isInterrupted() {
2560 const int nargs = 1+1; // receiver + boolean
2561 assert(nargs == arg_size(), "sanity");
2562 // Add a fast path to t.isInterrupted(clear_int):
2563 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2564 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2565 // So, in the common case that the interrupt bit is false,
2566 // we avoid making a call into the VM. Even if the interrupt bit
2567 // is true, if the clear_int argument is false, we avoid the VM call.
2568 // However, if the receiver is not currentThread, we must call the VM,
2569 // because there must be some locking done around the operation.
2570
2571 // We only go to the fast case code if we pass two guards.
2572 // Paths which do not pass are accumulated in the slow_region.
2573 RegionNode* slow_region = new (C, 1) RegionNode(1);
2574 record_for_igvn(slow_region);
2575 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2576 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2577 enum { no_int_result_path = 1,
2578 no_clear_result_path = 2,
2579 slow_result_path = 3
2580 };
2581
2582 // (a) Receiving thread must be the current thread.
2583 Node* rec_thr = argument(0);
2584 Node* tls_ptr = NULL;
2585 Node* cur_thr = generate_current_thread(tls_ptr);
2586 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2587 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2588
2589 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2590 if (!known_current_thread)
2591 generate_slow_guard(bol_thr, slow_region);
2592
2593 // (b) Interrupt bit on TLS must be false.
2594 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2595 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2596 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2597 // Set the control input on the field _interrupted read to prevent it floating up.
2598 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2599 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2600 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2601
2602 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2603
2604 // First fast path: if (!TLS._interrupted) return false;
2605 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2606 result_rgn->init_req(no_int_result_path, false_bit);
2607 result_val->init_req(no_int_result_path, intcon(0));
2608
2609 // drop through to next case
2610 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2611
2612 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2613 Node* clr_arg = argument(1);
2614 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2615 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2616 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2617
2618 // Second fast path: ... else if (!clear_int) return true;
2619 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2620 result_rgn->init_req(no_clear_result_path, false_arg);
2621 result_val->init_req(no_clear_result_path, intcon(1));
2622
2623 // drop through to next case
2624 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2625
2626 // (d) Otherwise, go to the slow path.
2627 slow_region->add_req(control());
2628 set_control( _gvn.transform(slow_region) );
2629
2630 if (stopped()) {
2631 // There is no slow path.
2632 result_rgn->init_req(slow_result_path, top());
2633 result_val->init_req(slow_result_path, top());
2634 } else {
2635 // non-virtual because it is a private non-static
2636 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2637
2638 Node* slow_val = set_results_for_java_call(slow_call);
2639 // this->control() comes from set_results_for_java_call
2640
2641 // If we know that the result of the slow call will be true, tell the optimizer!
2642 if (known_current_thread) slow_val = intcon(1);
2643
2644 Node* fast_io = slow_call->in(TypeFunc::I_O);
2645 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2646 // These two phis are pre-filled with copies of of the fast IO and Memory
2647 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2648 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2649
2650 result_rgn->init_req(slow_result_path, control());
2651 io_phi ->init_req(slow_result_path, i_o());
2652 mem_phi ->init_req(slow_result_path, reset_memory());
2653 result_val->init_req(slow_result_path, slow_val);
2654
2655 set_all_memory( _gvn.transform(mem_phi) );
2656 set_i_o( _gvn.transform(io_phi) );
2657 }
2658
2659 push_result(result_rgn, result_val);
2660 C->set_has_split_ifs(true); // Has chance for split-if optimization
2661
2662 return true;
2663 }
2664
2665 //---------------------------load_mirror_from_klass----------------------------
2666 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2667 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2668 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2669 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2670 }
2671
2672 //-----------------------load_klass_from_mirror_common-------------------------
2673 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2674 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2675 // and branch to the given path on the region.
2676 // If never_see_null, take an uncommon trap on null, so we can optimistically
2677 // compile for the non-null case.
2678 // If the region is NULL, force never_see_null = true.
2679 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2680 bool never_see_null,
2681 int nargs,
2682 RegionNode* region,
2683 int null_path,
2684 int offset) {
2685 if (region == NULL) never_see_null = true;
2686 Node* p = basic_plus_adr(mirror, offset);
2687 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2688 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2689 _sp += nargs; // any deopt will start just before call to enclosing method
2690 Node* null_ctl = top();
2691 kls = null_check_oop(kls, &null_ctl, never_see_null);
2692 if (region != NULL) {
2693 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2694 region->init_req(null_path, null_ctl);
2695 } else {
2696 assert(null_ctl == top(), "no loose ends");
2697 }
2698 _sp -= nargs;
2699 return kls;
2700 }
2701
2702 //--------------------(inline_native_Class_query helpers)---------------------
2703 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2704 // Fall through if (mods & mask) == bits, take the guard otherwise.
2705 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2706 // Branch around if the given klass has the given modifier bit set.
2707 // Like generate_guard, adds a new path onto the region.
2708 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2709 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2710 Node* mask = intcon(modifier_mask);
2711 Node* bits = intcon(modifier_bits);
2712 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2713 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2714 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2715 return generate_fair_guard(bol, region);
2716 }
2717 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2718 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2719 }
2720
2721 //-------------------------inline_native_Class_query-------------------
2722 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2723 int nargs = 1+0; // just the Class mirror, in most cases
2724 const Type* return_type = TypeInt::BOOL;
2725 Node* prim_return_value = top(); // what happens if it's a primitive class?
2726 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2727 bool expect_prim = false; // most of these guys expect to work on refs
2728
2729 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2730
2731 switch (id) {
2732 case vmIntrinsics::_isInstance:
2733 nargs = 1+1; // the Class mirror, plus the object getting queried about
2734 // nothing is an instance of a primitive type
2735 prim_return_value = intcon(0);
2736 break;
2737 case vmIntrinsics::_getModifiers:
2738 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2739 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2740 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2741 break;
2742 case vmIntrinsics::_isInterface:
2743 prim_return_value = intcon(0);
2744 break;
2745 case vmIntrinsics::_isArray:
2746 prim_return_value = intcon(0);
2747 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2748 break;
2749 case vmIntrinsics::_isPrimitive:
2750 prim_return_value = intcon(1);
2751 expect_prim = true; // obviously
2752 break;
2753 case vmIntrinsics::_getSuperclass:
2754 prim_return_value = null();
2755 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2756 break;
2757 case vmIntrinsics::_getComponentType:
2758 prim_return_value = null();
2759 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2760 break;
2761 case vmIntrinsics::_getClassAccessFlags:
2762 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2763 return_type = TypeInt::INT; // not bool! 6297094
2764 break;
2765 default:
2766 ShouldNotReachHere();
2767 }
2768
2769 Node* mirror = argument(0);
2770 Node* obj = (nargs <= 1)? top(): argument(1);
2771
2772 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2773 if (mirror_con == NULL) return false; // cannot happen?
2774
2775 #ifndef PRODUCT
2776 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2777 ciType* k = mirror_con->java_mirror_type();
2778 if (k) {
2779 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2780 k->print_name();
2781 tty->cr();
2782 }
2783 }
2784 #endif
2785
2786 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2787 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2788 record_for_igvn(region);
2789 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2790
2791 // The mirror will never be null of Reflection.getClassAccessFlags, however
2792 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2793 // if it is. See bug 4774291.
2794
2795 // For Reflection.getClassAccessFlags(), the null check occurs in
2796 // the wrong place; see inline_unsafe_access(), above, for a similar
2797 // situation.
2798 _sp += nargs; // set original stack for use by uncommon_trap
2799 mirror = do_null_check(mirror, T_OBJECT);
2800 _sp -= nargs;
2801 // If mirror or obj is dead, only null-path is taken.
2802 if (stopped()) return true;
2803
2804 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2805
2806 // Now load the mirror's klass metaobject, and null-check it.
2807 // Side-effects region with the control path if the klass is null.
2808 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2809 region, _prim_path);
2810 // If kls is null, we have a primitive mirror.
2811 phi->init_req(_prim_path, prim_return_value);
2812 if (stopped()) { push_result(region, phi); return true; }
2813
2814 Node* p; // handy temp
2815 Node* null_ctl;
2816
2817 // Now that we have the non-null klass, we can perform the real query.
2818 // For constant classes, the query will constant-fold in LoadNode::Value.
2819 Node* query_value = top();
2820 switch (id) {
2821 case vmIntrinsics::_isInstance:
2822 // nothing is an instance of a primitive type
2823 query_value = gen_instanceof(obj, kls);
2824 break;
2825
2826 case vmIntrinsics::_getModifiers:
2827 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2828 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2829 break;
2830
2831 case vmIntrinsics::_isInterface:
2832 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2833 if (generate_interface_guard(kls, region) != NULL)
2834 // A guard was added. If the guard is taken, it was an interface.
2835 phi->add_req(intcon(1));
2836 // If we fall through, it's a plain class.
2837 query_value = intcon(0);
2838 break;
2839
2840 case vmIntrinsics::_isArray:
2841 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2842 if (generate_array_guard(kls, region) != NULL)
2843 // A guard was added. If the guard is taken, it was an array.
2844 phi->add_req(intcon(1));
2845 // If we fall through, it's a plain class.
2846 query_value = intcon(0);
2847 break;
2848
2849 case vmIntrinsics::_isPrimitive:
2850 query_value = intcon(0); // "normal" path produces false
2851 break;
2852
2853 case vmIntrinsics::_getSuperclass:
2854 // The rules here are somewhat unfortunate, but we can still do better
2855 // with random logic than with a JNI call.
2856 // Interfaces store null or Object as _super, but must report null.
2857 // Arrays store an intermediate super as _super, but must report Object.
2858 // Other types can report the actual _super.
2859 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2860 if (generate_interface_guard(kls, region) != NULL)
2861 // A guard was added. If the guard is taken, it was an interface.
2862 phi->add_req(null());
2863 if (generate_array_guard(kls, region) != NULL)
2864 // A guard was added. If the guard is taken, it was an array.
2865 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2866 // If we fall through, it's a plain class. Get its _super.
2867 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2868 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2869 null_ctl = top();
2870 kls = null_check_oop(kls, &null_ctl);
2871 if (null_ctl != top()) {
2872 // If the guard is taken, Object.superClass is null (both klass and mirror).
2873 region->add_req(null_ctl);
2874 phi ->add_req(null());
2875 }
2876 if (!stopped()) {
2877 query_value = load_mirror_from_klass(kls);
2878 }
2879 break;
2880
2881 case vmIntrinsics::_getComponentType:
2882 if (generate_array_guard(kls, region) != NULL) {
2883 // Be sure to pin the oop load to the guard edge just created:
2884 Node* is_array_ctrl = region->in(region->req()-1);
2885 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
2886 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
2887 phi->add_req(cmo);
2888 }
2889 query_value = null(); // non-array case is null
2890 break;
2891
2892 case vmIntrinsics::_getClassAccessFlags:
2893 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2894 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2895 break;
2896
2897 default:
2898 ShouldNotReachHere();
2899 }
2900
2901 // Fall-through is the normal case of a query to a real class.
2902 phi->init_req(1, query_value);
2903 region->init_req(1, control());
2904
2905 push_result(region, phi);
2906 C->set_has_split_ifs(true); // Has chance for split-if optimization
2907
2908 return true;
2909 }
2910
2911 //--------------------------inline_native_subtype_check------------------------
2912 // This intrinsic takes the JNI calls out of the heart of
2913 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
2914 bool LibraryCallKit::inline_native_subtype_check() {
2915 int nargs = 1+1; // the Class mirror, plus the other class getting examined
2916
2917 // Pull both arguments off the stack.
2918 Node* args[2]; // two java.lang.Class mirrors: superc, subc
2919 args[0] = argument(0);
2920 args[1] = argument(1);
2921 Node* klasses[2]; // corresponding Klasses: superk, subk
2922 klasses[0] = klasses[1] = top();
2923
2924 enum {
2925 // A full decision tree on {superc is prim, subc is prim}:
2926 _prim_0_path = 1, // {P,N} => false
2927 // {P,P} & superc!=subc => false
2928 _prim_same_path, // {P,P} & superc==subc => true
2929 _prim_1_path, // {N,P} => false
2930 _ref_subtype_path, // {N,N} & subtype check wins => true
2931 _both_ref_path, // {N,N} & subtype check loses => false
2932 PATH_LIMIT
2933 };
2934
2935 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2936 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2937 record_for_igvn(region);
2938
2939 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
2940 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2941 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
2942
2943 // First null-check both mirrors and load each mirror's klass metaobject.
2944 int which_arg;
2945 for (which_arg = 0; which_arg <= 1; which_arg++) {
2946 Node* arg = args[which_arg];
2947 _sp += nargs; // set original stack for use by uncommon_trap
2948 arg = do_null_check(arg, T_OBJECT);
2949 _sp -= nargs;
2950 if (stopped()) break;
2951 args[which_arg] = _gvn.transform(arg);
2952
2953 Node* p = basic_plus_adr(arg, class_klass_offset);
2954 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
2955 klasses[which_arg] = _gvn.transform(kls);
2956 }
2957
2958 // Having loaded both klasses, test each for null.
2959 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2960 for (which_arg = 0; which_arg <= 1; which_arg++) {
2961 Node* kls = klasses[which_arg];
2962 Node* null_ctl = top();
2963 _sp += nargs; // set original stack for use by uncommon_trap
2964 kls = null_check_oop(kls, &null_ctl, never_see_null);
2965 _sp -= nargs;
2966 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
2967 region->init_req(prim_path, null_ctl);
2968 if (stopped()) break;
2969 klasses[which_arg] = kls;
2970 }
2971
2972 if (!stopped()) {
2973 // now we have two reference types, in klasses[0..1]
2974 Node* subk = klasses[1]; // the argument to isAssignableFrom
2975 Node* superk = klasses[0]; // the receiver
2976 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
2977 // now we have a successful reference subtype check
2978 region->set_req(_ref_subtype_path, control());
2979 }
2980
2981 // If both operands are primitive (both klasses null), then
2982 // we must return true when they are identical primitives.
2983 // It is convenient to test this after the first null klass check.
2984 set_control(region->in(_prim_0_path)); // go back to first null check
2985 if (!stopped()) {
2986 // Since superc is primitive, make a guard for the superc==subc case.
2987 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
2988 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
2989 generate_guard(bol_eq, region, PROB_FAIR);
2990 if (region->req() == PATH_LIMIT+1) {
2991 // A guard was added. If the added guard is taken, superc==subc.
2992 region->swap_edges(PATH_LIMIT, _prim_same_path);
2993 region->del_req(PATH_LIMIT);
2994 }
2995 region->set_req(_prim_0_path, control()); // Not equal after all.
2996 }
2997
2998 // these are the only paths that produce 'true':
2999 phi->set_req(_prim_same_path, intcon(1));
3000 phi->set_req(_ref_subtype_path, intcon(1));
3001
3002 // pull together the cases:
3003 assert(region->req() == PATH_LIMIT, "sane region");
3004 for (uint i = 1; i < region->req(); i++) {
3005 Node* ctl = region->in(i);
3006 if (ctl == NULL || ctl == top()) {
3007 region->set_req(i, top());
3008 phi ->set_req(i, top());
3009 } else if (phi->in(i) == NULL) {
3010 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3011 }
3012 }
3013
3014 set_control(_gvn.transform(region));
3015 push(_gvn.transform(phi));
3016
3017 return true;
3018 }
3019
3020 //---------------------generate_array_guard_common------------------------
3021 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3022 bool obj_array, bool not_array) {
3023 // If obj_array/non_array==false/false:
3024 // Branch around if the given klass is in fact an array (either obj or prim).
3025 // If obj_array/non_array==false/true:
3026 // Branch around if the given klass is not an array klass of any kind.
3027 // If obj_array/non_array==true/true:
3028 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3029 // If obj_array/non_array==true/false:
3030 // Branch around if the kls is an oop array (Object[] or subtype)
3031 //
3032 // Like generate_guard, adds a new path onto the region.
3033 jint layout_con = 0;
3034 Node* layout_val = get_layout_helper(kls, layout_con);
3035 if (layout_val == NULL) {
3036 bool query = (obj_array
3037 ? Klass::layout_helper_is_objArray(layout_con)
3038 : Klass::layout_helper_is_javaArray(layout_con));
3039 if (query == not_array) {
3040 return NULL; // never a branch
3041 } else { // always a branch
3042 Node* always_branch = control();
3043 if (region != NULL)
3044 region->add_req(always_branch);
3045 set_control(top());
3046 return always_branch;
3047 }
3048 }
3049 // Now test the correct condition.
3050 jint nval = (obj_array
3051 ? ((jint)Klass::_lh_array_tag_type_value
3052 << Klass::_lh_array_tag_shift)
3053 : Klass::_lh_neutral_value);
3054 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3055 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3056 // invert the test if we are looking for a non-array
3057 if (not_array) btest = BoolTest(btest).negate();
3058 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3059 return generate_fair_guard(bol, region);
3060 }
3061
3062
3063 //-----------------------inline_native_newArray--------------------------
3064 bool LibraryCallKit::inline_native_newArray() {
3065 int nargs = 2;
3066 Node* mirror = argument(0);
3067 Node* count_val = argument(1);
3068
3069 _sp += nargs; // set original stack for use by uncommon_trap
3070 mirror = do_null_check(mirror, T_OBJECT);
3071 _sp -= nargs;
3072 // If mirror or obj is dead, only null-path is taken.
3073 if (stopped()) return true;
3074
3075 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3076 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3077 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3078 TypeInstPtr::NOTNULL);
3079 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3080 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3081 TypePtr::BOTTOM);
3082
3083 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3084 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3085 nargs,
3086 result_reg, _slow_path);
3087 Node* normal_ctl = control();
3088 Node* no_array_ctl = result_reg->in(_slow_path);
3089
3090 // Generate code for the slow case. We make a call to newArray().
3091 set_control(no_array_ctl);
3092 if (!stopped()) {
3093 // Either the input type is void.class, or else the
3094 // array klass has not yet been cached. Either the
3095 // ensuing call will throw an exception, or else it
3096 // will cache the array klass for next time.
3097 PreserveJVMState pjvms(this);
3098 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3099 Node* slow_result = set_results_for_java_call(slow_call);
3100 // this->control() comes from set_results_for_java_call
3101 result_reg->set_req(_slow_path, control());
3102 result_val->set_req(_slow_path, slow_result);
3103 result_io ->set_req(_slow_path, i_o());
3104 result_mem->set_req(_slow_path, reset_memory());
3105 }
3106
3107 set_control(normal_ctl);
3108 if (!stopped()) {
3109 // Normal case: The array type has been cached in the java.lang.Class.
3110 // The following call works fine even if the array type is polymorphic.
3111 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3112 Node* obj = new_array(klass_node, count_val, nargs);
3113 result_reg->init_req(_normal_path, control());
3114 result_val->init_req(_normal_path, obj);
3115 result_io ->init_req(_normal_path, i_o());
3116 result_mem->init_req(_normal_path, reset_memory());
3117 }
3118
3119 // Return the combined state.
3120 set_i_o( _gvn.transform(result_io) );
3121 set_all_memory( _gvn.transform(result_mem) );
3122 push_result(result_reg, result_val);
3123 C->set_has_split_ifs(true); // Has chance for split-if optimization
3124
3125 return true;
3126 }
3127
3128 //----------------------inline_native_getLength--------------------------
3129 bool LibraryCallKit::inline_native_getLength() {
3130 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3131
3132 int nargs = 1;
3133 Node* array = argument(0);
3134
3135 _sp += nargs; // set original stack for use by uncommon_trap
3136 array = do_null_check(array, T_OBJECT);
3137 _sp -= nargs;
3138
3139 // If array is dead, only null-path is taken.
3140 if (stopped()) return true;
3141
3142 // Deoptimize if it is a non-array.
3143 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3144
3145 if (non_array != NULL) {
3146 PreserveJVMState pjvms(this);
3147 set_control(non_array);
3148 _sp += nargs; // push the arguments back on the stack
3149 uncommon_trap(Deoptimization::Reason_intrinsic,
3150 Deoptimization::Action_maybe_recompile);
3151 }
3152
3153 // If control is dead, only non-array-path is taken.
3154 if (stopped()) return true;
3155
3156 // The works fine even if the array type is polymorphic.
3157 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3158 push( load_array_length(array) );
3159
3160 C->set_has_split_ifs(true); // Has chance for split-if optimization
3161
3162 return true;
3163 }
3164
3165 //------------------------inline_array_copyOf----------------------------
3166 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3167 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3168
3169 // Restore the stack and pop off the arguments.
3170 int nargs = 3 + (is_copyOfRange? 1: 0);
3171 Node* original = argument(0);
3172 Node* start = is_copyOfRange? argument(1): intcon(0);
3173 Node* end = is_copyOfRange? argument(2): argument(1);
3174 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3175
3176 Node* newcopy;
3177
3178 //set the original stack and the reexecute bit for the interpreter to reexecute
3179 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3180 { PreserveReexecuteState preexecs(this);
3181 _sp += nargs;
3182 jvms()->set_should_reexecute(true);
3183
3184 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3185 original = do_null_check(original, T_OBJECT);
3186
3187 // Check if a null path was taken unconditionally.
3188 if (stopped()) return true;
3189
3190 Node* orig_length = load_array_length(original);
3191
3192 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3193 NULL, 0);
3194 klass_node = do_null_check(klass_node, T_OBJECT);
3195
3196 RegionNode* bailout = new (C, 1) RegionNode(1);
3197 record_for_igvn(bailout);
3198
3199 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3200 // Bail out if that is so.
3201 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3202 if (not_objArray != NULL) {
3203 // Improve the klass node's type from the new optimistic assumption:
3204 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3205 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3206 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3207 cast->init_req(0, control());
3208 klass_node = _gvn.transform(cast);
3209 }
3210
3211 // Bail out if either start or end is negative.
3212 generate_negative_guard(start, bailout, &start);
3213 generate_negative_guard(end, bailout, &end);
3214
3215 Node* length = end;
3216 if (_gvn.type(start) != TypeInt::ZERO) {
3217 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3218 }
3219
3220 // Bail out if length is negative.
3221 // ...Not needed, since the new_array will throw the right exception.
3222 //generate_negative_guard(length, bailout, &length);
3223
3224 if (bailout->req() > 1) {
3225 PreserveJVMState pjvms(this);
3226 set_control( _gvn.transform(bailout) );
3227 uncommon_trap(Deoptimization::Reason_intrinsic,
3228 Deoptimization::Action_maybe_recompile);
3229 }
3230
3231 if (!stopped()) {
3232
3233 // How many elements will we copy from the original?
3234 // The answer is MinI(orig_length - start, length).
3235 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3236 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3237
3238 const bool raw_mem_only = true;
3239 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3240
3241 // Generate a direct call to the right arraycopy function(s).
3242 // We know the copy is disjoint but we might not know if the
3243 // oop stores need checking.
3244 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3245 // This will fail a store-check if x contains any non-nulls.
3246 bool disjoint_bases = true;
3247 bool length_never_negative = true;
3248 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3249 original, start, newcopy, intcon(0), moved,
3250 disjoint_bases, length_never_negative);
3251 }
3252 } //original reexecute and sp are set back here
3253
3254 if(!stopped()) {
3255 push(newcopy);
3256 }
3257
3258 C->set_has_split_ifs(true); // Has chance for split-if optimization
3259
3260 return true;
3261 }
3262
3263
3264 //----------------------generate_virtual_guard---------------------------
3265 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3266 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3267 RegionNode* slow_region) {
3268 ciMethod* method = callee();
3269 int vtable_index = method->vtable_index();
3270 // Get the methodOop out of the appropriate vtable entry.
3271 int entry_offset = (instanceKlass::vtable_start_offset() +
3272 vtable_index*vtableEntry::size()) * wordSize +
3273 vtableEntry::method_offset_in_bytes();
3274 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3275 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3276
3277 // Compare the target method with the expected method (e.g., Object.hashCode).
3278 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3279
3280 Node* native_call = makecon(native_call_addr);
3281 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3282 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3283
3284 return generate_slow_guard(test_native, slow_region);
3285 }
3286
3287 //-----------------------generate_method_call----------------------------
3288 // Use generate_method_call to make a slow-call to the real
3289 // method if the fast path fails. An alternative would be to
3290 // use a stub like OptoRuntime::slow_arraycopy_Java.
3291 // This only works for expanding the current library call,
3292 // not another intrinsic. (E.g., don't use this for making an
3293 // arraycopy call inside of the copyOf intrinsic.)
3294 CallJavaNode*
3295 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3296 // When compiling the intrinsic method itself, do not use this technique.
3297 guarantee(callee() != C->method(), "cannot make slow-call to self");
3298
3299 ciMethod* method = callee();
3300 // ensure the JVMS we have will be correct for this call
3301 guarantee(method_id == method->intrinsic_id(), "must match");
3302
3303 const TypeFunc* tf = TypeFunc::make(method);
3304 int tfdc = tf->domain()->cnt();
3305 CallJavaNode* slow_call;
3306 if (is_static) {
3307 assert(!is_virtual, "");
3308 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3309 SharedRuntime::get_resolve_static_call_stub(),
3310 method, bci());
3311 } else if (is_virtual) {
3312 null_check_receiver(method);
3313 int vtable_index = methodOopDesc::invalid_vtable_index;
3314 if (UseInlineCaches) {
3315 // Suppress the vtable call
3316 } else {
3317 // hashCode and clone are not a miranda methods,
3318 // so the vtable index is fixed.
3319 // No need to use the linkResolver to get it.
3320 vtable_index = method->vtable_index();
3321 }
3322 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3323 SharedRuntime::get_resolve_virtual_call_stub(),
3324 method, vtable_index, bci());
3325 } else { // neither virtual nor static: opt_virtual
3326 null_check_receiver(method);
3327 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3328 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3329 method, bci());
3330 slow_call->set_optimized_virtual(true);
3331 }
3332 set_arguments_for_java_call(slow_call);
3333 set_edges_for_java_call(slow_call);
3334 return slow_call;
3335 }
3336
3337
3338 //------------------------------inline_native_hashcode--------------------
3339 // Build special case code for calls to hashCode on an object.
3340 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3341 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3342 assert(!(is_virtual && is_static), "either virtual, special, or static");
3343
3344 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3345
3346 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3347 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3348 TypeInt::INT);
3349 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3350 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3351 TypePtr::BOTTOM);
3352 Node* obj = NULL;
3353 if (!is_static) {
3354 // Check for hashing null object
3355 obj = null_check_receiver(callee());
3356 if (stopped()) return true; // unconditionally null
3357 result_reg->init_req(_null_path, top());
3358 result_val->init_req(_null_path, top());
3359 } else {
3360 // Do a null check, and return zero if null.
3361 // System.identityHashCode(null) == 0
3362 obj = argument(0);
3363 Node* null_ctl = top();
3364 obj = null_check_oop(obj, &null_ctl);
3365 result_reg->init_req(_null_path, null_ctl);
3366 result_val->init_req(_null_path, _gvn.intcon(0));
3367 }
3368
3369 // Unconditionally null? Then return right away.
3370 if (stopped()) {
3371 set_control( result_reg->in(_null_path) );
3372 if (!stopped())
3373 push( result_val ->in(_null_path) );
3374 return true;
3375 }
3376
3377 // After null check, get the object's klass.
3378 Node* obj_klass = load_object_klass(obj);
3379
3380 // This call may be virtual (invokevirtual) or bound (invokespecial).
3381 // For each case we generate slightly different code.
3382
3383 // We only go to the fast case code if we pass a number of guards. The
3384 // paths which do not pass are accumulated in the slow_region.
3385 RegionNode* slow_region = new (C, 1) RegionNode(1);
3386 record_for_igvn(slow_region);
3387
3388 // If this is a virtual call, we generate a funny guard. We pull out
3389 // the vtable entry corresponding to hashCode() from the target object.
3390 // If the target method which we are calling happens to be the native
3391 // Object hashCode() method, we pass the guard. We do not need this
3392 // guard for non-virtual calls -- the caller is known to be the native
3393 // Object hashCode().
3394 if (is_virtual) {
3395 generate_virtual_guard(obj_klass, slow_region);
3396 }
3397
3398 // Get the header out of the object, use LoadMarkNode when available
3399 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3400 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
3401 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
3402
3403 // Test the header to see if it is unlocked.
3404 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3405 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3406 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3407 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3408 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3409
3410 generate_slow_guard(test_unlocked, slow_region);
3411
3412 // Get the hash value and check to see that it has been properly assigned.
3413 // We depend on hash_mask being at most 32 bits and avoid the use of
3414 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3415 // vm: see markOop.hpp.
3416 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3417 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3418 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3419 // This hack lets the hash bits live anywhere in the mark object now, as long
3420 // as the shift drops the relevant bits into the low 32 bits. Note that
3421 // Java spec says that HashCode is an int so there's no point in capturing
3422 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3423 hshifted_header = ConvX2I(hshifted_header);
3424 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3425
3426 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3427 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3428 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3429
3430 generate_slow_guard(test_assigned, slow_region);
3431
3432 Node* init_mem = reset_memory();
3433 // fill in the rest of the null path:
3434 result_io ->init_req(_null_path, i_o());
3435 result_mem->init_req(_null_path, init_mem);
3436
3437 result_val->init_req(_fast_path, hash_val);
3438 result_reg->init_req(_fast_path, control());
3439 result_io ->init_req(_fast_path, i_o());
3440 result_mem->init_req(_fast_path, init_mem);
3441
3442 // Generate code for the slow case. We make a call to hashCode().
3443 set_control(_gvn.transform(slow_region));
3444 if (!stopped()) {
3445 // No need for PreserveJVMState, because we're using up the present state.
3446 set_all_memory(init_mem);
3447 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3448 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3449 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3450 Node* slow_result = set_results_for_java_call(slow_call);
3451 // this->control() comes from set_results_for_java_call
3452 result_reg->init_req(_slow_path, control());
3453 result_val->init_req(_slow_path, slow_result);
3454 result_io ->set_req(_slow_path, i_o());
3455 result_mem ->set_req(_slow_path, reset_memory());
3456 }
3457
3458 // Return the combined state.
3459 set_i_o( _gvn.transform(result_io) );
3460 set_all_memory( _gvn.transform(result_mem) );
3461 push_result(result_reg, result_val);
3462
3463 return true;
3464 }
3465
3466 //---------------------------inline_native_getClass----------------------------
3467 // Build special case code for calls to getClass on an object.
3468 bool LibraryCallKit::inline_native_getClass() {
3469 Node* obj = null_check_receiver(callee());
3470 if (stopped()) return true;
3471 push( load_mirror_from_klass(load_object_klass(obj)) );
3472 return true;
3473 }
3474
3475 //-----------------inline_native_Reflection_getCallerClass---------------------
3476 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3477 //
3478 // NOTE that this code must perform the same logic as
3479 // vframeStream::security_get_caller_frame in that it must skip
3480 // Method.invoke() and auxiliary frames.
3481
3482
3483
3484
3485 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3486 ciMethod* method = callee();
3487
3488 #ifndef PRODUCT
3489 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3490 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3491 }
3492 #endif
3493
3494 debug_only(int saved_sp = _sp);
3495
3496 // Argument words: (int depth)
3497 int nargs = 1;
3498
3499 _sp += nargs;
3500 Node* caller_depth_node = pop();
3501
3502 assert(saved_sp == _sp, "must have correct argument count");
3503
3504 // The depth value must be a constant in order for the runtime call
3505 // to be eliminated.
3506 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3507 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3508 #ifndef PRODUCT
3509 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3510 tty->print_cr(" Bailing out because caller depth was not a constant");
3511 }
3512 #endif
3513 return false;
3514 }
3515 // Note that the JVM state at this point does not include the
3516 // getCallerClass() frame which we are trying to inline. The
3517 // semantics of getCallerClass(), however, are that the "first"
3518 // frame is the getCallerClass() frame, so we subtract one from the
3519 // requested depth before continuing. We don't inline requests of
3520 // getCallerClass(0).
3521 int caller_depth = caller_depth_type->get_con() - 1;
3522 if (caller_depth < 0) {
3523 #ifndef PRODUCT
3524 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3525 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3526 }
3527 #endif
3528 return false;
3529 }
3530
3531 if (!jvms()->has_method()) {
3532 #ifndef PRODUCT
3533 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3534 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3535 }
3536 #endif
3537 return false;
3538 }
3539 int _depth = jvms()->depth(); // cache call chain depth
3540
3541 // Walk back up the JVM state to find the caller at the required
3542 // depth. NOTE that this code must perform the same logic as
3543 // vframeStream::security_get_caller_frame in that it must skip
3544 // Method.invoke() and auxiliary frames. Note also that depth is
3545 // 1-based (1 is the bottom of the inlining).
3546 int inlining_depth = _depth;
3547 JVMState* caller_jvms = NULL;
3548
3549 if (inlining_depth > 0) {
3550 caller_jvms = jvms();
3551 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3552 do {
3553 // The following if-tests should be performed in this order
3554 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3555 // Skip a Method.invoke() or auxiliary frame
3556 } else if (caller_depth > 0) {
3557 // Skip real frame
3558 --caller_depth;
3559 } else {
3560 // We're done: reached desired caller after skipping.
3561 break;
3562 }
3563 caller_jvms = caller_jvms->caller();
3564 --inlining_depth;
3565 } while (inlining_depth > 0);
3566 }
3567
3568 if (inlining_depth == 0) {
3569 #ifndef PRODUCT
3570 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3571 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3572 tty->print_cr(" JVM state at this point:");
3573 for (int i = _depth; i >= 1; i--) {
3574 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3575 }
3576 }
3577 #endif
3578 return false; // Reached end of inlining
3579 }
3580
3581 // Acquire method holder as java.lang.Class
3582 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3583 ciInstance* caller_mirror = caller_klass->java_mirror();
3584 // Push this as a constant
3585 push(makecon(TypeInstPtr::make(caller_mirror)));
3586 #ifndef PRODUCT
3587 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3588 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);
3589 tty->print_cr(" JVM state at this point:");
3590 for (int i = _depth; i >= 1; i--) {
3591 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3592 }
3593 }
3594 #endif
3595 return true;
3596 }
3597
3598 // Helper routine for above
3599 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3600 // Is this the Method.invoke method itself?
3601 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
3602 return true;
3603
3604 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3605 ciKlass* k = jvms->method()->holder();
3606 if (k->is_instance_klass()) {
3607 ciInstanceKlass* ik = k->as_instance_klass();
3608 for (; ik != NULL; ik = ik->super()) {
3609 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3610 ik == env()->find_system_klass(ik->name())) {
3611 return true;
3612 }
3613 }
3614 }
3615
3616 return false;
3617 }
3618
3619 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3620 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3621 // computing it since there is no lookup field by name function in the
3622 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3623 // Using a static variable here is safe even if we have multiple compilation
3624 // threads because the offset is constant. At worst the same offset will be
3625 // computed and stored multiple
3626
3627 bool LibraryCallKit::inline_native_AtomicLong_get() {
3628 // Restore the stack and pop off the argument
3629 _sp+=1;
3630 Node *obj = pop();
3631
3632 // get the offset of the "value" field. Since the CI interfaces
3633 // does not provide a way to look up a field by name, we scan the bytecodes
3634 // to get the field index. We expect the first 2 instructions of the method
3635 // to be:
3636 // 0 aload_0
3637 // 1 getfield "value"
3638 ciMethod* method = callee();
3639 if (value_field_offset == -1)
3640 {
3641 ciField* value_field;
3642 ciBytecodeStream iter(method);
3643 Bytecodes::Code bc = iter.next();
3644
3645 if ((bc != Bytecodes::_aload_0) &&
3646 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3647 return false;
3648 bc = iter.next();
3649 if (bc != Bytecodes::_getfield)
3650 return false;
3651 bool ignore;
3652 value_field = iter.get_field(ignore);
3653 value_field_offset = value_field->offset_in_bytes();
3654 }
3655
3656 // Null check without removing any arguments.
3657 _sp++;
3658 obj = do_null_check(obj, T_OBJECT);
3659 _sp--;
3660 // Check for locking null object
3661 if (stopped()) return true;
3662
3663 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3664 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3665 int alias_idx = C->get_alias_index(adr_type);
3666
3667 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3668
3669 push_pair(result);
3670
3671 return true;
3672 }
3673
3674 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3675 // Restore the stack and pop off the arguments
3676 _sp+=5;
3677 Node *newVal = pop_pair();
3678 Node *oldVal = pop_pair();
3679 Node *obj = pop();
3680
3681 // we need the offset of the "value" field which was computed when
3682 // inlining the get() method. Give up if we don't have it.
3683 if (value_field_offset == -1)
3684 return false;
3685
3686 // Null check without removing any arguments.
3687 _sp+=5;
3688 obj = do_null_check(obj, T_OBJECT);
3689 _sp-=5;
3690 // Check for locking null object
3691 if (stopped()) return true;
3692
3693 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3694 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3695 int alias_idx = C->get_alias_index(adr_type);
3696
3697 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3698 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3699 set_memory(store_proj, alias_idx);
3700 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3701
3702 Node *result;
3703 // CMove node is not used to be able fold a possible check code
3704 // after attemptUpdate() call. This code could be transformed
3705 // into CMove node by loop optimizations.
3706 {
3707 RegionNode *r = new (C, 3) RegionNode(3);
3708 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3709
3710 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3711 Node *iftrue = opt_iff(r, iff);
3712 r->init_req(1, iftrue);
3713 result->init_req(1, intcon(1));
3714 result->init_req(2, intcon(0));
3715
3716 set_control(_gvn.transform(r));
3717 record_for_igvn(r);
3718
3719 C->set_has_split_ifs(true); // Has chance for split-if optimization
3720 }
3721
3722 push(_gvn.transform(result));
3723 return true;
3724 }
3725
3726 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3727 // restore the arguments
3728 _sp += arg_size();
3729
3730 switch (id) {
3731 case vmIntrinsics::_floatToRawIntBits:
3732 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3733 break;
3734
3735 case vmIntrinsics::_intBitsToFloat:
3736 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3737 break;
3738
3739 case vmIntrinsics::_doubleToRawLongBits:
3740 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3741 break;
3742
3743 case vmIntrinsics::_longBitsToDouble:
3744 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3745 break;
3746
3747 case vmIntrinsics::_doubleToLongBits: {
3748 Node* value = pop_pair();
3749
3750 // two paths (plus control) merge in a wood
3751 RegionNode *r = new (C, 3) RegionNode(3);
3752 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3753
3754 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3755 // Build the boolean node
3756 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3757
3758 // Branch either way.
3759 // NaN case is less traveled, which makes all the difference.
3760 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3761 Node *opt_isnan = _gvn.transform(ifisnan);
3762 assert( opt_isnan->is_If(), "Expect an IfNode");
3763 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3764 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3765
3766 set_control(iftrue);
3767
3768 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3769 Node *slow_result = longcon(nan_bits); // return NaN
3770 phi->init_req(1, _gvn.transform( slow_result ));
3771 r->init_req(1, iftrue);
3772
3773 // Else fall through
3774 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3775 set_control(iffalse);
3776
3777 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3778 r->init_req(2, iffalse);
3779
3780 // Post merge
3781 set_control(_gvn.transform(r));
3782 record_for_igvn(r);
3783
3784 Node* result = _gvn.transform(phi);
3785 assert(result->bottom_type()->isa_long(), "must be");
3786 push_pair(result);
3787
3788 C->set_has_split_ifs(true); // Has chance for split-if optimization
3789
3790 break;
3791 }
3792
3793 case vmIntrinsics::_floatToIntBits: {
3794 Node* value = pop();
3795
3796 // two paths (plus control) merge in a wood
3797 RegionNode *r = new (C, 3) RegionNode(3);
3798 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3799
3800 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3801 // Build the boolean node
3802 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3803
3804 // Branch either way.
3805 // NaN case is less traveled, which makes all the difference.
3806 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3807 Node *opt_isnan = _gvn.transform(ifisnan);
3808 assert( opt_isnan->is_If(), "Expect an IfNode");
3809 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3810 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3811
3812 set_control(iftrue);
3813
3814 static const jint nan_bits = 0x7fc00000;
3815 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3816 phi->init_req(1, _gvn.transform( slow_result ));
3817 r->init_req(1, iftrue);
3818
3819 // Else fall through
3820 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3821 set_control(iffalse);
3822
3823 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3824 r->init_req(2, iffalse);
3825
3826 // Post merge
3827 set_control(_gvn.transform(r));
3828 record_for_igvn(r);
3829
3830 Node* result = _gvn.transform(phi);
3831 assert(result->bottom_type()->isa_int(), "must be");
3832 push(result);
3833
3834 C->set_has_split_ifs(true); // Has chance for split-if optimization
3835
3836 break;
3837 }
3838
3839 default:
3840 ShouldNotReachHere();
3841 }
3842
3843 return true;
3844 }
3845
3846 #ifdef _LP64
3847 #define XTOP ,top() /*additional argument*/
3848 #else //_LP64
3849 #define XTOP /*no additional argument*/
3850 #endif //_LP64
3851
3852 //----------------------inline_unsafe_copyMemory-------------------------
3853 bool LibraryCallKit::inline_unsafe_copyMemory() {
3854 if (callee()->is_static()) return false; // caller must have the capability!
3855 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3856 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3857 null_check_receiver(callee()); // check then ignore argument(0)
3858 if (stopped()) return true;
3859
3860 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3861
3862 Node* src_ptr = argument(1);
3863 Node* src_off = ConvL2X(argument(2));
3864 assert(argument(3)->is_top(), "2nd half of long");
3865 Node* dst_ptr = argument(4);
3866 Node* dst_off = ConvL2X(argument(5));
3867 assert(argument(6)->is_top(), "2nd half of long");
3868 Node* size = ConvL2X(argument(7));
3869 assert(argument(8)->is_top(), "2nd half of long");
3870
3871 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3872 "fieldOffset must be byte-scaled");
3873
3874 Node* src = make_unsafe_address(src_ptr, src_off);
3875 Node* dst = make_unsafe_address(dst_ptr, dst_off);
3876
3877 // Conservatively insert a memory barrier on all memory slices.
3878 // Do not let writes of the copy source or destination float below the copy.
3879 insert_mem_bar(Op_MemBarCPUOrder);
3880
3881 // Call it. Note that the length argument is not scaled.
3882 make_runtime_call(RC_LEAF|RC_NO_FP,
3883 OptoRuntime::fast_arraycopy_Type(),
3884 StubRoutines::unsafe_arraycopy(),
3885 "unsafe_arraycopy",
3886 TypeRawPtr::BOTTOM,
3887 src, dst, size XTOP);
3888
3889 // Do not let reads of the copy destination float above the copy.
3890 insert_mem_bar(Op_MemBarCPUOrder);
3891
3892 return true;
3893 }
3894
3895 //------------------------clone_coping-----------------------------------
3896 // Helper function for inline_native_clone.
3897 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
3898 assert(obj_size != NULL, "");
3899 Node* raw_obj = alloc_obj->in(1);
3900 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
3901
3902 if (ReduceBulkZeroing) {
3903 // We will be completely responsible for initializing this object -
3904 // mark Initialize node as complete.
3905 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
3906 // The object was just allocated - there should be no any stores!
3907 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
3908 }
3909
3910 // Cast to Object for arraycopy.
3911 // We can't use the original CheckCastPP since it should be moved
3912 // after the arraycopy to prevent stores flowing above it.
3913 Node* new_obj = new(C, 2) CheckCastPPNode(alloc_obj->in(0), raw_obj,
3914 TypeInstPtr::NOTNULL);
3915 new_obj = _gvn.transform(new_obj);
3916 // Substitute in the locally valid dest_oop.
3917 replace_in_map(alloc_obj, new_obj);
3918
3919 // Copy the fastest available way.
3920 // TODO: generate fields copies for small objects instead.
3921 Node* src = obj;
3922 Node* dest = new_obj;
3923 Node* size = _gvn.transform(obj_size);
3924
3925 // Exclude the header but include array length to copy by 8 bytes words.
3926 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
3927 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
3928 instanceOopDesc::base_offset_in_bytes();
3929 // base_off:
3930 // 8 - 32-bit VM
3931 // 12 - 64-bit VM, compressed oops
3932 // 16 - 64-bit VM, normal oops
3933 if (base_off % BytesPerLong != 0) {
3934 assert(UseCompressedOops, "");
3935 if (is_array) {
3936 // Exclude length to copy by 8 bytes words.
3937 base_off += sizeof(int);
3938 } else {
3939 // Include klass to copy by 8 bytes words.
3940 base_off = instanceOopDesc::klass_offset_in_bytes();
3941 }
3942 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
3943 }
3944 src = basic_plus_adr(src, base_off);
3945 dest = basic_plus_adr(dest, base_off);
3946
3947 // Compute the length also, if needed:
3948 Node* countx = size;
3949 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
3950 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
3951
3952 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
3953 bool disjoint_bases = true;
3954 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
3955 src, NULL, dest, NULL, countx);
3956
3957 // If necessary, emit some card marks afterwards. (Non-arrays only.)
3958 if (card_mark) {
3959 assert(!is_array, "");
3960 // Put in store barrier for any and all oops we are sticking
3961 // into this object. (We could avoid this if we could prove
3962 // that the object type contains no oop fields at all.)
3963 Node* no_particular_value = NULL;
3964 Node* no_particular_field = NULL;
3965 int raw_adr_idx = Compile::AliasIdxRaw;
3966 post_barrier(control(),
3967 memory(raw_adr_type),
3968 new_obj,
3969 no_particular_field,
3970 raw_adr_idx,
3971 no_particular_value,
3972 T_OBJECT,
3973 false);
3974 }
3975
3976 // Move the original CheckCastPP after arraycopy.
3977 _gvn.hash_delete(alloc_obj);
3978 alloc_obj->set_req(0, control());
3979 // Replace raw memory edge with new CheckCastPP to have a live oop
3980 // at safepoints instead of raw value.
3981 assert(new_obj->is_CheckCastPP() && new_obj->in(1) == alloc_obj->in(1), "sanity");
3982 alloc_obj->set_req(1, new_obj); // cast to the original type
3983 _gvn.hash_find_insert(alloc_obj); // put back into GVN table
3984 // Restore in the locally valid dest_oop.
3985 replace_in_map(new_obj, alloc_obj);
3986 }
3987
3988 //------------------------inline_native_clone----------------------------
3989 // Here are the simple edge cases:
3990 // null receiver => normal trap
3991 // virtual and clone was overridden => slow path to out-of-line clone
3992 // not cloneable or finalizer => slow path to out-of-line Object.clone
3993 //
3994 // The general case has two steps, allocation and copying.
3995 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
3996 //
3997 // Copying also has two cases, oop arrays and everything else.
3998 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
3999 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4000 //
4001 // These steps fold up nicely if and when the cloned object's klass
4002 // can be sharply typed as an object array, a type array, or an instance.
4003 //
4004 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4005 int nargs = 1;
4006 PhiNode* result_val;
4007
4008 //set the original stack and the reexecute bit for the interpreter to reexecute
4009 //the bytecode that invokes Object.clone if deoptimization happens
4010 { PreserveReexecuteState preexecs(this);
4011 jvms()->set_should_reexecute(true);
4012
4013 //null_check_receiver will adjust _sp (push and pop)
4014 Node* obj = null_check_receiver(callee());
4015 if (stopped()) return true;
4016
4017 _sp += nargs;
4018
4019 Node* obj_klass = load_object_klass(obj);
4020 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4021 const TypeOopPtr* toop = ((tklass != NULL)
4022 ? tklass->as_instance_type()
4023 : TypeInstPtr::NOTNULL);
4024
4025 // Conservatively insert a memory barrier on all memory slices.
4026 // Do not let writes into the original float below the clone.
4027 insert_mem_bar(Op_MemBarCPUOrder);
4028
4029 // paths into result_reg:
4030 enum {
4031 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4032 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4033 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4034 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4035 PATH_LIMIT
4036 };
4037 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4038 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4039 TypeInstPtr::NOTNULL);
4040 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4041 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4042 TypePtr::BOTTOM);
4043 record_for_igvn(result_reg);
4044
4045 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4046 int raw_adr_idx = Compile::AliasIdxRaw;
4047 const bool raw_mem_only = true;
4048
4049
4050 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4051 if (array_ctl != NULL) {
4052 // It's an array.
4053 PreserveJVMState pjvms(this);
4054 set_control(array_ctl);
4055 Node* obj_length = load_array_length(obj);
4056 Node* obj_size = NULL;
4057 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4058 raw_mem_only, &obj_size);
4059
4060 if (!use_ReduceInitialCardMarks()) {
4061 // If it is an oop array, it requires very special treatment,
4062 // because card marking is required on each card of the array.
4063 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4064 if (is_obja != NULL) {
4065 PreserveJVMState pjvms2(this);
4066 set_control(is_obja);
4067 // Generate a direct call to the right arraycopy function(s).
4068 bool disjoint_bases = true;
4069 bool length_never_negative = true;
4070 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4071 obj, intcon(0), alloc_obj, intcon(0),
4072 obj_length,
4073 disjoint_bases, length_never_negative);
4074 result_reg->init_req(_objArray_path, control());
4075 result_val->init_req(_objArray_path, alloc_obj);
4076 result_i_o ->set_req(_objArray_path, i_o());
4077 result_mem ->set_req(_objArray_path, reset_memory());
4078 }
4079 }
4080 // We can dispense with card marks if we know the allocation
4081 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4082 // causes the non-eden paths to simulate a fresh allocation,
4083 // insofar that no further card marks are required to initialize
4084 // the object.
4085
4086 // Otherwise, there are no card marks to worry about.
4087
4088 if (!stopped()) {
4089 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4090
4091 // Present the results of the copy.
4092 result_reg->init_req(_array_path, control());
4093 result_val->init_req(_array_path, alloc_obj);
4094 result_i_o ->set_req(_array_path, i_o());
4095 result_mem ->set_req(_array_path, reset_memory());
4096 }
4097 }
4098
4099 // We only go to the instance fast case code if we pass a number of guards.
4100 // The paths which do not pass are accumulated in the slow_region.
4101 RegionNode* slow_region = new (C, 1) RegionNode(1);
4102 record_for_igvn(slow_region);
4103 if (!stopped()) {
4104 // It's an instance (we did array above). Make the slow-path tests.
4105 // If this is a virtual call, we generate a funny guard. We grab
4106 // the vtable entry corresponding to clone() from the target object.
4107 // If the target method which we are calling happens to be the
4108 // Object clone() method, we pass the guard. We do not need this
4109 // guard for non-virtual calls; the caller is known to be the native
4110 // Object clone().
4111 if (is_virtual) {
4112 generate_virtual_guard(obj_klass, slow_region);
4113 }
4114
4115 // The object must be cloneable and must not have a finalizer.
4116 // Both of these conditions may be checked in a single test.
4117 // We could optimize the cloneable test further, but we don't care.
4118 generate_access_flags_guard(obj_klass,
4119 // Test both conditions:
4120 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4121 // Must be cloneable but not finalizer:
4122 JVM_ACC_IS_CLONEABLE,
4123 slow_region);
4124 }
4125
4126 if (!stopped()) {
4127 // It's an instance, and it passed the slow-path tests.
4128 PreserveJVMState pjvms(this);
4129 Node* obj_size = NULL;
4130 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4131
4132 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4133
4134 // Present the results of the slow call.
4135 result_reg->init_req(_instance_path, control());
4136 result_val->init_req(_instance_path, alloc_obj);
4137 result_i_o ->set_req(_instance_path, i_o());
4138 result_mem ->set_req(_instance_path, reset_memory());
4139 }
4140
4141 // Generate code for the slow case. We make a call to clone().
4142 set_control(_gvn.transform(slow_region));
4143 if (!stopped()) {
4144 PreserveJVMState pjvms(this);
4145 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4146 Node* slow_result = set_results_for_java_call(slow_call);
4147 // this->control() comes from set_results_for_java_call
4148 result_reg->init_req(_slow_path, control());
4149 result_val->init_req(_slow_path, slow_result);
4150 result_i_o ->set_req(_slow_path, i_o());
4151 result_mem ->set_req(_slow_path, reset_memory());
4152 }
4153
4154 // Return the combined state.
4155 set_control( _gvn.transform(result_reg) );
4156 set_i_o( _gvn.transform(result_i_o) );
4157 set_all_memory( _gvn.transform(result_mem) );
4158 } //original reexecute and sp are set back here
4159
4160 push(_gvn.transform(result_val));
4161
4162 return true;
4163 }
4164
4165
4166 // constants for computing the copy function
4167 enum {
4168 COPYFUNC_UNALIGNED = 0,
4169 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4170 COPYFUNC_CONJOINT = 0,
4171 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4172 };
4173
4174 // Note: The condition "disjoint" applies also for overlapping copies
4175 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4176 static address
4177 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
4178 int selector =
4179 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4180 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4181
4182 #define RETURN_STUB(xxx_arraycopy) { \
4183 name = #xxx_arraycopy; \
4184 return StubRoutines::xxx_arraycopy(); }
4185
4186 switch (t) {
4187 case T_BYTE:
4188 case T_BOOLEAN:
4189 switch (selector) {
4190 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4191 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4192 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4193 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4194 }
4195 case T_CHAR:
4196 case T_SHORT:
4197 switch (selector) {
4198 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4199 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4200 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4201 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4202 }
4203 case T_INT:
4204 case T_FLOAT:
4205 switch (selector) {
4206 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4207 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4208 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4209 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4210 }
4211 case T_DOUBLE:
4212 case T_LONG:
4213 switch (selector) {
4214 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4215 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4216 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4217 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4218 }
4219 case T_ARRAY:
4220 case T_OBJECT:
4221 switch (selector) {
4222 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
4223 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
4224 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
4225 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
4226 }
4227 default:
4228 ShouldNotReachHere();
4229 return NULL;
4230 }
4231
4232 #undef RETURN_STUB
4233 }
4234
4235 //------------------------------basictype2arraycopy----------------------------
4236 address LibraryCallKit::basictype2arraycopy(BasicType t,
4237 Node* src_offset,
4238 Node* dest_offset,
4239 bool disjoint_bases,
4240 const char* &name) {
4241 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4242 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4243
4244 bool aligned = false;
4245 bool disjoint = disjoint_bases;
4246
4247 // if the offsets are the same, we can treat the memory regions as
4248 // disjoint, because either the memory regions are in different arrays,
4249 // or they are identical (which we can treat as disjoint.) We can also
4250 // treat a copy with a destination index less that the source index
4251 // as disjoint since a low->high copy will work correctly in this case.
4252 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4253 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4254 // both indices are constants
4255 int s_offs = src_offset_inttype->get_con();
4256 int d_offs = dest_offset_inttype->get_con();
4257 int element_size = type2aelembytes(t);
4258 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4259 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4260 if (s_offs >= d_offs) disjoint = true;
4261 } else if (src_offset == dest_offset && src_offset != NULL) {
4262 // This can occur if the offsets are identical non-constants.
4263 disjoint = true;
4264 }
4265
4266 return select_arraycopy_function(t, aligned, disjoint, name);
4267 }
4268
4269
4270 //------------------------------inline_arraycopy-----------------------
4271 bool LibraryCallKit::inline_arraycopy() {
4272 // Restore the stack and pop off the arguments.
4273 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4274 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4275
4276 Node *src = argument(0);
4277 Node *src_offset = argument(1);
4278 Node *dest = argument(2);
4279 Node *dest_offset = argument(3);
4280 Node *length = argument(4);
4281
4282 // Compile time checks. If any of these checks cannot be verified at compile time,
4283 // we do not make a fast path for this call. Instead, we let the call remain as it
4284 // is. The checks we choose to mandate at compile time are:
4285 //
4286 // (1) src and dest are arrays.
4287 const Type* src_type = src->Value(&_gvn);
4288 const Type* dest_type = dest->Value(&_gvn);
4289 const TypeAryPtr* top_src = src_type->isa_aryptr();
4290 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4291 if (top_src == NULL || top_src->klass() == NULL ||
4292 top_dest == NULL || top_dest->klass() == NULL) {
4293 // Conservatively insert a memory barrier on all memory slices.
4294 // Do not let writes into the source float below the arraycopy.
4295 insert_mem_bar(Op_MemBarCPUOrder);
4296
4297 // Call StubRoutines::generic_arraycopy stub.
4298 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4299 src, src_offset, dest, dest_offset, length);
4300
4301 // Do not let reads from the destination float above the arraycopy.
4302 // Since we cannot type the arrays, we don't know which slices
4303 // might be affected. We could restrict this barrier only to those
4304 // memory slices which pertain to array elements--but don't bother.
4305 if (!InsertMemBarAfterArraycopy)
4306 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4307 insert_mem_bar(Op_MemBarCPUOrder);
4308 return true;
4309 }
4310
4311 // (2) src and dest arrays must have elements of the same BasicType
4312 // Figure out the size and type of the elements we will be copying.
4313 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4314 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4315 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4316 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4317
4318 if (src_elem != dest_elem || dest_elem == T_VOID) {
4319 // The component types are not the same or are not recognized. Punt.
4320 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4321 generate_slow_arraycopy(TypePtr::BOTTOM,
4322 src, src_offset, dest, dest_offset, length);
4323 return true;
4324 }
4325
4326 //---------------------------------------------------------------------------
4327 // We will make a fast path for this call to arraycopy.
4328
4329 // We have the following tests left to perform:
4330 //
4331 // (3) src and dest must not be null.
4332 // (4) src_offset must not be negative.
4333 // (5) dest_offset must not be negative.
4334 // (6) length must not be negative.
4335 // (7) src_offset + length must not exceed length of src.
4336 // (8) dest_offset + length must not exceed length of dest.
4337 // (9) each element of an oop array must be assignable
4338
4339 RegionNode* slow_region = new (C, 1) RegionNode(1);
4340 record_for_igvn(slow_region);
4341
4342 // (3) operands must not be null
4343 // We currently perform our null checks with the do_null_check routine.
4344 // This means that the null exceptions will be reported in the caller
4345 // rather than (correctly) reported inside of the native arraycopy call.
4346 // This should be corrected, given time. We do our null check with the
4347 // stack pointer restored.
4348 _sp += nargs;
4349 src = do_null_check(src, T_ARRAY);
4350 dest = do_null_check(dest, T_ARRAY);
4351 _sp -= nargs;
4352
4353 // (4) src_offset must not be negative.
4354 generate_negative_guard(src_offset, slow_region);
4355
4356 // (5) dest_offset must not be negative.
4357 generate_negative_guard(dest_offset, slow_region);
4358
4359 // (6) length must not be negative (moved to generate_arraycopy()).
4360 // generate_negative_guard(length, slow_region);
4361
4362 // (7) src_offset + length must not exceed length of src.
4363 generate_limit_guard(src_offset, length,
4364 load_array_length(src),
4365 slow_region);
4366
4367 // (8) dest_offset + length must not exceed length of dest.
4368 generate_limit_guard(dest_offset, length,
4369 load_array_length(dest),
4370 slow_region);
4371
4372 // (9) each element of an oop array must be assignable
4373 // The generate_arraycopy subroutine checks this.
4374
4375 // This is where the memory effects are placed:
4376 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4377 generate_arraycopy(adr_type, dest_elem,
4378 src, src_offset, dest, dest_offset, length,
4379 false, false, slow_region);
4380
4381 return true;
4382 }
4383
4384 //-----------------------------generate_arraycopy----------------------
4385 // Generate an optimized call to arraycopy.
4386 // Caller must guard against non-arrays.
4387 // Caller must determine a common array basic-type for both arrays.
4388 // Caller must validate offsets against array bounds.
4389 // The slow_region has already collected guard failure paths
4390 // (such as out of bounds length or non-conformable array types).
4391 // The generated code has this shape, in general:
4392 //
4393 // if (length == 0) return // via zero_path
4394 // slowval = -1
4395 // if (types unknown) {
4396 // slowval = call generic copy loop
4397 // if (slowval == 0) return // via checked_path
4398 // } else if (indexes in bounds) {
4399 // if ((is object array) && !(array type check)) {
4400 // slowval = call checked copy loop
4401 // if (slowval == 0) return // via checked_path
4402 // } else {
4403 // call bulk copy loop
4404 // return // via fast_path
4405 // }
4406 // }
4407 // // adjust params for remaining work:
4408 // if (slowval != -1) {
4409 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4410 // }
4411 // slow_region:
4412 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4413 // return // via slow_call_path
4414 //
4415 // This routine is used from several intrinsics: System.arraycopy,
4416 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4417 //
4418 void
4419 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4420 BasicType basic_elem_type,
4421 Node* src, Node* src_offset,
4422 Node* dest, Node* dest_offset,
4423 Node* copy_length,
4424 bool disjoint_bases,
4425 bool length_never_negative,
4426 RegionNode* slow_region) {
4427
4428 if (slow_region == NULL) {
4429 slow_region = new(C,1) RegionNode(1);
4430 record_for_igvn(slow_region);
4431 }
4432
4433 Node* original_dest = dest;
4434 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4435 bool must_clear_dest = false;
4436
4437 // See if this is the initialization of a newly-allocated array.
4438 // If so, we will take responsibility here for initializing it to zero.
4439 // (Note: Because tightly_coupled_allocation performs checks on the
4440 // out-edges of the dest, we need to avoid making derived pointers
4441 // from it until we have checked its uses.)
4442 if (ReduceBulkZeroing
4443 && !ZeroTLAB // pointless if already zeroed
4444 && basic_elem_type != T_CONFLICT // avoid corner case
4445 && !_gvn.eqv_uncast(src, dest)
4446 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4447 != NULL)
4448 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4449 && alloc->maybe_set_complete(&_gvn)) {
4450 // "You break it, you buy it."
4451 InitializeNode* init = alloc->initialization();
4452 assert(init->is_complete(), "we just did this");
4453 assert(dest->is_CheckCastPP(), "sanity");
4454 assert(dest->in(0)->in(0) == init, "dest pinned");
4455
4456 // Cast to Object for arraycopy.
4457 // We can't use the original CheckCastPP since it should be moved
4458 // after the arraycopy to prevent stores flowing above it.
4459 Node* new_obj = new(C, 2) CheckCastPPNode(dest->in(0), dest->in(1),
4460 TypeInstPtr::NOTNULL);
4461 dest = _gvn.transform(new_obj);
4462 // Substitute in the locally valid dest_oop.
4463 replace_in_map(original_dest, dest);
4464 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4465 // From this point on, every exit path is responsible for
4466 // initializing any non-copied parts of the object to zero.
4467 must_clear_dest = true;
4468 } else {
4469 // No zeroing elimination here.
4470 alloc = NULL;
4471 //original_dest = dest;
4472 //must_clear_dest = false;
4473 }
4474
4475 // Results are placed here:
4476 enum { fast_path = 1, // normal void-returning assembly stub
4477 checked_path = 2, // special assembly stub with cleanup
4478 slow_call_path = 3, // something went wrong; call the VM
4479 zero_path = 4, // bypass when length of copy is zero
4480 bcopy_path = 5, // copy primitive array by 64-bit blocks
4481 PATH_LIMIT = 6
4482 };
4483 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4484 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4485 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4486 record_for_igvn(result_region);
4487 _gvn.set_type_bottom(result_i_o);
4488 _gvn.set_type_bottom(result_memory);
4489 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4490
4491 // The slow_control path:
4492 Node* slow_control;
4493 Node* slow_i_o = i_o();
4494 Node* slow_mem = memory(adr_type);
4495 debug_only(slow_control = (Node*) badAddress);
4496
4497 // Checked control path:
4498 Node* checked_control = top();
4499 Node* checked_mem = NULL;
4500 Node* checked_i_o = NULL;
4501 Node* checked_value = NULL;
4502
4503 if (basic_elem_type == T_CONFLICT) {
4504 assert(!must_clear_dest, "");
4505 Node* cv = generate_generic_arraycopy(adr_type,
4506 src, src_offset, dest, dest_offset,
4507 copy_length);
4508 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4509 checked_control = control();
4510 checked_i_o = i_o();
4511 checked_mem = memory(adr_type);
4512 checked_value = cv;
4513 set_control(top()); // no fast path
4514 }
4515
4516 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4517 if (not_pos != NULL) {
4518 PreserveJVMState pjvms(this);
4519 set_control(not_pos);
4520
4521 // (6) length must not be negative.
4522 if (!length_never_negative) {
4523 generate_negative_guard(copy_length, slow_region);
4524 }
4525
4526 // copy_length is 0.
4527 if (!stopped() && must_clear_dest) {
4528 Node* dest_length = alloc->in(AllocateNode::ALength);
4529 if (_gvn.eqv_uncast(copy_length, dest_length)
4530 || _gvn.find_int_con(dest_length, 1) <= 0) {
4531 // There is no zeroing to do. No need for a secondary raw memory barrier.
4532 } else {
4533 // Clear the whole thing since there are no source elements to copy.
4534 generate_clear_array(adr_type, dest, basic_elem_type,
4535 intcon(0), NULL,
4536 alloc->in(AllocateNode::AllocSize));
4537 // Use a secondary InitializeNode as raw memory barrier.
4538 // Currently it is needed only on this path since other
4539 // paths have stub or runtime calls as raw memory barriers.
4540 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4541 Compile::AliasIdxRaw,
4542 top())->as_Initialize();
4543 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4544 }
4545 }
4546
4547 // Present the results of the fast call.
4548 result_region->init_req(zero_path, control());
4549 result_i_o ->init_req(zero_path, i_o());
4550 result_memory->init_req(zero_path, memory(adr_type));
4551 }
4552
4553 if (!stopped() && must_clear_dest) {
4554 // We have to initialize the *uncopied* part of the array to zero.
4555 // The copy destination is the slice dest[off..off+len]. The other slices
4556 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4557 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4558 Node* dest_length = alloc->in(AllocateNode::ALength);
4559 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4560 copy_length) );
4561
4562 // If there is a head section that needs zeroing, do it now.
4563 if (find_int_con(dest_offset, -1) != 0) {
4564 generate_clear_array(adr_type, dest, basic_elem_type,
4565 intcon(0), dest_offset,
4566 NULL);
4567 }
4568
4569 // Next, perform a dynamic check on the tail length.
4570 // It is often zero, and we can win big if we prove this.
4571 // There are two wins: Avoid generating the ClearArray
4572 // with its attendant messy index arithmetic, and upgrade
4573 // the copy to a more hardware-friendly word size of 64 bits.
4574 Node* tail_ctl = NULL;
4575 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4576 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4577 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4578 tail_ctl = generate_slow_guard(bol_lt, NULL);
4579 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4580 }
4581
4582 // At this point, let's assume there is no tail.
4583 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4584 // There is no tail. Try an upgrade to a 64-bit copy.
4585 bool didit = false;
4586 { PreserveJVMState pjvms(this);
4587 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4588 src, src_offset, dest, dest_offset,
4589 dest_size);
4590 if (didit) {
4591 // Present the results of the block-copying fast call.
4592 result_region->init_req(bcopy_path, control());
4593 result_i_o ->init_req(bcopy_path, i_o());
4594 result_memory->init_req(bcopy_path, memory(adr_type));
4595 }
4596 }
4597 if (didit)
4598 set_control(top()); // no regular fast path
4599 }
4600
4601 // Clear the tail, if any.
4602 if (tail_ctl != NULL) {
4603 Node* notail_ctl = stopped() ? NULL : control();
4604 set_control(tail_ctl);
4605 if (notail_ctl == NULL) {
4606 generate_clear_array(adr_type, dest, basic_elem_type,
4607 dest_tail, NULL,
4608 dest_size);
4609 } else {
4610 // Make a local merge.
4611 Node* done_ctl = new(C,3) RegionNode(3);
4612 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4613 done_ctl->init_req(1, notail_ctl);
4614 done_mem->init_req(1, memory(adr_type));
4615 generate_clear_array(adr_type, dest, basic_elem_type,
4616 dest_tail, NULL,
4617 dest_size);
4618 done_ctl->init_req(2, control());
4619 done_mem->init_req(2, memory(adr_type));
4620 set_control( _gvn.transform(done_ctl) );
4621 set_memory( _gvn.transform(done_mem), adr_type );
4622 }
4623 }
4624 }
4625
4626 BasicType copy_type = basic_elem_type;
4627 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4628 if (!stopped() && copy_type == T_OBJECT) {
4629 // If src and dest have compatible element types, we can copy bits.
4630 // Types S[] and D[] are compatible if D is a supertype of S.
4631 //
4632 // If they are not, we will use checked_oop_disjoint_arraycopy,
4633 // which performs a fast optimistic per-oop check, and backs off
4634 // further to JVM_ArrayCopy on the first per-oop check that fails.
4635 // (Actually, we don't move raw bits only; the GC requires card marks.)
4636
4637 // Get the klassOop for both src and dest
4638 Node* src_klass = load_object_klass(src);
4639 Node* dest_klass = load_object_klass(dest);
4640
4641 // Generate the subtype check.
4642 // This might fold up statically, or then again it might not.
4643 //
4644 // Non-static example: Copying List<String>.elements to a new String[].
4645 // The backing store for a List<String> is always an Object[],
4646 // but its elements are always type String, if the generic types
4647 // are correct at the source level.
4648 //
4649 // Test S[] against D[], not S against D, because (probably)
4650 // the secondary supertype cache is less busy for S[] than S.
4651 // This usually only matters when D is an interface.
4652 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4653 // Plug failing path into checked_oop_disjoint_arraycopy
4654 if (not_subtype_ctrl != top()) {
4655 PreserveJVMState pjvms(this);
4656 set_control(not_subtype_ctrl);
4657 // (At this point we can assume disjoint_bases, since types differ.)
4658 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4659 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4660 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4661 Node* dest_elem_klass = _gvn.transform(n1);
4662 Node* cv = generate_checkcast_arraycopy(adr_type,
4663 dest_elem_klass,
4664 src, src_offset, dest, dest_offset,
4665 copy_length);
4666 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4667 checked_control = control();
4668 checked_i_o = i_o();
4669 checked_mem = memory(adr_type);
4670 checked_value = cv;
4671 }
4672 // At this point we know we do not need type checks on oop stores.
4673
4674 // Let's see if we need card marks:
4675 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4676 // If we do not need card marks, copy using the jint or jlong stub.
4677 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4678 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4679 "sizes agree");
4680 }
4681 }
4682
4683 if (!stopped()) {
4684 // Generate the fast path, if possible.
4685 PreserveJVMState pjvms(this);
4686 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4687 src, src_offset, dest, dest_offset,
4688 ConvI2X(copy_length));
4689
4690 // Present the results of the fast call.
4691 result_region->init_req(fast_path, control());
4692 result_i_o ->init_req(fast_path, i_o());
4693 result_memory->init_req(fast_path, memory(adr_type));
4694 }
4695
4696 // Here are all the slow paths up to this point, in one bundle:
4697 slow_control = top();
4698 if (slow_region != NULL)
4699 slow_control = _gvn.transform(slow_region);
4700 debug_only(slow_region = (RegionNode*)badAddress);
4701
4702 set_control(checked_control);
4703 if (!stopped()) {
4704 // Clean up after the checked call.
4705 // The returned value is either 0 or -1^K,
4706 // where K = number of partially transferred array elements.
4707 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4708 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4709 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4710
4711 // If it is 0, we are done, so transfer to the end.
4712 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4713 result_region->init_req(checked_path, checks_done);
4714 result_i_o ->init_req(checked_path, checked_i_o);
4715 result_memory->init_req(checked_path, checked_mem);
4716
4717 // If it is not zero, merge into the slow call.
4718 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4719 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4720 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4721 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4722 record_for_igvn(slow_reg2);
4723 slow_reg2 ->init_req(1, slow_control);
4724 slow_i_o2 ->init_req(1, slow_i_o);
4725 slow_mem2 ->init_req(1, slow_mem);
4726 slow_reg2 ->init_req(2, control());
4727 slow_i_o2 ->init_req(2, checked_i_o);
4728 slow_mem2 ->init_req(2, checked_mem);
4729
4730 slow_control = _gvn.transform(slow_reg2);
4731 slow_i_o = _gvn.transform(slow_i_o2);
4732 slow_mem = _gvn.transform(slow_mem2);
4733
4734 if (alloc != NULL) {
4735 // We'll restart from the very beginning, after zeroing the whole thing.
4736 // This can cause double writes, but that's OK since dest is brand new.
4737 // So we ignore the low 31 bits of the value returned from the stub.
4738 } else {
4739 // We must continue the copy exactly where it failed, or else
4740 // another thread might see the wrong number of writes to dest.
4741 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4742 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4743 slow_offset->init_req(1, intcon(0));
4744 slow_offset->init_req(2, checked_offset);
4745 slow_offset = _gvn.transform(slow_offset);
4746
4747 // Adjust the arguments by the conditionally incoming offset.
4748 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4749 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4750 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4751
4752 // Tweak the node variables to adjust the code produced below:
4753 src_offset = src_off_plus;
4754 dest_offset = dest_off_plus;
4755 copy_length = length_minus;
4756 }
4757 }
4758
4759 set_control(slow_control);
4760 if (!stopped()) {
4761 // Generate the slow path, if needed.
4762 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4763
4764 set_memory(slow_mem, adr_type);
4765 set_i_o(slow_i_o);
4766
4767 if (must_clear_dest) {
4768 generate_clear_array(adr_type, dest, basic_elem_type,
4769 intcon(0), NULL,
4770 alloc->in(AllocateNode::AllocSize));
4771 }
4772
4773 generate_slow_arraycopy(adr_type,
4774 src, src_offset, dest, dest_offset,
4775 copy_length);
4776
4777 result_region->init_req(slow_call_path, control());
4778 result_i_o ->init_req(slow_call_path, i_o());
4779 result_memory->init_req(slow_call_path, memory(adr_type));
4780 }
4781
4782 // Remove unused edges.
4783 for (uint i = 1; i < result_region->req(); i++) {
4784 if (result_region->in(i) == NULL)
4785 result_region->init_req(i, top());
4786 }
4787
4788 // Finished; return the combined state.
4789 set_control( _gvn.transform(result_region) );
4790 set_i_o( _gvn.transform(result_i_o) );
4791 set_memory( _gvn.transform(result_memory), adr_type );
4792
4793 if (dest != original_dest) {
4794 // Pin the "finished" array node after the arraycopy/zeroing operations.
4795 _gvn.hash_delete(original_dest);
4796 original_dest->set_req(0, control());
4797 // Replace raw memory edge with new CheckCastPP to have a live oop
4798 // at safepoints instead of raw value.
4799 assert(dest->is_CheckCastPP() && dest->in(1) == original_dest->in(1), "sanity");
4800 original_dest->set_req(1, dest); // cast to the original type
4801 _gvn.hash_find_insert(original_dest); // put back into GVN table
4802 // Restore in the locally valid dest_oop.
4803 replace_in_map(dest, original_dest);
4804 }
4805 // The memory edges above are precise in order to model effects around
4806 // array copies accurately to allow value numbering of field loads around
4807 // arraycopy. Such field loads, both before and after, are common in Java
4808 // collections and similar classes involving header/array data structures.
4809 //
4810 // But with low number of register or when some registers are used or killed
4811 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4812 // The next memory barrier is added to avoid it. If the arraycopy can be
4813 // optimized away (which it can, sometimes) then we can manually remove
4814 // the membar also.
4815 if (InsertMemBarAfterArraycopy)
4816 insert_mem_bar(Op_MemBarCPUOrder);
4817 }
4818
4819
4820 // Helper function which determines if an arraycopy immediately follows
4821 // an allocation, with no intervening tests or other escapes for the object.
4822 AllocateArrayNode*
4823 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4824 RegionNode* slow_region) {
4825 if (stopped()) return NULL; // no fast path
4826 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4827
4828 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4829 if (alloc == NULL) return NULL;
4830
4831 Node* rawmem = memory(Compile::AliasIdxRaw);
4832 // Is the allocation's memory state untouched?
4833 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4834 // Bail out if there have been raw-memory effects since the allocation.
4835 // (Example: There might have been a call or safepoint.)
4836 return NULL;
4837 }
4838 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4839 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4840 return NULL;
4841 }
4842
4843 // There must be no unexpected observers of this allocation.
4844 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4845 Node* obs = ptr->fast_out(i);
4846 if (obs != this->map()) {
4847 return NULL;
4848 }
4849 }
4850
4851 // This arraycopy must unconditionally follow the allocation of the ptr.
4852 Node* alloc_ctl = ptr->in(0);
4853 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4854
4855 Node* ctl = control();
4856 while (ctl != alloc_ctl) {
4857 // There may be guards which feed into the slow_region.
4858 // Any other control flow means that we might not get a chance
4859 // to finish initializing the allocated object.
4860 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4861 IfNode* iff = ctl->in(0)->as_If();
4862 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4863 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4864 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4865 ctl = iff->in(0); // This test feeds the known slow_region.
4866 continue;
4867 }
4868 // One more try: Various low-level checks bottom out in
4869 // uncommon traps. If the debug-info of the trap omits
4870 // any reference to the allocation, as we've already
4871 // observed, then there can be no objection to the trap.
4872 bool found_trap = false;
4873 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4874 Node* obs = not_ctl->fast_out(j);
4875 if (obs->in(0) == not_ctl && obs->is_Call() &&
4876 (obs->as_Call()->entry_point() ==
4877 SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
4878 found_trap = true; break;
4879 }
4880 }
4881 if (found_trap) {
4882 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4883 continue;
4884 }
4885 }
4886 return NULL;
4887 }
4888
4889 // If we get this far, we have an allocation which immediately
4890 // precedes the arraycopy, and we can take over zeroing the new object.
4891 // The arraycopy will finish the initialization, and provide
4892 // a new control state to which we will anchor the destination pointer.
4893
4894 return alloc;
4895 }
4896
4897 // Helper for initialization of arrays, creating a ClearArray.
4898 // It writes zero bits in [start..end), within the body of an array object.
4899 // The memory effects are all chained onto the 'adr_type' alias category.
4900 //
4901 // Since the object is otherwise uninitialized, we are free
4902 // to put a little "slop" around the edges of the cleared area,
4903 // as long as it does not go back into the array's header,
4904 // or beyond the array end within the heap.
4905 //
4906 // The lower edge can be rounded down to the nearest jint and the
4907 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4908 //
4909 // Arguments:
4910 // adr_type memory slice where writes are generated
4911 // dest oop of the destination array
4912 // basic_elem_type element type of the destination
4913 // slice_idx array index of first element to store
4914 // slice_len number of elements to store (or NULL)
4915 // dest_size total size in bytes of the array object
4916 //
4917 // Exactly one of slice_len or dest_size must be non-NULL.
4918 // If dest_size is non-NULL, zeroing extends to the end of the object.
4919 // If slice_len is non-NULL, the slice_idx value must be a constant.
4920 void
4921 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
4922 Node* dest,
4923 BasicType basic_elem_type,
4924 Node* slice_idx,
4925 Node* slice_len,
4926 Node* dest_size) {
4927 // one or the other but not both of slice_len and dest_size:
4928 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
4929 if (slice_len == NULL) slice_len = top();
4930 if (dest_size == NULL) dest_size = top();
4931
4932 // operate on this memory slice:
4933 Node* mem = memory(adr_type); // memory slice to operate on
4934
4935 // scaling and rounding of indexes:
4936 int scale = exact_log2(type2aelembytes(basic_elem_type));
4937 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
4938 int clear_low = (-1 << scale) & (BytesPerInt - 1);
4939 int bump_bit = (-1 << scale) & BytesPerInt;
4940
4941 // determine constant starts and ends
4942 const intptr_t BIG_NEG = -128;
4943 assert(BIG_NEG + 2*abase < 0, "neg enough");
4944 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
4945 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
4946 if (slice_len_con == 0) {
4947 return; // nothing to do here
4948 }
4949 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
4950 intptr_t end_con = find_intptr_t_con(dest_size, -1);
4951 if (slice_idx_con >= 0 && slice_len_con >= 0) {
4952 assert(end_con < 0, "not two cons");
4953 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
4954 BytesPerLong);
4955 }
4956
4957 if (start_con >= 0 && end_con >= 0) {
4958 // Constant start and end. Simple.
4959 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4960 start_con, end_con, &_gvn);
4961 } else if (start_con >= 0 && dest_size != top()) {
4962 // Constant start, pre-rounded end after the tail of the array.
4963 Node* end = dest_size;
4964 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4965 start_con, end, &_gvn);
4966 } else if (start_con >= 0 && slice_len != top()) {
4967 // Constant start, non-constant end. End needs rounding up.
4968 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
4969 intptr_t end_base = abase + (slice_idx_con << scale);
4970 int end_round = (-1 << scale) & (BytesPerLong - 1);
4971 Node* end = ConvI2X(slice_len);
4972 if (scale != 0)
4973 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
4974 end_base += end_round;
4975 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
4976 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
4977 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4978 start_con, end, &_gvn);
4979 } else if (start_con < 0 && dest_size != top()) {
4980 // Non-constant start, pre-rounded end after the tail of the array.
4981 // This is almost certainly a "round-to-end" operation.
4982 Node* start = slice_idx;
4983 start = ConvI2X(start);
4984 if (scale != 0)
4985 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
4986 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
4987 if ((bump_bit | clear_low) != 0) {
4988 int to_clear = (bump_bit | clear_low);
4989 // Align up mod 8, then store a jint zero unconditionally
4990 // just before the mod-8 boundary.
4991 if (((abase + bump_bit) & ~to_clear) - bump_bit
4992 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
4993 bump_bit = 0;
4994 assert((abase & to_clear) == 0, "array base must be long-aligned");
4995 } else {
4996 // Bump 'start' up to (or past) the next jint boundary:
4997 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
4998 assert((abase & clear_low) == 0, "array base must be int-aligned");
4999 }
5000 // Round bumped 'start' down to jlong boundary in body of array.
5001 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5002 if (bump_bit != 0) {
5003 // Store a zero to the immediately preceding jint:
5004 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5005 Node* p1 = basic_plus_adr(dest, x1);
5006 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5007 mem = _gvn.transform(mem);
5008 }
5009 }
5010 Node* end = dest_size; // pre-rounded
5011 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5012 start, end, &_gvn);
5013 } else {
5014 // Non-constant start, unrounded non-constant end.
5015 // (Nobody zeroes a random midsection of an array using this routine.)
5016 ShouldNotReachHere(); // fix caller
5017 }
5018
5019 // Done.
5020 set_memory(mem, adr_type);
5021 }
5022
5023
5024 bool
5025 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5026 BasicType basic_elem_type,
5027 AllocateNode* alloc,
5028 Node* src, Node* src_offset,
5029 Node* dest, Node* dest_offset,
5030 Node* dest_size) {
5031 // See if there is an advantage from block transfer.
5032 int scale = exact_log2(type2aelembytes(basic_elem_type));
5033 if (scale >= LogBytesPerLong)
5034 return false; // it is already a block transfer
5035
5036 // Look at the alignment of the starting offsets.
5037 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5038 const intptr_t BIG_NEG = -128;
5039 assert(BIG_NEG + 2*abase < 0, "neg enough");
5040
5041 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5042 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5043 if (src_off < 0 || dest_off < 0)
5044 // At present, we can only understand constants.
5045 return false;
5046
5047 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5048 // Non-aligned; too bad.
5049 // One more chance: Pick off an initial 32-bit word.
5050 // This is a common case, since abase can be odd mod 8.
5051 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5052 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5053 Node* sptr = basic_plus_adr(src, src_off);
5054 Node* dptr = basic_plus_adr(dest, dest_off);
5055 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5056 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5057 src_off += BytesPerInt;
5058 dest_off += BytesPerInt;
5059 } else {
5060 return false;
5061 }
5062 }
5063 assert(src_off % BytesPerLong == 0, "");
5064 assert(dest_off % BytesPerLong == 0, "");
5065
5066 // Do this copy by giant steps.
5067 Node* sptr = basic_plus_adr(src, src_off);
5068 Node* dptr = basic_plus_adr(dest, dest_off);
5069 Node* countx = dest_size;
5070 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5071 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5072
5073 bool disjoint_bases = true; // since alloc != NULL
5074 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5075 sptr, NULL, dptr, NULL, countx);
5076
5077 return true;
5078 }
5079
5080
5081 // Helper function; generates code for the slow case.
5082 // We make a call to a runtime method which emulates the native method,
5083 // but without the native wrapper overhead.
5084 void
5085 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5086 Node* src, Node* src_offset,
5087 Node* dest, Node* dest_offset,
5088 Node* copy_length) {
5089 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5090 OptoRuntime::slow_arraycopy_Type(),
5091 OptoRuntime::slow_arraycopy_Java(),
5092 "slow_arraycopy", adr_type,
5093 src, src_offset, dest, dest_offset,
5094 copy_length);
5095
5096 // Handle exceptions thrown by this fellow:
5097 make_slow_call_ex(call, env()->Throwable_klass(), false);
5098 }
5099
5100 // Helper function; generates code for cases requiring runtime checks.
5101 Node*
5102 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5103 Node* dest_elem_klass,
5104 Node* src, Node* src_offset,
5105 Node* dest, Node* dest_offset,
5106 Node* copy_length) {
5107 if (stopped()) return NULL;
5108
5109 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
5110 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5111 return NULL;
5112 }
5113
5114 // Pick out the parameters required to perform a store-check
5115 // for the target array. This is an optimistic check. It will
5116 // look in each non-null element's class, at the desired klass's
5117 // super_check_offset, for the desired klass.
5118 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5119 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5120 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
5121 Node* check_offset = _gvn.transform(n3);
5122 Node* check_value = dest_elem_klass;
5123
5124 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5125 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5126
5127 // (We know the arrays are never conjoint, because their types differ.)
5128 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5129 OptoRuntime::checkcast_arraycopy_Type(),
5130 copyfunc_addr, "checkcast_arraycopy", adr_type,
5131 // five arguments, of which two are
5132 // intptr_t (jlong in LP64)
5133 src_start, dest_start,
5134 copy_length XTOP,
5135 check_offset XTOP,
5136 check_value);
5137
5138 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5139 }
5140
5141
5142 // Helper function; generates code for cases requiring runtime checks.
5143 Node*
5144 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5145 Node* src, Node* src_offset,
5146 Node* dest, Node* dest_offset,
5147 Node* copy_length) {
5148 if (stopped()) return NULL;
5149
5150 address copyfunc_addr = StubRoutines::generic_arraycopy();
5151 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5152 return NULL;
5153 }
5154
5155 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5156 OptoRuntime::generic_arraycopy_Type(),
5157 copyfunc_addr, "generic_arraycopy", adr_type,
5158 src, src_offset, dest, dest_offset, copy_length);
5159
5160 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5161 }
5162
5163 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5164 void
5165 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5166 BasicType basic_elem_type,
5167 bool disjoint_bases,
5168 Node* src, Node* src_offset,
5169 Node* dest, Node* dest_offset,
5170 Node* copy_length) {
5171 if (stopped()) return; // nothing to do
5172
5173 Node* src_start = src;
5174 Node* dest_start = dest;
5175 if (src_offset != NULL || dest_offset != NULL) {
5176 assert(src_offset != NULL && dest_offset != NULL, "");
5177 src_start = array_element_address(src, src_offset, basic_elem_type);
5178 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5179 }
5180
5181 // Figure out which arraycopy runtime method to call.
5182 const char* copyfunc_name = "arraycopy";
5183 address copyfunc_addr =
5184 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5185 disjoint_bases, copyfunc_name);
5186
5187 // Call it. Note that the count_ix value is not scaled to a byte-size.
5188 make_runtime_call(RC_LEAF|RC_NO_FP,
5189 OptoRuntime::fast_arraycopy_Type(),
5190 copyfunc_addr, copyfunc_name, adr_type,
5191 src_start, dest_start, copy_length XTOP);
5192 }