Wed, 21 May 2008 13:46:23 -0700
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
Summary: Add LoadNKlass and CMoveN nodes, use CmpN and ConN nodes to generate narrow oops compare instructions.
Reviewed-by: never, rasbold
1 /*
2 * Copyright 1999-2007 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 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_library_call.cpp.incl"
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;
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 };
49 // Local helper class for LibraryIntrinsic:
50 class LibraryCallKit : public GraphKit {
51 private:
52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
54 public:
55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
56 : GraphKit(caller),
57 _intrinsic(intrinsic)
58 {
59 }
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(); }
69 bool try_to_inline();
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 }
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 Node* pop_math_arg();
140 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
141 bool inline_math_native(vmIntrinsics::ID id);
142 bool inline_trig(vmIntrinsics::ID id);
143 bool inline_trans(vmIntrinsics::ID id);
144 bool inline_abs(vmIntrinsics::ID id);
145 bool inline_sqrt(vmIntrinsics::ID id);
146 bool inline_pow(vmIntrinsics::ID id);
147 bool inline_exp(vmIntrinsics::ID id);
148 bool inline_min_max(vmIntrinsics::ID id);
149 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
150 // This returns Type::AnyPtr, RawPtr, or OopPtr.
151 int classify_unsafe_addr(Node* &base, Node* &offset);
152 Node* make_unsafe_address(Node* base, Node* offset);
153 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
154 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
155 bool inline_unsafe_allocate();
156 bool inline_unsafe_copyMemory();
157 bool inline_native_currentThread();
158 bool inline_native_time_funcs(bool isNano);
159 bool inline_native_isInterrupted();
160 bool inline_native_Class_query(vmIntrinsics::ID id);
161 bool inline_native_subtype_check();
163 bool inline_native_newArray();
164 bool inline_native_getLength();
165 bool inline_array_copyOf(bool is_copyOfRange);
166 bool inline_native_clone(bool is_virtual);
167 bool inline_native_Reflection_getCallerClass();
168 bool inline_native_AtomicLong_get();
169 bool inline_native_AtomicLong_attemptUpdate();
170 bool is_method_invoke_or_aux_frame(JVMState* jvms);
171 // Helper function for inlining native object hash method
172 bool inline_native_hashcode(bool is_virtual, bool is_static);
173 bool inline_native_getClass();
175 // Helper functions for inlining arraycopy
176 bool inline_arraycopy();
177 void generate_arraycopy(const TypePtr* adr_type,
178 BasicType basic_elem_type,
179 Node* src, Node* src_offset,
180 Node* dest, Node* dest_offset,
181 Node* copy_length,
182 int nargs, // arguments on stack for debug info
183 bool disjoint_bases = false,
184 bool length_never_negative = false,
185 RegionNode* slow_region = NULL);
186 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
187 RegionNode* slow_region);
188 void generate_clear_array(const TypePtr* adr_type,
189 Node* dest,
190 BasicType basic_elem_type,
191 Node* slice_off,
192 Node* slice_len,
193 Node* slice_end);
194 bool generate_block_arraycopy(const TypePtr* adr_type,
195 BasicType basic_elem_type,
196 AllocateNode* alloc,
197 Node* src, Node* src_offset,
198 Node* dest, Node* dest_offset,
199 Node* dest_size);
200 void generate_slow_arraycopy(const TypePtr* adr_type,
201 Node* src, Node* src_offset,
202 Node* dest, Node* dest_offset,
203 Node* copy_length,
204 int nargs);
205 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
206 Node* dest_elem_klass,
207 Node* src, Node* src_offset,
208 Node* dest, Node* dest_offset,
209 Node* copy_length, int nargs);
210 Node* generate_generic_arraycopy(const TypePtr* adr_type,
211 Node* src, Node* src_offset,
212 Node* dest, Node* dest_offset,
213 Node* copy_length, int nargs);
214 void generate_unchecked_arraycopy(const TypePtr* adr_type,
215 BasicType basic_elem_type,
216 bool disjoint_bases,
217 Node* src, Node* src_offset,
218 Node* dest, Node* dest_offset,
219 Node* copy_length);
220 bool inline_unsafe_CAS(BasicType type);
221 bool inline_unsafe_ordered_store(BasicType type);
222 bool inline_fp_conversions(vmIntrinsics::ID id);
223 bool inline_reverseBytes(vmIntrinsics::ID id);
224 };
227 //---------------------------make_vm_intrinsic----------------------------
228 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
229 vmIntrinsics::ID id = m->intrinsic_id();
230 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
232 if (DisableIntrinsic[0] != '\0'
233 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
234 // disabled by a user request on the command line:
235 // example: -XX:DisableIntrinsic=_hashCode,_getClass
236 return NULL;
237 }
239 if (!m->is_loaded()) {
240 // do not attempt to inline unloaded methods
241 return NULL;
242 }
244 // Only a few intrinsics implement a virtual dispatch.
245 // They are expensive calls which are also frequently overridden.
246 if (is_virtual) {
247 switch (id) {
248 case vmIntrinsics::_hashCode:
249 case vmIntrinsics::_clone:
250 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
251 break;
252 default:
253 return NULL;
254 }
255 }
257 // -XX:-InlineNatives disables nearly all intrinsics:
258 if (!InlineNatives) {
259 switch (id) {
260 case vmIntrinsics::_indexOf:
261 case vmIntrinsics::_compareTo:
262 break; // InlineNatives does not control String.compareTo
263 default:
264 return NULL;
265 }
266 }
268 switch (id) {
269 case vmIntrinsics::_compareTo:
270 if (!SpecialStringCompareTo) return NULL;
271 break;
272 case vmIntrinsics::_indexOf:
273 if (!SpecialStringIndexOf) return NULL;
274 break;
275 case vmIntrinsics::_arraycopy:
276 if (!InlineArrayCopy) return NULL;
277 break;
278 case vmIntrinsics::_copyMemory:
279 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
280 if (!InlineArrayCopy) return NULL;
281 break;
282 case vmIntrinsics::_hashCode:
283 if (!InlineObjectHash) return NULL;
284 break;
285 case vmIntrinsics::_clone:
286 case vmIntrinsics::_copyOf:
287 case vmIntrinsics::_copyOfRange:
288 if (!InlineObjectCopy) return NULL;
289 // These also use the arraycopy intrinsic mechanism:
290 if (!InlineArrayCopy) return NULL;
291 break;
292 case vmIntrinsics::_checkIndex:
293 // We do not intrinsify this. The optimizer does fine with it.
294 return NULL;
296 case vmIntrinsics::_get_AtomicLong:
297 case vmIntrinsics::_attemptUpdate:
298 if (!InlineAtomicLong) return NULL;
299 break;
301 case vmIntrinsics::_Object_init:
302 case vmIntrinsics::_invoke:
303 // We do not intrinsify these; they are marked for other purposes.
304 return NULL;
306 case vmIntrinsics::_getCallerClass:
307 if (!UseNewReflection) return NULL;
308 if (!InlineReflectionGetCallerClass) return NULL;
309 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
310 break;
312 default:
313 break;
314 }
316 // -XX:-InlineClassNatives disables natives from the Class class.
317 // The flag applies to all reflective calls, notably Array.newArray
318 // (visible to Java programmers as Array.newInstance).
319 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
320 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
321 if (!InlineClassNatives) return NULL;
322 }
324 // -XX:-InlineThreadNatives disables natives from the Thread class.
325 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
326 if (!InlineThreadNatives) return NULL;
327 }
329 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
330 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
331 m->holder()->name() == ciSymbol::java_lang_Float() ||
332 m->holder()->name() == ciSymbol::java_lang_Double()) {
333 if (!InlineMathNatives) return NULL;
334 }
336 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
337 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
338 if (!InlineUnsafeOps) return NULL;
339 }
341 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
342 }
344 //----------------------register_library_intrinsics-----------------------
345 // Initialize this file's data structures, for each Compile instance.
346 void Compile::register_library_intrinsics() {
347 // Nothing to do here.
348 }
350 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
351 LibraryCallKit kit(jvms, this);
352 Compile* C = kit.C;
353 int nodes = C->unique();
354 #ifndef PRODUCT
355 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
356 char buf[1000];
357 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
358 tty->print_cr("Intrinsic %s", str);
359 }
360 #endif
361 if (kit.try_to_inline()) {
362 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
363 tty->print("Inlining intrinsic %s%s at bci:%d in",
364 vmIntrinsics::name_at(intrinsic_id()),
365 (is_virtual() ? " (virtual)" : ""), kit.bci());
366 kit.caller()->print_short_name(tty);
367 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
368 }
369 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
370 if (C->log()) {
371 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
372 vmIntrinsics::name_at(intrinsic_id()),
373 (is_virtual() ? " virtual='1'" : ""),
374 C->unique() - nodes);
375 }
376 return kit.transfer_exceptions_into_jvms();
377 }
379 if (PrintIntrinsics) {
380 switch (intrinsic_id()) {
381 case vmIntrinsics::_invoke:
382 case vmIntrinsics::_Object_init:
383 // We do not expect to inline these, so do not produce any noise about them.
384 break;
385 default:
386 tty->print("Did not inline intrinsic %s%s at bci:%d in",
387 vmIntrinsics::name_at(intrinsic_id()),
388 (is_virtual() ? " (virtual)" : ""), kit.bci());
389 kit.caller()->print_short_name(tty);
390 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
391 }
392 }
393 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
394 return NULL;
395 }
397 bool LibraryCallKit::try_to_inline() {
398 // Handle symbolic names for otherwise undistinguished boolean switches:
399 const bool is_store = true;
400 const bool is_native_ptr = true;
401 const bool is_static = true;
403 switch (intrinsic_id()) {
404 case vmIntrinsics::_hashCode:
405 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
406 case vmIntrinsics::_identityHashCode:
407 return inline_native_hashcode(/*!virtual*/ false, is_static);
408 case vmIntrinsics::_getClass:
409 return inline_native_getClass();
411 case vmIntrinsics::_dsin:
412 case vmIntrinsics::_dcos:
413 case vmIntrinsics::_dtan:
414 case vmIntrinsics::_dabs:
415 case vmIntrinsics::_datan2:
416 case vmIntrinsics::_dsqrt:
417 case vmIntrinsics::_dexp:
418 case vmIntrinsics::_dlog:
419 case vmIntrinsics::_dlog10:
420 case vmIntrinsics::_dpow:
421 return inline_math_native(intrinsic_id());
423 case vmIntrinsics::_min:
424 case vmIntrinsics::_max:
425 return inline_min_max(intrinsic_id());
427 case vmIntrinsics::_arraycopy:
428 return inline_arraycopy();
430 case vmIntrinsics::_compareTo:
431 return inline_string_compareTo();
432 case vmIntrinsics::_indexOf:
433 return inline_string_indexOf();
435 case vmIntrinsics::_getObject:
436 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
437 case vmIntrinsics::_getBoolean:
438 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
439 case vmIntrinsics::_getByte:
440 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
441 case vmIntrinsics::_getShort:
442 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
443 case vmIntrinsics::_getChar:
444 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
445 case vmIntrinsics::_getInt:
446 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
447 case vmIntrinsics::_getLong:
448 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
449 case vmIntrinsics::_getFloat:
450 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
451 case vmIntrinsics::_getDouble:
452 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
454 case vmIntrinsics::_putObject:
455 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
456 case vmIntrinsics::_putBoolean:
457 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
458 case vmIntrinsics::_putByte:
459 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
460 case vmIntrinsics::_putShort:
461 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
462 case vmIntrinsics::_putChar:
463 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
464 case vmIntrinsics::_putInt:
465 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
466 case vmIntrinsics::_putLong:
467 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
468 case vmIntrinsics::_putFloat:
469 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
470 case vmIntrinsics::_putDouble:
471 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
473 case vmIntrinsics::_getByte_raw:
474 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
475 case vmIntrinsics::_getShort_raw:
476 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
477 case vmIntrinsics::_getChar_raw:
478 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
479 case vmIntrinsics::_getInt_raw:
480 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
481 case vmIntrinsics::_getLong_raw:
482 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
483 case vmIntrinsics::_getFloat_raw:
484 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
485 case vmIntrinsics::_getDouble_raw:
486 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
487 case vmIntrinsics::_getAddress_raw:
488 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
490 case vmIntrinsics::_putByte_raw:
491 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
492 case vmIntrinsics::_putShort_raw:
493 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
494 case vmIntrinsics::_putChar_raw:
495 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
496 case vmIntrinsics::_putInt_raw:
497 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
498 case vmIntrinsics::_putLong_raw:
499 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
500 case vmIntrinsics::_putFloat_raw:
501 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
502 case vmIntrinsics::_putDouble_raw:
503 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
504 case vmIntrinsics::_putAddress_raw:
505 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
507 case vmIntrinsics::_getObjectVolatile:
508 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
509 case vmIntrinsics::_getBooleanVolatile:
510 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
511 case vmIntrinsics::_getByteVolatile:
512 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
513 case vmIntrinsics::_getShortVolatile:
514 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
515 case vmIntrinsics::_getCharVolatile:
516 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
517 case vmIntrinsics::_getIntVolatile:
518 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
519 case vmIntrinsics::_getLongVolatile:
520 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
521 case vmIntrinsics::_getFloatVolatile:
522 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
523 case vmIntrinsics::_getDoubleVolatile:
524 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
526 case vmIntrinsics::_putObjectVolatile:
527 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
528 case vmIntrinsics::_putBooleanVolatile:
529 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
530 case vmIntrinsics::_putByteVolatile:
531 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
532 case vmIntrinsics::_putShortVolatile:
533 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
534 case vmIntrinsics::_putCharVolatile:
535 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
536 case vmIntrinsics::_putIntVolatile:
537 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
538 case vmIntrinsics::_putLongVolatile:
539 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
540 case vmIntrinsics::_putFloatVolatile:
541 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
542 case vmIntrinsics::_putDoubleVolatile:
543 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
545 case vmIntrinsics::_prefetchRead:
546 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
547 case vmIntrinsics::_prefetchWrite:
548 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
549 case vmIntrinsics::_prefetchReadStatic:
550 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
551 case vmIntrinsics::_prefetchWriteStatic:
552 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
554 case vmIntrinsics::_compareAndSwapObject:
555 return inline_unsafe_CAS(T_OBJECT);
556 case vmIntrinsics::_compareAndSwapInt:
557 return inline_unsafe_CAS(T_INT);
558 case vmIntrinsics::_compareAndSwapLong:
559 return inline_unsafe_CAS(T_LONG);
561 case vmIntrinsics::_putOrderedObject:
562 return inline_unsafe_ordered_store(T_OBJECT);
563 case vmIntrinsics::_putOrderedInt:
564 return inline_unsafe_ordered_store(T_INT);
565 case vmIntrinsics::_putOrderedLong:
566 return inline_unsafe_ordered_store(T_LONG);
568 case vmIntrinsics::_currentThread:
569 return inline_native_currentThread();
570 case vmIntrinsics::_isInterrupted:
571 return inline_native_isInterrupted();
573 case vmIntrinsics::_currentTimeMillis:
574 return inline_native_time_funcs(false);
575 case vmIntrinsics::_nanoTime:
576 return inline_native_time_funcs(true);
577 case vmIntrinsics::_allocateInstance:
578 return inline_unsafe_allocate();
579 case vmIntrinsics::_copyMemory:
580 return inline_unsafe_copyMemory();
581 case vmIntrinsics::_newArray:
582 return inline_native_newArray();
583 case vmIntrinsics::_getLength:
584 return inline_native_getLength();
585 case vmIntrinsics::_copyOf:
586 return inline_array_copyOf(false);
587 case vmIntrinsics::_copyOfRange:
588 return inline_array_copyOf(true);
589 case vmIntrinsics::_clone:
590 return inline_native_clone(intrinsic()->is_virtual());
592 case vmIntrinsics::_isAssignableFrom:
593 return inline_native_subtype_check();
595 case vmIntrinsics::_isInstance:
596 case vmIntrinsics::_getModifiers:
597 case vmIntrinsics::_isInterface:
598 case vmIntrinsics::_isArray:
599 case vmIntrinsics::_isPrimitive:
600 case vmIntrinsics::_getSuperclass:
601 case vmIntrinsics::_getComponentType:
602 case vmIntrinsics::_getClassAccessFlags:
603 return inline_native_Class_query(intrinsic_id());
605 case vmIntrinsics::_floatToRawIntBits:
606 case vmIntrinsics::_floatToIntBits:
607 case vmIntrinsics::_intBitsToFloat:
608 case vmIntrinsics::_doubleToRawLongBits:
609 case vmIntrinsics::_doubleToLongBits:
610 case vmIntrinsics::_longBitsToDouble:
611 return inline_fp_conversions(intrinsic_id());
613 case vmIntrinsics::_reverseBytes_i:
614 case vmIntrinsics::_reverseBytes_l:
615 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
617 case vmIntrinsics::_get_AtomicLong:
618 return inline_native_AtomicLong_get();
619 case vmIntrinsics::_attemptUpdate:
620 return inline_native_AtomicLong_attemptUpdate();
622 case vmIntrinsics::_getCallerClass:
623 return inline_native_Reflection_getCallerClass();
625 default:
626 // If you get here, it may be that someone has added a new intrinsic
627 // to the list in vmSymbols.hpp without implementing it here.
628 #ifndef PRODUCT
629 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
630 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
631 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
632 }
633 #endif
634 return false;
635 }
636 }
638 //------------------------------push_result------------------------------
639 // Helper function for finishing intrinsics.
640 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
641 record_for_igvn(region);
642 set_control(_gvn.transform(region));
643 BasicType value_type = value->type()->basic_type();
644 push_node(value_type, _gvn.transform(value));
645 }
647 //------------------------------generate_guard---------------------------
648 // Helper function for generating guarded fast-slow graph structures.
649 // The given 'test', if true, guards a slow path. If the test fails
650 // then a fast path can be taken. (We generally hope it fails.)
651 // In all cases, GraphKit::control() is updated to the fast path.
652 // The returned value represents the control for the slow path.
653 // The return value is never 'top'; it is either a valid control
654 // or NULL if it is obvious that the slow path can never be taken.
655 // Also, if region and the slow control are not NULL, the slow edge
656 // is appended to the region.
657 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
658 if (stopped()) {
659 // Already short circuited.
660 return NULL;
661 }
663 // Build an if node and its projections.
664 // If test is true we take the slow path, which we assume is uncommon.
665 if (_gvn.type(test) == TypeInt::ZERO) {
666 // The slow branch is never taken. No need to build this guard.
667 return NULL;
668 }
670 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
672 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
673 if (if_slow == top()) {
674 // The slow branch is never taken. No need to build this guard.
675 return NULL;
676 }
678 if (region != NULL)
679 region->add_req(if_slow);
681 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
682 set_control(if_fast);
684 return if_slow;
685 }
687 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
688 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
689 }
690 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
691 return generate_guard(test, region, PROB_FAIR);
692 }
694 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
695 Node* *pos_index) {
696 if (stopped())
697 return NULL; // already stopped
698 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
699 return NULL; // index is already adequately typed
700 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
701 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
702 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
703 if (is_neg != NULL && pos_index != NULL) {
704 // Emulate effect of Parse::adjust_map_after_if.
705 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
706 ccast->set_req(0, control());
707 (*pos_index) = _gvn.transform(ccast);
708 }
709 return is_neg;
710 }
712 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
713 Node* *pos_index) {
714 if (stopped())
715 return NULL; // already stopped
716 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
717 return NULL; // index is already adequately typed
718 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
719 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
720 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
721 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
722 if (is_notp != NULL && pos_index != NULL) {
723 // Emulate effect of Parse::adjust_map_after_if.
724 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
725 ccast->set_req(0, control());
726 (*pos_index) = _gvn.transform(ccast);
727 }
728 return is_notp;
729 }
731 // Make sure that 'position' is a valid limit index, in [0..length].
732 // There are two equivalent plans for checking this:
733 // A. (offset + copyLength) unsigned<= arrayLength
734 // B. offset <= (arrayLength - copyLength)
735 // We require that all of the values above, except for the sum and
736 // difference, are already known to be non-negative.
737 // Plan A is robust in the face of overflow, if offset and copyLength
738 // are both hugely positive.
739 //
740 // Plan B is less direct and intuitive, but it does not overflow at
741 // all, since the difference of two non-negatives is always
742 // representable. Whenever Java methods must perform the equivalent
743 // check they generally use Plan B instead of Plan A.
744 // For the moment we use Plan A.
745 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
746 Node* subseq_length,
747 Node* array_length,
748 RegionNode* region) {
749 if (stopped())
750 return NULL; // already stopped
751 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
752 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
753 return NULL; // common case of whole-array copy
754 Node* last = subseq_length;
755 if (!zero_offset) // last += offset
756 last = _gvn.transform( new (C, 3) AddINode(last, offset));
757 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
758 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
759 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
760 return is_over;
761 }
764 //--------------------------generate_current_thread--------------------
765 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
766 ciKlass* thread_klass = env()->Thread_klass();
767 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
768 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
769 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
770 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
771 tls_output = thread;
772 return threadObj;
773 }
776 //------------------------------inline_string_compareTo------------------------
777 bool LibraryCallKit::inline_string_compareTo() {
779 const int value_offset = java_lang_String::value_offset_in_bytes();
780 const int count_offset = java_lang_String::count_offset_in_bytes();
781 const int offset_offset = java_lang_String::offset_offset_in_bytes();
783 _sp += 2;
784 Node *argument = pop(); // pop non-receiver first: it was pushed second
785 Node *receiver = pop();
787 // Null check on self without removing any arguments. The argument
788 // null check technically happens in the wrong place, which can lead to
789 // invalid stack traces when string compare is inlined into a method
790 // which handles NullPointerExceptions.
791 _sp += 2;
792 receiver = do_null_check(receiver, T_OBJECT);
793 argument = do_null_check(argument, T_OBJECT);
794 _sp -= 2;
795 if (stopped()) {
796 return true;
797 }
799 ciInstanceKlass* klass = env()->String_klass();
800 const TypeInstPtr* string_type =
801 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
803 Node* compare =
804 _gvn.transform(new (C, 7) StrCompNode(
805 control(),
806 memory(TypeAryPtr::CHARS),
807 memory(string_type->add_offset(value_offset)),
808 memory(string_type->add_offset(count_offset)),
809 memory(string_type->add_offset(offset_offset)),
810 receiver,
811 argument));
812 push(compare);
813 return true;
814 }
816 // Java version of String.indexOf(constant string)
817 // class StringDecl {
818 // StringDecl(char[] ca) {
819 // offset = 0;
820 // count = ca.length;
821 // value = ca;
822 // }
823 // int offset;
824 // int count;
825 // char[] value;
826 // }
827 //
828 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
829 // int targetOffset, int cache_i, int md2) {
830 // int cache = cache_i;
831 // int sourceOffset = string_object.offset;
832 // int sourceCount = string_object.count;
833 // int targetCount = target_object.length;
834 //
835 // int targetCountLess1 = targetCount - 1;
836 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
837 //
838 // char[] source = string_object.value;
839 // char[] target = target_object;
840 // int lastChar = target[targetCountLess1];
841 //
842 // outer_loop:
843 // for (int i = sourceOffset; i < sourceEnd; ) {
844 // int src = source[i + targetCountLess1];
845 // if (src == lastChar) {
846 // // With random strings and a 4-character alphabet,
847 // // reverse matching at this point sets up 0.8% fewer
848 // // frames, but (paradoxically) makes 0.3% more probes.
849 // // Since those probes are nearer the lastChar probe,
850 // // there is may be a net D$ win with reverse matching.
851 // // But, reversing loop inhibits unroll of inner loop
852 // // for unknown reason. So, does running outer loop from
853 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
854 // for (int j = 0; j < targetCountLess1; j++) {
855 // if (target[targetOffset + j] != source[i+j]) {
856 // if ((cache & (1 << source[i+j])) == 0) {
857 // if (md2 < j+1) {
858 // i += j+1;
859 // continue outer_loop;
860 // }
861 // }
862 // i += md2;
863 // continue outer_loop;
864 // }
865 // }
866 // return i - sourceOffset;
867 // }
868 // if ((cache & (1 << src)) == 0) {
869 // i += targetCountLess1;
870 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
871 // i++;
872 // }
873 // return -1;
874 // }
876 //------------------------------string_indexOf------------------------
877 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
878 jint cache_i, jint md2_i) {
880 Node* no_ctrl = NULL;
881 float likely = PROB_LIKELY(0.9);
882 float unlikely = PROB_UNLIKELY(0.9);
884 const int value_offset = java_lang_String::value_offset_in_bytes();
885 const int count_offset = java_lang_String::count_offset_in_bytes();
886 const int offset_offset = java_lang_String::offset_offset_in_bytes();
888 ciInstanceKlass* klass = env()->String_klass();
889 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
890 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
892 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
893 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
894 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
895 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
896 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
897 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
899 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
900 jint target_length = target_array->length();
901 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
902 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
904 IdealKit kit(gvn(), control(), merged_memory());
905 #define __ kit.
906 Node* zero = __ ConI(0);
907 Node* one = __ ConI(1);
908 Node* cache = __ ConI(cache_i);
909 Node* md2 = __ ConI(md2_i);
910 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
911 Node* targetCount = __ ConI(target_length);
912 Node* targetCountLess1 = __ ConI(target_length - 1);
913 Node* targetOffset = __ ConI(targetOffset_i);
914 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
916 IdealVariable rtn(kit), i(kit), j(kit); __ declares_done();
917 Node* outer_loop = __ make_label(2 /* goto */);
918 Node* return_ = __ make_label(1);
920 __ set(rtn,__ ConI(-1));
921 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
922 Node* i2 = __ AddI(__ value(i), targetCountLess1);
923 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
924 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
925 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
926 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
927 Node* tpj = __ AddI(targetOffset, __ value(j));
928 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
929 Node* ipj = __ AddI(__ value(i), __ value(j));
930 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
931 __ if_then(targ, BoolTest::ne, src2); {
932 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
933 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
934 __ increment(i, __ AddI(__ value(j), one));
935 __ goto_(outer_loop);
936 } __ end_if(); __ dead(j);
937 }__ end_if(); __ dead(j);
938 __ increment(i, md2);
939 __ goto_(outer_loop);
940 }__ end_if();
941 __ increment(j, one);
942 }__ end_loop(); __ dead(j);
943 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
944 __ goto_(return_);
945 }__ end_if();
946 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
947 __ increment(i, targetCountLess1);
948 }__ end_if();
949 __ increment(i, one);
950 __ bind(outer_loop);
951 }__ end_loop(); __ dead(i);
952 __ bind(return_);
953 __ drain_delay_transform();
955 set_control(__ ctrl());
956 Node* result = __ value(rtn);
957 #undef __
958 C->set_has_loops(true);
959 return result;
960 }
963 //------------------------------inline_string_indexOf------------------------
964 bool LibraryCallKit::inline_string_indexOf() {
966 _sp += 2;
967 Node *argument = pop(); // pop non-receiver first: it was pushed second
968 Node *receiver = pop();
970 // don't intrinsify is argument isn't a constant string.
971 if (!argument->is_Con()) {
972 return false;
973 }
974 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
975 if (str_type == NULL) {
976 return false;
977 }
978 ciInstanceKlass* klass = env()->String_klass();
979 ciObject* str_const = str_type->const_oop();
980 if (str_const == NULL || str_const->klass() != klass) {
981 return false;
982 }
983 ciInstance* str = str_const->as_instance();
984 assert(str != NULL, "must be instance");
986 const int value_offset = java_lang_String::value_offset_in_bytes();
987 const int count_offset = java_lang_String::count_offset_in_bytes();
988 const int offset_offset = java_lang_String::offset_offset_in_bytes();
990 ciObject* v = str->field_value_by_offset(value_offset).as_object();
991 int o = str->field_value_by_offset(offset_offset).as_int();
992 int c = str->field_value_by_offset(count_offset).as_int();
993 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
995 // constant strings have no offset and count == length which
996 // simplifies the resulting code somewhat so lets optimize for that.
997 if (o != 0 || c != pat->length()) {
998 return false;
999 }
1001 // Null check on self without removing any arguments. The argument
1002 // null check technically happens in the wrong place, which can lead to
1003 // invalid stack traces when string compare is inlined into a method
1004 // which handles NullPointerExceptions.
1005 _sp += 2;
1006 receiver = do_null_check(receiver, T_OBJECT);
1007 // No null check on the argument is needed since it's a constant String oop.
1008 _sp -= 2;
1009 if (stopped()) {
1010 return true;
1011 }
1013 // The null string as a pattern always returns 0 (match at beginning of string)
1014 if (c == 0) {
1015 push(intcon(0));
1016 return true;
1017 }
1019 jchar lastChar = pat->char_at(o + (c - 1));
1020 int cache = 0;
1021 int i;
1022 for (i = 0; i < c - 1; i++) {
1023 assert(i < pat->length(), "out of range");
1024 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1025 }
1027 int md2 = c;
1028 for (i = 0; i < c - 1; i++) {
1029 assert(i < pat->length(), "out of range");
1030 if (pat->char_at(o + i) == lastChar) {
1031 md2 = (c - 1) - i;
1032 }
1033 }
1035 Node* result = string_indexOf(receiver, pat, o, cache, md2);
1036 push(result);
1037 return true;
1038 }
1040 //--------------------------pop_math_arg--------------------------------
1041 // Pop a double argument to a math function from the stack
1042 // rounding it if necessary.
1043 Node * LibraryCallKit::pop_math_arg() {
1044 Node *arg = pop_pair();
1045 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1046 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1047 return arg;
1048 }
1050 //------------------------------inline_trig----------------------------------
1051 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1052 // argument reduction which will turn into a fast/slow diamond.
1053 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1054 _sp += arg_size(); // restore stack pointer
1055 Node* arg = pop_math_arg();
1056 Node* trig = NULL;
1058 switch (id) {
1059 case vmIntrinsics::_dsin:
1060 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1061 break;
1062 case vmIntrinsics::_dcos:
1063 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1064 break;
1065 case vmIntrinsics::_dtan:
1066 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1067 break;
1068 default:
1069 assert(false, "bad intrinsic was passed in");
1070 return false;
1071 }
1073 // Rounding required? Check for argument reduction!
1074 if( Matcher::strict_fp_requires_explicit_rounding ) {
1076 static const double pi_4 = 0.7853981633974483;
1077 static const double neg_pi_4 = -0.7853981633974483;
1078 // pi/2 in 80-bit extended precision
1079 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1080 // -pi/2 in 80-bit extended precision
1081 // 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};
1082 // Cutoff value for using this argument reduction technique
1083 //static const double pi_2_minus_epsilon = 1.564660403643354;
1084 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1086 // Pseudocode for sin:
1087 // if (x <= Math.PI / 4.0) {
1088 // if (x >= -Math.PI / 4.0) return fsin(x);
1089 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1090 // } else {
1091 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1092 // }
1093 // return StrictMath.sin(x);
1095 // Pseudocode for cos:
1096 // if (x <= Math.PI / 4.0) {
1097 // if (x >= -Math.PI / 4.0) return fcos(x);
1098 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1099 // } else {
1100 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1101 // }
1102 // return StrictMath.cos(x);
1104 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1105 // requires a special machine instruction to load it. Instead we'll try
1106 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1107 // probably do the math inside the SIN encoding.
1109 // Make the merge point
1110 RegionNode *r = new (C, 3) RegionNode(3);
1111 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1113 // Flatten arg so we need only 1 test
1114 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1115 // Node for PI/4 constant
1116 Node *pi4 = makecon(TypeD::make(pi_4));
1117 // Check PI/4 : abs(arg)
1118 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1119 // Check: If PI/4 < abs(arg) then go slow
1120 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1121 // Branch either way
1122 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1123 set_control(opt_iff(r,iff));
1125 // Set fast path result
1126 phi->init_req(2,trig);
1128 // Slow path - non-blocking leaf call
1129 Node* call = NULL;
1130 switch (id) {
1131 case vmIntrinsics::_dsin:
1132 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1133 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1134 "Sin", NULL, arg, top());
1135 break;
1136 case vmIntrinsics::_dcos:
1137 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1138 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1139 "Cos", NULL, arg, top());
1140 break;
1141 case vmIntrinsics::_dtan:
1142 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1143 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1144 "Tan", NULL, arg, top());
1145 break;
1146 }
1147 assert(control()->in(0) == call, "");
1148 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1149 r->init_req(1,control());
1150 phi->init_req(1,slow_result);
1152 // Post-merge
1153 set_control(_gvn.transform(r));
1154 record_for_igvn(r);
1155 trig = _gvn.transform(phi);
1157 C->set_has_split_ifs(true); // Has chance for split-if optimization
1158 }
1159 // Push result back on JVM stack
1160 push_pair(trig);
1161 return true;
1162 }
1164 //------------------------------inline_sqrt-------------------------------------
1165 // Inline square root instruction, if possible.
1166 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1167 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1168 _sp += arg_size(); // restore stack pointer
1169 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1170 return true;
1171 }
1173 //------------------------------inline_abs-------------------------------------
1174 // Inline absolute value instruction, if possible.
1175 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1176 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1177 _sp += arg_size(); // restore stack pointer
1178 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1179 return true;
1180 }
1182 //------------------------------inline_exp-------------------------------------
1183 // Inline exp instructions, if possible. The Intel hardware only misses
1184 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1185 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1186 assert(id == vmIntrinsics::_dexp, "Not exp");
1188 // If this inlining ever returned NaN in the past, we do not intrinsify it
1189 // every again. NaN results requires StrictMath.exp handling.
1190 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1192 // Do not intrinsify on older platforms which lack cmove.
1193 if (ConditionalMoveLimit == 0) return false;
1195 _sp += arg_size(); // restore stack pointer
1196 Node *x = pop_math_arg();
1197 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1199 //-------------------
1200 //result=(result.isNaN())? StrictMath::exp():result;
1201 // Check: If isNaN() by checking result!=result? then go to Strict Math
1202 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1203 // Build the boolean node
1204 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1206 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1207 // End the current control-flow path
1208 push_pair(x);
1209 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1210 // to handle. Recompile without intrinsifying Math.exp
1211 uncommon_trap(Deoptimization::Reason_intrinsic,
1212 Deoptimization::Action_make_not_entrant);
1213 }
1215 C->set_has_split_ifs(true); // Has chance for split-if optimization
1217 push_pair(result);
1219 return true;
1220 }
1222 //------------------------------inline_pow-------------------------------------
1223 // Inline power instructions, if possible.
1224 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1225 assert(id == vmIntrinsics::_dpow, "Not pow");
1227 // If this inlining ever returned NaN in the past, we do not intrinsify it
1228 // every again. NaN results requires StrictMath.pow handling.
1229 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1231 // Do not intrinsify on older platforms which lack cmove.
1232 if (ConditionalMoveLimit == 0) return false;
1234 // Pseudocode for pow
1235 // if (x <= 0.0) {
1236 // if ((double)((int)y)==y) { // if y is int
1237 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1238 // } else {
1239 // result = NaN;
1240 // }
1241 // } else {
1242 // result = DPow(x,y);
1243 // }
1244 // if (result != result)? {
1245 // ucommon_trap();
1246 // }
1247 // return result;
1249 _sp += arg_size(); // restore stack pointer
1250 Node* y = pop_math_arg();
1251 Node* x = pop_math_arg();
1253 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1255 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1256 // inside of something) then skip the fancy tests and just check for
1257 // NaN result.
1258 Node *result = NULL;
1259 if( jvms()->depth() >= 1 ) {
1260 result = fast_result;
1261 } else {
1263 // Set the merge point for If node with condition of (x <= 0.0)
1264 // There are four possible paths to region node and phi node
1265 RegionNode *r = new (C, 4) RegionNode(4);
1266 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1268 // Build the first if node: if (x <= 0.0)
1269 // Node for 0 constant
1270 Node *zeronode = makecon(TypeD::ZERO);
1271 // Check x:0
1272 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1273 // Check: If (x<=0) then go complex path
1274 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1275 // Branch either way
1276 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1277 Node *opt_test = _gvn.transform(if1);
1278 //assert( opt_test->is_If(), "Expect an IfNode");
1279 IfNode *opt_if1 = (IfNode*)opt_test;
1280 // Fast path taken; set region slot 3
1281 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1282 r->init_req(3,fast_taken); // Capture fast-control
1284 // Fast path not-taken, i.e. slow path
1285 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1287 // Set fast path result
1288 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1289 phi->init_req(3, fast_result);
1291 // Complex path
1292 // Build the second if node (if y is int)
1293 // Node for (int)y
1294 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1295 // Node for (double)((int) y)
1296 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1297 // Check (double)((int) y) : y
1298 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1299 // Check if (y isn't int) then go to slow path
1301 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1302 // Branch eith way
1303 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1304 Node *slow_path = opt_iff(r,if2); // Set region path 2
1306 // Calculate DPow(abs(x), y)*(1 & (int)y)
1307 // Node for constant 1
1308 Node *conone = intcon(1);
1309 // 1& (int)y
1310 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1311 // zero node
1312 Node *conzero = intcon(0);
1313 // Check (1&(int)y)==0?
1314 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1315 // Check if (1&(int)y)!=0?, if so the result is negative
1316 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1317 // abs(x)
1318 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1319 // abs(x)^y
1320 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1321 // -abs(x)^y
1322 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1323 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1324 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1325 // Set complex path fast result
1326 phi->init_req(2, signresult);
1328 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1329 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1330 r->init_req(1,slow_path);
1331 phi->init_req(1,slow_result);
1333 // Post merge
1334 set_control(_gvn.transform(r));
1335 record_for_igvn(r);
1336 result=_gvn.transform(phi);
1337 }
1339 //-------------------
1340 //result=(result.isNaN())? uncommon_trap():result;
1341 // Check: If isNaN() by checking result!=result? then go to Strict Math
1342 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1343 // Build the boolean node
1344 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1346 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1347 // End the current control-flow path
1348 push_pair(x);
1349 push_pair(y);
1350 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1351 // to handle. Recompile without intrinsifying Math.pow.
1352 uncommon_trap(Deoptimization::Reason_intrinsic,
1353 Deoptimization::Action_make_not_entrant);
1354 }
1356 C->set_has_split_ifs(true); // Has chance for split-if optimization
1358 push_pair(result);
1360 return true;
1361 }
1363 //------------------------------inline_trans-------------------------------------
1364 // Inline transcendental instructions, if possible. The Intel hardware gets
1365 // these right, no funny corner cases missed.
1366 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1367 _sp += arg_size(); // restore stack pointer
1368 Node* arg = pop_math_arg();
1369 Node* trans = NULL;
1371 switch (id) {
1372 case vmIntrinsics::_dlog:
1373 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1374 break;
1375 case vmIntrinsics::_dlog10:
1376 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1377 break;
1378 default:
1379 assert(false, "bad intrinsic was passed in");
1380 return false;
1381 }
1383 // Push result back on JVM stack
1384 push_pair(trans);
1385 return true;
1386 }
1388 //------------------------------runtime_math-----------------------------
1389 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1390 Node* a = NULL;
1391 Node* b = NULL;
1393 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1394 "must be (DD)D or (D)D type");
1396 // Inputs
1397 _sp += arg_size(); // restore stack pointer
1398 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1399 b = pop_math_arg();
1400 }
1401 a = pop_math_arg();
1403 const TypePtr* no_memory_effects = NULL;
1404 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1405 no_memory_effects,
1406 a, top(), b, b ? top() : NULL);
1407 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1408 #ifdef ASSERT
1409 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1410 assert(value_top == top(), "second value must be top");
1411 #endif
1413 push_pair(value);
1414 return true;
1415 }
1417 //------------------------------inline_math_native-----------------------------
1418 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1419 switch (id) {
1420 // These intrinsics are not properly supported on all hardware
1421 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1422 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1423 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1424 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1425 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1426 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1428 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1429 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1430 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1431 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1433 // These intrinsics are supported on all hardware
1434 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1435 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1437 // These intrinsics don't work on X86. The ad implementation doesn't
1438 // handle NaN's properly. Instead of returning infinity, the ad
1439 // implementation returns a NaN on overflow. See bug: 6304089
1440 // Once the ad implementations are fixed, change the code below
1441 // to match the intrinsics above
1443 case vmIntrinsics::_dexp: return
1444 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1445 case vmIntrinsics::_dpow: return
1446 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1448 // These intrinsics are not yet correctly implemented
1449 case vmIntrinsics::_datan2:
1450 return false;
1452 default:
1453 ShouldNotReachHere();
1454 return false;
1455 }
1456 }
1458 static bool is_simple_name(Node* n) {
1459 return (n->req() == 1 // constant
1460 || (n->is_Type() && n->as_Type()->type()->singleton())
1461 || n->is_Proj() // parameter or return value
1462 || n->is_Phi() // local of some sort
1463 );
1464 }
1466 //----------------------------inline_min_max-----------------------------------
1467 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1468 push(generate_min_max(id, argument(0), argument(1)));
1470 return true;
1471 }
1473 Node*
1474 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1475 // These are the candidate return value:
1476 Node* xvalue = x0;
1477 Node* yvalue = y0;
1479 if (xvalue == yvalue) {
1480 return xvalue;
1481 }
1483 bool want_max = (id == vmIntrinsics::_max);
1485 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1486 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1487 if (txvalue == NULL || tyvalue == NULL) return top();
1488 // This is not really necessary, but it is consistent with a
1489 // hypothetical MaxINode::Value method:
1490 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1492 // %%% This folding logic should (ideally) be in a different place.
1493 // Some should be inside IfNode, and there to be a more reliable
1494 // transformation of ?: style patterns into cmoves. We also want
1495 // more powerful optimizations around cmove and min/max.
1497 // Try to find a dominating comparison of these guys.
1498 // It can simplify the index computation for Arrays.copyOf
1499 // and similar uses of System.arraycopy.
1500 // First, compute the normalized version of CmpI(x, y).
1501 int cmp_op = Op_CmpI;
1502 Node* xkey = xvalue;
1503 Node* ykey = yvalue;
1504 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1505 if (ideal_cmpxy->is_Cmp()) {
1506 // E.g., if we have CmpI(length - offset, count),
1507 // it might idealize to CmpI(length, count + offset)
1508 cmp_op = ideal_cmpxy->Opcode();
1509 xkey = ideal_cmpxy->in(1);
1510 ykey = ideal_cmpxy->in(2);
1511 }
1513 // Start by locating any relevant comparisons.
1514 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1515 Node* cmpxy = NULL;
1516 Node* cmpyx = NULL;
1517 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1518 Node* cmp = start_from->fast_out(k);
1519 if (cmp->outcnt() > 0 && // must have prior uses
1520 cmp->in(0) == NULL && // must be context-independent
1521 cmp->Opcode() == cmp_op) { // right kind of compare
1522 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1523 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1524 }
1525 }
1527 const int NCMPS = 2;
1528 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1529 int cmpn;
1530 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1531 if (cmps[cmpn] != NULL) break; // find a result
1532 }
1533 if (cmpn < NCMPS) {
1534 // Look for a dominating test that tells us the min and max.
1535 int depth = 0; // Limit search depth for speed
1536 Node* dom = control();
1537 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1538 if (++depth >= 100) break;
1539 Node* ifproj = dom;
1540 if (!ifproj->is_Proj()) continue;
1541 Node* iff = ifproj->in(0);
1542 if (!iff->is_If()) continue;
1543 Node* bol = iff->in(1);
1544 if (!bol->is_Bool()) continue;
1545 Node* cmp = bol->in(1);
1546 if (cmp == NULL) continue;
1547 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1548 if (cmps[cmpn] == cmp) break;
1549 if (cmpn == NCMPS) continue;
1550 BoolTest::mask btest = bol->as_Bool()->_test._test;
1551 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1552 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1553 // At this point, we know that 'x btest y' is true.
1554 switch (btest) {
1555 case BoolTest::eq:
1556 // They are proven equal, so we can collapse the min/max.
1557 // Either value is the answer. Choose the simpler.
1558 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1559 return yvalue;
1560 return xvalue;
1561 case BoolTest::lt: // x < y
1562 case BoolTest::le: // x <= y
1563 return (want_max ? yvalue : xvalue);
1564 case BoolTest::gt: // x > y
1565 case BoolTest::ge: // x >= y
1566 return (want_max ? xvalue : yvalue);
1567 }
1568 }
1569 }
1571 // We failed to find a dominating test.
1572 // Let's pick a test that might GVN with prior tests.
1573 Node* best_bol = NULL;
1574 BoolTest::mask best_btest = BoolTest::illegal;
1575 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1576 Node* cmp = cmps[cmpn];
1577 if (cmp == NULL) continue;
1578 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1579 Node* bol = cmp->fast_out(j);
1580 if (!bol->is_Bool()) continue;
1581 BoolTest::mask btest = bol->as_Bool()->_test._test;
1582 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1583 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1584 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1585 best_bol = bol->as_Bool();
1586 best_btest = btest;
1587 }
1588 }
1589 }
1591 Node* answer_if_true = NULL;
1592 Node* answer_if_false = NULL;
1593 switch (best_btest) {
1594 default:
1595 if (cmpxy == NULL)
1596 cmpxy = ideal_cmpxy;
1597 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1598 // and fall through:
1599 case BoolTest::lt: // x < y
1600 case BoolTest::le: // x <= y
1601 answer_if_true = (want_max ? yvalue : xvalue);
1602 answer_if_false = (want_max ? xvalue : yvalue);
1603 break;
1604 case BoolTest::gt: // x > y
1605 case BoolTest::ge: // x >= y
1606 answer_if_true = (want_max ? xvalue : yvalue);
1607 answer_if_false = (want_max ? yvalue : xvalue);
1608 break;
1609 }
1611 jint hi, lo;
1612 if (want_max) {
1613 // We can sharpen the minimum.
1614 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1615 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1616 } else {
1617 // We can sharpen the maximum.
1618 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1619 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1620 }
1622 // Use a flow-free graph structure, to avoid creating excess control edges
1623 // which could hinder other optimizations.
1624 // Since Math.min/max is often used with arraycopy, we want
1625 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1626 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1627 answer_if_false, answer_if_true,
1628 TypeInt::make(lo, hi, widen));
1630 return _gvn.transform(cmov);
1632 /*
1633 // This is not as desirable as it may seem, since Min and Max
1634 // nodes do not have a full set of optimizations.
1635 // And they would interfere, anyway, with 'if' optimizations
1636 // and with CMoveI canonical forms.
1637 switch (id) {
1638 case vmIntrinsics::_min:
1639 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1640 case vmIntrinsics::_max:
1641 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1642 default:
1643 ShouldNotReachHere();
1644 }
1645 */
1646 }
1648 inline int
1649 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1650 const TypePtr* base_type = TypePtr::NULL_PTR;
1651 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1652 if (base_type == NULL) {
1653 // Unknown type.
1654 return Type::AnyPtr;
1655 } else if (base_type == TypePtr::NULL_PTR) {
1656 // Since this is a NULL+long form, we have to switch to a rawptr.
1657 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1658 offset = MakeConX(0);
1659 return Type::RawPtr;
1660 } else if (base_type->base() == Type::RawPtr) {
1661 return Type::RawPtr;
1662 } else if (base_type->isa_oopptr()) {
1663 // Base is never null => always a heap address.
1664 if (base_type->ptr() == TypePtr::NotNull) {
1665 return Type::OopPtr;
1666 }
1667 // Offset is small => always a heap address.
1668 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1669 if (offset_type != NULL &&
1670 base_type->offset() == 0 && // (should always be?)
1671 offset_type->_lo >= 0 &&
1672 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1673 return Type::OopPtr;
1674 }
1675 // Otherwise, it might either be oop+off or NULL+addr.
1676 return Type::AnyPtr;
1677 } else {
1678 // No information:
1679 return Type::AnyPtr;
1680 }
1681 }
1683 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1684 int kind = classify_unsafe_addr(base, offset);
1685 if (kind == Type::RawPtr) {
1686 return basic_plus_adr(top(), base, offset);
1687 } else {
1688 return basic_plus_adr(base, offset);
1689 }
1690 }
1692 //----------------------------inline_reverseBytes_int/long-------------------
1693 // inline Int.reverseBytes(int)
1694 // inline Long.reverseByes(long)
1695 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
1696 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
1697 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
1698 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
1699 _sp += arg_size(); // restore stack pointer
1700 switch (id) {
1701 case vmIntrinsics::_reverseBytes_i:
1702 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
1703 break;
1704 case vmIntrinsics::_reverseBytes_l:
1705 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
1706 break;
1707 default:
1708 ;
1709 }
1710 return true;
1711 }
1713 //----------------------------inline_unsafe_access----------------------------
1715 const static BasicType T_ADDRESS_HOLDER = T_LONG;
1717 // Interpret Unsafe.fieldOffset cookies correctly:
1718 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
1720 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
1721 if (callee()->is_static()) return false; // caller must have the capability!
1723 #ifndef PRODUCT
1724 {
1725 ResourceMark rm;
1726 // Check the signatures.
1727 ciSignature* sig = signature();
1728 #ifdef ASSERT
1729 if (!is_store) {
1730 // Object getObject(Object base, int/long offset), etc.
1731 BasicType rtype = sig->return_type()->basic_type();
1732 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
1733 rtype = T_ADDRESS; // it is really a C void*
1734 assert(rtype == type, "getter must return the expected value");
1735 if (!is_native_ptr) {
1736 assert(sig->count() == 2, "oop getter has 2 arguments");
1737 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
1738 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
1739 } else {
1740 assert(sig->count() == 1, "native getter has 1 argument");
1741 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
1742 }
1743 } else {
1744 // void putObject(Object base, int/long offset, Object x), etc.
1745 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
1746 if (!is_native_ptr) {
1747 assert(sig->count() == 3, "oop putter has 3 arguments");
1748 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
1749 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
1750 } else {
1751 assert(sig->count() == 2, "native putter has 2 arguments");
1752 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
1753 }
1754 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
1755 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
1756 vtype = T_ADDRESS; // it is really a C void*
1757 assert(vtype == type, "putter must accept the expected value");
1758 }
1759 #endif // ASSERT
1760 }
1761 #endif //PRODUCT
1763 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1765 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
1767 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
1768 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
1770 debug_only(int saved_sp = _sp);
1771 _sp += nargs;
1773 Node* val;
1774 debug_only(val = (Node*)(uintptr_t)-1);
1777 if (is_store) {
1778 // Get the value being stored. (Pop it first; it was pushed last.)
1779 switch (type) {
1780 case T_DOUBLE:
1781 case T_LONG:
1782 case T_ADDRESS:
1783 val = pop_pair();
1784 break;
1785 default:
1786 val = pop();
1787 }
1788 }
1790 // Build address expression. See the code in inline_unsafe_prefetch.
1791 Node *adr;
1792 Node *heap_base_oop = top();
1793 if (!is_native_ptr) {
1794 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
1795 Node* offset = pop_pair();
1796 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
1797 Node* base = pop();
1798 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1799 // to be plain byte offsets, which are also the same as those accepted
1800 // by oopDesc::field_base.
1801 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1802 "fieldOffset must be byte-scaled");
1803 // 32-bit machines ignore the high half!
1804 offset = ConvL2X(offset);
1805 adr = make_unsafe_address(base, offset);
1806 heap_base_oop = base;
1807 } else {
1808 Node* ptr = pop_pair();
1809 // Adjust Java long to machine word:
1810 ptr = ConvL2X(ptr);
1811 adr = make_unsafe_address(NULL, ptr);
1812 }
1814 // Pop receiver last: it was pushed first.
1815 Node *receiver = pop();
1817 assert(saved_sp == _sp, "must have correct argument count");
1819 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1821 // First guess at the value type.
1822 const Type *value_type = Type::get_const_basic_type(type);
1824 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
1825 // there was not enough information to nail it down.
1826 Compile::AliasType* alias_type = C->alias_type(adr_type);
1827 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
1829 // We will need memory barriers unless we can determine a unique
1830 // alias category for this reference. (Note: If for some reason
1831 // the barriers get omitted and the unsafe reference begins to "pollute"
1832 // the alias analysis of the rest of the graph, either Compile::can_alias
1833 // or Compile::must_alias will throw a diagnostic assert.)
1834 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
1836 if (!is_store && type == T_OBJECT) {
1837 // Attempt to infer a sharper value type from the offset and base type.
1838 ciKlass* sharpened_klass = NULL;
1840 // See if it is an instance field, with an object type.
1841 if (alias_type->field() != NULL) {
1842 assert(!is_native_ptr, "native pointer op cannot use a java address");
1843 if (alias_type->field()->type()->is_klass()) {
1844 sharpened_klass = alias_type->field()->type()->as_klass();
1845 }
1846 }
1848 // See if it is a narrow oop array.
1849 if (adr_type->isa_aryptr()) {
1850 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) {
1851 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
1852 if (elem_type != NULL) {
1853 sharpened_klass = elem_type->klass();
1854 }
1855 }
1856 }
1858 if (sharpened_klass != NULL) {
1859 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
1861 // Sharpen the value type.
1862 value_type = tjp;
1864 #ifndef PRODUCT
1865 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
1866 tty->print(" from base type: "); adr_type->dump();
1867 tty->print(" sharpened value: "); value_type->dump();
1868 }
1869 #endif
1870 }
1871 }
1873 // Null check on self without removing any arguments. The argument
1874 // null check technically happens in the wrong place, which can lead to
1875 // invalid stack traces when the primitive is inlined into a method
1876 // which handles NullPointerExceptions.
1877 _sp += nargs;
1878 do_null_check(receiver, T_OBJECT);
1879 _sp -= nargs;
1880 if (stopped()) {
1881 return true;
1882 }
1883 // Heap pointers get a null-check from the interpreter,
1884 // as a courtesy. However, this is not guaranteed by Unsafe,
1885 // and it is not possible to fully distinguish unintended nulls
1886 // from intended ones in this API.
1888 if (is_volatile) {
1889 // We need to emit leading and trailing CPU membars (see below) in
1890 // addition to memory membars when is_volatile. This is a little
1891 // too strong, but avoids the need to insert per-alias-type
1892 // volatile membars (for stores; compare Parse::do_put_xxx), which
1893 // we cannot do effctively here because we probably only have a
1894 // rough approximation of type.
1895 need_mem_bar = true;
1896 // For Stores, place a memory ordering barrier now.
1897 if (is_store)
1898 insert_mem_bar(Op_MemBarRelease);
1899 }
1901 // Memory barrier to prevent normal and 'unsafe' accesses from
1902 // bypassing each other. Happens after null checks, so the
1903 // exception paths do not take memory state from the memory barrier,
1904 // so there's no problems making a strong assert about mixing users
1905 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
1906 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
1907 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
1909 if (!is_store) {
1910 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
1911 // load value and push onto stack
1912 switch (type) {
1913 case T_BOOLEAN:
1914 case T_CHAR:
1915 case T_BYTE:
1916 case T_SHORT:
1917 case T_INT:
1918 case T_FLOAT:
1919 case T_OBJECT:
1920 push( p );
1921 break;
1922 case T_ADDRESS:
1923 // Cast to an int type.
1924 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
1925 p = ConvX2L(p);
1926 push_pair(p);
1927 break;
1928 case T_DOUBLE:
1929 case T_LONG:
1930 push_pair( p );
1931 break;
1932 default: ShouldNotReachHere();
1933 }
1934 } else {
1935 // place effect of store into memory
1936 switch (type) {
1937 case T_DOUBLE:
1938 val = dstore_rounding(val);
1939 break;
1940 case T_ADDRESS:
1941 // Repackage the long as a pointer.
1942 val = ConvL2X(val);
1943 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
1944 break;
1945 }
1947 if (type != T_OBJECT ) {
1948 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
1949 } else {
1950 // Possibly an oop being stored to Java heap or native memory
1951 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
1952 // oop to Java heap.
1953 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
1954 } else {
1956 // We can't tell at compile time if we are storing in the Java heap or outside
1957 // of it. So we need to emit code to conditionally do the proper type of
1958 // store.
1960 IdealKit kit(gvn(), control(), merged_memory());
1961 kit.declares_done();
1962 // QQQ who knows what probability is here??
1963 kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
1964 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
1965 } kit.else_(); {
1966 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
1967 } kit.end_if();
1968 }
1969 }
1970 }
1972 if (is_volatile) {
1973 if (!is_store)
1974 insert_mem_bar(Op_MemBarAcquire);
1975 else
1976 insert_mem_bar(Op_MemBarVolatile);
1977 }
1979 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
1981 return true;
1982 }
1984 //----------------------------inline_unsafe_prefetch----------------------------
1986 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
1987 #ifndef PRODUCT
1988 {
1989 ResourceMark rm;
1990 // Check the signatures.
1991 ciSignature* sig = signature();
1992 #ifdef ASSERT
1993 // Object getObject(Object base, int/long offset), etc.
1994 BasicType rtype = sig->return_type()->basic_type();
1995 if (!is_native_ptr) {
1996 assert(sig->count() == 2, "oop prefetch has 2 arguments");
1997 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
1998 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
1999 } else {
2000 assert(sig->count() == 1, "native prefetch has 1 argument");
2001 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2002 }
2003 #endif // ASSERT
2004 }
2005 #endif // !PRODUCT
2007 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2009 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2010 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2012 debug_only(int saved_sp = _sp);
2013 _sp += nargs;
2015 // Build address expression. See the code in inline_unsafe_access.
2016 Node *adr;
2017 if (!is_native_ptr) {
2018 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2019 Node* offset = pop_pair();
2020 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2021 Node* base = pop();
2022 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2023 // to be plain byte offsets, which are also the same as those accepted
2024 // by oopDesc::field_base.
2025 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2026 "fieldOffset must be byte-scaled");
2027 // 32-bit machines ignore the high half!
2028 offset = ConvL2X(offset);
2029 adr = make_unsafe_address(base, offset);
2030 } else {
2031 Node* ptr = pop_pair();
2032 // Adjust Java long to machine word:
2033 ptr = ConvL2X(ptr);
2034 adr = make_unsafe_address(NULL, ptr);
2035 }
2037 if (is_static) {
2038 assert(saved_sp == _sp, "must have correct argument count");
2039 } else {
2040 // Pop receiver last: it was pushed first.
2041 Node *receiver = pop();
2042 assert(saved_sp == _sp, "must have correct argument count");
2044 // Null check on self without removing any arguments. The argument
2045 // null check technically happens in the wrong place, which can lead to
2046 // invalid stack traces when the primitive is inlined into a method
2047 // which handles NullPointerExceptions.
2048 _sp += nargs;
2049 do_null_check(receiver, T_OBJECT);
2050 _sp -= nargs;
2051 if (stopped()) {
2052 return true;
2053 }
2054 }
2056 // Generate the read or write prefetch
2057 Node *prefetch;
2058 if (is_store) {
2059 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2060 } else {
2061 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2062 }
2063 prefetch->init_req(0, control());
2064 set_i_o(_gvn.transform(prefetch));
2066 return true;
2067 }
2069 //----------------------------inline_unsafe_CAS----------------------------
2071 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2072 // This basic scheme here is the same as inline_unsafe_access, but
2073 // differs in enough details that combining them would make the code
2074 // overly confusing. (This is a true fact! I originally combined
2075 // them, but even I was confused by it!) As much code/comments as
2076 // possible are retained from inline_unsafe_access though to make
2077 // the correspondances clearer. - dl
2079 if (callee()->is_static()) return false; // caller must have the capability!
2081 #ifndef PRODUCT
2082 {
2083 ResourceMark rm;
2084 // Check the signatures.
2085 ciSignature* sig = signature();
2086 #ifdef ASSERT
2087 BasicType rtype = sig->return_type()->basic_type();
2088 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2089 assert(sig->count() == 4, "CAS has 4 arguments");
2090 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2091 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2092 #endif // ASSERT
2093 }
2094 #endif //PRODUCT
2096 // number of stack slots per value argument (1 or 2)
2097 int type_words = type2size[type];
2099 // Cannot inline wide CAS on machines that don't support it natively
2100 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2101 return false;
2103 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2105 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2106 int nargs = 1 + 1 + 2 + type_words + type_words;
2108 // pop arguments: newval, oldval, offset, base, and receiver
2109 debug_only(int saved_sp = _sp);
2110 _sp += nargs;
2111 Node* newval = (type_words == 1) ? pop() : pop_pair();
2112 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2113 Node *offset = pop_pair();
2114 Node *base = pop();
2115 Node *receiver = pop();
2116 assert(saved_sp == _sp, "must have correct argument count");
2118 // Null check receiver.
2119 _sp += nargs;
2120 do_null_check(receiver, T_OBJECT);
2121 _sp -= nargs;
2122 if (stopped()) {
2123 return true;
2124 }
2126 // Build field offset expression.
2127 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2128 // to be plain byte offsets, which are also the same as those accepted
2129 // by oopDesc::field_base.
2130 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2131 // 32-bit machines ignore the high half of long offsets
2132 offset = ConvL2X(offset);
2133 Node* adr = make_unsafe_address(base, offset);
2134 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2136 // (Unlike inline_unsafe_access, there seems no point in trying
2137 // to refine types. Just use the coarse types here.
2138 const Type *value_type = Type::get_const_basic_type(type);
2139 Compile::AliasType* alias_type = C->alias_type(adr_type);
2140 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2141 int alias_idx = C->get_alias_index(adr_type);
2143 // Memory-model-wise, a CAS acts like a little synchronized block,
2144 // so needs barriers on each side. These don't't translate into
2145 // actual barriers on most machines, but we still need rest of
2146 // compiler to respect ordering.
2148 insert_mem_bar(Op_MemBarRelease);
2149 insert_mem_bar(Op_MemBarCPUOrder);
2151 // 4984716: MemBars must be inserted before this
2152 // memory node in order to avoid a false
2153 // dependency which will confuse the scheduler.
2154 Node *mem = memory(alias_idx);
2156 // For now, we handle only those cases that actually exist: ints,
2157 // longs, and Object. Adding others should be straightforward.
2158 Node* cas;
2159 switch(type) {
2160 case T_INT:
2161 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2162 break;
2163 case T_LONG:
2164 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2165 break;
2166 case T_OBJECT:
2167 // reference stores need a store barrier.
2168 // (They don't if CAS fails, but it isn't worth checking.)
2169 pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT);
2170 #ifdef _LP64
2171 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2172 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2173 EncodePNode::encode(&_gvn, newval),
2174 EncodePNode::encode(&_gvn, oldval)));
2175 } else
2176 #endif
2177 {
2178 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2179 }
2180 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2181 break;
2182 default:
2183 ShouldNotReachHere();
2184 break;
2185 }
2187 // SCMemProjNodes represent the memory state of CAS. Their main
2188 // role is to prevent CAS nodes from being optimized away when their
2189 // results aren't used.
2190 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2191 set_memory(proj, alias_idx);
2193 // Add the trailing membar surrounding the access
2194 insert_mem_bar(Op_MemBarCPUOrder);
2195 insert_mem_bar(Op_MemBarAcquire);
2197 push(cas);
2198 return true;
2199 }
2201 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2202 // This is another variant of inline_unsafe_access, differing in
2203 // that it always issues store-store ("release") barrier and ensures
2204 // store-atomicity (which only matters for "long").
2206 if (callee()->is_static()) return false; // caller must have the capability!
2208 #ifndef PRODUCT
2209 {
2210 ResourceMark rm;
2211 // Check the signatures.
2212 ciSignature* sig = signature();
2213 #ifdef ASSERT
2214 BasicType rtype = sig->return_type()->basic_type();
2215 assert(rtype == T_VOID, "must return void");
2216 assert(sig->count() == 3, "has 3 arguments");
2217 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2218 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2219 #endif // ASSERT
2220 }
2221 #endif //PRODUCT
2223 // number of stack slots per value argument (1 or 2)
2224 int type_words = type2size[type];
2226 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2228 // Argument words: "this" plus oop plus offset plus value;
2229 int nargs = 1 + 1 + 2 + type_words;
2231 // pop arguments: val, offset, base, and receiver
2232 debug_only(int saved_sp = _sp);
2233 _sp += nargs;
2234 Node* val = (type_words == 1) ? pop() : pop_pair();
2235 Node *offset = pop_pair();
2236 Node *base = pop();
2237 Node *receiver = pop();
2238 assert(saved_sp == _sp, "must have correct argument count");
2240 // Null check receiver.
2241 _sp += nargs;
2242 do_null_check(receiver, T_OBJECT);
2243 _sp -= nargs;
2244 if (stopped()) {
2245 return true;
2246 }
2248 // Build field offset expression.
2249 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2250 // 32-bit machines ignore the high half of long offsets
2251 offset = ConvL2X(offset);
2252 Node* adr = make_unsafe_address(base, offset);
2253 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2254 const Type *value_type = Type::get_const_basic_type(type);
2255 Compile::AliasType* alias_type = C->alias_type(adr_type);
2257 insert_mem_bar(Op_MemBarRelease);
2258 insert_mem_bar(Op_MemBarCPUOrder);
2259 // Ensure that the store is atomic for longs:
2260 bool require_atomic_access = true;
2261 Node* store;
2262 if (type == T_OBJECT) // reference stores need a store barrier.
2263 store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type);
2264 else {
2265 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2266 }
2267 insert_mem_bar(Op_MemBarCPUOrder);
2268 return true;
2269 }
2271 bool LibraryCallKit::inline_unsafe_allocate() {
2272 if (callee()->is_static()) return false; // caller must have the capability!
2273 int nargs = 1 + 1;
2274 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2275 null_check_receiver(callee()); // check then ignore argument(0)
2276 _sp += nargs; // set original stack for use by uncommon_trap
2277 Node* cls = do_null_check(argument(1), T_OBJECT);
2278 _sp -= nargs;
2279 if (stopped()) return true;
2281 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2282 _sp += nargs; // set original stack for use by uncommon_trap
2283 kls = do_null_check(kls, T_OBJECT);
2284 _sp -= nargs;
2285 if (stopped()) return true; // argument was like int.class
2287 // Note: The argument might still be an illegal value like
2288 // Serializable.class or Object[].class. The runtime will handle it.
2289 // But we must make an explicit check for initialization.
2290 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2291 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2292 Node* bits = intcon(instanceKlass::fully_initialized);
2293 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2294 // The 'test' is non-zero if we need to take a slow path.
2296 Node* obj = new_instance(kls, test);
2297 push(obj);
2299 return true;
2300 }
2302 //------------------------inline_native_time_funcs--------------
2303 // inline code for System.currentTimeMillis() and System.nanoTime()
2304 // these have the same type and signature
2305 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2306 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2307 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2308 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2309 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2310 const TypePtr* no_memory_effects = NULL;
2311 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2312 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2313 #ifdef ASSERT
2314 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2315 assert(value_top == top(), "second value must be top");
2316 #endif
2317 push_pair(value);
2318 return true;
2319 }
2321 //------------------------inline_native_currentThread------------------
2322 bool LibraryCallKit::inline_native_currentThread() {
2323 Node* junk = NULL;
2324 push(generate_current_thread(junk));
2325 return true;
2326 }
2328 //------------------------inline_native_isInterrupted------------------
2329 bool LibraryCallKit::inline_native_isInterrupted() {
2330 const int nargs = 1+1; // receiver + boolean
2331 assert(nargs == arg_size(), "sanity");
2332 // Add a fast path to t.isInterrupted(clear_int):
2333 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2334 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2335 // So, in the common case that the interrupt bit is false,
2336 // we avoid making a call into the VM. Even if the interrupt bit
2337 // is true, if the clear_int argument is false, we avoid the VM call.
2338 // However, if the receiver is not currentThread, we must call the VM,
2339 // because there must be some locking done around the operation.
2341 // We only go to the fast case code if we pass two guards.
2342 // Paths which do not pass are accumulated in the slow_region.
2343 RegionNode* slow_region = new (C, 1) RegionNode(1);
2344 record_for_igvn(slow_region);
2345 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2346 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2347 enum { no_int_result_path = 1,
2348 no_clear_result_path = 2,
2349 slow_result_path = 3
2350 };
2352 // (a) Receiving thread must be the current thread.
2353 Node* rec_thr = argument(0);
2354 Node* tls_ptr = NULL;
2355 Node* cur_thr = generate_current_thread(tls_ptr);
2356 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2357 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2359 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2360 if (!known_current_thread)
2361 generate_slow_guard(bol_thr, slow_region);
2363 // (b) Interrupt bit on TLS must be false.
2364 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2365 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2366 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2367 Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT);
2368 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2369 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2371 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2373 // First fast path: if (!TLS._interrupted) return false;
2374 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2375 result_rgn->init_req(no_int_result_path, false_bit);
2376 result_val->init_req(no_int_result_path, intcon(0));
2378 // drop through to next case
2379 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2381 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2382 Node* clr_arg = argument(1);
2383 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2384 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2385 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2387 // Second fast path: ... else if (!clear_int) return true;
2388 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2389 result_rgn->init_req(no_clear_result_path, false_arg);
2390 result_val->init_req(no_clear_result_path, intcon(1));
2392 // drop through to next case
2393 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2395 // (d) Otherwise, go to the slow path.
2396 slow_region->add_req(control());
2397 set_control( _gvn.transform(slow_region) );
2399 if (stopped()) {
2400 // There is no slow path.
2401 result_rgn->init_req(slow_result_path, top());
2402 result_val->init_req(slow_result_path, top());
2403 } else {
2404 // non-virtual because it is a private non-static
2405 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2407 Node* slow_val = set_results_for_java_call(slow_call);
2408 // this->control() comes from set_results_for_java_call
2410 // If we know that the result of the slow call will be true, tell the optimizer!
2411 if (known_current_thread) slow_val = intcon(1);
2413 Node* fast_io = slow_call->in(TypeFunc::I_O);
2414 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2415 // These two phis are pre-filled with copies of of the fast IO and Memory
2416 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2417 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2419 result_rgn->init_req(slow_result_path, control());
2420 io_phi ->init_req(slow_result_path, i_o());
2421 mem_phi ->init_req(slow_result_path, reset_memory());
2422 result_val->init_req(slow_result_path, slow_val);
2424 set_all_memory( _gvn.transform(mem_phi) );
2425 set_i_o( _gvn.transform(io_phi) );
2426 }
2428 push_result(result_rgn, result_val);
2429 C->set_has_split_ifs(true); // Has chance for split-if optimization
2431 return true;
2432 }
2434 //---------------------------load_mirror_from_klass----------------------------
2435 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2436 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2437 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2438 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2439 }
2441 //-----------------------load_klass_from_mirror_common-------------------------
2442 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2443 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2444 // and branch to the given path on the region.
2445 // If never_see_null, take an uncommon trap on null, so we can optimistically
2446 // compile for the non-null case.
2447 // If the region is NULL, force never_see_null = true.
2448 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2449 bool never_see_null,
2450 int nargs,
2451 RegionNode* region,
2452 int null_path,
2453 int offset) {
2454 if (region == NULL) never_see_null = true;
2455 Node* p = basic_plus_adr(mirror, offset);
2456 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2457 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2458 _sp += nargs; // any deopt will start just before call to enclosing method
2459 Node* null_ctl = top();
2460 kls = null_check_oop(kls, &null_ctl, never_see_null);
2461 if (region != NULL) {
2462 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2463 region->init_req(null_path, null_ctl);
2464 } else {
2465 assert(null_ctl == top(), "no loose ends");
2466 }
2467 _sp -= nargs;
2468 return kls;
2469 }
2471 //--------------------(inline_native_Class_query helpers)---------------------
2472 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2473 // Fall through if (mods & mask) == bits, take the guard otherwise.
2474 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2475 // Branch around if the given klass has the given modifier bit set.
2476 // Like generate_guard, adds a new path onto the region.
2477 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2478 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2479 Node* mask = intcon(modifier_mask);
2480 Node* bits = intcon(modifier_bits);
2481 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2482 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2483 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2484 return generate_fair_guard(bol, region);
2485 }
2486 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2487 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2488 }
2490 //-------------------------inline_native_Class_query-------------------
2491 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2492 int nargs = 1+0; // just the Class mirror, in most cases
2493 const Type* return_type = TypeInt::BOOL;
2494 Node* prim_return_value = top(); // what happens if it's a primitive class?
2495 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2496 bool expect_prim = false; // most of these guys expect to work on refs
2498 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2500 switch (id) {
2501 case vmIntrinsics::_isInstance:
2502 nargs = 1+1; // the Class mirror, plus the object getting queried about
2503 // nothing is an instance of a primitive type
2504 prim_return_value = intcon(0);
2505 break;
2506 case vmIntrinsics::_getModifiers:
2507 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2508 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2509 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2510 break;
2511 case vmIntrinsics::_isInterface:
2512 prim_return_value = intcon(0);
2513 break;
2514 case vmIntrinsics::_isArray:
2515 prim_return_value = intcon(0);
2516 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2517 break;
2518 case vmIntrinsics::_isPrimitive:
2519 prim_return_value = intcon(1);
2520 expect_prim = true; // obviously
2521 break;
2522 case vmIntrinsics::_getSuperclass:
2523 prim_return_value = null();
2524 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2525 break;
2526 case vmIntrinsics::_getComponentType:
2527 prim_return_value = null();
2528 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2529 break;
2530 case vmIntrinsics::_getClassAccessFlags:
2531 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2532 return_type = TypeInt::INT; // not bool! 6297094
2533 break;
2534 default:
2535 ShouldNotReachHere();
2536 }
2538 Node* mirror = argument(0);
2539 Node* obj = (nargs <= 1)? top(): argument(1);
2541 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2542 if (mirror_con == NULL) return false; // cannot happen?
2544 #ifndef PRODUCT
2545 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2546 ciType* k = mirror_con->java_mirror_type();
2547 if (k) {
2548 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2549 k->print_name();
2550 tty->cr();
2551 }
2552 }
2553 #endif
2555 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2556 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2557 record_for_igvn(region);
2558 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2560 // The mirror will never be null of Reflection.getClassAccessFlags, however
2561 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2562 // if it is. See bug 4774291.
2564 // For Reflection.getClassAccessFlags(), the null check occurs in
2565 // the wrong place; see inline_unsafe_access(), above, for a similar
2566 // situation.
2567 _sp += nargs; // set original stack for use by uncommon_trap
2568 mirror = do_null_check(mirror, T_OBJECT);
2569 _sp -= nargs;
2570 // If mirror or obj is dead, only null-path is taken.
2571 if (stopped()) return true;
2573 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2575 // Now load the mirror's klass metaobject, and null-check it.
2576 // Side-effects region with the control path if the klass is null.
2577 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2578 region, _prim_path);
2579 // If kls is null, we have a primitive mirror.
2580 phi->init_req(_prim_path, prim_return_value);
2581 if (stopped()) { push_result(region, phi); return true; }
2583 Node* p; // handy temp
2584 Node* null_ctl;
2586 // Now that we have the non-null klass, we can perform the real query.
2587 // For constant classes, the query will constant-fold in LoadNode::Value.
2588 Node* query_value = top();
2589 switch (id) {
2590 case vmIntrinsics::_isInstance:
2591 // nothing is an instance of a primitive type
2592 query_value = gen_instanceof(obj, kls);
2593 break;
2595 case vmIntrinsics::_getModifiers:
2596 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2597 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2598 break;
2600 case vmIntrinsics::_isInterface:
2601 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2602 if (generate_interface_guard(kls, region) != NULL)
2603 // A guard was added. If the guard is taken, it was an interface.
2604 phi->add_req(intcon(1));
2605 // If we fall through, it's a plain class.
2606 query_value = intcon(0);
2607 break;
2609 case vmIntrinsics::_isArray:
2610 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2611 if (generate_array_guard(kls, region) != NULL)
2612 // A guard was added. If the guard is taken, it was an array.
2613 phi->add_req(intcon(1));
2614 // If we fall through, it's a plain class.
2615 query_value = intcon(0);
2616 break;
2618 case vmIntrinsics::_isPrimitive:
2619 query_value = intcon(0); // "normal" path produces false
2620 break;
2622 case vmIntrinsics::_getSuperclass:
2623 // The rules here are somewhat unfortunate, but we can still do better
2624 // with random logic than with a JNI call.
2625 // Interfaces store null or Object as _super, but must report null.
2626 // Arrays store an intermediate super as _super, but must report Object.
2627 // Other types can report the actual _super.
2628 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2629 if (generate_interface_guard(kls, region) != NULL)
2630 // A guard was added. If the guard is taken, it was an interface.
2631 phi->add_req(null());
2632 if (generate_array_guard(kls, region) != NULL)
2633 // A guard was added. If the guard is taken, it was an array.
2634 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2635 // If we fall through, it's a plain class. Get its _super.
2636 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2637 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2638 null_ctl = top();
2639 kls = null_check_oop(kls, &null_ctl);
2640 if (null_ctl != top()) {
2641 // If the guard is taken, Object.superClass is null (both klass and mirror).
2642 region->add_req(null_ctl);
2643 phi ->add_req(null());
2644 }
2645 if (!stopped()) {
2646 query_value = load_mirror_from_klass(kls);
2647 }
2648 break;
2650 case vmIntrinsics::_getComponentType:
2651 if (generate_array_guard(kls, region) != NULL) {
2652 // Be sure to pin the oop load to the guard edge just created:
2653 Node* is_array_ctrl = region->in(region->req()-1);
2654 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
2655 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
2656 phi->add_req(cmo);
2657 }
2658 query_value = null(); // non-array case is null
2659 break;
2661 case vmIntrinsics::_getClassAccessFlags:
2662 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2663 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2664 break;
2666 default:
2667 ShouldNotReachHere();
2668 }
2670 // Fall-through is the normal case of a query to a real class.
2671 phi->init_req(1, query_value);
2672 region->init_req(1, control());
2674 push_result(region, phi);
2675 C->set_has_split_ifs(true); // Has chance for split-if optimization
2677 return true;
2678 }
2680 //--------------------------inline_native_subtype_check------------------------
2681 // This intrinsic takes the JNI calls out of the heart of
2682 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
2683 bool LibraryCallKit::inline_native_subtype_check() {
2684 int nargs = 1+1; // the Class mirror, plus the other class getting examined
2686 // Pull both arguments off the stack.
2687 Node* args[2]; // two java.lang.Class mirrors: superc, subc
2688 args[0] = argument(0);
2689 args[1] = argument(1);
2690 Node* klasses[2]; // corresponding Klasses: superk, subk
2691 klasses[0] = klasses[1] = top();
2693 enum {
2694 // A full decision tree on {superc is prim, subc is prim}:
2695 _prim_0_path = 1, // {P,N} => false
2696 // {P,P} & superc!=subc => false
2697 _prim_same_path, // {P,P} & superc==subc => true
2698 _prim_1_path, // {N,P} => false
2699 _ref_subtype_path, // {N,N} & subtype check wins => true
2700 _both_ref_path, // {N,N} & subtype check loses => false
2701 PATH_LIMIT
2702 };
2704 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2705 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2706 record_for_igvn(region);
2708 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
2709 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2710 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
2712 // First null-check both mirrors and load each mirror's klass metaobject.
2713 int which_arg;
2714 for (which_arg = 0; which_arg <= 1; which_arg++) {
2715 Node* arg = args[which_arg];
2716 _sp += nargs; // set original stack for use by uncommon_trap
2717 arg = do_null_check(arg, T_OBJECT);
2718 _sp -= nargs;
2719 if (stopped()) break;
2720 args[which_arg] = _gvn.transform(arg);
2722 Node* p = basic_plus_adr(arg, class_klass_offset);
2723 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
2724 klasses[which_arg] = _gvn.transform(kls);
2725 }
2727 // Having loaded both klasses, test each for null.
2728 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2729 for (which_arg = 0; which_arg <= 1; which_arg++) {
2730 Node* kls = klasses[which_arg];
2731 Node* null_ctl = top();
2732 _sp += nargs; // set original stack for use by uncommon_trap
2733 kls = null_check_oop(kls, &null_ctl, never_see_null);
2734 _sp -= nargs;
2735 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
2736 region->init_req(prim_path, null_ctl);
2737 if (stopped()) break;
2738 klasses[which_arg] = kls;
2739 }
2741 if (!stopped()) {
2742 // now we have two reference types, in klasses[0..1]
2743 Node* subk = klasses[1]; // the argument to isAssignableFrom
2744 Node* superk = klasses[0]; // the receiver
2745 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
2746 // now we have a successful reference subtype check
2747 region->set_req(_ref_subtype_path, control());
2748 }
2750 // If both operands are primitive (both klasses null), then
2751 // we must return true when they are identical primitives.
2752 // It is convenient to test this after the first null klass check.
2753 set_control(region->in(_prim_0_path)); // go back to first null check
2754 if (!stopped()) {
2755 // Since superc is primitive, make a guard for the superc==subc case.
2756 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
2757 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
2758 generate_guard(bol_eq, region, PROB_FAIR);
2759 if (region->req() == PATH_LIMIT+1) {
2760 // A guard was added. If the added guard is taken, superc==subc.
2761 region->swap_edges(PATH_LIMIT, _prim_same_path);
2762 region->del_req(PATH_LIMIT);
2763 }
2764 region->set_req(_prim_0_path, control()); // Not equal after all.
2765 }
2767 // these are the only paths that produce 'true':
2768 phi->set_req(_prim_same_path, intcon(1));
2769 phi->set_req(_ref_subtype_path, intcon(1));
2771 // pull together the cases:
2772 assert(region->req() == PATH_LIMIT, "sane region");
2773 for (uint i = 1; i < region->req(); i++) {
2774 Node* ctl = region->in(i);
2775 if (ctl == NULL || ctl == top()) {
2776 region->set_req(i, top());
2777 phi ->set_req(i, top());
2778 } else if (phi->in(i) == NULL) {
2779 phi->set_req(i, intcon(0)); // all other paths produce 'false'
2780 }
2781 }
2783 set_control(_gvn.transform(region));
2784 push(_gvn.transform(phi));
2786 return true;
2787 }
2789 //---------------------generate_array_guard_common------------------------
2790 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
2791 bool obj_array, bool not_array) {
2792 // If obj_array/non_array==false/false:
2793 // Branch around if the given klass is in fact an array (either obj or prim).
2794 // If obj_array/non_array==false/true:
2795 // Branch around if the given klass is not an array klass of any kind.
2796 // If obj_array/non_array==true/true:
2797 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
2798 // If obj_array/non_array==true/false:
2799 // Branch around if the kls is an oop array (Object[] or subtype)
2800 //
2801 // Like generate_guard, adds a new path onto the region.
2802 jint layout_con = 0;
2803 Node* layout_val = get_layout_helper(kls, layout_con);
2804 if (layout_val == NULL) {
2805 bool query = (obj_array
2806 ? Klass::layout_helper_is_objArray(layout_con)
2807 : Klass::layout_helper_is_javaArray(layout_con));
2808 if (query == not_array) {
2809 return NULL; // never a branch
2810 } else { // always a branch
2811 Node* always_branch = control();
2812 if (region != NULL)
2813 region->add_req(always_branch);
2814 set_control(top());
2815 return always_branch;
2816 }
2817 }
2818 // Now test the correct condition.
2819 jint nval = (obj_array
2820 ? ((jint)Klass::_lh_array_tag_type_value
2821 << Klass::_lh_array_tag_shift)
2822 : Klass::_lh_neutral_value);
2823 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
2824 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
2825 // invert the test if we are looking for a non-array
2826 if (not_array) btest = BoolTest(btest).negate();
2827 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
2828 return generate_fair_guard(bol, region);
2829 }
2832 //-----------------------inline_native_newArray--------------------------
2833 bool LibraryCallKit::inline_native_newArray() {
2834 int nargs = 2;
2835 Node* mirror = argument(0);
2836 Node* count_val = argument(1);
2838 _sp += nargs; // set original stack for use by uncommon_trap
2839 mirror = do_null_check(mirror, T_OBJECT);
2840 _sp -= nargs;
2841 // If mirror or obj is dead, only null-path is taken.
2842 if (stopped()) return true;
2844 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
2845 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2846 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
2847 TypeInstPtr::NOTNULL);
2848 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
2849 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
2850 TypePtr::BOTTOM);
2852 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2853 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
2854 nargs,
2855 result_reg, _slow_path);
2856 Node* normal_ctl = control();
2857 Node* no_array_ctl = result_reg->in(_slow_path);
2859 // Generate code for the slow case. We make a call to newArray().
2860 set_control(no_array_ctl);
2861 if (!stopped()) {
2862 // Either the input type is void.class, or else the
2863 // array klass has not yet been cached. Either the
2864 // ensuing call will throw an exception, or else it
2865 // will cache the array klass for next time.
2866 PreserveJVMState pjvms(this);
2867 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
2868 Node* slow_result = set_results_for_java_call(slow_call);
2869 // this->control() comes from set_results_for_java_call
2870 result_reg->set_req(_slow_path, control());
2871 result_val->set_req(_slow_path, slow_result);
2872 result_io ->set_req(_slow_path, i_o());
2873 result_mem->set_req(_slow_path, reset_memory());
2874 }
2876 set_control(normal_ctl);
2877 if (!stopped()) {
2878 // Normal case: The array type has been cached in the java.lang.Class.
2879 // The following call works fine even if the array type is polymorphic.
2880 // It could be a dynamic mix of int[], boolean[], Object[], etc.
2881 _sp += nargs; // set original stack for use by uncommon_trap
2882 Node* obj = new_array(klass_node, count_val);
2883 _sp -= nargs;
2884 result_reg->init_req(_normal_path, control());
2885 result_val->init_req(_normal_path, obj);
2886 result_io ->init_req(_normal_path, i_o());
2887 result_mem->init_req(_normal_path, reset_memory());
2888 }
2890 // Return the combined state.
2891 set_i_o( _gvn.transform(result_io) );
2892 set_all_memory( _gvn.transform(result_mem) );
2893 push_result(result_reg, result_val);
2894 C->set_has_split_ifs(true); // Has chance for split-if optimization
2896 return true;
2897 }
2899 //----------------------inline_native_getLength--------------------------
2900 bool LibraryCallKit::inline_native_getLength() {
2901 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
2903 int nargs = 1;
2904 Node* array = argument(0);
2906 _sp += nargs; // set original stack for use by uncommon_trap
2907 array = do_null_check(array, T_OBJECT);
2908 _sp -= nargs;
2910 // If array is dead, only null-path is taken.
2911 if (stopped()) return true;
2913 // Deoptimize if it is a non-array.
2914 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
2916 if (non_array != NULL) {
2917 PreserveJVMState pjvms(this);
2918 set_control(non_array);
2919 _sp += nargs; // push the arguments back on the stack
2920 uncommon_trap(Deoptimization::Reason_intrinsic,
2921 Deoptimization::Action_maybe_recompile);
2922 }
2924 // If control is dead, only non-array-path is taken.
2925 if (stopped()) return true;
2927 // The works fine even if the array type is polymorphic.
2928 // It could be a dynamic mix of int[], boolean[], Object[], etc.
2929 push( load_array_length(array) );
2931 C->set_has_split_ifs(true); // Has chance for split-if optimization
2933 return true;
2934 }
2936 //------------------------inline_array_copyOf----------------------------
2937 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
2938 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
2940 // Restore the stack and pop off the arguments.
2941 int nargs = 3 + (is_copyOfRange? 1: 0);
2942 Node* original = argument(0);
2943 Node* start = is_copyOfRange? argument(1): intcon(0);
2944 Node* end = is_copyOfRange? argument(2): argument(1);
2945 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
2947 _sp += nargs; // set original stack for use by uncommon_trap
2948 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
2949 original = do_null_check(original, T_OBJECT);
2950 _sp -= nargs;
2952 // Check if a null path was taken unconditionally.
2953 if (stopped()) return true;
2955 Node* orig_length = load_array_length(original);
2957 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs,
2958 NULL, 0);
2959 _sp += nargs; // set original stack for use by uncommon_trap
2960 klass_node = do_null_check(klass_node, T_OBJECT);
2961 _sp -= nargs;
2963 RegionNode* bailout = new (C, 1) RegionNode(1);
2964 record_for_igvn(bailout);
2966 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
2967 // Bail out if that is so.
2968 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
2969 if (not_objArray != NULL) {
2970 // Improve the klass node's type from the new optimistic assumption:
2971 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
2972 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2973 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
2974 cast->init_req(0, control());
2975 klass_node = _gvn.transform(cast);
2976 }
2978 // Bail out if either start or end is negative.
2979 generate_negative_guard(start, bailout, &start);
2980 generate_negative_guard(end, bailout, &end);
2982 Node* length = end;
2983 if (_gvn.type(start) != TypeInt::ZERO) {
2984 length = _gvn.transform( new (C, 3) SubINode(end, start) );
2985 }
2987 // Bail out if length is negative.
2988 // ...Not needed, since the new_array will throw the right exception.
2989 //generate_negative_guard(length, bailout, &length);
2991 if (bailout->req() > 1) {
2992 PreserveJVMState pjvms(this);
2993 set_control( _gvn.transform(bailout) );
2994 _sp += nargs; // push the arguments back on the stack
2995 uncommon_trap(Deoptimization::Reason_intrinsic,
2996 Deoptimization::Action_maybe_recompile);
2997 }
2999 if (!stopped()) {
3000 // How many elements will we copy from the original?
3001 // The answer is MinI(orig_length - start, length).
3002 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3003 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3005 _sp += nargs; // set original stack for use by uncommon_trap
3006 Node* newcopy = new_array(klass_node, length);
3007 _sp -= nargs;
3009 // Generate a direct call to the right arraycopy function(s).
3010 // We know the copy is disjoint but we might not know if the
3011 // oop stores need checking.
3012 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3013 // This will fail a store-check if x contains any non-nulls.
3014 bool disjoint_bases = true;
3015 bool length_never_negative = true;
3016 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3017 original, start, newcopy, intcon(0), moved,
3018 nargs, disjoint_bases, length_never_negative);
3020 push(newcopy);
3021 }
3023 C->set_has_split_ifs(true); // Has chance for split-if optimization
3025 return true;
3026 }
3029 //----------------------generate_virtual_guard---------------------------
3030 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3031 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3032 RegionNode* slow_region) {
3033 ciMethod* method = callee();
3034 int vtable_index = method->vtable_index();
3035 // Get the methodOop out of the appropriate vtable entry.
3036 int entry_offset = (instanceKlass::vtable_start_offset() +
3037 vtable_index*vtableEntry::size()) * wordSize +
3038 vtableEntry::method_offset_in_bytes();
3039 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3040 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3042 // Compare the target method with the expected method (e.g., Object.hashCode).
3043 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3045 Node* native_call = makecon(native_call_addr);
3046 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3047 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3049 return generate_slow_guard(test_native, slow_region);
3050 }
3052 //-----------------------generate_method_call----------------------------
3053 // Use generate_method_call to make a slow-call to the real
3054 // method if the fast path fails. An alternative would be to
3055 // use a stub like OptoRuntime::slow_arraycopy_Java.
3056 // This only works for expanding the current library call,
3057 // not another intrinsic. (E.g., don't use this for making an
3058 // arraycopy call inside of the copyOf intrinsic.)
3059 CallJavaNode*
3060 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3061 // When compiling the intrinsic method itself, do not use this technique.
3062 guarantee(callee() != C->method(), "cannot make slow-call to self");
3064 ciMethod* method = callee();
3065 // ensure the JVMS we have will be correct for this call
3066 guarantee(method_id == method->intrinsic_id(), "must match");
3068 const TypeFunc* tf = TypeFunc::make(method);
3069 int tfdc = tf->domain()->cnt();
3070 CallJavaNode* slow_call;
3071 if (is_static) {
3072 assert(!is_virtual, "");
3073 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3074 SharedRuntime::get_resolve_static_call_stub(),
3075 method, bci());
3076 } else if (is_virtual) {
3077 null_check_receiver(method);
3078 int vtable_index = methodOopDesc::invalid_vtable_index;
3079 if (UseInlineCaches) {
3080 // Suppress the vtable call
3081 } else {
3082 // hashCode and clone are not a miranda methods,
3083 // so the vtable index is fixed.
3084 // No need to use the linkResolver to get it.
3085 vtable_index = method->vtable_index();
3086 }
3087 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3088 SharedRuntime::get_resolve_virtual_call_stub(),
3089 method, vtable_index, bci());
3090 } else { // neither virtual nor static: opt_virtual
3091 null_check_receiver(method);
3092 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3093 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3094 method, bci());
3095 slow_call->set_optimized_virtual(true);
3096 }
3097 set_arguments_for_java_call(slow_call);
3098 set_edges_for_java_call(slow_call);
3099 return slow_call;
3100 }
3103 //------------------------------inline_native_hashcode--------------------
3104 // Build special case code for calls to hashCode on an object.
3105 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3106 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3107 assert(!(is_virtual && is_static), "either virtual, special, or static");
3109 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3111 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3112 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3113 TypeInt::INT);
3114 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3115 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3116 TypePtr::BOTTOM);
3117 Node* obj = NULL;
3118 if (!is_static) {
3119 // Check for hashing null object
3120 obj = null_check_receiver(callee());
3121 if (stopped()) return true; // unconditionally null
3122 result_reg->init_req(_null_path, top());
3123 result_val->init_req(_null_path, top());
3124 } else {
3125 // Do a null check, and return zero if null.
3126 // System.identityHashCode(null) == 0
3127 obj = argument(0);
3128 Node* null_ctl = top();
3129 obj = null_check_oop(obj, &null_ctl);
3130 result_reg->init_req(_null_path, null_ctl);
3131 result_val->init_req(_null_path, _gvn.intcon(0));
3132 }
3134 // Unconditionally null? Then return right away.
3135 if (stopped()) {
3136 set_control( result_reg->in(_null_path) );
3137 if (!stopped())
3138 push( result_val ->in(_null_path) );
3139 return true;
3140 }
3142 // After null check, get the object's klass.
3143 Node* obj_klass = load_object_klass(obj);
3145 // This call may be virtual (invokevirtual) or bound (invokespecial).
3146 // For each case we generate slightly different code.
3148 // We only go to the fast case code if we pass a number of guards. The
3149 // paths which do not pass are accumulated in the slow_region.
3150 RegionNode* slow_region = new (C, 1) RegionNode(1);
3151 record_for_igvn(slow_region);
3153 // If this is a virtual call, we generate a funny guard. We pull out
3154 // the vtable entry corresponding to hashCode() from the target object.
3155 // If the target method which we are calling happens to be the native
3156 // Object hashCode() method, we pass the guard. We do not need this
3157 // guard for non-virtual calls -- the caller is known to be the native
3158 // Object hashCode().
3159 if (is_virtual) {
3160 generate_virtual_guard(obj_klass, slow_region);
3161 }
3163 // Get the header out of the object, use LoadMarkNode when available
3164 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3165 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
3166 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
3168 // Test the header to see if it is unlocked.
3169 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3170 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3171 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3172 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3173 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3175 generate_slow_guard(test_unlocked, slow_region);
3177 // Get the hash value and check to see that it has been properly assigned.
3178 // We depend on hash_mask being at most 32 bits and avoid the use of
3179 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3180 // vm: see markOop.hpp.
3181 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3182 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3183 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3184 // This hack lets the hash bits live anywhere in the mark object now, as long
3185 // as the shift drops the relevent bits into the low 32 bits. Note that
3186 // Java spec says that HashCode is an int so there's no point in capturing
3187 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3188 hshifted_header = ConvX2I(hshifted_header);
3189 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3191 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3192 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3193 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3195 generate_slow_guard(test_assigned, slow_region);
3197 Node* init_mem = reset_memory();
3198 // fill in the rest of the null path:
3199 result_io ->init_req(_null_path, i_o());
3200 result_mem->init_req(_null_path, init_mem);
3202 result_val->init_req(_fast_path, hash_val);
3203 result_reg->init_req(_fast_path, control());
3204 result_io ->init_req(_fast_path, i_o());
3205 result_mem->init_req(_fast_path, init_mem);
3207 // Generate code for the slow case. We make a call to hashCode().
3208 set_control(_gvn.transform(slow_region));
3209 if (!stopped()) {
3210 // No need for PreserveJVMState, because we're using up the present state.
3211 set_all_memory(init_mem);
3212 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3213 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3214 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3215 Node* slow_result = set_results_for_java_call(slow_call);
3216 // this->control() comes from set_results_for_java_call
3217 result_reg->init_req(_slow_path, control());
3218 result_val->init_req(_slow_path, slow_result);
3219 result_io ->set_req(_slow_path, i_o());
3220 result_mem ->set_req(_slow_path, reset_memory());
3221 }
3223 // Return the combined state.
3224 set_i_o( _gvn.transform(result_io) );
3225 set_all_memory( _gvn.transform(result_mem) );
3226 push_result(result_reg, result_val);
3228 return true;
3229 }
3231 //---------------------------inline_native_getClass----------------------------
3232 // Build special case code for calls to hashCode on an object.
3233 bool LibraryCallKit::inline_native_getClass() {
3234 Node* obj = null_check_receiver(callee());
3235 if (stopped()) return true;
3236 push( load_mirror_from_klass(load_object_klass(obj)) );
3237 return true;
3238 }
3240 //-----------------inline_native_Reflection_getCallerClass---------------------
3241 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3242 //
3243 // NOTE that this code must perform the same logic as
3244 // vframeStream::security_get_caller_frame in that it must skip
3245 // Method.invoke() and auxiliary frames.
3250 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3251 ciMethod* method = callee();
3253 #ifndef PRODUCT
3254 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3255 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3256 }
3257 #endif
3259 debug_only(int saved_sp = _sp);
3261 // Argument words: (int depth)
3262 int nargs = 1;
3264 _sp += nargs;
3265 Node* caller_depth_node = pop();
3267 assert(saved_sp == _sp, "must have correct argument count");
3269 // The depth value must be a constant in order for the runtime call
3270 // to be eliminated.
3271 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3272 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3273 #ifndef PRODUCT
3274 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3275 tty->print_cr(" Bailing out because caller depth was not a constant");
3276 }
3277 #endif
3278 return false;
3279 }
3280 // Note that the JVM state at this point does not include the
3281 // getCallerClass() frame which we are trying to inline. The
3282 // semantics of getCallerClass(), however, are that the "first"
3283 // frame is the getCallerClass() frame, so we subtract one from the
3284 // requested depth before continuing. We don't inline requests of
3285 // getCallerClass(0).
3286 int caller_depth = caller_depth_type->get_con() - 1;
3287 if (caller_depth < 0) {
3288 #ifndef PRODUCT
3289 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3290 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3291 }
3292 #endif
3293 return false;
3294 }
3296 if (!jvms()->has_method()) {
3297 #ifndef PRODUCT
3298 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3299 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3300 }
3301 #endif
3302 return false;
3303 }
3304 int _depth = jvms()->depth(); // cache call chain depth
3306 // Walk back up the JVM state to find the caller at the required
3307 // depth. NOTE that this code must perform the same logic as
3308 // vframeStream::security_get_caller_frame in that it must skip
3309 // Method.invoke() and auxiliary frames. Note also that depth is
3310 // 1-based (1 is the bottom of the inlining).
3311 int inlining_depth = _depth;
3312 JVMState* caller_jvms = NULL;
3314 if (inlining_depth > 0) {
3315 caller_jvms = jvms();
3316 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3317 do {
3318 // The following if-tests should be performed in this order
3319 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3320 // Skip a Method.invoke() or auxiliary frame
3321 } else if (caller_depth > 0) {
3322 // Skip real frame
3323 --caller_depth;
3324 } else {
3325 // We're done: reached desired caller after skipping.
3326 break;
3327 }
3328 caller_jvms = caller_jvms->caller();
3329 --inlining_depth;
3330 } while (inlining_depth > 0);
3331 }
3333 if (inlining_depth == 0) {
3334 #ifndef PRODUCT
3335 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3336 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3337 tty->print_cr(" JVM state at this point:");
3338 for (int i = _depth; i >= 1; i--) {
3339 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3340 }
3341 }
3342 #endif
3343 return false; // Reached end of inlining
3344 }
3346 // Acquire method holder as java.lang.Class
3347 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3348 ciInstance* caller_mirror = caller_klass->java_mirror();
3349 // Push this as a constant
3350 push(makecon(TypeInstPtr::make(caller_mirror)));
3351 #ifndef PRODUCT
3352 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3353 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);
3354 tty->print_cr(" JVM state at this point:");
3355 for (int i = _depth; i >= 1; i--) {
3356 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3357 }
3358 }
3359 #endif
3360 return true;
3361 }
3363 // Helper routine for above
3364 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3365 // Is this the Method.invoke method itself?
3366 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
3367 return true;
3369 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3370 ciKlass* k = jvms->method()->holder();
3371 if (k->is_instance_klass()) {
3372 ciInstanceKlass* ik = k->as_instance_klass();
3373 for (; ik != NULL; ik = ik->super()) {
3374 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3375 ik == env()->find_system_klass(ik->name())) {
3376 return true;
3377 }
3378 }
3379 }
3381 return false;
3382 }
3384 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3385 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3386 // computing it since there is no lookup field by name function in the
3387 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3388 // Using a static variable here is safe even if we have multiple compilation
3389 // threads because the offset is constant. At worst the same offset will be
3390 // computed and stored multiple
3392 bool LibraryCallKit::inline_native_AtomicLong_get() {
3393 // Restore the stack and pop off the argument
3394 _sp+=1;
3395 Node *obj = pop();
3397 // get the offset of the "value" field. Since the CI interfaces
3398 // does not provide a way to look up a field by name, we scan the bytecodes
3399 // to get the field index. We expect the first 2 instructions of the method
3400 // to be:
3401 // 0 aload_0
3402 // 1 getfield "value"
3403 ciMethod* method = callee();
3404 if (value_field_offset == -1)
3405 {
3406 ciField* value_field;
3407 ciBytecodeStream iter(method);
3408 Bytecodes::Code bc = iter.next();
3410 if ((bc != Bytecodes::_aload_0) &&
3411 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3412 return false;
3413 bc = iter.next();
3414 if (bc != Bytecodes::_getfield)
3415 return false;
3416 bool ignore;
3417 value_field = iter.get_field(ignore);
3418 value_field_offset = value_field->offset_in_bytes();
3419 }
3421 // Null check without removing any arguments.
3422 _sp++;
3423 obj = do_null_check(obj, T_OBJECT);
3424 _sp--;
3425 // Check for locking null object
3426 if (stopped()) return true;
3428 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3429 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3430 int alias_idx = C->get_alias_index(adr_type);
3432 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3434 push_pair(result);
3436 return true;
3437 }
3439 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3440 // Restore the stack and pop off the arguments
3441 _sp+=5;
3442 Node *newVal = pop_pair();
3443 Node *oldVal = pop_pair();
3444 Node *obj = pop();
3446 // we need the offset of the "value" field which was computed when
3447 // inlining the get() method. Give up if we don't have it.
3448 if (value_field_offset == -1)
3449 return false;
3451 // Null check without removing any arguments.
3452 _sp+=5;
3453 obj = do_null_check(obj, T_OBJECT);
3454 _sp-=5;
3455 // Check for locking null object
3456 if (stopped()) return true;
3458 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3459 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3460 int alias_idx = C->get_alias_index(adr_type);
3462 Node *result = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3463 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(result));
3464 set_memory(store_proj, alias_idx);
3466 push(result);
3467 return true;
3468 }
3470 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3471 // restore the arguments
3472 _sp += arg_size();
3474 switch (id) {
3475 case vmIntrinsics::_floatToRawIntBits:
3476 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3477 break;
3479 case vmIntrinsics::_intBitsToFloat:
3480 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3481 break;
3483 case vmIntrinsics::_doubleToRawLongBits:
3484 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3485 break;
3487 case vmIntrinsics::_longBitsToDouble:
3488 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3489 break;
3491 case vmIntrinsics::_doubleToLongBits: {
3492 Node* value = pop_pair();
3494 // two paths (plus control) merge in a wood
3495 RegionNode *r = new (C, 3) RegionNode(3);
3496 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3498 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3499 // Build the boolean node
3500 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3502 // Branch either way.
3503 // NaN case is less traveled, which makes all the difference.
3504 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3505 Node *opt_isnan = _gvn.transform(ifisnan);
3506 assert( opt_isnan->is_If(), "Expect an IfNode");
3507 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3508 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3510 set_control(iftrue);
3512 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3513 Node *slow_result = longcon(nan_bits); // return NaN
3514 phi->init_req(1, _gvn.transform( slow_result ));
3515 r->init_req(1, iftrue);
3517 // Else fall through
3518 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3519 set_control(iffalse);
3521 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3522 r->init_req(2, iffalse);
3524 // Post merge
3525 set_control(_gvn.transform(r));
3526 record_for_igvn(r);
3528 Node* result = _gvn.transform(phi);
3529 assert(result->bottom_type()->isa_long(), "must be");
3530 push_pair(result);
3532 C->set_has_split_ifs(true); // Has chance for split-if optimization
3534 break;
3535 }
3537 case vmIntrinsics::_floatToIntBits: {
3538 Node* value = pop();
3540 // two paths (plus control) merge in a wood
3541 RegionNode *r = new (C, 3) RegionNode(3);
3542 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3544 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3545 // Build the boolean node
3546 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3548 // Branch either way.
3549 // NaN case is less traveled, which makes all the difference.
3550 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3551 Node *opt_isnan = _gvn.transform(ifisnan);
3552 assert( opt_isnan->is_If(), "Expect an IfNode");
3553 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3554 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3556 set_control(iftrue);
3558 static const jint nan_bits = 0x7fc00000;
3559 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3560 phi->init_req(1, _gvn.transform( slow_result ));
3561 r->init_req(1, iftrue);
3563 // Else fall through
3564 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3565 set_control(iffalse);
3567 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3568 r->init_req(2, iffalse);
3570 // Post merge
3571 set_control(_gvn.transform(r));
3572 record_for_igvn(r);
3574 Node* result = _gvn.transform(phi);
3575 assert(result->bottom_type()->isa_int(), "must be");
3576 push(result);
3578 C->set_has_split_ifs(true); // Has chance for split-if optimization
3580 break;
3581 }
3583 default:
3584 ShouldNotReachHere();
3585 }
3587 return true;
3588 }
3590 #ifdef _LP64
3591 #define XTOP ,top() /*additional argument*/
3592 #else //_LP64
3593 #define XTOP /*no additional argument*/
3594 #endif //_LP64
3596 //----------------------inline_unsafe_copyMemory-------------------------
3597 bool LibraryCallKit::inline_unsafe_copyMemory() {
3598 if (callee()->is_static()) return false; // caller must have the capability!
3599 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3600 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3601 null_check_receiver(callee()); // check then ignore argument(0)
3602 if (stopped()) return true;
3604 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3606 Node* src_ptr = argument(1);
3607 Node* src_off = ConvL2X(argument(2));
3608 assert(argument(3)->is_top(), "2nd half of long");
3609 Node* dst_ptr = argument(4);
3610 Node* dst_off = ConvL2X(argument(5));
3611 assert(argument(6)->is_top(), "2nd half of long");
3612 Node* size = ConvL2X(argument(7));
3613 assert(argument(8)->is_top(), "2nd half of long");
3615 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3616 "fieldOffset must be byte-scaled");
3618 Node* src = make_unsafe_address(src_ptr, src_off);
3619 Node* dst = make_unsafe_address(dst_ptr, dst_off);
3621 // Conservatively insert a memory barrier on all memory slices.
3622 // Do not let writes of the copy source or destination float below the copy.
3623 insert_mem_bar(Op_MemBarCPUOrder);
3625 // Call it. Note that the length argument is not scaled.
3626 make_runtime_call(RC_LEAF|RC_NO_FP,
3627 OptoRuntime::fast_arraycopy_Type(),
3628 StubRoutines::unsafe_arraycopy(),
3629 "unsafe_arraycopy",
3630 TypeRawPtr::BOTTOM,
3631 src, dst, size XTOP);
3633 // Do not let reads of the copy destination float above the copy.
3634 insert_mem_bar(Op_MemBarCPUOrder);
3636 return true;
3637 }
3640 //------------------------inline_native_clone----------------------------
3641 // Here are the simple edge cases:
3642 // null receiver => normal trap
3643 // virtual and clone was overridden => slow path to out-of-line clone
3644 // not cloneable or finalizer => slow path to out-of-line Object.clone
3645 //
3646 // The general case has two steps, allocation and copying.
3647 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
3648 //
3649 // Copying also has two cases, oop arrays and everything else.
3650 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
3651 // Everything else uses the tight inline loop supplied by CopyArrayNode.
3652 //
3653 // These steps fold up nicely if and when the cloned object's klass
3654 // can be sharply typed as an object array, a type array, or an instance.
3655 //
3656 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
3657 int nargs = 1;
3658 Node* obj = null_check_receiver(callee());
3659 if (stopped()) return true;
3660 Node* obj_klass = load_object_klass(obj);
3661 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
3662 const TypeOopPtr* toop = ((tklass != NULL)
3663 ? tklass->as_instance_type()
3664 : TypeInstPtr::NOTNULL);
3666 // Conservatively insert a memory barrier on all memory slices.
3667 // Do not let writes into the original float below the clone.
3668 insert_mem_bar(Op_MemBarCPUOrder);
3670 // paths into result_reg:
3671 enum {
3672 _slow_path = 1, // out-of-line call to clone method (virtual or not)
3673 _objArray_path, // plain allocation, plus arrayof_oop_arraycopy
3674 _fast_path, // plain allocation, plus a CopyArray operation
3675 PATH_LIMIT
3676 };
3677 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3678 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3679 TypeInstPtr::NOTNULL);
3680 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3681 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3682 TypePtr::BOTTOM);
3683 record_for_igvn(result_reg);
3685 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
3686 int raw_adr_idx = Compile::AliasIdxRaw;
3687 const bool raw_mem_only = true;
3689 // paths into alloc_reg (on the fast path, just before the CopyArray):
3690 enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT };
3691 RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT);
3692 PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type);
3693 PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X);
3694 PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO);
3695 PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY,
3696 raw_adr_type);
3697 record_for_igvn(alloc_reg);
3699 bool card_mark = false; // (see below)
3701 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
3702 if (array_ctl != NULL) {
3703 // It's an array.
3704 PreserveJVMState pjvms(this);
3705 set_control(array_ctl);
3706 Node* obj_length = load_array_length(obj);
3707 Node* obj_size = NULL;
3708 _sp += nargs; // set original stack for use by uncommon_trap
3709 Node* alloc_obj = new_array(obj_klass, obj_length,
3710 raw_mem_only, &obj_size);
3711 _sp -= nargs;
3712 assert(obj_size != NULL, "");
3713 Node* raw_obj = alloc_obj->in(1);
3714 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
3715 if (ReduceBulkZeroing) {
3716 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
3717 if (alloc != NULL) {
3718 // We will be completely responsible for initializing this object.
3719 alloc->maybe_set_complete(&_gvn);
3720 }
3721 }
3723 if (!use_ReduceInitialCardMarks()) {
3724 // If it is an oop array, it requires very special treatment,
3725 // because card marking is required on each card of the array.
3726 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
3727 if (is_obja != NULL) {
3728 PreserveJVMState pjvms2(this);
3729 set_control(is_obja);
3730 // Generate a direct call to the right arraycopy function(s).
3731 bool disjoint_bases = true;
3732 bool length_never_negative = true;
3733 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3734 obj, intcon(0), alloc_obj, intcon(0),
3735 obj_length, nargs,
3736 disjoint_bases, length_never_negative);
3737 result_reg->init_req(_objArray_path, control());
3738 result_val->init_req(_objArray_path, alloc_obj);
3739 result_i_o ->set_req(_objArray_path, i_o());
3740 result_mem ->set_req(_objArray_path, reset_memory());
3741 }
3742 }
3743 // We can dispense with card marks if we know the allocation
3744 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
3745 // causes the non-eden paths to simulate a fresh allocation,
3746 // insofar that no further card marks are required to initialize
3747 // the object.
3749 // Otherwise, there are no card marks to worry about.
3750 alloc_val->init_req(_typeArray_alloc, raw_obj);
3751 alloc_siz->init_req(_typeArray_alloc, obj_size);
3752 alloc_reg->init_req(_typeArray_alloc, control());
3753 alloc_i_o->init_req(_typeArray_alloc, i_o());
3754 alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type));
3755 }
3757 // We only go to the fast case code if we pass a number of guards.
3758 // The paths which do not pass are accumulated in the slow_region.
3759 RegionNode* slow_region = new (C, 1) RegionNode(1);
3760 record_for_igvn(slow_region);
3761 if (!stopped()) {
3762 // It's an instance. Make the slow-path tests.
3763 // If this is a virtual call, we generate a funny guard. We grab
3764 // the vtable entry corresponding to clone() from the target object.
3765 // If the target method which we are calling happens to be the
3766 // Object clone() method, we pass the guard. We do not need this
3767 // guard for non-virtual calls; the caller is known to be the native
3768 // Object clone().
3769 if (is_virtual) {
3770 generate_virtual_guard(obj_klass, slow_region);
3771 }
3773 // The object must be cloneable and must not have a finalizer.
3774 // Both of these conditions may be checked in a single test.
3775 // We could optimize the cloneable test further, but we don't care.
3776 generate_access_flags_guard(obj_klass,
3777 // Test both conditions:
3778 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
3779 // Must be cloneable but not finalizer:
3780 JVM_ACC_IS_CLONEABLE,
3781 slow_region);
3782 }
3784 if (!stopped()) {
3785 // It's an instance, and it passed the slow-path tests.
3786 PreserveJVMState pjvms(this);
3787 Node* obj_size = NULL;
3788 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
3789 assert(obj_size != NULL, "");
3790 Node* raw_obj = alloc_obj->in(1);
3791 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
3792 if (ReduceBulkZeroing) {
3793 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
3794 if (alloc != NULL && !alloc->maybe_set_complete(&_gvn))
3795 alloc = NULL;
3796 }
3797 if (!use_ReduceInitialCardMarks()) {
3798 // Put in store barrier for any and all oops we are sticking
3799 // into this object. (We could avoid this if we could prove
3800 // that the object type contains no oop fields at all.)
3801 card_mark = true;
3802 }
3803 alloc_val->init_req(_instance_alloc, raw_obj);
3804 alloc_siz->init_req(_instance_alloc, obj_size);
3805 alloc_reg->init_req(_instance_alloc, control());
3806 alloc_i_o->init_req(_instance_alloc, i_o());
3807 alloc_mem->init_req(_instance_alloc, memory(raw_adr_type));
3808 }
3810 // Generate code for the slow case. We make a call to clone().
3811 set_control(_gvn.transform(slow_region));
3812 if (!stopped()) {
3813 PreserveJVMState pjvms(this);
3814 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
3815 Node* slow_result = set_results_for_java_call(slow_call);
3816 // this->control() comes from set_results_for_java_call
3817 result_reg->init_req(_slow_path, control());
3818 result_val->init_req(_slow_path, slow_result);
3819 result_i_o ->set_req(_slow_path, i_o());
3820 result_mem ->set_req(_slow_path, reset_memory());
3821 }
3823 // The object is allocated, as an array and/or an instance. Now copy it.
3824 set_control( _gvn.transform(alloc_reg) );
3825 set_i_o( _gvn.transform(alloc_i_o) );
3826 set_memory( _gvn.transform(alloc_mem), raw_adr_type );
3827 Node* raw_obj = _gvn.transform(alloc_val);
3829 if (!stopped()) {
3830 // Copy the fastest available way.
3831 // (No need for PreserveJVMState, since we're using it all up now.)
3832 // TODO: generate fields/elements copies for small objects instead.
3833 Node* src = obj;
3834 Node* dest = raw_obj;
3835 Node* size = _gvn.transform(alloc_siz);
3837 // Exclude the header.
3838 int base_off = instanceOopDesc::base_offset_in_bytes();
3839 if (UseCompressedOops) {
3840 assert(base_off % BytesPerLong != 0, "base with compressed oops");
3841 // With compressed oops base_offset_in_bytes is 12 which creates
3842 // the gap since countx is rounded by 8 bytes below.
3843 // Copy klass and the gap.
3844 base_off = instanceOopDesc::klass_offset_in_bytes();
3845 }
3846 src = basic_plus_adr(src, base_off);
3847 dest = basic_plus_adr(dest, base_off);
3849 // Compute the length also, if needed:
3850 Node* countx = size;
3851 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
3852 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
3854 // Select an appropriate instruction to initialize the range.
3855 // The CopyArray instruction (if supported) can be optimized
3856 // into a discrete set of scalar loads and stores.
3857 bool disjoint_bases = true;
3858 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
3859 src, NULL, dest, NULL, countx);
3861 // Now that the object is properly initialized, type it as an oop.
3862 // Use a secondary InitializeNode memory barrier.
3863 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx,
3864 raw_obj)->as_Initialize();
3865 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
3866 Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj,
3867 TypeInstPtr::NOTNULL);
3868 new_obj = _gvn.transform(new_obj);
3870 // If necessary, emit some card marks afterwards. (Non-arrays only.)
3871 if (card_mark) {
3872 Node* no_particular_value = NULL;
3873 Node* no_particular_field = NULL;
3874 post_barrier(control(),
3875 memory(raw_adr_type),
3876 new_obj,
3877 no_particular_field,
3878 raw_adr_idx,
3879 no_particular_value,
3880 T_OBJECT,
3881 false);
3882 }
3883 // Present the results of the slow call.
3884 result_reg->init_req(_fast_path, control());
3885 result_val->init_req(_fast_path, new_obj);
3886 result_i_o ->set_req(_fast_path, i_o());
3887 result_mem ->set_req(_fast_path, reset_memory());
3888 }
3890 // Return the combined state.
3891 set_control( _gvn.transform(result_reg) );
3892 set_i_o( _gvn.transform(result_i_o) );
3893 set_all_memory( _gvn.transform(result_mem) );
3895 // Cast the result to a sharper type, since we know what clone does.
3896 Node* new_obj = _gvn.transform(result_val);
3897 Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop);
3898 push(_gvn.transform(cast));
3900 return true;
3901 }
3904 // constants for computing the copy function
3905 enum {
3906 COPYFUNC_UNALIGNED = 0,
3907 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
3908 COPYFUNC_CONJOINT = 0,
3909 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
3910 };
3912 // Note: The condition "disjoint" applies also for overlapping copies
3913 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
3914 static address
3915 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
3916 int selector =
3917 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
3918 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
3920 #define RETURN_STUB(xxx_arraycopy) { \
3921 name = #xxx_arraycopy; \
3922 return StubRoutines::xxx_arraycopy(); }
3924 switch (t) {
3925 case T_BYTE:
3926 case T_BOOLEAN:
3927 switch (selector) {
3928 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
3929 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
3930 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
3931 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
3932 }
3933 case T_CHAR:
3934 case T_SHORT:
3935 switch (selector) {
3936 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
3937 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
3938 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
3939 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
3940 }
3941 case T_INT:
3942 case T_FLOAT:
3943 switch (selector) {
3944 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
3945 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
3946 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
3947 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
3948 }
3949 case T_DOUBLE:
3950 case T_LONG:
3951 switch (selector) {
3952 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
3953 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
3954 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
3955 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
3956 }
3957 case T_ARRAY:
3958 case T_OBJECT:
3959 switch (selector) {
3960 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
3961 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
3962 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
3963 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
3964 }
3965 default:
3966 ShouldNotReachHere();
3967 return NULL;
3968 }
3970 #undef RETURN_STUB
3971 }
3973 //------------------------------basictype2arraycopy----------------------------
3974 address LibraryCallKit::basictype2arraycopy(BasicType t,
3975 Node* src_offset,
3976 Node* dest_offset,
3977 bool disjoint_bases,
3978 const char* &name) {
3979 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
3980 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
3982 bool aligned = false;
3983 bool disjoint = disjoint_bases;
3985 // if the offsets are the same, we can treat the memory regions as
3986 // disjoint, because either the memory regions are in different arrays,
3987 // or they are identical (which we can treat as disjoint.) We can also
3988 // treat a copy with a destination index less that the source index
3989 // as disjoint since a low->high copy will work correctly in this case.
3990 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
3991 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
3992 // both indices are constants
3993 int s_offs = src_offset_inttype->get_con();
3994 int d_offs = dest_offset_inttype->get_con();
3995 int element_size = type2aelembytes(t);
3996 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
3997 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
3998 if (s_offs >= d_offs) disjoint = true;
3999 } else if (src_offset == dest_offset && src_offset != NULL) {
4000 // This can occur if the offsets are identical non-constants.
4001 disjoint = true;
4002 }
4004 return select_arraycopy_function(t, aligned, disjoint, name);
4005 }
4008 //------------------------------inline_arraycopy-----------------------
4009 bool LibraryCallKit::inline_arraycopy() {
4010 // Restore the stack and pop off the arguments.
4011 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4012 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4014 Node *src = argument(0);
4015 Node *src_offset = argument(1);
4016 Node *dest = argument(2);
4017 Node *dest_offset = argument(3);
4018 Node *length = argument(4);
4020 // Compile time checks. If any of these checks cannot be verified at compile time,
4021 // we do not make a fast path for this call. Instead, we let the call remain as it
4022 // is. The checks we choose to mandate at compile time are:
4023 //
4024 // (1) src and dest are arrays.
4025 const Type* src_type = src->Value(&_gvn);
4026 const Type* dest_type = dest->Value(&_gvn);
4027 const TypeAryPtr* top_src = src_type->isa_aryptr();
4028 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4029 if (top_src == NULL || top_src->klass() == NULL ||
4030 top_dest == NULL || top_dest->klass() == NULL) {
4031 // Conservatively insert a memory barrier on all memory slices.
4032 // Do not let writes into the source float below the arraycopy.
4033 insert_mem_bar(Op_MemBarCPUOrder);
4035 // Call StubRoutines::generic_arraycopy stub.
4036 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4037 src, src_offset, dest, dest_offset, length,
4038 nargs);
4040 // Do not let reads from the destination float above the arraycopy.
4041 // Since we cannot type the arrays, we don't know which slices
4042 // might be affected. We could restrict this barrier only to those
4043 // memory slices which pertain to array elements--but don't bother.
4044 if (!InsertMemBarAfterArraycopy)
4045 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4046 insert_mem_bar(Op_MemBarCPUOrder);
4047 return true;
4048 }
4050 // (2) src and dest arrays must have elements of the same BasicType
4051 // Figure out the size and type of the elements we will be copying.
4052 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4053 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4054 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4055 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4057 if (src_elem != dest_elem || dest_elem == T_VOID) {
4058 // The component types are not the same or are not recognized. Punt.
4059 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4060 generate_slow_arraycopy(TypePtr::BOTTOM,
4061 src, src_offset, dest, dest_offset, length,
4062 nargs);
4063 return true;
4064 }
4066 //---------------------------------------------------------------------------
4067 // We will make a fast path for this call to arraycopy.
4069 // We have the following tests left to perform:
4070 //
4071 // (3) src and dest must not be null.
4072 // (4) src_offset must not be negative.
4073 // (5) dest_offset must not be negative.
4074 // (6) length must not be negative.
4075 // (7) src_offset + length must not exceed length of src.
4076 // (8) dest_offset + length must not exceed length of dest.
4077 // (9) each element of an oop array must be assignable
4079 RegionNode* slow_region = new (C, 1) RegionNode(1);
4080 record_for_igvn(slow_region);
4082 // (3) operands must not be null
4083 // We currently perform our null checks with the do_null_check routine.
4084 // This means that the null exceptions will be reported in the caller
4085 // rather than (correctly) reported inside of the native arraycopy call.
4086 // This should be corrected, given time. We do our null check with the
4087 // stack pointer restored.
4088 _sp += nargs;
4089 src = do_null_check(src, T_ARRAY);
4090 dest = do_null_check(dest, T_ARRAY);
4091 _sp -= nargs;
4093 // (4) src_offset must not be negative.
4094 generate_negative_guard(src_offset, slow_region);
4096 // (5) dest_offset must not be negative.
4097 generate_negative_guard(dest_offset, slow_region);
4099 // (6) length must not be negative (moved to generate_arraycopy()).
4100 // generate_negative_guard(length, slow_region);
4102 // (7) src_offset + length must not exceed length of src.
4103 generate_limit_guard(src_offset, length,
4104 load_array_length(src),
4105 slow_region);
4107 // (8) dest_offset + length must not exceed length of dest.
4108 generate_limit_guard(dest_offset, length,
4109 load_array_length(dest),
4110 slow_region);
4112 // (9) each element of an oop array must be assignable
4113 // The generate_arraycopy subroutine checks this.
4115 // This is where the memory effects are placed:
4116 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4117 generate_arraycopy(adr_type, dest_elem,
4118 src, src_offset, dest, dest_offset, length,
4119 nargs, false, false, slow_region);
4121 return true;
4122 }
4124 //-----------------------------generate_arraycopy----------------------
4125 // Generate an optimized call to arraycopy.
4126 // Caller must guard against non-arrays.
4127 // Caller must determine a common array basic-type for both arrays.
4128 // Caller must validate offsets against array bounds.
4129 // The slow_region has already collected guard failure paths
4130 // (such as out of bounds length or non-conformable array types).
4131 // The generated code has this shape, in general:
4132 //
4133 // if (length == 0) return // via zero_path
4134 // slowval = -1
4135 // if (types unknown) {
4136 // slowval = call generic copy loop
4137 // if (slowval == 0) return // via checked_path
4138 // } else if (indexes in bounds) {
4139 // if ((is object array) && !(array type check)) {
4140 // slowval = call checked copy loop
4141 // if (slowval == 0) return // via checked_path
4142 // } else {
4143 // call bulk copy loop
4144 // return // via fast_path
4145 // }
4146 // }
4147 // // adjust params for remaining work:
4148 // if (slowval != -1) {
4149 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4150 // }
4151 // slow_region:
4152 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4153 // return // via slow_call_path
4154 //
4155 // This routine is used from several intrinsics: System.arraycopy,
4156 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4157 //
4158 void
4159 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4160 BasicType basic_elem_type,
4161 Node* src, Node* src_offset,
4162 Node* dest, Node* dest_offset,
4163 Node* copy_length,
4164 int nargs,
4165 bool disjoint_bases,
4166 bool length_never_negative,
4167 RegionNode* slow_region) {
4169 if (slow_region == NULL) {
4170 slow_region = new(C,1) RegionNode(1);
4171 record_for_igvn(slow_region);
4172 }
4174 Node* original_dest = dest;
4175 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4176 Node* raw_dest = NULL; // used before zeroing, if needed
4177 bool must_clear_dest = false;
4179 // See if this is the initialization of a newly-allocated array.
4180 // If so, we will take responsibility here for initializing it to zero.
4181 // (Note: Because tightly_coupled_allocation performs checks on the
4182 // out-edges of the dest, we need to avoid making derived pointers
4183 // from it until we have checked its uses.)
4184 if (ReduceBulkZeroing
4185 && !ZeroTLAB // pointless if already zeroed
4186 && basic_elem_type != T_CONFLICT // avoid corner case
4187 && !_gvn.eqv_uncast(src, dest)
4188 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4189 != NULL)
4190 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4191 && alloc->maybe_set_complete(&_gvn)) {
4192 // "You break it, you buy it."
4193 InitializeNode* init = alloc->initialization();
4194 assert(init->is_complete(), "we just did this");
4195 assert(dest->Opcode() == Op_CheckCastPP, "sanity");
4196 assert(dest->in(0)->in(0) == init, "dest pinned");
4197 raw_dest = dest->in(1); // grab the raw pointer!
4198 original_dest = dest;
4199 dest = raw_dest;
4200 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4201 // Decouple the original InitializeNode, turning it into a simple membar.
4202 // We will build a new one at the end of this routine.
4203 init->set_req(InitializeNode::RawAddress, top());
4204 // From this point on, every exit path is responsible for
4205 // initializing any non-copied parts of the object to zero.
4206 must_clear_dest = true;
4207 } else {
4208 // No zeroing elimination here.
4209 alloc = NULL;
4210 //original_dest = dest;
4211 //must_clear_dest = false;
4212 }
4214 // Results are placed here:
4215 enum { fast_path = 1, // normal void-returning assembly stub
4216 checked_path = 2, // special assembly stub with cleanup
4217 slow_call_path = 3, // something went wrong; call the VM
4218 zero_path = 4, // bypass when length of copy is zero
4219 bcopy_path = 5, // copy primitive array by 64-bit blocks
4220 PATH_LIMIT = 6
4221 };
4222 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4223 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4224 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4225 record_for_igvn(result_region);
4226 _gvn.set_type_bottom(result_i_o);
4227 _gvn.set_type_bottom(result_memory);
4228 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4230 // The slow_control path:
4231 Node* slow_control;
4232 Node* slow_i_o = i_o();
4233 Node* slow_mem = memory(adr_type);
4234 debug_only(slow_control = (Node*) badAddress);
4236 // Checked control path:
4237 Node* checked_control = top();
4238 Node* checked_mem = NULL;
4239 Node* checked_i_o = NULL;
4240 Node* checked_value = NULL;
4242 if (basic_elem_type == T_CONFLICT) {
4243 assert(!must_clear_dest, "");
4244 Node* cv = generate_generic_arraycopy(adr_type,
4245 src, src_offset, dest, dest_offset,
4246 copy_length, nargs);
4247 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4248 checked_control = control();
4249 checked_i_o = i_o();
4250 checked_mem = memory(adr_type);
4251 checked_value = cv;
4252 set_control(top()); // no fast path
4253 }
4255 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4256 if (not_pos != NULL) {
4257 PreserveJVMState pjvms(this);
4258 set_control(not_pos);
4260 // (6) length must not be negative.
4261 if (!length_never_negative) {
4262 generate_negative_guard(copy_length, slow_region);
4263 }
4265 if (!stopped() && must_clear_dest) {
4266 Node* dest_length = alloc->in(AllocateNode::ALength);
4267 if (_gvn.eqv_uncast(copy_length, dest_length)
4268 || _gvn.find_int_con(dest_length, 1) <= 0) {
4269 // There is no zeroing to do.
4270 } else {
4271 // Clear the whole thing since there are no source elements to copy.
4272 generate_clear_array(adr_type, dest, basic_elem_type,
4273 intcon(0), NULL,
4274 alloc->in(AllocateNode::AllocSize));
4275 }
4276 }
4278 // Present the results of the fast call.
4279 result_region->init_req(zero_path, control());
4280 result_i_o ->init_req(zero_path, i_o());
4281 result_memory->init_req(zero_path, memory(adr_type));
4282 }
4284 if (!stopped() && must_clear_dest) {
4285 // We have to initialize the *uncopied* part of the array to zero.
4286 // The copy destination is the slice dest[off..off+len]. The other slices
4287 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4288 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4289 Node* dest_length = alloc->in(AllocateNode::ALength);
4290 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4291 copy_length) );
4293 // If there is a head section that needs zeroing, do it now.
4294 if (find_int_con(dest_offset, -1) != 0) {
4295 generate_clear_array(adr_type, dest, basic_elem_type,
4296 intcon(0), dest_offset,
4297 NULL);
4298 }
4300 // Next, perform a dynamic check on the tail length.
4301 // It is often zero, and we can win big if we prove this.
4302 // There are two wins: Avoid generating the ClearArray
4303 // with its attendant messy index arithmetic, and upgrade
4304 // the copy to a more hardware-friendly word size of 64 bits.
4305 Node* tail_ctl = NULL;
4306 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4307 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4308 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4309 tail_ctl = generate_slow_guard(bol_lt, NULL);
4310 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4311 }
4313 // At this point, let's assume there is no tail.
4314 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4315 // There is no tail. Try an upgrade to a 64-bit copy.
4316 bool didit = false;
4317 { PreserveJVMState pjvms(this);
4318 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4319 src, src_offset, dest, dest_offset,
4320 dest_size);
4321 if (didit) {
4322 // Present the results of the block-copying fast call.
4323 result_region->init_req(bcopy_path, control());
4324 result_i_o ->init_req(bcopy_path, i_o());
4325 result_memory->init_req(bcopy_path, memory(adr_type));
4326 }
4327 }
4328 if (didit)
4329 set_control(top()); // no regular fast path
4330 }
4332 // Clear the tail, if any.
4333 if (tail_ctl != NULL) {
4334 Node* notail_ctl = stopped() ? NULL : control();
4335 set_control(tail_ctl);
4336 if (notail_ctl == NULL) {
4337 generate_clear_array(adr_type, dest, basic_elem_type,
4338 dest_tail, NULL,
4339 dest_size);
4340 } else {
4341 // Make a local merge.
4342 Node* done_ctl = new(C,3) RegionNode(3);
4343 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4344 done_ctl->init_req(1, notail_ctl);
4345 done_mem->init_req(1, memory(adr_type));
4346 generate_clear_array(adr_type, dest, basic_elem_type,
4347 dest_tail, NULL,
4348 dest_size);
4349 done_ctl->init_req(2, control());
4350 done_mem->init_req(2, memory(adr_type));
4351 set_control( _gvn.transform(done_ctl) );
4352 set_memory( _gvn.transform(done_mem), adr_type );
4353 }
4354 }
4355 }
4357 BasicType copy_type = basic_elem_type;
4358 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4359 if (!stopped() && copy_type == T_OBJECT) {
4360 // If src and dest have compatible element types, we can copy bits.
4361 // Types S[] and D[] are compatible if D is a supertype of S.
4362 //
4363 // If they are not, we will use checked_oop_disjoint_arraycopy,
4364 // which performs a fast optimistic per-oop check, and backs off
4365 // further to JVM_ArrayCopy on the first per-oop check that fails.
4366 // (Actually, we don't move raw bits only; the GC requires card marks.)
4368 // Get the klassOop for both src and dest
4369 Node* src_klass = load_object_klass(src);
4370 Node* dest_klass = load_object_klass(dest);
4372 // Generate the subtype check.
4373 // This might fold up statically, or then again it might not.
4374 //
4375 // Non-static example: Copying List<String>.elements to a new String[].
4376 // The backing store for a List<String> is always an Object[],
4377 // but its elements are always type String, if the generic types
4378 // are correct at the source level.
4379 //
4380 // Test S[] against D[], not S against D, because (probably)
4381 // the secondary supertype cache is less busy for S[] than S.
4382 // This usually only matters when D is an interface.
4383 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4384 // Plug failing path into checked_oop_disjoint_arraycopy
4385 if (not_subtype_ctrl != top()) {
4386 PreserveJVMState pjvms(this);
4387 set_control(not_subtype_ctrl);
4388 // (At this point we can assume disjoint_bases, since types differ.)
4389 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4390 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4391 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4392 Node* dest_elem_klass = _gvn.transform(n1);
4393 Node* cv = generate_checkcast_arraycopy(adr_type,
4394 dest_elem_klass,
4395 src, src_offset, dest, dest_offset,
4396 copy_length,
4397 nargs);
4398 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4399 checked_control = control();
4400 checked_i_o = i_o();
4401 checked_mem = memory(adr_type);
4402 checked_value = cv;
4403 }
4404 // At this point we know we do not need type checks on oop stores.
4406 // Let's see if we need card marks:
4407 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4408 // If we do not need card marks, copy using the jint or jlong stub.
4409 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4410 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4411 "sizes agree");
4412 }
4413 }
4415 if (!stopped()) {
4416 // Generate the fast path, if possible.
4417 PreserveJVMState pjvms(this);
4418 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4419 src, src_offset, dest, dest_offset,
4420 ConvI2X(copy_length));
4422 // Present the results of the fast call.
4423 result_region->init_req(fast_path, control());
4424 result_i_o ->init_req(fast_path, i_o());
4425 result_memory->init_req(fast_path, memory(adr_type));
4426 }
4428 // Here are all the slow paths up to this point, in one bundle:
4429 slow_control = top();
4430 if (slow_region != NULL)
4431 slow_control = _gvn.transform(slow_region);
4432 debug_only(slow_region = (RegionNode*)badAddress);
4434 set_control(checked_control);
4435 if (!stopped()) {
4436 // Clean up after the checked call.
4437 // The returned value is either 0 or -1^K,
4438 // where K = number of partially transferred array elements.
4439 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4440 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4441 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4443 // If it is 0, we are done, so transfer to the end.
4444 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4445 result_region->init_req(checked_path, checks_done);
4446 result_i_o ->init_req(checked_path, checked_i_o);
4447 result_memory->init_req(checked_path, checked_mem);
4449 // If it is not zero, merge into the slow call.
4450 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4451 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4452 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4453 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4454 record_for_igvn(slow_reg2);
4455 slow_reg2 ->init_req(1, slow_control);
4456 slow_i_o2 ->init_req(1, slow_i_o);
4457 slow_mem2 ->init_req(1, slow_mem);
4458 slow_reg2 ->init_req(2, control());
4459 slow_i_o2 ->init_req(2, i_o());
4460 slow_mem2 ->init_req(2, memory(adr_type));
4462 slow_control = _gvn.transform(slow_reg2);
4463 slow_i_o = _gvn.transform(slow_i_o2);
4464 slow_mem = _gvn.transform(slow_mem2);
4466 if (alloc != NULL) {
4467 // We'll restart from the very beginning, after zeroing the whole thing.
4468 // This can cause double writes, but that's OK since dest is brand new.
4469 // So we ignore the low 31 bits of the value returned from the stub.
4470 } else {
4471 // We must continue the copy exactly where it failed, or else
4472 // another thread might see the wrong number of writes to dest.
4473 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4474 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4475 slow_offset->init_req(1, intcon(0));
4476 slow_offset->init_req(2, checked_offset);
4477 slow_offset = _gvn.transform(slow_offset);
4479 // Adjust the arguments by the conditionally incoming offset.
4480 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4481 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4482 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4484 // Tweak the node variables to adjust the code produced below:
4485 src_offset = src_off_plus;
4486 dest_offset = dest_off_plus;
4487 copy_length = length_minus;
4488 }
4489 }
4491 set_control(slow_control);
4492 if (!stopped()) {
4493 // Generate the slow path, if needed.
4494 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4496 set_memory(slow_mem, adr_type);
4497 set_i_o(slow_i_o);
4499 if (must_clear_dest) {
4500 generate_clear_array(adr_type, dest, basic_elem_type,
4501 intcon(0), NULL,
4502 alloc->in(AllocateNode::AllocSize));
4503 }
4505 if (dest != original_dest) {
4506 // Promote from rawptr to oop, so it looks right in the call's GC map.
4507 dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest,
4508 TypeInstPtr::NOTNULL) );
4510 // Edit the call's debug-info to avoid referring to original_dest.
4511 // (The problem with original_dest is that it isn't ready until
4512 // after the InitializeNode completes, but this stuff is before.)
4513 // Substitute in the locally valid dest_oop.
4514 replace_in_map(original_dest, dest);
4515 }
4517 generate_slow_arraycopy(adr_type,
4518 src, src_offset, dest, dest_offset,
4519 copy_length, nargs);
4521 result_region->init_req(slow_call_path, control());
4522 result_i_o ->init_req(slow_call_path, i_o());
4523 result_memory->init_req(slow_call_path, memory(adr_type));
4524 }
4526 // Remove unused edges.
4527 for (uint i = 1; i < result_region->req(); i++) {
4528 if (result_region->in(i) == NULL)
4529 result_region->init_req(i, top());
4530 }
4532 // Finished; return the combined state.
4533 set_control( _gvn.transform(result_region) );
4534 set_i_o( _gvn.transform(result_i_o) );
4535 set_memory( _gvn.transform(result_memory), adr_type );
4537 if (dest != original_dest) {
4538 // Pin the "finished" array node after the arraycopy/zeroing operations.
4539 // Use a secondary InitializeNode memory barrier.
4540 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4541 Compile::AliasIdxRaw,
4542 raw_dest)->as_Initialize();
4543 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4544 _gvn.hash_delete(original_dest);
4545 original_dest->set_req(0, control());
4546 _gvn.hash_find_insert(original_dest); // put back into GVN table
4547 }
4549 // The memory edges above are precise in order to model effects around
4550 // array copyies accurately to allow value numbering of field loads around
4551 // arraycopy. Such field loads, both before and after, are common in Java
4552 // collections and similar classes involving header/array data structures.
4553 //
4554 // But with low number of register or when some registers are used or killed
4555 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4556 // The next memory barrier is added to avoid it. If the arraycopy can be
4557 // optimized away (which it can, sometimes) then we can manually remove
4558 // the membar also.
4559 if (InsertMemBarAfterArraycopy)
4560 insert_mem_bar(Op_MemBarCPUOrder);
4561 }
4564 // Helper function which determines if an arraycopy immediately follows
4565 // an allocation, with no intervening tests or other escapes for the object.
4566 AllocateArrayNode*
4567 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4568 RegionNode* slow_region) {
4569 if (stopped()) return NULL; // no fast path
4570 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4572 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4573 if (alloc == NULL) return NULL;
4575 Node* rawmem = memory(Compile::AliasIdxRaw);
4576 // Is the allocation's memory state untouched?
4577 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4578 // Bail out if there have been raw-memory effects since the allocation.
4579 // (Example: There might have been a call or safepoint.)
4580 return NULL;
4581 }
4582 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4583 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4584 return NULL;
4585 }
4587 // There must be no unexpected observers of this allocation.
4588 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4589 Node* obs = ptr->fast_out(i);
4590 if (obs != this->map()) {
4591 return NULL;
4592 }
4593 }
4595 // This arraycopy must unconditionally follow the allocation of the ptr.
4596 Node* alloc_ctl = ptr->in(0);
4597 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4599 Node* ctl = control();
4600 while (ctl != alloc_ctl) {
4601 // There may be guards which feed into the slow_region.
4602 // Any other control flow means that we might not get a chance
4603 // to finish initializing the allocated object.
4604 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4605 IfNode* iff = ctl->in(0)->as_If();
4606 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4607 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4608 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4609 ctl = iff->in(0); // This test feeds the known slow_region.
4610 continue;
4611 }
4612 // One more try: Various low-level checks bottom out in
4613 // uncommon traps. If the debug-info of the trap omits
4614 // any reference to the allocation, as we've already
4615 // observed, then there can be no objection to the trap.
4616 bool found_trap = false;
4617 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4618 Node* obs = not_ctl->fast_out(j);
4619 if (obs->in(0) == not_ctl && obs->is_Call() &&
4620 (obs->as_Call()->entry_point() ==
4621 SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
4622 found_trap = true; break;
4623 }
4624 }
4625 if (found_trap) {
4626 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4627 continue;
4628 }
4629 }
4630 return NULL;
4631 }
4633 // If we get this far, we have an allocation which immediately
4634 // precedes the arraycopy, and we can take over zeroing the new object.
4635 // The arraycopy will finish the initialization, and provide
4636 // a new control state to which we will anchor the destination pointer.
4638 return alloc;
4639 }
4641 // Helper for initialization of arrays, creating a ClearArray.
4642 // It writes zero bits in [start..end), within the body of an array object.
4643 // The memory effects are all chained onto the 'adr_type' alias category.
4644 //
4645 // Since the object is otherwise uninitialized, we are free
4646 // to put a little "slop" around the edges of the cleared area,
4647 // as long as it does not go back into the array's header,
4648 // or beyond the array end within the heap.
4649 //
4650 // The lower edge can be rounded down to the nearest jint and the
4651 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4652 //
4653 // Arguments:
4654 // adr_type memory slice where writes are generated
4655 // dest oop of the destination array
4656 // basic_elem_type element type of the destination
4657 // slice_idx array index of first element to store
4658 // slice_len number of elements to store (or NULL)
4659 // dest_size total size in bytes of the array object
4660 //
4661 // Exactly one of slice_len or dest_size must be non-NULL.
4662 // If dest_size is non-NULL, zeroing extends to the end of the object.
4663 // If slice_len is non-NULL, the slice_idx value must be a constant.
4664 void
4665 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
4666 Node* dest,
4667 BasicType basic_elem_type,
4668 Node* slice_idx,
4669 Node* slice_len,
4670 Node* dest_size) {
4671 // one or the other but not both of slice_len and dest_size:
4672 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
4673 if (slice_len == NULL) slice_len = top();
4674 if (dest_size == NULL) dest_size = top();
4676 // operate on this memory slice:
4677 Node* mem = memory(adr_type); // memory slice to operate on
4679 // scaling and rounding of indexes:
4680 int scale = exact_log2(type2aelembytes(basic_elem_type));
4681 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
4682 int clear_low = (-1 << scale) & (BytesPerInt - 1);
4683 int bump_bit = (-1 << scale) & BytesPerInt;
4685 // determine constant starts and ends
4686 const intptr_t BIG_NEG = -128;
4687 assert(BIG_NEG + 2*abase < 0, "neg enough");
4688 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
4689 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
4690 if (slice_len_con == 0) {
4691 return; // nothing to do here
4692 }
4693 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
4694 intptr_t end_con = find_intptr_t_con(dest_size, -1);
4695 if (slice_idx_con >= 0 && slice_len_con >= 0) {
4696 assert(end_con < 0, "not two cons");
4697 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
4698 BytesPerLong);
4699 }
4701 if (start_con >= 0 && end_con >= 0) {
4702 // Constant start and end. Simple.
4703 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4704 start_con, end_con, &_gvn);
4705 } else if (start_con >= 0 && dest_size != top()) {
4706 // Constant start, pre-rounded end after the tail of the array.
4707 Node* end = dest_size;
4708 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4709 start_con, end, &_gvn);
4710 } else if (start_con >= 0 && slice_len != top()) {
4711 // Constant start, non-constant end. End needs rounding up.
4712 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
4713 intptr_t end_base = abase + (slice_idx_con << scale);
4714 int end_round = (-1 << scale) & (BytesPerLong - 1);
4715 Node* end = ConvI2X(slice_len);
4716 if (scale != 0)
4717 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
4718 end_base += end_round;
4719 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
4720 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
4721 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4722 start_con, end, &_gvn);
4723 } else if (start_con < 0 && dest_size != top()) {
4724 // Non-constant start, pre-rounded end after the tail of the array.
4725 // This is almost certainly a "round-to-end" operation.
4726 Node* start = slice_idx;
4727 start = ConvI2X(start);
4728 if (scale != 0)
4729 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
4730 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
4731 if ((bump_bit | clear_low) != 0) {
4732 int to_clear = (bump_bit | clear_low);
4733 // Align up mod 8, then store a jint zero unconditionally
4734 // just before the mod-8 boundary.
4735 if (((abase + bump_bit) & ~to_clear) - bump_bit
4736 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
4737 bump_bit = 0;
4738 assert((abase & to_clear) == 0, "array base must be long-aligned");
4739 } else {
4740 // Bump 'start' up to (or past) the next jint boundary:
4741 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
4742 assert((abase & clear_low) == 0, "array base must be int-aligned");
4743 }
4744 // Round bumped 'start' down to jlong boundary in body of array.
4745 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
4746 if (bump_bit != 0) {
4747 // Store a zero to the immediately preceding jint:
4748 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
4749 Node* p1 = basic_plus_adr(dest, x1);
4750 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
4751 mem = _gvn.transform(mem);
4752 }
4753 }
4754 Node* end = dest_size; // pre-rounded
4755 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4756 start, end, &_gvn);
4757 } else {
4758 // Non-constant start, unrounded non-constant end.
4759 // (Nobody zeroes a random midsection of an array using this routine.)
4760 ShouldNotReachHere(); // fix caller
4761 }
4763 // Done.
4764 set_memory(mem, adr_type);
4765 }
4768 bool
4769 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
4770 BasicType basic_elem_type,
4771 AllocateNode* alloc,
4772 Node* src, Node* src_offset,
4773 Node* dest, Node* dest_offset,
4774 Node* dest_size) {
4775 // See if there is an advantage from block transfer.
4776 int scale = exact_log2(type2aelembytes(basic_elem_type));
4777 if (scale >= LogBytesPerLong)
4778 return false; // it is already a block transfer
4780 // Look at the alignment of the starting offsets.
4781 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
4782 const intptr_t BIG_NEG = -128;
4783 assert(BIG_NEG + 2*abase < 0, "neg enough");
4785 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
4786 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
4787 if (src_off < 0 || dest_off < 0)
4788 // At present, we can only understand constants.
4789 return false;
4791 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
4792 // Non-aligned; too bad.
4793 // One more chance: Pick off an initial 32-bit word.
4794 // This is a common case, since abase can be odd mod 8.
4795 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
4796 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
4797 Node* sptr = basic_plus_adr(src, src_off);
4798 Node* dptr = basic_plus_adr(dest, dest_off);
4799 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
4800 store_to_memory(control(), dptr, sval, T_INT, adr_type);
4801 src_off += BytesPerInt;
4802 dest_off += BytesPerInt;
4803 } else {
4804 return false;
4805 }
4806 }
4807 assert(src_off % BytesPerLong == 0, "");
4808 assert(dest_off % BytesPerLong == 0, "");
4810 // Do this copy by giant steps.
4811 Node* sptr = basic_plus_adr(src, src_off);
4812 Node* dptr = basic_plus_adr(dest, dest_off);
4813 Node* countx = dest_size;
4814 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
4815 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
4817 bool disjoint_bases = true; // since alloc != NULL
4818 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
4819 sptr, NULL, dptr, NULL, countx);
4821 return true;
4822 }
4825 // Helper function; generates code for the slow case.
4826 // We make a call to a runtime method which emulates the native method,
4827 // but without the native wrapper overhead.
4828 void
4829 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
4830 Node* src, Node* src_offset,
4831 Node* dest, Node* dest_offset,
4832 Node* copy_length,
4833 int nargs) {
4834 _sp += nargs; // any deopt will start just before call to enclosing method
4835 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
4836 OptoRuntime::slow_arraycopy_Type(),
4837 OptoRuntime::slow_arraycopy_Java(),
4838 "slow_arraycopy", adr_type,
4839 src, src_offset, dest, dest_offset,
4840 copy_length);
4841 _sp -= nargs;
4843 // Handle exceptions thrown by this fellow:
4844 make_slow_call_ex(call, env()->Throwable_klass(), false);
4845 }
4847 // Helper function; generates code for cases requiring runtime checks.
4848 Node*
4849 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
4850 Node* dest_elem_klass,
4851 Node* src, Node* src_offset,
4852 Node* dest, Node* dest_offset,
4853 Node* copy_length,
4854 int nargs) {
4855 if (stopped()) return NULL;
4857 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
4858 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
4859 return NULL;
4860 }
4862 // Pick out the parameters required to perform a store-check
4863 // for the target array. This is an optimistic check. It will
4864 // look in each non-null element's class, at the desired klass's
4865 // super_check_offset, for the desired klass.
4866 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
4867 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
4868 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
4869 Node* check_offset = _gvn.transform(n3);
4870 Node* check_value = dest_elem_klass;
4872 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
4873 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
4875 // (We know the arrays are never conjoint, because their types differ.)
4876 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
4877 OptoRuntime::checkcast_arraycopy_Type(),
4878 copyfunc_addr, "checkcast_arraycopy", adr_type,
4879 // five arguments, of which two are
4880 // intptr_t (jlong in LP64)
4881 src_start, dest_start,
4882 copy_length XTOP,
4883 check_offset XTOP,
4884 check_value);
4886 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
4887 }
4890 // Helper function; generates code for cases requiring runtime checks.
4891 Node*
4892 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
4893 Node* src, Node* src_offset,
4894 Node* dest, Node* dest_offset,
4895 Node* copy_length,
4896 int nargs) {
4897 if (stopped()) return NULL;
4899 address copyfunc_addr = StubRoutines::generic_arraycopy();
4900 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
4901 return NULL;
4902 }
4904 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
4905 OptoRuntime::generic_arraycopy_Type(),
4906 copyfunc_addr, "generic_arraycopy", adr_type,
4907 src, src_offset, dest, dest_offset, copy_length);
4909 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
4910 }
4912 // Helper function; generates the fast out-of-line call to an arraycopy stub.
4913 void
4914 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
4915 BasicType basic_elem_type,
4916 bool disjoint_bases,
4917 Node* src, Node* src_offset,
4918 Node* dest, Node* dest_offset,
4919 Node* copy_length) {
4920 if (stopped()) return; // nothing to do
4922 Node* src_start = src;
4923 Node* dest_start = dest;
4924 if (src_offset != NULL || dest_offset != NULL) {
4925 assert(src_offset != NULL && dest_offset != NULL, "");
4926 src_start = array_element_address(src, src_offset, basic_elem_type);
4927 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
4928 }
4930 // Figure out which arraycopy runtime method to call.
4931 const char* copyfunc_name = "arraycopy";
4932 address copyfunc_addr =
4933 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
4934 disjoint_bases, copyfunc_name);
4936 // Call it. Note that the count_ix value is not scaled to a byte-size.
4937 make_runtime_call(RC_LEAF|RC_NO_FP,
4938 OptoRuntime::fast_arraycopy_Type(),
4939 copyfunc_addr, copyfunc_name, adr_type,
4940 src_start, dest_start, copy_length XTOP);
4941 }