Thu, 29 May 2008 16:22:09 -0700
6695049: (coll) Create an x86 intrinsic for Arrays.equals
Summary: Intrinsify java/util/Arrays.equals(char[], char[])
Reviewed-by: kvn, never
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_array_equals();
167 bool inline_native_clone(bool is_virtual);
168 bool inline_native_Reflection_getCallerClass();
169 bool inline_native_AtomicLong_get();
170 bool inline_native_AtomicLong_attemptUpdate();
171 bool is_method_invoke_or_aux_frame(JVMState* jvms);
172 // Helper function for inlining native object hash method
173 bool inline_native_hashcode(bool is_virtual, bool is_static);
174 bool inline_native_getClass();
176 // Helper functions for inlining arraycopy
177 bool inline_arraycopy();
178 void generate_arraycopy(const TypePtr* adr_type,
179 BasicType basic_elem_type,
180 Node* src, Node* src_offset,
181 Node* dest, Node* dest_offset,
182 Node* copy_length,
183 int nargs, // arguments on stack for debug info
184 bool disjoint_bases = false,
185 bool length_never_negative = false,
186 RegionNode* slow_region = NULL);
187 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
188 RegionNode* slow_region);
189 void generate_clear_array(const TypePtr* adr_type,
190 Node* dest,
191 BasicType basic_elem_type,
192 Node* slice_off,
193 Node* slice_len,
194 Node* slice_end);
195 bool generate_block_arraycopy(const TypePtr* adr_type,
196 BasicType basic_elem_type,
197 AllocateNode* alloc,
198 Node* src, Node* src_offset,
199 Node* dest, Node* dest_offset,
200 Node* dest_size);
201 void generate_slow_arraycopy(const TypePtr* adr_type,
202 Node* src, Node* src_offset,
203 Node* dest, Node* dest_offset,
204 Node* copy_length,
205 int nargs);
206 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
207 Node* dest_elem_klass,
208 Node* src, Node* src_offset,
209 Node* dest, Node* dest_offset,
210 Node* copy_length, int nargs);
211 Node* generate_generic_arraycopy(const TypePtr* adr_type,
212 Node* src, Node* src_offset,
213 Node* dest, Node* dest_offset,
214 Node* copy_length, int nargs);
215 void generate_unchecked_arraycopy(const TypePtr* adr_type,
216 BasicType basic_elem_type,
217 bool disjoint_bases,
218 Node* src, Node* src_offset,
219 Node* dest, Node* dest_offset,
220 Node* copy_length);
221 bool inline_unsafe_CAS(BasicType type);
222 bool inline_unsafe_ordered_store(BasicType type);
223 bool inline_fp_conversions(vmIntrinsics::ID id);
224 bool inline_reverseBytes(vmIntrinsics::ID id);
225 };
228 //---------------------------make_vm_intrinsic----------------------------
229 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
230 vmIntrinsics::ID id = m->intrinsic_id();
231 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
233 if (DisableIntrinsic[0] != '\0'
234 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
235 // disabled by a user request on the command line:
236 // example: -XX:DisableIntrinsic=_hashCode,_getClass
237 return NULL;
238 }
240 if (!m->is_loaded()) {
241 // do not attempt to inline unloaded methods
242 return NULL;
243 }
245 // Only a few intrinsics implement a virtual dispatch.
246 // They are expensive calls which are also frequently overridden.
247 if (is_virtual) {
248 switch (id) {
249 case vmIntrinsics::_hashCode:
250 case vmIntrinsics::_clone:
251 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
252 break;
253 default:
254 return NULL;
255 }
256 }
258 // -XX:-InlineNatives disables nearly all intrinsics:
259 if (!InlineNatives) {
260 switch (id) {
261 case vmIntrinsics::_indexOf:
262 case vmIntrinsics::_compareTo:
263 case vmIntrinsics::_equalsC:
264 break; // InlineNatives does not control String.compareTo
265 default:
266 return NULL;
267 }
268 }
270 switch (id) {
271 case vmIntrinsics::_compareTo:
272 if (!SpecialStringCompareTo) return NULL;
273 break;
274 case vmIntrinsics::_indexOf:
275 if (!SpecialStringIndexOf) return NULL;
276 break;
277 case vmIntrinsics::_equalsC:
278 if (!SpecialArraysEquals) return NULL;
279 break;
280 case vmIntrinsics::_arraycopy:
281 if (!InlineArrayCopy) return NULL;
282 break;
283 case vmIntrinsics::_copyMemory:
284 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
285 if (!InlineArrayCopy) return NULL;
286 break;
287 case vmIntrinsics::_hashCode:
288 if (!InlineObjectHash) return NULL;
289 break;
290 case vmIntrinsics::_clone:
291 case vmIntrinsics::_copyOf:
292 case vmIntrinsics::_copyOfRange:
293 if (!InlineObjectCopy) return NULL;
294 // These also use the arraycopy intrinsic mechanism:
295 if (!InlineArrayCopy) return NULL;
296 break;
297 case vmIntrinsics::_checkIndex:
298 // We do not intrinsify this. The optimizer does fine with it.
299 return NULL;
301 case vmIntrinsics::_get_AtomicLong:
302 case vmIntrinsics::_attemptUpdate:
303 if (!InlineAtomicLong) return NULL;
304 break;
306 case vmIntrinsics::_Object_init:
307 case vmIntrinsics::_invoke:
308 // We do not intrinsify these; they are marked for other purposes.
309 return NULL;
311 case vmIntrinsics::_getCallerClass:
312 if (!UseNewReflection) return NULL;
313 if (!InlineReflectionGetCallerClass) return NULL;
314 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
315 break;
317 default:
318 break;
319 }
321 // -XX:-InlineClassNatives disables natives from the Class class.
322 // The flag applies to all reflective calls, notably Array.newArray
323 // (visible to Java programmers as Array.newInstance).
324 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
325 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
326 if (!InlineClassNatives) return NULL;
327 }
329 // -XX:-InlineThreadNatives disables natives from the Thread class.
330 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
331 if (!InlineThreadNatives) return NULL;
332 }
334 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
335 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
336 m->holder()->name() == ciSymbol::java_lang_Float() ||
337 m->holder()->name() == ciSymbol::java_lang_Double()) {
338 if (!InlineMathNatives) return NULL;
339 }
341 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
342 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
343 if (!InlineUnsafeOps) return NULL;
344 }
346 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
347 }
349 //----------------------register_library_intrinsics-----------------------
350 // Initialize this file's data structures, for each Compile instance.
351 void Compile::register_library_intrinsics() {
352 // Nothing to do here.
353 }
355 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
356 LibraryCallKit kit(jvms, this);
357 Compile* C = kit.C;
358 int nodes = C->unique();
359 #ifndef PRODUCT
360 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
361 char buf[1000];
362 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
363 tty->print_cr("Intrinsic %s", str);
364 }
365 #endif
366 if (kit.try_to_inline()) {
367 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
368 tty->print("Inlining intrinsic %s%s at bci:%d in",
369 vmIntrinsics::name_at(intrinsic_id()),
370 (is_virtual() ? " (virtual)" : ""), kit.bci());
371 kit.caller()->print_short_name(tty);
372 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
373 }
374 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
375 if (C->log()) {
376 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
377 vmIntrinsics::name_at(intrinsic_id()),
378 (is_virtual() ? " virtual='1'" : ""),
379 C->unique() - nodes);
380 }
381 return kit.transfer_exceptions_into_jvms();
382 }
384 if (PrintIntrinsics) {
385 switch (intrinsic_id()) {
386 case vmIntrinsics::_invoke:
387 case vmIntrinsics::_Object_init:
388 // We do not expect to inline these, so do not produce any noise about them.
389 break;
390 default:
391 tty->print("Did not inline intrinsic %s%s at bci:%d in",
392 vmIntrinsics::name_at(intrinsic_id()),
393 (is_virtual() ? " (virtual)" : ""), kit.bci());
394 kit.caller()->print_short_name(tty);
395 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
396 }
397 }
398 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
399 return NULL;
400 }
402 bool LibraryCallKit::try_to_inline() {
403 // Handle symbolic names for otherwise undistinguished boolean switches:
404 const bool is_store = true;
405 const bool is_native_ptr = true;
406 const bool is_static = true;
408 switch (intrinsic_id()) {
409 case vmIntrinsics::_hashCode:
410 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
411 case vmIntrinsics::_identityHashCode:
412 return inline_native_hashcode(/*!virtual*/ false, is_static);
413 case vmIntrinsics::_getClass:
414 return inline_native_getClass();
416 case vmIntrinsics::_dsin:
417 case vmIntrinsics::_dcos:
418 case vmIntrinsics::_dtan:
419 case vmIntrinsics::_dabs:
420 case vmIntrinsics::_datan2:
421 case vmIntrinsics::_dsqrt:
422 case vmIntrinsics::_dexp:
423 case vmIntrinsics::_dlog:
424 case vmIntrinsics::_dlog10:
425 case vmIntrinsics::_dpow:
426 return inline_math_native(intrinsic_id());
428 case vmIntrinsics::_min:
429 case vmIntrinsics::_max:
430 return inline_min_max(intrinsic_id());
432 case vmIntrinsics::_arraycopy:
433 return inline_arraycopy();
435 case vmIntrinsics::_compareTo:
436 return inline_string_compareTo();
437 case vmIntrinsics::_indexOf:
438 return inline_string_indexOf();
440 case vmIntrinsics::_getObject:
441 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
442 case vmIntrinsics::_getBoolean:
443 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
444 case vmIntrinsics::_getByte:
445 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
446 case vmIntrinsics::_getShort:
447 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
448 case vmIntrinsics::_getChar:
449 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
450 case vmIntrinsics::_getInt:
451 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
452 case vmIntrinsics::_getLong:
453 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
454 case vmIntrinsics::_getFloat:
455 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
456 case vmIntrinsics::_getDouble:
457 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
459 case vmIntrinsics::_putObject:
460 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
461 case vmIntrinsics::_putBoolean:
462 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
463 case vmIntrinsics::_putByte:
464 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
465 case vmIntrinsics::_putShort:
466 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
467 case vmIntrinsics::_putChar:
468 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
469 case vmIntrinsics::_putInt:
470 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
471 case vmIntrinsics::_putLong:
472 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
473 case vmIntrinsics::_putFloat:
474 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
475 case vmIntrinsics::_putDouble:
476 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
478 case vmIntrinsics::_getByte_raw:
479 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
480 case vmIntrinsics::_getShort_raw:
481 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
482 case vmIntrinsics::_getChar_raw:
483 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
484 case vmIntrinsics::_getInt_raw:
485 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
486 case vmIntrinsics::_getLong_raw:
487 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
488 case vmIntrinsics::_getFloat_raw:
489 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
490 case vmIntrinsics::_getDouble_raw:
491 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
492 case vmIntrinsics::_getAddress_raw:
493 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
495 case vmIntrinsics::_putByte_raw:
496 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
497 case vmIntrinsics::_putShort_raw:
498 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
499 case vmIntrinsics::_putChar_raw:
500 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
501 case vmIntrinsics::_putInt_raw:
502 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
503 case vmIntrinsics::_putLong_raw:
504 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
505 case vmIntrinsics::_putFloat_raw:
506 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
507 case vmIntrinsics::_putDouble_raw:
508 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
509 case vmIntrinsics::_putAddress_raw:
510 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
512 case vmIntrinsics::_getObjectVolatile:
513 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
514 case vmIntrinsics::_getBooleanVolatile:
515 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
516 case vmIntrinsics::_getByteVolatile:
517 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
518 case vmIntrinsics::_getShortVolatile:
519 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
520 case vmIntrinsics::_getCharVolatile:
521 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
522 case vmIntrinsics::_getIntVolatile:
523 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
524 case vmIntrinsics::_getLongVolatile:
525 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
526 case vmIntrinsics::_getFloatVolatile:
527 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
528 case vmIntrinsics::_getDoubleVolatile:
529 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
531 case vmIntrinsics::_putObjectVolatile:
532 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
533 case vmIntrinsics::_putBooleanVolatile:
534 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
535 case vmIntrinsics::_putByteVolatile:
536 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
537 case vmIntrinsics::_putShortVolatile:
538 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
539 case vmIntrinsics::_putCharVolatile:
540 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
541 case vmIntrinsics::_putIntVolatile:
542 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
543 case vmIntrinsics::_putLongVolatile:
544 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
545 case vmIntrinsics::_putFloatVolatile:
546 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
547 case vmIntrinsics::_putDoubleVolatile:
548 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
550 case vmIntrinsics::_prefetchRead:
551 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
552 case vmIntrinsics::_prefetchWrite:
553 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
554 case vmIntrinsics::_prefetchReadStatic:
555 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
556 case vmIntrinsics::_prefetchWriteStatic:
557 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
559 case vmIntrinsics::_compareAndSwapObject:
560 return inline_unsafe_CAS(T_OBJECT);
561 case vmIntrinsics::_compareAndSwapInt:
562 return inline_unsafe_CAS(T_INT);
563 case vmIntrinsics::_compareAndSwapLong:
564 return inline_unsafe_CAS(T_LONG);
566 case vmIntrinsics::_putOrderedObject:
567 return inline_unsafe_ordered_store(T_OBJECT);
568 case vmIntrinsics::_putOrderedInt:
569 return inline_unsafe_ordered_store(T_INT);
570 case vmIntrinsics::_putOrderedLong:
571 return inline_unsafe_ordered_store(T_LONG);
573 case vmIntrinsics::_currentThread:
574 return inline_native_currentThread();
575 case vmIntrinsics::_isInterrupted:
576 return inline_native_isInterrupted();
578 case vmIntrinsics::_currentTimeMillis:
579 return inline_native_time_funcs(false);
580 case vmIntrinsics::_nanoTime:
581 return inline_native_time_funcs(true);
582 case vmIntrinsics::_allocateInstance:
583 return inline_unsafe_allocate();
584 case vmIntrinsics::_copyMemory:
585 return inline_unsafe_copyMemory();
586 case vmIntrinsics::_newArray:
587 return inline_native_newArray();
588 case vmIntrinsics::_getLength:
589 return inline_native_getLength();
590 case vmIntrinsics::_copyOf:
591 return inline_array_copyOf(false);
592 case vmIntrinsics::_copyOfRange:
593 return inline_array_copyOf(true);
594 case vmIntrinsics::_equalsC:
595 return inline_array_equals();
596 case vmIntrinsics::_clone:
597 return inline_native_clone(intrinsic()->is_virtual());
599 case vmIntrinsics::_isAssignableFrom:
600 return inline_native_subtype_check();
602 case vmIntrinsics::_isInstance:
603 case vmIntrinsics::_getModifiers:
604 case vmIntrinsics::_isInterface:
605 case vmIntrinsics::_isArray:
606 case vmIntrinsics::_isPrimitive:
607 case vmIntrinsics::_getSuperclass:
608 case vmIntrinsics::_getComponentType:
609 case vmIntrinsics::_getClassAccessFlags:
610 return inline_native_Class_query(intrinsic_id());
612 case vmIntrinsics::_floatToRawIntBits:
613 case vmIntrinsics::_floatToIntBits:
614 case vmIntrinsics::_intBitsToFloat:
615 case vmIntrinsics::_doubleToRawLongBits:
616 case vmIntrinsics::_doubleToLongBits:
617 case vmIntrinsics::_longBitsToDouble:
618 return inline_fp_conversions(intrinsic_id());
620 case vmIntrinsics::_reverseBytes_i:
621 case vmIntrinsics::_reverseBytes_l:
622 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
624 case vmIntrinsics::_get_AtomicLong:
625 return inline_native_AtomicLong_get();
626 case vmIntrinsics::_attemptUpdate:
627 return inline_native_AtomicLong_attemptUpdate();
629 case vmIntrinsics::_getCallerClass:
630 return inline_native_Reflection_getCallerClass();
632 default:
633 // If you get here, it may be that someone has added a new intrinsic
634 // to the list in vmSymbols.hpp without implementing it here.
635 #ifndef PRODUCT
636 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
637 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
638 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
639 }
640 #endif
641 return false;
642 }
643 }
645 //------------------------------push_result------------------------------
646 // Helper function for finishing intrinsics.
647 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
648 record_for_igvn(region);
649 set_control(_gvn.transform(region));
650 BasicType value_type = value->type()->basic_type();
651 push_node(value_type, _gvn.transform(value));
652 }
654 //------------------------------generate_guard---------------------------
655 // Helper function for generating guarded fast-slow graph structures.
656 // The given 'test', if true, guards a slow path. If the test fails
657 // then a fast path can be taken. (We generally hope it fails.)
658 // In all cases, GraphKit::control() is updated to the fast path.
659 // The returned value represents the control for the slow path.
660 // The return value is never 'top'; it is either a valid control
661 // or NULL if it is obvious that the slow path can never be taken.
662 // Also, if region and the slow control are not NULL, the slow edge
663 // is appended to the region.
664 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
665 if (stopped()) {
666 // Already short circuited.
667 return NULL;
668 }
670 // Build an if node and its projections.
671 // If test is true we take the slow path, which we assume is uncommon.
672 if (_gvn.type(test) == TypeInt::ZERO) {
673 // The slow branch is never taken. No need to build this guard.
674 return NULL;
675 }
677 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
679 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
680 if (if_slow == top()) {
681 // The slow branch is never taken. No need to build this guard.
682 return NULL;
683 }
685 if (region != NULL)
686 region->add_req(if_slow);
688 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
689 set_control(if_fast);
691 return if_slow;
692 }
694 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
695 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
696 }
697 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
698 return generate_guard(test, region, PROB_FAIR);
699 }
701 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
702 Node* *pos_index) {
703 if (stopped())
704 return NULL; // already stopped
705 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
706 return NULL; // index is already adequately typed
707 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
708 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
709 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
710 if (is_neg != NULL && pos_index != NULL) {
711 // Emulate effect of Parse::adjust_map_after_if.
712 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
713 ccast->set_req(0, control());
714 (*pos_index) = _gvn.transform(ccast);
715 }
716 return is_neg;
717 }
719 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
720 Node* *pos_index) {
721 if (stopped())
722 return NULL; // already stopped
723 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
724 return NULL; // index is already adequately typed
725 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
726 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
727 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
728 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
729 if (is_notp != NULL && pos_index != NULL) {
730 // Emulate effect of Parse::adjust_map_after_if.
731 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
732 ccast->set_req(0, control());
733 (*pos_index) = _gvn.transform(ccast);
734 }
735 return is_notp;
736 }
738 // Make sure that 'position' is a valid limit index, in [0..length].
739 // There are two equivalent plans for checking this:
740 // A. (offset + copyLength) unsigned<= arrayLength
741 // B. offset <= (arrayLength - copyLength)
742 // We require that all of the values above, except for the sum and
743 // difference, are already known to be non-negative.
744 // Plan A is robust in the face of overflow, if offset and copyLength
745 // are both hugely positive.
746 //
747 // Plan B is less direct and intuitive, but it does not overflow at
748 // all, since the difference of two non-negatives is always
749 // representable. Whenever Java methods must perform the equivalent
750 // check they generally use Plan B instead of Plan A.
751 // For the moment we use Plan A.
752 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
753 Node* subseq_length,
754 Node* array_length,
755 RegionNode* region) {
756 if (stopped())
757 return NULL; // already stopped
758 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
759 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
760 return NULL; // common case of whole-array copy
761 Node* last = subseq_length;
762 if (!zero_offset) // last += offset
763 last = _gvn.transform( new (C, 3) AddINode(last, offset));
764 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
765 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
766 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
767 return is_over;
768 }
771 //--------------------------generate_current_thread--------------------
772 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
773 ciKlass* thread_klass = env()->Thread_klass();
774 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
775 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
776 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
777 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
778 tls_output = thread;
779 return threadObj;
780 }
783 //------------------------------inline_string_compareTo------------------------
784 bool LibraryCallKit::inline_string_compareTo() {
786 const int value_offset = java_lang_String::value_offset_in_bytes();
787 const int count_offset = java_lang_String::count_offset_in_bytes();
788 const int offset_offset = java_lang_String::offset_offset_in_bytes();
790 _sp += 2;
791 Node *argument = pop(); // pop non-receiver first: it was pushed second
792 Node *receiver = pop();
794 // Null check on self without removing any arguments. The argument
795 // null check technically happens in the wrong place, which can lead to
796 // invalid stack traces when string compare is inlined into a method
797 // which handles NullPointerExceptions.
798 _sp += 2;
799 receiver = do_null_check(receiver, T_OBJECT);
800 argument = do_null_check(argument, T_OBJECT);
801 _sp -= 2;
802 if (stopped()) {
803 return true;
804 }
806 ciInstanceKlass* klass = env()->String_klass();
807 const TypeInstPtr* string_type =
808 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
810 Node* compare =
811 _gvn.transform(new (C, 7) StrCompNode(
812 control(),
813 memory(TypeAryPtr::CHARS),
814 memory(string_type->add_offset(value_offset)),
815 memory(string_type->add_offset(count_offset)),
816 memory(string_type->add_offset(offset_offset)),
817 receiver,
818 argument));
819 push(compare);
820 return true;
821 }
823 //------------------------------inline_array_equals----------------------------
824 bool LibraryCallKit::inline_array_equals() {
826 _sp += 2;
827 Node *argument2 = pop();
828 Node *argument1 = pop();
830 Node* equals =
831 _gvn.transform(new (C, 3) AryEqNode(control(),
832 argument1,
833 argument2)
834 );
835 push(equals);
836 return true;
837 }
839 // Java version of String.indexOf(constant string)
840 // class StringDecl {
841 // StringDecl(char[] ca) {
842 // offset = 0;
843 // count = ca.length;
844 // value = ca;
845 // }
846 // int offset;
847 // int count;
848 // char[] value;
849 // }
850 //
851 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
852 // int targetOffset, int cache_i, int md2) {
853 // int cache = cache_i;
854 // int sourceOffset = string_object.offset;
855 // int sourceCount = string_object.count;
856 // int targetCount = target_object.length;
857 //
858 // int targetCountLess1 = targetCount - 1;
859 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
860 //
861 // char[] source = string_object.value;
862 // char[] target = target_object;
863 // int lastChar = target[targetCountLess1];
864 //
865 // outer_loop:
866 // for (int i = sourceOffset; i < sourceEnd; ) {
867 // int src = source[i + targetCountLess1];
868 // if (src == lastChar) {
869 // // With random strings and a 4-character alphabet,
870 // // reverse matching at this point sets up 0.8% fewer
871 // // frames, but (paradoxically) makes 0.3% more probes.
872 // // Since those probes are nearer the lastChar probe,
873 // // there is may be a net D$ win with reverse matching.
874 // // But, reversing loop inhibits unroll of inner loop
875 // // for unknown reason. So, does running outer loop from
876 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
877 // for (int j = 0; j < targetCountLess1; j++) {
878 // if (target[targetOffset + j] != source[i+j]) {
879 // if ((cache & (1 << source[i+j])) == 0) {
880 // if (md2 < j+1) {
881 // i += j+1;
882 // continue outer_loop;
883 // }
884 // }
885 // i += md2;
886 // continue outer_loop;
887 // }
888 // }
889 // return i - sourceOffset;
890 // }
891 // if ((cache & (1 << src)) == 0) {
892 // i += targetCountLess1;
893 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
894 // i++;
895 // }
896 // return -1;
897 // }
899 //------------------------------string_indexOf------------------------
900 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
901 jint cache_i, jint md2_i) {
903 Node* no_ctrl = NULL;
904 float likely = PROB_LIKELY(0.9);
905 float unlikely = PROB_UNLIKELY(0.9);
907 const int value_offset = java_lang_String::value_offset_in_bytes();
908 const int count_offset = java_lang_String::count_offset_in_bytes();
909 const int offset_offset = java_lang_String::offset_offset_in_bytes();
911 ciInstanceKlass* klass = env()->String_klass();
912 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
913 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
915 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
916 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
917 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
918 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
919 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
920 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
922 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
923 jint target_length = target_array->length();
924 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
925 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
927 IdealKit kit(gvn(), control(), merged_memory());
928 #define __ kit.
929 Node* zero = __ ConI(0);
930 Node* one = __ ConI(1);
931 Node* cache = __ ConI(cache_i);
932 Node* md2 = __ ConI(md2_i);
933 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
934 Node* targetCount = __ ConI(target_length);
935 Node* targetCountLess1 = __ ConI(target_length - 1);
936 Node* targetOffset = __ ConI(targetOffset_i);
937 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
939 IdealVariable rtn(kit), i(kit), j(kit); __ declares_done();
940 Node* outer_loop = __ make_label(2 /* goto */);
941 Node* return_ = __ make_label(1);
943 __ set(rtn,__ ConI(-1));
944 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
945 Node* i2 = __ AddI(__ value(i), targetCountLess1);
946 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
947 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
948 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
949 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
950 Node* tpj = __ AddI(targetOffset, __ value(j));
951 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
952 Node* ipj = __ AddI(__ value(i), __ value(j));
953 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
954 __ if_then(targ, BoolTest::ne, src2); {
955 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
956 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
957 __ increment(i, __ AddI(__ value(j), one));
958 __ goto_(outer_loop);
959 } __ end_if(); __ dead(j);
960 }__ end_if(); __ dead(j);
961 __ increment(i, md2);
962 __ goto_(outer_loop);
963 }__ end_if();
964 __ increment(j, one);
965 }__ end_loop(); __ dead(j);
966 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
967 __ goto_(return_);
968 }__ end_if();
969 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
970 __ increment(i, targetCountLess1);
971 }__ end_if();
972 __ increment(i, one);
973 __ bind(outer_loop);
974 }__ end_loop(); __ dead(i);
975 __ bind(return_);
976 __ drain_delay_transform();
978 set_control(__ ctrl());
979 Node* result = __ value(rtn);
980 #undef __
981 C->set_has_loops(true);
982 return result;
983 }
986 //------------------------------inline_string_indexOf------------------------
987 bool LibraryCallKit::inline_string_indexOf() {
989 _sp += 2;
990 Node *argument = pop(); // pop non-receiver first: it was pushed second
991 Node *receiver = pop();
993 // don't intrinsify is argument isn't a constant string.
994 if (!argument->is_Con()) {
995 return false;
996 }
997 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
998 if (str_type == NULL) {
999 return false;
1000 }
1001 ciInstanceKlass* klass = env()->String_klass();
1002 ciObject* str_const = str_type->const_oop();
1003 if (str_const == NULL || str_const->klass() != klass) {
1004 return false;
1005 }
1006 ciInstance* str = str_const->as_instance();
1007 assert(str != NULL, "must be instance");
1009 const int value_offset = java_lang_String::value_offset_in_bytes();
1010 const int count_offset = java_lang_String::count_offset_in_bytes();
1011 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1013 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1014 int o = str->field_value_by_offset(offset_offset).as_int();
1015 int c = str->field_value_by_offset(count_offset).as_int();
1016 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1018 // constant strings have no offset and count == length which
1019 // simplifies the resulting code somewhat so lets optimize for that.
1020 if (o != 0 || c != pat->length()) {
1021 return false;
1022 }
1024 // Null check on self without removing any arguments. The argument
1025 // null check technically happens in the wrong place, which can lead to
1026 // invalid stack traces when string compare is inlined into a method
1027 // which handles NullPointerExceptions.
1028 _sp += 2;
1029 receiver = do_null_check(receiver, T_OBJECT);
1030 // No null check on the argument is needed since it's a constant String oop.
1031 _sp -= 2;
1032 if (stopped()) {
1033 return true;
1034 }
1036 // The null string as a pattern always returns 0 (match at beginning of string)
1037 if (c == 0) {
1038 push(intcon(0));
1039 return true;
1040 }
1042 jchar lastChar = pat->char_at(o + (c - 1));
1043 int cache = 0;
1044 int i;
1045 for (i = 0; i < c - 1; i++) {
1046 assert(i < pat->length(), "out of range");
1047 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1048 }
1050 int md2 = c;
1051 for (i = 0; i < c - 1; i++) {
1052 assert(i < pat->length(), "out of range");
1053 if (pat->char_at(o + i) == lastChar) {
1054 md2 = (c - 1) - i;
1055 }
1056 }
1058 Node* result = string_indexOf(receiver, pat, o, cache, md2);
1059 push(result);
1060 return true;
1061 }
1063 //--------------------------pop_math_arg--------------------------------
1064 // Pop a double argument to a math function from the stack
1065 // rounding it if necessary.
1066 Node * LibraryCallKit::pop_math_arg() {
1067 Node *arg = pop_pair();
1068 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1069 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1070 return arg;
1071 }
1073 //------------------------------inline_trig----------------------------------
1074 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1075 // argument reduction which will turn into a fast/slow diamond.
1076 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1077 _sp += arg_size(); // restore stack pointer
1078 Node* arg = pop_math_arg();
1079 Node* trig = NULL;
1081 switch (id) {
1082 case vmIntrinsics::_dsin:
1083 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1084 break;
1085 case vmIntrinsics::_dcos:
1086 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1087 break;
1088 case vmIntrinsics::_dtan:
1089 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1090 break;
1091 default:
1092 assert(false, "bad intrinsic was passed in");
1093 return false;
1094 }
1096 // Rounding required? Check for argument reduction!
1097 if( Matcher::strict_fp_requires_explicit_rounding ) {
1099 static const double pi_4 = 0.7853981633974483;
1100 static const double neg_pi_4 = -0.7853981633974483;
1101 // pi/2 in 80-bit extended precision
1102 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1103 // -pi/2 in 80-bit extended precision
1104 // 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};
1105 // Cutoff value for using this argument reduction technique
1106 //static const double pi_2_minus_epsilon = 1.564660403643354;
1107 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1109 // Pseudocode for sin:
1110 // if (x <= Math.PI / 4.0) {
1111 // if (x >= -Math.PI / 4.0) return fsin(x);
1112 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1113 // } else {
1114 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1115 // }
1116 // return StrictMath.sin(x);
1118 // Pseudocode for cos:
1119 // if (x <= Math.PI / 4.0) {
1120 // if (x >= -Math.PI / 4.0) return fcos(x);
1121 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1122 // } else {
1123 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1124 // }
1125 // return StrictMath.cos(x);
1127 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1128 // requires a special machine instruction to load it. Instead we'll try
1129 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1130 // probably do the math inside the SIN encoding.
1132 // Make the merge point
1133 RegionNode *r = new (C, 3) RegionNode(3);
1134 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1136 // Flatten arg so we need only 1 test
1137 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1138 // Node for PI/4 constant
1139 Node *pi4 = makecon(TypeD::make(pi_4));
1140 // Check PI/4 : abs(arg)
1141 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1142 // Check: If PI/4 < abs(arg) then go slow
1143 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1144 // Branch either way
1145 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1146 set_control(opt_iff(r,iff));
1148 // Set fast path result
1149 phi->init_req(2,trig);
1151 // Slow path - non-blocking leaf call
1152 Node* call = NULL;
1153 switch (id) {
1154 case vmIntrinsics::_dsin:
1155 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1156 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1157 "Sin", NULL, arg, top());
1158 break;
1159 case vmIntrinsics::_dcos:
1160 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1161 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1162 "Cos", NULL, arg, top());
1163 break;
1164 case vmIntrinsics::_dtan:
1165 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1166 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1167 "Tan", NULL, arg, top());
1168 break;
1169 }
1170 assert(control()->in(0) == call, "");
1171 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1172 r->init_req(1,control());
1173 phi->init_req(1,slow_result);
1175 // Post-merge
1176 set_control(_gvn.transform(r));
1177 record_for_igvn(r);
1178 trig = _gvn.transform(phi);
1180 C->set_has_split_ifs(true); // Has chance for split-if optimization
1181 }
1182 // Push result back on JVM stack
1183 push_pair(trig);
1184 return true;
1185 }
1187 //------------------------------inline_sqrt-------------------------------------
1188 // Inline square root instruction, if possible.
1189 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1190 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1191 _sp += arg_size(); // restore stack pointer
1192 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1193 return true;
1194 }
1196 //------------------------------inline_abs-------------------------------------
1197 // Inline absolute value instruction, if possible.
1198 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1199 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1200 _sp += arg_size(); // restore stack pointer
1201 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1202 return true;
1203 }
1205 //------------------------------inline_exp-------------------------------------
1206 // Inline exp instructions, if possible. The Intel hardware only misses
1207 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1208 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1209 assert(id == vmIntrinsics::_dexp, "Not exp");
1211 // If this inlining ever returned NaN in the past, we do not intrinsify it
1212 // every again. NaN results requires StrictMath.exp handling.
1213 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1215 // Do not intrinsify on older platforms which lack cmove.
1216 if (ConditionalMoveLimit == 0) return false;
1218 _sp += arg_size(); // restore stack pointer
1219 Node *x = pop_math_arg();
1220 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1222 //-------------------
1223 //result=(result.isNaN())? StrictMath::exp():result;
1224 // Check: If isNaN() by checking result!=result? then go to Strict Math
1225 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1226 // Build the boolean node
1227 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1229 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1230 // End the current control-flow path
1231 push_pair(x);
1232 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1233 // to handle. Recompile without intrinsifying Math.exp
1234 uncommon_trap(Deoptimization::Reason_intrinsic,
1235 Deoptimization::Action_make_not_entrant);
1236 }
1238 C->set_has_split_ifs(true); // Has chance for split-if optimization
1240 push_pair(result);
1242 return true;
1243 }
1245 //------------------------------inline_pow-------------------------------------
1246 // Inline power instructions, if possible.
1247 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1248 assert(id == vmIntrinsics::_dpow, "Not pow");
1250 // If this inlining ever returned NaN in the past, we do not intrinsify it
1251 // every again. NaN results requires StrictMath.pow handling.
1252 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1254 // Do not intrinsify on older platforms which lack cmove.
1255 if (ConditionalMoveLimit == 0) return false;
1257 // Pseudocode for pow
1258 // if (x <= 0.0) {
1259 // if ((double)((int)y)==y) { // if y is int
1260 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1261 // } else {
1262 // result = NaN;
1263 // }
1264 // } else {
1265 // result = DPow(x,y);
1266 // }
1267 // if (result != result)? {
1268 // ucommon_trap();
1269 // }
1270 // return result;
1272 _sp += arg_size(); // restore stack pointer
1273 Node* y = pop_math_arg();
1274 Node* x = pop_math_arg();
1276 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1278 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1279 // inside of something) then skip the fancy tests and just check for
1280 // NaN result.
1281 Node *result = NULL;
1282 if( jvms()->depth() >= 1 ) {
1283 result = fast_result;
1284 } else {
1286 // Set the merge point for If node with condition of (x <= 0.0)
1287 // There are four possible paths to region node and phi node
1288 RegionNode *r = new (C, 4) RegionNode(4);
1289 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1291 // Build the first if node: if (x <= 0.0)
1292 // Node for 0 constant
1293 Node *zeronode = makecon(TypeD::ZERO);
1294 // Check x:0
1295 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1296 // Check: If (x<=0) then go complex path
1297 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1298 // Branch either way
1299 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1300 Node *opt_test = _gvn.transform(if1);
1301 //assert( opt_test->is_If(), "Expect an IfNode");
1302 IfNode *opt_if1 = (IfNode*)opt_test;
1303 // Fast path taken; set region slot 3
1304 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1305 r->init_req(3,fast_taken); // Capture fast-control
1307 // Fast path not-taken, i.e. slow path
1308 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1310 // Set fast path result
1311 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1312 phi->init_req(3, fast_result);
1314 // Complex path
1315 // Build the second if node (if y is int)
1316 // Node for (int)y
1317 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1318 // Node for (double)((int) y)
1319 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1320 // Check (double)((int) y) : y
1321 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1322 // Check if (y isn't int) then go to slow path
1324 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1325 // Branch eith way
1326 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1327 Node *slow_path = opt_iff(r,if2); // Set region path 2
1329 // Calculate DPow(abs(x), y)*(1 & (int)y)
1330 // Node for constant 1
1331 Node *conone = intcon(1);
1332 // 1& (int)y
1333 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1334 // zero node
1335 Node *conzero = intcon(0);
1336 // Check (1&(int)y)==0?
1337 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1338 // Check if (1&(int)y)!=0?, if so the result is negative
1339 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1340 // abs(x)
1341 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1342 // abs(x)^y
1343 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1344 // -abs(x)^y
1345 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1346 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1347 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1348 // Set complex path fast result
1349 phi->init_req(2, signresult);
1351 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1352 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1353 r->init_req(1,slow_path);
1354 phi->init_req(1,slow_result);
1356 // Post merge
1357 set_control(_gvn.transform(r));
1358 record_for_igvn(r);
1359 result=_gvn.transform(phi);
1360 }
1362 //-------------------
1363 //result=(result.isNaN())? uncommon_trap():result;
1364 // Check: If isNaN() by checking result!=result? then go to Strict Math
1365 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1366 // Build the boolean node
1367 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1369 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1370 // End the current control-flow path
1371 push_pair(x);
1372 push_pair(y);
1373 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1374 // to handle. Recompile without intrinsifying Math.pow.
1375 uncommon_trap(Deoptimization::Reason_intrinsic,
1376 Deoptimization::Action_make_not_entrant);
1377 }
1379 C->set_has_split_ifs(true); // Has chance for split-if optimization
1381 push_pair(result);
1383 return true;
1384 }
1386 //------------------------------inline_trans-------------------------------------
1387 // Inline transcendental instructions, if possible. The Intel hardware gets
1388 // these right, no funny corner cases missed.
1389 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1390 _sp += arg_size(); // restore stack pointer
1391 Node* arg = pop_math_arg();
1392 Node* trans = NULL;
1394 switch (id) {
1395 case vmIntrinsics::_dlog:
1396 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1397 break;
1398 case vmIntrinsics::_dlog10:
1399 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1400 break;
1401 default:
1402 assert(false, "bad intrinsic was passed in");
1403 return false;
1404 }
1406 // Push result back on JVM stack
1407 push_pair(trans);
1408 return true;
1409 }
1411 //------------------------------runtime_math-----------------------------
1412 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1413 Node* a = NULL;
1414 Node* b = NULL;
1416 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1417 "must be (DD)D or (D)D type");
1419 // Inputs
1420 _sp += arg_size(); // restore stack pointer
1421 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1422 b = pop_math_arg();
1423 }
1424 a = pop_math_arg();
1426 const TypePtr* no_memory_effects = NULL;
1427 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1428 no_memory_effects,
1429 a, top(), b, b ? top() : NULL);
1430 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1431 #ifdef ASSERT
1432 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1433 assert(value_top == top(), "second value must be top");
1434 #endif
1436 push_pair(value);
1437 return true;
1438 }
1440 //------------------------------inline_math_native-----------------------------
1441 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1442 switch (id) {
1443 // These intrinsics are not properly supported on all hardware
1444 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1445 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1446 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1447 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1448 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1449 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1451 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1452 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1453 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1454 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1456 // These intrinsics are supported on all hardware
1457 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1458 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1460 // These intrinsics don't work on X86. The ad implementation doesn't
1461 // handle NaN's properly. Instead of returning infinity, the ad
1462 // implementation returns a NaN on overflow. See bug: 6304089
1463 // Once the ad implementations are fixed, change the code below
1464 // to match the intrinsics above
1466 case vmIntrinsics::_dexp: return
1467 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1468 case vmIntrinsics::_dpow: return
1469 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1471 // These intrinsics are not yet correctly implemented
1472 case vmIntrinsics::_datan2:
1473 return false;
1475 default:
1476 ShouldNotReachHere();
1477 return false;
1478 }
1479 }
1481 static bool is_simple_name(Node* n) {
1482 return (n->req() == 1 // constant
1483 || (n->is_Type() && n->as_Type()->type()->singleton())
1484 || n->is_Proj() // parameter or return value
1485 || n->is_Phi() // local of some sort
1486 );
1487 }
1489 //----------------------------inline_min_max-----------------------------------
1490 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1491 push(generate_min_max(id, argument(0), argument(1)));
1493 return true;
1494 }
1496 Node*
1497 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1498 // These are the candidate return value:
1499 Node* xvalue = x0;
1500 Node* yvalue = y0;
1502 if (xvalue == yvalue) {
1503 return xvalue;
1504 }
1506 bool want_max = (id == vmIntrinsics::_max);
1508 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1509 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1510 if (txvalue == NULL || tyvalue == NULL) return top();
1511 // This is not really necessary, but it is consistent with a
1512 // hypothetical MaxINode::Value method:
1513 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1515 // %%% This folding logic should (ideally) be in a different place.
1516 // Some should be inside IfNode, and there to be a more reliable
1517 // transformation of ?: style patterns into cmoves. We also want
1518 // more powerful optimizations around cmove and min/max.
1520 // Try to find a dominating comparison of these guys.
1521 // It can simplify the index computation for Arrays.copyOf
1522 // and similar uses of System.arraycopy.
1523 // First, compute the normalized version of CmpI(x, y).
1524 int cmp_op = Op_CmpI;
1525 Node* xkey = xvalue;
1526 Node* ykey = yvalue;
1527 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1528 if (ideal_cmpxy->is_Cmp()) {
1529 // E.g., if we have CmpI(length - offset, count),
1530 // it might idealize to CmpI(length, count + offset)
1531 cmp_op = ideal_cmpxy->Opcode();
1532 xkey = ideal_cmpxy->in(1);
1533 ykey = ideal_cmpxy->in(2);
1534 }
1536 // Start by locating any relevant comparisons.
1537 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1538 Node* cmpxy = NULL;
1539 Node* cmpyx = NULL;
1540 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1541 Node* cmp = start_from->fast_out(k);
1542 if (cmp->outcnt() > 0 && // must have prior uses
1543 cmp->in(0) == NULL && // must be context-independent
1544 cmp->Opcode() == cmp_op) { // right kind of compare
1545 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1546 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1547 }
1548 }
1550 const int NCMPS = 2;
1551 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1552 int cmpn;
1553 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1554 if (cmps[cmpn] != NULL) break; // find a result
1555 }
1556 if (cmpn < NCMPS) {
1557 // Look for a dominating test that tells us the min and max.
1558 int depth = 0; // Limit search depth for speed
1559 Node* dom = control();
1560 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1561 if (++depth >= 100) break;
1562 Node* ifproj = dom;
1563 if (!ifproj->is_Proj()) continue;
1564 Node* iff = ifproj->in(0);
1565 if (!iff->is_If()) continue;
1566 Node* bol = iff->in(1);
1567 if (!bol->is_Bool()) continue;
1568 Node* cmp = bol->in(1);
1569 if (cmp == NULL) continue;
1570 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1571 if (cmps[cmpn] == cmp) break;
1572 if (cmpn == NCMPS) continue;
1573 BoolTest::mask btest = bol->as_Bool()->_test._test;
1574 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1575 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1576 // At this point, we know that 'x btest y' is true.
1577 switch (btest) {
1578 case BoolTest::eq:
1579 // They are proven equal, so we can collapse the min/max.
1580 // Either value is the answer. Choose the simpler.
1581 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1582 return yvalue;
1583 return xvalue;
1584 case BoolTest::lt: // x < y
1585 case BoolTest::le: // x <= y
1586 return (want_max ? yvalue : xvalue);
1587 case BoolTest::gt: // x > y
1588 case BoolTest::ge: // x >= y
1589 return (want_max ? xvalue : yvalue);
1590 }
1591 }
1592 }
1594 // We failed to find a dominating test.
1595 // Let's pick a test that might GVN with prior tests.
1596 Node* best_bol = NULL;
1597 BoolTest::mask best_btest = BoolTest::illegal;
1598 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1599 Node* cmp = cmps[cmpn];
1600 if (cmp == NULL) continue;
1601 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1602 Node* bol = cmp->fast_out(j);
1603 if (!bol->is_Bool()) continue;
1604 BoolTest::mask btest = bol->as_Bool()->_test._test;
1605 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1606 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1607 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1608 best_bol = bol->as_Bool();
1609 best_btest = btest;
1610 }
1611 }
1612 }
1614 Node* answer_if_true = NULL;
1615 Node* answer_if_false = NULL;
1616 switch (best_btest) {
1617 default:
1618 if (cmpxy == NULL)
1619 cmpxy = ideal_cmpxy;
1620 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1621 // and fall through:
1622 case BoolTest::lt: // x < y
1623 case BoolTest::le: // x <= y
1624 answer_if_true = (want_max ? yvalue : xvalue);
1625 answer_if_false = (want_max ? xvalue : yvalue);
1626 break;
1627 case BoolTest::gt: // x > y
1628 case BoolTest::ge: // x >= y
1629 answer_if_true = (want_max ? xvalue : yvalue);
1630 answer_if_false = (want_max ? yvalue : xvalue);
1631 break;
1632 }
1634 jint hi, lo;
1635 if (want_max) {
1636 // We can sharpen the minimum.
1637 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1638 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1639 } else {
1640 // We can sharpen the maximum.
1641 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1642 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1643 }
1645 // Use a flow-free graph structure, to avoid creating excess control edges
1646 // which could hinder other optimizations.
1647 // Since Math.min/max is often used with arraycopy, we want
1648 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1649 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1650 answer_if_false, answer_if_true,
1651 TypeInt::make(lo, hi, widen));
1653 return _gvn.transform(cmov);
1655 /*
1656 // This is not as desirable as it may seem, since Min and Max
1657 // nodes do not have a full set of optimizations.
1658 // And they would interfere, anyway, with 'if' optimizations
1659 // and with CMoveI canonical forms.
1660 switch (id) {
1661 case vmIntrinsics::_min:
1662 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1663 case vmIntrinsics::_max:
1664 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1665 default:
1666 ShouldNotReachHere();
1667 }
1668 */
1669 }
1671 inline int
1672 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1673 const TypePtr* base_type = TypePtr::NULL_PTR;
1674 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1675 if (base_type == NULL) {
1676 // Unknown type.
1677 return Type::AnyPtr;
1678 } else if (base_type == TypePtr::NULL_PTR) {
1679 // Since this is a NULL+long form, we have to switch to a rawptr.
1680 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1681 offset = MakeConX(0);
1682 return Type::RawPtr;
1683 } else if (base_type->base() == Type::RawPtr) {
1684 return Type::RawPtr;
1685 } else if (base_type->isa_oopptr()) {
1686 // Base is never null => always a heap address.
1687 if (base_type->ptr() == TypePtr::NotNull) {
1688 return Type::OopPtr;
1689 }
1690 // Offset is small => always a heap address.
1691 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1692 if (offset_type != NULL &&
1693 base_type->offset() == 0 && // (should always be?)
1694 offset_type->_lo >= 0 &&
1695 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1696 return Type::OopPtr;
1697 }
1698 // Otherwise, it might either be oop+off or NULL+addr.
1699 return Type::AnyPtr;
1700 } else {
1701 // No information:
1702 return Type::AnyPtr;
1703 }
1704 }
1706 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1707 int kind = classify_unsafe_addr(base, offset);
1708 if (kind == Type::RawPtr) {
1709 return basic_plus_adr(top(), base, offset);
1710 } else {
1711 return basic_plus_adr(base, offset);
1712 }
1713 }
1715 //----------------------------inline_reverseBytes_int/long-------------------
1716 // inline Int.reverseBytes(int)
1717 // inline Long.reverseByes(long)
1718 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
1719 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
1720 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
1721 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
1722 _sp += arg_size(); // restore stack pointer
1723 switch (id) {
1724 case vmIntrinsics::_reverseBytes_i:
1725 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
1726 break;
1727 case vmIntrinsics::_reverseBytes_l:
1728 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
1729 break;
1730 default:
1731 ;
1732 }
1733 return true;
1734 }
1736 //----------------------------inline_unsafe_access----------------------------
1738 const static BasicType T_ADDRESS_HOLDER = T_LONG;
1740 // Interpret Unsafe.fieldOffset cookies correctly:
1741 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
1743 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
1744 if (callee()->is_static()) return false; // caller must have the capability!
1746 #ifndef PRODUCT
1747 {
1748 ResourceMark rm;
1749 // Check the signatures.
1750 ciSignature* sig = signature();
1751 #ifdef ASSERT
1752 if (!is_store) {
1753 // Object getObject(Object base, int/long offset), etc.
1754 BasicType rtype = sig->return_type()->basic_type();
1755 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
1756 rtype = T_ADDRESS; // it is really a C void*
1757 assert(rtype == type, "getter must return the expected value");
1758 if (!is_native_ptr) {
1759 assert(sig->count() == 2, "oop getter has 2 arguments");
1760 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
1761 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
1762 } else {
1763 assert(sig->count() == 1, "native getter has 1 argument");
1764 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
1765 }
1766 } else {
1767 // void putObject(Object base, int/long offset, Object x), etc.
1768 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
1769 if (!is_native_ptr) {
1770 assert(sig->count() == 3, "oop putter has 3 arguments");
1771 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
1772 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
1773 } else {
1774 assert(sig->count() == 2, "native putter has 2 arguments");
1775 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
1776 }
1777 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
1778 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
1779 vtype = T_ADDRESS; // it is really a C void*
1780 assert(vtype == type, "putter must accept the expected value");
1781 }
1782 #endif // ASSERT
1783 }
1784 #endif //PRODUCT
1786 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1788 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
1790 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
1791 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
1793 debug_only(int saved_sp = _sp);
1794 _sp += nargs;
1796 Node* val;
1797 debug_only(val = (Node*)(uintptr_t)-1);
1800 if (is_store) {
1801 // Get the value being stored. (Pop it first; it was pushed last.)
1802 switch (type) {
1803 case T_DOUBLE:
1804 case T_LONG:
1805 case T_ADDRESS:
1806 val = pop_pair();
1807 break;
1808 default:
1809 val = pop();
1810 }
1811 }
1813 // Build address expression. See the code in inline_unsafe_prefetch.
1814 Node *adr;
1815 Node *heap_base_oop = top();
1816 if (!is_native_ptr) {
1817 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
1818 Node* offset = pop_pair();
1819 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
1820 Node* base = pop();
1821 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1822 // to be plain byte offsets, which are also the same as those accepted
1823 // by oopDesc::field_base.
1824 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1825 "fieldOffset must be byte-scaled");
1826 // 32-bit machines ignore the high half!
1827 offset = ConvL2X(offset);
1828 adr = make_unsafe_address(base, offset);
1829 heap_base_oop = base;
1830 } else {
1831 Node* ptr = pop_pair();
1832 // Adjust Java long to machine word:
1833 ptr = ConvL2X(ptr);
1834 adr = make_unsafe_address(NULL, ptr);
1835 }
1837 // Pop receiver last: it was pushed first.
1838 Node *receiver = pop();
1840 assert(saved_sp == _sp, "must have correct argument count");
1842 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1844 // First guess at the value type.
1845 const Type *value_type = Type::get_const_basic_type(type);
1847 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
1848 // there was not enough information to nail it down.
1849 Compile::AliasType* alias_type = C->alias_type(adr_type);
1850 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
1852 // We will need memory barriers unless we can determine a unique
1853 // alias category for this reference. (Note: If for some reason
1854 // the barriers get omitted and the unsafe reference begins to "pollute"
1855 // the alias analysis of the rest of the graph, either Compile::can_alias
1856 // or Compile::must_alias will throw a diagnostic assert.)
1857 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
1859 if (!is_store && type == T_OBJECT) {
1860 // Attempt to infer a sharper value type from the offset and base type.
1861 ciKlass* sharpened_klass = NULL;
1863 // See if it is an instance field, with an object type.
1864 if (alias_type->field() != NULL) {
1865 assert(!is_native_ptr, "native pointer op cannot use a java address");
1866 if (alias_type->field()->type()->is_klass()) {
1867 sharpened_klass = alias_type->field()->type()->as_klass();
1868 }
1869 }
1871 // See if it is a narrow oop array.
1872 if (adr_type->isa_aryptr()) {
1873 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) {
1874 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
1875 if (elem_type != NULL) {
1876 sharpened_klass = elem_type->klass();
1877 }
1878 }
1879 }
1881 if (sharpened_klass != NULL) {
1882 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
1884 // Sharpen the value type.
1885 value_type = tjp;
1887 #ifndef PRODUCT
1888 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
1889 tty->print(" from base type: "); adr_type->dump();
1890 tty->print(" sharpened value: "); value_type->dump();
1891 }
1892 #endif
1893 }
1894 }
1896 // Null check on self without removing any arguments. The argument
1897 // null check technically happens in the wrong place, which can lead to
1898 // invalid stack traces when the primitive is inlined into a method
1899 // which handles NullPointerExceptions.
1900 _sp += nargs;
1901 do_null_check(receiver, T_OBJECT);
1902 _sp -= nargs;
1903 if (stopped()) {
1904 return true;
1905 }
1906 // Heap pointers get a null-check from the interpreter,
1907 // as a courtesy. However, this is not guaranteed by Unsafe,
1908 // and it is not possible to fully distinguish unintended nulls
1909 // from intended ones in this API.
1911 if (is_volatile) {
1912 // We need to emit leading and trailing CPU membars (see below) in
1913 // addition to memory membars when is_volatile. This is a little
1914 // too strong, but avoids the need to insert per-alias-type
1915 // volatile membars (for stores; compare Parse::do_put_xxx), which
1916 // we cannot do effctively here because we probably only have a
1917 // rough approximation of type.
1918 need_mem_bar = true;
1919 // For Stores, place a memory ordering barrier now.
1920 if (is_store)
1921 insert_mem_bar(Op_MemBarRelease);
1922 }
1924 // Memory barrier to prevent normal and 'unsafe' accesses from
1925 // bypassing each other. Happens after null checks, so the
1926 // exception paths do not take memory state from the memory barrier,
1927 // so there's no problems making a strong assert about mixing users
1928 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
1929 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
1930 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
1932 if (!is_store) {
1933 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
1934 // load value and push onto stack
1935 switch (type) {
1936 case T_BOOLEAN:
1937 case T_CHAR:
1938 case T_BYTE:
1939 case T_SHORT:
1940 case T_INT:
1941 case T_FLOAT:
1942 case T_OBJECT:
1943 push( p );
1944 break;
1945 case T_ADDRESS:
1946 // Cast to an int type.
1947 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
1948 p = ConvX2L(p);
1949 push_pair(p);
1950 break;
1951 case T_DOUBLE:
1952 case T_LONG:
1953 push_pair( p );
1954 break;
1955 default: ShouldNotReachHere();
1956 }
1957 } else {
1958 // place effect of store into memory
1959 switch (type) {
1960 case T_DOUBLE:
1961 val = dstore_rounding(val);
1962 break;
1963 case T_ADDRESS:
1964 // Repackage the long as a pointer.
1965 val = ConvL2X(val);
1966 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
1967 break;
1968 }
1970 if (type != T_OBJECT ) {
1971 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
1972 } else {
1973 // Possibly an oop being stored to Java heap or native memory
1974 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
1975 // oop to Java heap.
1976 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
1977 } else {
1979 // We can't tell at compile time if we are storing in the Java heap or outside
1980 // of it. So we need to emit code to conditionally do the proper type of
1981 // store.
1983 IdealKit kit(gvn(), control(), merged_memory());
1984 kit.declares_done();
1985 // QQQ who knows what probability is here??
1986 kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
1987 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
1988 } kit.else_(); {
1989 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
1990 } kit.end_if();
1991 }
1992 }
1993 }
1995 if (is_volatile) {
1996 if (!is_store)
1997 insert_mem_bar(Op_MemBarAcquire);
1998 else
1999 insert_mem_bar(Op_MemBarVolatile);
2000 }
2002 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2004 return true;
2005 }
2007 //----------------------------inline_unsafe_prefetch----------------------------
2009 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2010 #ifndef PRODUCT
2011 {
2012 ResourceMark rm;
2013 // Check the signatures.
2014 ciSignature* sig = signature();
2015 #ifdef ASSERT
2016 // Object getObject(Object base, int/long offset), etc.
2017 BasicType rtype = sig->return_type()->basic_type();
2018 if (!is_native_ptr) {
2019 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2020 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2021 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2022 } else {
2023 assert(sig->count() == 1, "native prefetch has 1 argument");
2024 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2025 }
2026 #endif // ASSERT
2027 }
2028 #endif // !PRODUCT
2030 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2032 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2033 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2035 debug_only(int saved_sp = _sp);
2036 _sp += nargs;
2038 // Build address expression. See the code in inline_unsafe_access.
2039 Node *adr;
2040 if (!is_native_ptr) {
2041 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2042 Node* offset = pop_pair();
2043 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2044 Node* base = pop();
2045 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2046 // to be plain byte offsets, which are also the same as those accepted
2047 // by oopDesc::field_base.
2048 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2049 "fieldOffset must be byte-scaled");
2050 // 32-bit machines ignore the high half!
2051 offset = ConvL2X(offset);
2052 adr = make_unsafe_address(base, offset);
2053 } else {
2054 Node* ptr = pop_pair();
2055 // Adjust Java long to machine word:
2056 ptr = ConvL2X(ptr);
2057 adr = make_unsafe_address(NULL, ptr);
2058 }
2060 if (is_static) {
2061 assert(saved_sp == _sp, "must have correct argument count");
2062 } else {
2063 // Pop receiver last: it was pushed first.
2064 Node *receiver = pop();
2065 assert(saved_sp == _sp, "must have correct argument count");
2067 // Null check on self without removing any arguments. The argument
2068 // null check technically happens in the wrong place, which can lead to
2069 // invalid stack traces when the primitive is inlined into a method
2070 // which handles NullPointerExceptions.
2071 _sp += nargs;
2072 do_null_check(receiver, T_OBJECT);
2073 _sp -= nargs;
2074 if (stopped()) {
2075 return true;
2076 }
2077 }
2079 // Generate the read or write prefetch
2080 Node *prefetch;
2081 if (is_store) {
2082 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2083 } else {
2084 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2085 }
2086 prefetch->init_req(0, control());
2087 set_i_o(_gvn.transform(prefetch));
2089 return true;
2090 }
2092 //----------------------------inline_unsafe_CAS----------------------------
2094 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2095 // This basic scheme here is the same as inline_unsafe_access, but
2096 // differs in enough details that combining them would make the code
2097 // overly confusing. (This is a true fact! I originally combined
2098 // them, but even I was confused by it!) As much code/comments as
2099 // possible are retained from inline_unsafe_access though to make
2100 // the correspondances clearer. - dl
2102 if (callee()->is_static()) return false; // caller must have the capability!
2104 #ifndef PRODUCT
2105 {
2106 ResourceMark rm;
2107 // Check the signatures.
2108 ciSignature* sig = signature();
2109 #ifdef ASSERT
2110 BasicType rtype = sig->return_type()->basic_type();
2111 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2112 assert(sig->count() == 4, "CAS has 4 arguments");
2113 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2114 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2115 #endif // ASSERT
2116 }
2117 #endif //PRODUCT
2119 // number of stack slots per value argument (1 or 2)
2120 int type_words = type2size[type];
2122 // Cannot inline wide CAS on machines that don't support it natively
2123 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2124 return false;
2126 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2128 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2129 int nargs = 1 + 1 + 2 + type_words + type_words;
2131 // pop arguments: newval, oldval, offset, base, and receiver
2132 debug_only(int saved_sp = _sp);
2133 _sp += nargs;
2134 Node* newval = (type_words == 1) ? pop() : pop_pair();
2135 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2136 Node *offset = pop_pair();
2137 Node *base = pop();
2138 Node *receiver = pop();
2139 assert(saved_sp == _sp, "must have correct argument count");
2141 // Null check receiver.
2142 _sp += nargs;
2143 do_null_check(receiver, T_OBJECT);
2144 _sp -= nargs;
2145 if (stopped()) {
2146 return true;
2147 }
2149 // Build field offset expression.
2150 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2151 // to be plain byte offsets, which are also the same as those accepted
2152 // by oopDesc::field_base.
2153 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2154 // 32-bit machines ignore the high half of long offsets
2155 offset = ConvL2X(offset);
2156 Node* adr = make_unsafe_address(base, offset);
2157 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2159 // (Unlike inline_unsafe_access, there seems no point in trying
2160 // to refine types. Just use the coarse types here.
2161 const Type *value_type = Type::get_const_basic_type(type);
2162 Compile::AliasType* alias_type = C->alias_type(adr_type);
2163 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2164 int alias_idx = C->get_alias_index(adr_type);
2166 // Memory-model-wise, a CAS acts like a little synchronized block,
2167 // so needs barriers on each side. These don't't translate into
2168 // actual barriers on most machines, but we still need rest of
2169 // compiler to respect ordering.
2171 insert_mem_bar(Op_MemBarRelease);
2172 insert_mem_bar(Op_MemBarCPUOrder);
2174 // 4984716: MemBars must be inserted before this
2175 // memory node in order to avoid a false
2176 // dependency which will confuse the scheduler.
2177 Node *mem = memory(alias_idx);
2179 // For now, we handle only those cases that actually exist: ints,
2180 // longs, and Object. Adding others should be straightforward.
2181 Node* cas;
2182 switch(type) {
2183 case T_INT:
2184 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2185 break;
2186 case T_LONG:
2187 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2188 break;
2189 case T_OBJECT:
2190 // reference stores need a store barrier.
2191 // (They don't if CAS fails, but it isn't worth checking.)
2192 pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT);
2193 #ifdef _LP64
2194 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2195 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2196 EncodePNode::encode(&_gvn, newval),
2197 EncodePNode::encode(&_gvn, oldval)));
2198 } else
2199 #endif
2200 {
2201 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2202 }
2203 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2204 break;
2205 default:
2206 ShouldNotReachHere();
2207 break;
2208 }
2210 // SCMemProjNodes represent the memory state of CAS. Their main
2211 // role is to prevent CAS nodes from being optimized away when their
2212 // results aren't used.
2213 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2214 set_memory(proj, alias_idx);
2216 // Add the trailing membar surrounding the access
2217 insert_mem_bar(Op_MemBarCPUOrder);
2218 insert_mem_bar(Op_MemBarAcquire);
2220 push(cas);
2221 return true;
2222 }
2224 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2225 // This is another variant of inline_unsafe_access, differing in
2226 // that it always issues store-store ("release") barrier and ensures
2227 // store-atomicity (which only matters for "long").
2229 if (callee()->is_static()) return false; // caller must have the capability!
2231 #ifndef PRODUCT
2232 {
2233 ResourceMark rm;
2234 // Check the signatures.
2235 ciSignature* sig = signature();
2236 #ifdef ASSERT
2237 BasicType rtype = sig->return_type()->basic_type();
2238 assert(rtype == T_VOID, "must return void");
2239 assert(sig->count() == 3, "has 3 arguments");
2240 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2241 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2242 #endif // ASSERT
2243 }
2244 #endif //PRODUCT
2246 // number of stack slots per value argument (1 or 2)
2247 int type_words = type2size[type];
2249 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2251 // Argument words: "this" plus oop plus offset plus value;
2252 int nargs = 1 + 1 + 2 + type_words;
2254 // pop arguments: val, offset, base, and receiver
2255 debug_only(int saved_sp = _sp);
2256 _sp += nargs;
2257 Node* val = (type_words == 1) ? pop() : pop_pair();
2258 Node *offset = pop_pair();
2259 Node *base = pop();
2260 Node *receiver = pop();
2261 assert(saved_sp == _sp, "must have correct argument count");
2263 // Null check receiver.
2264 _sp += nargs;
2265 do_null_check(receiver, T_OBJECT);
2266 _sp -= nargs;
2267 if (stopped()) {
2268 return true;
2269 }
2271 // Build field offset expression.
2272 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2273 // 32-bit machines ignore the high half of long offsets
2274 offset = ConvL2X(offset);
2275 Node* adr = make_unsafe_address(base, offset);
2276 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2277 const Type *value_type = Type::get_const_basic_type(type);
2278 Compile::AliasType* alias_type = C->alias_type(adr_type);
2280 insert_mem_bar(Op_MemBarRelease);
2281 insert_mem_bar(Op_MemBarCPUOrder);
2282 // Ensure that the store is atomic for longs:
2283 bool require_atomic_access = true;
2284 Node* store;
2285 if (type == T_OBJECT) // reference stores need a store barrier.
2286 store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type);
2287 else {
2288 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2289 }
2290 insert_mem_bar(Op_MemBarCPUOrder);
2291 return true;
2292 }
2294 bool LibraryCallKit::inline_unsafe_allocate() {
2295 if (callee()->is_static()) return false; // caller must have the capability!
2296 int nargs = 1 + 1;
2297 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2298 null_check_receiver(callee()); // check then ignore argument(0)
2299 _sp += nargs; // set original stack for use by uncommon_trap
2300 Node* cls = do_null_check(argument(1), T_OBJECT);
2301 _sp -= nargs;
2302 if (stopped()) return true;
2304 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2305 _sp += nargs; // set original stack for use by uncommon_trap
2306 kls = do_null_check(kls, T_OBJECT);
2307 _sp -= nargs;
2308 if (stopped()) return true; // argument was like int.class
2310 // Note: The argument might still be an illegal value like
2311 // Serializable.class or Object[].class. The runtime will handle it.
2312 // But we must make an explicit check for initialization.
2313 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2314 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2315 Node* bits = intcon(instanceKlass::fully_initialized);
2316 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2317 // The 'test' is non-zero if we need to take a slow path.
2319 Node* obj = new_instance(kls, test);
2320 push(obj);
2322 return true;
2323 }
2325 //------------------------inline_native_time_funcs--------------
2326 // inline code for System.currentTimeMillis() and System.nanoTime()
2327 // these have the same type and signature
2328 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2329 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2330 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2331 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2332 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2333 const TypePtr* no_memory_effects = NULL;
2334 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2335 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2336 #ifdef ASSERT
2337 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2338 assert(value_top == top(), "second value must be top");
2339 #endif
2340 push_pair(value);
2341 return true;
2342 }
2344 //------------------------inline_native_currentThread------------------
2345 bool LibraryCallKit::inline_native_currentThread() {
2346 Node* junk = NULL;
2347 push(generate_current_thread(junk));
2348 return true;
2349 }
2351 //------------------------inline_native_isInterrupted------------------
2352 bool LibraryCallKit::inline_native_isInterrupted() {
2353 const int nargs = 1+1; // receiver + boolean
2354 assert(nargs == arg_size(), "sanity");
2355 // Add a fast path to t.isInterrupted(clear_int):
2356 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2357 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2358 // So, in the common case that the interrupt bit is false,
2359 // we avoid making a call into the VM. Even if the interrupt bit
2360 // is true, if the clear_int argument is false, we avoid the VM call.
2361 // However, if the receiver is not currentThread, we must call the VM,
2362 // because there must be some locking done around the operation.
2364 // We only go to the fast case code if we pass two guards.
2365 // Paths which do not pass are accumulated in the slow_region.
2366 RegionNode* slow_region = new (C, 1) RegionNode(1);
2367 record_for_igvn(slow_region);
2368 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2369 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2370 enum { no_int_result_path = 1,
2371 no_clear_result_path = 2,
2372 slow_result_path = 3
2373 };
2375 // (a) Receiving thread must be the current thread.
2376 Node* rec_thr = argument(0);
2377 Node* tls_ptr = NULL;
2378 Node* cur_thr = generate_current_thread(tls_ptr);
2379 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2380 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2382 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2383 if (!known_current_thread)
2384 generate_slow_guard(bol_thr, slow_region);
2386 // (b) Interrupt bit on TLS must be false.
2387 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2388 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2389 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2390 Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT);
2391 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2392 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2394 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2396 // First fast path: if (!TLS._interrupted) return false;
2397 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2398 result_rgn->init_req(no_int_result_path, false_bit);
2399 result_val->init_req(no_int_result_path, intcon(0));
2401 // drop through to next case
2402 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2404 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2405 Node* clr_arg = argument(1);
2406 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2407 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2408 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2410 // Second fast path: ... else if (!clear_int) return true;
2411 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2412 result_rgn->init_req(no_clear_result_path, false_arg);
2413 result_val->init_req(no_clear_result_path, intcon(1));
2415 // drop through to next case
2416 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2418 // (d) Otherwise, go to the slow path.
2419 slow_region->add_req(control());
2420 set_control( _gvn.transform(slow_region) );
2422 if (stopped()) {
2423 // There is no slow path.
2424 result_rgn->init_req(slow_result_path, top());
2425 result_val->init_req(slow_result_path, top());
2426 } else {
2427 // non-virtual because it is a private non-static
2428 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2430 Node* slow_val = set_results_for_java_call(slow_call);
2431 // this->control() comes from set_results_for_java_call
2433 // If we know that the result of the slow call will be true, tell the optimizer!
2434 if (known_current_thread) slow_val = intcon(1);
2436 Node* fast_io = slow_call->in(TypeFunc::I_O);
2437 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2438 // These two phis are pre-filled with copies of of the fast IO and Memory
2439 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2440 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2442 result_rgn->init_req(slow_result_path, control());
2443 io_phi ->init_req(slow_result_path, i_o());
2444 mem_phi ->init_req(slow_result_path, reset_memory());
2445 result_val->init_req(slow_result_path, slow_val);
2447 set_all_memory( _gvn.transform(mem_phi) );
2448 set_i_o( _gvn.transform(io_phi) );
2449 }
2451 push_result(result_rgn, result_val);
2452 C->set_has_split_ifs(true); // Has chance for split-if optimization
2454 return true;
2455 }
2457 //---------------------------load_mirror_from_klass----------------------------
2458 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2459 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2460 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2461 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2462 }
2464 //-----------------------load_klass_from_mirror_common-------------------------
2465 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2466 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2467 // and branch to the given path on the region.
2468 // If never_see_null, take an uncommon trap on null, so we can optimistically
2469 // compile for the non-null case.
2470 // If the region is NULL, force never_see_null = true.
2471 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2472 bool never_see_null,
2473 int nargs,
2474 RegionNode* region,
2475 int null_path,
2476 int offset) {
2477 if (region == NULL) never_see_null = true;
2478 Node* p = basic_plus_adr(mirror, offset);
2479 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2480 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2481 _sp += nargs; // any deopt will start just before call to enclosing method
2482 Node* null_ctl = top();
2483 kls = null_check_oop(kls, &null_ctl, never_see_null);
2484 if (region != NULL) {
2485 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2486 region->init_req(null_path, null_ctl);
2487 } else {
2488 assert(null_ctl == top(), "no loose ends");
2489 }
2490 _sp -= nargs;
2491 return kls;
2492 }
2494 //--------------------(inline_native_Class_query helpers)---------------------
2495 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2496 // Fall through if (mods & mask) == bits, take the guard otherwise.
2497 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2498 // Branch around if the given klass has the given modifier bit set.
2499 // Like generate_guard, adds a new path onto the region.
2500 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2501 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2502 Node* mask = intcon(modifier_mask);
2503 Node* bits = intcon(modifier_bits);
2504 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2505 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2506 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2507 return generate_fair_guard(bol, region);
2508 }
2509 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2510 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2511 }
2513 //-------------------------inline_native_Class_query-------------------
2514 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2515 int nargs = 1+0; // just the Class mirror, in most cases
2516 const Type* return_type = TypeInt::BOOL;
2517 Node* prim_return_value = top(); // what happens if it's a primitive class?
2518 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2519 bool expect_prim = false; // most of these guys expect to work on refs
2521 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2523 switch (id) {
2524 case vmIntrinsics::_isInstance:
2525 nargs = 1+1; // the Class mirror, plus the object getting queried about
2526 // nothing is an instance of a primitive type
2527 prim_return_value = intcon(0);
2528 break;
2529 case vmIntrinsics::_getModifiers:
2530 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2531 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2532 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2533 break;
2534 case vmIntrinsics::_isInterface:
2535 prim_return_value = intcon(0);
2536 break;
2537 case vmIntrinsics::_isArray:
2538 prim_return_value = intcon(0);
2539 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2540 break;
2541 case vmIntrinsics::_isPrimitive:
2542 prim_return_value = intcon(1);
2543 expect_prim = true; // obviously
2544 break;
2545 case vmIntrinsics::_getSuperclass:
2546 prim_return_value = null();
2547 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2548 break;
2549 case vmIntrinsics::_getComponentType:
2550 prim_return_value = null();
2551 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2552 break;
2553 case vmIntrinsics::_getClassAccessFlags:
2554 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2555 return_type = TypeInt::INT; // not bool! 6297094
2556 break;
2557 default:
2558 ShouldNotReachHere();
2559 }
2561 Node* mirror = argument(0);
2562 Node* obj = (nargs <= 1)? top(): argument(1);
2564 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2565 if (mirror_con == NULL) return false; // cannot happen?
2567 #ifndef PRODUCT
2568 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2569 ciType* k = mirror_con->java_mirror_type();
2570 if (k) {
2571 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2572 k->print_name();
2573 tty->cr();
2574 }
2575 }
2576 #endif
2578 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2579 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2580 record_for_igvn(region);
2581 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2583 // The mirror will never be null of Reflection.getClassAccessFlags, however
2584 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2585 // if it is. See bug 4774291.
2587 // For Reflection.getClassAccessFlags(), the null check occurs in
2588 // the wrong place; see inline_unsafe_access(), above, for a similar
2589 // situation.
2590 _sp += nargs; // set original stack for use by uncommon_trap
2591 mirror = do_null_check(mirror, T_OBJECT);
2592 _sp -= nargs;
2593 // If mirror or obj is dead, only null-path is taken.
2594 if (stopped()) return true;
2596 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2598 // Now load the mirror's klass metaobject, and null-check it.
2599 // Side-effects region with the control path if the klass is null.
2600 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2601 region, _prim_path);
2602 // If kls is null, we have a primitive mirror.
2603 phi->init_req(_prim_path, prim_return_value);
2604 if (stopped()) { push_result(region, phi); return true; }
2606 Node* p; // handy temp
2607 Node* null_ctl;
2609 // Now that we have the non-null klass, we can perform the real query.
2610 // For constant classes, the query will constant-fold in LoadNode::Value.
2611 Node* query_value = top();
2612 switch (id) {
2613 case vmIntrinsics::_isInstance:
2614 // nothing is an instance of a primitive type
2615 query_value = gen_instanceof(obj, kls);
2616 break;
2618 case vmIntrinsics::_getModifiers:
2619 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2620 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2621 break;
2623 case vmIntrinsics::_isInterface:
2624 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2625 if (generate_interface_guard(kls, region) != NULL)
2626 // A guard was added. If the guard is taken, it was an interface.
2627 phi->add_req(intcon(1));
2628 // If we fall through, it's a plain class.
2629 query_value = intcon(0);
2630 break;
2632 case vmIntrinsics::_isArray:
2633 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2634 if (generate_array_guard(kls, region) != NULL)
2635 // A guard was added. If the guard is taken, it was an array.
2636 phi->add_req(intcon(1));
2637 // If we fall through, it's a plain class.
2638 query_value = intcon(0);
2639 break;
2641 case vmIntrinsics::_isPrimitive:
2642 query_value = intcon(0); // "normal" path produces false
2643 break;
2645 case vmIntrinsics::_getSuperclass:
2646 // The rules here are somewhat unfortunate, but we can still do better
2647 // with random logic than with a JNI call.
2648 // Interfaces store null or Object as _super, but must report null.
2649 // Arrays store an intermediate super as _super, but must report Object.
2650 // Other types can report the actual _super.
2651 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2652 if (generate_interface_guard(kls, region) != NULL)
2653 // A guard was added. If the guard is taken, it was an interface.
2654 phi->add_req(null());
2655 if (generate_array_guard(kls, region) != NULL)
2656 // A guard was added. If the guard is taken, it was an array.
2657 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2658 // If we fall through, it's a plain class. Get its _super.
2659 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2660 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2661 null_ctl = top();
2662 kls = null_check_oop(kls, &null_ctl);
2663 if (null_ctl != top()) {
2664 // If the guard is taken, Object.superClass is null (both klass and mirror).
2665 region->add_req(null_ctl);
2666 phi ->add_req(null());
2667 }
2668 if (!stopped()) {
2669 query_value = load_mirror_from_klass(kls);
2670 }
2671 break;
2673 case vmIntrinsics::_getComponentType:
2674 if (generate_array_guard(kls, region) != NULL) {
2675 // Be sure to pin the oop load to the guard edge just created:
2676 Node* is_array_ctrl = region->in(region->req()-1);
2677 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
2678 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
2679 phi->add_req(cmo);
2680 }
2681 query_value = null(); // non-array case is null
2682 break;
2684 case vmIntrinsics::_getClassAccessFlags:
2685 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2686 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2687 break;
2689 default:
2690 ShouldNotReachHere();
2691 }
2693 // Fall-through is the normal case of a query to a real class.
2694 phi->init_req(1, query_value);
2695 region->init_req(1, control());
2697 push_result(region, phi);
2698 C->set_has_split_ifs(true); // Has chance for split-if optimization
2700 return true;
2701 }
2703 //--------------------------inline_native_subtype_check------------------------
2704 // This intrinsic takes the JNI calls out of the heart of
2705 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
2706 bool LibraryCallKit::inline_native_subtype_check() {
2707 int nargs = 1+1; // the Class mirror, plus the other class getting examined
2709 // Pull both arguments off the stack.
2710 Node* args[2]; // two java.lang.Class mirrors: superc, subc
2711 args[0] = argument(0);
2712 args[1] = argument(1);
2713 Node* klasses[2]; // corresponding Klasses: superk, subk
2714 klasses[0] = klasses[1] = top();
2716 enum {
2717 // A full decision tree on {superc is prim, subc is prim}:
2718 _prim_0_path = 1, // {P,N} => false
2719 // {P,P} & superc!=subc => false
2720 _prim_same_path, // {P,P} & superc==subc => true
2721 _prim_1_path, // {N,P} => false
2722 _ref_subtype_path, // {N,N} & subtype check wins => true
2723 _both_ref_path, // {N,N} & subtype check loses => false
2724 PATH_LIMIT
2725 };
2727 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2728 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2729 record_for_igvn(region);
2731 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
2732 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2733 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
2735 // First null-check both mirrors and load each mirror's klass metaobject.
2736 int which_arg;
2737 for (which_arg = 0; which_arg <= 1; which_arg++) {
2738 Node* arg = args[which_arg];
2739 _sp += nargs; // set original stack for use by uncommon_trap
2740 arg = do_null_check(arg, T_OBJECT);
2741 _sp -= nargs;
2742 if (stopped()) break;
2743 args[which_arg] = _gvn.transform(arg);
2745 Node* p = basic_plus_adr(arg, class_klass_offset);
2746 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
2747 klasses[which_arg] = _gvn.transform(kls);
2748 }
2750 // Having loaded both klasses, test each for null.
2751 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2752 for (which_arg = 0; which_arg <= 1; which_arg++) {
2753 Node* kls = klasses[which_arg];
2754 Node* null_ctl = top();
2755 _sp += nargs; // set original stack for use by uncommon_trap
2756 kls = null_check_oop(kls, &null_ctl, never_see_null);
2757 _sp -= nargs;
2758 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
2759 region->init_req(prim_path, null_ctl);
2760 if (stopped()) break;
2761 klasses[which_arg] = kls;
2762 }
2764 if (!stopped()) {
2765 // now we have two reference types, in klasses[0..1]
2766 Node* subk = klasses[1]; // the argument to isAssignableFrom
2767 Node* superk = klasses[0]; // the receiver
2768 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
2769 // now we have a successful reference subtype check
2770 region->set_req(_ref_subtype_path, control());
2771 }
2773 // If both operands are primitive (both klasses null), then
2774 // we must return true when they are identical primitives.
2775 // It is convenient to test this after the first null klass check.
2776 set_control(region->in(_prim_0_path)); // go back to first null check
2777 if (!stopped()) {
2778 // Since superc is primitive, make a guard for the superc==subc case.
2779 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
2780 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
2781 generate_guard(bol_eq, region, PROB_FAIR);
2782 if (region->req() == PATH_LIMIT+1) {
2783 // A guard was added. If the added guard is taken, superc==subc.
2784 region->swap_edges(PATH_LIMIT, _prim_same_path);
2785 region->del_req(PATH_LIMIT);
2786 }
2787 region->set_req(_prim_0_path, control()); // Not equal after all.
2788 }
2790 // these are the only paths that produce 'true':
2791 phi->set_req(_prim_same_path, intcon(1));
2792 phi->set_req(_ref_subtype_path, intcon(1));
2794 // pull together the cases:
2795 assert(region->req() == PATH_LIMIT, "sane region");
2796 for (uint i = 1; i < region->req(); i++) {
2797 Node* ctl = region->in(i);
2798 if (ctl == NULL || ctl == top()) {
2799 region->set_req(i, top());
2800 phi ->set_req(i, top());
2801 } else if (phi->in(i) == NULL) {
2802 phi->set_req(i, intcon(0)); // all other paths produce 'false'
2803 }
2804 }
2806 set_control(_gvn.transform(region));
2807 push(_gvn.transform(phi));
2809 return true;
2810 }
2812 //---------------------generate_array_guard_common------------------------
2813 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
2814 bool obj_array, bool not_array) {
2815 // If obj_array/non_array==false/false:
2816 // Branch around if the given klass is in fact an array (either obj or prim).
2817 // If obj_array/non_array==false/true:
2818 // Branch around if the given klass is not an array klass of any kind.
2819 // If obj_array/non_array==true/true:
2820 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
2821 // If obj_array/non_array==true/false:
2822 // Branch around if the kls is an oop array (Object[] or subtype)
2823 //
2824 // Like generate_guard, adds a new path onto the region.
2825 jint layout_con = 0;
2826 Node* layout_val = get_layout_helper(kls, layout_con);
2827 if (layout_val == NULL) {
2828 bool query = (obj_array
2829 ? Klass::layout_helper_is_objArray(layout_con)
2830 : Klass::layout_helper_is_javaArray(layout_con));
2831 if (query == not_array) {
2832 return NULL; // never a branch
2833 } else { // always a branch
2834 Node* always_branch = control();
2835 if (region != NULL)
2836 region->add_req(always_branch);
2837 set_control(top());
2838 return always_branch;
2839 }
2840 }
2841 // Now test the correct condition.
2842 jint nval = (obj_array
2843 ? ((jint)Klass::_lh_array_tag_type_value
2844 << Klass::_lh_array_tag_shift)
2845 : Klass::_lh_neutral_value);
2846 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
2847 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
2848 // invert the test if we are looking for a non-array
2849 if (not_array) btest = BoolTest(btest).negate();
2850 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
2851 return generate_fair_guard(bol, region);
2852 }
2855 //-----------------------inline_native_newArray--------------------------
2856 bool LibraryCallKit::inline_native_newArray() {
2857 int nargs = 2;
2858 Node* mirror = argument(0);
2859 Node* count_val = argument(1);
2861 _sp += nargs; // set original stack for use by uncommon_trap
2862 mirror = do_null_check(mirror, T_OBJECT);
2863 _sp -= nargs;
2864 // If mirror or obj is dead, only null-path is taken.
2865 if (stopped()) return true;
2867 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
2868 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2869 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
2870 TypeInstPtr::NOTNULL);
2871 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
2872 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
2873 TypePtr::BOTTOM);
2875 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2876 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
2877 nargs,
2878 result_reg, _slow_path);
2879 Node* normal_ctl = control();
2880 Node* no_array_ctl = result_reg->in(_slow_path);
2882 // Generate code for the slow case. We make a call to newArray().
2883 set_control(no_array_ctl);
2884 if (!stopped()) {
2885 // Either the input type is void.class, or else the
2886 // array klass has not yet been cached. Either the
2887 // ensuing call will throw an exception, or else it
2888 // will cache the array klass for next time.
2889 PreserveJVMState pjvms(this);
2890 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
2891 Node* slow_result = set_results_for_java_call(slow_call);
2892 // this->control() comes from set_results_for_java_call
2893 result_reg->set_req(_slow_path, control());
2894 result_val->set_req(_slow_path, slow_result);
2895 result_io ->set_req(_slow_path, i_o());
2896 result_mem->set_req(_slow_path, reset_memory());
2897 }
2899 set_control(normal_ctl);
2900 if (!stopped()) {
2901 // Normal case: The array type has been cached in the java.lang.Class.
2902 // The following call works fine even if the array type is polymorphic.
2903 // It could be a dynamic mix of int[], boolean[], Object[], etc.
2904 _sp += nargs; // set original stack for use by uncommon_trap
2905 Node* obj = new_array(klass_node, count_val);
2906 _sp -= nargs;
2907 result_reg->init_req(_normal_path, control());
2908 result_val->init_req(_normal_path, obj);
2909 result_io ->init_req(_normal_path, i_o());
2910 result_mem->init_req(_normal_path, reset_memory());
2911 }
2913 // Return the combined state.
2914 set_i_o( _gvn.transform(result_io) );
2915 set_all_memory( _gvn.transform(result_mem) );
2916 push_result(result_reg, result_val);
2917 C->set_has_split_ifs(true); // Has chance for split-if optimization
2919 return true;
2920 }
2922 //----------------------inline_native_getLength--------------------------
2923 bool LibraryCallKit::inline_native_getLength() {
2924 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
2926 int nargs = 1;
2927 Node* array = argument(0);
2929 _sp += nargs; // set original stack for use by uncommon_trap
2930 array = do_null_check(array, T_OBJECT);
2931 _sp -= nargs;
2933 // If array is dead, only null-path is taken.
2934 if (stopped()) return true;
2936 // Deoptimize if it is a non-array.
2937 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
2939 if (non_array != NULL) {
2940 PreserveJVMState pjvms(this);
2941 set_control(non_array);
2942 _sp += nargs; // push the arguments back on the stack
2943 uncommon_trap(Deoptimization::Reason_intrinsic,
2944 Deoptimization::Action_maybe_recompile);
2945 }
2947 // If control is dead, only non-array-path is taken.
2948 if (stopped()) return true;
2950 // The works fine even if the array type is polymorphic.
2951 // It could be a dynamic mix of int[], boolean[], Object[], etc.
2952 push( load_array_length(array) );
2954 C->set_has_split_ifs(true); // Has chance for split-if optimization
2956 return true;
2957 }
2959 //------------------------inline_array_copyOf----------------------------
2960 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
2961 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
2963 // Restore the stack and pop off the arguments.
2964 int nargs = 3 + (is_copyOfRange? 1: 0);
2965 Node* original = argument(0);
2966 Node* start = is_copyOfRange? argument(1): intcon(0);
2967 Node* end = is_copyOfRange? argument(2): argument(1);
2968 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
2970 _sp += nargs; // set original stack for use by uncommon_trap
2971 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
2972 original = do_null_check(original, T_OBJECT);
2973 _sp -= nargs;
2975 // Check if a null path was taken unconditionally.
2976 if (stopped()) return true;
2978 Node* orig_length = load_array_length(original);
2980 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs,
2981 NULL, 0);
2982 _sp += nargs; // set original stack for use by uncommon_trap
2983 klass_node = do_null_check(klass_node, T_OBJECT);
2984 _sp -= nargs;
2986 RegionNode* bailout = new (C, 1) RegionNode(1);
2987 record_for_igvn(bailout);
2989 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
2990 // Bail out if that is so.
2991 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
2992 if (not_objArray != NULL) {
2993 // Improve the klass node's type from the new optimistic assumption:
2994 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
2995 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2996 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
2997 cast->init_req(0, control());
2998 klass_node = _gvn.transform(cast);
2999 }
3001 // Bail out if either start or end is negative.
3002 generate_negative_guard(start, bailout, &start);
3003 generate_negative_guard(end, bailout, &end);
3005 Node* length = end;
3006 if (_gvn.type(start) != TypeInt::ZERO) {
3007 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3008 }
3010 // Bail out if length is negative.
3011 // ...Not needed, since the new_array will throw the right exception.
3012 //generate_negative_guard(length, bailout, &length);
3014 if (bailout->req() > 1) {
3015 PreserveJVMState pjvms(this);
3016 set_control( _gvn.transform(bailout) );
3017 _sp += nargs; // push the arguments back on the stack
3018 uncommon_trap(Deoptimization::Reason_intrinsic,
3019 Deoptimization::Action_maybe_recompile);
3020 }
3022 if (!stopped()) {
3023 // How many elements will we copy from the original?
3024 // The answer is MinI(orig_length - start, length).
3025 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3026 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3028 _sp += nargs; // set original stack for use by uncommon_trap
3029 Node* newcopy = new_array(klass_node, length);
3030 _sp -= nargs;
3032 // Generate a direct call to the right arraycopy function(s).
3033 // We know the copy is disjoint but we might not know if the
3034 // oop stores need checking.
3035 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3036 // This will fail a store-check if x contains any non-nulls.
3037 bool disjoint_bases = true;
3038 bool length_never_negative = true;
3039 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3040 original, start, newcopy, intcon(0), moved,
3041 nargs, disjoint_bases, length_never_negative);
3043 push(newcopy);
3044 }
3046 C->set_has_split_ifs(true); // Has chance for split-if optimization
3048 return true;
3049 }
3052 //----------------------generate_virtual_guard---------------------------
3053 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3054 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3055 RegionNode* slow_region) {
3056 ciMethod* method = callee();
3057 int vtable_index = method->vtable_index();
3058 // Get the methodOop out of the appropriate vtable entry.
3059 int entry_offset = (instanceKlass::vtable_start_offset() +
3060 vtable_index*vtableEntry::size()) * wordSize +
3061 vtableEntry::method_offset_in_bytes();
3062 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3063 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3065 // Compare the target method with the expected method (e.g., Object.hashCode).
3066 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3068 Node* native_call = makecon(native_call_addr);
3069 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3070 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3072 return generate_slow_guard(test_native, slow_region);
3073 }
3075 //-----------------------generate_method_call----------------------------
3076 // Use generate_method_call to make a slow-call to the real
3077 // method if the fast path fails. An alternative would be to
3078 // use a stub like OptoRuntime::slow_arraycopy_Java.
3079 // This only works for expanding the current library call,
3080 // not another intrinsic. (E.g., don't use this for making an
3081 // arraycopy call inside of the copyOf intrinsic.)
3082 CallJavaNode*
3083 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3084 // When compiling the intrinsic method itself, do not use this technique.
3085 guarantee(callee() != C->method(), "cannot make slow-call to self");
3087 ciMethod* method = callee();
3088 // ensure the JVMS we have will be correct for this call
3089 guarantee(method_id == method->intrinsic_id(), "must match");
3091 const TypeFunc* tf = TypeFunc::make(method);
3092 int tfdc = tf->domain()->cnt();
3093 CallJavaNode* slow_call;
3094 if (is_static) {
3095 assert(!is_virtual, "");
3096 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3097 SharedRuntime::get_resolve_static_call_stub(),
3098 method, bci());
3099 } else if (is_virtual) {
3100 null_check_receiver(method);
3101 int vtable_index = methodOopDesc::invalid_vtable_index;
3102 if (UseInlineCaches) {
3103 // Suppress the vtable call
3104 } else {
3105 // hashCode and clone are not a miranda methods,
3106 // so the vtable index is fixed.
3107 // No need to use the linkResolver to get it.
3108 vtable_index = method->vtable_index();
3109 }
3110 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3111 SharedRuntime::get_resolve_virtual_call_stub(),
3112 method, vtable_index, bci());
3113 } else { // neither virtual nor static: opt_virtual
3114 null_check_receiver(method);
3115 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3116 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3117 method, bci());
3118 slow_call->set_optimized_virtual(true);
3119 }
3120 set_arguments_for_java_call(slow_call);
3121 set_edges_for_java_call(slow_call);
3122 return slow_call;
3123 }
3126 //------------------------------inline_native_hashcode--------------------
3127 // Build special case code for calls to hashCode on an object.
3128 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3129 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3130 assert(!(is_virtual && is_static), "either virtual, special, or static");
3132 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3134 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3135 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3136 TypeInt::INT);
3137 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3138 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3139 TypePtr::BOTTOM);
3140 Node* obj = NULL;
3141 if (!is_static) {
3142 // Check for hashing null object
3143 obj = null_check_receiver(callee());
3144 if (stopped()) return true; // unconditionally null
3145 result_reg->init_req(_null_path, top());
3146 result_val->init_req(_null_path, top());
3147 } else {
3148 // Do a null check, and return zero if null.
3149 // System.identityHashCode(null) == 0
3150 obj = argument(0);
3151 Node* null_ctl = top();
3152 obj = null_check_oop(obj, &null_ctl);
3153 result_reg->init_req(_null_path, null_ctl);
3154 result_val->init_req(_null_path, _gvn.intcon(0));
3155 }
3157 // Unconditionally null? Then return right away.
3158 if (stopped()) {
3159 set_control( result_reg->in(_null_path) );
3160 if (!stopped())
3161 push( result_val ->in(_null_path) );
3162 return true;
3163 }
3165 // After null check, get the object's klass.
3166 Node* obj_klass = load_object_klass(obj);
3168 // This call may be virtual (invokevirtual) or bound (invokespecial).
3169 // For each case we generate slightly different code.
3171 // We only go to the fast case code if we pass a number of guards. The
3172 // paths which do not pass are accumulated in the slow_region.
3173 RegionNode* slow_region = new (C, 1) RegionNode(1);
3174 record_for_igvn(slow_region);
3176 // If this is a virtual call, we generate a funny guard. We pull out
3177 // the vtable entry corresponding to hashCode() from the target object.
3178 // If the target method which we are calling happens to be the native
3179 // Object hashCode() method, we pass the guard. We do not need this
3180 // guard for non-virtual calls -- the caller is known to be the native
3181 // Object hashCode().
3182 if (is_virtual) {
3183 generate_virtual_guard(obj_klass, slow_region);
3184 }
3186 // Get the header out of the object, use LoadMarkNode when available
3187 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3188 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
3189 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
3191 // Test the header to see if it is unlocked.
3192 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3193 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3194 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3195 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3196 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3198 generate_slow_guard(test_unlocked, slow_region);
3200 // Get the hash value and check to see that it has been properly assigned.
3201 // We depend on hash_mask being at most 32 bits and avoid the use of
3202 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3203 // vm: see markOop.hpp.
3204 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3205 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3206 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3207 // This hack lets the hash bits live anywhere in the mark object now, as long
3208 // as the shift drops the relevent bits into the low 32 bits. Note that
3209 // Java spec says that HashCode is an int so there's no point in capturing
3210 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3211 hshifted_header = ConvX2I(hshifted_header);
3212 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3214 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3215 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3216 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3218 generate_slow_guard(test_assigned, slow_region);
3220 Node* init_mem = reset_memory();
3221 // fill in the rest of the null path:
3222 result_io ->init_req(_null_path, i_o());
3223 result_mem->init_req(_null_path, init_mem);
3225 result_val->init_req(_fast_path, hash_val);
3226 result_reg->init_req(_fast_path, control());
3227 result_io ->init_req(_fast_path, i_o());
3228 result_mem->init_req(_fast_path, init_mem);
3230 // Generate code for the slow case. We make a call to hashCode().
3231 set_control(_gvn.transform(slow_region));
3232 if (!stopped()) {
3233 // No need for PreserveJVMState, because we're using up the present state.
3234 set_all_memory(init_mem);
3235 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3236 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3237 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3238 Node* slow_result = set_results_for_java_call(slow_call);
3239 // this->control() comes from set_results_for_java_call
3240 result_reg->init_req(_slow_path, control());
3241 result_val->init_req(_slow_path, slow_result);
3242 result_io ->set_req(_slow_path, i_o());
3243 result_mem ->set_req(_slow_path, reset_memory());
3244 }
3246 // Return the combined state.
3247 set_i_o( _gvn.transform(result_io) );
3248 set_all_memory( _gvn.transform(result_mem) );
3249 push_result(result_reg, result_val);
3251 return true;
3252 }
3254 //---------------------------inline_native_getClass----------------------------
3255 // Build special case code for calls to hashCode on an object.
3256 bool LibraryCallKit::inline_native_getClass() {
3257 Node* obj = null_check_receiver(callee());
3258 if (stopped()) return true;
3259 push( load_mirror_from_klass(load_object_klass(obj)) );
3260 return true;
3261 }
3263 //-----------------inline_native_Reflection_getCallerClass---------------------
3264 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3265 //
3266 // NOTE that this code must perform the same logic as
3267 // vframeStream::security_get_caller_frame in that it must skip
3268 // Method.invoke() and auxiliary frames.
3273 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3274 ciMethod* method = callee();
3276 #ifndef PRODUCT
3277 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3278 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3279 }
3280 #endif
3282 debug_only(int saved_sp = _sp);
3284 // Argument words: (int depth)
3285 int nargs = 1;
3287 _sp += nargs;
3288 Node* caller_depth_node = pop();
3290 assert(saved_sp == _sp, "must have correct argument count");
3292 // The depth value must be a constant in order for the runtime call
3293 // to be eliminated.
3294 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3295 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3296 #ifndef PRODUCT
3297 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3298 tty->print_cr(" Bailing out because caller depth was not a constant");
3299 }
3300 #endif
3301 return false;
3302 }
3303 // Note that the JVM state at this point does not include the
3304 // getCallerClass() frame which we are trying to inline. The
3305 // semantics of getCallerClass(), however, are that the "first"
3306 // frame is the getCallerClass() frame, so we subtract one from the
3307 // requested depth before continuing. We don't inline requests of
3308 // getCallerClass(0).
3309 int caller_depth = caller_depth_type->get_con() - 1;
3310 if (caller_depth < 0) {
3311 #ifndef PRODUCT
3312 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3313 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3314 }
3315 #endif
3316 return false;
3317 }
3319 if (!jvms()->has_method()) {
3320 #ifndef PRODUCT
3321 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3322 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3323 }
3324 #endif
3325 return false;
3326 }
3327 int _depth = jvms()->depth(); // cache call chain depth
3329 // Walk back up the JVM state to find the caller at the required
3330 // depth. NOTE that this code must perform the same logic as
3331 // vframeStream::security_get_caller_frame in that it must skip
3332 // Method.invoke() and auxiliary frames. Note also that depth is
3333 // 1-based (1 is the bottom of the inlining).
3334 int inlining_depth = _depth;
3335 JVMState* caller_jvms = NULL;
3337 if (inlining_depth > 0) {
3338 caller_jvms = jvms();
3339 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3340 do {
3341 // The following if-tests should be performed in this order
3342 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3343 // Skip a Method.invoke() or auxiliary frame
3344 } else if (caller_depth > 0) {
3345 // Skip real frame
3346 --caller_depth;
3347 } else {
3348 // We're done: reached desired caller after skipping.
3349 break;
3350 }
3351 caller_jvms = caller_jvms->caller();
3352 --inlining_depth;
3353 } while (inlining_depth > 0);
3354 }
3356 if (inlining_depth == 0) {
3357 #ifndef PRODUCT
3358 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3359 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3360 tty->print_cr(" JVM state at this point:");
3361 for (int i = _depth; i >= 1; i--) {
3362 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3363 }
3364 }
3365 #endif
3366 return false; // Reached end of inlining
3367 }
3369 // Acquire method holder as java.lang.Class
3370 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3371 ciInstance* caller_mirror = caller_klass->java_mirror();
3372 // Push this as a constant
3373 push(makecon(TypeInstPtr::make(caller_mirror)));
3374 #ifndef PRODUCT
3375 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3376 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);
3377 tty->print_cr(" JVM state at this point:");
3378 for (int i = _depth; i >= 1; i--) {
3379 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3380 }
3381 }
3382 #endif
3383 return true;
3384 }
3386 // Helper routine for above
3387 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3388 // Is this the Method.invoke method itself?
3389 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
3390 return true;
3392 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3393 ciKlass* k = jvms->method()->holder();
3394 if (k->is_instance_klass()) {
3395 ciInstanceKlass* ik = k->as_instance_klass();
3396 for (; ik != NULL; ik = ik->super()) {
3397 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3398 ik == env()->find_system_klass(ik->name())) {
3399 return true;
3400 }
3401 }
3402 }
3404 return false;
3405 }
3407 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3408 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3409 // computing it since there is no lookup field by name function in the
3410 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3411 // Using a static variable here is safe even if we have multiple compilation
3412 // threads because the offset is constant. At worst the same offset will be
3413 // computed and stored multiple
3415 bool LibraryCallKit::inline_native_AtomicLong_get() {
3416 // Restore the stack and pop off the argument
3417 _sp+=1;
3418 Node *obj = pop();
3420 // get the offset of the "value" field. Since the CI interfaces
3421 // does not provide a way to look up a field by name, we scan the bytecodes
3422 // to get the field index. We expect the first 2 instructions of the method
3423 // to be:
3424 // 0 aload_0
3425 // 1 getfield "value"
3426 ciMethod* method = callee();
3427 if (value_field_offset == -1)
3428 {
3429 ciField* value_field;
3430 ciBytecodeStream iter(method);
3431 Bytecodes::Code bc = iter.next();
3433 if ((bc != Bytecodes::_aload_0) &&
3434 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3435 return false;
3436 bc = iter.next();
3437 if (bc != Bytecodes::_getfield)
3438 return false;
3439 bool ignore;
3440 value_field = iter.get_field(ignore);
3441 value_field_offset = value_field->offset_in_bytes();
3442 }
3444 // Null check without removing any arguments.
3445 _sp++;
3446 obj = do_null_check(obj, T_OBJECT);
3447 _sp--;
3448 // Check for locking null object
3449 if (stopped()) return true;
3451 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3452 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3453 int alias_idx = C->get_alias_index(adr_type);
3455 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3457 push_pair(result);
3459 return true;
3460 }
3462 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3463 // Restore the stack and pop off the arguments
3464 _sp+=5;
3465 Node *newVal = pop_pair();
3466 Node *oldVal = pop_pair();
3467 Node *obj = pop();
3469 // we need the offset of the "value" field which was computed when
3470 // inlining the get() method. Give up if we don't have it.
3471 if (value_field_offset == -1)
3472 return false;
3474 // Null check without removing any arguments.
3475 _sp+=5;
3476 obj = do_null_check(obj, T_OBJECT);
3477 _sp-=5;
3478 // Check for locking null object
3479 if (stopped()) return true;
3481 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3482 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3483 int alias_idx = C->get_alias_index(adr_type);
3485 Node *result = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3486 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(result));
3487 set_memory(store_proj, alias_idx);
3489 push(result);
3490 return true;
3491 }
3493 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3494 // restore the arguments
3495 _sp += arg_size();
3497 switch (id) {
3498 case vmIntrinsics::_floatToRawIntBits:
3499 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3500 break;
3502 case vmIntrinsics::_intBitsToFloat:
3503 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3504 break;
3506 case vmIntrinsics::_doubleToRawLongBits:
3507 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3508 break;
3510 case vmIntrinsics::_longBitsToDouble:
3511 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3512 break;
3514 case vmIntrinsics::_doubleToLongBits: {
3515 Node* value = pop_pair();
3517 // two paths (plus control) merge in a wood
3518 RegionNode *r = new (C, 3) RegionNode(3);
3519 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3521 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3522 // Build the boolean node
3523 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3525 // Branch either way.
3526 // NaN case is less traveled, which makes all the difference.
3527 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3528 Node *opt_isnan = _gvn.transform(ifisnan);
3529 assert( opt_isnan->is_If(), "Expect an IfNode");
3530 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3531 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3533 set_control(iftrue);
3535 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3536 Node *slow_result = longcon(nan_bits); // return NaN
3537 phi->init_req(1, _gvn.transform( slow_result ));
3538 r->init_req(1, iftrue);
3540 // Else fall through
3541 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3542 set_control(iffalse);
3544 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3545 r->init_req(2, iffalse);
3547 // Post merge
3548 set_control(_gvn.transform(r));
3549 record_for_igvn(r);
3551 Node* result = _gvn.transform(phi);
3552 assert(result->bottom_type()->isa_long(), "must be");
3553 push_pair(result);
3555 C->set_has_split_ifs(true); // Has chance for split-if optimization
3557 break;
3558 }
3560 case vmIntrinsics::_floatToIntBits: {
3561 Node* value = pop();
3563 // two paths (plus control) merge in a wood
3564 RegionNode *r = new (C, 3) RegionNode(3);
3565 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3567 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3568 // Build the boolean node
3569 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3571 // Branch either way.
3572 // NaN case is less traveled, which makes all the difference.
3573 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3574 Node *opt_isnan = _gvn.transform(ifisnan);
3575 assert( opt_isnan->is_If(), "Expect an IfNode");
3576 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3577 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3579 set_control(iftrue);
3581 static const jint nan_bits = 0x7fc00000;
3582 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3583 phi->init_req(1, _gvn.transform( slow_result ));
3584 r->init_req(1, iftrue);
3586 // Else fall through
3587 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3588 set_control(iffalse);
3590 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3591 r->init_req(2, iffalse);
3593 // Post merge
3594 set_control(_gvn.transform(r));
3595 record_for_igvn(r);
3597 Node* result = _gvn.transform(phi);
3598 assert(result->bottom_type()->isa_int(), "must be");
3599 push(result);
3601 C->set_has_split_ifs(true); // Has chance for split-if optimization
3603 break;
3604 }
3606 default:
3607 ShouldNotReachHere();
3608 }
3610 return true;
3611 }
3613 #ifdef _LP64
3614 #define XTOP ,top() /*additional argument*/
3615 #else //_LP64
3616 #define XTOP /*no additional argument*/
3617 #endif //_LP64
3619 //----------------------inline_unsafe_copyMemory-------------------------
3620 bool LibraryCallKit::inline_unsafe_copyMemory() {
3621 if (callee()->is_static()) return false; // caller must have the capability!
3622 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3623 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3624 null_check_receiver(callee()); // check then ignore argument(0)
3625 if (stopped()) return true;
3627 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3629 Node* src_ptr = argument(1);
3630 Node* src_off = ConvL2X(argument(2));
3631 assert(argument(3)->is_top(), "2nd half of long");
3632 Node* dst_ptr = argument(4);
3633 Node* dst_off = ConvL2X(argument(5));
3634 assert(argument(6)->is_top(), "2nd half of long");
3635 Node* size = ConvL2X(argument(7));
3636 assert(argument(8)->is_top(), "2nd half of long");
3638 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3639 "fieldOffset must be byte-scaled");
3641 Node* src = make_unsafe_address(src_ptr, src_off);
3642 Node* dst = make_unsafe_address(dst_ptr, dst_off);
3644 // Conservatively insert a memory barrier on all memory slices.
3645 // Do not let writes of the copy source or destination float below the copy.
3646 insert_mem_bar(Op_MemBarCPUOrder);
3648 // Call it. Note that the length argument is not scaled.
3649 make_runtime_call(RC_LEAF|RC_NO_FP,
3650 OptoRuntime::fast_arraycopy_Type(),
3651 StubRoutines::unsafe_arraycopy(),
3652 "unsafe_arraycopy",
3653 TypeRawPtr::BOTTOM,
3654 src, dst, size XTOP);
3656 // Do not let reads of the copy destination float above the copy.
3657 insert_mem_bar(Op_MemBarCPUOrder);
3659 return true;
3660 }
3663 //------------------------inline_native_clone----------------------------
3664 // Here are the simple edge cases:
3665 // null receiver => normal trap
3666 // virtual and clone was overridden => slow path to out-of-line clone
3667 // not cloneable or finalizer => slow path to out-of-line Object.clone
3668 //
3669 // The general case has two steps, allocation and copying.
3670 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
3671 //
3672 // Copying also has two cases, oop arrays and everything else.
3673 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
3674 // Everything else uses the tight inline loop supplied by CopyArrayNode.
3675 //
3676 // These steps fold up nicely if and when the cloned object's klass
3677 // can be sharply typed as an object array, a type array, or an instance.
3678 //
3679 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
3680 int nargs = 1;
3681 Node* obj = null_check_receiver(callee());
3682 if (stopped()) return true;
3683 Node* obj_klass = load_object_klass(obj);
3684 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
3685 const TypeOopPtr* toop = ((tklass != NULL)
3686 ? tklass->as_instance_type()
3687 : TypeInstPtr::NOTNULL);
3689 // Conservatively insert a memory barrier on all memory slices.
3690 // Do not let writes into the original float below the clone.
3691 insert_mem_bar(Op_MemBarCPUOrder);
3693 // paths into result_reg:
3694 enum {
3695 _slow_path = 1, // out-of-line call to clone method (virtual or not)
3696 _objArray_path, // plain allocation, plus arrayof_oop_arraycopy
3697 _fast_path, // plain allocation, plus a CopyArray operation
3698 PATH_LIMIT
3699 };
3700 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3701 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3702 TypeInstPtr::NOTNULL);
3703 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3704 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3705 TypePtr::BOTTOM);
3706 record_for_igvn(result_reg);
3708 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
3709 int raw_adr_idx = Compile::AliasIdxRaw;
3710 const bool raw_mem_only = true;
3712 // paths into alloc_reg (on the fast path, just before the CopyArray):
3713 enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT };
3714 RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT);
3715 PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type);
3716 PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X);
3717 PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO);
3718 PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY,
3719 raw_adr_type);
3720 record_for_igvn(alloc_reg);
3722 bool card_mark = false; // (see below)
3724 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
3725 if (array_ctl != NULL) {
3726 // It's an array.
3727 PreserveJVMState pjvms(this);
3728 set_control(array_ctl);
3729 Node* obj_length = load_array_length(obj);
3730 Node* obj_size = NULL;
3731 _sp += nargs; // set original stack for use by uncommon_trap
3732 Node* alloc_obj = new_array(obj_klass, obj_length,
3733 raw_mem_only, &obj_size);
3734 _sp -= nargs;
3735 assert(obj_size != NULL, "");
3736 Node* raw_obj = alloc_obj->in(1);
3737 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
3738 if (ReduceBulkZeroing) {
3739 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
3740 if (alloc != NULL) {
3741 // We will be completely responsible for initializing this object.
3742 alloc->maybe_set_complete(&_gvn);
3743 }
3744 }
3746 if (!use_ReduceInitialCardMarks()) {
3747 // If it is an oop array, it requires very special treatment,
3748 // because card marking is required on each card of the array.
3749 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
3750 if (is_obja != NULL) {
3751 PreserveJVMState pjvms2(this);
3752 set_control(is_obja);
3753 // Generate a direct call to the right arraycopy function(s).
3754 bool disjoint_bases = true;
3755 bool length_never_negative = true;
3756 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3757 obj, intcon(0), alloc_obj, intcon(0),
3758 obj_length, nargs,
3759 disjoint_bases, length_never_negative);
3760 result_reg->init_req(_objArray_path, control());
3761 result_val->init_req(_objArray_path, alloc_obj);
3762 result_i_o ->set_req(_objArray_path, i_o());
3763 result_mem ->set_req(_objArray_path, reset_memory());
3764 }
3765 }
3766 // We can dispense with card marks if we know the allocation
3767 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
3768 // causes the non-eden paths to simulate a fresh allocation,
3769 // insofar that no further card marks are required to initialize
3770 // the object.
3772 // Otherwise, there are no card marks to worry about.
3773 alloc_val->init_req(_typeArray_alloc, raw_obj);
3774 alloc_siz->init_req(_typeArray_alloc, obj_size);
3775 alloc_reg->init_req(_typeArray_alloc, control());
3776 alloc_i_o->init_req(_typeArray_alloc, i_o());
3777 alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type));
3778 }
3780 // We only go to the fast case code if we pass a number of guards.
3781 // The paths which do not pass are accumulated in the slow_region.
3782 RegionNode* slow_region = new (C, 1) RegionNode(1);
3783 record_for_igvn(slow_region);
3784 if (!stopped()) {
3785 // It's an instance. Make the slow-path tests.
3786 // If this is a virtual call, we generate a funny guard. We grab
3787 // the vtable entry corresponding to clone() from the target object.
3788 // If the target method which we are calling happens to be the
3789 // Object clone() method, we pass the guard. We do not need this
3790 // guard for non-virtual calls; the caller is known to be the native
3791 // Object clone().
3792 if (is_virtual) {
3793 generate_virtual_guard(obj_klass, slow_region);
3794 }
3796 // The object must be cloneable and must not have a finalizer.
3797 // Both of these conditions may be checked in a single test.
3798 // We could optimize the cloneable test further, but we don't care.
3799 generate_access_flags_guard(obj_klass,
3800 // Test both conditions:
3801 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
3802 // Must be cloneable but not finalizer:
3803 JVM_ACC_IS_CLONEABLE,
3804 slow_region);
3805 }
3807 if (!stopped()) {
3808 // It's an instance, and it passed the slow-path tests.
3809 PreserveJVMState pjvms(this);
3810 Node* obj_size = NULL;
3811 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
3812 assert(obj_size != NULL, "");
3813 Node* raw_obj = alloc_obj->in(1);
3814 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
3815 if (ReduceBulkZeroing) {
3816 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
3817 if (alloc != NULL && !alloc->maybe_set_complete(&_gvn))
3818 alloc = NULL;
3819 }
3820 if (!use_ReduceInitialCardMarks()) {
3821 // Put in store barrier for any and all oops we are sticking
3822 // into this object. (We could avoid this if we could prove
3823 // that the object type contains no oop fields at all.)
3824 card_mark = true;
3825 }
3826 alloc_val->init_req(_instance_alloc, raw_obj);
3827 alloc_siz->init_req(_instance_alloc, obj_size);
3828 alloc_reg->init_req(_instance_alloc, control());
3829 alloc_i_o->init_req(_instance_alloc, i_o());
3830 alloc_mem->init_req(_instance_alloc, memory(raw_adr_type));
3831 }
3833 // Generate code for the slow case. We make a call to clone().
3834 set_control(_gvn.transform(slow_region));
3835 if (!stopped()) {
3836 PreserveJVMState pjvms(this);
3837 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
3838 Node* slow_result = set_results_for_java_call(slow_call);
3839 // this->control() comes from set_results_for_java_call
3840 result_reg->init_req(_slow_path, control());
3841 result_val->init_req(_slow_path, slow_result);
3842 result_i_o ->set_req(_slow_path, i_o());
3843 result_mem ->set_req(_slow_path, reset_memory());
3844 }
3846 // The object is allocated, as an array and/or an instance. Now copy it.
3847 set_control( _gvn.transform(alloc_reg) );
3848 set_i_o( _gvn.transform(alloc_i_o) );
3849 set_memory( _gvn.transform(alloc_mem), raw_adr_type );
3850 Node* raw_obj = _gvn.transform(alloc_val);
3852 if (!stopped()) {
3853 // Copy the fastest available way.
3854 // (No need for PreserveJVMState, since we're using it all up now.)
3855 // TODO: generate fields/elements copies for small objects instead.
3856 Node* src = obj;
3857 Node* dest = raw_obj;
3858 Node* size = _gvn.transform(alloc_siz);
3860 // Exclude the header.
3861 int base_off = instanceOopDesc::base_offset_in_bytes();
3862 if (UseCompressedOops) {
3863 assert(base_off % BytesPerLong != 0, "base with compressed oops");
3864 // With compressed oops base_offset_in_bytes is 12 which creates
3865 // the gap since countx is rounded by 8 bytes below.
3866 // Copy klass and the gap.
3867 base_off = instanceOopDesc::klass_offset_in_bytes();
3868 }
3869 src = basic_plus_adr(src, base_off);
3870 dest = basic_plus_adr(dest, base_off);
3872 // Compute the length also, if needed:
3873 Node* countx = size;
3874 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
3875 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
3877 // Select an appropriate instruction to initialize the range.
3878 // The CopyArray instruction (if supported) can be optimized
3879 // into a discrete set of scalar loads and stores.
3880 bool disjoint_bases = true;
3881 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
3882 src, NULL, dest, NULL, countx);
3884 // Now that the object is properly initialized, type it as an oop.
3885 // Use a secondary InitializeNode memory barrier.
3886 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx,
3887 raw_obj)->as_Initialize();
3888 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
3889 Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj,
3890 TypeInstPtr::NOTNULL);
3891 new_obj = _gvn.transform(new_obj);
3893 // If necessary, emit some card marks afterwards. (Non-arrays only.)
3894 if (card_mark) {
3895 Node* no_particular_value = NULL;
3896 Node* no_particular_field = NULL;
3897 post_barrier(control(),
3898 memory(raw_adr_type),
3899 new_obj,
3900 no_particular_field,
3901 raw_adr_idx,
3902 no_particular_value,
3903 T_OBJECT,
3904 false);
3905 }
3906 // Present the results of the slow call.
3907 result_reg->init_req(_fast_path, control());
3908 result_val->init_req(_fast_path, new_obj);
3909 result_i_o ->set_req(_fast_path, i_o());
3910 result_mem ->set_req(_fast_path, reset_memory());
3911 }
3913 // Return the combined state.
3914 set_control( _gvn.transform(result_reg) );
3915 set_i_o( _gvn.transform(result_i_o) );
3916 set_all_memory( _gvn.transform(result_mem) );
3918 // Cast the result to a sharper type, since we know what clone does.
3919 Node* new_obj = _gvn.transform(result_val);
3920 Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop);
3921 push(_gvn.transform(cast));
3923 return true;
3924 }
3927 // constants for computing the copy function
3928 enum {
3929 COPYFUNC_UNALIGNED = 0,
3930 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
3931 COPYFUNC_CONJOINT = 0,
3932 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
3933 };
3935 // Note: The condition "disjoint" applies also for overlapping copies
3936 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
3937 static address
3938 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
3939 int selector =
3940 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
3941 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
3943 #define RETURN_STUB(xxx_arraycopy) { \
3944 name = #xxx_arraycopy; \
3945 return StubRoutines::xxx_arraycopy(); }
3947 switch (t) {
3948 case T_BYTE:
3949 case T_BOOLEAN:
3950 switch (selector) {
3951 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
3952 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
3953 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
3954 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
3955 }
3956 case T_CHAR:
3957 case T_SHORT:
3958 switch (selector) {
3959 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
3960 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
3961 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
3962 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
3963 }
3964 case T_INT:
3965 case T_FLOAT:
3966 switch (selector) {
3967 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
3968 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
3969 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
3970 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
3971 }
3972 case T_DOUBLE:
3973 case T_LONG:
3974 switch (selector) {
3975 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
3976 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
3977 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
3978 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
3979 }
3980 case T_ARRAY:
3981 case T_OBJECT:
3982 switch (selector) {
3983 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
3984 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
3985 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
3986 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
3987 }
3988 default:
3989 ShouldNotReachHere();
3990 return NULL;
3991 }
3993 #undef RETURN_STUB
3994 }
3996 //------------------------------basictype2arraycopy----------------------------
3997 address LibraryCallKit::basictype2arraycopy(BasicType t,
3998 Node* src_offset,
3999 Node* dest_offset,
4000 bool disjoint_bases,
4001 const char* &name) {
4002 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4003 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4005 bool aligned = false;
4006 bool disjoint = disjoint_bases;
4008 // if the offsets are the same, we can treat the memory regions as
4009 // disjoint, because either the memory regions are in different arrays,
4010 // or they are identical (which we can treat as disjoint.) We can also
4011 // treat a copy with a destination index less that the source index
4012 // as disjoint since a low->high copy will work correctly in this case.
4013 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4014 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4015 // both indices are constants
4016 int s_offs = src_offset_inttype->get_con();
4017 int d_offs = dest_offset_inttype->get_con();
4018 int element_size = type2aelembytes(t);
4019 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4020 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4021 if (s_offs >= d_offs) disjoint = true;
4022 } else if (src_offset == dest_offset && src_offset != NULL) {
4023 // This can occur if the offsets are identical non-constants.
4024 disjoint = true;
4025 }
4027 return select_arraycopy_function(t, aligned, disjoint, name);
4028 }
4031 //------------------------------inline_arraycopy-----------------------
4032 bool LibraryCallKit::inline_arraycopy() {
4033 // Restore the stack and pop off the arguments.
4034 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4035 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4037 Node *src = argument(0);
4038 Node *src_offset = argument(1);
4039 Node *dest = argument(2);
4040 Node *dest_offset = argument(3);
4041 Node *length = argument(4);
4043 // Compile time checks. If any of these checks cannot be verified at compile time,
4044 // we do not make a fast path for this call. Instead, we let the call remain as it
4045 // is. The checks we choose to mandate at compile time are:
4046 //
4047 // (1) src and dest are arrays.
4048 const Type* src_type = src->Value(&_gvn);
4049 const Type* dest_type = dest->Value(&_gvn);
4050 const TypeAryPtr* top_src = src_type->isa_aryptr();
4051 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4052 if (top_src == NULL || top_src->klass() == NULL ||
4053 top_dest == NULL || top_dest->klass() == NULL) {
4054 // Conservatively insert a memory barrier on all memory slices.
4055 // Do not let writes into the source float below the arraycopy.
4056 insert_mem_bar(Op_MemBarCPUOrder);
4058 // Call StubRoutines::generic_arraycopy stub.
4059 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4060 src, src_offset, dest, dest_offset, length,
4061 nargs);
4063 // Do not let reads from the destination float above the arraycopy.
4064 // Since we cannot type the arrays, we don't know which slices
4065 // might be affected. We could restrict this barrier only to those
4066 // memory slices which pertain to array elements--but don't bother.
4067 if (!InsertMemBarAfterArraycopy)
4068 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4069 insert_mem_bar(Op_MemBarCPUOrder);
4070 return true;
4071 }
4073 // (2) src and dest arrays must have elements of the same BasicType
4074 // Figure out the size and type of the elements we will be copying.
4075 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4076 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4077 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4078 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4080 if (src_elem != dest_elem || dest_elem == T_VOID) {
4081 // The component types are not the same or are not recognized. Punt.
4082 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4083 generate_slow_arraycopy(TypePtr::BOTTOM,
4084 src, src_offset, dest, dest_offset, length,
4085 nargs);
4086 return true;
4087 }
4089 //---------------------------------------------------------------------------
4090 // We will make a fast path for this call to arraycopy.
4092 // We have the following tests left to perform:
4093 //
4094 // (3) src and dest must not be null.
4095 // (4) src_offset must not be negative.
4096 // (5) dest_offset must not be negative.
4097 // (6) length must not be negative.
4098 // (7) src_offset + length must not exceed length of src.
4099 // (8) dest_offset + length must not exceed length of dest.
4100 // (9) each element of an oop array must be assignable
4102 RegionNode* slow_region = new (C, 1) RegionNode(1);
4103 record_for_igvn(slow_region);
4105 // (3) operands must not be null
4106 // We currently perform our null checks with the do_null_check routine.
4107 // This means that the null exceptions will be reported in the caller
4108 // rather than (correctly) reported inside of the native arraycopy call.
4109 // This should be corrected, given time. We do our null check with the
4110 // stack pointer restored.
4111 _sp += nargs;
4112 src = do_null_check(src, T_ARRAY);
4113 dest = do_null_check(dest, T_ARRAY);
4114 _sp -= nargs;
4116 // (4) src_offset must not be negative.
4117 generate_negative_guard(src_offset, slow_region);
4119 // (5) dest_offset must not be negative.
4120 generate_negative_guard(dest_offset, slow_region);
4122 // (6) length must not be negative (moved to generate_arraycopy()).
4123 // generate_negative_guard(length, slow_region);
4125 // (7) src_offset + length must not exceed length of src.
4126 generate_limit_guard(src_offset, length,
4127 load_array_length(src),
4128 slow_region);
4130 // (8) dest_offset + length must not exceed length of dest.
4131 generate_limit_guard(dest_offset, length,
4132 load_array_length(dest),
4133 slow_region);
4135 // (9) each element of an oop array must be assignable
4136 // The generate_arraycopy subroutine checks this.
4138 // This is where the memory effects are placed:
4139 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4140 generate_arraycopy(adr_type, dest_elem,
4141 src, src_offset, dest, dest_offset, length,
4142 nargs, false, false, slow_region);
4144 return true;
4145 }
4147 //-----------------------------generate_arraycopy----------------------
4148 // Generate an optimized call to arraycopy.
4149 // Caller must guard against non-arrays.
4150 // Caller must determine a common array basic-type for both arrays.
4151 // Caller must validate offsets against array bounds.
4152 // The slow_region has already collected guard failure paths
4153 // (such as out of bounds length or non-conformable array types).
4154 // The generated code has this shape, in general:
4155 //
4156 // if (length == 0) return // via zero_path
4157 // slowval = -1
4158 // if (types unknown) {
4159 // slowval = call generic copy loop
4160 // if (slowval == 0) return // via checked_path
4161 // } else if (indexes in bounds) {
4162 // if ((is object array) && !(array type check)) {
4163 // slowval = call checked copy loop
4164 // if (slowval == 0) return // via checked_path
4165 // } else {
4166 // call bulk copy loop
4167 // return // via fast_path
4168 // }
4169 // }
4170 // // adjust params for remaining work:
4171 // if (slowval != -1) {
4172 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4173 // }
4174 // slow_region:
4175 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4176 // return // via slow_call_path
4177 //
4178 // This routine is used from several intrinsics: System.arraycopy,
4179 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4180 //
4181 void
4182 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4183 BasicType basic_elem_type,
4184 Node* src, Node* src_offset,
4185 Node* dest, Node* dest_offset,
4186 Node* copy_length,
4187 int nargs,
4188 bool disjoint_bases,
4189 bool length_never_negative,
4190 RegionNode* slow_region) {
4192 if (slow_region == NULL) {
4193 slow_region = new(C,1) RegionNode(1);
4194 record_for_igvn(slow_region);
4195 }
4197 Node* original_dest = dest;
4198 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4199 Node* raw_dest = NULL; // used before zeroing, if needed
4200 bool must_clear_dest = false;
4202 // See if this is the initialization of a newly-allocated array.
4203 // If so, we will take responsibility here for initializing it to zero.
4204 // (Note: Because tightly_coupled_allocation performs checks on the
4205 // out-edges of the dest, we need to avoid making derived pointers
4206 // from it until we have checked its uses.)
4207 if (ReduceBulkZeroing
4208 && !ZeroTLAB // pointless if already zeroed
4209 && basic_elem_type != T_CONFLICT // avoid corner case
4210 && !_gvn.eqv_uncast(src, dest)
4211 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4212 != NULL)
4213 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4214 && alloc->maybe_set_complete(&_gvn)) {
4215 // "You break it, you buy it."
4216 InitializeNode* init = alloc->initialization();
4217 assert(init->is_complete(), "we just did this");
4218 assert(dest->Opcode() == Op_CheckCastPP, "sanity");
4219 assert(dest->in(0)->in(0) == init, "dest pinned");
4220 raw_dest = dest->in(1); // grab the raw pointer!
4221 original_dest = dest;
4222 dest = raw_dest;
4223 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4224 // Decouple the original InitializeNode, turning it into a simple membar.
4225 // We will build a new one at the end of this routine.
4226 init->set_req(InitializeNode::RawAddress, top());
4227 // From this point on, every exit path is responsible for
4228 // initializing any non-copied parts of the object to zero.
4229 must_clear_dest = true;
4230 } else {
4231 // No zeroing elimination here.
4232 alloc = NULL;
4233 //original_dest = dest;
4234 //must_clear_dest = false;
4235 }
4237 // Results are placed here:
4238 enum { fast_path = 1, // normal void-returning assembly stub
4239 checked_path = 2, // special assembly stub with cleanup
4240 slow_call_path = 3, // something went wrong; call the VM
4241 zero_path = 4, // bypass when length of copy is zero
4242 bcopy_path = 5, // copy primitive array by 64-bit blocks
4243 PATH_LIMIT = 6
4244 };
4245 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4246 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4247 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4248 record_for_igvn(result_region);
4249 _gvn.set_type_bottom(result_i_o);
4250 _gvn.set_type_bottom(result_memory);
4251 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4253 // The slow_control path:
4254 Node* slow_control;
4255 Node* slow_i_o = i_o();
4256 Node* slow_mem = memory(adr_type);
4257 debug_only(slow_control = (Node*) badAddress);
4259 // Checked control path:
4260 Node* checked_control = top();
4261 Node* checked_mem = NULL;
4262 Node* checked_i_o = NULL;
4263 Node* checked_value = NULL;
4265 if (basic_elem_type == T_CONFLICT) {
4266 assert(!must_clear_dest, "");
4267 Node* cv = generate_generic_arraycopy(adr_type,
4268 src, src_offset, dest, dest_offset,
4269 copy_length, nargs);
4270 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4271 checked_control = control();
4272 checked_i_o = i_o();
4273 checked_mem = memory(adr_type);
4274 checked_value = cv;
4275 set_control(top()); // no fast path
4276 }
4278 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4279 if (not_pos != NULL) {
4280 PreserveJVMState pjvms(this);
4281 set_control(not_pos);
4283 // (6) length must not be negative.
4284 if (!length_never_negative) {
4285 generate_negative_guard(copy_length, slow_region);
4286 }
4288 if (!stopped() && must_clear_dest) {
4289 Node* dest_length = alloc->in(AllocateNode::ALength);
4290 if (_gvn.eqv_uncast(copy_length, dest_length)
4291 || _gvn.find_int_con(dest_length, 1) <= 0) {
4292 // There is no zeroing to do.
4293 } else {
4294 // Clear the whole thing since there are no source elements to copy.
4295 generate_clear_array(adr_type, dest, basic_elem_type,
4296 intcon(0), NULL,
4297 alloc->in(AllocateNode::AllocSize));
4298 }
4299 }
4301 // Present the results of the fast call.
4302 result_region->init_req(zero_path, control());
4303 result_i_o ->init_req(zero_path, i_o());
4304 result_memory->init_req(zero_path, memory(adr_type));
4305 }
4307 if (!stopped() && must_clear_dest) {
4308 // We have to initialize the *uncopied* part of the array to zero.
4309 // The copy destination is the slice dest[off..off+len]. The other slices
4310 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4311 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4312 Node* dest_length = alloc->in(AllocateNode::ALength);
4313 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4314 copy_length) );
4316 // If there is a head section that needs zeroing, do it now.
4317 if (find_int_con(dest_offset, -1) != 0) {
4318 generate_clear_array(adr_type, dest, basic_elem_type,
4319 intcon(0), dest_offset,
4320 NULL);
4321 }
4323 // Next, perform a dynamic check on the tail length.
4324 // It is often zero, and we can win big if we prove this.
4325 // There are two wins: Avoid generating the ClearArray
4326 // with its attendant messy index arithmetic, and upgrade
4327 // the copy to a more hardware-friendly word size of 64 bits.
4328 Node* tail_ctl = NULL;
4329 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4330 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4331 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4332 tail_ctl = generate_slow_guard(bol_lt, NULL);
4333 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4334 }
4336 // At this point, let's assume there is no tail.
4337 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4338 // There is no tail. Try an upgrade to a 64-bit copy.
4339 bool didit = false;
4340 { PreserveJVMState pjvms(this);
4341 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4342 src, src_offset, dest, dest_offset,
4343 dest_size);
4344 if (didit) {
4345 // Present the results of the block-copying fast call.
4346 result_region->init_req(bcopy_path, control());
4347 result_i_o ->init_req(bcopy_path, i_o());
4348 result_memory->init_req(bcopy_path, memory(adr_type));
4349 }
4350 }
4351 if (didit)
4352 set_control(top()); // no regular fast path
4353 }
4355 // Clear the tail, if any.
4356 if (tail_ctl != NULL) {
4357 Node* notail_ctl = stopped() ? NULL : control();
4358 set_control(tail_ctl);
4359 if (notail_ctl == NULL) {
4360 generate_clear_array(adr_type, dest, basic_elem_type,
4361 dest_tail, NULL,
4362 dest_size);
4363 } else {
4364 // Make a local merge.
4365 Node* done_ctl = new(C,3) RegionNode(3);
4366 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4367 done_ctl->init_req(1, notail_ctl);
4368 done_mem->init_req(1, memory(adr_type));
4369 generate_clear_array(adr_type, dest, basic_elem_type,
4370 dest_tail, NULL,
4371 dest_size);
4372 done_ctl->init_req(2, control());
4373 done_mem->init_req(2, memory(adr_type));
4374 set_control( _gvn.transform(done_ctl) );
4375 set_memory( _gvn.transform(done_mem), adr_type );
4376 }
4377 }
4378 }
4380 BasicType copy_type = basic_elem_type;
4381 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4382 if (!stopped() && copy_type == T_OBJECT) {
4383 // If src and dest have compatible element types, we can copy bits.
4384 // Types S[] and D[] are compatible if D is a supertype of S.
4385 //
4386 // If they are not, we will use checked_oop_disjoint_arraycopy,
4387 // which performs a fast optimistic per-oop check, and backs off
4388 // further to JVM_ArrayCopy on the first per-oop check that fails.
4389 // (Actually, we don't move raw bits only; the GC requires card marks.)
4391 // Get the klassOop for both src and dest
4392 Node* src_klass = load_object_klass(src);
4393 Node* dest_klass = load_object_klass(dest);
4395 // Generate the subtype check.
4396 // This might fold up statically, or then again it might not.
4397 //
4398 // Non-static example: Copying List<String>.elements to a new String[].
4399 // The backing store for a List<String> is always an Object[],
4400 // but its elements are always type String, if the generic types
4401 // are correct at the source level.
4402 //
4403 // Test S[] against D[], not S against D, because (probably)
4404 // the secondary supertype cache is less busy for S[] than S.
4405 // This usually only matters when D is an interface.
4406 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4407 // Plug failing path into checked_oop_disjoint_arraycopy
4408 if (not_subtype_ctrl != top()) {
4409 PreserveJVMState pjvms(this);
4410 set_control(not_subtype_ctrl);
4411 // (At this point we can assume disjoint_bases, since types differ.)
4412 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4413 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4414 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4415 Node* dest_elem_klass = _gvn.transform(n1);
4416 Node* cv = generate_checkcast_arraycopy(adr_type,
4417 dest_elem_klass,
4418 src, src_offset, dest, dest_offset,
4419 copy_length,
4420 nargs);
4421 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4422 checked_control = control();
4423 checked_i_o = i_o();
4424 checked_mem = memory(adr_type);
4425 checked_value = cv;
4426 }
4427 // At this point we know we do not need type checks on oop stores.
4429 // Let's see if we need card marks:
4430 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4431 // If we do not need card marks, copy using the jint or jlong stub.
4432 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4433 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4434 "sizes agree");
4435 }
4436 }
4438 if (!stopped()) {
4439 // Generate the fast path, if possible.
4440 PreserveJVMState pjvms(this);
4441 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4442 src, src_offset, dest, dest_offset,
4443 ConvI2X(copy_length));
4445 // Present the results of the fast call.
4446 result_region->init_req(fast_path, control());
4447 result_i_o ->init_req(fast_path, i_o());
4448 result_memory->init_req(fast_path, memory(adr_type));
4449 }
4451 // Here are all the slow paths up to this point, in one bundle:
4452 slow_control = top();
4453 if (slow_region != NULL)
4454 slow_control = _gvn.transform(slow_region);
4455 debug_only(slow_region = (RegionNode*)badAddress);
4457 set_control(checked_control);
4458 if (!stopped()) {
4459 // Clean up after the checked call.
4460 // The returned value is either 0 or -1^K,
4461 // where K = number of partially transferred array elements.
4462 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4463 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4464 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4466 // If it is 0, we are done, so transfer to the end.
4467 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4468 result_region->init_req(checked_path, checks_done);
4469 result_i_o ->init_req(checked_path, checked_i_o);
4470 result_memory->init_req(checked_path, checked_mem);
4472 // If it is not zero, merge into the slow call.
4473 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4474 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4475 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4476 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4477 record_for_igvn(slow_reg2);
4478 slow_reg2 ->init_req(1, slow_control);
4479 slow_i_o2 ->init_req(1, slow_i_o);
4480 slow_mem2 ->init_req(1, slow_mem);
4481 slow_reg2 ->init_req(2, control());
4482 slow_i_o2 ->init_req(2, i_o());
4483 slow_mem2 ->init_req(2, memory(adr_type));
4485 slow_control = _gvn.transform(slow_reg2);
4486 slow_i_o = _gvn.transform(slow_i_o2);
4487 slow_mem = _gvn.transform(slow_mem2);
4489 if (alloc != NULL) {
4490 // We'll restart from the very beginning, after zeroing the whole thing.
4491 // This can cause double writes, but that's OK since dest is brand new.
4492 // So we ignore the low 31 bits of the value returned from the stub.
4493 } else {
4494 // We must continue the copy exactly where it failed, or else
4495 // another thread might see the wrong number of writes to dest.
4496 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4497 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4498 slow_offset->init_req(1, intcon(0));
4499 slow_offset->init_req(2, checked_offset);
4500 slow_offset = _gvn.transform(slow_offset);
4502 // Adjust the arguments by the conditionally incoming offset.
4503 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4504 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4505 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4507 // Tweak the node variables to adjust the code produced below:
4508 src_offset = src_off_plus;
4509 dest_offset = dest_off_plus;
4510 copy_length = length_minus;
4511 }
4512 }
4514 set_control(slow_control);
4515 if (!stopped()) {
4516 // Generate the slow path, if needed.
4517 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4519 set_memory(slow_mem, adr_type);
4520 set_i_o(slow_i_o);
4522 if (must_clear_dest) {
4523 generate_clear_array(adr_type, dest, basic_elem_type,
4524 intcon(0), NULL,
4525 alloc->in(AllocateNode::AllocSize));
4526 }
4528 if (dest != original_dest) {
4529 // Promote from rawptr to oop, so it looks right in the call's GC map.
4530 dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest,
4531 TypeInstPtr::NOTNULL) );
4533 // Edit the call's debug-info to avoid referring to original_dest.
4534 // (The problem with original_dest is that it isn't ready until
4535 // after the InitializeNode completes, but this stuff is before.)
4536 // Substitute in the locally valid dest_oop.
4537 replace_in_map(original_dest, dest);
4538 }
4540 generate_slow_arraycopy(adr_type,
4541 src, src_offset, dest, dest_offset,
4542 copy_length, nargs);
4544 result_region->init_req(slow_call_path, control());
4545 result_i_o ->init_req(slow_call_path, i_o());
4546 result_memory->init_req(slow_call_path, memory(adr_type));
4547 }
4549 // Remove unused edges.
4550 for (uint i = 1; i < result_region->req(); i++) {
4551 if (result_region->in(i) == NULL)
4552 result_region->init_req(i, top());
4553 }
4555 // Finished; return the combined state.
4556 set_control( _gvn.transform(result_region) );
4557 set_i_o( _gvn.transform(result_i_o) );
4558 set_memory( _gvn.transform(result_memory), adr_type );
4560 if (dest != original_dest) {
4561 // Pin the "finished" array node after the arraycopy/zeroing operations.
4562 // Use a secondary InitializeNode memory barrier.
4563 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4564 Compile::AliasIdxRaw,
4565 raw_dest)->as_Initialize();
4566 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4567 _gvn.hash_delete(original_dest);
4568 original_dest->set_req(0, control());
4569 _gvn.hash_find_insert(original_dest); // put back into GVN table
4570 }
4572 // The memory edges above are precise in order to model effects around
4573 // array copyies accurately to allow value numbering of field loads around
4574 // arraycopy. Such field loads, both before and after, are common in Java
4575 // collections and similar classes involving header/array data structures.
4576 //
4577 // But with low number of register or when some registers are used or killed
4578 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4579 // The next memory barrier is added to avoid it. If the arraycopy can be
4580 // optimized away (which it can, sometimes) then we can manually remove
4581 // the membar also.
4582 if (InsertMemBarAfterArraycopy)
4583 insert_mem_bar(Op_MemBarCPUOrder);
4584 }
4587 // Helper function which determines if an arraycopy immediately follows
4588 // an allocation, with no intervening tests or other escapes for the object.
4589 AllocateArrayNode*
4590 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4591 RegionNode* slow_region) {
4592 if (stopped()) return NULL; // no fast path
4593 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4595 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4596 if (alloc == NULL) return NULL;
4598 Node* rawmem = memory(Compile::AliasIdxRaw);
4599 // Is the allocation's memory state untouched?
4600 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4601 // Bail out if there have been raw-memory effects since the allocation.
4602 // (Example: There might have been a call or safepoint.)
4603 return NULL;
4604 }
4605 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4606 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4607 return NULL;
4608 }
4610 // There must be no unexpected observers of this allocation.
4611 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4612 Node* obs = ptr->fast_out(i);
4613 if (obs != this->map()) {
4614 return NULL;
4615 }
4616 }
4618 // This arraycopy must unconditionally follow the allocation of the ptr.
4619 Node* alloc_ctl = ptr->in(0);
4620 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4622 Node* ctl = control();
4623 while (ctl != alloc_ctl) {
4624 // There may be guards which feed into the slow_region.
4625 // Any other control flow means that we might not get a chance
4626 // to finish initializing the allocated object.
4627 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4628 IfNode* iff = ctl->in(0)->as_If();
4629 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4630 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4631 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4632 ctl = iff->in(0); // This test feeds the known slow_region.
4633 continue;
4634 }
4635 // One more try: Various low-level checks bottom out in
4636 // uncommon traps. If the debug-info of the trap omits
4637 // any reference to the allocation, as we've already
4638 // observed, then there can be no objection to the trap.
4639 bool found_trap = false;
4640 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4641 Node* obs = not_ctl->fast_out(j);
4642 if (obs->in(0) == not_ctl && obs->is_Call() &&
4643 (obs->as_Call()->entry_point() ==
4644 SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
4645 found_trap = true; break;
4646 }
4647 }
4648 if (found_trap) {
4649 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4650 continue;
4651 }
4652 }
4653 return NULL;
4654 }
4656 // If we get this far, we have an allocation which immediately
4657 // precedes the arraycopy, and we can take over zeroing the new object.
4658 // The arraycopy will finish the initialization, and provide
4659 // a new control state to which we will anchor the destination pointer.
4661 return alloc;
4662 }
4664 // Helper for initialization of arrays, creating a ClearArray.
4665 // It writes zero bits in [start..end), within the body of an array object.
4666 // The memory effects are all chained onto the 'adr_type' alias category.
4667 //
4668 // Since the object is otherwise uninitialized, we are free
4669 // to put a little "slop" around the edges of the cleared area,
4670 // as long as it does not go back into the array's header,
4671 // or beyond the array end within the heap.
4672 //
4673 // The lower edge can be rounded down to the nearest jint and the
4674 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4675 //
4676 // Arguments:
4677 // adr_type memory slice where writes are generated
4678 // dest oop of the destination array
4679 // basic_elem_type element type of the destination
4680 // slice_idx array index of first element to store
4681 // slice_len number of elements to store (or NULL)
4682 // dest_size total size in bytes of the array object
4683 //
4684 // Exactly one of slice_len or dest_size must be non-NULL.
4685 // If dest_size is non-NULL, zeroing extends to the end of the object.
4686 // If slice_len is non-NULL, the slice_idx value must be a constant.
4687 void
4688 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
4689 Node* dest,
4690 BasicType basic_elem_type,
4691 Node* slice_idx,
4692 Node* slice_len,
4693 Node* dest_size) {
4694 // one or the other but not both of slice_len and dest_size:
4695 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
4696 if (slice_len == NULL) slice_len = top();
4697 if (dest_size == NULL) dest_size = top();
4699 // operate on this memory slice:
4700 Node* mem = memory(adr_type); // memory slice to operate on
4702 // scaling and rounding of indexes:
4703 int scale = exact_log2(type2aelembytes(basic_elem_type));
4704 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
4705 int clear_low = (-1 << scale) & (BytesPerInt - 1);
4706 int bump_bit = (-1 << scale) & BytesPerInt;
4708 // determine constant starts and ends
4709 const intptr_t BIG_NEG = -128;
4710 assert(BIG_NEG + 2*abase < 0, "neg enough");
4711 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
4712 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
4713 if (slice_len_con == 0) {
4714 return; // nothing to do here
4715 }
4716 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
4717 intptr_t end_con = find_intptr_t_con(dest_size, -1);
4718 if (slice_idx_con >= 0 && slice_len_con >= 0) {
4719 assert(end_con < 0, "not two cons");
4720 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
4721 BytesPerLong);
4722 }
4724 if (start_con >= 0 && end_con >= 0) {
4725 // Constant start and end. Simple.
4726 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4727 start_con, end_con, &_gvn);
4728 } else if (start_con >= 0 && dest_size != top()) {
4729 // Constant start, pre-rounded end after the tail of the array.
4730 Node* end = dest_size;
4731 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4732 start_con, end, &_gvn);
4733 } else if (start_con >= 0 && slice_len != top()) {
4734 // Constant start, non-constant end. End needs rounding up.
4735 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
4736 intptr_t end_base = abase + (slice_idx_con << scale);
4737 int end_round = (-1 << scale) & (BytesPerLong - 1);
4738 Node* end = ConvI2X(slice_len);
4739 if (scale != 0)
4740 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
4741 end_base += end_round;
4742 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
4743 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
4744 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4745 start_con, end, &_gvn);
4746 } else if (start_con < 0 && dest_size != top()) {
4747 // Non-constant start, pre-rounded end after the tail of the array.
4748 // This is almost certainly a "round-to-end" operation.
4749 Node* start = slice_idx;
4750 start = ConvI2X(start);
4751 if (scale != 0)
4752 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
4753 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
4754 if ((bump_bit | clear_low) != 0) {
4755 int to_clear = (bump_bit | clear_low);
4756 // Align up mod 8, then store a jint zero unconditionally
4757 // just before the mod-8 boundary.
4758 if (((abase + bump_bit) & ~to_clear) - bump_bit
4759 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
4760 bump_bit = 0;
4761 assert((abase & to_clear) == 0, "array base must be long-aligned");
4762 } else {
4763 // Bump 'start' up to (or past) the next jint boundary:
4764 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
4765 assert((abase & clear_low) == 0, "array base must be int-aligned");
4766 }
4767 // Round bumped 'start' down to jlong boundary in body of array.
4768 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
4769 if (bump_bit != 0) {
4770 // Store a zero to the immediately preceding jint:
4771 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
4772 Node* p1 = basic_plus_adr(dest, x1);
4773 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
4774 mem = _gvn.transform(mem);
4775 }
4776 }
4777 Node* end = dest_size; // pre-rounded
4778 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4779 start, end, &_gvn);
4780 } else {
4781 // Non-constant start, unrounded non-constant end.
4782 // (Nobody zeroes a random midsection of an array using this routine.)
4783 ShouldNotReachHere(); // fix caller
4784 }
4786 // Done.
4787 set_memory(mem, adr_type);
4788 }
4791 bool
4792 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
4793 BasicType basic_elem_type,
4794 AllocateNode* alloc,
4795 Node* src, Node* src_offset,
4796 Node* dest, Node* dest_offset,
4797 Node* dest_size) {
4798 // See if there is an advantage from block transfer.
4799 int scale = exact_log2(type2aelembytes(basic_elem_type));
4800 if (scale >= LogBytesPerLong)
4801 return false; // it is already a block transfer
4803 // Look at the alignment of the starting offsets.
4804 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
4805 const intptr_t BIG_NEG = -128;
4806 assert(BIG_NEG + 2*abase < 0, "neg enough");
4808 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
4809 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
4810 if (src_off < 0 || dest_off < 0)
4811 // At present, we can only understand constants.
4812 return false;
4814 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
4815 // Non-aligned; too bad.
4816 // One more chance: Pick off an initial 32-bit word.
4817 // This is a common case, since abase can be odd mod 8.
4818 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
4819 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
4820 Node* sptr = basic_plus_adr(src, src_off);
4821 Node* dptr = basic_plus_adr(dest, dest_off);
4822 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
4823 store_to_memory(control(), dptr, sval, T_INT, adr_type);
4824 src_off += BytesPerInt;
4825 dest_off += BytesPerInt;
4826 } else {
4827 return false;
4828 }
4829 }
4830 assert(src_off % BytesPerLong == 0, "");
4831 assert(dest_off % BytesPerLong == 0, "");
4833 // Do this copy by giant steps.
4834 Node* sptr = basic_plus_adr(src, src_off);
4835 Node* dptr = basic_plus_adr(dest, dest_off);
4836 Node* countx = dest_size;
4837 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
4838 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
4840 bool disjoint_bases = true; // since alloc != NULL
4841 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
4842 sptr, NULL, dptr, NULL, countx);
4844 return true;
4845 }
4848 // Helper function; generates code for the slow case.
4849 // We make a call to a runtime method which emulates the native method,
4850 // but without the native wrapper overhead.
4851 void
4852 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
4853 Node* src, Node* src_offset,
4854 Node* dest, Node* dest_offset,
4855 Node* copy_length,
4856 int nargs) {
4857 _sp += nargs; // any deopt will start just before call to enclosing method
4858 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
4859 OptoRuntime::slow_arraycopy_Type(),
4860 OptoRuntime::slow_arraycopy_Java(),
4861 "slow_arraycopy", adr_type,
4862 src, src_offset, dest, dest_offset,
4863 copy_length);
4864 _sp -= nargs;
4866 // Handle exceptions thrown by this fellow:
4867 make_slow_call_ex(call, env()->Throwable_klass(), false);
4868 }
4870 // Helper function; generates code for cases requiring runtime checks.
4871 Node*
4872 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
4873 Node* dest_elem_klass,
4874 Node* src, Node* src_offset,
4875 Node* dest, Node* dest_offset,
4876 Node* copy_length,
4877 int nargs) {
4878 if (stopped()) return NULL;
4880 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
4881 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
4882 return NULL;
4883 }
4885 // Pick out the parameters required to perform a store-check
4886 // for the target array. This is an optimistic check. It will
4887 // look in each non-null element's class, at the desired klass's
4888 // super_check_offset, for the desired klass.
4889 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
4890 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
4891 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
4892 Node* check_offset = _gvn.transform(n3);
4893 Node* check_value = dest_elem_klass;
4895 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
4896 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
4898 // (We know the arrays are never conjoint, because their types differ.)
4899 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
4900 OptoRuntime::checkcast_arraycopy_Type(),
4901 copyfunc_addr, "checkcast_arraycopy", adr_type,
4902 // five arguments, of which two are
4903 // intptr_t (jlong in LP64)
4904 src_start, dest_start,
4905 copy_length XTOP,
4906 check_offset XTOP,
4907 check_value);
4909 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
4910 }
4913 // Helper function; generates code for cases requiring runtime checks.
4914 Node*
4915 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
4916 Node* src, Node* src_offset,
4917 Node* dest, Node* dest_offset,
4918 Node* copy_length,
4919 int nargs) {
4920 if (stopped()) return NULL;
4922 address copyfunc_addr = StubRoutines::generic_arraycopy();
4923 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
4924 return NULL;
4925 }
4927 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
4928 OptoRuntime::generic_arraycopy_Type(),
4929 copyfunc_addr, "generic_arraycopy", adr_type,
4930 src, src_offset, dest, dest_offset, copy_length);
4932 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
4933 }
4935 // Helper function; generates the fast out-of-line call to an arraycopy stub.
4936 void
4937 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
4938 BasicType basic_elem_type,
4939 bool disjoint_bases,
4940 Node* src, Node* src_offset,
4941 Node* dest, Node* dest_offset,
4942 Node* copy_length) {
4943 if (stopped()) return; // nothing to do
4945 Node* src_start = src;
4946 Node* dest_start = dest;
4947 if (src_offset != NULL || dest_offset != NULL) {
4948 assert(src_offset != NULL && dest_offset != NULL, "");
4949 src_start = array_element_address(src, src_offset, basic_elem_type);
4950 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
4951 }
4953 // Figure out which arraycopy runtime method to call.
4954 const char* copyfunc_name = "arraycopy";
4955 address copyfunc_addr =
4956 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
4957 disjoint_bases, copyfunc_name);
4959 // Call it. Note that the count_ix value is not scaled to a byte-size.
4960 make_runtime_call(RC_LEAF|RC_NO_FP,
4961 OptoRuntime::fast_arraycopy_Type(),
4962 copyfunc_addr, copyfunc_name, adr_type,
4963 src_start, dest_start, copy_length XTOP);
4964 }