Mon, 27 Jul 2009 06:15:29 -0700
Merge
1 /*
2 * Copyright 1999-2009 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
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 bool inline_string_equals();
140 Node* pop_math_arg();
141 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
142 bool inline_math_native(vmIntrinsics::ID id);
143 bool inline_trig(vmIntrinsics::ID id);
144 bool inline_trans(vmIntrinsics::ID id);
145 bool inline_abs(vmIntrinsics::ID id);
146 bool inline_sqrt(vmIntrinsics::ID id);
147 bool inline_pow(vmIntrinsics::ID id);
148 bool inline_exp(vmIntrinsics::ID id);
149 bool inline_min_max(vmIntrinsics::ID id);
150 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
151 // This returns Type::AnyPtr, RawPtr, or OopPtr.
152 int classify_unsafe_addr(Node* &base, Node* &offset);
153 Node* make_unsafe_address(Node* base, Node* offset);
154 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
155 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
156 bool inline_unsafe_allocate();
157 bool inline_unsafe_copyMemory();
158 bool inline_native_currentThread();
159 bool inline_native_time_funcs(bool isNano);
160 bool inline_native_isInterrupted();
161 bool inline_native_Class_query(vmIntrinsics::ID id);
162 bool inline_native_subtype_check();
164 bool inline_native_newArray();
165 bool inline_native_getLength();
166 bool inline_array_copyOf(bool is_copyOfRange);
167 bool inline_array_equals();
168 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
169 bool inline_native_clone(bool is_virtual);
170 bool inline_native_Reflection_getCallerClass();
171 bool inline_native_AtomicLong_get();
172 bool inline_native_AtomicLong_attemptUpdate();
173 bool is_method_invoke_or_aux_frame(JVMState* jvms);
174 // Helper function for inlining native object hash method
175 bool inline_native_hashcode(bool is_virtual, bool is_static);
176 bool inline_native_getClass();
178 // Helper functions for inlining arraycopy
179 bool inline_arraycopy();
180 void generate_arraycopy(const TypePtr* adr_type,
181 BasicType basic_elem_type,
182 Node* src, Node* src_offset,
183 Node* dest, Node* dest_offset,
184 Node* copy_length,
185 bool disjoint_bases = false,
186 bool length_never_negative = false,
187 RegionNode* slow_region = NULL);
188 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
189 RegionNode* slow_region);
190 void generate_clear_array(const TypePtr* adr_type,
191 Node* dest,
192 BasicType basic_elem_type,
193 Node* slice_off,
194 Node* slice_len,
195 Node* slice_end);
196 bool generate_block_arraycopy(const TypePtr* adr_type,
197 BasicType basic_elem_type,
198 AllocateNode* alloc,
199 Node* src, Node* src_offset,
200 Node* dest, Node* dest_offset,
201 Node* dest_size);
202 void generate_slow_arraycopy(const TypePtr* adr_type,
203 Node* src, Node* src_offset,
204 Node* dest, Node* dest_offset,
205 Node* copy_length);
206 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
207 Node* dest_elem_klass,
208 Node* src, Node* src_offset,
209 Node* dest, Node* dest_offset,
210 Node* copy_length);
211 Node* generate_generic_arraycopy(const TypePtr* adr_type,
212 Node* src, Node* src_offset,
213 Node* dest, Node* dest_offset,
214 Node* copy_length);
215 void generate_unchecked_arraycopy(const TypePtr* adr_type,
216 BasicType basic_elem_type,
217 bool disjoint_bases,
218 Node* src, Node* src_offset,
219 Node* dest, Node* dest_offset,
220 Node* copy_length);
221 bool inline_unsafe_CAS(BasicType type);
222 bool inline_unsafe_ordered_store(BasicType type);
223 bool inline_fp_conversions(vmIntrinsics::ID id);
224 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
225 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
226 bool inline_bitCount(vmIntrinsics::ID id);
227 bool inline_reverseBytes(vmIntrinsics::ID id);
228 };
231 //---------------------------make_vm_intrinsic----------------------------
232 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
233 vmIntrinsics::ID id = m->intrinsic_id();
234 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
236 if (DisableIntrinsic[0] != '\0'
237 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
238 // disabled by a user request on the command line:
239 // example: -XX:DisableIntrinsic=_hashCode,_getClass
240 return NULL;
241 }
243 if (!m->is_loaded()) {
244 // do not attempt to inline unloaded methods
245 return NULL;
246 }
248 // Only a few intrinsics implement a virtual dispatch.
249 // They are expensive calls which are also frequently overridden.
250 if (is_virtual) {
251 switch (id) {
252 case vmIntrinsics::_hashCode:
253 case vmIntrinsics::_clone:
254 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
255 break;
256 default:
257 return NULL;
258 }
259 }
261 // -XX:-InlineNatives disables nearly all intrinsics:
262 if (!InlineNatives) {
263 switch (id) {
264 case vmIntrinsics::_indexOf:
265 case vmIntrinsics::_compareTo:
266 case vmIntrinsics::_equals:
267 case vmIntrinsics::_equalsC:
268 break; // InlineNatives does not control String.compareTo
269 default:
270 return NULL;
271 }
272 }
274 switch (id) {
275 case vmIntrinsics::_compareTo:
276 if (!SpecialStringCompareTo) return NULL;
277 break;
278 case vmIntrinsics::_indexOf:
279 if (!SpecialStringIndexOf) return NULL;
280 break;
281 case vmIntrinsics::_equals:
282 if (!SpecialStringEquals) return NULL;
283 break;
284 case vmIntrinsics::_equalsC:
285 if (!SpecialArraysEquals) return NULL;
286 break;
287 case vmIntrinsics::_arraycopy:
288 if (!InlineArrayCopy) return NULL;
289 break;
290 case vmIntrinsics::_copyMemory:
291 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
292 if (!InlineArrayCopy) return NULL;
293 break;
294 case vmIntrinsics::_hashCode:
295 if (!InlineObjectHash) return NULL;
296 break;
297 case vmIntrinsics::_clone:
298 case vmIntrinsics::_copyOf:
299 case vmIntrinsics::_copyOfRange:
300 if (!InlineObjectCopy) return NULL;
301 // These also use the arraycopy intrinsic mechanism:
302 if (!InlineArrayCopy) return NULL;
303 break;
304 case vmIntrinsics::_checkIndex:
305 // We do not intrinsify this. The optimizer does fine with it.
306 return NULL;
308 case vmIntrinsics::_get_AtomicLong:
309 case vmIntrinsics::_attemptUpdate:
310 if (!InlineAtomicLong) return NULL;
311 break;
313 case vmIntrinsics::_getCallerClass:
314 if (!UseNewReflection) return NULL;
315 if (!InlineReflectionGetCallerClass) return NULL;
316 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
317 break;
319 case vmIntrinsics::_bitCount_i:
320 case vmIntrinsics::_bitCount_l:
321 if (!UsePopCountInstruction) return NULL;
322 break;
324 default:
325 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
326 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
327 break;
328 }
330 // -XX:-InlineClassNatives disables natives from the Class class.
331 // The flag applies to all reflective calls, notably Array.newArray
332 // (visible to Java programmers as Array.newInstance).
333 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
334 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
335 if (!InlineClassNatives) return NULL;
336 }
338 // -XX:-InlineThreadNatives disables natives from the Thread class.
339 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
340 if (!InlineThreadNatives) return NULL;
341 }
343 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
344 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
345 m->holder()->name() == ciSymbol::java_lang_Float() ||
346 m->holder()->name() == ciSymbol::java_lang_Double()) {
347 if (!InlineMathNatives) return NULL;
348 }
350 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
351 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
352 if (!InlineUnsafeOps) return NULL;
353 }
355 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
356 }
358 //----------------------register_library_intrinsics-----------------------
359 // Initialize this file's data structures, for each Compile instance.
360 void Compile::register_library_intrinsics() {
361 // Nothing to do here.
362 }
364 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
365 LibraryCallKit kit(jvms, this);
366 Compile* C = kit.C;
367 int nodes = C->unique();
368 #ifndef PRODUCT
369 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
370 char buf[1000];
371 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
372 tty->print_cr("Intrinsic %s", str);
373 }
374 #endif
375 if (kit.try_to_inline()) {
376 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
377 tty->print("Inlining intrinsic %s%s at bci:%d in",
378 vmIntrinsics::name_at(intrinsic_id()),
379 (is_virtual() ? " (virtual)" : ""), kit.bci());
380 kit.caller()->print_short_name(tty);
381 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
382 }
383 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
384 if (C->log()) {
385 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
386 vmIntrinsics::name_at(intrinsic_id()),
387 (is_virtual() ? " virtual='1'" : ""),
388 C->unique() - nodes);
389 }
390 return kit.transfer_exceptions_into_jvms();
391 }
393 if (PrintIntrinsics) {
394 tty->print("Did not inline intrinsic %s%s at bci:%d in",
395 vmIntrinsics::name_at(intrinsic_id()),
396 (is_virtual() ? " (virtual)" : ""), kit.bci());
397 kit.caller()->print_short_name(tty);
398 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
399 }
400 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
401 return NULL;
402 }
404 bool LibraryCallKit::try_to_inline() {
405 // Handle symbolic names for otherwise undistinguished boolean switches:
406 const bool is_store = true;
407 const bool is_native_ptr = true;
408 const bool is_static = true;
410 switch (intrinsic_id()) {
411 case vmIntrinsics::_hashCode:
412 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
413 case vmIntrinsics::_identityHashCode:
414 return inline_native_hashcode(/*!virtual*/ false, is_static);
415 case vmIntrinsics::_getClass:
416 return inline_native_getClass();
418 case vmIntrinsics::_dsin:
419 case vmIntrinsics::_dcos:
420 case vmIntrinsics::_dtan:
421 case vmIntrinsics::_dabs:
422 case vmIntrinsics::_datan2:
423 case vmIntrinsics::_dsqrt:
424 case vmIntrinsics::_dexp:
425 case vmIntrinsics::_dlog:
426 case vmIntrinsics::_dlog10:
427 case vmIntrinsics::_dpow:
428 return inline_math_native(intrinsic_id());
430 case vmIntrinsics::_min:
431 case vmIntrinsics::_max:
432 return inline_min_max(intrinsic_id());
434 case vmIntrinsics::_arraycopy:
435 return inline_arraycopy();
437 case vmIntrinsics::_compareTo:
438 return inline_string_compareTo();
439 case vmIntrinsics::_indexOf:
440 return inline_string_indexOf();
441 case vmIntrinsics::_equals:
442 return inline_string_equals();
444 case vmIntrinsics::_getObject:
445 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
446 case vmIntrinsics::_getBoolean:
447 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
448 case vmIntrinsics::_getByte:
449 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
450 case vmIntrinsics::_getShort:
451 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
452 case vmIntrinsics::_getChar:
453 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
454 case vmIntrinsics::_getInt:
455 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
456 case vmIntrinsics::_getLong:
457 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
458 case vmIntrinsics::_getFloat:
459 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
460 case vmIntrinsics::_getDouble:
461 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
463 case vmIntrinsics::_putObject:
464 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
465 case vmIntrinsics::_putBoolean:
466 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
467 case vmIntrinsics::_putByte:
468 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
469 case vmIntrinsics::_putShort:
470 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
471 case vmIntrinsics::_putChar:
472 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
473 case vmIntrinsics::_putInt:
474 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
475 case vmIntrinsics::_putLong:
476 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
477 case vmIntrinsics::_putFloat:
478 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
479 case vmIntrinsics::_putDouble:
480 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
482 case vmIntrinsics::_getByte_raw:
483 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
484 case vmIntrinsics::_getShort_raw:
485 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
486 case vmIntrinsics::_getChar_raw:
487 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
488 case vmIntrinsics::_getInt_raw:
489 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
490 case vmIntrinsics::_getLong_raw:
491 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
492 case vmIntrinsics::_getFloat_raw:
493 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
494 case vmIntrinsics::_getDouble_raw:
495 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
496 case vmIntrinsics::_getAddress_raw:
497 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
499 case vmIntrinsics::_putByte_raw:
500 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
501 case vmIntrinsics::_putShort_raw:
502 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
503 case vmIntrinsics::_putChar_raw:
504 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
505 case vmIntrinsics::_putInt_raw:
506 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
507 case vmIntrinsics::_putLong_raw:
508 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
509 case vmIntrinsics::_putFloat_raw:
510 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
511 case vmIntrinsics::_putDouble_raw:
512 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
513 case vmIntrinsics::_putAddress_raw:
514 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
516 case vmIntrinsics::_getObjectVolatile:
517 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
518 case vmIntrinsics::_getBooleanVolatile:
519 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
520 case vmIntrinsics::_getByteVolatile:
521 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
522 case vmIntrinsics::_getShortVolatile:
523 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
524 case vmIntrinsics::_getCharVolatile:
525 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
526 case vmIntrinsics::_getIntVolatile:
527 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
528 case vmIntrinsics::_getLongVolatile:
529 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
530 case vmIntrinsics::_getFloatVolatile:
531 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
532 case vmIntrinsics::_getDoubleVolatile:
533 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
535 case vmIntrinsics::_putObjectVolatile:
536 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
537 case vmIntrinsics::_putBooleanVolatile:
538 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
539 case vmIntrinsics::_putByteVolatile:
540 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
541 case vmIntrinsics::_putShortVolatile:
542 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
543 case vmIntrinsics::_putCharVolatile:
544 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
545 case vmIntrinsics::_putIntVolatile:
546 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
547 case vmIntrinsics::_putLongVolatile:
548 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
549 case vmIntrinsics::_putFloatVolatile:
550 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
551 case vmIntrinsics::_putDoubleVolatile:
552 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
554 case vmIntrinsics::_prefetchRead:
555 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
556 case vmIntrinsics::_prefetchWrite:
557 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
558 case vmIntrinsics::_prefetchReadStatic:
559 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
560 case vmIntrinsics::_prefetchWriteStatic:
561 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
563 case vmIntrinsics::_compareAndSwapObject:
564 return inline_unsafe_CAS(T_OBJECT);
565 case vmIntrinsics::_compareAndSwapInt:
566 return inline_unsafe_CAS(T_INT);
567 case vmIntrinsics::_compareAndSwapLong:
568 return inline_unsafe_CAS(T_LONG);
570 case vmIntrinsics::_putOrderedObject:
571 return inline_unsafe_ordered_store(T_OBJECT);
572 case vmIntrinsics::_putOrderedInt:
573 return inline_unsafe_ordered_store(T_INT);
574 case vmIntrinsics::_putOrderedLong:
575 return inline_unsafe_ordered_store(T_LONG);
577 case vmIntrinsics::_currentThread:
578 return inline_native_currentThread();
579 case vmIntrinsics::_isInterrupted:
580 return inline_native_isInterrupted();
582 case vmIntrinsics::_currentTimeMillis:
583 return inline_native_time_funcs(false);
584 case vmIntrinsics::_nanoTime:
585 return inline_native_time_funcs(true);
586 case vmIntrinsics::_allocateInstance:
587 return inline_unsafe_allocate();
588 case vmIntrinsics::_copyMemory:
589 return inline_unsafe_copyMemory();
590 case vmIntrinsics::_newArray:
591 return inline_native_newArray();
592 case vmIntrinsics::_getLength:
593 return inline_native_getLength();
594 case vmIntrinsics::_copyOf:
595 return inline_array_copyOf(false);
596 case vmIntrinsics::_copyOfRange:
597 return inline_array_copyOf(true);
598 case vmIntrinsics::_equalsC:
599 return inline_array_equals();
600 case vmIntrinsics::_clone:
601 return inline_native_clone(intrinsic()->is_virtual());
603 case vmIntrinsics::_isAssignableFrom:
604 return inline_native_subtype_check();
606 case vmIntrinsics::_isInstance:
607 case vmIntrinsics::_getModifiers:
608 case vmIntrinsics::_isInterface:
609 case vmIntrinsics::_isArray:
610 case vmIntrinsics::_isPrimitive:
611 case vmIntrinsics::_getSuperclass:
612 case vmIntrinsics::_getComponentType:
613 case vmIntrinsics::_getClassAccessFlags:
614 return inline_native_Class_query(intrinsic_id());
616 case vmIntrinsics::_floatToRawIntBits:
617 case vmIntrinsics::_floatToIntBits:
618 case vmIntrinsics::_intBitsToFloat:
619 case vmIntrinsics::_doubleToRawLongBits:
620 case vmIntrinsics::_doubleToLongBits:
621 case vmIntrinsics::_longBitsToDouble:
622 return inline_fp_conversions(intrinsic_id());
624 case vmIntrinsics::_numberOfLeadingZeros_i:
625 case vmIntrinsics::_numberOfLeadingZeros_l:
626 return inline_numberOfLeadingZeros(intrinsic_id());
628 case vmIntrinsics::_numberOfTrailingZeros_i:
629 case vmIntrinsics::_numberOfTrailingZeros_l:
630 return inline_numberOfTrailingZeros(intrinsic_id());
632 case vmIntrinsics::_bitCount_i:
633 case vmIntrinsics::_bitCount_l:
634 return inline_bitCount(intrinsic_id());
636 case vmIntrinsics::_reverseBytes_i:
637 case vmIntrinsics::_reverseBytes_l:
638 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
640 case vmIntrinsics::_get_AtomicLong:
641 return inline_native_AtomicLong_get();
642 case vmIntrinsics::_attemptUpdate:
643 return inline_native_AtomicLong_attemptUpdate();
645 case vmIntrinsics::_getCallerClass:
646 return inline_native_Reflection_getCallerClass();
648 default:
649 // If you get here, it may be that someone has added a new intrinsic
650 // to the list in vmSymbols.hpp without implementing it here.
651 #ifndef PRODUCT
652 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
653 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
654 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
655 }
656 #endif
657 return false;
658 }
659 }
661 //------------------------------push_result------------------------------
662 // Helper function for finishing intrinsics.
663 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
664 record_for_igvn(region);
665 set_control(_gvn.transform(region));
666 BasicType value_type = value->type()->basic_type();
667 push_node(value_type, _gvn.transform(value));
668 }
670 //------------------------------generate_guard---------------------------
671 // Helper function for generating guarded fast-slow graph structures.
672 // The given 'test', if true, guards a slow path. If the test fails
673 // then a fast path can be taken. (We generally hope it fails.)
674 // In all cases, GraphKit::control() is updated to the fast path.
675 // The returned value represents the control for the slow path.
676 // The return value is never 'top'; it is either a valid control
677 // or NULL if it is obvious that the slow path can never be taken.
678 // Also, if region and the slow control are not NULL, the slow edge
679 // is appended to the region.
680 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
681 if (stopped()) {
682 // Already short circuited.
683 return NULL;
684 }
686 // Build an if node and its projections.
687 // If test is true we take the slow path, which we assume is uncommon.
688 if (_gvn.type(test) == TypeInt::ZERO) {
689 // The slow branch is never taken. No need to build this guard.
690 return NULL;
691 }
693 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
695 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
696 if (if_slow == top()) {
697 // The slow branch is never taken. No need to build this guard.
698 return NULL;
699 }
701 if (region != NULL)
702 region->add_req(if_slow);
704 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
705 set_control(if_fast);
707 return if_slow;
708 }
710 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
711 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
712 }
713 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
714 return generate_guard(test, region, PROB_FAIR);
715 }
717 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
718 Node* *pos_index) {
719 if (stopped())
720 return NULL; // already stopped
721 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
722 return NULL; // index is already adequately typed
723 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
724 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
725 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
726 if (is_neg != NULL && pos_index != NULL) {
727 // Emulate effect of Parse::adjust_map_after_if.
728 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
729 ccast->set_req(0, control());
730 (*pos_index) = _gvn.transform(ccast);
731 }
732 return is_neg;
733 }
735 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
736 Node* *pos_index) {
737 if (stopped())
738 return NULL; // already stopped
739 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
740 return NULL; // index is already adequately typed
741 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
742 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
743 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
744 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
745 if (is_notp != NULL && pos_index != NULL) {
746 // Emulate effect of Parse::adjust_map_after_if.
747 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
748 ccast->set_req(0, control());
749 (*pos_index) = _gvn.transform(ccast);
750 }
751 return is_notp;
752 }
754 // Make sure that 'position' is a valid limit index, in [0..length].
755 // There are two equivalent plans for checking this:
756 // A. (offset + copyLength) unsigned<= arrayLength
757 // B. offset <= (arrayLength - copyLength)
758 // We require that all of the values above, except for the sum and
759 // difference, are already known to be non-negative.
760 // Plan A is robust in the face of overflow, if offset and copyLength
761 // are both hugely positive.
762 //
763 // Plan B is less direct and intuitive, but it does not overflow at
764 // all, since the difference of two non-negatives is always
765 // representable. Whenever Java methods must perform the equivalent
766 // check they generally use Plan B instead of Plan A.
767 // For the moment we use Plan A.
768 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
769 Node* subseq_length,
770 Node* array_length,
771 RegionNode* region) {
772 if (stopped())
773 return NULL; // already stopped
774 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
775 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
776 return NULL; // common case of whole-array copy
777 Node* last = subseq_length;
778 if (!zero_offset) // last += offset
779 last = _gvn.transform( new (C, 3) AddINode(last, offset));
780 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
781 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
782 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
783 return is_over;
784 }
787 //--------------------------generate_current_thread--------------------
788 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
789 ciKlass* thread_klass = env()->Thread_klass();
790 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
791 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
792 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
793 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
794 tls_output = thread;
795 return threadObj;
796 }
799 //------------------------------inline_string_compareTo------------------------
800 bool LibraryCallKit::inline_string_compareTo() {
802 if (!Matcher::has_match_rule(Op_StrComp)) return false;
804 const int value_offset = java_lang_String::value_offset_in_bytes();
805 const int count_offset = java_lang_String::count_offset_in_bytes();
806 const int offset_offset = java_lang_String::offset_offset_in_bytes();
808 _sp += 2;
809 Node *argument = pop(); // pop non-receiver first: it was pushed second
810 Node *receiver = pop();
812 // Null check on self without removing any arguments. The argument
813 // null check technically happens in the wrong place, which can lead to
814 // invalid stack traces when string compare is inlined into a method
815 // which handles NullPointerExceptions.
816 _sp += 2;
817 receiver = do_null_check(receiver, T_OBJECT);
818 argument = do_null_check(argument, T_OBJECT);
819 _sp -= 2;
820 if (stopped()) {
821 return true;
822 }
824 ciInstanceKlass* klass = env()->String_klass();
825 const TypeInstPtr* string_type =
826 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
828 Node* compare =
829 _gvn.transform(new (C, 7) StrCompNode(
830 control(),
831 memory(TypeAryPtr::CHARS),
832 memory(string_type->add_offset(value_offset)),
833 memory(string_type->add_offset(count_offset)),
834 memory(string_type->add_offset(offset_offset)),
835 receiver,
836 argument));
837 push(compare);
838 return true;
839 }
841 //------------------------------inline_string_equals------------------------
842 bool LibraryCallKit::inline_string_equals() {
844 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
846 const int value_offset = java_lang_String::value_offset_in_bytes();
847 const int count_offset = java_lang_String::count_offset_in_bytes();
848 const int offset_offset = java_lang_String::offset_offset_in_bytes();
850 _sp += 2;
851 Node* argument = pop(); // pop non-receiver first: it was pushed second
852 Node* receiver = pop();
854 // Null check on self without removing any arguments. The argument
855 // null check technically happens in the wrong place, which can lead to
856 // invalid stack traces when string compare is inlined into a method
857 // which handles NullPointerExceptions.
858 _sp += 2;
859 receiver = do_null_check(receiver, T_OBJECT);
860 //should not do null check for argument for String.equals(), because spec
861 //allows to specify NULL as argument.
862 _sp -= 2;
864 if (stopped()) {
865 return true;
866 }
868 // get String klass for instanceOf
869 ciInstanceKlass* klass = env()->String_klass();
871 // two paths (plus control) merge
872 RegionNode* region = new (C, 3) RegionNode(3);
873 Node* phi = new (C, 3) PhiNode(region, TypeInt::BOOL);
875 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
876 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
877 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
879 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
881 Node* if_true = _gvn.transform(new (C, 1) IfTrueNode(iff));
882 set_control(if_true);
884 const TypeInstPtr* string_type =
885 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
887 // instanceOf == true
888 Node* equals =
889 _gvn.transform(new (C, 7) StrEqualsNode(
890 control(),
891 memory(TypeAryPtr::CHARS),
892 memory(string_type->add_offset(value_offset)),
893 memory(string_type->add_offset(count_offset)),
894 memory(string_type->add_offset(offset_offset)),
895 receiver,
896 argument));
898 phi->init_req(1, _gvn.transform(equals));
899 region->init_req(1, if_true);
901 //instanceOf == false, fallthrough
902 Node* if_false = _gvn.transform(new (C, 1) IfFalseNode(iff));
903 set_control(if_false);
905 phi->init_req(2, _gvn.transform(intcon(0)));
906 region->init_req(2, if_false);
908 // post merge
909 set_control(_gvn.transform(region));
910 record_for_igvn(region);
912 push(_gvn.transform(phi));
914 return true;
915 }
917 //------------------------------inline_array_equals----------------------------
918 bool LibraryCallKit::inline_array_equals() {
920 if (!Matcher::has_match_rule(Op_AryEq)) return false;
922 _sp += 2;
923 Node *argument2 = pop();
924 Node *argument1 = pop();
926 Node* equals =
927 _gvn.transform(new (C, 3) AryEqNode(control(),
928 argument1,
929 argument2)
930 );
931 push(equals);
932 return true;
933 }
935 // Java version of String.indexOf(constant string)
936 // class StringDecl {
937 // StringDecl(char[] ca) {
938 // offset = 0;
939 // count = ca.length;
940 // value = ca;
941 // }
942 // int offset;
943 // int count;
944 // char[] value;
945 // }
946 //
947 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
948 // int targetOffset, int cache_i, int md2) {
949 // int cache = cache_i;
950 // int sourceOffset = string_object.offset;
951 // int sourceCount = string_object.count;
952 // int targetCount = target_object.length;
953 //
954 // int targetCountLess1 = targetCount - 1;
955 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
956 //
957 // char[] source = string_object.value;
958 // char[] target = target_object;
959 // int lastChar = target[targetCountLess1];
960 //
961 // outer_loop:
962 // for (int i = sourceOffset; i < sourceEnd; ) {
963 // int src = source[i + targetCountLess1];
964 // if (src == lastChar) {
965 // // With random strings and a 4-character alphabet,
966 // // reverse matching at this point sets up 0.8% fewer
967 // // frames, but (paradoxically) makes 0.3% more probes.
968 // // Since those probes are nearer the lastChar probe,
969 // // there is may be a net D$ win with reverse matching.
970 // // But, reversing loop inhibits unroll of inner loop
971 // // for unknown reason. So, does running outer loop from
972 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
973 // for (int j = 0; j < targetCountLess1; j++) {
974 // if (target[targetOffset + j] != source[i+j]) {
975 // if ((cache & (1 << source[i+j])) == 0) {
976 // if (md2 < j+1) {
977 // i += j+1;
978 // continue outer_loop;
979 // }
980 // }
981 // i += md2;
982 // continue outer_loop;
983 // }
984 // }
985 // return i - sourceOffset;
986 // }
987 // if ((cache & (1 << src)) == 0) {
988 // i += targetCountLess1;
989 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
990 // i++;
991 // }
992 // return -1;
993 // }
995 //------------------------------string_indexOf------------------------
996 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
997 jint cache_i, jint md2_i) {
999 Node* no_ctrl = NULL;
1000 float likely = PROB_LIKELY(0.9);
1001 float unlikely = PROB_UNLIKELY(0.9);
1003 const int value_offset = java_lang_String::value_offset_in_bytes();
1004 const int count_offset = java_lang_String::count_offset_in_bytes();
1005 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1007 ciInstanceKlass* klass = env()->String_klass();
1008 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1009 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1011 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1012 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1013 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1014 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1015 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1016 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1018 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
1019 jint target_length = target_array->length();
1020 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1021 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1023 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1024 #define __ kit.
1025 Node* zero = __ ConI(0);
1026 Node* one = __ ConI(1);
1027 Node* cache = __ ConI(cache_i);
1028 Node* md2 = __ ConI(md2_i);
1029 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1030 Node* targetCount = __ ConI(target_length);
1031 Node* targetCountLess1 = __ ConI(target_length - 1);
1032 Node* targetOffset = __ ConI(targetOffset_i);
1033 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1035 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1036 Node* outer_loop = __ make_label(2 /* goto */);
1037 Node* return_ = __ make_label(1);
1039 __ set(rtn,__ ConI(-1));
1040 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
1041 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1042 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1043 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1044 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1045 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
1046 Node* tpj = __ AddI(targetOffset, __ value(j));
1047 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1048 Node* ipj = __ AddI(__ value(i), __ value(j));
1049 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1050 __ if_then(targ, BoolTest::ne, src2); {
1051 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1052 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1053 __ increment(i, __ AddI(__ value(j), one));
1054 __ goto_(outer_loop);
1055 } __ end_if(); __ dead(j);
1056 }__ end_if(); __ dead(j);
1057 __ increment(i, md2);
1058 __ goto_(outer_loop);
1059 }__ end_if();
1060 __ increment(j, one);
1061 }__ end_loop(); __ dead(j);
1062 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1063 __ goto_(return_);
1064 }__ end_if();
1065 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1066 __ increment(i, targetCountLess1);
1067 }__ end_if();
1068 __ increment(i, one);
1069 __ bind(outer_loop);
1070 }__ end_loop(); __ dead(i);
1071 __ bind(return_);
1073 // Final sync IdealKit and GraphKit.
1074 sync_kit(kit);
1075 Node* result = __ value(rtn);
1076 #undef __
1077 C->set_has_loops(true);
1078 return result;
1079 }
1081 //------------------------------inline_string_indexOf------------------------
1082 bool LibraryCallKit::inline_string_indexOf() {
1084 const int value_offset = java_lang_String::value_offset_in_bytes();
1085 const int count_offset = java_lang_String::count_offset_in_bytes();
1086 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1088 _sp += 2;
1089 Node *argument = pop(); // pop non-receiver first: it was pushed second
1090 Node *receiver = pop();
1092 Node* result;
1093 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1094 UseSSE42Intrinsics) {
1095 // Generate SSE4.2 version of indexOf
1096 // We currently only have match rules that use SSE4.2
1098 // Null check on self without removing any arguments. The argument
1099 // null check technically happens in the wrong place, which can lead to
1100 // invalid stack traces when string compare is inlined into a method
1101 // which handles NullPointerExceptions.
1102 _sp += 2;
1103 receiver = do_null_check(receiver, T_OBJECT);
1104 argument = do_null_check(argument, T_OBJECT);
1105 _sp -= 2;
1107 if (stopped()) {
1108 return true;
1109 }
1111 ciInstanceKlass* klass = env()->String_klass();
1112 const TypeInstPtr* string_type =
1113 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1115 result =
1116 _gvn.transform(new (C, 7)
1117 StrIndexOfNode(control(),
1118 memory(TypeAryPtr::CHARS),
1119 memory(string_type->add_offset(value_offset)),
1120 memory(string_type->add_offset(count_offset)),
1121 memory(string_type->add_offset(offset_offset)),
1122 receiver,
1123 argument));
1124 } else { //Use LibraryCallKit::string_indexOf
1125 // don't intrinsify is argument isn't a constant string.
1126 if (!argument->is_Con()) {
1127 return false;
1128 }
1129 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1130 if (str_type == NULL) {
1131 return false;
1132 }
1133 ciInstanceKlass* klass = env()->String_klass();
1134 ciObject* str_const = str_type->const_oop();
1135 if (str_const == NULL || str_const->klass() != klass) {
1136 return false;
1137 }
1138 ciInstance* str = str_const->as_instance();
1139 assert(str != NULL, "must be instance");
1141 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1142 int o = str->field_value_by_offset(offset_offset).as_int();
1143 int c = str->field_value_by_offset(count_offset).as_int();
1144 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1146 // constant strings have no offset and count == length which
1147 // simplifies the resulting code somewhat so lets optimize for that.
1148 if (o != 0 || c != pat->length()) {
1149 return false;
1150 }
1152 // Null check on self without removing any arguments. The argument
1153 // null check technically happens in the wrong place, which can lead to
1154 // invalid stack traces when string compare is inlined into a method
1155 // which handles NullPointerExceptions.
1156 _sp += 2;
1157 receiver = do_null_check(receiver, T_OBJECT);
1158 // No null check on the argument is needed since it's a constant String oop.
1159 _sp -= 2;
1160 if (stopped()) {
1161 return true;
1162 }
1164 // The null string as a pattern always returns 0 (match at beginning of string)
1165 if (c == 0) {
1166 push(intcon(0));
1167 return true;
1168 }
1170 // Generate default indexOf
1171 jchar lastChar = pat->char_at(o + (c - 1));
1172 int cache = 0;
1173 int i;
1174 for (i = 0; i < c - 1; i++) {
1175 assert(i < pat->length(), "out of range");
1176 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1177 }
1179 int md2 = c;
1180 for (i = 0; i < c - 1; i++) {
1181 assert(i < pat->length(), "out of range");
1182 if (pat->char_at(o + i) == lastChar) {
1183 md2 = (c - 1) - i;
1184 }
1185 }
1187 result = string_indexOf(receiver, pat, o, cache, md2);
1188 }
1190 push(result);
1191 return true;
1192 }
1194 //--------------------------pop_math_arg--------------------------------
1195 // Pop a double argument to a math function from the stack
1196 // rounding it if necessary.
1197 Node * LibraryCallKit::pop_math_arg() {
1198 Node *arg = pop_pair();
1199 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1200 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1201 return arg;
1202 }
1204 //------------------------------inline_trig----------------------------------
1205 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1206 // argument reduction which will turn into a fast/slow diamond.
1207 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1208 _sp += arg_size(); // restore stack pointer
1209 Node* arg = pop_math_arg();
1210 Node* trig = NULL;
1212 switch (id) {
1213 case vmIntrinsics::_dsin:
1214 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1215 break;
1216 case vmIntrinsics::_dcos:
1217 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1218 break;
1219 case vmIntrinsics::_dtan:
1220 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1221 break;
1222 default:
1223 assert(false, "bad intrinsic was passed in");
1224 return false;
1225 }
1227 // Rounding required? Check for argument reduction!
1228 if( Matcher::strict_fp_requires_explicit_rounding ) {
1230 static const double pi_4 = 0.7853981633974483;
1231 static const double neg_pi_4 = -0.7853981633974483;
1232 // pi/2 in 80-bit extended precision
1233 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1234 // -pi/2 in 80-bit extended precision
1235 // 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};
1236 // Cutoff value for using this argument reduction technique
1237 //static const double pi_2_minus_epsilon = 1.564660403643354;
1238 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1240 // Pseudocode for sin:
1241 // if (x <= Math.PI / 4.0) {
1242 // if (x >= -Math.PI / 4.0) return fsin(x);
1243 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1244 // } else {
1245 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1246 // }
1247 // return StrictMath.sin(x);
1249 // Pseudocode for cos:
1250 // if (x <= Math.PI / 4.0) {
1251 // if (x >= -Math.PI / 4.0) return fcos(x);
1252 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1253 // } else {
1254 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1255 // }
1256 // return StrictMath.cos(x);
1258 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1259 // requires a special machine instruction to load it. Instead we'll try
1260 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1261 // probably do the math inside the SIN encoding.
1263 // Make the merge point
1264 RegionNode *r = new (C, 3) RegionNode(3);
1265 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1267 // Flatten arg so we need only 1 test
1268 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1269 // Node for PI/4 constant
1270 Node *pi4 = makecon(TypeD::make(pi_4));
1271 // Check PI/4 : abs(arg)
1272 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1273 // Check: If PI/4 < abs(arg) then go slow
1274 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1275 // Branch either way
1276 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1277 set_control(opt_iff(r,iff));
1279 // Set fast path result
1280 phi->init_req(2,trig);
1282 // Slow path - non-blocking leaf call
1283 Node* call = NULL;
1284 switch (id) {
1285 case vmIntrinsics::_dsin:
1286 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1287 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1288 "Sin", NULL, arg, top());
1289 break;
1290 case vmIntrinsics::_dcos:
1291 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1292 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1293 "Cos", NULL, arg, top());
1294 break;
1295 case vmIntrinsics::_dtan:
1296 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1297 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1298 "Tan", NULL, arg, top());
1299 break;
1300 }
1301 assert(control()->in(0) == call, "");
1302 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1303 r->init_req(1,control());
1304 phi->init_req(1,slow_result);
1306 // Post-merge
1307 set_control(_gvn.transform(r));
1308 record_for_igvn(r);
1309 trig = _gvn.transform(phi);
1311 C->set_has_split_ifs(true); // Has chance for split-if optimization
1312 }
1313 // Push result back on JVM stack
1314 push_pair(trig);
1315 return true;
1316 }
1318 //------------------------------inline_sqrt-------------------------------------
1319 // Inline square root instruction, if possible.
1320 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1321 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1322 _sp += arg_size(); // restore stack pointer
1323 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1324 return true;
1325 }
1327 //------------------------------inline_abs-------------------------------------
1328 // Inline absolute value instruction, if possible.
1329 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1330 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1331 _sp += arg_size(); // restore stack pointer
1332 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1333 return true;
1334 }
1336 //------------------------------inline_exp-------------------------------------
1337 // Inline exp instructions, if possible. The Intel hardware only misses
1338 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1339 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1340 assert(id == vmIntrinsics::_dexp, "Not exp");
1342 // If this inlining ever returned NaN in the past, we do not intrinsify it
1343 // every again. NaN results requires StrictMath.exp handling.
1344 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1346 // Do not intrinsify on older platforms which lack cmove.
1347 if (ConditionalMoveLimit == 0) return false;
1349 _sp += arg_size(); // restore stack pointer
1350 Node *x = pop_math_arg();
1351 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1353 //-------------------
1354 //result=(result.isNaN())? StrictMath::exp():result;
1355 // Check: If isNaN() by checking result!=result? then go to Strict Math
1356 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1357 // Build the boolean node
1358 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1360 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1361 // End the current control-flow path
1362 push_pair(x);
1363 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1364 // to handle. Recompile without intrinsifying Math.exp
1365 uncommon_trap(Deoptimization::Reason_intrinsic,
1366 Deoptimization::Action_make_not_entrant);
1367 }
1369 C->set_has_split_ifs(true); // Has chance for split-if optimization
1371 push_pair(result);
1373 return true;
1374 }
1376 //------------------------------inline_pow-------------------------------------
1377 // Inline power instructions, if possible.
1378 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1379 assert(id == vmIntrinsics::_dpow, "Not pow");
1381 // If this inlining ever returned NaN in the past, we do not intrinsify it
1382 // every again. NaN results requires StrictMath.pow handling.
1383 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1385 // Do not intrinsify on older platforms which lack cmove.
1386 if (ConditionalMoveLimit == 0) return false;
1388 // Pseudocode for pow
1389 // if (x <= 0.0) {
1390 // if ((double)((int)y)==y) { // if y is int
1391 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1392 // } else {
1393 // result = NaN;
1394 // }
1395 // } else {
1396 // result = DPow(x,y);
1397 // }
1398 // if (result != result)? {
1399 // uncommon_trap();
1400 // }
1401 // return result;
1403 _sp += arg_size(); // restore stack pointer
1404 Node* y = pop_math_arg();
1405 Node* x = pop_math_arg();
1407 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1409 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1410 // inside of something) then skip the fancy tests and just check for
1411 // NaN result.
1412 Node *result = NULL;
1413 if( jvms()->depth() >= 1 ) {
1414 result = fast_result;
1415 } else {
1417 // Set the merge point for If node with condition of (x <= 0.0)
1418 // There are four possible paths to region node and phi node
1419 RegionNode *r = new (C, 4) RegionNode(4);
1420 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1422 // Build the first if node: if (x <= 0.0)
1423 // Node for 0 constant
1424 Node *zeronode = makecon(TypeD::ZERO);
1425 // Check x:0
1426 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1427 // Check: If (x<=0) then go complex path
1428 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1429 // Branch either way
1430 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1431 Node *opt_test = _gvn.transform(if1);
1432 //assert( opt_test->is_If(), "Expect an IfNode");
1433 IfNode *opt_if1 = (IfNode*)opt_test;
1434 // Fast path taken; set region slot 3
1435 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1436 r->init_req(3,fast_taken); // Capture fast-control
1438 // Fast path not-taken, i.e. slow path
1439 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1441 // Set fast path result
1442 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1443 phi->init_req(3, fast_result);
1445 // Complex path
1446 // Build the second if node (if y is int)
1447 // Node for (int)y
1448 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1449 // Node for (double)((int) y)
1450 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1451 // Check (double)((int) y) : y
1452 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1453 // Check if (y isn't int) then go to slow path
1455 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1456 // Branch either way
1457 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1458 Node *slow_path = opt_iff(r,if2); // Set region path 2
1460 // Calculate DPow(abs(x), y)*(1 & (int)y)
1461 // Node for constant 1
1462 Node *conone = intcon(1);
1463 // 1& (int)y
1464 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1465 // zero node
1466 Node *conzero = intcon(0);
1467 // Check (1&(int)y)==0?
1468 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1469 // Check if (1&(int)y)!=0?, if so the result is negative
1470 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1471 // abs(x)
1472 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1473 // abs(x)^y
1474 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1475 // -abs(x)^y
1476 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1477 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1478 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1479 // Set complex path fast result
1480 phi->init_req(2, signresult);
1482 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1483 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1484 r->init_req(1,slow_path);
1485 phi->init_req(1,slow_result);
1487 // Post merge
1488 set_control(_gvn.transform(r));
1489 record_for_igvn(r);
1490 result=_gvn.transform(phi);
1491 }
1493 //-------------------
1494 //result=(result.isNaN())? uncommon_trap():result;
1495 // Check: If isNaN() by checking result!=result? then go to Strict Math
1496 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1497 // Build the boolean node
1498 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1500 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1501 // End the current control-flow path
1502 push_pair(x);
1503 push_pair(y);
1504 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1505 // to handle. Recompile without intrinsifying Math.pow.
1506 uncommon_trap(Deoptimization::Reason_intrinsic,
1507 Deoptimization::Action_make_not_entrant);
1508 }
1510 C->set_has_split_ifs(true); // Has chance for split-if optimization
1512 push_pair(result);
1514 return true;
1515 }
1517 //------------------------------inline_trans-------------------------------------
1518 // Inline transcendental instructions, if possible. The Intel hardware gets
1519 // these right, no funny corner cases missed.
1520 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1521 _sp += arg_size(); // restore stack pointer
1522 Node* arg = pop_math_arg();
1523 Node* trans = NULL;
1525 switch (id) {
1526 case vmIntrinsics::_dlog:
1527 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1528 break;
1529 case vmIntrinsics::_dlog10:
1530 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1531 break;
1532 default:
1533 assert(false, "bad intrinsic was passed in");
1534 return false;
1535 }
1537 // Push result back on JVM stack
1538 push_pair(trans);
1539 return true;
1540 }
1542 //------------------------------runtime_math-----------------------------
1543 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1544 Node* a = NULL;
1545 Node* b = NULL;
1547 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1548 "must be (DD)D or (D)D type");
1550 // Inputs
1551 _sp += arg_size(); // restore stack pointer
1552 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1553 b = pop_math_arg();
1554 }
1555 a = pop_math_arg();
1557 const TypePtr* no_memory_effects = NULL;
1558 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1559 no_memory_effects,
1560 a, top(), b, b ? top() : NULL);
1561 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1562 #ifdef ASSERT
1563 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1564 assert(value_top == top(), "second value must be top");
1565 #endif
1567 push_pair(value);
1568 return true;
1569 }
1571 //------------------------------inline_math_native-----------------------------
1572 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1573 switch (id) {
1574 // These intrinsics are not properly supported on all hardware
1575 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1576 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1577 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1578 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1579 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1580 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1582 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1583 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1584 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1585 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1587 // These intrinsics are supported on all hardware
1588 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1589 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1591 // These intrinsics don't work on X86. The ad implementation doesn't
1592 // handle NaN's properly. Instead of returning infinity, the ad
1593 // implementation returns a NaN on overflow. See bug: 6304089
1594 // Once the ad implementations are fixed, change the code below
1595 // to match the intrinsics above
1597 case vmIntrinsics::_dexp: return
1598 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1599 case vmIntrinsics::_dpow: return
1600 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1602 // These intrinsics are not yet correctly implemented
1603 case vmIntrinsics::_datan2:
1604 return false;
1606 default:
1607 ShouldNotReachHere();
1608 return false;
1609 }
1610 }
1612 static bool is_simple_name(Node* n) {
1613 return (n->req() == 1 // constant
1614 || (n->is_Type() && n->as_Type()->type()->singleton())
1615 || n->is_Proj() // parameter or return value
1616 || n->is_Phi() // local of some sort
1617 );
1618 }
1620 //----------------------------inline_min_max-----------------------------------
1621 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1622 push(generate_min_max(id, argument(0), argument(1)));
1624 return true;
1625 }
1627 Node*
1628 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1629 // These are the candidate return value:
1630 Node* xvalue = x0;
1631 Node* yvalue = y0;
1633 if (xvalue == yvalue) {
1634 return xvalue;
1635 }
1637 bool want_max = (id == vmIntrinsics::_max);
1639 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1640 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1641 if (txvalue == NULL || tyvalue == NULL) return top();
1642 // This is not really necessary, but it is consistent with a
1643 // hypothetical MaxINode::Value method:
1644 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1646 // %%% This folding logic should (ideally) be in a different place.
1647 // Some should be inside IfNode, and there to be a more reliable
1648 // transformation of ?: style patterns into cmoves. We also want
1649 // more powerful optimizations around cmove and min/max.
1651 // Try to find a dominating comparison of these guys.
1652 // It can simplify the index computation for Arrays.copyOf
1653 // and similar uses of System.arraycopy.
1654 // First, compute the normalized version of CmpI(x, y).
1655 int cmp_op = Op_CmpI;
1656 Node* xkey = xvalue;
1657 Node* ykey = yvalue;
1658 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1659 if (ideal_cmpxy->is_Cmp()) {
1660 // E.g., if we have CmpI(length - offset, count),
1661 // it might idealize to CmpI(length, count + offset)
1662 cmp_op = ideal_cmpxy->Opcode();
1663 xkey = ideal_cmpxy->in(1);
1664 ykey = ideal_cmpxy->in(2);
1665 }
1667 // Start by locating any relevant comparisons.
1668 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1669 Node* cmpxy = NULL;
1670 Node* cmpyx = NULL;
1671 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1672 Node* cmp = start_from->fast_out(k);
1673 if (cmp->outcnt() > 0 && // must have prior uses
1674 cmp->in(0) == NULL && // must be context-independent
1675 cmp->Opcode() == cmp_op) { // right kind of compare
1676 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1677 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1678 }
1679 }
1681 const int NCMPS = 2;
1682 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1683 int cmpn;
1684 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1685 if (cmps[cmpn] != NULL) break; // find a result
1686 }
1687 if (cmpn < NCMPS) {
1688 // Look for a dominating test that tells us the min and max.
1689 int depth = 0; // Limit search depth for speed
1690 Node* dom = control();
1691 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1692 if (++depth >= 100) break;
1693 Node* ifproj = dom;
1694 if (!ifproj->is_Proj()) continue;
1695 Node* iff = ifproj->in(0);
1696 if (!iff->is_If()) continue;
1697 Node* bol = iff->in(1);
1698 if (!bol->is_Bool()) continue;
1699 Node* cmp = bol->in(1);
1700 if (cmp == NULL) continue;
1701 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1702 if (cmps[cmpn] == cmp) break;
1703 if (cmpn == NCMPS) continue;
1704 BoolTest::mask btest = bol->as_Bool()->_test._test;
1705 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1706 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1707 // At this point, we know that 'x btest y' is true.
1708 switch (btest) {
1709 case BoolTest::eq:
1710 // They are proven equal, so we can collapse the min/max.
1711 // Either value is the answer. Choose the simpler.
1712 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1713 return yvalue;
1714 return xvalue;
1715 case BoolTest::lt: // x < y
1716 case BoolTest::le: // x <= y
1717 return (want_max ? yvalue : xvalue);
1718 case BoolTest::gt: // x > y
1719 case BoolTest::ge: // x >= y
1720 return (want_max ? xvalue : yvalue);
1721 }
1722 }
1723 }
1725 // We failed to find a dominating test.
1726 // Let's pick a test that might GVN with prior tests.
1727 Node* best_bol = NULL;
1728 BoolTest::mask best_btest = BoolTest::illegal;
1729 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1730 Node* cmp = cmps[cmpn];
1731 if (cmp == NULL) continue;
1732 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1733 Node* bol = cmp->fast_out(j);
1734 if (!bol->is_Bool()) continue;
1735 BoolTest::mask btest = bol->as_Bool()->_test._test;
1736 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1737 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1738 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1739 best_bol = bol->as_Bool();
1740 best_btest = btest;
1741 }
1742 }
1743 }
1745 Node* answer_if_true = NULL;
1746 Node* answer_if_false = NULL;
1747 switch (best_btest) {
1748 default:
1749 if (cmpxy == NULL)
1750 cmpxy = ideal_cmpxy;
1751 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1752 // and fall through:
1753 case BoolTest::lt: // x < y
1754 case BoolTest::le: // x <= y
1755 answer_if_true = (want_max ? yvalue : xvalue);
1756 answer_if_false = (want_max ? xvalue : yvalue);
1757 break;
1758 case BoolTest::gt: // x > y
1759 case BoolTest::ge: // x >= y
1760 answer_if_true = (want_max ? xvalue : yvalue);
1761 answer_if_false = (want_max ? yvalue : xvalue);
1762 break;
1763 }
1765 jint hi, lo;
1766 if (want_max) {
1767 // We can sharpen the minimum.
1768 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1769 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1770 } else {
1771 // We can sharpen the maximum.
1772 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1773 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1774 }
1776 // Use a flow-free graph structure, to avoid creating excess control edges
1777 // which could hinder other optimizations.
1778 // Since Math.min/max is often used with arraycopy, we want
1779 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1780 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1781 answer_if_false, answer_if_true,
1782 TypeInt::make(lo, hi, widen));
1784 return _gvn.transform(cmov);
1786 /*
1787 // This is not as desirable as it may seem, since Min and Max
1788 // nodes do not have a full set of optimizations.
1789 // And they would interfere, anyway, with 'if' optimizations
1790 // and with CMoveI canonical forms.
1791 switch (id) {
1792 case vmIntrinsics::_min:
1793 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1794 case vmIntrinsics::_max:
1795 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1796 default:
1797 ShouldNotReachHere();
1798 }
1799 */
1800 }
1802 inline int
1803 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1804 const TypePtr* base_type = TypePtr::NULL_PTR;
1805 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1806 if (base_type == NULL) {
1807 // Unknown type.
1808 return Type::AnyPtr;
1809 } else if (base_type == TypePtr::NULL_PTR) {
1810 // Since this is a NULL+long form, we have to switch to a rawptr.
1811 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1812 offset = MakeConX(0);
1813 return Type::RawPtr;
1814 } else if (base_type->base() == Type::RawPtr) {
1815 return Type::RawPtr;
1816 } else if (base_type->isa_oopptr()) {
1817 // Base is never null => always a heap address.
1818 if (base_type->ptr() == TypePtr::NotNull) {
1819 return Type::OopPtr;
1820 }
1821 // Offset is small => always a heap address.
1822 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1823 if (offset_type != NULL &&
1824 base_type->offset() == 0 && // (should always be?)
1825 offset_type->_lo >= 0 &&
1826 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1827 return Type::OopPtr;
1828 }
1829 // Otherwise, it might either be oop+off or NULL+addr.
1830 return Type::AnyPtr;
1831 } else {
1832 // No information:
1833 return Type::AnyPtr;
1834 }
1835 }
1837 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1838 int kind = classify_unsafe_addr(base, offset);
1839 if (kind == Type::RawPtr) {
1840 return basic_plus_adr(top(), base, offset);
1841 } else {
1842 return basic_plus_adr(base, offset);
1843 }
1844 }
1846 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1847 // inline int Integer.numberOfLeadingZeros(int)
1848 // inline int Long.numberOfLeadingZeros(long)
1849 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1850 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1851 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1852 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1853 _sp += arg_size(); // restore stack pointer
1854 switch (id) {
1855 case vmIntrinsics::_numberOfLeadingZeros_i:
1856 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1857 break;
1858 case vmIntrinsics::_numberOfLeadingZeros_l:
1859 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1860 break;
1861 default:
1862 ShouldNotReachHere();
1863 }
1864 return true;
1865 }
1867 //-------------------inline_numberOfTrailingZeros_int/long----------------------
1868 // inline int Integer.numberOfTrailingZeros(int)
1869 // inline int Long.numberOfTrailingZeros(long)
1870 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
1871 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
1872 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
1873 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
1874 _sp += arg_size(); // restore stack pointer
1875 switch (id) {
1876 case vmIntrinsics::_numberOfTrailingZeros_i:
1877 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
1878 break;
1879 case vmIntrinsics::_numberOfTrailingZeros_l:
1880 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
1881 break;
1882 default:
1883 ShouldNotReachHere();
1884 }
1885 return true;
1886 }
1888 //----------------------------inline_bitCount_int/long-----------------------
1889 // inline int Integer.bitCount(int)
1890 // inline int Long.bitCount(long)
1891 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
1892 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
1893 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
1894 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
1895 _sp += arg_size(); // restore stack pointer
1896 switch (id) {
1897 case vmIntrinsics::_bitCount_i:
1898 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
1899 break;
1900 case vmIntrinsics::_bitCount_l:
1901 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
1902 break;
1903 default:
1904 ShouldNotReachHere();
1905 }
1906 return true;
1907 }
1909 //----------------------------inline_reverseBytes_int/long-------------------
1910 // inline Integer.reverseBytes(int)
1911 // inline Long.reverseBytes(long)
1912 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
1913 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
1914 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
1915 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
1916 _sp += arg_size(); // restore stack pointer
1917 switch (id) {
1918 case vmIntrinsics::_reverseBytes_i:
1919 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
1920 break;
1921 case vmIntrinsics::_reverseBytes_l:
1922 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
1923 break;
1924 default:
1925 ;
1926 }
1927 return true;
1928 }
1930 //----------------------------inline_unsafe_access----------------------------
1932 const static BasicType T_ADDRESS_HOLDER = T_LONG;
1934 // Interpret Unsafe.fieldOffset cookies correctly:
1935 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
1937 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
1938 if (callee()->is_static()) return false; // caller must have the capability!
1940 #ifndef PRODUCT
1941 {
1942 ResourceMark rm;
1943 // Check the signatures.
1944 ciSignature* sig = signature();
1945 #ifdef ASSERT
1946 if (!is_store) {
1947 // Object getObject(Object base, int/long offset), etc.
1948 BasicType rtype = sig->return_type()->basic_type();
1949 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
1950 rtype = T_ADDRESS; // it is really a C void*
1951 assert(rtype == type, "getter must return the expected value");
1952 if (!is_native_ptr) {
1953 assert(sig->count() == 2, "oop getter has 2 arguments");
1954 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
1955 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
1956 } else {
1957 assert(sig->count() == 1, "native getter has 1 argument");
1958 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
1959 }
1960 } else {
1961 // void putObject(Object base, int/long offset, Object x), etc.
1962 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
1963 if (!is_native_ptr) {
1964 assert(sig->count() == 3, "oop putter has 3 arguments");
1965 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
1966 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
1967 } else {
1968 assert(sig->count() == 2, "native putter has 2 arguments");
1969 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
1970 }
1971 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
1972 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
1973 vtype = T_ADDRESS; // it is really a C void*
1974 assert(vtype == type, "putter must accept the expected value");
1975 }
1976 #endif // ASSERT
1977 }
1978 #endif //PRODUCT
1980 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1982 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
1984 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
1985 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
1987 debug_only(int saved_sp = _sp);
1988 _sp += nargs;
1990 Node* val;
1991 debug_only(val = (Node*)(uintptr_t)-1);
1994 if (is_store) {
1995 // Get the value being stored. (Pop it first; it was pushed last.)
1996 switch (type) {
1997 case T_DOUBLE:
1998 case T_LONG:
1999 case T_ADDRESS:
2000 val = pop_pair();
2001 break;
2002 default:
2003 val = pop();
2004 }
2005 }
2007 // Build address expression. See the code in inline_unsafe_prefetch.
2008 Node *adr;
2009 Node *heap_base_oop = top();
2010 if (!is_native_ptr) {
2011 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2012 Node* offset = pop_pair();
2013 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2014 Node* base = pop();
2015 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2016 // to be plain byte offsets, which are also the same as those accepted
2017 // by oopDesc::field_base.
2018 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2019 "fieldOffset must be byte-scaled");
2020 // 32-bit machines ignore the high half!
2021 offset = ConvL2X(offset);
2022 adr = make_unsafe_address(base, offset);
2023 heap_base_oop = base;
2024 } else {
2025 Node* ptr = pop_pair();
2026 // Adjust Java long to machine word:
2027 ptr = ConvL2X(ptr);
2028 adr = make_unsafe_address(NULL, ptr);
2029 }
2031 // Pop receiver last: it was pushed first.
2032 Node *receiver = pop();
2034 assert(saved_sp == _sp, "must have correct argument count");
2036 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2038 // First guess at the value type.
2039 const Type *value_type = Type::get_const_basic_type(type);
2041 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2042 // there was not enough information to nail it down.
2043 Compile::AliasType* alias_type = C->alias_type(adr_type);
2044 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2046 // We will need memory barriers unless we can determine a unique
2047 // alias category for this reference. (Note: If for some reason
2048 // the barriers get omitted and the unsafe reference begins to "pollute"
2049 // the alias analysis of the rest of the graph, either Compile::can_alias
2050 // or Compile::must_alias will throw a diagnostic assert.)
2051 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2053 if (!is_store && type == T_OBJECT) {
2054 // Attempt to infer a sharper value type from the offset and base type.
2055 ciKlass* sharpened_klass = NULL;
2057 // See if it is an instance field, with an object type.
2058 if (alias_type->field() != NULL) {
2059 assert(!is_native_ptr, "native pointer op cannot use a java address");
2060 if (alias_type->field()->type()->is_klass()) {
2061 sharpened_klass = alias_type->field()->type()->as_klass();
2062 }
2063 }
2065 // See if it is a narrow oop array.
2066 if (adr_type->isa_aryptr()) {
2067 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2068 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2069 if (elem_type != NULL) {
2070 sharpened_klass = elem_type->klass();
2071 }
2072 }
2073 }
2075 if (sharpened_klass != NULL) {
2076 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2078 // Sharpen the value type.
2079 value_type = tjp;
2081 #ifndef PRODUCT
2082 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2083 tty->print(" from base type: "); adr_type->dump();
2084 tty->print(" sharpened value: "); value_type->dump();
2085 }
2086 #endif
2087 }
2088 }
2090 // Null check on self without removing any arguments. The argument
2091 // null check technically happens in the wrong place, which can lead to
2092 // invalid stack traces when the primitive is inlined into a method
2093 // which handles NullPointerExceptions.
2094 _sp += nargs;
2095 do_null_check(receiver, T_OBJECT);
2096 _sp -= nargs;
2097 if (stopped()) {
2098 return true;
2099 }
2100 // Heap pointers get a null-check from the interpreter,
2101 // as a courtesy. However, this is not guaranteed by Unsafe,
2102 // and it is not possible to fully distinguish unintended nulls
2103 // from intended ones in this API.
2105 if (is_volatile) {
2106 // We need to emit leading and trailing CPU membars (see below) in
2107 // addition to memory membars when is_volatile. This is a little
2108 // too strong, but avoids the need to insert per-alias-type
2109 // volatile membars (for stores; compare Parse::do_put_xxx), which
2110 // we cannot do effectively here because we probably only have a
2111 // rough approximation of type.
2112 need_mem_bar = true;
2113 // For Stores, place a memory ordering barrier now.
2114 if (is_store)
2115 insert_mem_bar(Op_MemBarRelease);
2116 }
2118 // Memory barrier to prevent normal and 'unsafe' accesses from
2119 // bypassing each other. Happens after null checks, so the
2120 // exception paths do not take memory state from the memory barrier,
2121 // so there's no problems making a strong assert about mixing users
2122 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2123 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2124 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2126 if (!is_store) {
2127 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2128 // load value and push onto stack
2129 switch (type) {
2130 case T_BOOLEAN:
2131 case T_CHAR:
2132 case T_BYTE:
2133 case T_SHORT:
2134 case T_INT:
2135 case T_FLOAT:
2136 case T_OBJECT:
2137 push( p );
2138 break;
2139 case T_ADDRESS:
2140 // Cast to an int type.
2141 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2142 p = ConvX2L(p);
2143 push_pair(p);
2144 break;
2145 case T_DOUBLE:
2146 case T_LONG:
2147 push_pair( p );
2148 break;
2149 default: ShouldNotReachHere();
2150 }
2151 } else {
2152 // place effect of store into memory
2153 switch (type) {
2154 case T_DOUBLE:
2155 val = dstore_rounding(val);
2156 break;
2157 case T_ADDRESS:
2158 // Repackage the long as a pointer.
2159 val = ConvL2X(val);
2160 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2161 break;
2162 }
2164 if (type != T_OBJECT ) {
2165 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2166 } else {
2167 // Possibly an oop being stored to Java heap or native memory
2168 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2169 // oop to Java heap.
2170 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2171 } else {
2172 // We can't tell at compile time if we are storing in the Java heap or outside
2173 // of it. So we need to emit code to conditionally do the proper type of
2174 // store.
2176 IdealKit ideal(gvn(), control(), merged_memory());
2177 #define __ ideal.
2178 // QQQ who knows what probability is here??
2179 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2180 // Sync IdealKit and graphKit.
2181 set_all_memory( __ merged_memory());
2182 set_control(__ ctrl());
2183 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2184 // Update IdealKit memory.
2185 __ set_all_memory(merged_memory());
2186 __ set_ctrl(control());
2187 } __ else_(); {
2188 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2189 } __ end_if();
2190 // Final sync IdealKit and GraphKit.
2191 sync_kit(ideal);
2192 #undef __
2193 }
2194 }
2195 }
2197 if (is_volatile) {
2198 if (!is_store)
2199 insert_mem_bar(Op_MemBarAcquire);
2200 else
2201 insert_mem_bar(Op_MemBarVolatile);
2202 }
2204 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2206 return true;
2207 }
2209 //----------------------------inline_unsafe_prefetch----------------------------
2211 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2212 #ifndef PRODUCT
2213 {
2214 ResourceMark rm;
2215 // Check the signatures.
2216 ciSignature* sig = signature();
2217 #ifdef ASSERT
2218 // Object getObject(Object base, int/long offset), etc.
2219 BasicType rtype = sig->return_type()->basic_type();
2220 if (!is_native_ptr) {
2221 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2222 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2223 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2224 } else {
2225 assert(sig->count() == 1, "native prefetch has 1 argument");
2226 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2227 }
2228 #endif // ASSERT
2229 }
2230 #endif // !PRODUCT
2232 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2234 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2235 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2237 debug_only(int saved_sp = _sp);
2238 _sp += nargs;
2240 // Build address expression. See the code in inline_unsafe_access.
2241 Node *adr;
2242 if (!is_native_ptr) {
2243 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2244 Node* offset = pop_pair();
2245 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2246 Node* base = pop();
2247 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2248 // to be plain byte offsets, which are also the same as those accepted
2249 // by oopDesc::field_base.
2250 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2251 "fieldOffset must be byte-scaled");
2252 // 32-bit machines ignore the high half!
2253 offset = ConvL2X(offset);
2254 adr = make_unsafe_address(base, offset);
2255 } else {
2256 Node* ptr = pop_pair();
2257 // Adjust Java long to machine word:
2258 ptr = ConvL2X(ptr);
2259 adr = make_unsafe_address(NULL, ptr);
2260 }
2262 if (is_static) {
2263 assert(saved_sp == _sp, "must have correct argument count");
2264 } else {
2265 // Pop receiver last: it was pushed first.
2266 Node *receiver = pop();
2267 assert(saved_sp == _sp, "must have correct argument count");
2269 // Null check on self without removing any arguments. The argument
2270 // null check technically happens in the wrong place, which can lead to
2271 // invalid stack traces when the primitive is inlined into a method
2272 // which handles NullPointerExceptions.
2273 _sp += nargs;
2274 do_null_check(receiver, T_OBJECT);
2275 _sp -= nargs;
2276 if (stopped()) {
2277 return true;
2278 }
2279 }
2281 // Generate the read or write prefetch
2282 Node *prefetch;
2283 if (is_store) {
2284 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2285 } else {
2286 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2287 }
2288 prefetch->init_req(0, control());
2289 set_i_o(_gvn.transform(prefetch));
2291 return true;
2292 }
2294 //----------------------------inline_unsafe_CAS----------------------------
2296 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2297 // This basic scheme here is the same as inline_unsafe_access, but
2298 // differs in enough details that combining them would make the code
2299 // overly confusing. (This is a true fact! I originally combined
2300 // them, but even I was confused by it!) As much code/comments as
2301 // possible are retained from inline_unsafe_access though to make
2302 // the correspondences clearer. - dl
2304 if (callee()->is_static()) return false; // caller must have the capability!
2306 #ifndef PRODUCT
2307 {
2308 ResourceMark rm;
2309 // Check the signatures.
2310 ciSignature* sig = signature();
2311 #ifdef ASSERT
2312 BasicType rtype = sig->return_type()->basic_type();
2313 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2314 assert(sig->count() == 4, "CAS has 4 arguments");
2315 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2316 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2317 #endif // ASSERT
2318 }
2319 #endif //PRODUCT
2321 // number of stack slots per value argument (1 or 2)
2322 int type_words = type2size[type];
2324 // Cannot inline wide CAS on machines that don't support it natively
2325 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2326 return false;
2328 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2330 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2331 int nargs = 1 + 1 + 2 + type_words + type_words;
2333 // pop arguments: newval, oldval, offset, base, and receiver
2334 debug_only(int saved_sp = _sp);
2335 _sp += nargs;
2336 Node* newval = (type_words == 1) ? pop() : pop_pair();
2337 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2338 Node *offset = pop_pair();
2339 Node *base = pop();
2340 Node *receiver = pop();
2341 assert(saved_sp == _sp, "must have correct argument count");
2343 // Null check receiver.
2344 _sp += nargs;
2345 do_null_check(receiver, T_OBJECT);
2346 _sp -= nargs;
2347 if (stopped()) {
2348 return true;
2349 }
2351 // Build field offset expression.
2352 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2353 // to be plain byte offsets, which are also the same as those accepted
2354 // by oopDesc::field_base.
2355 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2356 // 32-bit machines ignore the high half of long offsets
2357 offset = ConvL2X(offset);
2358 Node* adr = make_unsafe_address(base, offset);
2359 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2361 // (Unlike inline_unsafe_access, there seems no point in trying
2362 // to refine types. Just use the coarse types here.
2363 const Type *value_type = Type::get_const_basic_type(type);
2364 Compile::AliasType* alias_type = C->alias_type(adr_type);
2365 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2366 int alias_idx = C->get_alias_index(adr_type);
2368 // Memory-model-wise, a CAS acts like a little synchronized block,
2369 // so needs barriers on each side. These don't translate into
2370 // actual barriers on most machines, but we still need rest of
2371 // compiler to respect ordering.
2373 insert_mem_bar(Op_MemBarRelease);
2374 insert_mem_bar(Op_MemBarCPUOrder);
2376 // 4984716: MemBars must be inserted before this
2377 // memory node in order to avoid a false
2378 // dependency which will confuse the scheduler.
2379 Node *mem = memory(alias_idx);
2381 // For now, we handle only those cases that actually exist: ints,
2382 // longs, and Object. Adding others should be straightforward.
2383 Node* cas;
2384 switch(type) {
2385 case T_INT:
2386 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2387 break;
2388 case T_LONG:
2389 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2390 break;
2391 case T_OBJECT:
2392 // reference stores need a store barrier.
2393 // (They don't if CAS fails, but it isn't worth checking.)
2394 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2395 #ifdef _LP64
2396 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2397 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2398 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2399 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2400 newval_enc, oldval_enc));
2401 } else
2402 #endif
2403 {
2404 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2405 }
2406 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2407 break;
2408 default:
2409 ShouldNotReachHere();
2410 break;
2411 }
2413 // SCMemProjNodes represent the memory state of CAS. Their main
2414 // role is to prevent CAS nodes from being optimized away when their
2415 // results aren't used.
2416 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2417 set_memory(proj, alias_idx);
2419 // Add the trailing membar surrounding the access
2420 insert_mem_bar(Op_MemBarCPUOrder);
2421 insert_mem_bar(Op_MemBarAcquire);
2423 push(cas);
2424 return true;
2425 }
2427 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2428 // This is another variant of inline_unsafe_access, differing in
2429 // that it always issues store-store ("release") barrier and ensures
2430 // store-atomicity (which only matters for "long").
2432 if (callee()->is_static()) return false; // caller must have the capability!
2434 #ifndef PRODUCT
2435 {
2436 ResourceMark rm;
2437 // Check the signatures.
2438 ciSignature* sig = signature();
2439 #ifdef ASSERT
2440 BasicType rtype = sig->return_type()->basic_type();
2441 assert(rtype == T_VOID, "must return void");
2442 assert(sig->count() == 3, "has 3 arguments");
2443 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2444 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2445 #endif // ASSERT
2446 }
2447 #endif //PRODUCT
2449 // number of stack slots per value argument (1 or 2)
2450 int type_words = type2size[type];
2452 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2454 // Argument words: "this" plus oop plus offset plus value;
2455 int nargs = 1 + 1 + 2 + type_words;
2457 // pop arguments: val, offset, base, and receiver
2458 debug_only(int saved_sp = _sp);
2459 _sp += nargs;
2460 Node* val = (type_words == 1) ? pop() : pop_pair();
2461 Node *offset = pop_pair();
2462 Node *base = pop();
2463 Node *receiver = pop();
2464 assert(saved_sp == _sp, "must have correct argument count");
2466 // Null check receiver.
2467 _sp += nargs;
2468 do_null_check(receiver, T_OBJECT);
2469 _sp -= nargs;
2470 if (stopped()) {
2471 return true;
2472 }
2474 // Build field offset expression.
2475 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2476 // 32-bit machines ignore the high half of long offsets
2477 offset = ConvL2X(offset);
2478 Node* adr = make_unsafe_address(base, offset);
2479 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2480 const Type *value_type = Type::get_const_basic_type(type);
2481 Compile::AliasType* alias_type = C->alias_type(adr_type);
2483 insert_mem_bar(Op_MemBarRelease);
2484 insert_mem_bar(Op_MemBarCPUOrder);
2485 // Ensure that the store is atomic for longs:
2486 bool require_atomic_access = true;
2487 Node* store;
2488 if (type == T_OBJECT) // reference stores need a store barrier.
2489 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2490 else {
2491 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2492 }
2493 insert_mem_bar(Op_MemBarCPUOrder);
2494 return true;
2495 }
2497 bool LibraryCallKit::inline_unsafe_allocate() {
2498 if (callee()->is_static()) return false; // caller must have the capability!
2499 int nargs = 1 + 1;
2500 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2501 null_check_receiver(callee()); // check then ignore argument(0)
2502 _sp += nargs; // set original stack for use by uncommon_trap
2503 Node* cls = do_null_check(argument(1), T_OBJECT);
2504 _sp -= nargs;
2505 if (stopped()) return true;
2507 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2508 _sp += nargs; // set original stack for use by uncommon_trap
2509 kls = do_null_check(kls, T_OBJECT);
2510 _sp -= nargs;
2511 if (stopped()) return true; // argument was like int.class
2513 // Note: The argument might still be an illegal value like
2514 // Serializable.class or Object[].class. The runtime will handle it.
2515 // But we must make an explicit check for initialization.
2516 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2517 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2518 Node* bits = intcon(instanceKlass::fully_initialized);
2519 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2520 // The 'test' is non-zero if we need to take a slow path.
2522 Node* obj = new_instance(kls, test);
2523 push(obj);
2525 return true;
2526 }
2528 //------------------------inline_native_time_funcs--------------
2529 // inline code for System.currentTimeMillis() and System.nanoTime()
2530 // these have the same type and signature
2531 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2532 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2533 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2534 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2535 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2536 const TypePtr* no_memory_effects = NULL;
2537 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2538 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2539 #ifdef ASSERT
2540 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2541 assert(value_top == top(), "second value must be top");
2542 #endif
2543 push_pair(value);
2544 return true;
2545 }
2547 //------------------------inline_native_currentThread------------------
2548 bool LibraryCallKit::inline_native_currentThread() {
2549 Node* junk = NULL;
2550 push(generate_current_thread(junk));
2551 return true;
2552 }
2554 //------------------------inline_native_isInterrupted------------------
2555 bool LibraryCallKit::inline_native_isInterrupted() {
2556 const int nargs = 1+1; // receiver + boolean
2557 assert(nargs == arg_size(), "sanity");
2558 // Add a fast path to t.isInterrupted(clear_int):
2559 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2560 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2561 // So, in the common case that the interrupt bit is false,
2562 // we avoid making a call into the VM. Even if the interrupt bit
2563 // is true, if the clear_int argument is false, we avoid the VM call.
2564 // However, if the receiver is not currentThread, we must call the VM,
2565 // because there must be some locking done around the operation.
2567 // We only go to the fast case code if we pass two guards.
2568 // Paths which do not pass are accumulated in the slow_region.
2569 RegionNode* slow_region = new (C, 1) RegionNode(1);
2570 record_for_igvn(slow_region);
2571 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2572 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2573 enum { no_int_result_path = 1,
2574 no_clear_result_path = 2,
2575 slow_result_path = 3
2576 };
2578 // (a) Receiving thread must be the current thread.
2579 Node* rec_thr = argument(0);
2580 Node* tls_ptr = NULL;
2581 Node* cur_thr = generate_current_thread(tls_ptr);
2582 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2583 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2585 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2586 if (!known_current_thread)
2587 generate_slow_guard(bol_thr, slow_region);
2589 // (b) Interrupt bit on TLS must be false.
2590 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2591 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2592 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2593 // Set the control input on the field _interrupted read to prevent it floating up.
2594 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2595 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2596 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2598 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2600 // First fast path: if (!TLS._interrupted) return false;
2601 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2602 result_rgn->init_req(no_int_result_path, false_bit);
2603 result_val->init_req(no_int_result_path, intcon(0));
2605 // drop through to next case
2606 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2608 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2609 Node* clr_arg = argument(1);
2610 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2611 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2612 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2614 // Second fast path: ... else if (!clear_int) return true;
2615 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2616 result_rgn->init_req(no_clear_result_path, false_arg);
2617 result_val->init_req(no_clear_result_path, intcon(1));
2619 // drop through to next case
2620 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2622 // (d) Otherwise, go to the slow path.
2623 slow_region->add_req(control());
2624 set_control( _gvn.transform(slow_region) );
2626 if (stopped()) {
2627 // There is no slow path.
2628 result_rgn->init_req(slow_result_path, top());
2629 result_val->init_req(slow_result_path, top());
2630 } else {
2631 // non-virtual because it is a private non-static
2632 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2634 Node* slow_val = set_results_for_java_call(slow_call);
2635 // this->control() comes from set_results_for_java_call
2637 // If we know that the result of the slow call will be true, tell the optimizer!
2638 if (known_current_thread) slow_val = intcon(1);
2640 Node* fast_io = slow_call->in(TypeFunc::I_O);
2641 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2642 // These two phis are pre-filled with copies of of the fast IO and Memory
2643 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2644 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2646 result_rgn->init_req(slow_result_path, control());
2647 io_phi ->init_req(slow_result_path, i_o());
2648 mem_phi ->init_req(slow_result_path, reset_memory());
2649 result_val->init_req(slow_result_path, slow_val);
2651 set_all_memory( _gvn.transform(mem_phi) );
2652 set_i_o( _gvn.transform(io_phi) );
2653 }
2655 push_result(result_rgn, result_val);
2656 C->set_has_split_ifs(true); // Has chance for split-if optimization
2658 return true;
2659 }
2661 //---------------------------load_mirror_from_klass----------------------------
2662 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2663 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2664 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2665 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2666 }
2668 //-----------------------load_klass_from_mirror_common-------------------------
2669 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2670 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2671 // and branch to the given path on the region.
2672 // If never_see_null, take an uncommon trap on null, so we can optimistically
2673 // compile for the non-null case.
2674 // If the region is NULL, force never_see_null = true.
2675 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2676 bool never_see_null,
2677 int nargs,
2678 RegionNode* region,
2679 int null_path,
2680 int offset) {
2681 if (region == NULL) never_see_null = true;
2682 Node* p = basic_plus_adr(mirror, offset);
2683 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2684 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2685 _sp += nargs; // any deopt will start just before call to enclosing method
2686 Node* null_ctl = top();
2687 kls = null_check_oop(kls, &null_ctl, never_see_null);
2688 if (region != NULL) {
2689 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2690 region->init_req(null_path, null_ctl);
2691 } else {
2692 assert(null_ctl == top(), "no loose ends");
2693 }
2694 _sp -= nargs;
2695 return kls;
2696 }
2698 //--------------------(inline_native_Class_query helpers)---------------------
2699 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2700 // Fall through if (mods & mask) == bits, take the guard otherwise.
2701 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2702 // Branch around if the given klass has the given modifier bit set.
2703 // Like generate_guard, adds a new path onto the region.
2704 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2705 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2706 Node* mask = intcon(modifier_mask);
2707 Node* bits = intcon(modifier_bits);
2708 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2709 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2710 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2711 return generate_fair_guard(bol, region);
2712 }
2713 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2714 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2715 }
2717 //-------------------------inline_native_Class_query-------------------
2718 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2719 int nargs = 1+0; // just the Class mirror, in most cases
2720 const Type* return_type = TypeInt::BOOL;
2721 Node* prim_return_value = top(); // what happens if it's a primitive class?
2722 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2723 bool expect_prim = false; // most of these guys expect to work on refs
2725 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2727 switch (id) {
2728 case vmIntrinsics::_isInstance:
2729 nargs = 1+1; // the Class mirror, plus the object getting queried about
2730 // nothing is an instance of a primitive type
2731 prim_return_value = intcon(0);
2732 break;
2733 case vmIntrinsics::_getModifiers:
2734 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2735 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2736 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2737 break;
2738 case vmIntrinsics::_isInterface:
2739 prim_return_value = intcon(0);
2740 break;
2741 case vmIntrinsics::_isArray:
2742 prim_return_value = intcon(0);
2743 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2744 break;
2745 case vmIntrinsics::_isPrimitive:
2746 prim_return_value = intcon(1);
2747 expect_prim = true; // obviously
2748 break;
2749 case vmIntrinsics::_getSuperclass:
2750 prim_return_value = null();
2751 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2752 break;
2753 case vmIntrinsics::_getComponentType:
2754 prim_return_value = null();
2755 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2756 break;
2757 case vmIntrinsics::_getClassAccessFlags:
2758 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2759 return_type = TypeInt::INT; // not bool! 6297094
2760 break;
2761 default:
2762 ShouldNotReachHere();
2763 }
2765 Node* mirror = argument(0);
2766 Node* obj = (nargs <= 1)? top(): argument(1);
2768 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2769 if (mirror_con == NULL) return false; // cannot happen?
2771 #ifndef PRODUCT
2772 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2773 ciType* k = mirror_con->java_mirror_type();
2774 if (k) {
2775 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2776 k->print_name();
2777 tty->cr();
2778 }
2779 }
2780 #endif
2782 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2783 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2784 record_for_igvn(region);
2785 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2787 // The mirror will never be null of Reflection.getClassAccessFlags, however
2788 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2789 // if it is. See bug 4774291.
2791 // For Reflection.getClassAccessFlags(), the null check occurs in
2792 // the wrong place; see inline_unsafe_access(), above, for a similar
2793 // situation.
2794 _sp += nargs; // set original stack for use by uncommon_trap
2795 mirror = do_null_check(mirror, T_OBJECT);
2796 _sp -= nargs;
2797 // If mirror or obj is dead, only null-path is taken.
2798 if (stopped()) return true;
2800 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2802 // Now load the mirror's klass metaobject, and null-check it.
2803 // Side-effects region with the control path if the klass is null.
2804 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2805 region, _prim_path);
2806 // If kls is null, we have a primitive mirror.
2807 phi->init_req(_prim_path, prim_return_value);
2808 if (stopped()) { push_result(region, phi); return true; }
2810 Node* p; // handy temp
2811 Node* null_ctl;
2813 // Now that we have the non-null klass, we can perform the real query.
2814 // For constant classes, the query will constant-fold in LoadNode::Value.
2815 Node* query_value = top();
2816 switch (id) {
2817 case vmIntrinsics::_isInstance:
2818 // nothing is an instance of a primitive type
2819 query_value = gen_instanceof(obj, kls);
2820 break;
2822 case vmIntrinsics::_getModifiers:
2823 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2824 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2825 break;
2827 case vmIntrinsics::_isInterface:
2828 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2829 if (generate_interface_guard(kls, region) != NULL)
2830 // A guard was added. If the guard is taken, it was an interface.
2831 phi->add_req(intcon(1));
2832 // If we fall through, it's a plain class.
2833 query_value = intcon(0);
2834 break;
2836 case vmIntrinsics::_isArray:
2837 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2838 if (generate_array_guard(kls, region) != NULL)
2839 // A guard was added. If the guard is taken, it was an array.
2840 phi->add_req(intcon(1));
2841 // If we fall through, it's a plain class.
2842 query_value = intcon(0);
2843 break;
2845 case vmIntrinsics::_isPrimitive:
2846 query_value = intcon(0); // "normal" path produces false
2847 break;
2849 case vmIntrinsics::_getSuperclass:
2850 // The rules here are somewhat unfortunate, but we can still do better
2851 // with random logic than with a JNI call.
2852 // Interfaces store null or Object as _super, but must report null.
2853 // Arrays store an intermediate super as _super, but must report Object.
2854 // Other types can report the actual _super.
2855 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2856 if (generate_interface_guard(kls, region) != NULL)
2857 // A guard was added. If the guard is taken, it was an interface.
2858 phi->add_req(null());
2859 if (generate_array_guard(kls, region) != NULL)
2860 // A guard was added. If the guard is taken, it was an array.
2861 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2862 // If we fall through, it's a plain class. Get its _super.
2863 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2864 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2865 null_ctl = top();
2866 kls = null_check_oop(kls, &null_ctl);
2867 if (null_ctl != top()) {
2868 // If the guard is taken, Object.superClass is null (both klass and mirror).
2869 region->add_req(null_ctl);
2870 phi ->add_req(null());
2871 }
2872 if (!stopped()) {
2873 query_value = load_mirror_from_klass(kls);
2874 }
2875 break;
2877 case vmIntrinsics::_getComponentType:
2878 if (generate_array_guard(kls, region) != NULL) {
2879 // Be sure to pin the oop load to the guard edge just created:
2880 Node* is_array_ctrl = region->in(region->req()-1);
2881 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
2882 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
2883 phi->add_req(cmo);
2884 }
2885 query_value = null(); // non-array case is null
2886 break;
2888 case vmIntrinsics::_getClassAccessFlags:
2889 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2890 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2891 break;
2893 default:
2894 ShouldNotReachHere();
2895 }
2897 // Fall-through is the normal case of a query to a real class.
2898 phi->init_req(1, query_value);
2899 region->init_req(1, control());
2901 push_result(region, phi);
2902 C->set_has_split_ifs(true); // Has chance for split-if optimization
2904 return true;
2905 }
2907 //--------------------------inline_native_subtype_check------------------------
2908 // This intrinsic takes the JNI calls out of the heart of
2909 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
2910 bool LibraryCallKit::inline_native_subtype_check() {
2911 int nargs = 1+1; // the Class mirror, plus the other class getting examined
2913 // Pull both arguments off the stack.
2914 Node* args[2]; // two java.lang.Class mirrors: superc, subc
2915 args[0] = argument(0);
2916 args[1] = argument(1);
2917 Node* klasses[2]; // corresponding Klasses: superk, subk
2918 klasses[0] = klasses[1] = top();
2920 enum {
2921 // A full decision tree on {superc is prim, subc is prim}:
2922 _prim_0_path = 1, // {P,N} => false
2923 // {P,P} & superc!=subc => false
2924 _prim_same_path, // {P,P} & superc==subc => true
2925 _prim_1_path, // {N,P} => false
2926 _ref_subtype_path, // {N,N} & subtype check wins => true
2927 _both_ref_path, // {N,N} & subtype check loses => false
2928 PATH_LIMIT
2929 };
2931 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2932 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2933 record_for_igvn(region);
2935 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
2936 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2937 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
2939 // First null-check both mirrors and load each mirror's klass metaobject.
2940 int which_arg;
2941 for (which_arg = 0; which_arg <= 1; which_arg++) {
2942 Node* arg = args[which_arg];
2943 _sp += nargs; // set original stack for use by uncommon_trap
2944 arg = do_null_check(arg, T_OBJECT);
2945 _sp -= nargs;
2946 if (stopped()) break;
2947 args[which_arg] = _gvn.transform(arg);
2949 Node* p = basic_plus_adr(arg, class_klass_offset);
2950 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
2951 klasses[which_arg] = _gvn.transform(kls);
2952 }
2954 // Having loaded both klasses, test each for null.
2955 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2956 for (which_arg = 0; which_arg <= 1; which_arg++) {
2957 Node* kls = klasses[which_arg];
2958 Node* null_ctl = top();
2959 _sp += nargs; // set original stack for use by uncommon_trap
2960 kls = null_check_oop(kls, &null_ctl, never_see_null);
2961 _sp -= nargs;
2962 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
2963 region->init_req(prim_path, null_ctl);
2964 if (stopped()) break;
2965 klasses[which_arg] = kls;
2966 }
2968 if (!stopped()) {
2969 // now we have two reference types, in klasses[0..1]
2970 Node* subk = klasses[1]; // the argument to isAssignableFrom
2971 Node* superk = klasses[0]; // the receiver
2972 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
2973 // now we have a successful reference subtype check
2974 region->set_req(_ref_subtype_path, control());
2975 }
2977 // If both operands are primitive (both klasses null), then
2978 // we must return true when they are identical primitives.
2979 // It is convenient to test this after the first null klass check.
2980 set_control(region->in(_prim_0_path)); // go back to first null check
2981 if (!stopped()) {
2982 // Since superc is primitive, make a guard for the superc==subc case.
2983 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
2984 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
2985 generate_guard(bol_eq, region, PROB_FAIR);
2986 if (region->req() == PATH_LIMIT+1) {
2987 // A guard was added. If the added guard is taken, superc==subc.
2988 region->swap_edges(PATH_LIMIT, _prim_same_path);
2989 region->del_req(PATH_LIMIT);
2990 }
2991 region->set_req(_prim_0_path, control()); // Not equal after all.
2992 }
2994 // these are the only paths that produce 'true':
2995 phi->set_req(_prim_same_path, intcon(1));
2996 phi->set_req(_ref_subtype_path, intcon(1));
2998 // pull together the cases:
2999 assert(region->req() == PATH_LIMIT, "sane region");
3000 for (uint i = 1; i < region->req(); i++) {
3001 Node* ctl = region->in(i);
3002 if (ctl == NULL || ctl == top()) {
3003 region->set_req(i, top());
3004 phi ->set_req(i, top());
3005 } else if (phi->in(i) == NULL) {
3006 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3007 }
3008 }
3010 set_control(_gvn.transform(region));
3011 push(_gvn.transform(phi));
3013 return true;
3014 }
3016 //---------------------generate_array_guard_common------------------------
3017 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3018 bool obj_array, bool not_array) {
3019 // If obj_array/non_array==false/false:
3020 // Branch around if the given klass is in fact an array (either obj or prim).
3021 // If obj_array/non_array==false/true:
3022 // Branch around if the given klass is not an array klass of any kind.
3023 // If obj_array/non_array==true/true:
3024 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3025 // If obj_array/non_array==true/false:
3026 // Branch around if the kls is an oop array (Object[] or subtype)
3027 //
3028 // Like generate_guard, adds a new path onto the region.
3029 jint layout_con = 0;
3030 Node* layout_val = get_layout_helper(kls, layout_con);
3031 if (layout_val == NULL) {
3032 bool query = (obj_array
3033 ? Klass::layout_helper_is_objArray(layout_con)
3034 : Klass::layout_helper_is_javaArray(layout_con));
3035 if (query == not_array) {
3036 return NULL; // never a branch
3037 } else { // always a branch
3038 Node* always_branch = control();
3039 if (region != NULL)
3040 region->add_req(always_branch);
3041 set_control(top());
3042 return always_branch;
3043 }
3044 }
3045 // Now test the correct condition.
3046 jint nval = (obj_array
3047 ? ((jint)Klass::_lh_array_tag_type_value
3048 << Klass::_lh_array_tag_shift)
3049 : Klass::_lh_neutral_value);
3050 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3051 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3052 // invert the test if we are looking for a non-array
3053 if (not_array) btest = BoolTest(btest).negate();
3054 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3055 return generate_fair_guard(bol, region);
3056 }
3059 //-----------------------inline_native_newArray--------------------------
3060 bool LibraryCallKit::inline_native_newArray() {
3061 int nargs = 2;
3062 Node* mirror = argument(0);
3063 Node* count_val = argument(1);
3065 _sp += nargs; // set original stack for use by uncommon_trap
3066 mirror = do_null_check(mirror, T_OBJECT);
3067 _sp -= nargs;
3068 // If mirror or obj is dead, only null-path is taken.
3069 if (stopped()) return true;
3071 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3072 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3073 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3074 TypeInstPtr::NOTNULL);
3075 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3076 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3077 TypePtr::BOTTOM);
3079 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3080 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3081 nargs,
3082 result_reg, _slow_path);
3083 Node* normal_ctl = control();
3084 Node* no_array_ctl = result_reg->in(_slow_path);
3086 // Generate code for the slow case. We make a call to newArray().
3087 set_control(no_array_ctl);
3088 if (!stopped()) {
3089 // Either the input type is void.class, or else the
3090 // array klass has not yet been cached. Either the
3091 // ensuing call will throw an exception, or else it
3092 // will cache the array klass for next time.
3093 PreserveJVMState pjvms(this);
3094 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3095 Node* slow_result = set_results_for_java_call(slow_call);
3096 // this->control() comes from set_results_for_java_call
3097 result_reg->set_req(_slow_path, control());
3098 result_val->set_req(_slow_path, slow_result);
3099 result_io ->set_req(_slow_path, i_o());
3100 result_mem->set_req(_slow_path, reset_memory());
3101 }
3103 set_control(normal_ctl);
3104 if (!stopped()) {
3105 // Normal case: The array type has been cached in the java.lang.Class.
3106 // The following call works fine even if the array type is polymorphic.
3107 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3108 Node* obj = new_array(klass_node, count_val, nargs);
3109 result_reg->init_req(_normal_path, control());
3110 result_val->init_req(_normal_path, obj);
3111 result_io ->init_req(_normal_path, i_o());
3112 result_mem->init_req(_normal_path, reset_memory());
3113 }
3115 // Return the combined state.
3116 set_i_o( _gvn.transform(result_io) );
3117 set_all_memory( _gvn.transform(result_mem) );
3118 push_result(result_reg, result_val);
3119 C->set_has_split_ifs(true); // Has chance for split-if optimization
3121 return true;
3122 }
3124 //----------------------inline_native_getLength--------------------------
3125 bool LibraryCallKit::inline_native_getLength() {
3126 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3128 int nargs = 1;
3129 Node* array = argument(0);
3131 _sp += nargs; // set original stack for use by uncommon_trap
3132 array = do_null_check(array, T_OBJECT);
3133 _sp -= nargs;
3135 // If array is dead, only null-path is taken.
3136 if (stopped()) return true;
3138 // Deoptimize if it is a non-array.
3139 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3141 if (non_array != NULL) {
3142 PreserveJVMState pjvms(this);
3143 set_control(non_array);
3144 _sp += nargs; // push the arguments back on the stack
3145 uncommon_trap(Deoptimization::Reason_intrinsic,
3146 Deoptimization::Action_maybe_recompile);
3147 }
3149 // If control is dead, only non-array-path is taken.
3150 if (stopped()) return true;
3152 // The works fine even if the array type is polymorphic.
3153 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3154 push( load_array_length(array) );
3156 C->set_has_split_ifs(true); // Has chance for split-if optimization
3158 return true;
3159 }
3161 //------------------------inline_array_copyOf----------------------------
3162 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3163 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3165 // Restore the stack and pop off the arguments.
3166 int nargs = 3 + (is_copyOfRange? 1: 0);
3167 Node* original = argument(0);
3168 Node* start = is_copyOfRange? argument(1): intcon(0);
3169 Node* end = is_copyOfRange? argument(2): argument(1);
3170 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3172 _sp += nargs; // set original stack for use by uncommon_trap
3173 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3174 original = do_null_check(original, T_OBJECT);
3175 _sp -= nargs;
3177 // Check if a null path was taken unconditionally.
3178 if (stopped()) return true;
3180 Node* orig_length = load_array_length(original);
3182 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs,
3183 NULL, 0);
3184 _sp += nargs; // set original stack for use by uncommon_trap
3185 klass_node = do_null_check(klass_node, T_OBJECT);
3186 _sp -= nargs;
3188 RegionNode* bailout = new (C, 1) RegionNode(1);
3189 record_for_igvn(bailout);
3191 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3192 // Bail out if that is so.
3193 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3194 if (not_objArray != NULL) {
3195 // Improve the klass node's type from the new optimistic assumption:
3196 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3197 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3198 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3199 cast->init_req(0, control());
3200 klass_node = _gvn.transform(cast);
3201 }
3203 // Bail out if either start or end is negative.
3204 generate_negative_guard(start, bailout, &start);
3205 generate_negative_guard(end, bailout, &end);
3207 Node* length = end;
3208 if (_gvn.type(start) != TypeInt::ZERO) {
3209 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3210 }
3212 // Bail out if length is negative.
3213 // ...Not needed, since the new_array will throw the right exception.
3214 //generate_negative_guard(length, bailout, &length);
3216 if (bailout->req() > 1) {
3217 PreserveJVMState pjvms(this);
3218 set_control( _gvn.transform(bailout) );
3219 _sp += nargs; // push the arguments back on the stack
3220 uncommon_trap(Deoptimization::Reason_intrinsic,
3221 Deoptimization::Action_maybe_recompile);
3222 }
3224 if (!stopped()) {
3225 // How many elements will we copy from the original?
3226 // The answer is MinI(orig_length - start, length).
3227 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3228 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3230 const bool raw_mem_only = true;
3231 Node* newcopy = new_array(klass_node, length, nargs, raw_mem_only);
3233 // Generate a direct call to the right arraycopy function(s).
3234 // We know the copy is disjoint but we might not know if the
3235 // oop stores need checking.
3236 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3237 // This will fail a store-check if x contains any non-nulls.
3238 bool disjoint_bases = true;
3239 bool length_never_negative = true;
3240 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3241 original, start, newcopy, intcon(0), moved,
3242 disjoint_bases, length_never_negative);
3244 push(newcopy);
3245 }
3247 C->set_has_split_ifs(true); // Has chance for split-if optimization
3249 return true;
3250 }
3253 //----------------------generate_virtual_guard---------------------------
3254 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3255 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3256 RegionNode* slow_region) {
3257 ciMethod* method = callee();
3258 int vtable_index = method->vtable_index();
3259 // Get the methodOop out of the appropriate vtable entry.
3260 int entry_offset = (instanceKlass::vtable_start_offset() +
3261 vtable_index*vtableEntry::size()) * wordSize +
3262 vtableEntry::method_offset_in_bytes();
3263 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3264 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3266 // Compare the target method with the expected method (e.g., Object.hashCode).
3267 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3269 Node* native_call = makecon(native_call_addr);
3270 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3271 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3273 return generate_slow_guard(test_native, slow_region);
3274 }
3276 //-----------------------generate_method_call----------------------------
3277 // Use generate_method_call to make a slow-call to the real
3278 // method if the fast path fails. An alternative would be to
3279 // use a stub like OptoRuntime::slow_arraycopy_Java.
3280 // This only works for expanding the current library call,
3281 // not another intrinsic. (E.g., don't use this for making an
3282 // arraycopy call inside of the copyOf intrinsic.)
3283 CallJavaNode*
3284 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3285 // When compiling the intrinsic method itself, do not use this technique.
3286 guarantee(callee() != C->method(), "cannot make slow-call to self");
3288 ciMethod* method = callee();
3289 // ensure the JVMS we have will be correct for this call
3290 guarantee(method_id == method->intrinsic_id(), "must match");
3292 const TypeFunc* tf = TypeFunc::make(method);
3293 int tfdc = tf->domain()->cnt();
3294 CallJavaNode* slow_call;
3295 if (is_static) {
3296 assert(!is_virtual, "");
3297 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3298 SharedRuntime::get_resolve_static_call_stub(),
3299 method, bci());
3300 } else if (is_virtual) {
3301 null_check_receiver(method);
3302 int vtable_index = methodOopDesc::invalid_vtable_index;
3303 if (UseInlineCaches) {
3304 // Suppress the vtable call
3305 } else {
3306 // hashCode and clone are not a miranda methods,
3307 // so the vtable index is fixed.
3308 // No need to use the linkResolver to get it.
3309 vtable_index = method->vtable_index();
3310 }
3311 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3312 SharedRuntime::get_resolve_virtual_call_stub(),
3313 method, vtable_index, bci());
3314 } else { // neither virtual nor static: opt_virtual
3315 null_check_receiver(method);
3316 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3317 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3318 method, bci());
3319 slow_call->set_optimized_virtual(true);
3320 }
3321 set_arguments_for_java_call(slow_call);
3322 set_edges_for_java_call(slow_call);
3323 return slow_call;
3324 }
3327 //------------------------------inline_native_hashcode--------------------
3328 // Build special case code for calls to hashCode on an object.
3329 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3330 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3331 assert(!(is_virtual && is_static), "either virtual, special, or static");
3333 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3335 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3336 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3337 TypeInt::INT);
3338 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3339 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3340 TypePtr::BOTTOM);
3341 Node* obj = NULL;
3342 if (!is_static) {
3343 // Check for hashing null object
3344 obj = null_check_receiver(callee());
3345 if (stopped()) return true; // unconditionally null
3346 result_reg->init_req(_null_path, top());
3347 result_val->init_req(_null_path, top());
3348 } else {
3349 // Do a null check, and return zero if null.
3350 // System.identityHashCode(null) == 0
3351 obj = argument(0);
3352 Node* null_ctl = top();
3353 obj = null_check_oop(obj, &null_ctl);
3354 result_reg->init_req(_null_path, null_ctl);
3355 result_val->init_req(_null_path, _gvn.intcon(0));
3356 }
3358 // Unconditionally null? Then return right away.
3359 if (stopped()) {
3360 set_control( result_reg->in(_null_path) );
3361 if (!stopped())
3362 push( result_val ->in(_null_path) );
3363 return true;
3364 }
3366 // After null check, get the object's klass.
3367 Node* obj_klass = load_object_klass(obj);
3369 // This call may be virtual (invokevirtual) or bound (invokespecial).
3370 // For each case we generate slightly different code.
3372 // We only go to the fast case code if we pass a number of guards. The
3373 // paths which do not pass are accumulated in the slow_region.
3374 RegionNode* slow_region = new (C, 1) RegionNode(1);
3375 record_for_igvn(slow_region);
3377 // If this is a virtual call, we generate a funny guard. We pull out
3378 // the vtable entry corresponding to hashCode() from the target object.
3379 // If the target method which we are calling happens to be the native
3380 // Object hashCode() method, we pass the guard. We do not need this
3381 // guard for non-virtual calls -- the caller is known to be the native
3382 // Object hashCode().
3383 if (is_virtual) {
3384 generate_virtual_guard(obj_klass, slow_region);
3385 }
3387 // Get the header out of the object, use LoadMarkNode when available
3388 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3389 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
3390 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
3392 // Test the header to see if it is unlocked.
3393 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3394 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3395 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3396 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3397 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3399 generate_slow_guard(test_unlocked, slow_region);
3401 // Get the hash value and check to see that it has been properly assigned.
3402 // We depend on hash_mask being at most 32 bits and avoid the use of
3403 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3404 // vm: see markOop.hpp.
3405 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3406 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3407 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3408 // This hack lets the hash bits live anywhere in the mark object now, as long
3409 // as the shift drops the relevant bits into the low 32 bits. Note that
3410 // Java spec says that HashCode is an int so there's no point in capturing
3411 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3412 hshifted_header = ConvX2I(hshifted_header);
3413 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3415 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3416 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3417 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3419 generate_slow_guard(test_assigned, slow_region);
3421 Node* init_mem = reset_memory();
3422 // fill in the rest of the null path:
3423 result_io ->init_req(_null_path, i_o());
3424 result_mem->init_req(_null_path, init_mem);
3426 result_val->init_req(_fast_path, hash_val);
3427 result_reg->init_req(_fast_path, control());
3428 result_io ->init_req(_fast_path, i_o());
3429 result_mem->init_req(_fast_path, init_mem);
3431 // Generate code for the slow case. We make a call to hashCode().
3432 set_control(_gvn.transform(slow_region));
3433 if (!stopped()) {
3434 // No need for PreserveJVMState, because we're using up the present state.
3435 set_all_memory(init_mem);
3436 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3437 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3438 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3439 Node* slow_result = set_results_for_java_call(slow_call);
3440 // this->control() comes from set_results_for_java_call
3441 result_reg->init_req(_slow_path, control());
3442 result_val->init_req(_slow_path, slow_result);
3443 result_io ->set_req(_slow_path, i_o());
3444 result_mem ->set_req(_slow_path, reset_memory());
3445 }
3447 // Return the combined state.
3448 set_i_o( _gvn.transform(result_io) );
3449 set_all_memory( _gvn.transform(result_mem) );
3450 push_result(result_reg, result_val);
3452 return true;
3453 }
3455 //---------------------------inline_native_getClass----------------------------
3456 // Build special case code for calls to getClass on an object.
3457 bool LibraryCallKit::inline_native_getClass() {
3458 Node* obj = null_check_receiver(callee());
3459 if (stopped()) return true;
3460 push( load_mirror_from_klass(load_object_klass(obj)) );
3461 return true;
3462 }
3464 //-----------------inline_native_Reflection_getCallerClass---------------------
3465 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3466 //
3467 // NOTE that this code must perform the same logic as
3468 // vframeStream::security_get_caller_frame in that it must skip
3469 // Method.invoke() and auxiliary frames.
3474 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3475 ciMethod* method = callee();
3477 #ifndef PRODUCT
3478 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3479 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3480 }
3481 #endif
3483 debug_only(int saved_sp = _sp);
3485 // Argument words: (int depth)
3486 int nargs = 1;
3488 _sp += nargs;
3489 Node* caller_depth_node = pop();
3491 assert(saved_sp == _sp, "must have correct argument count");
3493 // The depth value must be a constant in order for the runtime call
3494 // to be eliminated.
3495 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3496 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3497 #ifndef PRODUCT
3498 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3499 tty->print_cr(" Bailing out because caller depth was not a constant");
3500 }
3501 #endif
3502 return false;
3503 }
3504 // Note that the JVM state at this point does not include the
3505 // getCallerClass() frame which we are trying to inline. The
3506 // semantics of getCallerClass(), however, are that the "first"
3507 // frame is the getCallerClass() frame, so we subtract one from the
3508 // requested depth before continuing. We don't inline requests of
3509 // getCallerClass(0).
3510 int caller_depth = caller_depth_type->get_con() - 1;
3511 if (caller_depth < 0) {
3512 #ifndef PRODUCT
3513 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3514 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3515 }
3516 #endif
3517 return false;
3518 }
3520 if (!jvms()->has_method()) {
3521 #ifndef PRODUCT
3522 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3523 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3524 }
3525 #endif
3526 return false;
3527 }
3528 int _depth = jvms()->depth(); // cache call chain depth
3530 // Walk back up the JVM state to find the caller at the required
3531 // depth. NOTE that this code must perform the same logic as
3532 // vframeStream::security_get_caller_frame in that it must skip
3533 // Method.invoke() and auxiliary frames. Note also that depth is
3534 // 1-based (1 is the bottom of the inlining).
3535 int inlining_depth = _depth;
3536 JVMState* caller_jvms = NULL;
3538 if (inlining_depth > 0) {
3539 caller_jvms = jvms();
3540 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3541 do {
3542 // The following if-tests should be performed in this order
3543 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3544 // Skip a Method.invoke() or auxiliary frame
3545 } else if (caller_depth > 0) {
3546 // Skip real frame
3547 --caller_depth;
3548 } else {
3549 // We're done: reached desired caller after skipping.
3550 break;
3551 }
3552 caller_jvms = caller_jvms->caller();
3553 --inlining_depth;
3554 } while (inlining_depth > 0);
3555 }
3557 if (inlining_depth == 0) {
3558 #ifndef PRODUCT
3559 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3560 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3561 tty->print_cr(" JVM state at this point:");
3562 for (int i = _depth; i >= 1; i--) {
3563 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3564 }
3565 }
3566 #endif
3567 return false; // Reached end of inlining
3568 }
3570 // Acquire method holder as java.lang.Class
3571 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3572 ciInstance* caller_mirror = caller_klass->java_mirror();
3573 // Push this as a constant
3574 push(makecon(TypeInstPtr::make(caller_mirror)));
3575 #ifndef PRODUCT
3576 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3577 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);
3578 tty->print_cr(" JVM state at this point:");
3579 for (int i = _depth; i >= 1; i--) {
3580 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3581 }
3582 }
3583 #endif
3584 return true;
3585 }
3587 // Helper routine for above
3588 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3589 // Is this the Method.invoke method itself?
3590 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
3591 return true;
3593 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3594 ciKlass* k = jvms->method()->holder();
3595 if (k->is_instance_klass()) {
3596 ciInstanceKlass* ik = k->as_instance_klass();
3597 for (; ik != NULL; ik = ik->super()) {
3598 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3599 ik == env()->find_system_klass(ik->name())) {
3600 return true;
3601 }
3602 }
3603 }
3605 return false;
3606 }
3608 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3609 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3610 // computing it since there is no lookup field by name function in the
3611 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3612 // Using a static variable here is safe even if we have multiple compilation
3613 // threads because the offset is constant. At worst the same offset will be
3614 // computed and stored multiple
3616 bool LibraryCallKit::inline_native_AtomicLong_get() {
3617 // Restore the stack and pop off the argument
3618 _sp+=1;
3619 Node *obj = pop();
3621 // get the offset of the "value" field. Since the CI interfaces
3622 // does not provide a way to look up a field by name, we scan the bytecodes
3623 // to get the field index. We expect the first 2 instructions of the method
3624 // to be:
3625 // 0 aload_0
3626 // 1 getfield "value"
3627 ciMethod* method = callee();
3628 if (value_field_offset == -1)
3629 {
3630 ciField* value_field;
3631 ciBytecodeStream iter(method);
3632 Bytecodes::Code bc = iter.next();
3634 if ((bc != Bytecodes::_aload_0) &&
3635 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3636 return false;
3637 bc = iter.next();
3638 if (bc != Bytecodes::_getfield)
3639 return false;
3640 bool ignore;
3641 value_field = iter.get_field(ignore);
3642 value_field_offset = value_field->offset_in_bytes();
3643 }
3645 // Null check without removing any arguments.
3646 _sp++;
3647 obj = do_null_check(obj, T_OBJECT);
3648 _sp--;
3649 // Check for locking null object
3650 if (stopped()) return true;
3652 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3653 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3654 int alias_idx = C->get_alias_index(adr_type);
3656 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3658 push_pair(result);
3660 return true;
3661 }
3663 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3664 // Restore the stack and pop off the arguments
3665 _sp+=5;
3666 Node *newVal = pop_pair();
3667 Node *oldVal = pop_pair();
3668 Node *obj = pop();
3670 // we need the offset of the "value" field which was computed when
3671 // inlining the get() method. Give up if we don't have it.
3672 if (value_field_offset == -1)
3673 return false;
3675 // Null check without removing any arguments.
3676 _sp+=5;
3677 obj = do_null_check(obj, T_OBJECT);
3678 _sp-=5;
3679 // Check for locking null object
3680 if (stopped()) return true;
3682 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3683 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3684 int alias_idx = C->get_alias_index(adr_type);
3686 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3687 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3688 set_memory(store_proj, alias_idx);
3689 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3691 Node *result;
3692 // CMove node is not used to be able fold a possible check code
3693 // after attemptUpdate() call. This code could be transformed
3694 // into CMove node by loop optimizations.
3695 {
3696 RegionNode *r = new (C, 3) RegionNode(3);
3697 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3699 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3700 Node *iftrue = opt_iff(r, iff);
3701 r->init_req(1, iftrue);
3702 result->init_req(1, intcon(1));
3703 result->init_req(2, intcon(0));
3705 set_control(_gvn.transform(r));
3706 record_for_igvn(r);
3708 C->set_has_split_ifs(true); // Has chance for split-if optimization
3709 }
3711 push(_gvn.transform(result));
3712 return true;
3713 }
3715 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3716 // restore the arguments
3717 _sp += arg_size();
3719 switch (id) {
3720 case vmIntrinsics::_floatToRawIntBits:
3721 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3722 break;
3724 case vmIntrinsics::_intBitsToFloat:
3725 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3726 break;
3728 case vmIntrinsics::_doubleToRawLongBits:
3729 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3730 break;
3732 case vmIntrinsics::_longBitsToDouble:
3733 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3734 break;
3736 case vmIntrinsics::_doubleToLongBits: {
3737 Node* value = pop_pair();
3739 // two paths (plus control) merge in a wood
3740 RegionNode *r = new (C, 3) RegionNode(3);
3741 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3743 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3744 // Build the boolean node
3745 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3747 // Branch either way.
3748 // NaN case is less traveled, which makes all the difference.
3749 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3750 Node *opt_isnan = _gvn.transform(ifisnan);
3751 assert( opt_isnan->is_If(), "Expect an IfNode");
3752 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3753 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3755 set_control(iftrue);
3757 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3758 Node *slow_result = longcon(nan_bits); // return NaN
3759 phi->init_req(1, _gvn.transform( slow_result ));
3760 r->init_req(1, iftrue);
3762 // Else fall through
3763 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3764 set_control(iffalse);
3766 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3767 r->init_req(2, iffalse);
3769 // Post merge
3770 set_control(_gvn.transform(r));
3771 record_for_igvn(r);
3773 Node* result = _gvn.transform(phi);
3774 assert(result->bottom_type()->isa_long(), "must be");
3775 push_pair(result);
3777 C->set_has_split_ifs(true); // Has chance for split-if optimization
3779 break;
3780 }
3782 case vmIntrinsics::_floatToIntBits: {
3783 Node* value = pop();
3785 // two paths (plus control) merge in a wood
3786 RegionNode *r = new (C, 3) RegionNode(3);
3787 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3789 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3790 // Build the boolean node
3791 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3793 // Branch either way.
3794 // NaN case is less traveled, which makes all the difference.
3795 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3796 Node *opt_isnan = _gvn.transform(ifisnan);
3797 assert( opt_isnan->is_If(), "Expect an IfNode");
3798 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3799 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3801 set_control(iftrue);
3803 static const jint nan_bits = 0x7fc00000;
3804 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3805 phi->init_req(1, _gvn.transform( slow_result ));
3806 r->init_req(1, iftrue);
3808 // Else fall through
3809 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3810 set_control(iffalse);
3812 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3813 r->init_req(2, iffalse);
3815 // Post merge
3816 set_control(_gvn.transform(r));
3817 record_for_igvn(r);
3819 Node* result = _gvn.transform(phi);
3820 assert(result->bottom_type()->isa_int(), "must be");
3821 push(result);
3823 C->set_has_split_ifs(true); // Has chance for split-if optimization
3825 break;
3826 }
3828 default:
3829 ShouldNotReachHere();
3830 }
3832 return true;
3833 }
3835 #ifdef _LP64
3836 #define XTOP ,top() /*additional argument*/
3837 #else //_LP64
3838 #define XTOP /*no additional argument*/
3839 #endif //_LP64
3841 //----------------------inline_unsafe_copyMemory-------------------------
3842 bool LibraryCallKit::inline_unsafe_copyMemory() {
3843 if (callee()->is_static()) return false; // caller must have the capability!
3844 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3845 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3846 null_check_receiver(callee()); // check then ignore argument(0)
3847 if (stopped()) return true;
3849 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3851 Node* src_ptr = argument(1);
3852 Node* src_off = ConvL2X(argument(2));
3853 assert(argument(3)->is_top(), "2nd half of long");
3854 Node* dst_ptr = argument(4);
3855 Node* dst_off = ConvL2X(argument(5));
3856 assert(argument(6)->is_top(), "2nd half of long");
3857 Node* size = ConvL2X(argument(7));
3858 assert(argument(8)->is_top(), "2nd half of long");
3860 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3861 "fieldOffset must be byte-scaled");
3863 Node* src = make_unsafe_address(src_ptr, src_off);
3864 Node* dst = make_unsafe_address(dst_ptr, dst_off);
3866 // Conservatively insert a memory barrier on all memory slices.
3867 // Do not let writes of the copy source or destination float below the copy.
3868 insert_mem_bar(Op_MemBarCPUOrder);
3870 // Call it. Note that the length argument is not scaled.
3871 make_runtime_call(RC_LEAF|RC_NO_FP,
3872 OptoRuntime::fast_arraycopy_Type(),
3873 StubRoutines::unsafe_arraycopy(),
3874 "unsafe_arraycopy",
3875 TypeRawPtr::BOTTOM,
3876 src, dst, size XTOP);
3878 // Do not let reads of the copy destination float above the copy.
3879 insert_mem_bar(Op_MemBarCPUOrder);
3881 return true;
3882 }
3884 //------------------------clone_coping-----------------------------------
3885 // Helper function for inline_native_clone.
3886 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
3887 assert(obj_size != NULL, "");
3888 Node* raw_obj = alloc_obj->in(1);
3889 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
3891 if (ReduceBulkZeroing) {
3892 // We will be completely responsible for initializing this object -
3893 // mark Initialize node as complete.
3894 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
3895 // The object was just allocated - there should be no any stores!
3896 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
3897 }
3899 // Cast to Object for arraycopy.
3900 // We can't use the original CheckCastPP since it should be moved
3901 // after the arraycopy to prevent stores flowing above it.
3902 Node* new_obj = new(C, 2) CheckCastPPNode(alloc_obj->in(0), raw_obj,
3903 TypeInstPtr::NOTNULL);
3904 new_obj = _gvn.transform(new_obj);
3905 // Substitute in the locally valid dest_oop.
3906 replace_in_map(alloc_obj, new_obj);
3908 // Copy the fastest available way.
3909 // TODO: generate fields copies for small objects instead.
3910 Node* src = obj;
3911 Node* dest = new_obj;
3912 Node* size = _gvn.transform(obj_size);
3914 // Exclude the header but include array length to copy by 8 bytes words.
3915 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
3916 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
3917 instanceOopDesc::base_offset_in_bytes();
3918 // base_off:
3919 // 8 - 32-bit VM
3920 // 12 - 64-bit VM, compressed oops
3921 // 16 - 64-bit VM, normal oops
3922 if (base_off % BytesPerLong != 0) {
3923 assert(UseCompressedOops, "");
3924 if (is_array) {
3925 // Exclude length to copy by 8 bytes words.
3926 base_off += sizeof(int);
3927 } else {
3928 // Include klass to copy by 8 bytes words.
3929 base_off = instanceOopDesc::klass_offset_in_bytes();
3930 }
3931 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
3932 }
3933 src = basic_plus_adr(src, base_off);
3934 dest = basic_plus_adr(dest, base_off);
3936 // Compute the length also, if needed:
3937 Node* countx = size;
3938 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
3939 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
3941 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
3942 bool disjoint_bases = true;
3943 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
3944 src, NULL, dest, NULL, countx);
3946 // If necessary, emit some card marks afterwards. (Non-arrays only.)
3947 if (card_mark) {
3948 assert(!is_array, "");
3949 // Put in store barrier for any and all oops we are sticking
3950 // into this object. (We could avoid this if we could prove
3951 // that the object type contains no oop fields at all.)
3952 Node* no_particular_value = NULL;
3953 Node* no_particular_field = NULL;
3954 int raw_adr_idx = Compile::AliasIdxRaw;
3955 post_barrier(control(),
3956 memory(raw_adr_type),
3957 new_obj,
3958 no_particular_field,
3959 raw_adr_idx,
3960 no_particular_value,
3961 T_OBJECT,
3962 false);
3963 }
3965 // Move the original CheckCastPP after arraycopy.
3966 _gvn.hash_delete(alloc_obj);
3967 alloc_obj->set_req(0, control());
3968 // Replace raw memory edge with new CheckCastPP to have a live oop
3969 // at safepoints instead of raw value.
3970 assert(new_obj->is_CheckCastPP() && new_obj->in(1) == alloc_obj->in(1), "sanity");
3971 alloc_obj->set_req(1, new_obj); // cast to the original type
3972 _gvn.hash_find_insert(alloc_obj); // put back into GVN table
3973 // Restore in the locally valid dest_oop.
3974 replace_in_map(new_obj, alloc_obj);
3975 }
3977 //------------------------inline_native_clone----------------------------
3978 // Here are the simple edge cases:
3979 // null receiver => normal trap
3980 // virtual and clone was overridden => slow path to out-of-line clone
3981 // not cloneable or finalizer => slow path to out-of-line Object.clone
3982 //
3983 // The general case has two steps, allocation and copying.
3984 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
3985 //
3986 // Copying also has two cases, oop arrays and everything else.
3987 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
3988 // Everything else uses the tight inline loop supplied by CopyArrayNode.
3989 //
3990 // These steps fold up nicely if and when the cloned object's klass
3991 // can be sharply typed as an object array, a type array, or an instance.
3992 //
3993 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
3994 int nargs = 1;
3995 Node* obj = null_check_receiver(callee());
3996 if (stopped()) return true;
3997 Node* obj_klass = load_object_klass(obj);
3998 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
3999 const TypeOopPtr* toop = ((tklass != NULL)
4000 ? tklass->as_instance_type()
4001 : TypeInstPtr::NOTNULL);
4003 // Conservatively insert a memory barrier on all memory slices.
4004 // Do not let writes into the original float below the clone.
4005 insert_mem_bar(Op_MemBarCPUOrder);
4007 // paths into result_reg:
4008 enum {
4009 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4010 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4011 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4012 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4013 PATH_LIMIT
4014 };
4015 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4016 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4017 TypeInstPtr::NOTNULL);
4018 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4019 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4020 TypePtr::BOTTOM);
4021 record_for_igvn(result_reg);
4023 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4024 int raw_adr_idx = Compile::AliasIdxRaw;
4025 const bool raw_mem_only = true;
4027 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4028 if (array_ctl != NULL) {
4029 // It's an array.
4030 PreserveJVMState pjvms(this);
4031 set_control(array_ctl);
4032 Node* obj_length = load_array_length(obj);
4033 Node* obj_size = NULL;
4034 Node* alloc_obj = new_array(obj_klass, obj_length, nargs,
4035 raw_mem_only, &obj_size);
4037 if (!use_ReduceInitialCardMarks()) {
4038 // If it is an oop array, it requires very special treatment,
4039 // because card marking is required on each card of the array.
4040 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4041 if (is_obja != NULL) {
4042 PreserveJVMState pjvms2(this);
4043 set_control(is_obja);
4044 // Generate a direct call to the right arraycopy function(s).
4045 bool disjoint_bases = true;
4046 bool length_never_negative = true;
4047 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4048 obj, intcon(0), alloc_obj, intcon(0),
4049 obj_length,
4050 disjoint_bases, length_never_negative);
4051 result_reg->init_req(_objArray_path, control());
4052 result_val->init_req(_objArray_path, alloc_obj);
4053 result_i_o ->set_req(_objArray_path, i_o());
4054 result_mem ->set_req(_objArray_path, reset_memory());
4055 }
4056 }
4057 // We can dispense with card marks if we know the allocation
4058 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4059 // causes the non-eden paths to simulate a fresh allocation,
4060 // insofar that no further card marks are required to initialize
4061 // the object.
4063 // Otherwise, there are no card marks to worry about.
4065 if (!stopped()) {
4066 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4068 // Present the results of the copy.
4069 result_reg->init_req(_array_path, control());
4070 result_val->init_req(_array_path, alloc_obj);
4071 result_i_o ->set_req(_array_path, i_o());
4072 result_mem ->set_req(_array_path, reset_memory());
4073 }
4074 }
4076 // We only go to the instance fast case code if we pass a number of guards.
4077 // The paths which do not pass are accumulated in the slow_region.
4078 RegionNode* slow_region = new (C, 1) RegionNode(1);
4079 record_for_igvn(slow_region);
4080 if (!stopped()) {
4081 // It's an instance (we did array above). Make the slow-path tests.
4082 // If this is a virtual call, we generate a funny guard. We grab
4083 // the vtable entry corresponding to clone() from the target object.
4084 // If the target method which we are calling happens to be the
4085 // Object clone() method, we pass the guard. We do not need this
4086 // guard for non-virtual calls; the caller is known to be the native
4087 // Object clone().
4088 if (is_virtual) {
4089 generate_virtual_guard(obj_klass, slow_region);
4090 }
4092 // The object must be cloneable and must not have a finalizer.
4093 // Both of these conditions may be checked in a single test.
4094 // We could optimize the cloneable test further, but we don't care.
4095 generate_access_flags_guard(obj_klass,
4096 // Test both conditions:
4097 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4098 // Must be cloneable but not finalizer:
4099 JVM_ACC_IS_CLONEABLE,
4100 slow_region);
4101 }
4103 if (!stopped()) {
4104 // It's an instance, and it passed the slow-path tests.
4105 PreserveJVMState pjvms(this);
4106 Node* obj_size = NULL;
4107 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4109 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4111 // Present the results of the slow call.
4112 result_reg->init_req(_instance_path, control());
4113 result_val->init_req(_instance_path, alloc_obj);
4114 result_i_o ->set_req(_instance_path, i_o());
4115 result_mem ->set_req(_instance_path, reset_memory());
4116 }
4118 // Generate code for the slow case. We make a call to clone().
4119 set_control(_gvn.transform(slow_region));
4120 if (!stopped()) {
4121 PreserveJVMState pjvms(this);
4122 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4123 Node* slow_result = set_results_for_java_call(slow_call);
4124 // this->control() comes from set_results_for_java_call
4125 result_reg->init_req(_slow_path, control());
4126 result_val->init_req(_slow_path, slow_result);
4127 result_i_o ->set_req(_slow_path, i_o());
4128 result_mem ->set_req(_slow_path, reset_memory());
4129 }
4131 // Return the combined state.
4132 set_control( _gvn.transform(result_reg) );
4133 set_i_o( _gvn.transform(result_i_o) );
4134 set_all_memory( _gvn.transform(result_mem) );
4136 push(_gvn.transform(result_val));
4138 return true;
4139 }
4142 // constants for computing the copy function
4143 enum {
4144 COPYFUNC_UNALIGNED = 0,
4145 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4146 COPYFUNC_CONJOINT = 0,
4147 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4148 };
4150 // Note: The condition "disjoint" applies also for overlapping copies
4151 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4152 static address
4153 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
4154 int selector =
4155 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4156 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4158 #define RETURN_STUB(xxx_arraycopy) { \
4159 name = #xxx_arraycopy; \
4160 return StubRoutines::xxx_arraycopy(); }
4162 switch (t) {
4163 case T_BYTE:
4164 case T_BOOLEAN:
4165 switch (selector) {
4166 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4167 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4168 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4169 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4170 }
4171 case T_CHAR:
4172 case T_SHORT:
4173 switch (selector) {
4174 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4175 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4176 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4177 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4178 }
4179 case T_INT:
4180 case T_FLOAT:
4181 switch (selector) {
4182 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4183 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4184 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4185 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4186 }
4187 case T_DOUBLE:
4188 case T_LONG:
4189 switch (selector) {
4190 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4191 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4192 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4193 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4194 }
4195 case T_ARRAY:
4196 case T_OBJECT:
4197 switch (selector) {
4198 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
4199 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
4200 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
4201 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
4202 }
4203 default:
4204 ShouldNotReachHere();
4205 return NULL;
4206 }
4208 #undef RETURN_STUB
4209 }
4211 //------------------------------basictype2arraycopy----------------------------
4212 address LibraryCallKit::basictype2arraycopy(BasicType t,
4213 Node* src_offset,
4214 Node* dest_offset,
4215 bool disjoint_bases,
4216 const char* &name) {
4217 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4218 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4220 bool aligned = false;
4221 bool disjoint = disjoint_bases;
4223 // if the offsets are the same, we can treat the memory regions as
4224 // disjoint, because either the memory regions are in different arrays,
4225 // or they are identical (which we can treat as disjoint.) We can also
4226 // treat a copy with a destination index less that the source index
4227 // as disjoint since a low->high copy will work correctly in this case.
4228 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4229 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4230 // both indices are constants
4231 int s_offs = src_offset_inttype->get_con();
4232 int d_offs = dest_offset_inttype->get_con();
4233 int element_size = type2aelembytes(t);
4234 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4235 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4236 if (s_offs >= d_offs) disjoint = true;
4237 } else if (src_offset == dest_offset && src_offset != NULL) {
4238 // This can occur if the offsets are identical non-constants.
4239 disjoint = true;
4240 }
4242 return select_arraycopy_function(t, aligned, disjoint, name);
4243 }
4246 //------------------------------inline_arraycopy-----------------------
4247 bool LibraryCallKit::inline_arraycopy() {
4248 // Restore the stack and pop off the arguments.
4249 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4250 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4252 Node *src = argument(0);
4253 Node *src_offset = argument(1);
4254 Node *dest = argument(2);
4255 Node *dest_offset = argument(3);
4256 Node *length = argument(4);
4258 // Compile time checks. If any of these checks cannot be verified at compile time,
4259 // we do not make a fast path for this call. Instead, we let the call remain as it
4260 // is. The checks we choose to mandate at compile time are:
4261 //
4262 // (1) src and dest are arrays.
4263 const Type* src_type = src->Value(&_gvn);
4264 const Type* dest_type = dest->Value(&_gvn);
4265 const TypeAryPtr* top_src = src_type->isa_aryptr();
4266 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4267 if (top_src == NULL || top_src->klass() == NULL ||
4268 top_dest == NULL || top_dest->klass() == NULL) {
4269 // Conservatively insert a memory barrier on all memory slices.
4270 // Do not let writes into the source float below the arraycopy.
4271 insert_mem_bar(Op_MemBarCPUOrder);
4273 // Call StubRoutines::generic_arraycopy stub.
4274 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4275 src, src_offset, dest, dest_offset, length);
4277 // Do not let reads from the destination float above the arraycopy.
4278 // Since we cannot type the arrays, we don't know which slices
4279 // might be affected. We could restrict this barrier only to those
4280 // memory slices which pertain to array elements--but don't bother.
4281 if (!InsertMemBarAfterArraycopy)
4282 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4283 insert_mem_bar(Op_MemBarCPUOrder);
4284 return true;
4285 }
4287 // (2) src and dest arrays must have elements of the same BasicType
4288 // Figure out the size and type of the elements we will be copying.
4289 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4290 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4291 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4292 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4294 if (src_elem != dest_elem || dest_elem == T_VOID) {
4295 // The component types are not the same or are not recognized. Punt.
4296 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4297 generate_slow_arraycopy(TypePtr::BOTTOM,
4298 src, src_offset, dest, dest_offset, length);
4299 return true;
4300 }
4302 //---------------------------------------------------------------------------
4303 // We will make a fast path for this call to arraycopy.
4305 // We have the following tests left to perform:
4306 //
4307 // (3) src and dest must not be null.
4308 // (4) src_offset must not be negative.
4309 // (5) dest_offset must not be negative.
4310 // (6) length must not be negative.
4311 // (7) src_offset + length must not exceed length of src.
4312 // (8) dest_offset + length must not exceed length of dest.
4313 // (9) each element of an oop array must be assignable
4315 RegionNode* slow_region = new (C, 1) RegionNode(1);
4316 record_for_igvn(slow_region);
4318 // (3) operands must not be null
4319 // We currently perform our null checks with the do_null_check routine.
4320 // This means that the null exceptions will be reported in the caller
4321 // rather than (correctly) reported inside of the native arraycopy call.
4322 // This should be corrected, given time. We do our null check with the
4323 // stack pointer restored.
4324 _sp += nargs;
4325 src = do_null_check(src, T_ARRAY);
4326 dest = do_null_check(dest, T_ARRAY);
4327 _sp -= nargs;
4329 // (4) src_offset must not be negative.
4330 generate_negative_guard(src_offset, slow_region);
4332 // (5) dest_offset must not be negative.
4333 generate_negative_guard(dest_offset, slow_region);
4335 // (6) length must not be negative (moved to generate_arraycopy()).
4336 // generate_negative_guard(length, slow_region);
4338 // (7) src_offset + length must not exceed length of src.
4339 generate_limit_guard(src_offset, length,
4340 load_array_length(src),
4341 slow_region);
4343 // (8) dest_offset + length must not exceed length of dest.
4344 generate_limit_guard(dest_offset, length,
4345 load_array_length(dest),
4346 slow_region);
4348 // (9) each element of an oop array must be assignable
4349 // The generate_arraycopy subroutine checks this.
4351 // This is where the memory effects are placed:
4352 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4353 generate_arraycopy(adr_type, dest_elem,
4354 src, src_offset, dest, dest_offset, length,
4355 false, false, slow_region);
4357 return true;
4358 }
4360 //-----------------------------generate_arraycopy----------------------
4361 // Generate an optimized call to arraycopy.
4362 // Caller must guard against non-arrays.
4363 // Caller must determine a common array basic-type for both arrays.
4364 // Caller must validate offsets against array bounds.
4365 // The slow_region has already collected guard failure paths
4366 // (such as out of bounds length or non-conformable array types).
4367 // The generated code has this shape, in general:
4368 //
4369 // if (length == 0) return // via zero_path
4370 // slowval = -1
4371 // if (types unknown) {
4372 // slowval = call generic copy loop
4373 // if (slowval == 0) return // via checked_path
4374 // } else if (indexes in bounds) {
4375 // if ((is object array) && !(array type check)) {
4376 // slowval = call checked copy loop
4377 // if (slowval == 0) return // via checked_path
4378 // } else {
4379 // call bulk copy loop
4380 // return // via fast_path
4381 // }
4382 // }
4383 // // adjust params for remaining work:
4384 // if (slowval != -1) {
4385 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4386 // }
4387 // slow_region:
4388 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4389 // return // via slow_call_path
4390 //
4391 // This routine is used from several intrinsics: System.arraycopy,
4392 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4393 //
4394 void
4395 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4396 BasicType basic_elem_type,
4397 Node* src, Node* src_offset,
4398 Node* dest, Node* dest_offset,
4399 Node* copy_length,
4400 bool disjoint_bases,
4401 bool length_never_negative,
4402 RegionNode* slow_region) {
4404 if (slow_region == NULL) {
4405 slow_region = new(C,1) RegionNode(1);
4406 record_for_igvn(slow_region);
4407 }
4409 Node* original_dest = dest;
4410 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4411 bool must_clear_dest = false;
4413 // See if this is the initialization of a newly-allocated array.
4414 // If so, we will take responsibility here for initializing it to zero.
4415 // (Note: Because tightly_coupled_allocation performs checks on the
4416 // out-edges of the dest, we need to avoid making derived pointers
4417 // from it until we have checked its uses.)
4418 if (ReduceBulkZeroing
4419 && !ZeroTLAB // pointless if already zeroed
4420 && basic_elem_type != T_CONFLICT // avoid corner case
4421 && !_gvn.eqv_uncast(src, dest)
4422 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4423 != NULL)
4424 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4425 && alloc->maybe_set_complete(&_gvn)) {
4426 // "You break it, you buy it."
4427 InitializeNode* init = alloc->initialization();
4428 assert(init->is_complete(), "we just did this");
4429 assert(dest->is_CheckCastPP(), "sanity");
4430 assert(dest->in(0)->in(0) == init, "dest pinned");
4432 // Cast to Object for arraycopy.
4433 // We can't use the original CheckCastPP since it should be moved
4434 // after the arraycopy to prevent stores flowing above it.
4435 Node* new_obj = new(C, 2) CheckCastPPNode(dest->in(0), dest->in(1),
4436 TypeInstPtr::NOTNULL);
4437 dest = _gvn.transform(new_obj);
4438 // Substitute in the locally valid dest_oop.
4439 replace_in_map(original_dest, dest);
4440 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4441 // From this point on, every exit path is responsible for
4442 // initializing any non-copied parts of the object to zero.
4443 must_clear_dest = true;
4444 } else {
4445 // No zeroing elimination here.
4446 alloc = NULL;
4447 //original_dest = dest;
4448 //must_clear_dest = false;
4449 }
4451 // Results are placed here:
4452 enum { fast_path = 1, // normal void-returning assembly stub
4453 checked_path = 2, // special assembly stub with cleanup
4454 slow_call_path = 3, // something went wrong; call the VM
4455 zero_path = 4, // bypass when length of copy is zero
4456 bcopy_path = 5, // copy primitive array by 64-bit blocks
4457 PATH_LIMIT = 6
4458 };
4459 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4460 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4461 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4462 record_for_igvn(result_region);
4463 _gvn.set_type_bottom(result_i_o);
4464 _gvn.set_type_bottom(result_memory);
4465 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4467 // The slow_control path:
4468 Node* slow_control;
4469 Node* slow_i_o = i_o();
4470 Node* slow_mem = memory(adr_type);
4471 debug_only(slow_control = (Node*) badAddress);
4473 // Checked control path:
4474 Node* checked_control = top();
4475 Node* checked_mem = NULL;
4476 Node* checked_i_o = NULL;
4477 Node* checked_value = NULL;
4479 if (basic_elem_type == T_CONFLICT) {
4480 assert(!must_clear_dest, "");
4481 Node* cv = generate_generic_arraycopy(adr_type,
4482 src, src_offset, dest, dest_offset,
4483 copy_length);
4484 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4485 checked_control = control();
4486 checked_i_o = i_o();
4487 checked_mem = memory(adr_type);
4488 checked_value = cv;
4489 set_control(top()); // no fast path
4490 }
4492 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4493 if (not_pos != NULL) {
4494 PreserveJVMState pjvms(this);
4495 set_control(not_pos);
4497 // (6) length must not be negative.
4498 if (!length_never_negative) {
4499 generate_negative_guard(copy_length, slow_region);
4500 }
4502 // copy_length is 0.
4503 if (!stopped() && must_clear_dest) {
4504 Node* dest_length = alloc->in(AllocateNode::ALength);
4505 if (_gvn.eqv_uncast(copy_length, dest_length)
4506 || _gvn.find_int_con(dest_length, 1) <= 0) {
4507 // There is no zeroing to do. No need for a secondary raw memory barrier.
4508 } else {
4509 // Clear the whole thing since there are no source elements to copy.
4510 generate_clear_array(adr_type, dest, basic_elem_type,
4511 intcon(0), NULL,
4512 alloc->in(AllocateNode::AllocSize));
4513 // Use a secondary InitializeNode as raw memory barrier.
4514 // Currently it is needed only on this path since other
4515 // paths have stub or runtime calls as raw memory barriers.
4516 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4517 Compile::AliasIdxRaw,
4518 top())->as_Initialize();
4519 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4520 }
4521 }
4523 // Present the results of the fast call.
4524 result_region->init_req(zero_path, control());
4525 result_i_o ->init_req(zero_path, i_o());
4526 result_memory->init_req(zero_path, memory(adr_type));
4527 }
4529 if (!stopped() && must_clear_dest) {
4530 // We have to initialize the *uncopied* part of the array to zero.
4531 // The copy destination is the slice dest[off..off+len]. The other slices
4532 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4533 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4534 Node* dest_length = alloc->in(AllocateNode::ALength);
4535 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4536 copy_length) );
4538 // If there is a head section that needs zeroing, do it now.
4539 if (find_int_con(dest_offset, -1) != 0) {
4540 generate_clear_array(adr_type, dest, basic_elem_type,
4541 intcon(0), dest_offset,
4542 NULL);
4543 }
4545 // Next, perform a dynamic check on the tail length.
4546 // It is often zero, and we can win big if we prove this.
4547 // There are two wins: Avoid generating the ClearArray
4548 // with its attendant messy index arithmetic, and upgrade
4549 // the copy to a more hardware-friendly word size of 64 bits.
4550 Node* tail_ctl = NULL;
4551 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4552 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4553 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4554 tail_ctl = generate_slow_guard(bol_lt, NULL);
4555 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4556 }
4558 // At this point, let's assume there is no tail.
4559 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4560 // There is no tail. Try an upgrade to a 64-bit copy.
4561 bool didit = false;
4562 { PreserveJVMState pjvms(this);
4563 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4564 src, src_offset, dest, dest_offset,
4565 dest_size);
4566 if (didit) {
4567 // Present the results of the block-copying fast call.
4568 result_region->init_req(bcopy_path, control());
4569 result_i_o ->init_req(bcopy_path, i_o());
4570 result_memory->init_req(bcopy_path, memory(adr_type));
4571 }
4572 }
4573 if (didit)
4574 set_control(top()); // no regular fast path
4575 }
4577 // Clear the tail, if any.
4578 if (tail_ctl != NULL) {
4579 Node* notail_ctl = stopped() ? NULL : control();
4580 set_control(tail_ctl);
4581 if (notail_ctl == NULL) {
4582 generate_clear_array(adr_type, dest, basic_elem_type,
4583 dest_tail, NULL,
4584 dest_size);
4585 } else {
4586 // Make a local merge.
4587 Node* done_ctl = new(C,3) RegionNode(3);
4588 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4589 done_ctl->init_req(1, notail_ctl);
4590 done_mem->init_req(1, memory(adr_type));
4591 generate_clear_array(adr_type, dest, basic_elem_type,
4592 dest_tail, NULL,
4593 dest_size);
4594 done_ctl->init_req(2, control());
4595 done_mem->init_req(2, memory(adr_type));
4596 set_control( _gvn.transform(done_ctl) );
4597 set_memory( _gvn.transform(done_mem), adr_type );
4598 }
4599 }
4600 }
4602 BasicType copy_type = basic_elem_type;
4603 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4604 if (!stopped() && copy_type == T_OBJECT) {
4605 // If src and dest have compatible element types, we can copy bits.
4606 // Types S[] and D[] are compatible if D is a supertype of S.
4607 //
4608 // If they are not, we will use checked_oop_disjoint_arraycopy,
4609 // which performs a fast optimistic per-oop check, and backs off
4610 // further to JVM_ArrayCopy on the first per-oop check that fails.
4611 // (Actually, we don't move raw bits only; the GC requires card marks.)
4613 // Get the klassOop for both src and dest
4614 Node* src_klass = load_object_klass(src);
4615 Node* dest_klass = load_object_klass(dest);
4617 // Generate the subtype check.
4618 // This might fold up statically, or then again it might not.
4619 //
4620 // Non-static example: Copying List<String>.elements to a new String[].
4621 // The backing store for a List<String> is always an Object[],
4622 // but its elements are always type String, if the generic types
4623 // are correct at the source level.
4624 //
4625 // Test S[] against D[], not S against D, because (probably)
4626 // the secondary supertype cache is less busy for S[] than S.
4627 // This usually only matters when D is an interface.
4628 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4629 // Plug failing path into checked_oop_disjoint_arraycopy
4630 if (not_subtype_ctrl != top()) {
4631 PreserveJVMState pjvms(this);
4632 set_control(not_subtype_ctrl);
4633 // (At this point we can assume disjoint_bases, since types differ.)
4634 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4635 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4636 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4637 Node* dest_elem_klass = _gvn.transform(n1);
4638 Node* cv = generate_checkcast_arraycopy(adr_type,
4639 dest_elem_klass,
4640 src, src_offset, dest, dest_offset,
4641 copy_length);
4642 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4643 checked_control = control();
4644 checked_i_o = i_o();
4645 checked_mem = memory(adr_type);
4646 checked_value = cv;
4647 }
4648 // At this point we know we do not need type checks on oop stores.
4650 // Let's see if we need card marks:
4651 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4652 // If we do not need card marks, copy using the jint or jlong stub.
4653 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4654 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4655 "sizes agree");
4656 }
4657 }
4659 if (!stopped()) {
4660 // Generate the fast path, if possible.
4661 PreserveJVMState pjvms(this);
4662 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4663 src, src_offset, dest, dest_offset,
4664 ConvI2X(copy_length));
4666 // Present the results of the fast call.
4667 result_region->init_req(fast_path, control());
4668 result_i_o ->init_req(fast_path, i_o());
4669 result_memory->init_req(fast_path, memory(adr_type));
4670 }
4672 // Here are all the slow paths up to this point, in one bundle:
4673 slow_control = top();
4674 if (slow_region != NULL)
4675 slow_control = _gvn.transform(slow_region);
4676 debug_only(slow_region = (RegionNode*)badAddress);
4678 set_control(checked_control);
4679 if (!stopped()) {
4680 // Clean up after the checked call.
4681 // The returned value is either 0 or -1^K,
4682 // where K = number of partially transferred array elements.
4683 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4684 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4685 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4687 // If it is 0, we are done, so transfer to the end.
4688 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4689 result_region->init_req(checked_path, checks_done);
4690 result_i_o ->init_req(checked_path, checked_i_o);
4691 result_memory->init_req(checked_path, checked_mem);
4693 // If it is not zero, merge into the slow call.
4694 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4695 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4696 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4697 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4698 record_for_igvn(slow_reg2);
4699 slow_reg2 ->init_req(1, slow_control);
4700 slow_i_o2 ->init_req(1, slow_i_o);
4701 slow_mem2 ->init_req(1, slow_mem);
4702 slow_reg2 ->init_req(2, control());
4703 slow_i_o2 ->init_req(2, checked_i_o);
4704 slow_mem2 ->init_req(2, checked_mem);
4706 slow_control = _gvn.transform(slow_reg2);
4707 slow_i_o = _gvn.transform(slow_i_o2);
4708 slow_mem = _gvn.transform(slow_mem2);
4710 if (alloc != NULL) {
4711 // We'll restart from the very beginning, after zeroing the whole thing.
4712 // This can cause double writes, but that's OK since dest is brand new.
4713 // So we ignore the low 31 bits of the value returned from the stub.
4714 } else {
4715 // We must continue the copy exactly where it failed, or else
4716 // another thread might see the wrong number of writes to dest.
4717 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4718 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4719 slow_offset->init_req(1, intcon(0));
4720 slow_offset->init_req(2, checked_offset);
4721 slow_offset = _gvn.transform(slow_offset);
4723 // Adjust the arguments by the conditionally incoming offset.
4724 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4725 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4726 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4728 // Tweak the node variables to adjust the code produced below:
4729 src_offset = src_off_plus;
4730 dest_offset = dest_off_plus;
4731 copy_length = length_minus;
4732 }
4733 }
4735 set_control(slow_control);
4736 if (!stopped()) {
4737 // Generate the slow path, if needed.
4738 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4740 set_memory(slow_mem, adr_type);
4741 set_i_o(slow_i_o);
4743 if (must_clear_dest) {
4744 generate_clear_array(adr_type, dest, basic_elem_type,
4745 intcon(0), NULL,
4746 alloc->in(AllocateNode::AllocSize));
4747 }
4749 generate_slow_arraycopy(adr_type,
4750 src, src_offset, dest, dest_offset,
4751 copy_length);
4753 result_region->init_req(slow_call_path, control());
4754 result_i_o ->init_req(slow_call_path, i_o());
4755 result_memory->init_req(slow_call_path, memory(adr_type));
4756 }
4758 // Remove unused edges.
4759 for (uint i = 1; i < result_region->req(); i++) {
4760 if (result_region->in(i) == NULL)
4761 result_region->init_req(i, top());
4762 }
4764 // Finished; return the combined state.
4765 set_control( _gvn.transform(result_region) );
4766 set_i_o( _gvn.transform(result_i_o) );
4767 set_memory( _gvn.transform(result_memory), adr_type );
4769 if (dest != original_dest) {
4770 // Pin the "finished" array node after the arraycopy/zeroing operations.
4771 _gvn.hash_delete(original_dest);
4772 original_dest->set_req(0, control());
4773 // Replace raw memory edge with new CheckCastPP to have a live oop
4774 // at safepoints instead of raw value.
4775 assert(dest->is_CheckCastPP() && dest->in(1) == original_dest->in(1), "sanity");
4776 original_dest->set_req(1, dest); // cast to the original type
4777 _gvn.hash_find_insert(original_dest); // put back into GVN table
4778 // Restore in the locally valid dest_oop.
4779 replace_in_map(dest, original_dest);
4780 }
4781 // The memory edges above are precise in order to model effects around
4782 // array copies accurately to allow value numbering of field loads around
4783 // arraycopy. Such field loads, both before and after, are common in Java
4784 // collections and similar classes involving header/array data structures.
4785 //
4786 // But with low number of register or when some registers are used or killed
4787 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4788 // The next memory barrier is added to avoid it. If the arraycopy can be
4789 // optimized away (which it can, sometimes) then we can manually remove
4790 // the membar also.
4791 if (InsertMemBarAfterArraycopy)
4792 insert_mem_bar(Op_MemBarCPUOrder);
4793 }
4796 // Helper function which determines if an arraycopy immediately follows
4797 // an allocation, with no intervening tests or other escapes for the object.
4798 AllocateArrayNode*
4799 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4800 RegionNode* slow_region) {
4801 if (stopped()) return NULL; // no fast path
4802 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4804 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4805 if (alloc == NULL) return NULL;
4807 Node* rawmem = memory(Compile::AliasIdxRaw);
4808 // Is the allocation's memory state untouched?
4809 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4810 // Bail out if there have been raw-memory effects since the allocation.
4811 // (Example: There might have been a call or safepoint.)
4812 return NULL;
4813 }
4814 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4815 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4816 return NULL;
4817 }
4819 // There must be no unexpected observers of this allocation.
4820 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4821 Node* obs = ptr->fast_out(i);
4822 if (obs != this->map()) {
4823 return NULL;
4824 }
4825 }
4827 // This arraycopy must unconditionally follow the allocation of the ptr.
4828 Node* alloc_ctl = ptr->in(0);
4829 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4831 Node* ctl = control();
4832 while (ctl != alloc_ctl) {
4833 // There may be guards which feed into the slow_region.
4834 // Any other control flow means that we might not get a chance
4835 // to finish initializing the allocated object.
4836 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4837 IfNode* iff = ctl->in(0)->as_If();
4838 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4839 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4840 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4841 ctl = iff->in(0); // This test feeds the known slow_region.
4842 continue;
4843 }
4844 // One more try: Various low-level checks bottom out in
4845 // uncommon traps. If the debug-info of the trap omits
4846 // any reference to the allocation, as we've already
4847 // observed, then there can be no objection to the trap.
4848 bool found_trap = false;
4849 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4850 Node* obs = not_ctl->fast_out(j);
4851 if (obs->in(0) == not_ctl && obs->is_Call() &&
4852 (obs->as_Call()->entry_point() ==
4853 SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
4854 found_trap = true; break;
4855 }
4856 }
4857 if (found_trap) {
4858 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4859 continue;
4860 }
4861 }
4862 return NULL;
4863 }
4865 // If we get this far, we have an allocation which immediately
4866 // precedes the arraycopy, and we can take over zeroing the new object.
4867 // The arraycopy will finish the initialization, and provide
4868 // a new control state to which we will anchor the destination pointer.
4870 return alloc;
4871 }
4873 // Helper for initialization of arrays, creating a ClearArray.
4874 // It writes zero bits in [start..end), within the body of an array object.
4875 // The memory effects are all chained onto the 'adr_type' alias category.
4876 //
4877 // Since the object is otherwise uninitialized, we are free
4878 // to put a little "slop" around the edges of the cleared area,
4879 // as long as it does not go back into the array's header,
4880 // or beyond the array end within the heap.
4881 //
4882 // The lower edge can be rounded down to the nearest jint and the
4883 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4884 //
4885 // Arguments:
4886 // adr_type memory slice where writes are generated
4887 // dest oop of the destination array
4888 // basic_elem_type element type of the destination
4889 // slice_idx array index of first element to store
4890 // slice_len number of elements to store (or NULL)
4891 // dest_size total size in bytes of the array object
4892 //
4893 // Exactly one of slice_len or dest_size must be non-NULL.
4894 // If dest_size is non-NULL, zeroing extends to the end of the object.
4895 // If slice_len is non-NULL, the slice_idx value must be a constant.
4896 void
4897 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
4898 Node* dest,
4899 BasicType basic_elem_type,
4900 Node* slice_idx,
4901 Node* slice_len,
4902 Node* dest_size) {
4903 // one or the other but not both of slice_len and dest_size:
4904 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
4905 if (slice_len == NULL) slice_len = top();
4906 if (dest_size == NULL) dest_size = top();
4908 // operate on this memory slice:
4909 Node* mem = memory(adr_type); // memory slice to operate on
4911 // scaling and rounding of indexes:
4912 int scale = exact_log2(type2aelembytes(basic_elem_type));
4913 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
4914 int clear_low = (-1 << scale) & (BytesPerInt - 1);
4915 int bump_bit = (-1 << scale) & BytesPerInt;
4917 // determine constant starts and ends
4918 const intptr_t BIG_NEG = -128;
4919 assert(BIG_NEG + 2*abase < 0, "neg enough");
4920 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
4921 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
4922 if (slice_len_con == 0) {
4923 return; // nothing to do here
4924 }
4925 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
4926 intptr_t end_con = find_intptr_t_con(dest_size, -1);
4927 if (slice_idx_con >= 0 && slice_len_con >= 0) {
4928 assert(end_con < 0, "not two cons");
4929 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
4930 BytesPerLong);
4931 }
4933 if (start_con >= 0 && end_con >= 0) {
4934 // Constant start and end. Simple.
4935 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4936 start_con, end_con, &_gvn);
4937 } else if (start_con >= 0 && dest_size != top()) {
4938 // Constant start, pre-rounded end after the tail of the array.
4939 Node* end = dest_size;
4940 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4941 start_con, end, &_gvn);
4942 } else if (start_con >= 0 && slice_len != top()) {
4943 // Constant start, non-constant end. End needs rounding up.
4944 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
4945 intptr_t end_base = abase + (slice_idx_con << scale);
4946 int end_round = (-1 << scale) & (BytesPerLong - 1);
4947 Node* end = ConvI2X(slice_len);
4948 if (scale != 0)
4949 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
4950 end_base += end_round;
4951 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
4952 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
4953 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4954 start_con, end, &_gvn);
4955 } else if (start_con < 0 && dest_size != top()) {
4956 // Non-constant start, pre-rounded end after the tail of the array.
4957 // This is almost certainly a "round-to-end" operation.
4958 Node* start = slice_idx;
4959 start = ConvI2X(start);
4960 if (scale != 0)
4961 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
4962 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
4963 if ((bump_bit | clear_low) != 0) {
4964 int to_clear = (bump_bit | clear_low);
4965 // Align up mod 8, then store a jint zero unconditionally
4966 // just before the mod-8 boundary.
4967 if (((abase + bump_bit) & ~to_clear) - bump_bit
4968 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
4969 bump_bit = 0;
4970 assert((abase & to_clear) == 0, "array base must be long-aligned");
4971 } else {
4972 // Bump 'start' up to (or past) the next jint boundary:
4973 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
4974 assert((abase & clear_low) == 0, "array base must be int-aligned");
4975 }
4976 // Round bumped 'start' down to jlong boundary in body of array.
4977 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
4978 if (bump_bit != 0) {
4979 // Store a zero to the immediately preceding jint:
4980 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
4981 Node* p1 = basic_plus_adr(dest, x1);
4982 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
4983 mem = _gvn.transform(mem);
4984 }
4985 }
4986 Node* end = dest_size; // pre-rounded
4987 mem = ClearArrayNode::clear_memory(control(), mem, dest,
4988 start, end, &_gvn);
4989 } else {
4990 // Non-constant start, unrounded non-constant end.
4991 // (Nobody zeroes a random midsection of an array using this routine.)
4992 ShouldNotReachHere(); // fix caller
4993 }
4995 // Done.
4996 set_memory(mem, adr_type);
4997 }
5000 bool
5001 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5002 BasicType basic_elem_type,
5003 AllocateNode* alloc,
5004 Node* src, Node* src_offset,
5005 Node* dest, Node* dest_offset,
5006 Node* dest_size) {
5007 // See if there is an advantage from block transfer.
5008 int scale = exact_log2(type2aelembytes(basic_elem_type));
5009 if (scale >= LogBytesPerLong)
5010 return false; // it is already a block transfer
5012 // Look at the alignment of the starting offsets.
5013 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5014 const intptr_t BIG_NEG = -128;
5015 assert(BIG_NEG + 2*abase < 0, "neg enough");
5017 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5018 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5019 if (src_off < 0 || dest_off < 0)
5020 // At present, we can only understand constants.
5021 return false;
5023 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5024 // Non-aligned; too bad.
5025 // One more chance: Pick off an initial 32-bit word.
5026 // This is a common case, since abase can be odd mod 8.
5027 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5028 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5029 Node* sptr = basic_plus_adr(src, src_off);
5030 Node* dptr = basic_plus_adr(dest, dest_off);
5031 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5032 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5033 src_off += BytesPerInt;
5034 dest_off += BytesPerInt;
5035 } else {
5036 return false;
5037 }
5038 }
5039 assert(src_off % BytesPerLong == 0, "");
5040 assert(dest_off % BytesPerLong == 0, "");
5042 // Do this copy by giant steps.
5043 Node* sptr = basic_plus_adr(src, src_off);
5044 Node* dptr = basic_plus_adr(dest, dest_off);
5045 Node* countx = dest_size;
5046 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5047 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5049 bool disjoint_bases = true; // since alloc != NULL
5050 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5051 sptr, NULL, dptr, NULL, countx);
5053 return true;
5054 }
5057 // Helper function; generates code for the slow case.
5058 // We make a call to a runtime method which emulates the native method,
5059 // but without the native wrapper overhead.
5060 void
5061 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5062 Node* src, Node* src_offset,
5063 Node* dest, Node* dest_offset,
5064 Node* copy_length) {
5065 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5066 OptoRuntime::slow_arraycopy_Type(),
5067 OptoRuntime::slow_arraycopy_Java(),
5068 "slow_arraycopy", adr_type,
5069 src, src_offset, dest, dest_offset,
5070 copy_length);
5072 // Handle exceptions thrown by this fellow:
5073 make_slow_call_ex(call, env()->Throwable_klass(), false);
5074 }
5076 // Helper function; generates code for cases requiring runtime checks.
5077 Node*
5078 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5079 Node* dest_elem_klass,
5080 Node* src, Node* src_offset,
5081 Node* dest, Node* dest_offset,
5082 Node* copy_length) {
5083 if (stopped()) return NULL;
5085 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
5086 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5087 return NULL;
5088 }
5090 // Pick out the parameters required to perform a store-check
5091 // for the target array. This is an optimistic check. It will
5092 // look in each non-null element's class, at the desired klass's
5093 // super_check_offset, for the desired klass.
5094 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5095 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5096 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
5097 Node* check_offset = _gvn.transform(n3);
5098 Node* check_value = dest_elem_klass;
5100 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5101 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5103 // (We know the arrays are never conjoint, because their types differ.)
5104 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5105 OptoRuntime::checkcast_arraycopy_Type(),
5106 copyfunc_addr, "checkcast_arraycopy", adr_type,
5107 // five arguments, of which two are
5108 // intptr_t (jlong in LP64)
5109 src_start, dest_start,
5110 copy_length XTOP,
5111 check_offset XTOP,
5112 check_value);
5114 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5115 }
5118 // Helper function; generates code for cases requiring runtime checks.
5119 Node*
5120 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5121 Node* src, Node* src_offset,
5122 Node* dest, Node* dest_offset,
5123 Node* copy_length) {
5124 if (stopped()) return NULL;
5126 address copyfunc_addr = StubRoutines::generic_arraycopy();
5127 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5128 return NULL;
5129 }
5131 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5132 OptoRuntime::generic_arraycopy_Type(),
5133 copyfunc_addr, "generic_arraycopy", adr_type,
5134 src, src_offset, dest, dest_offset, copy_length);
5136 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5137 }
5139 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5140 void
5141 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5142 BasicType basic_elem_type,
5143 bool disjoint_bases,
5144 Node* src, Node* src_offset,
5145 Node* dest, Node* dest_offset,
5146 Node* copy_length) {
5147 if (stopped()) return; // nothing to do
5149 Node* src_start = src;
5150 Node* dest_start = dest;
5151 if (src_offset != NULL || dest_offset != NULL) {
5152 assert(src_offset != NULL && dest_offset != NULL, "");
5153 src_start = array_element_address(src, src_offset, basic_elem_type);
5154 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5155 }
5157 // Figure out which arraycopy runtime method to call.
5158 const char* copyfunc_name = "arraycopy";
5159 address copyfunc_addr =
5160 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5161 disjoint_bases, copyfunc_name);
5163 // Call it. Note that the count_ix value is not scaled to a byte-size.
5164 make_runtime_call(RC_LEAF|RC_NO_FP,
5165 OptoRuntime::fast_arraycopy_Type(),
5166 copyfunc_addr, copyfunc_name, adr_type,
5167 src_start, dest_start, copy_length XTOP);
5168 }