|
1 /* |
|
2 * Copyright 1999-2007 Sun Microsystems, Inc. All Rights Reserved. |
|
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
|
4 * |
|
5 * This code is free software; you can redistribute it and/or modify it |
|
6 * under the terms of the GNU General Public License version 2 only, as |
|
7 * published by the Free Software Foundation. |
|
8 * |
|
9 * This code is distributed in the hope that it will be useful, but WITHOUT |
|
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
|
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
|
12 * version 2 for more details (a copy is included in the LICENSE file that |
|
13 * accompanied this code). |
|
14 * |
|
15 * You should have received a copy of the GNU General Public License version |
|
16 * 2 along with this work; if not, write to the Free Software Foundation, |
|
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
|
18 * |
|
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or |
|
21 * have any questions. |
|
22 * |
|
23 */ |
|
24 |
|
25 #include "incls/_precompiled.incl" |
|
26 #include "incls/_library_call.cpp.incl" |
|
27 |
|
28 class LibraryIntrinsic : public InlineCallGenerator { |
|
29 // Extend the set of intrinsics known to the runtime: |
|
30 public: |
|
31 private: |
|
32 bool _is_virtual; |
|
33 vmIntrinsics::ID _intrinsic_id; |
|
34 |
|
35 public: |
|
36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) |
|
37 : InlineCallGenerator(m), |
|
38 _is_virtual(is_virtual), |
|
39 _intrinsic_id(id) |
|
40 { |
|
41 } |
|
42 virtual bool is_intrinsic() const { return true; } |
|
43 virtual bool is_virtual() const { return _is_virtual; } |
|
44 virtual JVMState* generate(JVMState* jvms); |
|
45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } |
|
46 }; |
|
47 |
|
48 |
|
49 // Local helper class for LibraryIntrinsic: |
|
50 class LibraryCallKit : public GraphKit { |
|
51 private: |
|
52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called |
|
53 |
|
54 public: |
|
55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) |
|
56 : GraphKit(caller), |
|
57 _intrinsic(intrinsic) |
|
58 { |
|
59 } |
|
60 |
|
61 ciMethod* caller() const { return jvms()->method(); } |
|
62 int bci() const { return jvms()->bci(); } |
|
63 LibraryIntrinsic* intrinsic() const { return _intrinsic; } |
|
64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } |
|
65 ciMethod* callee() const { return _intrinsic->method(); } |
|
66 ciSignature* signature() const { return callee()->signature(); } |
|
67 int arg_size() const { return callee()->arg_size(); } |
|
68 |
|
69 bool try_to_inline(); |
|
70 |
|
71 // Helper functions to inline natives |
|
72 void push_result(RegionNode* region, PhiNode* value); |
|
73 Node* generate_guard(Node* test, RegionNode* region, float true_prob); |
|
74 Node* generate_slow_guard(Node* test, RegionNode* region); |
|
75 Node* generate_fair_guard(Node* test, RegionNode* region); |
|
76 Node* generate_negative_guard(Node* index, RegionNode* region, |
|
77 // resulting CastII of index: |
|
78 Node* *pos_index = NULL); |
|
79 Node* generate_nonpositive_guard(Node* index, bool never_negative, |
|
80 // resulting CastII of index: |
|
81 Node* *pos_index = NULL); |
|
82 Node* generate_limit_guard(Node* offset, Node* subseq_length, |
|
83 Node* array_length, |
|
84 RegionNode* region); |
|
85 Node* generate_current_thread(Node* &tls_output); |
|
86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, |
|
87 bool disjoint_bases, const char* &name); |
|
88 Node* load_mirror_from_klass(Node* klass); |
|
89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, |
|
90 int nargs, |
|
91 RegionNode* region, int null_path, |
|
92 int offset); |
|
93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, |
|
94 RegionNode* region, int null_path) { |
|
95 int offset = java_lang_Class::klass_offset_in_bytes(); |
|
96 return load_klass_from_mirror_common(mirror, never_see_null, nargs, |
|
97 region, null_path, |
|
98 offset); |
|
99 } |
|
100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, |
|
101 int nargs, |
|
102 RegionNode* region, int null_path) { |
|
103 int offset = java_lang_Class::array_klass_offset_in_bytes(); |
|
104 return load_klass_from_mirror_common(mirror, never_see_null, nargs, |
|
105 region, null_path, |
|
106 offset); |
|
107 } |
|
108 Node* generate_access_flags_guard(Node* kls, |
|
109 int modifier_mask, int modifier_bits, |
|
110 RegionNode* region); |
|
111 Node* generate_interface_guard(Node* kls, RegionNode* region); |
|
112 Node* generate_array_guard(Node* kls, RegionNode* region) { |
|
113 return generate_array_guard_common(kls, region, false, false); |
|
114 } |
|
115 Node* generate_non_array_guard(Node* kls, RegionNode* region) { |
|
116 return generate_array_guard_common(kls, region, false, true); |
|
117 } |
|
118 Node* generate_objArray_guard(Node* kls, RegionNode* region) { |
|
119 return generate_array_guard_common(kls, region, true, false); |
|
120 } |
|
121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { |
|
122 return generate_array_guard_common(kls, region, true, true); |
|
123 } |
|
124 Node* generate_array_guard_common(Node* kls, RegionNode* region, |
|
125 bool obj_array, bool not_array); |
|
126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); |
|
127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, |
|
128 bool is_virtual = false, bool is_static = false); |
|
129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { |
|
130 return generate_method_call(method_id, false, true); |
|
131 } |
|
132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { |
|
133 return generate_method_call(method_id, true, false); |
|
134 } |
|
135 |
|
136 bool inline_string_compareTo(); |
|
137 bool inline_string_indexOf(); |
|
138 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); |
|
139 Node* pop_math_arg(); |
|
140 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); |
|
141 bool inline_math_native(vmIntrinsics::ID id); |
|
142 bool inline_trig(vmIntrinsics::ID id); |
|
143 bool inline_trans(vmIntrinsics::ID id); |
|
144 bool inline_abs(vmIntrinsics::ID id); |
|
145 bool inline_sqrt(vmIntrinsics::ID id); |
|
146 bool inline_pow(vmIntrinsics::ID id); |
|
147 bool inline_exp(vmIntrinsics::ID id); |
|
148 bool inline_min_max(vmIntrinsics::ID id); |
|
149 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); |
|
150 // This returns Type::AnyPtr, RawPtr, or OopPtr. |
|
151 int classify_unsafe_addr(Node* &base, Node* &offset); |
|
152 Node* make_unsafe_address(Node* base, Node* offset); |
|
153 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); |
|
154 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); |
|
155 bool inline_unsafe_allocate(); |
|
156 bool inline_unsafe_copyMemory(); |
|
157 bool inline_native_currentThread(); |
|
158 bool inline_native_time_funcs(bool isNano); |
|
159 bool inline_native_isInterrupted(); |
|
160 bool inline_native_Class_query(vmIntrinsics::ID id); |
|
161 bool inline_native_subtype_check(); |
|
162 |
|
163 bool inline_native_newArray(); |
|
164 bool inline_native_getLength(); |
|
165 bool inline_array_copyOf(bool is_copyOfRange); |
|
166 bool inline_native_clone(bool is_virtual); |
|
167 bool inline_native_Reflection_getCallerClass(); |
|
168 bool inline_native_AtomicLong_get(); |
|
169 bool inline_native_AtomicLong_attemptUpdate(); |
|
170 bool is_method_invoke_or_aux_frame(JVMState* jvms); |
|
171 // Helper function for inlining native object hash method |
|
172 bool inline_native_hashcode(bool is_virtual, bool is_static); |
|
173 bool inline_native_getClass(); |
|
174 |
|
175 // Helper functions for inlining arraycopy |
|
176 bool inline_arraycopy(); |
|
177 void generate_arraycopy(const TypePtr* adr_type, |
|
178 BasicType basic_elem_type, |
|
179 Node* src, Node* src_offset, |
|
180 Node* dest, Node* dest_offset, |
|
181 Node* copy_length, |
|
182 int nargs, // arguments on stack for debug info |
|
183 bool disjoint_bases = false, |
|
184 bool length_never_negative = false, |
|
185 RegionNode* slow_region = NULL); |
|
186 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, |
|
187 RegionNode* slow_region); |
|
188 void generate_clear_array(const TypePtr* adr_type, |
|
189 Node* dest, |
|
190 BasicType basic_elem_type, |
|
191 Node* slice_off, |
|
192 Node* slice_len, |
|
193 Node* slice_end); |
|
194 bool generate_block_arraycopy(const TypePtr* adr_type, |
|
195 BasicType basic_elem_type, |
|
196 AllocateNode* alloc, |
|
197 Node* src, Node* src_offset, |
|
198 Node* dest, Node* dest_offset, |
|
199 Node* dest_size); |
|
200 void generate_slow_arraycopy(const TypePtr* adr_type, |
|
201 Node* src, Node* src_offset, |
|
202 Node* dest, Node* dest_offset, |
|
203 Node* copy_length, |
|
204 int nargs); |
|
205 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, |
|
206 Node* dest_elem_klass, |
|
207 Node* src, Node* src_offset, |
|
208 Node* dest, Node* dest_offset, |
|
209 Node* copy_length, int nargs); |
|
210 Node* generate_generic_arraycopy(const TypePtr* adr_type, |
|
211 Node* src, Node* src_offset, |
|
212 Node* dest, Node* dest_offset, |
|
213 Node* copy_length, int nargs); |
|
214 void generate_unchecked_arraycopy(const TypePtr* adr_type, |
|
215 BasicType basic_elem_type, |
|
216 bool disjoint_bases, |
|
217 Node* src, Node* src_offset, |
|
218 Node* dest, Node* dest_offset, |
|
219 Node* copy_length); |
|
220 bool inline_unsafe_CAS(BasicType type); |
|
221 bool inline_unsafe_ordered_store(BasicType type); |
|
222 bool inline_fp_conversions(vmIntrinsics::ID id); |
|
223 bool inline_reverseBytes(vmIntrinsics::ID id); |
|
224 }; |
|
225 |
|
226 |
|
227 //---------------------------make_vm_intrinsic---------------------------- |
|
228 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { |
|
229 vmIntrinsics::ID id = m->intrinsic_id(); |
|
230 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); |
|
231 |
|
232 if (DisableIntrinsic[0] != '\0' |
|
233 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { |
|
234 // disabled by a user request on the command line: |
|
235 // example: -XX:DisableIntrinsic=_hashCode,_getClass |
|
236 return NULL; |
|
237 } |
|
238 |
|
239 if (!m->is_loaded()) { |
|
240 // do not attempt to inline unloaded methods |
|
241 return NULL; |
|
242 } |
|
243 |
|
244 // Only a few intrinsics implement a virtual dispatch. |
|
245 // They are expensive calls which are also frequently overridden. |
|
246 if (is_virtual) { |
|
247 switch (id) { |
|
248 case vmIntrinsics::_hashCode: |
|
249 case vmIntrinsics::_clone: |
|
250 // OK, Object.hashCode and Object.clone intrinsics come in both flavors |
|
251 break; |
|
252 default: |
|
253 return NULL; |
|
254 } |
|
255 } |
|
256 |
|
257 // -XX:-InlineNatives disables nearly all intrinsics: |
|
258 if (!InlineNatives) { |
|
259 switch (id) { |
|
260 case vmIntrinsics::_indexOf: |
|
261 case vmIntrinsics::_compareTo: |
|
262 break; // InlineNatives does not control String.compareTo |
|
263 default: |
|
264 return NULL; |
|
265 } |
|
266 } |
|
267 |
|
268 switch (id) { |
|
269 case vmIntrinsics::_compareTo: |
|
270 if (!SpecialStringCompareTo) return NULL; |
|
271 break; |
|
272 case vmIntrinsics::_indexOf: |
|
273 if (!SpecialStringIndexOf) return NULL; |
|
274 break; |
|
275 case vmIntrinsics::_arraycopy: |
|
276 if (!InlineArrayCopy) return NULL; |
|
277 break; |
|
278 case vmIntrinsics::_copyMemory: |
|
279 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; |
|
280 if (!InlineArrayCopy) return NULL; |
|
281 break; |
|
282 case vmIntrinsics::_hashCode: |
|
283 if (!InlineObjectHash) return NULL; |
|
284 break; |
|
285 case vmIntrinsics::_clone: |
|
286 case vmIntrinsics::_copyOf: |
|
287 case vmIntrinsics::_copyOfRange: |
|
288 if (!InlineObjectCopy) return NULL; |
|
289 // These also use the arraycopy intrinsic mechanism: |
|
290 if (!InlineArrayCopy) return NULL; |
|
291 break; |
|
292 case vmIntrinsics::_checkIndex: |
|
293 // We do not intrinsify this. The optimizer does fine with it. |
|
294 return NULL; |
|
295 |
|
296 case vmIntrinsics::_get_AtomicLong: |
|
297 case vmIntrinsics::_attemptUpdate: |
|
298 if (!InlineAtomicLong) return NULL; |
|
299 break; |
|
300 |
|
301 case vmIntrinsics::_Object_init: |
|
302 case vmIntrinsics::_invoke: |
|
303 // We do not intrinsify these; they are marked for other purposes. |
|
304 return NULL; |
|
305 |
|
306 case vmIntrinsics::_getCallerClass: |
|
307 if (!UseNewReflection) return NULL; |
|
308 if (!InlineReflectionGetCallerClass) return NULL; |
|
309 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; |
|
310 break; |
|
311 |
|
312 default: |
|
313 break; |
|
314 } |
|
315 |
|
316 // -XX:-InlineClassNatives disables natives from the Class class. |
|
317 // The flag applies to all reflective calls, notably Array.newArray |
|
318 // (visible to Java programmers as Array.newInstance). |
|
319 if (m->holder()->name() == ciSymbol::java_lang_Class() || |
|
320 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { |
|
321 if (!InlineClassNatives) return NULL; |
|
322 } |
|
323 |
|
324 // -XX:-InlineThreadNatives disables natives from the Thread class. |
|
325 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { |
|
326 if (!InlineThreadNatives) return NULL; |
|
327 } |
|
328 |
|
329 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. |
|
330 if (m->holder()->name() == ciSymbol::java_lang_Math() || |
|
331 m->holder()->name() == ciSymbol::java_lang_Float() || |
|
332 m->holder()->name() == ciSymbol::java_lang_Double()) { |
|
333 if (!InlineMathNatives) return NULL; |
|
334 } |
|
335 |
|
336 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. |
|
337 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { |
|
338 if (!InlineUnsafeOps) return NULL; |
|
339 } |
|
340 |
|
341 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); |
|
342 } |
|
343 |
|
344 //----------------------register_library_intrinsics----------------------- |
|
345 // Initialize this file's data structures, for each Compile instance. |
|
346 void Compile::register_library_intrinsics() { |
|
347 // Nothing to do here. |
|
348 } |
|
349 |
|
350 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { |
|
351 LibraryCallKit kit(jvms, this); |
|
352 Compile* C = kit.C; |
|
353 int nodes = C->unique(); |
|
354 #ifndef PRODUCT |
|
355 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { |
|
356 char buf[1000]; |
|
357 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); |
|
358 tty->print_cr("Intrinsic %s", str); |
|
359 } |
|
360 #endif |
|
361 if (kit.try_to_inline()) { |
|
362 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { |
|
363 tty->print("Inlining intrinsic %s%s at bci:%d in", |
|
364 vmIntrinsics::name_at(intrinsic_id()), |
|
365 (is_virtual() ? " (virtual)" : ""), kit.bci()); |
|
366 kit.caller()->print_short_name(tty); |
|
367 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); |
|
368 } |
|
369 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); |
|
370 if (C->log()) { |
|
371 C->log()->elem("intrinsic id='%s'%s nodes='%d'", |
|
372 vmIntrinsics::name_at(intrinsic_id()), |
|
373 (is_virtual() ? " virtual='1'" : ""), |
|
374 C->unique() - nodes); |
|
375 } |
|
376 return kit.transfer_exceptions_into_jvms(); |
|
377 } |
|
378 |
|
379 if (PrintIntrinsics) { |
|
380 switch (intrinsic_id()) { |
|
381 case vmIntrinsics::_invoke: |
|
382 case vmIntrinsics::_Object_init: |
|
383 // We do not expect to inline these, so do not produce any noise about them. |
|
384 break; |
|
385 default: |
|
386 tty->print("Did not inline intrinsic %s%s at bci:%d in", |
|
387 vmIntrinsics::name_at(intrinsic_id()), |
|
388 (is_virtual() ? " (virtual)" : ""), kit.bci()); |
|
389 kit.caller()->print_short_name(tty); |
|
390 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); |
|
391 } |
|
392 } |
|
393 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); |
|
394 return NULL; |
|
395 } |
|
396 |
|
397 bool LibraryCallKit::try_to_inline() { |
|
398 // Handle symbolic names for otherwise undistinguished boolean switches: |
|
399 const bool is_store = true; |
|
400 const bool is_native_ptr = true; |
|
401 const bool is_static = true; |
|
402 |
|
403 switch (intrinsic_id()) { |
|
404 case vmIntrinsics::_hashCode: |
|
405 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); |
|
406 case vmIntrinsics::_identityHashCode: |
|
407 return inline_native_hashcode(/*!virtual*/ false, is_static); |
|
408 case vmIntrinsics::_getClass: |
|
409 return inline_native_getClass(); |
|
410 |
|
411 case vmIntrinsics::_dsin: |
|
412 case vmIntrinsics::_dcos: |
|
413 case vmIntrinsics::_dtan: |
|
414 case vmIntrinsics::_dabs: |
|
415 case vmIntrinsics::_datan2: |
|
416 case vmIntrinsics::_dsqrt: |
|
417 case vmIntrinsics::_dexp: |
|
418 case vmIntrinsics::_dlog: |
|
419 case vmIntrinsics::_dlog10: |
|
420 case vmIntrinsics::_dpow: |
|
421 return inline_math_native(intrinsic_id()); |
|
422 |
|
423 case vmIntrinsics::_min: |
|
424 case vmIntrinsics::_max: |
|
425 return inline_min_max(intrinsic_id()); |
|
426 |
|
427 case vmIntrinsics::_arraycopy: |
|
428 return inline_arraycopy(); |
|
429 |
|
430 case vmIntrinsics::_compareTo: |
|
431 return inline_string_compareTo(); |
|
432 case vmIntrinsics::_indexOf: |
|
433 return inline_string_indexOf(); |
|
434 |
|
435 case vmIntrinsics::_getObject: |
|
436 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); |
|
437 case vmIntrinsics::_getBoolean: |
|
438 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); |
|
439 case vmIntrinsics::_getByte: |
|
440 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); |
|
441 case vmIntrinsics::_getShort: |
|
442 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); |
|
443 case vmIntrinsics::_getChar: |
|
444 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); |
|
445 case vmIntrinsics::_getInt: |
|
446 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); |
|
447 case vmIntrinsics::_getLong: |
|
448 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); |
|
449 case vmIntrinsics::_getFloat: |
|
450 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); |
|
451 case vmIntrinsics::_getDouble: |
|
452 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); |
|
453 |
|
454 case vmIntrinsics::_putObject: |
|
455 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); |
|
456 case vmIntrinsics::_putBoolean: |
|
457 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); |
|
458 case vmIntrinsics::_putByte: |
|
459 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); |
|
460 case vmIntrinsics::_putShort: |
|
461 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); |
|
462 case vmIntrinsics::_putChar: |
|
463 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); |
|
464 case vmIntrinsics::_putInt: |
|
465 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); |
|
466 case vmIntrinsics::_putLong: |
|
467 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); |
|
468 case vmIntrinsics::_putFloat: |
|
469 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); |
|
470 case vmIntrinsics::_putDouble: |
|
471 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); |
|
472 |
|
473 case vmIntrinsics::_getByte_raw: |
|
474 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); |
|
475 case vmIntrinsics::_getShort_raw: |
|
476 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); |
|
477 case vmIntrinsics::_getChar_raw: |
|
478 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); |
|
479 case vmIntrinsics::_getInt_raw: |
|
480 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); |
|
481 case vmIntrinsics::_getLong_raw: |
|
482 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); |
|
483 case vmIntrinsics::_getFloat_raw: |
|
484 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); |
|
485 case vmIntrinsics::_getDouble_raw: |
|
486 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); |
|
487 case vmIntrinsics::_getAddress_raw: |
|
488 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); |
|
489 |
|
490 case vmIntrinsics::_putByte_raw: |
|
491 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); |
|
492 case vmIntrinsics::_putShort_raw: |
|
493 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); |
|
494 case vmIntrinsics::_putChar_raw: |
|
495 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); |
|
496 case vmIntrinsics::_putInt_raw: |
|
497 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); |
|
498 case vmIntrinsics::_putLong_raw: |
|
499 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); |
|
500 case vmIntrinsics::_putFloat_raw: |
|
501 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); |
|
502 case vmIntrinsics::_putDouble_raw: |
|
503 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); |
|
504 case vmIntrinsics::_putAddress_raw: |
|
505 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); |
|
506 |
|
507 case vmIntrinsics::_getObjectVolatile: |
|
508 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); |
|
509 case vmIntrinsics::_getBooleanVolatile: |
|
510 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); |
|
511 case vmIntrinsics::_getByteVolatile: |
|
512 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); |
|
513 case vmIntrinsics::_getShortVolatile: |
|
514 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); |
|
515 case vmIntrinsics::_getCharVolatile: |
|
516 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); |
|
517 case vmIntrinsics::_getIntVolatile: |
|
518 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); |
|
519 case vmIntrinsics::_getLongVolatile: |
|
520 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); |
|
521 case vmIntrinsics::_getFloatVolatile: |
|
522 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); |
|
523 case vmIntrinsics::_getDoubleVolatile: |
|
524 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); |
|
525 |
|
526 case vmIntrinsics::_putObjectVolatile: |
|
527 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); |
|
528 case vmIntrinsics::_putBooleanVolatile: |
|
529 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); |
|
530 case vmIntrinsics::_putByteVolatile: |
|
531 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); |
|
532 case vmIntrinsics::_putShortVolatile: |
|
533 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); |
|
534 case vmIntrinsics::_putCharVolatile: |
|
535 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); |
|
536 case vmIntrinsics::_putIntVolatile: |
|
537 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); |
|
538 case vmIntrinsics::_putLongVolatile: |
|
539 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); |
|
540 case vmIntrinsics::_putFloatVolatile: |
|
541 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); |
|
542 case vmIntrinsics::_putDoubleVolatile: |
|
543 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); |
|
544 |
|
545 case vmIntrinsics::_prefetchRead: |
|
546 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); |
|
547 case vmIntrinsics::_prefetchWrite: |
|
548 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); |
|
549 case vmIntrinsics::_prefetchReadStatic: |
|
550 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); |
|
551 case vmIntrinsics::_prefetchWriteStatic: |
|
552 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); |
|
553 |
|
554 case vmIntrinsics::_compareAndSwapObject: |
|
555 return inline_unsafe_CAS(T_OBJECT); |
|
556 case vmIntrinsics::_compareAndSwapInt: |
|
557 return inline_unsafe_CAS(T_INT); |
|
558 case vmIntrinsics::_compareAndSwapLong: |
|
559 return inline_unsafe_CAS(T_LONG); |
|
560 |
|
561 case vmIntrinsics::_putOrderedObject: |
|
562 return inline_unsafe_ordered_store(T_OBJECT); |
|
563 case vmIntrinsics::_putOrderedInt: |
|
564 return inline_unsafe_ordered_store(T_INT); |
|
565 case vmIntrinsics::_putOrderedLong: |
|
566 return inline_unsafe_ordered_store(T_LONG); |
|
567 |
|
568 case vmIntrinsics::_currentThread: |
|
569 return inline_native_currentThread(); |
|
570 case vmIntrinsics::_isInterrupted: |
|
571 return inline_native_isInterrupted(); |
|
572 |
|
573 case vmIntrinsics::_currentTimeMillis: |
|
574 return inline_native_time_funcs(false); |
|
575 case vmIntrinsics::_nanoTime: |
|
576 return inline_native_time_funcs(true); |
|
577 case vmIntrinsics::_allocateInstance: |
|
578 return inline_unsafe_allocate(); |
|
579 case vmIntrinsics::_copyMemory: |
|
580 return inline_unsafe_copyMemory(); |
|
581 case vmIntrinsics::_newArray: |
|
582 return inline_native_newArray(); |
|
583 case vmIntrinsics::_getLength: |
|
584 return inline_native_getLength(); |
|
585 case vmIntrinsics::_copyOf: |
|
586 return inline_array_copyOf(false); |
|
587 case vmIntrinsics::_copyOfRange: |
|
588 return inline_array_copyOf(true); |
|
589 case vmIntrinsics::_clone: |
|
590 return inline_native_clone(intrinsic()->is_virtual()); |
|
591 |
|
592 case vmIntrinsics::_isAssignableFrom: |
|
593 return inline_native_subtype_check(); |
|
594 |
|
595 case vmIntrinsics::_isInstance: |
|
596 case vmIntrinsics::_getModifiers: |
|
597 case vmIntrinsics::_isInterface: |
|
598 case vmIntrinsics::_isArray: |
|
599 case vmIntrinsics::_isPrimitive: |
|
600 case vmIntrinsics::_getSuperclass: |
|
601 case vmIntrinsics::_getComponentType: |
|
602 case vmIntrinsics::_getClassAccessFlags: |
|
603 return inline_native_Class_query(intrinsic_id()); |
|
604 |
|
605 case vmIntrinsics::_floatToRawIntBits: |
|
606 case vmIntrinsics::_floatToIntBits: |
|
607 case vmIntrinsics::_intBitsToFloat: |
|
608 case vmIntrinsics::_doubleToRawLongBits: |
|
609 case vmIntrinsics::_doubleToLongBits: |
|
610 case vmIntrinsics::_longBitsToDouble: |
|
611 return inline_fp_conversions(intrinsic_id()); |
|
612 |
|
613 case vmIntrinsics::_reverseBytes_i: |
|
614 case vmIntrinsics::_reverseBytes_l: |
|
615 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); |
|
616 |
|
617 case vmIntrinsics::_get_AtomicLong: |
|
618 return inline_native_AtomicLong_get(); |
|
619 case vmIntrinsics::_attemptUpdate: |
|
620 return inline_native_AtomicLong_attemptUpdate(); |
|
621 |
|
622 case vmIntrinsics::_getCallerClass: |
|
623 return inline_native_Reflection_getCallerClass(); |
|
624 |
|
625 default: |
|
626 // If you get here, it may be that someone has added a new intrinsic |
|
627 // to the list in vmSymbols.hpp without implementing it here. |
|
628 #ifndef PRODUCT |
|
629 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { |
|
630 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", |
|
631 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); |
|
632 } |
|
633 #endif |
|
634 return false; |
|
635 } |
|
636 } |
|
637 |
|
638 //------------------------------push_result------------------------------ |
|
639 // Helper function for finishing intrinsics. |
|
640 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { |
|
641 record_for_igvn(region); |
|
642 set_control(_gvn.transform(region)); |
|
643 BasicType value_type = value->type()->basic_type(); |
|
644 push_node(value_type, _gvn.transform(value)); |
|
645 } |
|
646 |
|
647 //------------------------------generate_guard--------------------------- |
|
648 // Helper function for generating guarded fast-slow graph structures. |
|
649 // The given 'test', if true, guards a slow path. If the test fails |
|
650 // then a fast path can be taken. (We generally hope it fails.) |
|
651 // In all cases, GraphKit::control() is updated to the fast path. |
|
652 // The returned value represents the control for the slow path. |
|
653 // The return value is never 'top'; it is either a valid control |
|
654 // or NULL if it is obvious that the slow path can never be taken. |
|
655 // Also, if region and the slow control are not NULL, the slow edge |
|
656 // is appended to the region. |
|
657 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { |
|
658 if (stopped()) { |
|
659 // Already short circuited. |
|
660 return NULL; |
|
661 } |
|
662 |
|
663 // Build an if node and its projections. |
|
664 // If test is true we take the slow path, which we assume is uncommon. |
|
665 if (_gvn.type(test) == TypeInt::ZERO) { |
|
666 // The slow branch is never taken. No need to build this guard. |
|
667 return NULL; |
|
668 } |
|
669 |
|
670 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); |
|
671 |
|
672 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); |
|
673 if (if_slow == top()) { |
|
674 // The slow branch is never taken. No need to build this guard. |
|
675 return NULL; |
|
676 } |
|
677 |
|
678 if (region != NULL) |
|
679 region->add_req(if_slow); |
|
680 |
|
681 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); |
|
682 set_control(if_fast); |
|
683 |
|
684 return if_slow; |
|
685 } |
|
686 |
|
687 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { |
|
688 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); |
|
689 } |
|
690 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { |
|
691 return generate_guard(test, region, PROB_FAIR); |
|
692 } |
|
693 |
|
694 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, |
|
695 Node* *pos_index) { |
|
696 if (stopped()) |
|
697 return NULL; // already stopped |
|
698 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] |
|
699 return NULL; // index is already adequately typed |
|
700 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); |
|
701 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); |
|
702 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); |
|
703 if (is_neg != NULL && pos_index != NULL) { |
|
704 // Emulate effect of Parse::adjust_map_after_if. |
|
705 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); |
|
706 ccast->set_req(0, control()); |
|
707 (*pos_index) = _gvn.transform(ccast); |
|
708 } |
|
709 return is_neg; |
|
710 } |
|
711 |
|
712 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, |
|
713 Node* *pos_index) { |
|
714 if (stopped()) |
|
715 return NULL; // already stopped |
|
716 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] |
|
717 return NULL; // index is already adequately typed |
|
718 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); |
|
719 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); |
|
720 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); |
|
721 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); |
|
722 if (is_notp != NULL && pos_index != NULL) { |
|
723 // Emulate effect of Parse::adjust_map_after_if. |
|
724 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); |
|
725 ccast->set_req(0, control()); |
|
726 (*pos_index) = _gvn.transform(ccast); |
|
727 } |
|
728 return is_notp; |
|
729 } |
|
730 |
|
731 // Make sure that 'position' is a valid limit index, in [0..length]. |
|
732 // There are two equivalent plans for checking this: |
|
733 // A. (offset + copyLength) unsigned<= arrayLength |
|
734 // B. offset <= (arrayLength - copyLength) |
|
735 // We require that all of the values above, except for the sum and |
|
736 // difference, are already known to be non-negative. |
|
737 // Plan A is robust in the face of overflow, if offset and copyLength |
|
738 // are both hugely positive. |
|
739 // |
|
740 // Plan B is less direct and intuitive, but it does not overflow at |
|
741 // all, since the difference of two non-negatives is always |
|
742 // representable. Whenever Java methods must perform the equivalent |
|
743 // check they generally use Plan B instead of Plan A. |
|
744 // For the moment we use Plan A. |
|
745 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, |
|
746 Node* subseq_length, |
|
747 Node* array_length, |
|
748 RegionNode* region) { |
|
749 if (stopped()) |
|
750 return NULL; // already stopped |
|
751 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; |
|
752 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length)) |
|
753 return NULL; // common case of whole-array copy |
|
754 Node* last = subseq_length; |
|
755 if (!zero_offset) // last += offset |
|
756 last = _gvn.transform( new (C, 3) AddINode(last, offset)); |
|
757 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); |
|
758 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); |
|
759 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); |
|
760 return is_over; |
|
761 } |
|
762 |
|
763 |
|
764 //--------------------------generate_current_thread-------------------- |
|
765 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { |
|
766 ciKlass* thread_klass = env()->Thread_klass(); |
|
767 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); |
|
768 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); |
|
769 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); |
|
770 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); |
|
771 tls_output = thread; |
|
772 return threadObj; |
|
773 } |
|
774 |
|
775 |
|
776 //------------------------------inline_string_compareTo------------------------ |
|
777 bool LibraryCallKit::inline_string_compareTo() { |
|
778 |
|
779 const int value_offset = java_lang_String::value_offset_in_bytes(); |
|
780 const int count_offset = java_lang_String::count_offset_in_bytes(); |
|
781 const int offset_offset = java_lang_String::offset_offset_in_bytes(); |
|
782 |
|
783 _sp += 2; |
|
784 Node *argument = pop(); // pop non-receiver first: it was pushed second |
|
785 Node *receiver = pop(); |
|
786 |
|
787 // Null check on self without removing any arguments. The argument |
|
788 // null check technically happens in the wrong place, which can lead to |
|
789 // invalid stack traces when string compare is inlined into a method |
|
790 // which handles NullPointerExceptions. |
|
791 _sp += 2; |
|
792 receiver = do_null_check(receiver, T_OBJECT); |
|
793 argument = do_null_check(argument, T_OBJECT); |
|
794 _sp -= 2; |
|
795 if (stopped()) { |
|
796 return true; |
|
797 } |
|
798 |
|
799 ciInstanceKlass* klass = env()->String_klass(); |
|
800 const TypeInstPtr* string_type = |
|
801 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); |
|
802 |
|
803 Node* compare = |
|
804 _gvn.transform(new (C, 7) StrCompNode( |
|
805 control(), |
|
806 memory(TypeAryPtr::CHARS), |
|
807 memory(string_type->add_offset(value_offset)), |
|
808 memory(string_type->add_offset(count_offset)), |
|
809 memory(string_type->add_offset(offset_offset)), |
|
810 receiver, |
|
811 argument)); |
|
812 push(compare); |
|
813 return true; |
|
814 } |
|
815 |
|
816 // Java version of String.indexOf(constant string) |
|
817 // class StringDecl { |
|
818 // StringDecl(char[] ca) { |
|
819 // offset = 0; |
|
820 // count = ca.length; |
|
821 // value = ca; |
|
822 // } |
|
823 // int offset; |
|
824 // int count; |
|
825 // char[] value; |
|
826 // } |
|
827 // |
|
828 // static int string_indexOf_J(StringDecl string_object, char[] target_object, |
|
829 // int targetOffset, int cache_i, int md2) { |
|
830 // int cache = cache_i; |
|
831 // int sourceOffset = string_object.offset; |
|
832 // int sourceCount = string_object.count; |
|
833 // int targetCount = target_object.length; |
|
834 // |
|
835 // int targetCountLess1 = targetCount - 1; |
|
836 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; |
|
837 // |
|
838 // char[] source = string_object.value; |
|
839 // char[] target = target_object; |
|
840 // int lastChar = target[targetCountLess1]; |
|
841 // |
|
842 // outer_loop: |
|
843 // for (int i = sourceOffset; i < sourceEnd; ) { |
|
844 // int src = source[i + targetCountLess1]; |
|
845 // if (src == lastChar) { |
|
846 // // With random strings and a 4-character alphabet, |
|
847 // // reverse matching at this point sets up 0.8% fewer |
|
848 // // frames, but (paradoxically) makes 0.3% more probes. |
|
849 // // Since those probes are nearer the lastChar probe, |
|
850 // // there is may be a net D$ win with reverse matching. |
|
851 // // But, reversing loop inhibits unroll of inner loop |
|
852 // // for unknown reason. So, does running outer loop from |
|
853 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) |
|
854 // for (int j = 0; j < targetCountLess1; j++) { |
|
855 // if (target[targetOffset + j] != source[i+j]) { |
|
856 // if ((cache & (1 << source[i+j])) == 0) { |
|
857 // if (md2 < j+1) { |
|
858 // i += j+1; |
|
859 // continue outer_loop; |
|
860 // } |
|
861 // } |
|
862 // i += md2; |
|
863 // continue outer_loop; |
|
864 // } |
|
865 // } |
|
866 // return i - sourceOffset; |
|
867 // } |
|
868 // if ((cache & (1 << src)) == 0) { |
|
869 // i += targetCountLess1; |
|
870 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. |
|
871 // i++; |
|
872 // } |
|
873 // return -1; |
|
874 // } |
|
875 |
|
876 //------------------------------string_indexOf------------------------ |
|
877 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, |
|
878 jint cache_i, jint md2_i) { |
|
879 |
|
880 Node* no_ctrl = NULL; |
|
881 float likely = PROB_LIKELY(0.9); |
|
882 float unlikely = PROB_UNLIKELY(0.9); |
|
883 |
|
884 const int value_offset = java_lang_String::value_offset_in_bytes(); |
|
885 const int count_offset = java_lang_String::count_offset_in_bytes(); |
|
886 const int offset_offset = java_lang_String::offset_offset_in_bytes(); |
|
887 |
|
888 ciInstanceKlass* klass = env()->String_klass(); |
|
889 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); |
|
890 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); |
|
891 |
|
892 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); |
|
893 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); |
|
894 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); |
|
895 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); |
|
896 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); |
|
897 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); |
|
898 |
|
899 Node* target = _gvn.transform(ConPNode::make(C, target_array)); |
|
900 jint target_length = target_array->length(); |
|
901 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); |
|
902 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); |
|
903 |
|
904 IdealKit kit(gvn(), control(), merged_memory()); |
|
905 #define __ kit. |
|
906 Node* zero = __ ConI(0); |
|
907 Node* one = __ ConI(1); |
|
908 Node* cache = __ ConI(cache_i); |
|
909 Node* md2 = __ ConI(md2_i); |
|
910 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); |
|
911 Node* targetCount = __ ConI(target_length); |
|
912 Node* targetCountLess1 = __ ConI(target_length - 1); |
|
913 Node* targetOffset = __ ConI(targetOffset_i); |
|
914 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); |
|
915 |
|
916 IdealVariable rtn(kit), i(kit), j(kit); __ declares_done(); |
|
917 Node* outer_loop = __ make_label(2 /* goto */); |
|
918 Node* return_ = __ make_label(1); |
|
919 |
|
920 __ set(rtn,__ ConI(-1)); |
|
921 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); { |
|
922 Node* i2 = __ AddI(__ value(i), targetCountLess1); |
|
923 // pin to prohibit loading of "next iteration" value which may SEGV (rare) |
|
924 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); |
|
925 __ if_then(src, BoolTest::eq, lastChar, unlikely); { |
|
926 __ loop(j, zero, BoolTest::lt, targetCountLess1); { |
|
927 Node* tpj = __ AddI(targetOffset, __ value(j)); |
|
928 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); |
|
929 Node* ipj = __ AddI(__ value(i), __ value(j)); |
|
930 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); |
|
931 __ if_then(targ, BoolTest::ne, src2); { |
|
932 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { |
|
933 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { |
|
934 __ increment(i, __ AddI(__ value(j), one)); |
|
935 __ goto_(outer_loop); |
|
936 } __ end_if(); __ dead(j); |
|
937 }__ end_if(); __ dead(j); |
|
938 __ increment(i, md2); |
|
939 __ goto_(outer_loop); |
|
940 }__ end_if(); |
|
941 __ increment(j, one); |
|
942 }__ end_loop(); __ dead(j); |
|
943 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); |
|
944 __ goto_(return_); |
|
945 }__ end_if(); |
|
946 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { |
|
947 __ increment(i, targetCountLess1); |
|
948 }__ end_if(); |
|
949 __ increment(i, one); |
|
950 __ bind(outer_loop); |
|
951 }__ end_loop(); __ dead(i); |
|
952 __ bind(return_); |
|
953 __ drain_delay_transform(); |
|
954 |
|
955 set_control(__ ctrl()); |
|
956 Node* result = __ value(rtn); |
|
957 #undef __ |
|
958 C->set_has_loops(true); |
|
959 return result; |
|
960 } |
|
961 |
|
962 |
|
963 //------------------------------inline_string_indexOf------------------------ |
|
964 bool LibraryCallKit::inline_string_indexOf() { |
|
965 |
|
966 _sp += 2; |
|
967 Node *argument = pop(); // pop non-receiver first: it was pushed second |
|
968 Node *receiver = pop(); |
|
969 |
|
970 // don't intrinsify is argument isn't a constant string. |
|
971 if (!argument->is_Con()) { |
|
972 return false; |
|
973 } |
|
974 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); |
|
975 if (str_type == NULL) { |
|
976 return false; |
|
977 } |
|
978 ciInstanceKlass* klass = env()->String_klass(); |
|
979 ciObject* str_const = str_type->const_oop(); |
|
980 if (str_const == NULL || str_const->klass() != klass) { |
|
981 return false; |
|
982 } |
|
983 ciInstance* str = str_const->as_instance(); |
|
984 assert(str != NULL, "must be instance"); |
|
985 |
|
986 const int value_offset = java_lang_String::value_offset_in_bytes(); |
|
987 const int count_offset = java_lang_String::count_offset_in_bytes(); |
|
988 const int offset_offset = java_lang_String::offset_offset_in_bytes(); |
|
989 |
|
990 ciObject* v = str->field_value_by_offset(value_offset).as_object(); |
|
991 int o = str->field_value_by_offset(offset_offset).as_int(); |
|
992 int c = str->field_value_by_offset(count_offset).as_int(); |
|
993 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array |
|
994 |
|
995 // constant strings have no offset and count == length which |
|
996 // simplifies the resulting code somewhat so lets optimize for that. |
|
997 if (o != 0 || c != pat->length()) { |
|
998 return false; |
|
999 } |
|
1000 |
|
1001 // Null check on self without removing any arguments. The argument |
|
1002 // null check technically happens in the wrong place, which can lead to |
|
1003 // invalid stack traces when string compare is inlined into a method |
|
1004 // which handles NullPointerExceptions. |
|
1005 _sp += 2; |
|
1006 receiver = do_null_check(receiver, T_OBJECT); |
|
1007 // No null check on the argument is needed since it's a constant String oop. |
|
1008 _sp -= 2; |
|
1009 if (stopped()) { |
|
1010 return true; |
|
1011 } |
|
1012 |
|
1013 // The null string as a pattern always returns 0 (match at beginning of string) |
|
1014 if (c == 0) { |
|
1015 push(intcon(0)); |
|
1016 return true; |
|
1017 } |
|
1018 |
|
1019 jchar lastChar = pat->char_at(o + (c - 1)); |
|
1020 int cache = 0; |
|
1021 int i; |
|
1022 for (i = 0; i < c - 1; i++) { |
|
1023 assert(i < pat->length(), "out of range"); |
|
1024 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); |
|
1025 } |
|
1026 |
|
1027 int md2 = c; |
|
1028 for (i = 0; i < c - 1; i++) { |
|
1029 assert(i < pat->length(), "out of range"); |
|
1030 if (pat->char_at(o + i) == lastChar) { |
|
1031 md2 = (c - 1) - i; |
|
1032 } |
|
1033 } |
|
1034 |
|
1035 Node* result = string_indexOf(receiver, pat, o, cache, md2); |
|
1036 push(result); |
|
1037 return true; |
|
1038 } |
|
1039 |
|
1040 //--------------------------pop_math_arg-------------------------------- |
|
1041 // Pop a double argument to a math function from the stack |
|
1042 // rounding it if necessary. |
|
1043 Node * LibraryCallKit::pop_math_arg() { |
|
1044 Node *arg = pop_pair(); |
|
1045 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) |
|
1046 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); |
|
1047 return arg; |
|
1048 } |
|
1049 |
|
1050 //------------------------------inline_trig---------------------------------- |
|
1051 // Inline sin/cos/tan instructions, if possible. If rounding is required, do |
|
1052 // argument reduction which will turn into a fast/slow diamond. |
|
1053 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { |
|
1054 _sp += arg_size(); // restore stack pointer |
|
1055 Node* arg = pop_math_arg(); |
|
1056 Node* trig = NULL; |
|
1057 |
|
1058 switch (id) { |
|
1059 case vmIntrinsics::_dsin: |
|
1060 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); |
|
1061 break; |
|
1062 case vmIntrinsics::_dcos: |
|
1063 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); |
|
1064 break; |
|
1065 case vmIntrinsics::_dtan: |
|
1066 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); |
|
1067 break; |
|
1068 default: |
|
1069 assert(false, "bad intrinsic was passed in"); |
|
1070 return false; |
|
1071 } |
|
1072 |
|
1073 // Rounding required? Check for argument reduction! |
|
1074 if( Matcher::strict_fp_requires_explicit_rounding ) { |
|
1075 |
|
1076 static const double pi_4 = 0.7853981633974483; |
|
1077 static const double neg_pi_4 = -0.7853981633974483; |
|
1078 // pi/2 in 80-bit extended precision |
|
1079 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; |
|
1080 // -pi/2 in 80-bit extended precision |
|
1081 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00}; |
|
1082 // Cutoff value for using this argument reduction technique |
|
1083 //static const double pi_2_minus_epsilon = 1.564660403643354; |
|
1084 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; |
|
1085 |
|
1086 // Pseudocode for sin: |
|
1087 // if (x <= Math.PI / 4.0) { |
|
1088 // if (x >= -Math.PI / 4.0) return fsin(x); |
|
1089 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); |
|
1090 // } else { |
|
1091 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); |
|
1092 // } |
|
1093 // return StrictMath.sin(x); |
|
1094 |
|
1095 // Pseudocode for cos: |
|
1096 // if (x <= Math.PI / 4.0) { |
|
1097 // if (x >= -Math.PI / 4.0) return fcos(x); |
|
1098 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); |
|
1099 // } else { |
|
1100 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); |
|
1101 // } |
|
1102 // return StrictMath.cos(x); |
|
1103 |
|
1104 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it |
|
1105 // requires a special machine instruction to load it. Instead we'll try |
|
1106 // the 'easy' case. If we really need the extra range +/- PI/2 we'll |
|
1107 // probably do the math inside the SIN encoding. |
|
1108 |
|
1109 // Make the merge point |
|
1110 RegionNode *r = new (C, 3) RegionNode(3); |
|
1111 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); |
|
1112 |
|
1113 // Flatten arg so we need only 1 test |
|
1114 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); |
|
1115 // Node for PI/4 constant |
|
1116 Node *pi4 = makecon(TypeD::make(pi_4)); |
|
1117 // Check PI/4 : abs(arg) |
|
1118 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); |
|
1119 // Check: If PI/4 < abs(arg) then go slow |
|
1120 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); |
|
1121 // Branch either way |
|
1122 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
|
1123 set_control(opt_iff(r,iff)); |
|
1124 |
|
1125 // Set fast path result |
|
1126 phi->init_req(2,trig); |
|
1127 |
|
1128 // Slow path - non-blocking leaf call |
|
1129 Node* call = NULL; |
|
1130 switch (id) { |
|
1131 case vmIntrinsics::_dsin: |
|
1132 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), |
|
1133 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), |
|
1134 "Sin", NULL, arg, top()); |
|
1135 break; |
|
1136 case vmIntrinsics::_dcos: |
|
1137 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), |
|
1138 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), |
|
1139 "Cos", NULL, arg, top()); |
|
1140 break; |
|
1141 case vmIntrinsics::_dtan: |
|
1142 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), |
|
1143 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), |
|
1144 "Tan", NULL, arg, top()); |
|
1145 break; |
|
1146 } |
|
1147 assert(control()->in(0) == call, ""); |
|
1148 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); |
|
1149 r->init_req(1,control()); |
|
1150 phi->init_req(1,slow_result); |
|
1151 |
|
1152 // Post-merge |
|
1153 set_control(_gvn.transform(r)); |
|
1154 record_for_igvn(r); |
|
1155 trig = _gvn.transform(phi); |
|
1156 |
|
1157 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
1158 } |
|
1159 // Push result back on JVM stack |
|
1160 push_pair(trig); |
|
1161 return true; |
|
1162 } |
|
1163 |
|
1164 //------------------------------inline_sqrt------------------------------------- |
|
1165 // Inline square root instruction, if possible. |
|
1166 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { |
|
1167 assert(id == vmIntrinsics::_dsqrt, "Not square root"); |
|
1168 _sp += arg_size(); // restore stack pointer |
|
1169 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); |
|
1170 return true; |
|
1171 } |
|
1172 |
|
1173 //------------------------------inline_abs------------------------------------- |
|
1174 // Inline absolute value instruction, if possible. |
|
1175 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { |
|
1176 assert(id == vmIntrinsics::_dabs, "Not absolute value"); |
|
1177 _sp += arg_size(); // restore stack pointer |
|
1178 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); |
|
1179 return true; |
|
1180 } |
|
1181 |
|
1182 //------------------------------inline_exp------------------------------------- |
|
1183 // Inline exp instructions, if possible. The Intel hardware only misses |
|
1184 // really odd corner cases (+/- Infinity). Just uncommon-trap them. |
|
1185 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { |
|
1186 assert(id == vmIntrinsics::_dexp, "Not exp"); |
|
1187 |
|
1188 // If this inlining ever returned NaN in the past, we do not intrinsify it |
|
1189 // every again. NaN results requires StrictMath.exp handling. |
|
1190 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
1191 |
|
1192 // Do not intrinsify on older platforms which lack cmove. |
|
1193 if (ConditionalMoveLimit == 0) return false; |
|
1194 |
|
1195 _sp += arg_size(); // restore stack pointer |
|
1196 Node *x = pop_math_arg(); |
|
1197 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); |
|
1198 |
|
1199 //------------------- |
|
1200 //result=(result.isNaN())? StrictMath::exp():result; |
|
1201 // Check: If isNaN() by checking result!=result? then go to Strict Math |
|
1202 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); |
|
1203 // Build the boolean node |
|
1204 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); |
|
1205 |
|
1206 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); |
|
1207 // End the current control-flow path |
|
1208 push_pair(x); |
|
1209 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp |
|
1210 // to handle. Recompile without intrinsifying Math.exp |
|
1211 uncommon_trap(Deoptimization::Reason_intrinsic, |
|
1212 Deoptimization::Action_make_not_entrant); |
|
1213 } |
|
1214 |
|
1215 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
1216 |
|
1217 push_pair(result); |
|
1218 |
|
1219 return true; |
|
1220 } |
|
1221 |
|
1222 //------------------------------inline_pow------------------------------------- |
|
1223 // Inline power instructions, if possible. |
|
1224 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { |
|
1225 assert(id == vmIntrinsics::_dpow, "Not pow"); |
|
1226 |
|
1227 // If this inlining ever returned NaN in the past, we do not intrinsify it |
|
1228 // every again. NaN results requires StrictMath.pow handling. |
|
1229 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
1230 |
|
1231 // Do not intrinsify on older platforms which lack cmove. |
|
1232 if (ConditionalMoveLimit == 0) return false; |
|
1233 |
|
1234 // Pseudocode for pow |
|
1235 // if (x <= 0.0) { |
|
1236 // if ((double)((int)y)==y) { // if y is int |
|
1237 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) |
|
1238 // } else { |
|
1239 // result = NaN; |
|
1240 // } |
|
1241 // } else { |
|
1242 // result = DPow(x,y); |
|
1243 // } |
|
1244 // if (result != result)? { |
|
1245 // ucommon_trap(); |
|
1246 // } |
|
1247 // return result; |
|
1248 |
|
1249 _sp += arg_size(); // restore stack pointer |
|
1250 Node* y = pop_math_arg(); |
|
1251 Node* x = pop_math_arg(); |
|
1252 |
|
1253 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); |
|
1254 |
|
1255 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow |
|
1256 // inside of something) then skip the fancy tests and just check for |
|
1257 // NaN result. |
|
1258 Node *result = NULL; |
|
1259 if( jvms()->depth() >= 1 ) { |
|
1260 result = fast_result; |
|
1261 } else { |
|
1262 |
|
1263 // Set the merge point for If node with condition of (x <= 0.0) |
|
1264 // There are four possible paths to region node and phi node |
|
1265 RegionNode *r = new (C, 4) RegionNode(4); |
|
1266 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); |
|
1267 |
|
1268 // Build the first if node: if (x <= 0.0) |
|
1269 // Node for 0 constant |
|
1270 Node *zeronode = makecon(TypeD::ZERO); |
|
1271 // Check x:0 |
|
1272 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); |
|
1273 // Check: If (x<=0) then go complex path |
|
1274 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); |
|
1275 // Branch either way |
|
1276 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); |
|
1277 Node *opt_test = _gvn.transform(if1); |
|
1278 //assert( opt_test->is_If(), "Expect an IfNode"); |
|
1279 IfNode *opt_if1 = (IfNode*)opt_test; |
|
1280 // Fast path taken; set region slot 3 |
|
1281 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); |
|
1282 r->init_req(3,fast_taken); // Capture fast-control |
|
1283 |
|
1284 // Fast path not-taken, i.e. slow path |
|
1285 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); |
|
1286 |
|
1287 // Set fast path result |
|
1288 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); |
|
1289 phi->init_req(3, fast_result); |
|
1290 |
|
1291 // Complex path |
|
1292 // Build the second if node (if y is int) |
|
1293 // Node for (int)y |
|
1294 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); |
|
1295 // Node for (double)((int) y) |
|
1296 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); |
|
1297 // Check (double)((int) y) : y |
|
1298 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); |
|
1299 // Check if (y isn't int) then go to slow path |
|
1300 |
|
1301 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); |
|
1302 // Branch eith way |
|
1303 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); |
|
1304 Node *slow_path = opt_iff(r,if2); // Set region path 2 |
|
1305 |
|
1306 // Calculate DPow(abs(x), y)*(1 & (int)y) |
|
1307 // Node for constant 1 |
|
1308 Node *conone = intcon(1); |
|
1309 // 1& (int)y |
|
1310 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); |
|
1311 // zero node |
|
1312 Node *conzero = intcon(0); |
|
1313 // Check (1&(int)y)==0? |
|
1314 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); |
|
1315 // Check if (1&(int)y)!=0?, if so the result is negative |
|
1316 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); |
|
1317 // abs(x) |
|
1318 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); |
|
1319 // abs(x)^y |
|
1320 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); |
|
1321 // -abs(x)^y |
|
1322 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); |
|
1323 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) |
|
1324 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); |
|
1325 // Set complex path fast result |
|
1326 phi->init_req(2, signresult); |
|
1327 |
|
1328 static const jlong nan_bits = CONST64(0x7ff8000000000000); |
|
1329 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN |
|
1330 r->init_req(1,slow_path); |
|
1331 phi->init_req(1,slow_result); |
|
1332 |
|
1333 // Post merge |
|
1334 set_control(_gvn.transform(r)); |
|
1335 record_for_igvn(r); |
|
1336 result=_gvn.transform(phi); |
|
1337 } |
|
1338 |
|
1339 //------------------- |
|
1340 //result=(result.isNaN())? uncommon_trap():result; |
|
1341 // Check: If isNaN() by checking result!=result? then go to Strict Math |
|
1342 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); |
|
1343 // Build the boolean node |
|
1344 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); |
|
1345 |
|
1346 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); |
|
1347 // End the current control-flow path |
|
1348 push_pair(x); |
|
1349 push_pair(y); |
|
1350 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow |
|
1351 // to handle. Recompile without intrinsifying Math.pow. |
|
1352 uncommon_trap(Deoptimization::Reason_intrinsic, |
|
1353 Deoptimization::Action_make_not_entrant); |
|
1354 } |
|
1355 |
|
1356 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
1357 |
|
1358 push_pair(result); |
|
1359 |
|
1360 return true; |
|
1361 } |
|
1362 |
|
1363 //------------------------------inline_trans------------------------------------- |
|
1364 // Inline transcendental instructions, if possible. The Intel hardware gets |
|
1365 // these right, no funny corner cases missed. |
|
1366 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { |
|
1367 _sp += arg_size(); // restore stack pointer |
|
1368 Node* arg = pop_math_arg(); |
|
1369 Node* trans = NULL; |
|
1370 |
|
1371 switch (id) { |
|
1372 case vmIntrinsics::_dlog: |
|
1373 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); |
|
1374 break; |
|
1375 case vmIntrinsics::_dlog10: |
|
1376 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); |
|
1377 break; |
|
1378 default: |
|
1379 assert(false, "bad intrinsic was passed in"); |
|
1380 return false; |
|
1381 } |
|
1382 |
|
1383 // Push result back on JVM stack |
|
1384 push_pair(trans); |
|
1385 return true; |
|
1386 } |
|
1387 |
|
1388 //------------------------------runtime_math----------------------------- |
|
1389 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { |
|
1390 Node* a = NULL; |
|
1391 Node* b = NULL; |
|
1392 |
|
1393 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), |
|
1394 "must be (DD)D or (D)D type"); |
|
1395 |
|
1396 // Inputs |
|
1397 _sp += arg_size(); // restore stack pointer |
|
1398 if (call_type == OptoRuntime::Math_DD_D_Type()) { |
|
1399 b = pop_math_arg(); |
|
1400 } |
|
1401 a = pop_math_arg(); |
|
1402 |
|
1403 const TypePtr* no_memory_effects = NULL; |
|
1404 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, |
|
1405 no_memory_effects, |
|
1406 a, top(), b, b ? top() : NULL); |
|
1407 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); |
|
1408 #ifdef ASSERT |
|
1409 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); |
|
1410 assert(value_top == top(), "second value must be top"); |
|
1411 #endif |
|
1412 |
|
1413 push_pair(value); |
|
1414 return true; |
|
1415 } |
|
1416 |
|
1417 //------------------------------inline_math_native----------------------------- |
|
1418 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { |
|
1419 switch (id) { |
|
1420 // These intrinsics are not properly supported on all hardware |
|
1421 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : |
|
1422 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); |
|
1423 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : |
|
1424 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); |
|
1425 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : |
|
1426 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); |
|
1427 |
|
1428 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : |
|
1429 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); |
|
1430 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : |
|
1431 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); |
|
1432 |
|
1433 // These intrinsics are supported on all hardware |
|
1434 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; |
|
1435 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; |
|
1436 |
|
1437 // These intrinsics don't work on X86. The ad implementation doesn't |
|
1438 // handle NaN's properly. Instead of returning infinity, the ad |
|
1439 // implementation returns a NaN on overflow. See bug: 6304089 |
|
1440 // Once the ad implementations are fixed, change the code below |
|
1441 // to match the intrinsics above |
|
1442 |
|
1443 case vmIntrinsics::_dexp: return |
|
1444 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); |
|
1445 case vmIntrinsics::_dpow: return |
|
1446 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); |
|
1447 |
|
1448 // These intrinsics are not yet correctly implemented |
|
1449 case vmIntrinsics::_datan2: |
|
1450 return false; |
|
1451 |
|
1452 default: |
|
1453 ShouldNotReachHere(); |
|
1454 return false; |
|
1455 } |
|
1456 } |
|
1457 |
|
1458 static bool is_simple_name(Node* n) { |
|
1459 return (n->req() == 1 // constant |
|
1460 || (n->is_Type() && n->as_Type()->type()->singleton()) |
|
1461 || n->is_Proj() // parameter or return value |
|
1462 || n->is_Phi() // local of some sort |
|
1463 ); |
|
1464 } |
|
1465 |
|
1466 //----------------------------inline_min_max----------------------------------- |
|
1467 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { |
|
1468 push(generate_min_max(id, argument(0), argument(1))); |
|
1469 |
|
1470 return true; |
|
1471 } |
|
1472 |
|
1473 Node* |
|
1474 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { |
|
1475 // These are the candidate return value: |
|
1476 Node* xvalue = x0; |
|
1477 Node* yvalue = y0; |
|
1478 |
|
1479 if (xvalue == yvalue) { |
|
1480 return xvalue; |
|
1481 } |
|
1482 |
|
1483 bool want_max = (id == vmIntrinsics::_max); |
|
1484 |
|
1485 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); |
|
1486 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); |
|
1487 if (txvalue == NULL || tyvalue == NULL) return top(); |
|
1488 // This is not really necessary, but it is consistent with a |
|
1489 // hypothetical MaxINode::Value method: |
|
1490 int widen = MAX2(txvalue->_widen, tyvalue->_widen); |
|
1491 |
|
1492 // %%% This folding logic should (ideally) be in a different place. |
|
1493 // Some should be inside IfNode, and there to be a more reliable |
|
1494 // transformation of ?: style patterns into cmoves. We also want |
|
1495 // more powerful optimizations around cmove and min/max. |
|
1496 |
|
1497 // Try to find a dominating comparison of these guys. |
|
1498 // It can simplify the index computation for Arrays.copyOf |
|
1499 // and similar uses of System.arraycopy. |
|
1500 // First, compute the normalized version of CmpI(x, y). |
|
1501 int cmp_op = Op_CmpI; |
|
1502 Node* xkey = xvalue; |
|
1503 Node* ykey = yvalue; |
|
1504 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); |
|
1505 if (ideal_cmpxy->is_Cmp()) { |
|
1506 // E.g., if we have CmpI(length - offset, count), |
|
1507 // it might idealize to CmpI(length, count + offset) |
|
1508 cmp_op = ideal_cmpxy->Opcode(); |
|
1509 xkey = ideal_cmpxy->in(1); |
|
1510 ykey = ideal_cmpxy->in(2); |
|
1511 } |
|
1512 |
|
1513 // Start by locating any relevant comparisons. |
|
1514 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; |
|
1515 Node* cmpxy = NULL; |
|
1516 Node* cmpyx = NULL; |
|
1517 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { |
|
1518 Node* cmp = start_from->fast_out(k); |
|
1519 if (cmp->outcnt() > 0 && // must have prior uses |
|
1520 cmp->in(0) == NULL && // must be context-independent |
|
1521 cmp->Opcode() == cmp_op) { // right kind of compare |
|
1522 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; |
|
1523 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; |
|
1524 } |
|
1525 } |
|
1526 |
|
1527 const int NCMPS = 2; |
|
1528 Node* cmps[NCMPS] = { cmpxy, cmpyx }; |
|
1529 int cmpn; |
|
1530 for (cmpn = 0; cmpn < NCMPS; cmpn++) { |
|
1531 if (cmps[cmpn] != NULL) break; // find a result |
|
1532 } |
|
1533 if (cmpn < NCMPS) { |
|
1534 // Look for a dominating test that tells us the min and max. |
|
1535 int depth = 0; // Limit search depth for speed |
|
1536 Node* dom = control(); |
|
1537 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { |
|
1538 if (++depth >= 100) break; |
|
1539 Node* ifproj = dom; |
|
1540 if (!ifproj->is_Proj()) continue; |
|
1541 Node* iff = ifproj->in(0); |
|
1542 if (!iff->is_If()) continue; |
|
1543 Node* bol = iff->in(1); |
|
1544 if (!bol->is_Bool()) continue; |
|
1545 Node* cmp = bol->in(1); |
|
1546 if (cmp == NULL) continue; |
|
1547 for (cmpn = 0; cmpn < NCMPS; cmpn++) |
|
1548 if (cmps[cmpn] == cmp) break; |
|
1549 if (cmpn == NCMPS) continue; |
|
1550 BoolTest::mask btest = bol->as_Bool()->_test._test; |
|
1551 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); |
|
1552 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); |
|
1553 // At this point, we know that 'x btest y' is true. |
|
1554 switch (btest) { |
|
1555 case BoolTest::eq: |
|
1556 // They are proven equal, so we can collapse the min/max. |
|
1557 // Either value is the answer. Choose the simpler. |
|
1558 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) |
|
1559 return yvalue; |
|
1560 return xvalue; |
|
1561 case BoolTest::lt: // x < y |
|
1562 case BoolTest::le: // x <= y |
|
1563 return (want_max ? yvalue : xvalue); |
|
1564 case BoolTest::gt: // x > y |
|
1565 case BoolTest::ge: // x >= y |
|
1566 return (want_max ? xvalue : yvalue); |
|
1567 } |
|
1568 } |
|
1569 } |
|
1570 |
|
1571 // We failed to find a dominating test. |
|
1572 // Let's pick a test that might GVN with prior tests. |
|
1573 Node* best_bol = NULL; |
|
1574 BoolTest::mask best_btest = BoolTest::illegal; |
|
1575 for (cmpn = 0; cmpn < NCMPS; cmpn++) { |
|
1576 Node* cmp = cmps[cmpn]; |
|
1577 if (cmp == NULL) continue; |
|
1578 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { |
|
1579 Node* bol = cmp->fast_out(j); |
|
1580 if (!bol->is_Bool()) continue; |
|
1581 BoolTest::mask btest = bol->as_Bool()->_test._test; |
|
1582 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; |
|
1583 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); |
|
1584 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { |
|
1585 best_bol = bol->as_Bool(); |
|
1586 best_btest = btest; |
|
1587 } |
|
1588 } |
|
1589 } |
|
1590 |
|
1591 Node* answer_if_true = NULL; |
|
1592 Node* answer_if_false = NULL; |
|
1593 switch (best_btest) { |
|
1594 default: |
|
1595 if (cmpxy == NULL) |
|
1596 cmpxy = ideal_cmpxy; |
|
1597 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); |
|
1598 // and fall through: |
|
1599 case BoolTest::lt: // x < y |
|
1600 case BoolTest::le: // x <= y |
|
1601 answer_if_true = (want_max ? yvalue : xvalue); |
|
1602 answer_if_false = (want_max ? xvalue : yvalue); |
|
1603 break; |
|
1604 case BoolTest::gt: // x > y |
|
1605 case BoolTest::ge: // x >= y |
|
1606 answer_if_true = (want_max ? xvalue : yvalue); |
|
1607 answer_if_false = (want_max ? yvalue : xvalue); |
|
1608 break; |
|
1609 } |
|
1610 |
|
1611 jint hi, lo; |
|
1612 if (want_max) { |
|
1613 // We can sharpen the minimum. |
|
1614 hi = MAX2(txvalue->_hi, tyvalue->_hi); |
|
1615 lo = MAX2(txvalue->_lo, tyvalue->_lo); |
|
1616 } else { |
|
1617 // We can sharpen the maximum. |
|
1618 hi = MIN2(txvalue->_hi, tyvalue->_hi); |
|
1619 lo = MIN2(txvalue->_lo, tyvalue->_lo); |
|
1620 } |
|
1621 |
|
1622 // Use a flow-free graph structure, to avoid creating excess control edges |
|
1623 // which could hinder other optimizations. |
|
1624 // Since Math.min/max is often used with arraycopy, we want |
|
1625 // tightly_coupled_allocation to be able to see beyond min/max expressions. |
|
1626 Node* cmov = CMoveNode::make(C, NULL, best_bol, |
|
1627 answer_if_false, answer_if_true, |
|
1628 TypeInt::make(lo, hi, widen)); |
|
1629 |
|
1630 return _gvn.transform(cmov); |
|
1631 |
|
1632 /* |
|
1633 // This is not as desirable as it may seem, since Min and Max |
|
1634 // nodes do not have a full set of optimizations. |
|
1635 // And they would interfere, anyway, with 'if' optimizations |
|
1636 // and with CMoveI canonical forms. |
|
1637 switch (id) { |
|
1638 case vmIntrinsics::_min: |
|
1639 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; |
|
1640 case vmIntrinsics::_max: |
|
1641 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; |
|
1642 default: |
|
1643 ShouldNotReachHere(); |
|
1644 } |
|
1645 */ |
|
1646 } |
|
1647 |
|
1648 inline int |
|
1649 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { |
|
1650 const TypePtr* base_type = TypePtr::NULL_PTR; |
|
1651 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); |
|
1652 if (base_type == NULL) { |
|
1653 // Unknown type. |
|
1654 return Type::AnyPtr; |
|
1655 } else if (base_type == TypePtr::NULL_PTR) { |
|
1656 // Since this is a NULL+long form, we have to switch to a rawptr. |
|
1657 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); |
|
1658 offset = MakeConX(0); |
|
1659 return Type::RawPtr; |
|
1660 } else if (base_type->base() == Type::RawPtr) { |
|
1661 return Type::RawPtr; |
|
1662 } else if (base_type->isa_oopptr()) { |
|
1663 // Base is never null => always a heap address. |
|
1664 if (base_type->ptr() == TypePtr::NotNull) { |
|
1665 return Type::OopPtr; |
|
1666 } |
|
1667 // Offset is small => always a heap address. |
|
1668 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); |
|
1669 if (offset_type != NULL && |
|
1670 base_type->offset() == 0 && // (should always be?) |
|
1671 offset_type->_lo >= 0 && |
|
1672 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { |
|
1673 return Type::OopPtr; |
|
1674 } |
|
1675 // Otherwise, it might either be oop+off or NULL+addr. |
|
1676 return Type::AnyPtr; |
|
1677 } else { |
|
1678 // No information: |
|
1679 return Type::AnyPtr; |
|
1680 } |
|
1681 } |
|
1682 |
|
1683 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { |
|
1684 int kind = classify_unsafe_addr(base, offset); |
|
1685 if (kind == Type::RawPtr) { |
|
1686 return basic_plus_adr(top(), base, offset); |
|
1687 } else { |
|
1688 return basic_plus_adr(base, offset); |
|
1689 } |
|
1690 } |
|
1691 |
|
1692 //----------------------------inline_reverseBytes_int/long------------------- |
|
1693 // inline Int.reverseBytes(int) |
|
1694 // inline Long.reverseByes(long) |
|
1695 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { |
|
1696 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes"); |
|
1697 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; |
|
1698 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; |
|
1699 _sp += arg_size(); // restore stack pointer |
|
1700 switch (id) { |
|
1701 case vmIntrinsics::_reverseBytes_i: |
|
1702 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); |
|
1703 break; |
|
1704 case vmIntrinsics::_reverseBytes_l: |
|
1705 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); |
|
1706 break; |
|
1707 default: |
|
1708 ; |
|
1709 } |
|
1710 return true; |
|
1711 } |
|
1712 |
|
1713 //----------------------------inline_unsafe_access---------------------------- |
|
1714 |
|
1715 const static BasicType T_ADDRESS_HOLDER = T_LONG; |
|
1716 |
|
1717 // Interpret Unsafe.fieldOffset cookies correctly: |
|
1718 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); |
|
1719 |
|
1720 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { |
|
1721 if (callee()->is_static()) return false; // caller must have the capability! |
|
1722 |
|
1723 #ifndef PRODUCT |
|
1724 { |
|
1725 ResourceMark rm; |
|
1726 // Check the signatures. |
|
1727 ciSignature* sig = signature(); |
|
1728 #ifdef ASSERT |
|
1729 if (!is_store) { |
|
1730 // Object getObject(Object base, int/long offset), etc. |
|
1731 BasicType rtype = sig->return_type()->basic_type(); |
|
1732 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) |
|
1733 rtype = T_ADDRESS; // it is really a C void* |
|
1734 assert(rtype == type, "getter must return the expected value"); |
|
1735 if (!is_native_ptr) { |
|
1736 assert(sig->count() == 2, "oop getter has 2 arguments"); |
|
1737 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); |
|
1738 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); |
|
1739 } else { |
|
1740 assert(sig->count() == 1, "native getter has 1 argument"); |
|
1741 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); |
|
1742 } |
|
1743 } else { |
|
1744 // void putObject(Object base, int/long offset, Object x), etc. |
|
1745 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); |
|
1746 if (!is_native_ptr) { |
|
1747 assert(sig->count() == 3, "oop putter has 3 arguments"); |
|
1748 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); |
|
1749 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); |
|
1750 } else { |
|
1751 assert(sig->count() == 2, "native putter has 2 arguments"); |
|
1752 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); |
|
1753 } |
|
1754 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); |
|
1755 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) |
|
1756 vtype = T_ADDRESS; // it is really a C void* |
|
1757 assert(vtype == type, "putter must accept the expected value"); |
|
1758 } |
|
1759 #endif // ASSERT |
|
1760 } |
|
1761 #endif //PRODUCT |
|
1762 |
|
1763 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
1764 |
|
1765 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; |
|
1766 |
|
1767 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words |
|
1768 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); |
|
1769 |
|
1770 debug_only(int saved_sp = _sp); |
|
1771 _sp += nargs; |
|
1772 |
|
1773 Node* val; |
|
1774 debug_only(val = (Node*)(uintptr_t)-1); |
|
1775 |
|
1776 |
|
1777 if (is_store) { |
|
1778 // Get the value being stored. (Pop it first; it was pushed last.) |
|
1779 switch (type) { |
|
1780 case T_DOUBLE: |
|
1781 case T_LONG: |
|
1782 case T_ADDRESS: |
|
1783 val = pop_pair(); |
|
1784 break; |
|
1785 default: |
|
1786 val = pop(); |
|
1787 } |
|
1788 } |
|
1789 |
|
1790 // Build address expression. See the code in inline_unsafe_prefetch. |
|
1791 Node *adr; |
|
1792 Node *heap_base_oop = top(); |
|
1793 if (!is_native_ptr) { |
|
1794 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset |
|
1795 Node* offset = pop_pair(); |
|
1796 // The base is either a Java object or a value produced by Unsafe.staticFieldBase |
|
1797 Node* base = pop(); |
|
1798 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
|
1799 // to be plain byte offsets, which are also the same as those accepted |
|
1800 // by oopDesc::field_base. |
|
1801 assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
|
1802 "fieldOffset must be byte-scaled"); |
|
1803 // 32-bit machines ignore the high half! |
|
1804 offset = ConvL2X(offset); |
|
1805 adr = make_unsafe_address(base, offset); |
|
1806 heap_base_oop = base; |
|
1807 } else { |
|
1808 Node* ptr = pop_pair(); |
|
1809 // Adjust Java long to machine word: |
|
1810 ptr = ConvL2X(ptr); |
|
1811 adr = make_unsafe_address(NULL, ptr); |
|
1812 } |
|
1813 |
|
1814 // Pop receiver last: it was pushed first. |
|
1815 Node *receiver = pop(); |
|
1816 |
|
1817 assert(saved_sp == _sp, "must have correct argument count"); |
|
1818 |
|
1819 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); |
|
1820 |
|
1821 // First guess at the value type. |
|
1822 const Type *value_type = Type::get_const_basic_type(type); |
|
1823 |
|
1824 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, |
|
1825 // there was not enough information to nail it down. |
|
1826 Compile::AliasType* alias_type = C->alias_type(adr_type); |
|
1827 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); |
|
1828 |
|
1829 // We will need memory barriers unless we can determine a unique |
|
1830 // alias category for this reference. (Note: If for some reason |
|
1831 // the barriers get omitted and the unsafe reference begins to "pollute" |
|
1832 // the alias analysis of the rest of the graph, either Compile::can_alias |
|
1833 // or Compile::must_alias will throw a diagnostic assert.) |
|
1834 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); |
|
1835 |
|
1836 if (!is_store && type == T_OBJECT) { |
|
1837 // Attempt to infer a sharper value type from the offset and base type. |
|
1838 ciKlass* sharpened_klass = NULL; |
|
1839 |
|
1840 // See if it is an instance field, with an object type. |
|
1841 if (alias_type->field() != NULL) { |
|
1842 assert(!is_native_ptr, "native pointer op cannot use a java address"); |
|
1843 if (alias_type->field()->type()->is_klass()) { |
|
1844 sharpened_klass = alias_type->field()->type()->as_klass(); |
|
1845 } |
|
1846 } |
|
1847 |
|
1848 // See if it is a narrow oop array. |
|
1849 if (adr_type->isa_aryptr()) { |
|
1850 if (adr_type->offset() >= objArrayOopDesc::header_size() * wordSize) { |
|
1851 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); |
|
1852 if (elem_type != NULL) { |
|
1853 sharpened_klass = elem_type->klass(); |
|
1854 } |
|
1855 } |
|
1856 } |
|
1857 |
|
1858 if (sharpened_klass != NULL) { |
|
1859 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); |
|
1860 |
|
1861 // Sharpen the value type. |
|
1862 value_type = tjp; |
|
1863 |
|
1864 #ifndef PRODUCT |
|
1865 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { |
|
1866 tty->print(" from base type: "); adr_type->dump(); |
|
1867 tty->print(" sharpened value: "); value_type->dump(); |
|
1868 } |
|
1869 #endif |
|
1870 } |
|
1871 } |
|
1872 |
|
1873 // Null check on self without removing any arguments. The argument |
|
1874 // null check technically happens in the wrong place, which can lead to |
|
1875 // invalid stack traces when the primitive is inlined into a method |
|
1876 // which handles NullPointerExceptions. |
|
1877 _sp += nargs; |
|
1878 do_null_check(receiver, T_OBJECT); |
|
1879 _sp -= nargs; |
|
1880 if (stopped()) { |
|
1881 return true; |
|
1882 } |
|
1883 // Heap pointers get a null-check from the interpreter, |
|
1884 // as a courtesy. However, this is not guaranteed by Unsafe, |
|
1885 // and it is not possible to fully distinguish unintended nulls |
|
1886 // from intended ones in this API. |
|
1887 |
|
1888 if (is_volatile) { |
|
1889 // We need to emit leading and trailing CPU membars (see below) in |
|
1890 // addition to memory membars when is_volatile. This is a little |
|
1891 // too strong, but avoids the need to insert per-alias-type |
|
1892 // volatile membars (for stores; compare Parse::do_put_xxx), which |
|
1893 // we cannot do effctively here because we probably only have a |
|
1894 // rough approximation of type. |
|
1895 need_mem_bar = true; |
|
1896 // For Stores, place a memory ordering barrier now. |
|
1897 if (is_store) |
|
1898 insert_mem_bar(Op_MemBarRelease); |
|
1899 } |
|
1900 |
|
1901 // Memory barrier to prevent normal and 'unsafe' accesses from |
|
1902 // bypassing each other. Happens after null checks, so the |
|
1903 // exception paths do not take memory state from the memory barrier, |
|
1904 // so there's no problems making a strong assert about mixing users |
|
1905 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar |
|
1906 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. |
|
1907 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); |
|
1908 |
|
1909 if (!is_store) { |
|
1910 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); |
|
1911 // load value and push onto stack |
|
1912 switch (type) { |
|
1913 case T_BOOLEAN: |
|
1914 case T_CHAR: |
|
1915 case T_BYTE: |
|
1916 case T_SHORT: |
|
1917 case T_INT: |
|
1918 case T_FLOAT: |
|
1919 case T_OBJECT: |
|
1920 push( p ); |
|
1921 break; |
|
1922 case T_ADDRESS: |
|
1923 // Cast to an int type. |
|
1924 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); |
|
1925 p = ConvX2L(p); |
|
1926 push_pair(p); |
|
1927 break; |
|
1928 case T_DOUBLE: |
|
1929 case T_LONG: |
|
1930 push_pair( p ); |
|
1931 break; |
|
1932 default: ShouldNotReachHere(); |
|
1933 } |
|
1934 } else { |
|
1935 // place effect of store into memory |
|
1936 switch (type) { |
|
1937 case T_DOUBLE: |
|
1938 val = dstore_rounding(val); |
|
1939 break; |
|
1940 case T_ADDRESS: |
|
1941 // Repackage the long as a pointer. |
|
1942 val = ConvL2X(val); |
|
1943 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); |
|
1944 break; |
|
1945 } |
|
1946 |
|
1947 if (type != T_OBJECT ) { |
|
1948 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); |
|
1949 } else { |
|
1950 // Possibly an oop being stored to Java heap or native memory |
|
1951 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { |
|
1952 // oop to Java heap. |
|
1953 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); |
|
1954 } else { |
|
1955 |
|
1956 // We can't tell at compile time if we are storing in the Java heap or outside |
|
1957 // of it. So we need to emit code to conditionally do the proper type of |
|
1958 // store. |
|
1959 |
|
1960 IdealKit kit(gvn(), control(), merged_memory()); |
|
1961 kit.declares_done(); |
|
1962 // QQQ who knows what probability is here?? |
|
1963 kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { |
|
1964 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); |
|
1965 } kit.else_(); { |
|
1966 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); |
|
1967 } kit.end_if(); |
|
1968 } |
|
1969 } |
|
1970 } |
|
1971 |
|
1972 if (is_volatile) { |
|
1973 if (!is_store) |
|
1974 insert_mem_bar(Op_MemBarAcquire); |
|
1975 else |
|
1976 insert_mem_bar(Op_MemBarVolatile); |
|
1977 } |
|
1978 |
|
1979 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); |
|
1980 |
|
1981 return true; |
|
1982 } |
|
1983 |
|
1984 //----------------------------inline_unsafe_prefetch---------------------------- |
|
1985 |
|
1986 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { |
|
1987 #ifndef PRODUCT |
|
1988 { |
|
1989 ResourceMark rm; |
|
1990 // Check the signatures. |
|
1991 ciSignature* sig = signature(); |
|
1992 #ifdef ASSERT |
|
1993 // Object getObject(Object base, int/long offset), etc. |
|
1994 BasicType rtype = sig->return_type()->basic_type(); |
|
1995 if (!is_native_ptr) { |
|
1996 assert(sig->count() == 2, "oop prefetch has 2 arguments"); |
|
1997 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); |
|
1998 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); |
|
1999 } else { |
|
2000 assert(sig->count() == 1, "native prefetch has 1 argument"); |
|
2001 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); |
|
2002 } |
|
2003 #endif // ASSERT |
|
2004 } |
|
2005 #endif // !PRODUCT |
|
2006 |
|
2007 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
2008 |
|
2009 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args |
|
2010 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); |
|
2011 |
|
2012 debug_only(int saved_sp = _sp); |
|
2013 _sp += nargs; |
|
2014 |
|
2015 // Build address expression. See the code in inline_unsafe_access. |
|
2016 Node *adr; |
|
2017 if (!is_native_ptr) { |
|
2018 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset |
|
2019 Node* offset = pop_pair(); |
|
2020 // The base is either a Java object or a value produced by Unsafe.staticFieldBase |
|
2021 Node* base = pop(); |
|
2022 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
|
2023 // to be plain byte offsets, which are also the same as those accepted |
|
2024 // by oopDesc::field_base. |
|
2025 assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
|
2026 "fieldOffset must be byte-scaled"); |
|
2027 // 32-bit machines ignore the high half! |
|
2028 offset = ConvL2X(offset); |
|
2029 adr = make_unsafe_address(base, offset); |
|
2030 } else { |
|
2031 Node* ptr = pop_pair(); |
|
2032 // Adjust Java long to machine word: |
|
2033 ptr = ConvL2X(ptr); |
|
2034 adr = make_unsafe_address(NULL, ptr); |
|
2035 } |
|
2036 |
|
2037 if (is_static) { |
|
2038 assert(saved_sp == _sp, "must have correct argument count"); |
|
2039 } else { |
|
2040 // Pop receiver last: it was pushed first. |
|
2041 Node *receiver = pop(); |
|
2042 assert(saved_sp == _sp, "must have correct argument count"); |
|
2043 |
|
2044 // Null check on self without removing any arguments. The argument |
|
2045 // null check technically happens in the wrong place, which can lead to |
|
2046 // invalid stack traces when the primitive is inlined into a method |
|
2047 // which handles NullPointerExceptions. |
|
2048 _sp += nargs; |
|
2049 do_null_check(receiver, T_OBJECT); |
|
2050 _sp -= nargs; |
|
2051 if (stopped()) { |
|
2052 return true; |
|
2053 } |
|
2054 } |
|
2055 |
|
2056 // Generate the read or write prefetch |
|
2057 Node *prefetch; |
|
2058 if (is_store) { |
|
2059 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); |
|
2060 } else { |
|
2061 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); |
|
2062 } |
|
2063 prefetch->init_req(0, control()); |
|
2064 set_i_o(_gvn.transform(prefetch)); |
|
2065 |
|
2066 return true; |
|
2067 } |
|
2068 |
|
2069 //----------------------------inline_unsafe_CAS---------------------------- |
|
2070 |
|
2071 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { |
|
2072 // This basic scheme here is the same as inline_unsafe_access, but |
|
2073 // differs in enough details that combining them would make the code |
|
2074 // overly confusing. (This is a true fact! I originally combined |
|
2075 // them, but even I was confused by it!) As much code/comments as |
|
2076 // possible are retained from inline_unsafe_access though to make |
|
2077 // the correspondances clearer. - dl |
|
2078 |
|
2079 if (callee()->is_static()) return false; // caller must have the capability! |
|
2080 |
|
2081 #ifndef PRODUCT |
|
2082 { |
|
2083 ResourceMark rm; |
|
2084 // Check the signatures. |
|
2085 ciSignature* sig = signature(); |
|
2086 #ifdef ASSERT |
|
2087 BasicType rtype = sig->return_type()->basic_type(); |
|
2088 assert(rtype == T_BOOLEAN, "CAS must return boolean"); |
|
2089 assert(sig->count() == 4, "CAS has 4 arguments"); |
|
2090 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); |
|
2091 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); |
|
2092 #endif // ASSERT |
|
2093 } |
|
2094 #endif //PRODUCT |
|
2095 |
|
2096 // number of stack slots per value argument (1 or 2) |
|
2097 int type_words = type2size[type]; |
|
2098 |
|
2099 // Cannot inline wide CAS on machines that don't support it natively |
|
2100 if (type2aelembytes[type] > BytesPerInt && !VM_Version::supports_cx8()) |
|
2101 return false; |
|
2102 |
|
2103 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
2104 |
|
2105 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; |
|
2106 int nargs = 1 + 1 + 2 + type_words + type_words; |
|
2107 |
|
2108 // pop arguments: newval, oldval, offset, base, and receiver |
|
2109 debug_only(int saved_sp = _sp); |
|
2110 _sp += nargs; |
|
2111 Node* newval = (type_words == 1) ? pop() : pop_pair(); |
|
2112 Node* oldval = (type_words == 1) ? pop() : pop_pair(); |
|
2113 Node *offset = pop_pair(); |
|
2114 Node *base = pop(); |
|
2115 Node *receiver = pop(); |
|
2116 assert(saved_sp == _sp, "must have correct argument count"); |
|
2117 |
|
2118 // Null check receiver. |
|
2119 _sp += nargs; |
|
2120 do_null_check(receiver, T_OBJECT); |
|
2121 _sp -= nargs; |
|
2122 if (stopped()) { |
|
2123 return true; |
|
2124 } |
|
2125 |
|
2126 // Build field offset expression. |
|
2127 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
|
2128 // to be plain byte offsets, which are also the same as those accepted |
|
2129 // by oopDesc::field_base. |
|
2130 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); |
|
2131 // 32-bit machines ignore the high half of long offsets |
|
2132 offset = ConvL2X(offset); |
|
2133 Node* adr = make_unsafe_address(base, offset); |
|
2134 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); |
|
2135 |
|
2136 // (Unlike inline_unsafe_access, there seems no point in trying |
|
2137 // to refine types. Just use the coarse types here. |
|
2138 const Type *value_type = Type::get_const_basic_type(type); |
|
2139 Compile::AliasType* alias_type = C->alias_type(adr_type); |
|
2140 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); |
|
2141 int alias_idx = C->get_alias_index(adr_type); |
|
2142 |
|
2143 // Memory-model-wise, a CAS acts like a little synchronized block, |
|
2144 // so needs barriers on each side. These don't't translate into |
|
2145 // actual barriers on most machines, but we still need rest of |
|
2146 // compiler to respect ordering. |
|
2147 |
|
2148 insert_mem_bar(Op_MemBarRelease); |
|
2149 insert_mem_bar(Op_MemBarCPUOrder); |
|
2150 |
|
2151 // 4984716: MemBars must be inserted before this |
|
2152 // memory node in order to avoid a false |
|
2153 // dependency which will confuse the scheduler. |
|
2154 Node *mem = memory(alias_idx); |
|
2155 |
|
2156 // For now, we handle only those cases that actually exist: ints, |
|
2157 // longs, and Object. Adding others should be straightforward. |
|
2158 Node* cas; |
|
2159 switch(type) { |
|
2160 case T_INT: |
|
2161 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); |
|
2162 break; |
|
2163 case T_LONG: |
|
2164 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); |
|
2165 break; |
|
2166 case T_OBJECT: |
|
2167 // reference stores need a store barrier. |
|
2168 // (They don't if CAS fails, but it isn't worth checking.) |
|
2169 pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT); |
|
2170 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); |
|
2171 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); |
|
2172 break; |
|
2173 default: |
|
2174 ShouldNotReachHere(); |
|
2175 break; |
|
2176 } |
|
2177 |
|
2178 // SCMemProjNodes represent the memory state of CAS. Their main |
|
2179 // role is to prevent CAS nodes from being optimized away when their |
|
2180 // results aren't used. |
|
2181 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); |
|
2182 set_memory(proj, alias_idx); |
|
2183 |
|
2184 // Add the trailing membar surrounding the access |
|
2185 insert_mem_bar(Op_MemBarCPUOrder); |
|
2186 insert_mem_bar(Op_MemBarAcquire); |
|
2187 |
|
2188 push(cas); |
|
2189 return true; |
|
2190 } |
|
2191 |
|
2192 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { |
|
2193 // This is another variant of inline_unsafe_access, differing in |
|
2194 // that it always issues store-store ("release") barrier and ensures |
|
2195 // store-atomicity (which only matters for "long"). |
|
2196 |
|
2197 if (callee()->is_static()) return false; // caller must have the capability! |
|
2198 |
|
2199 #ifndef PRODUCT |
|
2200 { |
|
2201 ResourceMark rm; |
|
2202 // Check the signatures. |
|
2203 ciSignature* sig = signature(); |
|
2204 #ifdef ASSERT |
|
2205 BasicType rtype = sig->return_type()->basic_type(); |
|
2206 assert(rtype == T_VOID, "must return void"); |
|
2207 assert(sig->count() == 3, "has 3 arguments"); |
|
2208 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); |
|
2209 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); |
|
2210 #endif // ASSERT |
|
2211 } |
|
2212 #endif //PRODUCT |
|
2213 |
|
2214 // number of stack slots per value argument (1 or 2) |
|
2215 int type_words = type2size[type]; |
|
2216 |
|
2217 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
2218 |
|
2219 // Argument words: "this" plus oop plus offset plus value; |
|
2220 int nargs = 1 + 1 + 2 + type_words; |
|
2221 |
|
2222 // pop arguments: val, offset, base, and receiver |
|
2223 debug_only(int saved_sp = _sp); |
|
2224 _sp += nargs; |
|
2225 Node* val = (type_words == 1) ? pop() : pop_pair(); |
|
2226 Node *offset = pop_pair(); |
|
2227 Node *base = pop(); |
|
2228 Node *receiver = pop(); |
|
2229 assert(saved_sp == _sp, "must have correct argument count"); |
|
2230 |
|
2231 // Null check receiver. |
|
2232 _sp += nargs; |
|
2233 do_null_check(receiver, T_OBJECT); |
|
2234 _sp -= nargs; |
|
2235 if (stopped()) { |
|
2236 return true; |
|
2237 } |
|
2238 |
|
2239 // Build field offset expression. |
|
2240 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); |
|
2241 // 32-bit machines ignore the high half of long offsets |
|
2242 offset = ConvL2X(offset); |
|
2243 Node* adr = make_unsafe_address(base, offset); |
|
2244 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); |
|
2245 const Type *value_type = Type::get_const_basic_type(type); |
|
2246 Compile::AliasType* alias_type = C->alias_type(adr_type); |
|
2247 |
|
2248 insert_mem_bar(Op_MemBarRelease); |
|
2249 insert_mem_bar(Op_MemBarCPUOrder); |
|
2250 // Ensure that the store is atomic for longs: |
|
2251 bool require_atomic_access = true; |
|
2252 Node* store; |
|
2253 if (type == T_OBJECT) // reference stores need a store barrier. |
|
2254 store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type); |
|
2255 else { |
|
2256 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); |
|
2257 } |
|
2258 insert_mem_bar(Op_MemBarCPUOrder); |
|
2259 return true; |
|
2260 } |
|
2261 |
|
2262 bool LibraryCallKit::inline_unsafe_allocate() { |
|
2263 if (callee()->is_static()) return false; // caller must have the capability! |
|
2264 int nargs = 1 + 1; |
|
2265 assert(signature()->size() == nargs-1, "alloc has 1 argument"); |
|
2266 null_check_receiver(callee()); // check then ignore argument(0) |
|
2267 _sp += nargs; // set original stack for use by uncommon_trap |
|
2268 Node* cls = do_null_check(argument(1), T_OBJECT); |
|
2269 _sp -= nargs; |
|
2270 if (stopped()) return true; |
|
2271 |
|
2272 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); |
|
2273 _sp += nargs; // set original stack for use by uncommon_trap |
|
2274 kls = do_null_check(kls, T_OBJECT); |
|
2275 _sp -= nargs; |
|
2276 if (stopped()) return true; // argument was like int.class |
|
2277 |
|
2278 // Note: The argument might still be an illegal value like |
|
2279 // Serializable.class or Object[].class. The runtime will handle it. |
|
2280 // But we must make an explicit check for initialization. |
|
2281 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc)); |
|
2282 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT); |
|
2283 Node* bits = intcon(instanceKlass::fully_initialized); |
|
2284 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); |
|
2285 // The 'test' is non-zero if we need to take a slow path. |
|
2286 |
|
2287 Node* obj = new_instance(kls, test); |
|
2288 push(obj); |
|
2289 |
|
2290 return true; |
|
2291 } |
|
2292 |
|
2293 //------------------------inline_native_time_funcs-------------- |
|
2294 // inline code for System.currentTimeMillis() and System.nanoTime() |
|
2295 // these have the same type and signature |
|
2296 bool LibraryCallKit::inline_native_time_funcs(bool isNano) { |
|
2297 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : |
|
2298 CAST_FROM_FN_PTR(address, os::javaTimeMillis); |
|
2299 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; |
|
2300 const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); |
|
2301 const TypePtr* no_memory_effects = NULL; |
|
2302 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); |
|
2303 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); |
|
2304 #ifdef ASSERT |
|
2305 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); |
|
2306 assert(value_top == top(), "second value must be top"); |
|
2307 #endif |
|
2308 push_pair(value); |
|
2309 return true; |
|
2310 } |
|
2311 |
|
2312 //------------------------inline_native_currentThread------------------ |
|
2313 bool LibraryCallKit::inline_native_currentThread() { |
|
2314 Node* junk = NULL; |
|
2315 push(generate_current_thread(junk)); |
|
2316 return true; |
|
2317 } |
|
2318 |
|
2319 //------------------------inline_native_isInterrupted------------------ |
|
2320 bool LibraryCallKit::inline_native_isInterrupted() { |
|
2321 const int nargs = 1+1; // receiver + boolean |
|
2322 assert(nargs == arg_size(), "sanity"); |
|
2323 // Add a fast path to t.isInterrupted(clear_int): |
|
2324 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) |
|
2325 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) |
|
2326 // So, in the common case that the interrupt bit is false, |
|
2327 // we avoid making a call into the VM. Even if the interrupt bit |
|
2328 // is true, if the clear_int argument is false, we avoid the VM call. |
|
2329 // However, if the receiver is not currentThread, we must call the VM, |
|
2330 // because there must be some locking done around the operation. |
|
2331 |
|
2332 // We only go to the fast case code if we pass two guards. |
|
2333 // Paths which do not pass are accumulated in the slow_region. |
|
2334 RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
2335 record_for_igvn(slow_region); |
|
2336 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow |
|
2337 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); |
|
2338 enum { no_int_result_path = 1, |
|
2339 no_clear_result_path = 2, |
|
2340 slow_result_path = 3 |
|
2341 }; |
|
2342 |
|
2343 // (a) Receiving thread must be the current thread. |
|
2344 Node* rec_thr = argument(0); |
|
2345 Node* tls_ptr = NULL; |
|
2346 Node* cur_thr = generate_current_thread(tls_ptr); |
|
2347 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); |
|
2348 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); |
|
2349 |
|
2350 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); |
|
2351 if (!known_current_thread) |
|
2352 generate_slow_guard(bol_thr, slow_region); |
|
2353 |
|
2354 // (b) Interrupt bit on TLS must be false. |
|
2355 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); |
|
2356 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); |
|
2357 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); |
|
2358 Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT); |
|
2359 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); |
|
2360 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); |
|
2361 |
|
2362 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); |
|
2363 |
|
2364 // First fast path: if (!TLS._interrupted) return false; |
|
2365 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); |
|
2366 result_rgn->init_req(no_int_result_path, false_bit); |
|
2367 result_val->init_req(no_int_result_path, intcon(0)); |
|
2368 |
|
2369 // drop through to next case |
|
2370 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); |
|
2371 |
|
2372 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. |
|
2373 Node* clr_arg = argument(1); |
|
2374 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); |
|
2375 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); |
|
2376 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); |
|
2377 |
|
2378 // Second fast path: ... else if (!clear_int) return true; |
|
2379 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); |
|
2380 result_rgn->init_req(no_clear_result_path, false_arg); |
|
2381 result_val->init_req(no_clear_result_path, intcon(1)); |
|
2382 |
|
2383 // drop through to next case |
|
2384 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); |
|
2385 |
|
2386 // (d) Otherwise, go to the slow path. |
|
2387 slow_region->add_req(control()); |
|
2388 set_control( _gvn.transform(slow_region) ); |
|
2389 |
|
2390 if (stopped()) { |
|
2391 // There is no slow path. |
|
2392 result_rgn->init_req(slow_result_path, top()); |
|
2393 result_val->init_req(slow_result_path, top()); |
|
2394 } else { |
|
2395 // non-virtual because it is a private non-static |
|
2396 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); |
|
2397 |
|
2398 Node* slow_val = set_results_for_java_call(slow_call); |
|
2399 // this->control() comes from set_results_for_java_call |
|
2400 |
|
2401 // If we know that the result of the slow call will be true, tell the optimizer! |
|
2402 if (known_current_thread) slow_val = intcon(1); |
|
2403 |
|
2404 Node* fast_io = slow_call->in(TypeFunc::I_O); |
|
2405 Node* fast_mem = slow_call->in(TypeFunc::Memory); |
|
2406 // These two phis are pre-filled with copies of of the fast IO and Memory |
|
2407 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); |
|
2408 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); |
|
2409 |
|
2410 result_rgn->init_req(slow_result_path, control()); |
|
2411 io_phi ->init_req(slow_result_path, i_o()); |
|
2412 mem_phi ->init_req(slow_result_path, reset_memory()); |
|
2413 result_val->init_req(slow_result_path, slow_val); |
|
2414 |
|
2415 set_all_memory( _gvn.transform(mem_phi) ); |
|
2416 set_i_o( _gvn.transform(io_phi) ); |
|
2417 } |
|
2418 |
|
2419 push_result(result_rgn, result_val); |
|
2420 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2421 |
|
2422 return true; |
|
2423 } |
|
2424 |
|
2425 //---------------------------load_mirror_from_klass---------------------------- |
|
2426 // Given a klass oop, load its java mirror (a java.lang.Class oop). |
|
2427 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { |
|
2428 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc)); |
|
2429 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); |
|
2430 } |
|
2431 |
|
2432 //-----------------------load_klass_from_mirror_common------------------------- |
|
2433 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. |
|
2434 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), |
|
2435 // and branch to the given path on the region. |
|
2436 // If never_see_null, take an uncommon trap on null, so we can optimistically |
|
2437 // compile for the non-null case. |
|
2438 // If the region is NULL, force never_see_null = true. |
|
2439 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, |
|
2440 bool never_see_null, |
|
2441 int nargs, |
|
2442 RegionNode* region, |
|
2443 int null_path, |
|
2444 int offset) { |
|
2445 if (region == NULL) never_see_null = true; |
|
2446 Node* p = basic_plus_adr(mirror, offset); |
|
2447 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; |
|
2448 Node* kls = _gvn.transform(new (C, 3) LoadKlassNode(0, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); |
|
2449 _sp += nargs; // any deopt will start just before call to enclosing method |
|
2450 Node* null_ctl = top(); |
|
2451 kls = null_check_oop(kls, &null_ctl, never_see_null); |
|
2452 if (region != NULL) { |
|
2453 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). |
|
2454 region->init_req(null_path, null_ctl); |
|
2455 } else { |
|
2456 assert(null_ctl == top(), "no loose ends"); |
|
2457 } |
|
2458 _sp -= nargs; |
|
2459 return kls; |
|
2460 } |
|
2461 |
|
2462 //--------------------(inline_native_Class_query helpers)--------------------- |
|
2463 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. |
|
2464 // Fall through if (mods & mask) == bits, take the guard otherwise. |
|
2465 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { |
|
2466 // Branch around if the given klass has the given modifier bit set. |
|
2467 // Like generate_guard, adds a new path onto the region. |
|
2468 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); |
|
2469 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); |
|
2470 Node* mask = intcon(modifier_mask); |
|
2471 Node* bits = intcon(modifier_bits); |
|
2472 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); |
|
2473 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); |
|
2474 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); |
|
2475 return generate_fair_guard(bol, region); |
|
2476 } |
|
2477 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { |
|
2478 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); |
|
2479 } |
|
2480 |
|
2481 //-------------------------inline_native_Class_query------------------- |
|
2482 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { |
|
2483 int nargs = 1+0; // just the Class mirror, in most cases |
|
2484 const Type* return_type = TypeInt::BOOL; |
|
2485 Node* prim_return_value = top(); // what happens if it's a primitive class? |
|
2486 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
|
2487 bool expect_prim = false; // most of these guys expect to work on refs |
|
2488 |
|
2489 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; |
|
2490 |
|
2491 switch (id) { |
|
2492 case vmIntrinsics::_isInstance: |
|
2493 nargs = 1+1; // the Class mirror, plus the object getting queried about |
|
2494 // nothing is an instance of a primitive type |
|
2495 prim_return_value = intcon(0); |
|
2496 break; |
|
2497 case vmIntrinsics::_getModifiers: |
|
2498 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); |
|
2499 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); |
|
2500 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); |
|
2501 break; |
|
2502 case vmIntrinsics::_isInterface: |
|
2503 prim_return_value = intcon(0); |
|
2504 break; |
|
2505 case vmIntrinsics::_isArray: |
|
2506 prim_return_value = intcon(0); |
|
2507 expect_prim = true; // cf. ObjectStreamClass.getClassSignature |
|
2508 break; |
|
2509 case vmIntrinsics::_isPrimitive: |
|
2510 prim_return_value = intcon(1); |
|
2511 expect_prim = true; // obviously |
|
2512 break; |
|
2513 case vmIntrinsics::_getSuperclass: |
|
2514 prim_return_value = null(); |
|
2515 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); |
|
2516 break; |
|
2517 case vmIntrinsics::_getComponentType: |
|
2518 prim_return_value = null(); |
|
2519 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); |
|
2520 break; |
|
2521 case vmIntrinsics::_getClassAccessFlags: |
|
2522 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); |
|
2523 return_type = TypeInt::INT; // not bool! 6297094 |
|
2524 break; |
|
2525 default: |
|
2526 ShouldNotReachHere(); |
|
2527 } |
|
2528 |
|
2529 Node* mirror = argument(0); |
|
2530 Node* obj = (nargs <= 1)? top(): argument(1); |
|
2531 |
|
2532 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); |
|
2533 if (mirror_con == NULL) return false; // cannot happen? |
|
2534 |
|
2535 #ifndef PRODUCT |
|
2536 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { |
|
2537 ciType* k = mirror_con->java_mirror_type(); |
|
2538 if (k) { |
|
2539 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); |
|
2540 k->print_name(); |
|
2541 tty->cr(); |
|
2542 } |
|
2543 } |
|
2544 #endif |
|
2545 |
|
2546 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). |
|
2547 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
2548 record_for_igvn(region); |
|
2549 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); |
|
2550 |
|
2551 // The mirror will never be null of Reflection.getClassAccessFlags, however |
|
2552 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE |
|
2553 // if it is. See bug 4774291. |
|
2554 |
|
2555 // For Reflection.getClassAccessFlags(), the null check occurs in |
|
2556 // the wrong place; see inline_unsafe_access(), above, for a similar |
|
2557 // situation. |
|
2558 _sp += nargs; // set original stack for use by uncommon_trap |
|
2559 mirror = do_null_check(mirror, T_OBJECT); |
|
2560 _sp -= nargs; |
|
2561 // If mirror or obj is dead, only null-path is taken. |
|
2562 if (stopped()) return true; |
|
2563 |
|
2564 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) |
|
2565 |
|
2566 // Now load the mirror's klass metaobject, and null-check it. |
|
2567 // Side-effects region with the control path if the klass is null. |
|
2568 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, |
|
2569 region, _prim_path); |
|
2570 // If kls is null, we have a primitive mirror. |
|
2571 phi->init_req(_prim_path, prim_return_value); |
|
2572 if (stopped()) { push_result(region, phi); return true; } |
|
2573 |
|
2574 Node* p; // handy temp |
|
2575 Node* null_ctl; |
|
2576 |
|
2577 // Now that we have the non-null klass, we can perform the real query. |
|
2578 // For constant classes, the query will constant-fold in LoadNode::Value. |
|
2579 Node* query_value = top(); |
|
2580 switch (id) { |
|
2581 case vmIntrinsics::_isInstance: |
|
2582 // nothing is an instance of a primitive type |
|
2583 query_value = gen_instanceof(obj, kls); |
|
2584 break; |
|
2585 |
|
2586 case vmIntrinsics::_getModifiers: |
|
2587 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc)); |
|
2588 query_value = make_load(NULL, p, TypeInt::INT, T_INT); |
|
2589 break; |
|
2590 |
|
2591 case vmIntrinsics::_isInterface: |
|
2592 // (To verify this code sequence, check the asserts in JVM_IsInterface.) |
|
2593 if (generate_interface_guard(kls, region) != NULL) |
|
2594 // A guard was added. If the guard is taken, it was an interface. |
|
2595 phi->add_req(intcon(1)); |
|
2596 // If we fall through, it's a plain class. |
|
2597 query_value = intcon(0); |
|
2598 break; |
|
2599 |
|
2600 case vmIntrinsics::_isArray: |
|
2601 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) |
|
2602 if (generate_array_guard(kls, region) != NULL) |
|
2603 // A guard was added. If the guard is taken, it was an array. |
|
2604 phi->add_req(intcon(1)); |
|
2605 // If we fall through, it's a plain class. |
|
2606 query_value = intcon(0); |
|
2607 break; |
|
2608 |
|
2609 case vmIntrinsics::_isPrimitive: |
|
2610 query_value = intcon(0); // "normal" path produces false |
|
2611 break; |
|
2612 |
|
2613 case vmIntrinsics::_getSuperclass: |
|
2614 // The rules here are somewhat unfortunate, but we can still do better |
|
2615 // with random logic than with a JNI call. |
|
2616 // Interfaces store null or Object as _super, but must report null. |
|
2617 // Arrays store an intermediate super as _super, but must report Object. |
|
2618 // Other types can report the actual _super. |
|
2619 // (To verify this code sequence, check the asserts in JVM_IsInterface.) |
|
2620 if (generate_interface_guard(kls, region) != NULL) |
|
2621 // A guard was added. If the guard is taken, it was an interface. |
|
2622 phi->add_req(null()); |
|
2623 if (generate_array_guard(kls, region) != NULL) |
|
2624 // A guard was added. If the guard is taken, it was an array. |
|
2625 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); |
|
2626 // If we fall through, it's a plain class. Get its _super. |
|
2627 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc)); |
|
2628 kls = _gvn.transform(new (C, 3) LoadKlassNode(0, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); |
|
2629 null_ctl = top(); |
|
2630 kls = null_check_oop(kls, &null_ctl); |
|
2631 if (null_ctl != top()) { |
|
2632 // If the guard is taken, Object.superClass is null (both klass and mirror). |
|
2633 region->add_req(null_ctl); |
|
2634 phi ->add_req(null()); |
|
2635 } |
|
2636 if (!stopped()) { |
|
2637 query_value = load_mirror_from_klass(kls); |
|
2638 } |
|
2639 break; |
|
2640 |
|
2641 case vmIntrinsics::_getComponentType: |
|
2642 if (generate_array_guard(kls, region) != NULL) { |
|
2643 // Be sure to pin the oop load to the guard edge just created: |
|
2644 Node* is_array_ctrl = region->in(region->req()-1); |
|
2645 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc)); |
|
2646 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); |
|
2647 phi->add_req(cmo); |
|
2648 } |
|
2649 query_value = null(); // non-array case is null |
|
2650 break; |
|
2651 |
|
2652 case vmIntrinsics::_getClassAccessFlags: |
|
2653 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); |
|
2654 query_value = make_load(NULL, p, TypeInt::INT, T_INT); |
|
2655 break; |
|
2656 |
|
2657 default: |
|
2658 ShouldNotReachHere(); |
|
2659 } |
|
2660 |
|
2661 // Fall-through is the normal case of a query to a real class. |
|
2662 phi->init_req(1, query_value); |
|
2663 region->init_req(1, control()); |
|
2664 |
|
2665 push_result(region, phi); |
|
2666 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2667 |
|
2668 return true; |
|
2669 } |
|
2670 |
|
2671 //--------------------------inline_native_subtype_check------------------------ |
|
2672 // This intrinsic takes the JNI calls out of the heart of |
|
2673 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. |
|
2674 bool LibraryCallKit::inline_native_subtype_check() { |
|
2675 int nargs = 1+1; // the Class mirror, plus the other class getting examined |
|
2676 |
|
2677 // Pull both arguments off the stack. |
|
2678 Node* args[2]; // two java.lang.Class mirrors: superc, subc |
|
2679 args[0] = argument(0); |
|
2680 args[1] = argument(1); |
|
2681 Node* klasses[2]; // corresponding Klasses: superk, subk |
|
2682 klasses[0] = klasses[1] = top(); |
|
2683 |
|
2684 enum { |
|
2685 // A full decision tree on {superc is prim, subc is prim}: |
|
2686 _prim_0_path = 1, // {P,N} => false |
|
2687 // {P,P} & superc!=subc => false |
|
2688 _prim_same_path, // {P,P} & superc==subc => true |
|
2689 _prim_1_path, // {N,P} => false |
|
2690 _ref_subtype_path, // {N,N} & subtype check wins => true |
|
2691 _both_ref_path, // {N,N} & subtype check loses => false |
|
2692 PATH_LIMIT |
|
2693 }; |
|
2694 |
|
2695 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
2696 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); |
|
2697 record_for_igvn(region); |
|
2698 |
|
2699 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads |
|
2700 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; |
|
2701 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); |
|
2702 |
|
2703 // First null-check both mirrors and load each mirror's klass metaobject. |
|
2704 int which_arg; |
|
2705 for (which_arg = 0; which_arg <= 1; which_arg++) { |
|
2706 Node* arg = args[which_arg]; |
|
2707 _sp += nargs; // set original stack for use by uncommon_trap |
|
2708 arg = do_null_check(arg, T_OBJECT); |
|
2709 _sp -= nargs; |
|
2710 if (stopped()) break; |
|
2711 args[which_arg] = _gvn.transform(arg); |
|
2712 |
|
2713 Node* p = basic_plus_adr(arg, class_klass_offset); |
|
2714 Node* kls = new (C, 3) LoadKlassNode(0, immutable_memory(), p, adr_type, kls_type); |
|
2715 klasses[which_arg] = _gvn.transform(kls); |
|
2716 } |
|
2717 |
|
2718 // Having loaded both klasses, test each for null. |
|
2719 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
|
2720 for (which_arg = 0; which_arg <= 1; which_arg++) { |
|
2721 Node* kls = klasses[which_arg]; |
|
2722 Node* null_ctl = top(); |
|
2723 _sp += nargs; // set original stack for use by uncommon_trap |
|
2724 kls = null_check_oop(kls, &null_ctl, never_see_null); |
|
2725 _sp -= nargs; |
|
2726 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); |
|
2727 region->init_req(prim_path, null_ctl); |
|
2728 if (stopped()) break; |
|
2729 klasses[which_arg] = kls; |
|
2730 } |
|
2731 |
|
2732 if (!stopped()) { |
|
2733 // now we have two reference types, in klasses[0..1] |
|
2734 Node* subk = klasses[1]; // the argument to isAssignableFrom |
|
2735 Node* superk = klasses[0]; // the receiver |
|
2736 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); |
|
2737 // now we have a successful reference subtype check |
|
2738 region->set_req(_ref_subtype_path, control()); |
|
2739 } |
|
2740 |
|
2741 // If both operands are primitive (both klasses null), then |
|
2742 // we must return true when they are identical primitives. |
|
2743 // It is convenient to test this after the first null klass check. |
|
2744 set_control(region->in(_prim_0_path)); // go back to first null check |
|
2745 if (!stopped()) { |
|
2746 // Since superc is primitive, make a guard for the superc==subc case. |
|
2747 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); |
|
2748 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); |
|
2749 generate_guard(bol_eq, region, PROB_FAIR); |
|
2750 if (region->req() == PATH_LIMIT+1) { |
|
2751 // A guard was added. If the added guard is taken, superc==subc. |
|
2752 region->swap_edges(PATH_LIMIT, _prim_same_path); |
|
2753 region->del_req(PATH_LIMIT); |
|
2754 } |
|
2755 region->set_req(_prim_0_path, control()); // Not equal after all. |
|
2756 } |
|
2757 |
|
2758 // these are the only paths that produce 'true': |
|
2759 phi->set_req(_prim_same_path, intcon(1)); |
|
2760 phi->set_req(_ref_subtype_path, intcon(1)); |
|
2761 |
|
2762 // pull together the cases: |
|
2763 assert(region->req() == PATH_LIMIT, "sane region"); |
|
2764 for (uint i = 1; i < region->req(); i++) { |
|
2765 Node* ctl = region->in(i); |
|
2766 if (ctl == NULL || ctl == top()) { |
|
2767 region->set_req(i, top()); |
|
2768 phi ->set_req(i, top()); |
|
2769 } else if (phi->in(i) == NULL) { |
|
2770 phi->set_req(i, intcon(0)); // all other paths produce 'false' |
|
2771 } |
|
2772 } |
|
2773 |
|
2774 set_control(_gvn.transform(region)); |
|
2775 push(_gvn.transform(phi)); |
|
2776 |
|
2777 return true; |
|
2778 } |
|
2779 |
|
2780 //---------------------generate_array_guard_common------------------------ |
|
2781 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, |
|
2782 bool obj_array, bool not_array) { |
|
2783 // If obj_array/non_array==false/false: |
|
2784 // Branch around if the given klass is in fact an array (either obj or prim). |
|
2785 // If obj_array/non_array==false/true: |
|
2786 // Branch around if the given klass is not an array klass of any kind. |
|
2787 // If obj_array/non_array==true/true: |
|
2788 // Branch around if the kls is not an oop array (kls is int[], String, etc.) |
|
2789 // If obj_array/non_array==true/false: |
|
2790 // Branch around if the kls is an oop array (Object[] or subtype) |
|
2791 // |
|
2792 // Like generate_guard, adds a new path onto the region. |
|
2793 jint layout_con = 0; |
|
2794 Node* layout_val = get_layout_helper(kls, layout_con); |
|
2795 if (layout_val == NULL) { |
|
2796 bool query = (obj_array |
|
2797 ? Klass::layout_helper_is_objArray(layout_con) |
|
2798 : Klass::layout_helper_is_javaArray(layout_con)); |
|
2799 if (query == not_array) { |
|
2800 return NULL; // never a branch |
|
2801 } else { // always a branch |
|
2802 Node* always_branch = control(); |
|
2803 if (region != NULL) |
|
2804 region->add_req(always_branch); |
|
2805 set_control(top()); |
|
2806 return always_branch; |
|
2807 } |
|
2808 } |
|
2809 // Now test the correct condition. |
|
2810 jint nval = (obj_array |
|
2811 ? ((jint)Klass::_lh_array_tag_type_value |
|
2812 << Klass::_lh_array_tag_shift) |
|
2813 : Klass::_lh_neutral_value); |
|
2814 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); |
|
2815 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array |
|
2816 // invert the test if we are looking for a non-array |
|
2817 if (not_array) btest = BoolTest(btest).negate(); |
|
2818 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); |
|
2819 return generate_fair_guard(bol, region); |
|
2820 } |
|
2821 |
|
2822 |
|
2823 //-----------------------inline_native_newArray-------------------------- |
|
2824 bool LibraryCallKit::inline_native_newArray() { |
|
2825 int nargs = 2; |
|
2826 Node* mirror = argument(0); |
|
2827 Node* count_val = argument(1); |
|
2828 |
|
2829 _sp += nargs; // set original stack for use by uncommon_trap |
|
2830 mirror = do_null_check(mirror, T_OBJECT); |
|
2831 _sp -= nargs; |
|
2832 |
|
2833 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; |
|
2834 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
2835 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, |
|
2836 TypeInstPtr::NOTNULL); |
|
2837 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); |
|
2838 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, |
|
2839 TypePtr::BOTTOM); |
|
2840 |
|
2841 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
|
2842 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, |
|
2843 nargs, |
|
2844 result_reg, _slow_path); |
|
2845 Node* normal_ctl = control(); |
|
2846 Node* no_array_ctl = result_reg->in(_slow_path); |
|
2847 |
|
2848 // Generate code for the slow case. We make a call to newArray(). |
|
2849 set_control(no_array_ctl); |
|
2850 if (!stopped()) { |
|
2851 // Either the input type is void.class, or else the |
|
2852 // array klass has not yet been cached. Either the |
|
2853 // ensuing call will throw an exception, or else it |
|
2854 // will cache the array klass for next time. |
|
2855 PreserveJVMState pjvms(this); |
|
2856 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); |
|
2857 Node* slow_result = set_results_for_java_call(slow_call); |
|
2858 // this->control() comes from set_results_for_java_call |
|
2859 result_reg->set_req(_slow_path, control()); |
|
2860 result_val->set_req(_slow_path, slow_result); |
|
2861 result_io ->set_req(_slow_path, i_o()); |
|
2862 result_mem->set_req(_slow_path, reset_memory()); |
|
2863 } |
|
2864 |
|
2865 set_control(normal_ctl); |
|
2866 if (!stopped()) { |
|
2867 // Normal case: The array type has been cached in the java.lang.Class. |
|
2868 // The following call works fine even if the array type is polymorphic. |
|
2869 // It could be a dynamic mix of int[], boolean[], Object[], etc. |
|
2870 _sp += nargs; // set original stack for use by uncommon_trap |
|
2871 Node* obj = new_array(klass_node, count_val); |
|
2872 _sp -= nargs; |
|
2873 result_reg->init_req(_normal_path, control()); |
|
2874 result_val->init_req(_normal_path, obj); |
|
2875 result_io ->init_req(_normal_path, i_o()); |
|
2876 result_mem->init_req(_normal_path, reset_memory()); |
|
2877 } |
|
2878 |
|
2879 // Return the combined state. |
|
2880 set_i_o( _gvn.transform(result_io) ); |
|
2881 set_all_memory( _gvn.transform(result_mem) ); |
|
2882 push_result(result_reg, result_val); |
|
2883 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2884 |
|
2885 return true; |
|
2886 } |
|
2887 |
|
2888 //----------------------inline_native_getLength-------------------------- |
|
2889 bool LibraryCallKit::inline_native_getLength() { |
|
2890 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
2891 |
|
2892 int nargs = 1; |
|
2893 Node* array = argument(0); |
|
2894 |
|
2895 _sp += nargs; // set original stack for use by uncommon_trap |
|
2896 array = do_null_check(array, T_OBJECT); |
|
2897 _sp -= nargs; |
|
2898 |
|
2899 // If array is dead, only null-path is taken. |
|
2900 if (stopped()) return true; |
|
2901 |
|
2902 // Deoptimize if it is a non-array. |
|
2903 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); |
|
2904 |
|
2905 if (non_array != NULL) { |
|
2906 PreserveJVMState pjvms(this); |
|
2907 set_control(non_array); |
|
2908 _sp += nargs; // push the arguments back on the stack |
|
2909 uncommon_trap(Deoptimization::Reason_intrinsic, |
|
2910 Deoptimization::Action_maybe_recompile); |
|
2911 } |
|
2912 |
|
2913 // If control is dead, only non-array-path is taken. |
|
2914 if (stopped()) return true; |
|
2915 |
|
2916 // The works fine even if the array type is polymorphic. |
|
2917 // It could be a dynamic mix of int[], boolean[], Object[], etc. |
|
2918 push( load_array_length(array) ); |
|
2919 |
|
2920 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2921 |
|
2922 return true; |
|
2923 } |
|
2924 |
|
2925 //------------------------inline_array_copyOf---------------------------- |
|
2926 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { |
|
2927 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
2928 |
|
2929 // Restore the stack and pop off the arguments. |
|
2930 int nargs = 3 + (is_copyOfRange? 1: 0); |
|
2931 Node* original = argument(0); |
|
2932 Node* start = is_copyOfRange? argument(1): intcon(0); |
|
2933 Node* end = is_copyOfRange? argument(2): argument(1); |
|
2934 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); |
|
2935 |
|
2936 _sp += nargs; // set original stack for use by uncommon_trap |
|
2937 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); |
|
2938 original = do_null_check(original, T_OBJECT); |
|
2939 _sp -= nargs; |
|
2940 |
|
2941 // Check if a null path was taken unconditionally. |
|
2942 if (stopped()) return true; |
|
2943 |
|
2944 Node* orig_length = load_array_length(original); |
|
2945 |
|
2946 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs, |
|
2947 NULL, 0); |
|
2948 _sp += nargs; // set original stack for use by uncommon_trap |
|
2949 klass_node = do_null_check(klass_node, T_OBJECT); |
|
2950 _sp -= nargs; |
|
2951 |
|
2952 RegionNode* bailout = new (C, 1) RegionNode(1); |
|
2953 record_for_igvn(bailout); |
|
2954 |
|
2955 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. |
|
2956 // Bail out if that is so. |
|
2957 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); |
|
2958 if (not_objArray != NULL) { |
|
2959 // Improve the klass node's type from the new optimistic assumption: |
|
2960 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); |
|
2961 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); |
|
2962 Node* cast = new (C, 2) CastPPNode(klass_node, akls); |
|
2963 cast->init_req(0, control()); |
|
2964 klass_node = _gvn.transform(cast); |
|
2965 } |
|
2966 |
|
2967 // Bail out if either start or end is negative. |
|
2968 generate_negative_guard(start, bailout, &start); |
|
2969 generate_negative_guard(end, bailout, &end); |
|
2970 |
|
2971 Node* length = end; |
|
2972 if (_gvn.type(start) != TypeInt::ZERO) { |
|
2973 length = _gvn.transform( new (C, 3) SubINode(end, start) ); |
|
2974 } |
|
2975 |
|
2976 // Bail out if length is negative. |
|
2977 // ...Not needed, since the new_array will throw the right exception. |
|
2978 //generate_negative_guard(length, bailout, &length); |
|
2979 |
|
2980 if (bailout->req() > 1) { |
|
2981 PreserveJVMState pjvms(this); |
|
2982 set_control( _gvn.transform(bailout) ); |
|
2983 _sp += nargs; // push the arguments back on the stack |
|
2984 uncommon_trap(Deoptimization::Reason_intrinsic, |
|
2985 Deoptimization::Action_maybe_recompile); |
|
2986 } |
|
2987 |
|
2988 if (!stopped()) { |
|
2989 // How many elements will we copy from the original? |
|
2990 // The answer is MinI(orig_length - start, length). |
|
2991 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); |
|
2992 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); |
|
2993 |
|
2994 _sp += nargs; // set original stack for use by uncommon_trap |
|
2995 Node* newcopy = new_array(klass_node, length); |
|
2996 _sp -= nargs; |
|
2997 |
|
2998 // Generate a direct call to the right arraycopy function(s). |
|
2999 // We know the copy is disjoint but we might not know if the |
|
3000 // oop stores need checking. |
|
3001 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). |
|
3002 // This will fail a store-check if x contains any non-nulls. |
|
3003 bool disjoint_bases = true; |
|
3004 bool length_never_negative = true; |
|
3005 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, |
|
3006 original, start, newcopy, intcon(0), moved, |
|
3007 nargs, disjoint_bases, length_never_negative); |
|
3008 |
|
3009 push(newcopy); |
|
3010 } |
|
3011 |
|
3012 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
3013 |
|
3014 return true; |
|
3015 } |
|
3016 |
|
3017 |
|
3018 //----------------------generate_virtual_guard--------------------------- |
|
3019 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. |
|
3020 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, |
|
3021 RegionNode* slow_region) { |
|
3022 ciMethod* method = callee(); |
|
3023 int vtable_index = method->vtable_index(); |
|
3024 // Get the methodOop out of the appropriate vtable entry. |
|
3025 int entry_offset = (instanceKlass::vtable_start_offset() + |
|
3026 vtable_index*vtableEntry::size()) * wordSize + |
|
3027 vtableEntry::method_offset_in_bytes(); |
|
3028 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); |
|
3029 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); |
|
3030 |
|
3031 // Compare the target method with the expected method (e.g., Object.hashCode). |
|
3032 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); |
|
3033 |
|
3034 Node* native_call = makecon(native_call_addr); |
|
3035 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); |
|
3036 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); |
|
3037 |
|
3038 return generate_slow_guard(test_native, slow_region); |
|
3039 } |
|
3040 |
|
3041 //-----------------------generate_method_call---------------------------- |
|
3042 // Use generate_method_call to make a slow-call to the real |
|
3043 // method if the fast path fails. An alternative would be to |
|
3044 // use a stub like OptoRuntime::slow_arraycopy_Java. |
|
3045 // This only works for expanding the current library call, |
|
3046 // not another intrinsic. (E.g., don't use this for making an |
|
3047 // arraycopy call inside of the copyOf intrinsic.) |
|
3048 CallJavaNode* |
|
3049 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { |
|
3050 // When compiling the intrinsic method itself, do not use this technique. |
|
3051 guarantee(callee() != C->method(), "cannot make slow-call to self"); |
|
3052 |
|
3053 ciMethod* method = callee(); |
|
3054 // ensure the JVMS we have will be correct for this call |
|
3055 guarantee(method_id == method->intrinsic_id(), "must match"); |
|
3056 |
|
3057 const TypeFunc* tf = TypeFunc::make(method); |
|
3058 int tfdc = tf->domain()->cnt(); |
|
3059 CallJavaNode* slow_call; |
|
3060 if (is_static) { |
|
3061 assert(!is_virtual, ""); |
|
3062 slow_call = new(C, tfdc) CallStaticJavaNode(tf, |
|
3063 SharedRuntime::get_resolve_static_call_stub(), |
|
3064 method, bci()); |
|
3065 } else if (is_virtual) { |
|
3066 null_check_receiver(method); |
|
3067 int vtable_index = methodOopDesc::invalid_vtable_index; |
|
3068 if (UseInlineCaches) { |
|
3069 // Suppress the vtable call |
|
3070 } else { |
|
3071 // hashCode and clone are not a miranda methods, |
|
3072 // so the vtable index is fixed. |
|
3073 // No need to use the linkResolver to get it. |
|
3074 vtable_index = method->vtable_index(); |
|
3075 } |
|
3076 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, |
|
3077 SharedRuntime::get_resolve_virtual_call_stub(), |
|
3078 method, vtable_index, bci()); |
|
3079 } else { // neither virtual nor static: opt_virtual |
|
3080 null_check_receiver(method); |
|
3081 slow_call = new(C, tfdc) CallStaticJavaNode(tf, |
|
3082 SharedRuntime::get_resolve_opt_virtual_call_stub(), |
|
3083 method, bci()); |
|
3084 slow_call->set_optimized_virtual(true); |
|
3085 } |
|
3086 set_arguments_for_java_call(slow_call); |
|
3087 set_edges_for_java_call(slow_call); |
|
3088 return slow_call; |
|
3089 } |
|
3090 |
|
3091 |
|
3092 //------------------------------inline_native_hashcode-------------------- |
|
3093 // Build special case code for calls to hashCode on an object. |
|
3094 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { |
|
3095 assert(is_static == callee()->is_static(), "correct intrinsic selection"); |
|
3096 assert(!(is_virtual && is_static), "either virtual, special, or static"); |
|
3097 |
|
3098 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; |
|
3099 |
|
3100 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
3101 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, |
|
3102 TypeInt::INT); |
|
3103 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); |
|
3104 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, |
|
3105 TypePtr::BOTTOM); |
|
3106 Node* obj = NULL; |
|
3107 if (!is_static) { |
|
3108 // Check for hashing null object |
|
3109 obj = null_check_receiver(callee()); |
|
3110 if (stopped()) return true; // unconditionally null |
|
3111 result_reg->init_req(_null_path, top()); |
|
3112 result_val->init_req(_null_path, top()); |
|
3113 } else { |
|
3114 // Do a null check, and return zero if null. |
|
3115 // System.identityHashCode(null) == 0 |
|
3116 obj = argument(0); |
|
3117 Node* null_ctl = top(); |
|
3118 obj = null_check_oop(obj, &null_ctl); |
|
3119 result_reg->init_req(_null_path, null_ctl); |
|
3120 result_val->init_req(_null_path, _gvn.intcon(0)); |
|
3121 } |
|
3122 |
|
3123 // Unconditionally null? Then return right away. |
|
3124 if (stopped()) { |
|
3125 set_control( result_reg->in(_null_path) ); |
|
3126 if (!stopped()) |
|
3127 push( result_val ->in(_null_path) ); |
|
3128 return true; |
|
3129 } |
|
3130 |
|
3131 // After null check, get the object's klass. |
|
3132 Node* obj_klass = load_object_klass(obj); |
|
3133 |
|
3134 // This call may be virtual (invokevirtual) or bound (invokespecial). |
|
3135 // For each case we generate slightly different code. |
|
3136 |
|
3137 // We only go to the fast case code if we pass a number of guards. The |
|
3138 // paths which do not pass are accumulated in the slow_region. |
|
3139 RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
3140 record_for_igvn(slow_region); |
|
3141 |
|
3142 // If this is a virtual call, we generate a funny guard. We pull out |
|
3143 // the vtable entry corresponding to hashCode() from the target object. |
|
3144 // If the target method which we are calling happens to be the native |
|
3145 // Object hashCode() method, we pass the guard. We do not need this |
|
3146 // guard for non-virtual calls -- the caller is known to be the native |
|
3147 // Object hashCode(). |
|
3148 if (is_virtual) { |
|
3149 generate_virtual_guard(obj_klass, slow_region); |
|
3150 } |
|
3151 |
|
3152 // Get the header out of the object, use LoadMarkNode when available |
|
3153 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); |
|
3154 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS); |
|
3155 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) ); |
|
3156 |
|
3157 // Test the header to see if it is unlocked. |
|
3158 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); |
|
3159 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); |
|
3160 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); |
|
3161 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); |
|
3162 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); |
|
3163 |
|
3164 generate_slow_guard(test_unlocked, slow_region); |
|
3165 |
|
3166 // Get the hash value and check to see that it has been properly assigned. |
|
3167 // We depend on hash_mask being at most 32 bits and avoid the use of |
|
3168 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit |
|
3169 // vm: see markOop.hpp. |
|
3170 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); |
|
3171 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); |
|
3172 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); |
|
3173 // This hack lets the hash bits live anywhere in the mark object now, as long |
|
3174 // as the shift drops the relevent bits into the low 32 bits. Note that |
|
3175 // Java spec says that HashCode is an int so there's no point in capturing |
|
3176 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). |
|
3177 hshifted_header = ConvX2I(hshifted_header); |
|
3178 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); |
|
3179 |
|
3180 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); |
|
3181 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); |
|
3182 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); |
|
3183 |
|
3184 generate_slow_guard(test_assigned, slow_region); |
|
3185 |
|
3186 Node* init_mem = reset_memory(); |
|
3187 // fill in the rest of the null path: |
|
3188 result_io ->init_req(_null_path, i_o()); |
|
3189 result_mem->init_req(_null_path, init_mem); |
|
3190 |
|
3191 result_val->init_req(_fast_path, hash_val); |
|
3192 result_reg->init_req(_fast_path, control()); |
|
3193 result_io ->init_req(_fast_path, i_o()); |
|
3194 result_mem->init_req(_fast_path, init_mem); |
|
3195 |
|
3196 // Generate code for the slow case. We make a call to hashCode(). |
|
3197 set_control(_gvn.transform(slow_region)); |
|
3198 if (!stopped()) { |
|
3199 // No need for PreserveJVMState, because we're using up the present state. |
|
3200 set_all_memory(init_mem); |
|
3201 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; |
|
3202 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; |
|
3203 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); |
|
3204 Node* slow_result = set_results_for_java_call(slow_call); |
|
3205 // this->control() comes from set_results_for_java_call |
|
3206 result_reg->init_req(_slow_path, control()); |
|
3207 result_val->init_req(_slow_path, slow_result); |
|
3208 result_io ->set_req(_slow_path, i_o()); |
|
3209 result_mem ->set_req(_slow_path, reset_memory()); |
|
3210 } |
|
3211 |
|
3212 // Return the combined state. |
|
3213 set_i_o( _gvn.transform(result_io) ); |
|
3214 set_all_memory( _gvn.transform(result_mem) ); |
|
3215 push_result(result_reg, result_val); |
|
3216 |
|
3217 return true; |
|
3218 } |
|
3219 |
|
3220 //---------------------------inline_native_getClass---------------------------- |
|
3221 // Build special case code for calls to hashCode on an object. |
|
3222 bool LibraryCallKit::inline_native_getClass() { |
|
3223 Node* obj = null_check_receiver(callee()); |
|
3224 if (stopped()) return true; |
|
3225 push( load_mirror_from_klass(load_object_klass(obj)) ); |
|
3226 return true; |
|
3227 } |
|
3228 |
|
3229 //-----------------inline_native_Reflection_getCallerClass--------------------- |
|
3230 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. |
|
3231 // |
|
3232 // NOTE that this code must perform the same logic as |
|
3233 // vframeStream::security_get_caller_frame in that it must skip |
|
3234 // Method.invoke() and auxiliary frames. |
|
3235 |
|
3236 |
|
3237 |
|
3238 |
|
3239 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { |
|
3240 ciMethod* method = callee(); |
|
3241 |
|
3242 #ifndef PRODUCT |
|
3243 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3244 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); |
|
3245 } |
|
3246 #endif |
|
3247 |
|
3248 debug_only(int saved_sp = _sp); |
|
3249 |
|
3250 // Argument words: (int depth) |
|
3251 int nargs = 1; |
|
3252 |
|
3253 _sp += nargs; |
|
3254 Node* caller_depth_node = pop(); |
|
3255 |
|
3256 assert(saved_sp == _sp, "must have correct argument count"); |
|
3257 |
|
3258 // The depth value must be a constant in order for the runtime call |
|
3259 // to be eliminated. |
|
3260 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); |
|
3261 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { |
|
3262 #ifndef PRODUCT |
|
3263 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3264 tty->print_cr(" Bailing out because caller depth was not a constant"); |
|
3265 } |
|
3266 #endif |
|
3267 return false; |
|
3268 } |
|
3269 // Note that the JVM state at this point does not include the |
|
3270 // getCallerClass() frame which we are trying to inline. The |
|
3271 // semantics of getCallerClass(), however, are that the "first" |
|
3272 // frame is the getCallerClass() frame, so we subtract one from the |
|
3273 // requested depth before continuing. We don't inline requests of |
|
3274 // getCallerClass(0). |
|
3275 int caller_depth = caller_depth_type->get_con() - 1; |
|
3276 if (caller_depth < 0) { |
|
3277 #ifndef PRODUCT |
|
3278 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3279 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); |
|
3280 } |
|
3281 #endif |
|
3282 return false; |
|
3283 } |
|
3284 |
|
3285 if (!jvms()->has_method()) { |
|
3286 #ifndef PRODUCT |
|
3287 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3288 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); |
|
3289 } |
|
3290 #endif |
|
3291 return false; |
|
3292 } |
|
3293 int _depth = jvms()->depth(); // cache call chain depth |
|
3294 |
|
3295 // Walk back up the JVM state to find the caller at the required |
|
3296 // depth. NOTE that this code must perform the same logic as |
|
3297 // vframeStream::security_get_caller_frame in that it must skip |
|
3298 // Method.invoke() and auxiliary frames. Note also that depth is |
|
3299 // 1-based (1 is the bottom of the inlining). |
|
3300 int inlining_depth = _depth; |
|
3301 JVMState* caller_jvms = NULL; |
|
3302 |
|
3303 if (inlining_depth > 0) { |
|
3304 caller_jvms = jvms(); |
|
3305 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); |
|
3306 do { |
|
3307 // The following if-tests should be performed in this order |
|
3308 if (is_method_invoke_or_aux_frame(caller_jvms)) { |
|
3309 // Skip a Method.invoke() or auxiliary frame |
|
3310 } else if (caller_depth > 0) { |
|
3311 // Skip real frame |
|
3312 --caller_depth; |
|
3313 } else { |
|
3314 // We're done: reached desired caller after skipping. |
|
3315 break; |
|
3316 } |
|
3317 caller_jvms = caller_jvms->caller(); |
|
3318 --inlining_depth; |
|
3319 } while (inlining_depth > 0); |
|
3320 } |
|
3321 |
|
3322 if (inlining_depth == 0) { |
|
3323 #ifndef PRODUCT |
|
3324 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3325 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); |
|
3326 tty->print_cr(" JVM state at this point:"); |
|
3327 for (int i = _depth; i >= 1; i--) { |
|
3328 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); |
|
3329 } |
|
3330 } |
|
3331 #endif |
|
3332 return false; // Reached end of inlining |
|
3333 } |
|
3334 |
|
3335 // Acquire method holder as java.lang.Class |
|
3336 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); |
|
3337 ciInstance* caller_mirror = caller_klass->java_mirror(); |
|
3338 // Push this as a constant |
|
3339 push(makecon(TypeInstPtr::make(caller_mirror))); |
|
3340 #ifndef PRODUCT |
|
3341 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3342 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); |
|
3343 tty->print_cr(" JVM state at this point:"); |
|
3344 for (int i = _depth; i >= 1; i--) { |
|
3345 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); |
|
3346 } |
|
3347 } |
|
3348 #endif |
|
3349 return true; |
|
3350 } |
|
3351 |
|
3352 // Helper routine for above |
|
3353 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { |
|
3354 // Is this the Method.invoke method itself? |
|
3355 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke) |
|
3356 return true; |
|
3357 |
|
3358 // Is this a helper, defined somewhere underneath MethodAccessorImpl. |
|
3359 ciKlass* k = jvms->method()->holder(); |
|
3360 if (k->is_instance_klass()) { |
|
3361 ciInstanceKlass* ik = k->as_instance_klass(); |
|
3362 for (; ik != NULL; ik = ik->super()) { |
|
3363 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && |
|
3364 ik == env()->find_system_klass(ik->name())) { |
|
3365 return true; |
|
3366 } |
|
3367 } |
|
3368 } |
|
3369 |
|
3370 return false; |
|
3371 } |
|
3372 |
|
3373 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by |
|
3374 // inline_native_AtomicLong_attemptUpdate() but it has no way of |
|
3375 // computing it since there is no lookup field by name function in the |
|
3376 // CI interface. This is computed and set by inline_native_AtomicLong_get(). |
|
3377 // Using a static variable here is safe even if we have multiple compilation |
|
3378 // threads because the offset is constant. At worst the same offset will be |
|
3379 // computed and stored multiple |
|
3380 |
|
3381 bool LibraryCallKit::inline_native_AtomicLong_get() { |
|
3382 // Restore the stack and pop off the argument |
|
3383 _sp+=1; |
|
3384 Node *obj = pop(); |
|
3385 |
|
3386 // get the offset of the "value" field. Since the CI interfaces |
|
3387 // does not provide a way to look up a field by name, we scan the bytecodes |
|
3388 // to get the field index. We expect the first 2 instructions of the method |
|
3389 // to be: |
|
3390 // 0 aload_0 |
|
3391 // 1 getfield "value" |
|
3392 ciMethod* method = callee(); |
|
3393 if (value_field_offset == -1) |
|
3394 { |
|
3395 ciField* value_field; |
|
3396 ciBytecodeStream iter(method); |
|
3397 Bytecodes::Code bc = iter.next(); |
|
3398 |
|
3399 if ((bc != Bytecodes::_aload_0) && |
|
3400 ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) |
|
3401 return false; |
|
3402 bc = iter.next(); |
|
3403 if (bc != Bytecodes::_getfield) |
|
3404 return false; |
|
3405 bool ignore; |
|
3406 value_field = iter.get_field(ignore); |
|
3407 value_field_offset = value_field->offset_in_bytes(); |
|
3408 } |
|
3409 |
|
3410 // Null check without removing any arguments. |
|
3411 _sp++; |
|
3412 obj = do_null_check(obj, T_OBJECT); |
|
3413 _sp--; |
|
3414 // Check for locking null object |
|
3415 if (stopped()) return true; |
|
3416 |
|
3417 Node *adr = basic_plus_adr(obj, obj, value_field_offset); |
|
3418 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); |
|
3419 int alias_idx = C->get_alias_index(adr_type); |
|
3420 |
|
3421 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); |
|
3422 |
|
3423 push_pair(result); |
|
3424 |
|
3425 return true; |
|
3426 } |
|
3427 |
|
3428 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { |
|
3429 // Restore the stack and pop off the arguments |
|
3430 _sp+=5; |
|
3431 Node *newVal = pop_pair(); |
|
3432 Node *oldVal = pop_pair(); |
|
3433 Node *obj = pop(); |
|
3434 |
|
3435 // we need the offset of the "value" field which was computed when |
|
3436 // inlining the get() method. Give up if we don't have it. |
|
3437 if (value_field_offset == -1) |
|
3438 return false; |
|
3439 |
|
3440 // Null check without removing any arguments. |
|
3441 _sp+=5; |
|
3442 obj = do_null_check(obj, T_OBJECT); |
|
3443 _sp-=5; |
|
3444 // Check for locking null object |
|
3445 if (stopped()) return true; |
|
3446 |
|
3447 Node *adr = basic_plus_adr(obj, obj, value_field_offset); |
|
3448 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); |
|
3449 int alias_idx = C->get_alias_index(adr_type); |
|
3450 |
|
3451 Node *result = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); |
|
3452 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(result)); |
|
3453 set_memory(store_proj, alias_idx); |
|
3454 |
|
3455 push(result); |
|
3456 return true; |
|
3457 } |
|
3458 |
|
3459 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { |
|
3460 // restore the arguments |
|
3461 _sp += arg_size(); |
|
3462 |
|
3463 switch (id) { |
|
3464 case vmIntrinsics::_floatToRawIntBits: |
|
3465 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); |
|
3466 break; |
|
3467 |
|
3468 case vmIntrinsics::_intBitsToFloat: |
|
3469 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); |
|
3470 break; |
|
3471 |
|
3472 case vmIntrinsics::_doubleToRawLongBits: |
|
3473 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); |
|
3474 break; |
|
3475 |
|
3476 case vmIntrinsics::_longBitsToDouble: |
|
3477 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); |
|
3478 break; |
|
3479 |
|
3480 case vmIntrinsics::_doubleToLongBits: { |
|
3481 Node* value = pop_pair(); |
|
3482 |
|
3483 // two paths (plus control) merge in a wood |
|
3484 RegionNode *r = new (C, 3) RegionNode(3); |
|
3485 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); |
|
3486 |
|
3487 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); |
|
3488 // Build the boolean node |
|
3489 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); |
|
3490 |
|
3491 // Branch either way. |
|
3492 // NaN case is less traveled, which makes all the difference. |
|
3493 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
|
3494 Node *opt_isnan = _gvn.transform(ifisnan); |
|
3495 assert( opt_isnan->is_If(), "Expect an IfNode"); |
|
3496 IfNode *opt_ifisnan = (IfNode*)opt_isnan; |
|
3497 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); |
|
3498 |
|
3499 set_control(iftrue); |
|
3500 |
|
3501 static const jlong nan_bits = CONST64(0x7ff8000000000000); |
|
3502 Node *slow_result = longcon(nan_bits); // return NaN |
|
3503 phi->init_req(1, _gvn.transform( slow_result )); |
|
3504 r->init_req(1, iftrue); |
|
3505 |
|
3506 // Else fall through |
|
3507 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); |
|
3508 set_control(iffalse); |
|
3509 |
|
3510 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); |
|
3511 r->init_req(2, iffalse); |
|
3512 |
|
3513 // Post merge |
|
3514 set_control(_gvn.transform(r)); |
|
3515 record_for_igvn(r); |
|
3516 |
|
3517 Node* result = _gvn.transform(phi); |
|
3518 assert(result->bottom_type()->isa_long(), "must be"); |
|
3519 push_pair(result); |
|
3520 |
|
3521 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
3522 |
|
3523 break; |
|
3524 } |
|
3525 |
|
3526 case vmIntrinsics::_floatToIntBits: { |
|
3527 Node* value = pop(); |
|
3528 |
|
3529 // two paths (plus control) merge in a wood |
|
3530 RegionNode *r = new (C, 3) RegionNode(3); |
|
3531 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); |
|
3532 |
|
3533 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); |
|
3534 // Build the boolean node |
|
3535 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); |
|
3536 |
|
3537 // Branch either way. |
|
3538 // NaN case is less traveled, which makes all the difference. |
|
3539 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
|
3540 Node *opt_isnan = _gvn.transform(ifisnan); |
|
3541 assert( opt_isnan->is_If(), "Expect an IfNode"); |
|
3542 IfNode *opt_ifisnan = (IfNode*)opt_isnan; |
|
3543 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); |
|
3544 |
|
3545 set_control(iftrue); |
|
3546 |
|
3547 static const jint nan_bits = 0x7fc00000; |
|
3548 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN |
|
3549 phi->init_req(1, _gvn.transform( slow_result )); |
|
3550 r->init_req(1, iftrue); |
|
3551 |
|
3552 // Else fall through |
|
3553 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); |
|
3554 set_control(iffalse); |
|
3555 |
|
3556 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); |
|
3557 r->init_req(2, iffalse); |
|
3558 |
|
3559 // Post merge |
|
3560 set_control(_gvn.transform(r)); |
|
3561 record_for_igvn(r); |
|
3562 |
|
3563 Node* result = _gvn.transform(phi); |
|
3564 assert(result->bottom_type()->isa_int(), "must be"); |
|
3565 push(result); |
|
3566 |
|
3567 C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
3568 |
|
3569 break; |
|
3570 } |
|
3571 |
|
3572 default: |
|
3573 ShouldNotReachHere(); |
|
3574 } |
|
3575 |
|
3576 return true; |
|
3577 } |
|
3578 |
|
3579 #ifdef _LP64 |
|
3580 #define XTOP ,top() /*additional argument*/ |
|
3581 #else //_LP64 |
|
3582 #define XTOP /*no additional argument*/ |
|
3583 #endif //_LP64 |
|
3584 |
|
3585 //----------------------inline_unsafe_copyMemory------------------------- |
|
3586 bool LibraryCallKit::inline_unsafe_copyMemory() { |
|
3587 if (callee()->is_static()) return false; // caller must have the capability! |
|
3588 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) |
|
3589 assert(signature()->size() == nargs-1, "copy has 5 arguments"); |
|
3590 null_check_receiver(callee()); // check then ignore argument(0) |
|
3591 if (stopped()) return true; |
|
3592 |
|
3593 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
3594 |
|
3595 Node* src_ptr = argument(1); |
|
3596 Node* src_off = ConvL2X(argument(2)); |
|
3597 assert(argument(3)->is_top(), "2nd half of long"); |
|
3598 Node* dst_ptr = argument(4); |
|
3599 Node* dst_off = ConvL2X(argument(5)); |
|
3600 assert(argument(6)->is_top(), "2nd half of long"); |
|
3601 Node* size = ConvL2X(argument(7)); |
|
3602 assert(argument(8)->is_top(), "2nd half of long"); |
|
3603 |
|
3604 assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
|
3605 "fieldOffset must be byte-scaled"); |
|
3606 |
|
3607 Node* src = make_unsafe_address(src_ptr, src_off); |
|
3608 Node* dst = make_unsafe_address(dst_ptr, dst_off); |
|
3609 |
|
3610 // Conservatively insert a memory barrier on all memory slices. |
|
3611 // Do not let writes of the copy source or destination float below the copy. |
|
3612 insert_mem_bar(Op_MemBarCPUOrder); |
|
3613 |
|
3614 // Call it. Note that the length argument is not scaled. |
|
3615 make_runtime_call(RC_LEAF|RC_NO_FP, |
|
3616 OptoRuntime::fast_arraycopy_Type(), |
|
3617 StubRoutines::unsafe_arraycopy(), |
|
3618 "unsafe_arraycopy", |
|
3619 TypeRawPtr::BOTTOM, |
|
3620 src, dst, size XTOP); |
|
3621 |
|
3622 // Do not let reads of the copy destination float above the copy. |
|
3623 insert_mem_bar(Op_MemBarCPUOrder); |
|
3624 |
|
3625 return true; |
|
3626 } |
|
3627 |
|
3628 |
|
3629 //------------------------inline_native_clone---------------------------- |
|
3630 // Here are the simple edge cases: |
|
3631 // null receiver => normal trap |
|
3632 // virtual and clone was overridden => slow path to out-of-line clone |
|
3633 // not cloneable or finalizer => slow path to out-of-line Object.clone |
|
3634 // |
|
3635 // The general case has two steps, allocation and copying. |
|
3636 // Allocation has two cases, and uses GraphKit::new_instance or new_array. |
|
3637 // |
|
3638 // Copying also has two cases, oop arrays and everything else. |
|
3639 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). |
|
3640 // Everything else uses the tight inline loop supplied by CopyArrayNode. |
|
3641 // |
|
3642 // These steps fold up nicely if and when the cloned object's klass |
|
3643 // can be sharply typed as an object array, a type array, or an instance. |
|
3644 // |
|
3645 bool LibraryCallKit::inline_native_clone(bool is_virtual) { |
|
3646 int nargs = 1; |
|
3647 Node* obj = null_check_receiver(callee()); |
|
3648 if (stopped()) return true; |
|
3649 Node* obj_klass = load_object_klass(obj); |
|
3650 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); |
|
3651 const TypeOopPtr* toop = ((tklass != NULL) |
|
3652 ? tklass->as_instance_type() |
|
3653 : TypeInstPtr::NOTNULL); |
|
3654 |
|
3655 // Conservatively insert a memory barrier on all memory slices. |
|
3656 // Do not let writes into the original float below the clone. |
|
3657 insert_mem_bar(Op_MemBarCPUOrder); |
|
3658 |
|
3659 // paths into result_reg: |
|
3660 enum { |
|
3661 _slow_path = 1, // out-of-line call to clone method (virtual or not) |
|
3662 _objArray_path, // plain allocation, plus arrayof_oop_arraycopy |
|
3663 _fast_path, // plain allocation, plus a CopyArray operation |
|
3664 PATH_LIMIT |
|
3665 }; |
|
3666 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
3667 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, |
|
3668 TypeInstPtr::NOTNULL); |
|
3669 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); |
|
3670 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, |
|
3671 TypePtr::BOTTOM); |
|
3672 record_for_igvn(result_reg); |
|
3673 |
|
3674 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; |
|
3675 int raw_adr_idx = Compile::AliasIdxRaw; |
|
3676 const bool raw_mem_only = true; |
|
3677 |
|
3678 // paths into alloc_reg (on the fast path, just before the CopyArray): |
|
3679 enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT }; |
|
3680 RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT); |
|
3681 PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type); |
|
3682 PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X); |
|
3683 PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO); |
|
3684 PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY, |
|
3685 raw_adr_type); |
|
3686 record_for_igvn(alloc_reg); |
|
3687 |
|
3688 bool card_mark = false; // (see below) |
|
3689 |
|
3690 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); |
|
3691 if (array_ctl != NULL) { |
|
3692 // It's an array. |
|
3693 PreserveJVMState pjvms(this); |
|
3694 set_control(array_ctl); |
|
3695 Node* obj_length = load_array_length(obj); |
|
3696 Node* obj_size = NULL; |
|
3697 _sp += nargs; // set original stack for use by uncommon_trap |
|
3698 Node* alloc_obj = new_array(obj_klass, obj_length, |
|
3699 raw_mem_only, &obj_size); |
|
3700 _sp -= nargs; |
|
3701 assert(obj_size != NULL, ""); |
|
3702 Node* raw_obj = alloc_obj->in(1); |
|
3703 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); |
|
3704 if (ReduceBulkZeroing) { |
|
3705 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); |
|
3706 if (alloc != NULL) { |
|
3707 // We will be completely responsible for initializing this object. |
|
3708 alloc->maybe_set_complete(&_gvn); |
|
3709 } |
|
3710 } |
|
3711 |
|
3712 if (!use_ReduceInitialCardMarks()) { |
|
3713 // If it is an oop array, it requires very special treatment, |
|
3714 // because card marking is required on each card of the array. |
|
3715 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); |
|
3716 if (is_obja != NULL) { |
|
3717 PreserveJVMState pjvms2(this); |
|
3718 set_control(is_obja); |
|
3719 // Generate a direct call to the right arraycopy function(s). |
|
3720 bool disjoint_bases = true; |
|
3721 bool length_never_negative = true; |
|
3722 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, |
|
3723 obj, intcon(0), alloc_obj, intcon(0), |
|
3724 obj_length, nargs, |
|
3725 disjoint_bases, length_never_negative); |
|
3726 result_reg->init_req(_objArray_path, control()); |
|
3727 result_val->init_req(_objArray_path, alloc_obj); |
|
3728 result_i_o ->set_req(_objArray_path, i_o()); |
|
3729 result_mem ->set_req(_objArray_path, reset_memory()); |
|
3730 } |
|
3731 } |
|
3732 // We can dispense with card marks if we know the allocation |
|
3733 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks |
|
3734 // causes the non-eden paths to simulate a fresh allocation, |
|
3735 // insofar that no further card marks are required to initialize |
|
3736 // the object. |
|
3737 |
|
3738 // Otherwise, there are no card marks to worry about. |
|
3739 alloc_val->init_req(_typeArray_alloc, raw_obj); |
|
3740 alloc_siz->init_req(_typeArray_alloc, obj_size); |
|
3741 alloc_reg->init_req(_typeArray_alloc, control()); |
|
3742 alloc_i_o->init_req(_typeArray_alloc, i_o()); |
|
3743 alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type)); |
|
3744 } |
|
3745 |
|
3746 // We only go to the fast case code if we pass a number of guards. |
|
3747 // The paths which do not pass are accumulated in the slow_region. |
|
3748 RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
3749 record_for_igvn(slow_region); |
|
3750 if (!stopped()) { |
|
3751 // It's an instance. Make the slow-path tests. |
|
3752 // If this is a virtual call, we generate a funny guard. We grab |
|
3753 // the vtable entry corresponding to clone() from the target object. |
|
3754 // If the target method which we are calling happens to be the |
|
3755 // Object clone() method, we pass the guard. We do not need this |
|
3756 // guard for non-virtual calls; the caller is known to be the native |
|
3757 // Object clone(). |
|
3758 if (is_virtual) { |
|
3759 generate_virtual_guard(obj_klass, slow_region); |
|
3760 } |
|
3761 |
|
3762 // The object must be cloneable and must not have a finalizer. |
|
3763 // Both of these conditions may be checked in a single test. |
|
3764 // We could optimize the cloneable test further, but we don't care. |
|
3765 generate_access_flags_guard(obj_klass, |
|
3766 // Test both conditions: |
|
3767 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, |
|
3768 // Must be cloneable but not finalizer: |
|
3769 JVM_ACC_IS_CLONEABLE, |
|
3770 slow_region); |
|
3771 } |
|
3772 |
|
3773 if (!stopped()) { |
|
3774 // It's an instance, and it passed the slow-path tests. |
|
3775 PreserveJVMState pjvms(this); |
|
3776 Node* obj_size = NULL; |
|
3777 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size); |
|
3778 assert(obj_size != NULL, ""); |
|
3779 Node* raw_obj = alloc_obj->in(1); |
|
3780 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); |
|
3781 if (ReduceBulkZeroing) { |
|
3782 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); |
|
3783 if (alloc != NULL && !alloc->maybe_set_complete(&_gvn)) |
|
3784 alloc = NULL; |
|
3785 } |
|
3786 if (!use_ReduceInitialCardMarks()) { |
|
3787 // Put in store barrier for any and all oops we are sticking |
|
3788 // into this object. (We could avoid this if we could prove |
|
3789 // that the object type contains no oop fields at all.) |
|
3790 card_mark = true; |
|
3791 } |
|
3792 alloc_val->init_req(_instance_alloc, raw_obj); |
|
3793 alloc_siz->init_req(_instance_alloc, obj_size); |
|
3794 alloc_reg->init_req(_instance_alloc, control()); |
|
3795 alloc_i_o->init_req(_instance_alloc, i_o()); |
|
3796 alloc_mem->init_req(_instance_alloc, memory(raw_adr_type)); |
|
3797 } |
|
3798 |
|
3799 // Generate code for the slow case. We make a call to clone(). |
|
3800 set_control(_gvn.transform(slow_region)); |
|
3801 if (!stopped()) { |
|
3802 PreserveJVMState pjvms(this); |
|
3803 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); |
|
3804 Node* slow_result = set_results_for_java_call(slow_call); |
|
3805 // this->control() comes from set_results_for_java_call |
|
3806 result_reg->init_req(_slow_path, control()); |
|
3807 result_val->init_req(_slow_path, slow_result); |
|
3808 result_i_o ->set_req(_slow_path, i_o()); |
|
3809 result_mem ->set_req(_slow_path, reset_memory()); |
|
3810 } |
|
3811 |
|
3812 // The object is allocated, as an array and/or an instance. Now copy it. |
|
3813 set_control( _gvn.transform(alloc_reg) ); |
|
3814 set_i_o( _gvn.transform(alloc_i_o) ); |
|
3815 set_memory( _gvn.transform(alloc_mem), raw_adr_type ); |
|
3816 Node* raw_obj = _gvn.transform(alloc_val); |
|
3817 |
|
3818 if (!stopped()) { |
|
3819 // Copy the fastest available way. |
|
3820 // (No need for PreserveJVMState, since we're using it all up now.) |
|
3821 Node* src = obj; |
|
3822 Node* dest = raw_obj; |
|
3823 Node* end = dest; |
|
3824 Node* size = _gvn.transform(alloc_siz); |
|
3825 |
|
3826 // Exclude the header. |
|
3827 int base_off = sizeof(oopDesc); |
|
3828 src = basic_plus_adr(src, base_off); |
|
3829 dest = basic_plus_adr(dest, base_off); |
|
3830 end = basic_plus_adr(end, size); |
|
3831 |
|
3832 // Compute the length also, if needed: |
|
3833 Node* countx = size; |
|
3834 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); |
|
3835 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); |
|
3836 |
|
3837 // Select an appropriate instruction to initialize the range. |
|
3838 // The CopyArray instruction (if supported) can be optimized |
|
3839 // into a discrete set of scalar loads and stores. |
|
3840 bool disjoint_bases = true; |
|
3841 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, |
|
3842 src, NULL, dest, NULL, countx); |
|
3843 |
|
3844 // Now that the object is properly initialized, type it as an oop. |
|
3845 // Use a secondary InitializeNode memory barrier. |
|
3846 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx, |
|
3847 raw_obj)->as_Initialize(); |
|
3848 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) |
|
3849 Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj, |
|
3850 TypeInstPtr::NOTNULL); |
|
3851 new_obj = _gvn.transform(new_obj); |
|
3852 |
|
3853 // If necessary, emit some card marks afterwards. (Non-arrays only.) |
|
3854 if (card_mark) { |
|
3855 Node* no_particular_value = NULL; |
|
3856 Node* no_particular_field = NULL; |
|
3857 post_barrier(control(), |
|
3858 memory(raw_adr_type), |
|
3859 new_obj, |
|
3860 no_particular_field, |
|
3861 raw_adr_idx, |
|
3862 no_particular_value, |
|
3863 T_OBJECT, |
|
3864 false); |
|
3865 } |
|
3866 // Present the results of the slow call. |
|
3867 result_reg->init_req(_fast_path, control()); |
|
3868 result_val->init_req(_fast_path, new_obj); |
|
3869 result_i_o ->set_req(_fast_path, i_o()); |
|
3870 result_mem ->set_req(_fast_path, reset_memory()); |
|
3871 } |
|
3872 |
|
3873 // Return the combined state. |
|
3874 set_control( _gvn.transform(result_reg) ); |
|
3875 set_i_o( _gvn.transform(result_i_o) ); |
|
3876 set_all_memory( _gvn.transform(result_mem) ); |
|
3877 |
|
3878 // Cast the result to a sharper type, since we know what clone does. |
|
3879 Node* new_obj = _gvn.transform(result_val); |
|
3880 Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop); |
|
3881 push(_gvn.transform(cast)); |
|
3882 |
|
3883 return true; |
|
3884 } |
|
3885 |
|
3886 |
|
3887 // constants for computing the copy function |
|
3888 enum { |
|
3889 COPYFUNC_UNALIGNED = 0, |
|
3890 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize |
|
3891 COPYFUNC_CONJOINT = 0, |
|
3892 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend |
|
3893 }; |
|
3894 |
|
3895 // Note: The condition "disjoint" applies also for overlapping copies |
|
3896 // where an descending copy is permitted (i.e., dest_offset <= src_offset). |
|
3897 static address |
|
3898 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) { |
|
3899 int selector = |
|
3900 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) + |
|
3901 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT); |
|
3902 |
|
3903 #define RETURN_STUB(xxx_arraycopy) { \ |
|
3904 name = #xxx_arraycopy; \ |
|
3905 return StubRoutines::xxx_arraycopy(); } |
|
3906 |
|
3907 switch (t) { |
|
3908 case T_BYTE: |
|
3909 case T_BOOLEAN: |
|
3910 switch (selector) { |
|
3911 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy); |
|
3912 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy); |
|
3913 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy); |
|
3914 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy); |
|
3915 } |
|
3916 case T_CHAR: |
|
3917 case T_SHORT: |
|
3918 switch (selector) { |
|
3919 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy); |
|
3920 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy); |
|
3921 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy); |
|
3922 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy); |
|
3923 } |
|
3924 case T_INT: |
|
3925 case T_FLOAT: |
|
3926 switch (selector) { |
|
3927 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy); |
|
3928 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy); |
|
3929 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy); |
|
3930 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy); |
|
3931 } |
|
3932 case T_DOUBLE: |
|
3933 case T_LONG: |
|
3934 switch (selector) { |
|
3935 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy); |
|
3936 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy); |
|
3937 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy); |
|
3938 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy); |
|
3939 } |
|
3940 case T_ARRAY: |
|
3941 case T_OBJECT: |
|
3942 switch (selector) { |
|
3943 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy); |
|
3944 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy); |
|
3945 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy); |
|
3946 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy); |
|
3947 } |
|
3948 default: |
|
3949 ShouldNotReachHere(); |
|
3950 return NULL; |
|
3951 } |
|
3952 |
|
3953 #undef RETURN_STUB |
|
3954 } |
|
3955 |
|
3956 //------------------------------basictype2arraycopy---------------------------- |
|
3957 address LibraryCallKit::basictype2arraycopy(BasicType t, |
|
3958 Node* src_offset, |
|
3959 Node* dest_offset, |
|
3960 bool disjoint_bases, |
|
3961 const char* &name) { |
|
3962 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; |
|
3963 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; |
|
3964 |
|
3965 bool aligned = false; |
|
3966 bool disjoint = disjoint_bases; |
|
3967 |
|
3968 // if the offsets are the same, we can treat the memory regions as |
|
3969 // disjoint, because either the memory regions are in different arrays, |
|
3970 // or they are identical (which we can treat as disjoint.) We can also |
|
3971 // treat a copy with a destination index less that the source index |
|
3972 // as disjoint since a low->high copy will work correctly in this case. |
|
3973 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && |
|
3974 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { |
|
3975 // both indices are constants |
|
3976 int s_offs = src_offset_inttype->get_con(); |
|
3977 int d_offs = dest_offset_inttype->get_con(); |
|
3978 int element_size = type2aelembytes[t]; |
|
3979 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && |
|
3980 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); |
|
3981 if (s_offs >= d_offs) disjoint = true; |
|
3982 } else if (src_offset == dest_offset && src_offset != NULL) { |
|
3983 // This can occur if the offsets are identical non-constants. |
|
3984 disjoint = true; |
|
3985 } |
|
3986 |
|
3987 return select_arraycopy_function(t, aligned, disjoint, name); |
|
3988 } |
|
3989 |
|
3990 |
|
3991 //------------------------------inline_arraycopy----------------------- |
|
3992 bool LibraryCallKit::inline_arraycopy() { |
|
3993 // Restore the stack and pop off the arguments. |
|
3994 int nargs = 5; // 2 oops, 3 ints, no size_t or long |
|
3995 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); |
|
3996 |
|
3997 Node *src = argument(0); |
|
3998 Node *src_offset = argument(1); |
|
3999 Node *dest = argument(2); |
|
4000 Node *dest_offset = argument(3); |
|
4001 Node *length = argument(4); |
|
4002 |
|
4003 // Compile time checks. If any of these checks cannot be verified at compile time, |
|
4004 // we do not make a fast path for this call. Instead, we let the call remain as it |
|
4005 // is. The checks we choose to mandate at compile time are: |
|
4006 // |
|
4007 // (1) src and dest are arrays. |
|
4008 const Type* src_type = src->Value(&_gvn); |
|
4009 const Type* dest_type = dest->Value(&_gvn); |
|
4010 const TypeAryPtr* top_src = src_type->isa_aryptr(); |
|
4011 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); |
|
4012 if (top_src == NULL || top_src->klass() == NULL || |
|
4013 top_dest == NULL || top_dest->klass() == NULL) { |
|
4014 // Conservatively insert a memory barrier on all memory slices. |
|
4015 // Do not let writes into the source float below the arraycopy. |
|
4016 insert_mem_bar(Op_MemBarCPUOrder); |
|
4017 |
|
4018 // Call StubRoutines::generic_arraycopy stub. |
|
4019 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, |
|
4020 src, src_offset, dest, dest_offset, length, |
|
4021 nargs); |
|
4022 |
|
4023 // Do not let reads from the destination float above the arraycopy. |
|
4024 // Since we cannot type the arrays, we don't know which slices |
|
4025 // might be affected. We could restrict this barrier only to those |
|
4026 // memory slices which pertain to array elements--but don't bother. |
|
4027 if (!InsertMemBarAfterArraycopy) |
|
4028 // (If InsertMemBarAfterArraycopy, there is already one in place.) |
|
4029 insert_mem_bar(Op_MemBarCPUOrder); |
|
4030 return true; |
|
4031 } |
|
4032 |
|
4033 // (2) src and dest arrays must have elements of the same BasicType |
|
4034 // Figure out the size and type of the elements we will be copying. |
|
4035 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); |
|
4036 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); |
|
4037 if (src_elem == T_ARRAY) src_elem = T_OBJECT; |
|
4038 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; |
|
4039 |
|
4040 if (src_elem != dest_elem || dest_elem == T_VOID) { |
|
4041 // The component types are not the same or are not recognized. Punt. |
|
4042 // (But, avoid the native method wrapper to JVM_ArrayCopy.) |
|
4043 generate_slow_arraycopy(TypePtr::BOTTOM, |
|
4044 src, src_offset, dest, dest_offset, length, |
|
4045 nargs); |
|
4046 return true; |
|
4047 } |
|
4048 |
|
4049 //--------------------------------------------------------------------------- |
|
4050 // We will make a fast path for this call to arraycopy. |
|
4051 |
|
4052 // We have the following tests left to perform: |
|
4053 // |
|
4054 // (3) src and dest must not be null. |
|
4055 // (4) src_offset must not be negative. |
|
4056 // (5) dest_offset must not be negative. |
|
4057 // (6) length must not be negative. |
|
4058 // (7) src_offset + length must not exceed length of src. |
|
4059 // (8) dest_offset + length must not exceed length of dest. |
|
4060 // (9) each element of an oop array must be assignable |
|
4061 |
|
4062 RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
4063 record_for_igvn(slow_region); |
|
4064 |
|
4065 // (3) operands must not be null |
|
4066 // We currently perform our null checks with the do_null_check routine. |
|
4067 // This means that the null exceptions will be reported in the caller |
|
4068 // rather than (correctly) reported inside of the native arraycopy call. |
|
4069 // This should be corrected, given time. We do our null check with the |
|
4070 // stack pointer restored. |
|
4071 _sp += nargs; |
|
4072 src = do_null_check(src, T_ARRAY); |
|
4073 dest = do_null_check(dest, T_ARRAY); |
|
4074 _sp -= nargs; |
|
4075 |
|
4076 // (4) src_offset must not be negative. |
|
4077 generate_negative_guard(src_offset, slow_region); |
|
4078 |
|
4079 // (5) dest_offset must not be negative. |
|
4080 generate_negative_guard(dest_offset, slow_region); |
|
4081 |
|
4082 // (6) length must not be negative (moved to generate_arraycopy()). |
|
4083 // generate_negative_guard(length, slow_region); |
|
4084 |
|
4085 // (7) src_offset + length must not exceed length of src. |
|
4086 generate_limit_guard(src_offset, length, |
|
4087 load_array_length(src), |
|
4088 slow_region); |
|
4089 |
|
4090 // (8) dest_offset + length must not exceed length of dest. |
|
4091 generate_limit_guard(dest_offset, length, |
|
4092 load_array_length(dest), |
|
4093 slow_region); |
|
4094 |
|
4095 // (9) each element of an oop array must be assignable |
|
4096 // The generate_arraycopy subroutine checks this. |
|
4097 |
|
4098 // This is where the memory effects are placed: |
|
4099 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); |
|
4100 generate_arraycopy(adr_type, dest_elem, |
|
4101 src, src_offset, dest, dest_offset, length, |
|
4102 nargs, false, false, slow_region); |
|
4103 |
|
4104 return true; |
|
4105 } |
|
4106 |
|
4107 //-----------------------------generate_arraycopy---------------------- |
|
4108 // Generate an optimized call to arraycopy. |
|
4109 // Caller must guard against non-arrays. |
|
4110 // Caller must determine a common array basic-type for both arrays. |
|
4111 // Caller must validate offsets against array bounds. |
|
4112 // The slow_region has already collected guard failure paths |
|
4113 // (such as out of bounds length or non-conformable array types). |
|
4114 // The generated code has this shape, in general: |
|
4115 // |
|
4116 // if (length == 0) return // via zero_path |
|
4117 // slowval = -1 |
|
4118 // if (types unknown) { |
|
4119 // slowval = call generic copy loop |
|
4120 // if (slowval == 0) return // via checked_path |
|
4121 // } else if (indexes in bounds) { |
|
4122 // if ((is object array) && !(array type check)) { |
|
4123 // slowval = call checked copy loop |
|
4124 // if (slowval == 0) return // via checked_path |
|
4125 // } else { |
|
4126 // call bulk copy loop |
|
4127 // return // via fast_path |
|
4128 // } |
|
4129 // } |
|
4130 // // adjust params for remaining work: |
|
4131 // if (slowval != -1) { |
|
4132 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n |
|
4133 // } |
|
4134 // slow_region: |
|
4135 // call slow arraycopy(src, src_offset, dest, dest_offset, length) |
|
4136 // return // via slow_call_path |
|
4137 // |
|
4138 // This routine is used from several intrinsics: System.arraycopy, |
|
4139 // Object.clone (the array subcase), and Arrays.copyOf[Range]. |
|
4140 // |
|
4141 void |
|
4142 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, |
|
4143 BasicType basic_elem_type, |
|
4144 Node* src, Node* src_offset, |
|
4145 Node* dest, Node* dest_offset, |
|
4146 Node* copy_length, |
|
4147 int nargs, |
|
4148 bool disjoint_bases, |
|
4149 bool length_never_negative, |
|
4150 RegionNode* slow_region) { |
|
4151 |
|
4152 if (slow_region == NULL) { |
|
4153 slow_region = new(C,1) RegionNode(1); |
|
4154 record_for_igvn(slow_region); |
|
4155 } |
|
4156 |
|
4157 Node* original_dest = dest; |
|
4158 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed |
|
4159 Node* raw_dest = NULL; // used before zeroing, if needed |
|
4160 bool must_clear_dest = false; |
|
4161 |
|
4162 // See if this is the initialization of a newly-allocated array. |
|
4163 // If so, we will take responsibility here for initializing it to zero. |
|
4164 // (Note: Because tightly_coupled_allocation performs checks on the |
|
4165 // out-edges of the dest, we need to avoid making derived pointers |
|
4166 // from it until we have checked its uses.) |
|
4167 if (ReduceBulkZeroing |
|
4168 && !ZeroTLAB // pointless if already zeroed |
|
4169 && basic_elem_type != T_CONFLICT // avoid corner case |
|
4170 && !_gvn.eqv_uncast(src, dest) |
|
4171 && ((alloc = tightly_coupled_allocation(dest, slow_region)) |
|
4172 != NULL) |
|
4173 && alloc->maybe_set_complete(&_gvn)) { |
|
4174 // "You break it, you buy it." |
|
4175 InitializeNode* init = alloc->initialization(); |
|
4176 assert(init->is_complete(), "we just did this"); |
|
4177 assert(dest->Opcode() == Op_CheckCastPP, "sanity"); |
|
4178 assert(dest->in(0)->in(0) == init, "dest pinned"); |
|
4179 raw_dest = dest->in(1); // grab the raw pointer! |
|
4180 original_dest = dest; |
|
4181 dest = raw_dest; |
|
4182 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory |
|
4183 // Decouple the original InitializeNode, turning it into a simple membar. |
|
4184 // We will build a new one at the end of this routine. |
|
4185 init->set_req(InitializeNode::RawAddress, top()); |
|
4186 // From this point on, every exit path is responsible for |
|
4187 // initializing any non-copied parts of the object to zero. |
|
4188 must_clear_dest = true; |
|
4189 } else { |
|
4190 // No zeroing elimination here. |
|
4191 alloc = NULL; |
|
4192 //original_dest = dest; |
|
4193 //must_clear_dest = false; |
|
4194 } |
|
4195 |
|
4196 // Results are placed here: |
|
4197 enum { fast_path = 1, // normal void-returning assembly stub |
|
4198 checked_path = 2, // special assembly stub with cleanup |
|
4199 slow_call_path = 3, // something went wrong; call the VM |
|
4200 zero_path = 4, // bypass when length of copy is zero |
|
4201 bcopy_path = 5, // copy primitive array by 64-bit blocks |
|
4202 PATH_LIMIT = 6 |
|
4203 }; |
|
4204 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
4205 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); |
|
4206 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); |
|
4207 record_for_igvn(result_region); |
|
4208 _gvn.set_type_bottom(result_i_o); |
|
4209 _gvn.set_type_bottom(result_memory); |
|
4210 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); |
|
4211 |
|
4212 // The slow_control path: |
|
4213 Node* slow_control; |
|
4214 Node* slow_i_o = i_o(); |
|
4215 Node* slow_mem = memory(adr_type); |
|
4216 debug_only(slow_control = (Node*) badAddress); |
|
4217 |
|
4218 // Checked control path: |
|
4219 Node* checked_control = top(); |
|
4220 Node* checked_mem = NULL; |
|
4221 Node* checked_i_o = NULL; |
|
4222 Node* checked_value = NULL; |
|
4223 |
|
4224 if (basic_elem_type == T_CONFLICT) { |
|
4225 assert(!must_clear_dest, ""); |
|
4226 Node* cv = generate_generic_arraycopy(adr_type, |
|
4227 src, src_offset, dest, dest_offset, |
|
4228 copy_length, nargs); |
|
4229 if (cv == NULL) cv = intcon(-1); // failure (no stub available) |
|
4230 checked_control = control(); |
|
4231 checked_i_o = i_o(); |
|
4232 checked_mem = memory(adr_type); |
|
4233 checked_value = cv; |
|
4234 set_control(top()); // no fast path |
|
4235 } |
|
4236 |
|
4237 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); |
|
4238 if (not_pos != NULL) { |
|
4239 PreserveJVMState pjvms(this); |
|
4240 set_control(not_pos); |
|
4241 |
|
4242 // (6) length must not be negative. |
|
4243 if (!length_never_negative) { |
|
4244 generate_negative_guard(copy_length, slow_region); |
|
4245 } |
|
4246 |
|
4247 if (!stopped() && must_clear_dest) { |
|
4248 Node* dest_length = alloc->in(AllocateNode::ALength); |
|
4249 if (_gvn.eqv_uncast(copy_length, dest_length) |
|
4250 || _gvn.find_int_con(dest_length, 1) <= 0) { |
|
4251 // There is no zeroing to do. |
|
4252 } else { |
|
4253 // Clear the whole thing since there are no source elements to copy. |
|
4254 generate_clear_array(adr_type, dest, basic_elem_type, |
|
4255 intcon(0), NULL, |
|
4256 alloc->in(AllocateNode::AllocSize)); |
|
4257 } |
|
4258 } |
|
4259 |
|
4260 // Present the results of the fast call. |
|
4261 result_region->init_req(zero_path, control()); |
|
4262 result_i_o ->init_req(zero_path, i_o()); |
|
4263 result_memory->init_req(zero_path, memory(adr_type)); |
|
4264 } |
|
4265 |
|
4266 if (!stopped() && must_clear_dest) { |
|
4267 // We have to initialize the *uncopied* part of the array to zero. |
|
4268 // The copy destination is the slice dest[off..off+len]. The other slices |
|
4269 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. |
|
4270 Node* dest_size = alloc->in(AllocateNode::AllocSize); |
|
4271 Node* dest_length = alloc->in(AllocateNode::ALength); |
|
4272 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, |
|
4273 copy_length) ); |
|
4274 |
|
4275 // If there is a head section that needs zeroing, do it now. |
|
4276 if (find_int_con(dest_offset, -1) != 0) { |
|
4277 generate_clear_array(adr_type, dest, basic_elem_type, |
|
4278 intcon(0), dest_offset, |
|
4279 NULL); |
|
4280 } |
|
4281 |
|
4282 // Next, perform a dynamic check on the tail length. |
|
4283 // It is often zero, and we can win big if we prove this. |
|
4284 // There are two wins: Avoid generating the ClearArray |
|
4285 // with its attendant messy index arithmetic, and upgrade |
|
4286 // the copy to a more hardware-friendly word size of 64 bits. |
|
4287 Node* tail_ctl = NULL; |
|
4288 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) { |
|
4289 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); |
|
4290 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); |
|
4291 tail_ctl = generate_slow_guard(bol_lt, NULL); |
|
4292 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); |
|
4293 } |
|
4294 |
|
4295 // At this point, let's assume there is no tail. |
|
4296 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { |
|
4297 // There is no tail. Try an upgrade to a 64-bit copy. |
|
4298 bool didit = false; |
|
4299 { PreserveJVMState pjvms(this); |
|
4300 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, |
|
4301 src, src_offset, dest, dest_offset, |
|
4302 dest_size); |
|
4303 if (didit) { |
|
4304 // Present the results of the block-copying fast call. |
|
4305 result_region->init_req(bcopy_path, control()); |
|
4306 result_i_o ->init_req(bcopy_path, i_o()); |
|
4307 result_memory->init_req(bcopy_path, memory(adr_type)); |
|
4308 } |
|
4309 } |
|
4310 if (didit) |
|
4311 set_control(top()); // no regular fast path |
|
4312 } |
|
4313 |
|
4314 // Clear the tail, if any. |
|
4315 if (tail_ctl != NULL) { |
|
4316 Node* notail_ctl = stopped() ? NULL : control(); |
|
4317 set_control(tail_ctl); |
|
4318 if (notail_ctl == NULL) { |
|
4319 generate_clear_array(adr_type, dest, basic_elem_type, |
|
4320 dest_tail, NULL, |
|
4321 dest_size); |
|
4322 } else { |
|
4323 // Make a local merge. |
|
4324 Node* done_ctl = new(C,3) RegionNode(3); |
|
4325 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); |
|
4326 done_ctl->init_req(1, notail_ctl); |
|
4327 done_mem->init_req(1, memory(adr_type)); |
|
4328 generate_clear_array(adr_type, dest, basic_elem_type, |
|
4329 dest_tail, NULL, |
|
4330 dest_size); |
|
4331 done_ctl->init_req(2, control()); |
|
4332 done_mem->init_req(2, memory(adr_type)); |
|
4333 set_control( _gvn.transform(done_ctl) ); |
|
4334 set_memory( _gvn.transform(done_mem), adr_type ); |
|
4335 } |
|
4336 } |
|
4337 } |
|
4338 |
|
4339 BasicType copy_type = basic_elem_type; |
|
4340 assert(basic_elem_type != T_ARRAY, "caller must fix this"); |
|
4341 if (!stopped() && copy_type == T_OBJECT) { |
|
4342 // If src and dest have compatible element types, we can copy bits. |
|
4343 // Types S[] and D[] are compatible if D is a supertype of S. |
|
4344 // |
|
4345 // If they are not, we will use checked_oop_disjoint_arraycopy, |
|
4346 // which performs a fast optimistic per-oop check, and backs off |
|
4347 // further to JVM_ArrayCopy on the first per-oop check that fails. |
|
4348 // (Actually, we don't move raw bits only; the GC requires card marks.) |
|
4349 |
|
4350 // Get the klassOop for both src and dest |
|
4351 Node* src_klass = load_object_klass(src); |
|
4352 Node* dest_klass = load_object_klass(dest); |
|
4353 |
|
4354 // Generate the subtype check. |
|
4355 // This might fold up statically, or then again it might not. |
|
4356 // |
|
4357 // Non-static example: Copying List<String>.elements to a new String[]. |
|
4358 // The backing store for a List<String> is always an Object[], |
|
4359 // but its elements are always type String, if the generic types |
|
4360 // are correct at the source level. |
|
4361 // |
|
4362 // Test S[] against D[], not S against D, because (probably) |
|
4363 // the secondary supertype cache is less busy for S[] than S. |
|
4364 // This usually only matters when D is an interface. |
|
4365 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); |
|
4366 // Plug failing path into checked_oop_disjoint_arraycopy |
|
4367 if (not_subtype_ctrl != top()) { |
|
4368 PreserveJVMState pjvms(this); |
|
4369 set_control(not_subtype_ctrl); |
|
4370 // (At this point we can assume disjoint_bases, since types differ.) |
|
4371 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc); |
|
4372 Node* p1 = basic_plus_adr(dest_klass, ek_offset); |
|
4373 Node* n1 = new (C, 3) LoadKlassNode(0, immutable_memory(), p1, TypeRawPtr::BOTTOM); |
|
4374 Node* dest_elem_klass = _gvn.transform(n1); |
|
4375 Node* cv = generate_checkcast_arraycopy(adr_type, |
|
4376 dest_elem_klass, |
|
4377 src, src_offset, dest, dest_offset, |
|
4378 copy_length, |
|
4379 nargs); |
|
4380 if (cv == NULL) cv = intcon(-1); // failure (no stub available) |
|
4381 checked_control = control(); |
|
4382 checked_i_o = i_o(); |
|
4383 checked_mem = memory(adr_type); |
|
4384 checked_value = cv; |
|
4385 } |
|
4386 // At this point we know we do not need type checks on oop stores. |
|
4387 |
|
4388 // Let's see if we need card marks: |
|
4389 if (alloc != NULL && use_ReduceInitialCardMarks()) { |
|
4390 // If we do not need card marks, copy using the jint or jlong stub. |
|
4391 copy_type = LP64_ONLY(T_LONG) NOT_LP64(T_INT); |
|
4392 assert(type2aelembytes[basic_elem_type] == type2aelembytes[copy_type], |
|
4393 "sizes agree"); |
|
4394 } |
|
4395 } |
|
4396 |
|
4397 if (!stopped()) { |
|
4398 // Generate the fast path, if possible. |
|
4399 PreserveJVMState pjvms(this); |
|
4400 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, |
|
4401 src, src_offset, dest, dest_offset, |
|
4402 ConvI2X(copy_length)); |
|
4403 |
|
4404 // Present the results of the fast call. |
|
4405 result_region->init_req(fast_path, control()); |
|
4406 result_i_o ->init_req(fast_path, i_o()); |
|
4407 result_memory->init_req(fast_path, memory(adr_type)); |
|
4408 } |
|
4409 |
|
4410 // Here are all the slow paths up to this point, in one bundle: |
|
4411 slow_control = top(); |
|
4412 if (slow_region != NULL) |
|
4413 slow_control = _gvn.transform(slow_region); |
|
4414 debug_only(slow_region = (RegionNode*)badAddress); |
|
4415 |
|
4416 set_control(checked_control); |
|
4417 if (!stopped()) { |
|
4418 // Clean up after the checked call. |
|
4419 // The returned value is either 0 or -1^K, |
|
4420 // where K = number of partially transferred array elements. |
|
4421 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); |
|
4422 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); |
|
4423 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); |
|
4424 |
|
4425 // If it is 0, we are done, so transfer to the end. |
|
4426 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); |
|
4427 result_region->init_req(checked_path, checks_done); |
|
4428 result_i_o ->init_req(checked_path, checked_i_o); |
|
4429 result_memory->init_req(checked_path, checked_mem); |
|
4430 |
|
4431 // If it is not zero, merge into the slow call. |
|
4432 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); |
|
4433 RegionNode* slow_reg2 = new(C, 3) RegionNode(3); |
|
4434 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); |
|
4435 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); |
|
4436 record_for_igvn(slow_reg2); |
|
4437 slow_reg2 ->init_req(1, slow_control); |
|
4438 slow_i_o2 ->init_req(1, slow_i_o); |
|
4439 slow_mem2 ->init_req(1, slow_mem); |
|
4440 slow_reg2 ->init_req(2, control()); |
|
4441 slow_i_o2 ->init_req(2, i_o()); |
|
4442 slow_mem2 ->init_req(2, memory(adr_type)); |
|
4443 |
|
4444 slow_control = _gvn.transform(slow_reg2); |
|
4445 slow_i_o = _gvn.transform(slow_i_o2); |
|
4446 slow_mem = _gvn.transform(slow_mem2); |
|
4447 |
|
4448 if (alloc != NULL) { |
|
4449 // We'll restart from the very beginning, after zeroing the whole thing. |
|
4450 // This can cause double writes, but that's OK since dest is brand new. |
|
4451 // So we ignore the low 31 bits of the value returned from the stub. |
|
4452 } else { |
|
4453 // We must continue the copy exactly where it failed, or else |
|
4454 // another thread might see the wrong number of writes to dest. |
|
4455 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); |
|
4456 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); |
|
4457 slow_offset->init_req(1, intcon(0)); |
|
4458 slow_offset->init_req(2, checked_offset); |
|
4459 slow_offset = _gvn.transform(slow_offset); |
|
4460 |
|
4461 // Adjust the arguments by the conditionally incoming offset. |
|
4462 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); |
|
4463 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); |
|
4464 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); |
|
4465 |
|
4466 // Tweak the node variables to adjust the code produced below: |
|
4467 src_offset = src_off_plus; |
|
4468 dest_offset = dest_off_plus; |
|
4469 copy_length = length_minus; |
|
4470 } |
|
4471 } |
|
4472 |
|
4473 set_control(slow_control); |
|
4474 if (!stopped()) { |
|
4475 // Generate the slow path, if needed. |
|
4476 PreserveJVMState pjvms(this); // replace_in_map may trash the map |
|
4477 |
|
4478 set_memory(slow_mem, adr_type); |
|
4479 set_i_o(slow_i_o); |
|
4480 |
|
4481 if (must_clear_dest) { |
|
4482 generate_clear_array(adr_type, dest, basic_elem_type, |
|
4483 intcon(0), NULL, |
|
4484 alloc->in(AllocateNode::AllocSize)); |
|
4485 } |
|
4486 |
|
4487 if (dest != original_dest) { |
|
4488 // Promote from rawptr to oop, so it looks right in the call's GC map. |
|
4489 dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest, |
|
4490 TypeInstPtr::NOTNULL) ); |
|
4491 |
|
4492 // Edit the call's debug-info to avoid referring to original_dest. |
|
4493 // (The problem with original_dest is that it isn't ready until |
|
4494 // after the InitializeNode completes, but this stuff is before.) |
|
4495 // Substitute in the locally valid dest_oop. |
|
4496 replace_in_map(original_dest, dest); |
|
4497 } |
|
4498 |
|
4499 generate_slow_arraycopy(adr_type, |
|
4500 src, src_offset, dest, dest_offset, |
|
4501 copy_length, nargs); |
|
4502 |
|
4503 result_region->init_req(slow_call_path, control()); |
|
4504 result_i_o ->init_req(slow_call_path, i_o()); |
|
4505 result_memory->init_req(slow_call_path, memory(adr_type)); |
|
4506 } |
|
4507 |
|
4508 // Remove unused edges. |
|
4509 for (uint i = 1; i < result_region->req(); i++) { |
|
4510 if (result_region->in(i) == NULL) |
|
4511 result_region->init_req(i, top()); |
|
4512 } |
|
4513 |
|
4514 // Finished; return the combined state. |
|
4515 set_control( _gvn.transform(result_region) ); |
|
4516 set_i_o( _gvn.transform(result_i_o) ); |
|
4517 set_memory( _gvn.transform(result_memory), adr_type ); |
|
4518 |
|
4519 if (dest != original_dest) { |
|
4520 // Pin the "finished" array node after the arraycopy/zeroing operations. |
|
4521 // Use a secondary InitializeNode memory barrier. |
|
4522 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, |
|
4523 Compile::AliasIdxRaw, |
|
4524 raw_dest)->as_Initialize(); |
|
4525 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) |
|
4526 _gvn.hash_delete(original_dest); |
|
4527 original_dest->set_req(0, control()); |
|
4528 _gvn.hash_find_insert(original_dest); // put back into GVN table |
|
4529 } |
|
4530 |
|
4531 // The memory edges above are precise in order to model effects around |
|
4532 // array copyies accurately to allow value numbering of field loads around |
|
4533 // arraycopy. Such field loads, both before and after, are common in Java |
|
4534 // collections and similar classes involving header/array data structures. |
|
4535 // |
|
4536 // But with low number of register or when some registers are used or killed |
|
4537 // by arraycopy calls it causes registers spilling on stack. See 6544710. |
|
4538 // The next memory barrier is added to avoid it. If the arraycopy can be |
|
4539 // optimized away (which it can, sometimes) then we can manually remove |
|
4540 // the membar also. |
|
4541 if (InsertMemBarAfterArraycopy) |
|
4542 insert_mem_bar(Op_MemBarCPUOrder); |
|
4543 } |
|
4544 |
|
4545 |
|
4546 // Helper function which determines if an arraycopy immediately follows |
|
4547 // an allocation, with no intervening tests or other escapes for the object. |
|
4548 AllocateArrayNode* |
|
4549 LibraryCallKit::tightly_coupled_allocation(Node* ptr, |
|
4550 RegionNode* slow_region) { |
|
4551 if (stopped()) return NULL; // no fast path |
|
4552 if (C->AliasLevel() == 0) return NULL; // no MergeMems around |
|
4553 |
|
4554 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); |
|
4555 if (alloc == NULL) return NULL; |
|
4556 |
|
4557 Node* rawmem = memory(Compile::AliasIdxRaw); |
|
4558 // Is the allocation's memory state untouched? |
|
4559 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { |
|
4560 // Bail out if there have been raw-memory effects since the allocation. |
|
4561 // (Example: There might have been a call or safepoint.) |
|
4562 return NULL; |
|
4563 } |
|
4564 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); |
|
4565 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { |
|
4566 return NULL; |
|
4567 } |
|
4568 |
|
4569 // There must be no unexpected observers of this allocation. |
|
4570 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { |
|
4571 Node* obs = ptr->fast_out(i); |
|
4572 if (obs != this->map()) { |
|
4573 return NULL; |
|
4574 } |
|
4575 } |
|
4576 |
|
4577 // This arraycopy must unconditionally follow the allocation of the ptr. |
|
4578 Node* alloc_ctl = ptr->in(0); |
|
4579 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); |
|
4580 |
|
4581 Node* ctl = control(); |
|
4582 while (ctl != alloc_ctl) { |
|
4583 // There may be guards which feed into the slow_region. |
|
4584 // Any other control flow means that we might not get a chance |
|
4585 // to finish initializing the allocated object. |
|
4586 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { |
|
4587 IfNode* iff = ctl->in(0)->as_If(); |
|
4588 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); |
|
4589 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); |
|
4590 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { |
|
4591 ctl = iff->in(0); // This test feeds the known slow_region. |
|
4592 continue; |
|
4593 } |
|
4594 // One more try: Various low-level checks bottom out in |
|
4595 // uncommon traps. If the debug-info of the trap omits |
|
4596 // any reference to the allocation, as we've already |
|
4597 // observed, then there can be no objection to the trap. |
|
4598 bool found_trap = false; |
|
4599 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { |
|
4600 Node* obs = not_ctl->fast_out(j); |
|
4601 if (obs->in(0) == not_ctl && obs->is_Call() && |
|
4602 (obs->as_Call()->entry_point() == |
|
4603 SharedRuntime::uncommon_trap_blob()->instructions_begin())) { |
|
4604 found_trap = true; break; |
|
4605 } |
|
4606 } |
|
4607 if (found_trap) { |
|
4608 ctl = iff->in(0); // This test feeds a harmless uncommon trap. |
|
4609 continue; |
|
4610 } |
|
4611 } |
|
4612 return NULL; |
|
4613 } |
|
4614 |
|
4615 // If we get this far, we have an allocation which immediately |
|
4616 // precedes the arraycopy, and we can take over zeroing the new object. |
|
4617 // The arraycopy will finish the initialization, and provide |
|
4618 // a new control state to which we will anchor the destination pointer. |
|
4619 |
|
4620 return alloc; |
|
4621 } |
|
4622 |
|
4623 // Helper for initialization of arrays, creating a ClearArray. |
|
4624 // It writes zero bits in [start..end), within the body of an array object. |
|
4625 // The memory effects are all chained onto the 'adr_type' alias category. |
|
4626 // |
|
4627 // Since the object is otherwise uninitialized, we are free |
|
4628 // to put a little "slop" around the edges of the cleared area, |
|
4629 // as long as it does not go back into the array's header, |
|
4630 // or beyond the array end within the heap. |
|
4631 // |
|
4632 // The lower edge can be rounded down to the nearest jint and the |
|
4633 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. |
|
4634 // |
|
4635 // Arguments: |
|
4636 // adr_type memory slice where writes are generated |
|
4637 // dest oop of the destination array |
|
4638 // basic_elem_type element type of the destination |
|
4639 // slice_idx array index of first element to store |
|
4640 // slice_len number of elements to store (or NULL) |
|
4641 // dest_size total size in bytes of the array object |
|
4642 // |
|
4643 // Exactly one of slice_len or dest_size must be non-NULL. |
|
4644 // If dest_size is non-NULL, zeroing extends to the end of the object. |
|
4645 // If slice_len is non-NULL, the slice_idx value must be a constant. |
|
4646 void |
|
4647 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, |
|
4648 Node* dest, |
|
4649 BasicType basic_elem_type, |
|
4650 Node* slice_idx, |
|
4651 Node* slice_len, |
|
4652 Node* dest_size) { |
|
4653 // one or the other but not both of slice_len and dest_size: |
|
4654 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); |
|
4655 if (slice_len == NULL) slice_len = top(); |
|
4656 if (dest_size == NULL) dest_size = top(); |
|
4657 |
|
4658 // operate on this memory slice: |
|
4659 Node* mem = memory(adr_type); // memory slice to operate on |
|
4660 |
|
4661 // scaling and rounding of indexes: |
|
4662 int scale = exact_log2(type2aelembytes[basic_elem_type]); |
|
4663 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); |
|
4664 int clear_low = (-1 << scale) & (BytesPerInt - 1); |
|
4665 int bump_bit = (-1 << scale) & BytesPerInt; |
|
4666 |
|
4667 // determine constant starts and ends |
|
4668 const intptr_t BIG_NEG = -128; |
|
4669 assert(BIG_NEG + 2*abase < 0, "neg enough"); |
|
4670 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); |
|
4671 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); |
|
4672 if (slice_len_con == 0) { |
|
4673 return; // nothing to do here |
|
4674 } |
|
4675 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; |
|
4676 intptr_t end_con = find_intptr_t_con(dest_size, -1); |
|
4677 if (slice_idx_con >= 0 && slice_len_con >= 0) { |
|
4678 assert(end_con < 0, "not two cons"); |
|
4679 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), |
|
4680 BytesPerLong); |
|
4681 } |
|
4682 |
|
4683 if (start_con >= 0 && end_con >= 0) { |
|
4684 // Constant start and end. Simple. |
|
4685 mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4686 start_con, end_con, &_gvn); |
|
4687 } else if (start_con >= 0 && dest_size != top()) { |
|
4688 // Constant start, pre-rounded end after the tail of the array. |
|
4689 Node* end = dest_size; |
|
4690 mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4691 start_con, end, &_gvn); |
|
4692 } else if (start_con >= 0 && slice_len != top()) { |
|
4693 // Constant start, non-constant end. End needs rounding up. |
|
4694 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) |
|
4695 intptr_t end_base = abase + (slice_idx_con << scale); |
|
4696 int end_round = (-1 << scale) & (BytesPerLong - 1); |
|
4697 Node* end = ConvI2X(slice_len); |
|
4698 if (scale != 0) |
|
4699 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); |
|
4700 end_base += end_round; |
|
4701 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); |
|
4702 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); |
|
4703 mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4704 start_con, end, &_gvn); |
|
4705 } else if (start_con < 0 && dest_size != top()) { |
|
4706 // Non-constant start, pre-rounded end after the tail of the array. |
|
4707 // This is almost certainly a "round-to-end" operation. |
|
4708 Node* start = slice_idx; |
|
4709 start = ConvI2X(start); |
|
4710 if (scale != 0) |
|
4711 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); |
|
4712 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); |
|
4713 if ((bump_bit | clear_low) != 0) { |
|
4714 int to_clear = (bump_bit | clear_low); |
|
4715 // Align up mod 8, then store a jint zero unconditionally |
|
4716 // just before the mod-8 boundary. |
|
4717 // This would only fail if the first array element were immediately |
|
4718 // after the length field, and were also at an even offset mod 8. |
|
4719 assert(((abase + bump_bit) & ~to_clear) - BytesPerInt |
|
4720 >= arrayOopDesc::length_offset_in_bytes() + BytesPerInt, |
|
4721 "store must not trash length field"); |
|
4722 |
|
4723 // Bump 'start' up to (or past) the next jint boundary: |
|
4724 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); |
|
4725 // Round bumped 'start' down to jlong boundary in body of array. |
|
4726 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); |
|
4727 // Store a zero to the immediately preceding jint: |
|
4728 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-BytesPerInt)) ); |
|
4729 Node* p1 = basic_plus_adr(dest, x1); |
|
4730 mem = StoreNode::make(C, control(), mem, p1, adr_type, intcon(0), T_INT); |
|
4731 mem = _gvn.transform(mem); |
|
4732 } |
|
4733 |
|
4734 Node* end = dest_size; // pre-rounded |
|
4735 mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4736 start, end, &_gvn); |
|
4737 } else { |
|
4738 // Non-constant start, unrounded non-constant end. |
|
4739 // (Nobody zeroes a random midsection of an array using this routine.) |
|
4740 ShouldNotReachHere(); // fix caller |
|
4741 } |
|
4742 |
|
4743 // Done. |
|
4744 set_memory(mem, adr_type); |
|
4745 } |
|
4746 |
|
4747 |
|
4748 bool |
|
4749 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, |
|
4750 BasicType basic_elem_type, |
|
4751 AllocateNode* alloc, |
|
4752 Node* src, Node* src_offset, |
|
4753 Node* dest, Node* dest_offset, |
|
4754 Node* dest_size) { |
|
4755 // See if there is an advantage from block transfer. |
|
4756 int scale = exact_log2(type2aelembytes[basic_elem_type]); |
|
4757 if (scale >= LogBytesPerLong) |
|
4758 return false; // it is already a block transfer |
|
4759 |
|
4760 // Look at the alignment of the starting offsets. |
|
4761 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); |
|
4762 const intptr_t BIG_NEG = -128; |
|
4763 assert(BIG_NEG + 2*abase < 0, "neg enough"); |
|
4764 |
|
4765 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale); |
|
4766 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale); |
|
4767 if (src_off < 0 || dest_off < 0) |
|
4768 // At present, we can only understand constants. |
|
4769 return false; |
|
4770 |
|
4771 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { |
|
4772 // Non-aligned; too bad. |
|
4773 // One more chance: Pick off an initial 32-bit word. |
|
4774 // This is a common case, since abase can be odd mod 8. |
|
4775 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && |
|
4776 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { |
|
4777 Node* sptr = basic_plus_adr(src, src_off); |
|
4778 Node* dptr = basic_plus_adr(dest, dest_off); |
|
4779 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); |
|
4780 store_to_memory(control(), dptr, sval, T_INT, adr_type); |
|
4781 src_off += BytesPerInt; |
|
4782 dest_off += BytesPerInt; |
|
4783 } else { |
|
4784 return false; |
|
4785 } |
|
4786 } |
|
4787 assert(src_off % BytesPerLong == 0, ""); |
|
4788 assert(dest_off % BytesPerLong == 0, ""); |
|
4789 |
|
4790 // Do this copy by giant steps. |
|
4791 Node* sptr = basic_plus_adr(src, src_off); |
|
4792 Node* dptr = basic_plus_adr(dest, dest_off); |
|
4793 Node* countx = dest_size; |
|
4794 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); |
|
4795 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); |
|
4796 |
|
4797 bool disjoint_bases = true; // since alloc != NULL |
|
4798 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, |
|
4799 sptr, NULL, dptr, NULL, countx); |
|
4800 |
|
4801 return true; |
|
4802 } |
|
4803 |
|
4804 |
|
4805 // Helper function; generates code for the slow case. |
|
4806 // We make a call to a runtime method which emulates the native method, |
|
4807 // but without the native wrapper overhead. |
|
4808 void |
|
4809 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, |
|
4810 Node* src, Node* src_offset, |
|
4811 Node* dest, Node* dest_offset, |
|
4812 Node* copy_length, |
|
4813 int nargs) { |
|
4814 _sp += nargs; // any deopt will start just before call to enclosing method |
|
4815 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, |
|
4816 OptoRuntime::slow_arraycopy_Type(), |
|
4817 OptoRuntime::slow_arraycopy_Java(), |
|
4818 "slow_arraycopy", adr_type, |
|
4819 src, src_offset, dest, dest_offset, |
|
4820 copy_length); |
|
4821 _sp -= nargs; |
|
4822 |
|
4823 // Handle exceptions thrown by this fellow: |
|
4824 make_slow_call_ex(call, env()->Throwable_klass(), false); |
|
4825 } |
|
4826 |
|
4827 // Helper function; generates code for cases requiring runtime checks. |
|
4828 Node* |
|
4829 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, |
|
4830 Node* dest_elem_klass, |
|
4831 Node* src, Node* src_offset, |
|
4832 Node* dest, Node* dest_offset, |
|
4833 Node* copy_length, |
|
4834 int nargs) { |
|
4835 if (stopped()) return NULL; |
|
4836 |
|
4837 address copyfunc_addr = StubRoutines::checkcast_arraycopy(); |
|
4838 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. |
|
4839 return NULL; |
|
4840 } |
|
4841 |
|
4842 // Pick out the parameters required to perform a store-check |
|
4843 // for the target array. This is an optimistic check. It will |
|
4844 // look in each non-null element's class, at the desired klass's |
|
4845 // super_check_offset, for the desired klass. |
|
4846 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc); |
|
4847 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); |
|
4848 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM); |
|
4849 Node* check_offset = _gvn.transform(n3); |
|
4850 Node* check_value = dest_elem_klass; |
|
4851 |
|
4852 Node* src_start = array_element_address(src, src_offset, T_OBJECT); |
|
4853 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); |
|
4854 |
|
4855 // (We know the arrays are never conjoint, because their types differ.) |
|
4856 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
|
4857 OptoRuntime::checkcast_arraycopy_Type(), |
|
4858 copyfunc_addr, "checkcast_arraycopy", adr_type, |
|
4859 // five arguments, of which two are |
|
4860 // intptr_t (jlong in LP64) |
|
4861 src_start, dest_start, |
|
4862 copy_length XTOP, |
|
4863 check_offset XTOP, |
|
4864 check_value); |
|
4865 |
|
4866 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); |
|
4867 } |
|
4868 |
|
4869 |
|
4870 // Helper function; generates code for cases requiring runtime checks. |
|
4871 Node* |
|
4872 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, |
|
4873 Node* src, Node* src_offset, |
|
4874 Node* dest, Node* dest_offset, |
|
4875 Node* copy_length, |
|
4876 int nargs) { |
|
4877 if (stopped()) return NULL; |
|
4878 |
|
4879 address copyfunc_addr = StubRoutines::generic_arraycopy(); |
|
4880 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. |
|
4881 return NULL; |
|
4882 } |
|
4883 |
|
4884 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
|
4885 OptoRuntime::generic_arraycopy_Type(), |
|
4886 copyfunc_addr, "generic_arraycopy", adr_type, |
|
4887 src, src_offset, dest, dest_offset, copy_length); |
|
4888 |
|
4889 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); |
|
4890 } |
|
4891 |
|
4892 // Helper function; generates the fast out-of-line call to an arraycopy stub. |
|
4893 void |
|
4894 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, |
|
4895 BasicType basic_elem_type, |
|
4896 bool disjoint_bases, |
|
4897 Node* src, Node* src_offset, |
|
4898 Node* dest, Node* dest_offset, |
|
4899 Node* copy_length) { |
|
4900 if (stopped()) return; // nothing to do |
|
4901 |
|
4902 Node* src_start = src; |
|
4903 Node* dest_start = dest; |
|
4904 if (src_offset != NULL || dest_offset != NULL) { |
|
4905 assert(src_offset != NULL && dest_offset != NULL, ""); |
|
4906 src_start = array_element_address(src, src_offset, basic_elem_type); |
|
4907 dest_start = array_element_address(dest, dest_offset, basic_elem_type); |
|
4908 } |
|
4909 |
|
4910 // Figure out which arraycopy runtime method to call. |
|
4911 const char* copyfunc_name = "arraycopy"; |
|
4912 address copyfunc_addr = |
|
4913 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, |
|
4914 disjoint_bases, copyfunc_name); |
|
4915 |
|
4916 // Call it. Note that the count_ix value is not scaled to a byte-size. |
|
4917 make_runtime_call(RC_LEAF|RC_NO_FP, |
|
4918 OptoRuntime::fast_arraycopy_Type(), |
|
4919 copyfunc_addr, copyfunc_name, adr_type, |
|
4920 src_start, dest_start, copy_length XTOP); |
|
4921 } |