author | coleenp |
Wed, 14 Jan 2009 20:14:19 -0500 | |
changeset 1904 | 7aada8102b30 |
parent 1500 | bea9a90f3e8f |
child 2131 | 98f9cef66a34 |
permissions | -rw-r--r-- |
1 | 1 |
/* |
670 | 2 |
* Copyright 1999-2008 Sun Microsystems, Inc. All Rights Reserved. |
1 | 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); |
|
595
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
166 |
bool inline_array_equals(); |
1 | 167 |
bool inline_native_clone(bool is_virtual); |
168 |
bool inline_native_Reflection_getCallerClass(); |
|
169 |
bool inline_native_AtomicLong_get(); |
|
170 |
bool inline_native_AtomicLong_attemptUpdate(); |
|
171 |
bool is_method_invoke_or_aux_frame(JVMState* jvms); |
|
172 |
// Helper function for inlining native object hash method |
|
173 |
bool inline_native_hashcode(bool is_virtual, bool is_static); |
|
174 |
bool inline_native_getClass(); |
|
175 |
||
176 |
// Helper functions for inlining arraycopy |
|
177 |
bool inline_arraycopy(); |
|
178 |
void generate_arraycopy(const TypePtr* adr_type, |
|
179 |
BasicType basic_elem_type, |
|
180 |
Node* src, Node* src_offset, |
|
181 |
Node* dest, Node* dest_offset, |
|
182 |
Node* copy_length, |
|
183 |
int nargs, // arguments on stack for debug info |
|
184 |
bool disjoint_bases = false, |
|
185 |
bool length_never_negative = false, |
|
186 |
RegionNode* slow_region = NULL); |
|
187 |
AllocateArrayNode* tightly_coupled_allocation(Node* ptr, |
|
188 |
RegionNode* slow_region); |
|
189 |
void generate_clear_array(const TypePtr* adr_type, |
|
190 |
Node* dest, |
|
191 |
BasicType basic_elem_type, |
|
192 |
Node* slice_off, |
|
193 |
Node* slice_len, |
|
194 |
Node* slice_end); |
|
195 |
bool generate_block_arraycopy(const TypePtr* adr_type, |
|
196 |
BasicType basic_elem_type, |
|
197 |
AllocateNode* alloc, |
|
198 |
Node* src, Node* src_offset, |
|
199 |
Node* dest, Node* dest_offset, |
|
200 |
Node* dest_size); |
|
201 |
void generate_slow_arraycopy(const TypePtr* adr_type, |
|
202 |
Node* src, Node* src_offset, |
|
203 |
Node* dest, Node* dest_offset, |
|
204 |
Node* copy_length, |
|
205 |
int nargs); |
|
206 |
Node* generate_checkcast_arraycopy(const TypePtr* adr_type, |
|
207 |
Node* dest_elem_klass, |
|
208 |
Node* src, Node* src_offset, |
|
209 |
Node* dest, Node* dest_offset, |
|
210 |
Node* copy_length, int nargs); |
|
211 |
Node* generate_generic_arraycopy(const TypePtr* adr_type, |
|
212 |
Node* src, Node* src_offset, |
|
213 |
Node* dest, Node* dest_offset, |
|
214 |
Node* copy_length, int nargs); |
|
215 |
void generate_unchecked_arraycopy(const TypePtr* adr_type, |
|
216 |
BasicType basic_elem_type, |
|
217 |
bool disjoint_bases, |
|
218 |
Node* src, Node* src_offset, |
|
219 |
Node* dest, Node* dest_offset, |
|
220 |
Node* copy_length); |
|
221 |
bool inline_unsafe_CAS(BasicType type); |
|
222 |
bool inline_unsafe_ordered_store(BasicType type); |
|
223 |
bool inline_fp_conversions(vmIntrinsics::ID id); |
|
224 |
bool inline_reverseBytes(vmIntrinsics::ID id); |
|
225 |
}; |
|
226 |
||
227 |
||
228 |
//---------------------------make_vm_intrinsic---------------------------- |
|
229 |
CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { |
|
230 |
vmIntrinsics::ID id = m->intrinsic_id(); |
|
231 |
assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); |
|
232 |
||
233 |
if (DisableIntrinsic[0] != '\0' |
|
234 |
&& strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { |
|
235 |
// disabled by a user request on the command line: |
|
236 |
// example: -XX:DisableIntrinsic=_hashCode,_getClass |
|
237 |
return NULL; |
|
238 |
} |
|
239 |
||
240 |
if (!m->is_loaded()) { |
|
241 |
// do not attempt to inline unloaded methods |
|
242 |
return NULL; |
|
243 |
} |
|
244 |
||
245 |
// Only a few intrinsics implement a virtual dispatch. |
|
246 |
// They are expensive calls which are also frequently overridden. |
|
247 |
if (is_virtual) { |
|
248 |
switch (id) { |
|
249 |
case vmIntrinsics::_hashCode: |
|
250 |
case vmIntrinsics::_clone: |
|
251 |
// OK, Object.hashCode and Object.clone intrinsics come in both flavors |
|
252 |
break; |
|
253 |
default: |
|
254 |
return NULL; |
|
255 |
} |
|
256 |
} |
|
257 |
||
258 |
// -XX:-InlineNatives disables nearly all intrinsics: |
|
259 |
if (!InlineNatives) { |
|
260 |
switch (id) { |
|
261 |
case vmIntrinsics::_indexOf: |
|
262 |
case vmIntrinsics::_compareTo: |
|
595
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
263 |
case vmIntrinsics::_equalsC: |
1 | 264 |
break; // InlineNatives does not control String.compareTo |
265 |
default: |
|
266 |
return NULL; |
|
267 |
} |
|
268 |
} |
|
269 |
||
270 |
switch (id) { |
|
271 |
case vmIntrinsics::_compareTo: |
|
272 |
if (!SpecialStringCompareTo) return NULL; |
|
273 |
break; |
|
274 |
case vmIntrinsics::_indexOf: |
|
275 |
if (!SpecialStringIndexOf) return NULL; |
|
276 |
break; |
|
595
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
277 |
case vmIntrinsics::_equalsC: |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
278 |
if (!SpecialArraysEquals) return NULL; |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
279 |
break; |
1 | 280 |
case vmIntrinsics::_arraycopy: |
281 |
if (!InlineArrayCopy) return NULL; |
|
282 |
break; |
|
283 |
case vmIntrinsics::_copyMemory: |
|
284 |
if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; |
|
285 |
if (!InlineArrayCopy) return NULL; |
|
286 |
break; |
|
287 |
case vmIntrinsics::_hashCode: |
|
288 |
if (!InlineObjectHash) return NULL; |
|
289 |
break; |
|
290 |
case vmIntrinsics::_clone: |
|
291 |
case vmIntrinsics::_copyOf: |
|
292 |
case vmIntrinsics::_copyOfRange: |
|
293 |
if (!InlineObjectCopy) return NULL; |
|
294 |
// These also use the arraycopy intrinsic mechanism: |
|
295 |
if (!InlineArrayCopy) return NULL; |
|
296 |
break; |
|
297 |
case vmIntrinsics::_checkIndex: |
|
298 |
// We do not intrinsify this. The optimizer does fine with it. |
|
299 |
return NULL; |
|
300 |
||
301 |
case vmIntrinsics::_get_AtomicLong: |
|
302 |
case vmIntrinsics::_attemptUpdate: |
|
303 |
if (!InlineAtomicLong) return NULL; |
|
304 |
break; |
|
305 |
||
306 |
case vmIntrinsics::_Object_init: |
|
307 |
case vmIntrinsics::_invoke: |
|
308 |
// We do not intrinsify these; they are marked for other purposes. |
|
309 |
return NULL; |
|
310 |
||
311 |
case vmIntrinsics::_getCallerClass: |
|
312 |
if (!UseNewReflection) return NULL; |
|
313 |
if (!InlineReflectionGetCallerClass) return NULL; |
|
314 |
if (!JDK_Version::is_gte_jdk14x_version()) return NULL; |
|
315 |
break; |
|
316 |
||
317 |
default: |
|
318 |
break; |
|
319 |
} |
|
320 |
||
321 |
// -XX:-InlineClassNatives disables natives from the Class class. |
|
322 |
// The flag applies to all reflective calls, notably Array.newArray |
|
323 |
// (visible to Java programmers as Array.newInstance). |
|
324 |
if (m->holder()->name() == ciSymbol::java_lang_Class() || |
|
325 |
m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { |
|
326 |
if (!InlineClassNatives) return NULL; |
|
327 |
} |
|
328 |
||
329 |
// -XX:-InlineThreadNatives disables natives from the Thread class. |
|
330 |
if (m->holder()->name() == ciSymbol::java_lang_Thread()) { |
|
331 |
if (!InlineThreadNatives) return NULL; |
|
332 |
} |
|
333 |
||
334 |
// -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. |
|
335 |
if (m->holder()->name() == ciSymbol::java_lang_Math() || |
|
336 |
m->holder()->name() == ciSymbol::java_lang_Float() || |
|
337 |
m->holder()->name() == ciSymbol::java_lang_Double()) { |
|
338 |
if (!InlineMathNatives) return NULL; |
|
339 |
} |
|
340 |
||
341 |
// -XX:-InlineUnsafeOps disables natives from the Unsafe class. |
|
342 |
if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { |
|
343 |
if (!InlineUnsafeOps) return NULL; |
|
344 |
} |
|
345 |
||
346 |
return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); |
|
347 |
} |
|
348 |
||
349 |
//----------------------register_library_intrinsics----------------------- |
|
350 |
// Initialize this file's data structures, for each Compile instance. |
|
351 |
void Compile::register_library_intrinsics() { |
|
352 |
// Nothing to do here. |
|
353 |
} |
|
354 |
||
355 |
JVMState* LibraryIntrinsic::generate(JVMState* jvms) { |
|
356 |
LibraryCallKit kit(jvms, this); |
|
357 |
Compile* C = kit.C; |
|
358 |
int nodes = C->unique(); |
|
359 |
#ifndef PRODUCT |
|
360 |
if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { |
|
361 |
char buf[1000]; |
|
362 |
const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); |
|
363 |
tty->print_cr("Intrinsic %s", str); |
|
364 |
} |
|
365 |
#endif |
|
366 |
if (kit.try_to_inline()) { |
|
367 |
if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { |
|
368 |
tty->print("Inlining intrinsic %s%s at bci:%d in", |
|
369 |
vmIntrinsics::name_at(intrinsic_id()), |
|
370 |
(is_virtual() ? " (virtual)" : ""), kit.bci()); |
|
371 |
kit.caller()->print_short_name(tty); |
|
372 |
tty->print_cr(" (%d bytes)", kit.caller()->code_size()); |
|
373 |
} |
|
374 |
C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); |
|
375 |
if (C->log()) { |
|
376 |
C->log()->elem("intrinsic id='%s'%s nodes='%d'", |
|
377 |
vmIntrinsics::name_at(intrinsic_id()), |
|
378 |
(is_virtual() ? " virtual='1'" : ""), |
|
379 |
C->unique() - nodes); |
|
380 |
} |
|
381 |
return kit.transfer_exceptions_into_jvms(); |
|
382 |
} |
|
383 |
||
384 |
if (PrintIntrinsics) { |
|
385 |
switch (intrinsic_id()) { |
|
386 |
case vmIntrinsics::_invoke: |
|
387 |
case vmIntrinsics::_Object_init: |
|
388 |
// We do not expect to inline these, so do not produce any noise about them. |
|
389 |
break; |
|
390 |
default: |
|
391 |
tty->print("Did not inline intrinsic %s%s at bci:%d in", |
|
392 |
vmIntrinsics::name_at(intrinsic_id()), |
|
393 |
(is_virtual() ? " (virtual)" : ""), kit.bci()); |
|
394 |
kit.caller()->print_short_name(tty); |
|
395 |
tty->print_cr(" (%d bytes)", kit.caller()->code_size()); |
|
396 |
} |
|
397 |
} |
|
398 |
C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); |
|
399 |
return NULL; |
|
400 |
} |
|
401 |
||
402 |
bool LibraryCallKit::try_to_inline() { |
|
403 |
// Handle symbolic names for otherwise undistinguished boolean switches: |
|
404 |
const bool is_store = true; |
|
405 |
const bool is_native_ptr = true; |
|
406 |
const bool is_static = true; |
|
407 |
||
408 |
switch (intrinsic_id()) { |
|
409 |
case vmIntrinsics::_hashCode: |
|
410 |
return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); |
|
411 |
case vmIntrinsics::_identityHashCode: |
|
412 |
return inline_native_hashcode(/*!virtual*/ false, is_static); |
|
413 |
case vmIntrinsics::_getClass: |
|
414 |
return inline_native_getClass(); |
|
415 |
||
416 |
case vmIntrinsics::_dsin: |
|
417 |
case vmIntrinsics::_dcos: |
|
418 |
case vmIntrinsics::_dtan: |
|
419 |
case vmIntrinsics::_dabs: |
|
420 |
case vmIntrinsics::_datan2: |
|
421 |
case vmIntrinsics::_dsqrt: |
|
422 |
case vmIntrinsics::_dexp: |
|
423 |
case vmIntrinsics::_dlog: |
|
424 |
case vmIntrinsics::_dlog10: |
|
425 |
case vmIntrinsics::_dpow: |
|
426 |
return inline_math_native(intrinsic_id()); |
|
427 |
||
428 |
case vmIntrinsics::_min: |
|
429 |
case vmIntrinsics::_max: |
|
430 |
return inline_min_max(intrinsic_id()); |
|
431 |
||
432 |
case vmIntrinsics::_arraycopy: |
|
433 |
return inline_arraycopy(); |
|
434 |
||
435 |
case vmIntrinsics::_compareTo: |
|
436 |
return inline_string_compareTo(); |
|
437 |
case vmIntrinsics::_indexOf: |
|
438 |
return inline_string_indexOf(); |
|
439 |
||
440 |
case vmIntrinsics::_getObject: |
|
441 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); |
|
442 |
case vmIntrinsics::_getBoolean: |
|
443 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); |
|
444 |
case vmIntrinsics::_getByte: |
|
445 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); |
|
446 |
case vmIntrinsics::_getShort: |
|
447 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); |
|
448 |
case vmIntrinsics::_getChar: |
|
449 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); |
|
450 |
case vmIntrinsics::_getInt: |
|
451 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); |
|
452 |
case vmIntrinsics::_getLong: |
|
453 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); |
|
454 |
case vmIntrinsics::_getFloat: |
|
455 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); |
|
456 |
case vmIntrinsics::_getDouble: |
|
457 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); |
|
458 |
||
459 |
case vmIntrinsics::_putObject: |
|
460 |
return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); |
|
461 |
case vmIntrinsics::_putBoolean: |
|
462 |
return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); |
|
463 |
case vmIntrinsics::_putByte: |
|
464 |
return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); |
|
465 |
case vmIntrinsics::_putShort: |
|
466 |
return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); |
|
467 |
case vmIntrinsics::_putChar: |
|
468 |
return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); |
|
469 |
case vmIntrinsics::_putInt: |
|
470 |
return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); |
|
471 |
case vmIntrinsics::_putLong: |
|
472 |
return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); |
|
473 |
case vmIntrinsics::_putFloat: |
|
474 |
return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); |
|
475 |
case vmIntrinsics::_putDouble: |
|
476 |
return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); |
|
477 |
||
478 |
case vmIntrinsics::_getByte_raw: |
|
479 |
return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); |
|
480 |
case vmIntrinsics::_getShort_raw: |
|
481 |
return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); |
|
482 |
case vmIntrinsics::_getChar_raw: |
|
483 |
return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); |
|
484 |
case vmIntrinsics::_getInt_raw: |
|
485 |
return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); |
|
486 |
case vmIntrinsics::_getLong_raw: |
|
487 |
return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); |
|
488 |
case vmIntrinsics::_getFloat_raw: |
|
489 |
return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); |
|
490 |
case vmIntrinsics::_getDouble_raw: |
|
491 |
return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); |
|
492 |
case vmIntrinsics::_getAddress_raw: |
|
493 |
return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); |
|
494 |
||
495 |
case vmIntrinsics::_putByte_raw: |
|
496 |
return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); |
|
497 |
case vmIntrinsics::_putShort_raw: |
|
498 |
return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); |
|
499 |
case vmIntrinsics::_putChar_raw: |
|
500 |
return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); |
|
501 |
case vmIntrinsics::_putInt_raw: |
|
502 |
return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); |
|
503 |
case vmIntrinsics::_putLong_raw: |
|
504 |
return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); |
|
505 |
case vmIntrinsics::_putFloat_raw: |
|
506 |
return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); |
|
507 |
case vmIntrinsics::_putDouble_raw: |
|
508 |
return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); |
|
509 |
case vmIntrinsics::_putAddress_raw: |
|
510 |
return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); |
|
511 |
||
512 |
case vmIntrinsics::_getObjectVolatile: |
|
513 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); |
|
514 |
case vmIntrinsics::_getBooleanVolatile: |
|
515 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); |
|
516 |
case vmIntrinsics::_getByteVolatile: |
|
517 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); |
|
518 |
case vmIntrinsics::_getShortVolatile: |
|
519 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); |
|
520 |
case vmIntrinsics::_getCharVolatile: |
|
521 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); |
|
522 |
case vmIntrinsics::_getIntVolatile: |
|
523 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); |
|
524 |
case vmIntrinsics::_getLongVolatile: |
|
525 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); |
|
526 |
case vmIntrinsics::_getFloatVolatile: |
|
527 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); |
|
528 |
case vmIntrinsics::_getDoubleVolatile: |
|
529 |
return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); |
|
530 |
||
531 |
case vmIntrinsics::_putObjectVolatile: |
|
532 |
return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); |
|
533 |
case vmIntrinsics::_putBooleanVolatile: |
|
534 |
return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); |
|
535 |
case vmIntrinsics::_putByteVolatile: |
|
536 |
return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); |
|
537 |
case vmIntrinsics::_putShortVolatile: |
|
538 |
return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); |
|
539 |
case vmIntrinsics::_putCharVolatile: |
|
540 |
return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); |
|
541 |
case vmIntrinsics::_putIntVolatile: |
|
542 |
return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); |
|
543 |
case vmIntrinsics::_putLongVolatile: |
|
544 |
return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); |
|
545 |
case vmIntrinsics::_putFloatVolatile: |
|
546 |
return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); |
|
547 |
case vmIntrinsics::_putDoubleVolatile: |
|
548 |
return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); |
|
549 |
||
550 |
case vmIntrinsics::_prefetchRead: |
|
551 |
return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); |
|
552 |
case vmIntrinsics::_prefetchWrite: |
|
553 |
return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); |
|
554 |
case vmIntrinsics::_prefetchReadStatic: |
|
555 |
return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); |
|
556 |
case vmIntrinsics::_prefetchWriteStatic: |
|
557 |
return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); |
|
558 |
||
559 |
case vmIntrinsics::_compareAndSwapObject: |
|
560 |
return inline_unsafe_CAS(T_OBJECT); |
|
561 |
case vmIntrinsics::_compareAndSwapInt: |
|
562 |
return inline_unsafe_CAS(T_INT); |
|
563 |
case vmIntrinsics::_compareAndSwapLong: |
|
564 |
return inline_unsafe_CAS(T_LONG); |
|
565 |
||
566 |
case vmIntrinsics::_putOrderedObject: |
|
567 |
return inline_unsafe_ordered_store(T_OBJECT); |
|
568 |
case vmIntrinsics::_putOrderedInt: |
|
569 |
return inline_unsafe_ordered_store(T_INT); |
|
570 |
case vmIntrinsics::_putOrderedLong: |
|
571 |
return inline_unsafe_ordered_store(T_LONG); |
|
572 |
||
573 |
case vmIntrinsics::_currentThread: |
|
574 |
return inline_native_currentThread(); |
|
575 |
case vmIntrinsics::_isInterrupted: |
|
576 |
return inline_native_isInterrupted(); |
|
577 |
||
578 |
case vmIntrinsics::_currentTimeMillis: |
|
579 |
return inline_native_time_funcs(false); |
|
580 |
case vmIntrinsics::_nanoTime: |
|
581 |
return inline_native_time_funcs(true); |
|
582 |
case vmIntrinsics::_allocateInstance: |
|
583 |
return inline_unsafe_allocate(); |
|
584 |
case vmIntrinsics::_copyMemory: |
|
585 |
return inline_unsafe_copyMemory(); |
|
586 |
case vmIntrinsics::_newArray: |
|
587 |
return inline_native_newArray(); |
|
588 |
case vmIntrinsics::_getLength: |
|
589 |
return inline_native_getLength(); |
|
590 |
case vmIntrinsics::_copyOf: |
|
591 |
return inline_array_copyOf(false); |
|
592 |
case vmIntrinsics::_copyOfRange: |
|
593 |
return inline_array_copyOf(true); |
|
595
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
594 |
case vmIntrinsics::_equalsC: |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
595 |
return inline_array_equals(); |
1 | 596 |
case vmIntrinsics::_clone: |
597 |
return inline_native_clone(intrinsic()->is_virtual()); |
|
598 |
||
599 |
case vmIntrinsics::_isAssignableFrom: |
|
600 |
return inline_native_subtype_check(); |
|
601 |
||
602 |
case vmIntrinsics::_isInstance: |
|
603 |
case vmIntrinsics::_getModifiers: |
|
604 |
case vmIntrinsics::_isInterface: |
|
605 |
case vmIntrinsics::_isArray: |
|
606 |
case vmIntrinsics::_isPrimitive: |
|
607 |
case vmIntrinsics::_getSuperclass: |
|
608 |
case vmIntrinsics::_getComponentType: |
|
609 |
case vmIntrinsics::_getClassAccessFlags: |
|
610 |
return inline_native_Class_query(intrinsic_id()); |
|
611 |
||
612 |
case vmIntrinsics::_floatToRawIntBits: |
|
613 |
case vmIntrinsics::_floatToIntBits: |
|
614 |
case vmIntrinsics::_intBitsToFloat: |
|
615 |
case vmIntrinsics::_doubleToRawLongBits: |
|
616 |
case vmIntrinsics::_doubleToLongBits: |
|
617 |
case vmIntrinsics::_longBitsToDouble: |
|
618 |
return inline_fp_conversions(intrinsic_id()); |
|
619 |
||
620 |
case vmIntrinsics::_reverseBytes_i: |
|
621 |
case vmIntrinsics::_reverseBytes_l: |
|
622 |
return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); |
|
623 |
||
624 |
case vmIntrinsics::_get_AtomicLong: |
|
625 |
return inline_native_AtomicLong_get(); |
|
626 |
case vmIntrinsics::_attemptUpdate: |
|
627 |
return inline_native_AtomicLong_attemptUpdate(); |
|
628 |
||
629 |
case vmIntrinsics::_getCallerClass: |
|
630 |
return inline_native_Reflection_getCallerClass(); |
|
631 |
||
632 |
default: |
|
633 |
// If you get here, it may be that someone has added a new intrinsic |
|
634 |
// to the list in vmSymbols.hpp without implementing it here. |
|
635 |
#ifndef PRODUCT |
|
636 |
if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { |
|
637 |
tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", |
|
638 |
vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); |
|
639 |
} |
|
640 |
#endif |
|
641 |
return false; |
|
642 |
} |
|
643 |
} |
|
644 |
||
645 |
//------------------------------push_result------------------------------ |
|
646 |
// Helper function for finishing intrinsics. |
|
647 |
void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { |
|
648 |
record_for_igvn(region); |
|
649 |
set_control(_gvn.transform(region)); |
|
650 |
BasicType value_type = value->type()->basic_type(); |
|
651 |
push_node(value_type, _gvn.transform(value)); |
|
652 |
} |
|
653 |
||
654 |
//------------------------------generate_guard--------------------------- |
|
655 |
// Helper function for generating guarded fast-slow graph structures. |
|
656 |
// The given 'test', if true, guards a slow path. If the test fails |
|
657 |
// then a fast path can be taken. (We generally hope it fails.) |
|
658 |
// In all cases, GraphKit::control() is updated to the fast path. |
|
659 |
// The returned value represents the control for the slow path. |
|
660 |
// The return value is never 'top'; it is either a valid control |
|
661 |
// or NULL if it is obvious that the slow path can never be taken. |
|
662 |
// Also, if region and the slow control are not NULL, the slow edge |
|
663 |
// is appended to the region. |
|
664 |
Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { |
|
665 |
if (stopped()) { |
|
666 |
// Already short circuited. |
|
667 |
return NULL; |
|
668 |
} |
|
669 |
||
670 |
// Build an if node and its projections. |
|
671 |
// If test is true we take the slow path, which we assume is uncommon. |
|
672 |
if (_gvn.type(test) == TypeInt::ZERO) { |
|
673 |
// The slow branch is never taken. No need to build this guard. |
|
674 |
return NULL; |
|
675 |
} |
|
676 |
||
677 |
IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); |
|
678 |
||
679 |
Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); |
|
680 |
if (if_slow == top()) { |
|
681 |
// The slow branch is never taken. No need to build this guard. |
|
682 |
return NULL; |
|
683 |
} |
|
684 |
||
685 |
if (region != NULL) |
|
686 |
region->add_req(if_slow); |
|
687 |
||
688 |
Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); |
|
689 |
set_control(if_fast); |
|
690 |
||
691 |
return if_slow; |
|
692 |
} |
|
693 |
||
694 |
inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { |
|
695 |
return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); |
|
696 |
} |
|
697 |
inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { |
|
698 |
return generate_guard(test, region, PROB_FAIR); |
|
699 |
} |
|
700 |
||
701 |
inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, |
|
702 |
Node* *pos_index) { |
|
703 |
if (stopped()) |
|
704 |
return NULL; // already stopped |
|
705 |
if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] |
|
706 |
return NULL; // index is already adequately typed |
|
707 |
Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); |
|
708 |
Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); |
|
709 |
Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); |
|
710 |
if (is_neg != NULL && pos_index != NULL) { |
|
711 |
// Emulate effect of Parse::adjust_map_after_if. |
|
712 |
Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); |
|
713 |
ccast->set_req(0, control()); |
|
714 |
(*pos_index) = _gvn.transform(ccast); |
|
715 |
} |
|
716 |
return is_neg; |
|
717 |
} |
|
718 |
||
719 |
inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, |
|
720 |
Node* *pos_index) { |
|
721 |
if (stopped()) |
|
722 |
return NULL; // already stopped |
|
723 |
if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] |
|
724 |
return NULL; // index is already adequately typed |
|
725 |
Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); |
|
726 |
BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); |
|
727 |
Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); |
|
728 |
Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); |
|
729 |
if (is_notp != NULL && pos_index != NULL) { |
|
730 |
// Emulate effect of Parse::adjust_map_after_if. |
|
731 |
Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); |
|
732 |
ccast->set_req(0, control()); |
|
733 |
(*pos_index) = _gvn.transform(ccast); |
|
734 |
} |
|
735 |
return is_notp; |
|
736 |
} |
|
737 |
||
738 |
// Make sure that 'position' is a valid limit index, in [0..length]. |
|
739 |
// There are two equivalent plans for checking this: |
|
740 |
// A. (offset + copyLength) unsigned<= arrayLength |
|
741 |
// B. offset <= (arrayLength - copyLength) |
|
742 |
// We require that all of the values above, except for the sum and |
|
743 |
// difference, are already known to be non-negative. |
|
744 |
// Plan A is robust in the face of overflow, if offset and copyLength |
|
745 |
// are both hugely positive. |
|
746 |
// |
|
747 |
// Plan B is less direct and intuitive, but it does not overflow at |
|
748 |
// all, since the difference of two non-negatives is always |
|
749 |
// representable. Whenever Java methods must perform the equivalent |
|
750 |
// check they generally use Plan B instead of Plan A. |
|
751 |
// For the moment we use Plan A. |
|
752 |
inline Node* LibraryCallKit::generate_limit_guard(Node* offset, |
|
753 |
Node* subseq_length, |
|
754 |
Node* array_length, |
|
755 |
RegionNode* region) { |
|
756 |
if (stopped()) |
|
757 |
return NULL; // already stopped |
|
758 |
bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; |
|
759 |
if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length)) |
|
760 |
return NULL; // common case of whole-array copy |
|
761 |
Node* last = subseq_length; |
|
762 |
if (!zero_offset) // last += offset |
|
763 |
last = _gvn.transform( new (C, 3) AddINode(last, offset)); |
|
764 |
Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); |
|
765 |
Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); |
|
766 |
Node* is_over = generate_guard(bol_lt, region, PROB_MIN); |
|
767 |
return is_over; |
|
768 |
} |
|
769 |
||
770 |
||
771 |
//--------------------------generate_current_thread-------------------- |
|
772 |
Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { |
|
773 |
ciKlass* thread_klass = env()->Thread_klass(); |
|
774 |
const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); |
|
775 |
Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); |
|
776 |
Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); |
|
777 |
Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); |
|
778 |
tls_output = thread; |
|
779 |
return threadObj; |
|
780 |
} |
|
781 |
||
782 |
||
783 |
//------------------------------inline_string_compareTo------------------------ |
|
784 |
bool LibraryCallKit::inline_string_compareTo() { |
|
785 |
||
786 |
const int value_offset = java_lang_String::value_offset_in_bytes(); |
|
787 |
const int count_offset = java_lang_String::count_offset_in_bytes(); |
|
788 |
const int offset_offset = java_lang_String::offset_offset_in_bytes(); |
|
789 |
||
790 |
_sp += 2; |
|
791 |
Node *argument = pop(); // pop non-receiver first: it was pushed second |
|
792 |
Node *receiver = pop(); |
|
793 |
||
794 |
// Null check on self without removing any arguments. The argument |
|
795 |
// null check technically happens in the wrong place, which can lead to |
|
796 |
// invalid stack traces when string compare is inlined into a method |
|
797 |
// which handles NullPointerExceptions. |
|
798 |
_sp += 2; |
|
799 |
receiver = do_null_check(receiver, T_OBJECT); |
|
800 |
argument = do_null_check(argument, T_OBJECT); |
|
801 |
_sp -= 2; |
|
802 |
if (stopped()) { |
|
803 |
return true; |
|
804 |
} |
|
805 |
||
806 |
ciInstanceKlass* klass = env()->String_klass(); |
|
807 |
const TypeInstPtr* string_type = |
|
808 |
TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); |
|
809 |
||
810 |
Node* compare = |
|
811 |
_gvn.transform(new (C, 7) StrCompNode( |
|
812 |
control(), |
|
813 |
memory(TypeAryPtr::CHARS), |
|
814 |
memory(string_type->add_offset(value_offset)), |
|
815 |
memory(string_type->add_offset(count_offset)), |
|
816 |
memory(string_type->add_offset(offset_offset)), |
|
817 |
receiver, |
|
818 |
argument)); |
|
819 |
push(compare); |
|
820 |
return true; |
|
821 |
} |
|
822 |
||
595
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
823 |
//------------------------------inline_array_equals---------------------------- |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
824 |
bool LibraryCallKit::inline_array_equals() { |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
825 |
|
600
f7b4f99c0199
6709972: runThese failed with assert(false,"bad AD file")
rasbold
parents:
595
diff
changeset
|
826 |
if (!Matcher::has_match_rule(Op_AryEq)) return false; |
f7b4f99c0199
6709972: runThese failed with assert(false,"bad AD file")
rasbold
parents:
595
diff
changeset
|
827 |
|
595
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
828 |
_sp += 2; |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
829 |
Node *argument2 = pop(); |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
830 |
Node *argument1 = pop(); |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
831 |
|
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
832 |
Node* equals = |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
833 |
_gvn.transform(new (C, 3) AryEqNode(control(), |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
834 |
argument1, |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
835 |
argument2) |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
836 |
); |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
837 |
push(equals); |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
838 |
return true; |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
839 |
} |
a2be4c89de81
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
590
diff
changeset
|
840 |
|
1 | 841 |
// Java version of String.indexOf(constant string) |
842 |
// class StringDecl { |
|
843 |
// StringDecl(char[] ca) { |
|
844 |
// offset = 0; |
|
845 |
// count = ca.length; |
|
846 |
// value = ca; |
|
847 |
// } |
|
848 |
// int offset; |
|
849 |
// int count; |
|
850 |
// char[] value; |
|
851 |
// } |
|
852 |
// |
|
853 |
// static int string_indexOf_J(StringDecl string_object, char[] target_object, |
|
854 |
// int targetOffset, int cache_i, int md2) { |
|
855 |
// int cache = cache_i; |
|
856 |
// int sourceOffset = string_object.offset; |
|
857 |
// int sourceCount = string_object.count; |
|
858 |
// int targetCount = target_object.length; |
|
859 |
// |
|
860 |
// int targetCountLess1 = targetCount - 1; |
|
861 |
// int sourceEnd = sourceOffset + sourceCount - targetCountLess1; |
|
862 |
// |
|
863 |
// char[] source = string_object.value; |
|
864 |
// char[] target = target_object; |
|
865 |
// int lastChar = target[targetCountLess1]; |
|
866 |
// |
|
867 |
// outer_loop: |
|
868 |
// for (int i = sourceOffset; i < sourceEnd; ) { |
|
869 |
// int src = source[i + targetCountLess1]; |
|
870 |
// if (src == lastChar) { |
|
871 |
// // With random strings and a 4-character alphabet, |
|
872 |
// // reverse matching at this point sets up 0.8% fewer |
|
873 |
// // frames, but (paradoxically) makes 0.3% more probes. |
|
874 |
// // Since those probes are nearer the lastChar probe, |
|
875 |
// // there is may be a net D$ win with reverse matching. |
|
876 |
// // But, reversing loop inhibits unroll of inner loop |
|
877 |
// // for unknown reason. So, does running outer loop from |
|
878 |
// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) |
|
879 |
// for (int j = 0; j < targetCountLess1; j++) { |
|
880 |
// if (target[targetOffset + j] != source[i+j]) { |
|
881 |
// if ((cache & (1 << source[i+j])) == 0) { |
|
882 |
// if (md2 < j+1) { |
|
883 |
// i += j+1; |
|
884 |
// continue outer_loop; |
|
885 |
// } |
|
886 |
// } |
|
887 |
// i += md2; |
|
888 |
// continue outer_loop; |
|
889 |
// } |
|
890 |
// } |
|
891 |
// return i - sourceOffset; |
|
892 |
// } |
|
893 |
// if ((cache & (1 << src)) == 0) { |
|
894 |
// i += targetCountLess1; |
|
895 |
// } // using "i += targetCount;" and an "else i++;" causes a jump to jump. |
|
896 |
// i++; |
|
897 |
// } |
|
898 |
// return -1; |
|
899 |
// } |
|
900 |
||
901 |
//------------------------------string_indexOf------------------------ |
|
902 |
Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, |
|
903 |
jint cache_i, jint md2_i) { |
|
904 |
||
905 |
Node* no_ctrl = NULL; |
|
906 |
float likely = PROB_LIKELY(0.9); |
|
907 |
float unlikely = PROB_UNLIKELY(0.9); |
|
908 |
||
909 |
const int value_offset = java_lang_String::value_offset_in_bytes(); |
|
910 |
const int count_offset = java_lang_String::count_offset_in_bytes(); |
|
911 |
const int offset_offset = java_lang_String::offset_offset_in_bytes(); |
|
912 |
||
913 |
ciInstanceKlass* klass = env()->String_klass(); |
|
914 |
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); |
|
915 |
const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); |
|
916 |
||
917 |
Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); |
|
918 |
Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); |
|
919 |
Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); |
|
920 |
Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); |
|
921 |
Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); |
|
922 |
Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); |
|
923 |
||
590
2954744d7bba
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
589
diff
changeset
|
924 |
Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) ); |
1 | 925 |
jint target_length = target_array->length(); |
926 |
const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); |
|
927 |
const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); |
|
928 |
||
929 |
IdealKit kit(gvn(), control(), merged_memory()); |
|
930 |
#define __ kit. |
|
931 |
Node* zero = __ ConI(0); |
|
932 |
Node* one = __ ConI(1); |
|
933 |
Node* cache = __ ConI(cache_i); |
|
934 |
Node* md2 = __ ConI(md2_i); |
|
935 |
Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); |
|
936 |
Node* targetCount = __ ConI(target_length); |
|
937 |
Node* targetCountLess1 = __ ConI(target_length - 1); |
|
938 |
Node* targetOffset = __ ConI(targetOffset_i); |
|
939 |
Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); |
|
940 |
||
941 |
IdealVariable rtn(kit), i(kit), j(kit); __ declares_done(); |
|
942 |
Node* outer_loop = __ make_label(2 /* goto */); |
|
943 |
Node* return_ = __ make_label(1); |
|
944 |
||
945 |
__ set(rtn,__ ConI(-1)); |
|
946 |
__ loop(i, sourceOffset, BoolTest::lt, sourceEnd); { |
|
947 |
Node* i2 = __ AddI(__ value(i), targetCountLess1); |
|
948 |
// pin to prohibit loading of "next iteration" value which may SEGV (rare) |
|
949 |
Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); |
|
950 |
__ if_then(src, BoolTest::eq, lastChar, unlikely); { |
|
951 |
__ loop(j, zero, BoolTest::lt, targetCountLess1); { |
|
952 |
Node* tpj = __ AddI(targetOffset, __ value(j)); |
|
953 |
Node* targ = load_array_element(no_ctrl, target, tpj, target_type); |
|
954 |
Node* ipj = __ AddI(__ value(i), __ value(j)); |
|
955 |
Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); |
|
956 |
__ if_then(targ, BoolTest::ne, src2); { |
|
957 |
__ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { |
|
958 |
__ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { |
|
959 |
__ increment(i, __ AddI(__ value(j), one)); |
|
960 |
__ goto_(outer_loop); |
|
961 |
} __ end_if(); __ dead(j); |
|
962 |
}__ end_if(); __ dead(j); |
|
963 |
__ increment(i, md2); |
|
964 |
__ goto_(outer_loop); |
|
965 |
}__ end_if(); |
|
966 |
__ increment(j, one); |
|
967 |
}__ end_loop(); __ dead(j); |
|
968 |
__ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); |
|
969 |
__ goto_(return_); |
|
970 |
}__ end_if(); |
|
971 |
__ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { |
|
972 |
__ increment(i, targetCountLess1); |
|
973 |
}__ end_if(); |
|
974 |
__ increment(i, one); |
|
975 |
__ bind(outer_loop); |
|
976 |
}__ end_loop(); __ dead(i); |
|
977 |
__ bind(return_); |
|
978 |
__ drain_delay_transform(); |
|
979 |
||
980 |
set_control(__ ctrl()); |
|
981 |
Node* result = __ value(rtn); |
|
982 |
#undef __ |
|
983 |
C->set_has_loops(true); |
|
984 |
return result; |
|
985 |
} |
|
986 |
||
987 |
||
988 |
//------------------------------inline_string_indexOf------------------------ |
|
989 |
bool LibraryCallKit::inline_string_indexOf() { |
|
990 |
||
991 |
_sp += 2; |
|
992 |
Node *argument = pop(); // pop non-receiver first: it was pushed second |
|
993 |
Node *receiver = pop(); |
|
994 |
||
995 |
// don't intrinsify is argument isn't a constant string. |
|
996 |
if (!argument->is_Con()) { |
|
997 |
return false; |
|
998 |
} |
|
999 |
const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); |
|
1000 |
if (str_type == NULL) { |
|
1001 |
return false; |
|
1002 |
} |
|
1003 |
ciInstanceKlass* klass = env()->String_klass(); |
|
1004 |
ciObject* str_const = str_type->const_oop(); |
|
1005 |
if (str_const == NULL || str_const->klass() != klass) { |
|
1006 |
return false; |
|
1007 |
} |
|
1008 |
ciInstance* str = str_const->as_instance(); |
|
1009 |
assert(str != NULL, "must be instance"); |
|
1010 |
||
1011 |
const int value_offset = java_lang_String::value_offset_in_bytes(); |
|
1012 |
const int count_offset = java_lang_String::count_offset_in_bytes(); |
|
1013 |
const int offset_offset = java_lang_String::offset_offset_in_bytes(); |
|
1014 |
||
1015 |
ciObject* v = str->field_value_by_offset(value_offset).as_object(); |
|
1016 |
int o = str->field_value_by_offset(offset_offset).as_int(); |
|
1017 |
int c = str->field_value_by_offset(count_offset).as_int(); |
|
1018 |
ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array |
|
1019 |
||
1020 |
// constant strings have no offset and count == length which |
|
1021 |
// simplifies the resulting code somewhat so lets optimize for that. |
|
1022 |
if (o != 0 || c != pat->length()) { |
|
1023 |
return false; |
|
1024 |
} |
|
1025 |
||
1026 |
// Null check on self without removing any arguments. The argument |
|
1027 |
// null check technically happens in the wrong place, which can lead to |
|
1028 |
// invalid stack traces when string compare is inlined into a method |
|
1029 |
// which handles NullPointerExceptions. |
|
1030 |
_sp += 2; |
|
1031 |
receiver = do_null_check(receiver, T_OBJECT); |
|
1032 |
// No null check on the argument is needed since it's a constant String oop. |
|
1033 |
_sp -= 2; |
|
1034 |
if (stopped()) { |
|
1035 |
return true; |
|
1036 |
} |
|
1037 |
||
1038 |
// The null string as a pattern always returns 0 (match at beginning of string) |
|
1039 |
if (c == 0) { |
|
1040 |
push(intcon(0)); |
|
1041 |
return true; |
|
1042 |
} |
|
1043 |
||
1044 |
jchar lastChar = pat->char_at(o + (c - 1)); |
|
1045 |
int cache = 0; |
|
1046 |
int i; |
|
1047 |
for (i = 0; i < c - 1; i++) { |
|
1048 |
assert(i < pat->length(), "out of range"); |
|
1049 |
cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); |
|
1050 |
} |
|
1051 |
||
1052 |
int md2 = c; |
|
1053 |
for (i = 0; i < c - 1; i++) { |
|
1054 |
assert(i < pat->length(), "out of range"); |
|
1055 |
if (pat->char_at(o + i) == lastChar) { |
|
1056 |
md2 = (c - 1) - i; |
|
1057 |
} |
|
1058 |
} |
|
1059 |
||
1060 |
Node* result = string_indexOf(receiver, pat, o, cache, md2); |
|
1061 |
push(result); |
|
1062 |
return true; |
|
1063 |
} |
|
1064 |
||
1065 |
//--------------------------pop_math_arg-------------------------------- |
|
1066 |
// Pop a double argument to a math function from the stack |
|
1067 |
// rounding it if necessary. |
|
1068 |
Node * LibraryCallKit::pop_math_arg() { |
|
1069 |
Node *arg = pop_pair(); |
|
1070 |
if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) |
|
1071 |
arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); |
|
1072 |
return arg; |
|
1073 |
} |
|
1074 |
||
1075 |
//------------------------------inline_trig---------------------------------- |
|
1076 |
// Inline sin/cos/tan instructions, if possible. If rounding is required, do |
|
1077 |
// argument reduction which will turn into a fast/slow diamond. |
|
1078 |
bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { |
|
1079 |
_sp += arg_size(); // restore stack pointer |
|
1080 |
Node* arg = pop_math_arg(); |
|
1081 |
Node* trig = NULL; |
|
1082 |
||
1083 |
switch (id) { |
|
1084 |
case vmIntrinsics::_dsin: |
|
1085 |
trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); |
|
1086 |
break; |
|
1087 |
case vmIntrinsics::_dcos: |
|
1088 |
trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); |
|
1089 |
break; |
|
1090 |
case vmIntrinsics::_dtan: |
|
1091 |
trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); |
|
1092 |
break; |
|
1093 |
default: |
|
1094 |
assert(false, "bad intrinsic was passed in"); |
|
1095 |
return false; |
|
1096 |
} |
|
1097 |
||
1098 |
// Rounding required? Check for argument reduction! |
|
1099 |
if( Matcher::strict_fp_requires_explicit_rounding ) { |
|
1100 |
||
1101 |
static const double pi_4 = 0.7853981633974483; |
|
1102 |
static const double neg_pi_4 = -0.7853981633974483; |
|
1103 |
// pi/2 in 80-bit extended precision |
|
1104 |
// static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; |
|
1105 |
// -pi/2 in 80-bit extended precision |
|
1106 |
// 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}; |
|
1107 |
// Cutoff value for using this argument reduction technique |
|
1108 |
//static const double pi_2_minus_epsilon = 1.564660403643354; |
|
1109 |
//static const double neg_pi_2_plus_epsilon = -1.564660403643354; |
|
1110 |
||
1111 |
// Pseudocode for sin: |
|
1112 |
// if (x <= Math.PI / 4.0) { |
|
1113 |
// if (x >= -Math.PI / 4.0) return fsin(x); |
|
1114 |
// if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); |
|
1115 |
// } else { |
|
1116 |
// if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); |
|
1117 |
// } |
|
1118 |
// return StrictMath.sin(x); |
|
1119 |
||
1120 |
// Pseudocode for cos: |
|
1121 |
// if (x <= Math.PI / 4.0) { |
|
1122 |
// if (x >= -Math.PI / 4.0) return fcos(x); |
|
1123 |
// if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); |
|
1124 |
// } else { |
|
1125 |
// if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); |
|
1126 |
// } |
|
1127 |
// return StrictMath.cos(x); |
|
1128 |
||
1129 |
// Actually, sticking in an 80-bit Intel value into C2 will be tough; it |
|
1130 |
// requires a special machine instruction to load it. Instead we'll try |
|
1131 |
// the 'easy' case. If we really need the extra range +/- PI/2 we'll |
|
1132 |
// probably do the math inside the SIN encoding. |
|
1133 |
||
1134 |
// Make the merge point |
|
1135 |
RegionNode *r = new (C, 3) RegionNode(3); |
|
1136 |
Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); |
|
1137 |
||
1138 |
// Flatten arg so we need only 1 test |
|
1139 |
Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); |
|
1140 |
// Node for PI/4 constant |
|
1141 |
Node *pi4 = makecon(TypeD::make(pi_4)); |
|
1142 |
// Check PI/4 : abs(arg) |
|
1143 |
Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); |
|
1144 |
// Check: If PI/4 < abs(arg) then go slow |
|
1145 |
Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); |
|
1146 |
// Branch either way |
|
1147 |
IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
|
1148 |
set_control(opt_iff(r,iff)); |
|
1149 |
||
1150 |
// Set fast path result |
|
1151 |
phi->init_req(2,trig); |
|
1152 |
||
1153 |
// Slow path - non-blocking leaf call |
|
1154 |
Node* call = NULL; |
|
1155 |
switch (id) { |
|
1156 |
case vmIntrinsics::_dsin: |
|
1157 |
call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), |
|
1158 |
CAST_FROM_FN_PTR(address, SharedRuntime::dsin), |
|
1159 |
"Sin", NULL, arg, top()); |
|
1160 |
break; |
|
1161 |
case vmIntrinsics::_dcos: |
|
1162 |
call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), |
|
1163 |
CAST_FROM_FN_PTR(address, SharedRuntime::dcos), |
|
1164 |
"Cos", NULL, arg, top()); |
|
1165 |
break; |
|
1166 |
case vmIntrinsics::_dtan: |
|
1167 |
call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), |
|
1168 |
CAST_FROM_FN_PTR(address, SharedRuntime::dtan), |
|
1169 |
"Tan", NULL, arg, top()); |
|
1170 |
break; |
|
1171 |
} |
|
1172 |
assert(control()->in(0) == call, ""); |
|
1173 |
Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); |
|
1174 |
r->init_req(1,control()); |
|
1175 |
phi->init_req(1,slow_result); |
|
1176 |
||
1177 |
// Post-merge |
|
1178 |
set_control(_gvn.transform(r)); |
|
1179 |
record_for_igvn(r); |
|
1180 |
trig = _gvn.transform(phi); |
|
1181 |
||
1182 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
1183 |
} |
|
1184 |
// Push result back on JVM stack |
|
1185 |
push_pair(trig); |
|
1186 |
return true; |
|
1187 |
} |
|
1188 |
||
1189 |
//------------------------------inline_sqrt------------------------------------- |
|
1190 |
// Inline square root instruction, if possible. |
|
1191 |
bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { |
|
1192 |
assert(id == vmIntrinsics::_dsqrt, "Not square root"); |
|
1193 |
_sp += arg_size(); // restore stack pointer |
|
1194 |
push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); |
|
1195 |
return true; |
|
1196 |
} |
|
1197 |
||
1198 |
//------------------------------inline_abs------------------------------------- |
|
1199 |
// Inline absolute value instruction, if possible. |
|
1200 |
bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { |
|
1201 |
assert(id == vmIntrinsics::_dabs, "Not absolute value"); |
|
1202 |
_sp += arg_size(); // restore stack pointer |
|
1203 |
push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); |
|
1204 |
return true; |
|
1205 |
} |
|
1206 |
||
1207 |
//------------------------------inline_exp------------------------------------- |
|
1208 |
// Inline exp instructions, if possible. The Intel hardware only misses |
|
1209 |
// really odd corner cases (+/- Infinity). Just uncommon-trap them. |
|
1210 |
bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { |
|
1211 |
assert(id == vmIntrinsics::_dexp, "Not exp"); |
|
1212 |
||
1213 |
// If this inlining ever returned NaN in the past, we do not intrinsify it |
|
1214 |
// every again. NaN results requires StrictMath.exp handling. |
|
1215 |
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
1216 |
||
1217 |
// Do not intrinsify on older platforms which lack cmove. |
|
1218 |
if (ConditionalMoveLimit == 0) return false; |
|
1219 |
||
1220 |
_sp += arg_size(); // restore stack pointer |
|
1221 |
Node *x = pop_math_arg(); |
|
1222 |
Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); |
|
1223 |
||
1224 |
//------------------- |
|
1225 |
//result=(result.isNaN())? StrictMath::exp():result; |
|
1226 |
// Check: If isNaN() by checking result!=result? then go to Strict Math |
|
1227 |
Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); |
|
1228 |
// Build the boolean node |
|
1229 |
Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); |
|
1230 |
||
1231 |
{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); |
|
1232 |
// End the current control-flow path |
|
1233 |
push_pair(x); |
|
1234 |
// Math.exp intrinsic returned a NaN, which requires StrictMath.exp |
|
1235 |
// to handle. Recompile without intrinsifying Math.exp |
|
1236 |
uncommon_trap(Deoptimization::Reason_intrinsic, |
|
1237 |
Deoptimization::Action_make_not_entrant); |
|
1238 |
} |
|
1239 |
||
1240 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
1241 |
||
1242 |
push_pair(result); |
|
1243 |
||
1244 |
return true; |
|
1245 |
} |
|
1246 |
||
1247 |
//------------------------------inline_pow------------------------------------- |
|
1248 |
// Inline power instructions, if possible. |
|
1249 |
bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { |
|
1250 |
assert(id == vmIntrinsics::_dpow, "Not pow"); |
|
1251 |
||
1252 |
// If this inlining ever returned NaN in the past, we do not intrinsify it |
|
1253 |
// every again. NaN results requires StrictMath.pow handling. |
|
1254 |
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
1255 |
||
1256 |
// Do not intrinsify on older platforms which lack cmove. |
|
1257 |
if (ConditionalMoveLimit == 0) return false; |
|
1258 |
||
1259 |
// Pseudocode for pow |
|
1260 |
// if (x <= 0.0) { |
|
1261 |
// if ((double)((int)y)==y) { // if y is int |
|
1262 |
// result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) |
|
1263 |
// } else { |
|
1264 |
// result = NaN; |
|
1265 |
// } |
|
1266 |
// } else { |
|
1267 |
// result = DPow(x,y); |
|
1268 |
// } |
|
1269 |
// if (result != result)? { |
|
1270 |
// ucommon_trap(); |
|
1271 |
// } |
|
1272 |
// return result; |
|
1273 |
||
1274 |
_sp += arg_size(); // restore stack pointer |
|
1275 |
Node* y = pop_math_arg(); |
|
1276 |
Node* x = pop_math_arg(); |
|
1277 |
||
1278 |
Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); |
|
1279 |
||
1280 |
// Short form: if not top-level (i.e., Math.pow but inlining Math.pow |
|
1281 |
// inside of something) then skip the fancy tests and just check for |
|
1282 |
// NaN result. |
|
1283 |
Node *result = NULL; |
|
1284 |
if( jvms()->depth() >= 1 ) { |
|
1285 |
result = fast_result; |
|
1286 |
} else { |
|
1287 |
||
1288 |
// Set the merge point for If node with condition of (x <= 0.0) |
|
1289 |
// There are four possible paths to region node and phi node |
|
1290 |
RegionNode *r = new (C, 4) RegionNode(4); |
|
1291 |
Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); |
|
1292 |
||
1293 |
// Build the first if node: if (x <= 0.0) |
|
1294 |
// Node for 0 constant |
|
1295 |
Node *zeronode = makecon(TypeD::ZERO); |
|
1296 |
// Check x:0 |
|
1297 |
Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); |
|
1298 |
// Check: If (x<=0) then go complex path |
|
1299 |
Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); |
|
1300 |
// Branch either way |
|
1301 |
IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); |
|
1302 |
Node *opt_test = _gvn.transform(if1); |
|
1303 |
//assert( opt_test->is_If(), "Expect an IfNode"); |
|
1304 |
IfNode *opt_if1 = (IfNode*)opt_test; |
|
1305 |
// Fast path taken; set region slot 3 |
|
1306 |
Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); |
|
1307 |
r->init_req(3,fast_taken); // Capture fast-control |
|
1308 |
||
1309 |
// Fast path not-taken, i.e. slow path |
|
1310 |
Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); |
|
1311 |
||
1312 |
// Set fast path result |
|
1313 |
Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); |
|
1314 |
phi->init_req(3, fast_result); |
|
1315 |
||
1316 |
// Complex path |
|
1317 |
// Build the second if node (if y is int) |
|
1318 |
// Node for (int)y |
|
1319 |
Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); |
|
1320 |
// Node for (double)((int) y) |
|
1321 |
Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); |
|
1322 |
// Check (double)((int) y) : y |
|
1323 |
Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); |
|
1324 |
// Check if (y isn't int) then go to slow path |
|
1325 |
||
1326 |
Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); |
|
1327 |
// Branch eith way |
|
1328 |
IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); |
|
1329 |
Node *slow_path = opt_iff(r,if2); // Set region path 2 |
|
1330 |
||
1331 |
// Calculate DPow(abs(x), y)*(1 & (int)y) |
|
1332 |
// Node for constant 1 |
|
1333 |
Node *conone = intcon(1); |
|
1334 |
// 1& (int)y |
|
1335 |
Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); |
|
1336 |
// zero node |
|
1337 |
Node *conzero = intcon(0); |
|
1338 |
// Check (1&(int)y)==0? |
|
1339 |
Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); |
|
1340 |
// Check if (1&(int)y)!=0?, if so the result is negative |
|
1341 |
Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); |
|
1342 |
// abs(x) |
|
1343 |
Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); |
|
1344 |
// abs(x)^y |
|
1345 |
Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); |
|
1346 |
// -abs(x)^y |
|
1347 |
Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); |
|
1348 |
// (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) |
|
1349 |
Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); |
|
1350 |
// Set complex path fast result |
|
1351 |
phi->init_req(2, signresult); |
|
1352 |
||
1353 |
static const jlong nan_bits = CONST64(0x7ff8000000000000); |
|
1354 |
Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN |
|
1355 |
r->init_req(1,slow_path); |
|
1356 |
phi->init_req(1,slow_result); |
|
1357 |
||
1358 |
// Post merge |
|
1359 |
set_control(_gvn.transform(r)); |
|
1360 |
record_for_igvn(r); |
|
1361 |
result=_gvn.transform(phi); |
|
1362 |
} |
|
1363 |
||
1364 |
//------------------- |
|
1365 |
//result=(result.isNaN())? uncommon_trap():result; |
|
1366 |
// Check: If isNaN() by checking result!=result? then go to Strict Math |
|
1367 |
Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); |
|
1368 |
// Build the boolean node |
|
1369 |
Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); |
|
1370 |
||
1371 |
{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); |
|
1372 |
// End the current control-flow path |
|
1373 |
push_pair(x); |
|
1374 |
push_pair(y); |
|
1375 |
// Math.pow intrinsic returned a NaN, which requires StrictMath.pow |
|
1376 |
// to handle. Recompile without intrinsifying Math.pow. |
|
1377 |
uncommon_trap(Deoptimization::Reason_intrinsic, |
|
1378 |
Deoptimization::Action_make_not_entrant); |
|
1379 |
} |
|
1380 |
||
1381 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
1382 |
||
1383 |
push_pair(result); |
|
1384 |
||
1385 |
return true; |
|
1386 |
} |
|
1387 |
||
1388 |
//------------------------------inline_trans------------------------------------- |
|
1389 |
// Inline transcendental instructions, if possible. The Intel hardware gets |
|
1390 |
// these right, no funny corner cases missed. |
|
1391 |
bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { |
|
1392 |
_sp += arg_size(); // restore stack pointer |
|
1393 |
Node* arg = pop_math_arg(); |
|
1394 |
Node* trans = NULL; |
|
1395 |
||
1396 |
switch (id) { |
|
1397 |
case vmIntrinsics::_dlog: |
|
1398 |
trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); |
|
1399 |
break; |
|
1400 |
case vmIntrinsics::_dlog10: |
|
1401 |
trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); |
|
1402 |
break; |
|
1403 |
default: |
|
1404 |
assert(false, "bad intrinsic was passed in"); |
|
1405 |
return false; |
|
1406 |
} |
|
1407 |
||
1408 |
// Push result back on JVM stack |
|
1409 |
push_pair(trans); |
|
1410 |
return true; |
|
1411 |
} |
|
1412 |
||
1413 |
//------------------------------runtime_math----------------------------- |
|
1414 |
bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { |
|
1415 |
Node* a = NULL; |
|
1416 |
Node* b = NULL; |
|
1417 |
||
1418 |
assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), |
|
1419 |
"must be (DD)D or (D)D type"); |
|
1420 |
||
1421 |
// Inputs |
|
1422 |
_sp += arg_size(); // restore stack pointer |
|
1423 |
if (call_type == OptoRuntime::Math_DD_D_Type()) { |
|
1424 |
b = pop_math_arg(); |
|
1425 |
} |
|
1426 |
a = pop_math_arg(); |
|
1427 |
||
1428 |
const TypePtr* no_memory_effects = NULL; |
|
1429 |
Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, |
|
1430 |
no_memory_effects, |
|
1431 |
a, top(), b, b ? top() : NULL); |
|
1432 |
Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); |
|
1433 |
#ifdef ASSERT |
|
1434 |
Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); |
|
1435 |
assert(value_top == top(), "second value must be top"); |
|
1436 |
#endif |
|
1437 |
||
1438 |
push_pair(value); |
|
1439 |
return true; |
|
1440 |
} |
|
1441 |
||
1442 |
//------------------------------inline_math_native----------------------------- |
|
1443 |
bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { |
|
1444 |
switch (id) { |
|
1445 |
// These intrinsics are not properly supported on all hardware |
|
1446 |
case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : |
|
1447 |
runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); |
|
1448 |
case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : |
|
1449 |
runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); |
|
1450 |
case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : |
|
1451 |
runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); |
|
1452 |
||
1453 |
case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : |
|
1454 |
runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); |
|
1455 |
case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : |
|
1456 |
runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); |
|
1457 |
||
1458 |
// These intrinsics are supported on all hardware |
|
1459 |
case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; |
|
1460 |
case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; |
|
1461 |
||
1462 |
// These intrinsics don't work on X86. The ad implementation doesn't |
|
1463 |
// handle NaN's properly. Instead of returning infinity, the ad |
|
1464 |
// implementation returns a NaN on overflow. See bug: 6304089 |
|
1465 |
// Once the ad implementations are fixed, change the code below |
|
1466 |
// to match the intrinsics above |
|
1467 |
||
1468 |
case vmIntrinsics::_dexp: return |
|
1469 |
runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); |
|
1470 |
case vmIntrinsics::_dpow: return |
|
1471 |
runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); |
|
1472 |
||
1473 |
// These intrinsics are not yet correctly implemented |
|
1474 |
case vmIntrinsics::_datan2: |
|
1475 |
return false; |
|
1476 |
||
1477 |
default: |
|
1478 |
ShouldNotReachHere(); |
|
1479 |
return false; |
|
1480 |
} |
|
1481 |
} |
|
1482 |
||
1483 |
static bool is_simple_name(Node* n) { |
|
1484 |
return (n->req() == 1 // constant |
|
1485 |
|| (n->is_Type() && n->as_Type()->type()->singleton()) |
|
1486 |
|| n->is_Proj() // parameter or return value |
|
1487 |
|| n->is_Phi() // local of some sort |
|
1488 |
); |
|
1489 |
} |
|
1490 |
||
1491 |
//----------------------------inline_min_max----------------------------------- |
|
1492 |
bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { |
|
1493 |
push(generate_min_max(id, argument(0), argument(1))); |
|
1494 |
||
1495 |
return true; |
|
1496 |
} |
|
1497 |
||
1498 |
Node* |
|
1499 |
LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { |
|
1500 |
// These are the candidate return value: |
|
1501 |
Node* xvalue = x0; |
|
1502 |
Node* yvalue = y0; |
|
1503 |
||
1504 |
if (xvalue == yvalue) { |
|
1505 |
return xvalue; |
|
1506 |
} |
|
1507 |
||
1508 |
bool want_max = (id == vmIntrinsics::_max); |
|
1509 |
||
1510 |
const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); |
|
1511 |
const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); |
|
1512 |
if (txvalue == NULL || tyvalue == NULL) return top(); |
|
1513 |
// This is not really necessary, but it is consistent with a |
|
1514 |
// hypothetical MaxINode::Value method: |
|
1515 |
int widen = MAX2(txvalue->_widen, tyvalue->_widen); |
|
1516 |
||
1517 |
// %%% This folding logic should (ideally) be in a different place. |
|
1518 |
// Some should be inside IfNode, and there to be a more reliable |
|
1519 |
// transformation of ?: style patterns into cmoves. We also want |
|
1520 |
// more powerful optimizations around cmove and min/max. |
|
1521 |
||
1522 |
// Try to find a dominating comparison of these guys. |
|
1523 |
// It can simplify the index computation for Arrays.copyOf |
|
1524 |
// and similar uses of System.arraycopy. |
|
1525 |
// First, compute the normalized version of CmpI(x, y). |
|
1526 |
int cmp_op = Op_CmpI; |
|
1527 |
Node* xkey = xvalue; |
|
1528 |
Node* ykey = yvalue; |
|
1529 |
Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); |
|
1530 |
if (ideal_cmpxy->is_Cmp()) { |
|
1531 |
// E.g., if we have CmpI(length - offset, count), |
|
1532 |
// it might idealize to CmpI(length, count + offset) |
|
1533 |
cmp_op = ideal_cmpxy->Opcode(); |
|
1534 |
xkey = ideal_cmpxy->in(1); |
|
1535 |
ykey = ideal_cmpxy->in(2); |
|
1536 |
} |
|
1537 |
||
1538 |
// Start by locating any relevant comparisons. |
|
1539 |
Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; |
|
1540 |
Node* cmpxy = NULL; |
|
1541 |
Node* cmpyx = NULL; |
|
1542 |
for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { |
|
1543 |
Node* cmp = start_from->fast_out(k); |
|
1544 |
if (cmp->outcnt() > 0 && // must have prior uses |
|
1545 |
cmp->in(0) == NULL && // must be context-independent |
|
1546 |
cmp->Opcode() == cmp_op) { // right kind of compare |
|
1547 |
if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; |
|
1548 |
if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; |
|
1549 |
} |
|
1550 |
} |
|
1551 |
||
1552 |
const int NCMPS = 2; |
|
1553 |
Node* cmps[NCMPS] = { cmpxy, cmpyx }; |
|
1554 |
int cmpn; |
|
1555 |
for (cmpn = 0; cmpn < NCMPS; cmpn++) { |
|
1556 |
if (cmps[cmpn] != NULL) break; // find a result |
|
1557 |
} |
|
1558 |
if (cmpn < NCMPS) { |
|
1559 |
// Look for a dominating test that tells us the min and max. |
|
1560 |
int depth = 0; // Limit search depth for speed |
|
1561 |
Node* dom = control(); |
|
1562 |
for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { |
|
1563 |
if (++depth >= 100) break; |
|
1564 |
Node* ifproj = dom; |
|
1565 |
if (!ifproj->is_Proj()) continue; |
|
1566 |
Node* iff = ifproj->in(0); |
|
1567 |
if (!iff->is_If()) continue; |
|
1568 |
Node* bol = iff->in(1); |
|
1569 |
if (!bol->is_Bool()) continue; |
|
1570 |
Node* cmp = bol->in(1); |
|
1571 |
if (cmp == NULL) continue; |
|
1572 |
for (cmpn = 0; cmpn < NCMPS; cmpn++) |
|
1573 |
if (cmps[cmpn] == cmp) break; |
|
1574 |
if (cmpn == NCMPS) continue; |
|
1575 |
BoolTest::mask btest = bol->as_Bool()->_test._test; |
|
1576 |
if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); |
|
1577 |
if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); |
|
1578 |
// At this point, we know that 'x btest y' is true. |
|
1579 |
switch (btest) { |
|
1580 |
case BoolTest::eq: |
|
1581 |
// They are proven equal, so we can collapse the min/max. |
|
1582 |
// Either value is the answer. Choose the simpler. |
|
1583 |
if (is_simple_name(yvalue) && !is_simple_name(xvalue)) |
|
1584 |
return yvalue; |
|
1585 |
return xvalue; |
|
1586 |
case BoolTest::lt: // x < y |
|
1587 |
case BoolTest::le: // x <= y |
|
1588 |
return (want_max ? yvalue : xvalue); |
|
1589 |
case BoolTest::gt: // x > y |
|
1590 |
case BoolTest::ge: // x >= y |
|
1591 |
return (want_max ? xvalue : yvalue); |
|
1592 |
} |
|
1593 |
} |
|
1594 |
} |
|
1595 |
||
1596 |
// We failed to find a dominating test. |
|
1597 |
// Let's pick a test that might GVN with prior tests. |
|
1598 |
Node* best_bol = NULL; |
|
1599 |
BoolTest::mask best_btest = BoolTest::illegal; |
|
1600 |
for (cmpn = 0; cmpn < NCMPS; cmpn++) { |
|
1601 |
Node* cmp = cmps[cmpn]; |
|
1602 |
if (cmp == NULL) continue; |
|
1603 |
for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { |
|
1604 |
Node* bol = cmp->fast_out(j); |
|
1605 |
if (!bol->is_Bool()) continue; |
|
1606 |
BoolTest::mask btest = bol->as_Bool()->_test._test; |
|
1607 |
if (btest == BoolTest::eq || btest == BoolTest::ne) continue; |
|
1608 |
if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); |
|
1609 |
if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { |
|
1610 |
best_bol = bol->as_Bool(); |
|
1611 |
best_btest = btest; |
|
1612 |
} |
|
1613 |
} |
|
1614 |
} |
|
1615 |
||
1616 |
Node* answer_if_true = NULL; |
|
1617 |
Node* answer_if_false = NULL; |
|
1618 |
switch (best_btest) { |
|
1619 |
default: |
|
1620 |
if (cmpxy == NULL) |
|
1621 |
cmpxy = ideal_cmpxy; |
|
1622 |
best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); |
|
1623 |
// and fall through: |
|
1624 |
case BoolTest::lt: // x < y |
|
1625 |
case BoolTest::le: // x <= y |
|
1626 |
answer_if_true = (want_max ? yvalue : xvalue); |
|
1627 |
answer_if_false = (want_max ? xvalue : yvalue); |
|
1628 |
break; |
|
1629 |
case BoolTest::gt: // x > y |
|
1630 |
case BoolTest::ge: // x >= y |
|
1631 |
answer_if_true = (want_max ? xvalue : yvalue); |
|
1632 |
answer_if_false = (want_max ? yvalue : xvalue); |
|
1633 |
break; |
|
1634 |
} |
|
1635 |
||
1636 |
jint hi, lo; |
|
1637 |
if (want_max) { |
|
1638 |
// We can sharpen the minimum. |
|
1639 |
hi = MAX2(txvalue->_hi, tyvalue->_hi); |
|
1640 |
lo = MAX2(txvalue->_lo, tyvalue->_lo); |
|
1641 |
} else { |
|
1642 |
// We can sharpen the maximum. |
|
1643 |
hi = MIN2(txvalue->_hi, tyvalue->_hi); |
|
1644 |
lo = MIN2(txvalue->_lo, tyvalue->_lo); |
|
1645 |
} |
|
1646 |
||
1647 |
// Use a flow-free graph structure, to avoid creating excess control edges |
|
1648 |
// which could hinder other optimizations. |
|
1649 |
// Since Math.min/max is often used with arraycopy, we want |
|
1650 |
// tightly_coupled_allocation to be able to see beyond min/max expressions. |
|
1651 |
Node* cmov = CMoveNode::make(C, NULL, best_bol, |
|
1652 |
answer_if_false, answer_if_true, |
|
1653 |
TypeInt::make(lo, hi, widen)); |
|
1654 |
||
1655 |
return _gvn.transform(cmov); |
|
1656 |
||
1657 |
/* |
|
1658 |
// This is not as desirable as it may seem, since Min and Max |
|
1659 |
// nodes do not have a full set of optimizations. |
|
1660 |
// And they would interfere, anyway, with 'if' optimizations |
|
1661 |
// and with CMoveI canonical forms. |
|
1662 |
switch (id) { |
|
1663 |
case vmIntrinsics::_min: |
|
1664 |
result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; |
|
1665 |
case vmIntrinsics::_max: |
|
1666 |
result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; |
|
1667 |
default: |
|
1668 |
ShouldNotReachHere(); |
|
1669 |
} |
|
1670 |
*/ |
|
1671 |
} |
|
1672 |
||
1673 |
inline int |
|
1674 |
LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { |
|
1675 |
const TypePtr* base_type = TypePtr::NULL_PTR; |
|
1676 |
if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); |
|
1677 |
if (base_type == NULL) { |
|
1678 |
// Unknown type. |
|
1679 |
return Type::AnyPtr; |
|
1680 |
} else if (base_type == TypePtr::NULL_PTR) { |
|
1681 |
// Since this is a NULL+long form, we have to switch to a rawptr. |
|
1682 |
base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); |
|
1683 |
offset = MakeConX(0); |
|
1684 |
return Type::RawPtr; |
|
1685 |
} else if (base_type->base() == Type::RawPtr) { |
|
1686 |
return Type::RawPtr; |
|
1687 |
} else if (base_type->isa_oopptr()) { |
|
1688 |
// Base is never null => always a heap address. |
|
1689 |
if (base_type->ptr() == TypePtr::NotNull) { |
|
1690 |
return Type::OopPtr; |
|
1691 |
} |
|
1692 |
// Offset is small => always a heap address. |
|
1693 |
const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); |
|
1694 |
if (offset_type != NULL && |
|
1695 |
base_type->offset() == 0 && // (should always be?) |
|
1696 |
offset_type->_lo >= 0 && |
|
1697 |
!MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { |
|
1698 |
return Type::OopPtr; |
|
1699 |
} |
|
1700 |
// Otherwise, it might either be oop+off or NULL+addr. |
|
1701 |
return Type::AnyPtr; |
|
1702 |
} else { |
|
1703 |
// No information: |
|
1704 |
return Type::AnyPtr; |
|
1705 |
} |
|
1706 |
} |
|
1707 |
||
1708 |
inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { |
|
1709 |
int kind = classify_unsafe_addr(base, offset); |
|
1710 |
if (kind == Type::RawPtr) { |
|
1711 |
return basic_plus_adr(top(), base, offset); |
|
1712 |
} else { |
|
1713 |
return basic_plus_adr(base, offset); |
|
1714 |
} |
|
1715 |
} |
|
1716 |
||
1717 |
//----------------------------inline_reverseBytes_int/long------------------- |
|
1718 |
// inline Int.reverseBytes(int) |
|
1719 |
// inline Long.reverseByes(long) |
|
1720 |
bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { |
|
1721 |
assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes"); |
|
1722 |
if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; |
|
1723 |
if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; |
|
1724 |
_sp += arg_size(); // restore stack pointer |
|
1725 |
switch (id) { |
|
1726 |
case vmIntrinsics::_reverseBytes_i: |
|
1727 |
push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); |
|
1728 |
break; |
|
1729 |
case vmIntrinsics::_reverseBytes_l: |
|
1730 |
push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); |
|
1731 |
break; |
|
1732 |
default: |
|
1733 |
; |
|
1734 |
} |
|
1735 |
return true; |
|
1736 |
} |
|
1737 |
||
1738 |
//----------------------------inline_unsafe_access---------------------------- |
|
1739 |
||
1740 |
const static BasicType T_ADDRESS_HOLDER = T_LONG; |
|
1741 |
||
1742 |
// Interpret Unsafe.fieldOffset cookies correctly: |
|
1743 |
extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); |
|
1744 |
||
1745 |
bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { |
|
1746 |
if (callee()->is_static()) return false; // caller must have the capability! |
|
1747 |
||
1748 |
#ifndef PRODUCT |
|
1749 |
{ |
|
1750 |
ResourceMark rm; |
|
1751 |
// Check the signatures. |
|
1752 |
ciSignature* sig = signature(); |
|
1753 |
#ifdef ASSERT |
|
1754 |
if (!is_store) { |
|
1755 |
// Object getObject(Object base, int/long offset), etc. |
|
1756 |
BasicType rtype = sig->return_type()->basic_type(); |
|
1757 |
if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) |
|
1758 |
rtype = T_ADDRESS; // it is really a C void* |
|
1759 |
assert(rtype == type, "getter must return the expected value"); |
|
1760 |
if (!is_native_ptr) { |
|
1761 |
assert(sig->count() == 2, "oop getter has 2 arguments"); |
|
1762 |
assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); |
|
1763 |
assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); |
|
1764 |
} else { |
|
1765 |
assert(sig->count() == 1, "native getter has 1 argument"); |
|
1766 |
assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); |
|
1767 |
} |
|
1768 |
} else { |
|
1769 |
// void putObject(Object base, int/long offset, Object x), etc. |
|
1770 |
assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); |
|
1771 |
if (!is_native_ptr) { |
|
1772 |
assert(sig->count() == 3, "oop putter has 3 arguments"); |
|
1773 |
assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); |
|
1774 |
assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); |
|
1775 |
} else { |
|
1776 |
assert(sig->count() == 2, "native putter has 2 arguments"); |
|
1777 |
assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); |
|
1778 |
} |
|
1779 |
BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); |
|
1780 |
if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) |
|
1781 |
vtype = T_ADDRESS; // it is really a C void* |
|
1782 |
assert(vtype == type, "putter must accept the expected value"); |
|
1783 |
} |
|
1784 |
#endif // ASSERT |
|
1785 |
} |
|
1786 |
#endif //PRODUCT |
|
1787 |
||
1788 |
C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
1789 |
||
1790 |
int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; |
|
1791 |
||
1792 |
// Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words |
|
1793 |
int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); |
|
1794 |
||
1795 |
debug_only(int saved_sp = _sp); |
|
1796 |
_sp += nargs; |
|
1797 |
||
1798 |
Node* val; |
|
1799 |
debug_only(val = (Node*)(uintptr_t)-1); |
|
1800 |
||
1801 |
||
1802 |
if (is_store) { |
|
1803 |
// Get the value being stored. (Pop it first; it was pushed last.) |
|
1804 |
switch (type) { |
|
1805 |
case T_DOUBLE: |
|
1806 |
case T_LONG: |
|
1807 |
case T_ADDRESS: |
|
1808 |
val = pop_pair(); |
|
1809 |
break; |
|
1810 |
default: |
|
1811 |
val = pop(); |
|
1812 |
} |
|
1813 |
} |
|
1814 |
||
1815 |
// Build address expression. See the code in inline_unsafe_prefetch. |
|
1816 |
Node *adr; |
|
1817 |
Node *heap_base_oop = top(); |
|
1818 |
if (!is_native_ptr) { |
|
1819 |
// The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset |
|
1820 |
Node* offset = pop_pair(); |
|
1821 |
// The base is either a Java object or a value produced by Unsafe.staticFieldBase |
|
1822 |
Node* base = pop(); |
|
1823 |
// We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
|
1824 |
// to be plain byte offsets, which are also the same as those accepted |
|
1825 |
// by oopDesc::field_base. |
|
1826 |
assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
|
1827 |
"fieldOffset must be byte-scaled"); |
|
1828 |
// 32-bit machines ignore the high half! |
|
1829 |
offset = ConvL2X(offset); |
|
1830 |
adr = make_unsafe_address(base, offset); |
|
1831 |
heap_base_oop = base; |
|
1832 |
} else { |
|
1833 |
Node* ptr = pop_pair(); |
|
1834 |
// Adjust Java long to machine word: |
|
1835 |
ptr = ConvL2X(ptr); |
|
1836 |
adr = make_unsafe_address(NULL, ptr); |
|
1837 |
} |
|
1838 |
||
1839 |
// Pop receiver last: it was pushed first. |
|
1840 |
Node *receiver = pop(); |
|
1841 |
||
1842 |
assert(saved_sp == _sp, "must have correct argument count"); |
|
1843 |
||
1844 |
const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); |
|
1845 |
||
1846 |
// First guess at the value type. |
|
1847 |
const Type *value_type = Type::get_const_basic_type(type); |
|
1848 |
||
1849 |
// Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, |
|
1850 |
// there was not enough information to nail it down. |
|
1851 |
Compile::AliasType* alias_type = C->alias_type(adr_type); |
|
1852 |
assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); |
|
1853 |
||
1854 |
// We will need memory barriers unless we can determine a unique |
|
1855 |
// alias category for this reference. (Note: If for some reason |
|
1856 |
// the barriers get omitted and the unsafe reference begins to "pollute" |
|
1857 |
// the alias analysis of the rest of the graph, either Compile::can_alias |
|
1858 |
// or Compile::must_alias will throw a diagnostic assert.) |
|
1859 |
bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); |
|
1860 |
||
1861 |
if (!is_store && type == T_OBJECT) { |
|
1862 |
// Attempt to infer a sharper value type from the offset and base type. |
|
1863 |
ciKlass* sharpened_klass = NULL; |
|
1864 |
||
1865 |
// See if it is an instance field, with an object type. |
|
1866 |
if (alias_type->field() != NULL) { |
|
1867 |
assert(!is_native_ptr, "native pointer op cannot use a java address"); |
|
1868 |
if (alias_type->field()->type()->is_klass()) { |
|
1869 |
sharpened_klass = alias_type->field()->type()->as_klass(); |
|
1870 |
} |
|
1871 |
} |
|
1872 |
||
1873 |
// See if it is a narrow oop array. |
|
1874 |
if (adr_type->isa_aryptr()) { |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
1875 |
if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) { |
1 | 1876 |
const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); |
1877 |
if (elem_type != NULL) { |
|
1878 |
sharpened_klass = elem_type->klass(); |
|
1879 |
} |
|
1880 |
} |
|
1881 |
} |
|
1882 |
||
1883 |
if (sharpened_klass != NULL) { |
|
1884 |
const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); |
|
1885 |
||
1886 |
// Sharpen the value type. |
|
1887 |
value_type = tjp; |
|
1888 |
||
1889 |
#ifndef PRODUCT |
|
1890 |
if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { |
|
1891 |
tty->print(" from base type: "); adr_type->dump(); |
|
1892 |
tty->print(" sharpened value: "); value_type->dump(); |
|
1893 |
} |
|
1894 |
#endif |
|
1895 |
} |
|
1896 |
} |
|
1897 |
||
1898 |
// Null check on self without removing any arguments. The argument |
|
1899 |
// null check technically happens in the wrong place, which can lead to |
|
1900 |
// invalid stack traces when the primitive is inlined into a method |
|
1901 |
// which handles NullPointerExceptions. |
|
1902 |
_sp += nargs; |
|
1903 |
do_null_check(receiver, T_OBJECT); |
|
1904 |
_sp -= nargs; |
|
1905 |
if (stopped()) { |
|
1906 |
return true; |
|
1907 |
} |
|
1908 |
// Heap pointers get a null-check from the interpreter, |
|
1909 |
// as a courtesy. However, this is not guaranteed by Unsafe, |
|
1910 |
// and it is not possible to fully distinguish unintended nulls |
|
1911 |
// from intended ones in this API. |
|
1912 |
||
1913 |
if (is_volatile) { |
|
1914 |
// We need to emit leading and trailing CPU membars (see below) in |
|
1915 |
// addition to memory membars when is_volatile. This is a little |
|
1916 |
// too strong, but avoids the need to insert per-alias-type |
|
1917 |
// volatile membars (for stores; compare Parse::do_put_xxx), which |
|
1918 |
// we cannot do effctively here because we probably only have a |
|
1919 |
// rough approximation of type. |
|
1920 |
need_mem_bar = true; |
|
1921 |
// For Stores, place a memory ordering barrier now. |
|
1922 |
if (is_store) |
|
1923 |
insert_mem_bar(Op_MemBarRelease); |
|
1924 |
} |
|
1925 |
||
1926 |
// Memory barrier to prevent normal and 'unsafe' accesses from |
|
1927 |
// bypassing each other. Happens after null checks, so the |
|
1928 |
// exception paths do not take memory state from the memory barrier, |
|
1929 |
// so there's no problems making a strong assert about mixing users |
|
1930 |
// of safe & unsafe memory. Otherwise fails in a CTW of rt.jar |
|
1931 |
// around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. |
|
1932 |
if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); |
|
1933 |
||
1934 |
if (!is_store) { |
|
1935 |
Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); |
|
1936 |
// load value and push onto stack |
|
1937 |
switch (type) { |
|
1938 |
case T_BOOLEAN: |
|
1939 |
case T_CHAR: |
|
1940 |
case T_BYTE: |
|
1941 |
case T_SHORT: |
|
1942 |
case T_INT: |
|
1943 |
case T_FLOAT: |
|
1944 |
case T_OBJECT: |
|
1945 |
push( p ); |
|
1946 |
break; |
|
1947 |
case T_ADDRESS: |
|
1948 |
// Cast to an int type. |
|
1949 |
p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); |
|
1950 |
p = ConvX2L(p); |
|
1951 |
push_pair(p); |
|
1952 |
break; |
|
1953 |
case T_DOUBLE: |
|
1954 |
case T_LONG: |
|
1955 |
push_pair( p ); |
|
1956 |
break; |
|
1957 |
default: ShouldNotReachHere(); |
|
1958 |
} |
|
1959 |
} else { |
|
1960 |
// place effect of store into memory |
|
1961 |
switch (type) { |
|
1962 |
case T_DOUBLE: |
|
1963 |
val = dstore_rounding(val); |
|
1964 |
break; |
|
1965 |
case T_ADDRESS: |
|
1966 |
// Repackage the long as a pointer. |
|
1967 |
val = ConvL2X(val); |
|
1968 |
val = _gvn.transform( new (C, 2) CastX2PNode(val) ); |
|
1969 |
break; |
|
1970 |
} |
|
1971 |
||
1972 |
if (type != T_OBJECT ) { |
|
1973 |
(void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); |
|
1974 |
} else { |
|
1975 |
// Possibly an oop being stored to Java heap or native memory |
|
1976 |
if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { |
|
1977 |
// oop to Java heap. |
|
1978 |
(void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); |
|
1979 |
} else { |
|
1980 |
||
1981 |
// We can't tell at compile time if we are storing in the Java heap or outside |
|
1982 |
// of it. So we need to emit code to conditionally do the proper type of |
|
1983 |
// store. |
|
1984 |
||
1985 |
IdealKit kit(gvn(), control(), merged_memory()); |
|
1986 |
kit.declares_done(); |
|
1987 |
// QQQ who knows what probability is here?? |
|
1988 |
kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { |
|
1989 |
(void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); |
|
1990 |
} kit.else_(); { |
|
1991 |
(void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); |
|
1992 |
} kit.end_if(); |
|
1993 |
} |
|
1994 |
} |
|
1995 |
} |
|
1996 |
||
1997 |
if (is_volatile) { |
|
1998 |
if (!is_store) |
|
1999 |
insert_mem_bar(Op_MemBarAcquire); |
|
2000 |
else |
|
2001 |
insert_mem_bar(Op_MemBarVolatile); |
|
2002 |
} |
|
2003 |
||
2004 |
if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); |
|
2005 |
||
2006 |
return true; |
|
2007 |
} |
|
2008 |
||
2009 |
//----------------------------inline_unsafe_prefetch---------------------------- |
|
2010 |
||
2011 |
bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { |
|
2012 |
#ifndef PRODUCT |
|
2013 |
{ |
|
2014 |
ResourceMark rm; |
|
2015 |
// Check the signatures. |
|
2016 |
ciSignature* sig = signature(); |
|
2017 |
#ifdef ASSERT |
|
2018 |
// Object getObject(Object base, int/long offset), etc. |
|
2019 |
BasicType rtype = sig->return_type()->basic_type(); |
|
2020 |
if (!is_native_ptr) { |
|
2021 |
assert(sig->count() == 2, "oop prefetch has 2 arguments"); |
|
2022 |
assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); |
|
2023 |
assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); |
|
2024 |
} else { |
|
2025 |
assert(sig->count() == 1, "native prefetch has 1 argument"); |
|
2026 |
assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); |
|
2027 |
} |
|
2028 |
#endif // ASSERT |
|
2029 |
} |
|
2030 |
#endif // !PRODUCT |
|
2031 |
||
2032 |
C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
2033 |
||
2034 |
// Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args |
|
2035 |
int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); |
|
2036 |
||
2037 |
debug_only(int saved_sp = _sp); |
|
2038 |
_sp += nargs; |
|
2039 |
||
2040 |
// Build address expression. See the code in inline_unsafe_access. |
|
2041 |
Node *adr; |
|
2042 |
if (!is_native_ptr) { |
|
2043 |
// The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset |
|
2044 |
Node* offset = pop_pair(); |
|
2045 |
// The base is either a Java object or a value produced by Unsafe.staticFieldBase |
|
2046 |
Node* base = pop(); |
|
2047 |
// We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
|
2048 |
// to be plain byte offsets, which are also the same as those accepted |
|
2049 |
// by oopDesc::field_base. |
|
2050 |
assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
|
2051 |
"fieldOffset must be byte-scaled"); |
|
2052 |
// 32-bit machines ignore the high half! |
|
2053 |
offset = ConvL2X(offset); |
|
2054 |
adr = make_unsafe_address(base, offset); |
|
2055 |
} else { |
|
2056 |
Node* ptr = pop_pair(); |
|
2057 |
// Adjust Java long to machine word: |
|
2058 |
ptr = ConvL2X(ptr); |
|
2059 |
adr = make_unsafe_address(NULL, ptr); |
|
2060 |
} |
|
2061 |
||
2062 |
if (is_static) { |
|
2063 |
assert(saved_sp == _sp, "must have correct argument count"); |
|
2064 |
} else { |
|
2065 |
// Pop receiver last: it was pushed first. |
|
2066 |
Node *receiver = pop(); |
|
2067 |
assert(saved_sp == _sp, "must have correct argument count"); |
|
2068 |
||
2069 |
// Null check on self without removing any arguments. The argument |
|
2070 |
// null check technically happens in the wrong place, which can lead to |
|
2071 |
// invalid stack traces when the primitive is inlined into a method |
|
2072 |
// which handles NullPointerExceptions. |
|
2073 |
_sp += nargs; |
|
2074 |
do_null_check(receiver, T_OBJECT); |
|
2075 |
_sp -= nargs; |
|
2076 |
if (stopped()) { |
|
2077 |
return true; |
|
2078 |
} |
|
2079 |
} |
|
2080 |
||
2081 |
// Generate the read or write prefetch |
|
2082 |
Node *prefetch; |
|
2083 |
if (is_store) { |
|
2084 |
prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); |
|
2085 |
} else { |
|
2086 |
prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); |
|
2087 |
} |
|
2088 |
prefetch->init_req(0, control()); |
|
2089 |
set_i_o(_gvn.transform(prefetch)); |
|
2090 |
||
2091 |
return true; |
|
2092 |
} |
|
2093 |
||
2094 |
//----------------------------inline_unsafe_CAS---------------------------- |
|
2095 |
||
2096 |
bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { |
|
2097 |
// This basic scheme here is the same as inline_unsafe_access, but |
|
2098 |
// differs in enough details that combining them would make the code |
|
2099 |
// overly confusing. (This is a true fact! I originally combined |
|
2100 |
// them, but even I was confused by it!) As much code/comments as |
|
2101 |
// possible are retained from inline_unsafe_access though to make |
|
2102 |
// the correspondances clearer. - dl |
|
2103 |
||
2104 |
if (callee()->is_static()) return false; // caller must have the capability! |
|
2105 |
||
2106 |
#ifndef PRODUCT |
|
2107 |
{ |
|
2108 |
ResourceMark rm; |
|
2109 |
// Check the signatures. |
|
2110 |
ciSignature* sig = signature(); |
|
2111 |
#ifdef ASSERT |
|
2112 |
BasicType rtype = sig->return_type()->basic_type(); |
|
2113 |
assert(rtype == T_BOOLEAN, "CAS must return boolean"); |
|
2114 |
assert(sig->count() == 4, "CAS has 4 arguments"); |
|
2115 |
assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); |
|
2116 |
assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); |
|
2117 |
#endif // ASSERT |
|
2118 |
} |
|
2119 |
#endif //PRODUCT |
|
2120 |
||
2121 |
// number of stack slots per value argument (1 or 2) |
|
2122 |
int type_words = type2size[type]; |
|
2123 |
||
2124 |
// Cannot inline wide CAS on machines that don't support it natively |
|
202
dc13bf0e5d5d
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
1
diff
changeset
|
2125 |
if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) |
1 | 2126 |
return false; |
2127 |
||
2128 |
C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
2129 |
||
2130 |
// Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; |
|
2131 |
int nargs = 1 + 1 + 2 + type_words + type_words; |
|
2132 |
||
2133 |
// pop arguments: newval, oldval, offset, base, and receiver |
|
2134 |
debug_only(int saved_sp = _sp); |
|
2135 |
_sp += nargs; |
|
2136 |
Node* newval = (type_words == 1) ? pop() : pop_pair(); |
|
2137 |
Node* oldval = (type_words == 1) ? pop() : pop_pair(); |
|
2138 |
Node *offset = pop_pair(); |
|
2139 |
Node *base = pop(); |
|
2140 |
Node *receiver = pop(); |
|
2141 |
assert(saved_sp == _sp, "must have correct argument count"); |
|
2142 |
||
2143 |
// Null check receiver. |
|
2144 |
_sp += nargs; |
|
2145 |
do_null_check(receiver, T_OBJECT); |
|
2146 |
_sp -= nargs; |
|
2147 |
if (stopped()) { |
|
2148 |
return true; |
|
2149 |
} |
|
2150 |
||
2151 |
// Build field offset expression. |
|
2152 |
// We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
|
2153 |
// to be plain byte offsets, which are also the same as those accepted |
|
2154 |
// by oopDesc::field_base. |
|
2155 |
assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); |
|
2156 |
// 32-bit machines ignore the high half of long offsets |
|
2157 |
offset = ConvL2X(offset); |
|
2158 |
Node* adr = make_unsafe_address(base, offset); |
|
2159 |
const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); |
|
2160 |
||
2161 |
// (Unlike inline_unsafe_access, there seems no point in trying |
|
2162 |
// to refine types. Just use the coarse types here. |
|
2163 |
const Type *value_type = Type::get_const_basic_type(type); |
|
2164 |
Compile::AliasType* alias_type = C->alias_type(adr_type); |
|
2165 |
assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); |
|
2166 |
int alias_idx = C->get_alias_index(adr_type); |
|
2167 |
||
2168 |
// Memory-model-wise, a CAS acts like a little synchronized block, |
|
2169 |
// so needs barriers on each side. These don't't translate into |
|
2170 |
// actual barriers on most machines, but we still need rest of |
|
2171 |
// compiler to respect ordering. |
|
2172 |
||
2173 |
insert_mem_bar(Op_MemBarRelease); |
|
2174 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
2175 |
||
2176 |
// 4984716: MemBars must be inserted before this |
|
2177 |
// memory node in order to avoid a false |
|
2178 |
// dependency which will confuse the scheduler. |
|
2179 |
Node *mem = memory(alias_idx); |
|
2180 |
||
2181 |
// For now, we handle only those cases that actually exist: ints, |
|
2182 |
// longs, and Object. Adding others should be straightforward. |
|
2183 |
Node* cas; |
|
2184 |
switch(type) { |
|
2185 |
case T_INT: |
|
2186 |
cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); |
|
2187 |
break; |
|
2188 |
case T_LONG: |
|
2189 |
cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); |
|
2190 |
break; |
|
2191 |
case T_OBJECT: |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
2192 |
// reference stores need a store barrier. |
1 | 2193 |
// (They don't if CAS fails, but it isn't worth checking.) |
2194 |
pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT); |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
2195 |
#ifdef _LP64 |
589 | 2196 |
if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
767
64fb1fd7186d
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
600
diff
changeset
|
2197 |
Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); |
64fb1fd7186d
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
600
diff
changeset
|
2198 |
Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); |
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
2199 |
cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, |
767
64fb1fd7186d
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
600
diff
changeset
|
2200 |
newval_enc, oldval_enc)); |
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
2201 |
} else |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
2202 |
#endif |
767
64fb1fd7186d
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
600
diff
changeset
|
2203 |
{ |
64fb1fd7186d
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
600
diff
changeset
|
2204 |
cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); |
64fb1fd7186d
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
600
diff
changeset
|
2205 |
} |
1 | 2206 |
post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); |
2207 |
break; |
|
2208 |
default: |
|
2209 |
ShouldNotReachHere(); |
|
2210 |
break; |
|
2211 |
} |
|
2212 |
||
2213 |
// SCMemProjNodes represent the memory state of CAS. Their main |
|
2214 |
// role is to prevent CAS nodes from being optimized away when their |
|
2215 |
// results aren't used. |
|
2216 |
Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); |
|
2217 |
set_memory(proj, alias_idx); |
|
2218 |
||
2219 |
// Add the trailing membar surrounding the access |
|
2220 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
2221 |
insert_mem_bar(Op_MemBarAcquire); |
|
2222 |
||
2223 |
push(cas); |
|
2224 |
return true; |
|
2225 |
} |
|
2226 |
||
2227 |
bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { |
|
2228 |
// This is another variant of inline_unsafe_access, differing in |
|
2229 |
// that it always issues store-store ("release") barrier and ensures |
|
2230 |
// store-atomicity (which only matters for "long"). |
|
2231 |
||
2232 |
if (callee()->is_static()) return false; // caller must have the capability! |
|
2233 |
||
2234 |
#ifndef PRODUCT |
|
2235 |
{ |
|
2236 |
ResourceMark rm; |
|
2237 |
// Check the signatures. |
|
2238 |
ciSignature* sig = signature(); |
|
2239 |
#ifdef ASSERT |
|
2240 |
BasicType rtype = sig->return_type()->basic_type(); |
|
2241 |
assert(rtype == T_VOID, "must return void"); |
|
2242 |
assert(sig->count() == 3, "has 3 arguments"); |
|
2243 |
assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); |
|
2244 |
assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); |
|
2245 |
#endif // ASSERT |
|
2246 |
} |
|
2247 |
#endif //PRODUCT |
|
2248 |
||
2249 |
// number of stack slots per value argument (1 or 2) |
|
2250 |
int type_words = type2size[type]; |
|
2251 |
||
2252 |
C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
2253 |
||
2254 |
// Argument words: "this" plus oop plus offset plus value; |
|
2255 |
int nargs = 1 + 1 + 2 + type_words; |
|
2256 |
||
2257 |
// pop arguments: val, offset, base, and receiver |
|
2258 |
debug_only(int saved_sp = _sp); |
|
2259 |
_sp += nargs; |
|
2260 |
Node* val = (type_words == 1) ? pop() : pop_pair(); |
|
2261 |
Node *offset = pop_pair(); |
|
2262 |
Node *base = pop(); |
|
2263 |
Node *receiver = pop(); |
|
2264 |
assert(saved_sp == _sp, "must have correct argument count"); |
|
2265 |
||
2266 |
// Null check receiver. |
|
2267 |
_sp += nargs; |
|
2268 |
do_null_check(receiver, T_OBJECT); |
|
2269 |
_sp -= nargs; |
|
2270 |
if (stopped()) { |
|
2271 |
return true; |
|
2272 |
} |
|
2273 |
||
2274 |
// Build field offset expression. |
|
2275 |
assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); |
|
2276 |
// 32-bit machines ignore the high half of long offsets |
|
2277 |
offset = ConvL2X(offset); |
|
2278 |
Node* adr = make_unsafe_address(base, offset); |
|
2279 |
const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); |
|
2280 |
const Type *value_type = Type::get_const_basic_type(type); |
|
2281 |
Compile::AliasType* alias_type = C->alias_type(adr_type); |
|
2282 |
||
2283 |
insert_mem_bar(Op_MemBarRelease); |
|
2284 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
2285 |
// Ensure that the store is atomic for longs: |
|
2286 |
bool require_atomic_access = true; |
|
2287 |
Node* store; |
|
2288 |
if (type == T_OBJECT) // reference stores need a store barrier. |
|
2289 |
store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type); |
|
2290 |
else { |
|
2291 |
store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); |
|
2292 |
} |
|
2293 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
2294 |
return true; |
|
2295 |
} |
|
2296 |
||
2297 |
bool LibraryCallKit::inline_unsafe_allocate() { |
|
2298 |
if (callee()->is_static()) return false; // caller must have the capability! |
|
2299 |
int nargs = 1 + 1; |
|
2300 |
assert(signature()->size() == nargs-1, "alloc has 1 argument"); |
|
2301 |
null_check_receiver(callee()); // check then ignore argument(0) |
|
2302 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2303 |
Node* cls = do_null_check(argument(1), T_OBJECT); |
|
2304 |
_sp -= nargs; |
|
2305 |
if (stopped()) return true; |
|
2306 |
||
2307 |
Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); |
|
2308 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2309 |
kls = do_null_check(kls, T_OBJECT); |
|
2310 |
_sp -= nargs; |
|
2311 |
if (stopped()) return true; // argument was like int.class |
|
2312 |
||
2313 |
// Note: The argument might still be an illegal value like |
|
2314 |
// Serializable.class or Object[].class. The runtime will handle it. |
|
2315 |
// But we must make an explicit check for initialization. |
|
2316 |
Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc)); |
|
2317 |
Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT); |
|
2318 |
Node* bits = intcon(instanceKlass::fully_initialized); |
|
2319 |
Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); |
|
2320 |
// The 'test' is non-zero if we need to take a slow path. |
|
2321 |
||
2322 |
Node* obj = new_instance(kls, test); |
|
2323 |
push(obj); |
|
2324 |
||
2325 |
return true; |
|
2326 |
} |
|
2327 |
||
2328 |
//------------------------inline_native_time_funcs-------------- |
|
2329 |
// inline code for System.currentTimeMillis() and System.nanoTime() |
|
2330 |
// these have the same type and signature |
|
2331 |
bool LibraryCallKit::inline_native_time_funcs(bool isNano) { |
|
2332 |
address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : |
|
2333 |
CAST_FROM_FN_PTR(address, os::javaTimeMillis); |
|
2334 |
const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; |
|
2335 |
const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); |
|
2336 |
const TypePtr* no_memory_effects = NULL; |
|
2337 |
Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); |
|
2338 |
Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); |
|
2339 |
#ifdef ASSERT |
|
2340 |
Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); |
|
2341 |
assert(value_top == top(), "second value must be top"); |
|
2342 |
#endif |
|
2343 |
push_pair(value); |
|
2344 |
return true; |
|
2345 |
} |
|
2346 |
||
2347 |
//------------------------inline_native_currentThread------------------ |
|
2348 |
bool LibraryCallKit::inline_native_currentThread() { |
|
2349 |
Node* junk = NULL; |
|
2350 |
push(generate_current_thread(junk)); |
|
2351 |
return true; |
|
2352 |
} |
|
2353 |
||
2354 |
//------------------------inline_native_isInterrupted------------------ |
|
2355 |
bool LibraryCallKit::inline_native_isInterrupted() { |
|
2356 |
const int nargs = 1+1; // receiver + boolean |
|
2357 |
assert(nargs == arg_size(), "sanity"); |
|
2358 |
// Add a fast path to t.isInterrupted(clear_int): |
|
2359 |
// (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) |
|
2360 |
// ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) |
|
2361 |
// So, in the common case that the interrupt bit is false, |
|
2362 |
// we avoid making a call into the VM. Even if the interrupt bit |
|
2363 |
// is true, if the clear_int argument is false, we avoid the VM call. |
|
2364 |
// However, if the receiver is not currentThread, we must call the VM, |
|
2365 |
// because there must be some locking done around the operation. |
|
2366 |
||
2367 |
// We only go to the fast case code if we pass two guards. |
|
2368 |
// Paths which do not pass are accumulated in the slow_region. |
|
2369 |
RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
2370 |
record_for_igvn(slow_region); |
|
2371 |
RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow |
|
2372 |
PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); |
|
2373 |
enum { no_int_result_path = 1, |
|
2374 |
no_clear_result_path = 2, |
|
2375 |
slow_result_path = 3 |
|
2376 |
}; |
|
2377 |
||
2378 |
// (a) Receiving thread must be the current thread. |
|
2379 |
Node* rec_thr = argument(0); |
|
2380 |
Node* tls_ptr = NULL; |
|
2381 |
Node* cur_thr = generate_current_thread(tls_ptr); |
|
2382 |
Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); |
|
2383 |
Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); |
|
2384 |
||
2385 |
bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); |
|
2386 |
if (!known_current_thread) |
|
2387 |
generate_slow_guard(bol_thr, slow_region); |
|
2388 |
||
2389 |
// (b) Interrupt bit on TLS must be false. |
|
2390 |
Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); |
|
2391 |
Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); |
|
2392 |
p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); |
|
2393 |
Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT); |
|
2394 |
Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); |
|
2395 |
Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); |
|
2396 |
||
2397 |
IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); |
|
2398 |
||
2399 |
// First fast path: if (!TLS._interrupted) return false; |
|
2400 |
Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); |
|
2401 |
result_rgn->init_req(no_int_result_path, false_bit); |
|
2402 |
result_val->init_req(no_int_result_path, intcon(0)); |
|
2403 |
||
2404 |
// drop through to next case |
|
2405 |
set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); |
|
2406 |
||
2407 |
// (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. |
|
2408 |
Node* clr_arg = argument(1); |
|
2409 |
Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); |
|
2410 |
Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); |
|
2411 |
IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); |
|
2412 |
||
2413 |
// Second fast path: ... else if (!clear_int) return true; |
|
2414 |
Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); |
|
2415 |
result_rgn->init_req(no_clear_result_path, false_arg); |
|
2416 |
result_val->init_req(no_clear_result_path, intcon(1)); |
|
2417 |
||
2418 |
// drop through to next case |
|
2419 |
set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); |
|
2420 |
||
2421 |
// (d) Otherwise, go to the slow path. |
|
2422 |
slow_region->add_req(control()); |
|
2423 |
set_control( _gvn.transform(slow_region) ); |
|
2424 |
||
2425 |
if (stopped()) { |
|
2426 |
// There is no slow path. |
|
2427 |
result_rgn->init_req(slow_result_path, top()); |
|
2428 |
result_val->init_req(slow_result_path, top()); |
|
2429 |
} else { |
|
2430 |
// non-virtual because it is a private non-static |
|
2431 |
CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); |
|
2432 |
||
2433 |
Node* slow_val = set_results_for_java_call(slow_call); |
|
2434 |
// this->control() comes from set_results_for_java_call |
|
2435 |
||
2436 |
// If we know that the result of the slow call will be true, tell the optimizer! |
|
2437 |
if (known_current_thread) slow_val = intcon(1); |
|
2438 |
||
2439 |
Node* fast_io = slow_call->in(TypeFunc::I_O); |
|
2440 |
Node* fast_mem = slow_call->in(TypeFunc::Memory); |
|
2441 |
// These two phis are pre-filled with copies of of the fast IO and Memory |
|
2442 |
Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); |
|
2443 |
Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); |
|
2444 |
||
2445 |
result_rgn->init_req(slow_result_path, control()); |
|
2446 |
io_phi ->init_req(slow_result_path, i_o()); |
|
2447 |
mem_phi ->init_req(slow_result_path, reset_memory()); |
|
2448 |
result_val->init_req(slow_result_path, slow_val); |
|
2449 |
||
2450 |
set_all_memory( _gvn.transform(mem_phi) ); |
|
2451 |
set_i_o( _gvn.transform(io_phi) ); |
|
2452 |
} |
|
2453 |
||
2454 |
push_result(result_rgn, result_val); |
|
2455 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2456 |
||
2457 |
return true; |
|
2458 |
} |
|
2459 |
||
2460 |
//---------------------------load_mirror_from_klass---------------------------- |
|
2461 |
// Given a klass oop, load its java mirror (a java.lang.Class oop). |
|
2462 |
Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { |
|
2463 |
Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc)); |
|
2464 |
return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); |
|
2465 |
} |
|
2466 |
||
2467 |
//-----------------------load_klass_from_mirror_common------------------------- |
|
2468 |
// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. |
|
2469 |
// Test the klass oop for null (signifying a primitive Class like Integer.TYPE), |
|
2470 |
// and branch to the given path on the region. |
|
2471 |
// If never_see_null, take an uncommon trap on null, so we can optimistically |
|
2472 |
// compile for the non-null case. |
|
2473 |
// If the region is NULL, force never_see_null = true. |
|
2474 |
Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, |
|
2475 |
bool never_see_null, |
|
2476 |
int nargs, |
|
2477 |
RegionNode* region, |
|
2478 |
int null_path, |
|
2479 |
int offset) { |
|
2480 |
if (region == NULL) never_see_null = true; |
|
2481 |
Node* p = basic_plus_adr(mirror, offset); |
|
2482 |
const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; |
|
590
2954744d7bba
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
589
diff
changeset
|
2483 |
Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); |
1 | 2484 |
_sp += nargs; // any deopt will start just before call to enclosing method |
2485 |
Node* null_ctl = top(); |
|
2486 |
kls = null_check_oop(kls, &null_ctl, never_see_null); |
|
2487 |
if (region != NULL) { |
|
2488 |
// Set region->in(null_path) if the mirror is a primitive (e.g, int.class). |
|
2489 |
region->init_req(null_path, null_ctl); |
|
2490 |
} else { |
|
2491 |
assert(null_ctl == top(), "no loose ends"); |
|
2492 |
} |
|
2493 |
_sp -= nargs; |
|
2494 |
return kls; |
|
2495 |
} |
|
2496 |
||
2497 |
//--------------------(inline_native_Class_query helpers)--------------------- |
|
2498 |
// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. |
|
2499 |
// Fall through if (mods & mask) == bits, take the guard otherwise. |
|
2500 |
Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { |
|
2501 |
// Branch around if the given klass has the given modifier bit set. |
|
2502 |
// Like generate_guard, adds a new path onto the region. |
|
2503 |
Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); |
|
2504 |
Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); |
|
2505 |
Node* mask = intcon(modifier_mask); |
|
2506 |
Node* bits = intcon(modifier_bits); |
|
2507 |
Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); |
|
2508 |
Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); |
|
2509 |
Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); |
|
2510 |
return generate_fair_guard(bol, region); |
|
2511 |
} |
|
2512 |
Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { |
|
2513 |
return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); |
|
2514 |
} |
|
2515 |
||
2516 |
//-------------------------inline_native_Class_query------------------- |
|
2517 |
bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { |
|
2518 |
int nargs = 1+0; // just the Class mirror, in most cases |
|
2519 |
const Type* return_type = TypeInt::BOOL; |
|
2520 |
Node* prim_return_value = top(); // what happens if it's a primitive class? |
|
2521 |
bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
|
2522 |
bool expect_prim = false; // most of these guys expect to work on refs |
|
2523 |
||
2524 |
enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; |
|
2525 |
||
2526 |
switch (id) { |
|
2527 |
case vmIntrinsics::_isInstance: |
|
2528 |
nargs = 1+1; // the Class mirror, plus the object getting queried about |
|
2529 |
// nothing is an instance of a primitive type |
|
2530 |
prim_return_value = intcon(0); |
|
2531 |
break; |
|
2532 |
case vmIntrinsics::_getModifiers: |
|
2533 |
prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); |
|
2534 |
assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); |
|
2535 |
return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); |
|
2536 |
break; |
|
2537 |
case vmIntrinsics::_isInterface: |
|
2538 |
prim_return_value = intcon(0); |
|
2539 |
break; |
|
2540 |
case vmIntrinsics::_isArray: |
|
2541 |
prim_return_value = intcon(0); |
|
2542 |
expect_prim = true; // cf. ObjectStreamClass.getClassSignature |
|
2543 |
break; |
|
2544 |
case vmIntrinsics::_isPrimitive: |
|
2545 |
prim_return_value = intcon(1); |
|
2546 |
expect_prim = true; // obviously |
|
2547 |
break; |
|
2548 |
case vmIntrinsics::_getSuperclass: |
|
2549 |
prim_return_value = null(); |
|
2550 |
return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); |
|
2551 |
break; |
|
2552 |
case vmIntrinsics::_getComponentType: |
|
2553 |
prim_return_value = null(); |
|
2554 |
return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); |
|
2555 |
break; |
|
2556 |
case vmIntrinsics::_getClassAccessFlags: |
|
2557 |
prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); |
|
2558 |
return_type = TypeInt::INT; // not bool! 6297094 |
|
2559 |
break; |
|
2560 |
default: |
|
2561 |
ShouldNotReachHere(); |
|
2562 |
} |
|
2563 |
||
2564 |
Node* mirror = argument(0); |
|
2565 |
Node* obj = (nargs <= 1)? top(): argument(1); |
|
2566 |
||
2567 |
const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); |
|
2568 |
if (mirror_con == NULL) return false; // cannot happen? |
|
2569 |
||
2570 |
#ifndef PRODUCT |
|
2571 |
if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { |
|
2572 |
ciType* k = mirror_con->java_mirror_type(); |
|
2573 |
if (k) { |
|
2574 |
tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); |
|
2575 |
k->print_name(); |
|
2576 |
tty->cr(); |
|
2577 |
} |
|
2578 |
} |
|
2579 |
#endif |
|
2580 |
||
2581 |
// Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). |
|
2582 |
RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
2583 |
record_for_igvn(region); |
|
2584 |
PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); |
|
2585 |
||
2586 |
// The mirror will never be null of Reflection.getClassAccessFlags, however |
|
2587 |
// it may be null for Class.isInstance or Class.getModifiers. Throw a NPE |
|
2588 |
// if it is. See bug 4774291. |
|
2589 |
||
2590 |
// For Reflection.getClassAccessFlags(), the null check occurs in |
|
2591 |
// the wrong place; see inline_unsafe_access(), above, for a similar |
|
2592 |
// situation. |
|
2593 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2594 |
mirror = do_null_check(mirror, T_OBJECT); |
|
2595 |
_sp -= nargs; |
|
2596 |
// If mirror or obj is dead, only null-path is taken. |
|
2597 |
if (stopped()) return true; |
|
2598 |
||
2599 |
if (expect_prim) never_see_null = false; // expect nulls (meaning prims) |
|
2600 |
||
2601 |
// Now load the mirror's klass metaobject, and null-check it. |
|
2602 |
// Side-effects region with the control path if the klass is null. |
|
2603 |
Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, |
|
2604 |
region, _prim_path); |
|
2605 |
// If kls is null, we have a primitive mirror. |
|
2606 |
phi->init_req(_prim_path, prim_return_value); |
|
2607 |
if (stopped()) { push_result(region, phi); return true; } |
|
2608 |
||
2609 |
Node* p; // handy temp |
|
2610 |
Node* null_ctl; |
|
2611 |
||
2612 |
// Now that we have the non-null klass, we can perform the real query. |
|
2613 |
// For constant classes, the query will constant-fold in LoadNode::Value. |
|
2614 |
Node* query_value = top(); |
|
2615 |
switch (id) { |
|
2616 |
case vmIntrinsics::_isInstance: |
|
2617 |
// nothing is an instance of a primitive type |
|
2618 |
query_value = gen_instanceof(obj, kls); |
|
2619 |
break; |
|
2620 |
||
2621 |
case vmIntrinsics::_getModifiers: |
|
2622 |
p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc)); |
|
2623 |
query_value = make_load(NULL, p, TypeInt::INT, T_INT); |
|
2624 |
break; |
|
2625 |
||
2626 |
case vmIntrinsics::_isInterface: |
|
2627 |
// (To verify this code sequence, check the asserts in JVM_IsInterface.) |
|
2628 |
if (generate_interface_guard(kls, region) != NULL) |
|
2629 |
// A guard was added. If the guard is taken, it was an interface. |
|
2630 |
phi->add_req(intcon(1)); |
|
2631 |
// If we fall through, it's a plain class. |
|
2632 |
query_value = intcon(0); |
|
2633 |
break; |
|
2634 |
||
2635 |
case vmIntrinsics::_isArray: |
|
2636 |
// (To verify this code sequence, check the asserts in JVM_IsArrayClass.) |
|
2637 |
if (generate_array_guard(kls, region) != NULL) |
|
2638 |
// A guard was added. If the guard is taken, it was an array. |
|
2639 |
phi->add_req(intcon(1)); |
|
2640 |
// If we fall through, it's a plain class. |
|
2641 |
query_value = intcon(0); |
|
2642 |
break; |
|
2643 |
||
2644 |
case vmIntrinsics::_isPrimitive: |
|
2645 |
query_value = intcon(0); // "normal" path produces false |
|
2646 |
break; |
|
2647 |
||
2648 |
case vmIntrinsics::_getSuperclass: |
|
2649 |
// The rules here are somewhat unfortunate, but we can still do better |
|
2650 |
// with random logic than with a JNI call. |
|
2651 |
// Interfaces store null or Object as _super, but must report null. |
|
2652 |
// Arrays store an intermediate super as _super, but must report Object. |
|
2653 |
// Other types can report the actual _super. |
|
2654 |
// (To verify this code sequence, check the asserts in JVM_IsInterface.) |
|
2655 |
if (generate_interface_guard(kls, region) != NULL) |
|
2656 |
// A guard was added. If the guard is taken, it was an interface. |
|
2657 |
phi->add_req(null()); |
|
2658 |
if (generate_array_guard(kls, region) != NULL) |
|
2659 |
// A guard was added. If the guard is taken, it was an array. |
|
2660 |
phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); |
|
2661 |
// If we fall through, it's a plain class. Get its _super. |
|
2662 |
p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc)); |
|
590
2954744d7bba
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
589
diff
changeset
|
2663 |
kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); |
1 | 2664 |
null_ctl = top(); |
2665 |
kls = null_check_oop(kls, &null_ctl); |
|
2666 |
if (null_ctl != top()) { |
|
2667 |
// If the guard is taken, Object.superClass is null (both klass and mirror). |
|
2668 |
region->add_req(null_ctl); |
|
2669 |
phi ->add_req(null()); |
|
2670 |
} |
|
2671 |
if (!stopped()) { |
|
2672 |
query_value = load_mirror_from_klass(kls); |
|
2673 |
} |
|
2674 |
break; |
|
2675 |
||
2676 |
case vmIntrinsics::_getComponentType: |
|
2677 |
if (generate_array_guard(kls, region) != NULL) { |
|
2678 |
// Be sure to pin the oop load to the guard edge just created: |
|
2679 |
Node* is_array_ctrl = region->in(region->req()-1); |
|
2680 |
Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc)); |
|
2681 |
Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); |
|
2682 |
phi->add_req(cmo); |
|
2683 |
} |
|
2684 |
query_value = null(); // non-array case is null |
|
2685 |
break; |
|
2686 |
||
2687 |
case vmIntrinsics::_getClassAccessFlags: |
|
2688 |
p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); |
|
2689 |
query_value = make_load(NULL, p, TypeInt::INT, T_INT); |
|
2690 |
break; |
|
2691 |
||
2692 |
default: |
|
2693 |
ShouldNotReachHere(); |
|
2694 |
} |
|
2695 |
||
2696 |
// Fall-through is the normal case of a query to a real class. |
|
2697 |
phi->init_req(1, query_value); |
|
2698 |
region->init_req(1, control()); |
|
2699 |
||
2700 |
push_result(region, phi); |
|
2701 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2702 |
||
2703 |
return true; |
|
2704 |
} |
|
2705 |
||
2706 |
//--------------------------inline_native_subtype_check------------------------ |
|
2707 |
// This intrinsic takes the JNI calls out of the heart of |
|
2708 |
// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. |
|
2709 |
bool LibraryCallKit::inline_native_subtype_check() { |
|
2710 |
int nargs = 1+1; // the Class mirror, plus the other class getting examined |
|
2711 |
||
2712 |
// Pull both arguments off the stack. |
|
2713 |
Node* args[2]; // two java.lang.Class mirrors: superc, subc |
|
2714 |
args[0] = argument(0); |
|
2715 |
args[1] = argument(1); |
|
2716 |
Node* klasses[2]; // corresponding Klasses: superk, subk |
|
2717 |
klasses[0] = klasses[1] = top(); |
|
2718 |
||
2719 |
enum { |
|
2720 |
// A full decision tree on {superc is prim, subc is prim}: |
|
2721 |
_prim_0_path = 1, // {P,N} => false |
|
2722 |
// {P,P} & superc!=subc => false |
|
2723 |
_prim_same_path, // {P,P} & superc==subc => true |
|
2724 |
_prim_1_path, // {N,P} => false |
|
2725 |
_ref_subtype_path, // {N,N} & subtype check wins => true |
|
2726 |
_both_ref_path, // {N,N} & subtype check loses => false |
|
2727 |
PATH_LIMIT |
|
2728 |
}; |
|
2729 |
||
2730 |
RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
2731 |
Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); |
|
2732 |
record_for_igvn(region); |
|
2733 |
||
2734 |
const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads |
|
2735 |
const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; |
|
2736 |
int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); |
|
2737 |
||
2738 |
// First null-check both mirrors and load each mirror's klass metaobject. |
|
2739 |
int which_arg; |
|
2740 |
for (which_arg = 0; which_arg <= 1; which_arg++) { |
|
2741 |
Node* arg = args[which_arg]; |
|
2742 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2743 |
arg = do_null_check(arg, T_OBJECT); |
|
2744 |
_sp -= nargs; |
|
2745 |
if (stopped()) break; |
|
2746 |
args[which_arg] = _gvn.transform(arg); |
|
2747 |
||
2748 |
Node* p = basic_plus_adr(arg, class_klass_offset); |
|
590
2954744d7bba
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
589
diff
changeset
|
2749 |
Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); |
1 | 2750 |
klasses[which_arg] = _gvn.transform(kls); |
2751 |
} |
|
2752 |
||
2753 |
// Having loaded both klasses, test each for null. |
|
2754 |
bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
|
2755 |
for (which_arg = 0; which_arg <= 1; which_arg++) { |
|
2756 |
Node* kls = klasses[which_arg]; |
|
2757 |
Node* null_ctl = top(); |
|
2758 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2759 |
kls = null_check_oop(kls, &null_ctl, never_see_null); |
|
2760 |
_sp -= nargs; |
|
2761 |
int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); |
|
2762 |
region->init_req(prim_path, null_ctl); |
|
2763 |
if (stopped()) break; |
|
2764 |
klasses[which_arg] = kls; |
|
2765 |
} |
|
2766 |
||
2767 |
if (!stopped()) { |
|
2768 |
// now we have two reference types, in klasses[0..1] |
|
2769 |
Node* subk = klasses[1]; // the argument to isAssignableFrom |
|
2770 |
Node* superk = klasses[0]; // the receiver |
|
2771 |
region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); |
|
2772 |
// now we have a successful reference subtype check |
|
2773 |
region->set_req(_ref_subtype_path, control()); |
|
2774 |
} |
|
2775 |
||
2776 |
// If both operands are primitive (both klasses null), then |
|
2777 |
// we must return true when they are identical primitives. |
|
2778 |
// It is convenient to test this after the first null klass check. |
|
2779 |
set_control(region->in(_prim_0_path)); // go back to first null check |
|
2780 |
if (!stopped()) { |
|
2781 |
// Since superc is primitive, make a guard for the superc==subc case. |
|
2782 |
Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); |
|
2783 |
Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); |
|
2784 |
generate_guard(bol_eq, region, PROB_FAIR); |
|
2785 |
if (region->req() == PATH_LIMIT+1) { |
|
2786 |
// A guard was added. If the added guard is taken, superc==subc. |
|
2787 |
region->swap_edges(PATH_LIMIT, _prim_same_path); |
|
2788 |
region->del_req(PATH_LIMIT); |
|
2789 |
} |
|
2790 |
region->set_req(_prim_0_path, control()); // Not equal after all. |
|
2791 |
} |
|
2792 |
||
2793 |
// these are the only paths that produce 'true': |
|
2794 |
phi->set_req(_prim_same_path, intcon(1)); |
|
2795 |
phi->set_req(_ref_subtype_path, intcon(1)); |
|
2796 |
||
2797 |
// pull together the cases: |
|
2798 |
assert(region->req() == PATH_LIMIT, "sane region"); |
|
2799 |
for (uint i = 1; i < region->req(); i++) { |
|
2800 |
Node* ctl = region->in(i); |
|
2801 |
if (ctl == NULL || ctl == top()) { |
|
2802 |
region->set_req(i, top()); |
|
2803 |
phi ->set_req(i, top()); |
|
2804 |
} else if (phi->in(i) == NULL) { |
|
2805 |
phi->set_req(i, intcon(0)); // all other paths produce 'false' |
|
2806 |
} |
|
2807 |
} |
|
2808 |
||
2809 |
set_control(_gvn.transform(region)); |
|
2810 |
push(_gvn.transform(phi)); |
|
2811 |
||
2812 |
return true; |
|
2813 |
} |
|
2814 |
||
2815 |
//---------------------generate_array_guard_common------------------------ |
|
2816 |
Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, |
|
2817 |
bool obj_array, bool not_array) { |
|
2818 |
// If obj_array/non_array==false/false: |
|
2819 |
// Branch around if the given klass is in fact an array (either obj or prim). |
|
2820 |
// If obj_array/non_array==false/true: |
|
2821 |
// Branch around if the given klass is not an array klass of any kind. |
|
2822 |
// If obj_array/non_array==true/true: |
|
2823 |
// Branch around if the kls is not an oop array (kls is int[], String, etc.) |
|
2824 |
// If obj_array/non_array==true/false: |
|
2825 |
// Branch around if the kls is an oop array (Object[] or subtype) |
|
2826 |
// |
|
2827 |
// Like generate_guard, adds a new path onto the region. |
|
2828 |
jint layout_con = 0; |
|
2829 |
Node* layout_val = get_layout_helper(kls, layout_con); |
|
2830 |
if (layout_val == NULL) { |
|
2831 |
bool query = (obj_array |
|
2832 |
? Klass::layout_helper_is_objArray(layout_con) |
|
2833 |
: Klass::layout_helper_is_javaArray(layout_con)); |
|
2834 |
if (query == not_array) { |
|
2835 |
return NULL; // never a branch |
|
2836 |
} else { // always a branch |
|
2837 |
Node* always_branch = control(); |
|
2838 |
if (region != NULL) |
|
2839 |
region->add_req(always_branch); |
|
2840 |
set_control(top()); |
|
2841 |
return always_branch; |
|
2842 |
} |
|
2843 |
} |
|
2844 |
// Now test the correct condition. |
|
2845 |
jint nval = (obj_array |
|
2846 |
? ((jint)Klass::_lh_array_tag_type_value |
|
2847 |
<< Klass::_lh_array_tag_shift) |
|
2848 |
: Klass::_lh_neutral_value); |
|
2849 |
Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); |
|
2850 |
BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array |
|
2851 |
// invert the test if we are looking for a non-array |
|
2852 |
if (not_array) btest = BoolTest(btest).negate(); |
|
2853 |
Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); |
|
2854 |
return generate_fair_guard(bol, region); |
|
2855 |
} |
|
2856 |
||
2857 |
||
2858 |
//-----------------------inline_native_newArray-------------------------- |
|
2859 |
bool LibraryCallKit::inline_native_newArray() { |
|
2860 |
int nargs = 2; |
|
2861 |
Node* mirror = argument(0); |
|
2862 |
Node* count_val = argument(1); |
|
2863 |
||
2864 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2865 |
mirror = do_null_check(mirror, T_OBJECT); |
|
2866 |
_sp -= nargs; |
|
589 | 2867 |
// If mirror or obj is dead, only null-path is taken. |
2868 |
if (stopped()) return true; |
|
1 | 2869 |
|
2870 |
enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; |
|
2871 |
RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
2872 |
PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, |
|
2873 |
TypeInstPtr::NOTNULL); |
|
2874 |
PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); |
|
2875 |
PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, |
|
2876 |
TypePtr::BOTTOM); |
|
2877 |
||
2878 |
bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
|
2879 |
Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, |
|
2880 |
nargs, |
|
2881 |
result_reg, _slow_path); |
|
2882 |
Node* normal_ctl = control(); |
|
2883 |
Node* no_array_ctl = result_reg->in(_slow_path); |
|
2884 |
||
2885 |
// Generate code for the slow case. We make a call to newArray(). |
|
2886 |
set_control(no_array_ctl); |
|
2887 |
if (!stopped()) { |
|
2888 |
// Either the input type is void.class, or else the |
|
2889 |
// array klass has not yet been cached. Either the |
|
2890 |
// ensuing call will throw an exception, or else it |
|
2891 |
// will cache the array klass for next time. |
|
2892 |
PreserveJVMState pjvms(this); |
|
2893 |
CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); |
|
2894 |
Node* slow_result = set_results_for_java_call(slow_call); |
|
2895 |
// this->control() comes from set_results_for_java_call |
|
2896 |
result_reg->set_req(_slow_path, control()); |
|
2897 |
result_val->set_req(_slow_path, slow_result); |
|
2898 |
result_io ->set_req(_slow_path, i_o()); |
|
2899 |
result_mem->set_req(_slow_path, reset_memory()); |
|
2900 |
} |
|
2901 |
||
2902 |
set_control(normal_ctl); |
|
2903 |
if (!stopped()) { |
|
2904 |
// Normal case: The array type has been cached in the java.lang.Class. |
|
2905 |
// The following call works fine even if the array type is polymorphic. |
|
2906 |
// It could be a dynamic mix of int[], boolean[], Object[], etc. |
|
2907 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2908 |
Node* obj = new_array(klass_node, count_val); |
|
2909 |
_sp -= nargs; |
|
2910 |
result_reg->init_req(_normal_path, control()); |
|
2911 |
result_val->init_req(_normal_path, obj); |
|
2912 |
result_io ->init_req(_normal_path, i_o()); |
|
2913 |
result_mem->init_req(_normal_path, reset_memory()); |
|
2914 |
} |
|
2915 |
||
2916 |
// Return the combined state. |
|
2917 |
set_i_o( _gvn.transform(result_io) ); |
|
2918 |
set_all_memory( _gvn.transform(result_mem) ); |
|
2919 |
push_result(result_reg, result_val); |
|
2920 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2921 |
||
2922 |
return true; |
|
2923 |
} |
|
2924 |
||
2925 |
//----------------------inline_native_getLength-------------------------- |
|
2926 |
bool LibraryCallKit::inline_native_getLength() { |
|
2927 |
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
2928 |
||
2929 |
int nargs = 1; |
|
2930 |
Node* array = argument(0); |
|
2931 |
||
2932 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2933 |
array = do_null_check(array, T_OBJECT); |
|
2934 |
_sp -= nargs; |
|
2935 |
||
2936 |
// If array is dead, only null-path is taken. |
|
2937 |
if (stopped()) return true; |
|
2938 |
||
2939 |
// Deoptimize if it is a non-array. |
|
2940 |
Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); |
|
2941 |
||
2942 |
if (non_array != NULL) { |
|
2943 |
PreserveJVMState pjvms(this); |
|
2944 |
set_control(non_array); |
|
2945 |
_sp += nargs; // push the arguments back on the stack |
|
2946 |
uncommon_trap(Deoptimization::Reason_intrinsic, |
|
2947 |
Deoptimization::Action_maybe_recompile); |
|
2948 |
} |
|
2949 |
||
2950 |
// If control is dead, only non-array-path is taken. |
|
2951 |
if (stopped()) return true; |
|
2952 |
||
2953 |
// The works fine even if the array type is polymorphic. |
|
2954 |
// It could be a dynamic mix of int[], boolean[], Object[], etc. |
|
2955 |
push( load_array_length(array) ); |
|
2956 |
||
2957 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
2958 |
||
2959 |
return true; |
|
2960 |
} |
|
2961 |
||
2962 |
//------------------------inline_array_copyOf---------------------------- |
|
2963 |
bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { |
|
2964 |
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
|
2965 |
||
2966 |
// Restore the stack and pop off the arguments. |
|
2967 |
int nargs = 3 + (is_copyOfRange? 1: 0); |
|
2968 |
Node* original = argument(0); |
|
2969 |
Node* start = is_copyOfRange? argument(1): intcon(0); |
|
2970 |
Node* end = is_copyOfRange? argument(2): argument(1); |
|
2971 |
Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); |
|
2972 |
||
2973 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2974 |
array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); |
|
2975 |
original = do_null_check(original, T_OBJECT); |
|
2976 |
_sp -= nargs; |
|
2977 |
||
2978 |
// Check if a null path was taken unconditionally. |
|
2979 |
if (stopped()) return true; |
|
2980 |
||
2981 |
Node* orig_length = load_array_length(original); |
|
2982 |
||
2983 |
Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs, |
|
2984 |
NULL, 0); |
|
2985 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
2986 |
klass_node = do_null_check(klass_node, T_OBJECT); |
|
2987 |
_sp -= nargs; |
|
2988 |
||
2989 |
RegionNode* bailout = new (C, 1) RegionNode(1); |
|
2990 |
record_for_igvn(bailout); |
|
2991 |
||
2992 |
// Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. |
|
2993 |
// Bail out if that is so. |
|
2994 |
Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); |
|
2995 |
if (not_objArray != NULL) { |
|
2996 |
// Improve the klass node's type from the new optimistic assumption: |
|
2997 |
ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); |
|
2998 |
const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); |
|
2999 |
Node* cast = new (C, 2) CastPPNode(klass_node, akls); |
|
3000 |
cast->init_req(0, control()); |
|
3001 |
klass_node = _gvn.transform(cast); |
|
3002 |
} |
|
3003 |
||
3004 |
// Bail out if either start or end is negative. |
|
3005 |
generate_negative_guard(start, bailout, &start); |
|
3006 |
generate_negative_guard(end, bailout, &end); |
|
3007 |
||
3008 |
Node* length = end; |
|
3009 |
if (_gvn.type(start) != TypeInt::ZERO) { |
|
3010 |
length = _gvn.transform( new (C, 3) SubINode(end, start) ); |
|
3011 |
} |
|
3012 |
||
3013 |
// Bail out if length is negative. |
|
3014 |
// ...Not needed, since the new_array will throw the right exception. |
|
3015 |
//generate_negative_guard(length, bailout, &length); |
|
3016 |
||
3017 |
if (bailout->req() > 1) { |
|
3018 |
PreserveJVMState pjvms(this); |
|
3019 |
set_control( _gvn.transform(bailout) ); |
|
3020 |
_sp += nargs; // push the arguments back on the stack |
|
3021 |
uncommon_trap(Deoptimization::Reason_intrinsic, |
|
3022 |
Deoptimization::Action_maybe_recompile); |
|
3023 |
} |
|
3024 |
||
3025 |
if (!stopped()) { |
|
3026 |
// How many elements will we copy from the original? |
|
3027 |
// The answer is MinI(orig_length - start, length). |
|
3028 |
Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); |
|
3029 |
Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); |
|
3030 |
||
3031 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
3032 |
Node* newcopy = new_array(klass_node, length); |
|
3033 |
_sp -= nargs; |
|
3034 |
||
3035 |
// Generate a direct call to the right arraycopy function(s). |
|
3036 |
// We know the copy is disjoint but we might not know if the |
|
3037 |
// oop stores need checking. |
|
3038 |
// Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). |
|
3039 |
// This will fail a store-check if x contains any non-nulls. |
|
3040 |
bool disjoint_bases = true; |
|
3041 |
bool length_never_negative = true; |
|
3042 |
generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, |
|
3043 |
original, start, newcopy, intcon(0), moved, |
|
3044 |
nargs, disjoint_bases, length_never_negative); |
|
3045 |
||
3046 |
push(newcopy); |
|
3047 |
} |
|
3048 |
||
3049 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
3050 |
||
3051 |
return true; |
|
3052 |
} |
|
3053 |
||
3054 |
||
3055 |
//----------------------generate_virtual_guard--------------------------- |
|
3056 |
// Helper for hashCode and clone. Peeks inside the vtable to avoid a call. |
|
3057 |
Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, |
|
3058 |
RegionNode* slow_region) { |
|
3059 |
ciMethod* method = callee(); |
|
3060 |
int vtable_index = method->vtable_index(); |
|
3061 |
// Get the methodOop out of the appropriate vtable entry. |
|
3062 |
int entry_offset = (instanceKlass::vtable_start_offset() + |
|
3063 |
vtable_index*vtableEntry::size()) * wordSize + |
|
3064 |
vtableEntry::method_offset_in_bytes(); |
|
3065 |
Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); |
|
3066 |
Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); |
|
3067 |
||
3068 |
// Compare the target method with the expected method (e.g., Object.hashCode). |
|
3069 |
const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); |
|
3070 |
||
3071 |
Node* native_call = makecon(native_call_addr); |
|
3072 |
Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); |
|
3073 |
Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); |
|
3074 |
||
3075 |
return generate_slow_guard(test_native, slow_region); |
|
3076 |
} |
|
3077 |
||
3078 |
//-----------------------generate_method_call---------------------------- |
|
3079 |
// Use generate_method_call to make a slow-call to the real |
|
3080 |
// method if the fast path fails. An alternative would be to |
|
3081 |
// use a stub like OptoRuntime::slow_arraycopy_Java. |
|
3082 |
// This only works for expanding the current library call, |
|
3083 |
// not another intrinsic. (E.g., don't use this for making an |
|
3084 |
// arraycopy call inside of the copyOf intrinsic.) |
|
3085 |
CallJavaNode* |
|
3086 |
LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { |
|
3087 |
// When compiling the intrinsic method itself, do not use this technique. |
|
3088 |
guarantee(callee() != C->method(), "cannot make slow-call to self"); |
|
3089 |
||
3090 |
ciMethod* method = callee(); |
|
3091 |
// ensure the JVMS we have will be correct for this call |
|
3092 |
guarantee(method_id == method->intrinsic_id(), "must match"); |
|
3093 |
||
3094 |
const TypeFunc* tf = TypeFunc::make(method); |
|
3095 |
int tfdc = tf->domain()->cnt(); |
|
3096 |
CallJavaNode* slow_call; |
|
3097 |
if (is_static) { |
|
3098 |
assert(!is_virtual, ""); |
|
3099 |
slow_call = new(C, tfdc) CallStaticJavaNode(tf, |
|
3100 |
SharedRuntime::get_resolve_static_call_stub(), |
|
3101 |
method, bci()); |
|
3102 |
} else if (is_virtual) { |
|
3103 |
null_check_receiver(method); |
|
3104 |
int vtable_index = methodOopDesc::invalid_vtable_index; |
|
3105 |
if (UseInlineCaches) { |
|
3106 |
// Suppress the vtable call |
|
3107 |
} else { |
|
3108 |
// hashCode and clone are not a miranda methods, |
|
3109 |
// so the vtable index is fixed. |
|
3110 |
// No need to use the linkResolver to get it. |
|
3111 |
vtable_index = method->vtable_index(); |
|
3112 |
} |
|
3113 |
slow_call = new(C, tfdc) CallDynamicJavaNode(tf, |
|
3114 |
SharedRuntime::get_resolve_virtual_call_stub(), |
|
3115 |
method, vtable_index, bci()); |
|
3116 |
} else { // neither virtual nor static: opt_virtual |
|
3117 |
null_check_receiver(method); |
|
3118 |
slow_call = new(C, tfdc) CallStaticJavaNode(tf, |
|
3119 |
SharedRuntime::get_resolve_opt_virtual_call_stub(), |
|
3120 |
method, bci()); |
|
3121 |
slow_call->set_optimized_virtual(true); |
|
3122 |
} |
|
3123 |
set_arguments_for_java_call(slow_call); |
|
3124 |
set_edges_for_java_call(slow_call); |
|
3125 |
return slow_call; |
|
3126 |
} |
|
3127 |
||
3128 |
||
3129 |
//------------------------------inline_native_hashcode-------------------- |
|
3130 |
// Build special case code for calls to hashCode on an object. |
|
3131 |
bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { |
|
3132 |
assert(is_static == callee()->is_static(), "correct intrinsic selection"); |
|
3133 |
assert(!(is_virtual && is_static), "either virtual, special, or static"); |
|
3134 |
||
3135 |
enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; |
|
3136 |
||
3137 |
RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
3138 |
PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, |
|
3139 |
TypeInt::INT); |
|
3140 |
PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); |
|
3141 |
PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, |
|
3142 |
TypePtr::BOTTOM); |
|
3143 |
Node* obj = NULL; |
|
3144 |
if (!is_static) { |
|
3145 |
// Check for hashing null object |
|
3146 |
obj = null_check_receiver(callee()); |
|
3147 |
if (stopped()) return true; // unconditionally null |
|
3148 |
result_reg->init_req(_null_path, top()); |
|
3149 |
result_val->init_req(_null_path, top()); |
|
3150 |
} else { |
|
3151 |
// Do a null check, and return zero if null. |
|
3152 |
// System.identityHashCode(null) == 0 |
|
3153 |
obj = argument(0); |
|
3154 |
Node* null_ctl = top(); |
|
3155 |
obj = null_check_oop(obj, &null_ctl); |
|
3156 |
result_reg->init_req(_null_path, null_ctl); |
|
3157 |
result_val->init_req(_null_path, _gvn.intcon(0)); |
|
3158 |
} |
|
3159 |
||
3160 |
// Unconditionally null? Then return right away. |
|
3161 |
if (stopped()) { |
|
3162 |
set_control( result_reg->in(_null_path) ); |
|
3163 |
if (!stopped()) |
|
3164 |
push( result_val ->in(_null_path) ); |
|
3165 |
return true; |
|
3166 |
} |
|
3167 |
||
3168 |
// After null check, get the object's klass. |
|
3169 |
Node* obj_klass = load_object_klass(obj); |
|
3170 |
||
3171 |
// This call may be virtual (invokevirtual) or bound (invokespecial). |
|
3172 |
// For each case we generate slightly different code. |
|
3173 |
||
3174 |
// We only go to the fast case code if we pass a number of guards. The |
|
3175 |
// paths which do not pass are accumulated in the slow_region. |
|
3176 |
RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
3177 |
record_for_igvn(slow_region); |
|
3178 |
||
3179 |
// If this is a virtual call, we generate a funny guard. We pull out |
|
3180 |
// the vtable entry corresponding to hashCode() from the target object. |
|
3181 |
// If the target method which we are calling happens to be the native |
|
3182 |
// Object hashCode() method, we pass the guard. We do not need this |
|
3183 |
// guard for non-virtual calls -- the caller is known to be the native |
|
3184 |
// Object hashCode(). |
|
3185 |
if (is_virtual) { |
|
3186 |
generate_virtual_guard(obj_klass, slow_region); |
|
3187 |
} |
|
3188 |
||
3189 |
// Get the header out of the object, use LoadMarkNode when available |
|
3190 |
Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); |
|
3191 |
Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS); |
|
3192 |
header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) ); |
|
3193 |
||
3194 |
// Test the header to see if it is unlocked. |
|
3195 |
Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); |
|
3196 |
Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); |
|
3197 |
Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); |
|
3198 |
Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); |
|
3199 |
Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); |
|
3200 |
||
3201 |
generate_slow_guard(test_unlocked, slow_region); |
|
3202 |
||
3203 |
// Get the hash value and check to see that it has been properly assigned. |
|
3204 |
// We depend on hash_mask being at most 32 bits and avoid the use of |
|
3205 |
// hash_mask_in_place because it could be larger than 32 bits in a 64-bit |
|
3206 |
// vm: see markOop.hpp. |
|
3207 |
Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); |
|
3208 |
Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); |
|
3209 |
Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); |
|
3210 |
// This hack lets the hash bits live anywhere in the mark object now, as long |
|
3211 |
// as the shift drops the relevent bits into the low 32 bits. Note that |
|
3212 |
// Java spec says that HashCode is an int so there's no point in capturing |
|
3213 |
// an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). |
|
3214 |
hshifted_header = ConvX2I(hshifted_header); |
|
3215 |
Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); |
|
3216 |
||
3217 |
Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); |
|
3218 |
Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); |
|
3219 |
Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); |
|
3220 |
||
3221 |
generate_slow_guard(test_assigned, slow_region); |
|
3222 |
||
3223 |
Node* init_mem = reset_memory(); |
|
3224 |
// fill in the rest of the null path: |
|
3225 |
result_io ->init_req(_null_path, i_o()); |
|
3226 |
result_mem->init_req(_null_path, init_mem); |
|
3227 |
||
3228 |
result_val->init_req(_fast_path, hash_val); |
|
3229 |
result_reg->init_req(_fast_path, control()); |
|
3230 |
result_io ->init_req(_fast_path, i_o()); |
|
3231 |
result_mem->init_req(_fast_path, init_mem); |
|
3232 |
||
3233 |
// Generate code for the slow case. We make a call to hashCode(). |
|
3234 |
set_control(_gvn.transform(slow_region)); |
|
3235 |
if (!stopped()) { |
|
3236 |
// No need for PreserveJVMState, because we're using up the present state. |
|
3237 |
set_all_memory(init_mem); |
|
3238 |
vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; |
|
3239 |
if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; |
|
3240 |
CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); |
|
3241 |
Node* slow_result = set_results_for_java_call(slow_call); |
|
3242 |
// this->control() comes from set_results_for_java_call |
|
3243 |
result_reg->init_req(_slow_path, control()); |
|
3244 |
result_val->init_req(_slow_path, slow_result); |
|
3245 |
result_io ->set_req(_slow_path, i_o()); |
|
3246 |
result_mem ->set_req(_slow_path, reset_memory()); |
|
3247 |
} |
|
3248 |
||
3249 |
// Return the combined state. |
|
3250 |
set_i_o( _gvn.transform(result_io) ); |
|
3251 |
set_all_memory( _gvn.transform(result_mem) ); |
|
3252 |
push_result(result_reg, result_val); |
|
3253 |
||
3254 |
return true; |
|
3255 |
} |
|
3256 |
||
3257 |
//---------------------------inline_native_getClass---------------------------- |
|
3258 |
// Build special case code for calls to hashCode on an object. |
|
3259 |
bool LibraryCallKit::inline_native_getClass() { |
|
3260 |
Node* obj = null_check_receiver(callee()); |
|
3261 |
if (stopped()) return true; |
|
3262 |
push( load_mirror_from_klass(load_object_klass(obj)) ); |
|
3263 |
return true; |
|
3264 |
} |
|
3265 |
||
3266 |
//-----------------inline_native_Reflection_getCallerClass--------------------- |
|
3267 |
// In the presence of deep enough inlining, getCallerClass() becomes a no-op. |
|
3268 |
// |
|
3269 |
// NOTE that this code must perform the same logic as |
|
3270 |
// vframeStream::security_get_caller_frame in that it must skip |
|
3271 |
// Method.invoke() and auxiliary frames. |
|
3272 |
||
3273 |
||
3274 |
||
3275 |
||
3276 |
bool LibraryCallKit::inline_native_Reflection_getCallerClass() { |
|
3277 |
ciMethod* method = callee(); |
|
3278 |
||
3279 |
#ifndef PRODUCT |
|
3280 |
if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3281 |
tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); |
|
3282 |
} |
|
3283 |
#endif |
|
3284 |
||
3285 |
debug_only(int saved_sp = _sp); |
|
3286 |
||
3287 |
// Argument words: (int depth) |
|
3288 |
int nargs = 1; |
|
3289 |
||
3290 |
_sp += nargs; |
|
3291 |
Node* caller_depth_node = pop(); |
|
3292 |
||
3293 |
assert(saved_sp == _sp, "must have correct argument count"); |
|
3294 |
||
3295 |
// The depth value must be a constant in order for the runtime call |
|
3296 |
// to be eliminated. |
|
3297 |
const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); |
|
3298 |
if (caller_depth_type == NULL || !caller_depth_type->is_con()) { |
|
3299 |
#ifndef PRODUCT |
|
3300 |
if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3301 |
tty->print_cr(" Bailing out because caller depth was not a constant"); |
|
3302 |
} |
|
3303 |
#endif |
|
3304 |
return false; |
|
3305 |
} |
|
3306 |
// Note that the JVM state at this point does not include the |
|
3307 |
// getCallerClass() frame which we are trying to inline. The |
|
3308 |
// semantics of getCallerClass(), however, are that the "first" |
|
3309 |
// frame is the getCallerClass() frame, so we subtract one from the |
|
3310 |
// requested depth before continuing. We don't inline requests of |
|
3311 |
// getCallerClass(0). |
|
3312 |
int caller_depth = caller_depth_type->get_con() - 1; |
|
3313 |
if (caller_depth < 0) { |
|
3314 |
#ifndef PRODUCT |
|
3315 |
if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3316 |
tty->print_cr(" Bailing out because caller depth was %d", caller_depth); |
|
3317 |
} |
|
3318 |
#endif |
|
3319 |
return false; |
|
3320 |
} |
|
3321 |
||
3322 |
if (!jvms()->has_method()) { |
|
3323 |
#ifndef PRODUCT |
|
3324 |
if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3325 |
tty->print_cr(" Bailing out because intrinsic was inlined at top level"); |
|
3326 |
} |
|
3327 |
#endif |
|
3328 |
return false; |
|
3329 |
} |
|
3330 |
int _depth = jvms()->depth(); // cache call chain depth |
|
3331 |
||
3332 |
// Walk back up the JVM state to find the caller at the required |
|
3333 |
// depth. NOTE that this code must perform the same logic as |
|
3334 |
// vframeStream::security_get_caller_frame in that it must skip |
|
3335 |
// Method.invoke() and auxiliary frames. Note also that depth is |
|
3336 |
// 1-based (1 is the bottom of the inlining). |
|
3337 |
int inlining_depth = _depth; |
|
3338 |
JVMState* caller_jvms = NULL; |
|
3339 |
||
3340 |
if (inlining_depth > 0) { |
|
3341 |
caller_jvms = jvms(); |
|
3342 |
assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); |
|
3343 |
do { |
|
3344 |
// The following if-tests should be performed in this order |
|
3345 |
if (is_method_invoke_or_aux_frame(caller_jvms)) { |
|
3346 |
// Skip a Method.invoke() or auxiliary frame |
|
3347 |
} else if (caller_depth > 0) { |
|
3348 |
// Skip real frame |
|
3349 |
--caller_depth; |
|
3350 |
} else { |
|
3351 |
// We're done: reached desired caller after skipping. |
|
3352 |
break; |
|
3353 |
} |
|
3354 |
caller_jvms = caller_jvms->caller(); |
|
3355 |
--inlining_depth; |
|
3356 |
} while (inlining_depth > 0); |
|
3357 |
} |
|
3358 |
||
3359 |
if (inlining_depth == 0) { |
|
3360 |
#ifndef PRODUCT |
|
3361 |
if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3362 |
tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); |
|
3363 |
tty->print_cr(" JVM state at this point:"); |
|
3364 |
for (int i = _depth; i >= 1; i--) { |
|
3365 |
tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); |
|
3366 |
} |
|
3367 |
} |
|
3368 |
#endif |
|
3369 |
return false; // Reached end of inlining |
|
3370 |
} |
|
3371 |
||
3372 |
// Acquire method holder as java.lang.Class |
|
3373 |
ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); |
|
3374 |
ciInstance* caller_mirror = caller_klass->java_mirror(); |
|
3375 |
// Push this as a constant |
|
3376 |
push(makecon(TypeInstPtr::make(caller_mirror))); |
|
3377 |
#ifndef PRODUCT |
|
3378 |
if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { |
|
3379 |
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); |
|
3380 |
tty->print_cr(" JVM state at this point:"); |
|
3381 |
for (int i = _depth; i >= 1; i--) { |
|
3382 |
tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); |
|
3383 |
} |
|
3384 |
} |
|
3385 |
#endif |
|
3386 |
return true; |
|
3387 |
} |
|
3388 |
||
3389 |
// Helper routine for above |
|
3390 |
bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { |
|
3391 |
// Is this the Method.invoke method itself? |
|
3392 |
if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke) |
|
3393 |
return true; |
|
3394 |
||
3395 |
// Is this a helper, defined somewhere underneath MethodAccessorImpl. |
|
3396 |
ciKlass* k = jvms->method()->holder(); |
|
3397 |
if (k->is_instance_klass()) { |
|
3398 |
ciInstanceKlass* ik = k->as_instance_klass(); |
|
3399 |
for (; ik != NULL; ik = ik->super()) { |
|
3400 |
if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && |
|
3401 |
ik == env()->find_system_klass(ik->name())) { |
|
3402 |
return true; |
|
3403 |
} |
|
3404 |
} |
|
3405 |
} |
|
3406 |
||
3407 |
return false; |
|
3408 |
} |
|
3409 |
||
3410 |
static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by |
|
3411 |
// inline_native_AtomicLong_attemptUpdate() but it has no way of |
|
3412 |
// computing it since there is no lookup field by name function in the |
|
3413 |
// CI interface. This is computed and set by inline_native_AtomicLong_get(). |
|
3414 |
// Using a static variable here is safe even if we have multiple compilation |
|
3415 |
// threads because the offset is constant. At worst the same offset will be |
|
3416 |
// computed and stored multiple |
|
3417 |
||
3418 |
bool LibraryCallKit::inline_native_AtomicLong_get() { |
|
3419 |
// Restore the stack and pop off the argument |
|
3420 |
_sp+=1; |
|
3421 |
Node *obj = pop(); |
|
3422 |
||
3423 |
// get the offset of the "value" field. Since the CI interfaces |
|
3424 |
// does not provide a way to look up a field by name, we scan the bytecodes |
|
3425 |
// to get the field index. We expect the first 2 instructions of the method |
|
3426 |
// to be: |
|
3427 |
// 0 aload_0 |
|
3428 |
// 1 getfield "value" |
|
3429 |
ciMethod* method = callee(); |
|
3430 |
if (value_field_offset == -1) |
|
3431 |
{ |
|
3432 |
ciField* value_field; |
|
3433 |
ciBytecodeStream iter(method); |
|
3434 |
Bytecodes::Code bc = iter.next(); |
|
3435 |
||
3436 |
if ((bc != Bytecodes::_aload_0) && |
|
3437 |
((bc != Bytecodes::_aload) || (iter.get_index() != 0))) |
|
3438 |
return false; |
|
3439 |
bc = iter.next(); |
|
3440 |
if (bc != Bytecodes::_getfield) |
|
3441 |
return false; |
|
3442 |
bool ignore; |
|
3443 |
value_field = iter.get_field(ignore); |
|
3444 |
value_field_offset = value_field->offset_in_bytes(); |
|
3445 |
} |
|
3446 |
||
3447 |
// Null check without removing any arguments. |
|
3448 |
_sp++; |
|
3449 |
obj = do_null_check(obj, T_OBJECT); |
|
3450 |
_sp--; |
|
3451 |
// Check for locking null object |
|
3452 |
if (stopped()) return true; |
|
3453 |
||
3454 |
Node *adr = basic_plus_adr(obj, obj, value_field_offset); |
|
3455 |
const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); |
|
3456 |
int alias_idx = C->get_alias_index(adr_type); |
|
3457 |
||
3458 |
Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); |
|
3459 |
||
3460 |
push_pair(result); |
|
3461 |
||
3462 |
return true; |
|
3463 |
} |
|
3464 |
||
3465 |
bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { |
|
3466 |
// Restore the stack and pop off the arguments |
|
3467 |
_sp+=5; |
|
3468 |
Node *newVal = pop_pair(); |
|
3469 |
Node *oldVal = pop_pair(); |
|
3470 |
Node *obj = pop(); |
|
3471 |
||
3472 |
// we need the offset of the "value" field which was computed when |
|
3473 |
// inlining the get() method. Give up if we don't have it. |
|
3474 |
if (value_field_offset == -1) |
|
3475 |
return false; |
|
3476 |
||
3477 |
// Null check without removing any arguments. |
|
3478 |
_sp+=5; |
|
3479 |
obj = do_null_check(obj, T_OBJECT); |
|
3480 |
_sp-=5; |
|
3481 |
// Check for locking null object |
|
3482 |
if (stopped()) return true; |
|
3483 |
||
3484 |
Node *adr = basic_plus_adr(obj, obj, value_field_offset); |
|
3485 |
const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); |
|
3486 |
int alias_idx = C->get_alias_index(adr_type); |
|
3487 |
||
1500
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3488 |
Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3489 |
Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); |
1 | 3490 |
set_memory(store_proj, alias_idx); |
1500
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3491 |
Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3492 |
|
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3493 |
Node *result; |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3494 |
// CMove node is not used to be able fold a possible check code |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3495 |
// after attemptUpdate() call. This code could be transformed |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3496 |
// into CMove node by loop optimizations. |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3497 |
{ |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3498 |
RegionNode *r = new (C, 3) RegionNode(3); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3499 |
result = new (C, 3) PhiNode(r, TypeInt::BOOL); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3500 |
|
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3501 |
Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3502 |
Node *iftrue = opt_iff(r, iff); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3503 |
r->init_req(1, iftrue); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3504 |
result->init_req(1, intcon(1)); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3505 |
result->init_req(2, intcon(0)); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3506 |
|
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3507 |
set_control(_gvn.transform(r)); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3508 |
record_for_igvn(r); |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3509 |
|
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3510 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3511 |
} |
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3512 |
|
bea9a90f3e8f
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
781
diff
changeset
|
3513 |
push(_gvn.transform(result)); |
1 | 3514 |
return true; |
3515 |
} |
|
3516 |
||
3517 |
bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { |
|
3518 |
// restore the arguments |
|
3519 |
_sp += arg_size(); |
|
3520 |
||
3521 |
switch (id) { |
|
3522 |
case vmIntrinsics::_floatToRawIntBits: |
|
3523 |
push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); |
|
3524 |
break; |
|
3525 |
||
3526 |
case vmIntrinsics::_intBitsToFloat: |
|
3527 |
push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); |
|
3528 |
break; |
|
3529 |
||
3530 |
case vmIntrinsics::_doubleToRawLongBits: |
|
3531 |
push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); |
|
3532 |
break; |
|
3533 |
||
3534 |
case vmIntrinsics::_longBitsToDouble: |
|
3535 |
push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); |
|
3536 |
break; |
|
3537 |
||
3538 |
case vmIntrinsics::_doubleToLongBits: { |
|
3539 |
Node* value = pop_pair(); |
|
3540 |
||
3541 |
// two paths (plus control) merge in a wood |
|
3542 |
RegionNode *r = new (C, 3) RegionNode(3); |
|
3543 |
Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); |
|
3544 |
||
3545 |
Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); |
|
3546 |
// Build the boolean node |
|
3547 |
Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); |
|
3548 |
||
3549 |
// Branch either way. |
|
3550 |
// NaN case is less traveled, which makes all the difference. |
|
3551 |
IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
|
3552 |
Node *opt_isnan = _gvn.transform(ifisnan); |
|
3553 |
assert( opt_isnan->is_If(), "Expect an IfNode"); |
|
3554 |
IfNode *opt_ifisnan = (IfNode*)opt_isnan; |
|
3555 |
Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); |
|
3556 |
||
3557 |
set_control(iftrue); |
|
3558 |
||
3559 |
static const jlong nan_bits = CONST64(0x7ff8000000000000); |
|
3560 |
Node *slow_result = longcon(nan_bits); // return NaN |
|
3561 |
phi->init_req(1, _gvn.transform( slow_result )); |
|
3562 |
r->init_req(1, iftrue); |
|
3563 |
||
3564 |
// Else fall through |
|
3565 |
Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); |
|
3566 |
set_control(iffalse); |
|
3567 |
||
3568 |
phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); |
|
3569 |
r->init_req(2, iffalse); |
|
3570 |
||
3571 |
// Post merge |
|
3572 |
set_control(_gvn.transform(r)); |
|
3573 |
record_for_igvn(r); |
|
3574 |
||
3575 |
Node* result = _gvn.transform(phi); |
|
3576 |
assert(result->bottom_type()->isa_long(), "must be"); |
|
3577 |
push_pair(result); |
|
3578 |
||
3579 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
3580 |
||
3581 |
break; |
|
3582 |
} |
|
3583 |
||
3584 |
case vmIntrinsics::_floatToIntBits: { |
|
3585 |
Node* value = pop(); |
|
3586 |
||
3587 |
// two paths (plus control) merge in a wood |
|
3588 |
RegionNode *r = new (C, 3) RegionNode(3); |
|
3589 |
Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); |
|
3590 |
||
3591 |
Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); |
|
3592 |
// Build the boolean node |
|
3593 |
Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); |
|
3594 |
||
3595 |
// Branch either way. |
|
3596 |
// NaN case is less traveled, which makes all the difference. |
|
3597 |
IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
|
3598 |
Node *opt_isnan = _gvn.transform(ifisnan); |
|
3599 |
assert( opt_isnan->is_If(), "Expect an IfNode"); |
|
3600 |
IfNode *opt_ifisnan = (IfNode*)opt_isnan; |
|
3601 |
Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); |
|
3602 |
||
3603 |
set_control(iftrue); |
|
3604 |
||
3605 |
static const jint nan_bits = 0x7fc00000; |
|
3606 |
Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN |
|
3607 |
phi->init_req(1, _gvn.transform( slow_result )); |
|
3608 |
r->init_req(1, iftrue); |
|
3609 |
||
3610 |
// Else fall through |
|
3611 |
Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); |
|
3612 |
set_control(iffalse); |
|
3613 |
||
3614 |
phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); |
|
3615 |
r->init_req(2, iffalse); |
|
3616 |
||
3617 |
// Post merge |
|
3618 |
set_control(_gvn.transform(r)); |
|
3619 |
record_for_igvn(r); |
|
3620 |
||
3621 |
Node* result = _gvn.transform(phi); |
|
3622 |
assert(result->bottom_type()->isa_int(), "must be"); |
|
3623 |
push(result); |
|
3624 |
||
3625 |
C->set_has_split_ifs(true); // Has chance for split-if optimization |
|
3626 |
||
3627 |
break; |
|
3628 |
} |
|
3629 |
||
3630 |
default: |
|
3631 |
ShouldNotReachHere(); |
|
3632 |
} |
|
3633 |
||
3634 |
return true; |
|
3635 |
} |
|
3636 |
||
3637 |
#ifdef _LP64 |
|
3638 |
#define XTOP ,top() /*additional argument*/ |
|
3639 |
#else //_LP64 |
|
3640 |
#define XTOP /*no additional argument*/ |
|
3641 |
#endif //_LP64 |
|
3642 |
||
3643 |
//----------------------inline_unsafe_copyMemory------------------------- |
|
3644 |
bool LibraryCallKit::inline_unsafe_copyMemory() { |
|
3645 |
if (callee()->is_static()) return false; // caller must have the capability! |
|
3646 |
int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) |
|
3647 |
assert(signature()->size() == nargs-1, "copy has 5 arguments"); |
|
3648 |
null_check_receiver(callee()); // check then ignore argument(0) |
|
3649 |
if (stopped()) return true; |
|
3650 |
||
3651 |
C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
|
3652 |
||
3653 |
Node* src_ptr = argument(1); |
|
3654 |
Node* src_off = ConvL2X(argument(2)); |
|
3655 |
assert(argument(3)->is_top(), "2nd half of long"); |
|
3656 |
Node* dst_ptr = argument(4); |
|
3657 |
Node* dst_off = ConvL2X(argument(5)); |
|
3658 |
assert(argument(6)->is_top(), "2nd half of long"); |
|
3659 |
Node* size = ConvL2X(argument(7)); |
|
3660 |
assert(argument(8)->is_top(), "2nd half of long"); |
|
3661 |
||
3662 |
assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
|
3663 |
"fieldOffset must be byte-scaled"); |
|
3664 |
||
3665 |
Node* src = make_unsafe_address(src_ptr, src_off); |
|
3666 |
Node* dst = make_unsafe_address(dst_ptr, dst_off); |
|
3667 |
||
3668 |
// Conservatively insert a memory barrier on all memory slices. |
|
3669 |
// Do not let writes of the copy source or destination float below the copy. |
|
3670 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
3671 |
||
3672 |
// Call it. Note that the length argument is not scaled. |
|
3673 |
make_runtime_call(RC_LEAF|RC_NO_FP, |
|
3674 |
OptoRuntime::fast_arraycopy_Type(), |
|
3675 |
StubRoutines::unsafe_arraycopy(), |
|
3676 |
"unsafe_arraycopy", |
|
3677 |
TypeRawPtr::BOTTOM, |
|
3678 |
src, dst, size XTOP); |
|
3679 |
||
3680 |
// Do not let reads of the copy destination float above the copy. |
|
3681 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
3682 |
||
3683 |
return true; |
|
3684 |
} |
|
3685 |
||
3686 |
||
3687 |
//------------------------inline_native_clone---------------------------- |
|
3688 |
// Here are the simple edge cases: |
|
3689 |
// null receiver => normal trap |
|
3690 |
// virtual and clone was overridden => slow path to out-of-line clone |
|
3691 |
// not cloneable or finalizer => slow path to out-of-line Object.clone |
|
3692 |
// |
|
3693 |
// The general case has two steps, allocation and copying. |
|
3694 |
// Allocation has two cases, and uses GraphKit::new_instance or new_array. |
|
3695 |
// |
|
3696 |
// Copying also has two cases, oop arrays and everything else. |
|
3697 |
// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). |
|
3698 |
// Everything else uses the tight inline loop supplied by CopyArrayNode. |
|
3699 |
// |
|
3700 |
// These steps fold up nicely if and when the cloned object's klass |
|
3701 |
// can be sharply typed as an object array, a type array, or an instance. |
|
3702 |
// |
|
3703 |
bool LibraryCallKit::inline_native_clone(bool is_virtual) { |
|
3704 |
int nargs = 1; |
|
3705 |
Node* obj = null_check_receiver(callee()); |
|
3706 |
if (stopped()) return true; |
|
3707 |
Node* obj_klass = load_object_klass(obj); |
|
3708 |
const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); |
|
3709 |
const TypeOopPtr* toop = ((tklass != NULL) |
|
3710 |
? tklass->as_instance_type() |
|
3711 |
: TypeInstPtr::NOTNULL); |
|
3712 |
||
3713 |
// Conservatively insert a memory barrier on all memory slices. |
|
3714 |
// Do not let writes into the original float below the clone. |
|
3715 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
3716 |
||
3717 |
// paths into result_reg: |
|
3718 |
enum { |
|
3719 |
_slow_path = 1, // out-of-line call to clone method (virtual or not) |
|
3720 |
_objArray_path, // plain allocation, plus arrayof_oop_arraycopy |
|
3721 |
_fast_path, // plain allocation, plus a CopyArray operation |
|
3722 |
PATH_LIMIT |
|
3723 |
}; |
|
3724 |
RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
3725 |
PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, |
|
3726 |
TypeInstPtr::NOTNULL); |
|
3727 |
PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); |
|
3728 |
PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, |
|
3729 |
TypePtr::BOTTOM); |
|
3730 |
record_for_igvn(result_reg); |
|
3731 |
||
3732 |
const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; |
|
3733 |
int raw_adr_idx = Compile::AliasIdxRaw; |
|
3734 |
const bool raw_mem_only = true; |
|
3735 |
||
3736 |
// paths into alloc_reg (on the fast path, just before the CopyArray): |
|
3737 |
enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT }; |
|
3738 |
RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT); |
|
3739 |
PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type); |
|
3740 |
PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X); |
|
3741 |
PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO); |
|
3742 |
PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY, |
|
3743 |
raw_adr_type); |
|
3744 |
record_for_igvn(alloc_reg); |
|
3745 |
||
3746 |
bool card_mark = false; // (see below) |
|
3747 |
||
3748 |
Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); |
|
3749 |
if (array_ctl != NULL) { |
|
3750 |
// It's an array. |
|
3751 |
PreserveJVMState pjvms(this); |
|
3752 |
set_control(array_ctl); |
|
3753 |
Node* obj_length = load_array_length(obj); |
|
3754 |
Node* obj_size = NULL; |
|
3755 |
_sp += nargs; // set original stack for use by uncommon_trap |
|
3756 |
Node* alloc_obj = new_array(obj_klass, obj_length, |
|
3757 |
raw_mem_only, &obj_size); |
|
3758 |
_sp -= nargs; |
|
3759 |
assert(obj_size != NULL, ""); |
|
3760 |
Node* raw_obj = alloc_obj->in(1); |
|
3761 |
assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); |
|
3762 |
if (ReduceBulkZeroing) { |
|
3763 |
AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); |
|
3764 |
if (alloc != NULL) { |
|
3765 |
// We will be completely responsible for initializing this object. |
|
3766 |
alloc->maybe_set_complete(&_gvn); |
|
3767 |
} |
|
3768 |
} |
|
3769 |
||
3770 |
if (!use_ReduceInitialCardMarks()) { |
|
3771 |
// If it is an oop array, it requires very special treatment, |
|
3772 |
// because card marking is required on each card of the array. |
|
3773 |
Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); |
|
3774 |
if (is_obja != NULL) { |
|
3775 |
PreserveJVMState pjvms2(this); |
|
3776 |
set_control(is_obja); |
|
3777 |
// Generate a direct call to the right arraycopy function(s). |
|
3778 |
bool disjoint_bases = true; |
|
3779 |
bool length_never_negative = true; |
|
3780 |
generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, |
|
3781 |
obj, intcon(0), alloc_obj, intcon(0), |
|
3782 |
obj_length, nargs, |
|
3783 |
disjoint_bases, length_never_negative); |
|
3784 |
result_reg->init_req(_objArray_path, control()); |
|
3785 |
result_val->init_req(_objArray_path, alloc_obj); |
|
3786 |
result_i_o ->set_req(_objArray_path, i_o()); |
|
3787 |
result_mem ->set_req(_objArray_path, reset_memory()); |
|
3788 |
} |
|
3789 |
} |
|
3790 |
// We can dispense with card marks if we know the allocation |
|
3791 |
// comes out of eden (TLAB)... In fact, ReduceInitialCardMarks |
|
3792 |
// causes the non-eden paths to simulate a fresh allocation, |
|
3793 |
// insofar that no further card marks are required to initialize |
|
3794 |
// the object. |
|
3795 |
||
3796 |
// Otherwise, there are no card marks to worry about. |
|
3797 |
alloc_val->init_req(_typeArray_alloc, raw_obj); |
|
3798 |
alloc_siz->init_req(_typeArray_alloc, obj_size); |
|
3799 |
alloc_reg->init_req(_typeArray_alloc, control()); |
|
3800 |
alloc_i_o->init_req(_typeArray_alloc, i_o()); |
|
3801 |
alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type)); |
|
3802 |
} |
|
3803 |
||
3804 |
// We only go to the fast case code if we pass a number of guards. |
|
3805 |
// The paths which do not pass are accumulated in the slow_region. |
|
3806 |
RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
3807 |
record_for_igvn(slow_region); |
|
3808 |
if (!stopped()) { |
|
3809 |
// It's an instance. Make the slow-path tests. |
|
3810 |
// If this is a virtual call, we generate a funny guard. We grab |
|
3811 |
// the vtable entry corresponding to clone() from the target object. |
|
3812 |
// If the target method which we are calling happens to be the |
|
3813 |
// Object clone() method, we pass the guard. We do not need this |
|
3814 |
// guard for non-virtual calls; the caller is known to be the native |
|
3815 |
// Object clone(). |
|
3816 |
if (is_virtual) { |
|
3817 |
generate_virtual_guard(obj_klass, slow_region); |
|
3818 |
} |
|
3819 |
||
3820 |
// The object must be cloneable and must not have a finalizer. |
|
3821 |
// Both of these conditions may be checked in a single test. |
|
3822 |
// We could optimize the cloneable test further, but we don't care. |
|
3823 |
generate_access_flags_guard(obj_klass, |
|
3824 |
// Test both conditions: |
|
3825 |
JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, |
|
3826 |
// Must be cloneable but not finalizer: |
|
3827 |
JVM_ACC_IS_CLONEABLE, |
|
3828 |
slow_region); |
|
3829 |
} |
|
3830 |
||
3831 |
if (!stopped()) { |
|
3832 |
// It's an instance, and it passed the slow-path tests. |
|
3833 |
PreserveJVMState pjvms(this); |
|
3834 |
Node* obj_size = NULL; |
|
3835 |
Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size); |
|
3836 |
assert(obj_size != NULL, ""); |
|
3837 |
Node* raw_obj = alloc_obj->in(1); |
|
3838 |
assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); |
|
3839 |
if (ReduceBulkZeroing) { |
|
3840 |
AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); |
|
3841 |
if (alloc != NULL && !alloc->maybe_set_complete(&_gvn)) |
|
3842 |
alloc = NULL; |
|
3843 |
} |
|
3844 |
if (!use_ReduceInitialCardMarks()) { |
|
3845 |
// Put in store barrier for any and all oops we are sticking |
|
3846 |
// into this object. (We could avoid this if we could prove |
|
3847 |
// that the object type contains no oop fields at all.) |
|
3848 |
card_mark = true; |
|
3849 |
} |
|
3850 |
alloc_val->init_req(_instance_alloc, raw_obj); |
|
3851 |
alloc_siz->init_req(_instance_alloc, obj_size); |
|
3852 |
alloc_reg->init_req(_instance_alloc, control()); |
|
3853 |
alloc_i_o->init_req(_instance_alloc, i_o()); |
|
3854 |
alloc_mem->init_req(_instance_alloc, memory(raw_adr_type)); |
|
3855 |
} |
|
3856 |
||
3857 |
// Generate code for the slow case. We make a call to clone(). |
|
3858 |
set_control(_gvn.transform(slow_region)); |
|
3859 |
if (!stopped()) { |
|
3860 |
PreserveJVMState pjvms(this); |
|
3861 |
CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); |
|
3862 |
Node* slow_result = set_results_for_java_call(slow_call); |
|
3863 |
// this->control() comes from set_results_for_java_call |
|
3864 |
result_reg->init_req(_slow_path, control()); |
|
3865 |
result_val->init_req(_slow_path, slow_result); |
|
3866 |
result_i_o ->set_req(_slow_path, i_o()); |
|
3867 |
result_mem ->set_req(_slow_path, reset_memory()); |
|
3868 |
} |
|
3869 |
||
3870 |
// The object is allocated, as an array and/or an instance. Now copy it. |
|
3871 |
set_control( _gvn.transform(alloc_reg) ); |
|
3872 |
set_i_o( _gvn.transform(alloc_i_o) ); |
|
3873 |
set_memory( _gvn.transform(alloc_mem), raw_adr_type ); |
|
3874 |
Node* raw_obj = _gvn.transform(alloc_val); |
|
3875 |
||
3876 |
if (!stopped()) { |
|
3877 |
// Copy the fastest available way. |
|
3878 |
// (No need for PreserveJVMState, since we're using it all up now.) |
|
589 | 3879 |
// TODO: generate fields/elements copies for small objects instead. |
1 | 3880 |
Node* src = obj; |
3881 |
Node* dest = raw_obj; |
|
3882 |
Node* size = _gvn.transform(alloc_siz); |
|
3883 |
||
3884 |
// Exclude the header. |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
3885 |
int base_off = instanceOopDesc::base_offset_in_bytes(); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
3886 |
if (UseCompressedOops) { |
589 | 3887 |
assert(base_off % BytesPerLong != 0, "base with compressed oops"); |
3888 |
// With compressed oops base_offset_in_bytes is 12 which creates |
|
3889 |
// the gap since countx is rounded by 8 bytes below. |
|
3890 |
// Copy klass and the gap. |
|
3891 |
base_off = instanceOopDesc::klass_offset_in_bytes(); |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
3892 |
} |
1 | 3893 |
src = basic_plus_adr(src, base_off); |
3894 |
dest = basic_plus_adr(dest, base_off); |
|
3895 |
||
3896 |
// Compute the length also, if needed: |
|
3897 |
Node* countx = size; |
|
3898 |
countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); |
|
3899 |
countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); |
|
3900 |
||
3901 |
// Select an appropriate instruction to initialize the range. |
|
3902 |
// The CopyArray instruction (if supported) can be optimized |
|
3903 |
// into a discrete set of scalar loads and stores. |
|
3904 |
bool disjoint_bases = true; |
|
3905 |
generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, |
|
3906 |
src, NULL, dest, NULL, countx); |
|
3907 |
||
3908 |
// Now that the object is properly initialized, type it as an oop. |
|
3909 |
// Use a secondary InitializeNode memory barrier. |
|
3910 |
InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx, |
|
3911 |
raw_obj)->as_Initialize(); |
|
3912 |
init->set_complete(&_gvn); // (there is no corresponding AllocateNode) |
|
3913 |
Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj, |
|
3914 |
TypeInstPtr::NOTNULL); |
|
3915 |
new_obj = _gvn.transform(new_obj); |
|
3916 |
||
3917 |
// If necessary, emit some card marks afterwards. (Non-arrays only.) |
|
3918 |
if (card_mark) { |
|
3919 |
Node* no_particular_value = NULL; |
|
3920 |
Node* no_particular_field = NULL; |
|
3921 |
post_barrier(control(), |
|
3922 |
memory(raw_adr_type), |
|
3923 |
new_obj, |
|
3924 |
no_particular_field, |
|
3925 |
raw_adr_idx, |
|
3926 |
no_particular_value, |
|
3927 |
T_OBJECT, |
|
3928 |
false); |
|
3929 |
} |
|
3930 |
// Present the results of the slow call. |
|
3931 |
result_reg->init_req(_fast_path, control()); |
|
3932 |
result_val->init_req(_fast_path, new_obj); |
|
3933 |
result_i_o ->set_req(_fast_path, i_o()); |
|
3934 |
result_mem ->set_req(_fast_path, reset_memory()); |
|
3935 |
} |
|
3936 |
||
3937 |
// Return the combined state. |
|
3938 |
set_control( _gvn.transform(result_reg) ); |
|
3939 |
set_i_o( _gvn.transform(result_i_o) ); |
|
3940 |
set_all_memory( _gvn.transform(result_mem) ); |
|
3941 |
||
3942 |
// Cast the result to a sharper type, since we know what clone does. |
|
3943 |
Node* new_obj = _gvn.transform(result_val); |
|
3944 |
Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop); |
|
3945 |
push(_gvn.transform(cast)); |
|
3946 |
||
3947 |
return true; |
|
3948 |
} |
|
3949 |
||
3950 |
||
3951 |
// constants for computing the copy function |
|
3952 |
enum { |
|
3953 |
COPYFUNC_UNALIGNED = 0, |
|
3954 |
COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize |
|
3955 |
COPYFUNC_CONJOINT = 0, |
|
3956 |
COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend |
|
3957 |
}; |
|
3958 |
||
3959 |
// Note: The condition "disjoint" applies also for overlapping copies |
|
3960 |
// where an descending copy is permitted (i.e., dest_offset <= src_offset). |
|
3961 |
static address |
|
3962 |
select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) { |
|
3963 |
int selector = |
|
3964 |
(aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) + |
|
3965 |
(disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT); |
|
3966 |
||
3967 |
#define RETURN_STUB(xxx_arraycopy) { \ |
|
3968 |
name = #xxx_arraycopy; \ |
|
3969 |
return StubRoutines::xxx_arraycopy(); } |
|
3970 |
||
3971 |
switch (t) { |
|
3972 |
case T_BYTE: |
|
3973 |
case T_BOOLEAN: |
|
3974 |
switch (selector) { |
|
3975 |
case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy); |
|
3976 |
case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy); |
|
3977 |
case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy); |
|
3978 |
case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy); |
|
3979 |
} |
|
3980 |
case T_CHAR: |
|
3981 |
case T_SHORT: |
|
3982 |
switch (selector) { |
|
3983 |
case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy); |
|
3984 |
case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy); |
|
3985 |
case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy); |
|
3986 |
case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy); |
|
3987 |
} |
|
3988 |
case T_INT: |
|
3989 |
case T_FLOAT: |
|
3990 |
switch (selector) { |
|
3991 |
case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy); |
|
3992 |
case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy); |
|
3993 |
case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy); |
|
3994 |
case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy); |
|
3995 |
} |
|
3996 |
case T_DOUBLE: |
|
3997 |
case T_LONG: |
|
3998 |
switch (selector) { |
|
3999 |
case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy); |
|
4000 |
case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy); |
|
4001 |
case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy); |
|
4002 |
case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy); |
|
4003 |
} |
|
4004 |
case T_ARRAY: |
|
4005 |
case T_OBJECT: |
|
4006 |
switch (selector) { |
|
4007 |
case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy); |
|
4008 |
case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy); |
|
4009 |
case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy); |
|
4010 |
case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy); |
|
4011 |
} |
|
4012 |
default: |
|
4013 |
ShouldNotReachHere(); |
|
4014 |
return NULL; |
|
4015 |
} |
|
4016 |
||
4017 |
#undef RETURN_STUB |
|
4018 |
} |
|
4019 |
||
4020 |
//------------------------------basictype2arraycopy---------------------------- |
|
4021 |
address LibraryCallKit::basictype2arraycopy(BasicType t, |
|
4022 |
Node* src_offset, |
|
4023 |
Node* dest_offset, |
|
4024 |
bool disjoint_bases, |
|
4025 |
const char* &name) { |
|
4026 |
const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; |
|
4027 |
const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; |
|
4028 |
||
4029 |
bool aligned = false; |
|
4030 |
bool disjoint = disjoint_bases; |
|
4031 |
||
4032 |
// if the offsets are the same, we can treat the memory regions as |
|
4033 |
// disjoint, because either the memory regions are in different arrays, |
|
4034 |
// or they are identical (which we can treat as disjoint.) We can also |
|
4035 |
// treat a copy with a destination index less that the source index |
|
4036 |
// as disjoint since a low->high copy will work correctly in this case. |
|
4037 |
if (src_offset_inttype != NULL && src_offset_inttype->is_con() && |
|
4038 |
dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { |
|
4039 |
// both indices are constants |
|
4040 |
int s_offs = src_offset_inttype->get_con(); |
|
4041 |
int d_offs = dest_offset_inttype->get_con(); |
|
202
dc13bf0e5d5d
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
1
diff
changeset
|
4042 |
int element_size = type2aelembytes(t); |
1 | 4043 |
aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && |
4044 |
((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); |
|
4045 |
if (s_offs >= d_offs) disjoint = true; |
|
4046 |
} else if (src_offset == dest_offset && src_offset != NULL) { |
|
4047 |
// This can occur if the offsets are identical non-constants. |
|
4048 |
disjoint = true; |
|
4049 |
} |
|
4050 |
||
4051 |
return select_arraycopy_function(t, aligned, disjoint, name); |
|
4052 |
} |
|
4053 |
||
4054 |
||
4055 |
//------------------------------inline_arraycopy----------------------- |
|
4056 |
bool LibraryCallKit::inline_arraycopy() { |
|
4057 |
// Restore the stack and pop off the arguments. |
|
4058 |
int nargs = 5; // 2 oops, 3 ints, no size_t or long |
|
4059 |
assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); |
|
4060 |
||
4061 |
Node *src = argument(0); |
|
4062 |
Node *src_offset = argument(1); |
|
4063 |
Node *dest = argument(2); |
|
4064 |
Node *dest_offset = argument(3); |
|
4065 |
Node *length = argument(4); |
|
4066 |
||
4067 |
// Compile time checks. If any of these checks cannot be verified at compile time, |
|
4068 |
// we do not make a fast path for this call. Instead, we let the call remain as it |
|
4069 |
// is. The checks we choose to mandate at compile time are: |
|
4070 |
// |
|
4071 |
// (1) src and dest are arrays. |
|
4072 |
const Type* src_type = src->Value(&_gvn); |
|
4073 |
const Type* dest_type = dest->Value(&_gvn); |
|
4074 |
const TypeAryPtr* top_src = src_type->isa_aryptr(); |
|
4075 |
const TypeAryPtr* top_dest = dest_type->isa_aryptr(); |
|
4076 |
if (top_src == NULL || top_src->klass() == NULL || |
|
4077 |
top_dest == NULL || top_dest->klass() == NULL) { |
|
4078 |
// Conservatively insert a memory barrier on all memory slices. |
|
4079 |
// Do not let writes into the source float below the arraycopy. |
|
4080 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
4081 |
||
4082 |
// Call StubRoutines::generic_arraycopy stub. |
|
4083 |
generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, |
|
4084 |
src, src_offset, dest, dest_offset, length, |
|
4085 |
nargs); |
|
4086 |
||
4087 |
// Do not let reads from the destination float above the arraycopy. |
|
4088 |
// Since we cannot type the arrays, we don't know which slices |
|
4089 |
// might be affected. We could restrict this barrier only to those |
|
4090 |
// memory slices which pertain to array elements--but don't bother. |
|
4091 |
if (!InsertMemBarAfterArraycopy) |
|
4092 |
// (If InsertMemBarAfterArraycopy, there is already one in place.) |
|
4093 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
4094 |
return true; |
|
4095 |
} |
|
4096 |
||
4097 |
// (2) src and dest arrays must have elements of the same BasicType |
|
4098 |
// Figure out the size and type of the elements we will be copying. |
|
4099 |
BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); |
|
4100 |
BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); |
|
4101 |
if (src_elem == T_ARRAY) src_elem = T_OBJECT; |
|
4102 |
if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; |
|
4103 |
||
4104 |
if (src_elem != dest_elem || dest_elem == T_VOID) { |
|
4105 |
// The component types are not the same or are not recognized. Punt. |
|
4106 |
// (But, avoid the native method wrapper to JVM_ArrayCopy.) |
|
4107 |
generate_slow_arraycopy(TypePtr::BOTTOM, |
|
4108 |
src, src_offset, dest, dest_offset, length, |
|
4109 |
nargs); |
|
4110 |
return true; |
|
4111 |
} |
|
4112 |
||
4113 |
//--------------------------------------------------------------------------- |
|
4114 |
// We will make a fast path for this call to arraycopy. |
|
4115 |
||
4116 |
// We have the following tests left to perform: |
|
4117 |
// |
|
4118 |
// (3) src and dest must not be null. |
|
4119 |
// (4) src_offset must not be negative. |
|
4120 |
// (5) dest_offset must not be negative. |
|
4121 |
// (6) length must not be negative. |
|
4122 |
// (7) src_offset + length must not exceed length of src. |
|
4123 |
// (8) dest_offset + length must not exceed length of dest. |
|
4124 |
// (9) each element of an oop array must be assignable |
|
4125 |
||
4126 |
RegionNode* slow_region = new (C, 1) RegionNode(1); |
|
4127 |
record_for_igvn(slow_region); |
|
4128 |
||
4129 |
// (3) operands must not be null |
|
4130 |
// We currently perform our null checks with the do_null_check routine. |
|
4131 |
// This means that the null exceptions will be reported in the caller |
|
4132 |
// rather than (correctly) reported inside of the native arraycopy call. |
|
4133 |
// This should be corrected, given time. We do our null check with the |
|
4134 |
// stack pointer restored. |
|
4135 |
_sp += nargs; |
|
4136 |
src = do_null_check(src, T_ARRAY); |
|
4137 |
dest = do_null_check(dest, T_ARRAY); |
|
4138 |
_sp -= nargs; |
|
4139 |
||
4140 |
// (4) src_offset must not be negative. |
|
4141 |
generate_negative_guard(src_offset, slow_region); |
|
4142 |
||
4143 |
// (5) dest_offset must not be negative. |
|
4144 |
generate_negative_guard(dest_offset, slow_region); |
|
4145 |
||
4146 |
// (6) length must not be negative (moved to generate_arraycopy()). |
|
4147 |
// generate_negative_guard(length, slow_region); |
|
4148 |
||
4149 |
// (7) src_offset + length must not exceed length of src. |
|
4150 |
generate_limit_guard(src_offset, length, |
|
4151 |
load_array_length(src), |
|
4152 |
slow_region); |
|
4153 |
||
4154 |
// (8) dest_offset + length must not exceed length of dest. |
|
4155 |
generate_limit_guard(dest_offset, length, |
|
4156 |
load_array_length(dest), |
|
4157 |
slow_region); |
|
4158 |
||
4159 |
// (9) each element of an oop array must be assignable |
|
4160 |
// The generate_arraycopy subroutine checks this. |
|
4161 |
||
4162 |
// This is where the memory effects are placed: |
|
4163 |
const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); |
|
4164 |
generate_arraycopy(adr_type, dest_elem, |
|
4165 |
src, src_offset, dest, dest_offset, length, |
|
4166 |
nargs, false, false, slow_region); |
|
4167 |
||
4168 |
return true; |
|
4169 |
} |
|
4170 |
||
4171 |
//-----------------------------generate_arraycopy---------------------- |
|
4172 |
// Generate an optimized call to arraycopy. |
|
4173 |
// Caller must guard against non-arrays. |
|
4174 |
// Caller must determine a common array basic-type for both arrays. |
|
4175 |
// Caller must validate offsets against array bounds. |
|
4176 |
// The slow_region has already collected guard failure paths |
|
4177 |
// (such as out of bounds length or non-conformable array types). |
|
4178 |
// The generated code has this shape, in general: |
|
4179 |
// |
|
4180 |
// if (length == 0) return // via zero_path |
|
4181 |
// slowval = -1 |
|
4182 |
// if (types unknown) { |
|
4183 |
// slowval = call generic copy loop |
|
4184 |
// if (slowval == 0) return // via checked_path |
|
4185 |
// } else if (indexes in bounds) { |
|
4186 |
// if ((is object array) && !(array type check)) { |
|
4187 |
// slowval = call checked copy loop |
|
4188 |
// if (slowval == 0) return // via checked_path |
|
4189 |
// } else { |
|
4190 |
// call bulk copy loop |
|
4191 |
// return // via fast_path |
|
4192 |
// } |
|
4193 |
// } |
|
4194 |
// // adjust params for remaining work: |
|
4195 |
// if (slowval != -1) { |
|
4196 |
// n = -1^slowval; src_offset += n; dest_offset += n; length -= n |
|
4197 |
// } |
|
4198 |
// slow_region: |
|
4199 |
// call slow arraycopy(src, src_offset, dest, dest_offset, length) |
|
4200 |
// return // via slow_call_path |
|
4201 |
// |
|
4202 |
// This routine is used from several intrinsics: System.arraycopy, |
|
4203 |
// Object.clone (the array subcase), and Arrays.copyOf[Range]. |
|
4204 |
// |
|
4205 |
void |
|
4206 |
LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, |
|
4207 |
BasicType basic_elem_type, |
|
4208 |
Node* src, Node* src_offset, |
|
4209 |
Node* dest, Node* dest_offset, |
|
4210 |
Node* copy_length, |
|
4211 |
int nargs, |
|
4212 |
bool disjoint_bases, |
|
4213 |
bool length_never_negative, |
|
4214 |
RegionNode* slow_region) { |
|
4215 |
||
4216 |
if (slow_region == NULL) { |
|
4217 |
slow_region = new(C,1) RegionNode(1); |
|
4218 |
record_for_igvn(slow_region); |
|
4219 |
} |
|
4220 |
||
4221 |
Node* original_dest = dest; |
|
4222 |
AllocateArrayNode* alloc = NULL; // used for zeroing, if needed |
|
4223 |
Node* raw_dest = NULL; // used before zeroing, if needed |
|
4224 |
bool must_clear_dest = false; |
|
4225 |
||
4226 |
// See if this is the initialization of a newly-allocated array. |
|
4227 |
// If so, we will take responsibility here for initializing it to zero. |
|
4228 |
// (Note: Because tightly_coupled_allocation performs checks on the |
|
4229 |
// out-edges of the dest, we need to avoid making derived pointers |
|
4230 |
// from it until we have checked its uses.) |
|
4231 |
if (ReduceBulkZeroing |
|
4232 |
&& !ZeroTLAB // pointless if already zeroed |
|
4233 |
&& basic_elem_type != T_CONFLICT // avoid corner case |
|
4234 |
&& !_gvn.eqv_uncast(src, dest) |
|
4235 |
&& ((alloc = tightly_coupled_allocation(dest, slow_region)) |
|
4236 |
!= NULL) |
|
207
694076d0755f
6667581: Don't generate initialization (by 0) code for arrays with size 0
kvn
parents:
202
diff
changeset
|
4237 |
&& _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 |
1 | 4238 |
&& alloc->maybe_set_complete(&_gvn)) { |
4239 |
// "You break it, you buy it." |
|
4240 |
InitializeNode* init = alloc->initialization(); |
|
4241 |
assert(init->is_complete(), "we just did this"); |
|
4242 |
assert(dest->Opcode() == Op_CheckCastPP, "sanity"); |
|
4243 |
assert(dest->in(0)->in(0) == init, "dest pinned"); |
|
4244 |
raw_dest = dest->in(1); // grab the raw pointer! |
|
4245 |
original_dest = dest; |
|
4246 |
dest = raw_dest; |
|
4247 |
adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory |
|
4248 |
// Decouple the original InitializeNode, turning it into a simple membar. |
|
4249 |
// We will build a new one at the end of this routine. |
|
4250 |
init->set_req(InitializeNode::RawAddress, top()); |
|
4251 |
// From this point on, every exit path is responsible for |
|
4252 |
// initializing any non-copied parts of the object to zero. |
|
4253 |
must_clear_dest = true; |
|
4254 |
} else { |
|
4255 |
// No zeroing elimination here. |
|
4256 |
alloc = NULL; |
|
4257 |
//original_dest = dest; |
|
4258 |
//must_clear_dest = false; |
|
4259 |
} |
|
4260 |
||
4261 |
// Results are placed here: |
|
4262 |
enum { fast_path = 1, // normal void-returning assembly stub |
|
4263 |
checked_path = 2, // special assembly stub with cleanup |
|
4264 |
slow_call_path = 3, // something went wrong; call the VM |
|
4265 |
zero_path = 4, // bypass when length of copy is zero |
|
4266 |
bcopy_path = 5, // copy primitive array by 64-bit blocks |
|
4267 |
PATH_LIMIT = 6 |
|
4268 |
}; |
|
4269 |
RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
|
4270 |
PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); |
|
4271 |
PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); |
|
4272 |
record_for_igvn(result_region); |
|
4273 |
_gvn.set_type_bottom(result_i_o); |
|
4274 |
_gvn.set_type_bottom(result_memory); |
|
4275 |
assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); |
|
4276 |
||
4277 |
// The slow_control path: |
|
4278 |
Node* slow_control; |
|
4279 |
Node* slow_i_o = i_o(); |
|
4280 |
Node* slow_mem = memory(adr_type); |
|
4281 |
debug_only(slow_control = (Node*) badAddress); |
|
4282 |
||
4283 |
// Checked control path: |
|
4284 |
Node* checked_control = top(); |
|
4285 |
Node* checked_mem = NULL; |
|
4286 |
Node* checked_i_o = NULL; |
|
4287 |
Node* checked_value = NULL; |
|
4288 |
||
4289 |
if (basic_elem_type == T_CONFLICT) { |
|
4290 |
assert(!must_clear_dest, ""); |
|
4291 |
Node* cv = generate_generic_arraycopy(adr_type, |
|
4292 |
src, src_offset, dest, dest_offset, |
|
4293 |
copy_length, nargs); |
|
4294 |
if (cv == NULL) cv = intcon(-1); // failure (no stub available) |
|
4295 |
checked_control = control(); |
|
4296 |
checked_i_o = i_o(); |
|
4297 |
checked_mem = memory(adr_type); |
|
4298 |
checked_value = cv; |
|
4299 |
set_control(top()); // no fast path |
|
4300 |
} |
|
4301 |
||
4302 |
Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); |
|
4303 |
if (not_pos != NULL) { |
|
4304 |
PreserveJVMState pjvms(this); |
|
4305 |
set_control(not_pos); |
|
4306 |
||
4307 |
// (6) length must not be negative. |
|
4308 |
if (!length_never_negative) { |
|
4309 |
generate_negative_guard(copy_length, slow_region); |
|
4310 |
} |
|
4311 |
||
4312 |
if (!stopped() && must_clear_dest) { |
|
4313 |
Node* dest_length = alloc->in(AllocateNode::ALength); |
|
4314 |
if (_gvn.eqv_uncast(copy_length, dest_length) |
|
4315 |
|| _gvn.find_int_con(dest_length, 1) <= 0) { |
|
4316 |
// There is no zeroing to do. |
|
4317 |
} else { |
|
4318 |
// Clear the whole thing since there are no source elements to copy. |
|
4319 |
generate_clear_array(adr_type, dest, basic_elem_type, |
|
4320 |
intcon(0), NULL, |
|
4321 |
alloc->in(AllocateNode::AllocSize)); |
|
4322 |
} |
|
4323 |
} |
|
4324 |
||
4325 |
// Present the results of the fast call. |
|
4326 |
result_region->init_req(zero_path, control()); |
|
4327 |
result_i_o ->init_req(zero_path, i_o()); |
|
4328 |
result_memory->init_req(zero_path, memory(adr_type)); |
|
4329 |
} |
|
4330 |
||
4331 |
if (!stopped() && must_clear_dest) { |
|
4332 |
// We have to initialize the *uncopied* part of the array to zero. |
|
4333 |
// The copy destination is the slice dest[off..off+len]. The other slices |
|
4334 |
// are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. |
|
4335 |
Node* dest_size = alloc->in(AllocateNode::AllocSize); |
|
4336 |
Node* dest_length = alloc->in(AllocateNode::ALength); |
|
4337 |
Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, |
|
4338 |
copy_length) ); |
|
4339 |
||
4340 |
// If there is a head section that needs zeroing, do it now. |
|
4341 |
if (find_int_con(dest_offset, -1) != 0) { |
|
4342 |
generate_clear_array(adr_type, dest, basic_elem_type, |
|
4343 |
intcon(0), dest_offset, |
|
4344 |
NULL); |
|
4345 |
} |
|
4346 |
||
4347 |
// Next, perform a dynamic check on the tail length. |
|
4348 |
// It is often zero, and we can win big if we prove this. |
|
4349 |
// There are two wins: Avoid generating the ClearArray |
|
4350 |
// with its attendant messy index arithmetic, and upgrade |
|
4351 |
// the copy to a more hardware-friendly word size of 64 bits. |
|
4352 |
Node* tail_ctl = NULL; |
|
4353 |
if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) { |
|
4354 |
Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); |
|
4355 |
Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); |
|
4356 |
tail_ctl = generate_slow_guard(bol_lt, NULL); |
|
4357 |
assert(tail_ctl != NULL || !stopped(), "must be an outcome"); |
|
4358 |
} |
|
4359 |
||
4360 |
// At this point, let's assume there is no tail. |
|
4361 |
if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { |
|
4362 |
// There is no tail. Try an upgrade to a 64-bit copy. |
|
4363 |
bool didit = false; |
|
4364 |
{ PreserveJVMState pjvms(this); |
|
4365 |
didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, |
|
4366 |
src, src_offset, dest, dest_offset, |
|
4367 |
dest_size); |
|
4368 |
if (didit) { |
|
4369 |
// Present the results of the block-copying fast call. |
|
4370 |
result_region->init_req(bcopy_path, control()); |
|
4371 |
result_i_o ->init_req(bcopy_path, i_o()); |
|
4372 |
result_memory->init_req(bcopy_path, memory(adr_type)); |
|
4373 |
} |
|
4374 |
} |
|
4375 |
if (didit) |
|
4376 |
set_control(top()); // no regular fast path |
|
4377 |
} |
|
4378 |
||
4379 |
// Clear the tail, if any. |
|
4380 |
if (tail_ctl != NULL) { |
|
4381 |
Node* notail_ctl = stopped() ? NULL : control(); |
|
4382 |
set_control(tail_ctl); |
|
4383 |
if (notail_ctl == NULL) { |
|
4384 |
generate_clear_array(adr_type, dest, basic_elem_type, |
|
4385 |
dest_tail, NULL, |
|
4386 |
dest_size); |
|
4387 |
} else { |
|
4388 |
// Make a local merge. |
|
4389 |
Node* done_ctl = new(C,3) RegionNode(3); |
|
4390 |
Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); |
|
4391 |
done_ctl->init_req(1, notail_ctl); |
|
4392 |
done_mem->init_req(1, memory(adr_type)); |
|
4393 |
generate_clear_array(adr_type, dest, basic_elem_type, |
|
4394 |
dest_tail, NULL, |
|
4395 |
dest_size); |
|
4396 |
done_ctl->init_req(2, control()); |
|
4397 |
done_mem->init_req(2, memory(adr_type)); |
|
4398 |
set_control( _gvn.transform(done_ctl) ); |
|
4399 |
set_memory( _gvn.transform(done_mem), adr_type ); |
|
4400 |
} |
|
4401 |
} |
|
4402 |
} |
|
4403 |
||
4404 |
BasicType copy_type = basic_elem_type; |
|
4405 |
assert(basic_elem_type != T_ARRAY, "caller must fix this"); |
|
4406 |
if (!stopped() && copy_type == T_OBJECT) { |
|
4407 |
// If src and dest have compatible element types, we can copy bits. |
|
4408 |
// Types S[] and D[] are compatible if D is a supertype of S. |
|
4409 |
// |
|
4410 |
// If they are not, we will use checked_oop_disjoint_arraycopy, |
|
4411 |
// which performs a fast optimistic per-oop check, and backs off |
|
4412 |
// further to JVM_ArrayCopy on the first per-oop check that fails. |
|
4413 |
// (Actually, we don't move raw bits only; the GC requires card marks.) |
|
4414 |
||
4415 |
// Get the klassOop for both src and dest |
|
4416 |
Node* src_klass = load_object_klass(src); |
|
4417 |
Node* dest_klass = load_object_klass(dest); |
|
4418 |
||
4419 |
// Generate the subtype check. |
|
4420 |
// This might fold up statically, or then again it might not. |
|
4421 |
// |
|
4422 |
// Non-static example: Copying List<String>.elements to a new String[]. |
|
4423 |
// The backing store for a List<String> is always an Object[], |
|
4424 |
// but its elements are always type String, if the generic types |
|
4425 |
// are correct at the source level. |
|
4426 |
// |
|
4427 |
// Test S[] against D[], not S against D, because (probably) |
|
4428 |
// the secondary supertype cache is less busy for S[] than S. |
|
4429 |
// This usually only matters when D is an interface. |
|
4430 |
Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); |
|
4431 |
// Plug failing path into checked_oop_disjoint_arraycopy |
|
4432 |
if (not_subtype_ctrl != top()) { |
|
4433 |
PreserveJVMState pjvms(this); |
|
4434 |
set_control(not_subtype_ctrl); |
|
4435 |
// (At this point we can assume disjoint_bases, since types differ.) |
|
4436 |
int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc); |
|
4437 |
Node* p1 = basic_plus_adr(dest_klass, ek_offset); |
|
590
2954744d7bba
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
589
diff
changeset
|
4438 |
Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); |
1 | 4439 |
Node* dest_elem_klass = _gvn.transform(n1); |
4440 |
Node* cv = generate_checkcast_arraycopy(adr_type, |
|
4441 |
dest_elem_klass, |
|
4442 |
src, src_offset, dest, dest_offset, |
|
4443 |
copy_length, |
|
4444 |
nargs); |
|
4445 |
if (cv == NULL) cv = intcon(-1); // failure (no stub available) |
|
4446 |
checked_control = control(); |
|
4447 |
checked_i_o = i_o(); |
|
4448 |
checked_mem = memory(adr_type); |
|
4449 |
checked_value = cv; |
|
4450 |
} |
|
4451 |
// At this point we know we do not need type checks on oop stores. |
|
4452 |
||
4453 |
// Let's see if we need card marks: |
|
4454 |
if (alloc != NULL && use_ReduceInitialCardMarks()) { |
|
4455 |
// If we do not need card marks, copy using the jint or jlong stub. |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4456 |
copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); |
202
dc13bf0e5d5d
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
1
diff
changeset
|
4457 |
assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), |
1 | 4458 |
"sizes agree"); |
4459 |
} |
|
4460 |
} |
|
4461 |
||
4462 |
if (!stopped()) { |
|
4463 |
// Generate the fast path, if possible. |
|
4464 |
PreserveJVMState pjvms(this); |
|
4465 |
generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, |
|
4466 |
src, src_offset, dest, dest_offset, |
|
4467 |
ConvI2X(copy_length)); |
|
4468 |
||
4469 |
// Present the results of the fast call. |
|
4470 |
result_region->init_req(fast_path, control()); |
|
4471 |
result_i_o ->init_req(fast_path, i_o()); |
|
4472 |
result_memory->init_req(fast_path, memory(adr_type)); |
|
4473 |
} |
|
4474 |
||
4475 |
// Here are all the slow paths up to this point, in one bundle: |
|
4476 |
slow_control = top(); |
|
4477 |
if (slow_region != NULL) |
|
4478 |
slow_control = _gvn.transform(slow_region); |
|
4479 |
debug_only(slow_region = (RegionNode*)badAddress); |
|
4480 |
||
4481 |
set_control(checked_control); |
|
4482 |
if (!stopped()) { |
|
4483 |
// Clean up after the checked call. |
|
4484 |
// The returned value is either 0 or -1^K, |
|
4485 |
// where K = number of partially transferred array elements. |
|
4486 |
Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); |
|
4487 |
Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); |
|
4488 |
IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); |
|
4489 |
||
4490 |
// If it is 0, we are done, so transfer to the end. |
|
4491 |
Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); |
|
4492 |
result_region->init_req(checked_path, checks_done); |
|
4493 |
result_i_o ->init_req(checked_path, checked_i_o); |
|
4494 |
result_memory->init_req(checked_path, checked_mem); |
|
4495 |
||
4496 |
// If it is not zero, merge into the slow call. |
|
4497 |
set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); |
|
4498 |
RegionNode* slow_reg2 = new(C, 3) RegionNode(3); |
|
4499 |
PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); |
|
4500 |
PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); |
|
4501 |
record_for_igvn(slow_reg2); |
|
4502 |
slow_reg2 ->init_req(1, slow_control); |
|
4503 |
slow_i_o2 ->init_req(1, slow_i_o); |
|
4504 |
slow_mem2 ->init_req(1, slow_mem); |
|
4505 |
slow_reg2 ->init_req(2, control()); |
|
4506 |
slow_i_o2 ->init_req(2, i_o()); |
|
4507 |
slow_mem2 ->init_req(2, memory(adr_type)); |
|
4508 |
||
4509 |
slow_control = _gvn.transform(slow_reg2); |
|
4510 |
slow_i_o = _gvn.transform(slow_i_o2); |
|
4511 |
slow_mem = _gvn.transform(slow_mem2); |
|
4512 |
||
4513 |
if (alloc != NULL) { |
|
4514 |
// We'll restart from the very beginning, after zeroing the whole thing. |
|
4515 |
// This can cause double writes, but that's OK since dest is brand new. |
|
4516 |
// So we ignore the low 31 bits of the value returned from the stub. |
|
4517 |
} else { |
|
4518 |
// We must continue the copy exactly where it failed, or else |
|
4519 |
// another thread might see the wrong number of writes to dest. |
|
4520 |
Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); |
|
4521 |
Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); |
|
4522 |
slow_offset->init_req(1, intcon(0)); |
|
4523 |
slow_offset->init_req(2, checked_offset); |
|
4524 |
slow_offset = _gvn.transform(slow_offset); |
|
4525 |
||
4526 |
// Adjust the arguments by the conditionally incoming offset. |
|
4527 |
Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); |
|
4528 |
Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); |
|
4529 |
Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); |
|
4530 |
||
4531 |
// Tweak the node variables to adjust the code produced below: |
|
4532 |
src_offset = src_off_plus; |
|
4533 |
dest_offset = dest_off_plus; |
|
4534 |
copy_length = length_minus; |
|
4535 |
} |
|
4536 |
} |
|
4537 |
||
4538 |
set_control(slow_control); |
|
4539 |
if (!stopped()) { |
|
4540 |
// Generate the slow path, if needed. |
|
4541 |
PreserveJVMState pjvms(this); // replace_in_map may trash the map |
|
4542 |
||
4543 |
set_memory(slow_mem, adr_type); |
|
4544 |
set_i_o(slow_i_o); |
|
4545 |
||
4546 |
if (must_clear_dest) { |
|
4547 |
generate_clear_array(adr_type, dest, basic_elem_type, |
|
4548 |
intcon(0), NULL, |
|
4549 |
alloc->in(AllocateNode::AllocSize)); |
|
4550 |
} |
|
4551 |
||
4552 |
if (dest != original_dest) { |
|
4553 |
// Promote from rawptr to oop, so it looks right in the call's GC map. |
|
4554 |
dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest, |
|
4555 |
TypeInstPtr::NOTNULL) ); |
|
4556 |
||
4557 |
// Edit the call's debug-info to avoid referring to original_dest. |
|
4558 |
// (The problem with original_dest is that it isn't ready until |
|
4559 |
// after the InitializeNode completes, but this stuff is before.) |
|
4560 |
// Substitute in the locally valid dest_oop. |
|
4561 |
replace_in_map(original_dest, dest); |
|
4562 |
} |
|
4563 |
||
4564 |
generate_slow_arraycopy(adr_type, |
|
4565 |
src, src_offset, dest, dest_offset, |
|
4566 |
copy_length, nargs); |
|
4567 |
||
4568 |
result_region->init_req(slow_call_path, control()); |
|
4569 |
result_i_o ->init_req(slow_call_path, i_o()); |
|
4570 |
result_memory->init_req(slow_call_path, memory(adr_type)); |
|
4571 |
} |
|
4572 |
||
4573 |
// Remove unused edges. |
|
4574 |
for (uint i = 1; i < result_region->req(); i++) { |
|
4575 |
if (result_region->in(i) == NULL) |
|
4576 |
result_region->init_req(i, top()); |
|
4577 |
} |
|
4578 |
||
4579 |
// Finished; return the combined state. |
|
4580 |
set_control( _gvn.transform(result_region) ); |
|
4581 |
set_i_o( _gvn.transform(result_i_o) ); |
|
4582 |
set_memory( _gvn.transform(result_memory), adr_type ); |
|
4583 |
||
4584 |
if (dest != original_dest) { |
|
4585 |
// Pin the "finished" array node after the arraycopy/zeroing operations. |
|
4586 |
// Use a secondary InitializeNode memory barrier. |
|
4587 |
InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, |
|
4588 |
Compile::AliasIdxRaw, |
|
4589 |
raw_dest)->as_Initialize(); |
|
4590 |
init->set_complete(&_gvn); // (there is no corresponding AllocateNode) |
|
4591 |
_gvn.hash_delete(original_dest); |
|
4592 |
original_dest->set_req(0, control()); |
|
4593 |
_gvn.hash_find_insert(original_dest); // put back into GVN table |
|
4594 |
} |
|
4595 |
||
4596 |
// The memory edges above are precise in order to model effects around |
|
4597 |
// array copyies accurately to allow value numbering of field loads around |
|
4598 |
// arraycopy. Such field loads, both before and after, are common in Java |
|
4599 |
// collections and similar classes involving header/array data structures. |
|
4600 |
// |
|
4601 |
// But with low number of register or when some registers are used or killed |
|
4602 |
// by arraycopy calls it causes registers spilling on stack. See 6544710. |
|
4603 |
// The next memory barrier is added to avoid it. If the arraycopy can be |
|
4604 |
// optimized away (which it can, sometimes) then we can manually remove |
|
4605 |
// the membar also. |
|
4606 |
if (InsertMemBarAfterArraycopy) |
|
4607 |
insert_mem_bar(Op_MemBarCPUOrder); |
|
4608 |
} |
|
4609 |
||
4610 |
||
4611 |
// Helper function which determines if an arraycopy immediately follows |
|
4612 |
// an allocation, with no intervening tests or other escapes for the object. |
|
4613 |
AllocateArrayNode* |
|
4614 |
LibraryCallKit::tightly_coupled_allocation(Node* ptr, |
|
4615 |
RegionNode* slow_region) { |
|
4616 |
if (stopped()) return NULL; // no fast path |
|
4617 |
if (C->AliasLevel() == 0) return NULL; // no MergeMems around |
|
4618 |
||
4619 |
AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); |
|
4620 |
if (alloc == NULL) return NULL; |
|
4621 |
||
4622 |
Node* rawmem = memory(Compile::AliasIdxRaw); |
|
4623 |
// Is the allocation's memory state untouched? |
|
4624 |
if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { |
|
4625 |
// Bail out if there have been raw-memory effects since the allocation. |
|
4626 |
// (Example: There might have been a call or safepoint.) |
|
4627 |
return NULL; |
|
4628 |
} |
|
4629 |
rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); |
|
4630 |
if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { |
|
4631 |
return NULL; |
|
4632 |
} |
|
4633 |
||
4634 |
// There must be no unexpected observers of this allocation. |
|
4635 |
for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { |
|
4636 |
Node* obs = ptr->fast_out(i); |
|
4637 |
if (obs != this->map()) { |
|
4638 |
return NULL; |
|
4639 |
} |
|
4640 |
} |
|
4641 |
||
4642 |
// This arraycopy must unconditionally follow the allocation of the ptr. |
|
4643 |
Node* alloc_ctl = ptr->in(0); |
|
4644 |
assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); |
|
4645 |
||
4646 |
Node* ctl = control(); |
|
4647 |
while (ctl != alloc_ctl) { |
|
4648 |
// There may be guards which feed into the slow_region. |
|
4649 |
// Any other control flow means that we might not get a chance |
|
4650 |
// to finish initializing the allocated object. |
|
4651 |
if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { |
|
4652 |
IfNode* iff = ctl->in(0)->as_If(); |
|
4653 |
Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); |
|
4654 |
assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); |
|
4655 |
if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { |
|
4656 |
ctl = iff->in(0); // This test feeds the known slow_region. |
|
4657 |
continue; |
|
4658 |
} |
|
4659 |
// One more try: Various low-level checks bottom out in |
|
4660 |
// uncommon traps. If the debug-info of the trap omits |
|
4661 |
// any reference to the allocation, as we've already |
|
4662 |
// observed, then there can be no objection to the trap. |
|
4663 |
bool found_trap = false; |
|
4664 |
for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { |
|
4665 |
Node* obs = not_ctl->fast_out(j); |
|
4666 |
if (obs->in(0) == not_ctl && obs->is_Call() && |
|
4667 |
(obs->as_Call()->entry_point() == |
|
4668 |
SharedRuntime::uncommon_trap_blob()->instructions_begin())) { |
|
4669 |
found_trap = true; break; |
|
4670 |
} |
|
4671 |
} |
|
4672 |
if (found_trap) { |
|
4673 |
ctl = iff->in(0); // This test feeds a harmless uncommon trap. |
|
4674 |
continue; |
|
4675 |
} |
|
4676 |
} |
|
4677 |
return NULL; |
|
4678 |
} |
|
4679 |
||
4680 |
// If we get this far, we have an allocation which immediately |
|
4681 |
// precedes the arraycopy, and we can take over zeroing the new object. |
|
4682 |
// The arraycopy will finish the initialization, and provide |
|
4683 |
// a new control state to which we will anchor the destination pointer. |
|
4684 |
||
4685 |
return alloc; |
|
4686 |
} |
|
4687 |
||
4688 |
// Helper for initialization of arrays, creating a ClearArray. |
|
4689 |
// It writes zero bits in [start..end), within the body of an array object. |
|
4690 |
// The memory effects are all chained onto the 'adr_type' alias category. |
|
4691 |
// |
|
4692 |
// Since the object is otherwise uninitialized, we are free |
|
4693 |
// to put a little "slop" around the edges of the cleared area, |
|
4694 |
// as long as it does not go back into the array's header, |
|
4695 |
// or beyond the array end within the heap. |
|
4696 |
// |
|
4697 |
// The lower edge can be rounded down to the nearest jint and the |
|
4698 |
// upper edge can be rounded up to the nearest MinObjAlignmentInBytes. |
|
4699 |
// |
|
4700 |
// Arguments: |
|
4701 |
// adr_type memory slice where writes are generated |
|
4702 |
// dest oop of the destination array |
|
4703 |
// basic_elem_type element type of the destination |
|
4704 |
// slice_idx array index of first element to store |
|
4705 |
// slice_len number of elements to store (or NULL) |
|
4706 |
// dest_size total size in bytes of the array object |
|
4707 |
// |
|
4708 |
// Exactly one of slice_len or dest_size must be non-NULL. |
|
4709 |
// If dest_size is non-NULL, zeroing extends to the end of the object. |
|
4710 |
// If slice_len is non-NULL, the slice_idx value must be a constant. |
|
4711 |
void |
|
4712 |
LibraryCallKit::generate_clear_array(const TypePtr* adr_type, |
|
4713 |
Node* dest, |
|
4714 |
BasicType basic_elem_type, |
|
4715 |
Node* slice_idx, |
|
4716 |
Node* slice_len, |
|
4717 |
Node* dest_size) { |
|
4718 |
// one or the other but not both of slice_len and dest_size: |
|
4719 |
assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); |
|
4720 |
if (slice_len == NULL) slice_len = top(); |
|
4721 |
if (dest_size == NULL) dest_size = top(); |
|
4722 |
||
4723 |
// operate on this memory slice: |
|
4724 |
Node* mem = memory(adr_type); // memory slice to operate on |
|
4725 |
||
4726 |
// scaling and rounding of indexes: |
|
202
dc13bf0e5d5d
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
1
diff
changeset
|
4727 |
int scale = exact_log2(type2aelembytes(basic_elem_type)); |
1 | 4728 |
int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); |
4729 |
int clear_low = (-1 << scale) & (BytesPerInt - 1); |
|
4730 |
int bump_bit = (-1 << scale) & BytesPerInt; |
|
4731 |
||
4732 |
// determine constant starts and ends |
|
4733 |
const intptr_t BIG_NEG = -128; |
|
4734 |
assert(BIG_NEG + 2*abase < 0, "neg enough"); |
|
4735 |
intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); |
|
4736 |
intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); |
|
4737 |
if (slice_len_con == 0) { |
|
4738 |
return; // nothing to do here |
|
4739 |
} |
|
4740 |
intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; |
|
4741 |
intptr_t end_con = find_intptr_t_con(dest_size, -1); |
|
4742 |
if (slice_idx_con >= 0 && slice_len_con >= 0) { |
|
4743 |
assert(end_con < 0, "not two cons"); |
|
4744 |
end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), |
|
4745 |
BytesPerLong); |
|
4746 |
} |
|
4747 |
||
4748 |
if (start_con >= 0 && end_con >= 0) { |
|
4749 |
// Constant start and end. Simple. |
|
4750 |
mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4751 |
start_con, end_con, &_gvn); |
|
4752 |
} else if (start_con >= 0 && dest_size != top()) { |
|
4753 |
// Constant start, pre-rounded end after the tail of the array. |
|
4754 |
Node* end = dest_size; |
|
4755 |
mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4756 |
start_con, end, &_gvn); |
|
4757 |
} else if (start_con >= 0 && slice_len != top()) { |
|
4758 |
// Constant start, non-constant end. End needs rounding up. |
|
4759 |
// End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) |
|
4760 |
intptr_t end_base = abase + (slice_idx_con << scale); |
|
4761 |
int end_round = (-1 << scale) & (BytesPerLong - 1); |
|
4762 |
Node* end = ConvI2X(slice_len); |
|
4763 |
if (scale != 0) |
|
4764 |
end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); |
|
4765 |
end_base += end_round; |
|
4766 |
end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); |
|
4767 |
end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); |
|
4768 |
mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4769 |
start_con, end, &_gvn); |
|
4770 |
} else if (start_con < 0 && dest_size != top()) { |
|
4771 |
// Non-constant start, pre-rounded end after the tail of the array. |
|
4772 |
// This is almost certainly a "round-to-end" operation. |
|
4773 |
Node* start = slice_idx; |
|
4774 |
start = ConvI2X(start); |
|
4775 |
if (scale != 0) |
|
4776 |
start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); |
|
4777 |
start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); |
|
4778 |
if ((bump_bit | clear_low) != 0) { |
|
4779 |
int to_clear = (bump_bit | clear_low); |
|
4780 |
// Align up mod 8, then store a jint zero unconditionally |
|
4781 |
// just before the mod-8 boundary. |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4782 |
if (((abase + bump_bit) & ~to_clear) - bump_bit |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4783 |
< arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4784 |
bump_bit = 0; |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4785 |
assert((abase & to_clear) == 0, "array base must be long-aligned"); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4786 |
} else { |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4787 |
// Bump 'start' up to (or past) the next jint boundary: |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4788 |
start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4789 |
assert((abase & clear_low) == 0, "array base must be int-aligned"); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4790 |
} |
1 | 4791 |
// Round bumped 'start' down to jlong boundary in body of array. |
4792 |
start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); |
|
360
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4793 |
if (bump_bit != 0) { |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4794 |
// Store a zero to the immediately preceding jint: |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4795 |
Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4796 |
Node* p1 = basic_plus_adr(dest, x1); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4797 |
mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4798 |
mem = _gvn.transform(mem); |
21d113ecbf6a
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
207
diff
changeset
|
4799 |
} |
1 | 4800 |
} |
4801 |
Node* end = dest_size; // pre-rounded |
|
4802 |
mem = ClearArrayNode::clear_memory(control(), mem, dest, |
|
4803 |
start, end, &_gvn); |
|
4804 |
} else { |
|
4805 |
// Non-constant start, unrounded non-constant end. |
|
4806 |
// (Nobody zeroes a random midsection of an array using this routine.) |
|
4807 |
ShouldNotReachHere(); // fix caller |
|
4808 |
} |
|
4809 |
||
4810 |
// Done. |
|
4811 |
set_memory(mem, adr_type); |
|
4812 |
} |
|
4813 |
||
4814 |
||
4815 |
bool |
|
4816 |
LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, |
|
4817 |
BasicType basic_elem_type, |
|
4818 |
AllocateNode* alloc, |
|
4819 |
Node* src, Node* src_offset, |
|
4820 |
Node* dest, Node* dest_offset, |
|
4821 |
Node* dest_size) { |
|
4822 |
// See if there is an advantage from block transfer. |
|
202
dc13bf0e5d5d
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
1
diff
changeset
|
4823 |
int scale = exact_log2(type2aelembytes(basic_elem_type)); |
1 | 4824 |
if (scale >= LogBytesPerLong) |
4825 |
return false; // it is already a block transfer |
|
4826 |
||
4827 |
// Look at the alignment of the starting offsets. |
|
4828 |
int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); |
|
4829 |
const intptr_t BIG_NEG = -128; |
|
4830 |
assert(BIG_NEG + 2*abase < 0, "neg enough"); |
|
4831 |
||
4832 |
intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale); |
|
4833 |
intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale); |
|
4834 |
if (src_off < 0 || dest_off < 0) |
|
4835 |
// At present, we can only understand constants. |
|
4836 |
return false; |
|
4837 |
||
4838 |
if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { |
|
4839 |
// Non-aligned; too bad. |
|
4840 |
// One more chance: Pick off an initial 32-bit word. |
|
4841 |
// This is a common case, since abase can be odd mod 8. |
|
4842 |
if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && |
|
4843 |
((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { |
|
4844 |
Node* sptr = basic_plus_adr(src, src_off); |
|
4845 |
Node* dptr = basic_plus_adr(dest, dest_off); |
|
4846 |
Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); |
|
4847 |
store_to_memory(control(), dptr, sval, T_INT, adr_type); |
|
4848 |
src_off += BytesPerInt; |
|
4849 |
dest_off += BytesPerInt; |
|
4850 |
} else { |
|
4851 |
return false; |
|
4852 |
} |
|
4853 |
} |
|
4854 |
assert(src_off % BytesPerLong == 0, ""); |
|
4855 |
assert(dest_off % BytesPerLong == 0, ""); |
|
4856 |
||
4857 |
// Do this copy by giant steps. |
|
4858 |
Node* sptr = basic_plus_adr(src, src_off); |
|
4859 |
Node* dptr = basic_plus_adr(dest, dest_off); |
|
4860 |
Node* countx = dest_size; |
|
4861 |
countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); |
|
4862 |
countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); |
|
4863 |
||
4864 |
bool disjoint_bases = true; // since alloc != NULL |
|
4865 |
generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, |
|
4866 |
sptr, NULL, dptr, NULL, countx); |
|
4867 |
||
4868 |
return true; |
|
4869 |
} |
|
4870 |
||
4871 |
||
4872 |
// Helper function; generates code for the slow case. |
|
4873 |
// We make a call to a runtime method which emulates the native method, |
|
4874 |
// but without the native wrapper overhead. |
|
4875 |
void |
|
4876 |
LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, |
|
4877 |
Node* src, Node* src_offset, |
|
4878 |
Node* dest, Node* dest_offset, |
|
4879 |
Node* copy_length, |
|
4880 |
int nargs) { |
|
4881 |
_sp += nargs; // any deopt will start just before call to enclosing method |
|
4882 |
Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, |
|
4883 |
OptoRuntime::slow_arraycopy_Type(), |
|
4884 |
OptoRuntime::slow_arraycopy_Java(), |
|
4885 |
"slow_arraycopy", adr_type, |
|
4886 |
src, src_offset, dest, dest_offset, |
|
4887 |
copy_length); |
|
4888 |
_sp -= nargs; |
|
4889 |
||
4890 |
// Handle exceptions thrown by this fellow: |
|
4891 |
make_slow_call_ex(call, env()->Throwable_klass(), false); |
|
4892 |
} |
|
4893 |
||
4894 |
// Helper function; generates code for cases requiring runtime checks. |
|
4895 |
Node* |
|
4896 |
LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, |
|
4897 |
Node* dest_elem_klass, |
|
4898 |
Node* src, Node* src_offset, |
|
4899 |
Node* dest, Node* dest_offset, |
|
4900 |
Node* copy_length, |
|
4901 |
int nargs) { |
|
4902 |
if (stopped()) return NULL; |
|
4903 |
||
4904 |
address copyfunc_addr = StubRoutines::checkcast_arraycopy(); |
|
4905 |
if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. |
|
4906 |
return NULL; |
|
4907 |
} |
|
4908 |
||
4909 |
// Pick out the parameters required to perform a store-check |
|
4910 |
// for the target array. This is an optimistic check. It will |
|
4911 |
// look in each non-null element's class, at the desired klass's |
|
4912 |
// super_check_offset, for the desired klass. |
|
4913 |
int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc); |
|
4914 |
Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); |
|
4915 |
Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM); |
|
4916 |
Node* check_offset = _gvn.transform(n3); |
|
4917 |
Node* check_value = dest_elem_klass; |
|
4918 |
||
4919 |
Node* src_start = array_element_address(src, src_offset, T_OBJECT); |
|
4920 |
Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); |
|
4921 |
||
4922 |
// (We know the arrays are never conjoint, because their types differ.) |
|
4923 |
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
|
4924 |
OptoRuntime::checkcast_arraycopy_Type(), |
|
4925 |
copyfunc_addr, "checkcast_arraycopy", adr_type, |
|
4926 |
// five arguments, of which two are |
|
4927 |
// intptr_t (jlong in LP64) |
|
4928 |
src_start, dest_start, |
|
4929 |
copy_length XTOP, |
|
4930 |
check_offset XTOP, |
|
4931 |
check_value); |
|
4932 |
||
4933 |
return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); |
|
4934 |
} |
|
4935 |
||
4936 |
||
4937 |
// Helper function; generates code for cases requiring runtime checks. |
|
4938 |
Node* |
|
4939 |
LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, |
|
4940 |
Node* src, Node* src_offset, |
|
4941 |
Node* dest, Node* dest_offset, |
|
4942 |
Node* copy_length, |
|
4943 |
int nargs) { |
|
4944 |
if (stopped()) return NULL; |
|
4945 |
||
4946 |
address copyfunc_addr = StubRoutines::generic_arraycopy(); |
|
4947 |
if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. |
|
4948 |
return NULL; |
|
4949 |
} |
|
4950 |
||
4951 |
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
|
4952 |
OptoRuntime::generic_arraycopy_Type(), |
|
4953 |
copyfunc_addr, "generic_arraycopy", adr_type, |
|
4954 |
src, src_offset, dest, dest_offset, copy_length); |
|
4955 |
||
4956 |
return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); |
|
4957 |
} |
|
4958 |
||
4959 |
// Helper function; generates the fast out-of-line call to an arraycopy stub. |
|
4960 |
void |
|
4961 |
LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, |
|
4962 |
BasicType basic_elem_type, |
|
4963 |
bool disjoint_bases, |
|
4964 |
Node* src, Node* src_offset, |
|
4965 |
Node* dest, Node* dest_offset, |
|
4966 |
Node* copy_length) { |
|
4967 |
if (stopped()) return; // nothing to do |
|
4968 |
||
4969 |
Node* src_start = src; |
|
4970 |
Node* dest_start = dest; |
|
4971 |
if (src_offset != NULL || dest_offset != NULL) { |
|
4972 |
assert(src_offset != NULL && dest_offset != NULL, ""); |
|
4973 |
src_start = array_element_address(src, src_offset, basic_elem_type); |
|
4974 |
dest_start = array_element_address(dest, dest_offset, basic_elem_type); |
|
4975 |
} |
|
4976 |
||
4977 |
// Figure out which arraycopy runtime method to call. |
|
4978 |
const char* copyfunc_name = "arraycopy"; |
|
4979 |
address copyfunc_addr = |
|
4980 |
basictype2arraycopy(basic_elem_type, src_offset, dest_offset, |
|
4981 |
disjoint_bases, copyfunc_name); |
|
4982 |
||
4983 |
// Call it. Note that the count_ix value is not scaled to a byte-size. |
|
4984 |
make_runtime_call(RC_LEAF|RC_NO_FP, |
|
4985 |
OptoRuntime::fast_arraycopy_Type(), |
|
4986 |
copyfunc_addr, copyfunc_name, adr_type, |
|
4987 |
src_start, dest_start, copy_length XTOP); |
|
4988 |
} |