Tue, 11 May 2010 15:19:19 -0700
Merge
1 /*
2 * Copyright 1999-2010 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_library_call.cpp.incl"
28 class LibraryIntrinsic : public InlineCallGenerator {
29 // Extend the set of intrinsics known to the runtime:
30 public:
31 private:
32 bool _is_virtual;
33 vmIntrinsics::ID _intrinsic_id;
35 public:
36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
37 : InlineCallGenerator(m),
38 _is_virtual(is_virtual),
39 _intrinsic_id(id)
40 {
41 }
42 virtual bool is_intrinsic() const { return true; }
43 virtual bool is_virtual() const { return _is_virtual; }
44 virtual JVMState* generate(JVMState* jvms);
45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
46 };
49 // Local helper class for LibraryIntrinsic:
50 class LibraryCallKit : public GraphKit {
51 private:
52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
54 public:
55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
56 : GraphKit(caller),
57 _intrinsic(intrinsic)
58 {
59 }
61 ciMethod* caller() const { return jvms()->method(); }
62 int bci() const { return jvms()->bci(); }
63 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
65 ciMethod* callee() const { return _intrinsic->method(); }
66 ciSignature* signature() const { return callee()->signature(); }
67 int arg_size() const { return callee()->arg_size(); }
69 bool try_to_inline();
71 // Helper functions to inline natives
72 void push_result(RegionNode* region, PhiNode* value);
73 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
74 Node* generate_slow_guard(Node* test, RegionNode* region);
75 Node* generate_fair_guard(Node* test, RegionNode* region);
76 Node* generate_negative_guard(Node* index, RegionNode* region,
77 // resulting CastII of index:
78 Node* *pos_index = NULL);
79 Node* generate_nonpositive_guard(Node* index, bool never_negative,
80 // resulting CastII of index:
81 Node* *pos_index = NULL);
82 Node* generate_limit_guard(Node* offset, Node* subseq_length,
83 Node* array_length,
84 RegionNode* region);
85 Node* generate_current_thread(Node* &tls_output);
86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
87 bool disjoint_bases, const char* &name);
88 Node* load_mirror_from_klass(Node* klass);
89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
90 int nargs,
91 RegionNode* region, int null_path,
92 int offset);
93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
94 RegionNode* region, int null_path) {
95 int offset = java_lang_Class::klass_offset_in_bytes();
96 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
97 region, null_path,
98 offset);
99 }
100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
101 int nargs,
102 RegionNode* region, int null_path) {
103 int offset = java_lang_Class::array_klass_offset_in_bytes();
104 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
105 region, null_path,
106 offset);
107 }
108 Node* generate_access_flags_guard(Node* kls,
109 int modifier_mask, int modifier_bits,
110 RegionNode* region);
111 Node* generate_interface_guard(Node* kls, RegionNode* region);
112 Node* generate_array_guard(Node* kls, RegionNode* region) {
113 return generate_array_guard_common(kls, region, false, false);
114 }
115 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
116 return generate_array_guard_common(kls, region, false, true);
117 }
118 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
119 return generate_array_guard_common(kls, region, true, false);
120 }
121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
122 return generate_array_guard_common(kls, region, true, true);
123 }
124 Node* generate_array_guard_common(Node* kls, RegionNode* region,
125 bool obj_array, bool not_array);
126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
128 bool is_virtual = false, bool is_static = false);
129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
130 return generate_method_call(method_id, false, true);
131 }
132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
133 return generate_method_call(method_id, true, false);
134 }
136 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
137 bool inline_string_compareTo();
138 bool inline_string_indexOf();
139 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
140 bool inline_string_equals();
141 Node* pop_math_arg();
142 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
143 bool inline_math_native(vmIntrinsics::ID id);
144 bool inline_trig(vmIntrinsics::ID id);
145 bool inline_trans(vmIntrinsics::ID id);
146 bool inline_abs(vmIntrinsics::ID id);
147 bool inline_sqrt(vmIntrinsics::ID id);
148 bool inline_pow(vmIntrinsics::ID id);
149 bool inline_exp(vmIntrinsics::ID id);
150 bool inline_min_max(vmIntrinsics::ID id);
151 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
152 // This returns Type::AnyPtr, RawPtr, or OopPtr.
153 int classify_unsafe_addr(Node* &base, Node* &offset);
154 Node* make_unsafe_address(Node* base, Node* offset);
155 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
156 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
157 bool inline_unsafe_allocate();
158 bool inline_unsafe_copyMemory();
159 bool inline_native_currentThread();
160 bool inline_native_time_funcs(bool isNano);
161 bool inline_native_isInterrupted();
162 bool inline_native_Class_query(vmIntrinsics::ID id);
163 bool inline_native_subtype_check();
165 bool inline_native_newArray();
166 bool inline_native_getLength();
167 bool inline_array_copyOf(bool is_copyOfRange);
168 bool inline_array_equals();
169 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
170 bool inline_native_clone(bool is_virtual);
171 bool inline_native_Reflection_getCallerClass();
172 bool inline_native_AtomicLong_get();
173 bool inline_native_AtomicLong_attemptUpdate();
174 bool is_method_invoke_or_aux_frame(JVMState* jvms);
175 // Helper function for inlining native object hash method
176 bool inline_native_hashcode(bool is_virtual, bool is_static);
177 bool inline_native_getClass();
179 // Helper functions for inlining arraycopy
180 bool inline_arraycopy();
181 void generate_arraycopy(const TypePtr* adr_type,
182 BasicType basic_elem_type,
183 Node* src, Node* src_offset,
184 Node* dest, Node* dest_offset,
185 Node* copy_length,
186 bool disjoint_bases = false,
187 bool length_never_negative = false,
188 RegionNode* slow_region = NULL);
189 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
190 RegionNode* slow_region);
191 void generate_clear_array(const TypePtr* adr_type,
192 Node* dest,
193 BasicType basic_elem_type,
194 Node* slice_off,
195 Node* slice_len,
196 Node* slice_end);
197 bool generate_block_arraycopy(const TypePtr* adr_type,
198 BasicType basic_elem_type,
199 AllocateNode* alloc,
200 Node* src, Node* src_offset,
201 Node* dest, Node* dest_offset,
202 Node* dest_size);
203 void generate_slow_arraycopy(const TypePtr* adr_type,
204 Node* src, Node* src_offset,
205 Node* dest, Node* dest_offset,
206 Node* copy_length);
207 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
208 Node* dest_elem_klass,
209 Node* src, Node* src_offset,
210 Node* dest, Node* dest_offset,
211 Node* copy_length);
212 Node* generate_generic_arraycopy(const TypePtr* adr_type,
213 Node* src, Node* src_offset,
214 Node* dest, Node* dest_offset,
215 Node* copy_length);
216 void generate_unchecked_arraycopy(const TypePtr* adr_type,
217 BasicType basic_elem_type,
218 bool disjoint_bases,
219 Node* src, Node* src_offset,
220 Node* dest, Node* dest_offset,
221 Node* copy_length);
222 bool inline_unsafe_CAS(BasicType type);
223 bool inline_unsafe_ordered_store(BasicType type);
224 bool inline_fp_conversions(vmIntrinsics::ID id);
225 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
226 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
227 bool inline_bitCount(vmIntrinsics::ID id);
228 bool inline_reverseBytes(vmIntrinsics::ID id);
229 };
232 //---------------------------make_vm_intrinsic----------------------------
233 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
234 vmIntrinsics::ID id = m->intrinsic_id();
235 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
237 if (DisableIntrinsic[0] != '\0'
238 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
239 // disabled by a user request on the command line:
240 // example: -XX:DisableIntrinsic=_hashCode,_getClass
241 return NULL;
242 }
244 if (!m->is_loaded()) {
245 // do not attempt to inline unloaded methods
246 return NULL;
247 }
249 // Only a few intrinsics implement a virtual dispatch.
250 // They are expensive calls which are also frequently overridden.
251 if (is_virtual) {
252 switch (id) {
253 case vmIntrinsics::_hashCode:
254 case vmIntrinsics::_clone:
255 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
256 break;
257 default:
258 return NULL;
259 }
260 }
262 // -XX:-InlineNatives disables nearly all intrinsics:
263 if (!InlineNatives) {
264 switch (id) {
265 case vmIntrinsics::_indexOf:
266 case vmIntrinsics::_compareTo:
267 case vmIntrinsics::_equals:
268 case vmIntrinsics::_equalsC:
269 break; // InlineNatives does not control String.compareTo
270 default:
271 return NULL;
272 }
273 }
275 switch (id) {
276 case vmIntrinsics::_compareTo:
277 if (!SpecialStringCompareTo) return NULL;
278 break;
279 case vmIntrinsics::_indexOf:
280 if (!SpecialStringIndexOf) return NULL;
281 break;
282 case vmIntrinsics::_equals:
283 if (!SpecialStringEquals) return NULL;
284 break;
285 case vmIntrinsics::_equalsC:
286 if (!SpecialArraysEquals) return NULL;
287 break;
288 case vmIntrinsics::_arraycopy:
289 if (!InlineArrayCopy) return NULL;
290 break;
291 case vmIntrinsics::_copyMemory:
292 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
293 if (!InlineArrayCopy) return NULL;
294 break;
295 case vmIntrinsics::_hashCode:
296 if (!InlineObjectHash) return NULL;
297 break;
298 case vmIntrinsics::_clone:
299 case vmIntrinsics::_copyOf:
300 case vmIntrinsics::_copyOfRange:
301 if (!InlineObjectCopy) return NULL;
302 // These also use the arraycopy intrinsic mechanism:
303 if (!InlineArrayCopy) return NULL;
304 break;
305 case vmIntrinsics::_checkIndex:
306 // We do not intrinsify this. The optimizer does fine with it.
307 return NULL;
309 case vmIntrinsics::_get_AtomicLong:
310 case vmIntrinsics::_attemptUpdate:
311 if (!InlineAtomicLong) return NULL;
312 break;
314 case vmIntrinsics::_getCallerClass:
315 if (!UseNewReflection) return NULL;
316 if (!InlineReflectionGetCallerClass) return NULL;
317 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
318 break;
320 case vmIntrinsics::_bitCount_i:
321 case vmIntrinsics::_bitCount_l:
322 if (!UsePopCountInstruction) return NULL;
323 break;
325 default:
326 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
327 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
328 break;
329 }
331 // -XX:-InlineClassNatives disables natives from the Class class.
332 // The flag applies to all reflective calls, notably Array.newArray
333 // (visible to Java programmers as Array.newInstance).
334 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
335 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
336 if (!InlineClassNatives) return NULL;
337 }
339 // -XX:-InlineThreadNatives disables natives from the Thread class.
340 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
341 if (!InlineThreadNatives) return NULL;
342 }
344 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
345 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
346 m->holder()->name() == ciSymbol::java_lang_Float() ||
347 m->holder()->name() == ciSymbol::java_lang_Double()) {
348 if (!InlineMathNatives) return NULL;
349 }
351 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
352 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
353 if (!InlineUnsafeOps) return NULL;
354 }
356 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
357 }
359 //----------------------register_library_intrinsics-----------------------
360 // Initialize this file's data structures, for each Compile instance.
361 void Compile::register_library_intrinsics() {
362 // Nothing to do here.
363 }
365 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
366 LibraryCallKit kit(jvms, this);
367 Compile* C = kit.C;
368 int nodes = C->unique();
369 #ifndef PRODUCT
370 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
371 char buf[1000];
372 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
373 tty->print_cr("Intrinsic %s", str);
374 }
375 #endif
376 if (kit.try_to_inline()) {
377 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
378 tty->print("Inlining intrinsic %s%s at bci:%d in",
379 vmIntrinsics::name_at(intrinsic_id()),
380 (is_virtual() ? " (virtual)" : ""), kit.bci());
381 kit.caller()->print_short_name(tty);
382 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
383 }
384 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
385 if (C->log()) {
386 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
387 vmIntrinsics::name_at(intrinsic_id()),
388 (is_virtual() ? " virtual='1'" : ""),
389 C->unique() - nodes);
390 }
391 return kit.transfer_exceptions_into_jvms();
392 }
394 if (PrintIntrinsics) {
395 tty->print("Did not inline intrinsic %s%s at bci:%d in",
396 vmIntrinsics::name_at(intrinsic_id()),
397 (is_virtual() ? " (virtual)" : ""), kit.bci());
398 kit.caller()->print_short_name(tty);
399 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
400 }
401 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
402 return NULL;
403 }
405 bool LibraryCallKit::try_to_inline() {
406 // Handle symbolic names for otherwise undistinguished boolean switches:
407 const bool is_store = true;
408 const bool is_native_ptr = true;
409 const bool is_static = true;
411 switch (intrinsic_id()) {
412 case vmIntrinsics::_hashCode:
413 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
414 case vmIntrinsics::_identityHashCode:
415 return inline_native_hashcode(/*!virtual*/ false, is_static);
416 case vmIntrinsics::_getClass:
417 return inline_native_getClass();
419 case vmIntrinsics::_dsin:
420 case vmIntrinsics::_dcos:
421 case vmIntrinsics::_dtan:
422 case vmIntrinsics::_dabs:
423 case vmIntrinsics::_datan2:
424 case vmIntrinsics::_dsqrt:
425 case vmIntrinsics::_dexp:
426 case vmIntrinsics::_dlog:
427 case vmIntrinsics::_dlog10:
428 case vmIntrinsics::_dpow:
429 return inline_math_native(intrinsic_id());
431 case vmIntrinsics::_min:
432 case vmIntrinsics::_max:
433 return inline_min_max(intrinsic_id());
435 case vmIntrinsics::_arraycopy:
436 return inline_arraycopy();
438 case vmIntrinsics::_compareTo:
439 return inline_string_compareTo();
440 case vmIntrinsics::_indexOf:
441 return inline_string_indexOf();
442 case vmIntrinsics::_equals:
443 return inline_string_equals();
445 case vmIntrinsics::_getObject:
446 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
447 case vmIntrinsics::_getBoolean:
448 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
449 case vmIntrinsics::_getByte:
450 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
451 case vmIntrinsics::_getShort:
452 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
453 case vmIntrinsics::_getChar:
454 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
455 case vmIntrinsics::_getInt:
456 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
457 case vmIntrinsics::_getLong:
458 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
459 case vmIntrinsics::_getFloat:
460 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
461 case vmIntrinsics::_getDouble:
462 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
464 case vmIntrinsics::_putObject:
465 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
466 case vmIntrinsics::_putBoolean:
467 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
468 case vmIntrinsics::_putByte:
469 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
470 case vmIntrinsics::_putShort:
471 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
472 case vmIntrinsics::_putChar:
473 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
474 case vmIntrinsics::_putInt:
475 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
476 case vmIntrinsics::_putLong:
477 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
478 case vmIntrinsics::_putFloat:
479 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
480 case vmIntrinsics::_putDouble:
481 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
483 case vmIntrinsics::_getByte_raw:
484 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
485 case vmIntrinsics::_getShort_raw:
486 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
487 case vmIntrinsics::_getChar_raw:
488 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
489 case vmIntrinsics::_getInt_raw:
490 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
491 case vmIntrinsics::_getLong_raw:
492 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
493 case vmIntrinsics::_getFloat_raw:
494 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
495 case vmIntrinsics::_getDouble_raw:
496 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
497 case vmIntrinsics::_getAddress_raw:
498 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
500 case vmIntrinsics::_putByte_raw:
501 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
502 case vmIntrinsics::_putShort_raw:
503 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
504 case vmIntrinsics::_putChar_raw:
505 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
506 case vmIntrinsics::_putInt_raw:
507 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
508 case vmIntrinsics::_putLong_raw:
509 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
510 case vmIntrinsics::_putFloat_raw:
511 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
512 case vmIntrinsics::_putDouble_raw:
513 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
514 case vmIntrinsics::_putAddress_raw:
515 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
517 case vmIntrinsics::_getObjectVolatile:
518 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
519 case vmIntrinsics::_getBooleanVolatile:
520 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
521 case vmIntrinsics::_getByteVolatile:
522 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
523 case vmIntrinsics::_getShortVolatile:
524 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
525 case vmIntrinsics::_getCharVolatile:
526 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
527 case vmIntrinsics::_getIntVolatile:
528 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
529 case vmIntrinsics::_getLongVolatile:
530 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
531 case vmIntrinsics::_getFloatVolatile:
532 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
533 case vmIntrinsics::_getDoubleVolatile:
534 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
536 case vmIntrinsics::_putObjectVolatile:
537 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
538 case vmIntrinsics::_putBooleanVolatile:
539 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
540 case vmIntrinsics::_putByteVolatile:
541 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
542 case vmIntrinsics::_putShortVolatile:
543 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
544 case vmIntrinsics::_putCharVolatile:
545 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
546 case vmIntrinsics::_putIntVolatile:
547 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
548 case vmIntrinsics::_putLongVolatile:
549 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
550 case vmIntrinsics::_putFloatVolatile:
551 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
552 case vmIntrinsics::_putDoubleVolatile:
553 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
555 case vmIntrinsics::_prefetchRead:
556 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
557 case vmIntrinsics::_prefetchWrite:
558 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
559 case vmIntrinsics::_prefetchReadStatic:
560 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
561 case vmIntrinsics::_prefetchWriteStatic:
562 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
564 case vmIntrinsics::_compareAndSwapObject:
565 return inline_unsafe_CAS(T_OBJECT);
566 case vmIntrinsics::_compareAndSwapInt:
567 return inline_unsafe_CAS(T_INT);
568 case vmIntrinsics::_compareAndSwapLong:
569 return inline_unsafe_CAS(T_LONG);
571 case vmIntrinsics::_putOrderedObject:
572 return inline_unsafe_ordered_store(T_OBJECT);
573 case vmIntrinsics::_putOrderedInt:
574 return inline_unsafe_ordered_store(T_INT);
575 case vmIntrinsics::_putOrderedLong:
576 return inline_unsafe_ordered_store(T_LONG);
578 case vmIntrinsics::_currentThread:
579 return inline_native_currentThread();
580 case vmIntrinsics::_isInterrupted:
581 return inline_native_isInterrupted();
583 case vmIntrinsics::_currentTimeMillis:
584 return inline_native_time_funcs(false);
585 case vmIntrinsics::_nanoTime:
586 return inline_native_time_funcs(true);
587 case vmIntrinsics::_allocateInstance:
588 return inline_unsafe_allocate();
589 case vmIntrinsics::_copyMemory:
590 return inline_unsafe_copyMemory();
591 case vmIntrinsics::_newArray:
592 return inline_native_newArray();
593 case vmIntrinsics::_getLength:
594 return inline_native_getLength();
595 case vmIntrinsics::_copyOf:
596 return inline_array_copyOf(false);
597 case vmIntrinsics::_copyOfRange:
598 return inline_array_copyOf(true);
599 case vmIntrinsics::_equalsC:
600 return inline_array_equals();
601 case vmIntrinsics::_clone:
602 return inline_native_clone(intrinsic()->is_virtual());
604 case vmIntrinsics::_isAssignableFrom:
605 return inline_native_subtype_check();
607 case vmIntrinsics::_isInstance:
608 case vmIntrinsics::_getModifiers:
609 case vmIntrinsics::_isInterface:
610 case vmIntrinsics::_isArray:
611 case vmIntrinsics::_isPrimitive:
612 case vmIntrinsics::_getSuperclass:
613 case vmIntrinsics::_getComponentType:
614 case vmIntrinsics::_getClassAccessFlags:
615 return inline_native_Class_query(intrinsic_id());
617 case vmIntrinsics::_floatToRawIntBits:
618 case vmIntrinsics::_floatToIntBits:
619 case vmIntrinsics::_intBitsToFloat:
620 case vmIntrinsics::_doubleToRawLongBits:
621 case vmIntrinsics::_doubleToLongBits:
622 case vmIntrinsics::_longBitsToDouble:
623 return inline_fp_conversions(intrinsic_id());
625 case vmIntrinsics::_numberOfLeadingZeros_i:
626 case vmIntrinsics::_numberOfLeadingZeros_l:
627 return inline_numberOfLeadingZeros(intrinsic_id());
629 case vmIntrinsics::_numberOfTrailingZeros_i:
630 case vmIntrinsics::_numberOfTrailingZeros_l:
631 return inline_numberOfTrailingZeros(intrinsic_id());
633 case vmIntrinsics::_bitCount_i:
634 case vmIntrinsics::_bitCount_l:
635 return inline_bitCount(intrinsic_id());
637 case vmIntrinsics::_reverseBytes_i:
638 case vmIntrinsics::_reverseBytes_l:
639 case vmIntrinsics::_reverseBytes_s:
640 case vmIntrinsics::_reverseBytes_c:
641 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
643 case vmIntrinsics::_get_AtomicLong:
644 return inline_native_AtomicLong_get();
645 case vmIntrinsics::_attemptUpdate:
646 return inline_native_AtomicLong_attemptUpdate();
648 case vmIntrinsics::_getCallerClass:
649 return inline_native_Reflection_getCallerClass();
651 default:
652 // If you get here, it may be that someone has added a new intrinsic
653 // to the list in vmSymbols.hpp without implementing it here.
654 #ifndef PRODUCT
655 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
656 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
657 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
658 }
659 #endif
660 return false;
661 }
662 }
664 //------------------------------push_result------------------------------
665 // Helper function for finishing intrinsics.
666 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
667 record_for_igvn(region);
668 set_control(_gvn.transform(region));
669 BasicType value_type = value->type()->basic_type();
670 push_node(value_type, _gvn.transform(value));
671 }
673 //------------------------------generate_guard---------------------------
674 // Helper function for generating guarded fast-slow graph structures.
675 // The given 'test', if true, guards a slow path. If the test fails
676 // then a fast path can be taken. (We generally hope it fails.)
677 // In all cases, GraphKit::control() is updated to the fast path.
678 // The returned value represents the control for the slow path.
679 // The return value is never 'top'; it is either a valid control
680 // or NULL if it is obvious that the slow path can never be taken.
681 // Also, if region and the slow control are not NULL, the slow edge
682 // is appended to the region.
683 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
684 if (stopped()) {
685 // Already short circuited.
686 return NULL;
687 }
689 // Build an if node and its projections.
690 // If test is true we take the slow path, which we assume is uncommon.
691 if (_gvn.type(test) == TypeInt::ZERO) {
692 // The slow branch is never taken. No need to build this guard.
693 return NULL;
694 }
696 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
698 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
699 if (if_slow == top()) {
700 // The slow branch is never taken. No need to build this guard.
701 return NULL;
702 }
704 if (region != NULL)
705 region->add_req(if_slow);
707 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
708 set_control(if_fast);
710 return if_slow;
711 }
713 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
714 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
715 }
716 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
717 return generate_guard(test, region, PROB_FAIR);
718 }
720 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
721 Node* *pos_index) {
722 if (stopped())
723 return NULL; // already stopped
724 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
725 return NULL; // index is already adequately typed
726 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
727 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
728 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
729 if (is_neg != NULL && pos_index != NULL) {
730 // Emulate effect of Parse::adjust_map_after_if.
731 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
732 ccast->set_req(0, control());
733 (*pos_index) = _gvn.transform(ccast);
734 }
735 return is_neg;
736 }
738 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
739 Node* *pos_index) {
740 if (stopped())
741 return NULL; // already stopped
742 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
743 return NULL; // index is already adequately typed
744 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
745 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
746 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
747 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
748 if (is_notp != NULL && pos_index != NULL) {
749 // Emulate effect of Parse::adjust_map_after_if.
750 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
751 ccast->set_req(0, control());
752 (*pos_index) = _gvn.transform(ccast);
753 }
754 return is_notp;
755 }
757 // Make sure that 'position' is a valid limit index, in [0..length].
758 // There are two equivalent plans for checking this:
759 // A. (offset + copyLength) unsigned<= arrayLength
760 // B. offset <= (arrayLength - copyLength)
761 // We require that all of the values above, except for the sum and
762 // difference, are already known to be non-negative.
763 // Plan A is robust in the face of overflow, if offset and copyLength
764 // are both hugely positive.
765 //
766 // Plan B is less direct and intuitive, but it does not overflow at
767 // all, since the difference of two non-negatives is always
768 // representable. Whenever Java methods must perform the equivalent
769 // check they generally use Plan B instead of Plan A.
770 // For the moment we use Plan A.
771 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
772 Node* subseq_length,
773 Node* array_length,
774 RegionNode* region) {
775 if (stopped())
776 return NULL; // already stopped
777 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
778 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
779 return NULL; // common case of whole-array copy
780 Node* last = subseq_length;
781 if (!zero_offset) // last += offset
782 last = _gvn.transform( new (C, 3) AddINode(last, offset));
783 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
784 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
785 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
786 return is_over;
787 }
790 //--------------------------generate_current_thread--------------------
791 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
792 ciKlass* thread_klass = env()->Thread_klass();
793 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
794 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
795 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
796 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
797 tls_output = thread;
798 return threadObj;
799 }
802 //------------------------------make_string_method_node------------------------
803 // Helper method for String intrinsic finctions.
804 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
805 const int value_offset = java_lang_String::value_offset_in_bytes();
806 const int count_offset = java_lang_String::count_offset_in_bytes();
807 const int offset_offset = java_lang_String::offset_offset_in_bytes();
809 Node* no_ctrl = NULL;
811 ciInstanceKlass* klass = env()->String_klass();
812 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
814 const TypeAryPtr* value_type =
815 TypeAryPtr::make(TypePtr::NotNull,
816 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
817 ciTypeArrayKlass::make(T_CHAR), true, 0);
819 // Get start addr of string and substring
820 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
821 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
822 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
823 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
824 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
826 // Pin loads from String::equals() argument since it could be NULL.
827 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
828 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
829 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
830 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
831 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
832 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
834 Node* result = NULL;
835 switch (opcode) {
836 case Op_StrIndexOf:
837 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
838 str1_start, cnt1, str2_start, cnt2);
839 break;
840 case Op_StrComp:
841 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
842 str1_start, cnt1, str2_start, cnt2);
843 break;
844 case Op_StrEquals:
845 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
846 str1_start, str2_start, cnt1);
847 break;
848 default:
849 ShouldNotReachHere();
850 return NULL;
851 }
853 // All these intrinsics have checks.
854 C->set_has_split_ifs(true); // Has chance for split-if optimization
856 return _gvn.transform(result);
857 }
859 //------------------------------inline_string_compareTo------------------------
860 bool LibraryCallKit::inline_string_compareTo() {
862 if (!Matcher::has_match_rule(Op_StrComp)) return false;
864 const int value_offset = java_lang_String::value_offset_in_bytes();
865 const int count_offset = java_lang_String::count_offset_in_bytes();
866 const int offset_offset = java_lang_String::offset_offset_in_bytes();
868 _sp += 2;
869 Node *argument = pop(); // pop non-receiver first: it was pushed second
870 Node *receiver = pop();
872 // Null check on self without removing any arguments. The argument
873 // null check technically happens in the wrong place, which can lead to
874 // invalid stack traces when string compare is inlined into a method
875 // which handles NullPointerExceptions.
876 _sp += 2;
877 receiver = do_null_check(receiver, T_OBJECT);
878 argument = do_null_check(argument, T_OBJECT);
879 _sp -= 2;
880 if (stopped()) {
881 return true;
882 }
884 ciInstanceKlass* klass = env()->String_klass();
885 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
886 Node* no_ctrl = NULL;
888 // Get counts for string and argument
889 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
890 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
892 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
893 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
895 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
896 push(compare);
897 return true;
898 }
900 //------------------------------inline_string_equals------------------------
901 bool LibraryCallKit::inline_string_equals() {
903 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
905 const int value_offset = java_lang_String::value_offset_in_bytes();
906 const int count_offset = java_lang_String::count_offset_in_bytes();
907 const int offset_offset = java_lang_String::offset_offset_in_bytes();
909 _sp += 2;
910 Node* argument = pop(); // pop non-receiver first: it was pushed second
911 Node* receiver = pop();
913 // Null check on self without removing any arguments. The argument
914 // null check technically happens in the wrong place, which can lead to
915 // invalid stack traces when string compare is inlined into a method
916 // which handles NullPointerExceptions.
917 _sp += 2;
918 receiver = do_null_check(receiver, T_OBJECT);
919 //should not do null check for argument for String.equals(), because spec
920 //allows to specify NULL as argument.
921 _sp -= 2;
923 if (stopped()) {
924 return true;
925 }
927 // paths (plus control) merge
928 RegionNode* region = new (C, 5) RegionNode(5);
929 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
931 // does source == target string?
932 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
933 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
935 Node* if_eq = generate_slow_guard(bol, NULL);
936 if (if_eq != NULL) {
937 // receiver == argument
938 phi->init_req(2, intcon(1));
939 region->init_req(2, if_eq);
940 }
942 // get String klass for instanceOf
943 ciInstanceKlass* klass = env()->String_klass();
945 if (!stopped()) {
946 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
947 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
948 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
950 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
951 //instanceOf == true, fallthrough
953 if (inst_false != NULL) {
954 phi->init_req(3, intcon(0));
955 region->init_req(3, inst_false);
956 }
957 }
959 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
961 Node* no_ctrl = NULL;
962 Node* receiver_cnt;
963 Node* argument_cnt;
965 if (!stopped()) {
966 // Properly cast the argument to String
967 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
969 // Get counts for string and argument
970 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
971 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
973 // Pin load from argument string since it could be NULL.
974 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
975 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
977 // Check for receiver count != argument count
978 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
979 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
980 Node* if_ne = generate_slow_guard(bol, NULL);
981 if (if_ne != NULL) {
982 phi->init_req(4, intcon(0));
983 region->init_req(4, if_ne);
984 }
985 }
987 // Check for count == 0 is done by mach node StrEquals.
989 if (!stopped()) {
990 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
991 phi->init_req(1, equals);
992 region->init_req(1, control());
993 }
995 // post merge
996 set_control(_gvn.transform(region));
997 record_for_igvn(region);
999 push(_gvn.transform(phi));
1001 return true;
1002 }
1004 //------------------------------inline_array_equals----------------------------
1005 bool LibraryCallKit::inline_array_equals() {
1007 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1009 _sp += 2;
1010 Node *argument2 = pop();
1011 Node *argument1 = pop();
1013 Node* equals =
1014 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1015 argument1, argument2) );
1016 push(equals);
1017 return true;
1018 }
1020 // Java version of String.indexOf(constant string)
1021 // class StringDecl {
1022 // StringDecl(char[] ca) {
1023 // offset = 0;
1024 // count = ca.length;
1025 // value = ca;
1026 // }
1027 // int offset;
1028 // int count;
1029 // char[] value;
1030 // }
1031 //
1032 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1033 // int targetOffset, int cache_i, int md2) {
1034 // int cache = cache_i;
1035 // int sourceOffset = string_object.offset;
1036 // int sourceCount = string_object.count;
1037 // int targetCount = target_object.length;
1038 //
1039 // int targetCountLess1 = targetCount - 1;
1040 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1041 //
1042 // char[] source = string_object.value;
1043 // char[] target = target_object;
1044 // int lastChar = target[targetCountLess1];
1045 //
1046 // outer_loop:
1047 // for (int i = sourceOffset; i < sourceEnd; ) {
1048 // int src = source[i + targetCountLess1];
1049 // if (src == lastChar) {
1050 // // With random strings and a 4-character alphabet,
1051 // // reverse matching at this point sets up 0.8% fewer
1052 // // frames, but (paradoxically) makes 0.3% more probes.
1053 // // Since those probes are nearer the lastChar probe,
1054 // // there is may be a net D$ win with reverse matching.
1055 // // But, reversing loop inhibits unroll of inner loop
1056 // // for unknown reason. So, does running outer loop from
1057 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1058 // for (int j = 0; j < targetCountLess1; j++) {
1059 // if (target[targetOffset + j] != source[i+j]) {
1060 // if ((cache & (1 << source[i+j])) == 0) {
1061 // if (md2 < j+1) {
1062 // i += j+1;
1063 // continue outer_loop;
1064 // }
1065 // }
1066 // i += md2;
1067 // continue outer_loop;
1068 // }
1069 // }
1070 // return i - sourceOffset;
1071 // }
1072 // if ((cache & (1 << src)) == 0) {
1073 // i += targetCountLess1;
1074 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1075 // i++;
1076 // }
1077 // return -1;
1078 // }
1080 //------------------------------string_indexOf------------------------
1081 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1082 jint cache_i, jint md2_i) {
1084 Node* no_ctrl = NULL;
1085 float likely = PROB_LIKELY(0.9);
1086 float unlikely = PROB_UNLIKELY(0.9);
1088 const int value_offset = java_lang_String::value_offset_in_bytes();
1089 const int count_offset = java_lang_String::count_offset_in_bytes();
1090 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1092 ciInstanceKlass* klass = env()->String_klass();
1093 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1094 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1096 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1097 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1098 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1099 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1100 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1101 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1103 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
1104 jint target_length = target_array->length();
1105 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1106 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1108 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1109 #define __ kit.
1110 Node* zero = __ ConI(0);
1111 Node* one = __ ConI(1);
1112 Node* cache = __ ConI(cache_i);
1113 Node* md2 = __ ConI(md2_i);
1114 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1115 Node* targetCount = __ ConI(target_length);
1116 Node* targetCountLess1 = __ ConI(target_length - 1);
1117 Node* targetOffset = __ ConI(targetOffset_i);
1118 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1120 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1121 Node* outer_loop = __ make_label(2 /* goto */);
1122 Node* return_ = __ make_label(1);
1124 __ set(rtn,__ ConI(-1));
1125 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
1126 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1127 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1128 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1129 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1130 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
1131 Node* tpj = __ AddI(targetOffset, __ value(j));
1132 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1133 Node* ipj = __ AddI(__ value(i), __ value(j));
1134 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1135 __ if_then(targ, BoolTest::ne, src2); {
1136 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1137 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1138 __ increment(i, __ AddI(__ value(j), one));
1139 __ goto_(outer_loop);
1140 } __ end_if(); __ dead(j);
1141 }__ end_if(); __ dead(j);
1142 __ increment(i, md2);
1143 __ goto_(outer_loop);
1144 }__ end_if();
1145 __ increment(j, one);
1146 }__ end_loop(); __ dead(j);
1147 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1148 __ goto_(return_);
1149 }__ end_if();
1150 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1151 __ increment(i, targetCountLess1);
1152 }__ end_if();
1153 __ increment(i, one);
1154 __ bind(outer_loop);
1155 }__ end_loop(); __ dead(i);
1156 __ bind(return_);
1158 // Final sync IdealKit and GraphKit.
1159 sync_kit(kit);
1160 Node* result = __ value(rtn);
1161 #undef __
1162 C->set_has_loops(true);
1163 return result;
1164 }
1166 //------------------------------inline_string_indexOf------------------------
1167 bool LibraryCallKit::inline_string_indexOf() {
1169 const int value_offset = java_lang_String::value_offset_in_bytes();
1170 const int count_offset = java_lang_String::count_offset_in_bytes();
1171 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1173 _sp += 2;
1174 Node *argument = pop(); // pop non-receiver first: it was pushed second
1175 Node *receiver = pop();
1177 Node* result;
1178 // Disable the use of pcmpestri until it can be guaranteed that
1179 // the load doesn't cross into the uncommited space.
1180 if (false && Matcher::has_match_rule(Op_StrIndexOf) &&
1181 UseSSE42Intrinsics) {
1182 // Generate SSE4.2 version of indexOf
1183 // We currently only have match rules that use SSE4.2
1185 // Null check on self without removing any arguments. The argument
1186 // null check technically happens in the wrong place, which can lead to
1187 // invalid stack traces when string compare is inlined into a method
1188 // which handles NullPointerExceptions.
1189 _sp += 2;
1190 receiver = do_null_check(receiver, T_OBJECT);
1191 argument = do_null_check(argument, T_OBJECT);
1192 _sp -= 2;
1194 if (stopped()) {
1195 return true;
1196 }
1198 // Make the merge point
1199 RegionNode* result_rgn = new (C, 3) RegionNode(3);
1200 Node* result_phi = new (C, 3) PhiNode(result_rgn, TypeInt::INT);
1201 Node* no_ctrl = NULL;
1203 ciInstanceKlass* klass = env()->String_klass();
1204 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1206 // Get counts for string and substr
1207 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1208 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1210 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1211 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1213 // Check for substr count > string count
1214 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1215 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1216 Node* if_gt = generate_slow_guard(bol, NULL);
1217 if (if_gt != NULL) {
1218 result_phi->init_req(2, intcon(-1));
1219 result_rgn->init_req(2, if_gt);
1220 }
1222 if (!stopped()) {
1223 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1224 result_phi->init_req(1, result);
1225 result_rgn->init_req(1, control());
1226 }
1227 set_control(_gvn.transform(result_rgn));
1228 record_for_igvn(result_rgn);
1229 result = _gvn.transform(result_phi);
1231 } else { //Use LibraryCallKit::string_indexOf
1232 // don't intrinsify is argument isn't a constant string.
1233 if (!argument->is_Con()) {
1234 return false;
1235 }
1236 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1237 if (str_type == NULL) {
1238 return false;
1239 }
1240 ciInstanceKlass* klass = env()->String_klass();
1241 ciObject* str_const = str_type->const_oop();
1242 if (str_const == NULL || str_const->klass() != klass) {
1243 return false;
1244 }
1245 ciInstance* str = str_const->as_instance();
1246 assert(str != NULL, "must be instance");
1248 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1249 int o = str->field_value_by_offset(offset_offset).as_int();
1250 int c = str->field_value_by_offset(count_offset).as_int();
1251 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1253 // constant strings have no offset and count == length which
1254 // simplifies the resulting code somewhat so lets optimize for that.
1255 if (o != 0 || c != pat->length()) {
1256 return false;
1257 }
1259 // Null check on self without removing any arguments. The argument
1260 // null check technically happens in the wrong place, which can lead to
1261 // invalid stack traces when string compare is inlined into a method
1262 // which handles NullPointerExceptions.
1263 _sp += 2;
1264 receiver = do_null_check(receiver, T_OBJECT);
1265 // No null check on the argument is needed since it's a constant String oop.
1266 _sp -= 2;
1267 if (stopped()) {
1268 return true;
1269 }
1271 // The null string as a pattern always returns 0 (match at beginning of string)
1272 if (c == 0) {
1273 push(intcon(0));
1274 return true;
1275 }
1277 // Generate default indexOf
1278 jchar lastChar = pat->char_at(o + (c - 1));
1279 int cache = 0;
1280 int i;
1281 for (i = 0; i < c - 1; i++) {
1282 assert(i < pat->length(), "out of range");
1283 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1284 }
1286 int md2 = c;
1287 for (i = 0; i < c - 1; i++) {
1288 assert(i < pat->length(), "out of range");
1289 if (pat->char_at(o + i) == lastChar) {
1290 md2 = (c - 1) - i;
1291 }
1292 }
1294 result = string_indexOf(receiver, pat, o, cache, md2);
1295 }
1297 push(result);
1298 return true;
1299 }
1301 //--------------------------pop_math_arg--------------------------------
1302 // Pop a double argument to a math function from the stack
1303 // rounding it if necessary.
1304 Node * LibraryCallKit::pop_math_arg() {
1305 Node *arg = pop_pair();
1306 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1307 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1308 return arg;
1309 }
1311 //------------------------------inline_trig----------------------------------
1312 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1313 // argument reduction which will turn into a fast/slow diamond.
1314 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1315 _sp += arg_size(); // restore stack pointer
1316 Node* arg = pop_math_arg();
1317 Node* trig = NULL;
1319 switch (id) {
1320 case vmIntrinsics::_dsin:
1321 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1322 break;
1323 case vmIntrinsics::_dcos:
1324 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1325 break;
1326 case vmIntrinsics::_dtan:
1327 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1328 break;
1329 default:
1330 assert(false, "bad intrinsic was passed in");
1331 return false;
1332 }
1334 // Rounding required? Check for argument reduction!
1335 if( Matcher::strict_fp_requires_explicit_rounding ) {
1337 static const double pi_4 = 0.7853981633974483;
1338 static const double neg_pi_4 = -0.7853981633974483;
1339 // pi/2 in 80-bit extended precision
1340 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1341 // -pi/2 in 80-bit extended precision
1342 // 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};
1343 // Cutoff value for using this argument reduction technique
1344 //static const double pi_2_minus_epsilon = 1.564660403643354;
1345 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1347 // Pseudocode for sin:
1348 // if (x <= Math.PI / 4.0) {
1349 // if (x >= -Math.PI / 4.0) return fsin(x);
1350 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1351 // } else {
1352 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1353 // }
1354 // return StrictMath.sin(x);
1356 // Pseudocode for cos:
1357 // if (x <= Math.PI / 4.0) {
1358 // if (x >= -Math.PI / 4.0) return fcos(x);
1359 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1360 // } else {
1361 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1362 // }
1363 // return StrictMath.cos(x);
1365 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1366 // requires a special machine instruction to load it. Instead we'll try
1367 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1368 // probably do the math inside the SIN encoding.
1370 // Make the merge point
1371 RegionNode *r = new (C, 3) RegionNode(3);
1372 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1374 // Flatten arg so we need only 1 test
1375 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1376 // Node for PI/4 constant
1377 Node *pi4 = makecon(TypeD::make(pi_4));
1378 // Check PI/4 : abs(arg)
1379 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1380 // Check: If PI/4 < abs(arg) then go slow
1381 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1382 // Branch either way
1383 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1384 set_control(opt_iff(r,iff));
1386 // Set fast path result
1387 phi->init_req(2,trig);
1389 // Slow path - non-blocking leaf call
1390 Node* call = NULL;
1391 switch (id) {
1392 case vmIntrinsics::_dsin:
1393 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1394 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1395 "Sin", NULL, arg, top());
1396 break;
1397 case vmIntrinsics::_dcos:
1398 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1399 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1400 "Cos", NULL, arg, top());
1401 break;
1402 case vmIntrinsics::_dtan:
1403 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1404 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1405 "Tan", NULL, arg, top());
1406 break;
1407 }
1408 assert(control()->in(0) == call, "");
1409 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1410 r->init_req(1,control());
1411 phi->init_req(1,slow_result);
1413 // Post-merge
1414 set_control(_gvn.transform(r));
1415 record_for_igvn(r);
1416 trig = _gvn.transform(phi);
1418 C->set_has_split_ifs(true); // Has chance for split-if optimization
1419 }
1420 // Push result back on JVM stack
1421 push_pair(trig);
1422 return true;
1423 }
1425 //------------------------------inline_sqrt-------------------------------------
1426 // Inline square root instruction, if possible.
1427 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1428 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1429 _sp += arg_size(); // restore stack pointer
1430 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1431 return true;
1432 }
1434 //------------------------------inline_abs-------------------------------------
1435 // Inline absolute value instruction, if possible.
1436 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1437 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1438 _sp += arg_size(); // restore stack pointer
1439 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1440 return true;
1441 }
1443 //------------------------------inline_exp-------------------------------------
1444 // Inline exp instructions, if possible. The Intel hardware only misses
1445 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1446 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1447 assert(id == vmIntrinsics::_dexp, "Not exp");
1449 // If this inlining ever returned NaN in the past, we do not intrinsify it
1450 // every again. NaN results requires StrictMath.exp handling.
1451 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1453 // Do not intrinsify on older platforms which lack cmove.
1454 if (ConditionalMoveLimit == 0) return false;
1456 _sp += arg_size(); // restore stack pointer
1457 Node *x = pop_math_arg();
1458 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1460 //-------------------
1461 //result=(result.isNaN())? StrictMath::exp():result;
1462 // Check: If isNaN() by checking result!=result? then go to Strict Math
1463 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1464 // Build the boolean node
1465 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1467 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1468 // End the current control-flow path
1469 push_pair(x);
1470 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1471 // to handle. Recompile without intrinsifying Math.exp
1472 uncommon_trap(Deoptimization::Reason_intrinsic,
1473 Deoptimization::Action_make_not_entrant);
1474 }
1476 C->set_has_split_ifs(true); // Has chance for split-if optimization
1478 push_pair(result);
1480 return true;
1481 }
1483 //------------------------------inline_pow-------------------------------------
1484 // Inline power instructions, if possible.
1485 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1486 assert(id == vmIntrinsics::_dpow, "Not pow");
1488 // If this inlining ever returned NaN in the past, we do not intrinsify it
1489 // every again. NaN results requires StrictMath.pow handling.
1490 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1492 // Do not intrinsify on older platforms which lack cmove.
1493 if (ConditionalMoveLimit == 0) return false;
1495 // Pseudocode for pow
1496 // if (x <= 0.0) {
1497 // if ((double)((int)y)==y) { // if y is int
1498 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1499 // } else {
1500 // result = NaN;
1501 // }
1502 // } else {
1503 // result = DPow(x,y);
1504 // }
1505 // if (result != result)? {
1506 // uncommon_trap();
1507 // }
1508 // return result;
1510 _sp += arg_size(); // restore stack pointer
1511 Node* y = pop_math_arg();
1512 Node* x = pop_math_arg();
1514 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1516 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1517 // inside of something) then skip the fancy tests and just check for
1518 // NaN result.
1519 Node *result = NULL;
1520 if( jvms()->depth() >= 1 ) {
1521 result = fast_result;
1522 } else {
1524 // Set the merge point for If node with condition of (x <= 0.0)
1525 // There are four possible paths to region node and phi node
1526 RegionNode *r = new (C, 4) RegionNode(4);
1527 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1529 // Build the first if node: if (x <= 0.0)
1530 // Node for 0 constant
1531 Node *zeronode = makecon(TypeD::ZERO);
1532 // Check x:0
1533 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1534 // Check: If (x<=0) then go complex path
1535 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1536 // Branch either way
1537 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1538 Node *opt_test = _gvn.transform(if1);
1539 //assert( opt_test->is_If(), "Expect an IfNode");
1540 IfNode *opt_if1 = (IfNode*)opt_test;
1541 // Fast path taken; set region slot 3
1542 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1543 r->init_req(3,fast_taken); // Capture fast-control
1545 // Fast path not-taken, i.e. slow path
1546 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1548 // Set fast path result
1549 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1550 phi->init_req(3, fast_result);
1552 // Complex path
1553 // Build the second if node (if y is int)
1554 // Node for (int)y
1555 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1556 // Node for (double)((int) y)
1557 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1558 // Check (double)((int) y) : y
1559 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1560 // Check if (y isn't int) then go to slow path
1562 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1563 // Branch either way
1564 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1565 Node *slow_path = opt_iff(r,if2); // Set region path 2
1567 // Calculate DPow(abs(x), y)*(1 & (int)y)
1568 // Node for constant 1
1569 Node *conone = intcon(1);
1570 // 1& (int)y
1571 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1572 // zero node
1573 Node *conzero = intcon(0);
1574 // Check (1&(int)y)==0?
1575 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1576 // Check if (1&(int)y)!=0?, if so the result is negative
1577 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1578 // abs(x)
1579 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1580 // abs(x)^y
1581 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1582 // -abs(x)^y
1583 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1584 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1585 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1586 // Set complex path fast result
1587 phi->init_req(2, signresult);
1589 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1590 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1591 r->init_req(1,slow_path);
1592 phi->init_req(1,slow_result);
1594 // Post merge
1595 set_control(_gvn.transform(r));
1596 record_for_igvn(r);
1597 result=_gvn.transform(phi);
1598 }
1600 //-------------------
1601 //result=(result.isNaN())? uncommon_trap():result;
1602 // Check: If isNaN() by checking result!=result? then go to Strict Math
1603 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1604 // Build the boolean node
1605 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1607 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1608 // End the current control-flow path
1609 push_pair(x);
1610 push_pair(y);
1611 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1612 // to handle. Recompile without intrinsifying Math.pow.
1613 uncommon_trap(Deoptimization::Reason_intrinsic,
1614 Deoptimization::Action_make_not_entrant);
1615 }
1617 C->set_has_split_ifs(true); // Has chance for split-if optimization
1619 push_pair(result);
1621 return true;
1622 }
1624 //------------------------------inline_trans-------------------------------------
1625 // Inline transcendental instructions, if possible. The Intel hardware gets
1626 // these right, no funny corner cases missed.
1627 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1628 _sp += arg_size(); // restore stack pointer
1629 Node* arg = pop_math_arg();
1630 Node* trans = NULL;
1632 switch (id) {
1633 case vmIntrinsics::_dlog:
1634 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1635 break;
1636 case vmIntrinsics::_dlog10:
1637 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1638 break;
1639 default:
1640 assert(false, "bad intrinsic was passed in");
1641 return false;
1642 }
1644 // Push result back on JVM stack
1645 push_pair(trans);
1646 return true;
1647 }
1649 //------------------------------runtime_math-----------------------------
1650 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1651 Node* a = NULL;
1652 Node* b = NULL;
1654 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1655 "must be (DD)D or (D)D type");
1657 // Inputs
1658 _sp += arg_size(); // restore stack pointer
1659 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1660 b = pop_math_arg();
1661 }
1662 a = pop_math_arg();
1664 const TypePtr* no_memory_effects = NULL;
1665 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1666 no_memory_effects,
1667 a, top(), b, b ? top() : NULL);
1668 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1669 #ifdef ASSERT
1670 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1671 assert(value_top == top(), "second value must be top");
1672 #endif
1674 push_pair(value);
1675 return true;
1676 }
1678 //------------------------------inline_math_native-----------------------------
1679 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1680 switch (id) {
1681 // These intrinsics are not properly supported on all hardware
1682 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1683 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1684 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1685 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1686 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1687 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1689 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1690 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1691 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1692 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1694 // These intrinsics are supported on all hardware
1695 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1696 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1698 // These intrinsics don't work on X86. The ad implementation doesn't
1699 // handle NaN's properly. Instead of returning infinity, the ad
1700 // implementation returns a NaN on overflow. See bug: 6304089
1701 // Once the ad implementations are fixed, change the code below
1702 // to match the intrinsics above
1704 case vmIntrinsics::_dexp: return
1705 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1706 case vmIntrinsics::_dpow: return
1707 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1709 // These intrinsics are not yet correctly implemented
1710 case vmIntrinsics::_datan2:
1711 return false;
1713 default:
1714 ShouldNotReachHere();
1715 return false;
1716 }
1717 }
1719 static bool is_simple_name(Node* n) {
1720 return (n->req() == 1 // constant
1721 || (n->is_Type() && n->as_Type()->type()->singleton())
1722 || n->is_Proj() // parameter or return value
1723 || n->is_Phi() // local of some sort
1724 );
1725 }
1727 //----------------------------inline_min_max-----------------------------------
1728 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1729 push(generate_min_max(id, argument(0), argument(1)));
1731 return true;
1732 }
1734 Node*
1735 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1736 // These are the candidate return value:
1737 Node* xvalue = x0;
1738 Node* yvalue = y0;
1740 if (xvalue == yvalue) {
1741 return xvalue;
1742 }
1744 bool want_max = (id == vmIntrinsics::_max);
1746 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1747 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1748 if (txvalue == NULL || tyvalue == NULL) return top();
1749 // This is not really necessary, but it is consistent with a
1750 // hypothetical MaxINode::Value method:
1751 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1753 // %%% This folding logic should (ideally) be in a different place.
1754 // Some should be inside IfNode, and there to be a more reliable
1755 // transformation of ?: style patterns into cmoves. We also want
1756 // more powerful optimizations around cmove and min/max.
1758 // Try to find a dominating comparison of these guys.
1759 // It can simplify the index computation for Arrays.copyOf
1760 // and similar uses of System.arraycopy.
1761 // First, compute the normalized version of CmpI(x, y).
1762 int cmp_op = Op_CmpI;
1763 Node* xkey = xvalue;
1764 Node* ykey = yvalue;
1765 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1766 if (ideal_cmpxy->is_Cmp()) {
1767 // E.g., if we have CmpI(length - offset, count),
1768 // it might idealize to CmpI(length, count + offset)
1769 cmp_op = ideal_cmpxy->Opcode();
1770 xkey = ideal_cmpxy->in(1);
1771 ykey = ideal_cmpxy->in(2);
1772 }
1774 // Start by locating any relevant comparisons.
1775 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1776 Node* cmpxy = NULL;
1777 Node* cmpyx = NULL;
1778 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1779 Node* cmp = start_from->fast_out(k);
1780 if (cmp->outcnt() > 0 && // must have prior uses
1781 cmp->in(0) == NULL && // must be context-independent
1782 cmp->Opcode() == cmp_op) { // right kind of compare
1783 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1784 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1785 }
1786 }
1788 const int NCMPS = 2;
1789 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1790 int cmpn;
1791 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1792 if (cmps[cmpn] != NULL) break; // find a result
1793 }
1794 if (cmpn < NCMPS) {
1795 // Look for a dominating test that tells us the min and max.
1796 int depth = 0; // Limit search depth for speed
1797 Node* dom = control();
1798 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1799 if (++depth >= 100) break;
1800 Node* ifproj = dom;
1801 if (!ifproj->is_Proj()) continue;
1802 Node* iff = ifproj->in(0);
1803 if (!iff->is_If()) continue;
1804 Node* bol = iff->in(1);
1805 if (!bol->is_Bool()) continue;
1806 Node* cmp = bol->in(1);
1807 if (cmp == NULL) continue;
1808 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1809 if (cmps[cmpn] == cmp) break;
1810 if (cmpn == NCMPS) continue;
1811 BoolTest::mask btest = bol->as_Bool()->_test._test;
1812 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1813 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1814 // At this point, we know that 'x btest y' is true.
1815 switch (btest) {
1816 case BoolTest::eq:
1817 // They are proven equal, so we can collapse the min/max.
1818 // Either value is the answer. Choose the simpler.
1819 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1820 return yvalue;
1821 return xvalue;
1822 case BoolTest::lt: // x < y
1823 case BoolTest::le: // x <= y
1824 return (want_max ? yvalue : xvalue);
1825 case BoolTest::gt: // x > y
1826 case BoolTest::ge: // x >= y
1827 return (want_max ? xvalue : yvalue);
1828 }
1829 }
1830 }
1832 // We failed to find a dominating test.
1833 // Let's pick a test that might GVN with prior tests.
1834 Node* best_bol = NULL;
1835 BoolTest::mask best_btest = BoolTest::illegal;
1836 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1837 Node* cmp = cmps[cmpn];
1838 if (cmp == NULL) continue;
1839 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1840 Node* bol = cmp->fast_out(j);
1841 if (!bol->is_Bool()) continue;
1842 BoolTest::mask btest = bol->as_Bool()->_test._test;
1843 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1844 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1845 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1846 best_bol = bol->as_Bool();
1847 best_btest = btest;
1848 }
1849 }
1850 }
1852 Node* answer_if_true = NULL;
1853 Node* answer_if_false = NULL;
1854 switch (best_btest) {
1855 default:
1856 if (cmpxy == NULL)
1857 cmpxy = ideal_cmpxy;
1858 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1859 // and fall through:
1860 case BoolTest::lt: // x < y
1861 case BoolTest::le: // x <= y
1862 answer_if_true = (want_max ? yvalue : xvalue);
1863 answer_if_false = (want_max ? xvalue : yvalue);
1864 break;
1865 case BoolTest::gt: // x > y
1866 case BoolTest::ge: // x >= y
1867 answer_if_true = (want_max ? xvalue : yvalue);
1868 answer_if_false = (want_max ? yvalue : xvalue);
1869 break;
1870 }
1872 jint hi, lo;
1873 if (want_max) {
1874 // We can sharpen the minimum.
1875 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1876 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1877 } else {
1878 // We can sharpen the maximum.
1879 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1880 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1881 }
1883 // Use a flow-free graph structure, to avoid creating excess control edges
1884 // which could hinder other optimizations.
1885 // Since Math.min/max is often used with arraycopy, we want
1886 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1887 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1888 answer_if_false, answer_if_true,
1889 TypeInt::make(lo, hi, widen));
1891 return _gvn.transform(cmov);
1893 /*
1894 // This is not as desirable as it may seem, since Min and Max
1895 // nodes do not have a full set of optimizations.
1896 // And they would interfere, anyway, with 'if' optimizations
1897 // and with CMoveI canonical forms.
1898 switch (id) {
1899 case vmIntrinsics::_min:
1900 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1901 case vmIntrinsics::_max:
1902 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1903 default:
1904 ShouldNotReachHere();
1905 }
1906 */
1907 }
1909 inline int
1910 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1911 const TypePtr* base_type = TypePtr::NULL_PTR;
1912 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1913 if (base_type == NULL) {
1914 // Unknown type.
1915 return Type::AnyPtr;
1916 } else if (base_type == TypePtr::NULL_PTR) {
1917 // Since this is a NULL+long form, we have to switch to a rawptr.
1918 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1919 offset = MakeConX(0);
1920 return Type::RawPtr;
1921 } else if (base_type->base() == Type::RawPtr) {
1922 return Type::RawPtr;
1923 } else if (base_type->isa_oopptr()) {
1924 // Base is never null => always a heap address.
1925 if (base_type->ptr() == TypePtr::NotNull) {
1926 return Type::OopPtr;
1927 }
1928 // Offset is small => always a heap address.
1929 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1930 if (offset_type != NULL &&
1931 base_type->offset() == 0 && // (should always be?)
1932 offset_type->_lo >= 0 &&
1933 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1934 return Type::OopPtr;
1935 }
1936 // Otherwise, it might either be oop+off or NULL+addr.
1937 return Type::AnyPtr;
1938 } else {
1939 // No information:
1940 return Type::AnyPtr;
1941 }
1942 }
1944 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1945 int kind = classify_unsafe_addr(base, offset);
1946 if (kind == Type::RawPtr) {
1947 return basic_plus_adr(top(), base, offset);
1948 } else {
1949 return basic_plus_adr(base, offset);
1950 }
1951 }
1953 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1954 // inline int Integer.numberOfLeadingZeros(int)
1955 // inline int Long.numberOfLeadingZeros(long)
1956 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1957 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1958 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1959 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1960 _sp += arg_size(); // restore stack pointer
1961 switch (id) {
1962 case vmIntrinsics::_numberOfLeadingZeros_i:
1963 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1964 break;
1965 case vmIntrinsics::_numberOfLeadingZeros_l:
1966 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1967 break;
1968 default:
1969 ShouldNotReachHere();
1970 }
1971 return true;
1972 }
1974 //-------------------inline_numberOfTrailingZeros_int/long----------------------
1975 // inline int Integer.numberOfTrailingZeros(int)
1976 // inline int Long.numberOfTrailingZeros(long)
1977 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
1978 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
1979 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
1980 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
1981 _sp += arg_size(); // restore stack pointer
1982 switch (id) {
1983 case vmIntrinsics::_numberOfTrailingZeros_i:
1984 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
1985 break;
1986 case vmIntrinsics::_numberOfTrailingZeros_l:
1987 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
1988 break;
1989 default:
1990 ShouldNotReachHere();
1991 }
1992 return true;
1993 }
1995 //----------------------------inline_bitCount_int/long-----------------------
1996 // inline int Integer.bitCount(int)
1997 // inline int Long.bitCount(long)
1998 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
1999 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2000 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2001 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2002 _sp += arg_size(); // restore stack pointer
2003 switch (id) {
2004 case vmIntrinsics::_bitCount_i:
2005 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2006 break;
2007 case vmIntrinsics::_bitCount_l:
2008 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2009 break;
2010 default:
2011 ShouldNotReachHere();
2012 }
2013 return true;
2014 }
2016 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2017 // inline Integer.reverseBytes(int)
2018 // inline Long.reverseBytes(long)
2019 // inline Character.reverseBytes(char)
2020 // inline Short.reverseBytes(short)
2021 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2022 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2023 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2024 "not reverse Bytes");
2025 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2026 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2027 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2028 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2029 _sp += arg_size(); // restore stack pointer
2030 switch (id) {
2031 case vmIntrinsics::_reverseBytes_i:
2032 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2033 break;
2034 case vmIntrinsics::_reverseBytes_l:
2035 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2036 break;
2037 case vmIntrinsics::_reverseBytes_c:
2038 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2039 break;
2040 case vmIntrinsics::_reverseBytes_s:
2041 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2042 break;
2043 default:
2044 ;
2045 }
2046 return true;
2047 }
2049 //----------------------------inline_unsafe_access----------------------------
2051 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2053 // Interpret Unsafe.fieldOffset cookies correctly:
2054 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2056 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2057 if (callee()->is_static()) return false; // caller must have the capability!
2059 #ifndef PRODUCT
2060 {
2061 ResourceMark rm;
2062 // Check the signatures.
2063 ciSignature* sig = signature();
2064 #ifdef ASSERT
2065 if (!is_store) {
2066 // Object getObject(Object base, int/long offset), etc.
2067 BasicType rtype = sig->return_type()->basic_type();
2068 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2069 rtype = T_ADDRESS; // it is really a C void*
2070 assert(rtype == type, "getter must return the expected value");
2071 if (!is_native_ptr) {
2072 assert(sig->count() == 2, "oop getter has 2 arguments");
2073 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2074 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2075 } else {
2076 assert(sig->count() == 1, "native getter has 1 argument");
2077 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2078 }
2079 } else {
2080 // void putObject(Object base, int/long offset, Object x), etc.
2081 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2082 if (!is_native_ptr) {
2083 assert(sig->count() == 3, "oop putter has 3 arguments");
2084 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2085 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2086 } else {
2087 assert(sig->count() == 2, "native putter has 2 arguments");
2088 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2089 }
2090 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2091 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2092 vtype = T_ADDRESS; // it is really a C void*
2093 assert(vtype == type, "putter must accept the expected value");
2094 }
2095 #endif // ASSERT
2096 }
2097 #endif //PRODUCT
2099 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2101 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2103 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2104 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2106 debug_only(int saved_sp = _sp);
2107 _sp += nargs;
2109 Node* val;
2110 debug_only(val = (Node*)(uintptr_t)-1);
2113 if (is_store) {
2114 // Get the value being stored. (Pop it first; it was pushed last.)
2115 switch (type) {
2116 case T_DOUBLE:
2117 case T_LONG:
2118 case T_ADDRESS:
2119 val = pop_pair();
2120 break;
2121 default:
2122 val = pop();
2123 }
2124 }
2126 // Build address expression. See the code in inline_unsafe_prefetch.
2127 Node *adr;
2128 Node *heap_base_oop = top();
2129 if (!is_native_ptr) {
2130 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2131 Node* offset = pop_pair();
2132 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2133 Node* base = pop();
2134 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2135 // to be plain byte offsets, which are also the same as those accepted
2136 // by oopDesc::field_base.
2137 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2138 "fieldOffset must be byte-scaled");
2139 // 32-bit machines ignore the high half!
2140 offset = ConvL2X(offset);
2141 adr = make_unsafe_address(base, offset);
2142 heap_base_oop = base;
2143 } else {
2144 Node* ptr = pop_pair();
2145 // Adjust Java long to machine word:
2146 ptr = ConvL2X(ptr);
2147 adr = make_unsafe_address(NULL, ptr);
2148 }
2150 // Pop receiver last: it was pushed first.
2151 Node *receiver = pop();
2153 assert(saved_sp == _sp, "must have correct argument count");
2155 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2157 // First guess at the value type.
2158 const Type *value_type = Type::get_const_basic_type(type);
2160 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2161 // there was not enough information to nail it down.
2162 Compile::AliasType* alias_type = C->alias_type(adr_type);
2163 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2165 // We will need memory barriers unless we can determine a unique
2166 // alias category for this reference. (Note: If for some reason
2167 // the barriers get omitted and the unsafe reference begins to "pollute"
2168 // the alias analysis of the rest of the graph, either Compile::can_alias
2169 // or Compile::must_alias will throw a diagnostic assert.)
2170 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2172 if (!is_store && type == T_OBJECT) {
2173 // Attempt to infer a sharper value type from the offset and base type.
2174 ciKlass* sharpened_klass = NULL;
2176 // See if it is an instance field, with an object type.
2177 if (alias_type->field() != NULL) {
2178 assert(!is_native_ptr, "native pointer op cannot use a java address");
2179 if (alias_type->field()->type()->is_klass()) {
2180 sharpened_klass = alias_type->field()->type()->as_klass();
2181 }
2182 }
2184 // See if it is a narrow oop array.
2185 if (adr_type->isa_aryptr()) {
2186 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2187 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2188 if (elem_type != NULL) {
2189 sharpened_klass = elem_type->klass();
2190 }
2191 }
2192 }
2194 if (sharpened_klass != NULL) {
2195 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2197 // Sharpen the value type.
2198 value_type = tjp;
2200 #ifndef PRODUCT
2201 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2202 tty->print(" from base type: "); adr_type->dump();
2203 tty->print(" sharpened value: "); value_type->dump();
2204 }
2205 #endif
2206 }
2207 }
2209 // Null check on self without removing any arguments. The argument
2210 // null check technically happens in the wrong place, which can lead to
2211 // invalid stack traces when the primitive is inlined into a method
2212 // which handles NullPointerExceptions.
2213 _sp += nargs;
2214 do_null_check(receiver, T_OBJECT);
2215 _sp -= nargs;
2216 if (stopped()) {
2217 return true;
2218 }
2219 // Heap pointers get a null-check from the interpreter,
2220 // as a courtesy. However, this is not guaranteed by Unsafe,
2221 // and it is not possible to fully distinguish unintended nulls
2222 // from intended ones in this API.
2224 if (is_volatile) {
2225 // We need to emit leading and trailing CPU membars (see below) in
2226 // addition to memory membars when is_volatile. This is a little
2227 // too strong, but avoids the need to insert per-alias-type
2228 // volatile membars (for stores; compare Parse::do_put_xxx), which
2229 // we cannot do effectively here because we probably only have a
2230 // rough approximation of type.
2231 need_mem_bar = true;
2232 // For Stores, place a memory ordering barrier now.
2233 if (is_store)
2234 insert_mem_bar(Op_MemBarRelease);
2235 }
2237 // Memory barrier to prevent normal and 'unsafe' accesses from
2238 // bypassing each other. Happens after null checks, so the
2239 // exception paths do not take memory state from the memory barrier,
2240 // so there's no problems making a strong assert about mixing users
2241 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2242 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2243 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2245 if (!is_store) {
2246 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2247 // load value and push onto stack
2248 switch (type) {
2249 case T_BOOLEAN:
2250 case T_CHAR:
2251 case T_BYTE:
2252 case T_SHORT:
2253 case T_INT:
2254 case T_FLOAT:
2255 case T_OBJECT:
2256 push( p );
2257 break;
2258 case T_ADDRESS:
2259 // Cast to an int type.
2260 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2261 p = ConvX2L(p);
2262 push_pair(p);
2263 break;
2264 case T_DOUBLE:
2265 case T_LONG:
2266 push_pair( p );
2267 break;
2268 default: ShouldNotReachHere();
2269 }
2270 } else {
2271 // place effect of store into memory
2272 switch (type) {
2273 case T_DOUBLE:
2274 val = dstore_rounding(val);
2275 break;
2276 case T_ADDRESS:
2277 // Repackage the long as a pointer.
2278 val = ConvL2X(val);
2279 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2280 break;
2281 }
2283 if (type != T_OBJECT ) {
2284 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2285 } else {
2286 // Possibly an oop being stored to Java heap or native memory
2287 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2288 // oop to Java heap.
2289 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2290 } else {
2291 // We can't tell at compile time if we are storing in the Java heap or outside
2292 // of it. So we need to emit code to conditionally do the proper type of
2293 // store.
2295 IdealKit ideal(gvn(), control(), merged_memory());
2296 #define __ ideal.
2297 // QQQ who knows what probability is here??
2298 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2299 // Sync IdealKit and graphKit.
2300 set_all_memory( __ merged_memory());
2301 set_control(__ ctrl());
2302 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2303 // Update IdealKit memory.
2304 __ set_all_memory(merged_memory());
2305 __ set_ctrl(control());
2306 } __ else_(); {
2307 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2308 } __ end_if();
2309 // Final sync IdealKit and GraphKit.
2310 sync_kit(ideal);
2311 #undef __
2312 }
2313 }
2314 }
2316 if (is_volatile) {
2317 if (!is_store)
2318 insert_mem_bar(Op_MemBarAcquire);
2319 else
2320 insert_mem_bar(Op_MemBarVolatile);
2321 }
2323 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2325 return true;
2326 }
2328 //----------------------------inline_unsafe_prefetch----------------------------
2330 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2331 #ifndef PRODUCT
2332 {
2333 ResourceMark rm;
2334 // Check the signatures.
2335 ciSignature* sig = signature();
2336 #ifdef ASSERT
2337 // Object getObject(Object base, int/long offset), etc.
2338 BasicType rtype = sig->return_type()->basic_type();
2339 if (!is_native_ptr) {
2340 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2341 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2342 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2343 } else {
2344 assert(sig->count() == 1, "native prefetch has 1 argument");
2345 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2346 }
2347 #endif // ASSERT
2348 }
2349 #endif // !PRODUCT
2351 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2353 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2354 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2356 debug_only(int saved_sp = _sp);
2357 _sp += nargs;
2359 // Build address expression. See the code in inline_unsafe_access.
2360 Node *adr;
2361 if (!is_native_ptr) {
2362 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2363 Node* offset = pop_pair();
2364 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2365 Node* base = pop();
2366 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2367 // to be plain byte offsets, which are also the same as those accepted
2368 // by oopDesc::field_base.
2369 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2370 "fieldOffset must be byte-scaled");
2371 // 32-bit machines ignore the high half!
2372 offset = ConvL2X(offset);
2373 adr = make_unsafe_address(base, offset);
2374 } else {
2375 Node* ptr = pop_pair();
2376 // Adjust Java long to machine word:
2377 ptr = ConvL2X(ptr);
2378 adr = make_unsafe_address(NULL, ptr);
2379 }
2381 if (is_static) {
2382 assert(saved_sp == _sp, "must have correct argument count");
2383 } else {
2384 // Pop receiver last: it was pushed first.
2385 Node *receiver = pop();
2386 assert(saved_sp == _sp, "must have correct argument count");
2388 // Null check on self without removing any arguments. The argument
2389 // null check technically happens in the wrong place, which can lead to
2390 // invalid stack traces when the primitive is inlined into a method
2391 // which handles NullPointerExceptions.
2392 _sp += nargs;
2393 do_null_check(receiver, T_OBJECT);
2394 _sp -= nargs;
2395 if (stopped()) {
2396 return true;
2397 }
2398 }
2400 // Generate the read or write prefetch
2401 Node *prefetch;
2402 if (is_store) {
2403 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2404 } else {
2405 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2406 }
2407 prefetch->init_req(0, control());
2408 set_i_o(_gvn.transform(prefetch));
2410 return true;
2411 }
2413 //----------------------------inline_unsafe_CAS----------------------------
2415 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2416 // This basic scheme here is the same as inline_unsafe_access, but
2417 // differs in enough details that combining them would make the code
2418 // overly confusing. (This is a true fact! I originally combined
2419 // them, but even I was confused by it!) As much code/comments as
2420 // possible are retained from inline_unsafe_access though to make
2421 // the correspondences clearer. - dl
2423 if (callee()->is_static()) return false; // caller must have the capability!
2425 #ifndef PRODUCT
2426 {
2427 ResourceMark rm;
2428 // Check the signatures.
2429 ciSignature* sig = signature();
2430 #ifdef ASSERT
2431 BasicType rtype = sig->return_type()->basic_type();
2432 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2433 assert(sig->count() == 4, "CAS has 4 arguments");
2434 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2435 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2436 #endif // ASSERT
2437 }
2438 #endif //PRODUCT
2440 // number of stack slots per value argument (1 or 2)
2441 int type_words = type2size[type];
2443 // Cannot inline wide CAS on machines that don't support it natively
2444 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2445 return false;
2447 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2449 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2450 int nargs = 1 + 1 + 2 + type_words + type_words;
2452 // pop arguments: newval, oldval, offset, base, and receiver
2453 debug_only(int saved_sp = _sp);
2454 _sp += nargs;
2455 Node* newval = (type_words == 1) ? pop() : pop_pair();
2456 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2457 Node *offset = pop_pair();
2458 Node *base = pop();
2459 Node *receiver = pop();
2460 assert(saved_sp == _sp, "must have correct argument count");
2462 // Null check receiver.
2463 _sp += nargs;
2464 do_null_check(receiver, T_OBJECT);
2465 _sp -= nargs;
2466 if (stopped()) {
2467 return true;
2468 }
2470 // Build field offset expression.
2471 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2472 // to be plain byte offsets, which are also the same as those accepted
2473 // by oopDesc::field_base.
2474 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2475 // 32-bit machines ignore the high half of long offsets
2476 offset = ConvL2X(offset);
2477 Node* adr = make_unsafe_address(base, offset);
2478 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2480 // (Unlike inline_unsafe_access, there seems no point in trying
2481 // to refine types. Just use the coarse types here.
2482 const Type *value_type = Type::get_const_basic_type(type);
2483 Compile::AliasType* alias_type = C->alias_type(adr_type);
2484 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2485 int alias_idx = C->get_alias_index(adr_type);
2487 // Memory-model-wise, a CAS acts like a little synchronized block,
2488 // so needs barriers on each side. These don't translate into
2489 // actual barriers on most machines, but we still need rest of
2490 // compiler to respect ordering.
2492 insert_mem_bar(Op_MemBarRelease);
2493 insert_mem_bar(Op_MemBarCPUOrder);
2495 // 4984716: MemBars must be inserted before this
2496 // memory node in order to avoid a false
2497 // dependency which will confuse the scheduler.
2498 Node *mem = memory(alias_idx);
2500 // For now, we handle only those cases that actually exist: ints,
2501 // longs, and Object. Adding others should be straightforward.
2502 Node* cas;
2503 switch(type) {
2504 case T_INT:
2505 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2506 break;
2507 case T_LONG:
2508 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2509 break;
2510 case T_OBJECT:
2511 // reference stores need a store barrier.
2512 // (They don't if CAS fails, but it isn't worth checking.)
2513 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2514 #ifdef _LP64
2515 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2516 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2517 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2518 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2519 newval_enc, oldval_enc));
2520 } else
2521 #endif
2522 {
2523 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2524 }
2525 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2526 break;
2527 default:
2528 ShouldNotReachHere();
2529 break;
2530 }
2532 // SCMemProjNodes represent the memory state of CAS. Their main
2533 // role is to prevent CAS nodes from being optimized away when their
2534 // results aren't used.
2535 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2536 set_memory(proj, alias_idx);
2538 // Add the trailing membar surrounding the access
2539 insert_mem_bar(Op_MemBarCPUOrder);
2540 insert_mem_bar(Op_MemBarAcquire);
2542 push(cas);
2543 return true;
2544 }
2546 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2547 // This is another variant of inline_unsafe_access, differing in
2548 // that it always issues store-store ("release") barrier and ensures
2549 // store-atomicity (which only matters for "long").
2551 if (callee()->is_static()) return false; // caller must have the capability!
2553 #ifndef PRODUCT
2554 {
2555 ResourceMark rm;
2556 // Check the signatures.
2557 ciSignature* sig = signature();
2558 #ifdef ASSERT
2559 BasicType rtype = sig->return_type()->basic_type();
2560 assert(rtype == T_VOID, "must return void");
2561 assert(sig->count() == 3, "has 3 arguments");
2562 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2563 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2564 #endif // ASSERT
2565 }
2566 #endif //PRODUCT
2568 // number of stack slots per value argument (1 or 2)
2569 int type_words = type2size[type];
2571 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2573 // Argument words: "this" plus oop plus offset plus value;
2574 int nargs = 1 + 1 + 2 + type_words;
2576 // pop arguments: val, offset, base, and receiver
2577 debug_only(int saved_sp = _sp);
2578 _sp += nargs;
2579 Node* val = (type_words == 1) ? pop() : pop_pair();
2580 Node *offset = pop_pair();
2581 Node *base = pop();
2582 Node *receiver = pop();
2583 assert(saved_sp == _sp, "must have correct argument count");
2585 // Null check receiver.
2586 _sp += nargs;
2587 do_null_check(receiver, T_OBJECT);
2588 _sp -= nargs;
2589 if (stopped()) {
2590 return true;
2591 }
2593 // Build field offset expression.
2594 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2595 // 32-bit machines ignore the high half of long offsets
2596 offset = ConvL2X(offset);
2597 Node* adr = make_unsafe_address(base, offset);
2598 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2599 const Type *value_type = Type::get_const_basic_type(type);
2600 Compile::AliasType* alias_type = C->alias_type(adr_type);
2602 insert_mem_bar(Op_MemBarRelease);
2603 insert_mem_bar(Op_MemBarCPUOrder);
2604 // Ensure that the store is atomic for longs:
2605 bool require_atomic_access = true;
2606 Node* store;
2607 if (type == T_OBJECT) // reference stores need a store barrier.
2608 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2609 else {
2610 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2611 }
2612 insert_mem_bar(Op_MemBarCPUOrder);
2613 return true;
2614 }
2616 bool LibraryCallKit::inline_unsafe_allocate() {
2617 if (callee()->is_static()) return false; // caller must have the capability!
2618 int nargs = 1 + 1;
2619 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2620 null_check_receiver(callee()); // check then ignore argument(0)
2621 _sp += nargs; // set original stack for use by uncommon_trap
2622 Node* cls = do_null_check(argument(1), T_OBJECT);
2623 _sp -= nargs;
2624 if (stopped()) return true;
2626 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2627 _sp += nargs; // set original stack for use by uncommon_trap
2628 kls = do_null_check(kls, T_OBJECT);
2629 _sp -= nargs;
2630 if (stopped()) return true; // argument was like int.class
2632 // Note: The argument might still be an illegal value like
2633 // Serializable.class or Object[].class. The runtime will handle it.
2634 // But we must make an explicit check for initialization.
2635 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2636 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2637 Node* bits = intcon(instanceKlass::fully_initialized);
2638 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2639 // The 'test' is non-zero if we need to take a slow path.
2641 Node* obj = new_instance(kls, test);
2642 push(obj);
2644 return true;
2645 }
2647 //------------------------inline_native_time_funcs--------------
2648 // inline code for System.currentTimeMillis() and System.nanoTime()
2649 // these have the same type and signature
2650 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2651 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2652 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2653 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2654 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2655 const TypePtr* no_memory_effects = NULL;
2656 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2657 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2658 #ifdef ASSERT
2659 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2660 assert(value_top == top(), "second value must be top");
2661 #endif
2662 push_pair(value);
2663 return true;
2664 }
2666 //------------------------inline_native_currentThread------------------
2667 bool LibraryCallKit::inline_native_currentThread() {
2668 Node* junk = NULL;
2669 push(generate_current_thread(junk));
2670 return true;
2671 }
2673 //------------------------inline_native_isInterrupted------------------
2674 bool LibraryCallKit::inline_native_isInterrupted() {
2675 const int nargs = 1+1; // receiver + boolean
2676 assert(nargs == arg_size(), "sanity");
2677 // Add a fast path to t.isInterrupted(clear_int):
2678 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2679 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2680 // So, in the common case that the interrupt bit is false,
2681 // we avoid making a call into the VM. Even if the interrupt bit
2682 // is true, if the clear_int argument is false, we avoid the VM call.
2683 // However, if the receiver is not currentThread, we must call the VM,
2684 // because there must be some locking done around the operation.
2686 // We only go to the fast case code if we pass two guards.
2687 // Paths which do not pass are accumulated in the slow_region.
2688 RegionNode* slow_region = new (C, 1) RegionNode(1);
2689 record_for_igvn(slow_region);
2690 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2691 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2692 enum { no_int_result_path = 1,
2693 no_clear_result_path = 2,
2694 slow_result_path = 3
2695 };
2697 // (a) Receiving thread must be the current thread.
2698 Node* rec_thr = argument(0);
2699 Node* tls_ptr = NULL;
2700 Node* cur_thr = generate_current_thread(tls_ptr);
2701 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2702 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2704 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2705 if (!known_current_thread)
2706 generate_slow_guard(bol_thr, slow_region);
2708 // (b) Interrupt bit on TLS must be false.
2709 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2710 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2711 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2712 // Set the control input on the field _interrupted read to prevent it floating up.
2713 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2714 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2715 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2717 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2719 // First fast path: if (!TLS._interrupted) return false;
2720 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2721 result_rgn->init_req(no_int_result_path, false_bit);
2722 result_val->init_req(no_int_result_path, intcon(0));
2724 // drop through to next case
2725 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2727 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2728 Node* clr_arg = argument(1);
2729 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2730 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2731 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2733 // Second fast path: ... else if (!clear_int) return true;
2734 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2735 result_rgn->init_req(no_clear_result_path, false_arg);
2736 result_val->init_req(no_clear_result_path, intcon(1));
2738 // drop through to next case
2739 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2741 // (d) Otherwise, go to the slow path.
2742 slow_region->add_req(control());
2743 set_control( _gvn.transform(slow_region) );
2745 if (stopped()) {
2746 // There is no slow path.
2747 result_rgn->init_req(slow_result_path, top());
2748 result_val->init_req(slow_result_path, top());
2749 } else {
2750 // non-virtual because it is a private non-static
2751 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2753 Node* slow_val = set_results_for_java_call(slow_call);
2754 // this->control() comes from set_results_for_java_call
2756 // If we know that the result of the slow call will be true, tell the optimizer!
2757 if (known_current_thread) slow_val = intcon(1);
2759 Node* fast_io = slow_call->in(TypeFunc::I_O);
2760 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2761 // These two phis are pre-filled with copies of of the fast IO and Memory
2762 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2763 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2765 result_rgn->init_req(slow_result_path, control());
2766 io_phi ->init_req(slow_result_path, i_o());
2767 mem_phi ->init_req(slow_result_path, reset_memory());
2768 result_val->init_req(slow_result_path, slow_val);
2770 set_all_memory( _gvn.transform(mem_phi) );
2771 set_i_o( _gvn.transform(io_phi) );
2772 }
2774 push_result(result_rgn, result_val);
2775 C->set_has_split_ifs(true); // Has chance for split-if optimization
2777 return true;
2778 }
2780 //---------------------------load_mirror_from_klass----------------------------
2781 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2782 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2783 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2784 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2785 }
2787 //-----------------------load_klass_from_mirror_common-------------------------
2788 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2789 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2790 // and branch to the given path on the region.
2791 // If never_see_null, take an uncommon trap on null, so we can optimistically
2792 // compile for the non-null case.
2793 // If the region is NULL, force never_see_null = true.
2794 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2795 bool never_see_null,
2796 int nargs,
2797 RegionNode* region,
2798 int null_path,
2799 int offset) {
2800 if (region == NULL) never_see_null = true;
2801 Node* p = basic_plus_adr(mirror, offset);
2802 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2803 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2804 _sp += nargs; // any deopt will start just before call to enclosing method
2805 Node* null_ctl = top();
2806 kls = null_check_oop(kls, &null_ctl, never_see_null);
2807 if (region != NULL) {
2808 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2809 region->init_req(null_path, null_ctl);
2810 } else {
2811 assert(null_ctl == top(), "no loose ends");
2812 }
2813 _sp -= nargs;
2814 return kls;
2815 }
2817 //--------------------(inline_native_Class_query helpers)---------------------
2818 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2819 // Fall through if (mods & mask) == bits, take the guard otherwise.
2820 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2821 // Branch around if the given klass has the given modifier bit set.
2822 // Like generate_guard, adds a new path onto the region.
2823 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2824 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2825 Node* mask = intcon(modifier_mask);
2826 Node* bits = intcon(modifier_bits);
2827 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2828 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2829 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2830 return generate_fair_guard(bol, region);
2831 }
2832 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2833 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2834 }
2836 //-------------------------inline_native_Class_query-------------------
2837 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2838 int nargs = 1+0; // just the Class mirror, in most cases
2839 const Type* return_type = TypeInt::BOOL;
2840 Node* prim_return_value = top(); // what happens if it's a primitive class?
2841 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2842 bool expect_prim = false; // most of these guys expect to work on refs
2844 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2846 switch (id) {
2847 case vmIntrinsics::_isInstance:
2848 nargs = 1+1; // the Class mirror, plus the object getting queried about
2849 // nothing is an instance of a primitive type
2850 prim_return_value = intcon(0);
2851 break;
2852 case vmIntrinsics::_getModifiers:
2853 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2854 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2855 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2856 break;
2857 case vmIntrinsics::_isInterface:
2858 prim_return_value = intcon(0);
2859 break;
2860 case vmIntrinsics::_isArray:
2861 prim_return_value = intcon(0);
2862 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2863 break;
2864 case vmIntrinsics::_isPrimitive:
2865 prim_return_value = intcon(1);
2866 expect_prim = true; // obviously
2867 break;
2868 case vmIntrinsics::_getSuperclass:
2869 prim_return_value = null();
2870 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2871 break;
2872 case vmIntrinsics::_getComponentType:
2873 prim_return_value = null();
2874 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2875 break;
2876 case vmIntrinsics::_getClassAccessFlags:
2877 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2878 return_type = TypeInt::INT; // not bool! 6297094
2879 break;
2880 default:
2881 ShouldNotReachHere();
2882 }
2884 Node* mirror = argument(0);
2885 Node* obj = (nargs <= 1)? top(): argument(1);
2887 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2888 if (mirror_con == NULL) return false; // cannot happen?
2890 #ifndef PRODUCT
2891 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2892 ciType* k = mirror_con->java_mirror_type();
2893 if (k) {
2894 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2895 k->print_name();
2896 tty->cr();
2897 }
2898 }
2899 #endif
2901 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2902 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2903 record_for_igvn(region);
2904 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2906 // The mirror will never be null of Reflection.getClassAccessFlags, however
2907 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2908 // if it is. See bug 4774291.
2910 // For Reflection.getClassAccessFlags(), the null check occurs in
2911 // the wrong place; see inline_unsafe_access(), above, for a similar
2912 // situation.
2913 _sp += nargs; // set original stack for use by uncommon_trap
2914 mirror = do_null_check(mirror, T_OBJECT);
2915 _sp -= nargs;
2916 // If mirror or obj is dead, only null-path is taken.
2917 if (stopped()) return true;
2919 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2921 // Now load the mirror's klass metaobject, and null-check it.
2922 // Side-effects region with the control path if the klass is null.
2923 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2924 region, _prim_path);
2925 // If kls is null, we have a primitive mirror.
2926 phi->init_req(_prim_path, prim_return_value);
2927 if (stopped()) { push_result(region, phi); return true; }
2929 Node* p; // handy temp
2930 Node* null_ctl;
2932 // Now that we have the non-null klass, we can perform the real query.
2933 // For constant classes, the query will constant-fold in LoadNode::Value.
2934 Node* query_value = top();
2935 switch (id) {
2936 case vmIntrinsics::_isInstance:
2937 // nothing is an instance of a primitive type
2938 query_value = gen_instanceof(obj, kls);
2939 break;
2941 case vmIntrinsics::_getModifiers:
2942 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2943 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2944 break;
2946 case vmIntrinsics::_isInterface:
2947 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2948 if (generate_interface_guard(kls, region) != NULL)
2949 // A guard was added. If the guard is taken, it was an interface.
2950 phi->add_req(intcon(1));
2951 // If we fall through, it's a plain class.
2952 query_value = intcon(0);
2953 break;
2955 case vmIntrinsics::_isArray:
2956 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2957 if (generate_array_guard(kls, region) != NULL)
2958 // A guard was added. If the guard is taken, it was an array.
2959 phi->add_req(intcon(1));
2960 // If we fall through, it's a plain class.
2961 query_value = intcon(0);
2962 break;
2964 case vmIntrinsics::_isPrimitive:
2965 query_value = intcon(0); // "normal" path produces false
2966 break;
2968 case vmIntrinsics::_getSuperclass:
2969 // The rules here are somewhat unfortunate, but we can still do better
2970 // with random logic than with a JNI call.
2971 // Interfaces store null or Object as _super, but must report null.
2972 // Arrays store an intermediate super as _super, but must report Object.
2973 // Other types can report the actual _super.
2974 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2975 if (generate_interface_guard(kls, region) != NULL)
2976 // A guard was added. If the guard is taken, it was an interface.
2977 phi->add_req(null());
2978 if (generate_array_guard(kls, region) != NULL)
2979 // A guard was added. If the guard is taken, it was an array.
2980 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2981 // If we fall through, it's a plain class. Get its _super.
2982 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2983 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2984 null_ctl = top();
2985 kls = null_check_oop(kls, &null_ctl);
2986 if (null_ctl != top()) {
2987 // If the guard is taken, Object.superClass is null (both klass and mirror).
2988 region->add_req(null_ctl);
2989 phi ->add_req(null());
2990 }
2991 if (!stopped()) {
2992 query_value = load_mirror_from_klass(kls);
2993 }
2994 break;
2996 case vmIntrinsics::_getComponentType:
2997 if (generate_array_guard(kls, region) != NULL) {
2998 // Be sure to pin the oop load to the guard edge just created:
2999 Node* is_array_ctrl = region->in(region->req()-1);
3000 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3001 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3002 phi->add_req(cmo);
3003 }
3004 query_value = null(); // non-array case is null
3005 break;
3007 case vmIntrinsics::_getClassAccessFlags:
3008 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3009 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3010 break;
3012 default:
3013 ShouldNotReachHere();
3014 }
3016 // Fall-through is the normal case of a query to a real class.
3017 phi->init_req(1, query_value);
3018 region->init_req(1, control());
3020 push_result(region, phi);
3021 C->set_has_split_ifs(true); // Has chance for split-if optimization
3023 return true;
3024 }
3026 //--------------------------inline_native_subtype_check------------------------
3027 // This intrinsic takes the JNI calls out of the heart of
3028 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3029 bool LibraryCallKit::inline_native_subtype_check() {
3030 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3032 // Pull both arguments off the stack.
3033 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3034 args[0] = argument(0);
3035 args[1] = argument(1);
3036 Node* klasses[2]; // corresponding Klasses: superk, subk
3037 klasses[0] = klasses[1] = top();
3039 enum {
3040 // A full decision tree on {superc is prim, subc is prim}:
3041 _prim_0_path = 1, // {P,N} => false
3042 // {P,P} & superc!=subc => false
3043 _prim_same_path, // {P,P} & superc==subc => true
3044 _prim_1_path, // {N,P} => false
3045 _ref_subtype_path, // {N,N} & subtype check wins => true
3046 _both_ref_path, // {N,N} & subtype check loses => false
3047 PATH_LIMIT
3048 };
3050 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3051 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3052 record_for_igvn(region);
3054 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3055 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3056 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3058 // First null-check both mirrors and load each mirror's klass metaobject.
3059 int which_arg;
3060 for (which_arg = 0; which_arg <= 1; which_arg++) {
3061 Node* arg = args[which_arg];
3062 _sp += nargs; // set original stack for use by uncommon_trap
3063 arg = do_null_check(arg, T_OBJECT);
3064 _sp -= nargs;
3065 if (stopped()) break;
3066 args[which_arg] = _gvn.transform(arg);
3068 Node* p = basic_plus_adr(arg, class_klass_offset);
3069 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3070 klasses[which_arg] = _gvn.transform(kls);
3071 }
3073 // Having loaded both klasses, test each for null.
3074 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3075 for (which_arg = 0; which_arg <= 1; which_arg++) {
3076 Node* kls = klasses[which_arg];
3077 Node* null_ctl = top();
3078 _sp += nargs; // set original stack for use by uncommon_trap
3079 kls = null_check_oop(kls, &null_ctl, never_see_null);
3080 _sp -= nargs;
3081 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3082 region->init_req(prim_path, null_ctl);
3083 if (stopped()) break;
3084 klasses[which_arg] = kls;
3085 }
3087 if (!stopped()) {
3088 // now we have two reference types, in klasses[0..1]
3089 Node* subk = klasses[1]; // the argument to isAssignableFrom
3090 Node* superk = klasses[0]; // the receiver
3091 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3092 // now we have a successful reference subtype check
3093 region->set_req(_ref_subtype_path, control());
3094 }
3096 // If both operands are primitive (both klasses null), then
3097 // we must return true when they are identical primitives.
3098 // It is convenient to test this after the first null klass check.
3099 set_control(region->in(_prim_0_path)); // go back to first null check
3100 if (!stopped()) {
3101 // Since superc is primitive, make a guard for the superc==subc case.
3102 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3103 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3104 generate_guard(bol_eq, region, PROB_FAIR);
3105 if (region->req() == PATH_LIMIT+1) {
3106 // A guard was added. If the added guard is taken, superc==subc.
3107 region->swap_edges(PATH_LIMIT, _prim_same_path);
3108 region->del_req(PATH_LIMIT);
3109 }
3110 region->set_req(_prim_0_path, control()); // Not equal after all.
3111 }
3113 // these are the only paths that produce 'true':
3114 phi->set_req(_prim_same_path, intcon(1));
3115 phi->set_req(_ref_subtype_path, intcon(1));
3117 // pull together the cases:
3118 assert(region->req() == PATH_LIMIT, "sane region");
3119 for (uint i = 1; i < region->req(); i++) {
3120 Node* ctl = region->in(i);
3121 if (ctl == NULL || ctl == top()) {
3122 region->set_req(i, top());
3123 phi ->set_req(i, top());
3124 } else if (phi->in(i) == NULL) {
3125 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3126 }
3127 }
3129 set_control(_gvn.transform(region));
3130 push(_gvn.transform(phi));
3132 return true;
3133 }
3135 //---------------------generate_array_guard_common------------------------
3136 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3137 bool obj_array, bool not_array) {
3138 // If obj_array/non_array==false/false:
3139 // Branch around if the given klass is in fact an array (either obj or prim).
3140 // If obj_array/non_array==false/true:
3141 // Branch around if the given klass is not an array klass of any kind.
3142 // If obj_array/non_array==true/true:
3143 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3144 // If obj_array/non_array==true/false:
3145 // Branch around if the kls is an oop array (Object[] or subtype)
3146 //
3147 // Like generate_guard, adds a new path onto the region.
3148 jint layout_con = 0;
3149 Node* layout_val = get_layout_helper(kls, layout_con);
3150 if (layout_val == NULL) {
3151 bool query = (obj_array
3152 ? Klass::layout_helper_is_objArray(layout_con)
3153 : Klass::layout_helper_is_javaArray(layout_con));
3154 if (query == not_array) {
3155 return NULL; // never a branch
3156 } else { // always a branch
3157 Node* always_branch = control();
3158 if (region != NULL)
3159 region->add_req(always_branch);
3160 set_control(top());
3161 return always_branch;
3162 }
3163 }
3164 // Now test the correct condition.
3165 jint nval = (obj_array
3166 ? ((jint)Klass::_lh_array_tag_type_value
3167 << Klass::_lh_array_tag_shift)
3168 : Klass::_lh_neutral_value);
3169 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3170 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3171 // invert the test if we are looking for a non-array
3172 if (not_array) btest = BoolTest(btest).negate();
3173 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3174 return generate_fair_guard(bol, region);
3175 }
3178 //-----------------------inline_native_newArray--------------------------
3179 bool LibraryCallKit::inline_native_newArray() {
3180 int nargs = 2;
3181 Node* mirror = argument(0);
3182 Node* count_val = argument(1);
3184 _sp += nargs; // set original stack for use by uncommon_trap
3185 mirror = do_null_check(mirror, T_OBJECT);
3186 _sp -= nargs;
3187 // If mirror or obj is dead, only null-path is taken.
3188 if (stopped()) return true;
3190 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3191 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3192 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3193 TypeInstPtr::NOTNULL);
3194 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3195 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3196 TypePtr::BOTTOM);
3198 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3199 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3200 nargs,
3201 result_reg, _slow_path);
3202 Node* normal_ctl = control();
3203 Node* no_array_ctl = result_reg->in(_slow_path);
3205 // Generate code for the slow case. We make a call to newArray().
3206 set_control(no_array_ctl);
3207 if (!stopped()) {
3208 // Either the input type is void.class, or else the
3209 // array klass has not yet been cached. Either the
3210 // ensuing call will throw an exception, or else it
3211 // will cache the array klass for next time.
3212 PreserveJVMState pjvms(this);
3213 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3214 Node* slow_result = set_results_for_java_call(slow_call);
3215 // this->control() comes from set_results_for_java_call
3216 result_reg->set_req(_slow_path, control());
3217 result_val->set_req(_slow_path, slow_result);
3218 result_io ->set_req(_slow_path, i_o());
3219 result_mem->set_req(_slow_path, reset_memory());
3220 }
3222 set_control(normal_ctl);
3223 if (!stopped()) {
3224 // Normal case: The array type has been cached in the java.lang.Class.
3225 // The following call works fine even if the array type is polymorphic.
3226 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3227 Node* obj = new_array(klass_node, count_val, nargs);
3228 result_reg->init_req(_normal_path, control());
3229 result_val->init_req(_normal_path, obj);
3230 result_io ->init_req(_normal_path, i_o());
3231 result_mem->init_req(_normal_path, reset_memory());
3232 }
3234 // Return the combined state.
3235 set_i_o( _gvn.transform(result_io) );
3236 set_all_memory( _gvn.transform(result_mem) );
3237 push_result(result_reg, result_val);
3238 C->set_has_split_ifs(true); // Has chance for split-if optimization
3240 return true;
3241 }
3243 //----------------------inline_native_getLength--------------------------
3244 bool LibraryCallKit::inline_native_getLength() {
3245 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3247 int nargs = 1;
3248 Node* array = argument(0);
3250 _sp += nargs; // set original stack for use by uncommon_trap
3251 array = do_null_check(array, T_OBJECT);
3252 _sp -= nargs;
3254 // If array is dead, only null-path is taken.
3255 if (stopped()) return true;
3257 // Deoptimize if it is a non-array.
3258 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3260 if (non_array != NULL) {
3261 PreserveJVMState pjvms(this);
3262 set_control(non_array);
3263 _sp += nargs; // push the arguments back on the stack
3264 uncommon_trap(Deoptimization::Reason_intrinsic,
3265 Deoptimization::Action_maybe_recompile);
3266 }
3268 // If control is dead, only non-array-path is taken.
3269 if (stopped()) return true;
3271 // The works fine even if the array type is polymorphic.
3272 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3273 push( load_array_length(array) );
3275 C->set_has_split_ifs(true); // Has chance for split-if optimization
3277 return true;
3278 }
3280 //------------------------inline_array_copyOf----------------------------
3281 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3282 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3284 // Restore the stack and pop off the arguments.
3285 int nargs = 3 + (is_copyOfRange? 1: 0);
3286 Node* original = argument(0);
3287 Node* start = is_copyOfRange? argument(1): intcon(0);
3288 Node* end = is_copyOfRange? argument(2): argument(1);
3289 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3291 Node* newcopy;
3293 //set the original stack and the reexecute bit for the interpreter to reexecute
3294 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3295 { PreserveReexecuteState preexecs(this);
3296 _sp += nargs;
3297 jvms()->set_should_reexecute(true);
3299 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3300 original = do_null_check(original, T_OBJECT);
3302 // Check if a null path was taken unconditionally.
3303 if (stopped()) return true;
3305 Node* orig_length = load_array_length(original);
3307 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3308 NULL, 0);
3309 klass_node = do_null_check(klass_node, T_OBJECT);
3311 RegionNode* bailout = new (C, 1) RegionNode(1);
3312 record_for_igvn(bailout);
3314 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3315 // Bail out if that is so.
3316 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3317 if (not_objArray != NULL) {
3318 // Improve the klass node's type from the new optimistic assumption:
3319 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3320 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3321 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3322 cast->init_req(0, control());
3323 klass_node = _gvn.transform(cast);
3324 }
3326 // Bail out if either start or end is negative.
3327 generate_negative_guard(start, bailout, &start);
3328 generate_negative_guard(end, bailout, &end);
3330 Node* length = end;
3331 if (_gvn.type(start) != TypeInt::ZERO) {
3332 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3333 }
3335 // Bail out if length is negative.
3336 // ...Not needed, since the new_array will throw the right exception.
3337 //generate_negative_guard(length, bailout, &length);
3339 if (bailout->req() > 1) {
3340 PreserveJVMState pjvms(this);
3341 set_control( _gvn.transform(bailout) );
3342 uncommon_trap(Deoptimization::Reason_intrinsic,
3343 Deoptimization::Action_maybe_recompile);
3344 }
3346 if (!stopped()) {
3348 // How many elements will we copy from the original?
3349 // The answer is MinI(orig_length - start, length).
3350 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3351 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3353 const bool raw_mem_only = true;
3354 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3356 // Generate a direct call to the right arraycopy function(s).
3357 // We know the copy is disjoint but we might not know if the
3358 // oop stores need checking.
3359 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3360 // This will fail a store-check if x contains any non-nulls.
3361 bool disjoint_bases = true;
3362 bool length_never_negative = true;
3363 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3364 original, start, newcopy, intcon(0), moved,
3365 disjoint_bases, length_never_negative);
3366 }
3367 } //original reexecute and sp are set back here
3369 if(!stopped()) {
3370 push(newcopy);
3371 }
3373 C->set_has_split_ifs(true); // Has chance for split-if optimization
3375 return true;
3376 }
3379 //----------------------generate_virtual_guard---------------------------
3380 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3381 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3382 RegionNode* slow_region) {
3383 ciMethod* method = callee();
3384 int vtable_index = method->vtable_index();
3385 // Get the methodOop out of the appropriate vtable entry.
3386 int entry_offset = (instanceKlass::vtable_start_offset() +
3387 vtable_index*vtableEntry::size()) * wordSize +
3388 vtableEntry::method_offset_in_bytes();
3389 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3390 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3392 // Compare the target method with the expected method (e.g., Object.hashCode).
3393 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3395 Node* native_call = makecon(native_call_addr);
3396 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3397 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3399 return generate_slow_guard(test_native, slow_region);
3400 }
3402 //-----------------------generate_method_call----------------------------
3403 // Use generate_method_call to make a slow-call to the real
3404 // method if the fast path fails. An alternative would be to
3405 // use a stub like OptoRuntime::slow_arraycopy_Java.
3406 // This only works for expanding the current library call,
3407 // not another intrinsic. (E.g., don't use this for making an
3408 // arraycopy call inside of the copyOf intrinsic.)
3409 CallJavaNode*
3410 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3411 // When compiling the intrinsic method itself, do not use this technique.
3412 guarantee(callee() != C->method(), "cannot make slow-call to self");
3414 ciMethod* method = callee();
3415 // ensure the JVMS we have will be correct for this call
3416 guarantee(method_id == method->intrinsic_id(), "must match");
3418 const TypeFunc* tf = TypeFunc::make(method);
3419 int tfdc = tf->domain()->cnt();
3420 CallJavaNode* slow_call;
3421 if (is_static) {
3422 assert(!is_virtual, "");
3423 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3424 SharedRuntime::get_resolve_static_call_stub(),
3425 method, bci());
3426 } else if (is_virtual) {
3427 null_check_receiver(method);
3428 int vtable_index = methodOopDesc::invalid_vtable_index;
3429 if (UseInlineCaches) {
3430 // Suppress the vtable call
3431 } else {
3432 // hashCode and clone are not a miranda methods,
3433 // so the vtable index is fixed.
3434 // No need to use the linkResolver to get it.
3435 vtable_index = method->vtable_index();
3436 }
3437 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3438 SharedRuntime::get_resolve_virtual_call_stub(),
3439 method, vtable_index, bci());
3440 } else { // neither virtual nor static: opt_virtual
3441 null_check_receiver(method);
3442 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3443 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3444 method, bci());
3445 slow_call->set_optimized_virtual(true);
3446 }
3447 set_arguments_for_java_call(slow_call);
3448 set_edges_for_java_call(slow_call);
3449 return slow_call;
3450 }
3453 //------------------------------inline_native_hashcode--------------------
3454 // Build special case code for calls to hashCode on an object.
3455 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3456 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3457 assert(!(is_virtual && is_static), "either virtual, special, or static");
3459 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3461 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3462 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3463 TypeInt::INT);
3464 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3465 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3466 TypePtr::BOTTOM);
3467 Node* obj = NULL;
3468 if (!is_static) {
3469 // Check for hashing null object
3470 obj = null_check_receiver(callee());
3471 if (stopped()) return true; // unconditionally null
3472 result_reg->init_req(_null_path, top());
3473 result_val->init_req(_null_path, top());
3474 } else {
3475 // Do a null check, and return zero if null.
3476 // System.identityHashCode(null) == 0
3477 obj = argument(0);
3478 Node* null_ctl = top();
3479 obj = null_check_oop(obj, &null_ctl);
3480 result_reg->init_req(_null_path, null_ctl);
3481 result_val->init_req(_null_path, _gvn.intcon(0));
3482 }
3484 // Unconditionally null? Then return right away.
3485 if (stopped()) {
3486 set_control( result_reg->in(_null_path) );
3487 if (!stopped())
3488 push( result_val ->in(_null_path) );
3489 return true;
3490 }
3492 // After null check, get the object's klass.
3493 Node* obj_klass = load_object_klass(obj);
3495 // This call may be virtual (invokevirtual) or bound (invokespecial).
3496 // For each case we generate slightly different code.
3498 // We only go to the fast case code if we pass a number of guards. The
3499 // paths which do not pass are accumulated in the slow_region.
3500 RegionNode* slow_region = new (C, 1) RegionNode(1);
3501 record_for_igvn(slow_region);
3503 // If this is a virtual call, we generate a funny guard. We pull out
3504 // the vtable entry corresponding to hashCode() from the target object.
3505 // If the target method which we are calling happens to be the native
3506 // Object hashCode() method, we pass the guard. We do not need this
3507 // guard for non-virtual calls -- the caller is known to be the native
3508 // Object hashCode().
3509 if (is_virtual) {
3510 generate_virtual_guard(obj_klass, slow_region);
3511 }
3513 // Get the header out of the object, use LoadMarkNode when available
3514 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3515 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
3516 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
3518 // Test the header to see if it is unlocked.
3519 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3520 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3521 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3522 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3523 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3525 generate_slow_guard(test_unlocked, slow_region);
3527 // Get the hash value and check to see that it has been properly assigned.
3528 // We depend on hash_mask being at most 32 bits and avoid the use of
3529 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3530 // vm: see markOop.hpp.
3531 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3532 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3533 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3534 // This hack lets the hash bits live anywhere in the mark object now, as long
3535 // as the shift drops the relevant bits into the low 32 bits. Note that
3536 // Java spec says that HashCode is an int so there's no point in capturing
3537 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3538 hshifted_header = ConvX2I(hshifted_header);
3539 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3541 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3542 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3543 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3545 generate_slow_guard(test_assigned, slow_region);
3547 Node* init_mem = reset_memory();
3548 // fill in the rest of the null path:
3549 result_io ->init_req(_null_path, i_o());
3550 result_mem->init_req(_null_path, init_mem);
3552 result_val->init_req(_fast_path, hash_val);
3553 result_reg->init_req(_fast_path, control());
3554 result_io ->init_req(_fast_path, i_o());
3555 result_mem->init_req(_fast_path, init_mem);
3557 // Generate code for the slow case. We make a call to hashCode().
3558 set_control(_gvn.transform(slow_region));
3559 if (!stopped()) {
3560 // No need for PreserveJVMState, because we're using up the present state.
3561 set_all_memory(init_mem);
3562 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3563 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3564 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3565 Node* slow_result = set_results_for_java_call(slow_call);
3566 // this->control() comes from set_results_for_java_call
3567 result_reg->init_req(_slow_path, control());
3568 result_val->init_req(_slow_path, slow_result);
3569 result_io ->set_req(_slow_path, i_o());
3570 result_mem ->set_req(_slow_path, reset_memory());
3571 }
3573 // Return the combined state.
3574 set_i_o( _gvn.transform(result_io) );
3575 set_all_memory( _gvn.transform(result_mem) );
3576 push_result(result_reg, result_val);
3578 return true;
3579 }
3581 //---------------------------inline_native_getClass----------------------------
3582 // Build special case code for calls to getClass on an object.
3583 bool LibraryCallKit::inline_native_getClass() {
3584 Node* obj = null_check_receiver(callee());
3585 if (stopped()) return true;
3586 push( load_mirror_from_klass(load_object_klass(obj)) );
3587 return true;
3588 }
3590 //-----------------inline_native_Reflection_getCallerClass---------------------
3591 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3592 //
3593 // NOTE that this code must perform the same logic as
3594 // vframeStream::security_get_caller_frame in that it must skip
3595 // Method.invoke() and auxiliary frames.
3600 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3601 ciMethod* method = callee();
3603 #ifndef PRODUCT
3604 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3605 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3606 }
3607 #endif
3609 debug_only(int saved_sp = _sp);
3611 // Argument words: (int depth)
3612 int nargs = 1;
3614 _sp += nargs;
3615 Node* caller_depth_node = pop();
3617 assert(saved_sp == _sp, "must have correct argument count");
3619 // The depth value must be a constant in order for the runtime call
3620 // to be eliminated.
3621 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3622 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3623 #ifndef PRODUCT
3624 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3625 tty->print_cr(" Bailing out because caller depth was not a constant");
3626 }
3627 #endif
3628 return false;
3629 }
3630 // Note that the JVM state at this point does not include the
3631 // getCallerClass() frame which we are trying to inline. The
3632 // semantics of getCallerClass(), however, are that the "first"
3633 // frame is the getCallerClass() frame, so we subtract one from the
3634 // requested depth before continuing. We don't inline requests of
3635 // getCallerClass(0).
3636 int caller_depth = caller_depth_type->get_con() - 1;
3637 if (caller_depth < 0) {
3638 #ifndef PRODUCT
3639 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3640 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3641 }
3642 #endif
3643 return false;
3644 }
3646 if (!jvms()->has_method()) {
3647 #ifndef PRODUCT
3648 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3649 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3650 }
3651 #endif
3652 return false;
3653 }
3654 int _depth = jvms()->depth(); // cache call chain depth
3656 // Walk back up the JVM state to find the caller at the required
3657 // depth. NOTE that this code must perform the same logic as
3658 // vframeStream::security_get_caller_frame in that it must skip
3659 // Method.invoke() and auxiliary frames. Note also that depth is
3660 // 1-based (1 is the bottom of the inlining).
3661 int inlining_depth = _depth;
3662 JVMState* caller_jvms = NULL;
3664 if (inlining_depth > 0) {
3665 caller_jvms = jvms();
3666 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3667 do {
3668 // The following if-tests should be performed in this order
3669 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3670 // Skip a Method.invoke() or auxiliary frame
3671 } else if (caller_depth > 0) {
3672 // Skip real frame
3673 --caller_depth;
3674 } else {
3675 // We're done: reached desired caller after skipping.
3676 break;
3677 }
3678 caller_jvms = caller_jvms->caller();
3679 --inlining_depth;
3680 } while (inlining_depth > 0);
3681 }
3683 if (inlining_depth == 0) {
3684 #ifndef PRODUCT
3685 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3686 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3687 tty->print_cr(" JVM state at this point:");
3688 for (int i = _depth; i >= 1; i--) {
3689 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3690 }
3691 }
3692 #endif
3693 return false; // Reached end of inlining
3694 }
3696 // Acquire method holder as java.lang.Class
3697 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3698 ciInstance* caller_mirror = caller_klass->java_mirror();
3699 // Push this as a constant
3700 push(makecon(TypeInstPtr::make(caller_mirror)));
3701 #ifndef PRODUCT
3702 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3703 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);
3704 tty->print_cr(" JVM state at this point:");
3705 for (int i = _depth; i >= 1; i--) {
3706 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3707 }
3708 }
3709 #endif
3710 return true;
3711 }
3713 // Helper routine for above
3714 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3715 ciMethod* method = jvms->method();
3717 // Is this the Method.invoke method itself?
3718 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3719 return true;
3721 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3722 ciKlass* k = method->holder();
3723 if (k->is_instance_klass()) {
3724 ciInstanceKlass* ik = k->as_instance_klass();
3725 for (; ik != NULL; ik = ik->super()) {
3726 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3727 ik == env()->find_system_klass(ik->name())) {
3728 return true;
3729 }
3730 }
3731 }
3732 else if (method->is_method_handle_adapter()) {
3733 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3734 return true;
3735 }
3737 return false;
3738 }
3740 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3741 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3742 // computing it since there is no lookup field by name function in the
3743 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3744 // Using a static variable here is safe even if we have multiple compilation
3745 // threads because the offset is constant. At worst the same offset will be
3746 // computed and stored multiple
3748 bool LibraryCallKit::inline_native_AtomicLong_get() {
3749 // Restore the stack and pop off the argument
3750 _sp+=1;
3751 Node *obj = pop();
3753 // get the offset of the "value" field. Since the CI interfaces
3754 // does not provide a way to look up a field by name, we scan the bytecodes
3755 // to get the field index. We expect the first 2 instructions of the method
3756 // to be:
3757 // 0 aload_0
3758 // 1 getfield "value"
3759 ciMethod* method = callee();
3760 if (value_field_offset == -1)
3761 {
3762 ciField* value_field;
3763 ciBytecodeStream iter(method);
3764 Bytecodes::Code bc = iter.next();
3766 if ((bc != Bytecodes::_aload_0) &&
3767 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3768 return false;
3769 bc = iter.next();
3770 if (bc != Bytecodes::_getfield)
3771 return false;
3772 bool ignore;
3773 value_field = iter.get_field(ignore);
3774 value_field_offset = value_field->offset_in_bytes();
3775 }
3777 // Null check without removing any arguments.
3778 _sp++;
3779 obj = do_null_check(obj, T_OBJECT);
3780 _sp--;
3781 // Check for locking null object
3782 if (stopped()) return true;
3784 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3785 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3786 int alias_idx = C->get_alias_index(adr_type);
3788 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3790 push_pair(result);
3792 return true;
3793 }
3795 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3796 // Restore the stack and pop off the arguments
3797 _sp+=5;
3798 Node *newVal = pop_pair();
3799 Node *oldVal = pop_pair();
3800 Node *obj = pop();
3802 // we need the offset of the "value" field which was computed when
3803 // inlining the get() method. Give up if we don't have it.
3804 if (value_field_offset == -1)
3805 return false;
3807 // Null check without removing any arguments.
3808 _sp+=5;
3809 obj = do_null_check(obj, T_OBJECT);
3810 _sp-=5;
3811 // Check for locking null object
3812 if (stopped()) return true;
3814 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3815 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3816 int alias_idx = C->get_alias_index(adr_type);
3818 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3819 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3820 set_memory(store_proj, alias_idx);
3821 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3823 Node *result;
3824 // CMove node is not used to be able fold a possible check code
3825 // after attemptUpdate() call. This code could be transformed
3826 // into CMove node by loop optimizations.
3827 {
3828 RegionNode *r = new (C, 3) RegionNode(3);
3829 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3831 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3832 Node *iftrue = opt_iff(r, iff);
3833 r->init_req(1, iftrue);
3834 result->init_req(1, intcon(1));
3835 result->init_req(2, intcon(0));
3837 set_control(_gvn.transform(r));
3838 record_for_igvn(r);
3840 C->set_has_split_ifs(true); // Has chance for split-if optimization
3841 }
3843 push(_gvn.transform(result));
3844 return true;
3845 }
3847 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3848 // restore the arguments
3849 _sp += arg_size();
3851 switch (id) {
3852 case vmIntrinsics::_floatToRawIntBits:
3853 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3854 break;
3856 case vmIntrinsics::_intBitsToFloat:
3857 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3858 break;
3860 case vmIntrinsics::_doubleToRawLongBits:
3861 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3862 break;
3864 case vmIntrinsics::_longBitsToDouble:
3865 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3866 break;
3868 case vmIntrinsics::_doubleToLongBits: {
3869 Node* value = pop_pair();
3871 // two paths (plus control) merge in a wood
3872 RegionNode *r = new (C, 3) RegionNode(3);
3873 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3875 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3876 // Build the boolean node
3877 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3879 // Branch either way.
3880 // NaN case is less traveled, which makes all the difference.
3881 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3882 Node *opt_isnan = _gvn.transform(ifisnan);
3883 assert( opt_isnan->is_If(), "Expect an IfNode");
3884 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3885 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3887 set_control(iftrue);
3889 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3890 Node *slow_result = longcon(nan_bits); // return NaN
3891 phi->init_req(1, _gvn.transform( slow_result ));
3892 r->init_req(1, iftrue);
3894 // Else fall through
3895 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3896 set_control(iffalse);
3898 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3899 r->init_req(2, iffalse);
3901 // Post merge
3902 set_control(_gvn.transform(r));
3903 record_for_igvn(r);
3905 Node* result = _gvn.transform(phi);
3906 assert(result->bottom_type()->isa_long(), "must be");
3907 push_pair(result);
3909 C->set_has_split_ifs(true); // Has chance for split-if optimization
3911 break;
3912 }
3914 case vmIntrinsics::_floatToIntBits: {
3915 Node* value = pop();
3917 // two paths (plus control) merge in a wood
3918 RegionNode *r = new (C, 3) RegionNode(3);
3919 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3921 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3922 // Build the boolean node
3923 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3925 // Branch either way.
3926 // NaN case is less traveled, which makes all the difference.
3927 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3928 Node *opt_isnan = _gvn.transform(ifisnan);
3929 assert( opt_isnan->is_If(), "Expect an IfNode");
3930 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3931 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3933 set_control(iftrue);
3935 static const jint nan_bits = 0x7fc00000;
3936 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3937 phi->init_req(1, _gvn.transform( slow_result ));
3938 r->init_req(1, iftrue);
3940 // Else fall through
3941 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3942 set_control(iffalse);
3944 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3945 r->init_req(2, iffalse);
3947 // Post merge
3948 set_control(_gvn.transform(r));
3949 record_for_igvn(r);
3951 Node* result = _gvn.transform(phi);
3952 assert(result->bottom_type()->isa_int(), "must be");
3953 push(result);
3955 C->set_has_split_ifs(true); // Has chance for split-if optimization
3957 break;
3958 }
3960 default:
3961 ShouldNotReachHere();
3962 }
3964 return true;
3965 }
3967 #ifdef _LP64
3968 #define XTOP ,top() /*additional argument*/
3969 #else //_LP64
3970 #define XTOP /*no additional argument*/
3971 #endif //_LP64
3973 //----------------------inline_unsafe_copyMemory-------------------------
3974 bool LibraryCallKit::inline_unsafe_copyMemory() {
3975 if (callee()->is_static()) return false; // caller must have the capability!
3976 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3977 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3978 null_check_receiver(callee()); // check then ignore argument(0)
3979 if (stopped()) return true;
3981 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3983 Node* src_ptr = argument(1);
3984 Node* src_off = ConvL2X(argument(2));
3985 assert(argument(3)->is_top(), "2nd half of long");
3986 Node* dst_ptr = argument(4);
3987 Node* dst_off = ConvL2X(argument(5));
3988 assert(argument(6)->is_top(), "2nd half of long");
3989 Node* size = ConvL2X(argument(7));
3990 assert(argument(8)->is_top(), "2nd half of long");
3992 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3993 "fieldOffset must be byte-scaled");
3995 Node* src = make_unsafe_address(src_ptr, src_off);
3996 Node* dst = make_unsafe_address(dst_ptr, dst_off);
3998 // Conservatively insert a memory barrier on all memory slices.
3999 // Do not let writes of the copy source or destination float below the copy.
4000 insert_mem_bar(Op_MemBarCPUOrder);
4002 // Call it. Note that the length argument is not scaled.
4003 make_runtime_call(RC_LEAF|RC_NO_FP,
4004 OptoRuntime::fast_arraycopy_Type(),
4005 StubRoutines::unsafe_arraycopy(),
4006 "unsafe_arraycopy",
4007 TypeRawPtr::BOTTOM,
4008 src, dst, size XTOP);
4010 // Do not let reads of the copy destination float above the copy.
4011 insert_mem_bar(Op_MemBarCPUOrder);
4013 return true;
4014 }
4016 //------------------------clone_coping-----------------------------------
4017 // Helper function for inline_native_clone.
4018 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4019 assert(obj_size != NULL, "");
4020 Node* raw_obj = alloc_obj->in(1);
4021 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4023 if (ReduceBulkZeroing) {
4024 // We will be completely responsible for initializing this object -
4025 // mark Initialize node as complete.
4026 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4027 // The object was just allocated - there should be no any stores!
4028 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4029 }
4031 // Copy the fastest available way.
4032 // TODO: generate fields copies for small objects instead.
4033 Node* src = obj;
4034 Node* dest = alloc_obj;
4035 Node* size = _gvn.transform(obj_size);
4037 // Exclude the header but include array length to copy by 8 bytes words.
4038 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4039 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4040 instanceOopDesc::base_offset_in_bytes();
4041 // base_off:
4042 // 8 - 32-bit VM
4043 // 12 - 64-bit VM, compressed oops
4044 // 16 - 64-bit VM, normal oops
4045 if (base_off % BytesPerLong != 0) {
4046 assert(UseCompressedOops, "");
4047 if (is_array) {
4048 // Exclude length to copy by 8 bytes words.
4049 base_off += sizeof(int);
4050 } else {
4051 // Include klass to copy by 8 bytes words.
4052 base_off = instanceOopDesc::klass_offset_in_bytes();
4053 }
4054 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4055 }
4056 src = basic_plus_adr(src, base_off);
4057 dest = basic_plus_adr(dest, base_off);
4059 // Compute the length also, if needed:
4060 Node* countx = size;
4061 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4062 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4064 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4065 bool disjoint_bases = true;
4066 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4067 src, NULL, dest, NULL, countx);
4069 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4070 if (card_mark) {
4071 assert(!is_array, "");
4072 // Put in store barrier for any and all oops we are sticking
4073 // into this object. (We could avoid this if we could prove
4074 // that the object type contains no oop fields at all.)
4075 Node* no_particular_value = NULL;
4076 Node* no_particular_field = NULL;
4077 int raw_adr_idx = Compile::AliasIdxRaw;
4078 post_barrier(control(),
4079 memory(raw_adr_type),
4080 alloc_obj,
4081 no_particular_field,
4082 raw_adr_idx,
4083 no_particular_value,
4084 T_OBJECT,
4085 false);
4086 }
4088 // Do not let reads from the cloned object float above the arraycopy.
4089 insert_mem_bar(Op_MemBarCPUOrder);
4090 }
4092 //------------------------inline_native_clone----------------------------
4093 // Here are the simple edge cases:
4094 // null receiver => normal trap
4095 // virtual and clone was overridden => slow path to out-of-line clone
4096 // not cloneable or finalizer => slow path to out-of-line Object.clone
4097 //
4098 // The general case has two steps, allocation and copying.
4099 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4100 //
4101 // Copying also has two cases, oop arrays and everything else.
4102 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4103 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4104 //
4105 // These steps fold up nicely if and when the cloned object's klass
4106 // can be sharply typed as an object array, a type array, or an instance.
4107 //
4108 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4109 int nargs = 1;
4110 PhiNode* result_val;
4112 //set the original stack and the reexecute bit for the interpreter to reexecute
4113 //the bytecode that invokes Object.clone if deoptimization happens
4114 { PreserveReexecuteState preexecs(this);
4115 jvms()->set_should_reexecute(true);
4117 //null_check_receiver will adjust _sp (push and pop)
4118 Node* obj = null_check_receiver(callee());
4119 if (stopped()) return true;
4121 _sp += nargs;
4123 Node* obj_klass = load_object_klass(obj);
4124 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4125 const TypeOopPtr* toop = ((tklass != NULL)
4126 ? tklass->as_instance_type()
4127 : TypeInstPtr::NOTNULL);
4129 // Conservatively insert a memory barrier on all memory slices.
4130 // Do not let writes into the original float below the clone.
4131 insert_mem_bar(Op_MemBarCPUOrder);
4133 // paths into result_reg:
4134 enum {
4135 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4136 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4137 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4138 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4139 PATH_LIMIT
4140 };
4141 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4142 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4143 TypeInstPtr::NOTNULL);
4144 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4145 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4146 TypePtr::BOTTOM);
4147 record_for_igvn(result_reg);
4149 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4150 int raw_adr_idx = Compile::AliasIdxRaw;
4151 const bool raw_mem_only = true;
4154 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4155 if (array_ctl != NULL) {
4156 // It's an array.
4157 PreserveJVMState pjvms(this);
4158 set_control(array_ctl);
4159 Node* obj_length = load_array_length(obj);
4160 Node* obj_size = NULL;
4161 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4162 raw_mem_only, &obj_size);
4164 if (!use_ReduceInitialCardMarks()) {
4165 // If it is an oop array, it requires very special treatment,
4166 // because card marking is required on each card of the array.
4167 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4168 if (is_obja != NULL) {
4169 PreserveJVMState pjvms2(this);
4170 set_control(is_obja);
4171 // Generate a direct call to the right arraycopy function(s).
4172 bool disjoint_bases = true;
4173 bool length_never_negative = true;
4174 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4175 obj, intcon(0), alloc_obj, intcon(0),
4176 obj_length,
4177 disjoint_bases, length_never_negative);
4178 result_reg->init_req(_objArray_path, control());
4179 result_val->init_req(_objArray_path, alloc_obj);
4180 result_i_o ->set_req(_objArray_path, i_o());
4181 result_mem ->set_req(_objArray_path, reset_memory());
4182 }
4183 }
4184 // Otherwise, there are no card marks to worry about.
4185 // (We can dispense with card marks if we know the allocation
4186 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4187 // causes the non-eden paths to take compensating steps to
4188 // simulate a fresh allocation, so that no further
4189 // card marks are required in compiled code to initialize
4190 // the object.)
4192 if (!stopped()) {
4193 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4195 // Present the results of the copy.
4196 result_reg->init_req(_array_path, control());
4197 result_val->init_req(_array_path, alloc_obj);
4198 result_i_o ->set_req(_array_path, i_o());
4199 result_mem ->set_req(_array_path, reset_memory());
4200 }
4201 }
4203 // We only go to the instance fast case code if we pass a number of guards.
4204 // The paths which do not pass are accumulated in the slow_region.
4205 RegionNode* slow_region = new (C, 1) RegionNode(1);
4206 record_for_igvn(slow_region);
4207 if (!stopped()) {
4208 // It's an instance (we did array above). Make the slow-path tests.
4209 // If this is a virtual call, we generate a funny guard. We grab
4210 // the vtable entry corresponding to clone() from the target object.
4211 // If the target method which we are calling happens to be the
4212 // Object clone() method, we pass the guard. We do not need this
4213 // guard for non-virtual calls; the caller is known to be the native
4214 // Object clone().
4215 if (is_virtual) {
4216 generate_virtual_guard(obj_klass, slow_region);
4217 }
4219 // The object must be cloneable and must not have a finalizer.
4220 // Both of these conditions may be checked in a single test.
4221 // We could optimize the cloneable test further, but we don't care.
4222 generate_access_flags_guard(obj_klass,
4223 // Test both conditions:
4224 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4225 // Must be cloneable but not finalizer:
4226 JVM_ACC_IS_CLONEABLE,
4227 slow_region);
4228 }
4230 if (!stopped()) {
4231 // It's an instance, and it passed the slow-path tests.
4232 PreserveJVMState pjvms(this);
4233 Node* obj_size = NULL;
4234 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4236 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4238 // Present the results of the slow call.
4239 result_reg->init_req(_instance_path, control());
4240 result_val->init_req(_instance_path, alloc_obj);
4241 result_i_o ->set_req(_instance_path, i_o());
4242 result_mem ->set_req(_instance_path, reset_memory());
4243 }
4245 // Generate code for the slow case. We make a call to clone().
4246 set_control(_gvn.transform(slow_region));
4247 if (!stopped()) {
4248 PreserveJVMState pjvms(this);
4249 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4250 Node* slow_result = set_results_for_java_call(slow_call);
4251 // this->control() comes from set_results_for_java_call
4252 result_reg->init_req(_slow_path, control());
4253 result_val->init_req(_slow_path, slow_result);
4254 result_i_o ->set_req(_slow_path, i_o());
4255 result_mem ->set_req(_slow_path, reset_memory());
4256 }
4258 // Return the combined state.
4259 set_control( _gvn.transform(result_reg) );
4260 set_i_o( _gvn.transform(result_i_o) );
4261 set_all_memory( _gvn.transform(result_mem) );
4262 } //original reexecute and sp are set back here
4264 push(_gvn.transform(result_val));
4266 return true;
4267 }
4270 // constants for computing the copy function
4271 enum {
4272 COPYFUNC_UNALIGNED = 0,
4273 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4274 COPYFUNC_CONJOINT = 0,
4275 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4276 };
4278 // Note: The condition "disjoint" applies also for overlapping copies
4279 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4280 static address
4281 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
4282 int selector =
4283 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4284 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4286 #define RETURN_STUB(xxx_arraycopy) { \
4287 name = #xxx_arraycopy; \
4288 return StubRoutines::xxx_arraycopy(); }
4290 switch (t) {
4291 case T_BYTE:
4292 case T_BOOLEAN:
4293 switch (selector) {
4294 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4295 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4296 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4297 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4298 }
4299 case T_CHAR:
4300 case T_SHORT:
4301 switch (selector) {
4302 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4303 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4304 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4305 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4306 }
4307 case T_INT:
4308 case T_FLOAT:
4309 switch (selector) {
4310 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4311 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4312 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4313 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4314 }
4315 case T_DOUBLE:
4316 case T_LONG:
4317 switch (selector) {
4318 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4319 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4320 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4321 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4322 }
4323 case T_ARRAY:
4324 case T_OBJECT:
4325 switch (selector) {
4326 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
4327 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
4328 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
4329 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
4330 }
4331 default:
4332 ShouldNotReachHere();
4333 return NULL;
4334 }
4336 #undef RETURN_STUB
4337 }
4339 //------------------------------basictype2arraycopy----------------------------
4340 address LibraryCallKit::basictype2arraycopy(BasicType t,
4341 Node* src_offset,
4342 Node* dest_offset,
4343 bool disjoint_bases,
4344 const char* &name) {
4345 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4346 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4348 bool aligned = false;
4349 bool disjoint = disjoint_bases;
4351 // if the offsets are the same, we can treat the memory regions as
4352 // disjoint, because either the memory regions are in different arrays,
4353 // or they are identical (which we can treat as disjoint.) We can also
4354 // treat a copy with a destination index less that the source index
4355 // as disjoint since a low->high copy will work correctly in this case.
4356 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4357 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4358 // both indices are constants
4359 int s_offs = src_offset_inttype->get_con();
4360 int d_offs = dest_offset_inttype->get_con();
4361 int element_size = type2aelembytes(t);
4362 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4363 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4364 if (s_offs >= d_offs) disjoint = true;
4365 } else if (src_offset == dest_offset && src_offset != NULL) {
4366 // This can occur if the offsets are identical non-constants.
4367 disjoint = true;
4368 }
4370 return select_arraycopy_function(t, aligned, disjoint, name);
4371 }
4374 //------------------------------inline_arraycopy-----------------------
4375 bool LibraryCallKit::inline_arraycopy() {
4376 // Restore the stack and pop off the arguments.
4377 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4378 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4380 Node *src = argument(0);
4381 Node *src_offset = argument(1);
4382 Node *dest = argument(2);
4383 Node *dest_offset = argument(3);
4384 Node *length = argument(4);
4386 // Compile time checks. If any of these checks cannot be verified at compile time,
4387 // we do not make a fast path for this call. Instead, we let the call remain as it
4388 // is. The checks we choose to mandate at compile time are:
4389 //
4390 // (1) src and dest are arrays.
4391 const Type* src_type = src->Value(&_gvn);
4392 const Type* dest_type = dest->Value(&_gvn);
4393 const TypeAryPtr* top_src = src_type->isa_aryptr();
4394 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4395 if (top_src == NULL || top_src->klass() == NULL ||
4396 top_dest == NULL || top_dest->klass() == NULL) {
4397 // Conservatively insert a memory barrier on all memory slices.
4398 // Do not let writes into the source float below the arraycopy.
4399 insert_mem_bar(Op_MemBarCPUOrder);
4401 // Call StubRoutines::generic_arraycopy stub.
4402 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4403 src, src_offset, dest, dest_offset, length);
4405 // Do not let reads from the destination float above the arraycopy.
4406 // Since we cannot type the arrays, we don't know which slices
4407 // might be affected. We could restrict this barrier only to those
4408 // memory slices which pertain to array elements--but don't bother.
4409 if (!InsertMemBarAfterArraycopy)
4410 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4411 insert_mem_bar(Op_MemBarCPUOrder);
4412 return true;
4413 }
4415 // (2) src and dest arrays must have elements of the same BasicType
4416 // Figure out the size and type of the elements we will be copying.
4417 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4418 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4419 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4420 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4422 if (src_elem != dest_elem || dest_elem == T_VOID) {
4423 // The component types are not the same or are not recognized. Punt.
4424 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4425 generate_slow_arraycopy(TypePtr::BOTTOM,
4426 src, src_offset, dest, dest_offset, length);
4427 return true;
4428 }
4430 //---------------------------------------------------------------------------
4431 // We will make a fast path for this call to arraycopy.
4433 // We have the following tests left to perform:
4434 //
4435 // (3) src and dest must not be null.
4436 // (4) src_offset must not be negative.
4437 // (5) dest_offset must not be negative.
4438 // (6) length must not be negative.
4439 // (7) src_offset + length must not exceed length of src.
4440 // (8) dest_offset + length must not exceed length of dest.
4441 // (9) each element of an oop array must be assignable
4443 RegionNode* slow_region = new (C, 1) RegionNode(1);
4444 record_for_igvn(slow_region);
4446 // (3) operands must not be null
4447 // We currently perform our null checks with the do_null_check routine.
4448 // This means that the null exceptions will be reported in the caller
4449 // rather than (correctly) reported inside of the native arraycopy call.
4450 // This should be corrected, given time. We do our null check with the
4451 // stack pointer restored.
4452 _sp += nargs;
4453 src = do_null_check(src, T_ARRAY);
4454 dest = do_null_check(dest, T_ARRAY);
4455 _sp -= nargs;
4457 // (4) src_offset must not be negative.
4458 generate_negative_guard(src_offset, slow_region);
4460 // (5) dest_offset must not be negative.
4461 generate_negative_guard(dest_offset, slow_region);
4463 // (6) length must not be negative (moved to generate_arraycopy()).
4464 // generate_negative_guard(length, slow_region);
4466 // (7) src_offset + length must not exceed length of src.
4467 generate_limit_guard(src_offset, length,
4468 load_array_length(src),
4469 slow_region);
4471 // (8) dest_offset + length must not exceed length of dest.
4472 generate_limit_guard(dest_offset, length,
4473 load_array_length(dest),
4474 slow_region);
4476 // (9) each element of an oop array must be assignable
4477 // The generate_arraycopy subroutine checks this.
4479 // This is where the memory effects are placed:
4480 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4481 generate_arraycopy(adr_type, dest_elem,
4482 src, src_offset, dest, dest_offset, length,
4483 false, false, slow_region);
4485 return true;
4486 }
4488 //-----------------------------generate_arraycopy----------------------
4489 // Generate an optimized call to arraycopy.
4490 // Caller must guard against non-arrays.
4491 // Caller must determine a common array basic-type for both arrays.
4492 // Caller must validate offsets against array bounds.
4493 // The slow_region has already collected guard failure paths
4494 // (such as out of bounds length or non-conformable array types).
4495 // The generated code has this shape, in general:
4496 //
4497 // if (length == 0) return // via zero_path
4498 // slowval = -1
4499 // if (types unknown) {
4500 // slowval = call generic copy loop
4501 // if (slowval == 0) return // via checked_path
4502 // } else if (indexes in bounds) {
4503 // if ((is object array) && !(array type check)) {
4504 // slowval = call checked copy loop
4505 // if (slowval == 0) return // via checked_path
4506 // } else {
4507 // call bulk copy loop
4508 // return // via fast_path
4509 // }
4510 // }
4511 // // adjust params for remaining work:
4512 // if (slowval != -1) {
4513 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4514 // }
4515 // slow_region:
4516 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4517 // return // via slow_call_path
4518 //
4519 // This routine is used from several intrinsics: System.arraycopy,
4520 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4521 //
4522 void
4523 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4524 BasicType basic_elem_type,
4525 Node* src, Node* src_offset,
4526 Node* dest, Node* dest_offset,
4527 Node* copy_length,
4528 bool disjoint_bases,
4529 bool length_never_negative,
4530 RegionNode* slow_region) {
4532 if (slow_region == NULL) {
4533 slow_region = new(C,1) RegionNode(1);
4534 record_for_igvn(slow_region);
4535 }
4537 Node* original_dest = dest;
4538 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4539 bool must_clear_dest = false;
4541 // See if this is the initialization of a newly-allocated array.
4542 // If so, we will take responsibility here for initializing it to zero.
4543 // (Note: Because tightly_coupled_allocation performs checks on the
4544 // out-edges of the dest, we need to avoid making derived pointers
4545 // from it until we have checked its uses.)
4546 if (ReduceBulkZeroing
4547 && !ZeroTLAB // pointless if already zeroed
4548 && basic_elem_type != T_CONFLICT // avoid corner case
4549 && !_gvn.eqv_uncast(src, dest)
4550 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4551 != NULL)
4552 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4553 && alloc->maybe_set_complete(&_gvn)) {
4554 // "You break it, you buy it."
4555 InitializeNode* init = alloc->initialization();
4556 assert(init->is_complete(), "we just did this");
4557 assert(dest->is_CheckCastPP(), "sanity");
4558 assert(dest->in(0)->in(0) == init, "dest pinned");
4559 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4560 // From this point on, every exit path is responsible for
4561 // initializing any non-copied parts of the object to zero.
4562 must_clear_dest = true;
4563 } else {
4564 // No zeroing elimination here.
4565 alloc = NULL;
4566 //original_dest = dest;
4567 //must_clear_dest = false;
4568 }
4570 // Results are placed here:
4571 enum { fast_path = 1, // normal void-returning assembly stub
4572 checked_path = 2, // special assembly stub with cleanup
4573 slow_call_path = 3, // something went wrong; call the VM
4574 zero_path = 4, // bypass when length of copy is zero
4575 bcopy_path = 5, // copy primitive array by 64-bit blocks
4576 PATH_LIMIT = 6
4577 };
4578 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4579 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4580 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4581 record_for_igvn(result_region);
4582 _gvn.set_type_bottom(result_i_o);
4583 _gvn.set_type_bottom(result_memory);
4584 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4586 // The slow_control path:
4587 Node* slow_control;
4588 Node* slow_i_o = i_o();
4589 Node* slow_mem = memory(adr_type);
4590 debug_only(slow_control = (Node*) badAddress);
4592 // Checked control path:
4593 Node* checked_control = top();
4594 Node* checked_mem = NULL;
4595 Node* checked_i_o = NULL;
4596 Node* checked_value = NULL;
4598 if (basic_elem_type == T_CONFLICT) {
4599 assert(!must_clear_dest, "");
4600 Node* cv = generate_generic_arraycopy(adr_type,
4601 src, src_offset, dest, dest_offset,
4602 copy_length);
4603 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4604 checked_control = control();
4605 checked_i_o = i_o();
4606 checked_mem = memory(adr_type);
4607 checked_value = cv;
4608 set_control(top()); // no fast path
4609 }
4611 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4612 if (not_pos != NULL) {
4613 PreserveJVMState pjvms(this);
4614 set_control(not_pos);
4616 // (6) length must not be negative.
4617 if (!length_never_negative) {
4618 generate_negative_guard(copy_length, slow_region);
4619 }
4621 // copy_length is 0.
4622 if (!stopped() && must_clear_dest) {
4623 Node* dest_length = alloc->in(AllocateNode::ALength);
4624 if (_gvn.eqv_uncast(copy_length, dest_length)
4625 || _gvn.find_int_con(dest_length, 1) <= 0) {
4626 // There is no zeroing to do. No need for a secondary raw memory barrier.
4627 } else {
4628 // Clear the whole thing since there are no source elements to copy.
4629 generate_clear_array(adr_type, dest, basic_elem_type,
4630 intcon(0), NULL,
4631 alloc->in(AllocateNode::AllocSize));
4632 // Use a secondary InitializeNode as raw memory barrier.
4633 // Currently it is needed only on this path since other
4634 // paths have stub or runtime calls as raw memory barriers.
4635 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4636 Compile::AliasIdxRaw,
4637 top())->as_Initialize();
4638 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4639 }
4640 }
4642 // Present the results of the fast call.
4643 result_region->init_req(zero_path, control());
4644 result_i_o ->init_req(zero_path, i_o());
4645 result_memory->init_req(zero_path, memory(adr_type));
4646 }
4648 if (!stopped() && must_clear_dest) {
4649 // We have to initialize the *uncopied* part of the array to zero.
4650 // The copy destination is the slice dest[off..off+len]. The other slices
4651 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4652 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4653 Node* dest_length = alloc->in(AllocateNode::ALength);
4654 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4655 copy_length) );
4657 // If there is a head section that needs zeroing, do it now.
4658 if (find_int_con(dest_offset, -1) != 0) {
4659 generate_clear_array(adr_type, dest, basic_elem_type,
4660 intcon(0), dest_offset,
4661 NULL);
4662 }
4664 // Next, perform a dynamic check on the tail length.
4665 // It is often zero, and we can win big if we prove this.
4666 // There are two wins: Avoid generating the ClearArray
4667 // with its attendant messy index arithmetic, and upgrade
4668 // the copy to a more hardware-friendly word size of 64 bits.
4669 Node* tail_ctl = NULL;
4670 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4671 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4672 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4673 tail_ctl = generate_slow_guard(bol_lt, NULL);
4674 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4675 }
4677 // At this point, let's assume there is no tail.
4678 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4679 // There is no tail. Try an upgrade to a 64-bit copy.
4680 bool didit = false;
4681 { PreserveJVMState pjvms(this);
4682 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4683 src, src_offset, dest, dest_offset,
4684 dest_size);
4685 if (didit) {
4686 // Present the results of the block-copying fast call.
4687 result_region->init_req(bcopy_path, control());
4688 result_i_o ->init_req(bcopy_path, i_o());
4689 result_memory->init_req(bcopy_path, memory(adr_type));
4690 }
4691 }
4692 if (didit)
4693 set_control(top()); // no regular fast path
4694 }
4696 // Clear the tail, if any.
4697 if (tail_ctl != NULL) {
4698 Node* notail_ctl = stopped() ? NULL : control();
4699 set_control(tail_ctl);
4700 if (notail_ctl == NULL) {
4701 generate_clear_array(adr_type, dest, basic_elem_type,
4702 dest_tail, NULL,
4703 dest_size);
4704 } else {
4705 // Make a local merge.
4706 Node* done_ctl = new(C,3) RegionNode(3);
4707 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4708 done_ctl->init_req(1, notail_ctl);
4709 done_mem->init_req(1, memory(adr_type));
4710 generate_clear_array(adr_type, dest, basic_elem_type,
4711 dest_tail, NULL,
4712 dest_size);
4713 done_ctl->init_req(2, control());
4714 done_mem->init_req(2, memory(adr_type));
4715 set_control( _gvn.transform(done_ctl) );
4716 set_memory( _gvn.transform(done_mem), adr_type );
4717 }
4718 }
4719 }
4721 BasicType copy_type = basic_elem_type;
4722 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4723 if (!stopped() && copy_type == T_OBJECT) {
4724 // If src and dest have compatible element types, we can copy bits.
4725 // Types S[] and D[] are compatible if D is a supertype of S.
4726 //
4727 // If they are not, we will use checked_oop_disjoint_arraycopy,
4728 // which performs a fast optimistic per-oop check, and backs off
4729 // further to JVM_ArrayCopy on the first per-oop check that fails.
4730 // (Actually, we don't move raw bits only; the GC requires card marks.)
4732 // Get the klassOop for both src and dest
4733 Node* src_klass = load_object_klass(src);
4734 Node* dest_klass = load_object_klass(dest);
4736 // Generate the subtype check.
4737 // This might fold up statically, or then again it might not.
4738 //
4739 // Non-static example: Copying List<String>.elements to a new String[].
4740 // The backing store for a List<String> is always an Object[],
4741 // but its elements are always type String, if the generic types
4742 // are correct at the source level.
4743 //
4744 // Test S[] against D[], not S against D, because (probably)
4745 // the secondary supertype cache is less busy for S[] than S.
4746 // This usually only matters when D is an interface.
4747 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4748 // Plug failing path into checked_oop_disjoint_arraycopy
4749 if (not_subtype_ctrl != top()) {
4750 PreserveJVMState pjvms(this);
4751 set_control(not_subtype_ctrl);
4752 // (At this point we can assume disjoint_bases, since types differ.)
4753 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4754 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4755 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4756 Node* dest_elem_klass = _gvn.transform(n1);
4757 Node* cv = generate_checkcast_arraycopy(adr_type,
4758 dest_elem_klass,
4759 src, src_offset, dest, dest_offset,
4760 copy_length);
4761 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4762 checked_control = control();
4763 checked_i_o = i_o();
4764 checked_mem = memory(adr_type);
4765 checked_value = cv;
4766 }
4767 // At this point we know we do not need type checks on oop stores.
4769 // Let's see if we need card marks:
4770 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4771 // If we do not need card marks, copy using the jint or jlong stub.
4772 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4773 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4774 "sizes agree");
4775 }
4776 }
4778 if (!stopped()) {
4779 // Generate the fast path, if possible.
4780 PreserveJVMState pjvms(this);
4781 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4782 src, src_offset, dest, dest_offset,
4783 ConvI2X(copy_length));
4785 // Present the results of the fast call.
4786 result_region->init_req(fast_path, control());
4787 result_i_o ->init_req(fast_path, i_o());
4788 result_memory->init_req(fast_path, memory(adr_type));
4789 }
4791 // Here are all the slow paths up to this point, in one bundle:
4792 slow_control = top();
4793 if (slow_region != NULL)
4794 slow_control = _gvn.transform(slow_region);
4795 debug_only(slow_region = (RegionNode*)badAddress);
4797 set_control(checked_control);
4798 if (!stopped()) {
4799 // Clean up after the checked call.
4800 // The returned value is either 0 or -1^K,
4801 // where K = number of partially transferred array elements.
4802 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4803 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4804 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4806 // If it is 0, we are done, so transfer to the end.
4807 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4808 result_region->init_req(checked_path, checks_done);
4809 result_i_o ->init_req(checked_path, checked_i_o);
4810 result_memory->init_req(checked_path, checked_mem);
4812 // If it is not zero, merge into the slow call.
4813 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4814 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4815 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4816 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4817 record_for_igvn(slow_reg2);
4818 slow_reg2 ->init_req(1, slow_control);
4819 slow_i_o2 ->init_req(1, slow_i_o);
4820 slow_mem2 ->init_req(1, slow_mem);
4821 slow_reg2 ->init_req(2, control());
4822 slow_i_o2 ->init_req(2, checked_i_o);
4823 slow_mem2 ->init_req(2, checked_mem);
4825 slow_control = _gvn.transform(slow_reg2);
4826 slow_i_o = _gvn.transform(slow_i_o2);
4827 slow_mem = _gvn.transform(slow_mem2);
4829 if (alloc != NULL) {
4830 // We'll restart from the very beginning, after zeroing the whole thing.
4831 // This can cause double writes, but that's OK since dest is brand new.
4832 // So we ignore the low 31 bits of the value returned from the stub.
4833 } else {
4834 // We must continue the copy exactly where it failed, or else
4835 // another thread might see the wrong number of writes to dest.
4836 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4837 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4838 slow_offset->init_req(1, intcon(0));
4839 slow_offset->init_req(2, checked_offset);
4840 slow_offset = _gvn.transform(slow_offset);
4842 // Adjust the arguments by the conditionally incoming offset.
4843 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4844 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4845 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4847 // Tweak the node variables to adjust the code produced below:
4848 src_offset = src_off_plus;
4849 dest_offset = dest_off_plus;
4850 copy_length = length_minus;
4851 }
4852 }
4854 set_control(slow_control);
4855 if (!stopped()) {
4856 // Generate the slow path, if needed.
4857 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4859 set_memory(slow_mem, adr_type);
4860 set_i_o(slow_i_o);
4862 if (must_clear_dest) {
4863 generate_clear_array(adr_type, dest, basic_elem_type,
4864 intcon(0), NULL,
4865 alloc->in(AllocateNode::AllocSize));
4866 }
4868 generate_slow_arraycopy(adr_type,
4869 src, src_offset, dest, dest_offset,
4870 copy_length);
4872 result_region->init_req(slow_call_path, control());
4873 result_i_o ->init_req(slow_call_path, i_o());
4874 result_memory->init_req(slow_call_path, memory(adr_type));
4875 }
4877 // Remove unused edges.
4878 for (uint i = 1; i < result_region->req(); i++) {
4879 if (result_region->in(i) == NULL)
4880 result_region->init_req(i, top());
4881 }
4883 // Finished; return the combined state.
4884 set_control( _gvn.transform(result_region) );
4885 set_i_o( _gvn.transform(result_i_o) );
4886 set_memory( _gvn.transform(result_memory), adr_type );
4888 // The memory edges above are precise in order to model effects around
4889 // array copies accurately to allow value numbering of field loads around
4890 // arraycopy. Such field loads, both before and after, are common in Java
4891 // collections and similar classes involving header/array data structures.
4892 //
4893 // But with low number of register or when some registers are used or killed
4894 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4895 // The next memory barrier is added to avoid it. If the arraycopy can be
4896 // optimized away (which it can, sometimes) then we can manually remove
4897 // the membar also.
4898 //
4899 // Do not let reads from the cloned object float above the arraycopy.
4900 if (InsertMemBarAfterArraycopy || alloc != NULL)
4901 insert_mem_bar(Op_MemBarCPUOrder);
4902 }
4905 // Helper function which determines if an arraycopy immediately follows
4906 // an allocation, with no intervening tests or other escapes for the object.
4907 AllocateArrayNode*
4908 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4909 RegionNode* slow_region) {
4910 if (stopped()) return NULL; // no fast path
4911 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4913 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4914 if (alloc == NULL) return NULL;
4916 Node* rawmem = memory(Compile::AliasIdxRaw);
4917 // Is the allocation's memory state untouched?
4918 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4919 // Bail out if there have been raw-memory effects since the allocation.
4920 // (Example: There might have been a call or safepoint.)
4921 return NULL;
4922 }
4923 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4924 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4925 return NULL;
4926 }
4928 // There must be no unexpected observers of this allocation.
4929 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4930 Node* obs = ptr->fast_out(i);
4931 if (obs != this->map()) {
4932 return NULL;
4933 }
4934 }
4936 // This arraycopy must unconditionally follow the allocation of the ptr.
4937 Node* alloc_ctl = ptr->in(0);
4938 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4940 Node* ctl = control();
4941 while (ctl != alloc_ctl) {
4942 // There may be guards which feed into the slow_region.
4943 // Any other control flow means that we might not get a chance
4944 // to finish initializing the allocated object.
4945 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4946 IfNode* iff = ctl->in(0)->as_If();
4947 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4948 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4949 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4950 ctl = iff->in(0); // This test feeds the known slow_region.
4951 continue;
4952 }
4953 // One more try: Various low-level checks bottom out in
4954 // uncommon traps. If the debug-info of the trap omits
4955 // any reference to the allocation, as we've already
4956 // observed, then there can be no objection to the trap.
4957 bool found_trap = false;
4958 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4959 Node* obs = not_ctl->fast_out(j);
4960 if (obs->in(0) == not_ctl && obs->is_Call() &&
4961 (obs->as_Call()->entry_point() ==
4962 SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
4963 found_trap = true; break;
4964 }
4965 }
4966 if (found_trap) {
4967 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4968 continue;
4969 }
4970 }
4971 return NULL;
4972 }
4974 // If we get this far, we have an allocation which immediately
4975 // precedes the arraycopy, and we can take over zeroing the new object.
4976 // The arraycopy will finish the initialization, and provide
4977 // a new control state to which we will anchor the destination pointer.
4979 return alloc;
4980 }
4982 // Helper for initialization of arrays, creating a ClearArray.
4983 // It writes zero bits in [start..end), within the body of an array object.
4984 // The memory effects are all chained onto the 'adr_type' alias category.
4985 //
4986 // Since the object is otherwise uninitialized, we are free
4987 // to put a little "slop" around the edges of the cleared area,
4988 // as long as it does not go back into the array's header,
4989 // or beyond the array end within the heap.
4990 //
4991 // The lower edge can be rounded down to the nearest jint and the
4992 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4993 //
4994 // Arguments:
4995 // adr_type memory slice where writes are generated
4996 // dest oop of the destination array
4997 // basic_elem_type element type of the destination
4998 // slice_idx array index of first element to store
4999 // slice_len number of elements to store (or NULL)
5000 // dest_size total size in bytes of the array object
5001 //
5002 // Exactly one of slice_len or dest_size must be non-NULL.
5003 // If dest_size is non-NULL, zeroing extends to the end of the object.
5004 // If slice_len is non-NULL, the slice_idx value must be a constant.
5005 void
5006 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5007 Node* dest,
5008 BasicType basic_elem_type,
5009 Node* slice_idx,
5010 Node* slice_len,
5011 Node* dest_size) {
5012 // one or the other but not both of slice_len and dest_size:
5013 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5014 if (slice_len == NULL) slice_len = top();
5015 if (dest_size == NULL) dest_size = top();
5017 // operate on this memory slice:
5018 Node* mem = memory(adr_type); // memory slice to operate on
5020 // scaling and rounding of indexes:
5021 int scale = exact_log2(type2aelembytes(basic_elem_type));
5022 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5023 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5024 int bump_bit = (-1 << scale) & BytesPerInt;
5026 // determine constant starts and ends
5027 const intptr_t BIG_NEG = -128;
5028 assert(BIG_NEG + 2*abase < 0, "neg enough");
5029 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5030 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5031 if (slice_len_con == 0) {
5032 return; // nothing to do here
5033 }
5034 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5035 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5036 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5037 assert(end_con < 0, "not two cons");
5038 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5039 BytesPerLong);
5040 }
5042 if (start_con >= 0 && end_con >= 0) {
5043 // Constant start and end. Simple.
5044 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5045 start_con, end_con, &_gvn);
5046 } else if (start_con >= 0 && dest_size != top()) {
5047 // Constant start, pre-rounded end after the tail of the array.
5048 Node* end = dest_size;
5049 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5050 start_con, end, &_gvn);
5051 } else if (start_con >= 0 && slice_len != top()) {
5052 // Constant start, non-constant end. End needs rounding up.
5053 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5054 intptr_t end_base = abase + (slice_idx_con << scale);
5055 int end_round = (-1 << scale) & (BytesPerLong - 1);
5056 Node* end = ConvI2X(slice_len);
5057 if (scale != 0)
5058 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5059 end_base += end_round;
5060 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5061 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5062 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5063 start_con, end, &_gvn);
5064 } else if (start_con < 0 && dest_size != top()) {
5065 // Non-constant start, pre-rounded end after the tail of the array.
5066 // This is almost certainly a "round-to-end" operation.
5067 Node* start = slice_idx;
5068 start = ConvI2X(start);
5069 if (scale != 0)
5070 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5071 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5072 if ((bump_bit | clear_low) != 0) {
5073 int to_clear = (bump_bit | clear_low);
5074 // Align up mod 8, then store a jint zero unconditionally
5075 // just before the mod-8 boundary.
5076 if (((abase + bump_bit) & ~to_clear) - bump_bit
5077 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5078 bump_bit = 0;
5079 assert((abase & to_clear) == 0, "array base must be long-aligned");
5080 } else {
5081 // Bump 'start' up to (or past) the next jint boundary:
5082 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5083 assert((abase & clear_low) == 0, "array base must be int-aligned");
5084 }
5085 // Round bumped 'start' down to jlong boundary in body of array.
5086 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5087 if (bump_bit != 0) {
5088 // Store a zero to the immediately preceding jint:
5089 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5090 Node* p1 = basic_plus_adr(dest, x1);
5091 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5092 mem = _gvn.transform(mem);
5093 }
5094 }
5095 Node* end = dest_size; // pre-rounded
5096 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5097 start, end, &_gvn);
5098 } else {
5099 // Non-constant start, unrounded non-constant end.
5100 // (Nobody zeroes a random midsection of an array using this routine.)
5101 ShouldNotReachHere(); // fix caller
5102 }
5104 // Done.
5105 set_memory(mem, adr_type);
5106 }
5109 bool
5110 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5111 BasicType basic_elem_type,
5112 AllocateNode* alloc,
5113 Node* src, Node* src_offset,
5114 Node* dest, Node* dest_offset,
5115 Node* dest_size) {
5116 // See if there is an advantage from block transfer.
5117 int scale = exact_log2(type2aelembytes(basic_elem_type));
5118 if (scale >= LogBytesPerLong)
5119 return false; // it is already a block transfer
5121 // Look at the alignment of the starting offsets.
5122 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5123 const intptr_t BIG_NEG = -128;
5124 assert(BIG_NEG + 2*abase < 0, "neg enough");
5126 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5127 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5128 if (src_off < 0 || dest_off < 0)
5129 // At present, we can only understand constants.
5130 return false;
5132 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5133 // Non-aligned; too bad.
5134 // One more chance: Pick off an initial 32-bit word.
5135 // This is a common case, since abase can be odd mod 8.
5136 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5137 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5138 Node* sptr = basic_plus_adr(src, src_off);
5139 Node* dptr = basic_plus_adr(dest, dest_off);
5140 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5141 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5142 src_off += BytesPerInt;
5143 dest_off += BytesPerInt;
5144 } else {
5145 return false;
5146 }
5147 }
5148 assert(src_off % BytesPerLong == 0, "");
5149 assert(dest_off % BytesPerLong == 0, "");
5151 // Do this copy by giant steps.
5152 Node* sptr = basic_plus_adr(src, src_off);
5153 Node* dptr = basic_plus_adr(dest, dest_off);
5154 Node* countx = dest_size;
5155 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5156 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5158 bool disjoint_bases = true; // since alloc != NULL
5159 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5160 sptr, NULL, dptr, NULL, countx);
5162 return true;
5163 }
5166 // Helper function; generates code for the slow case.
5167 // We make a call to a runtime method which emulates the native method,
5168 // but without the native wrapper overhead.
5169 void
5170 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5171 Node* src, Node* src_offset,
5172 Node* dest, Node* dest_offset,
5173 Node* copy_length) {
5174 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5175 OptoRuntime::slow_arraycopy_Type(),
5176 OptoRuntime::slow_arraycopy_Java(),
5177 "slow_arraycopy", adr_type,
5178 src, src_offset, dest, dest_offset,
5179 copy_length);
5181 // Handle exceptions thrown by this fellow:
5182 make_slow_call_ex(call, env()->Throwable_klass(), false);
5183 }
5185 // Helper function; generates code for cases requiring runtime checks.
5186 Node*
5187 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5188 Node* dest_elem_klass,
5189 Node* src, Node* src_offset,
5190 Node* dest, Node* dest_offset,
5191 Node* copy_length) {
5192 if (stopped()) return NULL;
5194 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
5195 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5196 return NULL;
5197 }
5199 // Pick out the parameters required to perform a store-check
5200 // for the target array. This is an optimistic check. It will
5201 // look in each non-null element's class, at the desired klass's
5202 // super_check_offset, for the desired klass.
5203 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5204 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5205 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
5206 Node* check_offset = _gvn.transform(n3);
5207 Node* check_value = dest_elem_klass;
5209 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5210 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5212 // (We know the arrays are never conjoint, because their types differ.)
5213 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5214 OptoRuntime::checkcast_arraycopy_Type(),
5215 copyfunc_addr, "checkcast_arraycopy", adr_type,
5216 // five arguments, of which two are
5217 // intptr_t (jlong in LP64)
5218 src_start, dest_start,
5219 copy_length XTOP,
5220 check_offset XTOP,
5221 check_value);
5223 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5224 }
5227 // Helper function; generates code for cases requiring runtime checks.
5228 Node*
5229 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5230 Node* src, Node* src_offset,
5231 Node* dest, Node* dest_offset,
5232 Node* copy_length) {
5233 if (stopped()) return NULL;
5235 address copyfunc_addr = StubRoutines::generic_arraycopy();
5236 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5237 return NULL;
5238 }
5240 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5241 OptoRuntime::generic_arraycopy_Type(),
5242 copyfunc_addr, "generic_arraycopy", adr_type,
5243 src, src_offset, dest, dest_offset, copy_length);
5245 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5246 }
5248 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5249 void
5250 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5251 BasicType basic_elem_type,
5252 bool disjoint_bases,
5253 Node* src, Node* src_offset,
5254 Node* dest, Node* dest_offset,
5255 Node* copy_length) {
5256 if (stopped()) return; // nothing to do
5258 Node* src_start = src;
5259 Node* dest_start = dest;
5260 if (src_offset != NULL || dest_offset != NULL) {
5261 assert(src_offset != NULL && dest_offset != NULL, "");
5262 src_start = array_element_address(src, src_offset, basic_elem_type);
5263 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5264 }
5266 // Figure out which arraycopy runtime method to call.
5267 const char* copyfunc_name = "arraycopy";
5268 address copyfunc_addr =
5269 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5270 disjoint_bases, copyfunc_name);
5272 // Call it. Note that the count_ix value is not scaled to a byte-size.
5273 make_runtime_call(RC_LEAF|RC_NO_FP,
5274 OptoRuntime::fast_arraycopy_Type(),
5275 copyfunc_addr, copyfunc_name, adr_type,
5276 src_start, dest_start, copy_length XTOP);
5277 }