1 /*
2 * Copyright (c) 1999, 2010, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_library_call.cpp.incl"
27
28 class LibraryIntrinsic : public InlineCallGenerator {
29 // Extend the set of intrinsics known to the runtime:
30 public:
31 private:
32 bool _is_virtual;
33 vmIntrinsics::ID _intrinsic_id;
34
35 public:
36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
37 : InlineCallGenerator(m),
38 _is_virtual(is_virtual),
39 _intrinsic_id(id)
40 {
41 }
42 virtual bool is_intrinsic() const { return true; }
43 virtual bool is_virtual() const { return _is_virtual; }
44 virtual JVMState* generate(JVMState* jvms);
45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
46 };
47
48
49 // Local helper class for LibraryIntrinsic:
50 class LibraryCallKit : public GraphKit {
51 private:
52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
53
54 public:
55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
56 : GraphKit(caller),
57 _intrinsic(intrinsic)
58 {
59 }
60
61 ciMethod* caller() const { return jvms()->method(); }
62 int bci() const { return jvms()->bci(); }
63 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
65 ciMethod* callee() const { return _intrinsic->method(); }
66 ciSignature* signature() const { return callee()->signature(); }
67 int arg_size() const { return callee()->arg_size(); }
68
69 bool try_to_inline();
70
71 // Helper functions to inline natives
72 void push_result(RegionNode* region, PhiNode* value);
73 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
74 Node* generate_slow_guard(Node* test, RegionNode* region);
75 Node* generate_fair_guard(Node* test, RegionNode* region);
76 Node* generate_negative_guard(Node* index, RegionNode* region,
77 // resulting CastII of index:
78 Node* *pos_index = NULL);
79 Node* generate_nonpositive_guard(Node* index, bool never_negative,
80 // resulting CastII of index:
81 Node* *pos_index = NULL);
82 Node* generate_limit_guard(Node* offset, Node* subseq_length,
83 Node* array_length,
84 RegionNode* region);
85 Node* generate_current_thread(Node* &tls_output);
86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
87 bool disjoint_bases, const char* &name);
88 Node* load_mirror_from_klass(Node* klass);
89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
90 int nargs,
91 RegionNode* region, int null_path,
92 int offset);
93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
94 RegionNode* region, int null_path) {
95 int offset = java_lang_Class::klass_offset_in_bytes();
96 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
97 region, null_path,
98 offset);
99 }
100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
101 int nargs,
102 RegionNode* region, int null_path) {
103 int offset = java_lang_Class::array_klass_offset_in_bytes();
104 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
105 region, null_path,
106 offset);
107 }
108 Node* generate_access_flags_guard(Node* kls,
109 int modifier_mask, int modifier_bits,
110 RegionNode* region);
111 Node* generate_interface_guard(Node* kls, RegionNode* region);
112 Node* generate_array_guard(Node* kls, RegionNode* region) {
113 return generate_array_guard_common(kls, region, false, false);
114 }
115 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
116 return generate_array_guard_common(kls, region, false, true);
117 }
118 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
119 return generate_array_guard_common(kls, region, true, false);
120 }
121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
122 return generate_array_guard_common(kls, region, true, true);
123 }
124 Node* generate_array_guard_common(Node* kls, RegionNode* region,
125 bool obj_array, bool not_array);
126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
128 bool is_virtual = false, bool is_static = false);
129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
130 return generate_method_call(method_id, false, true);
131 }
132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
133 return generate_method_call(method_id, true, false);
134 }
135
136 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();
164
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();
178
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 };
230
231
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");
236
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 }
243
244 if (!m->is_loaded()) {
245 // do not attempt to inline unloaded methods
246 return NULL;
247 }
248
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 }
261
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 }
274
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;
308
309 case vmIntrinsics::_get_AtomicLong:
310 case vmIntrinsics::_attemptUpdate:
311 if (!InlineAtomicLong) return NULL;
312 break;
313
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;
319
320 case vmIntrinsics::_bitCount_i:
321 case vmIntrinsics::_bitCount_l:
322 if (!UsePopCountInstruction) return NULL;
323 break;
324
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 }
330
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 }
338
339 // -XX:-InlineThreadNatives disables natives from the Thread class.
340 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
341 if (!InlineThreadNatives) return NULL;
342 }
343
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 }
350
351 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
352 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
353 if (!InlineUnsafeOps) return NULL;
354 }
355
356 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
357 }
358
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 }
364
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 }
393
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 }
404
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;
410
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();
418
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());
430
431 case vmIntrinsics::_min:
432 case vmIntrinsics::_max:
433 return inline_min_max(intrinsic_id());
434
435 case vmIntrinsics::_arraycopy:
436 return inline_arraycopy();
437
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();
444
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);
463
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);
482
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);
499
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);
516
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);
535
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);
554
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);
563
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);
570
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);
577
578 case vmIntrinsics::_currentThread:
579 return inline_native_currentThread();
580 case vmIntrinsics::_isInterrupted:
581 return inline_native_isInterrupted();
582
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());
603
604 case vmIntrinsics::_isAssignableFrom:
605 return inline_native_subtype_check();
606
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());
616
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());
624
625 case vmIntrinsics::_numberOfLeadingZeros_i:
626 case vmIntrinsics::_numberOfLeadingZeros_l:
627 return inline_numberOfLeadingZeros(intrinsic_id());
628
629 case vmIntrinsics::_numberOfTrailingZeros_i:
630 case vmIntrinsics::_numberOfTrailingZeros_l:
631 return inline_numberOfTrailingZeros(intrinsic_id());
632
633 case vmIntrinsics::_bitCount_i:
634 case vmIntrinsics::_bitCount_l:
635 return inline_bitCount(intrinsic_id());
636
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());
642
643 case vmIntrinsics::_get_AtomicLong:
644 return inline_native_AtomicLong_get();
645 case vmIntrinsics::_attemptUpdate:
646 return inline_native_AtomicLong_attemptUpdate();
647
648 case vmIntrinsics::_getCallerClass:
649 return inline_native_Reflection_getCallerClass();
650
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 }
663
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 }
672
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 }
688
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 }
695
696 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
697
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 }
703
704 if (region != NULL)
705 region->add_req(if_slow);
706
707 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
708 set_control(if_fast);
709
710 return if_slow;
711 }
712
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 }
719
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 }
737
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 }
756
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 }
788
789
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 }
800
801
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();
808
809 Node* no_ctrl = NULL;
810
811 ciInstanceKlass* klass = env()->String_klass();
812 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
813
814 const TypeAryPtr* value_type =
815 TypeAryPtr::make(TypePtr::NotNull,
816 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
817 ciTypeArrayKlass::make(T_CHAR), true, 0);
818
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);
825
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);
833
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 }
852
853 // All these intrinsics have checks.
854 C->set_has_split_ifs(true); // Has chance for split-if optimization
855
856 return _gvn.transform(result);
857 }
858
859 //------------------------------inline_string_compareTo------------------------
860 bool LibraryCallKit::inline_string_compareTo() {
861
862 if (!Matcher::has_match_rule(Op_StrComp)) return false;
863
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();
867
868 _sp += 2;
869 Node *argument = pop(); // pop non-receiver first: it was pushed second
870 Node *receiver = pop();
871
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 }
883
884 ciInstanceKlass* klass = env()->String_klass();
885 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
886 Node* no_ctrl = NULL;
887
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));
891
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));
894
895 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
896 push(compare);
897 return true;
898 }
899
900 //------------------------------inline_string_equals------------------------
901 bool LibraryCallKit::inline_string_equals() {
902
903 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
904
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();
908
909 int nargs = 2;
910 _sp += nargs;
911 Node* argument = pop(); // pop non-receiver first: it was pushed second
912 Node* receiver = pop();
913
914 // Null check on self without removing any arguments. The argument
915 // null check technically happens in the wrong place, which can lead to
916 // invalid stack traces when string compare is inlined into a method
917 // which handles NullPointerExceptions.
918 _sp += nargs;
919 receiver = do_null_check(receiver, T_OBJECT);
920 //should not do null check for argument for String.equals(), because spec
921 //allows to specify NULL as argument.
922 _sp -= nargs;
923
924 if (stopped()) {
925 return true;
926 }
927
928 // paths (plus control) merge
929 RegionNode* region = new (C, 5) RegionNode(5);
930 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
931
932 // does source == target string?
933 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
934 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
935
936 Node* if_eq = generate_slow_guard(bol, NULL);
937 if (if_eq != NULL) {
938 // receiver == argument
939 phi->init_req(2, intcon(1));
940 region->init_req(2, if_eq);
941 }
942
943 // get String klass for instanceOf
944 ciInstanceKlass* klass = env()->String_klass();
945
946 if (!stopped()) {
947 _sp += nargs; // gen_instanceof might do an uncommon trap
948 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
949 _sp -= nargs;
950 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
951 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
952
953 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
954 //instanceOf == true, fallthrough
955
956 if (inst_false != NULL) {
957 phi->init_req(3, intcon(0));
958 region->init_req(3, inst_false);
959 }
960 }
961
962 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
963
964 Node* no_ctrl = NULL;
965 Node* receiver_cnt;
966 Node* argument_cnt;
967
968 if (!stopped()) {
969 // Properly cast the argument to String
970 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
971
972 // Get counts for string and argument
973 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
974 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
975
976 // Pin load from argument string since it could be NULL.
977 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
978 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
979
980 // Check for receiver count != argument count
981 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
982 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
983 Node* if_ne = generate_slow_guard(bol, NULL);
984 if (if_ne != NULL) {
985 phi->init_req(4, intcon(0));
986 region->init_req(4, if_ne);
987 }
988 }
989
990 // Check for count == 0 is done by mach node StrEquals.
991
992 if (!stopped()) {
993 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
994 phi->init_req(1, equals);
995 region->init_req(1, control());
996 }
997
998 // post merge
999 set_control(_gvn.transform(region));
1000 record_for_igvn(region);
1001
1002 push(_gvn.transform(phi));
1003
1004 return true;
1005 }
1006
1007 //------------------------------inline_array_equals----------------------------
1008 bool LibraryCallKit::inline_array_equals() {
1009
1010 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1011
1012 _sp += 2;
1013 Node *argument2 = pop();
1014 Node *argument1 = pop();
1015
1016 Node* equals =
1017 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1018 argument1, argument2) );
1019 push(equals);
1020 return true;
1021 }
1022
1023 // Java version of String.indexOf(constant string)
1024 // class StringDecl {
1025 // StringDecl(char[] ca) {
1026 // offset = 0;
1027 // count = ca.length;
1028 // value = ca;
1029 // }
1030 // int offset;
1031 // int count;
1032 // char[] value;
1033 // }
1034 //
1035 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1036 // int targetOffset, int cache_i, int md2) {
1037 // int cache = cache_i;
1038 // int sourceOffset = string_object.offset;
1039 // int sourceCount = string_object.count;
1040 // int targetCount = target_object.length;
1041 //
1042 // int targetCountLess1 = targetCount - 1;
1043 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1044 //
1045 // char[] source = string_object.value;
1046 // char[] target = target_object;
1047 // int lastChar = target[targetCountLess1];
1048 //
1049 // outer_loop:
1050 // for (int i = sourceOffset; i < sourceEnd; ) {
1051 // int src = source[i + targetCountLess1];
1052 // if (src == lastChar) {
1053 // // With random strings and a 4-character alphabet,
1054 // // reverse matching at this point sets up 0.8% fewer
1055 // // frames, but (paradoxically) makes 0.3% more probes.
1056 // // Since those probes are nearer the lastChar probe,
1057 // // there is may be a net D$ win with reverse matching.
1058 // // But, reversing loop inhibits unroll of inner loop
1059 // // for unknown reason. So, does running outer loop from
1060 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1061 // for (int j = 0; j < targetCountLess1; j++) {
1062 // if (target[targetOffset + j] != source[i+j]) {
1063 // if ((cache & (1 << source[i+j])) == 0) {
1064 // if (md2 < j+1) {
1065 // i += j+1;
1066 // continue outer_loop;
1067 // }
1068 // }
1069 // i += md2;
1070 // continue outer_loop;
1071 // }
1072 // }
1073 // return i - sourceOffset;
1074 // }
1075 // if ((cache & (1 << src)) == 0) {
1076 // i += targetCountLess1;
1077 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1078 // i++;
1079 // }
1080 // return -1;
1081 // }
1082
1083 //------------------------------string_indexOf------------------------
1084 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1085 jint cache_i, jint md2_i) {
1086
1087 Node* no_ctrl = NULL;
1088 float likely = PROB_LIKELY(0.9);
1089 float unlikely = PROB_UNLIKELY(0.9);
1090
1091 const int value_offset = java_lang_String::value_offset_in_bytes();
1092 const int count_offset = java_lang_String::count_offset_in_bytes();
1093 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1094
1095 ciInstanceKlass* klass = env()->String_klass();
1096 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1097 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1098
1099 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1100 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1101 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1102 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1103 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1104 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1105
1106 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
1107 jint target_length = target_array->length();
1108 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1109 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1110
1111 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1112 #define __ kit.
1113 Node* zero = __ ConI(0);
1114 Node* one = __ ConI(1);
1115 Node* cache = __ ConI(cache_i);
1116 Node* md2 = __ ConI(md2_i);
1117 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1118 Node* targetCount = __ ConI(target_length);
1119 Node* targetCountLess1 = __ ConI(target_length - 1);
1120 Node* targetOffset = __ ConI(targetOffset_i);
1121 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1122
1123 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1124 Node* outer_loop = __ make_label(2 /* goto */);
1125 Node* return_ = __ make_label(1);
1126
1127 __ set(rtn,__ ConI(-1));
1128 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
1129 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1130 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1131 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1132 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1133 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
1134 Node* tpj = __ AddI(targetOffset, __ value(j));
1135 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1136 Node* ipj = __ AddI(__ value(i), __ value(j));
1137 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1138 __ if_then(targ, BoolTest::ne, src2); {
1139 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1140 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1141 __ increment(i, __ AddI(__ value(j), one));
1142 __ goto_(outer_loop);
1143 } __ end_if(); __ dead(j);
1144 }__ end_if(); __ dead(j);
1145 __ increment(i, md2);
1146 __ goto_(outer_loop);
1147 }__ end_if();
1148 __ increment(j, one);
1149 }__ end_loop(); __ dead(j);
1150 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1151 __ goto_(return_);
1152 }__ end_if();
1153 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1154 __ increment(i, targetCountLess1);
1155 }__ end_if();
1156 __ increment(i, one);
1157 __ bind(outer_loop);
1158 }__ end_loop(); __ dead(i);
1159 __ bind(return_);
1160
1161 // Final sync IdealKit and GraphKit.
1162 sync_kit(kit);
1163 Node* result = __ value(rtn);
1164 #undef __
1165 C->set_has_loops(true);
1166 return result;
1167 }
1168
1169 //------------------------------inline_string_indexOf------------------------
1170 bool LibraryCallKit::inline_string_indexOf() {
1171
1172 const int value_offset = java_lang_String::value_offset_in_bytes();
1173 const int count_offset = java_lang_String::count_offset_in_bytes();
1174 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1175
1176 _sp += 2;
1177 Node *argument = pop(); // pop non-receiver first: it was pushed second
1178 Node *receiver = pop();
1179
1180 Node* result;
1181 // Disable the use of pcmpestri until it can be guaranteed that
1182 // the load doesn't cross into the uncommited space.
1183 if (false && Matcher::has_match_rule(Op_StrIndexOf) &&
1184 UseSSE42Intrinsics) {
1185 // Generate SSE4.2 version of indexOf
1186 // We currently only have match rules that use SSE4.2
1187
1188 // Null check on self without removing any arguments. The argument
1189 // null check technically happens in the wrong place, which can lead to
1190 // invalid stack traces when string compare is inlined into a method
1191 // which handles NullPointerExceptions.
1192 _sp += 2;
1193 receiver = do_null_check(receiver, T_OBJECT);
1194 argument = do_null_check(argument, T_OBJECT);
1195 _sp -= 2;
1196
1197 if (stopped()) {
1198 return true;
1199 }
1200
1201 // Make the merge point
1202 RegionNode* result_rgn = new (C, 3) RegionNode(3);
1203 Node* result_phi = new (C, 3) PhiNode(result_rgn, TypeInt::INT);
1204 Node* no_ctrl = NULL;
1205
1206 ciInstanceKlass* klass = env()->String_klass();
1207 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1208
1209 // Get counts for string and substr
1210 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1211 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1212
1213 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1214 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1215
1216 // Check for substr count > string count
1217 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1218 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1219 Node* if_gt = generate_slow_guard(bol, NULL);
1220 if (if_gt != NULL) {
1221 result_phi->init_req(2, intcon(-1));
1222 result_rgn->init_req(2, if_gt);
1223 }
1224
1225 if (!stopped()) {
1226 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1227 result_phi->init_req(1, result);
1228 result_rgn->init_req(1, control());
1229 }
1230 set_control(_gvn.transform(result_rgn));
1231 record_for_igvn(result_rgn);
1232 result = _gvn.transform(result_phi);
1233
1234 } else { //Use LibraryCallKit::string_indexOf
1235 // don't intrinsify is argument isn't a constant string.
1236 if (!argument->is_Con()) {
1237 return false;
1238 }
1239 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1240 if (str_type == NULL) {
1241 return false;
1242 }
1243 ciInstanceKlass* klass = env()->String_klass();
1244 ciObject* str_const = str_type->const_oop();
1245 if (str_const == NULL || str_const->klass() != klass) {
1246 return false;
1247 }
1248 ciInstance* str = str_const->as_instance();
1249 assert(str != NULL, "must be instance");
1250
1251 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1252 int o = str->field_value_by_offset(offset_offset).as_int();
1253 int c = str->field_value_by_offset(count_offset).as_int();
1254 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1255
1256 // constant strings have no offset and count == length which
1257 // simplifies the resulting code somewhat so lets optimize for that.
1258 if (o != 0 || c != pat->length()) {
1259 return false;
1260 }
1261
1262 // Null check on self without removing any arguments. The argument
1263 // null check technically happens in the wrong place, which can lead to
1264 // invalid stack traces when string compare is inlined into a method
1265 // which handles NullPointerExceptions.
1266 _sp += 2;
1267 receiver = do_null_check(receiver, T_OBJECT);
1268 // No null check on the argument is needed since it's a constant String oop.
1269 _sp -= 2;
1270 if (stopped()) {
1271 return true;
1272 }
1273
1274 // The null string as a pattern always returns 0 (match at beginning of string)
1275 if (c == 0) {
1276 push(intcon(0));
1277 return true;
1278 }
1279
1280 // Generate default indexOf
1281 jchar lastChar = pat->char_at(o + (c - 1));
1282 int cache = 0;
1283 int i;
1284 for (i = 0; i < c - 1; i++) {
1285 assert(i < pat->length(), "out of range");
1286 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1287 }
1288
1289 int md2 = c;
1290 for (i = 0; i < c - 1; i++) {
1291 assert(i < pat->length(), "out of range");
1292 if (pat->char_at(o + i) == lastChar) {
1293 md2 = (c - 1) - i;
1294 }
1295 }
1296
1297 result = string_indexOf(receiver, pat, o, cache, md2);
1298 }
1299
1300 push(result);
1301 return true;
1302 }
1303
1304 //--------------------------pop_math_arg--------------------------------
1305 // Pop a double argument to a math function from the stack
1306 // rounding it if necessary.
1307 Node * LibraryCallKit::pop_math_arg() {
1308 Node *arg = pop_pair();
1309 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1310 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1311 return arg;
1312 }
1313
1314 //------------------------------inline_trig----------------------------------
1315 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1316 // argument reduction which will turn into a fast/slow diamond.
1317 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1318 _sp += arg_size(); // restore stack pointer
1319 Node* arg = pop_math_arg();
1320 Node* trig = NULL;
1321
1322 switch (id) {
1323 case vmIntrinsics::_dsin:
1324 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1325 break;
1326 case vmIntrinsics::_dcos:
1327 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1328 break;
1329 case vmIntrinsics::_dtan:
1330 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1331 break;
1332 default:
1333 assert(false, "bad intrinsic was passed in");
1334 return false;
1335 }
1336
1337 // Rounding required? Check for argument reduction!
1338 if( Matcher::strict_fp_requires_explicit_rounding ) {
1339
1340 static const double pi_4 = 0.7853981633974483;
1341 static const double neg_pi_4 = -0.7853981633974483;
1342 // pi/2 in 80-bit extended precision
1343 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1344 // -pi/2 in 80-bit extended precision
1345 // 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};
1346 // Cutoff value for using this argument reduction technique
1347 //static const double pi_2_minus_epsilon = 1.564660403643354;
1348 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1349
1350 // Pseudocode for sin:
1351 // if (x <= Math.PI / 4.0) {
1352 // if (x >= -Math.PI / 4.0) return fsin(x);
1353 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1354 // } else {
1355 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1356 // }
1357 // return StrictMath.sin(x);
1358
1359 // Pseudocode for cos:
1360 // if (x <= Math.PI / 4.0) {
1361 // if (x >= -Math.PI / 4.0) return fcos(x);
1362 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1363 // } else {
1364 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1365 // }
1366 // return StrictMath.cos(x);
1367
1368 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1369 // requires a special machine instruction to load it. Instead we'll try
1370 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1371 // probably do the math inside the SIN encoding.
1372
1373 // Make the merge point
1374 RegionNode *r = new (C, 3) RegionNode(3);
1375 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1376
1377 // Flatten arg so we need only 1 test
1378 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1379 // Node for PI/4 constant
1380 Node *pi4 = makecon(TypeD::make(pi_4));
1381 // Check PI/4 : abs(arg)
1382 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1383 // Check: If PI/4 < abs(arg) then go slow
1384 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1385 // Branch either way
1386 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1387 set_control(opt_iff(r,iff));
1388
1389 // Set fast path result
1390 phi->init_req(2,trig);
1391
1392 // Slow path - non-blocking leaf call
1393 Node* call = NULL;
1394 switch (id) {
1395 case vmIntrinsics::_dsin:
1396 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1397 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1398 "Sin", NULL, arg, top());
1399 break;
1400 case vmIntrinsics::_dcos:
1401 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1402 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1403 "Cos", NULL, arg, top());
1404 break;
1405 case vmIntrinsics::_dtan:
1406 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1407 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1408 "Tan", NULL, arg, top());
1409 break;
1410 }
1411 assert(control()->in(0) == call, "");
1412 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1413 r->init_req(1,control());
1414 phi->init_req(1,slow_result);
1415
1416 // Post-merge
1417 set_control(_gvn.transform(r));
1418 record_for_igvn(r);
1419 trig = _gvn.transform(phi);
1420
1421 C->set_has_split_ifs(true); // Has chance for split-if optimization
1422 }
1423 // Push result back on JVM stack
1424 push_pair(trig);
1425 return true;
1426 }
1427
1428 //------------------------------inline_sqrt-------------------------------------
1429 // Inline square root instruction, if possible.
1430 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1431 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1432 _sp += arg_size(); // restore stack pointer
1433 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1434 return true;
1435 }
1436
1437 //------------------------------inline_abs-------------------------------------
1438 // Inline absolute value instruction, if possible.
1439 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1440 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1441 _sp += arg_size(); // restore stack pointer
1442 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1443 return true;
1444 }
1445
1446 //------------------------------inline_exp-------------------------------------
1447 // Inline exp instructions, if possible. The Intel hardware only misses
1448 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1449 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1450 assert(id == vmIntrinsics::_dexp, "Not exp");
1451
1452 // If this inlining ever returned NaN in the past, we do not intrinsify it
1453 // every again. NaN results requires StrictMath.exp handling.
1454 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1455
1456 // Do not intrinsify on older platforms which lack cmove.
1457 if (ConditionalMoveLimit == 0) return false;
1458
1459 _sp += arg_size(); // restore stack pointer
1460 Node *x = pop_math_arg();
1461 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1462
1463 //-------------------
1464 //result=(result.isNaN())? StrictMath::exp():result;
1465 // Check: If isNaN() by checking result!=result? then go to Strict Math
1466 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1467 // Build the boolean node
1468 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1469
1470 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1471 // End the current control-flow path
1472 push_pair(x);
1473 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1474 // to handle. Recompile without intrinsifying Math.exp
1475 uncommon_trap(Deoptimization::Reason_intrinsic,
1476 Deoptimization::Action_make_not_entrant);
1477 }
1478
1479 C->set_has_split_ifs(true); // Has chance for split-if optimization
1480
1481 push_pair(result);
1482
1483 return true;
1484 }
1485
1486 //------------------------------inline_pow-------------------------------------
1487 // Inline power instructions, if possible.
1488 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1489 assert(id == vmIntrinsics::_dpow, "Not pow");
1490
1491 // If this inlining ever returned NaN in the past, we do not intrinsify it
1492 // every again. NaN results requires StrictMath.pow handling.
1493 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1494
1495 // Do not intrinsify on older platforms which lack cmove.
1496 if (ConditionalMoveLimit == 0) return false;
1497
1498 // Pseudocode for pow
1499 // if (x <= 0.0) {
1500 // if ((double)((int)y)==y) { // if y is int
1501 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1502 // } else {
1503 // result = NaN;
1504 // }
1505 // } else {
1506 // result = DPow(x,y);
1507 // }
1508 // if (result != result)? {
1509 // uncommon_trap();
1510 // }
1511 // return result;
1512
1513 _sp += arg_size(); // restore stack pointer
1514 Node* y = pop_math_arg();
1515 Node* x = pop_math_arg();
1516
1517 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1518
1519 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1520 // inside of something) then skip the fancy tests and just check for
1521 // NaN result.
1522 Node *result = NULL;
1523 if( jvms()->depth() >= 1 ) {
1524 result = fast_result;
1525 } else {
1526
1527 // Set the merge point for If node with condition of (x <= 0.0)
1528 // There are four possible paths to region node and phi node
1529 RegionNode *r = new (C, 4) RegionNode(4);
1530 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1531
1532 // Build the first if node: if (x <= 0.0)
1533 // Node for 0 constant
1534 Node *zeronode = makecon(TypeD::ZERO);
1535 // Check x:0
1536 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1537 // Check: If (x<=0) then go complex path
1538 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1539 // Branch either way
1540 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1541 Node *opt_test = _gvn.transform(if1);
1542 //assert( opt_test->is_If(), "Expect an IfNode");
1543 IfNode *opt_if1 = (IfNode*)opt_test;
1544 // Fast path taken; set region slot 3
1545 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1546 r->init_req(3,fast_taken); // Capture fast-control
1547
1548 // Fast path not-taken, i.e. slow path
1549 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1550
1551 // Set fast path result
1552 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1553 phi->init_req(3, fast_result);
1554
1555 // Complex path
1556 // Build the second if node (if y is int)
1557 // Node for (int)y
1558 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1559 // Node for (double)((int) y)
1560 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1561 // Check (double)((int) y) : y
1562 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1563 // Check if (y isn't int) then go to slow path
1564
1565 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1566 // Branch either way
1567 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1568 Node *slow_path = opt_iff(r,if2); // Set region path 2
1569
1570 // Calculate DPow(abs(x), y)*(1 & (int)y)
1571 // Node for constant 1
1572 Node *conone = intcon(1);
1573 // 1& (int)y
1574 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1575 // zero node
1576 Node *conzero = intcon(0);
1577 // Check (1&(int)y)==0?
1578 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1579 // Check if (1&(int)y)!=0?, if so the result is negative
1580 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1581 // abs(x)
1582 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1583 // abs(x)^y
1584 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1585 // -abs(x)^y
1586 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1587 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1588 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1589 // Set complex path fast result
1590 phi->init_req(2, signresult);
1591
1592 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1593 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1594 r->init_req(1,slow_path);
1595 phi->init_req(1,slow_result);
1596
1597 // Post merge
1598 set_control(_gvn.transform(r));
1599 record_for_igvn(r);
1600 result=_gvn.transform(phi);
1601 }
1602
1603 //-------------------
1604 //result=(result.isNaN())? uncommon_trap():result;
1605 // Check: If isNaN() by checking result!=result? then go to Strict Math
1606 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1607 // Build the boolean node
1608 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1609
1610 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1611 // End the current control-flow path
1612 push_pair(x);
1613 push_pair(y);
1614 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1615 // to handle. Recompile without intrinsifying Math.pow.
1616 uncommon_trap(Deoptimization::Reason_intrinsic,
1617 Deoptimization::Action_make_not_entrant);
1618 }
1619
1620 C->set_has_split_ifs(true); // Has chance for split-if optimization
1621
1622 push_pair(result);
1623
1624 return true;
1625 }
1626
1627 //------------------------------inline_trans-------------------------------------
1628 // Inline transcendental instructions, if possible. The Intel hardware gets
1629 // these right, no funny corner cases missed.
1630 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1631 _sp += arg_size(); // restore stack pointer
1632 Node* arg = pop_math_arg();
1633 Node* trans = NULL;
1634
1635 switch (id) {
1636 case vmIntrinsics::_dlog:
1637 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1638 break;
1639 case vmIntrinsics::_dlog10:
1640 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1641 break;
1642 default:
1643 assert(false, "bad intrinsic was passed in");
1644 return false;
1645 }
1646
1647 // Push result back on JVM stack
1648 push_pair(trans);
1649 return true;
1650 }
1651
1652 //------------------------------runtime_math-----------------------------
1653 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1654 Node* a = NULL;
1655 Node* b = NULL;
1656
1657 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1658 "must be (DD)D or (D)D type");
1659
1660 // Inputs
1661 _sp += arg_size(); // restore stack pointer
1662 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1663 b = pop_math_arg();
1664 }
1665 a = pop_math_arg();
1666
1667 const TypePtr* no_memory_effects = NULL;
1668 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1669 no_memory_effects,
1670 a, top(), b, b ? top() : NULL);
1671 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1672 #ifdef ASSERT
1673 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1674 assert(value_top == top(), "second value must be top");
1675 #endif
1676
1677 push_pair(value);
1678 return true;
1679 }
1680
1681 //------------------------------inline_math_native-----------------------------
1682 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1683 switch (id) {
1684 // These intrinsics are not properly supported on all hardware
1685 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1686 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1687 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1688 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1689 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1690 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1691
1692 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1693 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1694 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1695 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1696
1697 // These intrinsics are supported on all hardware
1698 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1699 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1700
1701 // These intrinsics don't work on X86. The ad implementation doesn't
1702 // handle NaN's properly. Instead of returning infinity, the ad
1703 // implementation returns a NaN on overflow. See bug: 6304089
1704 // Once the ad implementations are fixed, change the code below
1705 // to match the intrinsics above
1706
1707 case vmIntrinsics::_dexp: return
1708 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1709 case vmIntrinsics::_dpow: return
1710 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1711
1712 // These intrinsics are not yet correctly implemented
1713 case vmIntrinsics::_datan2:
1714 return false;
1715
1716 default:
1717 ShouldNotReachHere();
1718 return false;
1719 }
1720 }
1721
1722 static bool is_simple_name(Node* n) {
1723 return (n->req() == 1 // constant
1724 || (n->is_Type() && n->as_Type()->type()->singleton())
1725 || n->is_Proj() // parameter or return value
1726 || n->is_Phi() // local of some sort
1727 );
1728 }
1729
1730 //----------------------------inline_min_max-----------------------------------
1731 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1732 push(generate_min_max(id, argument(0), argument(1)));
1733
1734 return true;
1735 }
1736
1737 Node*
1738 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1739 // These are the candidate return value:
1740 Node* xvalue = x0;
1741 Node* yvalue = y0;
1742
1743 if (xvalue == yvalue) {
1744 return xvalue;
1745 }
1746
1747 bool want_max = (id == vmIntrinsics::_max);
1748
1749 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1750 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1751 if (txvalue == NULL || tyvalue == NULL) return top();
1752 // This is not really necessary, but it is consistent with a
1753 // hypothetical MaxINode::Value method:
1754 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1755
1756 // %%% This folding logic should (ideally) be in a different place.
1757 // Some should be inside IfNode, and there to be a more reliable
1758 // transformation of ?: style patterns into cmoves. We also want
1759 // more powerful optimizations around cmove and min/max.
1760
1761 // Try to find a dominating comparison of these guys.
1762 // It can simplify the index computation for Arrays.copyOf
1763 // and similar uses of System.arraycopy.
1764 // First, compute the normalized version of CmpI(x, y).
1765 int cmp_op = Op_CmpI;
1766 Node* xkey = xvalue;
1767 Node* ykey = yvalue;
1768 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1769 if (ideal_cmpxy->is_Cmp()) {
1770 // E.g., if we have CmpI(length - offset, count),
1771 // it might idealize to CmpI(length, count + offset)
1772 cmp_op = ideal_cmpxy->Opcode();
1773 xkey = ideal_cmpxy->in(1);
1774 ykey = ideal_cmpxy->in(2);
1775 }
1776
1777 // Start by locating any relevant comparisons.
1778 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1779 Node* cmpxy = NULL;
1780 Node* cmpyx = NULL;
1781 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1782 Node* cmp = start_from->fast_out(k);
1783 if (cmp->outcnt() > 0 && // must have prior uses
1784 cmp->in(0) == NULL && // must be context-independent
1785 cmp->Opcode() == cmp_op) { // right kind of compare
1786 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1787 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1788 }
1789 }
1790
1791 const int NCMPS = 2;
1792 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1793 int cmpn;
1794 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1795 if (cmps[cmpn] != NULL) break; // find a result
1796 }
1797 if (cmpn < NCMPS) {
1798 // Look for a dominating test that tells us the min and max.
1799 int depth = 0; // Limit search depth for speed
1800 Node* dom = control();
1801 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1802 if (++depth >= 100) break;
1803 Node* ifproj = dom;
1804 if (!ifproj->is_Proj()) continue;
1805 Node* iff = ifproj->in(0);
1806 if (!iff->is_If()) continue;
1807 Node* bol = iff->in(1);
1808 if (!bol->is_Bool()) continue;
1809 Node* cmp = bol->in(1);
1810 if (cmp == NULL) continue;
1811 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1812 if (cmps[cmpn] == cmp) break;
1813 if (cmpn == NCMPS) continue;
1814 BoolTest::mask btest = bol->as_Bool()->_test._test;
1815 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1816 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1817 // At this point, we know that 'x btest y' is true.
1818 switch (btest) {
1819 case BoolTest::eq:
1820 // They are proven equal, so we can collapse the min/max.
1821 // Either value is the answer. Choose the simpler.
1822 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1823 return yvalue;
1824 return xvalue;
1825 case BoolTest::lt: // x < y
1826 case BoolTest::le: // x <= y
1827 return (want_max ? yvalue : xvalue);
1828 case BoolTest::gt: // x > y
1829 case BoolTest::ge: // x >= y
1830 return (want_max ? xvalue : yvalue);
1831 }
1832 }
1833 }
1834
1835 // We failed to find a dominating test.
1836 // Let's pick a test that might GVN with prior tests.
1837 Node* best_bol = NULL;
1838 BoolTest::mask best_btest = BoolTest::illegal;
1839 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1840 Node* cmp = cmps[cmpn];
1841 if (cmp == NULL) continue;
1842 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1843 Node* bol = cmp->fast_out(j);
1844 if (!bol->is_Bool()) continue;
1845 BoolTest::mask btest = bol->as_Bool()->_test._test;
1846 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1847 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1848 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1849 best_bol = bol->as_Bool();
1850 best_btest = btest;
1851 }
1852 }
1853 }
1854
1855 Node* answer_if_true = NULL;
1856 Node* answer_if_false = NULL;
1857 switch (best_btest) {
1858 default:
1859 if (cmpxy == NULL)
1860 cmpxy = ideal_cmpxy;
1861 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1862 // and fall through:
1863 case BoolTest::lt: // x < y
1864 case BoolTest::le: // x <= y
1865 answer_if_true = (want_max ? yvalue : xvalue);
1866 answer_if_false = (want_max ? xvalue : yvalue);
1867 break;
1868 case BoolTest::gt: // x > y
1869 case BoolTest::ge: // x >= y
1870 answer_if_true = (want_max ? xvalue : yvalue);
1871 answer_if_false = (want_max ? yvalue : xvalue);
1872 break;
1873 }
1874
1875 jint hi, lo;
1876 if (want_max) {
1877 // We can sharpen the minimum.
1878 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1879 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1880 } else {
1881 // We can sharpen the maximum.
1882 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1883 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1884 }
1885
1886 // Use a flow-free graph structure, to avoid creating excess control edges
1887 // which could hinder other optimizations.
1888 // Since Math.min/max is often used with arraycopy, we want
1889 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1890 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1891 answer_if_false, answer_if_true,
1892 TypeInt::make(lo, hi, widen));
1893
1894 return _gvn.transform(cmov);
1895
1896 /*
1897 // This is not as desirable as it may seem, since Min and Max
1898 // nodes do not have a full set of optimizations.
1899 // And they would interfere, anyway, with 'if' optimizations
1900 // and with CMoveI canonical forms.
1901 switch (id) {
1902 case vmIntrinsics::_min:
1903 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1904 case vmIntrinsics::_max:
1905 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1906 default:
1907 ShouldNotReachHere();
1908 }
1909 */
1910 }
1911
1912 inline int
1913 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1914 const TypePtr* base_type = TypePtr::NULL_PTR;
1915 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1916 if (base_type == NULL) {
1917 // Unknown type.
1918 return Type::AnyPtr;
1919 } else if (base_type == TypePtr::NULL_PTR) {
1920 // Since this is a NULL+long form, we have to switch to a rawptr.
1921 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1922 offset = MakeConX(0);
1923 return Type::RawPtr;
1924 } else if (base_type->base() == Type::RawPtr) {
1925 return Type::RawPtr;
1926 } else if (base_type->isa_oopptr()) {
1927 // Base is never null => always a heap address.
1928 if (base_type->ptr() == TypePtr::NotNull) {
1929 return Type::OopPtr;
1930 }
1931 // Offset is small => always a heap address.
1932 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1933 if (offset_type != NULL &&
1934 base_type->offset() == 0 && // (should always be?)
1935 offset_type->_lo >= 0 &&
1936 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1937 return Type::OopPtr;
1938 }
1939 // Otherwise, it might either be oop+off or NULL+addr.
1940 return Type::AnyPtr;
1941 } else {
1942 // No information:
1943 return Type::AnyPtr;
1944 }
1945 }
1946
1947 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1948 int kind = classify_unsafe_addr(base, offset);
1949 if (kind == Type::RawPtr) {
1950 return basic_plus_adr(top(), base, offset);
1951 } else {
1952 return basic_plus_adr(base, offset);
1953 }
1954 }
1955
1956 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1957 // inline int Integer.numberOfLeadingZeros(int)
1958 // inline int Long.numberOfLeadingZeros(long)
1959 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1960 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1961 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1962 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1963 _sp += arg_size(); // restore stack pointer
1964 switch (id) {
1965 case vmIntrinsics::_numberOfLeadingZeros_i:
1966 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1967 break;
1968 case vmIntrinsics::_numberOfLeadingZeros_l:
1969 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1970 break;
1971 default:
1972 ShouldNotReachHere();
1973 }
1974 return true;
1975 }
1976
1977 //-------------------inline_numberOfTrailingZeros_int/long----------------------
1978 // inline int Integer.numberOfTrailingZeros(int)
1979 // inline int Long.numberOfTrailingZeros(long)
1980 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
1981 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
1982 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
1983 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
1984 _sp += arg_size(); // restore stack pointer
1985 switch (id) {
1986 case vmIntrinsics::_numberOfTrailingZeros_i:
1987 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
1988 break;
1989 case vmIntrinsics::_numberOfTrailingZeros_l:
1990 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
1991 break;
1992 default:
1993 ShouldNotReachHere();
1994 }
1995 return true;
1996 }
1997
1998 //----------------------------inline_bitCount_int/long-----------------------
1999 // inline int Integer.bitCount(int)
2000 // inline int Long.bitCount(long)
2001 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2002 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2003 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2004 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2005 _sp += arg_size(); // restore stack pointer
2006 switch (id) {
2007 case vmIntrinsics::_bitCount_i:
2008 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2009 break;
2010 case vmIntrinsics::_bitCount_l:
2011 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2012 break;
2013 default:
2014 ShouldNotReachHere();
2015 }
2016 return true;
2017 }
2018
2019 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2020 // inline Integer.reverseBytes(int)
2021 // inline Long.reverseBytes(long)
2022 // inline Character.reverseBytes(char)
2023 // inline Short.reverseBytes(short)
2024 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2025 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2026 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2027 "not reverse Bytes");
2028 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2029 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2030 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2031 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2032 _sp += arg_size(); // restore stack pointer
2033 switch (id) {
2034 case vmIntrinsics::_reverseBytes_i:
2035 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2036 break;
2037 case vmIntrinsics::_reverseBytes_l:
2038 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2039 break;
2040 case vmIntrinsics::_reverseBytes_c:
2041 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2042 break;
2043 case vmIntrinsics::_reverseBytes_s:
2044 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2045 break;
2046 default:
2047 ;
2048 }
2049 return true;
2050 }
2051
2052 //----------------------------inline_unsafe_access----------------------------
2053
2054 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2055
2056 // Interpret Unsafe.fieldOffset cookies correctly:
2057 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2058
2059 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2060 if (callee()->is_static()) return false; // caller must have the capability!
2061
2062 #ifndef PRODUCT
2063 {
2064 ResourceMark rm;
2065 // Check the signatures.
2066 ciSignature* sig = signature();
2067 #ifdef ASSERT
2068 if (!is_store) {
2069 // Object getObject(Object base, int/long offset), etc.
2070 BasicType rtype = sig->return_type()->basic_type();
2071 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2072 rtype = T_ADDRESS; // it is really a C void*
2073 assert(rtype == type, "getter must return the expected value");
2074 if (!is_native_ptr) {
2075 assert(sig->count() == 2, "oop getter has 2 arguments");
2076 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2077 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2078 } else {
2079 assert(sig->count() == 1, "native getter has 1 argument");
2080 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2081 }
2082 } else {
2083 // void putObject(Object base, int/long offset, Object x), etc.
2084 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2085 if (!is_native_ptr) {
2086 assert(sig->count() == 3, "oop putter has 3 arguments");
2087 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2088 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2089 } else {
2090 assert(sig->count() == 2, "native putter has 2 arguments");
2091 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2092 }
2093 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2094 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2095 vtype = T_ADDRESS; // it is really a C void*
2096 assert(vtype == type, "putter must accept the expected value");
2097 }
2098 #endif // ASSERT
2099 }
2100 #endif //PRODUCT
2101
2102 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2103
2104 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2105
2106 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2107 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2108
2109 debug_only(int saved_sp = _sp);
2110 _sp += nargs;
2111
2112 Node* val;
2113 debug_only(val = (Node*)(uintptr_t)-1);
2114
2115
2116 if (is_store) {
2117 // Get the value being stored. (Pop it first; it was pushed last.)
2118 switch (type) {
2119 case T_DOUBLE:
2120 case T_LONG:
2121 case T_ADDRESS:
2122 val = pop_pair();
2123 break;
2124 default:
2125 val = pop();
2126 }
2127 }
2128
2129 // Build address expression. See the code in inline_unsafe_prefetch.
2130 Node *adr;
2131 Node *heap_base_oop = top();
2132 if (!is_native_ptr) {
2133 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2134 Node* offset = pop_pair();
2135 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2136 Node* base = pop();
2137 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2138 // to be plain byte offsets, which are also the same as those accepted
2139 // by oopDesc::field_base.
2140 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2141 "fieldOffset must be byte-scaled");
2142 // 32-bit machines ignore the high half!
2143 offset = ConvL2X(offset);
2144 adr = make_unsafe_address(base, offset);
2145 heap_base_oop = base;
2146 } else {
2147 Node* ptr = pop_pair();
2148 // Adjust Java long to machine word:
2149 ptr = ConvL2X(ptr);
2150 adr = make_unsafe_address(NULL, ptr);
2151 }
2152
2153 // Pop receiver last: it was pushed first.
2154 Node *receiver = pop();
2155
2156 assert(saved_sp == _sp, "must have correct argument count");
2157
2158 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2159
2160 // First guess at the value type.
2161 const Type *value_type = Type::get_const_basic_type(type);
2162
2163 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2164 // there was not enough information to nail it down.
2165 Compile::AliasType* alias_type = C->alias_type(adr_type);
2166 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2167
2168 // We will need memory barriers unless we can determine a unique
2169 // alias category for this reference. (Note: If for some reason
2170 // the barriers get omitted and the unsafe reference begins to "pollute"
2171 // the alias analysis of the rest of the graph, either Compile::can_alias
2172 // or Compile::must_alias will throw a diagnostic assert.)
2173 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2174
2175 if (!is_store && type == T_OBJECT) {
2176 // Attempt to infer a sharper value type from the offset and base type.
2177 ciKlass* sharpened_klass = NULL;
2178
2179 // See if it is an instance field, with an object type.
2180 if (alias_type->field() != NULL) {
2181 assert(!is_native_ptr, "native pointer op cannot use a java address");
2182 if (alias_type->field()->type()->is_klass()) {
2183 sharpened_klass = alias_type->field()->type()->as_klass();
2184 }
2185 }
2186
2187 // See if it is a narrow oop array.
2188 if (adr_type->isa_aryptr()) {
2189 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2190 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2191 if (elem_type != NULL) {
2192 sharpened_klass = elem_type->klass();
2193 }
2194 }
2195 }
2196
2197 if (sharpened_klass != NULL) {
2198 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2199
2200 // Sharpen the value type.
2201 value_type = tjp;
2202
2203 #ifndef PRODUCT
2204 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2205 tty->print(" from base type: "); adr_type->dump();
2206 tty->print(" sharpened value: "); value_type->dump();
2207 }
2208 #endif
2209 }
2210 }
2211
2212 // Null check on self without removing any arguments. The argument
2213 // null check technically happens in the wrong place, which can lead to
2214 // invalid stack traces when the primitive is inlined into a method
2215 // which handles NullPointerExceptions.
2216 _sp += nargs;
2217 do_null_check(receiver, T_OBJECT);
2218 _sp -= nargs;
2219 if (stopped()) {
2220 return true;
2221 }
2222 // Heap pointers get a null-check from the interpreter,
2223 // as a courtesy. However, this is not guaranteed by Unsafe,
2224 // and it is not possible to fully distinguish unintended nulls
2225 // from intended ones in this API.
2226
2227 if (is_volatile) {
2228 // We need to emit leading and trailing CPU membars (see below) in
2229 // addition to memory membars when is_volatile. This is a little
2230 // too strong, but avoids the need to insert per-alias-type
2231 // volatile membars (for stores; compare Parse::do_put_xxx), which
2232 // we cannot do effectively here because we probably only have a
2233 // rough approximation of type.
2234 need_mem_bar = true;
2235 // For Stores, place a memory ordering barrier now.
2236 if (is_store)
2237 insert_mem_bar(Op_MemBarRelease);
2238 }
2239
2240 // Memory barrier to prevent normal and 'unsafe' accesses from
2241 // bypassing each other. Happens after null checks, so the
2242 // exception paths do not take memory state from the memory barrier,
2243 // so there's no problems making a strong assert about mixing users
2244 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2245 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2246 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2247
2248 if (!is_store) {
2249 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2250 // load value and push onto stack
2251 switch (type) {
2252 case T_BOOLEAN:
2253 case T_CHAR:
2254 case T_BYTE:
2255 case T_SHORT:
2256 case T_INT:
2257 case T_FLOAT:
2258 case T_OBJECT:
2259 push( p );
2260 break;
2261 case T_ADDRESS:
2262 // Cast to an int type.
2263 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2264 p = ConvX2L(p);
2265 push_pair(p);
2266 break;
2267 case T_DOUBLE:
2268 case T_LONG:
2269 push_pair( p );
2270 break;
2271 default: ShouldNotReachHere();
2272 }
2273 } else {
2274 // place effect of store into memory
2275 switch (type) {
2276 case T_DOUBLE:
2277 val = dstore_rounding(val);
2278 break;
2279 case T_ADDRESS:
2280 // Repackage the long as a pointer.
2281 val = ConvL2X(val);
2282 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2283 break;
2284 }
2285
2286 if (type != T_OBJECT ) {
2287 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2288 } else {
2289 // Possibly an oop being stored to Java heap or native memory
2290 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2291 // oop to Java heap.
2292 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2293 } else {
2294 // We can't tell at compile time if we are storing in the Java heap or outside
2295 // of it. So we need to emit code to conditionally do the proper type of
2296 // store.
2297
2298 IdealKit ideal(gvn(), control(), merged_memory());
2299 #define __ ideal.
2300 // QQQ who knows what probability is here??
2301 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2302 // Sync IdealKit and graphKit.
2303 set_all_memory( __ merged_memory());
2304 set_control(__ ctrl());
2305 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2306 // Update IdealKit memory.
2307 __ set_all_memory(merged_memory());
2308 __ set_ctrl(control());
2309 } __ else_(); {
2310 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2311 } __ end_if();
2312 // Final sync IdealKit and GraphKit.
2313 sync_kit(ideal);
2314 #undef __
2315 }
2316 }
2317 }
2318
2319 if (is_volatile) {
2320 if (!is_store)
2321 insert_mem_bar(Op_MemBarAcquire);
2322 else
2323 insert_mem_bar(Op_MemBarVolatile);
2324 }
2325
2326 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2327
2328 return true;
2329 }
2330
2331 //----------------------------inline_unsafe_prefetch----------------------------
2332
2333 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2334 #ifndef PRODUCT
2335 {
2336 ResourceMark rm;
2337 // Check the signatures.
2338 ciSignature* sig = signature();
2339 #ifdef ASSERT
2340 // Object getObject(Object base, int/long offset), etc.
2341 BasicType rtype = sig->return_type()->basic_type();
2342 if (!is_native_ptr) {
2343 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2344 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2345 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2346 } else {
2347 assert(sig->count() == 1, "native prefetch has 1 argument");
2348 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2349 }
2350 #endif // ASSERT
2351 }
2352 #endif // !PRODUCT
2353
2354 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2355
2356 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2357 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2358
2359 debug_only(int saved_sp = _sp);
2360 _sp += nargs;
2361
2362 // Build address expression. See the code in inline_unsafe_access.
2363 Node *adr;
2364 if (!is_native_ptr) {
2365 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2366 Node* offset = pop_pair();
2367 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2368 Node* base = pop();
2369 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2370 // to be plain byte offsets, which are also the same as those accepted
2371 // by oopDesc::field_base.
2372 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2373 "fieldOffset must be byte-scaled");
2374 // 32-bit machines ignore the high half!
2375 offset = ConvL2X(offset);
2376 adr = make_unsafe_address(base, offset);
2377 } else {
2378 Node* ptr = pop_pair();
2379 // Adjust Java long to machine word:
2380 ptr = ConvL2X(ptr);
2381 adr = make_unsafe_address(NULL, ptr);
2382 }
2383
2384 if (is_static) {
2385 assert(saved_sp == _sp, "must have correct argument count");
2386 } else {
2387 // Pop receiver last: it was pushed first.
2388 Node *receiver = pop();
2389 assert(saved_sp == _sp, "must have correct argument count");
2390
2391 // Null check on self without removing any arguments. The argument
2392 // null check technically happens in the wrong place, which can lead to
2393 // invalid stack traces when the primitive is inlined into a method
2394 // which handles NullPointerExceptions.
2395 _sp += nargs;
2396 do_null_check(receiver, T_OBJECT);
2397 _sp -= nargs;
2398 if (stopped()) {
2399 return true;
2400 }
2401 }
2402
2403 // Generate the read or write prefetch
2404 Node *prefetch;
2405 if (is_store) {
2406 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2407 } else {
2408 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2409 }
2410 prefetch->init_req(0, control());
2411 set_i_o(_gvn.transform(prefetch));
2412
2413 return true;
2414 }
2415
2416 //----------------------------inline_unsafe_CAS----------------------------
2417
2418 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2419 // This basic scheme here is the same as inline_unsafe_access, but
2420 // differs in enough details that combining them would make the code
2421 // overly confusing. (This is a true fact! I originally combined
2422 // them, but even I was confused by it!) As much code/comments as
2423 // possible are retained from inline_unsafe_access though to make
2424 // the correspondences clearer. - dl
2425
2426 if (callee()->is_static()) return false; // caller must have the capability!
2427
2428 #ifndef PRODUCT
2429 {
2430 ResourceMark rm;
2431 // Check the signatures.
2432 ciSignature* sig = signature();
2433 #ifdef ASSERT
2434 BasicType rtype = sig->return_type()->basic_type();
2435 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2436 assert(sig->count() == 4, "CAS has 4 arguments");
2437 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2438 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2439 #endif // ASSERT
2440 }
2441 #endif //PRODUCT
2442
2443 // number of stack slots per value argument (1 or 2)
2444 int type_words = type2size[type];
2445
2446 // Cannot inline wide CAS on machines that don't support it natively
2447 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2448 return false;
2449
2450 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2451
2452 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2453 int nargs = 1 + 1 + 2 + type_words + type_words;
2454
2455 // pop arguments: newval, oldval, offset, base, and receiver
2456 debug_only(int saved_sp = _sp);
2457 _sp += nargs;
2458 Node* newval = (type_words == 1) ? pop() : pop_pair();
2459 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2460 Node *offset = pop_pair();
2461 Node *base = pop();
2462 Node *receiver = pop();
2463 assert(saved_sp == _sp, "must have correct argument count");
2464
2465 // Null check receiver.
2466 _sp += nargs;
2467 do_null_check(receiver, T_OBJECT);
2468 _sp -= nargs;
2469 if (stopped()) {
2470 return true;
2471 }
2472
2473 // Build field offset expression.
2474 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2475 // to be plain byte offsets, which are also the same as those accepted
2476 // by oopDesc::field_base.
2477 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2478 // 32-bit machines ignore the high half of long offsets
2479 offset = ConvL2X(offset);
2480 Node* adr = make_unsafe_address(base, offset);
2481 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2482
2483 // (Unlike inline_unsafe_access, there seems no point in trying
2484 // to refine types. Just use the coarse types here.
2485 const Type *value_type = Type::get_const_basic_type(type);
2486 Compile::AliasType* alias_type = C->alias_type(adr_type);
2487 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2488 int alias_idx = C->get_alias_index(adr_type);
2489
2490 // Memory-model-wise, a CAS acts like a little synchronized block,
2491 // so needs barriers on each side. These don't translate into
2492 // actual barriers on most machines, but we still need rest of
2493 // compiler to respect ordering.
2494
2495 insert_mem_bar(Op_MemBarRelease);
2496 insert_mem_bar(Op_MemBarCPUOrder);
2497
2498 // 4984716: MemBars must be inserted before this
2499 // memory node in order to avoid a false
2500 // dependency which will confuse the scheduler.
2501 Node *mem = memory(alias_idx);
2502
2503 // For now, we handle only those cases that actually exist: ints,
2504 // longs, and Object. Adding others should be straightforward.
2505 Node* cas;
2506 switch(type) {
2507 case T_INT:
2508 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2509 break;
2510 case T_LONG:
2511 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2512 break;
2513 case T_OBJECT:
2514 // reference stores need a store barrier.
2515 // (They don't if CAS fails, but it isn't worth checking.)
2516 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2517 #ifdef _LP64
2518 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2519 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2520 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2521 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2522 newval_enc, oldval_enc));
2523 } else
2524 #endif
2525 {
2526 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2527 }
2528 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2529 break;
2530 default:
2531 ShouldNotReachHere();
2532 break;
2533 }
2534
2535 // SCMemProjNodes represent the memory state of CAS. Their main
2536 // role is to prevent CAS nodes from being optimized away when their
2537 // results aren't used.
2538 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2539 set_memory(proj, alias_idx);
2540
2541 // Add the trailing membar surrounding the access
2542 insert_mem_bar(Op_MemBarCPUOrder);
2543 insert_mem_bar(Op_MemBarAcquire);
2544
2545 push(cas);
2546 return true;
2547 }
2548
2549 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2550 // This is another variant of inline_unsafe_access, differing in
2551 // that it always issues store-store ("release") barrier and ensures
2552 // store-atomicity (which only matters for "long").
2553
2554 if (callee()->is_static()) return false; // caller must have the capability!
2555
2556 #ifndef PRODUCT
2557 {
2558 ResourceMark rm;
2559 // Check the signatures.
2560 ciSignature* sig = signature();
2561 #ifdef ASSERT
2562 BasicType rtype = sig->return_type()->basic_type();
2563 assert(rtype == T_VOID, "must return void");
2564 assert(sig->count() == 3, "has 3 arguments");
2565 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2566 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2567 #endif // ASSERT
2568 }
2569 #endif //PRODUCT
2570
2571 // number of stack slots per value argument (1 or 2)
2572 int type_words = type2size[type];
2573
2574 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2575
2576 // Argument words: "this" plus oop plus offset plus value;
2577 int nargs = 1 + 1 + 2 + type_words;
2578
2579 // pop arguments: val, offset, base, and receiver
2580 debug_only(int saved_sp = _sp);
2581 _sp += nargs;
2582 Node* val = (type_words == 1) ? pop() : pop_pair();
2583 Node *offset = pop_pair();
2584 Node *base = pop();
2585 Node *receiver = pop();
2586 assert(saved_sp == _sp, "must have correct argument count");
2587
2588 // Null check receiver.
2589 _sp += nargs;
2590 do_null_check(receiver, T_OBJECT);
2591 _sp -= nargs;
2592 if (stopped()) {
2593 return true;
2594 }
2595
2596 // Build field offset expression.
2597 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2598 // 32-bit machines ignore the high half of long offsets
2599 offset = ConvL2X(offset);
2600 Node* adr = make_unsafe_address(base, offset);
2601 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2602 const Type *value_type = Type::get_const_basic_type(type);
2603 Compile::AliasType* alias_type = C->alias_type(adr_type);
2604
2605 insert_mem_bar(Op_MemBarRelease);
2606 insert_mem_bar(Op_MemBarCPUOrder);
2607 // Ensure that the store is atomic for longs:
2608 bool require_atomic_access = true;
2609 Node* store;
2610 if (type == T_OBJECT) // reference stores need a store barrier.
2611 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2612 else {
2613 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2614 }
2615 insert_mem_bar(Op_MemBarCPUOrder);
2616 return true;
2617 }
2618
2619 bool LibraryCallKit::inline_unsafe_allocate() {
2620 if (callee()->is_static()) return false; // caller must have the capability!
2621 int nargs = 1 + 1;
2622 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2623 null_check_receiver(callee()); // check then ignore argument(0)
2624 _sp += nargs; // set original stack for use by uncommon_trap
2625 Node* cls = do_null_check(argument(1), T_OBJECT);
2626 _sp -= nargs;
2627 if (stopped()) return true;
2628
2629 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2630 _sp += nargs; // set original stack for use by uncommon_trap
2631 kls = do_null_check(kls, T_OBJECT);
2632 _sp -= nargs;
2633 if (stopped()) return true; // argument was like int.class
2634
2635 // Note: The argument might still be an illegal value like
2636 // Serializable.class or Object[].class. The runtime will handle it.
2637 // But we must make an explicit check for initialization.
2638 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2639 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2640 Node* bits = intcon(instanceKlass::fully_initialized);
2641 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2642 // The 'test' is non-zero if we need to take a slow path.
2643
2644 Node* obj = new_instance(kls, test);
2645 push(obj);
2646
2647 return true;
2648 }
2649
2650 //------------------------inline_native_time_funcs--------------
2651 // inline code for System.currentTimeMillis() and System.nanoTime()
2652 // these have the same type and signature
2653 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2654 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2655 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2656 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2657 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2658 const TypePtr* no_memory_effects = NULL;
2659 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2660 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2661 #ifdef ASSERT
2662 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2663 assert(value_top == top(), "second value must be top");
2664 #endif
2665 push_pair(value);
2666 return true;
2667 }
2668
2669 //------------------------inline_native_currentThread------------------
2670 bool LibraryCallKit::inline_native_currentThread() {
2671 Node* junk = NULL;
2672 push(generate_current_thread(junk));
2673 return true;
2674 }
2675
2676 //------------------------inline_native_isInterrupted------------------
2677 bool LibraryCallKit::inline_native_isInterrupted() {
2678 const int nargs = 1+1; // receiver + boolean
2679 assert(nargs == arg_size(), "sanity");
2680 // Add a fast path to t.isInterrupted(clear_int):
2681 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2682 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2683 // So, in the common case that the interrupt bit is false,
2684 // we avoid making a call into the VM. Even if the interrupt bit
2685 // is true, if the clear_int argument is false, we avoid the VM call.
2686 // However, if the receiver is not currentThread, we must call the VM,
2687 // because there must be some locking done around the operation.
2688
2689 // We only go to the fast case code if we pass two guards.
2690 // Paths which do not pass are accumulated in the slow_region.
2691 RegionNode* slow_region = new (C, 1) RegionNode(1);
2692 record_for_igvn(slow_region);
2693 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2694 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2695 enum { no_int_result_path = 1,
2696 no_clear_result_path = 2,
2697 slow_result_path = 3
2698 };
2699
2700 // (a) Receiving thread must be the current thread.
2701 Node* rec_thr = argument(0);
2702 Node* tls_ptr = NULL;
2703 Node* cur_thr = generate_current_thread(tls_ptr);
2704 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2705 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2706
2707 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2708 if (!known_current_thread)
2709 generate_slow_guard(bol_thr, slow_region);
2710
2711 // (b) Interrupt bit on TLS must be false.
2712 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2713 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2714 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2715 // Set the control input on the field _interrupted read to prevent it floating up.
2716 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2717 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2718 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2719
2720 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2721
2722 // First fast path: if (!TLS._interrupted) return false;
2723 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2724 result_rgn->init_req(no_int_result_path, false_bit);
2725 result_val->init_req(no_int_result_path, intcon(0));
2726
2727 // drop through to next case
2728 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2729
2730 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2731 Node* clr_arg = argument(1);
2732 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2733 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2734 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2735
2736 // Second fast path: ... else if (!clear_int) return true;
2737 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2738 result_rgn->init_req(no_clear_result_path, false_arg);
2739 result_val->init_req(no_clear_result_path, intcon(1));
2740
2741 // drop through to next case
2742 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2743
2744 // (d) Otherwise, go to the slow path.
2745 slow_region->add_req(control());
2746 set_control( _gvn.transform(slow_region) );
2747
2748 if (stopped()) {
2749 // There is no slow path.
2750 result_rgn->init_req(slow_result_path, top());
2751 result_val->init_req(slow_result_path, top());
2752 } else {
2753 // non-virtual because it is a private non-static
2754 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2755
2756 Node* slow_val = set_results_for_java_call(slow_call);
2757 // this->control() comes from set_results_for_java_call
2758
2759 // If we know that the result of the slow call will be true, tell the optimizer!
2760 if (known_current_thread) slow_val = intcon(1);
2761
2762 Node* fast_io = slow_call->in(TypeFunc::I_O);
2763 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2764 // These two phis are pre-filled with copies of of the fast IO and Memory
2765 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2766 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2767
2768 result_rgn->init_req(slow_result_path, control());
2769 io_phi ->init_req(slow_result_path, i_o());
2770 mem_phi ->init_req(slow_result_path, reset_memory());
2771 result_val->init_req(slow_result_path, slow_val);
2772
2773 set_all_memory( _gvn.transform(mem_phi) );
2774 set_i_o( _gvn.transform(io_phi) );
2775 }
2776
2777 push_result(result_rgn, result_val);
2778 C->set_has_split_ifs(true); // Has chance for split-if optimization
2779
2780 return true;
2781 }
2782
2783 //---------------------------load_mirror_from_klass----------------------------
2784 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2785 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2786 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2787 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2788 }
2789
2790 //-----------------------load_klass_from_mirror_common-------------------------
2791 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2792 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2793 // and branch to the given path on the region.
2794 // If never_see_null, take an uncommon trap on null, so we can optimistically
2795 // compile for the non-null case.
2796 // If the region is NULL, force never_see_null = true.
2797 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2798 bool never_see_null,
2799 int nargs,
2800 RegionNode* region,
2801 int null_path,
2802 int offset) {
2803 if (region == NULL) never_see_null = true;
2804 Node* p = basic_plus_adr(mirror, offset);
2805 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2806 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2807 _sp += nargs; // any deopt will start just before call to enclosing method
2808 Node* null_ctl = top();
2809 kls = null_check_oop(kls, &null_ctl, never_see_null);
2810 if (region != NULL) {
2811 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2812 region->init_req(null_path, null_ctl);
2813 } else {
2814 assert(null_ctl == top(), "no loose ends");
2815 }
2816 _sp -= nargs;
2817 return kls;
2818 }
2819
2820 //--------------------(inline_native_Class_query helpers)---------------------
2821 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2822 // Fall through if (mods & mask) == bits, take the guard otherwise.
2823 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2824 // Branch around if the given klass has the given modifier bit set.
2825 // Like generate_guard, adds a new path onto the region.
2826 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2827 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2828 Node* mask = intcon(modifier_mask);
2829 Node* bits = intcon(modifier_bits);
2830 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2831 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2832 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2833 return generate_fair_guard(bol, region);
2834 }
2835 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2836 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2837 }
2838
2839 //-------------------------inline_native_Class_query-------------------
2840 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2841 int nargs = 1+0; // just the Class mirror, in most cases
2842 const Type* return_type = TypeInt::BOOL;
2843 Node* prim_return_value = top(); // what happens if it's a primitive class?
2844 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2845 bool expect_prim = false; // most of these guys expect to work on refs
2846
2847 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2848
2849 switch (id) {
2850 case vmIntrinsics::_isInstance:
2851 nargs = 1+1; // the Class mirror, plus the object getting queried about
2852 // nothing is an instance of a primitive type
2853 prim_return_value = intcon(0);
2854 break;
2855 case vmIntrinsics::_getModifiers:
2856 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2857 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2858 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2859 break;
2860 case vmIntrinsics::_isInterface:
2861 prim_return_value = intcon(0);
2862 break;
2863 case vmIntrinsics::_isArray:
2864 prim_return_value = intcon(0);
2865 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2866 break;
2867 case vmIntrinsics::_isPrimitive:
2868 prim_return_value = intcon(1);
2869 expect_prim = true; // obviously
2870 break;
2871 case vmIntrinsics::_getSuperclass:
2872 prim_return_value = null();
2873 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2874 break;
2875 case vmIntrinsics::_getComponentType:
2876 prim_return_value = null();
2877 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2878 break;
2879 case vmIntrinsics::_getClassAccessFlags:
2880 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2881 return_type = TypeInt::INT; // not bool! 6297094
2882 break;
2883 default:
2884 ShouldNotReachHere();
2885 }
2886
2887 Node* mirror = argument(0);
2888 Node* obj = (nargs <= 1)? top(): argument(1);
2889
2890 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2891 if (mirror_con == NULL) return false; // cannot happen?
2892
2893 #ifndef PRODUCT
2894 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2895 ciType* k = mirror_con->java_mirror_type();
2896 if (k) {
2897 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2898 k->print_name();
2899 tty->cr();
2900 }
2901 }
2902 #endif
2903
2904 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2905 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2906 record_for_igvn(region);
2907 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2908
2909 // The mirror will never be null of Reflection.getClassAccessFlags, however
2910 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2911 // if it is. See bug 4774291.
2912
2913 // For Reflection.getClassAccessFlags(), the null check occurs in
2914 // the wrong place; see inline_unsafe_access(), above, for a similar
2915 // situation.
2916 _sp += nargs; // set original stack for use by uncommon_trap
2917 mirror = do_null_check(mirror, T_OBJECT);
2918 _sp -= nargs;
2919 // If mirror or obj is dead, only null-path is taken.
2920 if (stopped()) return true;
2921
2922 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2923
2924 // Now load the mirror's klass metaobject, and null-check it.
2925 // Side-effects region with the control path if the klass is null.
2926 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2927 region, _prim_path);
2928 // If kls is null, we have a primitive mirror.
2929 phi->init_req(_prim_path, prim_return_value);
2930 if (stopped()) { push_result(region, phi); return true; }
2931
2932 Node* p; // handy temp
2933 Node* null_ctl;
2934
2935 // Now that we have the non-null klass, we can perform the real query.
2936 // For constant classes, the query will constant-fold in LoadNode::Value.
2937 Node* query_value = top();
2938 switch (id) {
2939 case vmIntrinsics::_isInstance:
2940 // nothing is an instance of a primitive type
2941 _sp += nargs; // gen_instanceof might do an uncommon trap
2942 query_value = gen_instanceof(obj, kls);
2943 _sp -= nargs;
2944 break;
2945
2946 case vmIntrinsics::_getModifiers:
2947 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2948 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2949 break;
2950
2951 case vmIntrinsics::_isInterface:
2952 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2953 if (generate_interface_guard(kls, region) != NULL)
2954 // A guard was added. If the guard is taken, it was an interface.
2955 phi->add_req(intcon(1));
2956 // If we fall through, it's a plain class.
2957 query_value = intcon(0);
2958 break;
2959
2960 case vmIntrinsics::_isArray:
2961 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2962 if (generate_array_guard(kls, region) != NULL)
2963 // A guard was added. If the guard is taken, it was an array.
2964 phi->add_req(intcon(1));
2965 // If we fall through, it's a plain class.
2966 query_value = intcon(0);
2967 break;
2968
2969 case vmIntrinsics::_isPrimitive:
2970 query_value = intcon(0); // "normal" path produces false
2971 break;
2972
2973 case vmIntrinsics::_getSuperclass:
2974 // The rules here are somewhat unfortunate, but we can still do better
2975 // with random logic than with a JNI call.
2976 // Interfaces store null or Object as _super, but must report null.
2977 // Arrays store an intermediate super as _super, but must report Object.
2978 // Other types can report the actual _super.
2979 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2980 if (generate_interface_guard(kls, region) != NULL)
2981 // A guard was added. If the guard is taken, it was an interface.
2982 phi->add_req(null());
2983 if (generate_array_guard(kls, region) != NULL)
2984 // A guard was added. If the guard is taken, it was an array.
2985 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2986 // If we fall through, it's a plain class. Get its _super.
2987 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2988 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2989 null_ctl = top();
2990 kls = null_check_oop(kls, &null_ctl);
2991 if (null_ctl != top()) {
2992 // If the guard is taken, Object.superClass is null (both klass and mirror).
2993 region->add_req(null_ctl);
2994 phi ->add_req(null());
2995 }
2996 if (!stopped()) {
2997 query_value = load_mirror_from_klass(kls);
2998 }
2999 break;
3000
3001 case vmIntrinsics::_getComponentType:
3002 if (generate_array_guard(kls, region) != NULL) {
3003 // Be sure to pin the oop load to the guard edge just created:
3004 Node* is_array_ctrl = region->in(region->req()-1);
3005 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3006 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3007 phi->add_req(cmo);
3008 }
3009 query_value = null(); // non-array case is null
3010 break;
3011
3012 case vmIntrinsics::_getClassAccessFlags:
3013 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3014 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3015 break;
3016
3017 default:
3018 ShouldNotReachHere();
3019 }
3020
3021 // Fall-through is the normal case of a query to a real class.
3022 phi->init_req(1, query_value);
3023 region->init_req(1, control());
3024
3025 push_result(region, phi);
3026 C->set_has_split_ifs(true); // Has chance for split-if optimization
3027
3028 return true;
3029 }
3030
3031 //--------------------------inline_native_subtype_check------------------------
3032 // This intrinsic takes the JNI calls out of the heart of
3033 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3034 bool LibraryCallKit::inline_native_subtype_check() {
3035 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3036
3037 // Pull both arguments off the stack.
3038 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3039 args[0] = argument(0);
3040 args[1] = argument(1);
3041 Node* klasses[2]; // corresponding Klasses: superk, subk
3042 klasses[0] = klasses[1] = top();
3043
3044 enum {
3045 // A full decision tree on {superc is prim, subc is prim}:
3046 _prim_0_path = 1, // {P,N} => false
3047 // {P,P} & superc!=subc => false
3048 _prim_same_path, // {P,P} & superc==subc => true
3049 _prim_1_path, // {N,P} => false
3050 _ref_subtype_path, // {N,N} & subtype check wins => true
3051 _both_ref_path, // {N,N} & subtype check loses => false
3052 PATH_LIMIT
3053 };
3054
3055 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3056 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3057 record_for_igvn(region);
3058
3059 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3060 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3061 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3062
3063 // First null-check both mirrors and load each mirror's klass metaobject.
3064 int which_arg;
3065 for (which_arg = 0; which_arg <= 1; which_arg++) {
3066 Node* arg = args[which_arg];
3067 _sp += nargs; // set original stack for use by uncommon_trap
3068 arg = do_null_check(arg, T_OBJECT);
3069 _sp -= nargs;
3070 if (stopped()) break;
3071 args[which_arg] = _gvn.transform(arg);
3072
3073 Node* p = basic_plus_adr(arg, class_klass_offset);
3074 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3075 klasses[which_arg] = _gvn.transform(kls);
3076 }
3077
3078 // Having loaded both klasses, test each for null.
3079 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3080 for (which_arg = 0; which_arg <= 1; which_arg++) {
3081 Node* kls = klasses[which_arg];
3082 Node* null_ctl = top();
3083 _sp += nargs; // set original stack for use by uncommon_trap
3084 kls = null_check_oop(kls, &null_ctl, never_see_null);
3085 _sp -= nargs;
3086 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3087 region->init_req(prim_path, null_ctl);
3088 if (stopped()) break;
3089 klasses[which_arg] = kls;
3090 }
3091
3092 if (!stopped()) {
3093 // now we have two reference types, in klasses[0..1]
3094 Node* subk = klasses[1]; // the argument to isAssignableFrom
3095 Node* superk = klasses[0]; // the receiver
3096 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3097 // now we have a successful reference subtype check
3098 region->set_req(_ref_subtype_path, control());
3099 }
3100
3101 // If both operands are primitive (both klasses null), then
3102 // we must return true when they are identical primitives.
3103 // It is convenient to test this after the first null klass check.
3104 set_control(region->in(_prim_0_path)); // go back to first null check
3105 if (!stopped()) {
3106 // Since superc is primitive, make a guard for the superc==subc case.
3107 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3108 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3109 generate_guard(bol_eq, region, PROB_FAIR);
3110 if (region->req() == PATH_LIMIT+1) {
3111 // A guard was added. If the added guard is taken, superc==subc.
3112 region->swap_edges(PATH_LIMIT, _prim_same_path);
3113 region->del_req(PATH_LIMIT);
3114 }
3115 region->set_req(_prim_0_path, control()); // Not equal after all.
3116 }
3117
3118 // these are the only paths that produce 'true':
3119 phi->set_req(_prim_same_path, intcon(1));
3120 phi->set_req(_ref_subtype_path, intcon(1));
3121
3122 // pull together the cases:
3123 assert(region->req() == PATH_LIMIT, "sane region");
3124 for (uint i = 1; i < region->req(); i++) {
3125 Node* ctl = region->in(i);
3126 if (ctl == NULL || ctl == top()) {
3127 region->set_req(i, top());
3128 phi ->set_req(i, top());
3129 } else if (phi->in(i) == NULL) {
3130 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3131 }
3132 }
3133
3134 set_control(_gvn.transform(region));
3135 push(_gvn.transform(phi));
3136
3137 return true;
3138 }
3139
3140 //---------------------generate_array_guard_common------------------------
3141 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3142 bool obj_array, bool not_array) {
3143 // If obj_array/non_array==false/false:
3144 // Branch around if the given klass is in fact an array (either obj or prim).
3145 // If obj_array/non_array==false/true:
3146 // Branch around if the given klass is not an array klass of any kind.
3147 // If obj_array/non_array==true/true:
3148 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3149 // If obj_array/non_array==true/false:
3150 // Branch around if the kls is an oop array (Object[] or subtype)
3151 //
3152 // Like generate_guard, adds a new path onto the region.
3153 jint layout_con = 0;
3154 Node* layout_val = get_layout_helper(kls, layout_con);
3155 if (layout_val == NULL) {
3156 bool query = (obj_array
3157 ? Klass::layout_helper_is_objArray(layout_con)
3158 : Klass::layout_helper_is_javaArray(layout_con));
3159 if (query == not_array) {
3160 return NULL; // never a branch
3161 } else { // always a branch
3162 Node* always_branch = control();
3163 if (region != NULL)
3164 region->add_req(always_branch);
3165 set_control(top());
3166 return always_branch;
3167 }
3168 }
3169 // Now test the correct condition.
3170 jint nval = (obj_array
3171 ? ((jint)Klass::_lh_array_tag_type_value
3172 << Klass::_lh_array_tag_shift)
3173 : Klass::_lh_neutral_value);
3174 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3175 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3176 // invert the test if we are looking for a non-array
3177 if (not_array) btest = BoolTest(btest).negate();
3178 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3179 return generate_fair_guard(bol, region);
3180 }
3181
3182
3183 //-----------------------inline_native_newArray--------------------------
3184 bool LibraryCallKit::inline_native_newArray() {
3185 int nargs = 2;
3186 Node* mirror = argument(0);
3187 Node* count_val = argument(1);
3188
3189 _sp += nargs; // set original stack for use by uncommon_trap
3190 mirror = do_null_check(mirror, T_OBJECT);
3191 _sp -= nargs;
3192 // If mirror or obj is dead, only null-path is taken.
3193 if (stopped()) return true;
3194
3195 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3196 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3197 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3198 TypeInstPtr::NOTNULL);
3199 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3200 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3201 TypePtr::BOTTOM);
3202
3203 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3204 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3205 nargs,
3206 result_reg, _slow_path);
3207 Node* normal_ctl = control();
3208 Node* no_array_ctl = result_reg->in(_slow_path);
3209
3210 // Generate code for the slow case. We make a call to newArray().
3211 set_control(no_array_ctl);
3212 if (!stopped()) {
3213 // Either the input type is void.class, or else the
3214 // array klass has not yet been cached. Either the
3215 // ensuing call will throw an exception, or else it
3216 // will cache the array klass for next time.
3217 PreserveJVMState pjvms(this);
3218 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3219 Node* slow_result = set_results_for_java_call(slow_call);
3220 // this->control() comes from set_results_for_java_call
3221 result_reg->set_req(_slow_path, control());
3222 result_val->set_req(_slow_path, slow_result);
3223 result_io ->set_req(_slow_path, i_o());
3224 result_mem->set_req(_slow_path, reset_memory());
3225 }
3226
3227 set_control(normal_ctl);
3228 if (!stopped()) {
3229 // Normal case: The array type has been cached in the java.lang.Class.
3230 // The following call works fine even if the array type is polymorphic.
3231 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3232 Node* obj = new_array(klass_node, count_val, nargs);
3233 result_reg->init_req(_normal_path, control());
3234 result_val->init_req(_normal_path, obj);
3235 result_io ->init_req(_normal_path, i_o());
3236 result_mem->init_req(_normal_path, reset_memory());
3237 }
3238
3239 // Return the combined state.
3240 set_i_o( _gvn.transform(result_io) );
3241 set_all_memory( _gvn.transform(result_mem) );
3242 push_result(result_reg, result_val);
3243 C->set_has_split_ifs(true); // Has chance for split-if optimization
3244
3245 return true;
3246 }
3247
3248 //----------------------inline_native_getLength--------------------------
3249 bool LibraryCallKit::inline_native_getLength() {
3250 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3251
3252 int nargs = 1;
3253 Node* array = argument(0);
3254
3255 _sp += nargs; // set original stack for use by uncommon_trap
3256 array = do_null_check(array, T_OBJECT);
3257 _sp -= nargs;
3258
3259 // If array is dead, only null-path is taken.
3260 if (stopped()) return true;
3261
3262 // Deoptimize if it is a non-array.
3263 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3264
3265 if (non_array != NULL) {
3266 PreserveJVMState pjvms(this);
3267 set_control(non_array);
3268 _sp += nargs; // push the arguments back on the stack
3269 uncommon_trap(Deoptimization::Reason_intrinsic,
3270 Deoptimization::Action_maybe_recompile);
3271 }
3272
3273 // If control is dead, only non-array-path is taken.
3274 if (stopped()) return true;
3275
3276 // The works fine even if the array type is polymorphic.
3277 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3278 push( load_array_length(array) );
3279
3280 C->set_has_split_ifs(true); // Has chance for split-if optimization
3281
3282 return true;
3283 }
3284
3285 //------------------------inline_array_copyOf----------------------------
3286 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3287 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3288
3289 // Restore the stack and pop off the arguments.
3290 int nargs = 3 + (is_copyOfRange? 1: 0);
3291 Node* original = argument(0);
3292 Node* start = is_copyOfRange? argument(1): intcon(0);
3293 Node* end = is_copyOfRange? argument(2): argument(1);
3294 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3295
3296 Node* newcopy;
3297
3298 //set the original stack and the reexecute bit for the interpreter to reexecute
3299 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3300 { PreserveReexecuteState preexecs(this);
3301 _sp += nargs;
3302 jvms()->set_should_reexecute(true);
3303
3304 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3305 original = do_null_check(original, T_OBJECT);
3306
3307 // Check if a null path was taken unconditionally.
3308 if (stopped()) return true;
3309
3310 Node* orig_length = load_array_length(original);
3311
3312 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3313 NULL, 0);
3314 klass_node = do_null_check(klass_node, T_OBJECT);
3315
3316 RegionNode* bailout = new (C, 1) RegionNode(1);
3317 record_for_igvn(bailout);
3318
3319 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3320 // Bail out if that is so.
3321 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3322 if (not_objArray != NULL) {
3323 // Improve the klass node's type from the new optimistic assumption:
3324 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3325 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3326 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3327 cast->init_req(0, control());
3328 klass_node = _gvn.transform(cast);
3329 }
3330
3331 // Bail out if either start or end is negative.
3332 generate_negative_guard(start, bailout, &start);
3333 generate_negative_guard(end, bailout, &end);
3334
3335 Node* length = end;
3336 if (_gvn.type(start) != TypeInt::ZERO) {
3337 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3338 }
3339
3340 // Bail out if length is negative.
3341 // ...Not needed, since the new_array will throw the right exception.
3342 //generate_negative_guard(length, bailout, &length);
3343
3344 if (bailout->req() > 1) {
3345 PreserveJVMState pjvms(this);
3346 set_control( _gvn.transform(bailout) );
3347 uncommon_trap(Deoptimization::Reason_intrinsic,
3348 Deoptimization::Action_maybe_recompile);
3349 }
3350
3351 if (!stopped()) {
3352
3353 // How many elements will we copy from the original?
3354 // The answer is MinI(orig_length - start, length).
3355 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3356 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3357
3358 const bool raw_mem_only = true;
3359 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3360
3361 // Generate a direct call to the right arraycopy function(s).
3362 // We know the copy is disjoint but we might not know if the
3363 // oop stores need checking.
3364 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3365 // This will fail a store-check if x contains any non-nulls.
3366 bool disjoint_bases = true;
3367 bool length_never_negative = true;
3368 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3369 original, start, newcopy, intcon(0), moved,
3370 disjoint_bases, length_never_negative);
3371 }
3372 } //original reexecute and sp are set back here
3373
3374 if(!stopped()) {
3375 push(newcopy);
3376 }
3377
3378 C->set_has_split_ifs(true); // Has chance for split-if optimization
3379
3380 return true;
3381 }
3382
3383
3384 //----------------------generate_virtual_guard---------------------------
3385 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3386 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3387 RegionNode* slow_region) {
3388 ciMethod* method = callee();
3389 int vtable_index = method->vtable_index();
3390 // Get the methodOop out of the appropriate vtable entry.
3391 int entry_offset = (instanceKlass::vtable_start_offset() +
3392 vtable_index*vtableEntry::size()) * wordSize +
3393 vtableEntry::method_offset_in_bytes();
3394 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3395 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3396
3397 // Compare the target method with the expected method (e.g., Object.hashCode).
3398 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3399
3400 Node* native_call = makecon(native_call_addr);
3401 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3402 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3403
3404 return generate_slow_guard(test_native, slow_region);
3405 }
3406
3407 //-----------------------generate_method_call----------------------------
3408 // Use generate_method_call to make a slow-call to the real
3409 // method if the fast path fails. An alternative would be to
3410 // use a stub like OptoRuntime::slow_arraycopy_Java.
3411 // This only works for expanding the current library call,
3412 // not another intrinsic. (E.g., don't use this for making an
3413 // arraycopy call inside of the copyOf intrinsic.)
3414 CallJavaNode*
3415 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3416 // When compiling the intrinsic method itself, do not use this technique.
3417 guarantee(callee() != C->method(), "cannot make slow-call to self");
3418
3419 ciMethod* method = callee();
3420 // ensure the JVMS we have will be correct for this call
3421 guarantee(method_id == method->intrinsic_id(), "must match");
3422
3423 const TypeFunc* tf = TypeFunc::make(method);
3424 int tfdc = tf->domain()->cnt();
3425 CallJavaNode* slow_call;
3426 if (is_static) {
3427 assert(!is_virtual, "");
3428 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3429 SharedRuntime::get_resolve_static_call_stub(),
3430 method, bci());
3431 } else if (is_virtual) {
3432 null_check_receiver(method);
3433 int vtable_index = methodOopDesc::invalid_vtable_index;
3434 if (UseInlineCaches) {
3435 // Suppress the vtable call
3436 } else {
3437 // hashCode and clone are not a miranda methods,
3438 // so the vtable index is fixed.
3439 // No need to use the linkResolver to get it.
3440 vtable_index = method->vtable_index();
3441 }
3442 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3443 SharedRuntime::get_resolve_virtual_call_stub(),
3444 method, vtable_index, bci());
3445 } else { // neither virtual nor static: opt_virtual
3446 null_check_receiver(method);
3447 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3448 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3449 method, bci());
3450 slow_call->set_optimized_virtual(true);
3451 }
3452 set_arguments_for_java_call(slow_call);
3453 set_edges_for_java_call(slow_call);
3454 return slow_call;
3455 }
3456
3457
3458 //------------------------------inline_native_hashcode--------------------
3459 // Build special case code for calls to hashCode on an object.
3460 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3461 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3462 assert(!(is_virtual && is_static), "either virtual, special, or static");
3463
3464 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3465
3466 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3467 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3468 TypeInt::INT);
3469 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3470 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3471 TypePtr::BOTTOM);
3472 Node* obj = NULL;
3473 if (!is_static) {
3474 // Check for hashing null object
3475 obj = null_check_receiver(callee());
3476 if (stopped()) return true; // unconditionally null
3477 result_reg->init_req(_null_path, top());
3478 result_val->init_req(_null_path, top());
3479 } else {
3480 // Do a null check, and return zero if null.
3481 // System.identityHashCode(null) == 0
3482 obj = argument(0);
3483 Node* null_ctl = top();
3484 obj = null_check_oop(obj, &null_ctl);
3485 result_reg->init_req(_null_path, null_ctl);
3486 result_val->init_req(_null_path, _gvn.intcon(0));
3487 }
3488
3489 // Unconditionally null? Then return right away.
3490 if (stopped()) {
3491 set_control( result_reg->in(_null_path) );
3492 if (!stopped())
3493 push( result_val ->in(_null_path) );
3494 return true;
3495 }
3496
3497 // After null check, get the object's klass.
3498 Node* obj_klass = load_object_klass(obj);
3499
3500 // This call may be virtual (invokevirtual) or bound (invokespecial).
3501 // For each case we generate slightly different code.
3502
3503 // We only go to the fast case code if we pass a number of guards. The
3504 // paths which do not pass are accumulated in the slow_region.
3505 RegionNode* slow_region = new (C, 1) RegionNode(1);
3506 record_for_igvn(slow_region);
3507
3508 // If this is a virtual call, we generate a funny guard. We pull out
3509 // the vtable entry corresponding to hashCode() from the target object.
3510 // If the target method which we are calling happens to be the native
3511 // Object hashCode() method, we pass the guard. We do not need this
3512 // guard for non-virtual calls -- the caller is known to be the native
3513 // Object hashCode().
3514 if (is_virtual) {
3515 generate_virtual_guard(obj_klass, slow_region);
3516 }
3517
3518 // Get the header out of the object, use LoadMarkNode when available
3519 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3520 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3521
3522 // Test the header to see if it is unlocked.
3523 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3524 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3525 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3526 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3527 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3528
3529 generate_slow_guard(test_unlocked, slow_region);
3530
3531 // Get the hash value and check to see that it has been properly assigned.
3532 // We depend on hash_mask being at most 32 bits and avoid the use of
3533 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3534 // vm: see markOop.hpp.
3535 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3536 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3537 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3538 // This hack lets the hash bits live anywhere in the mark object now, as long
3539 // as the shift drops the relevant bits into the low 32 bits. Note that
3540 // Java spec says that HashCode is an int so there's no point in capturing
3541 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3542 hshifted_header = ConvX2I(hshifted_header);
3543 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3544
3545 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3546 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3547 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3548
3549 generate_slow_guard(test_assigned, slow_region);
3550
3551 Node* init_mem = reset_memory();
3552 // fill in the rest of the null path:
3553 result_io ->init_req(_null_path, i_o());
3554 result_mem->init_req(_null_path, init_mem);
3555
3556 result_val->init_req(_fast_path, hash_val);
3557 result_reg->init_req(_fast_path, control());
3558 result_io ->init_req(_fast_path, i_o());
3559 result_mem->init_req(_fast_path, init_mem);
3560
3561 // Generate code for the slow case. We make a call to hashCode().
3562 set_control(_gvn.transform(slow_region));
3563 if (!stopped()) {
3564 // No need for PreserveJVMState, because we're using up the present state.
3565 set_all_memory(init_mem);
3566 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3567 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3568 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3569 Node* slow_result = set_results_for_java_call(slow_call);
3570 // this->control() comes from set_results_for_java_call
3571 result_reg->init_req(_slow_path, control());
3572 result_val->init_req(_slow_path, slow_result);
3573 result_io ->set_req(_slow_path, i_o());
3574 result_mem ->set_req(_slow_path, reset_memory());
3575 }
3576
3577 // Return the combined state.
3578 set_i_o( _gvn.transform(result_io) );
3579 set_all_memory( _gvn.transform(result_mem) );
3580 push_result(result_reg, result_val);
3581
3582 return true;
3583 }
3584
3585 //---------------------------inline_native_getClass----------------------------
3586 // Build special case code for calls to getClass on an object.
3587 bool LibraryCallKit::inline_native_getClass() {
3588 Node* obj = null_check_receiver(callee());
3589 if (stopped()) return true;
3590 push( load_mirror_from_klass(load_object_klass(obj)) );
3591 return true;
3592 }
3593
3594 //-----------------inline_native_Reflection_getCallerClass---------------------
3595 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3596 //
3597 // NOTE that this code must perform the same logic as
3598 // vframeStream::security_get_caller_frame in that it must skip
3599 // Method.invoke() and auxiliary frames.
3600
3601
3602
3603
3604 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3605 ciMethod* method = callee();
3606
3607 #ifndef PRODUCT
3608 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3609 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3610 }
3611 #endif
3612
3613 debug_only(int saved_sp = _sp);
3614
3615 // Argument words: (int depth)
3616 int nargs = 1;
3617
3618 _sp += nargs;
3619 Node* caller_depth_node = pop();
3620
3621 assert(saved_sp == _sp, "must have correct argument count");
3622
3623 // The depth value must be a constant in order for the runtime call
3624 // to be eliminated.
3625 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3626 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3627 #ifndef PRODUCT
3628 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3629 tty->print_cr(" Bailing out because caller depth was not a constant");
3630 }
3631 #endif
3632 return false;
3633 }
3634 // Note that the JVM state at this point does not include the
3635 // getCallerClass() frame which we are trying to inline. The
3636 // semantics of getCallerClass(), however, are that the "first"
3637 // frame is the getCallerClass() frame, so we subtract one from the
3638 // requested depth before continuing. We don't inline requests of
3639 // getCallerClass(0).
3640 int caller_depth = caller_depth_type->get_con() - 1;
3641 if (caller_depth < 0) {
3642 #ifndef PRODUCT
3643 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3644 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3645 }
3646 #endif
3647 return false;
3648 }
3649
3650 if (!jvms()->has_method()) {
3651 #ifndef PRODUCT
3652 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3653 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3654 }
3655 #endif
3656 return false;
3657 }
3658 int _depth = jvms()->depth(); // cache call chain depth
3659
3660 // Walk back up the JVM state to find the caller at the required
3661 // depth. NOTE that this code must perform the same logic as
3662 // vframeStream::security_get_caller_frame in that it must skip
3663 // Method.invoke() and auxiliary frames. Note also that depth is
3664 // 1-based (1 is the bottom of the inlining).
3665 int inlining_depth = _depth;
3666 JVMState* caller_jvms = NULL;
3667
3668 if (inlining_depth > 0) {
3669 caller_jvms = jvms();
3670 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3671 do {
3672 // The following if-tests should be performed in this order
3673 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3674 // Skip a Method.invoke() or auxiliary frame
3675 } else if (caller_depth > 0) {
3676 // Skip real frame
3677 --caller_depth;
3678 } else {
3679 // We're done: reached desired caller after skipping.
3680 break;
3681 }
3682 caller_jvms = caller_jvms->caller();
3683 --inlining_depth;
3684 } while (inlining_depth > 0);
3685 }
3686
3687 if (inlining_depth == 0) {
3688 #ifndef PRODUCT
3689 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3690 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3691 tty->print_cr(" JVM state at this point:");
3692 for (int i = _depth; i >= 1; i--) {
3693 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3694 }
3695 }
3696 #endif
3697 return false; // Reached end of inlining
3698 }
3699
3700 // Acquire method holder as java.lang.Class
3701 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3702 ciInstance* caller_mirror = caller_klass->java_mirror();
3703 // Push this as a constant
3704 push(makecon(TypeInstPtr::make(caller_mirror)));
3705 #ifndef PRODUCT
3706 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3707 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);
3708 tty->print_cr(" JVM state at this point:");
3709 for (int i = _depth; i >= 1; i--) {
3710 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3711 }
3712 }
3713 #endif
3714 return true;
3715 }
3716
3717 // Helper routine for above
3718 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3719 ciMethod* method = jvms->method();
3720
3721 // Is this the Method.invoke method itself?
3722 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3723 return true;
3724
3725 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3726 ciKlass* k = method->holder();
3727 if (k->is_instance_klass()) {
3728 ciInstanceKlass* ik = k->as_instance_klass();
3729 for (; ik != NULL; ik = ik->super()) {
3730 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3731 ik == env()->find_system_klass(ik->name())) {
3732 return true;
3733 }
3734 }
3735 }
3736 else if (method->is_method_handle_adapter()) {
3737 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3738 return true;
3739 }
3740
3741 return false;
3742 }
3743
3744 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3745 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3746 // computing it since there is no lookup field by name function in the
3747 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3748 // Using a static variable here is safe even if we have multiple compilation
3749 // threads because the offset is constant. At worst the same offset will be
3750 // computed and stored multiple
3751
3752 bool LibraryCallKit::inline_native_AtomicLong_get() {
3753 // Restore the stack and pop off the argument
3754 _sp+=1;
3755 Node *obj = pop();
3756
3757 // get the offset of the "value" field. Since the CI interfaces
3758 // does not provide a way to look up a field by name, we scan the bytecodes
3759 // to get the field index. We expect the first 2 instructions of the method
3760 // to be:
3761 // 0 aload_0
3762 // 1 getfield "value"
3763 ciMethod* method = callee();
3764 if (value_field_offset == -1)
3765 {
3766 ciField* value_field;
3767 ciBytecodeStream iter(method);
3768 Bytecodes::Code bc = iter.next();
3769
3770 if ((bc != Bytecodes::_aload_0) &&
3771 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3772 return false;
3773 bc = iter.next();
3774 if (bc != Bytecodes::_getfield)
3775 return false;
3776 bool ignore;
3777 value_field = iter.get_field(ignore);
3778 value_field_offset = value_field->offset_in_bytes();
3779 }
3780
3781 // Null check without removing any arguments.
3782 _sp++;
3783 obj = do_null_check(obj, T_OBJECT);
3784 _sp--;
3785 // Check for locking null object
3786 if (stopped()) return true;
3787
3788 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3789 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3790 int alias_idx = C->get_alias_index(adr_type);
3791
3792 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3793
3794 push_pair(result);
3795
3796 return true;
3797 }
3798
3799 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3800 // Restore the stack and pop off the arguments
3801 _sp+=5;
3802 Node *newVal = pop_pair();
3803 Node *oldVal = pop_pair();
3804 Node *obj = pop();
3805
3806 // we need the offset of the "value" field which was computed when
3807 // inlining the get() method. Give up if we don't have it.
3808 if (value_field_offset == -1)
3809 return false;
3810
3811 // Null check without removing any arguments.
3812 _sp+=5;
3813 obj = do_null_check(obj, T_OBJECT);
3814 _sp-=5;
3815 // Check for locking null object
3816 if (stopped()) return true;
3817
3818 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3819 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3820 int alias_idx = C->get_alias_index(adr_type);
3821
3822 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3823 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3824 set_memory(store_proj, alias_idx);
3825 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3826
3827 Node *result;
3828 // CMove node is not used to be able fold a possible check code
3829 // after attemptUpdate() call. This code could be transformed
3830 // into CMove node by loop optimizations.
3831 {
3832 RegionNode *r = new (C, 3) RegionNode(3);
3833 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3834
3835 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3836 Node *iftrue = opt_iff(r, iff);
3837 r->init_req(1, iftrue);
3838 result->init_req(1, intcon(1));
3839 result->init_req(2, intcon(0));
3840
3841 set_control(_gvn.transform(r));
3842 record_for_igvn(r);
3843
3844 C->set_has_split_ifs(true); // Has chance for split-if optimization
3845 }
3846
3847 push(_gvn.transform(result));
3848 return true;
3849 }
3850
3851 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3852 // restore the arguments
3853 _sp += arg_size();
3854
3855 switch (id) {
3856 case vmIntrinsics::_floatToRawIntBits:
3857 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3858 break;
3859
3860 case vmIntrinsics::_intBitsToFloat:
3861 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3862 break;
3863
3864 case vmIntrinsics::_doubleToRawLongBits:
3865 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3866 break;
3867
3868 case vmIntrinsics::_longBitsToDouble:
3869 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3870 break;
3871
3872 case vmIntrinsics::_doubleToLongBits: {
3873 Node* value = pop_pair();
3874
3875 // two paths (plus control) merge in a wood
3876 RegionNode *r = new (C, 3) RegionNode(3);
3877 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3878
3879 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3880 // Build the boolean node
3881 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3882
3883 // Branch either way.
3884 // NaN case is less traveled, which makes all the difference.
3885 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3886 Node *opt_isnan = _gvn.transform(ifisnan);
3887 assert( opt_isnan->is_If(), "Expect an IfNode");
3888 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3889 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3890
3891 set_control(iftrue);
3892
3893 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3894 Node *slow_result = longcon(nan_bits); // return NaN
3895 phi->init_req(1, _gvn.transform( slow_result ));
3896 r->init_req(1, iftrue);
3897
3898 // Else fall through
3899 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3900 set_control(iffalse);
3901
3902 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3903 r->init_req(2, iffalse);
3904
3905 // Post merge
3906 set_control(_gvn.transform(r));
3907 record_for_igvn(r);
3908
3909 Node* result = _gvn.transform(phi);
3910 assert(result->bottom_type()->isa_long(), "must be");
3911 push_pair(result);
3912
3913 C->set_has_split_ifs(true); // Has chance for split-if optimization
3914
3915 break;
3916 }
3917
3918 case vmIntrinsics::_floatToIntBits: {
3919 Node* value = pop();
3920
3921 // two paths (plus control) merge in a wood
3922 RegionNode *r = new (C, 3) RegionNode(3);
3923 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3924
3925 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3926 // Build the boolean node
3927 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3928
3929 // Branch either way.
3930 // NaN case is less traveled, which makes all the difference.
3931 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3932 Node *opt_isnan = _gvn.transform(ifisnan);
3933 assert( opt_isnan->is_If(), "Expect an IfNode");
3934 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3935 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3936
3937 set_control(iftrue);
3938
3939 static const jint nan_bits = 0x7fc00000;
3940 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3941 phi->init_req(1, _gvn.transform( slow_result ));
3942 r->init_req(1, iftrue);
3943
3944 // Else fall through
3945 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3946 set_control(iffalse);
3947
3948 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3949 r->init_req(2, iffalse);
3950
3951 // Post merge
3952 set_control(_gvn.transform(r));
3953 record_for_igvn(r);
3954
3955 Node* result = _gvn.transform(phi);
3956 assert(result->bottom_type()->isa_int(), "must be");
3957 push(result);
3958
3959 C->set_has_split_ifs(true); // Has chance for split-if optimization
3960
3961 break;
3962 }
3963
3964 default:
3965 ShouldNotReachHere();
3966 }
3967
3968 return true;
3969 }
3970
3971 #ifdef _LP64
3972 #define XTOP ,top() /*additional argument*/
3973 #else //_LP64
3974 #define XTOP /*no additional argument*/
3975 #endif //_LP64
3976
3977 //----------------------inline_unsafe_copyMemory-------------------------
3978 bool LibraryCallKit::inline_unsafe_copyMemory() {
3979 if (callee()->is_static()) return false; // caller must have the capability!
3980 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3981 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3982 null_check_receiver(callee()); // check then ignore argument(0)
3983 if (stopped()) return true;
3984
3985 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3986
3987 Node* src_ptr = argument(1);
3988 Node* src_off = ConvL2X(argument(2));
3989 assert(argument(3)->is_top(), "2nd half of long");
3990 Node* dst_ptr = argument(4);
3991 Node* dst_off = ConvL2X(argument(5));
3992 assert(argument(6)->is_top(), "2nd half of long");
3993 Node* size = ConvL2X(argument(7));
3994 assert(argument(8)->is_top(), "2nd half of long");
3995
3996 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3997 "fieldOffset must be byte-scaled");
3998
3999 Node* src = make_unsafe_address(src_ptr, src_off);
4000 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4001
4002 // Conservatively insert a memory barrier on all memory slices.
4003 // Do not let writes of the copy source or destination float below the copy.
4004 insert_mem_bar(Op_MemBarCPUOrder);
4005
4006 // Call it. Note that the length argument is not scaled.
4007 make_runtime_call(RC_LEAF|RC_NO_FP,
4008 OptoRuntime::fast_arraycopy_Type(),
4009 StubRoutines::unsafe_arraycopy(),
4010 "unsafe_arraycopy",
4011 TypeRawPtr::BOTTOM,
4012 src, dst, size XTOP);
4013
4014 // Do not let reads of the copy destination float above the copy.
4015 insert_mem_bar(Op_MemBarCPUOrder);
4016
4017 return true;
4018 }
4019
4020 //------------------------clone_coping-----------------------------------
4021 // Helper function for inline_native_clone.
4022 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4023 assert(obj_size != NULL, "");
4024 Node* raw_obj = alloc_obj->in(1);
4025 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4026
4027 if (ReduceBulkZeroing) {
4028 // We will be completely responsible for initializing this object -
4029 // mark Initialize node as complete.
4030 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4031 // The object was just allocated - there should be no any stores!
4032 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4033 }
4034
4035 // Copy the fastest available way.
4036 // TODO: generate fields copies for small objects instead.
4037 Node* src = obj;
4038 Node* dest = alloc_obj;
4039 Node* size = _gvn.transform(obj_size);
4040
4041 // Exclude the header but include array length to copy by 8 bytes words.
4042 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4043 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4044 instanceOopDesc::base_offset_in_bytes();
4045 // base_off:
4046 // 8 - 32-bit VM
4047 // 12 - 64-bit VM, compressed oops
4048 // 16 - 64-bit VM, normal oops
4049 if (base_off % BytesPerLong != 0) {
4050 assert(UseCompressedOops, "");
4051 if (is_array) {
4052 // Exclude length to copy by 8 bytes words.
4053 base_off += sizeof(int);
4054 } else {
4055 // Include klass to copy by 8 bytes words.
4056 base_off = instanceOopDesc::klass_offset_in_bytes();
4057 }
4058 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4059 }
4060 src = basic_plus_adr(src, base_off);
4061 dest = basic_plus_adr(dest, base_off);
4062
4063 // Compute the length also, if needed:
4064 Node* countx = size;
4065 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4066 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4067
4068 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4069 bool disjoint_bases = true;
4070 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4071 src, NULL, dest, NULL, countx);
4072
4073 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4074 if (card_mark) {
4075 assert(!is_array, "");
4076 // Put in store barrier for any and all oops we are sticking
4077 // into this object. (We could avoid this if we could prove
4078 // that the object type contains no oop fields at all.)
4079 Node* no_particular_value = NULL;
4080 Node* no_particular_field = NULL;
4081 int raw_adr_idx = Compile::AliasIdxRaw;
4082 post_barrier(control(),
4083 memory(raw_adr_type),
4084 alloc_obj,
4085 no_particular_field,
4086 raw_adr_idx,
4087 no_particular_value,
4088 T_OBJECT,
4089 false);
4090 }
4091
4092 // Do not let reads from the cloned object float above the arraycopy.
4093 insert_mem_bar(Op_MemBarCPUOrder);
4094 }
4095
4096 //------------------------inline_native_clone----------------------------
4097 // Here are the simple edge cases:
4098 // null receiver => normal trap
4099 // virtual and clone was overridden => slow path to out-of-line clone
4100 // not cloneable or finalizer => slow path to out-of-line Object.clone
4101 //
4102 // The general case has two steps, allocation and copying.
4103 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4104 //
4105 // Copying also has two cases, oop arrays and everything else.
4106 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4107 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4108 //
4109 // These steps fold up nicely if and when the cloned object's klass
4110 // can be sharply typed as an object array, a type array, or an instance.
4111 //
4112 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4113 int nargs = 1;
4114 PhiNode* result_val;
4115
4116 //set the original stack and the reexecute bit for the interpreter to reexecute
4117 //the bytecode that invokes Object.clone if deoptimization happens
4118 { PreserveReexecuteState preexecs(this);
4119 jvms()->set_should_reexecute(true);
4120
4121 //null_check_receiver will adjust _sp (push and pop)
4122 Node* obj = null_check_receiver(callee());
4123 if (stopped()) return true;
4124
4125 _sp += nargs;
4126
4127 Node* obj_klass = load_object_klass(obj);
4128 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4129 const TypeOopPtr* toop = ((tklass != NULL)
4130 ? tklass->as_instance_type()
4131 : TypeInstPtr::NOTNULL);
4132
4133 // Conservatively insert a memory barrier on all memory slices.
4134 // Do not let writes into the original float below the clone.
4135 insert_mem_bar(Op_MemBarCPUOrder);
4136
4137 // paths into result_reg:
4138 enum {
4139 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4140 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4141 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4142 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4143 PATH_LIMIT
4144 };
4145 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4146 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4147 TypeInstPtr::NOTNULL);
4148 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4149 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4150 TypePtr::BOTTOM);
4151 record_for_igvn(result_reg);
4152
4153 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4154 int raw_adr_idx = Compile::AliasIdxRaw;
4155 const bool raw_mem_only = true;
4156
4157
4158 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4159 if (array_ctl != NULL) {
4160 // It's an array.
4161 PreserveJVMState pjvms(this);
4162 set_control(array_ctl);
4163 Node* obj_length = load_array_length(obj);
4164 Node* obj_size = NULL;
4165 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4166 raw_mem_only, &obj_size);
4167
4168 if (!use_ReduceInitialCardMarks()) {
4169 // If it is an oop array, it requires very special treatment,
4170 // because card marking is required on each card of the array.
4171 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4172 if (is_obja != NULL) {
4173 PreserveJVMState pjvms2(this);
4174 set_control(is_obja);
4175 // Generate a direct call to the right arraycopy function(s).
4176 bool disjoint_bases = true;
4177 bool length_never_negative = true;
4178 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4179 obj, intcon(0), alloc_obj, intcon(0),
4180 obj_length,
4181 disjoint_bases, length_never_negative);
4182 result_reg->init_req(_objArray_path, control());
4183 result_val->init_req(_objArray_path, alloc_obj);
4184 result_i_o ->set_req(_objArray_path, i_o());
4185 result_mem ->set_req(_objArray_path, reset_memory());
4186 }
4187 }
4188 // Otherwise, there are no card marks to worry about.
4189 // (We can dispense with card marks if we know the allocation
4190 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4191 // causes the non-eden paths to take compensating steps to
4192 // simulate a fresh allocation, so that no further
4193 // card marks are required in compiled code to initialize
4194 // the object.)
4195
4196 if (!stopped()) {
4197 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4198
4199 // Present the results of the copy.
4200 result_reg->init_req(_array_path, control());
4201 result_val->init_req(_array_path, alloc_obj);
4202 result_i_o ->set_req(_array_path, i_o());
4203 result_mem ->set_req(_array_path, reset_memory());
4204 }
4205 }
4206
4207 // We only go to the instance fast case code if we pass a number of guards.
4208 // The paths which do not pass are accumulated in the slow_region.
4209 RegionNode* slow_region = new (C, 1) RegionNode(1);
4210 record_for_igvn(slow_region);
4211 if (!stopped()) {
4212 // It's an instance (we did array above). Make the slow-path tests.
4213 // If this is a virtual call, we generate a funny guard. We grab
4214 // the vtable entry corresponding to clone() from the target object.
4215 // If the target method which we are calling happens to be the
4216 // Object clone() method, we pass the guard. We do not need this
4217 // guard for non-virtual calls; the caller is known to be the native
4218 // Object clone().
4219 if (is_virtual) {
4220 generate_virtual_guard(obj_klass, slow_region);
4221 }
4222
4223 // The object must be cloneable and must not have a finalizer.
4224 // Both of these conditions may be checked in a single test.
4225 // We could optimize the cloneable test further, but we don't care.
4226 generate_access_flags_guard(obj_klass,
4227 // Test both conditions:
4228 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4229 // Must be cloneable but not finalizer:
4230 JVM_ACC_IS_CLONEABLE,
4231 slow_region);
4232 }
4233
4234 if (!stopped()) {
4235 // It's an instance, and it passed the slow-path tests.
4236 PreserveJVMState pjvms(this);
4237 Node* obj_size = NULL;
4238 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4239
4240 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4241
4242 // Present the results of the slow call.
4243 result_reg->init_req(_instance_path, control());
4244 result_val->init_req(_instance_path, alloc_obj);
4245 result_i_o ->set_req(_instance_path, i_o());
4246 result_mem ->set_req(_instance_path, reset_memory());
4247 }
4248
4249 // Generate code for the slow case. We make a call to clone().
4250 set_control(_gvn.transform(slow_region));
4251 if (!stopped()) {
4252 PreserveJVMState pjvms(this);
4253 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4254 Node* slow_result = set_results_for_java_call(slow_call);
4255 // this->control() comes from set_results_for_java_call
4256 result_reg->init_req(_slow_path, control());
4257 result_val->init_req(_slow_path, slow_result);
4258 result_i_o ->set_req(_slow_path, i_o());
4259 result_mem ->set_req(_slow_path, reset_memory());
4260 }
4261
4262 // Return the combined state.
4263 set_control( _gvn.transform(result_reg) );
4264 set_i_o( _gvn.transform(result_i_o) );
4265 set_all_memory( _gvn.transform(result_mem) );
4266 } //original reexecute and sp are set back here
4267
4268 push(_gvn.transform(result_val));
4269
4270 return true;
4271 }
4272
4273
4274 // constants for computing the copy function
4275 enum {
4276 COPYFUNC_UNALIGNED = 0,
4277 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4278 COPYFUNC_CONJOINT = 0,
4279 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4280 };
4281
4282 // Note: The condition "disjoint" applies also for overlapping copies
4283 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4284 static address
4285 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
4286 int selector =
4287 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4288 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4289
4290 #define RETURN_STUB(xxx_arraycopy) { \
4291 name = #xxx_arraycopy; \
4292 return StubRoutines::xxx_arraycopy(); }
4293
4294 switch (t) {
4295 case T_BYTE:
4296 case T_BOOLEAN:
4297 switch (selector) {
4298 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4299 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4300 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4301 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4302 }
4303 case T_CHAR:
4304 case T_SHORT:
4305 switch (selector) {
4306 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4307 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4308 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4309 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4310 }
4311 case T_INT:
4312 case T_FLOAT:
4313 switch (selector) {
4314 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4315 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4316 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4317 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4318 }
4319 case T_DOUBLE:
4320 case T_LONG:
4321 switch (selector) {
4322 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4323 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4324 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4325 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4326 }
4327 case T_ARRAY:
4328 case T_OBJECT:
4329 switch (selector) {
4330 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
4331 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
4332 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
4333 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
4334 }
4335 default:
4336 ShouldNotReachHere();
4337 return NULL;
4338 }
4339
4340 #undef RETURN_STUB
4341 }
4342
4343 //------------------------------basictype2arraycopy----------------------------
4344 address LibraryCallKit::basictype2arraycopy(BasicType t,
4345 Node* src_offset,
4346 Node* dest_offset,
4347 bool disjoint_bases,
4348 const char* &name) {
4349 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4350 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4351
4352 bool aligned = false;
4353 bool disjoint = disjoint_bases;
4354
4355 // if the offsets are the same, we can treat the memory regions as
4356 // disjoint, because either the memory regions are in different arrays,
4357 // or they are identical (which we can treat as disjoint.) We can also
4358 // treat a copy with a destination index less that the source index
4359 // as disjoint since a low->high copy will work correctly in this case.
4360 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4361 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4362 // both indices are constants
4363 int s_offs = src_offset_inttype->get_con();
4364 int d_offs = dest_offset_inttype->get_con();
4365 int element_size = type2aelembytes(t);
4366 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4367 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4368 if (s_offs >= d_offs) disjoint = true;
4369 } else if (src_offset == dest_offset && src_offset != NULL) {
4370 // This can occur if the offsets are identical non-constants.
4371 disjoint = true;
4372 }
4373
4374 return select_arraycopy_function(t, aligned, disjoint, name);
4375 }
4376
4377
4378 //------------------------------inline_arraycopy-----------------------
4379 bool LibraryCallKit::inline_arraycopy() {
4380 // Restore the stack and pop off the arguments.
4381 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4382 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4383
4384 Node *src = argument(0);
4385 Node *src_offset = argument(1);
4386 Node *dest = argument(2);
4387 Node *dest_offset = argument(3);
4388 Node *length = argument(4);
4389
4390 // Compile time checks. If any of these checks cannot be verified at compile time,
4391 // we do not make a fast path for this call. Instead, we let the call remain as it
4392 // is. The checks we choose to mandate at compile time are:
4393 //
4394 // (1) src and dest are arrays.
4395 const Type* src_type = src->Value(&_gvn);
4396 const Type* dest_type = dest->Value(&_gvn);
4397 const TypeAryPtr* top_src = src_type->isa_aryptr();
4398 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4399 if (top_src == NULL || top_src->klass() == NULL ||
4400 top_dest == NULL || top_dest->klass() == NULL) {
4401 // Conservatively insert a memory barrier on all memory slices.
4402 // Do not let writes into the source float below the arraycopy.
4403 insert_mem_bar(Op_MemBarCPUOrder);
4404
4405 // Call StubRoutines::generic_arraycopy stub.
4406 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4407 src, src_offset, dest, dest_offset, length);
4408
4409 // Do not let reads from the destination float above the arraycopy.
4410 // Since we cannot type the arrays, we don't know which slices
4411 // might be affected. We could restrict this barrier only to those
4412 // memory slices which pertain to array elements--but don't bother.
4413 if (!InsertMemBarAfterArraycopy)
4414 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4415 insert_mem_bar(Op_MemBarCPUOrder);
4416 return true;
4417 }
4418
4419 // (2) src and dest arrays must have elements of the same BasicType
4420 // Figure out the size and type of the elements we will be copying.
4421 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4422 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4423 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4424 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4425
4426 if (src_elem != dest_elem || dest_elem == T_VOID) {
4427 // The component types are not the same or are not recognized. Punt.
4428 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4429 generate_slow_arraycopy(TypePtr::BOTTOM,
4430 src, src_offset, dest, dest_offset, length);
4431 return true;
4432 }
4433
4434 //---------------------------------------------------------------------------
4435 // We will make a fast path for this call to arraycopy.
4436
4437 // We have the following tests left to perform:
4438 //
4439 // (3) src and dest must not be null.
4440 // (4) src_offset must not be negative.
4441 // (5) dest_offset must not be negative.
4442 // (6) length must not be negative.
4443 // (7) src_offset + length must not exceed length of src.
4444 // (8) dest_offset + length must not exceed length of dest.
4445 // (9) each element of an oop array must be assignable
4446
4447 RegionNode* slow_region = new (C, 1) RegionNode(1);
4448 record_for_igvn(slow_region);
4449
4450 // (3) operands must not be null
4451 // We currently perform our null checks with the do_null_check routine.
4452 // This means that the null exceptions will be reported in the caller
4453 // rather than (correctly) reported inside of the native arraycopy call.
4454 // This should be corrected, given time. We do our null check with the
4455 // stack pointer restored.
4456 _sp += nargs;
4457 src = do_null_check(src, T_ARRAY);
4458 dest = do_null_check(dest, T_ARRAY);
4459 _sp -= nargs;
4460
4461 // (4) src_offset must not be negative.
4462 generate_negative_guard(src_offset, slow_region);
4463
4464 // (5) dest_offset must not be negative.
4465 generate_negative_guard(dest_offset, slow_region);
4466
4467 // (6) length must not be negative (moved to generate_arraycopy()).
4468 // generate_negative_guard(length, slow_region);
4469
4470 // (7) src_offset + length must not exceed length of src.
4471 generate_limit_guard(src_offset, length,
4472 load_array_length(src),
4473 slow_region);
4474
4475 // (8) dest_offset + length must not exceed length of dest.
4476 generate_limit_guard(dest_offset, length,
4477 load_array_length(dest),
4478 slow_region);
4479
4480 // (9) each element of an oop array must be assignable
4481 // The generate_arraycopy subroutine checks this.
4482
4483 // This is where the memory effects are placed:
4484 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4485 generate_arraycopy(adr_type, dest_elem,
4486 src, src_offset, dest, dest_offset, length,
4487 false, false, slow_region);
4488
4489 return true;
4490 }
4491
4492 //-----------------------------generate_arraycopy----------------------
4493 // Generate an optimized call to arraycopy.
4494 // Caller must guard against non-arrays.
4495 // Caller must determine a common array basic-type for both arrays.
4496 // Caller must validate offsets against array bounds.
4497 // The slow_region has already collected guard failure paths
4498 // (such as out of bounds length or non-conformable array types).
4499 // The generated code has this shape, in general:
4500 //
4501 // if (length == 0) return // via zero_path
4502 // slowval = -1
4503 // if (types unknown) {
4504 // slowval = call generic copy loop
4505 // if (slowval == 0) return // via checked_path
4506 // } else if (indexes in bounds) {
4507 // if ((is object array) && !(array type check)) {
4508 // slowval = call checked copy loop
4509 // if (slowval == 0) return // via checked_path
4510 // } else {
4511 // call bulk copy loop
4512 // return // via fast_path
4513 // }
4514 // }
4515 // // adjust params for remaining work:
4516 // if (slowval != -1) {
4517 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4518 // }
4519 // slow_region:
4520 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4521 // return // via slow_call_path
4522 //
4523 // This routine is used from several intrinsics: System.arraycopy,
4524 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4525 //
4526 void
4527 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4528 BasicType basic_elem_type,
4529 Node* src, Node* src_offset,
4530 Node* dest, Node* dest_offset,
4531 Node* copy_length,
4532 bool disjoint_bases,
4533 bool length_never_negative,
4534 RegionNode* slow_region) {
4535
4536 if (slow_region == NULL) {
4537 slow_region = new(C,1) RegionNode(1);
4538 record_for_igvn(slow_region);
4539 }
4540
4541 Node* original_dest = dest;
4542 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4543 bool must_clear_dest = false;
4544
4545 // See if this is the initialization of a newly-allocated array.
4546 // If so, we will take responsibility here for initializing it to zero.
4547 // (Note: Because tightly_coupled_allocation performs checks on the
4548 // out-edges of the dest, we need to avoid making derived pointers
4549 // from it until we have checked its uses.)
4550 if (ReduceBulkZeroing
4551 && !ZeroTLAB // pointless if already zeroed
4552 && basic_elem_type != T_CONFLICT // avoid corner case
4553 && !_gvn.eqv_uncast(src, dest)
4554 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4555 != NULL)
4556 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4557 && alloc->maybe_set_complete(&_gvn)) {
4558 // "You break it, you buy it."
4559 InitializeNode* init = alloc->initialization();
4560 assert(init->is_complete(), "we just did this");
4561 assert(dest->is_CheckCastPP(), "sanity");
4562 assert(dest->in(0)->in(0) == init, "dest pinned");
4563 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4564 // From this point on, every exit path is responsible for
4565 // initializing any non-copied parts of the object to zero.
4566 must_clear_dest = true;
4567 } else {
4568 // No zeroing elimination here.
4569 alloc = NULL;
4570 //original_dest = dest;
4571 //must_clear_dest = false;
4572 }
4573
4574 // Results are placed here:
4575 enum { fast_path = 1, // normal void-returning assembly stub
4576 checked_path = 2, // special assembly stub with cleanup
4577 slow_call_path = 3, // something went wrong; call the VM
4578 zero_path = 4, // bypass when length of copy is zero
4579 bcopy_path = 5, // copy primitive array by 64-bit blocks
4580 PATH_LIMIT = 6
4581 };
4582 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4583 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4584 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4585 record_for_igvn(result_region);
4586 _gvn.set_type_bottom(result_i_o);
4587 _gvn.set_type_bottom(result_memory);
4588 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4589
4590 // The slow_control path:
4591 Node* slow_control;
4592 Node* slow_i_o = i_o();
4593 Node* slow_mem = memory(adr_type);
4594 debug_only(slow_control = (Node*) badAddress);
4595
4596 // Checked control path:
4597 Node* checked_control = top();
4598 Node* checked_mem = NULL;
4599 Node* checked_i_o = NULL;
4600 Node* checked_value = NULL;
4601
4602 if (basic_elem_type == T_CONFLICT) {
4603 assert(!must_clear_dest, "");
4604 Node* cv = generate_generic_arraycopy(adr_type,
4605 src, src_offset, dest, dest_offset,
4606 copy_length);
4607 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4608 checked_control = control();
4609 checked_i_o = i_o();
4610 checked_mem = memory(adr_type);
4611 checked_value = cv;
4612 set_control(top()); // no fast path
4613 }
4614
4615 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4616 if (not_pos != NULL) {
4617 PreserveJVMState pjvms(this);
4618 set_control(not_pos);
4619
4620 // (6) length must not be negative.
4621 if (!length_never_negative) {
4622 generate_negative_guard(copy_length, slow_region);
4623 }
4624
4625 // copy_length is 0.
4626 if (!stopped() && must_clear_dest) {
4627 Node* dest_length = alloc->in(AllocateNode::ALength);
4628 if (_gvn.eqv_uncast(copy_length, dest_length)
4629 || _gvn.find_int_con(dest_length, 1) <= 0) {
4630 // There is no zeroing to do. No need for a secondary raw memory barrier.
4631 } else {
4632 // Clear the whole thing since there are no source elements to copy.
4633 generate_clear_array(adr_type, dest, basic_elem_type,
4634 intcon(0), NULL,
4635 alloc->in(AllocateNode::AllocSize));
4636 // Use a secondary InitializeNode as raw memory barrier.
4637 // Currently it is needed only on this path since other
4638 // paths have stub or runtime calls as raw memory barriers.
4639 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4640 Compile::AliasIdxRaw,
4641 top())->as_Initialize();
4642 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4643 }
4644 }
4645
4646 // Present the results of the fast call.
4647 result_region->init_req(zero_path, control());
4648 result_i_o ->init_req(zero_path, i_o());
4649 result_memory->init_req(zero_path, memory(adr_type));
4650 }
4651
4652 if (!stopped() && must_clear_dest) {
4653 // We have to initialize the *uncopied* part of the array to zero.
4654 // The copy destination is the slice dest[off..off+len]. The other slices
4655 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4656 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4657 Node* dest_length = alloc->in(AllocateNode::ALength);
4658 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4659 copy_length) );
4660
4661 // If there is a head section that needs zeroing, do it now.
4662 if (find_int_con(dest_offset, -1) != 0) {
4663 generate_clear_array(adr_type, dest, basic_elem_type,
4664 intcon(0), dest_offset,
4665 NULL);
4666 }
4667
4668 // Next, perform a dynamic check on the tail length.
4669 // It is often zero, and we can win big if we prove this.
4670 // There are two wins: Avoid generating the ClearArray
4671 // with its attendant messy index arithmetic, and upgrade
4672 // the copy to a more hardware-friendly word size of 64 bits.
4673 Node* tail_ctl = NULL;
4674 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4675 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4676 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4677 tail_ctl = generate_slow_guard(bol_lt, NULL);
4678 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4679 }
4680
4681 // At this point, let's assume there is no tail.
4682 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4683 // There is no tail. Try an upgrade to a 64-bit copy.
4684 bool didit = false;
4685 { PreserveJVMState pjvms(this);
4686 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4687 src, src_offset, dest, dest_offset,
4688 dest_size);
4689 if (didit) {
4690 // Present the results of the block-copying fast call.
4691 result_region->init_req(bcopy_path, control());
4692 result_i_o ->init_req(bcopy_path, i_o());
4693 result_memory->init_req(bcopy_path, memory(adr_type));
4694 }
4695 }
4696 if (didit)
4697 set_control(top()); // no regular fast path
4698 }
4699
4700 // Clear the tail, if any.
4701 if (tail_ctl != NULL) {
4702 Node* notail_ctl = stopped() ? NULL : control();
4703 set_control(tail_ctl);
4704 if (notail_ctl == NULL) {
4705 generate_clear_array(adr_type, dest, basic_elem_type,
4706 dest_tail, NULL,
4707 dest_size);
4708 } else {
4709 // Make a local merge.
4710 Node* done_ctl = new(C,3) RegionNode(3);
4711 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4712 done_ctl->init_req(1, notail_ctl);
4713 done_mem->init_req(1, memory(adr_type));
4714 generate_clear_array(adr_type, dest, basic_elem_type,
4715 dest_tail, NULL,
4716 dest_size);
4717 done_ctl->init_req(2, control());
4718 done_mem->init_req(2, memory(adr_type));
4719 set_control( _gvn.transform(done_ctl) );
4720 set_memory( _gvn.transform(done_mem), adr_type );
4721 }
4722 }
4723 }
4724
4725 BasicType copy_type = basic_elem_type;
4726 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4727 if (!stopped() && copy_type == T_OBJECT) {
4728 // If src and dest have compatible element types, we can copy bits.
4729 // Types S[] and D[] are compatible if D is a supertype of S.
4730 //
4731 // If they are not, we will use checked_oop_disjoint_arraycopy,
4732 // which performs a fast optimistic per-oop check, and backs off
4733 // further to JVM_ArrayCopy on the first per-oop check that fails.
4734 // (Actually, we don't move raw bits only; the GC requires card marks.)
4735
4736 // Get the klassOop for both src and dest
4737 Node* src_klass = load_object_klass(src);
4738 Node* dest_klass = load_object_klass(dest);
4739
4740 // Generate the subtype check.
4741 // This might fold up statically, or then again it might not.
4742 //
4743 // Non-static example: Copying List<String>.elements to a new String[].
4744 // The backing store for a List<String> is always an Object[],
4745 // but its elements are always type String, if the generic types
4746 // are correct at the source level.
4747 //
4748 // Test S[] against D[], not S against D, because (probably)
4749 // the secondary supertype cache is less busy for S[] than S.
4750 // This usually only matters when D is an interface.
4751 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4752 // Plug failing path into checked_oop_disjoint_arraycopy
4753 if (not_subtype_ctrl != top()) {
4754 PreserveJVMState pjvms(this);
4755 set_control(not_subtype_ctrl);
4756 // (At this point we can assume disjoint_bases, since types differ.)
4757 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4758 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4759 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4760 Node* dest_elem_klass = _gvn.transform(n1);
4761 Node* cv = generate_checkcast_arraycopy(adr_type,
4762 dest_elem_klass,
4763 src, src_offset, dest, dest_offset,
4764 copy_length);
4765 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4766 checked_control = control();
4767 checked_i_o = i_o();
4768 checked_mem = memory(adr_type);
4769 checked_value = cv;
4770 }
4771 // At this point we know we do not need type checks on oop stores.
4772
4773 // Let's see if we need card marks:
4774 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4775 // If we do not need card marks, copy using the jint or jlong stub.
4776 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4777 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4778 "sizes agree");
4779 }
4780 }
4781
4782 if (!stopped()) {
4783 // Generate the fast path, if possible.
4784 PreserveJVMState pjvms(this);
4785 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4786 src, src_offset, dest, dest_offset,
4787 ConvI2X(copy_length));
4788
4789 // Present the results of the fast call.
4790 result_region->init_req(fast_path, control());
4791 result_i_o ->init_req(fast_path, i_o());
4792 result_memory->init_req(fast_path, memory(adr_type));
4793 }
4794
4795 // Here are all the slow paths up to this point, in one bundle:
4796 slow_control = top();
4797 if (slow_region != NULL)
4798 slow_control = _gvn.transform(slow_region);
4799 debug_only(slow_region = (RegionNode*)badAddress);
4800
4801 set_control(checked_control);
4802 if (!stopped()) {
4803 // Clean up after the checked call.
4804 // The returned value is either 0 or -1^K,
4805 // where K = number of partially transferred array elements.
4806 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4807 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4808 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4809
4810 // If it is 0, we are done, so transfer to the end.
4811 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4812 result_region->init_req(checked_path, checks_done);
4813 result_i_o ->init_req(checked_path, checked_i_o);
4814 result_memory->init_req(checked_path, checked_mem);
4815
4816 // If it is not zero, merge into the slow call.
4817 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4818 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4819 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4820 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4821 record_for_igvn(slow_reg2);
4822 slow_reg2 ->init_req(1, slow_control);
4823 slow_i_o2 ->init_req(1, slow_i_o);
4824 slow_mem2 ->init_req(1, slow_mem);
4825 slow_reg2 ->init_req(2, control());
4826 slow_i_o2 ->init_req(2, checked_i_o);
4827 slow_mem2 ->init_req(2, checked_mem);
4828
4829 slow_control = _gvn.transform(slow_reg2);
4830 slow_i_o = _gvn.transform(slow_i_o2);
4831 slow_mem = _gvn.transform(slow_mem2);
4832
4833 if (alloc != NULL) {
4834 // We'll restart from the very beginning, after zeroing the whole thing.
4835 // This can cause double writes, but that's OK since dest is brand new.
4836 // So we ignore the low 31 bits of the value returned from the stub.
4837 } else {
4838 // We must continue the copy exactly where it failed, or else
4839 // another thread might see the wrong number of writes to dest.
4840 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4841 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4842 slow_offset->init_req(1, intcon(0));
4843 slow_offset->init_req(2, checked_offset);
4844 slow_offset = _gvn.transform(slow_offset);
4845
4846 // Adjust the arguments by the conditionally incoming offset.
4847 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4848 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4849 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4850
4851 // Tweak the node variables to adjust the code produced below:
4852 src_offset = src_off_plus;
4853 dest_offset = dest_off_plus;
4854 copy_length = length_minus;
4855 }
4856 }
4857
4858 set_control(slow_control);
4859 if (!stopped()) {
4860 // Generate the slow path, if needed.
4861 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4862
4863 set_memory(slow_mem, adr_type);
4864 set_i_o(slow_i_o);
4865
4866 if (must_clear_dest) {
4867 generate_clear_array(adr_type, dest, basic_elem_type,
4868 intcon(0), NULL,
4869 alloc->in(AllocateNode::AllocSize));
4870 }
4871
4872 generate_slow_arraycopy(adr_type,
4873 src, src_offset, dest, dest_offset,
4874 copy_length);
4875
4876 result_region->init_req(slow_call_path, control());
4877 result_i_o ->init_req(slow_call_path, i_o());
4878 result_memory->init_req(slow_call_path, memory(adr_type));
4879 }
4880
4881 // Remove unused edges.
4882 for (uint i = 1; i < result_region->req(); i++) {
4883 if (result_region->in(i) == NULL)
4884 result_region->init_req(i, top());
4885 }
4886
4887 // Finished; return the combined state.
4888 set_control( _gvn.transform(result_region) );
4889 set_i_o( _gvn.transform(result_i_o) );
4890 set_memory( _gvn.transform(result_memory), adr_type );
4891
4892 // The memory edges above are precise in order to model effects around
4893 // array copies accurately to allow value numbering of field loads around
4894 // arraycopy. Such field loads, both before and after, are common in Java
4895 // collections and similar classes involving header/array data structures.
4896 //
4897 // But with low number of register or when some registers are used or killed
4898 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4899 // The next memory barrier is added to avoid it. If the arraycopy can be
4900 // optimized away (which it can, sometimes) then we can manually remove
4901 // the membar also.
4902 //
4903 // Do not let reads from the cloned object float above the arraycopy.
4904 if (InsertMemBarAfterArraycopy || alloc != NULL)
4905 insert_mem_bar(Op_MemBarCPUOrder);
4906 }
4907
4908
4909 // Helper function which determines if an arraycopy immediately follows
4910 // an allocation, with no intervening tests or other escapes for the object.
4911 AllocateArrayNode*
4912 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4913 RegionNode* slow_region) {
4914 if (stopped()) return NULL; // no fast path
4915 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4916
4917 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4918 if (alloc == NULL) return NULL;
4919
4920 Node* rawmem = memory(Compile::AliasIdxRaw);
4921 // Is the allocation's memory state untouched?
4922 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4923 // Bail out if there have been raw-memory effects since the allocation.
4924 // (Example: There might have been a call or safepoint.)
4925 return NULL;
4926 }
4927 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4928 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4929 return NULL;
4930 }
4931
4932 // There must be no unexpected observers of this allocation.
4933 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4934 Node* obs = ptr->fast_out(i);
4935 if (obs != this->map()) {
4936 return NULL;
4937 }
4938 }
4939
4940 // This arraycopy must unconditionally follow the allocation of the ptr.
4941 Node* alloc_ctl = ptr->in(0);
4942 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4943
4944 Node* ctl = control();
4945 while (ctl != alloc_ctl) {
4946 // There may be guards which feed into the slow_region.
4947 // Any other control flow means that we might not get a chance
4948 // to finish initializing the allocated object.
4949 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4950 IfNode* iff = ctl->in(0)->as_If();
4951 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4952 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4953 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4954 ctl = iff->in(0); // This test feeds the known slow_region.
4955 continue;
4956 }
4957 // One more try: Various low-level checks bottom out in
4958 // uncommon traps. If the debug-info of the trap omits
4959 // any reference to the allocation, as we've already
4960 // observed, then there can be no objection to the trap.
4961 bool found_trap = false;
4962 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4963 Node* obs = not_ctl->fast_out(j);
4964 if (obs->in(0) == not_ctl && obs->is_Call() &&
4965 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4966 found_trap = true; break;
4967 }
4968 }
4969 if (found_trap) {
4970 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4971 continue;
4972 }
4973 }
4974 return NULL;
4975 }
4976
4977 // If we get this far, we have an allocation which immediately
4978 // precedes the arraycopy, and we can take over zeroing the new object.
4979 // The arraycopy will finish the initialization, and provide
4980 // a new control state to which we will anchor the destination pointer.
4981
4982 return alloc;
4983 }
4984
4985 // Helper for initialization of arrays, creating a ClearArray.
4986 // It writes zero bits in [start..end), within the body of an array object.
4987 // The memory effects are all chained onto the 'adr_type' alias category.
4988 //
4989 // Since the object is otherwise uninitialized, we are free
4990 // to put a little "slop" around the edges of the cleared area,
4991 // as long as it does not go back into the array's header,
4992 // or beyond the array end within the heap.
4993 //
4994 // The lower edge can be rounded down to the nearest jint and the
4995 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4996 //
4997 // Arguments:
4998 // adr_type memory slice where writes are generated
4999 // dest oop of the destination array
5000 // basic_elem_type element type of the destination
5001 // slice_idx array index of first element to store
5002 // slice_len number of elements to store (or NULL)
5003 // dest_size total size in bytes of the array object
5004 //
5005 // Exactly one of slice_len or dest_size must be non-NULL.
5006 // If dest_size is non-NULL, zeroing extends to the end of the object.
5007 // If slice_len is non-NULL, the slice_idx value must be a constant.
5008 void
5009 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5010 Node* dest,
5011 BasicType basic_elem_type,
5012 Node* slice_idx,
5013 Node* slice_len,
5014 Node* dest_size) {
5015 // one or the other but not both of slice_len and dest_size:
5016 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5017 if (slice_len == NULL) slice_len = top();
5018 if (dest_size == NULL) dest_size = top();
5019
5020 // operate on this memory slice:
5021 Node* mem = memory(adr_type); // memory slice to operate on
5022
5023 // scaling and rounding of indexes:
5024 int scale = exact_log2(type2aelembytes(basic_elem_type));
5025 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5026 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5027 int bump_bit = (-1 << scale) & BytesPerInt;
5028
5029 // determine constant starts and ends
5030 const intptr_t BIG_NEG = -128;
5031 assert(BIG_NEG + 2*abase < 0, "neg enough");
5032 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5033 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5034 if (slice_len_con == 0) {
5035 return; // nothing to do here
5036 }
5037 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5038 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5039 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5040 assert(end_con < 0, "not two cons");
5041 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5042 BytesPerLong);
5043 }
5044
5045 if (start_con >= 0 && end_con >= 0) {
5046 // Constant start and end. Simple.
5047 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5048 start_con, end_con, &_gvn);
5049 } else if (start_con >= 0 && dest_size != top()) {
5050 // Constant start, pre-rounded end after the tail of the array.
5051 Node* end = dest_size;
5052 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5053 start_con, end, &_gvn);
5054 } else if (start_con >= 0 && slice_len != top()) {
5055 // Constant start, non-constant end. End needs rounding up.
5056 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5057 intptr_t end_base = abase + (slice_idx_con << scale);
5058 int end_round = (-1 << scale) & (BytesPerLong - 1);
5059 Node* end = ConvI2X(slice_len);
5060 if (scale != 0)
5061 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5062 end_base += end_round;
5063 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5064 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5065 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5066 start_con, end, &_gvn);
5067 } else if (start_con < 0 && dest_size != top()) {
5068 // Non-constant start, pre-rounded end after the tail of the array.
5069 // This is almost certainly a "round-to-end" operation.
5070 Node* start = slice_idx;
5071 start = ConvI2X(start);
5072 if (scale != 0)
5073 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5074 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5075 if ((bump_bit | clear_low) != 0) {
5076 int to_clear = (bump_bit | clear_low);
5077 // Align up mod 8, then store a jint zero unconditionally
5078 // just before the mod-8 boundary.
5079 if (((abase + bump_bit) & ~to_clear) - bump_bit
5080 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5081 bump_bit = 0;
5082 assert((abase & to_clear) == 0, "array base must be long-aligned");
5083 } else {
5084 // Bump 'start' up to (or past) the next jint boundary:
5085 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5086 assert((abase & clear_low) == 0, "array base must be int-aligned");
5087 }
5088 // Round bumped 'start' down to jlong boundary in body of array.
5089 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5090 if (bump_bit != 0) {
5091 // Store a zero to the immediately preceding jint:
5092 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5093 Node* p1 = basic_plus_adr(dest, x1);
5094 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5095 mem = _gvn.transform(mem);
5096 }
5097 }
5098 Node* end = dest_size; // pre-rounded
5099 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5100 start, end, &_gvn);
5101 } else {
5102 // Non-constant start, unrounded non-constant end.
5103 // (Nobody zeroes a random midsection of an array using this routine.)
5104 ShouldNotReachHere(); // fix caller
5105 }
5106
5107 // Done.
5108 set_memory(mem, adr_type);
5109 }
5110
5111
5112 bool
5113 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5114 BasicType basic_elem_type,
5115 AllocateNode* alloc,
5116 Node* src, Node* src_offset,
5117 Node* dest, Node* dest_offset,
5118 Node* dest_size) {
5119 // See if there is an advantage from block transfer.
5120 int scale = exact_log2(type2aelembytes(basic_elem_type));
5121 if (scale >= LogBytesPerLong)
5122 return false; // it is already a block transfer
5123
5124 // Look at the alignment of the starting offsets.
5125 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5126 const intptr_t BIG_NEG = -128;
5127 assert(BIG_NEG + 2*abase < 0, "neg enough");
5128
5129 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5130 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5131 if (src_off < 0 || dest_off < 0)
5132 // At present, we can only understand constants.
5133 return false;
5134
5135 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5136 // Non-aligned; too bad.
5137 // One more chance: Pick off an initial 32-bit word.
5138 // This is a common case, since abase can be odd mod 8.
5139 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5140 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5141 Node* sptr = basic_plus_adr(src, src_off);
5142 Node* dptr = basic_plus_adr(dest, dest_off);
5143 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5144 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5145 src_off += BytesPerInt;
5146 dest_off += BytesPerInt;
5147 } else {
5148 return false;
5149 }
5150 }
5151 assert(src_off % BytesPerLong == 0, "");
5152 assert(dest_off % BytesPerLong == 0, "");
5153
5154 // Do this copy by giant steps.
5155 Node* sptr = basic_plus_adr(src, src_off);
5156 Node* dptr = basic_plus_adr(dest, dest_off);
5157 Node* countx = dest_size;
5158 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5159 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5160
5161 bool disjoint_bases = true; // since alloc != NULL
5162 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5163 sptr, NULL, dptr, NULL, countx);
5164
5165 return true;
5166 }
5167
5168
5169 // Helper function; generates code for the slow case.
5170 // We make a call to a runtime method which emulates the native method,
5171 // but without the native wrapper overhead.
5172 void
5173 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5174 Node* src, Node* src_offset,
5175 Node* dest, Node* dest_offset,
5176 Node* copy_length) {
5177 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5178 OptoRuntime::slow_arraycopy_Type(),
5179 OptoRuntime::slow_arraycopy_Java(),
5180 "slow_arraycopy", adr_type,
5181 src, src_offset, dest, dest_offset,
5182 copy_length);
5183
5184 // Handle exceptions thrown by this fellow:
5185 make_slow_call_ex(call, env()->Throwable_klass(), false);
5186 }
5187
5188 // Helper function; generates code for cases requiring runtime checks.
5189 Node*
5190 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5191 Node* dest_elem_klass,
5192 Node* src, Node* src_offset,
5193 Node* dest, Node* dest_offset,
5194 Node* copy_length) {
5195 if (stopped()) return NULL;
5196
5197 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
5198 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5199 return NULL;
5200 }
5201
5202 // Pick out the parameters required to perform a store-check
5203 // for the target array. This is an optimistic check. It will
5204 // look in each non-null element's class, at the desired klass's
5205 // super_check_offset, for the desired klass.
5206 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5207 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5208 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5209 Node* check_offset = _gvn.transform(n3);
5210 Node* check_value = dest_elem_klass;
5211
5212 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5213 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5214
5215 // (We know the arrays are never conjoint, because their types differ.)
5216 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5217 OptoRuntime::checkcast_arraycopy_Type(),
5218 copyfunc_addr, "checkcast_arraycopy", adr_type,
5219 // five arguments, of which two are
5220 // intptr_t (jlong in LP64)
5221 src_start, dest_start,
5222 copy_length XTOP,
5223 check_offset XTOP,
5224 check_value);
5225
5226 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5227 }
5228
5229
5230 // Helper function; generates code for cases requiring runtime checks.
5231 Node*
5232 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5233 Node* src, Node* src_offset,
5234 Node* dest, Node* dest_offset,
5235 Node* copy_length) {
5236 if (stopped()) return NULL;
5237
5238 address copyfunc_addr = StubRoutines::generic_arraycopy();
5239 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5240 return NULL;
5241 }
5242
5243 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5244 OptoRuntime::generic_arraycopy_Type(),
5245 copyfunc_addr, "generic_arraycopy", adr_type,
5246 src, src_offset, dest, dest_offset, copy_length);
5247
5248 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5249 }
5250
5251 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5252 void
5253 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5254 BasicType basic_elem_type,
5255 bool disjoint_bases,
5256 Node* src, Node* src_offset,
5257 Node* dest, Node* dest_offset,
5258 Node* copy_length) {
5259 if (stopped()) return; // nothing to do
5260
5261 Node* src_start = src;
5262 Node* dest_start = dest;
5263 if (src_offset != NULL || dest_offset != NULL) {
5264 assert(src_offset != NULL && dest_offset != NULL, "");
5265 src_start = array_element_address(src, src_offset, basic_elem_type);
5266 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5267 }
5268
5269 // Figure out which arraycopy runtime method to call.
5270 const char* copyfunc_name = "arraycopy";
5271 address copyfunc_addr =
5272 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5273 disjoint_bases, copyfunc_name);
5274
5275 // Call it. Note that the count_ix value is not scaled to a byte-size.
5276 make_runtime_call(RC_LEAF|RC_NO_FP,
5277 OptoRuntime::fast_arraycopy_Type(),
5278 copyfunc_addr, copyfunc_name, adr_type,
5279 src_start, dest_start, copy_length XTOP);
5280 }