1 /*
2 * Copyright (c) 1999, 2011, 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 "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/parse.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/subnode.hpp"
39 #include "prims/nativeLookup.hpp"
40 #include "runtime/sharedRuntime.hpp"
41
42 class LibraryIntrinsic : public InlineCallGenerator {
43 // Extend the set of intrinsics known to the runtime:
44 public:
45 private:
46 bool _is_virtual;
47 vmIntrinsics::ID _intrinsic_id;
48
49 public:
50 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
51 : InlineCallGenerator(m),
52 _is_virtual(is_virtual),
53 _intrinsic_id(id)
54 {
55 }
56 virtual bool is_intrinsic() const { return true; }
57 virtual bool is_virtual() const { return _is_virtual; }
58 virtual JVMState* generate(JVMState* jvms);
59 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
60 };
61
62
63 // Local helper class for LibraryIntrinsic:
64 class LibraryCallKit : public GraphKit {
65 private:
66 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
67
68 public:
69 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
70 : GraphKit(caller),
71 _intrinsic(intrinsic)
72 {
73 }
74
75 ciMethod* caller() const { return jvms()->method(); }
76 int bci() const { return jvms()->bci(); }
77 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
79 ciMethod* callee() const { return _intrinsic->method(); }
80 ciSignature* signature() const { return callee()->signature(); }
81 int arg_size() const { return callee()->arg_size(); }
82
83 bool try_to_inline();
84
85 // Helper functions to inline natives
86 void push_result(RegionNode* region, PhiNode* value);
87 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
88 Node* generate_slow_guard(Node* test, RegionNode* region);
89 Node* generate_fair_guard(Node* test, RegionNode* region);
90 Node* generate_negative_guard(Node* index, RegionNode* region,
91 // resulting CastII of index:
92 Node* *pos_index = NULL);
93 Node* generate_nonpositive_guard(Node* index, bool never_negative,
94 // resulting CastII of index:
95 Node* *pos_index = NULL);
96 Node* generate_limit_guard(Node* offset, Node* subseq_length,
97 Node* array_length,
98 RegionNode* region);
99 Node* generate_current_thread(Node* &tls_output);
100 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
101 bool disjoint_bases, const char* &name, bool dest_uninitialized);
102 Node* load_mirror_from_klass(Node* klass);
103 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
104 int nargs,
105 RegionNode* region, int null_path,
106 int offset);
107 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
108 RegionNode* region, int null_path) {
109 int offset = java_lang_Class::klass_offset_in_bytes();
110 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
111 region, null_path,
112 offset);
113 }
114 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
115 int nargs,
116 RegionNode* region, int null_path) {
117 int offset = java_lang_Class::array_klass_offset_in_bytes();
118 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
119 region, null_path,
120 offset);
121 }
122 Node* generate_access_flags_guard(Node* kls,
123 int modifier_mask, int modifier_bits,
124 RegionNode* region);
125 Node* generate_interface_guard(Node* kls, RegionNode* region);
126 Node* generate_array_guard(Node* kls, RegionNode* region) {
127 return generate_array_guard_common(kls, region, false, false);
128 }
129 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
130 return generate_array_guard_common(kls, region, false, true);
131 }
132 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
133 return generate_array_guard_common(kls, region, true, false);
134 }
135 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
136 return generate_array_guard_common(kls, region, true, true);
137 }
138 Node* generate_array_guard_common(Node* kls, RegionNode* region,
139 bool obj_array, bool not_array);
140 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
141 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
142 bool is_virtual = false, bool is_static = false);
143 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
144 return generate_method_call(method_id, false, true);
145 }
146 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
147 return generate_method_call(method_id, true, false);
148 }
149
150 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
151 bool inline_string_compareTo();
152 bool inline_string_indexOf();
153 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
154 bool inline_string_equals();
155 Node* pop_math_arg();
156 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
157 bool inline_math_native(vmIntrinsics::ID id);
158 bool inline_trig(vmIntrinsics::ID id);
159 bool inline_trans(vmIntrinsics::ID id);
160 bool inline_abs(vmIntrinsics::ID id);
161 bool inline_sqrt(vmIntrinsics::ID id);
162 bool inline_pow(vmIntrinsics::ID id);
163 bool inline_exp(vmIntrinsics::ID id);
164 bool inline_min_max(vmIntrinsics::ID id);
165 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
166 // This returns Type::AnyPtr, RawPtr, or OopPtr.
167 int classify_unsafe_addr(Node* &base, Node* &offset);
168 Node* make_unsafe_address(Node* base, Node* offset);
169 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
170 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
171 bool inline_unsafe_allocate();
172 bool inline_unsafe_copyMemory();
173 bool inline_native_currentThread();
174 bool inline_native_time_funcs(bool isNano);
175 bool inline_native_isInterrupted();
176 bool inline_native_Class_query(vmIntrinsics::ID id);
177 bool inline_native_subtype_check();
178
179 bool inline_native_newArray();
180 bool inline_native_getLength();
181 bool inline_array_copyOf(bool is_copyOfRange);
182 bool inline_array_equals();
183 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
184 bool inline_native_clone(bool is_virtual);
185 bool inline_native_Reflection_getCallerClass();
186 bool inline_native_AtomicLong_get();
187 bool inline_native_AtomicLong_attemptUpdate();
188 bool is_method_invoke_or_aux_frame(JVMState* jvms);
189 // Helper function for inlining native object hash method
190 bool inline_native_hashcode(bool is_virtual, bool is_static);
191 bool inline_native_getClass();
192
193 // Helper functions for inlining arraycopy
194 bool inline_arraycopy();
195 void generate_arraycopy(const TypePtr* adr_type,
196 BasicType basic_elem_type,
197 Node* src, Node* src_offset,
198 Node* dest, Node* dest_offset,
199 Node* copy_length,
200 bool disjoint_bases = false,
201 bool length_never_negative = false,
202 RegionNode* slow_region = NULL);
203 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
204 RegionNode* slow_region);
205 void generate_clear_array(const TypePtr* adr_type,
206 Node* dest,
207 BasicType basic_elem_type,
208 Node* slice_off,
209 Node* slice_len,
210 Node* slice_end);
211 bool generate_block_arraycopy(const TypePtr* adr_type,
212 BasicType basic_elem_type,
213 AllocateNode* alloc,
214 Node* src, Node* src_offset,
215 Node* dest, Node* dest_offset,
216 Node* dest_size, bool dest_uninitialized);
217 void generate_slow_arraycopy(const TypePtr* adr_type,
218 Node* src, Node* src_offset,
219 Node* dest, Node* dest_offset,
220 Node* copy_length, bool dest_uninitialized);
221 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
222 Node* dest_elem_klass,
223 Node* src, Node* src_offset,
224 Node* dest, Node* dest_offset,
225 Node* copy_length, bool dest_uninitialized);
226 Node* generate_generic_arraycopy(const TypePtr* adr_type,
227 Node* src, Node* src_offset,
228 Node* dest, Node* dest_offset,
229 Node* copy_length, bool dest_uninitialized);
230 void generate_unchecked_arraycopy(const TypePtr* adr_type,
231 BasicType basic_elem_type,
232 bool disjoint_bases,
233 Node* src, Node* src_offset,
234 Node* dest, Node* dest_offset,
235 Node* copy_length, bool dest_uninitialized);
236 bool inline_unsafe_CAS(BasicType type);
237 bool inline_unsafe_ordered_store(BasicType type);
238 bool inline_fp_conversions(vmIntrinsics::ID id);
239 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
240 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
241 bool inline_bitCount(vmIntrinsics::ID id);
242 bool inline_reverseBytes(vmIntrinsics::ID id);
243 };
244
245
246 //---------------------------make_vm_intrinsic----------------------------
247 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
248 vmIntrinsics::ID id = m->intrinsic_id();
249 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
250
251 if (DisableIntrinsic[0] != '\0'
252 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
253 // disabled by a user request on the command line:
254 // example: -XX:DisableIntrinsic=_hashCode,_getClass
255 return NULL;
256 }
257
258 if (!m->is_loaded()) {
259 // do not attempt to inline unloaded methods
260 return NULL;
261 }
262
263 // Only a few intrinsics implement a virtual dispatch.
264 // They are expensive calls which are also frequently overridden.
265 if (is_virtual) {
266 switch (id) {
267 case vmIntrinsics::_hashCode:
268 case vmIntrinsics::_clone:
269 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
270 break;
271 default:
272 return NULL;
273 }
274 }
275
276 // -XX:-InlineNatives disables nearly all intrinsics:
277 if (!InlineNatives) {
278 switch (id) {
279 case vmIntrinsics::_indexOf:
280 case vmIntrinsics::_compareTo:
281 case vmIntrinsics::_equals:
282 case vmIntrinsics::_equalsC:
283 break; // InlineNatives does not control String.compareTo
284 default:
285 return NULL;
286 }
287 }
288
289 switch (id) {
290 case vmIntrinsics::_compareTo:
291 if (!SpecialStringCompareTo) return NULL;
292 break;
293 case vmIntrinsics::_indexOf:
294 if (!SpecialStringIndexOf) return NULL;
295 break;
296 case vmIntrinsics::_equals:
297 if (!SpecialStringEquals) return NULL;
298 break;
299 case vmIntrinsics::_equalsC:
300 if (!SpecialArraysEquals) return NULL;
301 break;
302 case vmIntrinsics::_arraycopy:
303 if (!InlineArrayCopy) return NULL;
304 break;
305 case vmIntrinsics::_copyMemory:
306 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
307 if (!InlineArrayCopy) return NULL;
308 break;
309 case vmIntrinsics::_hashCode:
310 if (!InlineObjectHash) return NULL;
311 break;
312 case vmIntrinsics::_clone:
313 case vmIntrinsics::_copyOf:
314 case vmIntrinsics::_copyOfRange:
315 if (!InlineObjectCopy) return NULL;
316 // These also use the arraycopy intrinsic mechanism:
317 if (!InlineArrayCopy) return NULL;
318 break;
319 case vmIntrinsics::_checkIndex:
320 // We do not intrinsify this. The optimizer does fine with it.
321 return NULL;
322
323 case vmIntrinsics::_get_AtomicLong:
324 case vmIntrinsics::_attemptUpdate:
325 if (!InlineAtomicLong) return NULL;
326 break;
327
328 case vmIntrinsics::_getCallerClass:
329 if (!UseNewReflection) return NULL;
330 if (!InlineReflectionGetCallerClass) return NULL;
331 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
332 break;
333
334 case vmIntrinsics::_bitCount_i:
335 case vmIntrinsics::_bitCount_l:
336 if (!UsePopCountInstruction) return NULL;
337 break;
338
339 default:
340 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
341 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
342 break;
343 }
344
345 // -XX:-InlineClassNatives disables natives from the Class class.
346 // The flag applies to all reflective calls, notably Array.newArray
347 // (visible to Java programmers as Array.newInstance).
348 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
349 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
350 if (!InlineClassNatives) return NULL;
351 }
352
353 // -XX:-InlineThreadNatives disables natives from the Thread class.
354 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
355 if (!InlineThreadNatives) return NULL;
356 }
357
358 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
359 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
360 m->holder()->name() == ciSymbol::java_lang_Float() ||
361 m->holder()->name() == ciSymbol::java_lang_Double()) {
362 if (!InlineMathNatives) return NULL;
363 }
364
365 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
366 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
367 if (!InlineUnsafeOps) return NULL;
368 }
369
370 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
371 }
372
373 //----------------------register_library_intrinsics-----------------------
374 // Initialize this file's data structures, for each Compile instance.
375 void Compile::register_library_intrinsics() {
376 // Nothing to do here.
377 }
378
379 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
380 LibraryCallKit kit(jvms, this);
381 Compile* C = kit.C;
382 int nodes = C->unique();
383 #ifndef PRODUCT
384 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
385 char buf[1000];
386 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
387 tty->print_cr("Intrinsic %s", str);
388 }
389 #endif
390 if (kit.try_to_inline()) {
391 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
392 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
393 }
394 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
395 if (C->log()) {
396 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
397 vmIntrinsics::name_at(intrinsic_id()),
398 (is_virtual() ? " virtual='1'" : ""),
399 C->unique() - nodes);
400 }
401 return kit.transfer_exceptions_into_jvms();
402 }
403
404 if (PrintIntrinsics) {
405 tty->print("Did not inline intrinsic %s%s at bci:%d in",
406 vmIntrinsics::name_at(intrinsic_id()),
407 (is_virtual() ? " (virtual)" : ""), kit.bci());
408 kit.caller()->print_short_name(tty);
409 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
410 }
411 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
412 return NULL;
413 }
414
415 bool LibraryCallKit::try_to_inline() {
416 // Handle symbolic names for otherwise undistinguished boolean switches:
417 const bool is_store = true;
418 const bool is_native_ptr = true;
419 const bool is_static = true;
420
421 switch (intrinsic_id()) {
422 case vmIntrinsics::_hashCode:
423 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
424 case vmIntrinsics::_identityHashCode:
425 return inline_native_hashcode(/*!virtual*/ false, is_static);
426 case vmIntrinsics::_getClass:
427 return inline_native_getClass();
428
429 case vmIntrinsics::_dsin:
430 case vmIntrinsics::_dcos:
431 case vmIntrinsics::_dtan:
432 case vmIntrinsics::_dabs:
433 case vmIntrinsics::_datan2:
434 case vmIntrinsics::_dsqrt:
435 case vmIntrinsics::_dexp:
436 case vmIntrinsics::_dlog:
437 case vmIntrinsics::_dlog10:
438 case vmIntrinsics::_dpow:
439 return inline_math_native(intrinsic_id());
440
441 case vmIntrinsics::_min:
442 case vmIntrinsics::_max:
443 return inline_min_max(intrinsic_id());
444
445 case vmIntrinsics::_arraycopy:
446 return inline_arraycopy();
447
448 case vmIntrinsics::_compareTo:
449 return inline_string_compareTo();
450 case vmIntrinsics::_indexOf:
451 return inline_string_indexOf();
452 case vmIntrinsics::_equals:
453 return inline_string_equals();
454
455 case vmIntrinsics::_getObject:
456 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
457 case vmIntrinsics::_getBoolean:
458 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
459 case vmIntrinsics::_getByte:
460 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
461 case vmIntrinsics::_getShort:
462 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
463 case vmIntrinsics::_getChar:
464 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
465 case vmIntrinsics::_getInt:
466 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
467 case vmIntrinsics::_getLong:
468 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
469 case vmIntrinsics::_getFloat:
470 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
471 case vmIntrinsics::_getDouble:
472 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
473
474 case vmIntrinsics::_putObject:
475 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
476 case vmIntrinsics::_putBoolean:
477 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
478 case vmIntrinsics::_putByte:
479 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
480 case vmIntrinsics::_putShort:
481 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
482 case vmIntrinsics::_putChar:
483 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
484 case vmIntrinsics::_putInt:
485 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
486 case vmIntrinsics::_putLong:
487 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
488 case vmIntrinsics::_putFloat:
489 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
490 case vmIntrinsics::_putDouble:
491 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
492
493 case vmIntrinsics::_getByte_raw:
494 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
495 case vmIntrinsics::_getShort_raw:
496 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
497 case vmIntrinsics::_getChar_raw:
498 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
499 case vmIntrinsics::_getInt_raw:
500 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
501 case vmIntrinsics::_getLong_raw:
502 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
503 case vmIntrinsics::_getFloat_raw:
504 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
505 case vmIntrinsics::_getDouble_raw:
506 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
507 case vmIntrinsics::_getAddress_raw:
508 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
509
510 case vmIntrinsics::_putByte_raw:
511 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
512 case vmIntrinsics::_putShort_raw:
513 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
514 case vmIntrinsics::_putChar_raw:
515 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
516 case vmIntrinsics::_putInt_raw:
517 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
518 case vmIntrinsics::_putLong_raw:
519 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
520 case vmIntrinsics::_putFloat_raw:
521 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
522 case vmIntrinsics::_putDouble_raw:
523 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
524 case vmIntrinsics::_putAddress_raw:
525 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
526
527 case vmIntrinsics::_getObjectVolatile:
528 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
529 case vmIntrinsics::_getBooleanVolatile:
530 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
531 case vmIntrinsics::_getByteVolatile:
532 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
533 case vmIntrinsics::_getShortVolatile:
534 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
535 case vmIntrinsics::_getCharVolatile:
536 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
537 case vmIntrinsics::_getIntVolatile:
538 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
539 case vmIntrinsics::_getLongVolatile:
540 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
541 case vmIntrinsics::_getFloatVolatile:
542 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
543 case vmIntrinsics::_getDoubleVolatile:
544 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
545
546 case vmIntrinsics::_putObjectVolatile:
547 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
548 case vmIntrinsics::_putBooleanVolatile:
549 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
550 case vmIntrinsics::_putByteVolatile:
551 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
552 case vmIntrinsics::_putShortVolatile:
553 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
554 case vmIntrinsics::_putCharVolatile:
555 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
556 case vmIntrinsics::_putIntVolatile:
557 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
558 case vmIntrinsics::_putLongVolatile:
559 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
560 case vmIntrinsics::_putFloatVolatile:
561 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
562 case vmIntrinsics::_putDoubleVolatile:
563 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
564
565 case vmIntrinsics::_prefetchRead:
566 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
567 case vmIntrinsics::_prefetchWrite:
568 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
569 case vmIntrinsics::_prefetchReadStatic:
570 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
571 case vmIntrinsics::_prefetchWriteStatic:
572 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
573
574 case vmIntrinsics::_compareAndSwapObject:
575 return inline_unsafe_CAS(T_OBJECT);
576 case vmIntrinsics::_compareAndSwapInt:
577 return inline_unsafe_CAS(T_INT);
578 case vmIntrinsics::_compareAndSwapLong:
579 return inline_unsafe_CAS(T_LONG);
580
581 case vmIntrinsics::_putOrderedObject:
582 return inline_unsafe_ordered_store(T_OBJECT);
583 case vmIntrinsics::_putOrderedInt:
584 return inline_unsafe_ordered_store(T_INT);
585 case vmIntrinsics::_putOrderedLong:
586 return inline_unsafe_ordered_store(T_LONG);
587
588 case vmIntrinsics::_currentThread:
589 return inline_native_currentThread();
590 case vmIntrinsics::_isInterrupted:
591 return inline_native_isInterrupted();
592
593 case vmIntrinsics::_currentTimeMillis:
594 return inline_native_time_funcs(false);
595 case vmIntrinsics::_nanoTime:
596 return inline_native_time_funcs(true);
597 case vmIntrinsics::_allocateInstance:
598 return inline_unsafe_allocate();
599 case vmIntrinsics::_copyMemory:
600 return inline_unsafe_copyMemory();
601 case vmIntrinsics::_newArray:
602 return inline_native_newArray();
603 case vmIntrinsics::_getLength:
604 return inline_native_getLength();
605 case vmIntrinsics::_copyOf:
606 return inline_array_copyOf(false);
607 case vmIntrinsics::_copyOfRange:
608 return inline_array_copyOf(true);
609 case vmIntrinsics::_equalsC:
610 return inline_array_equals();
611 case vmIntrinsics::_clone:
612 return inline_native_clone(intrinsic()->is_virtual());
613
614 case vmIntrinsics::_isAssignableFrom:
615 return inline_native_subtype_check();
616
617 case vmIntrinsics::_isInstance:
618 case vmIntrinsics::_getModifiers:
619 case vmIntrinsics::_isInterface:
620 case vmIntrinsics::_isArray:
621 case vmIntrinsics::_isPrimitive:
622 case vmIntrinsics::_getSuperclass:
623 case vmIntrinsics::_getComponentType:
624 case vmIntrinsics::_getClassAccessFlags:
625 return inline_native_Class_query(intrinsic_id());
626
627 case vmIntrinsics::_floatToRawIntBits:
628 case vmIntrinsics::_floatToIntBits:
629 case vmIntrinsics::_intBitsToFloat:
630 case vmIntrinsics::_doubleToRawLongBits:
631 case vmIntrinsics::_doubleToLongBits:
632 case vmIntrinsics::_longBitsToDouble:
633 return inline_fp_conversions(intrinsic_id());
634
635 case vmIntrinsics::_numberOfLeadingZeros_i:
636 case vmIntrinsics::_numberOfLeadingZeros_l:
637 return inline_numberOfLeadingZeros(intrinsic_id());
638
639 case vmIntrinsics::_numberOfTrailingZeros_i:
640 case vmIntrinsics::_numberOfTrailingZeros_l:
641 return inline_numberOfTrailingZeros(intrinsic_id());
642
643 case vmIntrinsics::_bitCount_i:
644 case vmIntrinsics::_bitCount_l:
645 return inline_bitCount(intrinsic_id());
646
647 case vmIntrinsics::_reverseBytes_i:
648 case vmIntrinsics::_reverseBytes_l:
649 case vmIntrinsics::_reverseBytes_s:
650 case vmIntrinsics::_reverseBytes_c:
651 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
652
653 case vmIntrinsics::_get_AtomicLong:
654 return inline_native_AtomicLong_get();
655 case vmIntrinsics::_attemptUpdate:
656 return inline_native_AtomicLong_attemptUpdate();
657
658 case vmIntrinsics::_getCallerClass:
659 return inline_native_Reflection_getCallerClass();
660
661 default:
662 // If you get here, it may be that someone has added a new intrinsic
663 // to the list in vmSymbols.hpp without implementing it here.
664 #ifndef PRODUCT
665 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
666 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
667 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
668 }
669 #endif
670 return false;
671 }
672 }
673
674 //------------------------------push_result------------------------------
675 // Helper function for finishing intrinsics.
676 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
677 record_for_igvn(region);
678 set_control(_gvn.transform(region));
679 BasicType value_type = value->type()->basic_type();
680 push_node(value_type, _gvn.transform(value));
681 }
682
683 //------------------------------generate_guard---------------------------
684 // Helper function for generating guarded fast-slow graph structures.
685 // The given 'test', if true, guards a slow path. If the test fails
686 // then a fast path can be taken. (We generally hope it fails.)
687 // In all cases, GraphKit::control() is updated to the fast path.
688 // The returned value represents the control for the slow path.
689 // The return value is never 'top'; it is either a valid control
690 // or NULL if it is obvious that the slow path can never be taken.
691 // Also, if region and the slow control are not NULL, the slow edge
692 // is appended to the region.
693 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
694 if (stopped()) {
695 // Already short circuited.
696 return NULL;
697 }
698
699 // Build an if node and its projections.
700 // If test is true we take the slow path, which we assume is uncommon.
701 if (_gvn.type(test) == TypeInt::ZERO) {
702 // The slow branch is never taken. No need to build this guard.
703 return NULL;
704 }
705
706 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
707
708 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
709 if (if_slow == top()) {
710 // The slow branch is never taken. No need to build this guard.
711 return NULL;
712 }
713
714 if (region != NULL)
715 region->add_req(if_slow);
716
717 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
718 set_control(if_fast);
719
720 return if_slow;
721 }
722
723 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
724 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
725 }
726 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
727 return generate_guard(test, region, PROB_FAIR);
728 }
729
730 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
731 Node* *pos_index) {
732 if (stopped())
733 return NULL; // already stopped
734 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
735 return NULL; // index is already adequately typed
736 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
737 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
738 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
739 if (is_neg != NULL && pos_index != NULL) {
740 // Emulate effect of Parse::adjust_map_after_if.
741 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
742 ccast->set_req(0, control());
743 (*pos_index) = _gvn.transform(ccast);
744 }
745 return is_neg;
746 }
747
748 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
749 Node* *pos_index) {
750 if (stopped())
751 return NULL; // already stopped
752 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
753 return NULL; // index is already adequately typed
754 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
755 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
756 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
757 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
758 if (is_notp != NULL && pos_index != NULL) {
759 // Emulate effect of Parse::adjust_map_after_if.
760 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
761 ccast->set_req(0, control());
762 (*pos_index) = _gvn.transform(ccast);
763 }
764 return is_notp;
765 }
766
767 // Make sure that 'position' is a valid limit index, in [0..length].
768 // There are two equivalent plans for checking this:
769 // A. (offset + copyLength) unsigned<= arrayLength
770 // B. offset <= (arrayLength - copyLength)
771 // We require that all of the values above, except for the sum and
772 // difference, are already known to be non-negative.
773 // Plan A is robust in the face of overflow, if offset and copyLength
774 // are both hugely positive.
775 //
776 // Plan B is less direct and intuitive, but it does not overflow at
777 // all, since the difference of two non-negatives is always
778 // representable. Whenever Java methods must perform the equivalent
779 // check they generally use Plan B instead of Plan A.
780 // For the moment we use Plan A.
781 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
782 Node* subseq_length,
783 Node* array_length,
784 RegionNode* region) {
785 if (stopped())
786 return NULL; // already stopped
787 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
788 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
789 return NULL; // common case of whole-array copy
790 Node* last = subseq_length;
791 if (!zero_offset) // last += offset
792 last = _gvn.transform( new (C, 3) AddINode(last, offset));
793 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
794 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
795 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
796 return is_over;
797 }
798
799
800 //--------------------------generate_current_thread--------------------
801 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
802 ciKlass* thread_klass = env()->Thread_klass();
803 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
804 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
805 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
806 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
807 tls_output = thread;
808 return threadObj;
809 }
810
811
812 //------------------------------make_string_method_node------------------------
813 // Helper method for String intrinsic finctions.
814 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
815 const int value_offset = java_lang_String::value_offset_in_bytes();
816 const int count_offset = java_lang_String::count_offset_in_bytes();
817 const int offset_offset = java_lang_String::offset_offset_in_bytes();
818
819 Node* no_ctrl = NULL;
820
821 ciInstanceKlass* klass = env()->String_klass();
822 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
823
824 const TypeAryPtr* value_type =
825 TypeAryPtr::make(TypePtr::NotNull,
826 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
827 ciTypeArrayKlass::make(T_CHAR), true, 0);
828
829 // Get start addr of string and substring
830 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
831 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
832 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
833 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
834 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
835
836 // Pin loads from String::equals() argument since it could be NULL.
837 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
838 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
839 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
840 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
841 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
842 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
843
844 Node* result = NULL;
845 switch (opcode) {
846 case Op_StrIndexOf:
847 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
848 str1_start, cnt1, str2_start, cnt2);
849 break;
850 case Op_StrComp:
851 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
852 str1_start, cnt1, str2_start, cnt2);
853 break;
854 case Op_StrEquals:
855 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
856 str1_start, str2_start, cnt1);
857 break;
858 default:
859 ShouldNotReachHere();
860 return NULL;
861 }
862
863 // All these intrinsics have checks.
864 C->set_has_split_ifs(true); // Has chance for split-if optimization
865
866 return _gvn.transform(result);
867 }
868
869 //------------------------------inline_string_compareTo------------------------
870 bool LibraryCallKit::inline_string_compareTo() {
871
872 if (!Matcher::has_match_rule(Op_StrComp)) return false;
873
874 const int value_offset = java_lang_String::value_offset_in_bytes();
875 const int count_offset = java_lang_String::count_offset_in_bytes();
876 const int offset_offset = java_lang_String::offset_offset_in_bytes();
877
878 _sp += 2;
879 Node *argument = pop(); // pop non-receiver first: it was pushed second
880 Node *receiver = pop();
881
882 // Null check on self without removing any arguments. The argument
883 // null check technically happens in the wrong place, which can lead to
884 // invalid stack traces when string compare is inlined into a method
885 // which handles NullPointerExceptions.
886 _sp += 2;
887 receiver = do_null_check(receiver, T_OBJECT);
888 argument = do_null_check(argument, T_OBJECT);
889 _sp -= 2;
890 if (stopped()) {
891 return true;
892 }
893
894 ciInstanceKlass* klass = env()->String_klass();
895 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
896 Node* no_ctrl = NULL;
897
898 // Get counts for string and argument
899 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
900 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
901
902 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
903 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
904
905 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
906 push(compare);
907 return true;
908 }
909
910 //------------------------------inline_string_equals------------------------
911 bool LibraryCallKit::inline_string_equals() {
912
913 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
914
915 const int value_offset = java_lang_String::value_offset_in_bytes();
916 const int count_offset = java_lang_String::count_offset_in_bytes();
917 const int offset_offset = java_lang_String::offset_offset_in_bytes();
918
919 int nargs = 2;
920 _sp += nargs;
921 Node* argument = pop(); // pop non-receiver first: it was pushed second
922 Node* receiver = pop();
923
924 // Null check on self without removing any arguments. The argument
925 // null check technically happens in the wrong place, which can lead to
926 // invalid stack traces when string compare is inlined into a method
927 // which handles NullPointerExceptions.
928 _sp += nargs;
929 receiver = do_null_check(receiver, T_OBJECT);
930 //should not do null check for argument for String.equals(), because spec
931 //allows to specify NULL as argument.
932 _sp -= nargs;
933
934 if (stopped()) {
935 return true;
936 }
937
938 // paths (plus control) merge
939 RegionNode* region = new (C, 5) RegionNode(5);
940 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
941
942 // does source == target string?
943 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
944 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
945
946 Node* if_eq = generate_slow_guard(bol, NULL);
947 if (if_eq != NULL) {
948 // receiver == argument
949 phi->init_req(2, intcon(1));
950 region->init_req(2, if_eq);
951 }
952
953 // get String klass for instanceOf
954 ciInstanceKlass* klass = env()->String_klass();
955
956 if (!stopped()) {
957 _sp += nargs; // gen_instanceof might do an uncommon trap
958 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
959 _sp -= nargs;
960 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
961 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
962
963 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
964 //instanceOf == true, fallthrough
965
966 if (inst_false != NULL) {
967 phi->init_req(3, intcon(0));
968 region->init_req(3, inst_false);
969 }
970 }
971
972 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
973
974 Node* no_ctrl = NULL;
975 Node* receiver_cnt;
976 Node* argument_cnt;
977
978 if (!stopped()) {
979 // Properly cast the argument to String
980 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
981
982 // Get counts for string and argument
983 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
984 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
985
986 // Pin load from argument string since it could be NULL.
987 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
988 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
989
990 // Check for receiver count != argument count
991 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
992 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
993 Node* if_ne = generate_slow_guard(bol, NULL);
994 if (if_ne != NULL) {
995 phi->init_req(4, intcon(0));
996 region->init_req(4, if_ne);
997 }
998 }
999
1000 // Check for count == 0 is done by mach node StrEquals.
1001
1002 if (!stopped()) {
1003 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1004 phi->init_req(1, equals);
1005 region->init_req(1, control());
1006 }
1007
1008 // post merge
1009 set_control(_gvn.transform(region));
1010 record_for_igvn(region);
1011
1012 push(_gvn.transform(phi));
1013
1014 return true;
1015 }
1016
1017 //------------------------------inline_array_equals----------------------------
1018 bool LibraryCallKit::inline_array_equals() {
1019
1020 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1021
1022 _sp += 2;
1023 Node *argument2 = pop();
1024 Node *argument1 = pop();
1025
1026 Node* equals =
1027 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1028 argument1, argument2) );
1029 push(equals);
1030 return true;
1031 }
1032
1033 // Java version of String.indexOf(constant string)
1034 // class StringDecl {
1035 // StringDecl(char[] ca) {
1036 // offset = 0;
1037 // count = ca.length;
1038 // value = ca;
1039 // }
1040 // int offset;
1041 // int count;
1042 // char[] value;
1043 // }
1044 //
1045 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1046 // int targetOffset, int cache_i, int md2) {
1047 // int cache = cache_i;
1048 // int sourceOffset = string_object.offset;
1049 // int sourceCount = string_object.count;
1050 // int targetCount = target_object.length;
1051 //
1052 // int targetCountLess1 = targetCount - 1;
1053 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1054 //
1055 // char[] source = string_object.value;
1056 // char[] target = target_object;
1057 // int lastChar = target[targetCountLess1];
1058 //
1059 // outer_loop:
1060 // for (int i = sourceOffset; i < sourceEnd; ) {
1061 // int src = source[i + targetCountLess1];
1062 // if (src == lastChar) {
1063 // // With random strings and a 4-character alphabet,
1064 // // reverse matching at this point sets up 0.8% fewer
1065 // // frames, but (paradoxically) makes 0.3% more probes.
1066 // // Since those probes are nearer the lastChar probe,
1067 // // there is may be a net D$ win with reverse matching.
1068 // // But, reversing loop inhibits unroll of inner loop
1069 // // for unknown reason. So, does running outer loop from
1070 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1071 // for (int j = 0; j < targetCountLess1; j++) {
1072 // if (target[targetOffset + j] != source[i+j]) {
1073 // if ((cache & (1 << source[i+j])) == 0) {
1074 // if (md2 < j+1) {
1075 // i += j+1;
1076 // continue outer_loop;
1077 // }
1078 // }
1079 // i += md2;
1080 // continue outer_loop;
1081 // }
1082 // }
1083 // return i - sourceOffset;
1084 // }
1085 // if ((cache & (1 << src)) == 0) {
1086 // i += targetCountLess1;
1087 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1088 // i++;
1089 // }
1090 // return -1;
1091 // }
1092
1093 //------------------------------string_indexOf------------------------
1094 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1095 jint cache_i, jint md2_i) {
1096
1097 Node* no_ctrl = NULL;
1098 float likely = PROB_LIKELY(0.9);
1099 float unlikely = PROB_UNLIKELY(0.9);
1100
1101 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1102
1103 const int value_offset = java_lang_String::value_offset_in_bytes();
1104 const int count_offset = java_lang_String::count_offset_in_bytes();
1105 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1106
1107 ciInstanceKlass* klass = env()->String_klass();
1108 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1109 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1110
1111 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1112 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1113 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1114 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1115 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1116 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1117
1118 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1119 jint target_length = target_array->length();
1120 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1121 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1122
1123 IdealKit kit(this, false, true);
1124 #define __ kit.
1125 Node* zero = __ ConI(0);
1126 Node* one = __ ConI(1);
1127 Node* cache = __ ConI(cache_i);
1128 Node* md2 = __ ConI(md2_i);
1129 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1130 Node* targetCount = __ ConI(target_length);
1131 Node* targetCountLess1 = __ ConI(target_length - 1);
1132 Node* targetOffset = __ ConI(targetOffset_i);
1133 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1134
1135 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1136 Node* outer_loop = __ make_label(2 /* goto */);
1137 Node* return_ = __ make_label(1);
1138
1139 __ set(rtn,__ ConI(-1));
1140 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1141 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1142 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1143 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1144 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1145 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1146 Node* tpj = __ AddI(targetOffset, __ value(j));
1147 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1148 Node* ipj = __ AddI(__ value(i), __ value(j));
1149 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1150 __ if_then(targ, BoolTest::ne, src2); {
1151 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1152 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1153 __ increment(i, __ AddI(__ value(j), one));
1154 __ goto_(outer_loop);
1155 } __ end_if(); __ dead(j);
1156 }__ end_if(); __ dead(j);
1157 __ increment(i, md2);
1158 __ goto_(outer_loop);
1159 }__ end_if();
1160 __ increment(j, one);
1161 }__ end_loop(); __ dead(j);
1162 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1163 __ goto_(return_);
1164 }__ end_if();
1165 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1166 __ increment(i, targetCountLess1);
1167 }__ end_if();
1168 __ increment(i, one);
1169 __ bind(outer_loop);
1170 }__ end_loop(); __ dead(i);
1171 __ bind(return_);
1172
1173 // Final sync IdealKit and GraphKit.
1174 sync_kit(kit);
1175 Node* result = __ value(rtn);
1176 #undef __
1177 C->set_has_loops(true);
1178 return result;
1179 }
1180
1181 //------------------------------inline_string_indexOf------------------------
1182 bool LibraryCallKit::inline_string_indexOf() {
1183
1184 const int value_offset = java_lang_String::value_offset_in_bytes();
1185 const int count_offset = java_lang_String::count_offset_in_bytes();
1186 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1187
1188 _sp += 2;
1189 Node *argument = pop(); // pop non-receiver first: it was pushed second
1190 Node *receiver = pop();
1191
1192 Node* result;
1193 // Disable the use of pcmpestri until it can be guaranteed that
1194 // the load doesn't cross into the uncommited space.
1195 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1196 UseSSE42Intrinsics) {
1197 // Generate SSE4.2 version of indexOf
1198 // We currently only have match rules that use SSE4.2
1199
1200 // Null check on self without removing any arguments. The argument
1201 // null check technically happens in the wrong place, which can lead to
1202 // invalid stack traces when string compare is inlined into a method
1203 // which handles NullPointerExceptions.
1204 _sp += 2;
1205 receiver = do_null_check(receiver, T_OBJECT);
1206 argument = do_null_check(argument, T_OBJECT);
1207 _sp -= 2;
1208
1209 if (stopped()) {
1210 return true;
1211 }
1212
1213 ciInstanceKlass* str_klass = env()->String_klass();
1214 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1215
1216 // Make the merge point
1217 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1218 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1219 Node* no_ctrl = NULL;
1220
1221 // Get counts for string and substr
1222 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1223 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1224
1225 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1226 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1227
1228 // Check for substr count > string count
1229 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1230 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1231 Node* if_gt = generate_slow_guard(bol, NULL);
1232 if (if_gt != NULL) {
1233 result_phi->init_req(2, intcon(-1));
1234 result_rgn->init_req(2, if_gt);
1235 }
1236
1237 if (!stopped()) {
1238 // Check for substr count == 0
1239 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1240 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1241 Node* if_zero = generate_slow_guard(bol, NULL);
1242 if (if_zero != NULL) {
1243 result_phi->init_req(3, intcon(0));
1244 result_rgn->init_req(3, if_zero);
1245 }
1246 }
1247
1248 if (!stopped()) {
1249 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1250 result_phi->init_req(1, result);
1251 result_rgn->init_req(1, control());
1252 }
1253 set_control(_gvn.transform(result_rgn));
1254 record_for_igvn(result_rgn);
1255 result = _gvn.transform(result_phi);
1256
1257 } else { // Use LibraryCallKit::string_indexOf
1258 // don't intrinsify if argument isn't a constant string.
1259 if (!argument->is_Con()) {
1260 return false;
1261 }
1262 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1263 if (str_type == NULL) {
1264 return false;
1265 }
1266 ciInstanceKlass* klass = env()->String_klass();
1267 ciObject* str_const = str_type->const_oop();
1268 if (str_const == NULL || str_const->klass() != klass) {
1269 return false;
1270 }
1271 ciInstance* str = str_const->as_instance();
1272 assert(str != NULL, "must be instance");
1273
1274 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1275 int o = str->field_value_by_offset(offset_offset).as_int();
1276 int c = str->field_value_by_offset(count_offset).as_int();
1277 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1278
1279 // constant strings have no offset and count == length which
1280 // simplifies the resulting code somewhat so lets optimize for that.
1281 if (o != 0 || c != pat->length()) {
1282 return false;
1283 }
1284
1285 // Null check on self without removing any arguments. The argument
1286 // null check technically happens in the wrong place, which can lead to
1287 // invalid stack traces when string compare is inlined into a method
1288 // which handles NullPointerExceptions.
1289 _sp += 2;
1290 receiver = do_null_check(receiver, T_OBJECT);
1291 // No null check on the argument is needed since it's a constant String oop.
1292 _sp -= 2;
1293 if (stopped()) {
1294 return true;
1295 }
1296
1297 // The null string as a pattern always returns 0 (match at beginning of string)
1298 if (c == 0) {
1299 push(intcon(0));
1300 return true;
1301 }
1302
1303 // Generate default indexOf
1304 jchar lastChar = pat->char_at(o + (c - 1));
1305 int cache = 0;
1306 int i;
1307 for (i = 0; i < c - 1; i++) {
1308 assert(i < pat->length(), "out of range");
1309 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1310 }
1311
1312 int md2 = c;
1313 for (i = 0; i < c - 1; i++) {
1314 assert(i < pat->length(), "out of range");
1315 if (pat->char_at(o + i) == lastChar) {
1316 md2 = (c - 1) - i;
1317 }
1318 }
1319
1320 result = string_indexOf(receiver, pat, o, cache, md2);
1321 }
1322
1323 push(result);
1324 return true;
1325 }
1326
1327 //--------------------------pop_math_arg--------------------------------
1328 // Pop a double argument to a math function from the stack
1329 // rounding it if necessary.
1330 Node * LibraryCallKit::pop_math_arg() {
1331 Node *arg = pop_pair();
1332 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1333 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1334 return arg;
1335 }
1336
1337 //------------------------------inline_trig----------------------------------
1338 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1339 // argument reduction which will turn into a fast/slow diamond.
1340 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1341 _sp += arg_size(); // restore stack pointer
1342 Node* arg = pop_math_arg();
1343 Node* trig = NULL;
1344
1345 switch (id) {
1346 case vmIntrinsics::_dsin:
1347 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1348 break;
1349 case vmIntrinsics::_dcos:
1350 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1351 break;
1352 case vmIntrinsics::_dtan:
1353 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1354 break;
1355 default:
1356 assert(false, "bad intrinsic was passed in");
1357 return false;
1358 }
1359
1360 // Rounding required? Check for argument reduction!
1361 if( Matcher::strict_fp_requires_explicit_rounding ) {
1362
1363 static const double pi_4 = 0.7853981633974483;
1364 static const double neg_pi_4 = -0.7853981633974483;
1365 // pi/2 in 80-bit extended precision
1366 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1367 // -pi/2 in 80-bit extended precision
1368 // 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};
1369 // Cutoff value for using this argument reduction technique
1370 //static const double pi_2_minus_epsilon = 1.564660403643354;
1371 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1372
1373 // Pseudocode for sin:
1374 // if (x <= Math.PI / 4.0) {
1375 // if (x >= -Math.PI / 4.0) return fsin(x);
1376 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1377 // } else {
1378 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1379 // }
1380 // return StrictMath.sin(x);
1381
1382 // Pseudocode for cos:
1383 // if (x <= Math.PI / 4.0) {
1384 // if (x >= -Math.PI / 4.0) return fcos(x);
1385 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1386 // } else {
1387 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1388 // }
1389 // return StrictMath.cos(x);
1390
1391 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1392 // requires a special machine instruction to load it. Instead we'll try
1393 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1394 // probably do the math inside the SIN encoding.
1395
1396 // Make the merge point
1397 RegionNode *r = new (C, 3) RegionNode(3);
1398 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1399
1400 // Flatten arg so we need only 1 test
1401 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1402 // Node for PI/4 constant
1403 Node *pi4 = makecon(TypeD::make(pi_4));
1404 // Check PI/4 : abs(arg)
1405 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1406 // Check: If PI/4 < abs(arg) then go slow
1407 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1408 // Branch either way
1409 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1410 set_control(opt_iff(r,iff));
1411
1412 // Set fast path result
1413 phi->init_req(2,trig);
1414
1415 // Slow path - non-blocking leaf call
1416 Node* call = NULL;
1417 switch (id) {
1418 case vmIntrinsics::_dsin:
1419 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1420 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1421 "Sin", NULL, arg, top());
1422 break;
1423 case vmIntrinsics::_dcos:
1424 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1425 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1426 "Cos", NULL, arg, top());
1427 break;
1428 case vmIntrinsics::_dtan:
1429 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1430 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1431 "Tan", NULL, arg, top());
1432 break;
1433 }
1434 assert(control()->in(0) == call, "");
1435 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1436 r->init_req(1,control());
1437 phi->init_req(1,slow_result);
1438
1439 // Post-merge
1440 set_control(_gvn.transform(r));
1441 record_for_igvn(r);
1442 trig = _gvn.transform(phi);
1443
1444 C->set_has_split_ifs(true); // Has chance for split-if optimization
1445 }
1446 // Push result back on JVM stack
1447 push_pair(trig);
1448 return true;
1449 }
1450
1451 //------------------------------inline_sqrt-------------------------------------
1452 // Inline square root instruction, if possible.
1453 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1454 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1455 _sp += arg_size(); // restore stack pointer
1456 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1457 return true;
1458 }
1459
1460 //------------------------------inline_abs-------------------------------------
1461 // Inline absolute value instruction, if possible.
1462 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1463 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1464 _sp += arg_size(); // restore stack pointer
1465 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1466 return true;
1467 }
1468
1469 //------------------------------inline_exp-------------------------------------
1470 // Inline exp instructions, if possible. The Intel hardware only misses
1471 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1472 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1473 assert(id == vmIntrinsics::_dexp, "Not exp");
1474
1475 // If this inlining ever returned NaN in the past, we do not intrinsify it
1476 // every again. NaN results requires StrictMath.exp handling.
1477 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1478
1479 // Do not intrinsify on older platforms which lack cmove.
1480 if (ConditionalMoveLimit == 0) return false;
1481
1482 _sp += arg_size(); // restore stack pointer
1483 Node *x = pop_math_arg();
1484 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1485
1486 //-------------------
1487 //result=(result.isNaN())? StrictMath::exp():result;
1488 // Check: If isNaN() by checking result!=result? then go to Strict Math
1489 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1490 // Build the boolean node
1491 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1492
1493 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1494 // End the current control-flow path
1495 push_pair(x);
1496 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1497 // to handle. Recompile without intrinsifying Math.exp
1498 uncommon_trap(Deoptimization::Reason_intrinsic,
1499 Deoptimization::Action_make_not_entrant);
1500 }
1501
1502 C->set_has_split_ifs(true); // Has chance for split-if optimization
1503
1504 push_pair(result);
1505
1506 return true;
1507 }
1508
1509 //------------------------------inline_pow-------------------------------------
1510 // Inline power instructions, if possible.
1511 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1512 assert(id == vmIntrinsics::_dpow, "Not pow");
1513
1514 // If this inlining ever returned NaN in the past, we do not intrinsify it
1515 // every again. NaN results requires StrictMath.pow handling.
1516 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1517
1518 // Do not intrinsify on older platforms which lack cmove.
1519 if (ConditionalMoveLimit == 0) return false;
1520
1521 // Pseudocode for pow
1522 // if (x <= 0.0) {
1523 // if ((double)((int)y)==y) { // if y is int
1524 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1525 // } else {
1526 // result = NaN;
1527 // }
1528 // } else {
1529 // result = DPow(x,y);
1530 // }
1531 // if (result != result)? {
1532 // uncommon_trap();
1533 // }
1534 // return result;
1535
1536 _sp += arg_size(); // restore stack pointer
1537 Node* y = pop_math_arg();
1538 Node* x = pop_math_arg();
1539
1540 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1541
1542 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1543 // inside of something) then skip the fancy tests and just check for
1544 // NaN result.
1545 Node *result = NULL;
1546 if( jvms()->depth() >= 1 ) {
1547 result = fast_result;
1548 } else {
1549
1550 // Set the merge point for If node with condition of (x <= 0.0)
1551 // There are four possible paths to region node and phi node
1552 RegionNode *r = new (C, 4) RegionNode(4);
1553 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1554
1555 // Build the first if node: if (x <= 0.0)
1556 // Node for 0 constant
1557 Node *zeronode = makecon(TypeD::ZERO);
1558 // Check x:0
1559 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1560 // Check: If (x<=0) then go complex path
1561 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1562 // Branch either way
1563 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1564 Node *opt_test = _gvn.transform(if1);
1565 //assert( opt_test->is_If(), "Expect an IfNode");
1566 IfNode *opt_if1 = (IfNode*)opt_test;
1567 // Fast path taken; set region slot 3
1568 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1569 r->init_req(3,fast_taken); // Capture fast-control
1570
1571 // Fast path not-taken, i.e. slow path
1572 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1573
1574 // Set fast path result
1575 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1576 phi->init_req(3, fast_result);
1577
1578 // Complex path
1579 // Build the second if node (if y is int)
1580 // Node for (int)y
1581 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1582 // Node for (double)((int) y)
1583 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1584 // Check (double)((int) y) : y
1585 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1586 // Check if (y isn't int) then go to slow path
1587
1588 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1589 // Branch either way
1590 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1591 Node *slow_path = opt_iff(r,if2); // Set region path 2
1592
1593 // Calculate DPow(abs(x), y)*(1 & (int)y)
1594 // Node for constant 1
1595 Node *conone = intcon(1);
1596 // 1& (int)y
1597 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1598 // zero node
1599 Node *conzero = intcon(0);
1600 // Check (1&(int)y)==0?
1601 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1602 // Check if (1&(int)y)!=0?, if so the result is negative
1603 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1604 // abs(x)
1605 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1606 // abs(x)^y
1607 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1608 // -abs(x)^y
1609 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1610 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1611 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1612 // Set complex path fast result
1613 phi->init_req(2, signresult);
1614
1615 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1616 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1617 r->init_req(1,slow_path);
1618 phi->init_req(1,slow_result);
1619
1620 // Post merge
1621 set_control(_gvn.transform(r));
1622 record_for_igvn(r);
1623 result=_gvn.transform(phi);
1624 }
1625
1626 //-------------------
1627 //result=(result.isNaN())? uncommon_trap():result;
1628 // Check: If isNaN() by checking result!=result? then go to Strict Math
1629 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1630 // Build the boolean node
1631 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1632
1633 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1634 // End the current control-flow path
1635 push_pair(x);
1636 push_pair(y);
1637 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1638 // to handle. Recompile without intrinsifying Math.pow.
1639 uncommon_trap(Deoptimization::Reason_intrinsic,
1640 Deoptimization::Action_make_not_entrant);
1641 }
1642
1643 C->set_has_split_ifs(true); // Has chance for split-if optimization
1644
1645 push_pair(result);
1646
1647 return true;
1648 }
1649
1650 //------------------------------inline_trans-------------------------------------
1651 // Inline transcendental instructions, if possible. The Intel hardware gets
1652 // these right, no funny corner cases missed.
1653 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1654 _sp += arg_size(); // restore stack pointer
1655 Node* arg = pop_math_arg();
1656 Node* trans = NULL;
1657
1658 switch (id) {
1659 case vmIntrinsics::_dlog:
1660 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1661 break;
1662 case vmIntrinsics::_dlog10:
1663 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1664 break;
1665 default:
1666 assert(false, "bad intrinsic was passed in");
1667 return false;
1668 }
1669
1670 // Push result back on JVM stack
1671 push_pair(trans);
1672 return true;
1673 }
1674
1675 //------------------------------runtime_math-----------------------------
1676 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1677 Node* a = NULL;
1678 Node* b = NULL;
1679
1680 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1681 "must be (DD)D or (D)D type");
1682
1683 // Inputs
1684 _sp += arg_size(); // restore stack pointer
1685 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1686 b = pop_math_arg();
1687 }
1688 a = pop_math_arg();
1689
1690 const TypePtr* no_memory_effects = NULL;
1691 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1692 no_memory_effects,
1693 a, top(), b, b ? top() : NULL);
1694 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1695 #ifdef ASSERT
1696 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1697 assert(value_top == top(), "second value must be top");
1698 #endif
1699
1700 push_pair(value);
1701 return true;
1702 }
1703
1704 //------------------------------inline_math_native-----------------------------
1705 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1706 switch (id) {
1707 // These intrinsics are not properly supported on all hardware
1708 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1709 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1710 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1711 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1712 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1713 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1714
1715 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1716 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1717 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1718 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1719
1720 // These intrinsics are supported on all hardware
1721 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1722 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1723
1724 // These intrinsics don't work on X86. The ad implementation doesn't
1725 // handle NaN's properly. Instead of returning infinity, the ad
1726 // implementation returns a NaN on overflow. See bug: 6304089
1727 // Once the ad implementations are fixed, change the code below
1728 // to match the intrinsics above
1729
1730 case vmIntrinsics::_dexp: return
1731 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1732 case vmIntrinsics::_dpow: return
1733 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1734
1735 // These intrinsics are not yet correctly implemented
1736 case vmIntrinsics::_datan2:
1737 return false;
1738
1739 default:
1740 ShouldNotReachHere();
1741 return false;
1742 }
1743 }
1744
1745 static bool is_simple_name(Node* n) {
1746 return (n->req() == 1 // constant
1747 || (n->is_Type() && n->as_Type()->type()->singleton())
1748 || n->is_Proj() // parameter or return value
1749 || n->is_Phi() // local of some sort
1750 );
1751 }
1752
1753 //----------------------------inline_min_max-----------------------------------
1754 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1755 push(generate_min_max(id, argument(0), argument(1)));
1756
1757 return true;
1758 }
1759
1760 Node*
1761 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1762 // These are the candidate return value:
1763 Node* xvalue = x0;
1764 Node* yvalue = y0;
1765
1766 if (xvalue == yvalue) {
1767 return xvalue;
1768 }
1769
1770 bool want_max = (id == vmIntrinsics::_max);
1771
1772 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1773 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1774 if (txvalue == NULL || tyvalue == NULL) return top();
1775 // This is not really necessary, but it is consistent with a
1776 // hypothetical MaxINode::Value method:
1777 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1778
1779 // %%% This folding logic should (ideally) be in a different place.
1780 // Some should be inside IfNode, and there to be a more reliable
1781 // transformation of ?: style patterns into cmoves. We also want
1782 // more powerful optimizations around cmove and min/max.
1783
1784 // Try to find a dominating comparison of these guys.
1785 // It can simplify the index computation for Arrays.copyOf
1786 // and similar uses of System.arraycopy.
1787 // First, compute the normalized version of CmpI(x, y).
1788 int cmp_op = Op_CmpI;
1789 Node* xkey = xvalue;
1790 Node* ykey = yvalue;
1791 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1792 if (ideal_cmpxy->is_Cmp()) {
1793 // E.g., if we have CmpI(length - offset, count),
1794 // it might idealize to CmpI(length, count + offset)
1795 cmp_op = ideal_cmpxy->Opcode();
1796 xkey = ideal_cmpxy->in(1);
1797 ykey = ideal_cmpxy->in(2);
1798 }
1799
1800 // Start by locating any relevant comparisons.
1801 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1802 Node* cmpxy = NULL;
1803 Node* cmpyx = NULL;
1804 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1805 Node* cmp = start_from->fast_out(k);
1806 if (cmp->outcnt() > 0 && // must have prior uses
1807 cmp->in(0) == NULL && // must be context-independent
1808 cmp->Opcode() == cmp_op) { // right kind of compare
1809 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1810 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1811 }
1812 }
1813
1814 const int NCMPS = 2;
1815 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1816 int cmpn;
1817 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1818 if (cmps[cmpn] != NULL) break; // find a result
1819 }
1820 if (cmpn < NCMPS) {
1821 // Look for a dominating test that tells us the min and max.
1822 int depth = 0; // Limit search depth for speed
1823 Node* dom = control();
1824 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1825 if (++depth >= 100) break;
1826 Node* ifproj = dom;
1827 if (!ifproj->is_Proj()) continue;
1828 Node* iff = ifproj->in(0);
1829 if (!iff->is_If()) continue;
1830 Node* bol = iff->in(1);
1831 if (!bol->is_Bool()) continue;
1832 Node* cmp = bol->in(1);
1833 if (cmp == NULL) continue;
1834 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1835 if (cmps[cmpn] == cmp) break;
1836 if (cmpn == NCMPS) continue;
1837 BoolTest::mask btest = bol->as_Bool()->_test._test;
1838 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1839 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1840 // At this point, we know that 'x btest y' is true.
1841 switch (btest) {
1842 case BoolTest::eq:
1843 // They are proven equal, so we can collapse the min/max.
1844 // Either value is the answer. Choose the simpler.
1845 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1846 return yvalue;
1847 return xvalue;
1848 case BoolTest::lt: // x < y
1849 case BoolTest::le: // x <= y
1850 return (want_max ? yvalue : xvalue);
1851 case BoolTest::gt: // x > y
1852 case BoolTest::ge: // x >= y
1853 return (want_max ? xvalue : yvalue);
1854 }
1855 }
1856 }
1857
1858 // We failed to find a dominating test.
1859 // Let's pick a test that might GVN with prior tests.
1860 Node* best_bol = NULL;
1861 BoolTest::mask best_btest = BoolTest::illegal;
1862 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1863 Node* cmp = cmps[cmpn];
1864 if (cmp == NULL) continue;
1865 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1866 Node* bol = cmp->fast_out(j);
1867 if (!bol->is_Bool()) continue;
1868 BoolTest::mask btest = bol->as_Bool()->_test._test;
1869 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1870 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1871 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1872 best_bol = bol->as_Bool();
1873 best_btest = btest;
1874 }
1875 }
1876 }
1877
1878 Node* answer_if_true = NULL;
1879 Node* answer_if_false = NULL;
1880 switch (best_btest) {
1881 default:
1882 if (cmpxy == NULL)
1883 cmpxy = ideal_cmpxy;
1884 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1885 // and fall through:
1886 case BoolTest::lt: // x < y
1887 case BoolTest::le: // x <= y
1888 answer_if_true = (want_max ? yvalue : xvalue);
1889 answer_if_false = (want_max ? xvalue : yvalue);
1890 break;
1891 case BoolTest::gt: // x > y
1892 case BoolTest::ge: // x >= y
1893 answer_if_true = (want_max ? xvalue : yvalue);
1894 answer_if_false = (want_max ? yvalue : xvalue);
1895 break;
1896 }
1897
1898 jint hi, lo;
1899 if (want_max) {
1900 // We can sharpen the minimum.
1901 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1902 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1903 } else {
1904 // We can sharpen the maximum.
1905 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1906 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1907 }
1908
1909 // Use a flow-free graph structure, to avoid creating excess control edges
1910 // which could hinder other optimizations.
1911 // Since Math.min/max is often used with arraycopy, we want
1912 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1913 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1914 answer_if_false, answer_if_true,
1915 TypeInt::make(lo, hi, widen));
1916
1917 return _gvn.transform(cmov);
1918
1919 /*
1920 // This is not as desirable as it may seem, since Min and Max
1921 // nodes do not have a full set of optimizations.
1922 // And they would interfere, anyway, with 'if' optimizations
1923 // and with CMoveI canonical forms.
1924 switch (id) {
1925 case vmIntrinsics::_min:
1926 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1927 case vmIntrinsics::_max:
1928 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1929 default:
1930 ShouldNotReachHere();
1931 }
1932 */
1933 }
1934
1935 inline int
1936 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1937 const TypePtr* base_type = TypePtr::NULL_PTR;
1938 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1939 if (base_type == NULL) {
1940 // Unknown type.
1941 return Type::AnyPtr;
1942 } else if (base_type == TypePtr::NULL_PTR) {
1943 // Since this is a NULL+long form, we have to switch to a rawptr.
1944 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1945 offset = MakeConX(0);
1946 return Type::RawPtr;
1947 } else if (base_type->base() == Type::RawPtr) {
1948 return Type::RawPtr;
1949 } else if (base_type->isa_oopptr()) {
1950 // Base is never null => always a heap address.
1951 if (base_type->ptr() == TypePtr::NotNull) {
1952 return Type::OopPtr;
1953 }
1954 // Offset is small => always a heap address.
1955 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1956 if (offset_type != NULL &&
1957 base_type->offset() == 0 && // (should always be?)
1958 offset_type->_lo >= 0 &&
1959 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1960 return Type::OopPtr;
1961 }
1962 // Otherwise, it might either be oop+off or NULL+addr.
1963 return Type::AnyPtr;
1964 } else {
1965 // No information:
1966 return Type::AnyPtr;
1967 }
1968 }
1969
1970 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1971 int kind = classify_unsafe_addr(base, offset);
1972 if (kind == Type::RawPtr) {
1973 return basic_plus_adr(top(), base, offset);
1974 } else {
1975 return basic_plus_adr(base, offset);
1976 }
1977 }
1978
1979 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1980 // inline int Integer.numberOfLeadingZeros(int)
1981 // inline int Long.numberOfLeadingZeros(long)
1982 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1983 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1984 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1985 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1986 _sp += arg_size(); // restore stack pointer
1987 switch (id) {
1988 case vmIntrinsics::_numberOfLeadingZeros_i:
1989 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1990 break;
1991 case vmIntrinsics::_numberOfLeadingZeros_l:
1992 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1993 break;
1994 default:
1995 ShouldNotReachHere();
1996 }
1997 return true;
1998 }
1999
2000 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2001 // inline int Integer.numberOfTrailingZeros(int)
2002 // inline int Long.numberOfTrailingZeros(long)
2003 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2004 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2005 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2006 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2007 _sp += arg_size(); // restore stack pointer
2008 switch (id) {
2009 case vmIntrinsics::_numberOfTrailingZeros_i:
2010 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2011 break;
2012 case vmIntrinsics::_numberOfTrailingZeros_l:
2013 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2014 break;
2015 default:
2016 ShouldNotReachHere();
2017 }
2018 return true;
2019 }
2020
2021 //----------------------------inline_bitCount_int/long-----------------------
2022 // inline int Integer.bitCount(int)
2023 // inline int Long.bitCount(long)
2024 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2025 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2026 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2027 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2028 _sp += arg_size(); // restore stack pointer
2029 switch (id) {
2030 case vmIntrinsics::_bitCount_i:
2031 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2032 break;
2033 case vmIntrinsics::_bitCount_l:
2034 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2035 break;
2036 default:
2037 ShouldNotReachHere();
2038 }
2039 return true;
2040 }
2041
2042 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2043 // inline Integer.reverseBytes(int)
2044 // inline Long.reverseBytes(long)
2045 // inline Character.reverseBytes(char)
2046 // inline Short.reverseBytes(short)
2047 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2048 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2049 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2050 "not reverse Bytes");
2051 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2052 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2053 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2054 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2055 _sp += arg_size(); // restore stack pointer
2056 switch (id) {
2057 case vmIntrinsics::_reverseBytes_i:
2058 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2059 break;
2060 case vmIntrinsics::_reverseBytes_l:
2061 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2062 break;
2063 case vmIntrinsics::_reverseBytes_c:
2064 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2065 break;
2066 case vmIntrinsics::_reverseBytes_s:
2067 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2068 break;
2069 default:
2070 ;
2071 }
2072 return true;
2073 }
2074
2075 //----------------------------inline_unsafe_access----------------------------
2076
2077 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2078
2079 // Interpret Unsafe.fieldOffset cookies correctly:
2080 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2081
2082 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2083 if (callee()->is_static()) return false; // caller must have the capability!
2084
2085 #ifndef PRODUCT
2086 {
2087 ResourceMark rm;
2088 // Check the signatures.
2089 ciSignature* sig = signature();
2090 #ifdef ASSERT
2091 if (!is_store) {
2092 // Object getObject(Object base, int/long offset), etc.
2093 BasicType rtype = sig->return_type()->basic_type();
2094 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2095 rtype = T_ADDRESS; // it is really a C void*
2096 assert(rtype == type, "getter must return the expected value");
2097 if (!is_native_ptr) {
2098 assert(sig->count() == 2, "oop getter has 2 arguments");
2099 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2100 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2101 } else {
2102 assert(sig->count() == 1, "native getter has 1 argument");
2103 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2104 }
2105 } else {
2106 // void putObject(Object base, int/long offset, Object x), etc.
2107 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2108 if (!is_native_ptr) {
2109 assert(sig->count() == 3, "oop putter has 3 arguments");
2110 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2111 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2112 } else {
2113 assert(sig->count() == 2, "native putter has 2 arguments");
2114 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2115 }
2116 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2117 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2118 vtype = T_ADDRESS; // it is really a C void*
2119 assert(vtype == type, "putter must accept the expected value");
2120 }
2121 #endif // ASSERT
2122 }
2123 #endif //PRODUCT
2124
2125 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2126
2127 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2128
2129 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2130 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2131
2132 debug_only(int saved_sp = _sp);
2133 _sp += nargs;
2134
2135 Node* val;
2136 debug_only(val = (Node*)(uintptr_t)-1);
2137
2138
2139 if (is_store) {
2140 // Get the value being stored. (Pop it first; it was pushed last.)
2141 switch (type) {
2142 case T_DOUBLE:
2143 case T_LONG:
2144 case T_ADDRESS:
2145 val = pop_pair();
2146 break;
2147 default:
2148 val = pop();
2149 }
2150 }
2151
2152 // Build address expression. See the code in inline_unsafe_prefetch.
2153 Node *adr;
2154 Node *heap_base_oop = top();
2155 if (!is_native_ptr) {
2156 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2157 Node* offset = pop_pair();
2158 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2159 Node* base = pop();
2160 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2161 // to be plain byte offsets, which are also the same as those accepted
2162 // by oopDesc::field_base.
2163 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2164 "fieldOffset must be byte-scaled");
2165 // 32-bit machines ignore the high half!
2166 offset = ConvL2X(offset);
2167 adr = make_unsafe_address(base, offset);
2168 heap_base_oop = base;
2169 } else {
2170 Node* ptr = pop_pair();
2171 // Adjust Java long to machine word:
2172 ptr = ConvL2X(ptr);
2173 adr = make_unsafe_address(NULL, ptr);
2174 }
2175
2176 // Pop receiver last: it was pushed first.
2177 Node *receiver = pop();
2178
2179 assert(saved_sp == _sp, "must have correct argument count");
2180
2181 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2182
2183 // First guess at the value type.
2184 const Type *value_type = Type::get_const_basic_type(type);
2185
2186 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2187 // there was not enough information to nail it down.
2188 Compile::AliasType* alias_type = C->alias_type(adr_type);
2189 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2190
2191 // We will need memory barriers unless we can determine a unique
2192 // alias category for this reference. (Note: If for some reason
2193 // the barriers get omitted and the unsafe reference begins to "pollute"
2194 // the alias analysis of the rest of the graph, either Compile::can_alias
2195 // or Compile::must_alias will throw a diagnostic assert.)
2196 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2197
2198 if (!is_store && type == T_OBJECT) {
2199 // Attempt to infer a sharper value type from the offset and base type.
2200 ciKlass* sharpened_klass = NULL;
2201
2202 // See if it is an instance field, with an object type.
2203 if (alias_type->field() != NULL) {
2204 assert(!is_native_ptr, "native pointer op cannot use a java address");
2205 if (alias_type->field()->type()->is_klass()) {
2206 sharpened_klass = alias_type->field()->type()->as_klass();
2207 }
2208 }
2209
2210 // See if it is a narrow oop array.
2211 if (adr_type->isa_aryptr()) {
2212 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2213 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2214 if (elem_type != NULL) {
2215 sharpened_klass = elem_type->klass();
2216 }
2217 }
2218 }
2219
2220 if (sharpened_klass != NULL) {
2221 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2222
2223 // Sharpen the value type.
2224 value_type = tjp;
2225
2226 #ifndef PRODUCT
2227 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2228 tty->print(" from base type: "); adr_type->dump();
2229 tty->print(" sharpened value: "); value_type->dump();
2230 }
2231 #endif
2232 }
2233 }
2234
2235 // Null check on self without removing any arguments. The argument
2236 // null check technically happens in the wrong place, which can lead to
2237 // invalid stack traces when the primitive is inlined into a method
2238 // which handles NullPointerExceptions.
2239 _sp += nargs;
2240 do_null_check(receiver, T_OBJECT);
2241 _sp -= nargs;
2242 if (stopped()) {
2243 return true;
2244 }
2245 // Heap pointers get a null-check from the interpreter,
2246 // as a courtesy. However, this is not guaranteed by Unsafe,
2247 // and it is not possible to fully distinguish unintended nulls
2248 // from intended ones in this API.
2249
2250 if (is_volatile) {
2251 // We need to emit leading and trailing CPU membars (see below) in
2252 // addition to memory membars when is_volatile. This is a little
2253 // too strong, but avoids the need to insert per-alias-type
2254 // volatile membars (for stores; compare Parse::do_put_xxx), which
2255 // we cannot do effectively here because we probably only have a
2256 // rough approximation of type.
2257 need_mem_bar = true;
2258 // For Stores, place a memory ordering barrier now.
2259 if (is_store)
2260 insert_mem_bar(Op_MemBarRelease);
2261 }
2262
2263 // Memory barrier to prevent normal and 'unsafe' accesses from
2264 // bypassing each other. Happens after null checks, so the
2265 // exception paths do not take memory state from the memory barrier,
2266 // so there's no problems making a strong assert about mixing users
2267 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2268 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2269 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2270
2271 if (!is_store) {
2272 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2273 // load value and push onto stack
2274 switch (type) {
2275 case T_BOOLEAN:
2276 case T_CHAR:
2277 case T_BYTE:
2278 case T_SHORT:
2279 case T_INT:
2280 case T_FLOAT:
2281 case T_OBJECT:
2282 push( p );
2283 break;
2284 case T_ADDRESS:
2285 // Cast to an int type.
2286 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2287 p = ConvX2L(p);
2288 push_pair(p);
2289 break;
2290 case T_DOUBLE:
2291 case T_LONG:
2292 push_pair( p );
2293 break;
2294 default: ShouldNotReachHere();
2295 }
2296 } else {
2297 // place effect of store into memory
2298 switch (type) {
2299 case T_DOUBLE:
2300 val = dstore_rounding(val);
2301 break;
2302 case T_ADDRESS:
2303 // Repackage the long as a pointer.
2304 val = ConvL2X(val);
2305 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2306 break;
2307 }
2308
2309 if (type != T_OBJECT ) {
2310 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2311 } else {
2312 // Possibly an oop being stored to Java heap or native memory
2313 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2314 // oop to Java heap.
2315 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2316 } else {
2317 // We can't tell at compile time if we are storing in the Java heap or outside
2318 // of it. So we need to emit code to conditionally do the proper type of
2319 // store.
2320
2321 IdealKit ideal(this);
2322 #define __ ideal.
2323 // QQQ who knows what probability is here??
2324 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2325 // Sync IdealKit and graphKit.
2326 sync_kit(ideal, false);
2327 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2328 // Update IdealKit memory.
2329 __ sync_kit(this);
2330 } __ else_(); {
2331 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2332 } __ end_if();
2333 // Final sync IdealKit and GraphKit.
2334 sync_kit(ideal);
2335 #undef __
2336 }
2337 }
2338 }
2339
2340 if (is_volatile) {
2341 if (!is_store)
2342 insert_mem_bar(Op_MemBarAcquire);
2343 else
2344 insert_mem_bar(Op_MemBarVolatile);
2345 }
2346
2347 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2348
2349 return true;
2350 }
2351
2352 //----------------------------inline_unsafe_prefetch----------------------------
2353
2354 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2355 #ifndef PRODUCT
2356 {
2357 ResourceMark rm;
2358 // Check the signatures.
2359 ciSignature* sig = signature();
2360 #ifdef ASSERT
2361 // Object getObject(Object base, int/long offset), etc.
2362 BasicType rtype = sig->return_type()->basic_type();
2363 if (!is_native_ptr) {
2364 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2365 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2366 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2367 } else {
2368 assert(sig->count() == 1, "native prefetch has 1 argument");
2369 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2370 }
2371 #endif // ASSERT
2372 }
2373 #endif // !PRODUCT
2374
2375 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2376
2377 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2378 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2379
2380 debug_only(int saved_sp = _sp);
2381 _sp += nargs;
2382
2383 // Build address expression. See the code in inline_unsafe_access.
2384 Node *adr;
2385 if (!is_native_ptr) {
2386 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2387 Node* offset = pop_pair();
2388 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2389 Node* base = pop();
2390 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2391 // to be plain byte offsets, which are also the same as those accepted
2392 // by oopDesc::field_base.
2393 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2394 "fieldOffset must be byte-scaled");
2395 // 32-bit machines ignore the high half!
2396 offset = ConvL2X(offset);
2397 adr = make_unsafe_address(base, offset);
2398 } else {
2399 Node* ptr = pop_pair();
2400 // Adjust Java long to machine word:
2401 ptr = ConvL2X(ptr);
2402 adr = make_unsafe_address(NULL, ptr);
2403 }
2404
2405 if (is_static) {
2406 assert(saved_sp == _sp, "must have correct argument count");
2407 } else {
2408 // Pop receiver last: it was pushed first.
2409 Node *receiver = pop();
2410 assert(saved_sp == _sp, "must have correct argument count");
2411
2412 // Null check on self without removing any arguments. The argument
2413 // null check technically happens in the wrong place, which can lead to
2414 // invalid stack traces when the primitive is inlined into a method
2415 // which handles NullPointerExceptions.
2416 _sp += nargs;
2417 do_null_check(receiver, T_OBJECT);
2418 _sp -= nargs;
2419 if (stopped()) {
2420 return true;
2421 }
2422 }
2423
2424 // Generate the read or write prefetch
2425 Node *prefetch;
2426 if (is_store) {
2427 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2428 } else {
2429 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2430 }
2431 prefetch->init_req(0, control());
2432 set_i_o(_gvn.transform(prefetch));
2433
2434 return true;
2435 }
2436
2437 //----------------------------inline_unsafe_CAS----------------------------
2438
2439 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2440 // This basic scheme here is the same as inline_unsafe_access, but
2441 // differs in enough details that combining them would make the code
2442 // overly confusing. (This is a true fact! I originally combined
2443 // them, but even I was confused by it!) As much code/comments as
2444 // possible are retained from inline_unsafe_access though to make
2445 // the correspondences clearer. - dl
2446
2447 if (callee()->is_static()) return false; // caller must have the capability!
2448
2449 #ifndef PRODUCT
2450 {
2451 ResourceMark rm;
2452 // Check the signatures.
2453 ciSignature* sig = signature();
2454 #ifdef ASSERT
2455 BasicType rtype = sig->return_type()->basic_type();
2456 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2457 assert(sig->count() == 4, "CAS has 4 arguments");
2458 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2459 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2460 #endif // ASSERT
2461 }
2462 #endif //PRODUCT
2463
2464 // number of stack slots per value argument (1 or 2)
2465 int type_words = type2size[type];
2466
2467 // Cannot inline wide CAS on machines that don't support it natively
2468 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2469 return false;
2470
2471 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2472
2473 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2474 int nargs = 1 + 1 + 2 + type_words + type_words;
2475
2476 // pop arguments: newval, oldval, offset, base, and receiver
2477 debug_only(int saved_sp = _sp);
2478 _sp += nargs;
2479 Node* newval = (type_words == 1) ? pop() : pop_pair();
2480 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2481 Node *offset = pop_pair();
2482 Node *base = pop();
2483 Node *receiver = pop();
2484 assert(saved_sp == _sp, "must have correct argument count");
2485
2486 // Null check receiver.
2487 _sp += nargs;
2488 do_null_check(receiver, T_OBJECT);
2489 _sp -= nargs;
2490 if (stopped()) {
2491 return true;
2492 }
2493
2494 // Build field offset expression.
2495 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2496 // to be plain byte offsets, which are also the same as those accepted
2497 // by oopDesc::field_base.
2498 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2499 // 32-bit machines ignore the high half of long offsets
2500 offset = ConvL2X(offset);
2501 Node* adr = make_unsafe_address(base, offset);
2502 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2503
2504 // (Unlike inline_unsafe_access, there seems no point in trying
2505 // to refine types. Just use the coarse types here.
2506 const Type *value_type = Type::get_const_basic_type(type);
2507 Compile::AliasType* alias_type = C->alias_type(adr_type);
2508 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2509 int alias_idx = C->get_alias_index(adr_type);
2510
2511 // Memory-model-wise, a CAS acts like a little synchronized block,
2512 // so needs barriers on each side. These don't translate into
2513 // actual barriers on most machines, but we still need rest of
2514 // compiler to respect ordering.
2515
2516 insert_mem_bar(Op_MemBarRelease);
2517 insert_mem_bar(Op_MemBarCPUOrder);
2518
2519 // 4984716: MemBars must be inserted before this
2520 // memory node in order to avoid a false
2521 // dependency which will confuse the scheduler.
2522 Node *mem = memory(alias_idx);
2523
2524 // For now, we handle only those cases that actually exist: ints,
2525 // longs, and Object. Adding others should be straightforward.
2526 Node* cas;
2527 switch(type) {
2528 case T_INT:
2529 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2530 break;
2531 case T_LONG:
2532 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2533 break;
2534 case T_OBJECT:
2535 // reference stores need a store barrier.
2536 // (They don't if CAS fails, but it isn't worth checking.)
2537 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2538 #ifdef _LP64
2539 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2540 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2541 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2542 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2543 newval_enc, oldval_enc));
2544 } else
2545 #endif
2546 {
2547 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2548 }
2549 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2550 break;
2551 default:
2552 ShouldNotReachHere();
2553 break;
2554 }
2555
2556 // SCMemProjNodes represent the memory state of CAS. Their main
2557 // role is to prevent CAS nodes from being optimized away when their
2558 // results aren't used.
2559 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2560 set_memory(proj, alias_idx);
2561
2562 // Add the trailing membar surrounding the access
2563 insert_mem_bar(Op_MemBarCPUOrder);
2564 insert_mem_bar(Op_MemBarAcquire);
2565
2566 push(cas);
2567 return true;
2568 }
2569
2570 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2571 // This is another variant of inline_unsafe_access, differing in
2572 // that it always issues store-store ("release") barrier and ensures
2573 // store-atomicity (which only matters for "long").
2574
2575 if (callee()->is_static()) return false; // caller must have the capability!
2576
2577 #ifndef PRODUCT
2578 {
2579 ResourceMark rm;
2580 // Check the signatures.
2581 ciSignature* sig = signature();
2582 #ifdef ASSERT
2583 BasicType rtype = sig->return_type()->basic_type();
2584 assert(rtype == T_VOID, "must return void");
2585 assert(sig->count() == 3, "has 3 arguments");
2586 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2587 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2588 #endif // ASSERT
2589 }
2590 #endif //PRODUCT
2591
2592 // number of stack slots per value argument (1 or 2)
2593 int type_words = type2size[type];
2594
2595 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2596
2597 // Argument words: "this" plus oop plus offset plus value;
2598 int nargs = 1 + 1 + 2 + type_words;
2599
2600 // pop arguments: val, offset, base, and receiver
2601 debug_only(int saved_sp = _sp);
2602 _sp += nargs;
2603 Node* val = (type_words == 1) ? pop() : pop_pair();
2604 Node *offset = pop_pair();
2605 Node *base = pop();
2606 Node *receiver = pop();
2607 assert(saved_sp == _sp, "must have correct argument count");
2608
2609 // Null check receiver.
2610 _sp += nargs;
2611 do_null_check(receiver, T_OBJECT);
2612 _sp -= nargs;
2613 if (stopped()) {
2614 return true;
2615 }
2616
2617 // Build field offset expression.
2618 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2619 // 32-bit machines ignore the high half of long offsets
2620 offset = ConvL2X(offset);
2621 Node* adr = make_unsafe_address(base, offset);
2622 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2623 const Type *value_type = Type::get_const_basic_type(type);
2624 Compile::AliasType* alias_type = C->alias_type(adr_type);
2625
2626 insert_mem_bar(Op_MemBarRelease);
2627 insert_mem_bar(Op_MemBarCPUOrder);
2628 // Ensure that the store is atomic for longs:
2629 bool require_atomic_access = true;
2630 Node* store;
2631 if (type == T_OBJECT) // reference stores need a store barrier.
2632 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2633 else {
2634 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2635 }
2636 insert_mem_bar(Op_MemBarCPUOrder);
2637 return true;
2638 }
2639
2640 bool LibraryCallKit::inline_unsafe_allocate() {
2641 if (callee()->is_static()) return false; // caller must have the capability!
2642 int nargs = 1 + 1;
2643 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2644 null_check_receiver(callee()); // check then ignore argument(0)
2645 _sp += nargs; // set original stack for use by uncommon_trap
2646 Node* cls = do_null_check(argument(1), T_OBJECT);
2647 _sp -= nargs;
2648 if (stopped()) return true;
2649
2650 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2651 _sp += nargs; // set original stack for use by uncommon_trap
2652 kls = do_null_check(kls, T_OBJECT);
2653 _sp -= nargs;
2654 if (stopped()) return true; // argument was like int.class
2655
2656 // Note: The argument might still be an illegal value like
2657 // Serializable.class or Object[].class. The runtime will handle it.
2658 // But we must make an explicit check for initialization.
2659 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2660 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2661 Node* bits = intcon(instanceKlass::fully_initialized);
2662 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2663 // The 'test' is non-zero if we need to take a slow path.
2664
2665 Node* obj = new_instance(kls, test);
2666 push(obj);
2667
2668 return true;
2669 }
2670
2671 //------------------------inline_native_time_funcs--------------
2672 // inline code for System.currentTimeMillis() and System.nanoTime()
2673 // these have the same type and signature
2674 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2675 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2676 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2677 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2678 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2679 const TypePtr* no_memory_effects = NULL;
2680 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2681 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2682 #ifdef ASSERT
2683 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2684 assert(value_top == top(), "second value must be top");
2685 #endif
2686 push_pair(value);
2687 return true;
2688 }
2689
2690 //------------------------inline_native_currentThread------------------
2691 bool LibraryCallKit::inline_native_currentThread() {
2692 Node* junk = NULL;
2693 push(generate_current_thread(junk));
2694 return true;
2695 }
2696
2697 //------------------------inline_native_isInterrupted------------------
2698 bool LibraryCallKit::inline_native_isInterrupted() {
2699 const int nargs = 1+1; // receiver + boolean
2700 assert(nargs == arg_size(), "sanity");
2701 // Add a fast path to t.isInterrupted(clear_int):
2702 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2703 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2704 // So, in the common case that the interrupt bit is false,
2705 // we avoid making a call into the VM. Even if the interrupt bit
2706 // is true, if the clear_int argument is false, we avoid the VM call.
2707 // However, if the receiver is not currentThread, we must call the VM,
2708 // because there must be some locking done around the operation.
2709
2710 // We only go to the fast case code if we pass two guards.
2711 // Paths which do not pass are accumulated in the slow_region.
2712 RegionNode* slow_region = new (C, 1) RegionNode(1);
2713 record_for_igvn(slow_region);
2714 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2715 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2716 enum { no_int_result_path = 1,
2717 no_clear_result_path = 2,
2718 slow_result_path = 3
2719 };
2720
2721 // (a) Receiving thread must be the current thread.
2722 Node* rec_thr = argument(0);
2723 Node* tls_ptr = NULL;
2724 Node* cur_thr = generate_current_thread(tls_ptr);
2725 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2726 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2727
2728 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2729 if (!known_current_thread)
2730 generate_slow_guard(bol_thr, slow_region);
2731
2732 // (b) Interrupt bit on TLS must be false.
2733 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2734 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2735 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2736 // Set the control input on the field _interrupted read to prevent it floating up.
2737 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2738 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2739 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2740
2741 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2742
2743 // First fast path: if (!TLS._interrupted) return false;
2744 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2745 result_rgn->init_req(no_int_result_path, false_bit);
2746 result_val->init_req(no_int_result_path, intcon(0));
2747
2748 // drop through to next case
2749 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2750
2751 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2752 Node* clr_arg = argument(1);
2753 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2754 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2755 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2756
2757 // Second fast path: ... else if (!clear_int) return true;
2758 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2759 result_rgn->init_req(no_clear_result_path, false_arg);
2760 result_val->init_req(no_clear_result_path, intcon(1));
2761
2762 // drop through to next case
2763 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2764
2765 // (d) Otherwise, go to the slow path.
2766 slow_region->add_req(control());
2767 set_control( _gvn.transform(slow_region) );
2768
2769 if (stopped()) {
2770 // There is no slow path.
2771 result_rgn->init_req(slow_result_path, top());
2772 result_val->init_req(slow_result_path, top());
2773 } else {
2774 // non-virtual because it is a private non-static
2775 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2776
2777 Node* slow_val = set_results_for_java_call(slow_call);
2778 // this->control() comes from set_results_for_java_call
2779
2780 // If we know that the result of the slow call will be true, tell the optimizer!
2781 if (known_current_thread) slow_val = intcon(1);
2782
2783 Node* fast_io = slow_call->in(TypeFunc::I_O);
2784 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2785 // These two phis are pre-filled with copies of of the fast IO and Memory
2786 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2787 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2788
2789 result_rgn->init_req(slow_result_path, control());
2790 io_phi ->init_req(slow_result_path, i_o());
2791 mem_phi ->init_req(slow_result_path, reset_memory());
2792 result_val->init_req(slow_result_path, slow_val);
2793
2794 set_all_memory( _gvn.transform(mem_phi) );
2795 set_i_o( _gvn.transform(io_phi) );
2796 }
2797
2798 push_result(result_rgn, result_val);
2799 C->set_has_split_ifs(true); // Has chance for split-if optimization
2800
2801 return true;
2802 }
2803
2804 //---------------------------load_mirror_from_klass----------------------------
2805 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2806 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2807 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2808 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2809 }
2810
2811 //-----------------------load_klass_from_mirror_common-------------------------
2812 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2813 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2814 // and branch to the given path on the region.
2815 // If never_see_null, take an uncommon trap on null, so we can optimistically
2816 // compile for the non-null case.
2817 // If the region is NULL, force never_see_null = true.
2818 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2819 bool never_see_null,
2820 int nargs,
2821 RegionNode* region,
2822 int null_path,
2823 int offset) {
2824 if (region == NULL) never_see_null = true;
2825 Node* p = basic_plus_adr(mirror, offset);
2826 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2827 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2828 _sp += nargs; // any deopt will start just before call to enclosing method
2829 Node* null_ctl = top();
2830 kls = null_check_oop(kls, &null_ctl, never_see_null);
2831 if (region != NULL) {
2832 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2833 region->init_req(null_path, null_ctl);
2834 } else {
2835 assert(null_ctl == top(), "no loose ends");
2836 }
2837 _sp -= nargs;
2838 return kls;
2839 }
2840
2841 //--------------------(inline_native_Class_query helpers)---------------------
2842 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2843 // Fall through if (mods & mask) == bits, take the guard otherwise.
2844 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2845 // Branch around if the given klass has the given modifier bit set.
2846 // Like generate_guard, adds a new path onto the region.
2847 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2848 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2849 Node* mask = intcon(modifier_mask);
2850 Node* bits = intcon(modifier_bits);
2851 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2852 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2853 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2854 return generate_fair_guard(bol, region);
2855 }
2856 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2857 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2858 }
2859
2860 //-------------------------inline_native_Class_query-------------------
2861 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2862 int nargs = 1+0; // just the Class mirror, in most cases
2863 const Type* return_type = TypeInt::BOOL;
2864 Node* prim_return_value = top(); // what happens if it's a primitive class?
2865 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2866 bool expect_prim = false; // most of these guys expect to work on refs
2867
2868 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2869
2870 switch (id) {
2871 case vmIntrinsics::_isInstance:
2872 nargs = 1+1; // the Class mirror, plus the object getting queried about
2873 // nothing is an instance of a primitive type
2874 prim_return_value = intcon(0);
2875 break;
2876 case vmIntrinsics::_getModifiers:
2877 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2878 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2879 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2880 break;
2881 case vmIntrinsics::_isInterface:
2882 prim_return_value = intcon(0);
2883 break;
2884 case vmIntrinsics::_isArray:
2885 prim_return_value = intcon(0);
2886 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2887 break;
2888 case vmIntrinsics::_isPrimitive:
2889 prim_return_value = intcon(1);
2890 expect_prim = true; // obviously
2891 break;
2892 case vmIntrinsics::_getSuperclass:
2893 prim_return_value = null();
2894 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2895 break;
2896 case vmIntrinsics::_getComponentType:
2897 prim_return_value = null();
2898 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2899 break;
2900 case vmIntrinsics::_getClassAccessFlags:
2901 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2902 return_type = TypeInt::INT; // not bool! 6297094
2903 break;
2904 default:
2905 ShouldNotReachHere();
2906 }
2907
2908 Node* mirror = argument(0);
2909 Node* obj = (nargs <= 1)? top(): argument(1);
2910
2911 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2912 if (mirror_con == NULL) return false; // cannot happen?
2913
2914 #ifndef PRODUCT
2915 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2916 ciType* k = mirror_con->java_mirror_type();
2917 if (k) {
2918 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2919 k->print_name();
2920 tty->cr();
2921 }
2922 }
2923 #endif
2924
2925 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2926 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2927 record_for_igvn(region);
2928 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2929
2930 // The mirror will never be null of Reflection.getClassAccessFlags, however
2931 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2932 // if it is. See bug 4774291.
2933
2934 // For Reflection.getClassAccessFlags(), the null check occurs in
2935 // the wrong place; see inline_unsafe_access(), above, for a similar
2936 // situation.
2937 _sp += nargs; // set original stack for use by uncommon_trap
2938 mirror = do_null_check(mirror, T_OBJECT);
2939 _sp -= nargs;
2940 // If mirror or obj is dead, only null-path is taken.
2941 if (stopped()) return true;
2942
2943 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2944
2945 // Now load the mirror's klass metaobject, and null-check it.
2946 // Side-effects region with the control path if the klass is null.
2947 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2948 region, _prim_path);
2949 // If kls is null, we have a primitive mirror.
2950 phi->init_req(_prim_path, prim_return_value);
2951 if (stopped()) { push_result(region, phi); return true; }
2952
2953 Node* p; // handy temp
2954 Node* null_ctl;
2955
2956 // Now that we have the non-null klass, we can perform the real query.
2957 // For constant classes, the query will constant-fold in LoadNode::Value.
2958 Node* query_value = top();
2959 switch (id) {
2960 case vmIntrinsics::_isInstance:
2961 // nothing is an instance of a primitive type
2962 _sp += nargs; // gen_instanceof might do an uncommon trap
2963 query_value = gen_instanceof(obj, kls);
2964 _sp -= nargs;
2965 break;
2966
2967 case vmIntrinsics::_getModifiers:
2968 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2969 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2970 break;
2971
2972 case vmIntrinsics::_isInterface:
2973 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2974 if (generate_interface_guard(kls, region) != NULL)
2975 // A guard was added. If the guard is taken, it was an interface.
2976 phi->add_req(intcon(1));
2977 // If we fall through, it's a plain class.
2978 query_value = intcon(0);
2979 break;
2980
2981 case vmIntrinsics::_isArray:
2982 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
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(intcon(1));
2986 // If we fall through, it's a plain class.
2987 query_value = intcon(0);
2988 break;
2989
2990 case vmIntrinsics::_isPrimitive:
2991 query_value = intcon(0); // "normal" path produces false
2992 break;
2993
2994 case vmIntrinsics::_getSuperclass:
2995 // The rules here are somewhat unfortunate, but we can still do better
2996 // with random logic than with a JNI call.
2997 // Interfaces store null or Object as _super, but must report null.
2998 // Arrays store an intermediate super as _super, but must report Object.
2999 // Other types can report the actual _super.
3000 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3001 if (generate_interface_guard(kls, region) != NULL)
3002 // A guard was added. If the guard is taken, it was an interface.
3003 phi->add_req(null());
3004 if (generate_array_guard(kls, region) != NULL)
3005 // A guard was added. If the guard is taken, it was an array.
3006 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3007 // If we fall through, it's a plain class. Get its _super.
3008 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
3009 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3010 null_ctl = top();
3011 kls = null_check_oop(kls, &null_ctl);
3012 if (null_ctl != top()) {
3013 // If the guard is taken, Object.superClass is null (both klass and mirror).
3014 region->add_req(null_ctl);
3015 phi ->add_req(null());
3016 }
3017 if (!stopped()) {
3018 query_value = load_mirror_from_klass(kls);
3019 }
3020 break;
3021
3022 case vmIntrinsics::_getComponentType:
3023 if (generate_array_guard(kls, region) != NULL) {
3024 // Be sure to pin the oop load to the guard edge just created:
3025 Node* is_array_ctrl = region->in(region->req()-1);
3026 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3027 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3028 phi->add_req(cmo);
3029 }
3030 query_value = null(); // non-array case is null
3031 break;
3032
3033 case vmIntrinsics::_getClassAccessFlags:
3034 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3035 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3036 break;
3037
3038 default:
3039 ShouldNotReachHere();
3040 }
3041
3042 // Fall-through is the normal case of a query to a real class.
3043 phi->init_req(1, query_value);
3044 region->init_req(1, control());
3045
3046 push_result(region, phi);
3047 C->set_has_split_ifs(true); // Has chance for split-if optimization
3048
3049 return true;
3050 }
3051
3052 //--------------------------inline_native_subtype_check------------------------
3053 // This intrinsic takes the JNI calls out of the heart of
3054 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3055 bool LibraryCallKit::inline_native_subtype_check() {
3056 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3057
3058 // Pull both arguments off the stack.
3059 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3060 args[0] = argument(0);
3061 args[1] = argument(1);
3062 Node* klasses[2]; // corresponding Klasses: superk, subk
3063 klasses[0] = klasses[1] = top();
3064
3065 enum {
3066 // A full decision tree on {superc is prim, subc is prim}:
3067 _prim_0_path = 1, // {P,N} => false
3068 // {P,P} & superc!=subc => false
3069 _prim_same_path, // {P,P} & superc==subc => true
3070 _prim_1_path, // {N,P} => false
3071 _ref_subtype_path, // {N,N} & subtype check wins => true
3072 _both_ref_path, // {N,N} & subtype check loses => false
3073 PATH_LIMIT
3074 };
3075
3076 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3077 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3078 record_for_igvn(region);
3079
3080 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3081 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3082 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3083
3084 // First null-check both mirrors and load each mirror's klass metaobject.
3085 int which_arg;
3086 for (which_arg = 0; which_arg <= 1; which_arg++) {
3087 Node* arg = args[which_arg];
3088 _sp += nargs; // set original stack for use by uncommon_trap
3089 arg = do_null_check(arg, T_OBJECT);
3090 _sp -= nargs;
3091 if (stopped()) break;
3092 args[which_arg] = _gvn.transform(arg);
3093
3094 Node* p = basic_plus_adr(arg, class_klass_offset);
3095 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3096 klasses[which_arg] = _gvn.transform(kls);
3097 }
3098
3099 // Having loaded both klasses, test each for null.
3100 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3101 for (which_arg = 0; which_arg <= 1; which_arg++) {
3102 Node* kls = klasses[which_arg];
3103 Node* null_ctl = top();
3104 _sp += nargs; // set original stack for use by uncommon_trap
3105 kls = null_check_oop(kls, &null_ctl, never_see_null);
3106 _sp -= nargs;
3107 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3108 region->init_req(prim_path, null_ctl);
3109 if (stopped()) break;
3110 klasses[which_arg] = kls;
3111 }
3112
3113 if (!stopped()) {
3114 // now we have two reference types, in klasses[0..1]
3115 Node* subk = klasses[1]; // the argument to isAssignableFrom
3116 Node* superk = klasses[0]; // the receiver
3117 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3118 // now we have a successful reference subtype check
3119 region->set_req(_ref_subtype_path, control());
3120 }
3121
3122 // If both operands are primitive (both klasses null), then
3123 // we must return true when they are identical primitives.
3124 // It is convenient to test this after the first null klass check.
3125 set_control(region->in(_prim_0_path)); // go back to first null check
3126 if (!stopped()) {
3127 // Since superc is primitive, make a guard for the superc==subc case.
3128 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3129 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3130 generate_guard(bol_eq, region, PROB_FAIR);
3131 if (region->req() == PATH_LIMIT+1) {
3132 // A guard was added. If the added guard is taken, superc==subc.
3133 region->swap_edges(PATH_LIMIT, _prim_same_path);
3134 region->del_req(PATH_LIMIT);
3135 }
3136 region->set_req(_prim_0_path, control()); // Not equal after all.
3137 }
3138
3139 // these are the only paths that produce 'true':
3140 phi->set_req(_prim_same_path, intcon(1));
3141 phi->set_req(_ref_subtype_path, intcon(1));
3142
3143 // pull together the cases:
3144 assert(region->req() == PATH_LIMIT, "sane region");
3145 for (uint i = 1; i < region->req(); i++) {
3146 Node* ctl = region->in(i);
3147 if (ctl == NULL || ctl == top()) {
3148 region->set_req(i, top());
3149 phi ->set_req(i, top());
3150 } else if (phi->in(i) == NULL) {
3151 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3152 }
3153 }
3154
3155 set_control(_gvn.transform(region));
3156 push(_gvn.transform(phi));
3157
3158 return true;
3159 }
3160
3161 //---------------------generate_array_guard_common------------------------
3162 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3163 bool obj_array, bool not_array) {
3164 // If obj_array/non_array==false/false:
3165 // Branch around if the given klass is in fact an array (either obj or prim).
3166 // If obj_array/non_array==false/true:
3167 // Branch around if the given klass is not an array klass of any kind.
3168 // If obj_array/non_array==true/true:
3169 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3170 // If obj_array/non_array==true/false:
3171 // Branch around if the kls is an oop array (Object[] or subtype)
3172 //
3173 // Like generate_guard, adds a new path onto the region.
3174 jint layout_con = 0;
3175 Node* layout_val = get_layout_helper(kls, layout_con);
3176 if (layout_val == NULL) {
3177 bool query = (obj_array
3178 ? Klass::layout_helper_is_objArray(layout_con)
3179 : Klass::layout_helper_is_javaArray(layout_con));
3180 if (query == not_array) {
3181 return NULL; // never a branch
3182 } else { // always a branch
3183 Node* always_branch = control();
3184 if (region != NULL)
3185 region->add_req(always_branch);
3186 set_control(top());
3187 return always_branch;
3188 }
3189 }
3190 // Now test the correct condition.
3191 jint nval = (obj_array
3192 ? ((jint)Klass::_lh_array_tag_type_value
3193 << Klass::_lh_array_tag_shift)
3194 : Klass::_lh_neutral_value);
3195 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3196 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3197 // invert the test if we are looking for a non-array
3198 if (not_array) btest = BoolTest(btest).negate();
3199 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3200 return generate_fair_guard(bol, region);
3201 }
3202
3203
3204 //-----------------------inline_native_newArray--------------------------
3205 bool LibraryCallKit::inline_native_newArray() {
3206 int nargs = 2;
3207 Node* mirror = argument(0);
3208 Node* count_val = argument(1);
3209
3210 _sp += nargs; // set original stack for use by uncommon_trap
3211 mirror = do_null_check(mirror, T_OBJECT);
3212 _sp -= nargs;
3213 // If mirror or obj is dead, only null-path is taken.
3214 if (stopped()) return true;
3215
3216 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3217 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3218 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3219 TypeInstPtr::NOTNULL);
3220 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3221 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3222 TypePtr::BOTTOM);
3223
3224 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3225 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3226 nargs,
3227 result_reg, _slow_path);
3228 Node* normal_ctl = control();
3229 Node* no_array_ctl = result_reg->in(_slow_path);
3230
3231 // Generate code for the slow case. We make a call to newArray().
3232 set_control(no_array_ctl);
3233 if (!stopped()) {
3234 // Either the input type is void.class, or else the
3235 // array klass has not yet been cached. Either the
3236 // ensuing call will throw an exception, or else it
3237 // will cache the array klass for next time.
3238 PreserveJVMState pjvms(this);
3239 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3240 Node* slow_result = set_results_for_java_call(slow_call);
3241 // this->control() comes from set_results_for_java_call
3242 result_reg->set_req(_slow_path, control());
3243 result_val->set_req(_slow_path, slow_result);
3244 result_io ->set_req(_slow_path, i_o());
3245 result_mem->set_req(_slow_path, reset_memory());
3246 }
3247
3248 set_control(normal_ctl);
3249 if (!stopped()) {
3250 // Normal case: The array type has been cached in the java.lang.Class.
3251 // The following call works fine even if the array type is polymorphic.
3252 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3253 Node* obj = new_array(klass_node, count_val, nargs);
3254 result_reg->init_req(_normal_path, control());
3255 result_val->init_req(_normal_path, obj);
3256 result_io ->init_req(_normal_path, i_o());
3257 result_mem->init_req(_normal_path, reset_memory());
3258 }
3259
3260 // Return the combined state.
3261 set_i_o( _gvn.transform(result_io) );
3262 set_all_memory( _gvn.transform(result_mem) );
3263 push_result(result_reg, result_val);
3264 C->set_has_split_ifs(true); // Has chance for split-if optimization
3265
3266 return true;
3267 }
3268
3269 //----------------------inline_native_getLength--------------------------
3270 bool LibraryCallKit::inline_native_getLength() {
3271 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3272
3273 int nargs = 1;
3274 Node* array = argument(0);
3275
3276 _sp += nargs; // set original stack for use by uncommon_trap
3277 array = do_null_check(array, T_OBJECT);
3278 _sp -= nargs;
3279
3280 // If array is dead, only null-path is taken.
3281 if (stopped()) return true;
3282
3283 // Deoptimize if it is a non-array.
3284 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3285
3286 if (non_array != NULL) {
3287 PreserveJVMState pjvms(this);
3288 set_control(non_array);
3289 _sp += nargs; // push the arguments back on the stack
3290 uncommon_trap(Deoptimization::Reason_intrinsic,
3291 Deoptimization::Action_maybe_recompile);
3292 }
3293
3294 // If control is dead, only non-array-path is taken.
3295 if (stopped()) return true;
3296
3297 // The works fine even if the array type is polymorphic.
3298 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3299 push( load_array_length(array) );
3300
3301 C->set_has_split_ifs(true); // Has chance for split-if optimization
3302
3303 return true;
3304 }
3305
3306 //------------------------inline_array_copyOf----------------------------
3307 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3308 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3309
3310 // Restore the stack and pop off the arguments.
3311 int nargs = 3 + (is_copyOfRange? 1: 0);
3312 Node* original = argument(0);
3313 Node* start = is_copyOfRange? argument(1): intcon(0);
3314 Node* end = is_copyOfRange? argument(2): argument(1);
3315 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3316
3317 Node* newcopy;
3318
3319 //set the original stack and the reexecute bit for the interpreter to reexecute
3320 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3321 { PreserveReexecuteState preexecs(this);
3322 _sp += nargs;
3323 jvms()->set_should_reexecute(true);
3324
3325 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3326 original = do_null_check(original, T_OBJECT);
3327
3328 // Check if a null path was taken unconditionally.
3329 if (stopped()) return true;
3330
3331 Node* orig_length = load_array_length(original);
3332
3333 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3334 NULL, 0);
3335 klass_node = do_null_check(klass_node, T_OBJECT);
3336
3337 RegionNode* bailout = new (C, 1) RegionNode(1);
3338 record_for_igvn(bailout);
3339
3340 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3341 // Bail out if that is so.
3342 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3343 if (not_objArray != NULL) {
3344 // Improve the klass node's type from the new optimistic assumption:
3345 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3346 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3347 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3348 cast->init_req(0, control());
3349 klass_node = _gvn.transform(cast);
3350 }
3351
3352 // Bail out if either start or end is negative.
3353 generate_negative_guard(start, bailout, &start);
3354 generate_negative_guard(end, bailout, &end);
3355
3356 Node* length = end;
3357 if (_gvn.type(start) != TypeInt::ZERO) {
3358 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3359 }
3360
3361 // Bail out if length is negative.
3362 // ...Not needed, since the new_array will throw the right exception.
3363 //generate_negative_guard(length, bailout, &length);
3364
3365 if (bailout->req() > 1) {
3366 PreserveJVMState pjvms(this);
3367 set_control( _gvn.transform(bailout) );
3368 uncommon_trap(Deoptimization::Reason_intrinsic,
3369 Deoptimization::Action_maybe_recompile);
3370 }
3371
3372 if (!stopped()) {
3373
3374 // How many elements will we copy from the original?
3375 // The answer is MinI(orig_length - start, length).
3376 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3377 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3378
3379 const bool raw_mem_only = true;
3380 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3381
3382 // Generate a direct call to the right arraycopy function(s).
3383 // We know the copy is disjoint but we might not know if the
3384 // oop stores need checking.
3385 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3386 // This will fail a store-check if x contains any non-nulls.
3387 bool disjoint_bases = true;
3388 bool length_never_negative = true;
3389 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3390 original, start, newcopy, intcon(0), moved,
3391 disjoint_bases, length_never_negative);
3392 }
3393 } //original reexecute and sp are set back here
3394
3395 if(!stopped()) {
3396 push(newcopy);
3397 }
3398
3399 C->set_has_split_ifs(true); // Has chance for split-if optimization
3400
3401 return true;
3402 }
3403
3404
3405 //----------------------generate_virtual_guard---------------------------
3406 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3407 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3408 RegionNode* slow_region) {
3409 ciMethod* method = callee();
3410 int vtable_index = method->vtable_index();
3411 // Get the methodOop out of the appropriate vtable entry.
3412 int entry_offset = (instanceKlass::vtable_start_offset() +
3413 vtable_index*vtableEntry::size()) * wordSize +
3414 vtableEntry::method_offset_in_bytes();
3415 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3416 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3417
3418 // Compare the target method with the expected method (e.g., Object.hashCode).
3419 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3420
3421 Node* native_call = makecon(native_call_addr);
3422 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3423 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3424
3425 return generate_slow_guard(test_native, slow_region);
3426 }
3427
3428 //-----------------------generate_method_call----------------------------
3429 // Use generate_method_call to make a slow-call to the real
3430 // method if the fast path fails. An alternative would be to
3431 // use a stub like OptoRuntime::slow_arraycopy_Java.
3432 // This only works for expanding the current library call,
3433 // not another intrinsic. (E.g., don't use this for making an
3434 // arraycopy call inside of the copyOf intrinsic.)
3435 CallJavaNode*
3436 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3437 // When compiling the intrinsic method itself, do not use this technique.
3438 guarantee(callee() != C->method(), "cannot make slow-call to self");
3439
3440 ciMethod* method = callee();
3441 // ensure the JVMS we have will be correct for this call
3442 guarantee(method_id == method->intrinsic_id(), "must match");
3443
3444 const TypeFunc* tf = TypeFunc::make(method);
3445 int tfdc = tf->domain()->cnt();
3446 CallJavaNode* slow_call;
3447 if (is_static) {
3448 assert(!is_virtual, "");
3449 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3450 SharedRuntime::get_resolve_static_call_stub(),
3451 method, bci());
3452 } else if (is_virtual) {
3453 null_check_receiver(method);
3454 int vtable_index = methodOopDesc::invalid_vtable_index;
3455 if (UseInlineCaches) {
3456 // Suppress the vtable call
3457 } else {
3458 // hashCode and clone are not a miranda methods,
3459 // so the vtable index is fixed.
3460 // No need to use the linkResolver to get it.
3461 vtable_index = method->vtable_index();
3462 }
3463 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3464 SharedRuntime::get_resolve_virtual_call_stub(),
3465 method, vtable_index, bci());
3466 } else { // neither virtual nor static: opt_virtual
3467 null_check_receiver(method);
3468 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3469 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3470 method, bci());
3471 slow_call->set_optimized_virtual(true);
3472 }
3473 set_arguments_for_java_call(slow_call);
3474 set_edges_for_java_call(slow_call);
3475 return slow_call;
3476 }
3477
3478
3479 //------------------------------inline_native_hashcode--------------------
3480 // Build special case code for calls to hashCode on an object.
3481 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3482 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3483 assert(!(is_virtual && is_static), "either virtual, special, or static");
3484
3485 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3486
3487 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3488 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3489 TypeInt::INT);
3490 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3491 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3492 TypePtr::BOTTOM);
3493 Node* obj = NULL;
3494 if (!is_static) {
3495 // Check for hashing null object
3496 obj = null_check_receiver(callee());
3497 if (stopped()) return true; // unconditionally null
3498 result_reg->init_req(_null_path, top());
3499 result_val->init_req(_null_path, top());
3500 } else {
3501 // Do a null check, and return zero if null.
3502 // System.identityHashCode(null) == 0
3503 obj = argument(0);
3504 Node* null_ctl = top();
3505 obj = null_check_oop(obj, &null_ctl);
3506 result_reg->init_req(_null_path, null_ctl);
3507 result_val->init_req(_null_path, _gvn.intcon(0));
3508 }
3509
3510 // Unconditionally null? Then return right away.
3511 if (stopped()) {
3512 set_control( result_reg->in(_null_path) );
3513 if (!stopped())
3514 push( result_val ->in(_null_path) );
3515 return true;
3516 }
3517
3518 // After null check, get the object's klass.
3519 Node* obj_klass = load_object_klass(obj);
3520
3521 // This call may be virtual (invokevirtual) or bound (invokespecial).
3522 // For each case we generate slightly different code.
3523
3524 // We only go to the fast case code if we pass a number of guards. The
3525 // paths which do not pass are accumulated in the slow_region.
3526 RegionNode* slow_region = new (C, 1) RegionNode(1);
3527 record_for_igvn(slow_region);
3528
3529 // If this is a virtual call, we generate a funny guard. We pull out
3530 // the vtable entry corresponding to hashCode() from the target object.
3531 // If the target method which we are calling happens to be the native
3532 // Object hashCode() method, we pass the guard. We do not need this
3533 // guard for non-virtual calls -- the caller is known to be the native
3534 // Object hashCode().
3535 if (is_virtual) {
3536 generate_virtual_guard(obj_klass, slow_region);
3537 }
3538
3539 // Get the header out of the object, use LoadMarkNode when available
3540 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3541 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3542
3543 // Test the header to see if it is unlocked.
3544 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3545 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3546 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3547 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3548 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3549
3550 generate_slow_guard(test_unlocked, slow_region);
3551
3552 // Get the hash value and check to see that it has been properly assigned.
3553 // We depend on hash_mask being at most 32 bits and avoid the use of
3554 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3555 // vm: see markOop.hpp.
3556 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3557 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3558 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3559 // This hack lets the hash bits live anywhere in the mark object now, as long
3560 // as the shift drops the relevant bits into the low 32 bits. Note that
3561 // Java spec says that HashCode is an int so there's no point in capturing
3562 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3563 hshifted_header = ConvX2I(hshifted_header);
3564 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3565
3566 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3567 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3568 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3569
3570 generate_slow_guard(test_assigned, slow_region);
3571
3572 Node* init_mem = reset_memory();
3573 // fill in the rest of the null path:
3574 result_io ->init_req(_null_path, i_o());
3575 result_mem->init_req(_null_path, init_mem);
3576
3577 result_val->init_req(_fast_path, hash_val);
3578 result_reg->init_req(_fast_path, control());
3579 result_io ->init_req(_fast_path, i_o());
3580 result_mem->init_req(_fast_path, init_mem);
3581
3582 // Generate code for the slow case. We make a call to hashCode().
3583 set_control(_gvn.transform(slow_region));
3584 if (!stopped()) {
3585 // No need for PreserveJVMState, because we're using up the present state.
3586 set_all_memory(init_mem);
3587 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3588 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3589 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3590 Node* slow_result = set_results_for_java_call(slow_call);
3591 // this->control() comes from set_results_for_java_call
3592 result_reg->init_req(_slow_path, control());
3593 result_val->init_req(_slow_path, slow_result);
3594 result_io ->set_req(_slow_path, i_o());
3595 result_mem ->set_req(_slow_path, reset_memory());
3596 }
3597
3598 // Return the combined state.
3599 set_i_o( _gvn.transform(result_io) );
3600 set_all_memory( _gvn.transform(result_mem) );
3601 push_result(result_reg, result_val);
3602
3603 return true;
3604 }
3605
3606 //---------------------------inline_native_getClass----------------------------
3607 // Build special case code for calls to getClass on an object.
3608 bool LibraryCallKit::inline_native_getClass() {
3609 Node* obj = null_check_receiver(callee());
3610 if (stopped()) return true;
3611 push( load_mirror_from_klass(load_object_klass(obj)) );
3612 return true;
3613 }
3614
3615 //-----------------inline_native_Reflection_getCallerClass---------------------
3616 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3617 //
3618 // NOTE that this code must perform the same logic as
3619 // vframeStream::security_get_caller_frame in that it must skip
3620 // Method.invoke() and auxiliary frames.
3621
3622
3623
3624
3625 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3626 ciMethod* method = callee();
3627
3628 #ifndef PRODUCT
3629 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3630 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3631 }
3632 #endif
3633
3634 debug_only(int saved_sp = _sp);
3635
3636 // Argument words: (int depth)
3637 int nargs = 1;
3638
3639 _sp += nargs;
3640 Node* caller_depth_node = pop();
3641
3642 assert(saved_sp == _sp, "must have correct argument count");
3643
3644 // The depth value must be a constant in order for the runtime call
3645 // to be eliminated.
3646 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3647 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3648 #ifndef PRODUCT
3649 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3650 tty->print_cr(" Bailing out because caller depth was not a constant");
3651 }
3652 #endif
3653 return false;
3654 }
3655 // Note that the JVM state at this point does not include the
3656 // getCallerClass() frame which we are trying to inline. The
3657 // semantics of getCallerClass(), however, are that the "first"
3658 // frame is the getCallerClass() frame, so we subtract one from the
3659 // requested depth before continuing. We don't inline requests of
3660 // getCallerClass(0).
3661 int caller_depth = caller_depth_type->get_con() - 1;
3662 if (caller_depth < 0) {
3663 #ifndef PRODUCT
3664 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3665 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3666 }
3667 #endif
3668 return false;
3669 }
3670
3671 if (!jvms()->has_method()) {
3672 #ifndef PRODUCT
3673 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3674 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3675 }
3676 #endif
3677 return false;
3678 }
3679 int _depth = jvms()->depth(); // cache call chain depth
3680
3681 // Walk back up the JVM state to find the caller at the required
3682 // depth. NOTE that this code must perform the same logic as
3683 // vframeStream::security_get_caller_frame in that it must skip
3684 // Method.invoke() and auxiliary frames. Note also that depth is
3685 // 1-based (1 is the bottom of the inlining).
3686 int inlining_depth = _depth;
3687 JVMState* caller_jvms = NULL;
3688
3689 if (inlining_depth > 0) {
3690 caller_jvms = jvms();
3691 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3692 do {
3693 // The following if-tests should be performed in this order
3694 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3695 // Skip a Method.invoke() or auxiliary frame
3696 } else if (caller_depth > 0) {
3697 // Skip real frame
3698 --caller_depth;
3699 } else {
3700 // We're done: reached desired caller after skipping.
3701 break;
3702 }
3703 caller_jvms = caller_jvms->caller();
3704 --inlining_depth;
3705 } while (inlining_depth > 0);
3706 }
3707
3708 if (inlining_depth == 0) {
3709 #ifndef PRODUCT
3710 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3711 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3712 tty->print_cr(" JVM state at this point:");
3713 for (int i = _depth; i >= 1; i--) {
3714 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3715 }
3716 }
3717 #endif
3718 return false; // Reached end of inlining
3719 }
3720
3721 // Acquire method holder as java.lang.Class
3722 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3723 ciInstance* caller_mirror = caller_klass->java_mirror();
3724 // Push this as a constant
3725 push(makecon(TypeInstPtr::make(caller_mirror)));
3726 #ifndef PRODUCT
3727 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3728 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);
3729 tty->print_cr(" JVM state at this point:");
3730 for (int i = _depth; i >= 1; i--) {
3731 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3732 }
3733 }
3734 #endif
3735 return true;
3736 }
3737
3738 // Helper routine for above
3739 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3740 ciMethod* method = jvms->method();
3741
3742 // Is this the Method.invoke method itself?
3743 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3744 return true;
3745
3746 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3747 ciKlass* k = method->holder();
3748 if (k->is_instance_klass()) {
3749 ciInstanceKlass* ik = k->as_instance_klass();
3750 for (; ik != NULL; ik = ik->super()) {
3751 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3752 ik == env()->find_system_klass(ik->name())) {
3753 return true;
3754 }
3755 }
3756 }
3757 else if (method->is_method_handle_adapter()) {
3758 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3759 return true;
3760 }
3761
3762 return false;
3763 }
3764
3765 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3766 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3767 // computing it since there is no lookup field by name function in the
3768 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3769 // Using a static variable here is safe even if we have multiple compilation
3770 // threads because the offset is constant. At worst the same offset will be
3771 // computed and stored multiple
3772
3773 bool LibraryCallKit::inline_native_AtomicLong_get() {
3774 // Restore the stack and pop off the argument
3775 _sp+=1;
3776 Node *obj = pop();
3777
3778 // get the offset of the "value" field. Since the CI interfaces
3779 // does not provide a way to look up a field by name, we scan the bytecodes
3780 // to get the field index. We expect the first 2 instructions of the method
3781 // to be:
3782 // 0 aload_0
3783 // 1 getfield "value"
3784 ciMethod* method = callee();
3785 if (value_field_offset == -1)
3786 {
3787 ciField* value_field;
3788 ciBytecodeStream iter(method);
3789 Bytecodes::Code bc = iter.next();
3790
3791 if ((bc != Bytecodes::_aload_0) &&
3792 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3793 return false;
3794 bc = iter.next();
3795 if (bc != Bytecodes::_getfield)
3796 return false;
3797 bool ignore;
3798 value_field = iter.get_field(ignore);
3799 value_field_offset = value_field->offset_in_bytes();
3800 }
3801
3802 // Null check without removing any arguments.
3803 _sp++;
3804 obj = do_null_check(obj, T_OBJECT);
3805 _sp--;
3806 // Check for locking null object
3807 if (stopped()) return true;
3808
3809 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3810 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3811 int alias_idx = C->get_alias_index(adr_type);
3812
3813 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3814
3815 push_pair(result);
3816
3817 return true;
3818 }
3819
3820 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3821 // Restore the stack and pop off the arguments
3822 _sp+=5;
3823 Node *newVal = pop_pair();
3824 Node *oldVal = pop_pair();
3825 Node *obj = pop();
3826
3827 // we need the offset of the "value" field which was computed when
3828 // inlining the get() method. Give up if we don't have it.
3829 if (value_field_offset == -1)
3830 return false;
3831
3832 // Null check without removing any arguments.
3833 _sp+=5;
3834 obj = do_null_check(obj, T_OBJECT);
3835 _sp-=5;
3836 // Check for locking null object
3837 if (stopped()) return true;
3838
3839 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3840 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3841 int alias_idx = C->get_alias_index(adr_type);
3842
3843 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3844 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3845 set_memory(store_proj, alias_idx);
3846 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3847
3848 Node *result;
3849 // CMove node is not used to be able fold a possible check code
3850 // after attemptUpdate() call. This code could be transformed
3851 // into CMove node by loop optimizations.
3852 {
3853 RegionNode *r = new (C, 3) RegionNode(3);
3854 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3855
3856 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3857 Node *iftrue = opt_iff(r, iff);
3858 r->init_req(1, iftrue);
3859 result->init_req(1, intcon(1));
3860 result->init_req(2, intcon(0));
3861
3862 set_control(_gvn.transform(r));
3863 record_for_igvn(r);
3864
3865 C->set_has_split_ifs(true); // Has chance for split-if optimization
3866 }
3867
3868 push(_gvn.transform(result));
3869 return true;
3870 }
3871
3872 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3873 // restore the arguments
3874 _sp += arg_size();
3875
3876 switch (id) {
3877 case vmIntrinsics::_floatToRawIntBits:
3878 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3879 break;
3880
3881 case vmIntrinsics::_intBitsToFloat:
3882 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3883 break;
3884
3885 case vmIntrinsics::_doubleToRawLongBits:
3886 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3887 break;
3888
3889 case vmIntrinsics::_longBitsToDouble:
3890 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3891 break;
3892
3893 case vmIntrinsics::_doubleToLongBits: {
3894 Node* value = pop_pair();
3895
3896 // two paths (plus control) merge in a wood
3897 RegionNode *r = new (C, 3) RegionNode(3);
3898 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3899
3900 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3901 // Build the boolean node
3902 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3903
3904 // Branch either way.
3905 // NaN case is less traveled, which makes all the difference.
3906 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3907 Node *opt_isnan = _gvn.transform(ifisnan);
3908 assert( opt_isnan->is_If(), "Expect an IfNode");
3909 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3910 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3911
3912 set_control(iftrue);
3913
3914 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3915 Node *slow_result = longcon(nan_bits); // return NaN
3916 phi->init_req(1, _gvn.transform( slow_result ));
3917 r->init_req(1, iftrue);
3918
3919 // Else fall through
3920 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3921 set_control(iffalse);
3922
3923 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3924 r->init_req(2, iffalse);
3925
3926 // Post merge
3927 set_control(_gvn.transform(r));
3928 record_for_igvn(r);
3929
3930 Node* result = _gvn.transform(phi);
3931 assert(result->bottom_type()->isa_long(), "must be");
3932 push_pair(result);
3933
3934 C->set_has_split_ifs(true); // Has chance for split-if optimization
3935
3936 break;
3937 }
3938
3939 case vmIntrinsics::_floatToIntBits: {
3940 Node* value = pop();
3941
3942 // two paths (plus control) merge in a wood
3943 RegionNode *r = new (C, 3) RegionNode(3);
3944 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3945
3946 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3947 // Build the boolean node
3948 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3949
3950 // Branch either way.
3951 // NaN case is less traveled, which makes all the difference.
3952 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3953 Node *opt_isnan = _gvn.transform(ifisnan);
3954 assert( opt_isnan->is_If(), "Expect an IfNode");
3955 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3956 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3957
3958 set_control(iftrue);
3959
3960 static const jint nan_bits = 0x7fc00000;
3961 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3962 phi->init_req(1, _gvn.transform( slow_result ));
3963 r->init_req(1, iftrue);
3964
3965 // Else fall through
3966 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3967 set_control(iffalse);
3968
3969 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3970 r->init_req(2, iffalse);
3971
3972 // Post merge
3973 set_control(_gvn.transform(r));
3974 record_for_igvn(r);
3975
3976 Node* result = _gvn.transform(phi);
3977 assert(result->bottom_type()->isa_int(), "must be");
3978 push(result);
3979
3980 C->set_has_split_ifs(true); // Has chance for split-if optimization
3981
3982 break;
3983 }
3984
3985 default:
3986 ShouldNotReachHere();
3987 }
3988
3989 return true;
3990 }
3991
3992 #ifdef _LP64
3993 #define XTOP ,top() /*additional argument*/
3994 #else //_LP64
3995 #define XTOP /*no additional argument*/
3996 #endif //_LP64
3997
3998 //----------------------inline_unsafe_copyMemory-------------------------
3999 bool LibraryCallKit::inline_unsafe_copyMemory() {
4000 if (callee()->is_static()) return false; // caller must have the capability!
4001 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4002 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4003 null_check_receiver(callee()); // check then ignore argument(0)
4004 if (stopped()) return true;
4005
4006 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4007
4008 Node* src_ptr = argument(1);
4009 Node* src_off = ConvL2X(argument(2));
4010 assert(argument(3)->is_top(), "2nd half of long");
4011 Node* dst_ptr = argument(4);
4012 Node* dst_off = ConvL2X(argument(5));
4013 assert(argument(6)->is_top(), "2nd half of long");
4014 Node* size = ConvL2X(argument(7));
4015 assert(argument(8)->is_top(), "2nd half of long");
4016
4017 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4018 "fieldOffset must be byte-scaled");
4019
4020 Node* src = make_unsafe_address(src_ptr, src_off);
4021 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4022
4023 // Conservatively insert a memory barrier on all memory slices.
4024 // Do not let writes of the copy source or destination float below the copy.
4025 insert_mem_bar(Op_MemBarCPUOrder);
4026
4027 // Call it. Note that the length argument is not scaled.
4028 make_runtime_call(RC_LEAF|RC_NO_FP,
4029 OptoRuntime::fast_arraycopy_Type(),
4030 StubRoutines::unsafe_arraycopy(),
4031 "unsafe_arraycopy",
4032 TypeRawPtr::BOTTOM,
4033 src, dst, size XTOP);
4034
4035 // Do not let reads of the copy destination float above the copy.
4036 insert_mem_bar(Op_MemBarCPUOrder);
4037
4038 return true;
4039 }
4040
4041 //------------------------clone_coping-----------------------------------
4042 // Helper function for inline_native_clone.
4043 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4044 assert(obj_size != NULL, "");
4045 Node* raw_obj = alloc_obj->in(1);
4046 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4047
4048 if (ReduceBulkZeroing) {
4049 // We will be completely responsible for initializing this object -
4050 // mark Initialize node as complete.
4051 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4052 // The object was just allocated - there should be no any stores!
4053 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4054 }
4055
4056 // Copy the fastest available way.
4057 // TODO: generate fields copies for small objects instead.
4058 Node* src = obj;
4059 Node* dest = alloc_obj;
4060 Node* size = _gvn.transform(obj_size);
4061
4062 // Exclude the header but include array length to copy by 8 bytes words.
4063 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4064 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4065 instanceOopDesc::base_offset_in_bytes();
4066 // base_off:
4067 // 8 - 32-bit VM
4068 // 12 - 64-bit VM, compressed oops
4069 // 16 - 64-bit VM, normal oops
4070 if (base_off % BytesPerLong != 0) {
4071 assert(UseCompressedOops, "");
4072 if (is_array) {
4073 // Exclude length to copy by 8 bytes words.
4074 base_off += sizeof(int);
4075 } else {
4076 // Include klass to copy by 8 bytes words.
4077 base_off = instanceOopDesc::klass_offset_in_bytes();
4078 }
4079 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4080 }
4081 src = basic_plus_adr(src, base_off);
4082 dest = basic_plus_adr(dest, base_off);
4083
4084 // Compute the length also, if needed:
4085 Node* countx = size;
4086 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4087 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4088
4089 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4090 bool disjoint_bases = true;
4091 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4092 src, NULL, dest, NULL, countx,
4093 /*dest_uninitialized*/true);
4094
4095 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4096 if (card_mark) {
4097 assert(!is_array, "");
4098 // Put in store barrier for any and all oops we are sticking
4099 // into this object. (We could avoid this if we could prove
4100 // that the object type contains no oop fields at all.)
4101 Node* no_particular_value = NULL;
4102 Node* no_particular_field = NULL;
4103 int raw_adr_idx = Compile::AliasIdxRaw;
4104 post_barrier(control(),
4105 memory(raw_adr_type),
4106 alloc_obj,
4107 no_particular_field,
4108 raw_adr_idx,
4109 no_particular_value,
4110 T_OBJECT,
4111 false);
4112 }
4113
4114 // Do not let reads from the cloned object float above the arraycopy.
4115 insert_mem_bar(Op_MemBarCPUOrder);
4116 }
4117
4118 //------------------------inline_native_clone----------------------------
4119 // Here are the simple edge cases:
4120 // null receiver => normal trap
4121 // virtual and clone was overridden => slow path to out-of-line clone
4122 // not cloneable or finalizer => slow path to out-of-line Object.clone
4123 //
4124 // The general case has two steps, allocation and copying.
4125 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4126 //
4127 // Copying also has two cases, oop arrays and everything else.
4128 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4129 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4130 //
4131 // These steps fold up nicely if and when the cloned object's klass
4132 // can be sharply typed as an object array, a type array, or an instance.
4133 //
4134 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4135 int nargs = 1;
4136 PhiNode* result_val;
4137
4138 //set the original stack and the reexecute bit for the interpreter to reexecute
4139 //the bytecode that invokes Object.clone if deoptimization happens
4140 { PreserveReexecuteState preexecs(this);
4141 jvms()->set_should_reexecute(true);
4142
4143 //null_check_receiver will adjust _sp (push and pop)
4144 Node* obj = null_check_receiver(callee());
4145 if (stopped()) return true;
4146
4147 _sp += nargs;
4148
4149 Node* obj_klass = load_object_klass(obj);
4150 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4151 const TypeOopPtr* toop = ((tklass != NULL)
4152 ? tklass->as_instance_type()
4153 : TypeInstPtr::NOTNULL);
4154
4155 // Conservatively insert a memory barrier on all memory slices.
4156 // Do not let writes into the original float below the clone.
4157 insert_mem_bar(Op_MemBarCPUOrder);
4158
4159 // paths into result_reg:
4160 enum {
4161 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4162 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4163 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4164 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4165 PATH_LIMIT
4166 };
4167 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4168 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4169 TypeInstPtr::NOTNULL);
4170 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4171 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4172 TypePtr::BOTTOM);
4173 record_for_igvn(result_reg);
4174
4175 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4176 int raw_adr_idx = Compile::AliasIdxRaw;
4177 const bool raw_mem_only = true;
4178
4179
4180 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4181 if (array_ctl != NULL) {
4182 // It's an array.
4183 PreserveJVMState pjvms(this);
4184 set_control(array_ctl);
4185 Node* obj_length = load_array_length(obj);
4186 Node* obj_size = NULL;
4187 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4188 raw_mem_only, &obj_size);
4189
4190 if (!use_ReduceInitialCardMarks()) {
4191 // If it is an oop array, it requires very special treatment,
4192 // because card marking is required on each card of the array.
4193 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4194 if (is_obja != NULL) {
4195 PreserveJVMState pjvms2(this);
4196 set_control(is_obja);
4197 // Generate a direct call to the right arraycopy function(s).
4198 bool disjoint_bases = true;
4199 bool length_never_negative = true;
4200 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4201 obj, intcon(0), alloc_obj, intcon(0),
4202 obj_length,
4203 disjoint_bases, length_never_negative);
4204 result_reg->init_req(_objArray_path, control());
4205 result_val->init_req(_objArray_path, alloc_obj);
4206 result_i_o ->set_req(_objArray_path, i_o());
4207 result_mem ->set_req(_objArray_path, reset_memory());
4208 }
4209 }
4210 // Otherwise, there are no card marks to worry about.
4211 // (We can dispense with card marks if we know the allocation
4212 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4213 // causes the non-eden paths to take compensating steps to
4214 // simulate a fresh allocation, so that no further
4215 // card marks are required in compiled code to initialize
4216 // the object.)
4217
4218 if (!stopped()) {
4219 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4220
4221 // Present the results of the copy.
4222 result_reg->init_req(_array_path, control());
4223 result_val->init_req(_array_path, alloc_obj);
4224 result_i_o ->set_req(_array_path, i_o());
4225 result_mem ->set_req(_array_path, reset_memory());
4226 }
4227 }
4228
4229 // We only go to the instance fast case code if we pass a number of guards.
4230 // The paths which do not pass are accumulated in the slow_region.
4231 RegionNode* slow_region = new (C, 1) RegionNode(1);
4232 record_for_igvn(slow_region);
4233 if (!stopped()) {
4234 // It's an instance (we did array above). Make the slow-path tests.
4235 // If this is a virtual call, we generate a funny guard. We grab
4236 // the vtable entry corresponding to clone() from the target object.
4237 // If the target method which we are calling happens to be the
4238 // Object clone() method, we pass the guard. We do not need this
4239 // guard for non-virtual calls; the caller is known to be the native
4240 // Object clone().
4241 if (is_virtual) {
4242 generate_virtual_guard(obj_klass, slow_region);
4243 }
4244
4245 // The object must be cloneable and must not have a finalizer.
4246 // Both of these conditions may be checked in a single test.
4247 // We could optimize the cloneable test further, but we don't care.
4248 generate_access_flags_guard(obj_klass,
4249 // Test both conditions:
4250 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4251 // Must be cloneable but not finalizer:
4252 JVM_ACC_IS_CLONEABLE,
4253 slow_region);
4254 }
4255
4256 if (!stopped()) {
4257 // It's an instance, and it passed the slow-path tests.
4258 PreserveJVMState pjvms(this);
4259 Node* obj_size = NULL;
4260 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4261
4262 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4263
4264 // Present the results of the slow call.
4265 result_reg->init_req(_instance_path, control());
4266 result_val->init_req(_instance_path, alloc_obj);
4267 result_i_o ->set_req(_instance_path, i_o());
4268 result_mem ->set_req(_instance_path, reset_memory());
4269 }
4270
4271 // Generate code for the slow case. We make a call to clone().
4272 set_control(_gvn.transform(slow_region));
4273 if (!stopped()) {
4274 PreserveJVMState pjvms(this);
4275 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4276 Node* slow_result = set_results_for_java_call(slow_call);
4277 // this->control() comes from set_results_for_java_call
4278 result_reg->init_req(_slow_path, control());
4279 result_val->init_req(_slow_path, slow_result);
4280 result_i_o ->set_req(_slow_path, i_o());
4281 result_mem ->set_req(_slow_path, reset_memory());
4282 }
4283
4284 // Return the combined state.
4285 set_control( _gvn.transform(result_reg) );
4286 set_i_o( _gvn.transform(result_i_o) );
4287 set_all_memory( _gvn.transform(result_mem) );
4288 } //original reexecute and sp are set back here
4289
4290 push(_gvn.transform(result_val));
4291
4292 return true;
4293 }
4294
4295
4296 // constants for computing the copy function
4297 enum {
4298 COPYFUNC_UNALIGNED = 0,
4299 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4300 COPYFUNC_CONJOINT = 0,
4301 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4302 };
4303
4304 // Note: The condition "disjoint" applies also for overlapping copies
4305 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4306 static address
4307 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name, bool dest_uninitialized) {
4308 int selector =
4309 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4310 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4311
4312 #define RETURN_STUB(xxx_arraycopy) { \
4313 name = #xxx_arraycopy; \
4314 return StubRoutines::xxx_arraycopy(); }
4315
4316 #define RETURN_STUB_PARM(xxx_arraycopy, parm) { \
4317 name = #xxx_arraycopy; \
4318 return StubRoutines::xxx_arraycopy(parm); }
4319
4320 switch (t) {
4321 case T_BYTE:
4322 case T_BOOLEAN:
4323 switch (selector) {
4324 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4325 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4326 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4327 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4328 }
4329 case T_CHAR:
4330 case T_SHORT:
4331 switch (selector) {
4332 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4333 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4334 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4335 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4336 }
4337 case T_INT:
4338 case T_FLOAT:
4339 switch (selector) {
4340 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4341 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4342 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4343 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4344 }
4345 case T_DOUBLE:
4346 case T_LONG:
4347 switch (selector) {
4348 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4349 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4350 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4351 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4352 }
4353 case T_ARRAY:
4354 case T_OBJECT:
4355 switch (selector) {
4356 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_arraycopy, dest_uninitialized);
4357 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_arraycopy, dest_uninitialized);
4358 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_disjoint_arraycopy, dest_uninitialized);
4359 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_disjoint_arraycopy, dest_uninitialized);
4360 }
4361 default:
4362 ShouldNotReachHere();
4363 return NULL;
4364 }
4365
4366 #undef RETURN_STUB
4367 #undef RETURN_STUB_PARM
4368 }
4369
4370 //------------------------------basictype2arraycopy----------------------------
4371 address LibraryCallKit::basictype2arraycopy(BasicType t,
4372 Node* src_offset,
4373 Node* dest_offset,
4374 bool disjoint_bases,
4375 const char* &name,
4376 bool dest_uninitialized) {
4377 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4378 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4379
4380 bool aligned = false;
4381 bool disjoint = disjoint_bases;
4382
4383 // if the offsets are the same, we can treat the memory regions as
4384 // disjoint, because either the memory regions are in different arrays,
4385 // or they are identical (which we can treat as disjoint.) We can also
4386 // treat a copy with a destination index less that the source index
4387 // as disjoint since a low->high copy will work correctly in this case.
4388 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4389 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4390 // both indices are constants
4391 int s_offs = src_offset_inttype->get_con();
4392 int d_offs = dest_offset_inttype->get_con();
4393 int element_size = type2aelembytes(t);
4394 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4395 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4396 if (s_offs >= d_offs) disjoint = true;
4397 } else if (src_offset == dest_offset && src_offset != NULL) {
4398 // This can occur if the offsets are identical non-constants.
4399 disjoint = true;
4400 }
4401
4402 return select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4403 }
4404
4405
4406 //------------------------------inline_arraycopy-----------------------
4407 bool LibraryCallKit::inline_arraycopy() {
4408 // Restore the stack and pop off the arguments.
4409 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4410 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4411
4412 Node *src = argument(0);
4413 Node *src_offset = argument(1);
4414 Node *dest = argument(2);
4415 Node *dest_offset = argument(3);
4416 Node *length = argument(4);
4417
4418 // Compile time checks. If any of these checks cannot be verified at compile time,
4419 // we do not make a fast path for this call. Instead, we let the call remain as it
4420 // is. The checks we choose to mandate at compile time are:
4421 //
4422 // (1) src and dest are arrays.
4423 const Type* src_type = src->Value(&_gvn);
4424 const Type* dest_type = dest->Value(&_gvn);
4425 const TypeAryPtr* top_src = src_type->isa_aryptr();
4426 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4427 if (top_src == NULL || top_src->klass() == NULL ||
4428 top_dest == NULL || top_dest->klass() == NULL) {
4429 // Conservatively insert a memory barrier on all memory slices.
4430 // Do not let writes into the source float below the arraycopy.
4431 insert_mem_bar(Op_MemBarCPUOrder);
4432
4433 // Call StubRoutines::generic_arraycopy stub.
4434 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4435 src, src_offset, dest, dest_offset, length);
4436
4437 // Do not let reads from the destination float above the arraycopy.
4438 // Since we cannot type the arrays, we don't know which slices
4439 // might be affected. We could restrict this barrier only to those
4440 // memory slices which pertain to array elements--but don't bother.
4441 if (!InsertMemBarAfterArraycopy)
4442 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4443 insert_mem_bar(Op_MemBarCPUOrder);
4444 return true;
4445 }
4446
4447 // (2) src and dest arrays must have elements of the same BasicType
4448 // Figure out the size and type of the elements we will be copying.
4449 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4450 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4451 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4452 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4453
4454 if (src_elem != dest_elem || dest_elem == T_VOID) {
4455 // The component types are not the same or are not recognized. Punt.
4456 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4457 generate_slow_arraycopy(TypePtr::BOTTOM,
4458 src, src_offset, dest, dest_offset, length,
4459 /*dest_uninitialized*/false);
4460 return true;
4461 }
4462
4463 //---------------------------------------------------------------------------
4464 // We will make a fast path for this call to arraycopy.
4465
4466 // We have the following tests left to perform:
4467 //
4468 // (3) src and dest must not be null.
4469 // (4) src_offset must not be negative.
4470 // (5) dest_offset must not be negative.
4471 // (6) length must not be negative.
4472 // (7) src_offset + length must not exceed length of src.
4473 // (8) dest_offset + length must not exceed length of dest.
4474 // (9) each element of an oop array must be assignable
4475
4476 RegionNode* slow_region = new (C, 1) RegionNode(1);
4477 record_for_igvn(slow_region);
4478
4479 // (3) operands must not be null
4480 // We currently perform our null checks with the do_null_check routine.
4481 // This means that the null exceptions will be reported in the caller
4482 // rather than (correctly) reported inside of the native arraycopy call.
4483 // This should be corrected, given time. We do our null check with the
4484 // stack pointer restored.
4485 _sp += nargs;
4486 src = do_null_check(src, T_ARRAY);
4487 dest = do_null_check(dest, T_ARRAY);
4488 _sp -= nargs;
4489
4490 // (4) src_offset must not be negative.
4491 generate_negative_guard(src_offset, slow_region);
4492
4493 // (5) dest_offset must not be negative.
4494 generate_negative_guard(dest_offset, slow_region);
4495
4496 // (6) length must not be negative (moved to generate_arraycopy()).
4497 // generate_negative_guard(length, slow_region);
4498
4499 // (7) src_offset + length must not exceed length of src.
4500 generate_limit_guard(src_offset, length,
4501 load_array_length(src),
4502 slow_region);
4503
4504 // (8) dest_offset + length must not exceed length of dest.
4505 generate_limit_guard(dest_offset, length,
4506 load_array_length(dest),
4507 slow_region);
4508
4509 // (9) each element of an oop array must be assignable
4510 // The generate_arraycopy subroutine checks this.
4511
4512 // This is where the memory effects are placed:
4513 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4514 generate_arraycopy(adr_type, dest_elem,
4515 src, src_offset, dest, dest_offset, length,
4516 false, false, slow_region);
4517
4518 return true;
4519 }
4520
4521 //-----------------------------generate_arraycopy----------------------
4522 // Generate an optimized call to arraycopy.
4523 // Caller must guard against non-arrays.
4524 // Caller must determine a common array basic-type for both arrays.
4525 // Caller must validate offsets against array bounds.
4526 // The slow_region has already collected guard failure paths
4527 // (such as out of bounds length or non-conformable array types).
4528 // The generated code has this shape, in general:
4529 //
4530 // if (length == 0) return // via zero_path
4531 // slowval = -1
4532 // if (types unknown) {
4533 // slowval = call generic copy loop
4534 // if (slowval == 0) return // via checked_path
4535 // } else if (indexes in bounds) {
4536 // if ((is object array) && !(array type check)) {
4537 // slowval = call checked copy loop
4538 // if (slowval == 0) return // via checked_path
4539 // } else {
4540 // call bulk copy loop
4541 // return // via fast_path
4542 // }
4543 // }
4544 // // adjust params for remaining work:
4545 // if (slowval != -1) {
4546 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4547 // }
4548 // slow_region:
4549 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4550 // return // via slow_call_path
4551 //
4552 // This routine is used from several intrinsics: System.arraycopy,
4553 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4554 //
4555 void
4556 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4557 BasicType basic_elem_type,
4558 Node* src, Node* src_offset,
4559 Node* dest, Node* dest_offset,
4560 Node* copy_length,
4561 bool disjoint_bases,
4562 bool length_never_negative,
4563 RegionNode* slow_region) {
4564
4565 if (slow_region == NULL) {
4566 slow_region = new(C,1) RegionNode(1);
4567 record_for_igvn(slow_region);
4568 }
4569
4570 Node* original_dest = dest;
4571 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4572 bool dest_uninitialized = false;
4573
4574 // See if this is the initialization of a newly-allocated array.
4575 // If so, we will take responsibility here for initializing it to zero.
4576 // (Note: Because tightly_coupled_allocation performs checks on the
4577 // out-edges of the dest, we need to avoid making derived pointers
4578 // from it until we have checked its uses.)
4579 if (ReduceBulkZeroing
4580 && !ZeroTLAB // pointless if already zeroed
4581 && basic_elem_type != T_CONFLICT // avoid corner case
4582 && !_gvn.eqv_uncast(src, dest)
4583 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4584 != NULL)
4585 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4586 && alloc->maybe_set_complete(&_gvn)) {
4587 // "You break it, you buy it."
4588 InitializeNode* init = alloc->initialization();
4589 assert(init->is_complete(), "we just did this");
4590 assert(dest->is_CheckCastPP(), "sanity");
4591 assert(dest->in(0)->in(0) == init, "dest pinned");
4592 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4593 // From this point on, every exit path is responsible for
4594 // initializing any non-copied parts of the object to zero.
4595 // Also, if this flag is set we make sure that arraycopy interacts properly
4596 // with G1, eliding pre-barriers. See CR 6627983.
4597 dest_uninitialized = true;
4598 } else {
4599 // No zeroing elimination here.
4600 alloc = NULL;
4601 //original_dest = dest;
4602 //dest_uninitialized = false;
4603 }
4604
4605 // Results are placed here:
4606 enum { fast_path = 1, // normal void-returning assembly stub
4607 checked_path = 2, // special assembly stub with cleanup
4608 slow_call_path = 3, // something went wrong; call the VM
4609 zero_path = 4, // bypass when length of copy is zero
4610 bcopy_path = 5, // copy primitive array by 64-bit blocks
4611 PATH_LIMIT = 6
4612 };
4613 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4614 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4615 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4616 record_for_igvn(result_region);
4617 _gvn.set_type_bottom(result_i_o);
4618 _gvn.set_type_bottom(result_memory);
4619 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4620
4621 // The slow_control path:
4622 Node* slow_control;
4623 Node* slow_i_o = i_o();
4624 Node* slow_mem = memory(adr_type);
4625 debug_only(slow_control = (Node*) badAddress);
4626
4627 // Checked control path:
4628 Node* checked_control = top();
4629 Node* checked_mem = NULL;
4630 Node* checked_i_o = NULL;
4631 Node* checked_value = NULL;
4632
4633 if (basic_elem_type == T_CONFLICT) {
4634 assert(!dest_uninitialized, "");
4635 Node* cv = generate_generic_arraycopy(adr_type,
4636 src, src_offset, dest, dest_offset,
4637 copy_length, dest_uninitialized);
4638 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4639 checked_control = control();
4640 checked_i_o = i_o();
4641 checked_mem = memory(adr_type);
4642 checked_value = cv;
4643 set_control(top()); // no fast path
4644 }
4645
4646 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4647 if (not_pos != NULL) {
4648 PreserveJVMState pjvms(this);
4649 set_control(not_pos);
4650
4651 // (6) length must not be negative.
4652 if (!length_never_negative) {
4653 generate_negative_guard(copy_length, slow_region);
4654 }
4655
4656 // copy_length is 0.
4657 if (!stopped() && dest_uninitialized) {
4658 Node* dest_length = alloc->in(AllocateNode::ALength);
4659 if (_gvn.eqv_uncast(copy_length, dest_length)
4660 || _gvn.find_int_con(dest_length, 1) <= 0) {
4661 // There is no zeroing to do. No need for a secondary raw memory barrier.
4662 } else {
4663 // Clear the whole thing since there are no source elements to copy.
4664 generate_clear_array(adr_type, dest, basic_elem_type,
4665 intcon(0), NULL,
4666 alloc->in(AllocateNode::AllocSize));
4667 // Use a secondary InitializeNode as raw memory barrier.
4668 // Currently it is needed only on this path since other
4669 // paths have stub or runtime calls as raw memory barriers.
4670 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4671 Compile::AliasIdxRaw,
4672 top())->as_Initialize();
4673 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4674 }
4675 }
4676
4677 // Present the results of the fast call.
4678 result_region->init_req(zero_path, control());
4679 result_i_o ->init_req(zero_path, i_o());
4680 result_memory->init_req(zero_path, memory(adr_type));
4681 }
4682
4683 if (!stopped() && dest_uninitialized) {
4684 // We have to initialize the *uncopied* part of the array to zero.
4685 // The copy destination is the slice dest[off..off+len]. The other slices
4686 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4687 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4688 Node* dest_length = alloc->in(AllocateNode::ALength);
4689 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4690 copy_length) );
4691
4692 // If there is a head section that needs zeroing, do it now.
4693 if (find_int_con(dest_offset, -1) != 0) {
4694 generate_clear_array(adr_type, dest, basic_elem_type,
4695 intcon(0), dest_offset,
4696 NULL);
4697 }
4698
4699 // Next, perform a dynamic check on the tail length.
4700 // It is often zero, and we can win big if we prove this.
4701 // There are two wins: Avoid generating the ClearArray
4702 // with its attendant messy index arithmetic, and upgrade
4703 // the copy to a more hardware-friendly word size of 64 bits.
4704 Node* tail_ctl = NULL;
4705 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4706 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4707 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4708 tail_ctl = generate_slow_guard(bol_lt, NULL);
4709 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4710 }
4711
4712 // At this point, let's assume there is no tail.
4713 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4714 // There is no tail. Try an upgrade to a 64-bit copy.
4715 bool didit = false;
4716 { PreserveJVMState pjvms(this);
4717 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4718 src, src_offset, dest, dest_offset,
4719 dest_size, dest_uninitialized);
4720 if (didit) {
4721 // Present the results of the block-copying fast call.
4722 result_region->init_req(bcopy_path, control());
4723 result_i_o ->init_req(bcopy_path, i_o());
4724 result_memory->init_req(bcopy_path, memory(adr_type));
4725 }
4726 }
4727 if (didit)
4728 set_control(top()); // no regular fast path
4729 }
4730
4731 // Clear the tail, if any.
4732 if (tail_ctl != NULL) {
4733 Node* notail_ctl = stopped() ? NULL : control();
4734 set_control(tail_ctl);
4735 if (notail_ctl == NULL) {
4736 generate_clear_array(adr_type, dest, basic_elem_type,
4737 dest_tail, NULL,
4738 dest_size);
4739 } else {
4740 // Make a local merge.
4741 Node* done_ctl = new(C,3) RegionNode(3);
4742 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4743 done_ctl->init_req(1, notail_ctl);
4744 done_mem->init_req(1, memory(adr_type));
4745 generate_clear_array(adr_type, dest, basic_elem_type,
4746 dest_tail, NULL,
4747 dest_size);
4748 done_ctl->init_req(2, control());
4749 done_mem->init_req(2, memory(adr_type));
4750 set_control( _gvn.transform(done_ctl) );
4751 set_memory( _gvn.transform(done_mem), adr_type );
4752 }
4753 }
4754 }
4755
4756 BasicType copy_type = basic_elem_type;
4757 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4758 if (!stopped() && copy_type == T_OBJECT) {
4759 // If src and dest have compatible element types, we can copy bits.
4760 // Types S[] and D[] are compatible if D is a supertype of S.
4761 //
4762 // If they are not, we will use checked_oop_disjoint_arraycopy,
4763 // which performs a fast optimistic per-oop check, and backs off
4764 // further to JVM_ArrayCopy on the first per-oop check that fails.
4765 // (Actually, we don't move raw bits only; the GC requires card marks.)
4766
4767 // Get the klassOop for both src and dest
4768 Node* src_klass = load_object_klass(src);
4769 Node* dest_klass = load_object_klass(dest);
4770
4771 // Generate the subtype check.
4772 // This might fold up statically, or then again it might not.
4773 //
4774 // Non-static example: Copying List<String>.elements to a new String[].
4775 // The backing store for a List<String> is always an Object[],
4776 // but its elements are always type String, if the generic types
4777 // are correct at the source level.
4778 //
4779 // Test S[] against D[], not S against D, because (probably)
4780 // the secondary supertype cache is less busy for S[] than S.
4781 // This usually only matters when D is an interface.
4782 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4783 // Plug failing path into checked_oop_disjoint_arraycopy
4784 if (not_subtype_ctrl != top()) {
4785 PreserveJVMState pjvms(this);
4786 set_control(not_subtype_ctrl);
4787 // (At this point we can assume disjoint_bases, since types differ.)
4788 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4789 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4790 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4791 Node* dest_elem_klass = _gvn.transform(n1);
4792 Node* cv = generate_checkcast_arraycopy(adr_type,
4793 dest_elem_klass,
4794 src, src_offset, dest, dest_offset,
4795 ConvI2X(copy_length), dest_uninitialized);
4796 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4797 checked_control = control();
4798 checked_i_o = i_o();
4799 checked_mem = memory(adr_type);
4800 checked_value = cv;
4801 }
4802 // At this point we know we do not need type checks on oop stores.
4803
4804 // Let's see if we need card marks:
4805 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4806 // If we do not need card marks, copy using the jint or jlong stub.
4807 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4808 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4809 "sizes agree");
4810 }
4811 }
4812
4813 if (!stopped()) {
4814 // Generate the fast path, if possible.
4815 PreserveJVMState pjvms(this);
4816 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4817 src, src_offset, dest, dest_offset,
4818 ConvI2X(copy_length), dest_uninitialized);
4819
4820 // Present the results of the fast call.
4821 result_region->init_req(fast_path, control());
4822 result_i_o ->init_req(fast_path, i_o());
4823 result_memory->init_req(fast_path, memory(adr_type));
4824 }
4825
4826 // Here are all the slow paths up to this point, in one bundle:
4827 slow_control = top();
4828 if (slow_region != NULL)
4829 slow_control = _gvn.transform(slow_region);
4830 debug_only(slow_region = (RegionNode*)badAddress);
4831
4832 set_control(checked_control);
4833 if (!stopped()) {
4834 // Clean up after the checked call.
4835 // The returned value is either 0 or -1^K,
4836 // where K = number of partially transferred array elements.
4837 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4838 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4839 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4840
4841 // If it is 0, we are done, so transfer to the end.
4842 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4843 result_region->init_req(checked_path, checks_done);
4844 result_i_o ->init_req(checked_path, checked_i_o);
4845 result_memory->init_req(checked_path, checked_mem);
4846
4847 // If it is not zero, merge into the slow call.
4848 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4849 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4850 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4851 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4852 record_for_igvn(slow_reg2);
4853 slow_reg2 ->init_req(1, slow_control);
4854 slow_i_o2 ->init_req(1, slow_i_o);
4855 slow_mem2 ->init_req(1, slow_mem);
4856 slow_reg2 ->init_req(2, control());
4857 slow_i_o2 ->init_req(2, checked_i_o);
4858 slow_mem2 ->init_req(2, checked_mem);
4859
4860 slow_control = _gvn.transform(slow_reg2);
4861 slow_i_o = _gvn.transform(slow_i_o2);
4862 slow_mem = _gvn.transform(slow_mem2);
4863
4864 if (alloc != NULL) {
4865 // We'll restart from the very beginning, after zeroing the whole thing.
4866 // This can cause double writes, but that's OK since dest is brand new.
4867 // So we ignore the low 31 bits of the value returned from the stub.
4868 } else {
4869 // We must continue the copy exactly where it failed, or else
4870 // another thread might see the wrong number of writes to dest.
4871 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4872 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4873 slow_offset->init_req(1, intcon(0));
4874 slow_offset->init_req(2, checked_offset);
4875 slow_offset = _gvn.transform(slow_offset);
4876
4877 // Adjust the arguments by the conditionally incoming offset.
4878 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4879 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4880 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4881
4882 // Tweak the node variables to adjust the code produced below:
4883 src_offset = src_off_plus;
4884 dest_offset = dest_off_plus;
4885 copy_length = length_minus;
4886 }
4887 }
4888
4889 set_control(slow_control);
4890 if (!stopped()) {
4891 // Generate the slow path, if needed.
4892 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4893
4894 set_memory(slow_mem, adr_type);
4895 set_i_o(slow_i_o);
4896
4897 if (dest_uninitialized) {
4898 generate_clear_array(adr_type, dest, basic_elem_type,
4899 intcon(0), NULL,
4900 alloc->in(AllocateNode::AllocSize));
4901 }
4902
4903 generate_slow_arraycopy(adr_type,
4904 src, src_offset, dest, dest_offset,
4905 copy_length, /*dest_uninitialized*/false);
4906
4907 result_region->init_req(slow_call_path, control());
4908 result_i_o ->init_req(slow_call_path, i_o());
4909 result_memory->init_req(slow_call_path, memory(adr_type));
4910 }
4911
4912 // Remove unused edges.
4913 for (uint i = 1; i < result_region->req(); i++) {
4914 if (result_region->in(i) == NULL)
4915 result_region->init_req(i, top());
4916 }
4917
4918 // Finished; return the combined state.
4919 set_control( _gvn.transform(result_region) );
4920 set_i_o( _gvn.transform(result_i_o) );
4921 set_memory( _gvn.transform(result_memory), adr_type );
4922
4923 // The memory edges above are precise in order to model effects around
4924 // array copies accurately to allow value numbering of field loads around
4925 // arraycopy. Such field loads, both before and after, are common in Java
4926 // collections and similar classes involving header/array data structures.
4927 //
4928 // But with low number of register or when some registers are used or killed
4929 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4930 // The next memory barrier is added to avoid it. If the arraycopy can be
4931 // optimized away (which it can, sometimes) then we can manually remove
4932 // the membar also.
4933 //
4934 // Do not let reads from the cloned object float above the arraycopy.
4935 if (InsertMemBarAfterArraycopy || alloc != NULL)
4936 insert_mem_bar(Op_MemBarCPUOrder);
4937 }
4938
4939
4940 // Helper function which determines if an arraycopy immediately follows
4941 // an allocation, with no intervening tests or other escapes for the object.
4942 AllocateArrayNode*
4943 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4944 RegionNode* slow_region) {
4945 if (stopped()) return NULL; // no fast path
4946 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4947
4948 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4949 if (alloc == NULL) return NULL;
4950
4951 Node* rawmem = memory(Compile::AliasIdxRaw);
4952 // Is the allocation's memory state untouched?
4953 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4954 // Bail out if there have been raw-memory effects since the allocation.
4955 // (Example: There might have been a call or safepoint.)
4956 return NULL;
4957 }
4958 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4959 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4960 return NULL;
4961 }
4962
4963 // There must be no unexpected observers of this allocation.
4964 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4965 Node* obs = ptr->fast_out(i);
4966 if (obs != this->map()) {
4967 return NULL;
4968 }
4969 }
4970
4971 // This arraycopy must unconditionally follow the allocation of the ptr.
4972 Node* alloc_ctl = ptr->in(0);
4973 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4974
4975 Node* ctl = control();
4976 while (ctl != alloc_ctl) {
4977 // There may be guards which feed into the slow_region.
4978 // Any other control flow means that we might not get a chance
4979 // to finish initializing the allocated object.
4980 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4981 IfNode* iff = ctl->in(0)->as_If();
4982 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4983 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4984 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4985 ctl = iff->in(0); // This test feeds the known slow_region.
4986 continue;
4987 }
4988 // One more try: Various low-level checks bottom out in
4989 // uncommon traps. If the debug-info of the trap omits
4990 // any reference to the allocation, as we've already
4991 // observed, then there can be no objection to the trap.
4992 bool found_trap = false;
4993 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4994 Node* obs = not_ctl->fast_out(j);
4995 if (obs->in(0) == not_ctl && obs->is_Call() &&
4996 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4997 found_trap = true; break;
4998 }
4999 }
5000 if (found_trap) {
5001 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5002 continue;
5003 }
5004 }
5005 return NULL;
5006 }
5007
5008 // If we get this far, we have an allocation which immediately
5009 // precedes the arraycopy, and we can take over zeroing the new object.
5010 // The arraycopy will finish the initialization, and provide
5011 // a new control state to which we will anchor the destination pointer.
5012
5013 return alloc;
5014 }
5015
5016 // Helper for initialization of arrays, creating a ClearArray.
5017 // It writes zero bits in [start..end), within the body of an array object.
5018 // The memory effects are all chained onto the 'adr_type' alias category.
5019 //
5020 // Since the object is otherwise uninitialized, we are free
5021 // to put a little "slop" around the edges of the cleared area,
5022 // as long as it does not go back into the array's header,
5023 // or beyond the array end within the heap.
5024 //
5025 // The lower edge can be rounded down to the nearest jint and the
5026 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5027 //
5028 // Arguments:
5029 // adr_type memory slice where writes are generated
5030 // dest oop of the destination array
5031 // basic_elem_type element type of the destination
5032 // slice_idx array index of first element to store
5033 // slice_len number of elements to store (or NULL)
5034 // dest_size total size in bytes of the array object
5035 //
5036 // Exactly one of slice_len or dest_size must be non-NULL.
5037 // If dest_size is non-NULL, zeroing extends to the end of the object.
5038 // If slice_len is non-NULL, the slice_idx value must be a constant.
5039 void
5040 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5041 Node* dest,
5042 BasicType basic_elem_type,
5043 Node* slice_idx,
5044 Node* slice_len,
5045 Node* dest_size) {
5046 // one or the other but not both of slice_len and dest_size:
5047 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5048 if (slice_len == NULL) slice_len = top();
5049 if (dest_size == NULL) dest_size = top();
5050
5051 // operate on this memory slice:
5052 Node* mem = memory(adr_type); // memory slice to operate on
5053
5054 // scaling and rounding of indexes:
5055 int scale = exact_log2(type2aelembytes(basic_elem_type));
5056 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5057 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5058 int bump_bit = (-1 << scale) & BytesPerInt;
5059
5060 // determine constant starts and ends
5061 const intptr_t BIG_NEG = -128;
5062 assert(BIG_NEG + 2*abase < 0, "neg enough");
5063 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5064 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5065 if (slice_len_con == 0) {
5066 return; // nothing to do here
5067 }
5068 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5069 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5070 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5071 assert(end_con < 0, "not two cons");
5072 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5073 BytesPerLong);
5074 }
5075
5076 if (start_con >= 0 && end_con >= 0) {
5077 // Constant start and end. Simple.
5078 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5079 start_con, end_con, &_gvn);
5080 } else if (start_con >= 0 && dest_size != top()) {
5081 // Constant start, pre-rounded end after the tail of the array.
5082 Node* end = dest_size;
5083 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5084 start_con, end, &_gvn);
5085 } else if (start_con >= 0 && slice_len != top()) {
5086 // Constant start, non-constant end. End needs rounding up.
5087 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5088 intptr_t end_base = abase + (slice_idx_con << scale);
5089 int end_round = (-1 << scale) & (BytesPerLong - 1);
5090 Node* end = ConvI2X(slice_len);
5091 if (scale != 0)
5092 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5093 end_base += end_round;
5094 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5095 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5096 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5097 start_con, end, &_gvn);
5098 } else if (start_con < 0 && dest_size != top()) {
5099 // Non-constant start, pre-rounded end after the tail of the array.
5100 // This is almost certainly a "round-to-end" operation.
5101 Node* start = slice_idx;
5102 start = ConvI2X(start);
5103 if (scale != 0)
5104 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5105 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5106 if ((bump_bit | clear_low) != 0) {
5107 int to_clear = (bump_bit | clear_low);
5108 // Align up mod 8, then store a jint zero unconditionally
5109 // just before the mod-8 boundary.
5110 if (((abase + bump_bit) & ~to_clear) - bump_bit
5111 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5112 bump_bit = 0;
5113 assert((abase & to_clear) == 0, "array base must be long-aligned");
5114 } else {
5115 // Bump 'start' up to (or past) the next jint boundary:
5116 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5117 assert((abase & clear_low) == 0, "array base must be int-aligned");
5118 }
5119 // Round bumped 'start' down to jlong boundary in body of array.
5120 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5121 if (bump_bit != 0) {
5122 // Store a zero to the immediately preceding jint:
5123 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5124 Node* p1 = basic_plus_adr(dest, x1);
5125 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5126 mem = _gvn.transform(mem);
5127 }
5128 }
5129 Node* end = dest_size; // pre-rounded
5130 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5131 start, end, &_gvn);
5132 } else {
5133 // Non-constant start, unrounded non-constant end.
5134 // (Nobody zeroes a random midsection of an array using this routine.)
5135 ShouldNotReachHere(); // fix caller
5136 }
5137
5138 // Done.
5139 set_memory(mem, adr_type);
5140 }
5141
5142
5143 bool
5144 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5145 BasicType basic_elem_type,
5146 AllocateNode* alloc,
5147 Node* src, Node* src_offset,
5148 Node* dest, Node* dest_offset,
5149 Node* dest_size, bool dest_uninitialized) {
5150 // See if there is an advantage from block transfer.
5151 int scale = exact_log2(type2aelembytes(basic_elem_type));
5152 if (scale >= LogBytesPerLong)
5153 return false; // it is already a block transfer
5154
5155 // Look at the alignment of the starting offsets.
5156 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5157 const intptr_t BIG_NEG = -128;
5158 assert(BIG_NEG + 2*abase < 0, "neg enough");
5159
5160 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5161 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5162 if (src_off < 0 || dest_off < 0)
5163 // At present, we can only understand constants.
5164 return false;
5165
5166 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5167 // Non-aligned; too bad.
5168 // One more chance: Pick off an initial 32-bit word.
5169 // This is a common case, since abase can be odd mod 8.
5170 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5171 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5172 Node* sptr = basic_plus_adr(src, src_off);
5173 Node* dptr = basic_plus_adr(dest, dest_off);
5174 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5175 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5176 src_off += BytesPerInt;
5177 dest_off += BytesPerInt;
5178 } else {
5179 return false;
5180 }
5181 }
5182 assert(src_off % BytesPerLong == 0, "");
5183 assert(dest_off % BytesPerLong == 0, "");
5184
5185 // Do this copy by giant steps.
5186 Node* sptr = basic_plus_adr(src, src_off);
5187 Node* dptr = basic_plus_adr(dest, dest_off);
5188 Node* countx = dest_size;
5189 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5190 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5191
5192 bool disjoint_bases = true; // since alloc != NULL
5193 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5194 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5195
5196 return true;
5197 }
5198
5199
5200 // Helper function; generates code for the slow case.
5201 // We make a call to a runtime method which emulates the native method,
5202 // but without the native wrapper overhead.
5203 void
5204 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5205 Node* src, Node* src_offset,
5206 Node* dest, Node* dest_offset,
5207 Node* copy_length, bool dest_uninitialized) {
5208 assert(!dest_uninitialized, "Invariant");
5209 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5210 OptoRuntime::slow_arraycopy_Type(),
5211 OptoRuntime::slow_arraycopy_Java(),
5212 "slow_arraycopy", adr_type,
5213 src, src_offset, dest, dest_offset,
5214 copy_length);
5215
5216 // Handle exceptions thrown by this fellow:
5217 make_slow_call_ex(call, env()->Throwable_klass(), false);
5218 }
5219
5220 // Helper function; generates code for cases requiring runtime checks.
5221 Node*
5222 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5223 Node* dest_elem_klass,
5224 Node* src, Node* src_offset,
5225 Node* dest, Node* dest_offset,
5226 Node* copy_length, bool dest_uninitialized) {
5227 if (stopped()) return NULL;
5228
5229 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5230 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5231 return NULL;
5232 }
5233
5234 // Pick out the parameters required to perform a store-check
5235 // for the target array. This is an optimistic check. It will
5236 // look in each non-null element's class, at the desired klass's
5237 // super_check_offset, for the desired klass.
5238 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5239 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5240 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5241 Node* check_offset = ConvI2X(_gvn.transform(n3));
5242 Node* check_value = dest_elem_klass;
5243
5244 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5245 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5246
5247 // (We know the arrays are never conjoint, because their types differ.)
5248 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5249 OptoRuntime::checkcast_arraycopy_Type(),
5250 copyfunc_addr, "checkcast_arraycopy", adr_type,
5251 // five arguments, of which two are
5252 // intptr_t (jlong in LP64)
5253 src_start, dest_start,
5254 copy_length XTOP,
5255 check_offset XTOP,
5256 check_value);
5257
5258 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5259 }
5260
5261
5262 // Helper function; generates code for cases requiring runtime checks.
5263 Node*
5264 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5265 Node* src, Node* src_offset,
5266 Node* dest, Node* dest_offset,
5267 Node* copy_length, bool dest_uninitialized) {
5268 assert(!dest_uninitialized, "Invariant");
5269 if (stopped()) return NULL;
5270 address copyfunc_addr = StubRoutines::generic_arraycopy();
5271 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5272 return NULL;
5273 }
5274
5275 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5276 OptoRuntime::generic_arraycopy_Type(),
5277 copyfunc_addr, "generic_arraycopy", adr_type,
5278 src, src_offset, dest, dest_offset, copy_length);
5279
5280 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5281 }
5282
5283 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5284 void
5285 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5286 BasicType basic_elem_type,
5287 bool disjoint_bases,
5288 Node* src, Node* src_offset,
5289 Node* dest, Node* dest_offset,
5290 Node* copy_length, bool dest_uninitialized) {
5291 if (stopped()) return; // nothing to do
5292
5293 Node* src_start = src;
5294 Node* dest_start = dest;
5295 if (src_offset != NULL || dest_offset != NULL) {
5296 assert(src_offset != NULL && dest_offset != NULL, "");
5297 src_start = array_element_address(src, src_offset, basic_elem_type);
5298 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5299 }
5300
5301 // Figure out which arraycopy runtime method to call.
5302 const char* copyfunc_name = "arraycopy";
5303 address copyfunc_addr =
5304 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5305 disjoint_bases, copyfunc_name, dest_uninitialized);
5306
5307 // Call it. Note that the count_ix value is not scaled to a byte-size.
5308 make_runtime_call(RC_LEAF|RC_NO_FP,
5309 OptoRuntime::fast_arraycopy_Type(),
5310 copyfunc_addr, copyfunc_name, adr_type,
5311 src_start, dest_start, copy_length XTOP);
5312 }