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 // Helper for inline_unsafe_access.
170 // Generates the guards that check whether the result of
171 // Unsafe.getObject should be recorded in an SATB log buffer.
172 void insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val);
173 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
174 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
175 bool inline_unsafe_allocate();
176 bool inline_unsafe_copyMemory();
177 bool inline_native_currentThread();
178 bool inline_native_time_funcs(bool isNano);
179 bool inline_native_isInterrupted();
180 bool inline_native_Class_query(vmIntrinsics::ID id);
181 bool inline_native_subtype_check();
182
183 bool inline_native_newArray();
184 bool inline_native_getLength();
185 bool inline_array_copyOf(bool is_copyOfRange);
186 bool inline_array_equals();
187 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
188 bool inline_native_clone(bool is_virtual);
189 bool inline_native_Reflection_getCallerClass();
190 bool inline_native_AtomicLong_get();
191 bool inline_native_AtomicLong_attemptUpdate();
192 bool is_method_invoke_or_aux_frame(JVMState* jvms);
193 // Helper function for inlining native object hash method
194 bool inline_native_hashcode(bool is_virtual, bool is_static);
195 bool inline_native_getClass();
196
197 // Helper functions for inlining arraycopy
198 bool inline_arraycopy();
199 void generate_arraycopy(const TypePtr* adr_type,
200 BasicType basic_elem_type,
201 Node* src, Node* src_offset,
202 Node* dest, Node* dest_offset,
203 Node* copy_length,
204 bool disjoint_bases = false,
205 bool length_never_negative = false,
206 RegionNode* slow_region = NULL);
207 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
208 RegionNode* slow_region);
209 void generate_clear_array(const TypePtr* adr_type,
210 Node* dest,
211 BasicType basic_elem_type,
212 Node* slice_off,
213 Node* slice_len,
214 Node* slice_end);
215 bool generate_block_arraycopy(const TypePtr* adr_type,
216 BasicType basic_elem_type,
217 AllocateNode* alloc,
218 Node* src, Node* src_offset,
219 Node* dest, Node* dest_offset,
220 Node* dest_size, bool dest_uninitialized);
221 void generate_slow_arraycopy(const TypePtr* adr_type,
222 Node* src, Node* src_offset,
223 Node* dest, Node* dest_offset,
224 Node* copy_length, bool dest_uninitialized);
225 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
226 Node* dest_elem_klass,
227 Node* src, Node* src_offset,
228 Node* dest, Node* dest_offset,
229 Node* copy_length, bool dest_uninitialized);
230 Node* generate_generic_arraycopy(const TypePtr* adr_type,
231 Node* src, Node* src_offset,
232 Node* dest, Node* dest_offset,
233 Node* copy_length, bool dest_uninitialized);
234 void generate_unchecked_arraycopy(const TypePtr* adr_type,
235 BasicType basic_elem_type,
236 bool disjoint_bases,
237 Node* src, Node* src_offset,
238 Node* dest, Node* dest_offset,
239 Node* copy_length, bool dest_uninitialized);
240 bool inline_unsafe_CAS(BasicType type);
241 bool inline_unsafe_ordered_store(BasicType type);
242 bool inline_fp_conversions(vmIntrinsics::ID id);
243 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
244 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
245 bool inline_bitCount(vmIntrinsics::ID id);
246 bool inline_reverseBytes(vmIntrinsics::ID id);
247
248 bool inline_reference_get();
249 };
250
251
252 //---------------------------make_vm_intrinsic----------------------------
253 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
254 vmIntrinsics::ID id = m->intrinsic_id();
255 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
256
257 if (DisableIntrinsic[0] != '\0'
258 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
259 // disabled by a user request on the command line:
260 // example: -XX:DisableIntrinsic=_hashCode,_getClass
261 return NULL;
262 }
263
264 if (!m->is_loaded()) {
265 // do not attempt to inline unloaded methods
266 return NULL;
267 }
268
269 // Only a few intrinsics implement a virtual dispatch.
270 // They are expensive calls which are also frequently overridden.
271 if (is_virtual) {
272 switch (id) {
273 case vmIntrinsics::_hashCode:
274 case vmIntrinsics::_clone:
275 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
276 break;
277 default:
278 return NULL;
279 }
280 }
281
282 // -XX:-InlineNatives disables nearly all intrinsics:
283 if (!InlineNatives) {
284 switch (id) {
285 case vmIntrinsics::_indexOf:
286 case vmIntrinsics::_compareTo:
287 case vmIntrinsics::_equals:
288 case vmIntrinsics::_equalsC:
289 break; // InlineNatives does not control String.compareTo
290 default:
291 return NULL;
292 }
293 }
294
295 switch (id) {
296 case vmIntrinsics::_compareTo:
297 if (!SpecialStringCompareTo) return NULL;
298 break;
299 case vmIntrinsics::_indexOf:
300 if (!SpecialStringIndexOf) return NULL;
301 break;
302 case vmIntrinsics::_equals:
303 if (!SpecialStringEquals) return NULL;
304 break;
305 case vmIntrinsics::_equalsC:
306 if (!SpecialArraysEquals) return NULL;
307 break;
308 case vmIntrinsics::_arraycopy:
309 if (!InlineArrayCopy) return NULL;
310 break;
311 case vmIntrinsics::_copyMemory:
312 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
313 if (!InlineArrayCopy) return NULL;
314 break;
315 case vmIntrinsics::_hashCode:
316 if (!InlineObjectHash) return NULL;
317 break;
318 case vmIntrinsics::_clone:
319 case vmIntrinsics::_copyOf:
320 case vmIntrinsics::_copyOfRange:
321 if (!InlineObjectCopy) return NULL;
322 // These also use the arraycopy intrinsic mechanism:
323 if (!InlineArrayCopy) return NULL;
324 break;
325 case vmIntrinsics::_checkIndex:
326 // We do not intrinsify this. The optimizer does fine with it.
327 return NULL;
328
329 case vmIntrinsics::_get_AtomicLong:
330 case vmIntrinsics::_attemptUpdate:
331 if (!InlineAtomicLong) return NULL;
332 break;
333
334 case vmIntrinsics::_getCallerClass:
335 if (!UseNewReflection) return NULL;
336 if (!InlineReflectionGetCallerClass) return NULL;
337 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
338 break;
339
340 case vmIntrinsics::_bitCount_i:
341 case vmIntrinsics::_bitCount_l:
342 if (!UsePopCountInstruction) return NULL;
343 break;
344
345 case vmIntrinsics::_Reference_get:
346 // It is only when G1 is enabled that we absolutely
347 // need to use the intrinsic version of Reference.get()
348 // so that the value in the referent field, if necessary,
349 // can be registered by the pre-barrier code.
350 if (!UseG1GC) return NULL;
351 break;
352
353 default:
354 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
355 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
356 break;
357 }
358
359 // -XX:-InlineClassNatives disables natives from the Class class.
360 // The flag applies to all reflective calls, notably Array.newArray
361 // (visible to Java programmers as Array.newInstance).
362 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
363 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
364 if (!InlineClassNatives) return NULL;
365 }
366
367 // -XX:-InlineThreadNatives disables natives from the Thread class.
368 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
369 if (!InlineThreadNatives) return NULL;
370 }
371
372 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
373 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
374 m->holder()->name() == ciSymbol::java_lang_Float() ||
375 m->holder()->name() == ciSymbol::java_lang_Double()) {
376 if (!InlineMathNatives) return NULL;
377 }
378
379 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
380 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
381 if (!InlineUnsafeOps) return NULL;
382 }
383
384 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
385 }
386
387 //----------------------register_library_intrinsics-----------------------
388 // Initialize this file's data structures, for each Compile instance.
389 void Compile::register_library_intrinsics() {
390 // Nothing to do here.
391 }
392
393 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
394 LibraryCallKit kit(jvms, this);
395 Compile* C = kit.C;
396 int nodes = C->unique();
397 #ifndef PRODUCT
398 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
399 char buf[1000];
400 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
401 tty->print_cr("Intrinsic %s", str);
402 }
403 #endif
404
405 if (kit.try_to_inline()) {
406 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
407 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
408 }
409 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
410 if (C->log()) {
411 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
412 vmIntrinsics::name_at(intrinsic_id()),
413 (is_virtual() ? " virtual='1'" : ""),
414 C->unique() - nodes);
415 }
416 return kit.transfer_exceptions_into_jvms();
417 }
418
419 if (PrintIntrinsics) {
420 if (jvms->has_method()) {
421 // Not a root compile.
422 tty->print("Did not inline intrinsic %s%s at bci:%d in",
423 vmIntrinsics::name_at(intrinsic_id()),
424 (is_virtual() ? " (virtual)" : ""), kit.bci());
425 kit.caller()->print_short_name(tty);
426 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
427 } else {
428 // Root compile
429 tty->print("Did not generate intrinsic %s%s at bci:%d in",
430 vmIntrinsics::name_at(intrinsic_id()),
431 (is_virtual() ? " (virtual)" : ""), kit.bci());
432 }
433 }
434 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
435 return NULL;
436 }
437
438 bool LibraryCallKit::try_to_inline() {
439 // Handle symbolic names for otherwise undistinguished boolean switches:
440 const bool is_store = true;
441 const bool is_native_ptr = true;
442 const bool is_static = true;
443
444 if (!jvms()->has_method()) {
445 // Root JVMState has a null method.
446 assert(map()->memory()->Opcode() == Op_Parm, "");
447 // Insert the memory aliasing node
448 set_all_memory(reset_memory());
449 }
450 assert(merged_memory(), "");
451
452 switch (intrinsic_id()) {
453 case vmIntrinsics::_hashCode:
454 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
455 case vmIntrinsics::_identityHashCode:
456 return inline_native_hashcode(/*!virtual*/ false, is_static);
457 case vmIntrinsics::_getClass:
458 return inline_native_getClass();
459
460 case vmIntrinsics::_dsin:
461 case vmIntrinsics::_dcos:
462 case vmIntrinsics::_dtan:
463 case vmIntrinsics::_dabs:
464 case vmIntrinsics::_datan2:
465 case vmIntrinsics::_dsqrt:
466 case vmIntrinsics::_dexp:
467 case vmIntrinsics::_dlog:
468 case vmIntrinsics::_dlog10:
469 case vmIntrinsics::_dpow:
470 return inline_math_native(intrinsic_id());
471
472 case vmIntrinsics::_min:
473 case vmIntrinsics::_max:
474 return inline_min_max(intrinsic_id());
475
476 case vmIntrinsics::_arraycopy:
477 return inline_arraycopy();
478
479 case vmIntrinsics::_compareTo:
480 return inline_string_compareTo();
481 case vmIntrinsics::_indexOf:
482 return inline_string_indexOf();
483 case vmIntrinsics::_equals:
484 return inline_string_equals();
485
486 case vmIntrinsics::_getObject:
487 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
488 case vmIntrinsics::_getBoolean:
489 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
490 case vmIntrinsics::_getByte:
491 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
492 case vmIntrinsics::_getShort:
493 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
494 case vmIntrinsics::_getChar:
495 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
496 case vmIntrinsics::_getInt:
497 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
498 case vmIntrinsics::_getLong:
499 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
500 case vmIntrinsics::_getFloat:
501 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
502 case vmIntrinsics::_getDouble:
503 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
504
505 case vmIntrinsics::_putObject:
506 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
507 case vmIntrinsics::_putBoolean:
508 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
509 case vmIntrinsics::_putByte:
510 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
511 case vmIntrinsics::_putShort:
512 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
513 case vmIntrinsics::_putChar:
514 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
515 case vmIntrinsics::_putInt:
516 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
517 case vmIntrinsics::_putLong:
518 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
519 case vmIntrinsics::_putFloat:
520 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
521 case vmIntrinsics::_putDouble:
522 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
523
524 case vmIntrinsics::_getByte_raw:
525 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
526 case vmIntrinsics::_getShort_raw:
527 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
528 case vmIntrinsics::_getChar_raw:
529 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
530 case vmIntrinsics::_getInt_raw:
531 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
532 case vmIntrinsics::_getLong_raw:
533 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
534 case vmIntrinsics::_getFloat_raw:
535 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
536 case vmIntrinsics::_getDouble_raw:
537 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
538 case vmIntrinsics::_getAddress_raw:
539 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
540
541 case vmIntrinsics::_putByte_raw:
542 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
543 case vmIntrinsics::_putShort_raw:
544 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
545 case vmIntrinsics::_putChar_raw:
546 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
547 case vmIntrinsics::_putInt_raw:
548 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
549 case vmIntrinsics::_putLong_raw:
550 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
551 case vmIntrinsics::_putFloat_raw:
552 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
553 case vmIntrinsics::_putDouble_raw:
554 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
555 case vmIntrinsics::_putAddress_raw:
556 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
557
558 case vmIntrinsics::_getObjectVolatile:
559 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
560 case vmIntrinsics::_getBooleanVolatile:
561 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
562 case vmIntrinsics::_getByteVolatile:
563 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
564 case vmIntrinsics::_getShortVolatile:
565 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
566 case vmIntrinsics::_getCharVolatile:
567 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
568 case vmIntrinsics::_getIntVolatile:
569 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
570 case vmIntrinsics::_getLongVolatile:
571 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
572 case vmIntrinsics::_getFloatVolatile:
573 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
574 case vmIntrinsics::_getDoubleVolatile:
575 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
576
577 case vmIntrinsics::_putObjectVolatile:
578 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
579 case vmIntrinsics::_putBooleanVolatile:
580 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
581 case vmIntrinsics::_putByteVolatile:
582 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
583 case vmIntrinsics::_putShortVolatile:
584 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
585 case vmIntrinsics::_putCharVolatile:
586 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
587 case vmIntrinsics::_putIntVolatile:
588 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
589 case vmIntrinsics::_putLongVolatile:
590 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
591 case vmIntrinsics::_putFloatVolatile:
592 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
593 case vmIntrinsics::_putDoubleVolatile:
594 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
595
596 case vmIntrinsics::_prefetchRead:
597 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
598 case vmIntrinsics::_prefetchWrite:
599 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
600 case vmIntrinsics::_prefetchReadStatic:
601 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
602 case vmIntrinsics::_prefetchWriteStatic:
603 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
604
605 case vmIntrinsics::_compareAndSwapObject:
606 return inline_unsafe_CAS(T_OBJECT);
607 case vmIntrinsics::_compareAndSwapInt:
608 return inline_unsafe_CAS(T_INT);
609 case vmIntrinsics::_compareAndSwapLong:
610 return inline_unsafe_CAS(T_LONG);
611
612 case vmIntrinsics::_putOrderedObject:
613 return inline_unsafe_ordered_store(T_OBJECT);
614 case vmIntrinsics::_putOrderedInt:
615 return inline_unsafe_ordered_store(T_INT);
616 case vmIntrinsics::_putOrderedLong:
617 return inline_unsafe_ordered_store(T_LONG);
618
619 case vmIntrinsics::_currentThread:
620 return inline_native_currentThread();
621 case vmIntrinsics::_isInterrupted:
622 return inline_native_isInterrupted();
623
624 case vmIntrinsics::_currentTimeMillis:
625 return inline_native_time_funcs(false);
626 case vmIntrinsics::_nanoTime:
627 return inline_native_time_funcs(true);
628 case vmIntrinsics::_allocateInstance:
629 return inline_unsafe_allocate();
630 case vmIntrinsics::_copyMemory:
631 return inline_unsafe_copyMemory();
632 case vmIntrinsics::_newArray:
633 return inline_native_newArray();
634 case vmIntrinsics::_getLength:
635 return inline_native_getLength();
636 case vmIntrinsics::_copyOf:
637 return inline_array_copyOf(false);
638 case vmIntrinsics::_copyOfRange:
639 return inline_array_copyOf(true);
640 case vmIntrinsics::_equalsC:
641 return inline_array_equals();
642 case vmIntrinsics::_clone:
643 return inline_native_clone(intrinsic()->is_virtual());
644
645 case vmIntrinsics::_isAssignableFrom:
646 return inline_native_subtype_check();
647
648 case vmIntrinsics::_isInstance:
649 case vmIntrinsics::_getModifiers:
650 case vmIntrinsics::_isInterface:
651 case vmIntrinsics::_isArray:
652 case vmIntrinsics::_isPrimitive:
653 case vmIntrinsics::_getSuperclass:
654 case vmIntrinsics::_getComponentType:
655 case vmIntrinsics::_getClassAccessFlags:
656 return inline_native_Class_query(intrinsic_id());
657
658 case vmIntrinsics::_floatToRawIntBits:
659 case vmIntrinsics::_floatToIntBits:
660 case vmIntrinsics::_intBitsToFloat:
661 case vmIntrinsics::_doubleToRawLongBits:
662 case vmIntrinsics::_doubleToLongBits:
663 case vmIntrinsics::_longBitsToDouble:
664 return inline_fp_conversions(intrinsic_id());
665
666 case vmIntrinsics::_numberOfLeadingZeros_i:
667 case vmIntrinsics::_numberOfLeadingZeros_l:
668 return inline_numberOfLeadingZeros(intrinsic_id());
669
670 case vmIntrinsics::_numberOfTrailingZeros_i:
671 case vmIntrinsics::_numberOfTrailingZeros_l:
672 return inline_numberOfTrailingZeros(intrinsic_id());
673
674 case vmIntrinsics::_bitCount_i:
675 case vmIntrinsics::_bitCount_l:
676 return inline_bitCount(intrinsic_id());
677
678 case vmIntrinsics::_reverseBytes_i:
679 case vmIntrinsics::_reverseBytes_l:
680 case vmIntrinsics::_reverseBytes_s:
681 case vmIntrinsics::_reverseBytes_c:
682 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
683
684 case vmIntrinsics::_get_AtomicLong:
685 return inline_native_AtomicLong_get();
686 case vmIntrinsics::_attemptUpdate:
687 return inline_native_AtomicLong_attemptUpdate();
688
689 case vmIntrinsics::_getCallerClass:
690 return inline_native_Reflection_getCallerClass();
691
692 case vmIntrinsics::_Reference_get:
693 return inline_reference_get();
694
695 default:
696 // If you get here, it may be that someone has added a new intrinsic
697 // to the list in vmSymbols.hpp without implementing it here.
698 #ifndef PRODUCT
699 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
700 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
701 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
702 }
703 #endif
704 return false;
705 }
706 }
707
708 //------------------------------push_result------------------------------
709 // Helper function for finishing intrinsics.
710 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
711 record_for_igvn(region);
712 set_control(_gvn.transform(region));
713 BasicType value_type = value->type()->basic_type();
714 push_node(value_type, _gvn.transform(value));
715 }
716
717 //------------------------------generate_guard---------------------------
718 // Helper function for generating guarded fast-slow graph structures.
719 // The given 'test', if true, guards a slow path. If the test fails
720 // then a fast path can be taken. (We generally hope it fails.)
721 // In all cases, GraphKit::control() is updated to the fast path.
722 // The returned value represents the control for the slow path.
723 // The return value is never 'top'; it is either a valid control
724 // or NULL if it is obvious that the slow path can never be taken.
725 // Also, if region and the slow control are not NULL, the slow edge
726 // is appended to the region.
727 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
728 if (stopped()) {
729 // Already short circuited.
730 return NULL;
731 }
732
733 // Build an if node and its projections.
734 // If test is true we take the slow path, which we assume is uncommon.
735 if (_gvn.type(test) == TypeInt::ZERO) {
736 // The slow branch is never taken. No need to build this guard.
737 return NULL;
738 }
739
740 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
741
742 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
743 if (if_slow == top()) {
744 // The slow branch is never taken. No need to build this guard.
745 return NULL;
746 }
747
748 if (region != NULL)
749 region->add_req(if_slow);
750
751 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
752 set_control(if_fast);
753
754 return if_slow;
755 }
756
757 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
758 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
759 }
760 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
761 return generate_guard(test, region, PROB_FAIR);
762 }
763
764 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
765 Node* *pos_index) {
766 if (stopped())
767 return NULL; // already stopped
768 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
769 return NULL; // index is already adequately typed
770 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
771 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
772 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
773 if (is_neg != NULL && pos_index != NULL) {
774 // Emulate effect of Parse::adjust_map_after_if.
775 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
776 ccast->set_req(0, control());
777 (*pos_index) = _gvn.transform(ccast);
778 }
779 return is_neg;
780 }
781
782 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
783 Node* *pos_index) {
784 if (stopped())
785 return NULL; // already stopped
786 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
787 return NULL; // index is already adequately typed
788 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
789 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
790 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
791 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
792 if (is_notp != NULL && pos_index != NULL) {
793 // Emulate effect of Parse::adjust_map_after_if.
794 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
795 ccast->set_req(0, control());
796 (*pos_index) = _gvn.transform(ccast);
797 }
798 return is_notp;
799 }
800
801 // Make sure that 'position' is a valid limit index, in [0..length].
802 // There are two equivalent plans for checking this:
803 // A. (offset + copyLength) unsigned<= arrayLength
804 // B. offset <= (arrayLength - copyLength)
805 // We require that all of the values above, except for the sum and
806 // difference, are already known to be non-negative.
807 // Plan A is robust in the face of overflow, if offset and copyLength
808 // are both hugely positive.
809 //
810 // Plan B is less direct and intuitive, but it does not overflow at
811 // all, since the difference of two non-negatives is always
812 // representable. Whenever Java methods must perform the equivalent
813 // check they generally use Plan B instead of Plan A.
814 // For the moment we use Plan A.
815 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
816 Node* subseq_length,
817 Node* array_length,
818 RegionNode* region) {
819 if (stopped())
820 return NULL; // already stopped
821 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
822 if (zero_offset && subseq_length->eqv_uncast(array_length))
823 return NULL; // common case of whole-array copy
824 Node* last = subseq_length;
825 if (!zero_offset) // last += offset
826 last = _gvn.transform( new (C, 3) AddINode(last, offset));
827 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
828 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
829 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
830 return is_over;
831 }
832
833
834 //--------------------------generate_current_thread--------------------
835 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
836 ciKlass* thread_klass = env()->Thread_klass();
837 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
838 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
839 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
840 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
841 tls_output = thread;
842 return threadObj;
843 }
844
845
846 //------------------------------make_string_method_node------------------------
847 // Helper method for String intrinsic finctions.
848 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
849 const int value_offset = java_lang_String::value_offset_in_bytes();
850 const int count_offset = java_lang_String::count_offset_in_bytes();
851 const int offset_offset = java_lang_String::offset_offset_in_bytes();
852
853 Node* no_ctrl = NULL;
854
855 ciInstanceKlass* klass = env()->String_klass();
856 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
857
858 const TypeAryPtr* value_type =
859 TypeAryPtr::make(TypePtr::NotNull,
860 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
861 ciTypeArrayKlass::make(T_CHAR), true, 0);
862
863 // Get start addr of string and substring
864 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
865 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
866 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
867 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
868 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
869
870 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
871 Node* str2_value = make_load(no_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
872 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
873 Node* str2_offset = make_load(no_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
874 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
875
876 Node* result = NULL;
877 switch (opcode) {
878 case Op_StrIndexOf:
879 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
880 str1_start, cnt1, str2_start, cnt2);
881 break;
882 case Op_StrComp:
883 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
884 str1_start, cnt1, str2_start, cnt2);
885 break;
886 case Op_StrEquals:
887 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
888 str1_start, str2_start, cnt1);
889 break;
890 default:
891 ShouldNotReachHere();
892 return NULL;
893 }
894
895 // All these intrinsics have checks.
896 C->set_has_split_ifs(true); // Has chance for split-if optimization
897
898 return _gvn.transform(result);
899 }
900
901 //------------------------------inline_string_compareTo------------------------
902 bool LibraryCallKit::inline_string_compareTo() {
903
904 if (!Matcher::has_match_rule(Op_StrComp)) return false;
905
906 const int value_offset = java_lang_String::value_offset_in_bytes();
907 const int count_offset = java_lang_String::count_offset_in_bytes();
908 const int offset_offset = java_lang_String::offset_offset_in_bytes();
909
910 _sp += 2;
911 Node *argument = pop(); // pop non-receiver first: it was pushed second
912 Node *receiver = pop();
913
914 // Null check on self without removing any arguments. The argument
915 // null check technically happens in the wrong place, which can lead to
916 // invalid stack traces when string compare is inlined into a method
917 // which handles NullPointerExceptions.
918 _sp += 2;
919 receiver = do_null_check(receiver, T_OBJECT);
920 argument = do_null_check(argument, T_OBJECT);
921 _sp -= 2;
922 if (stopped()) {
923 return true;
924 }
925
926 ciInstanceKlass* klass = env()->String_klass();
927 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
928 Node* no_ctrl = NULL;
929
930 // Get counts for string and argument
931 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
932 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
933
934 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
935 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
936
937 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
938 push(compare);
939 return true;
940 }
941
942 //------------------------------inline_string_equals------------------------
943 bool LibraryCallKit::inline_string_equals() {
944
945 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
946
947 const int value_offset = java_lang_String::value_offset_in_bytes();
948 const int count_offset = java_lang_String::count_offset_in_bytes();
949 const int offset_offset = java_lang_String::offset_offset_in_bytes();
950
951 int nargs = 2;
952 _sp += nargs;
953 Node* argument = pop(); // pop non-receiver first: it was pushed second
954 Node* receiver = pop();
955
956 // Null check on self without removing any arguments. The argument
957 // null check technically happens in the wrong place, which can lead to
958 // invalid stack traces when string compare is inlined into a method
959 // which handles NullPointerExceptions.
960 _sp += nargs;
961 receiver = do_null_check(receiver, T_OBJECT);
962 //should not do null check for argument for String.equals(), because spec
963 //allows to specify NULL as argument.
964 _sp -= nargs;
965
966 if (stopped()) {
967 return true;
968 }
969
970 // paths (plus control) merge
971 RegionNode* region = new (C, 5) RegionNode(5);
972 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
973
974 // does source == target string?
975 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
976 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
977
978 Node* if_eq = generate_slow_guard(bol, NULL);
979 if (if_eq != NULL) {
980 // receiver == argument
981 phi->init_req(2, intcon(1));
982 region->init_req(2, if_eq);
983 }
984
985 // get String klass for instanceOf
986 ciInstanceKlass* klass = env()->String_klass();
987
988 if (!stopped()) {
989 _sp += nargs; // gen_instanceof might do an uncommon trap
990 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
991 _sp -= nargs;
992 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
993 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
994
995 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
996 //instanceOf == true, fallthrough
997
998 if (inst_false != NULL) {
999 phi->init_req(3, intcon(0));
1000 region->init_req(3, inst_false);
1001 }
1002 }
1003
1004 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1005
1006 Node* no_ctrl = NULL;
1007 Node* receiver_cnt;
1008 Node* argument_cnt;
1009
1010 if (!stopped()) {
1011 // Properly cast the argument to String
1012 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
1013 // This path is taken only when argument's type is String:NotNull.
1014 argument = cast_not_null(argument, false);
1015
1016 // Get counts for string and argument
1017 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
1018 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1019
1020 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
1021 argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1022
1023 // Check for receiver count != argument count
1024 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
1025 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
1026 Node* if_ne = generate_slow_guard(bol, NULL);
1027 if (if_ne != NULL) {
1028 phi->init_req(4, intcon(0));
1029 region->init_req(4, if_ne);
1030 }
1031 }
1032
1033 // Check for count == 0 is done by mach node StrEquals.
1034
1035 if (!stopped()) {
1036 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1037 phi->init_req(1, equals);
1038 region->init_req(1, control());
1039 }
1040
1041 // post merge
1042 set_control(_gvn.transform(region));
1043 record_for_igvn(region);
1044
1045 push(_gvn.transform(phi));
1046
1047 return true;
1048 }
1049
1050 //------------------------------inline_array_equals----------------------------
1051 bool LibraryCallKit::inline_array_equals() {
1052
1053 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1054
1055 _sp += 2;
1056 Node *argument2 = pop();
1057 Node *argument1 = pop();
1058
1059 Node* equals =
1060 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1061 argument1, argument2) );
1062 push(equals);
1063 return true;
1064 }
1065
1066 // Java version of String.indexOf(constant string)
1067 // class StringDecl {
1068 // StringDecl(char[] ca) {
1069 // offset = 0;
1070 // count = ca.length;
1071 // value = ca;
1072 // }
1073 // int offset;
1074 // int count;
1075 // char[] value;
1076 // }
1077 //
1078 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1079 // int targetOffset, int cache_i, int md2) {
1080 // int cache = cache_i;
1081 // int sourceOffset = string_object.offset;
1082 // int sourceCount = string_object.count;
1083 // int targetCount = target_object.length;
1084 //
1085 // int targetCountLess1 = targetCount - 1;
1086 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1087 //
1088 // char[] source = string_object.value;
1089 // char[] target = target_object;
1090 // int lastChar = target[targetCountLess1];
1091 //
1092 // outer_loop:
1093 // for (int i = sourceOffset; i < sourceEnd; ) {
1094 // int src = source[i + targetCountLess1];
1095 // if (src == lastChar) {
1096 // // With random strings and a 4-character alphabet,
1097 // // reverse matching at this point sets up 0.8% fewer
1098 // // frames, but (paradoxically) makes 0.3% more probes.
1099 // // Since those probes are nearer the lastChar probe,
1100 // // there is may be a net D$ win with reverse matching.
1101 // // But, reversing loop inhibits unroll of inner loop
1102 // // for unknown reason. So, does running outer loop from
1103 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1104 // for (int j = 0; j < targetCountLess1; j++) {
1105 // if (target[targetOffset + j] != source[i+j]) {
1106 // if ((cache & (1 << source[i+j])) == 0) {
1107 // if (md2 < j+1) {
1108 // i += j+1;
1109 // continue outer_loop;
1110 // }
1111 // }
1112 // i += md2;
1113 // continue outer_loop;
1114 // }
1115 // }
1116 // return i - sourceOffset;
1117 // }
1118 // if ((cache & (1 << src)) == 0) {
1119 // i += targetCountLess1;
1120 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1121 // i++;
1122 // }
1123 // return -1;
1124 // }
1125
1126 //------------------------------string_indexOf------------------------
1127 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1128 jint cache_i, jint md2_i) {
1129
1130 Node* no_ctrl = NULL;
1131 float likely = PROB_LIKELY(0.9);
1132 float unlikely = PROB_UNLIKELY(0.9);
1133
1134 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1135
1136 const int value_offset = java_lang_String::value_offset_in_bytes();
1137 const int count_offset = java_lang_String::count_offset_in_bytes();
1138 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1139
1140 ciInstanceKlass* klass = env()->String_klass();
1141 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1142 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1143
1144 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1145 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1146 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1147 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1148 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1149 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1150
1151 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1152 jint target_length = target_array->length();
1153 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1154 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1155
1156 IdealKit kit(this, false, true);
1157 #define __ kit.
1158 Node* zero = __ ConI(0);
1159 Node* one = __ ConI(1);
1160 Node* cache = __ ConI(cache_i);
1161 Node* md2 = __ ConI(md2_i);
1162 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1163 Node* targetCount = __ ConI(target_length);
1164 Node* targetCountLess1 = __ ConI(target_length - 1);
1165 Node* targetOffset = __ ConI(targetOffset_i);
1166 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1167
1168 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1169 Node* outer_loop = __ make_label(2 /* goto */);
1170 Node* return_ = __ make_label(1);
1171
1172 __ set(rtn,__ ConI(-1));
1173 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1174 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1175 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1176 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1177 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1178 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1179 Node* tpj = __ AddI(targetOffset, __ value(j));
1180 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1181 Node* ipj = __ AddI(__ value(i), __ value(j));
1182 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1183 __ if_then(targ, BoolTest::ne, src2); {
1184 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1185 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1186 __ increment(i, __ AddI(__ value(j), one));
1187 __ goto_(outer_loop);
1188 } __ end_if(); __ dead(j);
1189 }__ end_if(); __ dead(j);
1190 __ increment(i, md2);
1191 __ goto_(outer_loop);
1192 }__ end_if();
1193 __ increment(j, one);
1194 }__ end_loop(); __ dead(j);
1195 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1196 __ goto_(return_);
1197 }__ end_if();
1198 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1199 __ increment(i, targetCountLess1);
1200 }__ end_if();
1201 __ increment(i, one);
1202 __ bind(outer_loop);
1203 }__ end_loop(); __ dead(i);
1204 __ bind(return_);
1205
1206 // Final sync IdealKit and GraphKit.
1207 final_sync(kit);
1208 Node* result = __ value(rtn);
1209 #undef __
1210 C->set_has_loops(true);
1211 return result;
1212 }
1213
1214 //------------------------------inline_string_indexOf------------------------
1215 bool LibraryCallKit::inline_string_indexOf() {
1216
1217 const int value_offset = java_lang_String::value_offset_in_bytes();
1218 const int count_offset = java_lang_String::count_offset_in_bytes();
1219 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1220
1221 _sp += 2;
1222 Node *argument = pop(); // pop non-receiver first: it was pushed second
1223 Node *receiver = pop();
1224
1225 Node* result;
1226 // Disable the use of pcmpestri until it can be guaranteed that
1227 // the load doesn't cross into the uncommited space.
1228 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1229 UseSSE42Intrinsics) {
1230 // Generate SSE4.2 version of indexOf
1231 // We currently only have match rules that use SSE4.2
1232
1233 // Null check on self without removing any arguments. The argument
1234 // null check technically happens in the wrong place, which can lead to
1235 // invalid stack traces when string compare is inlined into a method
1236 // which handles NullPointerExceptions.
1237 _sp += 2;
1238 receiver = do_null_check(receiver, T_OBJECT);
1239 argument = do_null_check(argument, T_OBJECT);
1240 _sp -= 2;
1241
1242 if (stopped()) {
1243 return true;
1244 }
1245
1246 ciInstanceKlass* str_klass = env()->String_klass();
1247 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1248
1249 // Make the merge point
1250 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1251 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1252 Node* no_ctrl = NULL;
1253
1254 // Get counts for string and substr
1255 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1256 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1257
1258 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1259 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1260
1261 // Check for substr count > string count
1262 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1263 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1264 Node* if_gt = generate_slow_guard(bol, NULL);
1265 if (if_gt != NULL) {
1266 result_phi->init_req(2, intcon(-1));
1267 result_rgn->init_req(2, if_gt);
1268 }
1269
1270 if (!stopped()) {
1271 // Check for substr count == 0
1272 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1273 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1274 Node* if_zero = generate_slow_guard(bol, NULL);
1275 if (if_zero != NULL) {
1276 result_phi->init_req(3, intcon(0));
1277 result_rgn->init_req(3, if_zero);
1278 }
1279 }
1280
1281 if (!stopped()) {
1282 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1283 result_phi->init_req(1, result);
1284 result_rgn->init_req(1, control());
1285 }
1286 set_control(_gvn.transform(result_rgn));
1287 record_for_igvn(result_rgn);
1288 result = _gvn.transform(result_phi);
1289
1290 } else { // Use LibraryCallKit::string_indexOf
1291 // don't intrinsify if argument isn't a constant string.
1292 if (!argument->is_Con()) {
1293 return false;
1294 }
1295 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1296 if (str_type == NULL) {
1297 return false;
1298 }
1299 ciInstanceKlass* klass = env()->String_klass();
1300 ciObject* str_const = str_type->const_oop();
1301 if (str_const == NULL || str_const->klass() != klass) {
1302 return false;
1303 }
1304 ciInstance* str = str_const->as_instance();
1305 assert(str != NULL, "must be instance");
1306
1307 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1308 int o = str->field_value_by_offset(offset_offset).as_int();
1309 int c = str->field_value_by_offset(count_offset).as_int();
1310 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1311
1312 // constant strings have no offset and count == length which
1313 // simplifies the resulting code somewhat so lets optimize for that.
1314 if (o != 0 || c != pat->length()) {
1315 return false;
1316 }
1317
1318 // Null check on self without removing any arguments. The argument
1319 // null check technically happens in the wrong place, which can lead to
1320 // invalid stack traces when string compare is inlined into a method
1321 // which handles NullPointerExceptions.
1322 _sp += 2;
1323 receiver = do_null_check(receiver, T_OBJECT);
1324 // No null check on the argument is needed since it's a constant String oop.
1325 _sp -= 2;
1326 if (stopped()) {
1327 return true;
1328 }
1329
1330 // The null string as a pattern always returns 0 (match at beginning of string)
1331 if (c == 0) {
1332 push(intcon(0));
1333 return true;
1334 }
1335
1336 // Generate default indexOf
1337 jchar lastChar = pat->char_at(o + (c - 1));
1338 int cache = 0;
1339 int i;
1340 for (i = 0; i < c - 1; i++) {
1341 assert(i < pat->length(), "out of range");
1342 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1343 }
1344
1345 int md2 = c;
1346 for (i = 0; i < c - 1; i++) {
1347 assert(i < pat->length(), "out of range");
1348 if (pat->char_at(o + i) == lastChar) {
1349 md2 = (c - 1) - i;
1350 }
1351 }
1352
1353 result = string_indexOf(receiver, pat, o, cache, md2);
1354 }
1355
1356 push(result);
1357 return true;
1358 }
1359
1360 //--------------------------pop_math_arg--------------------------------
1361 // Pop a double argument to a math function from the stack
1362 // rounding it if necessary.
1363 Node * LibraryCallKit::pop_math_arg() {
1364 Node *arg = pop_pair();
1365 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1366 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1367 return arg;
1368 }
1369
1370 //------------------------------inline_trig----------------------------------
1371 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1372 // argument reduction which will turn into a fast/slow diamond.
1373 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1374 _sp += arg_size(); // restore stack pointer
1375 Node* arg = pop_math_arg();
1376 Node* trig = NULL;
1377
1378 switch (id) {
1379 case vmIntrinsics::_dsin:
1380 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1381 break;
1382 case vmIntrinsics::_dcos:
1383 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1384 break;
1385 case vmIntrinsics::_dtan:
1386 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1387 break;
1388 default:
1389 assert(false, "bad intrinsic was passed in");
1390 return false;
1391 }
1392
1393 // Rounding required? Check for argument reduction!
1394 if( Matcher::strict_fp_requires_explicit_rounding ) {
1395
1396 static const double pi_4 = 0.7853981633974483;
1397 static const double neg_pi_4 = -0.7853981633974483;
1398 // pi/2 in 80-bit extended precision
1399 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1400 // -pi/2 in 80-bit extended precision
1401 // 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};
1402 // Cutoff value for using this argument reduction technique
1403 //static const double pi_2_minus_epsilon = 1.564660403643354;
1404 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1405
1406 // Pseudocode for sin:
1407 // if (x <= Math.PI / 4.0) {
1408 // if (x >= -Math.PI / 4.0) return fsin(x);
1409 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1410 // } else {
1411 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1412 // }
1413 // return StrictMath.sin(x);
1414
1415 // Pseudocode for cos:
1416 // if (x <= Math.PI / 4.0) {
1417 // if (x >= -Math.PI / 4.0) return fcos(x);
1418 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1419 // } else {
1420 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1421 // }
1422 // return StrictMath.cos(x);
1423
1424 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1425 // requires a special machine instruction to load it. Instead we'll try
1426 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1427 // probably do the math inside the SIN encoding.
1428
1429 // Make the merge point
1430 RegionNode *r = new (C, 3) RegionNode(3);
1431 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1432
1433 // Flatten arg so we need only 1 test
1434 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1435 // Node for PI/4 constant
1436 Node *pi4 = makecon(TypeD::make(pi_4));
1437 // Check PI/4 : abs(arg)
1438 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1439 // Check: If PI/4 < abs(arg) then go slow
1440 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1441 // Branch either way
1442 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1443 set_control(opt_iff(r,iff));
1444
1445 // Set fast path result
1446 phi->init_req(2,trig);
1447
1448 // Slow path - non-blocking leaf call
1449 Node* call = NULL;
1450 switch (id) {
1451 case vmIntrinsics::_dsin:
1452 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1453 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1454 "Sin", NULL, arg, top());
1455 break;
1456 case vmIntrinsics::_dcos:
1457 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1458 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1459 "Cos", NULL, arg, top());
1460 break;
1461 case vmIntrinsics::_dtan:
1462 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1463 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1464 "Tan", NULL, arg, top());
1465 break;
1466 }
1467 assert(control()->in(0) == call, "");
1468 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1469 r->init_req(1,control());
1470 phi->init_req(1,slow_result);
1471
1472 // Post-merge
1473 set_control(_gvn.transform(r));
1474 record_for_igvn(r);
1475 trig = _gvn.transform(phi);
1476
1477 C->set_has_split_ifs(true); // Has chance for split-if optimization
1478 }
1479 // Push result back on JVM stack
1480 push_pair(trig);
1481 return true;
1482 }
1483
1484 //------------------------------inline_sqrt-------------------------------------
1485 // Inline square root instruction, if possible.
1486 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1487 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1488 _sp += arg_size(); // restore stack pointer
1489 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1490 return true;
1491 }
1492
1493 //------------------------------inline_abs-------------------------------------
1494 // Inline absolute value instruction, if possible.
1495 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1496 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1497 _sp += arg_size(); // restore stack pointer
1498 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1499 return true;
1500 }
1501
1502 //------------------------------inline_exp-------------------------------------
1503 // Inline exp instructions, if possible. The Intel hardware only misses
1504 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1505 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1506 assert(id == vmIntrinsics::_dexp, "Not exp");
1507
1508 // If this inlining ever returned NaN in the past, we do not intrinsify it
1509 // every again. NaN results requires StrictMath.exp handling.
1510 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1511
1512 // Do not intrinsify on older platforms which lack cmove.
1513 if (ConditionalMoveLimit == 0) return false;
1514
1515 _sp += arg_size(); // restore stack pointer
1516 Node *x = pop_math_arg();
1517 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1518
1519 //-------------------
1520 //result=(result.isNaN())? StrictMath::exp():result;
1521 // Check: If isNaN() by checking result!=result? then go to Strict Math
1522 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1523 // Build the boolean node
1524 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1525
1526 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1527 // End the current control-flow path
1528 push_pair(x);
1529 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1530 // to handle. Recompile without intrinsifying Math.exp
1531 uncommon_trap(Deoptimization::Reason_intrinsic,
1532 Deoptimization::Action_make_not_entrant);
1533 }
1534
1535 C->set_has_split_ifs(true); // Has chance for split-if optimization
1536
1537 push_pair(result);
1538
1539 return true;
1540 }
1541
1542 //------------------------------inline_pow-------------------------------------
1543 // Inline power instructions, if possible.
1544 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1545 assert(id == vmIntrinsics::_dpow, "Not pow");
1546
1547 // If this inlining ever returned NaN in the past, we do not intrinsify it
1548 // every again. NaN results requires StrictMath.pow handling.
1549 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1550
1551 // Do not intrinsify on older platforms which lack cmove.
1552 if (ConditionalMoveLimit == 0) return false;
1553
1554 // Pseudocode for pow
1555 // if (x <= 0.0) {
1556 // if ((double)((int)y)==y) { // if y is int
1557 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1558 // } else {
1559 // result = NaN;
1560 // }
1561 // } else {
1562 // result = DPow(x,y);
1563 // }
1564 // if (result != result)? {
1565 // uncommon_trap();
1566 // }
1567 // return result;
1568
1569 _sp += arg_size(); // restore stack pointer
1570 Node* y = pop_math_arg();
1571 Node* x = pop_math_arg();
1572
1573 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1574
1575 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1576 // inside of something) then skip the fancy tests and just check for
1577 // NaN result.
1578 Node *result = NULL;
1579 if( jvms()->depth() >= 1 ) {
1580 result = fast_result;
1581 } else {
1582
1583 // Set the merge point for If node with condition of (x <= 0.0)
1584 // There are four possible paths to region node and phi node
1585 RegionNode *r = new (C, 4) RegionNode(4);
1586 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1587
1588 // Build the first if node: if (x <= 0.0)
1589 // Node for 0 constant
1590 Node *zeronode = makecon(TypeD::ZERO);
1591 // Check x:0
1592 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1593 // Check: If (x<=0) then go complex path
1594 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1595 // Branch either way
1596 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1597 Node *opt_test = _gvn.transform(if1);
1598 //assert( opt_test->is_If(), "Expect an IfNode");
1599 IfNode *opt_if1 = (IfNode*)opt_test;
1600 // Fast path taken; set region slot 3
1601 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1602 r->init_req(3,fast_taken); // Capture fast-control
1603
1604 // Fast path not-taken, i.e. slow path
1605 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1606
1607 // Set fast path result
1608 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1609 phi->init_req(3, fast_result);
1610
1611 // Complex path
1612 // Build the second if node (if y is int)
1613 // Node for (int)y
1614 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1615 // Node for (double)((int) y)
1616 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1617 // Check (double)((int) y) : y
1618 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1619 // Check if (y isn't int) then go to slow path
1620
1621 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1622 // Branch either way
1623 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1624 Node *slow_path = opt_iff(r,if2); // Set region path 2
1625
1626 // Calculate DPow(abs(x), y)*(1 & (int)y)
1627 // Node for constant 1
1628 Node *conone = intcon(1);
1629 // 1& (int)y
1630 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1631 // zero node
1632 Node *conzero = intcon(0);
1633 // Check (1&(int)y)==0?
1634 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1635 // Check if (1&(int)y)!=0?, if so the result is negative
1636 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1637 // abs(x)
1638 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1639 // abs(x)^y
1640 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1641 // -abs(x)^y
1642 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1643 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1644 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1645 // Set complex path fast result
1646 phi->init_req(2, signresult);
1647
1648 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1649 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1650 r->init_req(1,slow_path);
1651 phi->init_req(1,slow_result);
1652
1653 // Post merge
1654 set_control(_gvn.transform(r));
1655 record_for_igvn(r);
1656 result=_gvn.transform(phi);
1657 }
1658
1659 //-------------------
1660 //result=(result.isNaN())? uncommon_trap():result;
1661 // Check: If isNaN() by checking result!=result? then go to Strict Math
1662 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1663 // Build the boolean node
1664 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1665
1666 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1667 // End the current control-flow path
1668 push_pair(x);
1669 push_pair(y);
1670 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1671 // to handle. Recompile without intrinsifying Math.pow.
1672 uncommon_trap(Deoptimization::Reason_intrinsic,
1673 Deoptimization::Action_make_not_entrant);
1674 }
1675
1676 C->set_has_split_ifs(true); // Has chance for split-if optimization
1677
1678 push_pair(result);
1679
1680 return true;
1681 }
1682
1683 //------------------------------inline_trans-------------------------------------
1684 // Inline transcendental instructions, if possible. The Intel hardware gets
1685 // these right, no funny corner cases missed.
1686 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1687 _sp += arg_size(); // restore stack pointer
1688 Node* arg = pop_math_arg();
1689 Node* trans = NULL;
1690
1691 switch (id) {
1692 case vmIntrinsics::_dlog:
1693 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1694 break;
1695 case vmIntrinsics::_dlog10:
1696 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1697 break;
1698 default:
1699 assert(false, "bad intrinsic was passed in");
1700 return false;
1701 }
1702
1703 // Push result back on JVM stack
1704 push_pair(trans);
1705 return true;
1706 }
1707
1708 //------------------------------runtime_math-----------------------------
1709 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1710 Node* a = NULL;
1711 Node* b = NULL;
1712
1713 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1714 "must be (DD)D or (D)D type");
1715
1716 // Inputs
1717 _sp += arg_size(); // restore stack pointer
1718 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1719 b = pop_math_arg();
1720 }
1721 a = pop_math_arg();
1722
1723 const TypePtr* no_memory_effects = NULL;
1724 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1725 no_memory_effects,
1726 a, top(), b, b ? top() : NULL);
1727 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1728 #ifdef ASSERT
1729 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1730 assert(value_top == top(), "second value must be top");
1731 #endif
1732
1733 push_pair(value);
1734 return true;
1735 }
1736
1737 //------------------------------inline_math_native-----------------------------
1738 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1739 switch (id) {
1740 // These intrinsics are not properly supported on all hardware
1741 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1742 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1743 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1744 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1745 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1746 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1747
1748 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1749 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1750 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1751 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1752
1753 // These intrinsics are supported on all hardware
1754 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1755 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1756
1757 // These intrinsics don't work on X86. The ad implementation doesn't
1758 // handle NaN's properly. Instead of returning infinity, the ad
1759 // implementation returns a NaN on overflow. See bug: 6304089
1760 // Once the ad implementations are fixed, change the code below
1761 // to match the intrinsics above
1762
1763 case vmIntrinsics::_dexp: return
1764 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1765 case vmIntrinsics::_dpow: return
1766 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1767
1768 // These intrinsics are not yet correctly implemented
1769 case vmIntrinsics::_datan2:
1770 return false;
1771
1772 default:
1773 ShouldNotReachHere();
1774 return false;
1775 }
1776 }
1777
1778 static bool is_simple_name(Node* n) {
1779 return (n->req() == 1 // constant
1780 || (n->is_Type() && n->as_Type()->type()->singleton())
1781 || n->is_Proj() // parameter or return value
1782 || n->is_Phi() // local of some sort
1783 );
1784 }
1785
1786 //----------------------------inline_min_max-----------------------------------
1787 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1788 push(generate_min_max(id, argument(0), argument(1)));
1789
1790 return true;
1791 }
1792
1793 Node*
1794 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1795 // These are the candidate return value:
1796 Node* xvalue = x0;
1797 Node* yvalue = y0;
1798
1799 if (xvalue == yvalue) {
1800 return xvalue;
1801 }
1802
1803 bool want_max = (id == vmIntrinsics::_max);
1804
1805 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1806 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1807 if (txvalue == NULL || tyvalue == NULL) return top();
1808 // This is not really necessary, but it is consistent with a
1809 // hypothetical MaxINode::Value method:
1810 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1811
1812 // %%% This folding logic should (ideally) be in a different place.
1813 // Some should be inside IfNode, and there to be a more reliable
1814 // transformation of ?: style patterns into cmoves. We also want
1815 // more powerful optimizations around cmove and min/max.
1816
1817 // Try to find a dominating comparison of these guys.
1818 // It can simplify the index computation for Arrays.copyOf
1819 // and similar uses of System.arraycopy.
1820 // First, compute the normalized version of CmpI(x, y).
1821 int cmp_op = Op_CmpI;
1822 Node* xkey = xvalue;
1823 Node* ykey = yvalue;
1824 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1825 if (ideal_cmpxy->is_Cmp()) {
1826 // E.g., if we have CmpI(length - offset, count),
1827 // it might idealize to CmpI(length, count + offset)
1828 cmp_op = ideal_cmpxy->Opcode();
1829 xkey = ideal_cmpxy->in(1);
1830 ykey = ideal_cmpxy->in(2);
1831 }
1832
1833 // Start by locating any relevant comparisons.
1834 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1835 Node* cmpxy = NULL;
1836 Node* cmpyx = NULL;
1837 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1838 Node* cmp = start_from->fast_out(k);
1839 if (cmp->outcnt() > 0 && // must have prior uses
1840 cmp->in(0) == NULL && // must be context-independent
1841 cmp->Opcode() == cmp_op) { // right kind of compare
1842 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1843 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1844 }
1845 }
1846
1847 const int NCMPS = 2;
1848 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1849 int cmpn;
1850 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1851 if (cmps[cmpn] != NULL) break; // find a result
1852 }
1853 if (cmpn < NCMPS) {
1854 // Look for a dominating test that tells us the min and max.
1855 int depth = 0; // Limit search depth for speed
1856 Node* dom = control();
1857 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1858 if (++depth >= 100) break;
1859 Node* ifproj = dom;
1860 if (!ifproj->is_Proj()) continue;
1861 Node* iff = ifproj->in(0);
1862 if (!iff->is_If()) continue;
1863 Node* bol = iff->in(1);
1864 if (!bol->is_Bool()) continue;
1865 Node* cmp = bol->in(1);
1866 if (cmp == NULL) continue;
1867 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1868 if (cmps[cmpn] == cmp) break;
1869 if (cmpn == NCMPS) continue;
1870 BoolTest::mask btest = bol->as_Bool()->_test._test;
1871 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1872 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1873 // At this point, we know that 'x btest y' is true.
1874 switch (btest) {
1875 case BoolTest::eq:
1876 // They are proven equal, so we can collapse the min/max.
1877 // Either value is the answer. Choose the simpler.
1878 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1879 return yvalue;
1880 return xvalue;
1881 case BoolTest::lt: // x < y
1882 case BoolTest::le: // x <= y
1883 return (want_max ? yvalue : xvalue);
1884 case BoolTest::gt: // x > y
1885 case BoolTest::ge: // x >= y
1886 return (want_max ? xvalue : yvalue);
1887 }
1888 }
1889 }
1890
1891 // We failed to find a dominating test.
1892 // Let's pick a test that might GVN with prior tests.
1893 Node* best_bol = NULL;
1894 BoolTest::mask best_btest = BoolTest::illegal;
1895 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1896 Node* cmp = cmps[cmpn];
1897 if (cmp == NULL) continue;
1898 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1899 Node* bol = cmp->fast_out(j);
1900 if (!bol->is_Bool()) continue;
1901 BoolTest::mask btest = bol->as_Bool()->_test._test;
1902 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1903 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1904 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1905 best_bol = bol->as_Bool();
1906 best_btest = btest;
1907 }
1908 }
1909 }
1910
1911 Node* answer_if_true = NULL;
1912 Node* answer_if_false = NULL;
1913 switch (best_btest) {
1914 default:
1915 if (cmpxy == NULL)
1916 cmpxy = ideal_cmpxy;
1917 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1918 // and fall through:
1919 case BoolTest::lt: // x < y
1920 case BoolTest::le: // x <= y
1921 answer_if_true = (want_max ? yvalue : xvalue);
1922 answer_if_false = (want_max ? xvalue : yvalue);
1923 break;
1924 case BoolTest::gt: // x > y
1925 case BoolTest::ge: // x >= y
1926 answer_if_true = (want_max ? xvalue : yvalue);
1927 answer_if_false = (want_max ? yvalue : xvalue);
1928 break;
1929 }
1930
1931 jint hi, lo;
1932 if (want_max) {
1933 // We can sharpen the minimum.
1934 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1935 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1936 } else {
1937 // We can sharpen the maximum.
1938 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1939 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1940 }
1941
1942 // Use a flow-free graph structure, to avoid creating excess control edges
1943 // which could hinder other optimizations.
1944 // Since Math.min/max is often used with arraycopy, we want
1945 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1946 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1947 answer_if_false, answer_if_true,
1948 TypeInt::make(lo, hi, widen));
1949
1950 return _gvn.transform(cmov);
1951
1952 /*
1953 // This is not as desirable as it may seem, since Min and Max
1954 // nodes do not have a full set of optimizations.
1955 // And they would interfere, anyway, with 'if' optimizations
1956 // and with CMoveI canonical forms.
1957 switch (id) {
1958 case vmIntrinsics::_min:
1959 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1960 case vmIntrinsics::_max:
1961 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1962 default:
1963 ShouldNotReachHere();
1964 }
1965 */
1966 }
1967
1968 inline int
1969 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1970 const TypePtr* base_type = TypePtr::NULL_PTR;
1971 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1972 if (base_type == NULL) {
1973 // Unknown type.
1974 return Type::AnyPtr;
1975 } else if (base_type == TypePtr::NULL_PTR) {
1976 // Since this is a NULL+long form, we have to switch to a rawptr.
1977 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1978 offset = MakeConX(0);
1979 return Type::RawPtr;
1980 } else if (base_type->base() == Type::RawPtr) {
1981 return Type::RawPtr;
1982 } else if (base_type->isa_oopptr()) {
1983 // Base is never null => always a heap address.
1984 if (base_type->ptr() == TypePtr::NotNull) {
1985 return Type::OopPtr;
1986 }
1987 // Offset is small => always a heap address.
1988 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1989 if (offset_type != NULL &&
1990 base_type->offset() == 0 && // (should always be?)
1991 offset_type->_lo >= 0 &&
1992 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1993 return Type::OopPtr;
1994 }
1995 // Otherwise, it might either be oop+off or NULL+addr.
1996 return Type::AnyPtr;
1997 } else {
1998 // No information:
1999 return Type::AnyPtr;
2000 }
2001 }
2002
2003 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2004 int kind = classify_unsafe_addr(base, offset);
2005 if (kind == Type::RawPtr) {
2006 return basic_plus_adr(top(), base, offset);
2007 } else {
2008 return basic_plus_adr(base, offset);
2009 }
2010 }
2011
2012 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2013 // inline int Integer.numberOfLeadingZeros(int)
2014 // inline int Long.numberOfLeadingZeros(long)
2015 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2016 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2017 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2018 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2019 _sp += arg_size(); // restore stack pointer
2020 switch (id) {
2021 case vmIntrinsics::_numberOfLeadingZeros_i:
2022 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
2023 break;
2024 case vmIntrinsics::_numberOfLeadingZeros_l:
2025 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
2026 break;
2027 default:
2028 ShouldNotReachHere();
2029 }
2030 return true;
2031 }
2032
2033 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2034 // inline int Integer.numberOfTrailingZeros(int)
2035 // inline int Long.numberOfTrailingZeros(long)
2036 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2037 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2038 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2039 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2040 _sp += arg_size(); // restore stack pointer
2041 switch (id) {
2042 case vmIntrinsics::_numberOfTrailingZeros_i:
2043 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2044 break;
2045 case vmIntrinsics::_numberOfTrailingZeros_l:
2046 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2047 break;
2048 default:
2049 ShouldNotReachHere();
2050 }
2051 return true;
2052 }
2053
2054 //----------------------------inline_bitCount_int/long-----------------------
2055 // inline int Integer.bitCount(int)
2056 // inline int Long.bitCount(long)
2057 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2058 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2059 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2060 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2061 _sp += arg_size(); // restore stack pointer
2062 switch (id) {
2063 case vmIntrinsics::_bitCount_i:
2064 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2065 break;
2066 case vmIntrinsics::_bitCount_l:
2067 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2068 break;
2069 default:
2070 ShouldNotReachHere();
2071 }
2072 return true;
2073 }
2074
2075 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2076 // inline Integer.reverseBytes(int)
2077 // inline Long.reverseBytes(long)
2078 // inline Character.reverseBytes(char)
2079 // inline Short.reverseBytes(short)
2080 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2081 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2082 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2083 "not reverse Bytes");
2084 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2085 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2086 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2087 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2088 _sp += arg_size(); // restore stack pointer
2089 switch (id) {
2090 case vmIntrinsics::_reverseBytes_i:
2091 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2092 break;
2093 case vmIntrinsics::_reverseBytes_l:
2094 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2095 break;
2096 case vmIntrinsics::_reverseBytes_c:
2097 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2098 break;
2099 case vmIntrinsics::_reverseBytes_s:
2100 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2101 break;
2102 default:
2103 ;
2104 }
2105 return true;
2106 }
2107
2108 //----------------------------inline_unsafe_access----------------------------
2109
2110 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2111
2112 // Helper that guards and inserts a G1 pre-barrier.
2113 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) {
2114 assert(UseG1GC, "should not call this otherwise");
2115
2116 // We could be accessing the referent field of a reference object. If so, when G1
2117 // is enabled, we need to log the value in the referent field in an SATB buffer.
2118 // This routine performs some compile time filters and generates suitable
2119 // runtime filters that guard the pre-barrier code.
2120
2121 // Some compile time checks.
2122
2123 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2124 const TypeX* otype = offset->find_intptr_t_type();
2125 if (otype != NULL && otype->is_con() &&
2126 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2127 // Constant offset but not the reference_offset so just return
2128 return;
2129 }
2130
2131 // We only need to generate the runtime guards for instances.
2132 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2133 if (btype != NULL) {
2134 if (btype->isa_aryptr()) {
2135 // Array type so nothing to do
2136 return;
2137 }
2138
2139 const TypeInstPtr* itype = btype->isa_instptr();
2140 if (itype != NULL) {
2141 // Can the klass of base_oop be statically determined
2142 // to be _not_ a sub-class of Reference?
2143 ciKlass* klass = itype->klass();
2144 if (klass->is_subtype_of(env()->Reference_klass()) &&
2145 !env()->Reference_klass()->is_subtype_of(klass)) {
2146 return;
2147 }
2148 }
2149 }
2150
2151 // The compile time filters did not reject base_oop/offset so
2152 // we need to generate the following runtime filters
2153 //
2154 // if (offset == java_lang_ref_Reference::_reference_offset) {
2155 // if (base != null) {
2156 // if (klass(base)->reference_type() != REF_NONE)) {
2157 // pre_barrier(_, pre_val, ...);
2158 // }
2159 // }
2160 // }
2161
2162 float likely = PROB_LIKELY(0.999);
2163 float unlikely = PROB_UNLIKELY(0.999);
2164
2165 IdealKit ideal(this);
2166 #define __ ideal.
2167
2168 const int reference_type_offset = in_bytes(instanceKlass::reference_type_offset());
2169
2170 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2171
2172 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2173 __ if_then(base_oop, BoolTest::ne, null(), likely); {
2174
2175 // Update graphKit memory and control from IdealKit.
2176 sync_kit(ideal);
2177
2178 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2179 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2180
2181 // Update IdealKit memory and control from graphKit.
2182 __ sync_kit(this);
2183
2184 Node* one = __ ConI(1);
2185
2186 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2187
2188 // Update graphKit from IdeakKit.
2189 sync_kit(ideal);
2190
2191 // Use the pre-barrier to record the value in the referent field
2192 pre_barrier(false /* do_load */,
2193 __ ctrl(),
2194 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2195 pre_val /* pre_val */,
2196 T_OBJECT);
2197
2198 // Update IdealKit from graphKit.
2199 __ sync_kit(this);
2200
2201 } __ end_if(); // _ref_type != ref_none
2202 } __ end_if(); // base != NULL
2203 } __ end_if(); // offset == referent_offset
2204
2205 // Final sync IdealKit and GraphKit.
2206 final_sync(ideal);
2207 #undef __
2208 }
2209
2210
2211 // Interpret Unsafe.fieldOffset cookies correctly:
2212 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2213
2214 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2215 if (callee()->is_static()) return false; // caller must have the capability!
2216
2217 #ifndef PRODUCT
2218 {
2219 ResourceMark rm;
2220 // Check the signatures.
2221 ciSignature* sig = signature();
2222 #ifdef ASSERT
2223 if (!is_store) {
2224 // Object getObject(Object base, int/long offset), etc.
2225 BasicType rtype = sig->return_type()->basic_type();
2226 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2227 rtype = T_ADDRESS; // it is really a C void*
2228 assert(rtype == type, "getter must return the expected value");
2229 if (!is_native_ptr) {
2230 assert(sig->count() == 2, "oop getter has 2 arguments");
2231 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2232 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2233 } else {
2234 assert(sig->count() == 1, "native getter has 1 argument");
2235 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2236 }
2237 } else {
2238 // void putObject(Object base, int/long offset, Object x), etc.
2239 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2240 if (!is_native_ptr) {
2241 assert(sig->count() == 3, "oop putter has 3 arguments");
2242 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2243 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2244 } else {
2245 assert(sig->count() == 2, "native putter has 2 arguments");
2246 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2247 }
2248 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2249 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2250 vtype = T_ADDRESS; // it is really a C void*
2251 assert(vtype == type, "putter must accept the expected value");
2252 }
2253 #endif // ASSERT
2254 }
2255 #endif //PRODUCT
2256
2257 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2258
2259 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2260
2261 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2262 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2263
2264 debug_only(int saved_sp = _sp);
2265 _sp += nargs;
2266
2267 Node* val;
2268 debug_only(val = (Node*)(uintptr_t)-1);
2269
2270
2271 if (is_store) {
2272 // Get the value being stored. (Pop it first; it was pushed last.)
2273 switch (type) {
2274 case T_DOUBLE:
2275 case T_LONG:
2276 case T_ADDRESS:
2277 val = pop_pair();
2278 break;
2279 default:
2280 val = pop();
2281 }
2282 }
2283
2284 // Build address expression. See the code in inline_unsafe_prefetch.
2285 Node *adr;
2286 Node *heap_base_oop = top();
2287 Node* offset = top();
2288
2289 if (!is_native_ptr) {
2290 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2291 offset = pop_pair();
2292 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2293 Node* base = pop();
2294 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2295 // to be plain byte offsets, which are also the same as those accepted
2296 // by oopDesc::field_base.
2297 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2298 "fieldOffset must be byte-scaled");
2299 // 32-bit machines ignore the high half!
2300 offset = ConvL2X(offset);
2301 adr = make_unsafe_address(base, offset);
2302 heap_base_oop = base;
2303 } else {
2304 Node* ptr = pop_pair();
2305 // Adjust Java long to machine word:
2306 ptr = ConvL2X(ptr);
2307 adr = make_unsafe_address(NULL, ptr);
2308 }
2309
2310 // Pop receiver last: it was pushed first.
2311 Node *receiver = pop();
2312
2313 assert(saved_sp == _sp, "must have correct argument count");
2314
2315 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2316
2317 // First guess at the value type.
2318 const Type *value_type = Type::get_const_basic_type(type);
2319
2320 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2321 // there was not enough information to nail it down.
2322 Compile::AliasType* alias_type = C->alias_type(adr_type);
2323 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2324
2325 // We will need memory barriers unless we can determine a unique
2326 // alias category for this reference. (Note: If for some reason
2327 // the barriers get omitted and the unsafe reference begins to "pollute"
2328 // the alias analysis of the rest of the graph, either Compile::can_alias
2329 // or Compile::must_alias will throw a diagnostic assert.)
2330 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2331
2332 // If we are reading the value of the referent field of a Reference
2333 // object (either by using Unsafe directly or through reflection)
2334 // then, if G1 is enabled, we need to record the referent in an
2335 // SATB log buffer using the pre-barrier mechanism.
2336 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store &&
2337 offset != top() && heap_base_oop != top();
2338
2339 if (!is_store && type == T_OBJECT) {
2340 // Attempt to infer a sharper value type from the offset and base type.
2341 ciKlass* sharpened_klass = NULL;
2342
2343 // See if it is an instance field, with an object type.
2344 if (alias_type->field() != NULL) {
2345 assert(!is_native_ptr, "native pointer op cannot use a java address");
2346 if (alias_type->field()->type()->is_klass()) {
2347 sharpened_klass = alias_type->field()->type()->as_klass();
2348 }
2349 }
2350
2351 // See if it is a narrow oop array.
2352 if (adr_type->isa_aryptr()) {
2353 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2354 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2355 if (elem_type != NULL) {
2356 sharpened_klass = elem_type->klass();
2357 }
2358 }
2359 }
2360
2361 if (sharpened_klass != NULL) {
2362 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2363
2364 // Sharpen the value type.
2365 value_type = tjp;
2366
2367 #ifndef PRODUCT
2368 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2369 tty->print(" from base type: "); adr_type->dump();
2370 tty->print(" sharpened value: "); value_type->dump();
2371 }
2372 #endif
2373 }
2374 }
2375
2376 // Null check on self without removing any arguments. The argument
2377 // null check technically happens in the wrong place, which can lead to
2378 // invalid stack traces when the primitive is inlined into a method
2379 // which handles NullPointerExceptions.
2380 _sp += nargs;
2381 do_null_check(receiver, T_OBJECT);
2382 _sp -= nargs;
2383 if (stopped()) {
2384 return true;
2385 }
2386 // Heap pointers get a null-check from the interpreter,
2387 // as a courtesy. However, this is not guaranteed by Unsafe,
2388 // and it is not possible to fully distinguish unintended nulls
2389 // from intended ones in this API.
2390
2391 if (is_volatile) {
2392 // We need to emit leading and trailing CPU membars (see below) in
2393 // addition to memory membars when is_volatile. This is a little
2394 // too strong, but avoids the need to insert per-alias-type
2395 // volatile membars (for stores; compare Parse::do_put_xxx), which
2396 // we cannot do effectively here because we probably only have a
2397 // rough approximation of type.
2398 need_mem_bar = true;
2399 // For Stores, place a memory ordering barrier now.
2400 if (is_store)
2401 insert_mem_bar(Op_MemBarRelease);
2402 }
2403
2404 // Memory barrier to prevent normal and 'unsafe' accesses from
2405 // bypassing each other. Happens after null checks, so the
2406 // exception paths do not take memory state from the memory barrier,
2407 // so there's no problems making a strong assert about mixing users
2408 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2409 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2410 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2411
2412 if (!is_store) {
2413 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2414 // load value and push onto stack
2415 switch (type) {
2416 case T_BOOLEAN:
2417 case T_CHAR:
2418 case T_BYTE:
2419 case T_SHORT:
2420 case T_INT:
2421 case T_FLOAT:
2422 push(p);
2423 break;
2424 case T_OBJECT:
2425 if (need_read_barrier) {
2426 insert_g1_pre_barrier(heap_base_oop, offset, p);
2427 }
2428 push(p);
2429 break;
2430 case T_ADDRESS:
2431 // Cast to an int type.
2432 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2433 p = ConvX2L(p);
2434 push_pair(p);
2435 break;
2436 case T_DOUBLE:
2437 case T_LONG:
2438 push_pair( p );
2439 break;
2440 default: ShouldNotReachHere();
2441 }
2442 } else {
2443 // place effect of store into memory
2444 switch (type) {
2445 case T_DOUBLE:
2446 val = dstore_rounding(val);
2447 break;
2448 case T_ADDRESS:
2449 // Repackage the long as a pointer.
2450 val = ConvL2X(val);
2451 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2452 break;
2453 }
2454
2455 if (type != T_OBJECT ) {
2456 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2457 } else {
2458 // Possibly an oop being stored to Java heap or native memory
2459 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2460 // oop to Java heap.
2461 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2462 } else {
2463 // We can't tell at compile time if we are storing in the Java heap or outside
2464 // of it. So we need to emit code to conditionally do the proper type of
2465 // store.
2466
2467 IdealKit ideal(this);
2468 #define __ ideal.
2469 // QQQ who knows what probability is here??
2470 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2471 // Sync IdealKit and graphKit.
2472 sync_kit(ideal);
2473 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2474 // Update IdealKit memory.
2475 __ sync_kit(this);
2476 } __ else_(); {
2477 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2478 } __ end_if();
2479 // Final sync IdealKit and GraphKit.
2480 final_sync(ideal);
2481 #undef __
2482 }
2483 }
2484 }
2485
2486 if (is_volatile) {
2487 if (!is_store)
2488 insert_mem_bar(Op_MemBarAcquire);
2489 else
2490 insert_mem_bar(Op_MemBarVolatile);
2491 }
2492
2493 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2494
2495 return true;
2496 }
2497
2498 //----------------------------inline_unsafe_prefetch----------------------------
2499
2500 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2501 #ifndef PRODUCT
2502 {
2503 ResourceMark rm;
2504 // Check the signatures.
2505 ciSignature* sig = signature();
2506 #ifdef ASSERT
2507 // Object getObject(Object base, int/long offset), etc.
2508 BasicType rtype = sig->return_type()->basic_type();
2509 if (!is_native_ptr) {
2510 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2511 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2512 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2513 } else {
2514 assert(sig->count() == 1, "native prefetch has 1 argument");
2515 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2516 }
2517 #endif // ASSERT
2518 }
2519 #endif // !PRODUCT
2520
2521 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2522
2523 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2524 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2525
2526 debug_only(int saved_sp = _sp);
2527 _sp += nargs;
2528
2529 // Build address expression. See the code in inline_unsafe_access.
2530 Node *adr;
2531 if (!is_native_ptr) {
2532 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2533 Node* offset = pop_pair();
2534 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2535 Node* base = pop();
2536 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2537 // to be plain byte offsets, which are also the same as those accepted
2538 // by oopDesc::field_base.
2539 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2540 "fieldOffset must be byte-scaled");
2541 // 32-bit machines ignore the high half!
2542 offset = ConvL2X(offset);
2543 adr = make_unsafe_address(base, offset);
2544 } else {
2545 Node* ptr = pop_pair();
2546 // Adjust Java long to machine word:
2547 ptr = ConvL2X(ptr);
2548 adr = make_unsafe_address(NULL, ptr);
2549 }
2550
2551 if (is_static) {
2552 assert(saved_sp == _sp, "must have correct argument count");
2553 } else {
2554 // Pop receiver last: it was pushed first.
2555 Node *receiver = pop();
2556 assert(saved_sp == _sp, "must have correct argument count");
2557
2558 // Null check on self without removing any arguments. The argument
2559 // null check technically happens in the wrong place, which can lead to
2560 // invalid stack traces when the primitive is inlined into a method
2561 // which handles NullPointerExceptions.
2562 _sp += nargs;
2563 do_null_check(receiver, T_OBJECT);
2564 _sp -= nargs;
2565 if (stopped()) {
2566 return true;
2567 }
2568 }
2569
2570 // Generate the read or write prefetch
2571 Node *prefetch;
2572 if (is_store) {
2573 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2574 } else {
2575 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2576 }
2577 prefetch->init_req(0, control());
2578 set_i_o(_gvn.transform(prefetch));
2579
2580 return true;
2581 }
2582
2583 //----------------------------inline_unsafe_CAS----------------------------
2584
2585 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2586 // This basic scheme here is the same as inline_unsafe_access, but
2587 // differs in enough details that combining them would make the code
2588 // overly confusing. (This is a true fact! I originally combined
2589 // them, but even I was confused by it!) As much code/comments as
2590 // possible are retained from inline_unsafe_access though to make
2591 // the correspondences clearer. - dl
2592
2593 if (callee()->is_static()) return false; // caller must have the capability!
2594
2595 #ifndef PRODUCT
2596 {
2597 ResourceMark rm;
2598 // Check the signatures.
2599 ciSignature* sig = signature();
2600 #ifdef ASSERT
2601 BasicType rtype = sig->return_type()->basic_type();
2602 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2603 assert(sig->count() == 4, "CAS has 4 arguments");
2604 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2605 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2606 #endif // ASSERT
2607 }
2608 #endif //PRODUCT
2609
2610 // number of stack slots per value argument (1 or 2)
2611 int type_words = type2size[type];
2612
2613 // Cannot inline wide CAS on machines that don't support it natively
2614 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2615 return false;
2616
2617 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2618
2619 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2620 int nargs = 1 + 1 + 2 + type_words + type_words;
2621
2622 // pop arguments: newval, oldval, offset, base, and receiver
2623 debug_only(int saved_sp = _sp);
2624 _sp += nargs;
2625 Node* newval = (type_words == 1) ? pop() : pop_pair();
2626 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2627 Node *offset = pop_pair();
2628 Node *base = pop();
2629 Node *receiver = pop();
2630 assert(saved_sp == _sp, "must have correct argument count");
2631
2632 // Null check receiver.
2633 _sp += nargs;
2634 do_null_check(receiver, T_OBJECT);
2635 _sp -= nargs;
2636 if (stopped()) {
2637 return true;
2638 }
2639
2640 // Build field offset expression.
2641 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2642 // to be plain byte offsets, which are also the same as those accepted
2643 // by oopDesc::field_base.
2644 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2645 // 32-bit machines ignore the high half of long offsets
2646 offset = ConvL2X(offset);
2647 Node* adr = make_unsafe_address(base, offset);
2648 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2649
2650 // (Unlike inline_unsafe_access, there seems no point in trying
2651 // to refine types. Just use the coarse types here.
2652 const Type *value_type = Type::get_const_basic_type(type);
2653 Compile::AliasType* alias_type = C->alias_type(adr_type);
2654 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2655 int alias_idx = C->get_alias_index(adr_type);
2656
2657 // Memory-model-wise, a CAS acts like a little synchronized block,
2658 // so needs barriers on each side. These don't translate into
2659 // actual barriers on most machines, but we still need rest of
2660 // compiler to respect ordering.
2661
2662 insert_mem_bar(Op_MemBarRelease);
2663 insert_mem_bar(Op_MemBarCPUOrder);
2664
2665 // 4984716: MemBars must be inserted before this
2666 // memory node in order to avoid a false
2667 // dependency which will confuse the scheduler.
2668 Node *mem = memory(alias_idx);
2669
2670 // For now, we handle only those cases that actually exist: ints,
2671 // longs, and Object. Adding others should be straightforward.
2672 Node* cas;
2673 switch(type) {
2674 case T_INT:
2675 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2676 break;
2677 case T_LONG:
2678 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2679 break;
2680 case T_OBJECT:
2681 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2682 // could be delayed during Parse (for example, in adjust_map_after_if()).
2683 // Execute transformation here to avoid barrier generation in such case.
2684 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2685 newval = _gvn.makecon(TypePtr::NULL_PTR);
2686
2687 // Reference stores need a store barrier.
2688 // (They don't if CAS fails, but it isn't worth checking.)
2689 pre_barrier(true /* do_load*/,
2690 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2691 NULL /* pre_val*/,
2692 T_OBJECT);
2693 #ifdef _LP64
2694 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2695 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2696 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2697 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2698 newval_enc, oldval_enc));
2699 } else
2700 #endif
2701 {
2702 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2703 }
2704 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2705 break;
2706 default:
2707 ShouldNotReachHere();
2708 break;
2709 }
2710
2711 // SCMemProjNodes represent the memory state of CAS. Their main
2712 // role is to prevent CAS nodes from being optimized away when their
2713 // results aren't used.
2714 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2715 set_memory(proj, alias_idx);
2716
2717 // Add the trailing membar surrounding the access
2718 insert_mem_bar(Op_MemBarCPUOrder);
2719 insert_mem_bar(Op_MemBarAcquire);
2720
2721 push(cas);
2722 return true;
2723 }
2724
2725 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2726 // This is another variant of inline_unsafe_access, differing in
2727 // that it always issues store-store ("release") barrier and ensures
2728 // store-atomicity (which only matters for "long").
2729
2730 if (callee()->is_static()) return false; // caller must have the capability!
2731
2732 #ifndef PRODUCT
2733 {
2734 ResourceMark rm;
2735 // Check the signatures.
2736 ciSignature* sig = signature();
2737 #ifdef ASSERT
2738 BasicType rtype = sig->return_type()->basic_type();
2739 assert(rtype == T_VOID, "must return void");
2740 assert(sig->count() == 3, "has 3 arguments");
2741 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2742 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2743 #endif // ASSERT
2744 }
2745 #endif //PRODUCT
2746
2747 // number of stack slots per value argument (1 or 2)
2748 int type_words = type2size[type];
2749
2750 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2751
2752 // Argument words: "this" plus oop plus offset plus value;
2753 int nargs = 1 + 1 + 2 + type_words;
2754
2755 // pop arguments: val, offset, base, and receiver
2756 debug_only(int saved_sp = _sp);
2757 _sp += nargs;
2758 Node* val = (type_words == 1) ? pop() : pop_pair();
2759 Node *offset = pop_pair();
2760 Node *base = pop();
2761 Node *receiver = pop();
2762 assert(saved_sp == _sp, "must have correct argument count");
2763
2764 // Null check receiver.
2765 _sp += nargs;
2766 do_null_check(receiver, T_OBJECT);
2767 _sp -= nargs;
2768 if (stopped()) {
2769 return true;
2770 }
2771
2772 // Build field offset expression.
2773 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2774 // 32-bit machines ignore the high half of long offsets
2775 offset = ConvL2X(offset);
2776 Node* adr = make_unsafe_address(base, offset);
2777 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2778 const Type *value_type = Type::get_const_basic_type(type);
2779 Compile::AliasType* alias_type = C->alias_type(adr_type);
2780
2781 insert_mem_bar(Op_MemBarRelease);
2782 insert_mem_bar(Op_MemBarCPUOrder);
2783 // Ensure that the store is atomic for longs:
2784 bool require_atomic_access = true;
2785 Node* store;
2786 if (type == T_OBJECT) // reference stores need a store barrier.
2787 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2788 else {
2789 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2790 }
2791 insert_mem_bar(Op_MemBarCPUOrder);
2792 return true;
2793 }
2794
2795 bool LibraryCallKit::inline_unsafe_allocate() {
2796 if (callee()->is_static()) return false; // caller must have the capability!
2797 int nargs = 1 + 1;
2798 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2799 null_check_receiver(callee()); // check then ignore argument(0)
2800 _sp += nargs; // set original stack for use by uncommon_trap
2801 Node* cls = do_null_check(argument(1), T_OBJECT);
2802 _sp -= nargs;
2803 if (stopped()) return true;
2804
2805 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2806 _sp += nargs; // set original stack for use by uncommon_trap
2807 kls = do_null_check(kls, T_OBJECT);
2808 _sp -= nargs;
2809 if (stopped()) return true; // argument was like int.class
2810
2811 // Note: The argument might still be an illegal value like
2812 // Serializable.class or Object[].class. The runtime will handle it.
2813 // But we must make an explicit check for initialization.
2814 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset()));
2815 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler
2816 // can generate code to load it as unsigned byte.
2817 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2818 Node* bits = intcon(instanceKlass::fully_initialized);
2819 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2820 // The 'test' is non-zero if we need to take a slow path.
2821
2822 Node* obj = new_instance(kls, test);
2823 push(obj);
2824
2825 return true;
2826 }
2827
2828 //------------------------inline_native_time_funcs--------------
2829 // inline code for System.currentTimeMillis() and System.nanoTime()
2830 // these have the same type and signature
2831 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2832 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2833 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2834 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2835 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2836 const TypePtr* no_memory_effects = NULL;
2837 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2838 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2839 #ifdef ASSERT
2840 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2841 assert(value_top == top(), "second value must be top");
2842 #endif
2843 push_pair(value);
2844 return true;
2845 }
2846
2847 //------------------------inline_native_currentThread------------------
2848 bool LibraryCallKit::inline_native_currentThread() {
2849 Node* junk = NULL;
2850 push(generate_current_thread(junk));
2851 return true;
2852 }
2853
2854 //------------------------inline_native_isInterrupted------------------
2855 bool LibraryCallKit::inline_native_isInterrupted() {
2856 const int nargs = 1+1; // receiver + boolean
2857 assert(nargs == arg_size(), "sanity");
2858 // Add a fast path to t.isInterrupted(clear_int):
2859 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2860 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2861 // So, in the common case that the interrupt bit is false,
2862 // we avoid making a call into the VM. Even if the interrupt bit
2863 // is true, if the clear_int argument is false, we avoid the VM call.
2864 // However, if the receiver is not currentThread, we must call the VM,
2865 // because there must be some locking done around the operation.
2866
2867 // We only go to the fast case code if we pass two guards.
2868 // Paths which do not pass are accumulated in the slow_region.
2869 RegionNode* slow_region = new (C, 1) RegionNode(1);
2870 record_for_igvn(slow_region);
2871 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2872 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2873 enum { no_int_result_path = 1,
2874 no_clear_result_path = 2,
2875 slow_result_path = 3
2876 };
2877
2878 // (a) Receiving thread must be the current thread.
2879 Node* rec_thr = argument(0);
2880 Node* tls_ptr = NULL;
2881 Node* cur_thr = generate_current_thread(tls_ptr);
2882 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2883 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2884
2885 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2886 if (!known_current_thread)
2887 generate_slow_guard(bol_thr, slow_region);
2888
2889 // (b) Interrupt bit on TLS must be false.
2890 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2891 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2892 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2893 // Set the control input on the field _interrupted read to prevent it floating up.
2894 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2895 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2896 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2897
2898 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2899
2900 // First fast path: if (!TLS._interrupted) return false;
2901 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2902 result_rgn->init_req(no_int_result_path, false_bit);
2903 result_val->init_req(no_int_result_path, intcon(0));
2904
2905 // drop through to next case
2906 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2907
2908 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2909 Node* clr_arg = argument(1);
2910 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2911 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2912 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2913
2914 // Second fast path: ... else if (!clear_int) return true;
2915 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2916 result_rgn->init_req(no_clear_result_path, false_arg);
2917 result_val->init_req(no_clear_result_path, intcon(1));
2918
2919 // drop through to next case
2920 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2921
2922 // (d) Otherwise, go to the slow path.
2923 slow_region->add_req(control());
2924 set_control( _gvn.transform(slow_region) );
2925
2926 if (stopped()) {
2927 // There is no slow path.
2928 result_rgn->init_req(slow_result_path, top());
2929 result_val->init_req(slow_result_path, top());
2930 } else {
2931 // non-virtual because it is a private non-static
2932 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2933
2934 Node* slow_val = set_results_for_java_call(slow_call);
2935 // this->control() comes from set_results_for_java_call
2936
2937 // If we know that the result of the slow call will be true, tell the optimizer!
2938 if (known_current_thread) slow_val = intcon(1);
2939
2940 Node* fast_io = slow_call->in(TypeFunc::I_O);
2941 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2942 // These two phis are pre-filled with copies of of the fast IO and Memory
2943 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2944 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2945
2946 result_rgn->init_req(slow_result_path, control());
2947 io_phi ->init_req(slow_result_path, i_o());
2948 mem_phi ->init_req(slow_result_path, reset_memory());
2949 result_val->init_req(slow_result_path, slow_val);
2950
2951 set_all_memory( _gvn.transform(mem_phi) );
2952 set_i_o( _gvn.transform(io_phi) );
2953 }
2954
2955 push_result(result_rgn, result_val);
2956 C->set_has_split_ifs(true); // Has chance for split-if optimization
2957
2958 return true;
2959 }
2960
2961 //---------------------------load_mirror_from_klass----------------------------
2962 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2963 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2964 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
2965 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2966 }
2967
2968 //-----------------------load_klass_from_mirror_common-------------------------
2969 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2970 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2971 // and branch to the given path on the region.
2972 // If never_see_null, take an uncommon trap on null, so we can optimistically
2973 // compile for the non-null case.
2974 // If the region is NULL, force never_see_null = true.
2975 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2976 bool never_see_null,
2977 int nargs,
2978 RegionNode* region,
2979 int null_path,
2980 int offset) {
2981 if (region == NULL) never_see_null = true;
2982 Node* p = basic_plus_adr(mirror, offset);
2983 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2984 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2985 _sp += nargs; // any deopt will start just before call to enclosing method
2986 Node* null_ctl = top();
2987 kls = null_check_oop(kls, &null_ctl, never_see_null);
2988 if (region != NULL) {
2989 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2990 region->init_req(null_path, null_ctl);
2991 } else {
2992 assert(null_ctl == top(), "no loose ends");
2993 }
2994 _sp -= nargs;
2995 return kls;
2996 }
2997
2998 //--------------------(inline_native_Class_query helpers)---------------------
2999 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3000 // Fall through if (mods & mask) == bits, take the guard otherwise.
3001 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3002 // Branch around if the given klass has the given modifier bit set.
3003 // Like generate_guard, adds a new path onto the region.
3004 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3005 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3006 Node* mask = intcon(modifier_mask);
3007 Node* bits = intcon(modifier_bits);
3008 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
3009 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
3010 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
3011 return generate_fair_guard(bol, region);
3012 }
3013 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3014 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3015 }
3016
3017 //-------------------------inline_native_Class_query-------------------
3018 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3019 int nargs = 1+0; // just the Class mirror, in most cases
3020 const Type* return_type = TypeInt::BOOL;
3021 Node* prim_return_value = top(); // what happens if it's a primitive class?
3022 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3023 bool expect_prim = false; // most of these guys expect to work on refs
3024
3025 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3026
3027 switch (id) {
3028 case vmIntrinsics::_isInstance:
3029 nargs = 1+1; // the Class mirror, plus the object getting queried about
3030 // nothing is an instance of a primitive type
3031 prim_return_value = intcon(0);
3032 break;
3033 case vmIntrinsics::_getModifiers:
3034 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3035 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3036 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3037 break;
3038 case vmIntrinsics::_isInterface:
3039 prim_return_value = intcon(0);
3040 break;
3041 case vmIntrinsics::_isArray:
3042 prim_return_value = intcon(0);
3043 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3044 break;
3045 case vmIntrinsics::_isPrimitive:
3046 prim_return_value = intcon(1);
3047 expect_prim = true; // obviously
3048 break;
3049 case vmIntrinsics::_getSuperclass:
3050 prim_return_value = null();
3051 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3052 break;
3053 case vmIntrinsics::_getComponentType:
3054 prim_return_value = null();
3055 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3056 break;
3057 case vmIntrinsics::_getClassAccessFlags:
3058 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3059 return_type = TypeInt::INT; // not bool! 6297094
3060 break;
3061 default:
3062 ShouldNotReachHere();
3063 }
3064
3065 Node* mirror = argument(0);
3066 Node* obj = (nargs <= 1)? top(): argument(1);
3067
3068 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3069 if (mirror_con == NULL) return false; // cannot happen?
3070
3071 #ifndef PRODUCT
3072 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3073 ciType* k = mirror_con->java_mirror_type();
3074 if (k) {
3075 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3076 k->print_name();
3077 tty->cr();
3078 }
3079 }
3080 #endif
3081
3082 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3083 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3084 record_for_igvn(region);
3085 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
3086
3087 // The mirror will never be null of Reflection.getClassAccessFlags, however
3088 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3089 // if it is. See bug 4774291.
3090
3091 // For Reflection.getClassAccessFlags(), the null check occurs in
3092 // the wrong place; see inline_unsafe_access(), above, for a similar
3093 // situation.
3094 _sp += nargs; // set original stack for use by uncommon_trap
3095 mirror = do_null_check(mirror, T_OBJECT);
3096 _sp -= nargs;
3097 // If mirror or obj is dead, only null-path is taken.
3098 if (stopped()) return true;
3099
3100 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3101
3102 // Now load the mirror's klass metaobject, and null-check it.
3103 // Side-effects region with the control path if the klass is null.
3104 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3105 region, _prim_path);
3106 // If kls is null, we have a primitive mirror.
3107 phi->init_req(_prim_path, prim_return_value);
3108 if (stopped()) { push_result(region, phi); return true; }
3109
3110 Node* p; // handy temp
3111 Node* null_ctl;
3112
3113 // Now that we have the non-null klass, we can perform the real query.
3114 // For constant classes, the query will constant-fold in LoadNode::Value.
3115 Node* query_value = top();
3116 switch (id) {
3117 case vmIntrinsics::_isInstance:
3118 // nothing is an instance of a primitive type
3119 _sp += nargs; // gen_instanceof might do an uncommon trap
3120 query_value = gen_instanceof(obj, kls);
3121 _sp -= nargs;
3122 break;
3123
3124 case vmIntrinsics::_getModifiers:
3125 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3126 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3127 break;
3128
3129 case vmIntrinsics::_isInterface:
3130 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3131 if (generate_interface_guard(kls, region) != NULL)
3132 // A guard was added. If the guard is taken, it was an interface.
3133 phi->add_req(intcon(1));
3134 // If we fall through, it's a plain class.
3135 query_value = intcon(0);
3136 break;
3137
3138 case vmIntrinsics::_isArray:
3139 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3140 if (generate_array_guard(kls, region) != NULL)
3141 // A guard was added. If the guard is taken, it was an array.
3142 phi->add_req(intcon(1));
3143 // If we fall through, it's a plain class.
3144 query_value = intcon(0);
3145 break;
3146
3147 case vmIntrinsics::_isPrimitive:
3148 query_value = intcon(0); // "normal" path produces false
3149 break;
3150
3151 case vmIntrinsics::_getSuperclass:
3152 // The rules here are somewhat unfortunate, but we can still do better
3153 // with random logic than with a JNI call.
3154 // Interfaces store null or Object as _super, but must report null.
3155 // Arrays store an intermediate super as _super, but must report Object.
3156 // Other types can report the actual _super.
3157 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3158 if (generate_interface_guard(kls, region) != NULL)
3159 // A guard was added. If the guard is taken, it was an interface.
3160 phi->add_req(null());
3161 if (generate_array_guard(kls, region) != NULL)
3162 // A guard was added. If the guard is taken, it was an array.
3163 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3164 // If we fall through, it's a plain class. Get its _super.
3165 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3166 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3167 null_ctl = top();
3168 kls = null_check_oop(kls, &null_ctl);
3169 if (null_ctl != top()) {
3170 // If the guard is taken, Object.superClass is null (both klass and mirror).
3171 region->add_req(null_ctl);
3172 phi ->add_req(null());
3173 }
3174 if (!stopped()) {
3175 query_value = load_mirror_from_klass(kls);
3176 }
3177 break;
3178
3179 case vmIntrinsics::_getComponentType:
3180 if (generate_array_guard(kls, region) != NULL) {
3181 // Be sure to pin the oop load to the guard edge just created:
3182 Node* is_array_ctrl = region->in(region->req()-1);
3183 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3184 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3185 phi->add_req(cmo);
3186 }
3187 query_value = null(); // non-array case is null
3188 break;
3189
3190 case vmIntrinsics::_getClassAccessFlags:
3191 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3192 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3193 break;
3194
3195 default:
3196 ShouldNotReachHere();
3197 }
3198
3199 // Fall-through is the normal case of a query to a real class.
3200 phi->init_req(1, query_value);
3201 region->init_req(1, control());
3202
3203 push_result(region, phi);
3204 C->set_has_split_ifs(true); // Has chance for split-if optimization
3205
3206 return true;
3207 }
3208
3209 //--------------------------inline_native_subtype_check------------------------
3210 // This intrinsic takes the JNI calls out of the heart of
3211 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3212 bool LibraryCallKit::inline_native_subtype_check() {
3213 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3214
3215 // Pull both arguments off the stack.
3216 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3217 args[0] = argument(0);
3218 args[1] = argument(1);
3219 Node* klasses[2]; // corresponding Klasses: superk, subk
3220 klasses[0] = klasses[1] = top();
3221
3222 enum {
3223 // A full decision tree on {superc is prim, subc is prim}:
3224 _prim_0_path = 1, // {P,N} => false
3225 // {P,P} & superc!=subc => false
3226 _prim_same_path, // {P,P} & superc==subc => true
3227 _prim_1_path, // {N,P} => false
3228 _ref_subtype_path, // {N,N} & subtype check wins => true
3229 _both_ref_path, // {N,N} & subtype check loses => false
3230 PATH_LIMIT
3231 };
3232
3233 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3234 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3235 record_for_igvn(region);
3236
3237 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3238 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3239 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3240
3241 // First null-check both mirrors and load each mirror's klass metaobject.
3242 int which_arg;
3243 for (which_arg = 0; which_arg <= 1; which_arg++) {
3244 Node* arg = args[which_arg];
3245 _sp += nargs; // set original stack for use by uncommon_trap
3246 arg = do_null_check(arg, T_OBJECT);
3247 _sp -= nargs;
3248 if (stopped()) break;
3249 args[which_arg] = _gvn.transform(arg);
3250
3251 Node* p = basic_plus_adr(arg, class_klass_offset);
3252 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3253 klasses[which_arg] = _gvn.transform(kls);
3254 }
3255
3256 // Having loaded both klasses, test each for null.
3257 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3258 for (which_arg = 0; which_arg <= 1; which_arg++) {
3259 Node* kls = klasses[which_arg];
3260 Node* null_ctl = top();
3261 _sp += nargs; // set original stack for use by uncommon_trap
3262 kls = null_check_oop(kls, &null_ctl, never_see_null);
3263 _sp -= nargs;
3264 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3265 region->init_req(prim_path, null_ctl);
3266 if (stopped()) break;
3267 klasses[which_arg] = kls;
3268 }
3269
3270 if (!stopped()) {
3271 // now we have two reference types, in klasses[0..1]
3272 Node* subk = klasses[1]; // the argument to isAssignableFrom
3273 Node* superk = klasses[0]; // the receiver
3274 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3275 // now we have a successful reference subtype check
3276 region->set_req(_ref_subtype_path, control());
3277 }
3278
3279 // If both operands are primitive (both klasses null), then
3280 // we must return true when they are identical primitives.
3281 // It is convenient to test this after the first null klass check.
3282 set_control(region->in(_prim_0_path)); // go back to first null check
3283 if (!stopped()) {
3284 // Since superc is primitive, make a guard for the superc==subc case.
3285 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3286 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3287 generate_guard(bol_eq, region, PROB_FAIR);
3288 if (region->req() == PATH_LIMIT+1) {
3289 // A guard was added. If the added guard is taken, superc==subc.
3290 region->swap_edges(PATH_LIMIT, _prim_same_path);
3291 region->del_req(PATH_LIMIT);
3292 }
3293 region->set_req(_prim_0_path, control()); // Not equal after all.
3294 }
3295
3296 // these are the only paths that produce 'true':
3297 phi->set_req(_prim_same_path, intcon(1));
3298 phi->set_req(_ref_subtype_path, intcon(1));
3299
3300 // pull together the cases:
3301 assert(region->req() == PATH_LIMIT, "sane region");
3302 for (uint i = 1; i < region->req(); i++) {
3303 Node* ctl = region->in(i);
3304 if (ctl == NULL || ctl == top()) {
3305 region->set_req(i, top());
3306 phi ->set_req(i, top());
3307 } else if (phi->in(i) == NULL) {
3308 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3309 }
3310 }
3311
3312 set_control(_gvn.transform(region));
3313 push(_gvn.transform(phi));
3314
3315 return true;
3316 }
3317
3318 //---------------------generate_array_guard_common------------------------
3319 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3320 bool obj_array, bool not_array) {
3321 // If obj_array/non_array==false/false:
3322 // Branch around if the given klass is in fact an array (either obj or prim).
3323 // If obj_array/non_array==false/true:
3324 // Branch around if the given klass is not an array klass of any kind.
3325 // If obj_array/non_array==true/true:
3326 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3327 // If obj_array/non_array==true/false:
3328 // Branch around if the kls is an oop array (Object[] or subtype)
3329 //
3330 // Like generate_guard, adds a new path onto the region.
3331 jint layout_con = 0;
3332 Node* layout_val = get_layout_helper(kls, layout_con);
3333 if (layout_val == NULL) {
3334 bool query = (obj_array
3335 ? Klass::layout_helper_is_objArray(layout_con)
3336 : Klass::layout_helper_is_javaArray(layout_con));
3337 if (query == not_array) {
3338 return NULL; // never a branch
3339 } else { // always a branch
3340 Node* always_branch = control();
3341 if (region != NULL)
3342 region->add_req(always_branch);
3343 set_control(top());
3344 return always_branch;
3345 }
3346 }
3347 // Now test the correct condition.
3348 jint nval = (obj_array
3349 ? ((jint)Klass::_lh_array_tag_type_value
3350 << Klass::_lh_array_tag_shift)
3351 : Klass::_lh_neutral_value);
3352 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3353 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3354 // invert the test if we are looking for a non-array
3355 if (not_array) btest = BoolTest(btest).negate();
3356 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3357 return generate_fair_guard(bol, region);
3358 }
3359
3360
3361 //-----------------------inline_native_newArray--------------------------
3362 bool LibraryCallKit::inline_native_newArray() {
3363 int nargs = 2;
3364 Node* mirror = argument(0);
3365 Node* count_val = argument(1);
3366
3367 _sp += nargs; // set original stack for use by uncommon_trap
3368 mirror = do_null_check(mirror, T_OBJECT);
3369 _sp -= nargs;
3370 // If mirror or obj is dead, only null-path is taken.
3371 if (stopped()) return true;
3372
3373 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3374 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3375 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3376 TypeInstPtr::NOTNULL);
3377 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3378 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3379 TypePtr::BOTTOM);
3380
3381 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3382 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3383 nargs,
3384 result_reg, _slow_path);
3385 Node* normal_ctl = control();
3386 Node* no_array_ctl = result_reg->in(_slow_path);
3387
3388 // Generate code for the slow case. We make a call to newArray().
3389 set_control(no_array_ctl);
3390 if (!stopped()) {
3391 // Either the input type is void.class, or else the
3392 // array klass has not yet been cached. Either the
3393 // ensuing call will throw an exception, or else it
3394 // will cache the array klass for next time.
3395 PreserveJVMState pjvms(this);
3396 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3397 Node* slow_result = set_results_for_java_call(slow_call);
3398 // this->control() comes from set_results_for_java_call
3399 result_reg->set_req(_slow_path, control());
3400 result_val->set_req(_slow_path, slow_result);
3401 result_io ->set_req(_slow_path, i_o());
3402 result_mem->set_req(_slow_path, reset_memory());
3403 }
3404
3405 set_control(normal_ctl);
3406 if (!stopped()) {
3407 // Normal case: The array type has been cached in the java.lang.Class.
3408 // The following call works fine even if the array type is polymorphic.
3409 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3410 Node* obj = new_array(klass_node, count_val, nargs);
3411 result_reg->init_req(_normal_path, control());
3412 result_val->init_req(_normal_path, obj);
3413 result_io ->init_req(_normal_path, i_o());
3414 result_mem->init_req(_normal_path, reset_memory());
3415 }
3416
3417 // Return the combined state.
3418 set_i_o( _gvn.transform(result_io) );
3419 set_all_memory( _gvn.transform(result_mem) );
3420 push_result(result_reg, result_val);
3421 C->set_has_split_ifs(true); // Has chance for split-if optimization
3422
3423 return true;
3424 }
3425
3426 //----------------------inline_native_getLength--------------------------
3427 bool LibraryCallKit::inline_native_getLength() {
3428 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3429
3430 int nargs = 1;
3431 Node* array = argument(0);
3432
3433 _sp += nargs; // set original stack for use by uncommon_trap
3434 array = do_null_check(array, T_OBJECT);
3435 _sp -= nargs;
3436
3437 // If array is dead, only null-path is taken.
3438 if (stopped()) return true;
3439
3440 // Deoptimize if it is a non-array.
3441 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3442
3443 if (non_array != NULL) {
3444 PreserveJVMState pjvms(this);
3445 set_control(non_array);
3446 _sp += nargs; // push the arguments back on the stack
3447 uncommon_trap(Deoptimization::Reason_intrinsic,
3448 Deoptimization::Action_maybe_recompile);
3449 }
3450
3451 // If control is dead, only non-array-path is taken.
3452 if (stopped()) return true;
3453
3454 // The works fine even if the array type is polymorphic.
3455 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3456 push( load_array_length(array) );
3457
3458 C->set_has_split_ifs(true); // Has chance for split-if optimization
3459
3460 return true;
3461 }
3462
3463 //------------------------inline_array_copyOf----------------------------
3464 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3465 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3466
3467 // Restore the stack and pop off the arguments.
3468 int nargs = 3 + (is_copyOfRange? 1: 0);
3469 Node* original = argument(0);
3470 Node* start = is_copyOfRange? argument(1): intcon(0);
3471 Node* end = is_copyOfRange? argument(2): argument(1);
3472 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3473
3474 Node* newcopy;
3475
3476 //set the original stack and the reexecute bit for the interpreter to reexecute
3477 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3478 { PreserveReexecuteState preexecs(this);
3479 _sp += nargs;
3480 jvms()->set_should_reexecute(true);
3481
3482 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3483 original = do_null_check(original, T_OBJECT);
3484
3485 // Check if a null path was taken unconditionally.
3486 if (stopped()) return true;
3487
3488 Node* orig_length = load_array_length(original);
3489
3490 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3491 NULL, 0);
3492 klass_node = do_null_check(klass_node, T_OBJECT);
3493
3494 RegionNode* bailout = new (C, 1) RegionNode(1);
3495 record_for_igvn(bailout);
3496
3497 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3498 // Bail out if that is so.
3499 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3500 if (not_objArray != NULL) {
3501 // Improve the klass node's type from the new optimistic assumption:
3502 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3503 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3504 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3505 cast->init_req(0, control());
3506 klass_node = _gvn.transform(cast);
3507 }
3508
3509 // Bail out if either start or end is negative.
3510 generate_negative_guard(start, bailout, &start);
3511 generate_negative_guard(end, bailout, &end);
3512
3513 Node* length = end;
3514 if (_gvn.type(start) != TypeInt::ZERO) {
3515 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3516 }
3517
3518 // Bail out if length is negative.
3519 // ...Not needed, since the new_array will throw the right exception.
3520 //generate_negative_guard(length, bailout, &length);
3521
3522 if (bailout->req() > 1) {
3523 PreserveJVMState pjvms(this);
3524 set_control( _gvn.transform(bailout) );
3525 uncommon_trap(Deoptimization::Reason_intrinsic,
3526 Deoptimization::Action_maybe_recompile);
3527 }
3528
3529 if (!stopped()) {
3530
3531 // How many elements will we copy from the original?
3532 // The answer is MinI(orig_length - start, length).
3533 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3534 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3535
3536 newcopy = new_array(klass_node, length, 0);
3537
3538 // Generate a direct call to the right arraycopy function(s).
3539 // We know the copy is disjoint but we might not know if the
3540 // oop stores need checking.
3541 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3542 // This will fail a store-check if x contains any non-nulls.
3543 bool disjoint_bases = true;
3544 bool length_never_negative = true;
3545 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3546 original, start, newcopy, intcon(0), moved,
3547 disjoint_bases, length_never_negative);
3548 }
3549 } //original reexecute and sp are set back here
3550
3551 if(!stopped()) {
3552 push(newcopy);
3553 }
3554
3555 C->set_has_split_ifs(true); // Has chance for split-if optimization
3556
3557 return true;
3558 }
3559
3560
3561 //----------------------generate_virtual_guard---------------------------
3562 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3563 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3564 RegionNode* slow_region) {
3565 ciMethod* method = callee();
3566 int vtable_index = method->vtable_index();
3567 // Get the methodOop out of the appropriate vtable entry.
3568 int entry_offset = (instanceKlass::vtable_start_offset() +
3569 vtable_index*vtableEntry::size()) * wordSize +
3570 vtableEntry::method_offset_in_bytes();
3571 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3572 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3573
3574 // Compare the target method with the expected method (e.g., Object.hashCode).
3575 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3576
3577 Node* native_call = makecon(native_call_addr);
3578 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3579 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3580
3581 return generate_slow_guard(test_native, slow_region);
3582 }
3583
3584 //-----------------------generate_method_call----------------------------
3585 // Use generate_method_call to make a slow-call to the real
3586 // method if the fast path fails. An alternative would be to
3587 // use a stub like OptoRuntime::slow_arraycopy_Java.
3588 // This only works for expanding the current library call,
3589 // not another intrinsic. (E.g., don't use this for making an
3590 // arraycopy call inside of the copyOf intrinsic.)
3591 CallJavaNode*
3592 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3593 // When compiling the intrinsic method itself, do not use this technique.
3594 guarantee(callee() != C->method(), "cannot make slow-call to self");
3595
3596 ciMethod* method = callee();
3597 // ensure the JVMS we have will be correct for this call
3598 guarantee(method_id == method->intrinsic_id(), "must match");
3599
3600 const TypeFunc* tf = TypeFunc::make(method);
3601 int tfdc = tf->domain()->cnt();
3602 CallJavaNode* slow_call;
3603 if (is_static) {
3604 assert(!is_virtual, "");
3605 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3606 SharedRuntime::get_resolve_static_call_stub(),
3607 method, bci());
3608 } else if (is_virtual) {
3609 null_check_receiver(method);
3610 int vtable_index = methodOopDesc::invalid_vtable_index;
3611 if (UseInlineCaches) {
3612 // Suppress the vtable call
3613 } else {
3614 // hashCode and clone are not a miranda methods,
3615 // so the vtable index is fixed.
3616 // No need to use the linkResolver to get it.
3617 vtable_index = method->vtable_index();
3618 }
3619 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3620 SharedRuntime::get_resolve_virtual_call_stub(),
3621 method, vtable_index, bci());
3622 } else { // neither virtual nor static: opt_virtual
3623 null_check_receiver(method);
3624 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3625 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3626 method, bci());
3627 slow_call->set_optimized_virtual(true);
3628 }
3629 set_arguments_for_java_call(slow_call);
3630 set_edges_for_java_call(slow_call);
3631 return slow_call;
3632 }
3633
3634
3635 //------------------------------inline_native_hashcode--------------------
3636 // Build special case code for calls to hashCode on an object.
3637 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3638 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3639 assert(!(is_virtual && is_static), "either virtual, special, or static");
3640
3641 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3642
3643 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3644 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3645 TypeInt::INT);
3646 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3647 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3648 TypePtr::BOTTOM);
3649 Node* obj = NULL;
3650 if (!is_static) {
3651 // Check for hashing null object
3652 obj = null_check_receiver(callee());
3653 if (stopped()) return true; // unconditionally null
3654 result_reg->init_req(_null_path, top());
3655 result_val->init_req(_null_path, top());
3656 } else {
3657 // Do a null check, and return zero if null.
3658 // System.identityHashCode(null) == 0
3659 obj = argument(0);
3660 Node* null_ctl = top();
3661 obj = null_check_oop(obj, &null_ctl);
3662 result_reg->init_req(_null_path, null_ctl);
3663 result_val->init_req(_null_path, _gvn.intcon(0));
3664 }
3665
3666 // Unconditionally null? Then return right away.
3667 if (stopped()) {
3668 set_control( result_reg->in(_null_path) );
3669 if (!stopped())
3670 push( result_val ->in(_null_path) );
3671 return true;
3672 }
3673
3674 // After null check, get the object's klass.
3675 Node* obj_klass = load_object_klass(obj);
3676
3677 // This call may be virtual (invokevirtual) or bound (invokespecial).
3678 // For each case we generate slightly different code.
3679
3680 // We only go to the fast case code if we pass a number of guards. The
3681 // paths which do not pass are accumulated in the slow_region.
3682 RegionNode* slow_region = new (C, 1) RegionNode(1);
3683 record_for_igvn(slow_region);
3684
3685 // If this is a virtual call, we generate a funny guard. We pull out
3686 // the vtable entry corresponding to hashCode() from the target object.
3687 // If the target method which we are calling happens to be the native
3688 // Object hashCode() method, we pass the guard. We do not need this
3689 // guard for non-virtual calls -- the caller is known to be the native
3690 // Object hashCode().
3691 if (is_virtual) {
3692 generate_virtual_guard(obj_klass, slow_region);
3693 }
3694
3695 // Get the header out of the object, use LoadMarkNode when available
3696 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3697 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3698
3699 // Test the header to see if it is unlocked.
3700 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3701 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3702 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3703 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3704 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3705
3706 generate_slow_guard(test_unlocked, slow_region);
3707
3708 // Get the hash value and check to see that it has been properly assigned.
3709 // We depend on hash_mask being at most 32 bits and avoid the use of
3710 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3711 // vm: see markOop.hpp.
3712 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3713 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3714 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3715 // This hack lets the hash bits live anywhere in the mark object now, as long
3716 // as the shift drops the relevant bits into the low 32 bits. Note that
3717 // Java spec says that HashCode is an int so there's no point in capturing
3718 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3719 hshifted_header = ConvX2I(hshifted_header);
3720 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3721
3722 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3723 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3724 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3725
3726 generate_slow_guard(test_assigned, slow_region);
3727
3728 Node* init_mem = reset_memory();
3729 // fill in the rest of the null path:
3730 result_io ->init_req(_null_path, i_o());
3731 result_mem->init_req(_null_path, init_mem);
3732
3733 result_val->init_req(_fast_path, hash_val);
3734 result_reg->init_req(_fast_path, control());
3735 result_io ->init_req(_fast_path, i_o());
3736 result_mem->init_req(_fast_path, init_mem);
3737
3738 // Generate code for the slow case. We make a call to hashCode().
3739 set_control(_gvn.transform(slow_region));
3740 if (!stopped()) {
3741 // No need for PreserveJVMState, because we're using up the present state.
3742 set_all_memory(init_mem);
3743 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3744 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3745 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3746 Node* slow_result = set_results_for_java_call(slow_call);
3747 // this->control() comes from set_results_for_java_call
3748 result_reg->init_req(_slow_path, control());
3749 result_val->init_req(_slow_path, slow_result);
3750 result_io ->set_req(_slow_path, i_o());
3751 result_mem ->set_req(_slow_path, reset_memory());
3752 }
3753
3754 // Return the combined state.
3755 set_i_o( _gvn.transform(result_io) );
3756 set_all_memory( _gvn.transform(result_mem) );
3757 push_result(result_reg, result_val);
3758
3759 return true;
3760 }
3761
3762 //---------------------------inline_native_getClass----------------------------
3763 // Build special case code for calls to getClass on an object.
3764 bool LibraryCallKit::inline_native_getClass() {
3765 Node* obj = null_check_receiver(callee());
3766 if (stopped()) return true;
3767 push( load_mirror_from_klass(load_object_klass(obj)) );
3768 return true;
3769 }
3770
3771 //-----------------inline_native_Reflection_getCallerClass---------------------
3772 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3773 //
3774 // NOTE that this code must perform the same logic as
3775 // vframeStream::security_get_caller_frame in that it must skip
3776 // Method.invoke() and auxiliary frames.
3777
3778
3779
3780
3781 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3782 ciMethod* method = callee();
3783
3784 #ifndef PRODUCT
3785 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3786 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3787 }
3788 #endif
3789
3790 debug_only(int saved_sp = _sp);
3791
3792 // Argument words: (int depth)
3793 int nargs = 1;
3794
3795 _sp += nargs;
3796 Node* caller_depth_node = pop();
3797
3798 assert(saved_sp == _sp, "must have correct argument count");
3799
3800 // The depth value must be a constant in order for the runtime call
3801 // to be eliminated.
3802 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3803 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3804 #ifndef PRODUCT
3805 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3806 tty->print_cr(" Bailing out because caller depth was not a constant");
3807 }
3808 #endif
3809 return false;
3810 }
3811 // Note that the JVM state at this point does not include the
3812 // getCallerClass() frame which we are trying to inline. The
3813 // semantics of getCallerClass(), however, are that the "first"
3814 // frame is the getCallerClass() frame, so we subtract one from the
3815 // requested depth before continuing. We don't inline requests of
3816 // getCallerClass(0).
3817 int caller_depth = caller_depth_type->get_con() - 1;
3818 if (caller_depth < 0) {
3819 #ifndef PRODUCT
3820 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3821 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3822 }
3823 #endif
3824 return false;
3825 }
3826
3827 if (!jvms()->has_method()) {
3828 #ifndef PRODUCT
3829 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3830 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3831 }
3832 #endif
3833 return false;
3834 }
3835 int _depth = jvms()->depth(); // cache call chain depth
3836
3837 // Walk back up the JVM state to find the caller at the required
3838 // depth. NOTE that this code must perform the same logic as
3839 // vframeStream::security_get_caller_frame in that it must skip
3840 // Method.invoke() and auxiliary frames. Note also that depth is
3841 // 1-based (1 is the bottom of the inlining).
3842 int inlining_depth = _depth;
3843 JVMState* caller_jvms = NULL;
3844
3845 if (inlining_depth > 0) {
3846 caller_jvms = jvms();
3847 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3848 do {
3849 // The following if-tests should be performed in this order
3850 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3851 // Skip a Method.invoke() or auxiliary frame
3852 } else if (caller_depth > 0) {
3853 // Skip real frame
3854 --caller_depth;
3855 } else {
3856 // We're done: reached desired caller after skipping.
3857 break;
3858 }
3859 caller_jvms = caller_jvms->caller();
3860 --inlining_depth;
3861 } while (inlining_depth > 0);
3862 }
3863
3864 if (inlining_depth == 0) {
3865 #ifndef PRODUCT
3866 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3867 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3868 tty->print_cr(" JVM state at this point:");
3869 for (int i = _depth; i >= 1; i--) {
3870 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3871 }
3872 }
3873 #endif
3874 return false; // Reached end of inlining
3875 }
3876
3877 // Acquire method holder as java.lang.Class
3878 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3879 ciInstance* caller_mirror = caller_klass->java_mirror();
3880 // Push this as a constant
3881 push(makecon(TypeInstPtr::make(caller_mirror)));
3882 #ifndef PRODUCT
3883 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3884 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);
3885 tty->print_cr(" JVM state at this point:");
3886 for (int i = _depth; i >= 1; i--) {
3887 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3888 }
3889 }
3890 #endif
3891 return true;
3892 }
3893
3894 // Helper routine for above
3895 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3896 ciMethod* method = jvms->method();
3897
3898 // Is this the Method.invoke method itself?
3899 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3900 return true;
3901
3902 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3903 ciKlass* k = method->holder();
3904 if (k->is_instance_klass()) {
3905 ciInstanceKlass* ik = k->as_instance_klass();
3906 for (; ik != NULL; ik = ik->super()) {
3907 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3908 ik == env()->find_system_klass(ik->name())) {
3909 return true;
3910 }
3911 }
3912 }
3913 else if (method->is_method_handle_adapter()) {
3914 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3915 return true;
3916 }
3917
3918 return false;
3919 }
3920
3921 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3922 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3923 // computing it since there is no lookup field by name function in the
3924 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3925 // Using a static variable here is safe even if we have multiple compilation
3926 // threads because the offset is constant. At worst the same offset will be
3927 // computed and stored multiple
3928
3929 bool LibraryCallKit::inline_native_AtomicLong_get() {
3930 // Restore the stack and pop off the argument
3931 _sp+=1;
3932 Node *obj = pop();
3933
3934 // get the offset of the "value" field. Since the CI interfaces
3935 // does not provide a way to look up a field by name, we scan the bytecodes
3936 // to get the field index. We expect the first 2 instructions of the method
3937 // to be:
3938 // 0 aload_0
3939 // 1 getfield "value"
3940 ciMethod* method = callee();
3941 if (value_field_offset == -1)
3942 {
3943 ciField* value_field;
3944 ciBytecodeStream iter(method);
3945 Bytecodes::Code bc = iter.next();
3946
3947 if ((bc != Bytecodes::_aload_0) &&
3948 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3949 return false;
3950 bc = iter.next();
3951 if (bc != Bytecodes::_getfield)
3952 return false;
3953 bool ignore;
3954 value_field = iter.get_field(ignore);
3955 value_field_offset = value_field->offset_in_bytes();
3956 }
3957
3958 // Null check without removing any arguments.
3959 _sp++;
3960 obj = do_null_check(obj, T_OBJECT);
3961 _sp--;
3962 // Check for locking null object
3963 if (stopped()) return true;
3964
3965 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3966 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3967 int alias_idx = C->get_alias_index(adr_type);
3968
3969 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3970
3971 push_pair(result);
3972
3973 return true;
3974 }
3975
3976 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3977 // Restore the stack and pop off the arguments
3978 _sp+=5;
3979 Node *newVal = pop_pair();
3980 Node *oldVal = pop_pair();
3981 Node *obj = pop();
3982
3983 // we need the offset of the "value" field which was computed when
3984 // inlining the get() method. Give up if we don't have it.
3985 if (value_field_offset == -1)
3986 return false;
3987
3988 // Null check without removing any arguments.
3989 _sp+=5;
3990 obj = do_null_check(obj, T_OBJECT);
3991 _sp-=5;
3992 // Check for locking null object
3993 if (stopped()) return true;
3994
3995 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3996 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3997 int alias_idx = C->get_alias_index(adr_type);
3998
3999 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
4000 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
4001 set_memory(store_proj, alias_idx);
4002 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
4003
4004 Node *result;
4005 // CMove node is not used to be able fold a possible check code
4006 // after attemptUpdate() call. This code could be transformed
4007 // into CMove node by loop optimizations.
4008 {
4009 RegionNode *r = new (C, 3) RegionNode(3);
4010 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
4011
4012 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4013 Node *iftrue = opt_iff(r, iff);
4014 r->init_req(1, iftrue);
4015 result->init_req(1, intcon(1));
4016 result->init_req(2, intcon(0));
4017
4018 set_control(_gvn.transform(r));
4019 record_for_igvn(r);
4020
4021 C->set_has_split_ifs(true); // Has chance for split-if optimization
4022 }
4023
4024 push(_gvn.transform(result));
4025 return true;
4026 }
4027
4028 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4029 // restore the arguments
4030 _sp += arg_size();
4031
4032 switch (id) {
4033 case vmIntrinsics::_floatToRawIntBits:
4034 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
4035 break;
4036
4037 case vmIntrinsics::_intBitsToFloat:
4038 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
4039 break;
4040
4041 case vmIntrinsics::_doubleToRawLongBits:
4042 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
4043 break;
4044
4045 case vmIntrinsics::_longBitsToDouble:
4046 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
4047 break;
4048
4049 case vmIntrinsics::_doubleToLongBits: {
4050 Node* value = pop_pair();
4051
4052 // two paths (plus control) merge in a wood
4053 RegionNode *r = new (C, 3) RegionNode(3);
4054 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
4055
4056 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
4057 // Build the boolean node
4058 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4059
4060 // Branch either way.
4061 // NaN case is less traveled, which makes all the difference.
4062 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4063 Node *opt_isnan = _gvn.transform(ifisnan);
4064 assert( opt_isnan->is_If(), "Expect an IfNode");
4065 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4066 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4067
4068 set_control(iftrue);
4069
4070 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4071 Node *slow_result = longcon(nan_bits); // return NaN
4072 phi->init_req(1, _gvn.transform( slow_result ));
4073 r->init_req(1, iftrue);
4074
4075 // Else fall through
4076 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4077 set_control(iffalse);
4078
4079 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
4080 r->init_req(2, iffalse);
4081
4082 // Post merge
4083 set_control(_gvn.transform(r));
4084 record_for_igvn(r);
4085
4086 Node* result = _gvn.transform(phi);
4087 assert(result->bottom_type()->isa_long(), "must be");
4088 push_pair(result);
4089
4090 C->set_has_split_ifs(true); // Has chance for split-if optimization
4091
4092 break;
4093 }
4094
4095 case vmIntrinsics::_floatToIntBits: {
4096 Node* value = pop();
4097
4098 // two paths (plus control) merge in a wood
4099 RegionNode *r = new (C, 3) RegionNode(3);
4100 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
4101
4102 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
4103 // Build the boolean node
4104 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4105
4106 // Branch either way.
4107 // NaN case is less traveled, which makes all the difference.
4108 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4109 Node *opt_isnan = _gvn.transform(ifisnan);
4110 assert( opt_isnan->is_If(), "Expect an IfNode");
4111 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4112 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4113
4114 set_control(iftrue);
4115
4116 static const jint nan_bits = 0x7fc00000;
4117 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4118 phi->init_req(1, _gvn.transform( slow_result ));
4119 r->init_req(1, iftrue);
4120
4121 // Else fall through
4122 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4123 set_control(iffalse);
4124
4125 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
4126 r->init_req(2, iffalse);
4127
4128 // Post merge
4129 set_control(_gvn.transform(r));
4130 record_for_igvn(r);
4131
4132 Node* result = _gvn.transform(phi);
4133 assert(result->bottom_type()->isa_int(), "must be");
4134 push(result);
4135
4136 C->set_has_split_ifs(true); // Has chance for split-if optimization
4137
4138 break;
4139 }
4140
4141 default:
4142 ShouldNotReachHere();
4143 }
4144
4145 return true;
4146 }
4147
4148 #ifdef _LP64
4149 #define XTOP ,top() /*additional argument*/
4150 #else //_LP64
4151 #define XTOP /*no additional argument*/
4152 #endif //_LP64
4153
4154 //----------------------inline_unsafe_copyMemory-------------------------
4155 bool LibraryCallKit::inline_unsafe_copyMemory() {
4156 if (callee()->is_static()) return false; // caller must have the capability!
4157 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4158 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4159 null_check_receiver(callee()); // check then ignore argument(0)
4160 if (stopped()) return true;
4161
4162 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4163
4164 Node* src_ptr = argument(1);
4165 Node* src_off = ConvL2X(argument(2));
4166 assert(argument(3)->is_top(), "2nd half of long");
4167 Node* dst_ptr = argument(4);
4168 Node* dst_off = ConvL2X(argument(5));
4169 assert(argument(6)->is_top(), "2nd half of long");
4170 Node* size = ConvL2X(argument(7));
4171 assert(argument(8)->is_top(), "2nd half of long");
4172
4173 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4174 "fieldOffset must be byte-scaled");
4175
4176 Node* src = make_unsafe_address(src_ptr, src_off);
4177 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4178
4179 // Conservatively insert a memory barrier on all memory slices.
4180 // Do not let writes of the copy source or destination float below the copy.
4181 insert_mem_bar(Op_MemBarCPUOrder);
4182
4183 // Call it. Note that the length argument is not scaled.
4184 make_runtime_call(RC_LEAF|RC_NO_FP,
4185 OptoRuntime::fast_arraycopy_Type(),
4186 StubRoutines::unsafe_arraycopy(),
4187 "unsafe_arraycopy",
4188 TypeRawPtr::BOTTOM,
4189 src, dst, size XTOP);
4190
4191 // Do not let reads of the copy destination float above the copy.
4192 insert_mem_bar(Op_MemBarCPUOrder);
4193
4194 return true;
4195 }
4196
4197 //------------------------clone_coping-----------------------------------
4198 // Helper function for inline_native_clone.
4199 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4200 assert(obj_size != NULL, "");
4201 Node* raw_obj = alloc_obj->in(1);
4202 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4203
4204 AllocateNode* alloc = NULL;
4205 if (ReduceBulkZeroing) {
4206 // We will be completely responsible for initializing this object -
4207 // mark Initialize node as complete.
4208 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4209 // The object was just allocated - there should be no any stores!
4210 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4211 // Mark as complete_with_arraycopy so that on AllocateNode
4212 // expansion, we know this AllocateNode is initialized by an array
4213 // copy and a StoreStore barrier exists after the array copy.
4214 alloc->initialization()->set_complete_with_arraycopy();
4215 }
4216
4217 // Copy the fastest available way.
4218 // TODO: generate fields copies for small objects instead.
4219 Node* src = obj;
4220 Node* dest = alloc_obj;
4221 Node* size = _gvn.transform(obj_size);
4222
4223 // Exclude the header but include array length to copy by 8 bytes words.
4224 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4225 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4226 instanceOopDesc::base_offset_in_bytes();
4227 // base_off:
4228 // 8 - 32-bit VM
4229 // 12 - 64-bit VM, compressed oops
4230 // 16 - 64-bit VM, normal oops
4231 if (base_off % BytesPerLong != 0) {
4232 assert(UseCompressedOops, "");
4233 if (is_array) {
4234 // Exclude length to copy by 8 bytes words.
4235 base_off += sizeof(int);
4236 } else {
4237 // Include klass to copy by 8 bytes words.
4238 base_off = instanceOopDesc::klass_offset_in_bytes();
4239 }
4240 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4241 }
4242 src = basic_plus_adr(src, base_off);
4243 dest = basic_plus_adr(dest, base_off);
4244
4245 // Compute the length also, if needed:
4246 Node* countx = size;
4247 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4248 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4249
4250 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4251 bool disjoint_bases = true;
4252 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4253 src, NULL, dest, NULL, countx,
4254 /*dest_uninitialized*/true);
4255
4256 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4257 if (card_mark) {
4258 assert(!is_array, "");
4259 // Put in store barrier for any and all oops we are sticking
4260 // into this object. (We could avoid this if we could prove
4261 // that the object type contains no oop fields at all.)
4262 Node* no_particular_value = NULL;
4263 Node* no_particular_field = NULL;
4264 int raw_adr_idx = Compile::AliasIdxRaw;
4265 post_barrier(control(),
4266 memory(raw_adr_type),
4267 alloc_obj,
4268 no_particular_field,
4269 raw_adr_idx,
4270 no_particular_value,
4271 T_OBJECT,
4272 false);
4273 }
4274
4275 // Do not let reads from the cloned object float above the arraycopy.
4276 if (alloc != NULL) {
4277 // Do not let stores that initialize this object be reordered with
4278 // a subsequent store that would make this object accessible by
4279 // other threads.
4280 // Record what AllocateNode this StoreStore protects so that
4281 // escape analysis can go from the MemBarStoreStoreNode to the
4282 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4283 // based on the escape status of the AllocateNode.
4284 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4285 } else {
4286 insert_mem_bar(Op_MemBarCPUOrder);
4287 }
4288 }
4289
4290 //------------------------inline_native_clone----------------------------
4291 // Here are the simple edge cases:
4292 // null receiver => normal trap
4293 // virtual and clone was overridden => slow path to out-of-line clone
4294 // not cloneable or finalizer => slow path to out-of-line Object.clone
4295 //
4296 // The general case has two steps, allocation and copying.
4297 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4298 //
4299 // Copying also has two cases, oop arrays and everything else.
4300 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4301 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4302 //
4303 // These steps fold up nicely if and when the cloned object's klass
4304 // can be sharply typed as an object array, a type array, or an instance.
4305 //
4306 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4307 int nargs = 1;
4308 PhiNode* result_val;
4309
4310 //set the original stack and the reexecute bit for the interpreter to reexecute
4311 //the bytecode that invokes Object.clone if deoptimization happens
4312 { PreserveReexecuteState preexecs(this);
4313 jvms()->set_should_reexecute(true);
4314
4315 //null_check_receiver will adjust _sp (push and pop)
4316 Node* obj = null_check_receiver(callee());
4317 if (stopped()) return true;
4318
4319 _sp += nargs;
4320
4321 Node* obj_klass = load_object_klass(obj);
4322 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4323 const TypeOopPtr* toop = ((tklass != NULL)
4324 ? tklass->as_instance_type()
4325 : TypeInstPtr::NOTNULL);
4326
4327 // Conservatively insert a memory barrier on all memory slices.
4328 // Do not let writes into the original float below the clone.
4329 insert_mem_bar(Op_MemBarCPUOrder);
4330
4331 // paths into result_reg:
4332 enum {
4333 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4334 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4335 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4336 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4337 PATH_LIMIT
4338 };
4339 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4340 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4341 TypeInstPtr::NOTNULL);
4342 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4343 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4344 TypePtr::BOTTOM);
4345 record_for_igvn(result_reg);
4346
4347 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4348 int raw_adr_idx = Compile::AliasIdxRaw;
4349
4350 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4351 if (array_ctl != NULL) {
4352 // It's an array.
4353 PreserveJVMState pjvms(this);
4354 set_control(array_ctl);
4355 Node* obj_length = load_array_length(obj);
4356 Node* obj_size = NULL;
4357 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);
4358
4359 if (!use_ReduceInitialCardMarks()) {
4360 // If it is an oop array, it requires very special treatment,
4361 // because card marking is required on each card of the array.
4362 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4363 if (is_obja != NULL) {
4364 PreserveJVMState pjvms2(this);
4365 set_control(is_obja);
4366 // Generate a direct call to the right arraycopy function(s).
4367 bool disjoint_bases = true;
4368 bool length_never_negative = true;
4369 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4370 obj, intcon(0), alloc_obj, intcon(0),
4371 obj_length,
4372 disjoint_bases, length_never_negative);
4373 result_reg->init_req(_objArray_path, control());
4374 result_val->init_req(_objArray_path, alloc_obj);
4375 result_i_o ->set_req(_objArray_path, i_o());
4376 result_mem ->set_req(_objArray_path, reset_memory());
4377 }
4378 }
4379 // Otherwise, there are no card marks to worry about.
4380 // (We can dispense with card marks if we know the allocation
4381 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4382 // causes the non-eden paths to take compensating steps to
4383 // simulate a fresh allocation, so that no further
4384 // card marks are required in compiled code to initialize
4385 // the object.)
4386
4387 if (!stopped()) {
4388 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4389
4390 // Present the results of the copy.
4391 result_reg->init_req(_array_path, control());
4392 result_val->init_req(_array_path, alloc_obj);
4393 result_i_o ->set_req(_array_path, i_o());
4394 result_mem ->set_req(_array_path, reset_memory());
4395 }
4396 }
4397
4398 // We only go to the instance fast case code if we pass a number of guards.
4399 // The paths which do not pass are accumulated in the slow_region.
4400 RegionNode* slow_region = new (C, 1) RegionNode(1);
4401 record_for_igvn(slow_region);
4402 if (!stopped()) {
4403 // It's an instance (we did array above). Make the slow-path tests.
4404 // If this is a virtual call, we generate a funny guard. We grab
4405 // the vtable entry corresponding to clone() from the target object.
4406 // If the target method which we are calling happens to be the
4407 // Object clone() method, we pass the guard. We do not need this
4408 // guard for non-virtual calls; the caller is known to be the native
4409 // Object clone().
4410 if (is_virtual) {
4411 generate_virtual_guard(obj_klass, slow_region);
4412 }
4413
4414 // The object must be cloneable and must not have a finalizer.
4415 // Both of these conditions may be checked in a single test.
4416 // We could optimize the cloneable test further, but we don't care.
4417 generate_access_flags_guard(obj_klass,
4418 // Test both conditions:
4419 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4420 // Must be cloneable but not finalizer:
4421 JVM_ACC_IS_CLONEABLE,
4422 slow_region);
4423 }
4424
4425 if (!stopped()) {
4426 // It's an instance, and it passed the slow-path tests.
4427 PreserveJVMState pjvms(this);
4428 Node* obj_size = NULL;
4429 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4430
4431 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4432
4433 // Present the results of the slow call.
4434 result_reg->init_req(_instance_path, control());
4435 result_val->init_req(_instance_path, alloc_obj);
4436 result_i_o ->set_req(_instance_path, i_o());
4437 result_mem ->set_req(_instance_path, reset_memory());
4438 }
4439
4440 // Generate code for the slow case. We make a call to clone().
4441 set_control(_gvn.transform(slow_region));
4442 if (!stopped()) {
4443 PreserveJVMState pjvms(this);
4444 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4445 Node* slow_result = set_results_for_java_call(slow_call);
4446 // this->control() comes from set_results_for_java_call
4447 result_reg->init_req(_slow_path, control());
4448 result_val->init_req(_slow_path, slow_result);
4449 result_i_o ->set_req(_slow_path, i_o());
4450 result_mem ->set_req(_slow_path, reset_memory());
4451 }
4452
4453 // Return the combined state.
4454 set_control( _gvn.transform(result_reg) );
4455 set_i_o( _gvn.transform(result_i_o) );
4456 set_all_memory( _gvn.transform(result_mem) );
4457 } //original reexecute and sp are set back here
4458
4459 push(_gvn.transform(result_val));
4460
4461 return true;
4462 }
4463
4464 //------------------------------basictype2arraycopy----------------------------
4465 address LibraryCallKit::basictype2arraycopy(BasicType t,
4466 Node* src_offset,
4467 Node* dest_offset,
4468 bool disjoint_bases,
4469 const char* &name,
4470 bool dest_uninitialized) {
4471 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4472 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4473
4474 bool aligned = false;
4475 bool disjoint = disjoint_bases;
4476
4477 // if the offsets are the same, we can treat the memory regions as
4478 // disjoint, because either the memory regions are in different arrays,
4479 // or they are identical (which we can treat as disjoint.) We can also
4480 // treat a copy with a destination index less that the source index
4481 // as disjoint since a low->high copy will work correctly in this case.
4482 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4483 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4484 // both indices are constants
4485 int s_offs = src_offset_inttype->get_con();
4486 int d_offs = dest_offset_inttype->get_con();
4487 int element_size = type2aelembytes(t);
4488 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4489 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4490 if (s_offs >= d_offs) disjoint = true;
4491 } else if (src_offset == dest_offset && src_offset != NULL) {
4492 // This can occur if the offsets are identical non-constants.
4493 disjoint = true;
4494 }
4495
4496 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4497 }
4498
4499
4500 //------------------------------inline_arraycopy-----------------------
4501 bool LibraryCallKit::inline_arraycopy() {
4502 // Restore the stack and pop off the arguments.
4503 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4504 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4505
4506 Node *src = argument(0);
4507 Node *src_offset = argument(1);
4508 Node *dest = argument(2);
4509 Node *dest_offset = argument(3);
4510 Node *length = argument(4);
4511
4512 // Compile time checks. If any of these checks cannot be verified at compile time,
4513 // we do not make a fast path for this call. Instead, we let the call remain as it
4514 // is. The checks we choose to mandate at compile time are:
4515 //
4516 // (1) src and dest are arrays.
4517 const Type* src_type = src->Value(&_gvn);
4518 const Type* dest_type = dest->Value(&_gvn);
4519 const TypeAryPtr* top_src = src_type->isa_aryptr();
4520 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4521 if (top_src == NULL || top_src->klass() == NULL ||
4522 top_dest == NULL || top_dest->klass() == NULL) {
4523 // Conservatively insert a memory barrier on all memory slices.
4524 // Do not let writes into the source float below the arraycopy.
4525 insert_mem_bar(Op_MemBarCPUOrder);
4526
4527 // Call StubRoutines::generic_arraycopy stub.
4528 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4529 src, src_offset, dest, dest_offset, length);
4530
4531 // Do not let reads from the destination float above the arraycopy.
4532 // Since we cannot type the arrays, we don't know which slices
4533 // might be affected. We could restrict this barrier only to those
4534 // memory slices which pertain to array elements--but don't bother.
4535 if (!InsertMemBarAfterArraycopy)
4536 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4537 insert_mem_bar(Op_MemBarCPUOrder);
4538 return true;
4539 }
4540
4541 // (2) src and dest arrays must have elements of the same BasicType
4542 // Figure out the size and type of the elements we will be copying.
4543 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4544 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4545 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4546 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4547
4548 if (src_elem != dest_elem || dest_elem == T_VOID) {
4549 // The component types are not the same or are not recognized. Punt.
4550 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4551 generate_slow_arraycopy(TypePtr::BOTTOM,
4552 src, src_offset, dest, dest_offset, length,
4553 /*dest_uninitialized*/false);
4554 return true;
4555 }
4556
4557 //---------------------------------------------------------------------------
4558 // We will make a fast path for this call to arraycopy.
4559
4560 // We have the following tests left to perform:
4561 //
4562 // (3) src and dest must not be null.
4563 // (4) src_offset must not be negative.
4564 // (5) dest_offset must not be negative.
4565 // (6) length must not be negative.
4566 // (7) src_offset + length must not exceed length of src.
4567 // (8) dest_offset + length must not exceed length of dest.
4568 // (9) each element of an oop array must be assignable
4569
4570 RegionNode* slow_region = new (C, 1) RegionNode(1);
4571 record_for_igvn(slow_region);
4572
4573 // (3) operands must not be null
4574 // We currently perform our null checks with the do_null_check routine.
4575 // This means that the null exceptions will be reported in the caller
4576 // rather than (correctly) reported inside of the native arraycopy call.
4577 // This should be corrected, given time. We do our null check with the
4578 // stack pointer restored.
4579 _sp += nargs;
4580 src = do_null_check(src, T_ARRAY);
4581 dest = do_null_check(dest, T_ARRAY);
4582 _sp -= nargs;
4583
4584 // (4) src_offset must not be negative.
4585 generate_negative_guard(src_offset, slow_region);
4586
4587 // (5) dest_offset must not be negative.
4588 generate_negative_guard(dest_offset, slow_region);
4589
4590 // (6) length must not be negative (moved to generate_arraycopy()).
4591 // generate_negative_guard(length, slow_region);
4592
4593 // (7) src_offset + length must not exceed length of src.
4594 generate_limit_guard(src_offset, length,
4595 load_array_length(src),
4596 slow_region);
4597
4598 // (8) dest_offset + length must not exceed length of dest.
4599 generate_limit_guard(dest_offset, length,
4600 load_array_length(dest),
4601 slow_region);
4602
4603 // (9) each element of an oop array must be assignable
4604 // The generate_arraycopy subroutine checks this.
4605
4606 // This is where the memory effects are placed:
4607 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4608 generate_arraycopy(adr_type, dest_elem,
4609 src, src_offset, dest, dest_offset, length,
4610 false, false, slow_region);
4611
4612 return true;
4613 }
4614
4615 //-----------------------------generate_arraycopy----------------------
4616 // Generate an optimized call to arraycopy.
4617 // Caller must guard against non-arrays.
4618 // Caller must determine a common array basic-type for both arrays.
4619 // Caller must validate offsets against array bounds.
4620 // The slow_region has already collected guard failure paths
4621 // (such as out of bounds length or non-conformable array types).
4622 // The generated code has this shape, in general:
4623 //
4624 // if (length == 0) return // via zero_path
4625 // slowval = -1
4626 // if (types unknown) {
4627 // slowval = call generic copy loop
4628 // if (slowval == 0) return // via checked_path
4629 // } else if (indexes in bounds) {
4630 // if ((is object array) && !(array type check)) {
4631 // slowval = call checked copy loop
4632 // if (slowval == 0) return // via checked_path
4633 // } else {
4634 // call bulk copy loop
4635 // return // via fast_path
4636 // }
4637 // }
4638 // // adjust params for remaining work:
4639 // if (slowval != -1) {
4640 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4641 // }
4642 // slow_region:
4643 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4644 // return // via slow_call_path
4645 //
4646 // This routine is used from several intrinsics: System.arraycopy,
4647 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4648 //
4649 void
4650 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4651 BasicType basic_elem_type,
4652 Node* src, Node* src_offset,
4653 Node* dest, Node* dest_offset,
4654 Node* copy_length,
4655 bool disjoint_bases,
4656 bool length_never_negative,
4657 RegionNode* slow_region) {
4658
4659 if (slow_region == NULL) {
4660 slow_region = new(C,1) RegionNode(1);
4661 record_for_igvn(slow_region);
4662 }
4663
4664 Node* original_dest = dest;
4665 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4666 bool dest_uninitialized = false;
4667
4668 // See if this is the initialization of a newly-allocated array.
4669 // If so, we will take responsibility here for initializing it to zero.
4670 // (Note: Because tightly_coupled_allocation performs checks on the
4671 // out-edges of the dest, we need to avoid making derived pointers
4672 // from it until we have checked its uses.)
4673 if (ReduceBulkZeroing
4674 && !ZeroTLAB // pointless if already zeroed
4675 && basic_elem_type != T_CONFLICT // avoid corner case
4676 && !src->eqv_uncast(dest)
4677 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4678 != NULL)
4679 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4680 && alloc->maybe_set_complete(&_gvn)) {
4681 // "You break it, you buy it."
4682 InitializeNode* init = alloc->initialization();
4683 assert(init->is_complete(), "we just did this");
4684 init->set_complete_with_arraycopy();
4685 assert(dest->is_CheckCastPP(), "sanity");
4686 assert(dest->in(0)->in(0) == init, "dest pinned");
4687 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4688 // From this point on, every exit path is responsible for
4689 // initializing any non-copied parts of the object to zero.
4690 // Also, if this flag is set we make sure that arraycopy interacts properly
4691 // with G1, eliding pre-barriers. See CR 6627983.
4692 dest_uninitialized = true;
4693 } else {
4694 // No zeroing elimination here.
4695 alloc = NULL;
4696 //original_dest = dest;
4697 //dest_uninitialized = false;
4698 }
4699
4700 // Results are placed here:
4701 enum { fast_path = 1, // normal void-returning assembly stub
4702 checked_path = 2, // special assembly stub with cleanup
4703 slow_call_path = 3, // something went wrong; call the VM
4704 zero_path = 4, // bypass when length of copy is zero
4705 bcopy_path = 5, // copy primitive array by 64-bit blocks
4706 PATH_LIMIT = 6
4707 };
4708 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4709 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4710 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4711 record_for_igvn(result_region);
4712 _gvn.set_type_bottom(result_i_o);
4713 _gvn.set_type_bottom(result_memory);
4714 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4715
4716 // The slow_control path:
4717 Node* slow_control;
4718 Node* slow_i_o = i_o();
4719 Node* slow_mem = memory(adr_type);
4720 debug_only(slow_control = (Node*) badAddress);
4721
4722 // Checked control path:
4723 Node* checked_control = top();
4724 Node* checked_mem = NULL;
4725 Node* checked_i_o = NULL;
4726 Node* checked_value = NULL;
4727
4728 if (basic_elem_type == T_CONFLICT) {
4729 assert(!dest_uninitialized, "");
4730 Node* cv = generate_generic_arraycopy(adr_type,
4731 src, src_offset, dest, dest_offset,
4732 copy_length, dest_uninitialized);
4733 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4734 checked_control = control();
4735 checked_i_o = i_o();
4736 checked_mem = memory(adr_type);
4737 checked_value = cv;
4738 set_control(top()); // no fast path
4739 }
4740
4741 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4742 if (not_pos != NULL) {
4743 PreserveJVMState pjvms(this);
4744 set_control(not_pos);
4745
4746 // (6) length must not be negative.
4747 if (!length_never_negative) {
4748 generate_negative_guard(copy_length, slow_region);
4749 }
4750
4751 // copy_length is 0.
4752 if (!stopped() && dest_uninitialized) {
4753 Node* dest_length = alloc->in(AllocateNode::ALength);
4754 if (copy_length->eqv_uncast(dest_length)
4755 || _gvn.find_int_con(dest_length, 1) <= 0) {
4756 // There is no zeroing to do. No need for a secondary raw memory barrier.
4757 } else {
4758 // Clear the whole thing since there are no source elements to copy.
4759 generate_clear_array(adr_type, dest, basic_elem_type,
4760 intcon(0), NULL,
4761 alloc->in(AllocateNode::AllocSize));
4762 // Use a secondary InitializeNode as raw memory barrier.
4763 // Currently it is needed only on this path since other
4764 // paths have stub or runtime calls as raw memory barriers.
4765 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4766 Compile::AliasIdxRaw,
4767 top())->as_Initialize();
4768 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4769 }
4770 }
4771
4772 // Present the results of the fast call.
4773 result_region->init_req(zero_path, control());
4774 result_i_o ->init_req(zero_path, i_o());
4775 result_memory->init_req(zero_path, memory(adr_type));
4776 }
4777
4778 if (!stopped() && dest_uninitialized) {
4779 // We have to initialize the *uncopied* part of the array to zero.
4780 // The copy destination is the slice dest[off..off+len]. The other slices
4781 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4782 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4783 Node* dest_length = alloc->in(AllocateNode::ALength);
4784 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4785 copy_length) );
4786
4787 // If there is a head section that needs zeroing, do it now.
4788 if (find_int_con(dest_offset, -1) != 0) {
4789 generate_clear_array(adr_type, dest, basic_elem_type,
4790 intcon(0), dest_offset,
4791 NULL);
4792 }
4793
4794 // Next, perform a dynamic check on the tail length.
4795 // It is often zero, and we can win big if we prove this.
4796 // There are two wins: Avoid generating the ClearArray
4797 // with its attendant messy index arithmetic, and upgrade
4798 // the copy to a more hardware-friendly word size of 64 bits.
4799 Node* tail_ctl = NULL;
4800 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4801 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4802 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4803 tail_ctl = generate_slow_guard(bol_lt, NULL);
4804 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4805 }
4806
4807 // At this point, let's assume there is no tail.
4808 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4809 // There is no tail. Try an upgrade to a 64-bit copy.
4810 bool didit = false;
4811 { PreserveJVMState pjvms(this);
4812 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4813 src, src_offset, dest, dest_offset,
4814 dest_size, dest_uninitialized);
4815 if (didit) {
4816 // Present the results of the block-copying fast call.
4817 result_region->init_req(bcopy_path, control());
4818 result_i_o ->init_req(bcopy_path, i_o());
4819 result_memory->init_req(bcopy_path, memory(adr_type));
4820 }
4821 }
4822 if (didit)
4823 set_control(top()); // no regular fast path
4824 }
4825
4826 // Clear the tail, if any.
4827 if (tail_ctl != NULL) {
4828 Node* notail_ctl = stopped() ? NULL : control();
4829 set_control(tail_ctl);
4830 if (notail_ctl == NULL) {
4831 generate_clear_array(adr_type, dest, basic_elem_type,
4832 dest_tail, NULL,
4833 dest_size);
4834 } else {
4835 // Make a local merge.
4836 Node* done_ctl = new(C,3) RegionNode(3);
4837 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4838 done_ctl->init_req(1, notail_ctl);
4839 done_mem->init_req(1, memory(adr_type));
4840 generate_clear_array(adr_type, dest, basic_elem_type,
4841 dest_tail, NULL,
4842 dest_size);
4843 done_ctl->init_req(2, control());
4844 done_mem->init_req(2, memory(adr_type));
4845 set_control( _gvn.transform(done_ctl) );
4846 set_memory( _gvn.transform(done_mem), adr_type );
4847 }
4848 }
4849 }
4850
4851 BasicType copy_type = basic_elem_type;
4852 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4853 if (!stopped() && copy_type == T_OBJECT) {
4854 // If src and dest have compatible element types, we can copy bits.
4855 // Types S[] and D[] are compatible if D is a supertype of S.
4856 //
4857 // If they are not, we will use checked_oop_disjoint_arraycopy,
4858 // which performs a fast optimistic per-oop check, and backs off
4859 // further to JVM_ArrayCopy on the first per-oop check that fails.
4860 // (Actually, we don't move raw bits only; the GC requires card marks.)
4861
4862 // Get the klassOop for both src and dest
4863 Node* src_klass = load_object_klass(src);
4864 Node* dest_klass = load_object_klass(dest);
4865
4866 // Generate the subtype check.
4867 // This might fold up statically, or then again it might not.
4868 //
4869 // Non-static example: Copying List<String>.elements to a new String[].
4870 // The backing store for a List<String> is always an Object[],
4871 // but its elements are always type String, if the generic types
4872 // are correct at the source level.
4873 //
4874 // Test S[] against D[], not S against D, because (probably)
4875 // the secondary supertype cache is less busy for S[] than S.
4876 // This usually only matters when D is an interface.
4877 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4878 // Plug failing path into checked_oop_disjoint_arraycopy
4879 if (not_subtype_ctrl != top()) {
4880 PreserveJVMState pjvms(this);
4881 set_control(not_subtype_ctrl);
4882 // (At this point we can assume disjoint_bases, since types differ.)
4883 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
4884 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4885 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4886 Node* dest_elem_klass = _gvn.transform(n1);
4887 Node* cv = generate_checkcast_arraycopy(adr_type,
4888 dest_elem_klass,
4889 src, src_offset, dest, dest_offset,
4890 ConvI2X(copy_length), dest_uninitialized);
4891 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4892 checked_control = control();
4893 checked_i_o = i_o();
4894 checked_mem = memory(adr_type);
4895 checked_value = cv;
4896 }
4897 // At this point we know we do not need type checks on oop stores.
4898
4899 // Let's see if we need card marks:
4900 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4901 // If we do not need card marks, copy using the jint or jlong stub.
4902 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4903 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4904 "sizes agree");
4905 }
4906 }
4907
4908 if (!stopped()) {
4909 // Generate the fast path, if possible.
4910 PreserveJVMState pjvms(this);
4911 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4912 src, src_offset, dest, dest_offset,
4913 ConvI2X(copy_length), dest_uninitialized);
4914
4915 // Present the results of the fast call.
4916 result_region->init_req(fast_path, control());
4917 result_i_o ->init_req(fast_path, i_o());
4918 result_memory->init_req(fast_path, memory(adr_type));
4919 }
4920
4921 // Here are all the slow paths up to this point, in one bundle:
4922 slow_control = top();
4923 if (slow_region != NULL)
4924 slow_control = _gvn.transform(slow_region);
4925 debug_only(slow_region = (RegionNode*)badAddress);
4926
4927 set_control(checked_control);
4928 if (!stopped()) {
4929 // Clean up after the checked call.
4930 // The returned value is either 0 or -1^K,
4931 // where K = number of partially transferred array elements.
4932 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4933 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4934 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4935
4936 // If it is 0, we are done, so transfer to the end.
4937 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4938 result_region->init_req(checked_path, checks_done);
4939 result_i_o ->init_req(checked_path, checked_i_o);
4940 result_memory->init_req(checked_path, checked_mem);
4941
4942 // If it is not zero, merge into the slow call.
4943 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4944 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4945 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4946 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4947 record_for_igvn(slow_reg2);
4948 slow_reg2 ->init_req(1, slow_control);
4949 slow_i_o2 ->init_req(1, slow_i_o);
4950 slow_mem2 ->init_req(1, slow_mem);
4951 slow_reg2 ->init_req(2, control());
4952 slow_i_o2 ->init_req(2, checked_i_o);
4953 slow_mem2 ->init_req(2, checked_mem);
4954
4955 slow_control = _gvn.transform(slow_reg2);
4956 slow_i_o = _gvn.transform(slow_i_o2);
4957 slow_mem = _gvn.transform(slow_mem2);
4958
4959 if (alloc != NULL) {
4960 // We'll restart from the very beginning, after zeroing the whole thing.
4961 // This can cause double writes, but that's OK since dest is brand new.
4962 // So we ignore the low 31 bits of the value returned from the stub.
4963 } else {
4964 // We must continue the copy exactly where it failed, or else
4965 // another thread might see the wrong number of writes to dest.
4966 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4967 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4968 slow_offset->init_req(1, intcon(0));
4969 slow_offset->init_req(2, checked_offset);
4970 slow_offset = _gvn.transform(slow_offset);
4971
4972 // Adjust the arguments by the conditionally incoming offset.
4973 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4974 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4975 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4976
4977 // Tweak the node variables to adjust the code produced below:
4978 src_offset = src_off_plus;
4979 dest_offset = dest_off_plus;
4980 copy_length = length_minus;
4981 }
4982 }
4983
4984 set_control(slow_control);
4985 if (!stopped()) {
4986 // Generate the slow path, if needed.
4987 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4988
4989 set_memory(slow_mem, adr_type);
4990 set_i_o(slow_i_o);
4991
4992 if (dest_uninitialized) {
4993 generate_clear_array(adr_type, dest, basic_elem_type,
4994 intcon(0), NULL,
4995 alloc->in(AllocateNode::AllocSize));
4996 }
4997
4998 generate_slow_arraycopy(adr_type,
4999 src, src_offset, dest, dest_offset,
5000 copy_length, /*dest_uninitialized*/false);
5001
5002 result_region->init_req(slow_call_path, control());
5003 result_i_o ->init_req(slow_call_path, i_o());
5004 result_memory->init_req(slow_call_path, memory(adr_type));
5005 }
5006
5007 // Remove unused edges.
5008 for (uint i = 1; i < result_region->req(); i++) {
5009 if (result_region->in(i) == NULL)
5010 result_region->init_req(i, top());
5011 }
5012
5013 // Finished; return the combined state.
5014 set_control( _gvn.transform(result_region) );
5015 set_i_o( _gvn.transform(result_i_o) );
5016 set_memory( _gvn.transform(result_memory), adr_type );
5017
5018 // The memory edges above are precise in order to model effects around
5019 // array copies accurately to allow value numbering of field loads around
5020 // arraycopy. Such field loads, both before and after, are common in Java
5021 // collections and similar classes involving header/array data structures.
5022 //
5023 // But with low number of register or when some registers are used or killed
5024 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5025 // The next memory barrier is added to avoid it. If the arraycopy can be
5026 // optimized away (which it can, sometimes) then we can manually remove
5027 // the membar also.
5028 //
5029 // Do not let reads from the cloned object float above the arraycopy.
5030 if (alloc != NULL) {
5031 // Do not let stores that initialize this object be reordered with
5032 // a subsequent store that would make this object accessible by
5033 // other threads.
5034 // Record what AllocateNode this StoreStore protects so that
5035 // escape analysis can go from the MemBarStoreStoreNode to the
5036 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5037 // based on the escape status of the AllocateNode.
5038 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5039 } else if (InsertMemBarAfterArraycopy)
5040 insert_mem_bar(Op_MemBarCPUOrder);
5041 }
5042
5043
5044 // Helper function which determines if an arraycopy immediately follows
5045 // an allocation, with no intervening tests or other escapes for the object.
5046 AllocateArrayNode*
5047 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5048 RegionNode* slow_region) {
5049 if (stopped()) return NULL; // no fast path
5050 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5051
5052 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5053 if (alloc == NULL) return NULL;
5054
5055 Node* rawmem = memory(Compile::AliasIdxRaw);
5056 // Is the allocation's memory state untouched?
5057 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5058 // Bail out if there have been raw-memory effects since the allocation.
5059 // (Example: There might have been a call or safepoint.)
5060 return NULL;
5061 }
5062 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5063 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5064 return NULL;
5065 }
5066
5067 // There must be no unexpected observers of this allocation.
5068 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5069 Node* obs = ptr->fast_out(i);
5070 if (obs != this->map()) {
5071 return NULL;
5072 }
5073 }
5074
5075 // This arraycopy must unconditionally follow the allocation of the ptr.
5076 Node* alloc_ctl = ptr->in(0);
5077 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5078
5079 Node* ctl = control();
5080 while (ctl != alloc_ctl) {
5081 // There may be guards which feed into the slow_region.
5082 // Any other control flow means that we might not get a chance
5083 // to finish initializing the allocated object.
5084 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5085 IfNode* iff = ctl->in(0)->as_If();
5086 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5087 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5088 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5089 ctl = iff->in(0); // This test feeds the known slow_region.
5090 continue;
5091 }
5092 // One more try: Various low-level checks bottom out in
5093 // uncommon traps. If the debug-info of the trap omits
5094 // any reference to the allocation, as we've already
5095 // observed, then there can be no objection to the trap.
5096 bool found_trap = false;
5097 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5098 Node* obs = not_ctl->fast_out(j);
5099 if (obs->in(0) == not_ctl && obs->is_Call() &&
5100 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5101 found_trap = true; break;
5102 }
5103 }
5104 if (found_trap) {
5105 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5106 continue;
5107 }
5108 }
5109 return NULL;
5110 }
5111
5112 // If we get this far, we have an allocation which immediately
5113 // precedes the arraycopy, and we can take over zeroing the new object.
5114 // The arraycopy will finish the initialization, and provide
5115 // a new control state to which we will anchor the destination pointer.
5116
5117 return alloc;
5118 }
5119
5120 // Helper for initialization of arrays, creating a ClearArray.
5121 // It writes zero bits in [start..end), within the body of an array object.
5122 // The memory effects are all chained onto the 'adr_type' alias category.
5123 //
5124 // Since the object is otherwise uninitialized, we are free
5125 // to put a little "slop" around the edges of the cleared area,
5126 // as long as it does not go back into the array's header,
5127 // or beyond the array end within the heap.
5128 //
5129 // The lower edge can be rounded down to the nearest jint and the
5130 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5131 //
5132 // Arguments:
5133 // adr_type memory slice where writes are generated
5134 // dest oop of the destination array
5135 // basic_elem_type element type of the destination
5136 // slice_idx array index of first element to store
5137 // slice_len number of elements to store (or NULL)
5138 // dest_size total size in bytes of the array object
5139 //
5140 // Exactly one of slice_len or dest_size must be non-NULL.
5141 // If dest_size is non-NULL, zeroing extends to the end of the object.
5142 // If slice_len is non-NULL, the slice_idx value must be a constant.
5143 void
5144 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5145 Node* dest,
5146 BasicType basic_elem_type,
5147 Node* slice_idx,
5148 Node* slice_len,
5149 Node* dest_size) {
5150 // one or the other but not both of slice_len and dest_size:
5151 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5152 if (slice_len == NULL) slice_len = top();
5153 if (dest_size == NULL) dest_size = top();
5154
5155 // operate on this memory slice:
5156 Node* mem = memory(adr_type); // memory slice to operate on
5157
5158 // scaling and rounding of indexes:
5159 int scale = exact_log2(type2aelembytes(basic_elem_type));
5160 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5161 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5162 int bump_bit = (-1 << scale) & BytesPerInt;
5163
5164 // determine constant starts and ends
5165 const intptr_t BIG_NEG = -128;
5166 assert(BIG_NEG + 2*abase < 0, "neg enough");
5167 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5168 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5169 if (slice_len_con == 0) {
5170 return; // nothing to do here
5171 }
5172 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5173 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5174 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5175 assert(end_con < 0, "not two cons");
5176 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5177 BytesPerLong);
5178 }
5179
5180 if (start_con >= 0 && end_con >= 0) {
5181 // Constant start and end. Simple.
5182 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5183 start_con, end_con, &_gvn);
5184 } else if (start_con >= 0 && dest_size != top()) {
5185 // Constant start, pre-rounded end after the tail of the array.
5186 Node* end = dest_size;
5187 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5188 start_con, end, &_gvn);
5189 } else if (start_con >= 0 && slice_len != top()) {
5190 // Constant start, non-constant end. End needs rounding up.
5191 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5192 intptr_t end_base = abase + (slice_idx_con << scale);
5193 int end_round = (-1 << scale) & (BytesPerLong - 1);
5194 Node* end = ConvI2X(slice_len);
5195 if (scale != 0)
5196 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5197 end_base += end_round;
5198 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5199 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5200 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5201 start_con, end, &_gvn);
5202 } else if (start_con < 0 && dest_size != top()) {
5203 // Non-constant start, pre-rounded end after the tail of the array.
5204 // This is almost certainly a "round-to-end" operation.
5205 Node* start = slice_idx;
5206 start = ConvI2X(start);
5207 if (scale != 0)
5208 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5209 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5210 if ((bump_bit | clear_low) != 0) {
5211 int to_clear = (bump_bit | clear_low);
5212 // Align up mod 8, then store a jint zero unconditionally
5213 // just before the mod-8 boundary.
5214 if (((abase + bump_bit) & ~to_clear) - bump_bit
5215 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5216 bump_bit = 0;
5217 assert((abase & to_clear) == 0, "array base must be long-aligned");
5218 } else {
5219 // Bump 'start' up to (or past) the next jint boundary:
5220 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5221 assert((abase & clear_low) == 0, "array base must be int-aligned");
5222 }
5223 // Round bumped 'start' down to jlong boundary in body of array.
5224 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5225 if (bump_bit != 0) {
5226 // Store a zero to the immediately preceding jint:
5227 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5228 Node* p1 = basic_plus_adr(dest, x1);
5229 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5230 mem = _gvn.transform(mem);
5231 }
5232 }
5233 Node* end = dest_size; // pre-rounded
5234 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5235 start, end, &_gvn);
5236 } else {
5237 // Non-constant start, unrounded non-constant end.
5238 // (Nobody zeroes a random midsection of an array using this routine.)
5239 ShouldNotReachHere(); // fix caller
5240 }
5241
5242 // Done.
5243 set_memory(mem, adr_type);
5244 }
5245
5246
5247 bool
5248 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5249 BasicType basic_elem_type,
5250 AllocateNode* alloc,
5251 Node* src, Node* src_offset,
5252 Node* dest, Node* dest_offset,
5253 Node* dest_size, bool dest_uninitialized) {
5254 // See if there is an advantage from block transfer.
5255 int scale = exact_log2(type2aelembytes(basic_elem_type));
5256 if (scale >= LogBytesPerLong)
5257 return false; // it is already a block transfer
5258
5259 // Look at the alignment of the starting offsets.
5260 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5261
5262 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5263 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5264 if (src_off_con < 0 || dest_off_con < 0)
5265 // At present, we can only understand constants.
5266 return false;
5267
5268 intptr_t src_off = abase + (src_off_con << scale);
5269 intptr_t dest_off = abase + (dest_off_con << scale);
5270
5271 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5272 // Non-aligned; too bad.
5273 // One more chance: Pick off an initial 32-bit word.
5274 // This is a common case, since abase can be odd mod 8.
5275 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5276 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5277 Node* sptr = basic_plus_adr(src, src_off);
5278 Node* dptr = basic_plus_adr(dest, dest_off);
5279 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5280 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5281 src_off += BytesPerInt;
5282 dest_off += BytesPerInt;
5283 } else {
5284 return false;
5285 }
5286 }
5287 assert(src_off % BytesPerLong == 0, "");
5288 assert(dest_off % BytesPerLong == 0, "");
5289
5290 // Do this copy by giant steps.
5291 Node* sptr = basic_plus_adr(src, src_off);
5292 Node* dptr = basic_plus_adr(dest, dest_off);
5293 Node* countx = dest_size;
5294 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5295 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5296
5297 bool disjoint_bases = true; // since alloc != NULL
5298 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5299 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5300
5301 return true;
5302 }
5303
5304
5305 // Helper function; generates code for the slow case.
5306 // We make a call to a runtime method which emulates the native method,
5307 // but without the native wrapper overhead.
5308 void
5309 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5310 Node* src, Node* src_offset,
5311 Node* dest, Node* dest_offset,
5312 Node* copy_length, bool dest_uninitialized) {
5313 assert(!dest_uninitialized, "Invariant");
5314 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5315 OptoRuntime::slow_arraycopy_Type(),
5316 OptoRuntime::slow_arraycopy_Java(),
5317 "slow_arraycopy", adr_type,
5318 src, src_offset, dest, dest_offset,
5319 copy_length);
5320
5321 // Handle exceptions thrown by this fellow:
5322 make_slow_call_ex(call, env()->Throwable_klass(), false);
5323 }
5324
5325 // Helper function; generates code for cases requiring runtime checks.
5326 Node*
5327 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5328 Node* dest_elem_klass,
5329 Node* src, Node* src_offset,
5330 Node* dest, Node* dest_offset,
5331 Node* copy_length, bool dest_uninitialized) {
5332 if (stopped()) return NULL;
5333
5334 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5335 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5336 return NULL;
5337 }
5338
5339 // Pick out the parameters required to perform a store-check
5340 // for the target array. This is an optimistic check. It will
5341 // look in each non-null element's class, at the desired klass's
5342 // super_check_offset, for the desired klass.
5343 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5344 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5345 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5346 Node* check_offset = ConvI2X(_gvn.transform(n3));
5347 Node* check_value = dest_elem_klass;
5348
5349 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5350 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5351
5352 // (We know the arrays are never conjoint, because their types differ.)
5353 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5354 OptoRuntime::checkcast_arraycopy_Type(),
5355 copyfunc_addr, "checkcast_arraycopy", adr_type,
5356 // five arguments, of which two are
5357 // intptr_t (jlong in LP64)
5358 src_start, dest_start,
5359 copy_length XTOP,
5360 check_offset XTOP,
5361 check_value);
5362
5363 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5364 }
5365
5366
5367 // Helper function; generates code for cases requiring runtime checks.
5368 Node*
5369 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5370 Node* src, Node* src_offset,
5371 Node* dest, Node* dest_offset,
5372 Node* copy_length, bool dest_uninitialized) {
5373 assert(!dest_uninitialized, "Invariant");
5374 if (stopped()) return NULL;
5375 address copyfunc_addr = StubRoutines::generic_arraycopy();
5376 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5377 return NULL;
5378 }
5379
5380 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5381 OptoRuntime::generic_arraycopy_Type(),
5382 copyfunc_addr, "generic_arraycopy", adr_type,
5383 src, src_offset, dest, dest_offset, copy_length);
5384
5385 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5386 }
5387
5388 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5389 void
5390 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5391 BasicType basic_elem_type,
5392 bool disjoint_bases,
5393 Node* src, Node* src_offset,
5394 Node* dest, Node* dest_offset,
5395 Node* copy_length, bool dest_uninitialized) {
5396 if (stopped()) return; // nothing to do
5397
5398 Node* src_start = src;
5399 Node* dest_start = dest;
5400 if (src_offset != NULL || dest_offset != NULL) {
5401 assert(src_offset != NULL && dest_offset != NULL, "");
5402 src_start = array_element_address(src, src_offset, basic_elem_type);
5403 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5404 }
5405
5406 // Figure out which arraycopy runtime method to call.
5407 const char* copyfunc_name = "arraycopy";
5408 address copyfunc_addr =
5409 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5410 disjoint_bases, copyfunc_name, dest_uninitialized);
5411
5412 // Call it. Note that the count_ix value is not scaled to a byte-size.
5413 make_runtime_call(RC_LEAF|RC_NO_FP,
5414 OptoRuntime::fast_arraycopy_Type(),
5415 copyfunc_addr, copyfunc_name, adr_type,
5416 src_start, dest_start, copy_length XTOP);
5417 }
5418
5419 //----------------------------inline_reference_get----------------------------
5420
5421 bool LibraryCallKit::inline_reference_get() {
5422 const int nargs = 1; // self
5423
5424 guarantee(java_lang_ref_Reference::referent_offset > 0,
5425 "should have already been set");
5426
5427 int referent_offset = java_lang_ref_Reference::referent_offset;
5428
5429 // Restore the stack and pop off the argument
5430 _sp += nargs;
5431 Node *reference_obj = pop();
5432
5433 // Null check on self without removing any arguments.
5434 _sp += nargs;
5435 reference_obj = do_null_check(reference_obj, T_OBJECT);
5436 _sp -= nargs;;
5437
5438 if (stopped()) return true;
5439
5440 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5441
5442 ciInstanceKlass* klass = env()->Object_klass();
5443 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5444
5445 Node* no_ctrl = NULL;
5446 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5447
5448 // Use the pre-barrier to record the value in the referent field
5449 pre_barrier(false /* do_load */,
5450 control(),
5451 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5452 result /* pre_val */,
5453 T_OBJECT);
5454
5455 push(result);
5456 return true;
5457 }
5458