Thu, 29 Apr 2010 17:53:58 -0700
6948602: Disable use of SSE4.2 in String.indexOf intrinsic until 6942326 is fixed
Summary: Disable the use of pcmpestri until it can be guaranteed that the load doesn't cross in to the uncommited space. See 6942326.
Reviewed-by: never, kvn
1 /*
2 * Copyright 1999-2010 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_library_call.cpp.incl"
28 class LibraryIntrinsic : public InlineCallGenerator {
29 // Extend the set of intrinsics known to the runtime:
30 public:
31 private:
32 bool _is_virtual;
33 vmIntrinsics::ID _intrinsic_id;
35 public:
36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
37 : InlineCallGenerator(m),
38 _is_virtual(is_virtual),
39 _intrinsic_id(id)
40 {
41 }
42 virtual bool is_intrinsic() const { return true; }
43 virtual bool is_virtual() const { return _is_virtual; }
44 virtual JVMState* generate(JVMState* jvms);
45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
46 };
49 // Local helper class for LibraryIntrinsic:
50 class LibraryCallKit : public GraphKit {
51 private:
52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
54 public:
55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
56 : GraphKit(caller),
57 _intrinsic(intrinsic)
58 {
59 }
61 ciMethod* caller() const { return jvms()->method(); }
62 int bci() const { return jvms()->bci(); }
63 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
65 ciMethod* callee() const { return _intrinsic->method(); }
66 ciSignature* signature() const { return callee()->signature(); }
67 int arg_size() const { return callee()->arg_size(); }
69 bool try_to_inline();
71 // Helper functions to inline natives
72 void push_result(RegionNode* region, PhiNode* value);
73 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
74 Node* generate_slow_guard(Node* test, RegionNode* region);
75 Node* generate_fair_guard(Node* test, RegionNode* region);
76 Node* generate_negative_guard(Node* index, RegionNode* region,
77 // resulting CastII of index:
78 Node* *pos_index = NULL);
79 Node* generate_nonpositive_guard(Node* index, bool never_negative,
80 // resulting CastII of index:
81 Node* *pos_index = NULL);
82 Node* generate_limit_guard(Node* offset, Node* subseq_length,
83 Node* array_length,
84 RegionNode* region);
85 Node* generate_current_thread(Node* &tls_output);
86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
87 bool disjoint_bases, const char* &name);
88 Node* load_mirror_from_klass(Node* klass);
89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
90 int nargs,
91 RegionNode* region, int null_path,
92 int offset);
93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
94 RegionNode* region, int null_path) {
95 int offset = java_lang_Class::klass_offset_in_bytes();
96 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
97 region, null_path,
98 offset);
99 }
100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
101 int nargs,
102 RegionNode* region, int null_path) {
103 int offset = java_lang_Class::array_klass_offset_in_bytes();
104 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
105 region, null_path,
106 offset);
107 }
108 Node* generate_access_flags_guard(Node* kls,
109 int modifier_mask, int modifier_bits,
110 RegionNode* region);
111 Node* generate_interface_guard(Node* kls, RegionNode* region);
112 Node* generate_array_guard(Node* kls, RegionNode* region) {
113 return generate_array_guard_common(kls, region, false, false);
114 }
115 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
116 return generate_array_guard_common(kls, region, false, true);
117 }
118 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
119 return generate_array_guard_common(kls, region, true, false);
120 }
121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
122 return generate_array_guard_common(kls, region, true, true);
123 }
124 Node* generate_array_guard_common(Node* kls, RegionNode* region,
125 bool obj_array, bool not_array);
126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
128 bool is_virtual = false, bool is_static = false);
129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
130 return generate_method_call(method_id, false, true);
131 }
132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
133 return generate_method_call(method_id, true, false);
134 }
136 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
137 bool inline_string_compareTo();
138 bool inline_string_indexOf();
139 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
140 bool inline_string_equals();
141 Node* pop_math_arg();
142 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
143 bool inline_math_native(vmIntrinsics::ID id);
144 bool inline_trig(vmIntrinsics::ID id);
145 bool inline_trans(vmIntrinsics::ID id);
146 bool inline_abs(vmIntrinsics::ID id);
147 bool inline_sqrt(vmIntrinsics::ID id);
148 bool inline_pow(vmIntrinsics::ID id);
149 bool inline_exp(vmIntrinsics::ID id);
150 bool inline_min_max(vmIntrinsics::ID id);
151 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
152 // This returns Type::AnyPtr, RawPtr, or OopPtr.
153 int classify_unsafe_addr(Node* &base, Node* &offset);
154 Node* make_unsafe_address(Node* base, Node* offset);
155 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
156 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
157 bool inline_unsafe_allocate();
158 bool inline_unsafe_copyMemory();
159 bool inline_native_currentThread();
160 bool inline_native_time_funcs(bool isNano);
161 bool inline_native_isInterrupted();
162 bool inline_native_Class_query(vmIntrinsics::ID id);
163 bool inline_native_subtype_check();
165 bool inline_native_newArray();
166 bool inline_native_getLength();
167 bool inline_array_copyOf(bool is_copyOfRange);
168 bool inline_array_equals();
169 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
170 bool inline_native_clone(bool is_virtual);
171 bool inline_native_Reflection_getCallerClass();
172 bool inline_native_AtomicLong_get();
173 bool inline_native_AtomicLong_attemptUpdate();
174 bool is_method_invoke_or_aux_frame(JVMState* jvms);
175 // Helper function for inlining native object hash method
176 bool inline_native_hashcode(bool is_virtual, bool is_static);
177 bool inline_native_getClass();
179 // Helper functions for inlining arraycopy
180 bool inline_arraycopy();
181 void generate_arraycopy(const TypePtr* adr_type,
182 BasicType basic_elem_type,
183 Node* src, Node* src_offset,
184 Node* dest, Node* dest_offset,
185 Node* copy_length,
186 bool disjoint_bases = false,
187 bool length_never_negative = false,
188 RegionNode* slow_region = NULL);
189 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
190 RegionNode* slow_region);
191 void generate_clear_array(const TypePtr* adr_type,
192 Node* dest,
193 BasicType basic_elem_type,
194 Node* slice_off,
195 Node* slice_len,
196 Node* slice_end);
197 bool generate_block_arraycopy(const TypePtr* adr_type,
198 BasicType basic_elem_type,
199 AllocateNode* alloc,
200 Node* src, Node* src_offset,
201 Node* dest, Node* dest_offset,
202 Node* dest_size);
203 void generate_slow_arraycopy(const TypePtr* adr_type,
204 Node* src, Node* src_offset,
205 Node* dest, Node* dest_offset,
206 Node* copy_length);
207 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
208 Node* dest_elem_klass,
209 Node* src, Node* src_offset,
210 Node* dest, Node* dest_offset,
211 Node* copy_length);
212 Node* generate_generic_arraycopy(const TypePtr* adr_type,
213 Node* src, Node* src_offset,
214 Node* dest, Node* dest_offset,
215 Node* copy_length);
216 void generate_unchecked_arraycopy(const TypePtr* adr_type,
217 BasicType basic_elem_type,
218 bool disjoint_bases,
219 Node* src, Node* src_offset,
220 Node* dest, Node* dest_offset,
221 Node* copy_length);
222 bool inline_unsafe_CAS(BasicType type);
223 bool inline_unsafe_ordered_store(BasicType type);
224 bool inline_fp_conversions(vmIntrinsics::ID id);
225 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
226 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
227 bool inline_bitCount(vmIntrinsics::ID id);
228 bool inline_reverseBytes(vmIntrinsics::ID id);
229 };
232 //---------------------------make_vm_intrinsic----------------------------
233 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
234 vmIntrinsics::ID id = m->intrinsic_id();
235 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
237 if (DisableIntrinsic[0] != '\0'
238 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
239 // disabled by a user request on the command line:
240 // example: -XX:DisableIntrinsic=_hashCode,_getClass
241 return NULL;
242 }
244 if (!m->is_loaded()) {
245 // do not attempt to inline unloaded methods
246 return NULL;
247 }
249 // Only a few intrinsics implement a virtual dispatch.
250 // They are expensive calls which are also frequently overridden.
251 if (is_virtual) {
252 switch (id) {
253 case vmIntrinsics::_hashCode:
254 case vmIntrinsics::_clone:
255 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
256 break;
257 default:
258 return NULL;
259 }
260 }
262 // -XX:-InlineNatives disables nearly all intrinsics:
263 if (!InlineNatives) {
264 switch (id) {
265 case vmIntrinsics::_indexOf:
266 case vmIntrinsics::_compareTo:
267 case vmIntrinsics::_equals:
268 case vmIntrinsics::_equalsC:
269 break; // InlineNatives does not control String.compareTo
270 default:
271 return NULL;
272 }
273 }
275 switch (id) {
276 case vmIntrinsics::_compareTo:
277 if (!SpecialStringCompareTo) return NULL;
278 break;
279 case vmIntrinsics::_indexOf:
280 if (!SpecialStringIndexOf) return NULL;
281 break;
282 case vmIntrinsics::_equals:
283 if (!SpecialStringEquals) return NULL;
284 break;
285 case vmIntrinsics::_equalsC:
286 if (!SpecialArraysEquals) return NULL;
287 break;
288 case vmIntrinsics::_arraycopy:
289 if (!InlineArrayCopy) return NULL;
290 break;
291 case vmIntrinsics::_copyMemory:
292 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
293 if (!InlineArrayCopy) return NULL;
294 break;
295 case vmIntrinsics::_hashCode:
296 if (!InlineObjectHash) return NULL;
297 break;
298 case vmIntrinsics::_clone:
299 case vmIntrinsics::_copyOf:
300 case vmIntrinsics::_copyOfRange:
301 if (!InlineObjectCopy) return NULL;
302 // These also use the arraycopy intrinsic mechanism:
303 if (!InlineArrayCopy) return NULL;
304 break;
305 case vmIntrinsics::_checkIndex:
306 // We do not intrinsify this. The optimizer does fine with it.
307 return NULL;
309 case vmIntrinsics::_get_AtomicLong:
310 case vmIntrinsics::_attemptUpdate:
311 if (!InlineAtomicLong) return NULL;
312 break;
314 case vmIntrinsics::_getCallerClass:
315 if (!UseNewReflection) return NULL;
316 if (!InlineReflectionGetCallerClass) return NULL;
317 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
318 break;
320 case vmIntrinsics::_bitCount_i:
321 case vmIntrinsics::_bitCount_l:
322 if (!UsePopCountInstruction) return NULL;
323 break;
325 default:
326 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
327 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
328 break;
329 }
331 // -XX:-InlineClassNatives disables natives from the Class class.
332 // The flag applies to all reflective calls, notably Array.newArray
333 // (visible to Java programmers as Array.newInstance).
334 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
335 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
336 if (!InlineClassNatives) return NULL;
337 }
339 // -XX:-InlineThreadNatives disables natives from the Thread class.
340 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
341 if (!InlineThreadNatives) return NULL;
342 }
344 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
345 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
346 m->holder()->name() == ciSymbol::java_lang_Float() ||
347 m->holder()->name() == ciSymbol::java_lang_Double()) {
348 if (!InlineMathNatives) return NULL;
349 }
351 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
352 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
353 if (!InlineUnsafeOps) return NULL;
354 }
356 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
357 }
359 //----------------------register_library_intrinsics-----------------------
360 // Initialize this file's data structures, for each Compile instance.
361 void Compile::register_library_intrinsics() {
362 // Nothing to do here.
363 }
365 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
366 LibraryCallKit kit(jvms, this);
367 Compile* C = kit.C;
368 int nodes = C->unique();
369 #ifndef PRODUCT
370 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
371 char buf[1000];
372 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
373 tty->print_cr("Intrinsic %s", str);
374 }
375 #endif
376 if (kit.try_to_inline()) {
377 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
378 tty->print("Inlining intrinsic %s%s at bci:%d in",
379 vmIntrinsics::name_at(intrinsic_id()),
380 (is_virtual() ? " (virtual)" : ""), kit.bci());
381 kit.caller()->print_short_name(tty);
382 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
383 }
384 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
385 if (C->log()) {
386 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
387 vmIntrinsics::name_at(intrinsic_id()),
388 (is_virtual() ? " virtual='1'" : ""),
389 C->unique() - nodes);
390 }
391 return kit.transfer_exceptions_into_jvms();
392 }
394 if (PrintIntrinsics) {
395 tty->print("Did not inline intrinsic %s%s at bci:%d in",
396 vmIntrinsics::name_at(intrinsic_id()),
397 (is_virtual() ? " (virtual)" : ""), kit.bci());
398 kit.caller()->print_short_name(tty);
399 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
400 }
401 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
402 return NULL;
403 }
405 bool LibraryCallKit::try_to_inline() {
406 // Handle symbolic names for otherwise undistinguished boolean switches:
407 const bool is_store = true;
408 const bool is_native_ptr = true;
409 const bool is_static = true;
411 switch (intrinsic_id()) {
412 case vmIntrinsics::_hashCode:
413 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
414 case vmIntrinsics::_identityHashCode:
415 return inline_native_hashcode(/*!virtual*/ false, is_static);
416 case vmIntrinsics::_getClass:
417 return inline_native_getClass();
419 case vmIntrinsics::_dsin:
420 case vmIntrinsics::_dcos:
421 case vmIntrinsics::_dtan:
422 case vmIntrinsics::_dabs:
423 case vmIntrinsics::_datan2:
424 case vmIntrinsics::_dsqrt:
425 case vmIntrinsics::_dexp:
426 case vmIntrinsics::_dlog:
427 case vmIntrinsics::_dlog10:
428 case vmIntrinsics::_dpow:
429 return inline_math_native(intrinsic_id());
431 case vmIntrinsics::_min:
432 case vmIntrinsics::_max:
433 return inline_min_max(intrinsic_id());
435 case vmIntrinsics::_arraycopy:
436 return inline_arraycopy();
438 case vmIntrinsics::_compareTo:
439 return inline_string_compareTo();
440 case vmIntrinsics::_indexOf:
441 return inline_string_indexOf();
442 case vmIntrinsics::_equals:
443 return inline_string_equals();
445 case vmIntrinsics::_getObject:
446 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
447 case vmIntrinsics::_getBoolean:
448 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
449 case vmIntrinsics::_getByte:
450 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
451 case vmIntrinsics::_getShort:
452 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
453 case vmIntrinsics::_getChar:
454 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
455 case vmIntrinsics::_getInt:
456 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
457 case vmIntrinsics::_getLong:
458 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
459 case vmIntrinsics::_getFloat:
460 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
461 case vmIntrinsics::_getDouble:
462 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
464 case vmIntrinsics::_putObject:
465 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
466 case vmIntrinsics::_putBoolean:
467 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
468 case vmIntrinsics::_putByte:
469 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
470 case vmIntrinsics::_putShort:
471 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
472 case vmIntrinsics::_putChar:
473 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
474 case vmIntrinsics::_putInt:
475 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
476 case vmIntrinsics::_putLong:
477 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
478 case vmIntrinsics::_putFloat:
479 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
480 case vmIntrinsics::_putDouble:
481 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
483 case vmIntrinsics::_getByte_raw:
484 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
485 case vmIntrinsics::_getShort_raw:
486 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
487 case vmIntrinsics::_getChar_raw:
488 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
489 case vmIntrinsics::_getInt_raw:
490 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
491 case vmIntrinsics::_getLong_raw:
492 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
493 case vmIntrinsics::_getFloat_raw:
494 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
495 case vmIntrinsics::_getDouble_raw:
496 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
497 case vmIntrinsics::_getAddress_raw:
498 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
500 case vmIntrinsics::_putByte_raw:
501 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
502 case vmIntrinsics::_putShort_raw:
503 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
504 case vmIntrinsics::_putChar_raw:
505 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
506 case vmIntrinsics::_putInt_raw:
507 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
508 case vmIntrinsics::_putLong_raw:
509 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
510 case vmIntrinsics::_putFloat_raw:
511 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
512 case vmIntrinsics::_putDouble_raw:
513 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
514 case vmIntrinsics::_putAddress_raw:
515 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
517 case vmIntrinsics::_getObjectVolatile:
518 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
519 case vmIntrinsics::_getBooleanVolatile:
520 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
521 case vmIntrinsics::_getByteVolatile:
522 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
523 case vmIntrinsics::_getShortVolatile:
524 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
525 case vmIntrinsics::_getCharVolatile:
526 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
527 case vmIntrinsics::_getIntVolatile:
528 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
529 case vmIntrinsics::_getLongVolatile:
530 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
531 case vmIntrinsics::_getFloatVolatile:
532 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
533 case vmIntrinsics::_getDoubleVolatile:
534 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
536 case vmIntrinsics::_putObjectVolatile:
537 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
538 case vmIntrinsics::_putBooleanVolatile:
539 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
540 case vmIntrinsics::_putByteVolatile:
541 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
542 case vmIntrinsics::_putShortVolatile:
543 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
544 case vmIntrinsics::_putCharVolatile:
545 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
546 case vmIntrinsics::_putIntVolatile:
547 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
548 case vmIntrinsics::_putLongVolatile:
549 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
550 case vmIntrinsics::_putFloatVolatile:
551 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
552 case vmIntrinsics::_putDoubleVolatile:
553 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
555 case vmIntrinsics::_prefetchRead:
556 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
557 case vmIntrinsics::_prefetchWrite:
558 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
559 case vmIntrinsics::_prefetchReadStatic:
560 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
561 case vmIntrinsics::_prefetchWriteStatic:
562 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
564 case vmIntrinsics::_compareAndSwapObject:
565 return inline_unsafe_CAS(T_OBJECT);
566 case vmIntrinsics::_compareAndSwapInt:
567 return inline_unsafe_CAS(T_INT);
568 case vmIntrinsics::_compareAndSwapLong:
569 return inline_unsafe_CAS(T_LONG);
571 case vmIntrinsics::_putOrderedObject:
572 return inline_unsafe_ordered_store(T_OBJECT);
573 case vmIntrinsics::_putOrderedInt:
574 return inline_unsafe_ordered_store(T_INT);
575 case vmIntrinsics::_putOrderedLong:
576 return inline_unsafe_ordered_store(T_LONG);
578 case vmIntrinsics::_currentThread:
579 return inline_native_currentThread();
580 case vmIntrinsics::_isInterrupted:
581 return inline_native_isInterrupted();
583 case vmIntrinsics::_currentTimeMillis:
584 return inline_native_time_funcs(false);
585 case vmIntrinsics::_nanoTime:
586 return inline_native_time_funcs(true);
587 case vmIntrinsics::_allocateInstance:
588 return inline_unsafe_allocate();
589 case vmIntrinsics::_copyMemory:
590 return inline_unsafe_copyMemory();
591 case vmIntrinsics::_newArray:
592 return inline_native_newArray();
593 case vmIntrinsics::_getLength:
594 return inline_native_getLength();
595 case vmIntrinsics::_copyOf:
596 return inline_array_copyOf(false);
597 case vmIntrinsics::_copyOfRange:
598 return inline_array_copyOf(true);
599 case vmIntrinsics::_equalsC:
600 return inline_array_equals();
601 case vmIntrinsics::_clone:
602 return inline_native_clone(intrinsic()->is_virtual());
604 case vmIntrinsics::_isAssignableFrom:
605 return inline_native_subtype_check();
607 case vmIntrinsics::_isInstance:
608 case vmIntrinsics::_getModifiers:
609 case vmIntrinsics::_isInterface:
610 case vmIntrinsics::_isArray:
611 case vmIntrinsics::_isPrimitive:
612 case vmIntrinsics::_getSuperclass:
613 case vmIntrinsics::_getComponentType:
614 case vmIntrinsics::_getClassAccessFlags:
615 return inline_native_Class_query(intrinsic_id());
617 case vmIntrinsics::_floatToRawIntBits:
618 case vmIntrinsics::_floatToIntBits:
619 case vmIntrinsics::_intBitsToFloat:
620 case vmIntrinsics::_doubleToRawLongBits:
621 case vmIntrinsics::_doubleToLongBits:
622 case vmIntrinsics::_longBitsToDouble:
623 return inline_fp_conversions(intrinsic_id());
625 case vmIntrinsics::_numberOfLeadingZeros_i:
626 case vmIntrinsics::_numberOfLeadingZeros_l:
627 return inline_numberOfLeadingZeros(intrinsic_id());
629 case vmIntrinsics::_numberOfTrailingZeros_i:
630 case vmIntrinsics::_numberOfTrailingZeros_l:
631 return inline_numberOfTrailingZeros(intrinsic_id());
633 case vmIntrinsics::_bitCount_i:
634 case vmIntrinsics::_bitCount_l:
635 return inline_bitCount(intrinsic_id());
637 case vmIntrinsics::_reverseBytes_i:
638 case vmIntrinsics::_reverseBytes_l:
639 case vmIntrinsics::_reverseBytes_s:
640 case vmIntrinsics::_reverseBytes_c:
641 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
643 case vmIntrinsics::_get_AtomicLong:
644 return inline_native_AtomicLong_get();
645 case vmIntrinsics::_attemptUpdate:
646 return inline_native_AtomicLong_attemptUpdate();
648 case vmIntrinsics::_getCallerClass:
649 return inline_native_Reflection_getCallerClass();
651 default:
652 // If you get here, it may be that someone has added a new intrinsic
653 // to the list in vmSymbols.hpp without implementing it here.
654 #ifndef PRODUCT
655 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
656 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
657 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
658 }
659 #endif
660 return false;
661 }
662 }
664 //------------------------------push_result------------------------------
665 // Helper function for finishing intrinsics.
666 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
667 record_for_igvn(region);
668 set_control(_gvn.transform(region));
669 BasicType value_type = value->type()->basic_type();
670 push_node(value_type, _gvn.transform(value));
671 }
673 //------------------------------generate_guard---------------------------
674 // Helper function for generating guarded fast-slow graph structures.
675 // The given 'test', if true, guards a slow path. If the test fails
676 // then a fast path can be taken. (We generally hope it fails.)
677 // In all cases, GraphKit::control() is updated to the fast path.
678 // The returned value represents the control for the slow path.
679 // The return value is never 'top'; it is either a valid control
680 // or NULL if it is obvious that the slow path can never be taken.
681 // Also, if region and the slow control are not NULL, the slow edge
682 // is appended to the region.
683 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
684 if (stopped()) {
685 // Already short circuited.
686 return NULL;
687 }
689 // Build an if node and its projections.
690 // If test is true we take the slow path, which we assume is uncommon.
691 if (_gvn.type(test) == TypeInt::ZERO) {
692 // The slow branch is never taken. No need to build this guard.
693 return NULL;
694 }
696 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
698 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
699 if (if_slow == top()) {
700 // The slow branch is never taken. No need to build this guard.
701 return NULL;
702 }
704 if (region != NULL)
705 region->add_req(if_slow);
707 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
708 set_control(if_fast);
710 return if_slow;
711 }
713 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
714 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
715 }
716 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
717 return generate_guard(test, region, PROB_FAIR);
718 }
720 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
721 Node* *pos_index) {
722 if (stopped())
723 return NULL; // already stopped
724 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
725 return NULL; // index is already adequately typed
726 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
727 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
728 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
729 if (is_neg != NULL && pos_index != NULL) {
730 // Emulate effect of Parse::adjust_map_after_if.
731 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
732 ccast->set_req(0, control());
733 (*pos_index) = _gvn.transform(ccast);
734 }
735 return is_neg;
736 }
738 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
739 Node* *pos_index) {
740 if (stopped())
741 return NULL; // already stopped
742 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
743 return NULL; // index is already adequately typed
744 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
745 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
746 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
747 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
748 if (is_notp != NULL && pos_index != NULL) {
749 // Emulate effect of Parse::adjust_map_after_if.
750 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
751 ccast->set_req(0, control());
752 (*pos_index) = _gvn.transform(ccast);
753 }
754 return is_notp;
755 }
757 // Make sure that 'position' is a valid limit index, in [0..length].
758 // There are two equivalent plans for checking this:
759 // A. (offset + copyLength) unsigned<= arrayLength
760 // B. offset <= (arrayLength - copyLength)
761 // We require that all of the values above, except for the sum and
762 // difference, are already known to be non-negative.
763 // Plan A is robust in the face of overflow, if offset and copyLength
764 // are both hugely positive.
765 //
766 // Plan B is less direct and intuitive, but it does not overflow at
767 // all, since the difference of two non-negatives is always
768 // representable. Whenever Java methods must perform the equivalent
769 // check they generally use Plan B instead of Plan A.
770 // For the moment we use Plan A.
771 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
772 Node* subseq_length,
773 Node* array_length,
774 RegionNode* region) {
775 if (stopped())
776 return NULL; // already stopped
777 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
778 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
779 return NULL; // common case of whole-array copy
780 Node* last = subseq_length;
781 if (!zero_offset) // last += offset
782 last = _gvn.transform( new (C, 3) AddINode(last, offset));
783 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
784 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
785 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
786 return is_over;
787 }
790 //--------------------------generate_current_thread--------------------
791 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
792 ciKlass* thread_klass = env()->Thread_klass();
793 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
794 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
795 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
796 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
797 tls_output = thread;
798 return threadObj;
799 }
802 //------------------------------make_string_method_node------------------------
803 // Helper method for String intrinsic finctions.
804 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
805 const int value_offset = java_lang_String::value_offset_in_bytes();
806 const int count_offset = java_lang_String::count_offset_in_bytes();
807 const int offset_offset = java_lang_String::offset_offset_in_bytes();
809 Node* no_ctrl = NULL;
811 ciInstanceKlass* klass = env()->String_klass();
812 const TypeInstPtr* string_type =
813 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
815 const TypeAryPtr* value_type =
816 TypeAryPtr::make(TypePtr::NotNull,
817 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
818 ciTypeArrayKlass::make(T_CHAR), true, 0);
820 // Get start addr of string and substring
821 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
822 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
823 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
824 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
825 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
827 // Pin loads from String::equals() argument since it could be NULL.
828 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
829 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
830 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
831 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
832 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
833 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
835 Node* result = NULL;
836 switch (opcode) {
837 case Op_StrIndexOf:
838 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
839 str1_start, cnt1, str2_start, cnt2);
840 break;
841 case Op_StrComp:
842 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
843 str1_start, cnt1, str2_start, cnt2);
844 break;
845 case Op_StrEquals:
846 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
847 str1_start, str2_start, cnt1);
848 break;
849 default:
850 ShouldNotReachHere();
851 return NULL;
852 }
854 // All these intrinsics have checks.
855 C->set_has_split_ifs(true); // Has chance for split-if optimization
857 return _gvn.transform(result);
858 }
860 //------------------------------inline_string_compareTo------------------------
861 bool LibraryCallKit::inline_string_compareTo() {
863 if (!Matcher::has_match_rule(Op_StrComp)) return false;
865 const int value_offset = java_lang_String::value_offset_in_bytes();
866 const int count_offset = java_lang_String::count_offset_in_bytes();
867 const int offset_offset = java_lang_String::offset_offset_in_bytes();
869 _sp += 2;
870 Node *argument = pop(); // pop non-receiver first: it was pushed second
871 Node *receiver = pop();
873 // Null check on self without removing any arguments. The argument
874 // null check technically happens in the wrong place, which can lead to
875 // invalid stack traces when string compare is inlined into a method
876 // which handles NullPointerExceptions.
877 _sp += 2;
878 receiver = do_null_check(receiver, T_OBJECT);
879 argument = do_null_check(argument, T_OBJECT);
880 _sp -= 2;
881 if (stopped()) {
882 return true;
883 }
885 ciInstanceKlass* klass = env()->String_klass();
886 const TypeInstPtr* string_type =
887 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
888 Node* no_ctrl = NULL;
890 // Get counts for string and argument
891 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
892 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
894 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
895 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
897 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
898 push(compare);
899 return true;
900 }
902 //------------------------------inline_string_equals------------------------
903 bool LibraryCallKit::inline_string_equals() {
905 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
907 const int value_offset = java_lang_String::value_offset_in_bytes();
908 const int count_offset = java_lang_String::count_offset_in_bytes();
909 const int offset_offset = java_lang_String::offset_offset_in_bytes();
911 _sp += 2;
912 Node* argument = pop(); // pop non-receiver first: it was pushed second
913 Node* receiver = pop();
915 // Null check on self without removing any arguments. The argument
916 // null check technically happens in the wrong place, which can lead to
917 // invalid stack traces when string compare is inlined into a method
918 // which handles NullPointerExceptions.
919 _sp += 2;
920 receiver = do_null_check(receiver, T_OBJECT);
921 //should not do null check for argument for String.equals(), because spec
922 //allows to specify NULL as argument.
923 _sp -= 2;
925 if (stopped()) {
926 return true;
927 }
929 // paths (plus control) merge
930 RegionNode* region = new (C, 5) RegionNode(5);
931 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
933 // does source == target string?
934 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
935 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
937 Node* if_eq = generate_slow_guard(bol, NULL);
938 if (if_eq != NULL) {
939 // receiver == argument
940 phi->init_req(2, intcon(1));
941 region->init_req(2, if_eq);
942 }
944 // get String klass for instanceOf
945 ciInstanceKlass* klass = env()->String_klass();
947 if (!stopped()) {
948 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
949 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
950 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
952 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
953 //instanceOf == true, fallthrough
955 if (inst_false != NULL) {
956 phi->init_req(3, intcon(0));
957 region->init_req(3, inst_false);
958 }
959 }
961 const TypeInstPtr* string_type =
962 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
964 Node* no_ctrl = NULL;
965 Node* receiver_cnt;
966 Node* argument_cnt;
968 if (!stopped()) {
969 // Get counts for string and argument
970 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
971 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
973 // Pin load from argument string since it could be NULL.
974 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
975 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
977 // Check for receiver count != argument count
978 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
979 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
980 Node* if_ne = generate_slow_guard(bol, NULL);
981 if (if_ne != NULL) {
982 phi->init_req(4, intcon(0));
983 region->init_req(4, if_ne);
984 }
985 }
987 // Check for count == 0 is done by mach node StrEquals.
989 if (!stopped()) {
990 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
991 phi->init_req(1, equals);
992 region->init_req(1, control());
993 }
995 // post merge
996 set_control(_gvn.transform(region));
997 record_for_igvn(region);
999 push(_gvn.transform(phi));
1001 return true;
1002 }
1004 //------------------------------inline_array_equals----------------------------
1005 bool LibraryCallKit::inline_array_equals() {
1007 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1009 _sp += 2;
1010 Node *argument2 = pop();
1011 Node *argument1 = pop();
1013 Node* equals =
1014 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1015 argument1, argument2) );
1016 push(equals);
1017 return true;
1018 }
1020 // Java version of String.indexOf(constant string)
1021 // class StringDecl {
1022 // StringDecl(char[] ca) {
1023 // offset = 0;
1024 // count = ca.length;
1025 // value = ca;
1026 // }
1027 // int offset;
1028 // int count;
1029 // char[] value;
1030 // }
1031 //
1032 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1033 // int targetOffset, int cache_i, int md2) {
1034 // int cache = cache_i;
1035 // int sourceOffset = string_object.offset;
1036 // int sourceCount = string_object.count;
1037 // int targetCount = target_object.length;
1038 //
1039 // int targetCountLess1 = targetCount - 1;
1040 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1041 //
1042 // char[] source = string_object.value;
1043 // char[] target = target_object;
1044 // int lastChar = target[targetCountLess1];
1045 //
1046 // outer_loop:
1047 // for (int i = sourceOffset; i < sourceEnd; ) {
1048 // int src = source[i + targetCountLess1];
1049 // if (src == lastChar) {
1050 // // With random strings and a 4-character alphabet,
1051 // // reverse matching at this point sets up 0.8% fewer
1052 // // frames, but (paradoxically) makes 0.3% more probes.
1053 // // Since those probes are nearer the lastChar probe,
1054 // // there is may be a net D$ win with reverse matching.
1055 // // But, reversing loop inhibits unroll of inner loop
1056 // // for unknown reason. So, does running outer loop from
1057 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1058 // for (int j = 0; j < targetCountLess1; j++) {
1059 // if (target[targetOffset + j] != source[i+j]) {
1060 // if ((cache & (1 << source[i+j])) == 0) {
1061 // if (md2 < j+1) {
1062 // i += j+1;
1063 // continue outer_loop;
1064 // }
1065 // }
1066 // i += md2;
1067 // continue outer_loop;
1068 // }
1069 // }
1070 // return i - sourceOffset;
1071 // }
1072 // if ((cache & (1 << src)) == 0) {
1073 // i += targetCountLess1;
1074 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1075 // i++;
1076 // }
1077 // return -1;
1078 // }
1080 //------------------------------string_indexOf------------------------
1081 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1082 jint cache_i, jint md2_i) {
1084 Node* no_ctrl = NULL;
1085 float likely = PROB_LIKELY(0.9);
1086 float unlikely = PROB_UNLIKELY(0.9);
1088 const int value_offset = java_lang_String::value_offset_in_bytes();
1089 const int count_offset = java_lang_String::count_offset_in_bytes();
1090 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1092 ciInstanceKlass* klass = env()->String_klass();
1093 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1094 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1096 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1097 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1098 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1099 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1100 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1101 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1103 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
1104 jint target_length = target_array->length();
1105 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1106 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1108 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1109 #define __ kit.
1110 Node* zero = __ ConI(0);
1111 Node* one = __ ConI(1);
1112 Node* cache = __ ConI(cache_i);
1113 Node* md2 = __ ConI(md2_i);
1114 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1115 Node* targetCount = __ ConI(target_length);
1116 Node* targetCountLess1 = __ ConI(target_length - 1);
1117 Node* targetOffset = __ ConI(targetOffset_i);
1118 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1120 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1121 Node* outer_loop = __ make_label(2 /* goto */);
1122 Node* return_ = __ make_label(1);
1124 __ set(rtn,__ ConI(-1));
1125 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
1126 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1127 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1128 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1129 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1130 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
1131 Node* tpj = __ AddI(targetOffset, __ value(j));
1132 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1133 Node* ipj = __ AddI(__ value(i), __ value(j));
1134 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1135 __ if_then(targ, BoolTest::ne, src2); {
1136 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1137 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1138 __ increment(i, __ AddI(__ value(j), one));
1139 __ goto_(outer_loop);
1140 } __ end_if(); __ dead(j);
1141 }__ end_if(); __ dead(j);
1142 __ increment(i, md2);
1143 __ goto_(outer_loop);
1144 }__ end_if();
1145 __ increment(j, one);
1146 }__ end_loop(); __ dead(j);
1147 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1148 __ goto_(return_);
1149 }__ end_if();
1150 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1151 __ increment(i, targetCountLess1);
1152 }__ end_if();
1153 __ increment(i, one);
1154 __ bind(outer_loop);
1155 }__ end_loop(); __ dead(i);
1156 __ bind(return_);
1158 // Final sync IdealKit and GraphKit.
1159 sync_kit(kit);
1160 Node* result = __ value(rtn);
1161 #undef __
1162 C->set_has_loops(true);
1163 return result;
1164 }
1166 //------------------------------inline_string_indexOf------------------------
1167 bool LibraryCallKit::inline_string_indexOf() {
1169 const int value_offset = java_lang_String::value_offset_in_bytes();
1170 const int count_offset = java_lang_String::count_offset_in_bytes();
1171 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1173 _sp += 2;
1174 Node *argument = pop(); // pop non-receiver first: it was pushed second
1175 Node *receiver = pop();
1177 Node* result;
1178 // Disable the use of pcmpestri until it can be guaranteed that
1179 // the load doesn't cross into the uncommited space.
1180 if (false && Matcher::has_match_rule(Op_StrIndexOf) &&
1181 UseSSE42Intrinsics) {
1182 // Generate SSE4.2 version of indexOf
1183 // We currently only have match rules that use SSE4.2
1185 // Null check on self without removing any arguments. The argument
1186 // null check technically happens in the wrong place, which can lead to
1187 // invalid stack traces when string compare is inlined into a method
1188 // which handles NullPointerExceptions.
1189 _sp += 2;
1190 receiver = do_null_check(receiver, T_OBJECT);
1191 argument = do_null_check(argument, T_OBJECT);
1192 _sp -= 2;
1194 if (stopped()) {
1195 return true;
1196 }
1198 // Make the merge point
1199 RegionNode* result_rgn = new (C, 3) RegionNode(3);
1200 Node* result_phi = new (C, 3) PhiNode(result_rgn, TypeInt::INT);
1201 Node* no_ctrl = NULL;
1203 ciInstanceKlass* klass = env()->String_klass();
1204 const TypeInstPtr* string_type =
1205 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1207 // Get counts for string and substr
1208 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1209 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1211 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1212 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1214 // Check for substr count > string count
1215 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1216 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1217 Node* if_gt = generate_slow_guard(bol, NULL);
1218 if (if_gt != NULL) {
1219 result_phi->init_req(2, intcon(-1));
1220 result_rgn->init_req(2, if_gt);
1221 }
1223 if (!stopped()) {
1224 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1225 result_phi->init_req(1, result);
1226 result_rgn->init_req(1, control());
1227 }
1228 set_control(_gvn.transform(result_rgn));
1229 record_for_igvn(result_rgn);
1230 result = _gvn.transform(result_phi);
1232 } else { //Use LibraryCallKit::string_indexOf
1233 // don't intrinsify is argument isn't a constant string.
1234 if (!argument->is_Con()) {
1235 return false;
1236 }
1237 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1238 if (str_type == NULL) {
1239 return false;
1240 }
1241 ciInstanceKlass* klass = env()->String_klass();
1242 ciObject* str_const = str_type->const_oop();
1243 if (str_const == NULL || str_const->klass() != klass) {
1244 return false;
1245 }
1246 ciInstance* str = str_const->as_instance();
1247 assert(str != NULL, "must be instance");
1249 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1250 int o = str->field_value_by_offset(offset_offset).as_int();
1251 int c = str->field_value_by_offset(count_offset).as_int();
1252 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1254 // constant strings have no offset and count == length which
1255 // simplifies the resulting code somewhat so lets optimize for that.
1256 if (o != 0 || c != pat->length()) {
1257 return false;
1258 }
1260 // Null check on self without removing any arguments. The argument
1261 // null check technically happens in the wrong place, which can lead to
1262 // invalid stack traces when string compare is inlined into a method
1263 // which handles NullPointerExceptions.
1264 _sp += 2;
1265 receiver = do_null_check(receiver, T_OBJECT);
1266 // No null check on the argument is needed since it's a constant String oop.
1267 _sp -= 2;
1268 if (stopped()) {
1269 return true;
1270 }
1272 // The null string as a pattern always returns 0 (match at beginning of string)
1273 if (c == 0) {
1274 push(intcon(0));
1275 return true;
1276 }
1278 // Generate default indexOf
1279 jchar lastChar = pat->char_at(o + (c - 1));
1280 int cache = 0;
1281 int i;
1282 for (i = 0; i < c - 1; i++) {
1283 assert(i < pat->length(), "out of range");
1284 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1285 }
1287 int md2 = c;
1288 for (i = 0; i < c - 1; i++) {
1289 assert(i < pat->length(), "out of range");
1290 if (pat->char_at(o + i) == lastChar) {
1291 md2 = (c - 1) - i;
1292 }
1293 }
1295 result = string_indexOf(receiver, pat, o, cache, md2);
1296 }
1298 push(result);
1299 return true;
1300 }
1302 //--------------------------pop_math_arg--------------------------------
1303 // Pop a double argument to a math function from the stack
1304 // rounding it if necessary.
1305 Node * LibraryCallKit::pop_math_arg() {
1306 Node *arg = pop_pair();
1307 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1308 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1309 return arg;
1310 }
1312 //------------------------------inline_trig----------------------------------
1313 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1314 // argument reduction which will turn into a fast/slow diamond.
1315 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1316 _sp += arg_size(); // restore stack pointer
1317 Node* arg = pop_math_arg();
1318 Node* trig = NULL;
1320 switch (id) {
1321 case vmIntrinsics::_dsin:
1322 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1323 break;
1324 case vmIntrinsics::_dcos:
1325 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1326 break;
1327 case vmIntrinsics::_dtan:
1328 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1329 break;
1330 default:
1331 assert(false, "bad intrinsic was passed in");
1332 return false;
1333 }
1335 // Rounding required? Check for argument reduction!
1336 if( Matcher::strict_fp_requires_explicit_rounding ) {
1338 static const double pi_4 = 0.7853981633974483;
1339 static const double neg_pi_4 = -0.7853981633974483;
1340 // pi/2 in 80-bit extended precision
1341 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1342 // -pi/2 in 80-bit extended precision
1343 // 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};
1344 // Cutoff value for using this argument reduction technique
1345 //static const double pi_2_minus_epsilon = 1.564660403643354;
1346 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1348 // Pseudocode for sin:
1349 // if (x <= Math.PI / 4.0) {
1350 // if (x >= -Math.PI / 4.0) return fsin(x);
1351 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1352 // } else {
1353 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1354 // }
1355 // return StrictMath.sin(x);
1357 // Pseudocode for cos:
1358 // if (x <= Math.PI / 4.0) {
1359 // if (x >= -Math.PI / 4.0) return fcos(x);
1360 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1361 // } else {
1362 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1363 // }
1364 // return StrictMath.cos(x);
1366 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1367 // requires a special machine instruction to load it. Instead we'll try
1368 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1369 // probably do the math inside the SIN encoding.
1371 // Make the merge point
1372 RegionNode *r = new (C, 3) RegionNode(3);
1373 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1375 // Flatten arg so we need only 1 test
1376 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1377 // Node for PI/4 constant
1378 Node *pi4 = makecon(TypeD::make(pi_4));
1379 // Check PI/4 : abs(arg)
1380 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1381 // Check: If PI/4 < abs(arg) then go slow
1382 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1383 // Branch either way
1384 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1385 set_control(opt_iff(r,iff));
1387 // Set fast path result
1388 phi->init_req(2,trig);
1390 // Slow path - non-blocking leaf call
1391 Node* call = NULL;
1392 switch (id) {
1393 case vmIntrinsics::_dsin:
1394 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1395 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1396 "Sin", NULL, arg, top());
1397 break;
1398 case vmIntrinsics::_dcos:
1399 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1400 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1401 "Cos", NULL, arg, top());
1402 break;
1403 case vmIntrinsics::_dtan:
1404 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1405 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1406 "Tan", NULL, arg, top());
1407 break;
1408 }
1409 assert(control()->in(0) == call, "");
1410 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1411 r->init_req(1,control());
1412 phi->init_req(1,slow_result);
1414 // Post-merge
1415 set_control(_gvn.transform(r));
1416 record_for_igvn(r);
1417 trig = _gvn.transform(phi);
1419 C->set_has_split_ifs(true); // Has chance for split-if optimization
1420 }
1421 // Push result back on JVM stack
1422 push_pair(trig);
1423 return true;
1424 }
1426 //------------------------------inline_sqrt-------------------------------------
1427 // Inline square root instruction, if possible.
1428 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1429 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1430 _sp += arg_size(); // restore stack pointer
1431 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1432 return true;
1433 }
1435 //------------------------------inline_abs-------------------------------------
1436 // Inline absolute value instruction, if possible.
1437 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1438 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1439 _sp += arg_size(); // restore stack pointer
1440 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1441 return true;
1442 }
1444 //------------------------------inline_exp-------------------------------------
1445 // Inline exp instructions, if possible. The Intel hardware only misses
1446 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1447 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1448 assert(id == vmIntrinsics::_dexp, "Not exp");
1450 // If this inlining ever returned NaN in the past, we do not intrinsify it
1451 // every again. NaN results requires StrictMath.exp handling.
1452 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1454 // Do not intrinsify on older platforms which lack cmove.
1455 if (ConditionalMoveLimit == 0) return false;
1457 _sp += arg_size(); // restore stack pointer
1458 Node *x = pop_math_arg();
1459 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1461 //-------------------
1462 //result=(result.isNaN())? StrictMath::exp():result;
1463 // Check: If isNaN() by checking result!=result? then go to Strict Math
1464 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1465 // Build the boolean node
1466 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1468 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1469 // End the current control-flow path
1470 push_pair(x);
1471 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1472 // to handle. Recompile without intrinsifying Math.exp
1473 uncommon_trap(Deoptimization::Reason_intrinsic,
1474 Deoptimization::Action_make_not_entrant);
1475 }
1477 C->set_has_split_ifs(true); // Has chance for split-if optimization
1479 push_pair(result);
1481 return true;
1482 }
1484 //------------------------------inline_pow-------------------------------------
1485 // Inline power instructions, if possible.
1486 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1487 assert(id == vmIntrinsics::_dpow, "Not pow");
1489 // If this inlining ever returned NaN in the past, we do not intrinsify it
1490 // every again. NaN results requires StrictMath.pow handling.
1491 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1493 // Do not intrinsify on older platforms which lack cmove.
1494 if (ConditionalMoveLimit == 0) return false;
1496 // Pseudocode for pow
1497 // if (x <= 0.0) {
1498 // if ((double)((int)y)==y) { // if y is int
1499 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1500 // } else {
1501 // result = NaN;
1502 // }
1503 // } else {
1504 // result = DPow(x,y);
1505 // }
1506 // if (result != result)? {
1507 // uncommon_trap();
1508 // }
1509 // return result;
1511 _sp += arg_size(); // restore stack pointer
1512 Node* y = pop_math_arg();
1513 Node* x = pop_math_arg();
1515 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1517 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1518 // inside of something) then skip the fancy tests and just check for
1519 // NaN result.
1520 Node *result = NULL;
1521 if( jvms()->depth() >= 1 ) {
1522 result = fast_result;
1523 } else {
1525 // Set the merge point for If node with condition of (x <= 0.0)
1526 // There are four possible paths to region node and phi node
1527 RegionNode *r = new (C, 4) RegionNode(4);
1528 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1530 // Build the first if node: if (x <= 0.0)
1531 // Node for 0 constant
1532 Node *zeronode = makecon(TypeD::ZERO);
1533 // Check x:0
1534 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1535 // Check: If (x<=0) then go complex path
1536 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1537 // Branch either way
1538 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1539 Node *opt_test = _gvn.transform(if1);
1540 //assert( opt_test->is_If(), "Expect an IfNode");
1541 IfNode *opt_if1 = (IfNode*)opt_test;
1542 // Fast path taken; set region slot 3
1543 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1544 r->init_req(3,fast_taken); // Capture fast-control
1546 // Fast path not-taken, i.e. slow path
1547 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1549 // Set fast path result
1550 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1551 phi->init_req(3, fast_result);
1553 // Complex path
1554 // Build the second if node (if y is int)
1555 // Node for (int)y
1556 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1557 // Node for (double)((int) y)
1558 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1559 // Check (double)((int) y) : y
1560 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1561 // Check if (y isn't int) then go to slow path
1563 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1564 // Branch either way
1565 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1566 Node *slow_path = opt_iff(r,if2); // Set region path 2
1568 // Calculate DPow(abs(x), y)*(1 & (int)y)
1569 // Node for constant 1
1570 Node *conone = intcon(1);
1571 // 1& (int)y
1572 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1573 // zero node
1574 Node *conzero = intcon(0);
1575 // Check (1&(int)y)==0?
1576 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1577 // Check if (1&(int)y)!=0?, if so the result is negative
1578 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1579 // abs(x)
1580 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1581 // abs(x)^y
1582 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1583 // -abs(x)^y
1584 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1585 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1586 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1587 // Set complex path fast result
1588 phi->init_req(2, signresult);
1590 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1591 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1592 r->init_req(1,slow_path);
1593 phi->init_req(1,slow_result);
1595 // Post merge
1596 set_control(_gvn.transform(r));
1597 record_for_igvn(r);
1598 result=_gvn.transform(phi);
1599 }
1601 //-------------------
1602 //result=(result.isNaN())? uncommon_trap():result;
1603 // Check: If isNaN() by checking result!=result? then go to Strict Math
1604 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1605 // Build the boolean node
1606 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1608 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1609 // End the current control-flow path
1610 push_pair(x);
1611 push_pair(y);
1612 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1613 // to handle. Recompile without intrinsifying Math.pow.
1614 uncommon_trap(Deoptimization::Reason_intrinsic,
1615 Deoptimization::Action_make_not_entrant);
1616 }
1618 C->set_has_split_ifs(true); // Has chance for split-if optimization
1620 push_pair(result);
1622 return true;
1623 }
1625 //------------------------------inline_trans-------------------------------------
1626 // Inline transcendental instructions, if possible. The Intel hardware gets
1627 // these right, no funny corner cases missed.
1628 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1629 _sp += arg_size(); // restore stack pointer
1630 Node* arg = pop_math_arg();
1631 Node* trans = NULL;
1633 switch (id) {
1634 case vmIntrinsics::_dlog:
1635 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1636 break;
1637 case vmIntrinsics::_dlog10:
1638 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1639 break;
1640 default:
1641 assert(false, "bad intrinsic was passed in");
1642 return false;
1643 }
1645 // Push result back on JVM stack
1646 push_pair(trans);
1647 return true;
1648 }
1650 //------------------------------runtime_math-----------------------------
1651 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1652 Node* a = NULL;
1653 Node* b = NULL;
1655 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1656 "must be (DD)D or (D)D type");
1658 // Inputs
1659 _sp += arg_size(); // restore stack pointer
1660 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1661 b = pop_math_arg();
1662 }
1663 a = pop_math_arg();
1665 const TypePtr* no_memory_effects = NULL;
1666 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1667 no_memory_effects,
1668 a, top(), b, b ? top() : NULL);
1669 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1670 #ifdef ASSERT
1671 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1672 assert(value_top == top(), "second value must be top");
1673 #endif
1675 push_pair(value);
1676 return true;
1677 }
1679 //------------------------------inline_math_native-----------------------------
1680 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1681 switch (id) {
1682 // These intrinsics are not properly supported on all hardware
1683 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1684 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1685 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1686 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1687 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1688 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1690 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1691 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1692 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1693 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1695 // These intrinsics are supported on all hardware
1696 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1697 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1699 // These intrinsics don't work on X86. The ad implementation doesn't
1700 // handle NaN's properly. Instead of returning infinity, the ad
1701 // implementation returns a NaN on overflow. See bug: 6304089
1702 // Once the ad implementations are fixed, change the code below
1703 // to match the intrinsics above
1705 case vmIntrinsics::_dexp: return
1706 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1707 case vmIntrinsics::_dpow: return
1708 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1710 // These intrinsics are not yet correctly implemented
1711 case vmIntrinsics::_datan2:
1712 return false;
1714 default:
1715 ShouldNotReachHere();
1716 return false;
1717 }
1718 }
1720 static bool is_simple_name(Node* n) {
1721 return (n->req() == 1 // constant
1722 || (n->is_Type() && n->as_Type()->type()->singleton())
1723 || n->is_Proj() // parameter or return value
1724 || n->is_Phi() // local of some sort
1725 );
1726 }
1728 //----------------------------inline_min_max-----------------------------------
1729 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1730 push(generate_min_max(id, argument(0), argument(1)));
1732 return true;
1733 }
1735 Node*
1736 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1737 // These are the candidate return value:
1738 Node* xvalue = x0;
1739 Node* yvalue = y0;
1741 if (xvalue == yvalue) {
1742 return xvalue;
1743 }
1745 bool want_max = (id == vmIntrinsics::_max);
1747 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1748 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1749 if (txvalue == NULL || tyvalue == NULL) return top();
1750 // This is not really necessary, but it is consistent with a
1751 // hypothetical MaxINode::Value method:
1752 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1754 // %%% This folding logic should (ideally) be in a different place.
1755 // Some should be inside IfNode, and there to be a more reliable
1756 // transformation of ?: style patterns into cmoves. We also want
1757 // more powerful optimizations around cmove and min/max.
1759 // Try to find a dominating comparison of these guys.
1760 // It can simplify the index computation for Arrays.copyOf
1761 // and similar uses of System.arraycopy.
1762 // First, compute the normalized version of CmpI(x, y).
1763 int cmp_op = Op_CmpI;
1764 Node* xkey = xvalue;
1765 Node* ykey = yvalue;
1766 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1767 if (ideal_cmpxy->is_Cmp()) {
1768 // E.g., if we have CmpI(length - offset, count),
1769 // it might idealize to CmpI(length, count + offset)
1770 cmp_op = ideal_cmpxy->Opcode();
1771 xkey = ideal_cmpxy->in(1);
1772 ykey = ideal_cmpxy->in(2);
1773 }
1775 // Start by locating any relevant comparisons.
1776 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1777 Node* cmpxy = NULL;
1778 Node* cmpyx = NULL;
1779 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1780 Node* cmp = start_from->fast_out(k);
1781 if (cmp->outcnt() > 0 && // must have prior uses
1782 cmp->in(0) == NULL && // must be context-independent
1783 cmp->Opcode() == cmp_op) { // right kind of compare
1784 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1785 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1786 }
1787 }
1789 const int NCMPS = 2;
1790 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1791 int cmpn;
1792 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1793 if (cmps[cmpn] != NULL) break; // find a result
1794 }
1795 if (cmpn < NCMPS) {
1796 // Look for a dominating test that tells us the min and max.
1797 int depth = 0; // Limit search depth for speed
1798 Node* dom = control();
1799 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1800 if (++depth >= 100) break;
1801 Node* ifproj = dom;
1802 if (!ifproj->is_Proj()) continue;
1803 Node* iff = ifproj->in(0);
1804 if (!iff->is_If()) continue;
1805 Node* bol = iff->in(1);
1806 if (!bol->is_Bool()) continue;
1807 Node* cmp = bol->in(1);
1808 if (cmp == NULL) continue;
1809 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1810 if (cmps[cmpn] == cmp) break;
1811 if (cmpn == NCMPS) continue;
1812 BoolTest::mask btest = bol->as_Bool()->_test._test;
1813 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1814 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1815 // At this point, we know that 'x btest y' is true.
1816 switch (btest) {
1817 case BoolTest::eq:
1818 // They are proven equal, so we can collapse the min/max.
1819 // Either value is the answer. Choose the simpler.
1820 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1821 return yvalue;
1822 return xvalue;
1823 case BoolTest::lt: // x < y
1824 case BoolTest::le: // x <= y
1825 return (want_max ? yvalue : xvalue);
1826 case BoolTest::gt: // x > y
1827 case BoolTest::ge: // x >= y
1828 return (want_max ? xvalue : yvalue);
1829 }
1830 }
1831 }
1833 // We failed to find a dominating test.
1834 // Let's pick a test that might GVN with prior tests.
1835 Node* best_bol = NULL;
1836 BoolTest::mask best_btest = BoolTest::illegal;
1837 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1838 Node* cmp = cmps[cmpn];
1839 if (cmp == NULL) continue;
1840 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1841 Node* bol = cmp->fast_out(j);
1842 if (!bol->is_Bool()) continue;
1843 BoolTest::mask btest = bol->as_Bool()->_test._test;
1844 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1845 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1846 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1847 best_bol = bol->as_Bool();
1848 best_btest = btest;
1849 }
1850 }
1851 }
1853 Node* answer_if_true = NULL;
1854 Node* answer_if_false = NULL;
1855 switch (best_btest) {
1856 default:
1857 if (cmpxy == NULL)
1858 cmpxy = ideal_cmpxy;
1859 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1860 // and fall through:
1861 case BoolTest::lt: // x < y
1862 case BoolTest::le: // x <= y
1863 answer_if_true = (want_max ? yvalue : xvalue);
1864 answer_if_false = (want_max ? xvalue : yvalue);
1865 break;
1866 case BoolTest::gt: // x > y
1867 case BoolTest::ge: // x >= y
1868 answer_if_true = (want_max ? xvalue : yvalue);
1869 answer_if_false = (want_max ? yvalue : xvalue);
1870 break;
1871 }
1873 jint hi, lo;
1874 if (want_max) {
1875 // We can sharpen the minimum.
1876 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1877 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1878 } else {
1879 // We can sharpen the maximum.
1880 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1881 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1882 }
1884 // Use a flow-free graph structure, to avoid creating excess control edges
1885 // which could hinder other optimizations.
1886 // Since Math.min/max is often used with arraycopy, we want
1887 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1888 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1889 answer_if_false, answer_if_true,
1890 TypeInt::make(lo, hi, widen));
1892 return _gvn.transform(cmov);
1894 /*
1895 // This is not as desirable as it may seem, since Min and Max
1896 // nodes do not have a full set of optimizations.
1897 // And they would interfere, anyway, with 'if' optimizations
1898 // and with CMoveI canonical forms.
1899 switch (id) {
1900 case vmIntrinsics::_min:
1901 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1902 case vmIntrinsics::_max:
1903 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1904 default:
1905 ShouldNotReachHere();
1906 }
1907 */
1908 }
1910 inline int
1911 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1912 const TypePtr* base_type = TypePtr::NULL_PTR;
1913 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1914 if (base_type == NULL) {
1915 // Unknown type.
1916 return Type::AnyPtr;
1917 } else if (base_type == TypePtr::NULL_PTR) {
1918 // Since this is a NULL+long form, we have to switch to a rawptr.
1919 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1920 offset = MakeConX(0);
1921 return Type::RawPtr;
1922 } else if (base_type->base() == Type::RawPtr) {
1923 return Type::RawPtr;
1924 } else if (base_type->isa_oopptr()) {
1925 // Base is never null => always a heap address.
1926 if (base_type->ptr() == TypePtr::NotNull) {
1927 return Type::OopPtr;
1928 }
1929 // Offset is small => always a heap address.
1930 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1931 if (offset_type != NULL &&
1932 base_type->offset() == 0 && // (should always be?)
1933 offset_type->_lo >= 0 &&
1934 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1935 return Type::OopPtr;
1936 }
1937 // Otherwise, it might either be oop+off or NULL+addr.
1938 return Type::AnyPtr;
1939 } else {
1940 // No information:
1941 return Type::AnyPtr;
1942 }
1943 }
1945 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1946 int kind = classify_unsafe_addr(base, offset);
1947 if (kind == Type::RawPtr) {
1948 return basic_plus_adr(top(), base, offset);
1949 } else {
1950 return basic_plus_adr(base, offset);
1951 }
1952 }
1954 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1955 // inline int Integer.numberOfLeadingZeros(int)
1956 // inline int Long.numberOfLeadingZeros(long)
1957 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1958 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1959 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1960 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1961 _sp += arg_size(); // restore stack pointer
1962 switch (id) {
1963 case vmIntrinsics::_numberOfLeadingZeros_i:
1964 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1965 break;
1966 case vmIntrinsics::_numberOfLeadingZeros_l:
1967 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1968 break;
1969 default:
1970 ShouldNotReachHere();
1971 }
1972 return true;
1973 }
1975 //-------------------inline_numberOfTrailingZeros_int/long----------------------
1976 // inline int Integer.numberOfTrailingZeros(int)
1977 // inline int Long.numberOfTrailingZeros(long)
1978 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
1979 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
1980 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
1981 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
1982 _sp += arg_size(); // restore stack pointer
1983 switch (id) {
1984 case vmIntrinsics::_numberOfTrailingZeros_i:
1985 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
1986 break;
1987 case vmIntrinsics::_numberOfTrailingZeros_l:
1988 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
1989 break;
1990 default:
1991 ShouldNotReachHere();
1992 }
1993 return true;
1994 }
1996 //----------------------------inline_bitCount_int/long-----------------------
1997 // inline int Integer.bitCount(int)
1998 // inline int Long.bitCount(long)
1999 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2000 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2001 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2002 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2003 _sp += arg_size(); // restore stack pointer
2004 switch (id) {
2005 case vmIntrinsics::_bitCount_i:
2006 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2007 break;
2008 case vmIntrinsics::_bitCount_l:
2009 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2010 break;
2011 default:
2012 ShouldNotReachHere();
2013 }
2014 return true;
2015 }
2017 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2018 // inline Integer.reverseBytes(int)
2019 // inline Long.reverseBytes(long)
2020 // inline Character.reverseBytes(char)
2021 // inline Short.reverseBytes(short)
2022 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2023 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2024 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2025 "not reverse Bytes");
2026 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2027 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2028 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2029 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2030 _sp += arg_size(); // restore stack pointer
2031 switch (id) {
2032 case vmIntrinsics::_reverseBytes_i:
2033 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2034 break;
2035 case vmIntrinsics::_reverseBytes_l:
2036 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2037 break;
2038 case vmIntrinsics::_reverseBytes_c:
2039 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2040 break;
2041 case vmIntrinsics::_reverseBytes_s:
2042 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2043 break;
2044 default:
2045 ;
2046 }
2047 return true;
2048 }
2050 //----------------------------inline_unsafe_access----------------------------
2052 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2054 // Interpret Unsafe.fieldOffset cookies correctly:
2055 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2057 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2058 if (callee()->is_static()) return false; // caller must have the capability!
2060 #ifndef PRODUCT
2061 {
2062 ResourceMark rm;
2063 // Check the signatures.
2064 ciSignature* sig = signature();
2065 #ifdef ASSERT
2066 if (!is_store) {
2067 // Object getObject(Object base, int/long offset), etc.
2068 BasicType rtype = sig->return_type()->basic_type();
2069 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2070 rtype = T_ADDRESS; // it is really a C void*
2071 assert(rtype == type, "getter must return the expected value");
2072 if (!is_native_ptr) {
2073 assert(sig->count() == 2, "oop getter has 2 arguments");
2074 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2075 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2076 } else {
2077 assert(sig->count() == 1, "native getter has 1 argument");
2078 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2079 }
2080 } else {
2081 // void putObject(Object base, int/long offset, Object x), etc.
2082 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2083 if (!is_native_ptr) {
2084 assert(sig->count() == 3, "oop putter has 3 arguments");
2085 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2086 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2087 } else {
2088 assert(sig->count() == 2, "native putter has 2 arguments");
2089 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2090 }
2091 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2092 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2093 vtype = T_ADDRESS; // it is really a C void*
2094 assert(vtype == type, "putter must accept the expected value");
2095 }
2096 #endif // ASSERT
2097 }
2098 #endif //PRODUCT
2100 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2102 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2104 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2105 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2107 debug_only(int saved_sp = _sp);
2108 _sp += nargs;
2110 Node* val;
2111 debug_only(val = (Node*)(uintptr_t)-1);
2114 if (is_store) {
2115 // Get the value being stored. (Pop it first; it was pushed last.)
2116 switch (type) {
2117 case T_DOUBLE:
2118 case T_LONG:
2119 case T_ADDRESS:
2120 val = pop_pair();
2121 break;
2122 default:
2123 val = pop();
2124 }
2125 }
2127 // Build address expression. See the code in inline_unsafe_prefetch.
2128 Node *adr;
2129 Node *heap_base_oop = top();
2130 if (!is_native_ptr) {
2131 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2132 Node* offset = pop_pair();
2133 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2134 Node* base = pop();
2135 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2136 // to be plain byte offsets, which are also the same as those accepted
2137 // by oopDesc::field_base.
2138 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2139 "fieldOffset must be byte-scaled");
2140 // 32-bit machines ignore the high half!
2141 offset = ConvL2X(offset);
2142 adr = make_unsafe_address(base, offset);
2143 heap_base_oop = base;
2144 } else {
2145 Node* ptr = pop_pair();
2146 // Adjust Java long to machine word:
2147 ptr = ConvL2X(ptr);
2148 adr = make_unsafe_address(NULL, ptr);
2149 }
2151 // Pop receiver last: it was pushed first.
2152 Node *receiver = pop();
2154 assert(saved_sp == _sp, "must have correct argument count");
2156 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2158 // First guess at the value type.
2159 const Type *value_type = Type::get_const_basic_type(type);
2161 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2162 // there was not enough information to nail it down.
2163 Compile::AliasType* alias_type = C->alias_type(adr_type);
2164 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2166 // We will need memory barriers unless we can determine a unique
2167 // alias category for this reference. (Note: If for some reason
2168 // the barriers get omitted and the unsafe reference begins to "pollute"
2169 // the alias analysis of the rest of the graph, either Compile::can_alias
2170 // or Compile::must_alias will throw a diagnostic assert.)
2171 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2173 if (!is_store && type == T_OBJECT) {
2174 // Attempt to infer a sharper value type from the offset and base type.
2175 ciKlass* sharpened_klass = NULL;
2177 // See if it is an instance field, with an object type.
2178 if (alias_type->field() != NULL) {
2179 assert(!is_native_ptr, "native pointer op cannot use a java address");
2180 if (alias_type->field()->type()->is_klass()) {
2181 sharpened_klass = alias_type->field()->type()->as_klass();
2182 }
2183 }
2185 // See if it is a narrow oop array.
2186 if (adr_type->isa_aryptr()) {
2187 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2188 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2189 if (elem_type != NULL) {
2190 sharpened_klass = elem_type->klass();
2191 }
2192 }
2193 }
2195 if (sharpened_klass != NULL) {
2196 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2198 // Sharpen the value type.
2199 value_type = tjp;
2201 #ifndef PRODUCT
2202 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2203 tty->print(" from base type: "); adr_type->dump();
2204 tty->print(" sharpened value: "); value_type->dump();
2205 }
2206 #endif
2207 }
2208 }
2210 // Null check on self without removing any arguments. The argument
2211 // null check technically happens in the wrong place, which can lead to
2212 // invalid stack traces when the primitive is inlined into a method
2213 // which handles NullPointerExceptions.
2214 _sp += nargs;
2215 do_null_check(receiver, T_OBJECT);
2216 _sp -= nargs;
2217 if (stopped()) {
2218 return true;
2219 }
2220 // Heap pointers get a null-check from the interpreter,
2221 // as a courtesy. However, this is not guaranteed by Unsafe,
2222 // and it is not possible to fully distinguish unintended nulls
2223 // from intended ones in this API.
2225 if (is_volatile) {
2226 // We need to emit leading and trailing CPU membars (see below) in
2227 // addition to memory membars when is_volatile. This is a little
2228 // too strong, but avoids the need to insert per-alias-type
2229 // volatile membars (for stores; compare Parse::do_put_xxx), which
2230 // we cannot do effectively here because we probably only have a
2231 // rough approximation of type.
2232 need_mem_bar = true;
2233 // For Stores, place a memory ordering barrier now.
2234 if (is_store)
2235 insert_mem_bar(Op_MemBarRelease);
2236 }
2238 // Memory barrier to prevent normal and 'unsafe' accesses from
2239 // bypassing each other. Happens after null checks, so the
2240 // exception paths do not take memory state from the memory barrier,
2241 // so there's no problems making a strong assert about mixing users
2242 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2243 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2244 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2246 if (!is_store) {
2247 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2248 // load value and push onto stack
2249 switch (type) {
2250 case T_BOOLEAN:
2251 case T_CHAR:
2252 case T_BYTE:
2253 case T_SHORT:
2254 case T_INT:
2255 case T_FLOAT:
2256 case T_OBJECT:
2257 push( p );
2258 break;
2259 case T_ADDRESS:
2260 // Cast to an int type.
2261 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2262 p = ConvX2L(p);
2263 push_pair(p);
2264 break;
2265 case T_DOUBLE:
2266 case T_LONG:
2267 push_pair( p );
2268 break;
2269 default: ShouldNotReachHere();
2270 }
2271 } else {
2272 // place effect of store into memory
2273 switch (type) {
2274 case T_DOUBLE:
2275 val = dstore_rounding(val);
2276 break;
2277 case T_ADDRESS:
2278 // Repackage the long as a pointer.
2279 val = ConvL2X(val);
2280 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2281 break;
2282 }
2284 if (type != T_OBJECT ) {
2285 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2286 } else {
2287 // Possibly an oop being stored to Java heap or native memory
2288 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2289 // oop to Java heap.
2290 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2291 } else {
2292 // We can't tell at compile time if we are storing in the Java heap or outside
2293 // of it. So we need to emit code to conditionally do the proper type of
2294 // store.
2296 IdealKit ideal(gvn(), control(), merged_memory());
2297 #define __ ideal.
2298 // QQQ who knows what probability is here??
2299 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2300 // Sync IdealKit and graphKit.
2301 set_all_memory( __ merged_memory());
2302 set_control(__ ctrl());
2303 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2304 // Update IdealKit memory.
2305 __ set_all_memory(merged_memory());
2306 __ set_ctrl(control());
2307 } __ else_(); {
2308 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2309 } __ end_if();
2310 // Final sync IdealKit and GraphKit.
2311 sync_kit(ideal);
2312 #undef __
2313 }
2314 }
2315 }
2317 if (is_volatile) {
2318 if (!is_store)
2319 insert_mem_bar(Op_MemBarAcquire);
2320 else
2321 insert_mem_bar(Op_MemBarVolatile);
2322 }
2324 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2326 return true;
2327 }
2329 //----------------------------inline_unsafe_prefetch----------------------------
2331 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2332 #ifndef PRODUCT
2333 {
2334 ResourceMark rm;
2335 // Check the signatures.
2336 ciSignature* sig = signature();
2337 #ifdef ASSERT
2338 // Object getObject(Object base, int/long offset), etc.
2339 BasicType rtype = sig->return_type()->basic_type();
2340 if (!is_native_ptr) {
2341 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2342 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2343 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2344 } else {
2345 assert(sig->count() == 1, "native prefetch has 1 argument");
2346 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2347 }
2348 #endif // ASSERT
2349 }
2350 #endif // !PRODUCT
2352 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2354 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2355 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2357 debug_only(int saved_sp = _sp);
2358 _sp += nargs;
2360 // Build address expression. See the code in inline_unsafe_access.
2361 Node *adr;
2362 if (!is_native_ptr) {
2363 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2364 Node* offset = pop_pair();
2365 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2366 Node* base = pop();
2367 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2368 // to be plain byte offsets, which are also the same as those accepted
2369 // by oopDesc::field_base.
2370 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2371 "fieldOffset must be byte-scaled");
2372 // 32-bit machines ignore the high half!
2373 offset = ConvL2X(offset);
2374 adr = make_unsafe_address(base, offset);
2375 } else {
2376 Node* ptr = pop_pair();
2377 // Adjust Java long to machine word:
2378 ptr = ConvL2X(ptr);
2379 adr = make_unsafe_address(NULL, ptr);
2380 }
2382 if (is_static) {
2383 assert(saved_sp == _sp, "must have correct argument count");
2384 } else {
2385 // Pop receiver last: it was pushed first.
2386 Node *receiver = pop();
2387 assert(saved_sp == _sp, "must have correct argument count");
2389 // Null check on self without removing any arguments. The argument
2390 // null check technically happens in the wrong place, which can lead to
2391 // invalid stack traces when the primitive is inlined into a method
2392 // which handles NullPointerExceptions.
2393 _sp += nargs;
2394 do_null_check(receiver, T_OBJECT);
2395 _sp -= nargs;
2396 if (stopped()) {
2397 return true;
2398 }
2399 }
2401 // Generate the read or write prefetch
2402 Node *prefetch;
2403 if (is_store) {
2404 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2405 } else {
2406 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2407 }
2408 prefetch->init_req(0, control());
2409 set_i_o(_gvn.transform(prefetch));
2411 return true;
2412 }
2414 //----------------------------inline_unsafe_CAS----------------------------
2416 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2417 // This basic scheme here is the same as inline_unsafe_access, but
2418 // differs in enough details that combining them would make the code
2419 // overly confusing. (This is a true fact! I originally combined
2420 // them, but even I was confused by it!) As much code/comments as
2421 // possible are retained from inline_unsafe_access though to make
2422 // the correspondences clearer. - dl
2424 if (callee()->is_static()) return false; // caller must have the capability!
2426 #ifndef PRODUCT
2427 {
2428 ResourceMark rm;
2429 // Check the signatures.
2430 ciSignature* sig = signature();
2431 #ifdef ASSERT
2432 BasicType rtype = sig->return_type()->basic_type();
2433 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2434 assert(sig->count() == 4, "CAS has 4 arguments");
2435 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2436 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2437 #endif // ASSERT
2438 }
2439 #endif //PRODUCT
2441 // number of stack slots per value argument (1 or 2)
2442 int type_words = type2size[type];
2444 // Cannot inline wide CAS on machines that don't support it natively
2445 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2446 return false;
2448 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2450 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2451 int nargs = 1 + 1 + 2 + type_words + type_words;
2453 // pop arguments: newval, oldval, offset, base, and receiver
2454 debug_only(int saved_sp = _sp);
2455 _sp += nargs;
2456 Node* newval = (type_words == 1) ? pop() : pop_pair();
2457 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2458 Node *offset = pop_pair();
2459 Node *base = pop();
2460 Node *receiver = pop();
2461 assert(saved_sp == _sp, "must have correct argument count");
2463 // Null check receiver.
2464 _sp += nargs;
2465 do_null_check(receiver, T_OBJECT);
2466 _sp -= nargs;
2467 if (stopped()) {
2468 return true;
2469 }
2471 // Build field offset expression.
2472 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2473 // to be plain byte offsets, which are also the same as those accepted
2474 // by oopDesc::field_base.
2475 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2476 // 32-bit machines ignore the high half of long offsets
2477 offset = ConvL2X(offset);
2478 Node* adr = make_unsafe_address(base, offset);
2479 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2481 // (Unlike inline_unsafe_access, there seems no point in trying
2482 // to refine types. Just use the coarse types here.
2483 const Type *value_type = Type::get_const_basic_type(type);
2484 Compile::AliasType* alias_type = C->alias_type(adr_type);
2485 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2486 int alias_idx = C->get_alias_index(adr_type);
2488 // Memory-model-wise, a CAS acts like a little synchronized block,
2489 // so needs barriers on each side. These don't translate into
2490 // actual barriers on most machines, but we still need rest of
2491 // compiler to respect ordering.
2493 insert_mem_bar(Op_MemBarRelease);
2494 insert_mem_bar(Op_MemBarCPUOrder);
2496 // 4984716: MemBars must be inserted before this
2497 // memory node in order to avoid a false
2498 // dependency which will confuse the scheduler.
2499 Node *mem = memory(alias_idx);
2501 // For now, we handle only those cases that actually exist: ints,
2502 // longs, and Object. Adding others should be straightforward.
2503 Node* cas;
2504 switch(type) {
2505 case T_INT:
2506 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2507 break;
2508 case T_LONG:
2509 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2510 break;
2511 case T_OBJECT:
2512 // reference stores need a store barrier.
2513 // (They don't if CAS fails, but it isn't worth checking.)
2514 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2515 #ifdef _LP64
2516 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2517 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2518 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2519 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2520 newval_enc, oldval_enc));
2521 } else
2522 #endif
2523 {
2524 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2525 }
2526 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2527 break;
2528 default:
2529 ShouldNotReachHere();
2530 break;
2531 }
2533 // SCMemProjNodes represent the memory state of CAS. Their main
2534 // role is to prevent CAS nodes from being optimized away when their
2535 // results aren't used.
2536 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2537 set_memory(proj, alias_idx);
2539 // Add the trailing membar surrounding the access
2540 insert_mem_bar(Op_MemBarCPUOrder);
2541 insert_mem_bar(Op_MemBarAcquire);
2543 push(cas);
2544 return true;
2545 }
2547 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2548 // This is another variant of inline_unsafe_access, differing in
2549 // that it always issues store-store ("release") barrier and ensures
2550 // store-atomicity (which only matters for "long").
2552 if (callee()->is_static()) return false; // caller must have the capability!
2554 #ifndef PRODUCT
2555 {
2556 ResourceMark rm;
2557 // Check the signatures.
2558 ciSignature* sig = signature();
2559 #ifdef ASSERT
2560 BasicType rtype = sig->return_type()->basic_type();
2561 assert(rtype == T_VOID, "must return void");
2562 assert(sig->count() == 3, "has 3 arguments");
2563 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2564 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2565 #endif // ASSERT
2566 }
2567 #endif //PRODUCT
2569 // number of stack slots per value argument (1 or 2)
2570 int type_words = type2size[type];
2572 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2574 // Argument words: "this" plus oop plus offset plus value;
2575 int nargs = 1 + 1 + 2 + type_words;
2577 // pop arguments: val, offset, base, and receiver
2578 debug_only(int saved_sp = _sp);
2579 _sp += nargs;
2580 Node* val = (type_words == 1) ? pop() : pop_pair();
2581 Node *offset = pop_pair();
2582 Node *base = pop();
2583 Node *receiver = pop();
2584 assert(saved_sp == _sp, "must have correct argument count");
2586 // Null check receiver.
2587 _sp += nargs;
2588 do_null_check(receiver, T_OBJECT);
2589 _sp -= nargs;
2590 if (stopped()) {
2591 return true;
2592 }
2594 // Build field offset expression.
2595 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2596 // 32-bit machines ignore the high half of long offsets
2597 offset = ConvL2X(offset);
2598 Node* adr = make_unsafe_address(base, offset);
2599 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2600 const Type *value_type = Type::get_const_basic_type(type);
2601 Compile::AliasType* alias_type = C->alias_type(adr_type);
2603 insert_mem_bar(Op_MemBarRelease);
2604 insert_mem_bar(Op_MemBarCPUOrder);
2605 // Ensure that the store is atomic for longs:
2606 bool require_atomic_access = true;
2607 Node* store;
2608 if (type == T_OBJECT) // reference stores need a store barrier.
2609 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2610 else {
2611 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2612 }
2613 insert_mem_bar(Op_MemBarCPUOrder);
2614 return true;
2615 }
2617 bool LibraryCallKit::inline_unsafe_allocate() {
2618 if (callee()->is_static()) return false; // caller must have the capability!
2619 int nargs = 1 + 1;
2620 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2621 null_check_receiver(callee()); // check then ignore argument(0)
2622 _sp += nargs; // set original stack for use by uncommon_trap
2623 Node* cls = do_null_check(argument(1), T_OBJECT);
2624 _sp -= nargs;
2625 if (stopped()) return true;
2627 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2628 _sp += nargs; // set original stack for use by uncommon_trap
2629 kls = do_null_check(kls, T_OBJECT);
2630 _sp -= nargs;
2631 if (stopped()) return true; // argument was like int.class
2633 // Note: The argument might still be an illegal value like
2634 // Serializable.class or Object[].class. The runtime will handle it.
2635 // But we must make an explicit check for initialization.
2636 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2637 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2638 Node* bits = intcon(instanceKlass::fully_initialized);
2639 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2640 // The 'test' is non-zero if we need to take a slow path.
2642 Node* obj = new_instance(kls, test);
2643 push(obj);
2645 return true;
2646 }
2648 //------------------------inline_native_time_funcs--------------
2649 // inline code for System.currentTimeMillis() and System.nanoTime()
2650 // these have the same type and signature
2651 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2652 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2653 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2654 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2655 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2656 const TypePtr* no_memory_effects = NULL;
2657 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2658 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2659 #ifdef ASSERT
2660 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2661 assert(value_top == top(), "second value must be top");
2662 #endif
2663 push_pair(value);
2664 return true;
2665 }
2667 //------------------------inline_native_currentThread------------------
2668 bool LibraryCallKit::inline_native_currentThread() {
2669 Node* junk = NULL;
2670 push(generate_current_thread(junk));
2671 return true;
2672 }
2674 //------------------------inline_native_isInterrupted------------------
2675 bool LibraryCallKit::inline_native_isInterrupted() {
2676 const int nargs = 1+1; // receiver + boolean
2677 assert(nargs == arg_size(), "sanity");
2678 // Add a fast path to t.isInterrupted(clear_int):
2679 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2680 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2681 // So, in the common case that the interrupt bit is false,
2682 // we avoid making a call into the VM. Even if the interrupt bit
2683 // is true, if the clear_int argument is false, we avoid the VM call.
2684 // However, if the receiver is not currentThread, we must call the VM,
2685 // because there must be some locking done around the operation.
2687 // We only go to the fast case code if we pass two guards.
2688 // Paths which do not pass are accumulated in the slow_region.
2689 RegionNode* slow_region = new (C, 1) RegionNode(1);
2690 record_for_igvn(slow_region);
2691 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2692 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2693 enum { no_int_result_path = 1,
2694 no_clear_result_path = 2,
2695 slow_result_path = 3
2696 };
2698 // (a) Receiving thread must be the current thread.
2699 Node* rec_thr = argument(0);
2700 Node* tls_ptr = NULL;
2701 Node* cur_thr = generate_current_thread(tls_ptr);
2702 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2703 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2705 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2706 if (!known_current_thread)
2707 generate_slow_guard(bol_thr, slow_region);
2709 // (b) Interrupt bit on TLS must be false.
2710 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2711 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2712 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2713 // Set the control input on the field _interrupted read to prevent it floating up.
2714 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2715 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2716 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2718 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2720 // First fast path: if (!TLS._interrupted) return false;
2721 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2722 result_rgn->init_req(no_int_result_path, false_bit);
2723 result_val->init_req(no_int_result_path, intcon(0));
2725 // drop through to next case
2726 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2728 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2729 Node* clr_arg = argument(1);
2730 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2731 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2732 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2734 // Second fast path: ... else if (!clear_int) return true;
2735 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2736 result_rgn->init_req(no_clear_result_path, false_arg);
2737 result_val->init_req(no_clear_result_path, intcon(1));
2739 // drop through to next case
2740 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2742 // (d) Otherwise, go to the slow path.
2743 slow_region->add_req(control());
2744 set_control( _gvn.transform(slow_region) );
2746 if (stopped()) {
2747 // There is no slow path.
2748 result_rgn->init_req(slow_result_path, top());
2749 result_val->init_req(slow_result_path, top());
2750 } else {
2751 // non-virtual because it is a private non-static
2752 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2754 Node* slow_val = set_results_for_java_call(slow_call);
2755 // this->control() comes from set_results_for_java_call
2757 // If we know that the result of the slow call will be true, tell the optimizer!
2758 if (known_current_thread) slow_val = intcon(1);
2760 Node* fast_io = slow_call->in(TypeFunc::I_O);
2761 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2762 // These two phis are pre-filled with copies of of the fast IO and Memory
2763 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2764 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2766 result_rgn->init_req(slow_result_path, control());
2767 io_phi ->init_req(slow_result_path, i_o());
2768 mem_phi ->init_req(slow_result_path, reset_memory());
2769 result_val->init_req(slow_result_path, slow_val);
2771 set_all_memory( _gvn.transform(mem_phi) );
2772 set_i_o( _gvn.transform(io_phi) );
2773 }
2775 push_result(result_rgn, result_val);
2776 C->set_has_split_ifs(true); // Has chance for split-if optimization
2778 return true;
2779 }
2781 //---------------------------load_mirror_from_klass----------------------------
2782 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2783 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2784 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2785 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2786 }
2788 //-----------------------load_klass_from_mirror_common-------------------------
2789 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2790 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2791 // and branch to the given path on the region.
2792 // If never_see_null, take an uncommon trap on null, so we can optimistically
2793 // compile for the non-null case.
2794 // If the region is NULL, force never_see_null = true.
2795 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2796 bool never_see_null,
2797 int nargs,
2798 RegionNode* region,
2799 int null_path,
2800 int offset) {
2801 if (region == NULL) never_see_null = true;
2802 Node* p = basic_plus_adr(mirror, offset);
2803 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2804 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2805 _sp += nargs; // any deopt will start just before call to enclosing method
2806 Node* null_ctl = top();
2807 kls = null_check_oop(kls, &null_ctl, never_see_null);
2808 if (region != NULL) {
2809 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2810 region->init_req(null_path, null_ctl);
2811 } else {
2812 assert(null_ctl == top(), "no loose ends");
2813 }
2814 _sp -= nargs;
2815 return kls;
2816 }
2818 //--------------------(inline_native_Class_query helpers)---------------------
2819 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2820 // Fall through if (mods & mask) == bits, take the guard otherwise.
2821 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2822 // Branch around if the given klass has the given modifier bit set.
2823 // Like generate_guard, adds a new path onto the region.
2824 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2825 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2826 Node* mask = intcon(modifier_mask);
2827 Node* bits = intcon(modifier_bits);
2828 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2829 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2830 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2831 return generate_fair_guard(bol, region);
2832 }
2833 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2834 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2835 }
2837 //-------------------------inline_native_Class_query-------------------
2838 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2839 int nargs = 1+0; // just the Class mirror, in most cases
2840 const Type* return_type = TypeInt::BOOL;
2841 Node* prim_return_value = top(); // what happens if it's a primitive class?
2842 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2843 bool expect_prim = false; // most of these guys expect to work on refs
2845 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2847 switch (id) {
2848 case vmIntrinsics::_isInstance:
2849 nargs = 1+1; // the Class mirror, plus the object getting queried about
2850 // nothing is an instance of a primitive type
2851 prim_return_value = intcon(0);
2852 break;
2853 case vmIntrinsics::_getModifiers:
2854 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2855 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2856 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2857 break;
2858 case vmIntrinsics::_isInterface:
2859 prim_return_value = intcon(0);
2860 break;
2861 case vmIntrinsics::_isArray:
2862 prim_return_value = intcon(0);
2863 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2864 break;
2865 case vmIntrinsics::_isPrimitive:
2866 prim_return_value = intcon(1);
2867 expect_prim = true; // obviously
2868 break;
2869 case vmIntrinsics::_getSuperclass:
2870 prim_return_value = null();
2871 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2872 break;
2873 case vmIntrinsics::_getComponentType:
2874 prim_return_value = null();
2875 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2876 break;
2877 case vmIntrinsics::_getClassAccessFlags:
2878 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2879 return_type = TypeInt::INT; // not bool! 6297094
2880 break;
2881 default:
2882 ShouldNotReachHere();
2883 }
2885 Node* mirror = argument(0);
2886 Node* obj = (nargs <= 1)? top(): argument(1);
2888 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2889 if (mirror_con == NULL) return false; // cannot happen?
2891 #ifndef PRODUCT
2892 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2893 ciType* k = mirror_con->java_mirror_type();
2894 if (k) {
2895 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2896 k->print_name();
2897 tty->cr();
2898 }
2899 }
2900 #endif
2902 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2903 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2904 record_for_igvn(region);
2905 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2907 // The mirror will never be null of Reflection.getClassAccessFlags, however
2908 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2909 // if it is. See bug 4774291.
2911 // For Reflection.getClassAccessFlags(), the null check occurs in
2912 // the wrong place; see inline_unsafe_access(), above, for a similar
2913 // situation.
2914 _sp += nargs; // set original stack for use by uncommon_trap
2915 mirror = do_null_check(mirror, T_OBJECT);
2916 _sp -= nargs;
2917 // If mirror or obj is dead, only null-path is taken.
2918 if (stopped()) return true;
2920 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2922 // Now load the mirror's klass metaobject, and null-check it.
2923 // Side-effects region with the control path if the klass is null.
2924 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2925 region, _prim_path);
2926 // If kls is null, we have a primitive mirror.
2927 phi->init_req(_prim_path, prim_return_value);
2928 if (stopped()) { push_result(region, phi); return true; }
2930 Node* p; // handy temp
2931 Node* null_ctl;
2933 // Now that we have the non-null klass, we can perform the real query.
2934 // For constant classes, the query will constant-fold in LoadNode::Value.
2935 Node* query_value = top();
2936 switch (id) {
2937 case vmIntrinsics::_isInstance:
2938 // nothing is an instance of a primitive type
2939 query_value = gen_instanceof(obj, kls);
2940 break;
2942 case vmIntrinsics::_getModifiers:
2943 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2944 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2945 break;
2947 case vmIntrinsics::_isInterface:
2948 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2949 if (generate_interface_guard(kls, region) != NULL)
2950 // A guard was added. If the guard is taken, it was an interface.
2951 phi->add_req(intcon(1));
2952 // If we fall through, it's a plain class.
2953 query_value = intcon(0);
2954 break;
2956 case vmIntrinsics::_isArray:
2957 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2958 if (generate_array_guard(kls, region) != NULL)
2959 // A guard was added. If the guard is taken, it was an array.
2960 phi->add_req(intcon(1));
2961 // If we fall through, it's a plain class.
2962 query_value = intcon(0);
2963 break;
2965 case vmIntrinsics::_isPrimitive:
2966 query_value = intcon(0); // "normal" path produces false
2967 break;
2969 case vmIntrinsics::_getSuperclass:
2970 // The rules here are somewhat unfortunate, but we can still do better
2971 // with random logic than with a JNI call.
2972 // Interfaces store null or Object as _super, but must report null.
2973 // Arrays store an intermediate super as _super, but must report Object.
2974 // Other types can report the actual _super.
2975 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2976 if (generate_interface_guard(kls, region) != NULL)
2977 // A guard was added. If the guard is taken, it was an interface.
2978 phi->add_req(null());
2979 if (generate_array_guard(kls, region) != NULL)
2980 // A guard was added. If the guard is taken, it was an array.
2981 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2982 // If we fall through, it's a plain class. Get its _super.
2983 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2984 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2985 null_ctl = top();
2986 kls = null_check_oop(kls, &null_ctl);
2987 if (null_ctl != top()) {
2988 // If the guard is taken, Object.superClass is null (both klass and mirror).
2989 region->add_req(null_ctl);
2990 phi ->add_req(null());
2991 }
2992 if (!stopped()) {
2993 query_value = load_mirror_from_klass(kls);
2994 }
2995 break;
2997 case vmIntrinsics::_getComponentType:
2998 if (generate_array_guard(kls, region) != NULL) {
2999 // Be sure to pin the oop load to the guard edge just created:
3000 Node* is_array_ctrl = region->in(region->req()-1);
3001 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3002 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3003 phi->add_req(cmo);
3004 }
3005 query_value = null(); // non-array case is null
3006 break;
3008 case vmIntrinsics::_getClassAccessFlags:
3009 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3010 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3011 break;
3013 default:
3014 ShouldNotReachHere();
3015 }
3017 // Fall-through is the normal case of a query to a real class.
3018 phi->init_req(1, query_value);
3019 region->init_req(1, control());
3021 push_result(region, phi);
3022 C->set_has_split_ifs(true); // Has chance for split-if optimization
3024 return true;
3025 }
3027 //--------------------------inline_native_subtype_check------------------------
3028 // This intrinsic takes the JNI calls out of the heart of
3029 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3030 bool LibraryCallKit::inline_native_subtype_check() {
3031 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3033 // Pull both arguments off the stack.
3034 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3035 args[0] = argument(0);
3036 args[1] = argument(1);
3037 Node* klasses[2]; // corresponding Klasses: superk, subk
3038 klasses[0] = klasses[1] = top();
3040 enum {
3041 // A full decision tree on {superc is prim, subc is prim}:
3042 _prim_0_path = 1, // {P,N} => false
3043 // {P,P} & superc!=subc => false
3044 _prim_same_path, // {P,P} & superc==subc => true
3045 _prim_1_path, // {N,P} => false
3046 _ref_subtype_path, // {N,N} & subtype check wins => true
3047 _both_ref_path, // {N,N} & subtype check loses => false
3048 PATH_LIMIT
3049 };
3051 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3052 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3053 record_for_igvn(region);
3055 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3056 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3057 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3059 // First null-check both mirrors and load each mirror's klass metaobject.
3060 int which_arg;
3061 for (which_arg = 0; which_arg <= 1; which_arg++) {
3062 Node* arg = args[which_arg];
3063 _sp += nargs; // set original stack for use by uncommon_trap
3064 arg = do_null_check(arg, T_OBJECT);
3065 _sp -= nargs;
3066 if (stopped()) break;
3067 args[which_arg] = _gvn.transform(arg);
3069 Node* p = basic_plus_adr(arg, class_klass_offset);
3070 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3071 klasses[which_arg] = _gvn.transform(kls);
3072 }
3074 // Having loaded both klasses, test each for null.
3075 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3076 for (which_arg = 0; which_arg <= 1; which_arg++) {
3077 Node* kls = klasses[which_arg];
3078 Node* null_ctl = top();
3079 _sp += nargs; // set original stack for use by uncommon_trap
3080 kls = null_check_oop(kls, &null_ctl, never_see_null);
3081 _sp -= nargs;
3082 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3083 region->init_req(prim_path, null_ctl);
3084 if (stopped()) break;
3085 klasses[which_arg] = kls;
3086 }
3088 if (!stopped()) {
3089 // now we have two reference types, in klasses[0..1]
3090 Node* subk = klasses[1]; // the argument to isAssignableFrom
3091 Node* superk = klasses[0]; // the receiver
3092 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3093 // now we have a successful reference subtype check
3094 region->set_req(_ref_subtype_path, control());
3095 }
3097 // If both operands are primitive (both klasses null), then
3098 // we must return true when they are identical primitives.
3099 // It is convenient to test this after the first null klass check.
3100 set_control(region->in(_prim_0_path)); // go back to first null check
3101 if (!stopped()) {
3102 // Since superc is primitive, make a guard for the superc==subc case.
3103 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3104 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3105 generate_guard(bol_eq, region, PROB_FAIR);
3106 if (region->req() == PATH_LIMIT+1) {
3107 // A guard was added. If the added guard is taken, superc==subc.
3108 region->swap_edges(PATH_LIMIT, _prim_same_path);
3109 region->del_req(PATH_LIMIT);
3110 }
3111 region->set_req(_prim_0_path, control()); // Not equal after all.
3112 }
3114 // these are the only paths that produce 'true':
3115 phi->set_req(_prim_same_path, intcon(1));
3116 phi->set_req(_ref_subtype_path, intcon(1));
3118 // pull together the cases:
3119 assert(region->req() == PATH_LIMIT, "sane region");
3120 for (uint i = 1; i < region->req(); i++) {
3121 Node* ctl = region->in(i);
3122 if (ctl == NULL || ctl == top()) {
3123 region->set_req(i, top());
3124 phi ->set_req(i, top());
3125 } else if (phi->in(i) == NULL) {
3126 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3127 }
3128 }
3130 set_control(_gvn.transform(region));
3131 push(_gvn.transform(phi));
3133 return true;
3134 }
3136 //---------------------generate_array_guard_common------------------------
3137 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3138 bool obj_array, bool not_array) {
3139 // If obj_array/non_array==false/false:
3140 // Branch around if the given klass is in fact an array (either obj or prim).
3141 // If obj_array/non_array==false/true:
3142 // Branch around if the given klass is not an array klass of any kind.
3143 // If obj_array/non_array==true/true:
3144 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3145 // If obj_array/non_array==true/false:
3146 // Branch around if the kls is an oop array (Object[] or subtype)
3147 //
3148 // Like generate_guard, adds a new path onto the region.
3149 jint layout_con = 0;
3150 Node* layout_val = get_layout_helper(kls, layout_con);
3151 if (layout_val == NULL) {
3152 bool query = (obj_array
3153 ? Klass::layout_helper_is_objArray(layout_con)
3154 : Klass::layout_helper_is_javaArray(layout_con));
3155 if (query == not_array) {
3156 return NULL; // never a branch
3157 } else { // always a branch
3158 Node* always_branch = control();
3159 if (region != NULL)
3160 region->add_req(always_branch);
3161 set_control(top());
3162 return always_branch;
3163 }
3164 }
3165 // Now test the correct condition.
3166 jint nval = (obj_array
3167 ? ((jint)Klass::_lh_array_tag_type_value
3168 << Klass::_lh_array_tag_shift)
3169 : Klass::_lh_neutral_value);
3170 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3171 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3172 // invert the test if we are looking for a non-array
3173 if (not_array) btest = BoolTest(btest).negate();
3174 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3175 return generate_fair_guard(bol, region);
3176 }
3179 //-----------------------inline_native_newArray--------------------------
3180 bool LibraryCallKit::inline_native_newArray() {
3181 int nargs = 2;
3182 Node* mirror = argument(0);
3183 Node* count_val = argument(1);
3185 _sp += nargs; // set original stack for use by uncommon_trap
3186 mirror = do_null_check(mirror, T_OBJECT);
3187 _sp -= nargs;
3188 // If mirror or obj is dead, only null-path is taken.
3189 if (stopped()) return true;
3191 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3192 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3193 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3194 TypeInstPtr::NOTNULL);
3195 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3196 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3197 TypePtr::BOTTOM);
3199 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3200 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3201 nargs,
3202 result_reg, _slow_path);
3203 Node* normal_ctl = control();
3204 Node* no_array_ctl = result_reg->in(_slow_path);
3206 // Generate code for the slow case. We make a call to newArray().
3207 set_control(no_array_ctl);
3208 if (!stopped()) {
3209 // Either the input type is void.class, or else the
3210 // array klass has not yet been cached. Either the
3211 // ensuing call will throw an exception, or else it
3212 // will cache the array klass for next time.
3213 PreserveJVMState pjvms(this);
3214 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3215 Node* slow_result = set_results_for_java_call(slow_call);
3216 // this->control() comes from set_results_for_java_call
3217 result_reg->set_req(_slow_path, control());
3218 result_val->set_req(_slow_path, slow_result);
3219 result_io ->set_req(_slow_path, i_o());
3220 result_mem->set_req(_slow_path, reset_memory());
3221 }
3223 set_control(normal_ctl);
3224 if (!stopped()) {
3225 // Normal case: The array type has been cached in the java.lang.Class.
3226 // The following call works fine even if the array type is polymorphic.
3227 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3228 Node* obj = new_array(klass_node, count_val, nargs);
3229 result_reg->init_req(_normal_path, control());
3230 result_val->init_req(_normal_path, obj);
3231 result_io ->init_req(_normal_path, i_o());
3232 result_mem->init_req(_normal_path, reset_memory());
3233 }
3235 // Return the combined state.
3236 set_i_o( _gvn.transform(result_io) );
3237 set_all_memory( _gvn.transform(result_mem) );
3238 push_result(result_reg, result_val);
3239 C->set_has_split_ifs(true); // Has chance for split-if optimization
3241 return true;
3242 }
3244 //----------------------inline_native_getLength--------------------------
3245 bool LibraryCallKit::inline_native_getLength() {
3246 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3248 int nargs = 1;
3249 Node* array = argument(0);
3251 _sp += nargs; // set original stack for use by uncommon_trap
3252 array = do_null_check(array, T_OBJECT);
3253 _sp -= nargs;
3255 // If array is dead, only null-path is taken.
3256 if (stopped()) return true;
3258 // Deoptimize if it is a non-array.
3259 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3261 if (non_array != NULL) {
3262 PreserveJVMState pjvms(this);
3263 set_control(non_array);
3264 _sp += nargs; // push the arguments back on the stack
3265 uncommon_trap(Deoptimization::Reason_intrinsic,
3266 Deoptimization::Action_maybe_recompile);
3267 }
3269 // If control is dead, only non-array-path is taken.
3270 if (stopped()) return true;
3272 // The works fine even if the array type is polymorphic.
3273 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3274 push( load_array_length(array) );
3276 C->set_has_split_ifs(true); // Has chance for split-if optimization
3278 return true;
3279 }
3281 //------------------------inline_array_copyOf----------------------------
3282 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3283 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3285 // Restore the stack and pop off the arguments.
3286 int nargs = 3 + (is_copyOfRange? 1: 0);
3287 Node* original = argument(0);
3288 Node* start = is_copyOfRange? argument(1): intcon(0);
3289 Node* end = is_copyOfRange? argument(2): argument(1);
3290 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3292 Node* newcopy;
3294 //set the original stack and the reexecute bit for the interpreter to reexecute
3295 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3296 { PreserveReexecuteState preexecs(this);
3297 _sp += nargs;
3298 jvms()->set_should_reexecute(true);
3300 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3301 original = do_null_check(original, T_OBJECT);
3303 // Check if a null path was taken unconditionally.
3304 if (stopped()) return true;
3306 Node* orig_length = load_array_length(original);
3308 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3309 NULL, 0);
3310 klass_node = do_null_check(klass_node, T_OBJECT);
3312 RegionNode* bailout = new (C, 1) RegionNode(1);
3313 record_for_igvn(bailout);
3315 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3316 // Bail out if that is so.
3317 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3318 if (not_objArray != NULL) {
3319 // Improve the klass node's type from the new optimistic assumption:
3320 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3321 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3322 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3323 cast->init_req(0, control());
3324 klass_node = _gvn.transform(cast);
3325 }
3327 // Bail out if either start or end is negative.
3328 generate_negative_guard(start, bailout, &start);
3329 generate_negative_guard(end, bailout, &end);
3331 Node* length = end;
3332 if (_gvn.type(start) != TypeInt::ZERO) {
3333 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3334 }
3336 // Bail out if length is negative.
3337 // ...Not needed, since the new_array will throw the right exception.
3338 //generate_negative_guard(length, bailout, &length);
3340 if (bailout->req() > 1) {
3341 PreserveJVMState pjvms(this);
3342 set_control( _gvn.transform(bailout) );
3343 uncommon_trap(Deoptimization::Reason_intrinsic,
3344 Deoptimization::Action_maybe_recompile);
3345 }
3347 if (!stopped()) {
3349 // How many elements will we copy from the original?
3350 // The answer is MinI(orig_length - start, length).
3351 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3352 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3354 const bool raw_mem_only = true;
3355 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3357 // Generate a direct call to the right arraycopy function(s).
3358 // We know the copy is disjoint but we might not know if the
3359 // oop stores need checking.
3360 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3361 // This will fail a store-check if x contains any non-nulls.
3362 bool disjoint_bases = true;
3363 bool length_never_negative = true;
3364 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3365 original, start, newcopy, intcon(0), moved,
3366 disjoint_bases, length_never_negative);
3367 }
3368 } //original reexecute and sp are set back here
3370 if(!stopped()) {
3371 push(newcopy);
3372 }
3374 C->set_has_split_ifs(true); // Has chance for split-if optimization
3376 return true;
3377 }
3380 //----------------------generate_virtual_guard---------------------------
3381 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3382 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3383 RegionNode* slow_region) {
3384 ciMethod* method = callee();
3385 int vtable_index = method->vtable_index();
3386 // Get the methodOop out of the appropriate vtable entry.
3387 int entry_offset = (instanceKlass::vtable_start_offset() +
3388 vtable_index*vtableEntry::size()) * wordSize +
3389 vtableEntry::method_offset_in_bytes();
3390 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3391 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3393 // Compare the target method with the expected method (e.g., Object.hashCode).
3394 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3396 Node* native_call = makecon(native_call_addr);
3397 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3398 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3400 return generate_slow_guard(test_native, slow_region);
3401 }
3403 //-----------------------generate_method_call----------------------------
3404 // Use generate_method_call to make a slow-call to the real
3405 // method if the fast path fails. An alternative would be to
3406 // use a stub like OptoRuntime::slow_arraycopy_Java.
3407 // This only works for expanding the current library call,
3408 // not another intrinsic. (E.g., don't use this for making an
3409 // arraycopy call inside of the copyOf intrinsic.)
3410 CallJavaNode*
3411 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3412 // When compiling the intrinsic method itself, do not use this technique.
3413 guarantee(callee() != C->method(), "cannot make slow-call to self");
3415 ciMethod* method = callee();
3416 // ensure the JVMS we have will be correct for this call
3417 guarantee(method_id == method->intrinsic_id(), "must match");
3419 const TypeFunc* tf = TypeFunc::make(method);
3420 int tfdc = tf->domain()->cnt();
3421 CallJavaNode* slow_call;
3422 if (is_static) {
3423 assert(!is_virtual, "");
3424 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3425 SharedRuntime::get_resolve_static_call_stub(),
3426 method, bci());
3427 } else if (is_virtual) {
3428 null_check_receiver(method);
3429 int vtable_index = methodOopDesc::invalid_vtable_index;
3430 if (UseInlineCaches) {
3431 // Suppress the vtable call
3432 } else {
3433 // hashCode and clone are not a miranda methods,
3434 // so the vtable index is fixed.
3435 // No need to use the linkResolver to get it.
3436 vtable_index = method->vtable_index();
3437 }
3438 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3439 SharedRuntime::get_resolve_virtual_call_stub(),
3440 method, vtable_index, bci());
3441 } else { // neither virtual nor static: opt_virtual
3442 null_check_receiver(method);
3443 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3444 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3445 method, bci());
3446 slow_call->set_optimized_virtual(true);
3447 }
3448 set_arguments_for_java_call(slow_call);
3449 set_edges_for_java_call(slow_call);
3450 return slow_call;
3451 }
3454 //------------------------------inline_native_hashcode--------------------
3455 // Build special case code for calls to hashCode on an object.
3456 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3457 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3458 assert(!(is_virtual && is_static), "either virtual, special, or static");
3460 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3462 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3463 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3464 TypeInt::INT);
3465 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3466 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3467 TypePtr::BOTTOM);
3468 Node* obj = NULL;
3469 if (!is_static) {
3470 // Check for hashing null object
3471 obj = null_check_receiver(callee());
3472 if (stopped()) return true; // unconditionally null
3473 result_reg->init_req(_null_path, top());
3474 result_val->init_req(_null_path, top());
3475 } else {
3476 // Do a null check, and return zero if null.
3477 // System.identityHashCode(null) == 0
3478 obj = argument(0);
3479 Node* null_ctl = top();
3480 obj = null_check_oop(obj, &null_ctl);
3481 result_reg->init_req(_null_path, null_ctl);
3482 result_val->init_req(_null_path, _gvn.intcon(0));
3483 }
3485 // Unconditionally null? Then return right away.
3486 if (stopped()) {
3487 set_control( result_reg->in(_null_path) );
3488 if (!stopped())
3489 push( result_val ->in(_null_path) );
3490 return true;
3491 }
3493 // After null check, get the object's klass.
3494 Node* obj_klass = load_object_klass(obj);
3496 // This call may be virtual (invokevirtual) or bound (invokespecial).
3497 // For each case we generate slightly different code.
3499 // We only go to the fast case code if we pass a number of guards. The
3500 // paths which do not pass are accumulated in the slow_region.
3501 RegionNode* slow_region = new (C, 1) RegionNode(1);
3502 record_for_igvn(slow_region);
3504 // If this is a virtual call, we generate a funny guard. We pull out
3505 // the vtable entry corresponding to hashCode() from the target object.
3506 // If the target method which we are calling happens to be the native
3507 // Object hashCode() method, we pass the guard. We do not need this
3508 // guard for non-virtual calls -- the caller is known to be the native
3509 // Object hashCode().
3510 if (is_virtual) {
3511 generate_virtual_guard(obj_klass, slow_region);
3512 }
3514 // Get the header out of the object, use LoadMarkNode when available
3515 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3516 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
3517 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
3519 // Test the header to see if it is unlocked.
3520 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3521 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3522 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3523 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3524 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3526 generate_slow_guard(test_unlocked, slow_region);
3528 // Get the hash value and check to see that it has been properly assigned.
3529 // We depend on hash_mask being at most 32 bits and avoid the use of
3530 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3531 // vm: see markOop.hpp.
3532 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3533 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3534 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3535 // This hack lets the hash bits live anywhere in the mark object now, as long
3536 // as the shift drops the relevant bits into the low 32 bits. Note that
3537 // Java spec says that HashCode is an int so there's no point in capturing
3538 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3539 hshifted_header = ConvX2I(hshifted_header);
3540 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3542 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3543 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3544 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3546 generate_slow_guard(test_assigned, slow_region);
3548 Node* init_mem = reset_memory();
3549 // fill in the rest of the null path:
3550 result_io ->init_req(_null_path, i_o());
3551 result_mem->init_req(_null_path, init_mem);
3553 result_val->init_req(_fast_path, hash_val);
3554 result_reg->init_req(_fast_path, control());
3555 result_io ->init_req(_fast_path, i_o());
3556 result_mem->init_req(_fast_path, init_mem);
3558 // Generate code for the slow case. We make a call to hashCode().
3559 set_control(_gvn.transform(slow_region));
3560 if (!stopped()) {
3561 // No need for PreserveJVMState, because we're using up the present state.
3562 set_all_memory(init_mem);
3563 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3564 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3565 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3566 Node* slow_result = set_results_for_java_call(slow_call);
3567 // this->control() comes from set_results_for_java_call
3568 result_reg->init_req(_slow_path, control());
3569 result_val->init_req(_slow_path, slow_result);
3570 result_io ->set_req(_slow_path, i_o());
3571 result_mem ->set_req(_slow_path, reset_memory());
3572 }
3574 // Return the combined state.
3575 set_i_o( _gvn.transform(result_io) );
3576 set_all_memory( _gvn.transform(result_mem) );
3577 push_result(result_reg, result_val);
3579 return true;
3580 }
3582 //---------------------------inline_native_getClass----------------------------
3583 // Build special case code for calls to getClass on an object.
3584 bool LibraryCallKit::inline_native_getClass() {
3585 Node* obj = null_check_receiver(callee());
3586 if (stopped()) return true;
3587 push( load_mirror_from_klass(load_object_klass(obj)) );
3588 return true;
3589 }
3591 //-----------------inline_native_Reflection_getCallerClass---------------------
3592 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3593 //
3594 // NOTE that this code must perform the same logic as
3595 // vframeStream::security_get_caller_frame in that it must skip
3596 // Method.invoke() and auxiliary frames.
3601 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3602 ciMethod* method = callee();
3604 #ifndef PRODUCT
3605 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3606 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3607 }
3608 #endif
3610 debug_only(int saved_sp = _sp);
3612 // Argument words: (int depth)
3613 int nargs = 1;
3615 _sp += nargs;
3616 Node* caller_depth_node = pop();
3618 assert(saved_sp == _sp, "must have correct argument count");
3620 // The depth value must be a constant in order for the runtime call
3621 // to be eliminated.
3622 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3623 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3624 #ifndef PRODUCT
3625 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3626 tty->print_cr(" Bailing out because caller depth was not a constant");
3627 }
3628 #endif
3629 return false;
3630 }
3631 // Note that the JVM state at this point does not include the
3632 // getCallerClass() frame which we are trying to inline. The
3633 // semantics of getCallerClass(), however, are that the "first"
3634 // frame is the getCallerClass() frame, so we subtract one from the
3635 // requested depth before continuing. We don't inline requests of
3636 // getCallerClass(0).
3637 int caller_depth = caller_depth_type->get_con() - 1;
3638 if (caller_depth < 0) {
3639 #ifndef PRODUCT
3640 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3641 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3642 }
3643 #endif
3644 return false;
3645 }
3647 if (!jvms()->has_method()) {
3648 #ifndef PRODUCT
3649 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3650 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3651 }
3652 #endif
3653 return false;
3654 }
3655 int _depth = jvms()->depth(); // cache call chain depth
3657 // Walk back up the JVM state to find the caller at the required
3658 // depth. NOTE that this code must perform the same logic as
3659 // vframeStream::security_get_caller_frame in that it must skip
3660 // Method.invoke() and auxiliary frames. Note also that depth is
3661 // 1-based (1 is the bottom of the inlining).
3662 int inlining_depth = _depth;
3663 JVMState* caller_jvms = NULL;
3665 if (inlining_depth > 0) {
3666 caller_jvms = jvms();
3667 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3668 do {
3669 // The following if-tests should be performed in this order
3670 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3671 // Skip a Method.invoke() or auxiliary frame
3672 } else if (caller_depth > 0) {
3673 // Skip real frame
3674 --caller_depth;
3675 } else {
3676 // We're done: reached desired caller after skipping.
3677 break;
3678 }
3679 caller_jvms = caller_jvms->caller();
3680 --inlining_depth;
3681 } while (inlining_depth > 0);
3682 }
3684 if (inlining_depth == 0) {
3685 #ifndef PRODUCT
3686 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3687 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3688 tty->print_cr(" JVM state at this point:");
3689 for (int i = _depth; i >= 1; i--) {
3690 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3691 }
3692 }
3693 #endif
3694 return false; // Reached end of inlining
3695 }
3697 // Acquire method holder as java.lang.Class
3698 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3699 ciInstance* caller_mirror = caller_klass->java_mirror();
3700 // Push this as a constant
3701 push(makecon(TypeInstPtr::make(caller_mirror)));
3702 #ifndef PRODUCT
3703 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3704 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);
3705 tty->print_cr(" JVM state at this point:");
3706 for (int i = _depth; i >= 1; i--) {
3707 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3708 }
3709 }
3710 #endif
3711 return true;
3712 }
3714 // Helper routine for above
3715 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3716 ciMethod* method = jvms->method();
3718 // Is this the Method.invoke method itself?
3719 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3720 return true;
3722 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3723 ciKlass* k = method->holder();
3724 if (k->is_instance_klass()) {
3725 ciInstanceKlass* ik = k->as_instance_klass();
3726 for (; ik != NULL; ik = ik->super()) {
3727 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3728 ik == env()->find_system_klass(ik->name())) {
3729 return true;
3730 }
3731 }
3732 }
3733 else if (method->is_method_handle_adapter()) {
3734 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3735 return true;
3736 }
3738 return false;
3739 }
3741 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3742 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3743 // computing it since there is no lookup field by name function in the
3744 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3745 // Using a static variable here is safe even if we have multiple compilation
3746 // threads because the offset is constant. At worst the same offset will be
3747 // computed and stored multiple
3749 bool LibraryCallKit::inline_native_AtomicLong_get() {
3750 // Restore the stack and pop off the argument
3751 _sp+=1;
3752 Node *obj = pop();
3754 // get the offset of the "value" field. Since the CI interfaces
3755 // does not provide a way to look up a field by name, we scan the bytecodes
3756 // to get the field index. We expect the first 2 instructions of the method
3757 // to be:
3758 // 0 aload_0
3759 // 1 getfield "value"
3760 ciMethod* method = callee();
3761 if (value_field_offset == -1)
3762 {
3763 ciField* value_field;
3764 ciBytecodeStream iter(method);
3765 Bytecodes::Code bc = iter.next();
3767 if ((bc != Bytecodes::_aload_0) &&
3768 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3769 return false;
3770 bc = iter.next();
3771 if (bc != Bytecodes::_getfield)
3772 return false;
3773 bool ignore;
3774 value_field = iter.get_field(ignore);
3775 value_field_offset = value_field->offset_in_bytes();
3776 }
3778 // Null check without removing any arguments.
3779 _sp++;
3780 obj = do_null_check(obj, T_OBJECT);
3781 _sp--;
3782 // Check for locking null object
3783 if (stopped()) return true;
3785 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3786 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3787 int alias_idx = C->get_alias_index(adr_type);
3789 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3791 push_pair(result);
3793 return true;
3794 }
3796 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3797 // Restore the stack and pop off the arguments
3798 _sp+=5;
3799 Node *newVal = pop_pair();
3800 Node *oldVal = pop_pair();
3801 Node *obj = pop();
3803 // we need the offset of the "value" field which was computed when
3804 // inlining the get() method. Give up if we don't have it.
3805 if (value_field_offset == -1)
3806 return false;
3808 // Null check without removing any arguments.
3809 _sp+=5;
3810 obj = do_null_check(obj, T_OBJECT);
3811 _sp-=5;
3812 // Check for locking null object
3813 if (stopped()) return true;
3815 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3816 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3817 int alias_idx = C->get_alias_index(adr_type);
3819 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3820 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3821 set_memory(store_proj, alias_idx);
3822 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3824 Node *result;
3825 // CMove node is not used to be able fold a possible check code
3826 // after attemptUpdate() call. This code could be transformed
3827 // into CMove node by loop optimizations.
3828 {
3829 RegionNode *r = new (C, 3) RegionNode(3);
3830 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3832 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3833 Node *iftrue = opt_iff(r, iff);
3834 r->init_req(1, iftrue);
3835 result->init_req(1, intcon(1));
3836 result->init_req(2, intcon(0));
3838 set_control(_gvn.transform(r));
3839 record_for_igvn(r);
3841 C->set_has_split_ifs(true); // Has chance for split-if optimization
3842 }
3844 push(_gvn.transform(result));
3845 return true;
3846 }
3848 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3849 // restore the arguments
3850 _sp += arg_size();
3852 switch (id) {
3853 case vmIntrinsics::_floatToRawIntBits:
3854 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3855 break;
3857 case vmIntrinsics::_intBitsToFloat:
3858 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3859 break;
3861 case vmIntrinsics::_doubleToRawLongBits:
3862 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3863 break;
3865 case vmIntrinsics::_longBitsToDouble:
3866 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3867 break;
3869 case vmIntrinsics::_doubleToLongBits: {
3870 Node* value = pop_pair();
3872 // two paths (plus control) merge in a wood
3873 RegionNode *r = new (C, 3) RegionNode(3);
3874 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3876 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3877 // Build the boolean node
3878 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3880 // Branch either way.
3881 // NaN case is less traveled, which makes all the difference.
3882 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3883 Node *opt_isnan = _gvn.transform(ifisnan);
3884 assert( opt_isnan->is_If(), "Expect an IfNode");
3885 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3886 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3888 set_control(iftrue);
3890 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3891 Node *slow_result = longcon(nan_bits); // return NaN
3892 phi->init_req(1, _gvn.transform( slow_result ));
3893 r->init_req(1, iftrue);
3895 // Else fall through
3896 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3897 set_control(iffalse);
3899 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3900 r->init_req(2, iffalse);
3902 // Post merge
3903 set_control(_gvn.transform(r));
3904 record_for_igvn(r);
3906 Node* result = _gvn.transform(phi);
3907 assert(result->bottom_type()->isa_long(), "must be");
3908 push_pair(result);
3910 C->set_has_split_ifs(true); // Has chance for split-if optimization
3912 break;
3913 }
3915 case vmIntrinsics::_floatToIntBits: {
3916 Node* value = pop();
3918 // two paths (plus control) merge in a wood
3919 RegionNode *r = new (C, 3) RegionNode(3);
3920 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3922 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3923 // Build the boolean node
3924 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3926 // Branch either way.
3927 // NaN case is less traveled, which makes all the difference.
3928 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3929 Node *opt_isnan = _gvn.transform(ifisnan);
3930 assert( opt_isnan->is_If(), "Expect an IfNode");
3931 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3932 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3934 set_control(iftrue);
3936 static const jint nan_bits = 0x7fc00000;
3937 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3938 phi->init_req(1, _gvn.transform( slow_result ));
3939 r->init_req(1, iftrue);
3941 // Else fall through
3942 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3943 set_control(iffalse);
3945 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3946 r->init_req(2, iffalse);
3948 // Post merge
3949 set_control(_gvn.transform(r));
3950 record_for_igvn(r);
3952 Node* result = _gvn.transform(phi);
3953 assert(result->bottom_type()->isa_int(), "must be");
3954 push(result);
3956 C->set_has_split_ifs(true); // Has chance for split-if optimization
3958 break;
3959 }
3961 default:
3962 ShouldNotReachHere();
3963 }
3965 return true;
3966 }
3968 #ifdef _LP64
3969 #define XTOP ,top() /*additional argument*/
3970 #else //_LP64
3971 #define XTOP /*no additional argument*/
3972 #endif //_LP64
3974 //----------------------inline_unsafe_copyMemory-------------------------
3975 bool LibraryCallKit::inline_unsafe_copyMemory() {
3976 if (callee()->is_static()) return false; // caller must have the capability!
3977 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3978 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3979 null_check_receiver(callee()); // check then ignore argument(0)
3980 if (stopped()) return true;
3982 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3984 Node* src_ptr = argument(1);
3985 Node* src_off = ConvL2X(argument(2));
3986 assert(argument(3)->is_top(), "2nd half of long");
3987 Node* dst_ptr = argument(4);
3988 Node* dst_off = ConvL2X(argument(5));
3989 assert(argument(6)->is_top(), "2nd half of long");
3990 Node* size = ConvL2X(argument(7));
3991 assert(argument(8)->is_top(), "2nd half of long");
3993 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3994 "fieldOffset must be byte-scaled");
3996 Node* src = make_unsafe_address(src_ptr, src_off);
3997 Node* dst = make_unsafe_address(dst_ptr, dst_off);
3999 // Conservatively insert a memory barrier on all memory slices.
4000 // Do not let writes of the copy source or destination float below the copy.
4001 insert_mem_bar(Op_MemBarCPUOrder);
4003 // Call it. Note that the length argument is not scaled.
4004 make_runtime_call(RC_LEAF|RC_NO_FP,
4005 OptoRuntime::fast_arraycopy_Type(),
4006 StubRoutines::unsafe_arraycopy(),
4007 "unsafe_arraycopy",
4008 TypeRawPtr::BOTTOM,
4009 src, dst, size XTOP);
4011 // Do not let reads of the copy destination float above the copy.
4012 insert_mem_bar(Op_MemBarCPUOrder);
4014 return true;
4015 }
4017 //------------------------clone_coping-----------------------------------
4018 // Helper function for inline_native_clone.
4019 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4020 assert(obj_size != NULL, "");
4021 Node* raw_obj = alloc_obj->in(1);
4022 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4024 if (ReduceBulkZeroing) {
4025 // We will be completely responsible for initializing this object -
4026 // mark Initialize node as complete.
4027 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4028 // The object was just allocated - there should be no any stores!
4029 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4030 }
4032 // Copy the fastest available way.
4033 // TODO: generate fields copies for small objects instead.
4034 Node* src = obj;
4035 Node* dest = alloc_obj;
4036 Node* size = _gvn.transform(obj_size);
4038 // Exclude the header but include array length to copy by 8 bytes words.
4039 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4040 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4041 instanceOopDesc::base_offset_in_bytes();
4042 // base_off:
4043 // 8 - 32-bit VM
4044 // 12 - 64-bit VM, compressed oops
4045 // 16 - 64-bit VM, normal oops
4046 if (base_off % BytesPerLong != 0) {
4047 assert(UseCompressedOops, "");
4048 if (is_array) {
4049 // Exclude length to copy by 8 bytes words.
4050 base_off += sizeof(int);
4051 } else {
4052 // Include klass to copy by 8 bytes words.
4053 base_off = instanceOopDesc::klass_offset_in_bytes();
4054 }
4055 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4056 }
4057 src = basic_plus_adr(src, base_off);
4058 dest = basic_plus_adr(dest, base_off);
4060 // Compute the length also, if needed:
4061 Node* countx = size;
4062 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4063 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4065 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4066 bool disjoint_bases = true;
4067 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4068 src, NULL, dest, NULL, countx);
4070 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4071 if (card_mark) {
4072 assert(!is_array, "");
4073 // Put in store barrier for any and all oops we are sticking
4074 // into this object. (We could avoid this if we could prove
4075 // that the object type contains no oop fields at all.)
4076 Node* no_particular_value = NULL;
4077 Node* no_particular_field = NULL;
4078 int raw_adr_idx = Compile::AliasIdxRaw;
4079 post_barrier(control(),
4080 memory(raw_adr_type),
4081 alloc_obj,
4082 no_particular_field,
4083 raw_adr_idx,
4084 no_particular_value,
4085 T_OBJECT,
4086 false);
4087 }
4089 // Do not let reads from the cloned object float above the arraycopy.
4090 insert_mem_bar(Op_MemBarCPUOrder);
4091 }
4093 //------------------------inline_native_clone----------------------------
4094 // Here are the simple edge cases:
4095 // null receiver => normal trap
4096 // virtual and clone was overridden => slow path to out-of-line clone
4097 // not cloneable or finalizer => slow path to out-of-line Object.clone
4098 //
4099 // The general case has two steps, allocation and copying.
4100 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4101 //
4102 // Copying also has two cases, oop arrays and everything else.
4103 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4104 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4105 //
4106 // These steps fold up nicely if and when the cloned object's klass
4107 // can be sharply typed as an object array, a type array, or an instance.
4108 //
4109 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4110 int nargs = 1;
4111 PhiNode* result_val;
4113 //set the original stack and the reexecute bit for the interpreter to reexecute
4114 //the bytecode that invokes Object.clone if deoptimization happens
4115 { PreserveReexecuteState preexecs(this);
4116 jvms()->set_should_reexecute(true);
4118 //null_check_receiver will adjust _sp (push and pop)
4119 Node* obj = null_check_receiver(callee());
4120 if (stopped()) return true;
4122 _sp += nargs;
4124 Node* obj_klass = load_object_klass(obj);
4125 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4126 const TypeOopPtr* toop = ((tklass != NULL)
4127 ? tklass->as_instance_type()
4128 : TypeInstPtr::NOTNULL);
4130 // Conservatively insert a memory barrier on all memory slices.
4131 // Do not let writes into the original float below the clone.
4132 insert_mem_bar(Op_MemBarCPUOrder);
4134 // paths into result_reg:
4135 enum {
4136 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4137 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4138 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4139 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4140 PATH_LIMIT
4141 };
4142 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4143 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4144 TypeInstPtr::NOTNULL);
4145 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4146 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4147 TypePtr::BOTTOM);
4148 record_for_igvn(result_reg);
4150 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4151 int raw_adr_idx = Compile::AliasIdxRaw;
4152 const bool raw_mem_only = true;
4155 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4156 if (array_ctl != NULL) {
4157 // It's an array.
4158 PreserveJVMState pjvms(this);
4159 set_control(array_ctl);
4160 Node* obj_length = load_array_length(obj);
4161 Node* obj_size = NULL;
4162 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4163 raw_mem_only, &obj_size);
4165 if (!use_ReduceInitialCardMarks()) {
4166 // If it is an oop array, it requires very special treatment,
4167 // because card marking is required on each card of the array.
4168 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4169 if (is_obja != NULL) {
4170 PreserveJVMState pjvms2(this);
4171 set_control(is_obja);
4172 // Generate a direct call to the right arraycopy function(s).
4173 bool disjoint_bases = true;
4174 bool length_never_negative = true;
4175 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4176 obj, intcon(0), alloc_obj, intcon(0),
4177 obj_length,
4178 disjoint_bases, length_never_negative);
4179 result_reg->init_req(_objArray_path, control());
4180 result_val->init_req(_objArray_path, alloc_obj);
4181 result_i_o ->set_req(_objArray_path, i_o());
4182 result_mem ->set_req(_objArray_path, reset_memory());
4183 }
4184 }
4185 // Otherwise, there are no card marks to worry about.
4186 // (We can dispense with card marks if we know the allocation
4187 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4188 // causes the non-eden paths to take compensating steps to
4189 // simulate a fresh allocation, so that no further
4190 // card marks are required in compiled code to initialize
4191 // the object.)
4193 if (!stopped()) {
4194 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4196 // Present the results of the copy.
4197 result_reg->init_req(_array_path, control());
4198 result_val->init_req(_array_path, alloc_obj);
4199 result_i_o ->set_req(_array_path, i_o());
4200 result_mem ->set_req(_array_path, reset_memory());
4201 }
4202 }
4204 // We only go to the instance fast case code if we pass a number of guards.
4205 // The paths which do not pass are accumulated in the slow_region.
4206 RegionNode* slow_region = new (C, 1) RegionNode(1);
4207 record_for_igvn(slow_region);
4208 if (!stopped()) {
4209 // It's an instance (we did array above). Make the slow-path tests.
4210 // If this is a virtual call, we generate a funny guard. We grab
4211 // the vtable entry corresponding to clone() from the target object.
4212 // If the target method which we are calling happens to be the
4213 // Object clone() method, we pass the guard. We do not need this
4214 // guard for non-virtual calls; the caller is known to be the native
4215 // Object clone().
4216 if (is_virtual) {
4217 generate_virtual_guard(obj_klass, slow_region);
4218 }
4220 // The object must be cloneable and must not have a finalizer.
4221 // Both of these conditions may be checked in a single test.
4222 // We could optimize the cloneable test further, but we don't care.
4223 generate_access_flags_guard(obj_klass,
4224 // Test both conditions:
4225 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4226 // Must be cloneable but not finalizer:
4227 JVM_ACC_IS_CLONEABLE,
4228 slow_region);
4229 }
4231 if (!stopped()) {
4232 // It's an instance, and it passed the slow-path tests.
4233 PreserveJVMState pjvms(this);
4234 Node* obj_size = NULL;
4235 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4237 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4239 // Present the results of the slow call.
4240 result_reg->init_req(_instance_path, control());
4241 result_val->init_req(_instance_path, alloc_obj);
4242 result_i_o ->set_req(_instance_path, i_o());
4243 result_mem ->set_req(_instance_path, reset_memory());
4244 }
4246 // Generate code for the slow case. We make a call to clone().
4247 set_control(_gvn.transform(slow_region));
4248 if (!stopped()) {
4249 PreserveJVMState pjvms(this);
4250 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4251 Node* slow_result = set_results_for_java_call(slow_call);
4252 // this->control() comes from set_results_for_java_call
4253 result_reg->init_req(_slow_path, control());
4254 result_val->init_req(_slow_path, slow_result);
4255 result_i_o ->set_req(_slow_path, i_o());
4256 result_mem ->set_req(_slow_path, reset_memory());
4257 }
4259 // Return the combined state.
4260 set_control( _gvn.transform(result_reg) );
4261 set_i_o( _gvn.transform(result_i_o) );
4262 set_all_memory( _gvn.transform(result_mem) );
4263 } //original reexecute and sp are set back here
4265 push(_gvn.transform(result_val));
4267 return true;
4268 }
4271 // constants for computing the copy function
4272 enum {
4273 COPYFUNC_UNALIGNED = 0,
4274 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4275 COPYFUNC_CONJOINT = 0,
4276 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4277 };
4279 // Note: The condition "disjoint" applies also for overlapping copies
4280 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4281 static address
4282 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
4283 int selector =
4284 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4285 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4287 #define RETURN_STUB(xxx_arraycopy) { \
4288 name = #xxx_arraycopy; \
4289 return StubRoutines::xxx_arraycopy(); }
4291 switch (t) {
4292 case T_BYTE:
4293 case T_BOOLEAN:
4294 switch (selector) {
4295 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4296 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4297 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4298 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4299 }
4300 case T_CHAR:
4301 case T_SHORT:
4302 switch (selector) {
4303 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4304 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4305 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4306 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4307 }
4308 case T_INT:
4309 case T_FLOAT:
4310 switch (selector) {
4311 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4312 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4313 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4314 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4315 }
4316 case T_DOUBLE:
4317 case T_LONG:
4318 switch (selector) {
4319 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4320 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4321 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4322 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4323 }
4324 case T_ARRAY:
4325 case T_OBJECT:
4326 switch (selector) {
4327 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
4328 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
4329 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
4330 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
4331 }
4332 default:
4333 ShouldNotReachHere();
4334 return NULL;
4335 }
4337 #undef RETURN_STUB
4338 }
4340 //------------------------------basictype2arraycopy----------------------------
4341 address LibraryCallKit::basictype2arraycopy(BasicType t,
4342 Node* src_offset,
4343 Node* dest_offset,
4344 bool disjoint_bases,
4345 const char* &name) {
4346 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4347 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4349 bool aligned = false;
4350 bool disjoint = disjoint_bases;
4352 // if the offsets are the same, we can treat the memory regions as
4353 // disjoint, because either the memory regions are in different arrays,
4354 // or they are identical (which we can treat as disjoint.) We can also
4355 // treat a copy with a destination index less that the source index
4356 // as disjoint since a low->high copy will work correctly in this case.
4357 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4358 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4359 // both indices are constants
4360 int s_offs = src_offset_inttype->get_con();
4361 int d_offs = dest_offset_inttype->get_con();
4362 int element_size = type2aelembytes(t);
4363 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4364 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4365 if (s_offs >= d_offs) disjoint = true;
4366 } else if (src_offset == dest_offset && src_offset != NULL) {
4367 // This can occur if the offsets are identical non-constants.
4368 disjoint = true;
4369 }
4371 return select_arraycopy_function(t, aligned, disjoint, name);
4372 }
4375 //------------------------------inline_arraycopy-----------------------
4376 bool LibraryCallKit::inline_arraycopy() {
4377 // Restore the stack and pop off the arguments.
4378 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4379 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4381 Node *src = argument(0);
4382 Node *src_offset = argument(1);
4383 Node *dest = argument(2);
4384 Node *dest_offset = argument(3);
4385 Node *length = argument(4);
4387 // Compile time checks. If any of these checks cannot be verified at compile time,
4388 // we do not make a fast path for this call. Instead, we let the call remain as it
4389 // is. The checks we choose to mandate at compile time are:
4390 //
4391 // (1) src and dest are arrays.
4392 const Type* src_type = src->Value(&_gvn);
4393 const Type* dest_type = dest->Value(&_gvn);
4394 const TypeAryPtr* top_src = src_type->isa_aryptr();
4395 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4396 if (top_src == NULL || top_src->klass() == NULL ||
4397 top_dest == NULL || top_dest->klass() == NULL) {
4398 // Conservatively insert a memory barrier on all memory slices.
4399 // Do not let writes into the source float below the arraycopy.
4400 insert_mem_bar(Op_MemBarCPUOrder);
4402 // Call StubRoutines::generic_arraycopy stub.
4403 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4404 src, src_offset, dest, dest_offset, length);
4406 // Do not let reads from the destination float above the arraycopy.
4407 // Since we cannot type the arrays, we don't know which slices
4408 // might be affected. We could restrict this barrier only to those
4409 // memory slices which pertain to array elements--but don't bother.
4410 if (!InsertMemBarAfterArraycopy)
4411 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4412 insert_mem_bar(Op_MemBarCPUOrder);
4413 return true;
4414 }
4416 // (2) src and dest arrays must have elements of the same BasicType
4417 // Figure out the size and type of the elements we will be copying.
4418 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4419 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4420 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4421 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4423 if (src_elem != dest_elem || dest_elem == T_VOID) {
4424 // The component types are not the same or are not recognized. Punt.
4425 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4426 generate_slow_arraycopy(TypePtr::BOTTOM,
4427 src, src_offset, dest, dest_offset, length);
4428 return true;
4429 }
4431 //---------------------------------------------------------------------------
4432 // We will make a fast path for this call to arraycopy.
4434 // We have the following tests left to perform:
4435 //
4436 // (3) src and dest must not be null.
4437 // (4) src_offset must not be negative.
4438 // (5) dest_offset must not be negative.
4439 // (6) length must not be negative.
4440 // (7) src_offset + length must not exceed length of src.
4441 // (8) dest_offset + length must not exceed length of dest.
4442 // (9) each element of an oop array must be assignable
4444 RegionNode* slow_region = new (C, 1) RegionNode(1);
4445 record_for_igvn(slow_region);
4447 // (3) operands must not be null
4448 // We currently perform our null checks with the do_null_check routine.
4449 // This means that the null exceptions will be reported in the caller
4450 // rather than (correctly) reported inside of the native arraycopy call.
4451 // This should be corrected, given time. We do our null check with the
4452 // stack pointer restored.
4453 _sp += nargs;
4454 src = do_null_check(src, T_ARRAY);
4455 dest = do_null_check(dest, T_ARRAY);
4456 _sp -= nargs;
4458 // (4) src_offset must not be negative.
4459 generate_negative_guard(src_offset, slow_region);
4461 // (5) dest_offset must not be negative.
4462 generate_negative_guard(dest_offset, slow_region);
4464 // (6) length must not be negative (moved to generate_arraycopy()).
4465 // generate_negative_guard(length, slow_region);
4467 // (7) src_offset + length must not exceed length of src.
4468 generate_limit_guard(src_offset, length,
4469 load_array_length(src),
4470 slow_region);
4472 // (8) dest_offset + length must not exceed length of dest.
4473 generate_limit_guard(dest_offset, length,
4474 load_array_length(dest),
4475 slow_region);
4477 // (9) each element of an oop array must be assignable
4478 // The generate_arraycopy subroutine checks this.
4480 // This is where the memory effects are placed:
4481 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4482 generate_arraycopy(adr_type, dest_elem,
4483 src, src_offset, dest, dest_offset, length,
4484 false, false, slow_region);
4486 return true;
4487 }
4489 //-----------------------------generate_arraycopy----------------------
4490 // Generate an optimized call to arraycopy.
4491 // Caller must guard against non-arrays.
4492 // Caller must determine a common array basic-type for both arrays.
4493 // Caller must validate offsets against array bounds.
4494 // The slow_region has already collected guard failure paths
4495 // (such as out of bounds length or non-conformable array types).
4496 // The generated code has this shape, in general:
4497 //
4498 // if (length == 0) return // via zero_path
4499 // slowval = -1
4500 // if (types unknown) {
4501 // slowval = call generic copy loop
4502 // if (slowval == 0) return // via checked_path
4503 // } else if (indexes in bounds) {
4504 // if ((is object array) && !(array type check)) {
4505 // slowval = call checked copy loop
4506 // if (slowval == 0) return // via checked_path
4507 // } else {
4508 // call bulk copy loop
4509 // return // via fast_path
4510 // }
4511 // }
4512 // // adjust params for remaining work:
4513 // if (slowval != -1) {
4514 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4515 // }
4516 // slow_region:
4517 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4518 // return // via slow_call_path
4519 //
4520 // This routine is used from several intrinsics: System.arraycopy,
4521 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4522 //
4523 void
4524 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4525 BasicType basic_elem_type,
4526 Node* src, Node* src_offset,
4527 Node* dest, Node* dest_offset,
4528 Node* copy_length,
4529 bool disjoint_bases,
4530 bool length_never_negative,
4531 RegionNode* slow_region) {
4533 if (slow_region == NULL) {
4534 slow_region = new(C,1) RegionNode(1);
4535 record_for_igvn(slow_region);
4536 }
4538 Node* original_dest = dest;
4539 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4540 bool must_clear_dest = false;
4542 // See if this is the initialization of a newly-allocated array.
4543 // If so, we will take responsibility here for initializing it to zero.
4544 // (Note: Because tightly_coupled_allocation performs checks on the
4545 // out-edges of the dest, we need to avoid making derived pointers
4546 // from it until we have checked its uses.)
4547 if (ReduceBulkZeroing
4548 && !ZeroTLAB // pointless if already zeroed
4549 && basic_elem_type != T_CONFLICT // avoid corner case
4550 && !_gvn.eqv_uncast(src, dest)
4551 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4552 != NULL)
4553 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4554 && alloc->maybe_set_complete(&_gvn)) {
4555 // "You break it, you buy it."
4556 InitializeNode* init = alloc->initialization();
4557 assert(init->is_complete(), "we just did this");
4558 assert(dest->is_CheckCastPP(), "sanity");
4559 assert(dest->in(0)->in(0) == init, "dest pinned");
4560 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4561 // From this point on, every exit path is responsible for
4562 // initializing any non-copied parts of the object to zero.
4563 must_clear_dest = true;
4564 } else {
4565 // No zeroing elimination here.
4566 alloc = NULL;
4567 //original_dest = dest;
4568 //must_clear_dest = false;
4569 }
4571 // Results are placed here:
4572 enum { fast_path = 1, // normal void-returning assembly stub
4573 checked_path = 2, // special assembly stub with cleanup
4574 slow_call_path = 3, // something went wrong; call the VM
4575 zero_path = 4, // bypass when length of copy is zero
4576 bcopy_path = 5, // copy primitive array by 64-bit blocks
4577 PATH_LIMIT = 6
4578 };
4579 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4580 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4581 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4582 record_for_igvn(result_region);
4583 _gvn.set_type_bottom(result_i_o);
4584 _gvn.set_type_bottom(result_memory);
4585 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4587 // The slow_control path:
4588 Node* slow_control;
4589 Node* slow_i_o = i_o();
4590 Node* slow_mem = memory(adr_type);
4591 debug_only(slow_control = (Node*) badAddress);
4593 // Checked control path:
4594 Node* checked_control = top();
4595 Node* checked_mem = NULL;
4596 Node* checked_i_o = NULL;
4597 Node* checked_value = NULL;
4599 if (basic_elem_type == T_CONFLICT) {
4600 assert(!must_clear_dest, "");
4601 Node* cv = generate_generic_arraycopy(adr_type,
4602 src, src_offset, dest, dest_offset,
4603 copy_length);
4604 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4605 checked_control = control();
4606 checked_i_o = i_o();
4607 checked_mem = memory(adr_type);
4608 checked_value = cv;
4609 set_control(top()); // no fast path
4610 }
4612 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4613 if (not_pos != NULL) {
4614 PreserveJVMState pjvms(this);
4615 set_control(not_pos);
4617 // (6) length must not be negative.
4618 if (!length_never_negative) {
4619 generate_negative_guard(copy_length, slow_region);
4620 }
4622 // copy_length is 0.
4623 if (!stopped() && must_clear_dest) {
4624 Node* dest_length = alloc->in(AllocateNode::ALength);
4625 if (_gvn.eqv_uncast(copy_length, dest_length)
4626 || _gvn.find_int_con(dest_length, 1) <= 0) {
4627 // There is no zeroing to do. No need for a secondary raw memory barrier.
4628 } else {
4629 // Clear the whole thing since there are no source elements to copy.
4630 generate_clear_array(adr_type, dest, basic_elem_type,
4631 intcon(0), NULL,
4632 alloc->in(AllocateNode::AllocSize));
4633 // Use a secondary InitializeNode as raw memory barrier.
4634 // Currently it is needed only on this path since other
4635 // paths have stub or runtime calls as raw memory barriers.
4636 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4637 Compile::AliasIdxRaw,
4638 top())->as_Initialize();
4639 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4640 }
4641 }
4643 // Present the results of the fast call.
4644 result_region->init_req(zero_path, control());
4645 result_i_o ->init_req(zero_path, i_o());
4646 result_memory->init_req(zero_path, memory(adr_type));
4647 }
4649 if (!stopped() && must_clear_dest) {
4650 // We have to initialize the *uncopied* part of the array to zero.
4651 // The copy destination is the slice dest[off..off+len]. The other slices
4652 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4653 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4654 Node* dest_length = alloc->in(AllocateNode::ALength);
4655 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4656 copy_length) );
4658 // If there is a head section that needs zeroing, do it now.
4659 if (find_int_con(dest_offset, -1) != 0) {
4660 generate_clear_array(adr_type, dest, basic_elem_type,
4661 intcon(0), dest_offset,
4662 NULL);
4663 }
4665 // Next, perform a dynamic check on the tail length.
4666 // It is often zero, and we can win big if we prove this.
4667 // There are two wins: Avoid generating the ClearArray
4668 // with its attendant messy index arithmetic, and upgrade
4669 // the copy to a more hardware-friendly word size of 64 bits.
4670 Node* tail_ctl = NULL;
4671 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4672 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4673 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4674 tail_ctl = generate_slow_guard(bol_lt, NULL);
4675 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4676 }
4678 // At this point, let's assume there is no tail.
4679 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4680 // There is no tail. Try an upgrade to a 64-bit copy.
4681 bool didit = false;
4682 { PreserveJVMState pjvms(this);
4683 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4684 src, src_offset, dest, dest_offset,
4685 dest_size);
4686 if (didit) {
4687 // Present the results of the block-copying fast call.
4688 result_region->init_req(bcopy_path, control());
4689 result_i_o ->init_req(bcopy_path, i_o());
4690 result_memory->init_req(bcopy_path, memory(adr_type));
4691 }
4692 }
4693 if (didit)
4694 set_control(top()); // no regular fast path
4695 }
4697 // Clear the tail, if any.
4698 if (tail_ctl != NULL) {
4699 Node* notail_ctl = stopped() ? NULL : control();
4700 set_control(tail_ctl);
4701 if (notail_ctl == NULL) {
4702 generate_clear_array(adr_type, dest, basic_elem_type,
4703 dest_tail, NULL,
4704 dest_size);
4705 } else {
4706 // Make a local merge.
4707 Node* done_ctl = new(C,3) RegionNode(3);
4708 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4709 done_ctl->init_req(1, notail_ctl);
4710 done_mem->init_req(1, memory(adr_type));
4711 generate_clear_array(adr_type, dest, basic_elem_type,
4712 dest_tail, NULL,
4713 dest_size);
4714 done_ctl->init_req(2, control());
4715 done_mem->init_req(2, memory(adr_type));
4716 set_control( _gvn.transform(done_ctl) );
4717 set_memory( _gvn.transform(done_mem), adr_type );
4718 }
4719 }
4720 }
4722 BasicType copy_type = basic_elem_type;
4723 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4724 if (!stopped() && copy_type == T_OBJECT) {
4725 // If src and dest have compatible element types, we can copy bits.
4726 // Types S[] and D[] are compatible if D is a supertype of S.
4727 //
4728 // If they are not, we will use checked_oop_disjoint_arraycopy,
4729 // which performs a fast optimistic per-oop check, and backs off
4730 // further to JVM_ArrayCopy on the first per-oop check that fails.
4731 // (Actually, we don't move raw bits only; the GC requires card marks.)
4733 // Get the klassOop for both src and dest
4734 Node* src_klass = load_object_klass(src);
4735 Node* dest_klass = load_object_klass(dest);
4737 // Generate the subtype check.
4738 // This might fold up statically, or then again it might not.
4739 //
4740 // Non-static example: Copying List<String>.elements to a new String[].
4741 // The backing store for a List<String> is always an Object[],
4742 // but its elements are always type String, if the generic types
4743 // are correct at the source level.
4744 //
4745 // Test S[] against D[], not S against D, because (probably)
4746 // the secondary supertype cache is less busy for S[] than S.
4747 // This usually only matters when D is an interface.
4748 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4749 // Plug failing path into checked_oop_disjoint_arraycopy
4750 if (not_subtype_ctrl != top()) {
4751 PreserveJVMState pjvms(this);
4752 set_control(not_subtype_ctrl);
4753 // (At this point we can assume disjoint_bases, since types differ.)
4754 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4755 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4756 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4757 Node* dest_elem_klass = _gvn.transform(n1);
4758 Node* cv = generate_checkcast_arraycopy(adr_type,
4759 dest_elem_klass,
4760 src, src_offset, dest, dest_offset,
4761 copy_length);
4762 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4763 checked_control = control();
4764 checked_i_o = i_o();
4765 checked_mem = memory(adr_type);
4766 checked_value = cv;
4767 }
4768 // At this point we know we do not need type checks on oop stores.
4770 // Let's see if we need card marks:
4771 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4772 // If we do not need card marks, copy using the jint or jlong stub.
4773 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4774 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4775 "sizes agree");
4776 }
4777 }
4779 if (!stopped()) {
4780 // Generate the fast path, if possible.
4781 PreserveJVMState pjvms(this);
4782 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4783 src, src_offset, dest, dest_offset,
4784 ConvI2X(copy_length));
4786 // Present the results of the fast call.
4787 result_region->init_req(fast_path, control());
4788 result_i_o ->init_req(fast_path, i_o());
4789 result_memory->init_req(fast_path, memory(adr_type));
4790 }
4792 // Here are all the slow paths up to this point, in one bundle:
4793 slow_control = top();
4794 if (slow_region != NULL)
4795 slow_control = _gvn.transform(slow_region);
4796 debug_only(slow_region = (RegionNode*)badAddress);
4798 set_control(checked_control);
4799 if (!stopped()) {
4800 // Clean up after the checked call.
4801 // The returned value is either 0 or -1^K,
4802 // where K = number of partially transferred array elements.
4803 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4804 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4805 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4807 // If it is 0, we are done, so transfer to the end.
4808 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4809 result_region->init_req(checked_path, checks_done);
4810 result_i_o ->init_req(checked_path, checked_i_o);
4811 result_memory->init_req(checked_path, checked_mem);
4813 // If it is not zero, merge into the slow call.
4814 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4815 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4816 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4817 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4818 record_for_igvn(slow_reg2);
4819 slow_reg2 ->init_req(1, slow_control);
4820 slow_i_o2 ->init_req(1, slow_i_o);
4821 slow_mem2 ->init_req(1, slow_mem);
4822 slow_reg2 ->init_req(2, control());
4823 slow_i_o2 ->init_req(2, checked_i_o);
4824 slow_mem2 ->init_req(2, checked_mem);
4826 slow_control = _gvn.transform(slow_reg2);
4827 slow_i_o = _gvn.transform(slow_i_o2);
4828 slow_mem = _gvn.transform(slow_mem2);
4830 if (alloc != NULL) {
4831 // We'll restart from the very beginning, after zeroing the whole thing.
4832 // This can cause double writes, but that's OK since dest is brand new.
4833 // So we ignore the low 31 bits of the value returned from the stub.
4834 } else {
4835 // We must continue the copy exactly where it failed, or else
4836 // another thread might see the wrong number of writes to dest.
4837 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4838 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4839 slow_offset->init_req(1, intcon(0));
4840 slow_offset->init_req(2, checked_offset);
4841 slow_offset = _gvn.transform(slow_offset);
4843 // Adjust the arguments by the conditionally incoming offset.
4844 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4845 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4846 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4848 // Tweak the node variables to adjust the code produced below:
4849 src_offset = src_off_plus;
4850 dest_offset = dest_off_plus;
4851 copy_length = length_minus;
4852 }
4853 }
4855 set_control(slow_control);
4856 if (!stopped()) {
4857 // Generate the slow path, if needed.
4858 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4860 set_memory(slow_mem, adr_type);
4861 set_i_o(slow_i_o);
4863 if (must_clear_dest) {
4864 generate_clear_array(adr_type, dest, basic_elem_type,
4865 intcon(0), NULL,
4866 alloc->in(AllocateNode::AllocSize));
4867 }
4869 generate_slow_arraycopy(adr_type,
4870 src, src_offset, dest, dest_offset,
4871 copy_length);
4873 result_region->init_req(slow_call_path, control());
4874 result_i_o ->init_req(slow_call_path, i_o());
4875 result_memory->init_req(slow_call_path, memory(adr_type));
4876 }
4878 // Remove unused edges.
4879 for (uint i = 1; i < result_region->req(); i++) {
4880 if (result_region->in(i) == NULL)
4881 result_region->init_req(i, top());
4882 }
4884 // Finished; return the combined state.
4885 set_control( _gvn.transform(result_region) );
4886 set_i_o( _gvn.transform(result_i_o) );
4887 set_memory( _gvn.transform(result_memory), adr_type );
4889 // The memory edges above are precise in order to model effects around
4890 // array copies accurately to allow value numbering of field loads around
4891 // arraycopy. Such field loads, both before and after, are common in Java
4892 // collections and similar classes involving header/array data structures.
4893 //
4894 // But with low number of register or when some registers are used or killed
4895 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4896 // The next memory barrier is added to avoid it. If the arraycopy can be
4897 // optimized away (which it can, sometimes) then we can manually remove
4898 // the membar also.
4899 //
4900 // Do not let reads from the cloned object float above the arraycopy.
4901 if (InsertMemBarAfterArraycopy || alloc != NULL)
4902 insert_mem_bar(Op_MemBarCPUOrder);
4903 }
4906 // Helper function which determines if an arraycopy immediately follows
4907 // an allocation, with no intervening tests or other escapes for the object.
4908 AllocateArrayNode*
4909 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4910 RegionNode* slow_region) {
4911 if (stopped()) return NULL; // no fast path
4912 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4914 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4915 if (alloc == NULL) return NULL;
4917 Node* rawmem = memory(Compile::AliasIdxRaw);
4918 // Is the allocation's memory state untouched?
4919 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4920 // Bail out if there have been raw-memory effects since the allocation.
4921 // (Example: There might have been a call or safepoint.)
4922 return NULL;
4923 }
4924 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4925 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4926 return NULL;
4927 }
4929 // There must be no unexpected observers of this allocation.
4930 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4931 Node* obs = ptr->fast_out(i);
4932 if (obs != this->map()) {
4933 return NULL;
4934 }
4935 }
4937 // This arraycopy must unconditionally follow the allocation of the ptr.
4938 Node* alloc_ctl = ptr->in(0);
4939 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4941 Node* ctl = control();
4942 while (ctl != alloc_ctl) {
4943 // There may be guards which feed into the slow_region.
4944 // Any other control flow means that we might not get a chance
4945 // to finish initializing the allocated object.
4946 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4947 IfNode* iff = ctl->in(0)->as_If();
4948 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4949 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4950 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4951 ctl = iff->in(0); // This test feeds the known slow_region.
4952 continue;
4953 }
4954 // One more try: Various low-level checks bottom out in
4955 // uncommon traps. If the debug-info of the trap omits
4956 // any reference to the allocation, as we've already
4957 // observed, then there can be no objection to the trap.
4958 bool found_trap = false;
4959 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4960 Node* obs = not_ctl->fast_out(j);
4961 if (obs->in(0) == not_ctl && obs->is_Call() &&
4962 (obs->as_Call()->entry_point() ==
4963 SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
4964 found_trap = true; break;
4965 }
4966 }
4967 if (found_trap) {
4968 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4969 continue;
4970 }
4971 }
4972 return NULL;
4973 }
4975 // If we get this far, we have an allocation which immediately
4976 // precedes the arraycopy, and we can take over zeroing the new object.
4977 // The arraycopy will finish the initialization, and provide
4978 // a new control state to which we will anchor the destination pointer.
4980 return alloc;
4981 }
4983 // Helper for initialization of arrays, creating a ClearArray.
4984 // It writes zero bits in [start..end), within the body of an array object.
4985 // The memory effects are all chained onto the 'adr_type' alias category.
4986 //
4987 // Since the object is otherwise uninitialized, we are free
4988 // to put a little "slop" around the edges of the cleared area,
4989 // as long as it does not go back into the array's header,
4990 // or beyond the array end within the heap.
4991 //
4992 // The lower edge can be rounded down to the nearest jint and the
4993 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4994 //
4995 // Arguments:
4996 // adr_type memory slice where writes are generated
4997 // dest oop of the destination array
4998 // basic_elem_type element type of the destination
4999 // slice_idx array index of first element to store
5000 // slice_len number of elements to store (or NULL)
5001 // dest_size total size in bytes of the array object
5002 //
5003 // Exactly one of slice_len or dest_size must be non-NULL.
5004 // If dest_size is non-NULL, zeroing extends to the end of the object.
5005 // If slice_len is non-NULL, the slice_idx value must be a constant.
5006 void
5007 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5008 Node* dest,
5009 BasicType basic_elem_type,
5010 Node* slice_idx,
5011 Node* slice_len,
5012 Node* dest_size) {
5013 // one or the other but not both of slice_len and dest_size:
5014 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5015 if (slice_len == NULL) slice_len = top();
5016 if (dest_size == NULL) dest_size = top();
5018 // operate on this memory slice:
5019 Node* mem = memory(adr_type); // memory slice to operate on
5021 // scaling and rounding of indexes:
5022 int scale = exact_log2(type2aelembytes(basic_elem_type));
5023 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5024 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5025 int bump_bit = (-1 << scale) & BytesPerInt;
5027 // determine constant starts and ends
5028 const intptr_t BIG_NEG = -128;
5029 assert(BIG_NEG + 2*abase < 0, "neg enough");
5030 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5031 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5032 if (slice_len_con == 0) {
5033 return; // nothing to do here
5034 }
5035 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5036 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5037 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5038 assert(end_con < 0, "not two cons");
5039 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5040 BytesPerLong);
5041 }
5043 if (start_con >= 0 && end_con >= 0) {
5044 // Constant start and end. Simple.
5045 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5046 start_con, end_con, &_gvn);
5047 } else if (start_con >= 0 && dest_size != top()) {
5048 // Constant start, pre-rounded end after the tail of the array.
5049 Node* end = dest_size;
5050 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5051 start_con, end, &_gvn);
5052 } else if (start_con >= 0 && slice_len != top()) {
5053 // Constant start, non-constant end. End needs rounding up.
5054 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5055 intptr_t end_base = abase + (slice_idx_con << scale);
5056 int end_round = (-1 << scale) & (BytesPerLong - 1);
5057 Node* end = ConvI2X(slice_len);
5058 if (scale != 0)
5059 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5060 end_base += end_round;
5061 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5062 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5063 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5064 start_con, end, &_gvn);
5065 } else if (start_con < 0 && dest_size != top()) {
5066 // Non-constant start, pre-rounded end after the tail of the array.
5067 // This is almost certainly a "round-to-end" operation.
5068 Node* start = slice_idx;
5069 start = ConvI2X(start);
5070 if (scale != 0)
5071 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5072 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5073 if ((bump_bit | clear_low) != 0) {
5074 int to_clear = (bump_bit | clear_low);
5075 // Align up mod 8, then store a jint zero unconditionally
5076 // just before the mod-8 boundary.
5077 if (((abase + bump_bit) & ~to_clear) - bump_bit
5078 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5079 bump_bit = 0;
5080 assert((abase & to_clear) == 0, "array base must be long-aligned");
5081 } else {
5082 // Bump 'start' up to (or past) the next jint boundary:
5083 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5084 assert((abase & clear_low) == 0, "array base must be int-aligned");
5085 }
5086 // Round bumped 'start' down to jlong boundary in body of array.
5087 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5088 if (bump_bit != 0) {
5089 // Store a zero to the immediately preceding jint:
5090 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5091 Node* p1 = basic_plus_adr(dest, x1);
5092 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5093 mem = _gvn.transform(mem);
5094 }
5095 }
5096 Node* end = dest_size; // pre-rounded
5097 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5098 start, end, &_gvn);
5099 } else {
5100 // Non-constant start, unrounded non-constant end.
5101 // (Nobody zeroes a random midsection of an array using this routine.)
5102 ShouldNotReachHere(); // fix caller
5103 }
5105 // Done.
5106 set_memory(mem, adr_type);
5107 }
5110 bool
5111 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5112 BasicType basic_elem_type,
5113 AllocateNode* alloc,
5114 Node* src, Node* src_offset,
5115 Node* dest, Node* dest_offset,
5116 Node* dest_size) {
5117 // See if there is an advantage from block transfer.
5118 int scale = exact_log2(type2aelembytes(basic_elem_type));
5119 if (scale >= LogBytesPerLong)
5120 return false; // it is already a block transfer
5122 // Look at the alignment of the starting offsets.
5123 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5124 const intptr_t BIG_NEG = -128;
5125 assert(BIG_NEG + 2*abase < 0, "neg enough");
5127 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5128 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5129 if (src_off < 0 || dest_off < 0)
5130 // At present, we can only understand constants.
5131 return false;
5133 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5134 // Non-aligned; too bad.
5135 // One more chance: Pick off an initial 32-bit word.
5136 // This is a common case, since abase can be odd mod 8.
5137 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5138 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5139 Node* sptr = basic_plus_adr(src, src_off);
5140 Node* dptr = basic_plus_adr(dest, dest_off);
5141 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5142 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5143 src_off += BytesPerInt;
5144 dest_off += BytesPerInt;
5145 } else {
5146 return false;
5147 }
5148 }
5149 assert(src_off % BytesPerLong == 0, "");
5150 assert(dest_off % BytesPerLong == 0, "");
5152 // Do this copy by giant steps.
5153 Node* sptr = basic_plus_adr(src, src_off);
5154 Node* dptr = basic_plus_adr(dest, dest_off);
5155 Node* countx = dest_size;
5156 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5157 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5159 bool disjoint_bases = true; // since alloc != NULL
5160 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5161 sptr, NULL, dptr, NULL, countx);
5163 return true;
5164 }
5167 // Helper function; generates code for the slow case.
5168 // We make a call to a runtime method which emulates the native method,
5169 // but without the native wrapper overhead.
5170 void
5171 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5172 Node* src, Node* src_offset,
5173 Node* dest, Node* dest_offset,
5174 Node* copy_length) {
5175 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5176 OptoRuntime::slow_arraycopy_Type(),
5177 OptoRuntime::slow_arraycopy_Java(),
5178 "slow_arraycopy", adr_type,
5179 src, src_offset, dest, dest_offset,
5180 copy_length);
5182 // Handle exceptions thrown by this fellow:
5183 make_slow_call_ex(call, env()->Throwable_klass(), false);
5184 }
5186 // Helper function; generates code for cases requiring runtime checks.
5187 Node*
5188 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5189 Node* dest_elem_klass,
5190 Node* src, Node* src_offset,
5191 Node* dest, Node* dest_offset,
5192 Node* copy_length) {
5193 if (stopped()) return NULL;
5195 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
5196 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5197 return NULL;
5198 }
5200 // Pick out the parameters required to perform a store-check
5201 // for the target array. This is an optimistic check. It will
5202 // look in each non-null element's class, at the desired klass's
5203 // super_check_offset, for the desired klass.
5204 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5205 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5206 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
5207 Node* check_offset = _gvn.transform(n3);
5208 Node* check_value = dest_elem_klass;
5210 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5211 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5213 // (We know the arrays are never conjoint, because their types differ.)
5214 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5215 OptoRuntime::checkcast_arraycopy_Type(),
5216 copyfunc_addr, "checkcast_arraycopy", adr_type,
5217 // five arguments, of which two are
5218 // intptr_t (jlong in LP64)
5219 src_start, dest_start,
5220 copy_length XTOP,
5221 check_offset XTOP,
5222 check_value);
5224 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5225 }
5228 // Helper function; generates code for cases requiring runtime checks.
5229 Node*
5230 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5231 Node* src, Node* src_offset,
5232 Node* dest, Node* dest_offset,
5233 Node* copy_length) {
5234 if (stopped()) return NULL;
5236 address copyfunc_addr = StubRoutines::generic_arraycopy();
5237 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5238 return NULL;
5239 }
5241 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5242 OptoRuntime::generic_arraycopy_Type(),
5243 copyfunc_addr, "generic_arraycopy", adr_type,
5244 src, src_offset, dest, dest_offset, copy_length);
5246 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5247 }
5249 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5250 void
5251 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5252 BasicType basic_elem_type,
5253 bool disjoint_bases,
5254 Node* src, Node* src_offset,
5255 Node* dest, Node* dest_offset,
5256 Node* copy_length) {
5257 if (stopped()) return; // nothing to do
5259 Node* src_start = src;
5260 Node* dest_start = dest;
5261 if (src_offset != NULL || dest_offset != NULL) {
5262 assert(src_offset != NULL && dest_offset != NULL, "");
5263 src_start = array_element_address(src, src_offset, basic_elem_type);
5264 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5265 }
5267 // Figure out which arraycopy runtime method to call.
5268 const char* copyfunc_name = "arraycopy";
5269 address copyfunc_addr =
5270 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5271 disjoint_bases, copyfunc_name);
5273 // Call it. Note that the count_ix value is not scaled to a byte-size.
5274 make_runtime_call(RC_LEAF|RC_NO_FP,
5275 OptoRuntime::fast_arraycopy_Type(),
5276 copyfunc_addr, copyfunc_name, adr_type,
5277 src_start, dest_start, copy_length XTOP);
5278 }