Tue, 12 Feb 2013 12:56:11 +0100
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
Summary: Add support for expensive nodes.
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1999, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/parse.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/subnode.hpp"
39 #include "prims/nativeLookup.hpp"
40 #include "runtime/sharedRuntime.hpp"
42 class LibraryIntrinsic : public InlineCallGenerator {
43 // Extend the set of intrinsics known to the runtime:
44 public:
45 private:
46 bool _is_virtual;
47 bool _is_predicted;
48 vmIntrinsics::ID _intrinsic_id;
50 public:
51 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, vmIntrinsics::ID id)
52 : InlineCallGenerator(m),
53 _is_virtual(is_virtual),
54 _is_predicted(is_predicted),
55 _intrinsic_id(id)
56 {
57 }
58 virtual bool is_intrinsic() const { return true; }
59 virtual bool is_virtual() const { return _is_virtual; }
60 virtual bool is_predicted() const { return _is_predicted; }
61 virtual JVMState* generate(JVMState* jvms);
62 virtual Node* generate_predicate(JVMState* jvms);
63 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
64 };
67 // Local helper class for LibraryIntrinsic:
68 class LibraryCallKit : public GraphKit {
69 private:
70 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
71 Node* _result; // the result node, if any
72 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
74 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
76 public:
77 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
78 : GraphKit(jvms),
79 _intrinsic(intrinsic),
80 _result(NULL)
81 {
82 // Check if this is a root compile. In that case we don't have a caller.
83 if (!jvms->has_method()) {
84 _reexecute_sp = sp();
85 } else {
86 // Find out how many arguments the interpreter needs when deoptimizing
87 // and save the stack pointer value so it can used by uncommon_trap.
88 // We find the argument count by looking at the declared signature.
89 bool ignored_will_link;
90 ciSignature* declared_signature = NULL;
91 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
92 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
93 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
94 }
95 }
97 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
99 ciMethod* caller() const { return jvms()->method(); }
100 int bci() const { return jvms()->bci(); }
101 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
102 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
103 ciMethod* callee() const { return _intrinsic->method(); }
105 bool try_to_inline();
106 Node* try_to_predicate();
108 void push_result() {
109 // Push the result onto the stack.
110 if (!stopped() && result() != NULL) {
111 BasicType bt = result()->bottom_type()->basic_type();
112 push_node(bt, result());
113 }
114 }
116 private:
117 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
118 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
119 }
121 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
122 void set_result(RegionNode* region, PhiNode* value);
123 Node* result() { return _result; }
125 virtual int reexecute_sp() { return _reexecute_sp; }
127 // Helper functions to inline natives
128 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
129 Node* generate_slow_guard(Node* test, RegionNode* region);
130 Node* generate_fair_guard(Node* test, RegionNode* region);
131 Node* generate_negative_guard(Node* index, RegionNode* region,
132 // resulting CastII of index:
133 Node* *pos_index = NULL);
134 Node* generate_nonpositive_guard(Node* index, bool never_negative,
135 // resulting CastII of index:
136 Node* *pos_index = NULL);
137 Node* generate_limit_guard(Node* offset, Node* subseq_length,
138 Node* array_length,
139 RegionNode* region);
140 Node* generate_current_thread(Node* &tls_output);
141 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
142 bool disjoint_bases, const char* &name, bool dest_uninitialized);
143 Node* load_mirror_from_klass(Node* klass);
144 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
145 RegionNode* region, int null_path,
146 int offset);
147 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
148 RegionNode* region, int null_path) {
149 int offset = java_lang_Class::klass_offset_in_bytes();
150 return load_klass_from_mirror_common(mirror, never_see_null,
151 region, null_path,
152 offset);
153 }
154 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
155 RegionNode* region, int null_path) {
156 int offset = java_lang_Class::array_klass_offset_in_bytes();
157 return load_klass_from_mirror_common(mirror, never_see_null,
158 region, null_path,
159 offset);
160 }
161 Node* generate_access_flags_guard(Node* kls,
162 int modifier_mask, int modifier_bits,
163 RegionNode* region);
164 Node* generate_interface_guard(Node* kls, RegionNode* region);
165 Node* generate_array_guard(Node* kls, RegionNode* region) {
166 return generate_array_guard_common(kls, region, false, false);
167 }
168 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
169 return generate_array_guard_common(kls, region, false, true);
170 }
171 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
172 return generate_array_guard_common(kls, region, true, false);
173 }
174 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
175 return generate_array_guard_common(kls, region, true, true);
176 }
177 Node* generate_array_guard_common(Node* kls, RegionNode* region,
178 bool obj_array, bool not_array);
179 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
180 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
181 bool is_virtual = false, bool is_static = false);
182 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
183 return generate_method_call(method_id, false, true);
184 }
185 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
186 return generate_method_call(method_id, true, false);
187 }
188 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
190 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
191 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
192 bool inline_string_compareTo();
193 bool inline_string_indexOf();
194 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
195 bool inline_string_equals();
196 Node* round_double_node(Node* n);
197 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
198 bool inline_math_native(vmIntrinsics::ID id);
199 bool inline_trig(vmIntrinsics::ID id);
200 bool inline_math(vmIntrinsics::ID id);
201 bool inline_exp();
202 bool inline_pow();
203 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
204 bool inline_min_max(vmIntrinsics::ID id);
205 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
206 // This returns Type::AnyPtr, RawPtr, or OopPtr.
207 int classify_unsafe_addr(Node* &base, Node* &offset);
208 Node* make_unsafe_address(Node* base, Node* offset);
209 // Helper for inline_unsafe_access.
210 // Generates the guards that check whether the result of
211 // Unsafe.getObject should be recorded in an SATB log buffer.
212 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
213 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
214 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
215 bool inline_unsafe_allocate();
216 bool inline_unsafe_copyMemory();
217 bool inline_native_currentThread();
218 #ifdef TRACE_HAVE_INTRINSICS
219 bool inline_native_classID();
220 bool inline_native_threadID();
221 #endif
222 bool inline_native_time_funcs(address method, const char* funcName);
223 bool inline_native_isInterrupted();
224 bool inline_native_Class_query(vmIntrinsics::ID id);
225 bool inline_native_subtype_check();
227 bool inline_native_newArray();
228 bool inline_native_getLength();
229 bool inline_array_copyOf(bool is_copyOfRange);
230 bool inline_array_equals();
231 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
232 bool inline_native_clone(bool is_virtual);
233 bool inline_native_Reflection_getCallerClass();
234 bool is_method_invoke_or_aux_frame(JVMState* jvms);
235 // Helper function for inlining native object hash method
236 bool inline_native_hashcode(bool is_virtual, bool is_static);
237 bool inline_native_getClass();
239 // Helper functions for inlining arraycopy
240 bool inline_arraycopy();
241 void generate_arraycopy(const TypePtr* adr_type,
242 BasicType basic_elem_type,
243 Node* src, Node* src_offset,
244 Node* dest, Node* dest_offset,
245 Node* copy_length,
246 bool disjoint_bases = false,
247 bool length_never_negative = false,
248 RegionNode* slow_region = NULL);
249 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
250 RegionNode* slow_region);
251 void generate_clear_array(const TypePtr* adr_type,
252 Node* dest,
253 BasicType basic_elem_type,
254 Node* slice_off,
255 Node* slice_len,
256 Node* slice_end);
257 bool generate_block_arraycopy(const TypePtr* adr_type,
258 BasicType basic_elem_type,
259 AllocateNode* alloc,
260 Node* src, Node* src_offset,
261 Node* dest, Node* dest_offset,
262 Node* dest_size, bool dest_uninitialized);
263 void generate_slow_arraycopy(const TypePtr* adr_type,
264 Node* src, Node* src_offset,
265 Node* dest, Node* dest_offset,
266 Node* copy_length, bool dest_uninitialized);
267 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
268 Node* dest_elem_klass,
269 Node* src, Node* src_offset,
270 Node* dest, Node* dest_offset,
271 Node* copy_length, bool dest_uninitialized);
272 Node* generate_generic_arraycopy(const TypePtr* adr_type,
273 Node* src, Node* src_offset,
274 Node* dest, Node* dest_offset,
275 Node* copy_length, bool dest_uninitialized);
276 void generate_unchecked_arraycopy(const TypePtr* adr_type,
277 BasicType basic_elem_type,
278 bool disjoint_bases,
279 Node* src, Node* src_offset,
280 Node* dest, Node* dest_offset,
281 Node* copy_length, bool dest_uninitialized);
282 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
283 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
284 bool inline_unsafe_ordered_store(BasicType type);
285 bool inline_unsafe_fence(vmIntrinsics::ID id);
286 bool inline_fp_conversions(vmIntrinsics::ID id);
287 bool inline_number_methods(vmIntrinsics::ID id);
288 bool inline_reference_get();
289 bool inline_aescrypt_Block(vmIntrinsics::ID id);
290 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
291 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
292 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
293 bool inline_encodeISOArray();
294 };
297 //---------------------------make_vm_intrinsic----------------------------
298 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
299 vmIntrinsics::ID id = m->intrinsic_id();
300 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
302 if (DisableIntrinsic[0] != '\0'
303 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
304 // disabled by a user request on the command line:
305 // example: -XX:DisableIntrinsic=_hashCode,_getClass
306 return NULL;
307 }
309 if (!m->is_loaded()) {
310 // do not attempt to inline unloaded methods
311 return NULL;
312 }
314 // Only a few intrinsics implement a virtual dispatch.
315 // They are expensive calls which are also frequently overridden.
316 if (is_virtual) {
317 switch (id) {
318 case vmIntrinsics::_hashCode:
319 case vmIntrinsics::_clone:
320 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
321 break;
322 default:
323 return NULL;
324 }
325 }
327 // -XX:-InlineNatives disables nearly all intrinsics:
328 if (!InlineNatives) {
329 switch (id) {
330 case vmIntrinsics::_indexOf:
331 case vmIntrinsics::_compareTo:
332 case vmIntrinsics::_equals:
333 case vmIntrinsics::_equalsC:
334 case vmIntrinsics::_getAndAddInt:
335 case vmIntrinsics::_getAndAddLong:
336 case vmIntrinsics::_getAndSetInt:
337 case vmIntrinsics::_getAndSetLong:
338 case vmIntrinsics::_getAndSetObject:
339 case vmIntrinsics::_loadFence:
340 case vmIntrinsics::_storeFence:
341 case vmIntrinsics::_fullFence:
342 break; // InlineNatives does not control String.compareTo
343 case vmIntrinsics::_Reference_get:
344 break; // InlineNatives does not control Reference.get
345 default:
346 return NULL;
347 }
348 }
350 bool is_predicted = false;
352 switch (id) {
353 case vmIntrinsics::_compareTo:
354 if (!SpecialStringCompareTo) return NULL;
355 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
356 break;
357 case vmIntrinsics::_indexOf:
358 if (!SpecialStringIndexOf) return NULL;
359 break;
360 case vmIntrinsics::_equals:
361 if (!SpecialStringEquals) return NULL;
362 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
363 break;
364 case vmIntrinsics::_equalsC:
365 if (!SpecialArraysEquals) return NULL;
366 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
367 break;
368 case vmIntrinsics::_arraycopy:
369 if (!InlineArrayCopy) return NULL;
370 break;
371 case vmIntrinsics::_copyMemory:
372 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
373 if (!InlineArrayCopy) return NULL;
374 break;
375 case vmIntrinsics::_hashCode:
376 if (!InlineObjectHash) return NULL;
377 break;
378 case vmIntrinsics::_clone:
379 case vmIntrinsics::_copyOf:
380 case vmIntrinsics::_copyOfRange:
381 if (!InlineObjectCopy) return NULL;
382 // These also use the arraycopy intrinsic mechanism:
383 if (!InlineArrayCopy) return NULL;
384 break;
385 case vmIntrinsics::_encodeISOArray:
386 if (!SpecialEncodeISOArray) return NULL;
387 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
388 break;
389 case vmIntrinsics::_checkIndex:
390 // We do not intrinsify this. The optimizer does fine with it.
391 return NULL;
393 case vmIntrinsics::_getCallerClass:
394 if (!UseNewReflection) return NULL;
395 if (!InlineReflectionGetCallerClass) return NULL;
396 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
397 break;
399 case vmIntrinsics::_bitCount_i:
400 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
401 break;
403 case vmIntrinsics::_bitCount_l:
404 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
405 break;
407 case vmIntrinsics::_numberOfLeadingZeros_i:
408 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
409 break;
411 case vmIntrinsics::_numberOfLeadingZeros_l:
412 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
413 break;
415 case vmIntrinsics::_numberOfTrailingZeros_i:
416 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
417 break;
419 case vmIntrinsics::_numberOfTrailingZeros_l:
420 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
421 break;
423 case vmIntrinsics::_reverseBytes_c:
424 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
425 break;
426 case vmIntrinsics::_reverseBytes_s:
427 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
428 break;
429 case vmIntrinsics::_reverseBytes_i:
430 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
431 break;
432 case vmIntrinsics::_reverseBytes_l:
433 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
434 break;
436 case vmIntrinsics::_Reference_get:
437 // Use the intrinsic version of Reference.get() so that the value in
438 // the referent field can be registered by the G1 pre-barrier code.
439 // Also add memory barrier to prevent commoning reads from this field
440 // across safepoint since GC can change it value.
441 break;
443 case vmIntrinsics::_compareAndSwapObject:
444 #ifdef _LP64
445 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
446 #endif
447 break;
449 case vmIntrinsics::_compareAndSwapLong:
450 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
451 break;
453 case vmIntrinsics::_getAndAddInt:
454 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
455 break;
457 case vmIntrinsics::_getAndAddLong:
458 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
459 break;
461 case vmIntrinsics::_getAndSetInt:
462 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
463 break;
465 case vmIntrinsics::_getAndSetLong:
466 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
467 break;
469 case vmIntrinsics::_getAndSetObject:
470 #ifdef _LP64
471 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
472 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
473 break;
474 #else
475 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
476 break;
477 #endif
479 case vmIntrinsics::_aescrypt_encryptBlock:
480 case vmIntrinsics::_aescrypt_decryptBlock:
481 if (!UseAESIntrinsics) return NULL;
482 break;
484 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
485 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
486 if (!UseAESIntrinsics) return NULL;
487 // these two require the predicated logic
488 is_predicted = true;
489 break;
491 default:
492 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
493 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
494 break;
495 }
497 // -XX:-InlineClassNatives disables natives from the Class class.
498 // The flag applies to all reflective calls, notably Array.newArray
499 // (visible to Java programmers as Array.newInstance).
500 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
501 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
502 if (!InlineClassNatives) return NULL;
503 }
505 // -XX:-InlineThreadNatives disables natives from the Thread class.
506 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
507 if (!InlineThreadNatives) return NULL;
508 }
510 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
511 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
512 m->holder()->name() == ciSymbol::java_lang_Float() ||
513 m->holder()->name() == ciSymbol::java_lang_Double()) {
514 if (!InlineMathNatives) return NULL;
515 }
517 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
518 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
519 if (!InlineUnsafeOps) return NULL;
520 }
522 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id);
523 }
525 //----------------------register_library_intrinsics-----------------------
526 // Initialize this file's data structures, for each Compile instance.
527 void Compile::register_library_intrinsics() {
528 // Nothing to do here.
529 }
531 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
532 LibraryCallKit kit(jvms, this);
533 Compile* C = kit.C;
534 int nodes = C->unique();
535 #ifndef PRODUCT
536 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
537 char buf[1000];
538 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
539 tty->print_cr("Intrinsic %s", str);
540 }
541 #endif
542 ciMethod* callee = kit.callee();
543 const int bci = kit.bci();
545 // Try to inline the intrinsic.
546 if (kit.try_to_inline()) {
547 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
548 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
549 }
550 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
551 if (C->log()) {
552 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
553 vmIntrinsics::name_at(intrinsic_id()),
554 (is_virtual() ? " virtual='1'" : ""),
555 C->unique() - nodes);
556 }
557 // Push the result from the inlined method onto the stack.
558 kit.push_result();
559 return kit.transfer_exceptions_into_jvms();
560 }
562 // The intrinsic bailed out
563 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
564 if (jvms->has_method()) {
565 // Not a root compile.
566 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
567 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
568 } else {
569 // Root compile
570 tty->print("Did not generate intrinsic %s%s at bci:%d in",
571 vmIntrinsics::name_at(intrinsic_id()),
572 (is_virtual() ? " (virtual)" : ""), bci);
573 }
574 }
575 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
576 return NULL;
577 }
579 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
580 LibraryCallKit kit(jvms, this);
581 Compile* C = kit.C;
582 int nodes = C->unique();
583 #ifndef PRODUCT
584 assert(is_predicted(), "sanity");
585 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
586 char buf[1000];
587 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
588 tty->print_cr("Predicate for intrinsic %s", str);
589 }
590 #endif
591 ciMethod* callee = kit.callee();
592 const int bci = kit.bci();
594 Node* slow_ctl = kit.try_to_predicate();
595 if (!kit.failing()) {
596 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
597 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
598 }
599 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
600 if (C->log()) {
601 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
602 vmIntrinsics::name_at(intrinsic_id()),
603 (is_virtual() ? " virtual='1'" : ""),
604 C->unique() - nodes);
605 }
606 return slow_ctl; // Could be NULL if the check folds.
607 }
609 // The intrinsic bailed out
610 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
611 if (jvms->has_method()) {
612 // Not a root compile.
613 const char* msg = "failed to generate predicate for intrinsic";
614 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
615 } else {
616 // Root compile
617 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
618 vmIntrinsics::name_at(intrinsic_id()),
619 (is_virtual() ? " (virtual)" : ""), bci);
620 }
621 }
622 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
623 return NULL;
624 }
626 bool LibraryCallKit::try_to_inline() {
627 // Handle symbolic names for otherwise undistinguished boolean switches:
628 const bool is_store = true;
629 const bool is_native_ptr = true;
630 const bool is_static = true;
631 const bool is_volatile = true;
633 if (!jvms()->has_method()) {
634 // Root JVMState has a null method.
635 assert(map()->memory()->Opcode() == Op_Parm, "");
636 // Insert the memory aliasing node
637 set_all_memory(reset_memory());
638 }
639 assert(merged_memory(), "");
642 switch (intrinsic_id()) {
643 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
644 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
645 case vmIntrinsics::_getClass: return inline_native_getClass();
647 case vmIntrinsics::_dsin:
648 case vmIntrinsics::_dcos:
649 case vmIntrinsics::_dtan:
650 case vmIntrinsics::_dabs:
651 case vmIntrinsics::_datan2:
652 case vmIntrinsics::_dsqrt:
653 case vmIntrinsics::_dexp:
654 case vmIntrinsics::_dlog:
655 case vmIntrinsics::_dlog10:
656 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
658 case vmIntrinsics::_min:
659 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
661 case vmIntrinsics::_arraycopy: return inline_arraycopy();
663 case vmIntrinsics::_compareTo: return inline_string_compareTo();
664 case vmIntrinsics::_indexOf: return inline_string_indexOf();
665 case vmIntrinsics::_equals: return inline_string_equals();
667 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
668 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
669 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
670 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
671 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
672 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
673 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
674 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
675 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
677 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
678 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
679 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
680 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
681 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
682 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
683 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
684 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
685 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
687 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
688 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
689 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
690 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
691 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
692 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
693 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
694 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
696 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
697 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
698 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
699 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
700 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
701 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
702 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
703 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
705 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
706 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
707 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
708 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
709 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
710 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
711 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
712 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
713 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
715 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
716 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
717 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
718 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
719 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
720 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
721 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
722 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
723 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
725 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
726 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
727 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
728 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
730 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
731 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
732 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
734 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
735 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
736 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
738 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
739 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
740 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
741 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
742 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
744 case vmIntrinsics::_loadFence:
745 case vmIntrinsics::_storeFence:
746 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
748 case vmIntrinsics::_currentThread: return inline_native_currentThread();
749 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
751 #ifdef TRACE_HAVE_INTRINSICS
752 case vmIntrinsics::_classID: return inline_native_classID();
753 case vmIntrinsics::_threadID: return inline_native_threadID();
754 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
755 #endif
756 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
757 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
758 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
759 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
760 case vmIntrinsics::_newArray: return inline_native_newArray();
761 case vmIntrinsics::_getLength: return inline_native_getLength();
762 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
763 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
764 case vmIntrinsics::_equalsC: return inline_array_equals();
765 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
767 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
769 case vmIntrinsics::_isInstance:
770 case vmIntrinsics::_getModifiers:
771 case vmIntrinsics::_isInterface:
772 case vmIntrinsics::_isArray:
773 case vmIntrinsics::_isPrimitive:
774 case vmIntrinsics::_getSuperclass:
775 case vmIntrinsics::_getComponentType:
776 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
778 case vmIntrinsics::_floatToRawIntBits:
779 case vmIntrinsics::_floatToIntBits:
780 case vmIntrinsics::_intBitsToFloat:
781 case vmIntrinsics::_doubleToRawLongBits:
782 case vmIntrinsics::_doubleToLongBits:
783 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
785 case vmIntrinsics::_numberOfLeadingZeros_i:
786 case vmIntrinsics::_numberOfLeadingZeros_l:
787 case vmIntrinsics::_numberOfTrailingZeros_i:
788 case vmIntrinsics::_numberOfTrailingZeros_l:
789 case vmIntrinsics::_bitCount_i:
790 case vmIntrinsics::_bitCount_l:
791 case vmIntrinsics::_reverseBytes_i:
792 case vmIntrinsics::_reverseBytes_l:
793 case vmIntrinsics::_reverseBytes_s:
794 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
796 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
798 case vmIntrinsics::_Reference_get: return inline_reference_get();
800 case vmIntrinsics::_aescrypt_encryptBlock:
801 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
803 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
804 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
805 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
807 case vmIntrinsics::_encodeISOArray:
808 return inline_encodeISOArray();
810 default:
811 // If you get here, it may be that someone has added a new intrinsic
812 // to the list in vmSymbols.hpp without implementing it here.
813 #ifndef PRODUCT
814 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
815 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
816 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
817 }
818 #endif
819 return false;
820 }
821 }
823 Node* LibraryCallKit::try_to_predicate() {
824 if (!jvms()->has_method()) {
825 // Root JVMState has a null method.
826 assert(map()->memory()->Opcode() == Op_Parm, "");
827 // Insert the memory aliasing node
828 set_all_memory(reset_memory());
829 }
830 assert(merged_memory(), "");
832 switch (intrinsic_id()) {
833 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
834 return inline_cipherBlockChaining_AESCrypt_predicate(false);
835 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
836 return inline_cipherBlockChaining_AESCrypt_predicate(true);
838 default:
839 // If you get here, it may be that someone has added a new intrinsic
840 // to the list in vmSymbols.hpp without implementing it here.
841 #ifndef PRODUCT
842 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
843 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
844 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
845 }
846 #endif
847 Node* slow_ctl = control();
848 set_control(top()); // No fast path instrinsic
849 return slow_ctl;
850 }
851 }
853 //------------------------------set_result-------------------------------
854 // Helper function for finishing intrinsics.
855 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
856 record_for_igvn(region);
857 set_control(_gvn.transform(region));
858 set_result( _gvn.transform(value));
859 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
860 }
862 //------------------------------generate_guard---------------------------
863 // Helper function for generating guarded fast-slow graph structures.
864 // The given 'test', if true, guards a slow path. If the test fails
865 // then a fast path can be taken. (We generally hope it fails.)
866 // In all cases, GraphKit::control() is updated to the fast path.
867 // The returned value represents the control for the slow path.
868 // The return value is never 'top'; it is either a valid control
869 // or NULL if it is obvious that the slow path can never be taken.
870 // Also, if region and the slow control are not NULL, the slow edge
871 // is appended to the region.
872 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
873 if (stopped()) {
874 // Already short circuited.
875 return NULL;
876 }
878 // Build an if node and its projections.
879 // If test is true we take the slow path, which we assume is uncommon.
880 if (_gvn.type(test) == TypeInt::ZERO) {
881 // The slow branch is never taken. No need to build this guard.
882 return NULL;
883 }
885 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
887 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
888 if (if_slow == top()) {
889 // The slow branch is never taken. No need to build this guard.
890 return NULL;
891 }
893 if (region != NULL)
894 region->add_req(if_slow);
896 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
897 set_control(if_fast);
899 return if_slow;
900 }
902 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
903 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
904 }
905 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
906 return generate_guard(test, region, PROB_FAIR);
907 }
909 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
910 Node* *pos_index) {
911 if (stopped())
912 return NULL; // already stopped
913 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
914 return NULL; // index is already adequately typed
915 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
916 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
917 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
918 if (is_neg != NULL && pos_index != NULL) {
919 // Emulate effect of Parse::adjust_map_after_if.
920 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
921 ccast->set_req(0, control());
922 (*pos_index) = _gvn.transform(ccast);
923 }
924 return is_neg;
925 }
927 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
928 Node* *pos_index) {
929 if (stopped())
930 return NULL; // already stopped
931 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
932 return NULL; // index is already adequately typed
933 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
934 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
935 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
936 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
937 if (is_notp != NULL && pos_index != NULL) {
938 // Emulate effect of Parse::adjust_map_after_if.
939 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
940 ccast->set_req(0, control());
941 (*pos_index) = _gvn.transform(ccast);
942 }
943 return is_notp;
944 }
946 // Make sure that 'position' is a valid limit index, in [0..length].
947 // There are two equivalent plans for checking this:
948 // A. (offset + copyLength) unsigned<= arrayLength
949 // B. offset <= (arrayLength - copyLength)
950 // We require that all of the values above, except for the sum and
951 // difference, are already known to be non-negative.
952 // Plan A is robust in the face of overflow, if offset and copyLength
953 // are both hugely positive.
954 //
955 // Plan B is less direct and intuitive, but it does not overflow at
956 // all, since the difference of two non-negatives is always
957 // representable. Whenever Java methods must perform the equivalent
958 // check they generally use Plan B instead of Plan A.
959 // For the moment we use Plan A.
960 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
961 Node* subseq_length,
962 Node* array_length,
963 RegionNode* region) {
964 if (stopped())
965 return NULL; // already stopped
966 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
967 if (zero_offset && subseq_length->eqv_uncast(array_length))
968 return NULL; // common case of whole-array copy
969 Node* last = subseq_length;
970 if (!zero_offset) // last += offset
971 last = _gvn.transform( new (C) AddINode(last, offset));
972 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
973 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
974 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
975 return is_over;
976 }
979 //--------------------------generate_current_thread--------------------
980 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
981 ciKlass* thread_klass = env()->Thread_klass();
982 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
983 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
984 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
985 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
986 tls_output = thread;
987 return threadObj;
988 }
991 //------------------------------make_string_method_node------------------------
992 // Helper method for String intrinsic functions. This version is called
993 // with str1 and str2 pointing to String object nodes.
994 //
995 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
996 Node* no_ctrl = NULL;
998 // Get start addr of string
999 Node* str1_value = load_String_value(no_ctrl, str1);
1000 Node* str1_offset = load_String_offset(no_ctrl, str1);
1001 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1003 // Get length of string 1
1004 Node* str1_len = load_String_length(no_ctrl, str1);
1006 Node* str2_value = load_String_value(no_ctrl, str2);
1007 Node* str2_offset = load_String_offset(no_ctrl, str2);
1008 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1010 Node* str2_len = NULL;
1011 Node* result = NULL;
1013 switch (opcode) {
1014 case Op_StrIndexOf:
1015 // Get length of string 2
1016 str2_len = load_String_length(no_ctrl, str2);
1018 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1019 str1_start, str1_len, str2_start, str2_len);
1020 break;
1021 case Op_StrComp:
1022 // Get length of string 2
1023 str2_len = load_String_length(no_ctrl, str2);
1025 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1026 str1_start, str1_len, str2_start, str2_len);
1027 break;
1028 case Op_StrEquals:
1029 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1030 str1_start, str2_start, str1_len);
1031 break;
1032 default:
1033 ShouldNotReachHere();
1034 return NULL;
1035 }
1037 // All these intrinsics have checks.
1038 C->set_has_split_ifs(true); // Has chance for split-if optimization
1040 return _gvn.transform(result);
1041 }
1043 // Helper method for String intrinsic functions. This version is called
1044 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1045 // to Int nodes containing the lenghts of str1 and str2.
1046 //
1047 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1048 Node* result = NULL;
1049 switch (opcode) {
1050 case Op_StrIndexOf:
1051 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1052 str1_start, cnt1, str2_start, cnt2);
1053 break;
1054 case Op_StrComp:
1055 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1056 str1_start, cnt1, str2_start, cnt2);
1057 break;
1058 case Op_StrEquals:
1059 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1060 str1_start, str2_start, cnt1);
1061 break;
1062 default:
1063 ShouldNotReachHere();
1064 return NULL;
1065 }
1067 // All these intrinsics have checks.
1068 C->set_has_split_ifs(true); // Has chance for split-if optimization
1070 return _gvn.transform(result);
1071 }
1073 //------------------------------inline_string_compareTo------------------------
1074 // public int java.lang.String.compareTo(String anotherString);
1075 bool LibraryCallKit::inline_string_compareTo() {
1076 Node* receiver = null_check(argument(0));
1077 Node* arg = null_check(argument(1));
1078 if (stopped()) {
1079 return true;
1080 }
1081 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1082 return true;
1083 }
1085 //------------------------------inline_string_equals------------------------
1086 bool LibraryCallKit::inline_string_equals() {
1087 Node* receiver = null_check_receiver();
1088 // NOTE: Do not null check argument for String.equals() because spec
1089 // allows to specify NULL as argument.
1090 Node* argument = this->argument(1);
1091 if (stopped()) {
1092 return true;
1093 }
1095 // paths (plus control) merge
1096 RegionNode* region = new (C) RegionNode(5);
1097 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1099 // does source == target string?
1100 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1101 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1103 Node* if_eq = generate_slow_guard(bol, NULL);
1104 if (if_eq != NULL) {
1105 // receiver == argument
1106 phi->init_req(2, intcon(1));
1107 region->init_req(2, if_eq);
1108 }
1110 // get String klass for instanceOf
1111 ciInstanceKlass* klass = env()->String_klass();
1113 if (!stopped()) {
1114 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1115 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1116 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1118 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1119 //instanceOf == true, fallthrough
1121 if (inst_false != NULL) {
1122 phi->init_req(3, intcon(0));
1123 region->init_req(3, inst_false);
1124 }
1125 }
1127 if (!stopped()) {
1128 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1130 // Properly cast the argument to String
1131 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1132 // This path is taken only when argument's type is String:NotNull.
1133 argument = cast_not_null(argument, false);
1135 Node* no_ctrl = NULL;
1137 // Get start addr of receiver
1138 Node* receiver_val = load_String_value(no_ctrl, receiver);
1139 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1140 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1142 // Get length of receiver
1143 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1145 // Get start addr of argument
1146 Node* argument_val = load_String_value(no_ctrl, argument);
1147 Node* argument_offset = load_String_offset(no_ctrl, argument);
1148 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1150 // Get length of argument
1151 Node* argument_cnt = load_String_length(no_ctrl, argument);
1153 // Check for receiver count != argument count
1154 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1155 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1156 Node* if_ne = generate_slow_guard(bol, NULL);
1157 if (if_ne != NULL) {
1158 phi->init_req(4, intcon(0));
1159 region->init_req(4, if_ne);
1160 }
1162 // Check for count == 0 is done by assembler code for StrEquals.
1164 if (!stopped()) {
1165 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1166 phi->init_req(1, equals);
1167 region->init_req(1, control());
1168 }
1169 }
1171 // post merge
1172 set_control(_gvn.transform(region));
1173 record_for_igvn(region);
1175 set_result(_gvn.transform(phi));
1176 return true;
1177 }
1179 //------------------------------inline_array_equals----------------------------
1180 bool LibraryCallKit::inline_array_equals() {
1181 Node* arg1 = argument(0);
1182 Node* arg2 = argument(1);
1183 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1184 return true;
1185 }
1187 // Java version of String.indexOf(constant string)
1188 // class StringDecl {
1189 // StringDecl(char[] ca) {
1190 // offset = 0;
1191 // count = ca.length;
1192 // value = ca;
1193 // }
1194 // int offset;
1195 // int count;
1196 // char[] value;
1197 // }
1198 //
1199 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1200 // int targetOffset, int cache_i, int md2) {
1201 // int cache = cache_i;
1202 // int sourceOffset = string_object.offset;
1203 // int sourceCount = string_object.count;
1204 // int targetCount = target_object.length;
1205 //
1206 // int targetCountLess1 = targetCount - 1;
1207 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1208 //
1209 // char[] source = string_object.value;
1210 // char[] target = target_object;
1211 // int lastChar = target[targetCountLess1];
1212 //
1213 // outer_loop:
1214 // for (int i = sourceOffset; i < sourceEnd; ) {
1215 // int src = source[i + targetCountLess1];
1216 // if (src == lastChar) {
1217 // // With random strings and a 4-character alphabet,
1218 // // reverse matching at this point sets up 0.8% fewer
1219 // // frames, but (paradoxically) makes 0.3% more probes.
1220 // // Since those probes are nearer the lastChar probe,
1221 // // there is may be a net D$ win with reverse matching.
1222 // // But, reversing loop inhibits unroll of inner loop
1223 // // for unknown reason. So, does running outer loop from
1224 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1225 // for (int j = 0; j < targetCountLess1; j++) {
1226 // if (target[targetOffset + j] != source[i+j]) {
1227 // if ((cache & (1 << source[i+j])) == 0) {
1228 // if (md2 < j+1) {
1229 // i += j+1;
1230 // continue outer_loop;
1231 // }
1232 // }
1233 // i += md2;
1234 // continue outer_loop;
1235 // }
1236 // }
1237 // return i - sourceOffset;
1238 // }
1239 // if ((cache & (1 << src)) == 0) {
1240 // i += targetCountLess1;
1241 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1242 // i++;
1243 // }
1244 // return -1;
1245 // }
1247 //------------------------------string_indexOf------------------------
1248 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1249 jint cache_i, jint md2_i) {
1251 Node* no_ctrl = NULL;
1252 float likely = PROB_LIKELY(0.9);
1253 float unlikely = PROB_UNLIKELY(0.9);
1255 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1257 Node* source = load_String_value(no_ctrl, string_object);
1258 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1259 Node* sourceCount = load_String_length(no_ctrl, string_object);
1261 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1262 jint target_length = target_array->length();
1263 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1264 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1266 IdealKit kit(this, false, true);
1267 #define __ kit.
1268 Node* zero = __ ConI(0);
1269 Node* one = __ ConI(1);
1270 Node* cache = __ ConI(cache_i);
1271 Node* md2 = __ ConI(md2_i);
1272 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1273 Node* targetCount = __ ConI(target_length);
1274 Node* targetCountLess1 = __ ConI(target_length - 1);
1275 Node* targetOffset = __ ConI(targetOffset_i);
1276 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1278 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1279 Node* outer_loop = __ make_label(2 /* goto */);
1280 Node* return_ = __ make_label(1);
1282 __ set(rtn,__ ConI(-1));
1283 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1284 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1285 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1286 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1287 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1288 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1289 Node* tpj = __ AddI(targetOffset, __ value(j));
1290 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1291 Node* ipj = __ AddI(__ value(i), __ value(j));
1292 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1293 __ if_then(targ, BoolTest::ne, src2); {
1294 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1295 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1296 __ increment(i, __ AddI(__ value(j), one));
1297 __ goto_(outer_loop);
1298 } __ end_if(); __ dead(j);
1299 }__ end_if(); __ dead(j);
1300 __ increment(i, md2);
1301 __ goto_(outer_loop);
1302 }__ end_if();
1303 __ increment(j, one);
1304 }__ end_loop(); __ dead(j);
1305 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1306 __ goto_(return_);
1307 }__ end_if();
1308 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1309 __ increment(i, targetCountLess1);
1310 }__ end_if();
1311 __ increment(i, one);
1312 __ bind(outer_loop);
1313 }__ end_loop(); __ dead(i);
1314 __ bind(return_);
1316 // Final sync IdealKit and GraphKit.
1317 final_sync(kit);
1318 Node* result = __ value(rtn);
1319 #undef __
1320 C->set_has_loops(true);
1321 return result;
1322 }
1324 //------------------------------inline_string_indexOf------------------------
1325 bool LibraryCallKit::inline_string_indexOf() {
1326 Node* receiver = argument(0);
1327 Node* arg = argument(1);
1329 Node* result;
1330 // Disable the use of pcmpestri until it can be guaranteed that
1331 // the load doesn't cross into the uncommited space.
1332 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1333 UseSSE42Intrinsics) {
1334 // Generate SSE4.2 version of indexOf
1335 // We currently only have match rules that use SSE4.2
1337 receiver = null_check(receiver);
1338 arg = null_check(arg);
1339 if (stopped()) {
1340 return true;
1341 }
1343 ciInstanceKlass* str_klass = env()->String_klass();
1344 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1346 // Make the merge point
1347 RegionNode* result_rgn = new (C) RegionNode(4);
1348 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1349 Node* no_ctrl = NULL;
1351 // Get start addr of source string
1352 Node* source = load_String_value(no_ctrl, receiver);
1353 Node* source_offset = load_String_offset(no_ctrl, receiver);
1354 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1356 // Get length of source string
1357 Node* source_cnt = load_String_length(no_ctrl, receiver);
1359 // Get start addr of substring
1360 Node* substr = load_String_value(no_ctrl, arg);
1361 Node* substr_offset = load_String_offset(no_ctrl, arg);
1362 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1364 // Get length of source string
1365 Node* substr_cnt = load_String_length(no_ctrl, arg);
1367 // Check for substr count > string count
1368 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1369 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1370 Node* if_gt = generate_slow_guard(bol, NULL);
1371 if (if_gt != NULL) {
1372 result_phi->init_req(2, intcon(-1));
1373 result_rgn->init_req(2, if_gt);
1374 }
1376 if (!stopped()) {
1377 // Check for substr count == 0
1378 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1379 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1380 Node* if_zero = generate_slow_guard(bol, NULL);
1381 if (if_zero != NULL) {
1382 result_phi->init_req(3, intcon(0));
1383 result_rgn->init_req(3, if_zero);
1384 }
1385 }
1387 if (!stopped()) {
1388 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1389 result_phi->init_req(1, result);
1390 result_rgn->init_req(1, control());
1391 }
1392 set_control(_gvn.transform(result_rgn));
1393 record_for_igvn(result_rgn);
1394 result = _gvn.transform(result_phi);
1396 } else { // Use LibraryCallKit::string_indexOf
1397 // don't intrinsify if argument isn't a constant string.
1398 if (!arg->is_Con()) {
1399 return false;
1400 }
1401 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1402 if (str_type == NULL) {
1403 return false;
1404 }
1405 ciInstanceKlass* klass = env()->String_klass();
1406 ciObject* str_const = str_type->const_oop();
1407 if (str_const == NULL || str_const->klass() != klass) {
1408 return false;
1409 }
1410 ciInstance* str = str_const->as_instance();
1411 assert(str != NULL, "must be instance");
1413 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1414 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1416 int o;
1417 int c;
1418 if (java_lang_String::has_offset_field()) {
1419 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1420 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1421 } else {
1422 o = 0;
1423 c = pat->length();
1424 }
1426 // constant strings have no offset and count == length which
1427 // simplifies the resulting code somewhat so lets optimize for that.
1428 if (o != 0 || c != pat->length()) {
1429 return false;
1430 }
1432 receiver = null_check(receiver, T_OBJECT);
1433 // NOTE: No null check on the argument is needed since it's a constant String oop.
1434 if (stopped()) {
1435 return true;
1436 }
1438 // The null string as a pattern always returns 0 (match at beginning of string)
1439 if (c == 0) {
1440 set_result(intcon(0));
1441 return true;
1442 }
1444 // Generate default indexOf
1445 jchar lastChar = pat->char_at(o + (c - 1));
1446 int cache = 0;
1447 int i;
1448 for (i = 0; i < c - 1; i++) {
1449 assert(i < pat->length(), "out of range");
1450 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1451 }
1453 int md2 = c;
1454 for (i = 0; i < c - 1; i++) {
1455 assert(i < pat->length(), "out of range");
1456 if (pat->char_at(o + i) == lastChar) {
1457 md2 = (c - 1) - i;
1458 }
1459 }
1461 result = string_indexOf(receiver, pat, o, cache, md2);
1462 }
1463 set_result(result);
1464 return true;
1465 }
1467 //--------------------------round_double_node--------------------------------
1468 // Round a double node if necessary.
1469 Node* LibraryCallKit::round_double_node(Node* n) {
1470 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1471 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1472 return n;
1473 }
1475 //------------------------------inline_math-----------------------------------
1476 // public static double Math.abs(double)
1477 // public static double Math.sqrt(double)
1478 // public static double Math.log(double)
1479 // public static double Math.log10(double)
1480 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1481 Node* arg = round_double_node(argument(0));
1482 Node* n;
1483 switch (id) {
1484 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1485 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(0, arg); break;
1486 case vmIntrinsics::_dlog: n = new (C) LogDNode( arg); break;
1487 case vmIntrinsics::_dlog10: n = new (C) Log10DNode( arg); break;
1488 default: fatal_unexpected_iid(id); break;
1489 }
1490 set_result(_gvn.transform(n));
1491 return true;
1492 }
1494 //------------------------------inline_trig----------------------------------
1495 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1496 // argument reduction which will turn into a fast/slow diamond.
1497 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1498 Node* arg = round_double_node(argument(0));
1499 Node* n = NULL;
1501 switch (id) {
1502 case vmIntrinsics::_dsin: n = new (C) SinDNode(arg); break;
1503 case vmIntrinsics::_dcos: n = new (C) CosDNode(arg); break;
1504 case vmIntrinsics::_dtan: n = new (C) TanDNode(arg); break;
1505 default: fatal_unexpected_iid(id); break;
1506 }
1507 n = _gvn.transform(n);
1509 // Rounding required? Check for argument reduction!
1510 if (Matcher::strict_fp_requires_explicit_rounding) {
1511 static const double pi_4 = 0.7853981633974483;
1512 static const double neg_pi_4 = -0.7853981633974483;
1513 // pi/2 in 80-bit extended precision
1514 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1515 // -pi/2 in 80-bit extended precision
1516 // 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};
1517 // Cutoff value for using this argument reduction technique
1518 //static const double pi_2_minus_epsilon = 1.564660403643354;
1519 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1521 // Pseudocode for sin:
1522 // if (x <= Math.PI / 4.0) {
1523 // if (x >= -Math.PI / 4.0) return fsin(x);
1524 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1525 // } else {
1526 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1527 // }
1528 // return StrictMath.sin(x);
1530 // Pseudocode for cos:
1531 // if (x <= Math.PI / 4.0) {
1532 // if (x >= -Math.PI / 4.0) return fcos(x);
1533 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1534 // } else {
1535 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1536 // }
1537 // return StrictMath.cos(x);
1539 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1540 // requires a special machine instruction to load it. Instead we'll try
1541 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1542 // probably do the math inside the SIN encoding.
1544 // Make the merge point
1545 RegionNode* r = new (C) RegionNode(3);
1546 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1548 // Flatten arg so we need only 1 test
1549 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1550 // Node for PI/4 constant
1551 Node *pi4 = makecon(TypeD::make(pi_4));
1552 // Check PI/4 : abs(arg)
1553 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1554 // Check: If PI/4 < abs(arg) then go slow
1555 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1556 // Branch either way
1557 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1558 set_control(opt_iff(r,iff));
1560 // Set fast path result
1561 phi->init_req(2, n);
1563 // Slow path - non-blocking leaf call
1564 Node* call = NULL;
1565 switch (id) {
1566 case vmIntrinsics::_dsin:
1567 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1568 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1569 "Sin", NULL, arg, top());
1570 break;
1571 case vmIntrinsics::_dcos:
1572 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1573 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1574 "Cos", NULL, arg, top());
1575 break;
1576 case vmIntrinsics::_dtan:
1577 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1578 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1579 "Tan", NULL, arg, top());
1580 break;
1581 }
1582 assert(control()->in(0) == call, "");
1583 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1584 r->init_req(1, control());
1585 phi->init_req(1, slow_result);
1587 // Post-merge
1588 set_control(_gvn.transform(r));
1589 record_for_igvn(r);
1590 n = _gvn.transform(phi);
1592 C->set_has_split_ifs(true); // Has chance for split-if optimization
1593 }
1594 set_result(n);
1595 return true;
1596 }
1598 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1599 //-------------------
1600 //result=(result.isNaN())? funcAddr():result;
1601 // Check: If isNaN() by checking result!=result? then either trap
1602 // or go to runtime
1603 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1604 // Build the boolean node
1605 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1607 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1608 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1609 // The pow or exp intrinsic returned a NaN, which requires a call
1610 // to the runtime. Recompile with the runtime call.
1611 uncommon_trap(Deoptimization::Reason_intrinsic,
1612 Deoptimization::Action_make_not_entrant);
1613 }
1614 set_result(result);
1615 } else {
1616 // If this inlining ever returned NaN in the past, we compile a call
1617 // to the runtime to properly handle corner cases
1619 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1620 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1621 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1623 if (!if_slow->is_top()) {
1624 RegionNode* result_region = new (C) RegionNode(3);
1625 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1627 result_region->init_req(1, if_fast);
1628 result_val->init_req(1, result);
1630 set_control(if_slow);
1632 const TypePtr* no_memory_effects = NULL;
1633 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1634 no_memory_effects,
1635 x, top(), y, y ? top() : NULL);
1636 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1637 #ifdef ASSERT
1638 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1639 assert(value_top == top(), "second value must be top");
1640 #endif
1642 result_region->init_req(2, control());
1643 result_val->init_req(2, value);
1644 set_result(result_region, result_val);
1645 } else {
1646 set_result(result);
1647 }
1648 }
1649 }
1651 //------------------------------inline_exp-------------------------------------
1652 // Inline exp instructions, if possible. The Intel hardware only misses
1653 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1654 bool LibraryCallKit::inline_exp() {
1655 Node* arg = round_double_node(argument(0));
1656 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1658 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1660 C->set_has_split_ifs(true); // Has chance for split-if optimization
1661 return true;
1662 }
1664 //------------------------------inline_pow-------------------------------------
1665 // Inline power instructions, if possible.
1666 bool LibraryCallKit::inline_pow() {
1667 // Pseudocode for pow
1668 // if (x <= 0.0) {
1669 // long longy = (long)y;
1670 // if ((double)longy == y) { // if y is long
1671 // if (y + 1 == y) longy = 0; // huge number: even
1672 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1673 // } else {
1674 // result = NaN;
1675 // }
1676 // } else {
1677 // result = DPow(x,y);
1678 // }
1679 // if (result != result)? {
1680 // result = uncommon_trap() or runtime_call();
1681 // }
1682 // return result;
1684 Node* x = round_double_node(argument(0));
1685 Node* y = round_double_node(argument(2));
1687 Node* result = NULL;
1689 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1690 // Short form: skip the fancy tests and just check for NaN result.
1691 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1692 } else {
1693 // If this inlining ever returned NaN in the past, include all
1694 // checks + call to the runtime.
1696 // Set the merge point for If node with condition of (x <= 0.0)
1697 // There are four possible paths to region node and phi node
1698 RegionNode *r = new (C) RegionNode(4);
1699 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1701 // Build the first if node: if (x <= 0.0)
1702 // Node for 0 constant
1703 Node *zeronode = makecon(TypeD::ZERO);
1704 // Check x:0
1705 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1706 // Check: If (x<=0) then go complex path
1707 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1708 // Branch either way
1709 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1710 // Fast path taken; set region slot 3
1711 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1712 r->init_req(3,fast_taken); // Capture fast-control
1714 // Fast path not-taken, i.e. slow path
1715 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1717 // Set fast path result
1718 Node *fast_result = _gvn.transform( new (C) PowDNode(C, control(), x, y) );
1719 phi->init_req(3, fast_result);
1721 // Complex path
1722 // Build the second if node (if y is long)
1723 // Node for (long)y
1724 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1725 // Node for (double)((long) y)
1726 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1727 // Check (double)((long) y) : y
1728 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1729 // Check if (y isn't long) then go to slow path
1731 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1732 // Branch either way
1733 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1734 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1736 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1738 // Calculate DPow(abs(x), y)*(1 & (long)y)
1739 // Node for constant 1
1740 Node *conone = longcon(1);
1741 // 1& (long)y
1742 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1744 // A huge number is always even. Detect a huge number by checking
1745 // if y + 1 == y and set integer to be tested for parity to 0.
1746 // Required for corner case:
1747 // (long)9.223372036854776E18 = max_jlong
1748 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1749 // max_jlong is odd but 9.223372036854776E18 is even
1750 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1751 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1752 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1753 Node* correctedsign = NULL;
1754 if (ConditionalMoveLimit != 0) {
1755 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1756 } else {
1757 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1758 RegionNode *r = new (C) RegionNode(3);
1759 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1760 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1761 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1762 phi->init_req(1, signnode);
1763 phi->init_req(2, longcon(0));
1764 correctedsign = _gvn.transform(phi);
1765 ylong_path = _gvn.transform(r);
1766 record_for_igvn(r);
1767 }
1769 // zero node
1770 Node *conzero = longcon(0);
1771 // Check (1&(long)y)==0?
1772 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1773 // Check if (1&(long)y)!=0?, if so the result is negative
1774 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1775 // abs(x)
1776 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1777 // abs(x)^y
1778 Node *absxpowy = _gvn.transform( new (C) PowDNode(C, control(), absx, y) );
1779 // -abs(x)^y
1780 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1781 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1782 Node *signresult = NULL;
1783 if (ConditionalMoveLimit != 0) {
1784 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1785 } else {
1786 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1787 RegionNode *r = new (C) RegionNode(3);
1788 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1789 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1790 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1791 phi->init_req(1, absxpowy);
1792 phi->init_req(2, negabsxpowy);
1793 signresult = _gvn.transform(phi);
1794 ylong_path = _gvn.transform(r);
1795 record_for_igvn(r);
1796 }
1797 // Set complex path fast result
1798 r->init_req(2, ylong_path);
1799 phi->init_req(2, signresult);
1801 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1802 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1803 r->init_req(1,slow_path);
1804 phi->init_req(1,slow_result);
1806 // Post merge
1807 set_control(_gvn.transform(r));
1808 record_for_igvn(r);
1809 result = _gvn.transform(phi);
1810 }
1812 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1814 C->set_has_split_ifs(true); // Has chance for split-if optimization
1815 return true;
1816 }
1818 //------------------------------runtime_math-----------------------------
1819 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1820 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1821 "must be (DD)D or (D)D type");
1823 // Inputs
1824 Node* a = round_double_node(argument(0));
1825 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1827 const TypePtr* no_memory_effects = NULL;
1828 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1829 no_memory_effects,
1830 a, top(), b, b ? top() : NULL);
1831 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1832 #ifdef ASSERT
1833 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1834 assert(value_top == top(), "second value must be top");
1835 #endif
1837 set_result(value);
1838 return true;
1839 }
1841 //------------------------------inline_math_native-----------------------------
1842 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1843 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1844 switch (id) {
1845 // These intrinsics are not properly supported on all hardware
1846 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1847 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1848 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1849 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1850 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1851 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1853 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1854 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1855 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1856 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1858 // These intrinsics are supported on all hardware
1859 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1860 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1862 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1863 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1864 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1865 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1866 #undef FN_PTR
1868 // These intrinsics are not yet correctly implemented
1869 case vmIntrinsics::_datan2:
1870 return false;
1872 default:
1873 fatal_unexpected_iid(id);
1874 return false;
1875 }
1876 }
1878 static bool is_simple_name(Node* n) {
1879 return (n->req() == 1 // constant
1880 || (n->is_Type() && n->as_Type()->type()->singleton())
1881 || n->is_Proj() // parameter or return value
1882 || n->is_Phi() // local of some sort
1883 );
1884 }
1886 //----------------------------inline_min_max-----------------------------------
1887 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1888 set_result(generate_min_max(id, argument(0), argument(1)));
1889 return true;
1890 }
1892 Node*
1893 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1894 // These are the candidate return value:
1895 Node* xvalue = x0;
1896 Node* yvalue = y0;
1898 if (xvalue == yvalue) {
1899 return xvalue;
1900 }
1902 bool want_max = (id == vmIntrinsics::_max);
1904 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1905 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1906 if (txvalue == NULL || tyvalue == NULL) return top();
1907 // This is not really necessary, but it is consistent with a
1908 // hypothetical MaxINode::Value method:
1909 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1911 // %%% This folding logic should (ideally) be in a different place.
1912 // Some should be inside IfNode, and there to be a more reliable
1913 // transformation of ?: style patterns into cmoves. We also want
1914 // more powerful optimizations around cmove and min/max.
1916 // Try to find a dominating comparison of these guys.
1917 // It can simplify the index computation for Arrays.copyOf
1918 // and similar uses of System.arraycopy.
1919 // First, compute the normalized version of CmpI(x, y).
1920 int cmp_op = Op_CmpI;
1921 Node* xkey = xvalue;
1922 Node* ykey = yvalue;
1923 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
1924 if (ideal_cmpxy->is_Cmp()) {
1925 // E.g., if we have CmpI(length - offset, count),
1926 // it might idealize to CmpI(length, count + offset)
1927 cmp_op = ideal_cmpxy->Opcode();
1928 xkey = ideal_cmpxy->in(1);
1929 ykey = ideal_cmpxy->in(2);
1930 }
1932 // Start by locating any relevant comparisons.
1933 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1934 Node* cmpxy = NULL;
1935 Node* cmpyx = NULL;
1936 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1937 Node* cmp = start_from->fast_out(k);
1938 if (cmp->outcnt() > 0 && // must have prior uses
1939 cmp->in(0) == NULL && // must be context-independent
1940 cmp->Opcode() == cmp_op) { // right kind of compare
1941 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1942 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1943 }
1944 }
1946 const int NCMPS = 2;
1947 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1948 int cmpn;
1949 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1950 if (cmps[cmpn] != NULL) break; // find a result
1951 }
1952 if (cmpn < NCMPS) {
1953 // Look for a dominating test that tells us the min and max.
1954 int depth = 0; // Limit search depth for speed
1955 Node* dom = control();
1956 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1957 if (++depth >= 100) break;
1958 Node* ifproj = dom;
1959 if (!ifproj->is_Proj()) continue;
1960 Node* iff = ifproj->in(0);
1961 if (!iff->is_If()) continue;
1962 Node* bol = iff->in(1);
1963 if (!bol->is_Bool()) continue;
1964 Node* cmp = bol->in(1);
1965 if (cmp == NULL) continue;
1966 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1967 if (cmps[cmpn] == cmp) break;
1968 if (cmpn == NCMPS) continue;
1969 BoolTest::mask btest = bol->as_Bool()->_test._test;
1970 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1971 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1972 // At this point, we know that 'x btest y' is true.
1973 switch (btest) {
1974 case BoolTest::eq:
1975 // They are proven equal, so we can collapse the min/max.
1976 // Either value is the answer. Choose the simpler.
1977 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1978 return yvalue;
1979 return xvalue;
1980 case BoolTest::lt: // x < y
1981 case BoolTest::le: // x <= y
1982 return (want_max ? yvalue : xvalue);
1983 case BoolTest::gt: // x > y
1984 case BoolTest::ge: // x >= y
1985 return (want_max ? xvalue : yvalue);
1986 }
1987 }
1988 }
1990 // We failed to find a dominating test.
1991 // Let's pick a test that might GVN with prior tests.
1992 Node* best_bol = NULL;
1993 BoolTest::mask best_btest = BoolTest::illegal;
1994 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1995 Node* cmp = cmps[cmpn];
1996 if (cmp == NULL) continue;
1997 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1998 Node* bol = cmp->fast_out(j);
1999 if (!bol->is_Bool()) continue;
2000 BoolTest::mask btest = bol->as_Bool()->_test._test;
2001 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2002 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2003 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2004 best_bol = bol->as_Bool();
2005 best_btest = btest;
2006 }
2007 }
2008 }
2010 Node* answer_if_true = NULL;
2011 Node* answer_if_false = NULL;
2012 switch (best_btest) {
2013 default:
2014 if (cmpxy == NULL)
2015 cmpxy = ideal_cmpxy;
2016 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2017 // and fall through:
2018 case BoolTest::lt: // x < y
2019 case BoolTest::le: // x <= y
2020 answer_if_true = (want_max ? yvalue : xvalue);
2021 answer_if_false = (want_max ? xvalue : yvalue);
2022 break;
2023 case BoolTest::gt: // x > y
2024 case BoolTest::ge: // x >= y
2025 answer_if_true = (want_max ? xvalue : yvalue);
2026 answer_if_false = (want_max ? yvalue : xvalue);
2027 break;
2028 }
2030 jint hi, lo;
2031 if (want_max) {
2032 // We can sharpen the minimum.
2033 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2034 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2035 } else {
2036 // We can sharpen the maximum.
2037 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2038 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2039 }
2041 // Use a flow-free graph structure, to avoid creating excess control edges
2042 // which could hinder other optimizations.
2043 // Since Math.min/max is often used with arraycopy, we want
2044 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2045 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2046 answer_if_false, answer_if_true,
2047 TypeInt::make(lo, hi, widen));
2049 return _gvn.transform(cmov);
2051 /*
2052 // This is not as desirable as it may seem, since Min and Max
2053 // nodes do not have a full set of optimizations.
2054 // And they would interfere, anyway, with 'if' optimizations
2055 // and with CMoveI canonical forms.
2056 switch (id) {
2057 case vmIntrinsics::_min:
2058 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2059 case vmIntrinsics::_max:
2060 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2061 default:
2062 ShouldNotReachHere();
2063 }
2064 */
2065 }
2067 inline int
2068 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2069 const TypePtr* base_type = TypePtr::NULL_PTR;
2070 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2071 if (base_type == NULL) {
2072 // Unknown type.
2073 return Type::AnyPtr;
2074 } else if (base_type == TypePtr::NULL_PTR) {
2075 // Since this is a NULL+long form, we have to switch to a rawptr.
2076 base = _gvn.transform( new (C) CastX2PNode(offset) );
2077 offset = MakeConX(0);
2078 return Type::RawPtr;
2079 } else if (base_type->base() == Type::RawPtr) {
2080 return Type::RawPtr;
2081 } else if (base_type->isa_oopptr()) {
2082 // Base is never null => always a heap address.
2083 if (base_type->ptr() == TypePtr::NotNull) {
2084 return Type::OopPtr;
2085 }
2086 // Offset is small => always a heap address.
2087 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2088 if (offset_type != NULL &&
2089 base_type->offset() == 0 && // (should always be?)
2090 offset_type->_lo >= 0 &&
2091 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2092 return Type::OopPtr;
2093 }
2094 // Otherwise, it might either be oop+off or NULL+addr.
2095 return Type::AnyPtr;
2096 } else {
2097 // No information:
2098 return Type::AnyPtr;
2099 }
2100 }
2102 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2103 int kind = classify_unsafe_addr(base, offset);
2104 if (kind == Type::RawPtr) {
2105 return basic_plus_adr(top(), base, offset);
2106 } else {
2107 return basic_plus_adr(base, offset);
2108 }
2109 }
2111 //--------------------------inline_number_methods-----------------------------
2112 // inline int Integer.numberOfLeadingZeros(int)
2113 // inline int Long.numberOfLeadingZeros(long)
2114 //
2115 // inline int Integer.numberOfTrailingZeros(int)
2116 // inline int Long.numberOfTrailingZeros(long)
2117 //
2118 // inline int Integer.bitCount(int)
2119 // inline int Long.bitCount(long)
2120 //
2121 // inline char Character.reverseBytes(char)
2122 // inline short Short.reverseBytes(short)
2123 // inline int Integer.reverseBytes(int)
2124 // inline long Long.reverseBytes(long)
2125 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2126 Node* arg = argument(0);
2127 Node* n;
2128 switch (id) {
2129 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2130 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2131 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2132 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2133 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2134 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2135 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2136 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2137 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2138 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2139 default: fatal_unexpected_iid(id); break;
2140 }
2141 set_result(_gvn.transform(n));
2142 return true;
2143 }
2145 //----------------------------inline_unsafe_access----------------------------
2147 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2149 // Helper that guards and inserts a pre-barrier.
2150 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2151 Node* pre_val, bool need_mem_bar) {
2152 // We could be accessing the referent field of a reference object. If so, when G1
2153 // is enabled, we need to log the value in the referent field in an SATB buffer.
2154 // This routine performs some compile time filters and generates suitable
2155 // runtime filters that guard the pre-barrier code.
2156 // Also add memory barrier for non volatile load from the referent field
2157 // to prevent commoning of loads across safepoint.
2158 if (!UseG1GC && !need_mem_bar)
2159 return;
2161 // Some compile time checks.
2163 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2164 const TypeX* otype = offset->find_intptr_t_type();
2165 if (otype != NULL && otype->is_con() &&
2166 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2167 // Constant offset but not the reference_offset so just return
2168 return;
2169 }
2171 // We only need to generate the runtime guards for instances.
2172 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2173 if (btype != NULL) {
2174 if (btype->isa_aryptr()) {
2175 // Array type so nothing to do
2176 return;
2177 }
2179 const TypeInstPtr* itype = btype->isa_instptr();
2180 if (itype != NULL) {
2181 // Can the klass of base_oop be statically determined to be
2182 // _not_ a sub-class of Reference and _not_ Object?
2183 ciKlass* klass = itype->klass();
2184 if ( klass->is_loaded() &&
2185 !klass->is_subtype_of(env()->Reference_klass()) &&
2186 !env()->Object_klass()->is_subtype_of(klass)) {
2187 return;
2188 }
2189 }
2190 }
2192 // The compile time filters did not reject base_oop/offset so
2193 // we need to generate the following runtime filters
2194 //
2195 // if (offset == java_lang_ref_Reference::_reference_offset) {
2196 // if (instance_of(base, java.lang.ref.Reference)) {
2197 // pre_barrier(_, pre_val, ...);
2198 // }
2199 // }
2201 float likely = PROB_LIKELY( 0.999);
2202 float unlikely = PROB_UNLIKELY(0.999);
2204 IdealKit ideal(this);
2205 #define __ ideal.
2207 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2209 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2210 // Update graphKit memory and control from IdealKit.
2211 sync_kit(ideal);
2213 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2214 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2216 // Update IdealKit memory and control from graphKit.
2217 __ sync_kit(this);
2219 Node* one = __ ConI(1);
2220 // is_instof == 0 if base_oop == NULL
2221 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2223 // Update graphKit from IdeakKit.
2224 sync_kit(ideal);
2226 // Use the pre-barrier to record the value in the referent field
2227 pre_barrier(false /* do_load */,
2228 __ ctrl(),
2229 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2230 pre_val /* pre_val */,
2231 T_OBJECT);
2232 if (need_mem_bar) {
2233 // Add memory barrier to prevent commoning reads from this field
2234 // across safepoint since GC can change its value.
2235 insert_mem_bar(Op_MemBarCPUOrder);
2236 }
2237 // Update IdealKit from graphKit.
2238 __ sync_kit(this);
2240 } __ end_if(); // _ref_type != ref_none
2241 } __ end_if(); // offset == referent_offset
2243 // Final sync IdealKit and GraphKit.
2244 final_sync(ideal);
2245 #undef __
2246 }
2249 // Interpret Unsafe.fieldOffset cookies correctly:
2250 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2252 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2253 // Attempt to infer a sharper value type from the offset and base type.
2254 ciKlass* sharpened_klass = NULL;
2256 // See if it is an instance field, with an object type.
2257 if (alias_type->field() != NULL) {
2258 assert(!is_native_ptr, "native pointer op cannot use a java address");
2259 if (alias_type->field()->type()->is_klass()) {
2260 sharpened_klass = alias_type->field()->type()->as_klass();
2261 }
2262 }
2264 // See if it is a narrow oop array.
2265 if (adr_type->isa_aryptr()) {
2266 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2267 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2268 if (elem_type != NULL) {
2269 sharpened_klass = elem_type->klass();
2270 }
2271 }
2272 }
2274 // The sharpened class might be unloaded if there is no class loader
2275 // contraint in place.
2276 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2277 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2279 #ifndef PRODUCT
2280 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2281 tty->print(" from base type: "); adr_type->dump();
2282 tty->print(" sharpened value: "); tjp->dump();
2283 }
2284 #endif
2285 // Sharpen the value type.
2286 return tjp;
2287 }
2288 return NULL;
2289 }
2291 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2292 if (callee()->is_static()) return false; // caller must have the capability!
2294 #ifndef PRODUCT
2295 {
2296 ResourceMark rm;
2297 // Check the signatures.
2298 ciSignature* sig = callee()->signature();
2299 #ifdef ASSERT
2300 if (!is_store) {
2301 // Object getObject(Object base, int/long offset), etc.
2302 BasicType rtype = sig->return_type()->basic_type();
2303 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2304 rtype = T_ADDRESS; // it is really a C void*
2305 assert(rtype == type, "getter must return the expected value");
2306 if (!is_native_ptr) {
2307 assert(sig->count() == 2, "oop getter has 2 arguments");
2308 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2309 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2310 } else {
2311 assert(sig->count() == 1, "native getter has 1 argument");
2312 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2313 }
2314 } else {
2315 // void putObject(Object base, int/long offset, Object x), etc.
2316 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2317 if (!is_native_ptr) {
2318 assert(sig->count() == 3, "oop putter has 3 arguments");
2319 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2320 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2321 } else {
2322 assert(sig->count() == 2, "native putter has 2 arguments");
2323 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2324 }
2325 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2326 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2327 vtype = T_ADDRESS; // it is really a C void*
2328 assert(vtype == type, "putter must accept the expected value");
2329 }
2330 #endif // ASSERT
2331 }
2332 #endif //PRODUCT
2334 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2336 Node* receiver = argument(0); // type: oop
2338 // Build address expression. See the code in inline_unsafe_prefetch.
2339 Node* adr;
2340 Node* heap_base_oop = top();
2341 Node* offset = top();
2342 Node* val;
2344 if (!is_native_ptr) {
2345 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2346 Node* base = argument(1); // type: oop
2347 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2348 offset = argument(2); // type: long
2349 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2350 // to be plain byte offsets, which are also the same as those accepted
2351 // by oopDesc::field_base.
2352 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2353 "fieldOffset must be byte-scaled");
2354 // 32-bit machines ignore the high half!
2355 offset = ConvL2X(offset);
2356 adr = make_unsafe_address(base, offset);
2357 heap_base_oop = base;
2358 val = is_store ? argument(4) : NULL;
2359 } else {
2360 Node* ptr = argument(1); // type: long
2361 ptr = ConvL2X(ptr); // adjust Java long to machine word
2362 adr = make_unsafe_address(NULL, ptr);
2363 val = is_store ? argument(3) : NULL;
2364 }
2366 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2368 // First guess at the value type.
2369 const Type *value_type = Type::get_const_basic_type(type);
2371 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2372 // there was not enough information to nail it down.
2373 Compile::AliasType* alias_type = C->alias_type(adr_type);
2374 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2376 // We will need memory barriers unless we can determine a unique
2377 // alias category for this reference. (Note: If for some reason
2378 // the barriers get omitted and the unsafe reference begins to "pollute"
2379 // the alias analysis of the rest of the graph, either Compile::can_alias
2380 // or Compile::must_alias will throw a diagnostic assert.)
2381 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2383 // If we are reading the value of the referent field of a Reference
2384 // object (either by using Unsafe directly or through reflection)
2385 // then, if G1 is enabled, we need to record the referent in an
2386 // SATB log buffer using the pre-barrier mechanism.
2387 // Also we need to add memory barrier to prevent commoning reads
2388 // from this field across safepoint since GC can change its value.
2389 bool need_read_barrier = !is_native_ptr && !is_store &&
2390 offset != top() && heap_base_oop != top();
2392 if (!is_store && type == T_OBJECT) {
2393 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2394 if (tjp != NULL) {
2395 value_type = tjp;
2396 }
2397 }
2399 receiver = null_check(receiver);
2400 if (stopped()) {
2401 return true;
2402 }
2403 // Heap pointers get a null-check from the interpreter,
2404 // as a courtesy. However, this is not guaranteed by Unsafe,
2405 // and it is not possible to fully distinguish unintended nulls
2406 // from intended ones in this API.
2408 if (is_volatile) {
2409 // We need to emit leading and trailing CPU membars (see below) in
2410 // addition to memory membars when is_volatile. This is a little
2411 // too strong, but avoids the need to insert per-alias-type
2412 // volatile membars (for stores; compare Parse::do_put_xxx), which
2413 // we cannot do effectively here because we probably only have a
2414 // rough approximation of type.
2415 need_mem_bar = true;
2416 // For Stores, place a memory ordering barrier now.
2417 if (is_store)
2418 insert_mem_bar(Op_MemBarRelease);
2419 }
2421 // Memory barrier to prevent normal and 'unsafe' accesses from
2422 // bypassing each other. Happens after null checks, so the
2423 // exception paths do not take memory state from the memory barrier,
2424 // so there's no problems making a strong assert about mixing users
2425 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2426 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2427 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2429 if (!is_store) {
2430 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2431 // load value
2432 switch (type) {
2433 case T_BOOLEAN:
2434 case T_CHAR:
2435 case T_BYTE:
2436 case T_SHORT:
2437 case T_INT:
2438 case T_LONG:
2439 case T_FLOAT:
2440 case T_DOUBLE:
2441 break;
2442 case T_OBJECT:
2443 if (need_read_barrier) {
2444 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2445 }
2446 break;
2447 case T_ADDRESS:
2448 // Cast to an int type.
2449 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2450 p = ConvX2L(p);
2451 break;
2452 default:
2453 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2454 break;
2455 }
2456 // The load node has the control of the preceding MemBarCPUOrder. All
2457 // following nodes will have the control of the MemBarCPUOrder inserted at
2458 // the end of this method. So, pushing the load onto the stack at a later
2459 // point is fine.
2460 set_result(p);
2461 } else {
2462 // place effect of store into memory
2463 switch (type) {
2464 case T_DOUBLE:
2465 val = dstore_rounding(val);
2466 break;
2467 case T_ADDRESS:
2468 // Repackage the long as a pointer.
2469 val = ConvL2X(val);
2470 val = _gvn.transform( new (C) CastX2PNode(val) );
2471 break;
2472 }
2474 if (type != T_OBJECT ) {
2475 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2476 } else {
2477 // Possibly an oop being stored to Java heap or native memory
2478 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2479 // oop to Java heap.
2480 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2481 } else {
2482 // We can't tell at compile time if we are storing in the Java heap or outside
2483 // of it. So we need to emit code to conditionally do the proper type of
2484 // store.
2486 IdealKit ideal(this);
2487 #define __ ideal.
2488 // QQQ who knows what probability is here??
2489 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2490 // Sync IdealKit and graphKit.
2491 sync_kit(ideal);
2492 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2493 // Update IdealKit memory.
2494 __ sync_kit(this);
2495 } __ else_(); {
2496 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2497 } __ end_if();
2498 // Final sync IdealKit and GraphKit.
2499 final_sync(ideal);
2500 #undef __
2501 }
2502 }
2503 }
2505 if (is_volatile) {
2506 if (!is_store)
2507 insert_mem_bar(Op_MemBarAcquire);
2508 else
2509 insert_mem_bar(Op_MemBarVolatile);
2510 }
2512 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2514 return true;
2515 }
2517 //----------------------------inline_unsafe_prefetch----------------------------
2519 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2520 #ifndef PRODUCT
2521 {
2522 ResourceMark rm;
2523 // Check the signatures.
2524 ciSignature* sig = callee()->signature();
2525 #ifdef ASSERT
2526 // Object getObject(Object base, int/long offset), etc.
2527 BasicType rtype = sig->return_type()->basic_type();
2528 if (!is_native_ptr) {
2529 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2530 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2531 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2532 } else {
2533 assert(sig->count() == 1, "native prefetch has 1 argument");
2534 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2535 }
2536 #endif // ASSERT
2537 }
2538 #endif // !PRODUCT
2540 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2542 const int idx = is_static ? 0 : 1;
2543 if (!is_static) {
2544 null_check_receiver();
2545 if (stopped()) {
2546 return true;
2547 }
2548 }
2550 // Build address expression. See the code in inline_unsafe_access.
2551 Node *adr;
2552 if (!is_native_ptr) {
2553 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2554 Node* base = argument(idx + 0); // type: oop
2555 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2556 Node* offset = argument(idx + 1); // type: long
2557 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2558 // to be plain byte offsets, which are also the same as those accepted
2559 // by oopDesc::field_base.
2560 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2561 "fieldOffset must be byte-scaled");
2562 // 32-bit machines ignore the high half!
2563 offset = ConvL2X(offset);
2564 adr = make_unsafe_address(base, offset);
2565 } else {
2566 Node* ptr = argument(idx + 0); // type: long
2567 ptr = ConvL2X(ptr); // adjust Java long to machine word
2568 adr = make_unsafe_address(NULL, ptr);
2569 }
2571 // Generate the read or write prefetch
2572 Node *prefetch;
2573 if (is_store) {
2574 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2575 } else {
2576 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2577 }
2578 prefetch->init_req(0, control());
2579 set_i_o(_gvn.transform(prefetch));
2581 return true;
2582 }
2584 //----------------------------inline_unsafe_load_store----------------------------
2585 // This method serves a couple of different customers (depending on LoadStoreKind):
2586 //
2587 // LS_cmpxchg:
2588 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2589 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2590 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2591 //
2592 // LS_xadd:
2593 // public int getAndAddInt( Object o, long offset, int delta)
2594 // public long getAndAddLong(Object o, long offset, long delta)
2595 //
2596 // LS_xchg:
2597 // int getAndSet(Object o, long offset, int newValue)
2598 // long getAndSet(Object o, long offset, long newValue)
2599 // Object getAndSet(Object o, long offset, Object newValue)
2600 //
2601 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2602 // This basic scheme here is the same as inline_unsafe_access, but
2603 // differs in enough details that combining them would make the code
2604 // overly confusing. (This is a true fact! I originally combined
2605 // them, but even I was confused by it!) As much code/comments as
2606 // possible are retained from inline_unsafe_access though to make
2607 // the correspondences clearer. - dl
2609 if (callee()->is_static()) return false; // caller must have the capability!
2611 #ifndef PRODUCT
2612 BasicType rtype;
2613 {
2614 ResourceMark rm;
2615 // Check the signatures.
2616 ciSignature* sig = callee()->signature();
2617 rtype = sig->return_type()->basic_type();
2618 if (kind == LS_xadd || kind == LS_xchg) {
2619 // Check the signatures.
2620 #ifdef ASSERT
2621 assert(rtype == type, "get and set must return the expected type");
2622 assert(sig->count() == 3, "get and set has 3 arguments");
2623 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2624 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2625 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2626 #endif // ASSERT
2627 } else if (kind == LS_cmpxchg) {
2628 // Check the signatures.
2629 #ifdef ASSERT
2630 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2631 assert(sig->count() == 4, "CAS has 4 arguments");
2632 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2633 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2634 #endif // ASSERT
2635 } else {
2636 ShouldNotReachHere();
2637 }
2638 }
2639 #endif //PRODUCT
2641 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2643 // Get arguments:
2644 Node* receiver = NULL;
2645 Node* base = NULL;
2646 Node* offset = NULL;
2647 Node* oldval = NULL;
2648 Node* newval = NULL;
2649 if (kind == LS_cmpxchg) {
2650 const bool two_slot_type = type2size[type] == 2;
2651 receiver = argument(0); // type: oop
2652 base = argument(1); // type: oop
2653 offset = argument(2); // type: long
2654 oldval = argument(4); // type: oop, int, or long
2655 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2656 } else if (kind == LS_xadd || kind == LS_xchg){
2657 receiver = argument(0); // type: oop
2658 base = argument(1); // type: oop
2659 offset = argument(2); // type: long
2660 oldval = NULL;
2661 newval = argument(4); // type: oop, int, or long
2662 }
2664 // Null check receiver.
2665 receiver = null_check(receiver);
2666 if (stopped()) {
2667 return true;
2668 }
2670 // Build field offset expression.
2671 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2672 // to be plain byte offsets, which are also the same as those accepted
2673 // by oopDesc::field_base.
2674 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2675 // 32-bit machines ignore the high half of long offsets
2676 offset = ConvL2X(offset);
2677 Node* adr = make_unsafe_address(base, offset);
2678 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2680 // For CAS, unlike inline_unsafe_access, there seems no point in
2681 // trying to refine types. Just use the coarse types here.
2682 const Type *value_type = Type::get_const_basic_type(type);
2683 Compile::AliasType* alias_type = C->alias_type(adr_type);
2684 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2686 if (kind == LS_xchg && type == T_OBJECT) {
2687 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2688 if (tjp != NULL) {
2689 value_type = tjp;
2690 }
2691 }
2693 int alias_idx = C->get_alias_index(adr_type);
2695 // Memory-model-wise, a LoadStore acts like a little synchronized
2696 // block, so needs barriers on each side. These don't translate
2697 // into actual barriers on most machines, but we still need rest of
2698 // compiler to respect ordering.
2700 insert_mem_bar(Op_MemBarRelease);
2701 insert_mem_bar(Op_MemBarCPUOrder);
2703 // 4984716: MemBars must be inserted before this
2704 // memory node in order to avoid a false
2705 // dependency which will confuse the scheduler.
2706 Node *mem = memory(alias_idx);
2708 // For now, we handle only those cases that actually exist: ints,
2709 // longs, and Object. Adding others should be straightforward.
2710 Node* load_store;
2711 switch(type) {
2712 case T_INT:
2713 if (kind == LS_xadd) {
2714 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2715 } else if (kind == LS_xchg) {
2716 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2717 } else if (kind == LS_cmpxchg) {
2718 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2719 } else {
2720 ShouldNotReachHere();
2721 }
2722 break;
2723 case T_LONG:
2724 if (kind == LS_xadd) {
2725 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2726 } else if (kind == LS_xchg) {
2727 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2728 } else if (kind == LS_cmpxchg) {
2729 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2730 } else {
2731 ShouldNotReachHere();
2732 }
2733 break;
2734 case T_OBJECT:
2735 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2736 // could be delayed during Parse (for example, in adjust_map_after_if()).
2737 // Execute transformation here to avoid barrier generation in such case.
2738 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2739 newval = _gvn.makecon(TypePtr::NULL_PTR);
2741 // Reference stores need a store barrier.
2742 pre_barrier(true /* do_load*/,
2743 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2744 NULL /* pre_val*/,
2745 T_OBJECT);
2746 #ifdef _LP64
2747 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2748 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2749 if (kind == LS_xchg) {
2750 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2751 newval_enc, adr_type, value_type->make_narrowoop()));
2752 } else {
2753 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2754 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2755 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2756 newval_enc, oldval_enc));
2757 }
2758 } else
2759 #endif
2760 {
2761 if (kind == LS_xchg) {
2762 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2763 } else {
2764 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2765 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2766 }
2767 }
2768 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2769 break;
2770 default:
2771 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2772 break;
2773 }
2775 // SCMemProjNodes represent the memory state of a LoadStore. Their
2776 // main role is to prevent LoadStore nodes from being optimized away
2777 // when their results aren't used.
2778 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
2779 set_memory(proj, alias_idx);
2781 // Add the trailing membar surrounding the access
2782 insert_mem_bar(Op_MemBarCPUOrder);
2783 insert_mem_bar(Op_MemBarAcquire);
2785 #ifdef _LP64
2786 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
2787 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->bottom_type()->make_ptr()));
2788 }
2789 #endif
2791 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2792 set_result(load_store);
2793 return true;
2794 }
2796 //----------------------------inline_unsafe_ordered_store----------------------
2797 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2798 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2799 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2800 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2801 // This is another variant of inline_unsafe_access, differing in
2802 // that it always issues store-store ("release") barrier and ensures
2803 // store-atomicity (which only matters for "long").
2805 if (callee()->is_static()) return false; // caller must have the capability!
2807 #ifndef PRODUCT
2808 {
2809 ResourceMark rm;
2810 // Check the signatures.
2811 ciSignature* sig = callee()->signature();
2812 #ifdef ASSERT
2813 BasicType rtype = sig->return_type()->basic_type();
2814 assert(rtype == T_VOID, "must return void");
2815 assert(sig->count() == 3, "has 3 arguments");
2816 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2817 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2818 #endif // ASSERT
2819 }
2820 #endif //PRODUCT
2822 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2824 // Get arguments:
2825 Node* receiver = argument(0); // type: oop
2826 Node* base = argument(1); // type: oop
2827 Node* offset = argument(2); // type: long
2828 Node* val = argument(4); // type: oop, int, or long
2830 // Null check receiver.
2831 receiver = null_check(receiver);
2832 if (stopped()) {
2833 return true;
2834 }
2836 // Build field offset expression.
2837 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2838 // 32-bit machines ignore the high half of long offsets
2839 offset = ConvL2X(offset);
2840 Node* adr = make_unsafe_address(base, offset);
2841 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2842 const Type *value_type = Type::get_const_basic_type(type);
2843 Compile::AliasType* alias_type = C->alias_type(adr_type);
2845 insert_mem_bar(Op_MemBarRelease);
2846 insert_mem_bar(Op_MemBarCPUOrder);
2847 // Ensure that the store is atomic for longs:
2848 const bool require_atomic_access = true;
2849 Node* store;
2850 if (type == T_OBJECT) // reference stores need a store barrier.
2851 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2852 else {
2853 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2854 }
2855 insert_mem_bar(Op_MemBarCPUOrder);
2856 return true;
2857 }
2859 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2860 // Regardless of form, don't allow previous ld/st to move down,
2861 // then issue acquire, release, or volatile mem_bar.
2862 insert_mem_bar(Op_MemBarCPUOrder);
2863 switch(id) {
2864 case vmIntrinsics::_loadFence:
2865 insert_mem_bar(Op_MemBarAcquire);
2866 return true;
2867 case vmIntrinsics::_storeFence:
2868 insert_mem_bar(Op_MemBarRelease);
2869 return true;
2870 case vmIntrinsics::_fullFence:
2871 insert_mem_bar(Op_MemBarVolatile);
2872 return true;
2873 default:
2874 fatal_unexpected_iid(id);
2875 return false;
2876 }
2877 }
2879 //----------------------------inline_unsafe_allocate---------------------------
2880 // public native Object sun.mics.Unsafe.allocateInstance(Class<?> cls);
2881 bool LibraryCallKit::inline_unsafe_allocate() {
2882 if (callee()->is_static()) return false; // caller must have the capability!
2884 null_check_receiver(); // null-check, then ignore
2885 Node* cls = null_check(argument(1));
2886 if (stopped()) return true;
2888 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2889 kls = null_check(kls);
2890 if (stopped()) return true; // argument was like int.class
2892 // Note: The argument might still be an illegal value like
2893 // Serializable.class or Object[].class. The runtime will handle it.
2894 // But we must make an explicit check for initialization.
2895 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2896 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2897 // can generate code to load it as unsigned byte.
2898 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2899 Node* bits = intcon(InstanceKlass::fully_initialized);
2900 Node* test = _gvn.transform(new (C) SubINode(inst, bits));
2901 // The 'test' is non-zero if we need to take a slow path.
2903 Node* obj = new_instance(kls, test);
2904 set_result(obj);
2905 return true;
2906 }
2908 #ifdef TRACE_HAVE_INTRINSICS
2909 /*
2910 * oop -> myklass
2911 * myklass->trace_id |= USED
2912 * return myklass->trace_id & ~0x3
2913 */
2914 bool LibraryCallKit::inline_native_classID() {
2915 null_check_receiver(); // null-check, then ignore
2916 Node* cls = null_check(argument(1), T_OBJECT);
2917 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2918 kls = null_check(kls, T_OBJECT);
2919 ByteSize offset = TRACE_ID_OFFSET;
2920 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2921 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2922 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2923 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
2924 Node* clsused = longcon(0x01l); // set the class bit
2925 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
2927 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2928 store_to_memory(control(), insp, orl, T_LONG, adr_type);
2929 set_result(andl);
2930 return true;
2931 }
2933 bool LibraryCallKit::inline_native_threadID() {
2934 Node* tls_ptr = NULL;
2935 Node* cur_thr = generate_current_thread(tls_ptr);
2936 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2937 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2938 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2940 Node* threadid = NULL;
2941 size_t thread_id_size = OSThread::thread_id_size();
2942 if (thread_id_size == (size_t) BytesPerLong) {
2943 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
2944 } else if (thread_id_size == (size_t) BytesPerInt) {
2945 threadid = make_load(control(), p, TypeInt::INT, T_INT);
2946 } else {
2947 ShouldNotReachHere();
2948 }
2949 set_result(threadid);
2950 return true;
2951 }
2952 #endif
2954 //------------------------inline_native_time_funcs--------------
2955 // inline code for System.currentTimeMillis() and System.nanoTime()
2956 // these have the same type and signature
2957 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2958 const TypeFunc* tf = OptoRuntime::void_long_Type();
2959 const TypePtr* no_memory_effects = NULL;
2960 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2961 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
2962 #ifdef ASSERT
2963 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
2964 assert(value_top == top(), "second value must be top");
2965 #endif
2966 set_result(value);
2967 return true;
2968 }
2970 //------------------------inline_native_currentThread------------------
2971 bool LibraryCallKit::inline_native_currentThread() {
2972 Node* junk = NULL;
2973 set_result(generate_current_thread(junk));
2974 return true;
2975 }
2977 //------------------------inline_native_isInterrupted------------------
2978 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2979 bool LibraryCallKit::inline_native_isInterrupted() {
2980 // Add a fast path to t.isInterrupted(clear_int):
2981 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2982 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2983 // So, in the common case that the interrupt bit is false,
2984 // we avoid making a call into the VM. Even if the interrupt bit
2985 // is true, if the clear_int argument is false, we avoid the VM call.
2986 // However, if the receiver is not currentThread, we must call the VM,
2987 // because there must be some locking done around the operation.
2989 // We only go to the fast case code if we pass two guards.
2990 // Paths which do not pass are accumulated in the slow_region.
2992 enum {
2993 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2994 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2995 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2996 PATH_LIMIT
2997 };
2999 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3000 // out of the function.
3001 insert_mem_bar(Op_MemBarCPUOrder);
3003 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3004 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3006 RegionNode* slow_region = new (C) RegionNode(1);
3007 record_for_igvn(slow_region);
3009 // (a) Receiving thread must be the current thread.
3010 Node* rec_thr = argument(0);
3011 Node* tls_ptr = NULL;
3012 Node* cur_thr = generate_current_thread(tls_ptr);
3013 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
3014 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
3016 generate_slow_guard(bol_thr, slow_region);
3018 // (b) Interrupt bit on TLS must be false.
3019 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3020 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3021 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3023 // Set the control input on the field _interrupted read to prevent it floating up.
3024 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3025 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
3026 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
3028 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3030 // First fast path: if (!TLS._interrupted) return false;
3031 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
3032 result_rgn->init_req(no_int_result_path, false_bit);
3033 result_val->init_req(no_int_result_path, intcon(0));
3035 // drop through to next case
3036 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
3038 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3039 Node* clr_arg = argument(1);
3040 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
3041 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
3042 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3044 // Second fast path: ... else if (!clear_int) return true;
3045 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3046 result_rgn->init_req(no_clear_result_path, false_arg);
3047 result_val->init_req(no_clear_result_path, intcon(1));
3049 // drop through to next case
3050 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3052 // (d) Otherwise, go to the slow path.
3053 slow_region->add_req(control());
3054 set_control( _gvn.transform(slow_region) );
3056 if (stopped()) {
3057 // There is no slow path.
3058 result_rgn->init_req(slow_result_path, top());
3059 result_val->init_req(slow_result_path, top());
3060 } else {
3061 // non-virtual because it is a private non-static
3062 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3064 Node* slow_val = set_results_for_java_call(slow_call);
3065 // this->control() comes from set_results_for_java_call
3067 Node* fast_io = slow_call->in(TypeFunc::I_O);
3068 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3070 // These two phis are pre-filled with copies of of the fast IO and Memory
3071 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3072 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3074 result_rgn->init_req(slow_result_path, control());
3075 result_io ->init_req(slow_result_path, i_o());
3076 result_mem->init_req(slow_result_path, reset_memory());
3077 result_val->init_req(slow_result_path, slow_val);
3079 set_all_memory(_gvn.transform(result_mem));
3080 set_i_o( _gvn.transform(result_io));
3081 }
3083 C->set_has_split_ifs(true); // Has chance for split-if optimization
3084 set_result(result_rgn, result_val);
3085 return true;
3086 }
3088 //---------------------------load_mirror_from_klass----------------------------
3089 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3090 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3091 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3092 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3093 }
3095 //-----------------------load_klass_from_mirror_common-------------------------
3096 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3097 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3098 // and branch to the given path on the region.
3099 // If never_see_null, take an uncommon trap on null, so we can optimistically
3100 // compile for the non-null case.
3101 // If the region is NULL, force never_see_null = true.
3102 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3103 bool never_see_null,
3104 RegionNode* region,
3105 int null_path,
3106 int offset) {
3107 if (region == NULL) never_see_null = true;
3108 Node* p = basic_plus_adr(mirror, offset);
3109 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3110 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3111 Node* null_ctl = top();
3112 kls = null_check_oop(kls, &null_ctl, never_see_null);
3113 if (region != NULL) {
3114 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3115 region->init_req(null_path, null_ctl);
3116 } else {
3117 assert(null_ctl == top(), "no loose ends");
3118 }
3119 return kls;
3120 }
3122 //--------------------(inline_native_Class_query helpers)---------------------
3123 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3124 // Fall through if (mods & mask) == bits, take the guard otherwise.
3125 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3126 // Branch around if the given klass has the given modifier bit set.
3127 // Like generate_guard, adds a new path onto the region.
3128 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3129 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3130 Node* mask = intcon(modifier_mask);
3131 Node* bits = intcon(modifier_bits);
3132 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3133 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3134 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3135 return generate_fair_guard(bol, region);
3136 }
3137 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3138 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3139 }
3141 //-------------------------inline_native_Class_query-------------------
3142 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3143 const Type* return_type = TypeInt::BOOL;
3144 Node* prim_return_value = top(); // what happens if it's a primitive class?
3145 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3146 bool expect_prim = false; // most of these guys expect to work on refs
3148 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3150 Node* mirror = argument(0);
3151 Node* obj = top();
3153 switch (id) {
3154 case vmIntrinsics::_isInstance:
3155 // nothing is an instance of a primitive type
3156 prim_return_value = intcon(0);
3157 obj = argument(1);
3158 break;
3159 case vmIntrinsics::_getModifiers:
3160 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3161 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3162 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3163 break;
3164 case vmIntrinsics::_isInterface:
3165 prim_return_value = intcon(0);
3166 break;
3167 case vmIntrinsics::_isArray:
3168 prim_return_value = intcon(0);
3169 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3170 break;
3171 case vmIntrinsics::_isPrimitive:
3172 prim_return_value = intcon(1);
3173 expect_prim = true; // obviously
3174 break;
3175 case vmIntrinsics::_getSuperclass:
3176 prim_return_value = null();
3177 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3178 break;
3179 case vmIntrinsics::_getComponentType:
3180 prim_return_value = null();
3181 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3182 break;
3183 case vmIntrinsics::_getClassAccessFlags:
3184 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3185 return_type = TypeInt::INT; // not bool! 6297094
3186 break;
3187 default:
3188 fatal_unexpected_iid(id);
3189 break;
3190 }
3192 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3193 if (mirror_con == NULL) return false; // cannot happen?
3195 #ifndef PRODUCT
3196 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3197 ciType* k = mirror_con->java_mirror_type();
3198 if (k) {
3199 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3200 k->print_name();
3201 tty->cr();
3202 }
3203 }
3204 #endif
3206 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3207 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3208 record_for_igvn(region);
3209 PhiNode* phi = new (C) PhiNode(region, return_type);
3211 // The mirror will never be null of Reflection.getClassAccessFlags, however
3212 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3213 // if it is. See bug 4774291.
3215 // For Reflection.getClassAccessFlags(), the null check occurs in
3216 // the wrong place; see inline_unsafe_access(), above, for a similar
3217 // situation.
3218 mirror = null_check(mirror);
3219 // If mirror or obj is dead, only null-path is taken.
3220 if (stopped()) return true;
3222 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3224 // Now load the mirror's klass metaobject, and null-check it.
3225 // Side-effects region with the control path if the klass is null.
3226 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3227 // If kls is null, we have a primitive mirror.
3228 phi->init_req(_prim_path, prim_return_value);
3229 if (stopped()) { set_result(region, phi); return true; }
3231 Node* p; // handy temp
3232 Node* null_ctl;
3234 // Now that we have the non-null klass, we can perform the real query.
3235 // For constant classes, the query will constant-fold in LoadNode::Value.
3236 Node* query_value = top();
3237 switch (id) {
3238 case vmIntrinsics::_isInstance:
3239 // nothing is an instance of a primitive type
3240 query_value = gen_instanceof(obj, kls);
3241 break;
3243 case vmIntrinsics::_getModifiers:
3244 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3245 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3246 break;
3248 case vmIntrinsics::_isInterface:
3249 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3250 if (generate_interface_guard(kls, region) != NULL)
3251 // A guard was added. If the guard is taken, it was an interface.
3252 phi->add_req(intcon(1));
3253 // If we fall through, it's a plain class.
3254 query_value = intcon(0);
3255 break;
3257 case vmIntrinsics::_isArray:
3258 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3259 if (generate_array_guard(kls, region) != NULL)
3260 // A guard was added. If the guard is taken, it was an array.
3261 phi->add_req(intcon(1));
3262 // If we fall through, it's a plain class.
3263 query_value = intcon(0);
3264 break;
3266 case vmIntrinsics::_isPrimitive:
3267 query_value = intcon(0); // "normal" path produces false
3268 break;
3270 case vmIntrinsics::_getSuperclass:
3271 // The rules here are somewhat unfortunate, but we can still do better
3272 // with random logic than with a JNI call.
3273 // Interfaces store null or Object as _super, but must report null.
3274 // Arrays store an intermediate super as _super, but must report Object.
3275 // Other types can report the actual _super.
3276 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3277 if (generate_interface_guard(kls, region) != NULL)
3278 // A guard was added. If the guard is taken, it was an interface.
3279 phi->add_req(null());
3280 if (generate_array_guard(kls, region) != NULL)
3281 // A guard was added. If the guard is taken, it was an array.
3282 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3283 // If we fall through, it's a plain class. Get its _super.
3284 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3285 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3286 null_ctl = top();
3287 kls = null_check_oop(kls, &null_ctl);
3288 if (null_ctl != top()) {
3289 // If the guard is taken, Object.superClass is null (both klass and mirror).
3290 region->add_req(null_ctl);
3291 phi ->add_req(null());
3292 }
3293 if (!stopped()) {
3294 query_value = load_mirror_from_klass(kls);
3295 }
3296 break;
3298 case vmIntrinsics::_getComponentType:
3299 if (generate_array_guard(kls, region) != NULL) {
3300 // Be sure to pin the oop load to the guard edge just created:
3301 Node* is_array_ctrl = region->in(region->req()-1);
3302 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3303 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3304 phi->add_req(cmo);
3305 }
3306 query_value = null(); // non-array case is null
3307 break;
3309 case vmIntrinsics::_getClassAccessFlags:
3310 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3311 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3312 break;
3314 default:
3315 fatal_unexpected_iid(id);
3316 break;
3317 }
3319 // Fall-through is the normal case of a query to a real class.
3320 phi->init_req(1, query_value);
3321 region->init_req(1, control());
3323 C->set_has_split_ifs(true); // Has chance for split-if optimization
3324 set_result(region, phi);
3325 return true;
3326 }
3328 //--------------------------inline_native_subtype_check------------------------
3329 // This intrinsic takes the JNI calls out of the heart of
3330 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3331 bool LibraryCallKit::inline_native_subtype_check() {
3332 // Pull both arguments off the stack.
3333 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3334 args[0] = argument(0);
3335 args[1] = argument(1);
3336 Node* klasses[2]; // corresponding Klasses: superk, subk
3337 klasses[0] = klasses[1] = top();
3339 enum {
3340 // A full decision tree on {superc is prim, subc is prim}:
3341 _prim_0_path = 1, // {P,N} => false
3342 // {P,P} & superc!=subc => false
3343 _prim_same_path, // {P,P} & superc==subc => true
3344 _prim_1_path, // {N,P} => false
3345 _ref_subtype_path, // {N,N} & subtype check wins => true
3346 _both_ref_path, // {N,N} & subtype check loses => false
3347 PATH_LIMIT
3348 };
3350 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3351 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3352 record_for_igvn(region);
3354 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3355 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3356 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3358 // First null-check both mirrors and load each mirror's klass metaobject.
3359 int which_arg;
3360 for (which_arg = 0; which_arg <= 1; which_arg++) {
3361 Node* arg = args[which_arg];
3362 arg = null_check(arg);
3363 if (stopped()) break;
3364 args[which_arg] = arg;
3366 Node* p = basic_plus_adr(arg, class_klass_offset);
3367 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3368 klasses[which_arg] = _gvn.transform(kls);
3369 }
3371 // Having loaded both klasses, test each for null.
3372 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3373 for (which_arg = 0; which_arg <= 1; which_arg++) {
3374 Node* kls = klasses[which_arg];
3375 Node* null_ctl = top();
3376 kls = null_check_oop(kls, &null_ctl, never_see_null);
3377 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3378 region->init_req(prim_path, null_ctl);
3379 if (stopped()) break;
3380 klasses[which_arg] = kls;
3381 }
3383 if (!stopped()) {
3384 // now we have two reference types, in klasses[0..1]
3385 Node* subk = klasses[1]; // the argument to isAssignableFrom
3386 Node* superk = klasses[0]; // the receiver
3387 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3388 // now we have a successful reference subtype check
3389 region->set_req(_ref_subtype_path, control());
3390 }
3392 // If both operands are primitive (both klasses null), then
3393 // we must return true when they are identical primitives.
3394 // It is convenient to test this after the first null klass check.
3395 set_control(region->in(_prim_0_path)); // go back to first null check
3396 if (!stopped()) {
3397 // Since superc is primitive, make a guard for the superc==subc case.
3398 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3399 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3400 generate_guard(bol_eq, region, PROB_FAIR);
3401 if (region->req() == PATH_LIMIT+1) {
3402 // A guard was added. If the added guard is taken, superc==subc.
3403 region->swap_edges(PATH_LIMIT, _prim_same_path);
3404 region->del_req(PATH_LIMIT);
3405 }
3406 region->set_req(_prim_0_path, control()); // Not equal after all.
3407 }
3409 // these are the only paths that produce 'true':
3410 phi->set_req(_prim_same_path, intcon(1));
3411 phi->set_req(_ref_subtype_path, intcon(1));
3413 // pull together the cases:
3414 assert(region->req() == PATH_LIMIT, "sane region");
3415 for (uint i = 1; i < region->req(); i++) {
3416 Node* ctl = region->in(i);
3417 if (ctl == NULL || ctl == top()) {
3418 region->set_req(i, top());
3419 phi ->set_req(i, top());
3420 } else if (phi->in(i) == NULL) {
3421 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3422 }
3423 }
3425 set_control(_gvn.transform(region));
3426 set_result(_gvn.transform(phi));
3427 return true;
3428 }
3430 //---------------------generate_array_guard_common------------------------
3431 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3432 bool obj_array, bool not_array) {
3433 // If obj_array/non_array==false/false:
3434 // Branch around if the given klass is in fact an array (either obj or prim).
3435 // If obj_array/non_array==false/true:
3436 // Branch around if the given klass is not an array klass of any kind.
3437 // If obj_array/non_array==true/true:
3438 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3439 // If obj_array/non_array==true/false:
3440 // Branch around if the kls is an oop array (Object[] or subtype)
3441 //
3442 // Like generate_guard, adds a new path onto the region.
3443 jint layout_con = 0;
3444 Node* layout_val = get_layout_helper(kls, layout_con);
3445 if (layout_val == NULL) {
3446 bool query = (obj_array
3447 ? Klass::layout_helper_is_objArray(layout_con)
3448 : Klass::layout_helper_is_array(layout_con));
3449 if (query == not_array) {
3450 return NULL; // never a branch
3451 } else { // always a branch
3452 Node* always_branch = control();
3453 if (region != NULL)
3454 region->add_req(always_branch);
3455 set_control(top());
3456 return always_branch;
3457 }
3458 }
3459 // Now test the correct condition.
3460 jint nval = (obj_array
3461 ? ((jint)Klass::_lh_array_tag_type_value
3462 << Klass::_lh_array_tag_shift)
3463 : Klass::_lh_neutral_value);
3464 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3465 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3466 // invert the test if we are looking for a non-array
3467 if (not_array) btest = BoolTest(btest).negate();
3468 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3469 return generate_fair_guard(bol, region);
3470 }
3473 //-----------------------inline_native_newArray--------------------------
3474 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3475 bool LibraryCallKit::inline_native_newArray() {
3476 Node* mirror = argument(0);
3477 Node* count_val = argument(1);
3479 mirror = null_check(mirror);
3480 // If mirror or obj is dead, only null-path is taken.
3481 if (stopped()) return true;
3483 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3484 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3485 PhiNode* result_val = new(C) PhiNode(result_reg,
3486 TypeInstPtr::NOTNULL);
3487 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3488 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3489 TypePtr::BOTTOM);
3491 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3492 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3493 result_reg, _slow_path);
3494 Node* normal_ctl = control();
3495 Node* no_array_ctl = result_reg->in(_slow_path);
3497 // Generate code for the slow case. We make a call to newArray().
3498 set_control(no_array_ctl);
3499 if (!stopped()) {
3500 // Either the input type is void.class, or else the
3501 // array klass has not yet been cached. Either the
3502 // ensuing call will throw an exception, or else it
3503 // will cache the array klass for next time.
3504 PreserveJVMState pjvms(this);
3505 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3506 Node* slow_result = set_results_for_java_call(slow_call);
3507 // this->control() comes from set_results_for_java_call
3508 result_reg->set_req(_slow_path, control());
3509 result_val->set_req(_slow_path, slow_result);
3510 result_io ->set_req(_slow_path, i_o());
3511 result_mem->set_req(_slow_path, reset_memory());
3512 }
3514 set_control(normal_ctl);
3515 if (!stopped()) {
3516 // Normal case: The array type has been cached in the java.lang.Class.
3517 // The following call works fine even if the array type is polymorphic.
3518 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3519 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3520 result_reg->init_req(_normal_path, control());
3521 result_val->init_req(_normal_path, obj);
3522 result_io ->init_req(_normal_path, i_o());
3523 result_mem->init_req(_normal_path, reset_memory());
3524 }
3526 // Return the combined state.
3527 set_i_o( _gvn.transform(result_io) );
3528 set_all_memory( _gvn.transform(result_mem) );
3530 C->set_has_split_ifs(true); // Has chance for split-if optimization
3531 set_result(result_reg, result_val);
3532 return true;
3533 }
3535 //----------------------inline_native_getLength--------------------------
3536 // public static native int java.lang.reflect.Array.getLength(Object array);
3537 bool LibraryCallKit::inline_native_getLength() {
3538 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3540 Node* array = null_check(argument(0));
3541 // If array is dead, only null-path is taken.
3542 if (stopped()) return true;
3544 // Deoptimize if it is a non-array.
3545 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3547 if (non_array != NULL) {
3548 PreserveJVMState pjvms(this);
3549 set_control(non_array);
3550 uncommon_trap(Deoptimization::Reason_intrinsic,
3551 Deoptimization::Action_maybe_recompile);
3552 }
3554 // If control is dead, only non-array-path is taken.
3555 if (stopped()) return true;
3557 // The works fine even if the array type is polymorphic.
3558 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3559 Node* result = load_array_length(array);
3561 C->set_has_split_ifs(true); // Has chance for split-if optimization
3562 set_result(result);
3563 return true;
3564 }
3566 //------------------------inline_array_copyOf----------------------------
3567 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3568 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3569 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3570 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3572 // Get the arguments.
3573 Node* original = argument(0);
3574 Node* start = is_copyOfRange? argument(1): intcon(0);
3575 Node* end = is_copyOfRange? argument(2): argument(1);
3576 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3578 Node* newcopy;
3580 // Set the original stack and the reexecute bit for the interpreter to reexecute
3581 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3582 { PreserveReexecuteState preexecs(this);
3583 jvms()->set_should_reexecute(true);
3585 array_type_mirror = null_check(array_type_mirror);
3586 original = null_check(original);
3588 // Check if a null path was taken unconditionally.
3589 if (stopped()) return true;
3591 Node* orig_length = load_array_length(original);
3593 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3594 klass_node = null_check(klass_node);
3596 RegionNode* bailout = new (C) RegionNode(1);
3597 record_for_igvn(bailout);
3599 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3600 // Bail out if that is so.
3601 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3602 if (not_objArray != NULL) {
3603 // Improve the klass node's type from the new optimistic assumption:
3604 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3605 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3606 Node* cast = new (C) CastPPNode(klass_node, akls);
3607 cast->init_req(0, control());
3608 klass_node = _gvn.transform(cast);
3609 }
3611 // Bail out if either start or end is negative.
3612 generate_negative_guard(start, bailout, &start);
3613 generate_negative_guard(end, bailout, &end);
3615 Node* length = end;
3616 if (_gvn.type(start) != TypeInt::ZERO) {
3617 length = _gvn.transform(new (C) SubINode(end, start));
3618 }
3620 // Bail out if length is negative.
3621 // Without this the new_array would throw
3622 // NegativeArraySizeException but IllegalArgumentException is what
3623 // should be thrown
3624 generate_negative_guard(length, bailout, &length);
3626 if (bailout->req() > 1) {
3627 PreserveJVMState pjvms(this);
3628 set_control(_gvn.transform(bailout));
3629 uncommon_trap(Deoptimization::Reason_intrinsic,
3630 Deoptimization::Action_maybe_recompile);
3631 }
3633 if (!stopped()) {
3634 // How many elements will we copy from the original?
3635 // The answer is MinI(orig_length - start, length).
3636 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3637 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3639 newcopy = new_array(klass_node, length, 0); // no argments to push
3641 // Generate a direct call to the right arraycopy function(s).
3642 // We know the copy is disjoint but we might not know if the
3643 // oop stores need checking.
3644 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3645 // This will fail a store-check if x contains any non-nulls.
3646 bool disjoint_bases = true;
3647 // if start > orig_length then the length of the copy may be
3648 // negative.
3649 bool length_never_negative = !is_copyOfRange;
3650 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3651 original, start, newcopy, intcon(0), moved,
3652 disjoint_bases, length_never_negative);
3653 }
3654 } // original reexecute is set back here
3656 C->set_has_split_ifs(true); // Has chance for split-if optimization
3657 if (!stopped()) {
3658 set_result(newcopy);
3659 }
3660 return true;
3661 }
3664 //----------------------generate_virtual_guard---------------------------
3665 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3666 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3667 RegionNode* slow_region) {
3668 ciMethod* method = callee();
3669 int vtable_index = method->vtable_index();
3670 // Get the Method* out of the appropriate vtable entry.
3671 int entry_offset = (InstanceKlass::vtable_start_offset() +
3672 vtable_index*vtableEntry::size()) * wordSize +
3673 vtableEntry::method_offset_in_bytes();
3674 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3675 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3677 // Compare the target method with the expected method (e.g., Object.hashCode).
3678 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3680 Node* native_call = makecon(native_call_addr);
3681 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3682 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3684 return generate_slow_guard(test_native, slow_region);
3685 }
3687 //-----------------------generate_method_call----------------------------
3688 // Use generate_method_call to make a slow-call to the real
3689 // method if the fast path fails. An alternative would be to
3690 // use a stub like OptoRuntime::slow_arraycopy_Java.
3691 // This only works for expanding the current library call,
3692 // not another intrinsic. (E.g., don't use this for making an
3693 // arraycopy call inside of the copyOf intrinsic.)
3694 CallJavaNode*
3695 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3696 // When compiling the intrinsic method itself, do not use this technique.
3697 guarantee(callee() != C->method(), "cannot make slow-call to self");
3699 ciMethod* method = callee();
3700 // ensure the JVMS we have will be correct for this call
3701 guarantee(method_id == method->intrinsic_id(), "must match");
3703 const TypeFunc* tf = TypeFunc::make(method);
3704 CallJavaNode* slow_call;
3705 if (is_static) {
3706 assert(!is_virtual, "");
3707 slow_call = new(C) CallStaticJavaNode(tf,
3708 SharedRuntime::get_resolve_static_call_stub(),
3709 method, bci());
3710 } else if (is_virtual) {
3711 null_check_receiver();
3712 int vtable_index = Method::invalid_vtable_index;
3713 if (UseInlineCaches) {
3714 // Suppress the vtable call
3715 } else {
3716 // hashCode and clone are not a miranda methods,
3717 // so the vtable index is fixed.
3718 // No need to use the linkResolver to get it.
3719 vtable_index = method->vtable_index();
3720 }
3721 slow_call = new(C) CallDynamicJavaNode(tf,
3722 SharedRuntime::get_resolve_virtual_call_stub(),
3723 method, vtable_index, bci());
3724 } else { // neither virtual nor static: opt_virtual
3725 null_check_receiver();
3726 slow_call = new(C) CallStaticJavaNode(tf,
3727 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3728 method, bci());
3729 slow_call->set_optimized_virtual(true);
3730 }
3731 set_arguments_for_java_call(slow_call);
3732 set_edges_for_java_call(slow_call);
3733 return slow_call;
3734 }
3737 //------------------------------inline_native_hashcode--------------------
3738 // Build special case code for calls to hashCode on an object.
3739 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3740 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3741 assert(!(is_virtual && is_static), "either virtual, special, or static");
3743 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3745 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3746 PhiNode* result_val = new(C) PhiNode(result_reg,
3747 TypeInt::INT);
3748 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3749 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3750 TypePtr::BOTTOM);
3751 Node* obj = NULL;
3752 if (!is_static) {
3753 // Check for hashing null object
3754 obj = null_check_receiver();
3755 if (stopped()) return true; // unconditionally null
3756 result_reg->init_req(_null_path, top());
3757 result_val->init_req(_null_path, top());
3758 } else {
3759 // Do a null check, and return zero if null.
3760 // System.identityHashCode(null) == 0
3761 obj = argument(0);
3762 Node* null_ctl = top();
3763 obj = null_check_oop(obj, &null_ctl);
3764 result_reg->init_req(_null_path, null_ctl);
3765 result_val->init_req(_null_path, _gvn.intcon(0));
3766 }
3768 // Unconditionally null? Then return right away.
3769 if (stopped()) {
3770 set_control( result_reg->in(_null_path));
3771 if (!stopped())
3772 set_result(result_val->in(_null_path));
3773 return true;
3774 }
3776 // After null check, get the object's klass.
3777 Node* obj_klass = load_object_klass(obj);
3779 // This call may be virtual (invokevirtual) or bound (invokespecial).
3780 // For each case we generate slightly different code.
3782 // We only go to the fast case code if we pass a number of guards. The
3783 // paths which do not pass are accumulated in the slow_region.
3784 RegionNode* slow_region = new (C) RegionNode(1);
3785 record_for_igvn(slow_region);
3787 // If this is a virtual call, we generate a funny guard. We pull out
3788 // the vtable entry corresponding to hashCode() from the target object.
3789 // If the target method which we are calling happens to be the native
3790 // Object hashCode() method, we pass the guard. We do not need this
3791 // guard for non-virtual calls -- the caller is known to be the native
3792 // Object hashCode().
3793 if (is_virtual) {
3794 generate_virtual_guard(obj_klass, slow_region);
3795 }
3797 // Get the header out of the object, use LoadMarkNode when available
3798 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3799 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3801 // Test the header to see if it is unlocked.
3802 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3803 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) );
3804 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3805 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val));
3806 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) );
3808 generate_slow_guard(test_unlocked, slow_region);
3810 // Get the hash value and check to see that it has been properly assigned.
3811 // We depend on hash_mask being at most 32 bits and avoid the use of
3812 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3813 // vm: see markOop.hpp.
3814 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3815 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3816 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) );
3817 // This hack lets the hash bits live anywhere in the mark object now, as long
3818 // as the shift drops the relevant bits into the low 32 bits. Note that
3819 // Java spec says that HashCode is an int so there's no point in capturing
3820 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3821 hshifted_header = ConvX2I(hshifted_header);
3822 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) );
3824 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3825 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val));
3826 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) );
3828 generate_slow_guard(test_assigned, slow_region);
3830 Node* init_mem = reset_memory();
3831 // fill in the rest of the null path:
3832 result_io ->init_req(_null_path, i_o());
3833 result_mem->init_req(_null_path, init_mem);
3835 result_val->init_req(_fast_path, hash_val);
3836 result_reg->init_req(_fast_path, control());
3837 result_io ->init_req(_fast_path, i_o());
3838 result_mem->init_req(_fast_path, init_mem);
3840 // Generate code for the slow case. We make a call to hashCode().
3841 set_control(_gvn.transform(slow_region));
3842 if (!stopped()) {
3843 // No need for PreserveJVMState, because we're using up the present state.
3844 set_all_memory(init_mem);
3845 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3846 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3847 Node* slow_result = set_results_for_java_call(slow_call);
3848 // this->control() comes from set_results_for_java_call
3849 result_reg->init_req(_slow_path, control());
3850 result_val->init_req(_slow_path, slow_result);
3851 result_io ->set_req(_slow_path, i_o());
3852 result_mem ->set_req(_slow_path, reset_memory());
3853 }
3855 // Return the combined state.
3856 set_i_o( _gvn.transform(result_io) );
3857 set_all_memory( _gvn.transform(result_mem) );
3859 set_result(result_reg, result_val);
3860 return true;
3861 }
3863 //---------------------------inline_native_getClass----------------------------
3864 // public final native Class<?> java.lang.Object.getClass();
3865 //
3866 // Build special case code for calls to getClass on an object.
3867 bool LibraryCallKit::inline_native_getClass() {
3868 Node* obj = null_check_receiver();
3869 if (stopped()) return true;
3870 set_result(load_mirror_from_klass(load_object_klass(obj)));
3871 return true;
3872 }
3874 //-----------------inline_native_Reflection_getCallerClass---------------------
3875 // public static native Class<?> sun.reflect.Reflection.getCallerClass(int realFramesToSkip);
3876 //
3877 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3878 //
3879 // NOTE that this code must perform the same logic as
3880 // vframeStream::security_get_caller_frame in that it must skip
3881 // Method.invoke() and auxiliary frames.
3882 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3883 #ifndef PRODUCT
3884 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3885 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3886 }
3887 #endif
3889 Node* caller_depth_node = argument(0);
3891 // The depth value must be a constant in order for the runtime call
3892 // to be eliminated.
3893 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3894 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3895 #ifndef PRODUCT
3896 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3897 tty->print_cr(" Bailing out because caller depth was not a constant");
3898 }
3899 #endif
3900 return false;
3901 }
3902 // Note that the JVM state at this point does not include the
3903 // getCallerClass() frame which we are trying to inline. The
3904 // semantics of getCallerClass(), however, are that the "first"
3905 // frame is the getCallerClass() frame, so we subtract one from the
3906 // requested depth before continuing. We don't inline requests of
3907 // getCallerClass(0).
3908 int caller_depth = caller_depth_type->get_con() - 1;
3909 if (caller_depth < 0) {
3910 #ifndef PRODUCT
3911 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3912 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3913 }
3914 #endif
3915 return false;
3916 }
3918 if (!jvms()->has_method()) {
3919 #ifndef PRODUCT
3920 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3921 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3922 }
3923 #endif
3924 return false;
3925 }
3926 int _depth = jvms()->depth(); // cache call chain depth
3928 // Walk back up the JVM state to find the caller at the required
3929 // depth. NOTE that this code must perform the same logic as
3930 // vframeStream::security_get_caller_frame in that it must skip
3931 // Method.invoke() and auxiliary frames. Note also that depth is
3932 // 1-based (1 is the bottom of the inlining).
3933 int inlining_depth = _depth;
3934 JVMState* caller_jvms = NULL;
3936 if (inlining_depth > 0) {
3937 caller_jvms = jvms();
3938 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3939 do {
3940 // The following if-tests should be performed in this order
3941 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3942 // Skip a Method.invoke() or auxiliary frame
3943 } else if (caller_depth > 0) {
3944 // Skip real frame
3945 --caller_depth;
3946 } else {
3947 // We're done: reached desired caller after skipping.
3948 break;
3949 }
3950 caller_jvms = caller_jvms->caller();
3951 --inlining_depth;
3952 } while (inlining_depth > 0);
3953 }
3955 if (inlining_depth == 0) {
3956 #ifndef PRODUCT
3957 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3958 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3959 tty->print_cr(" JVM state at this point:");
3960 for (int i = _depth; i >= 1; i--) {
3961 ciMethod* m = jvms()->of_depth(i)->method();
3962 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3963 }
3964 }
3965 #endif
3966 return false; // Reached end of inlining
3967 }
3969 // Acquire method holder as java.lang.Class
3970 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3971 ciInstance* caller_mirror = caller_klass->java_mirror();
3973 // Push this as a constant
3974 set_result(makecon(TypeInstPtr::make(caller_mirror)));
3976 #ifndef PRODUCT
3977 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3978 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);
3979 tty->print_cr(" JVM state at this point:");
3980 for (int i = _depth; i >= 1; i--) {
3981 ciMethod* m = jvms()->of_depth(i)->method();
3982 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3983 }
3984 }
3985 #endif
3986 return true;
3987 }
3989 // Helper routine for above
3990 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3991 ciMethod* method = jvms->method();
3993 // Is this the Method.invoke method itself?
3994 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3995 return true;
3997 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3998 ciKlass* k = method->holder();
3999 if (k->is_instance_klass()) {
4000 ciInstanceKlass* ik = k->as_instance_klass();
4001 for (; ik != NULL; ik = ik->super()) {
4002 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
4003 ik == env()->find_system_klass(ik->name())) {
4004 return true;
4005 }
4006 }
4007 }
4008 else if (method->is_method_handle_intrinsic() ||
4009 method->is_compiled_lambda_form()) {
4010 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
4011 return true;
4012 }
4014 return false;
4015 }
4017 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4018 Node* arg = argument(0);
4019 Node* result;
4021 switch (id) {
4022 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4023 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4024 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4025 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4027 case vmIntrinsics::_doubleToLongBits: {
4028 // two paths (plus control) merge in a wood
4029 RegionNode *r = new (C) RegionNode(3);
4030 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4032 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4033 // Build the boolean node
4034 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4036 // Branch either way.
4037 // NaN case is less traveled, which makes all the difference.
4038 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4039 Node *opt_isnan = _gvn.transform(ifisnan);
4040 assert( opt_isnan->is_If(), "Expect an IfNode");
4041 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4042 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4044 set_control(iftrue);
4046 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4047 Node *slow_result = longcon(nan_bits); // return NaN
4048 phi->init_req(1, _gvn.transform( slow_result ));
4049 r->init_req(1, iftrue);
4051 // Else fall through
4052 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4053 set_control(iffalse);
4055 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4056 r->init_req(2, iffalse);
4058 // Post merge
4059 set_control(_gvn.transform(r));
4060 record_for_igvn(r);
4062 C->set_has_split_ifs(true); // Has chance for split-if optimization
4063 result = phi;
4064 assert(result->bottom_type()->isa_long(), "must be");
4065 break;
4066 }
4068 case vmIntrinsics::_floatToIntBits: {
4069 // two paths (plus control) merge in a wood
4070 RegionNode *r = new (C) RegionNode(3);
4071 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4073 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4074 // Build the boolean node
4075 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4077 // Branch either way.
4078 // NaN case is less traveled, which makes all the difference.
4079 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4080 Node *opt_isnan = _gvn.transform(ifisnan);
4081 assert( opt_isnan->is_If(), "Expect an IfNode");
4082 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4083 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4085 set_control(iftrue);
4087 static const jint nan_bits = 0x7fc00000;
4088 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4089 phi->init_req(1, _gvn.transform( slow_result ));
4090 r->init_req(1, iftrue);
4092 // Else fall through
4093 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4094 set_control(iffalse);
4096 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4097 r->init_req(2, iffalse);
4099 // Post merge
4100 set_control(_gvn.transform(r));
4101 record_for_igvn(r);
4103 C->set_has_split_ifs(true); // Has chance for split-if optimization
4104 result = phi;
4105 assert(result->bottom_type()->isa_int(), "must be");
4106 break;
4107 }
4109 default:
4110 fatal_unexpected_iid(id);
4111 break;
4112 }
4113 set_result(_gvn.transform(result));
4114 return true;
4115 }
4117 #ifdef _LP64
4118 #define XTOP ,top() /*additional argument*/
4119 #else //_LP64
4120 #define XTOP /*no additional argument*/
4121 #endif //_LP64
4123 //----------------------inline_unsafe_copyMemory-------------------------
4124 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4125 bool LibraryCallKit::inline_unsafe_copyMemory() {
4126 if (callee()->is_static()) return false; // caller must have the capability!
4127 null_check_receiver(); // null-check receiver
4128 if (stopped()) return true;
4130 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4132 Node* src_ptr = argument(1); // type: oop
4133 Node* src_off = ConvL2X(argument(2)); // type: long
4134 Node* dst_ptr = argument(4); // type: oop
4135 Node* dst_off = ConvL2X(argument(5)); // type: long
4136 Node* size = ConvL2X(argument(7)); // type: long
4138 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4139 "fieldOffset must be byte-scaled");
4141 Node* src = make_unsafe_address(src_ptr, src_off);
4142 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4144 // Conservatively insert a memory barrier on all memory slices.
4145 // Do not let writes of the copy source or destination float below the copy.
4146 insert_mem_bar(Op_MemBarCPUOrder);
4148 // Call it. Note that the length argument is not scaled.
4149 make_runtime_call(RC_LEAF|RC_NO_FP,
4150 OptoRuntime::fast_arraycopy_Type(),
4151 StubRoutines::unsafe_arraycopy(),
4152 "unsafe_arraycopy",
4153 TypeRawPtr::BOTTOM,
4154 src, dst, size XTOP);
4156 // Do not let reads of the copy destination float above the copy.
4157 insert_mem_bar(Op_MemBarCPUOrder);
4159 return true;
4160 }
4162 //------------------------clone_coping-----------------------------------
4163 // Helper function for inline_native_clone.
4164 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4165 assert(obj_size != NULL, "");
4166 Node* raw_obj = alloc_obj->in(1);
4167 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4169 AllocateNode* alloc = NULL;
4170 if (ReduceBulkZeroing) {
4171 // We will be completely responsible for initializing this object -
4172 // mark Initialize node as complete.
4173 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4174 // The object was just allocated - there should be no any stores!
4175 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4176 // Mark as complete_with_arraycopy so that on AllocateNode
4177 // expansion, we know this AllocateNode is initialized by an array
4178 // copy and a StoreStore barrier exists after the array copy.
4179 alloc->initialization()->set_complete_with_arraycopy();
4180 }
4182 // Copy the fastest available way.
4183 // TODO: generate fields copies for small objects instead.
4184 Node* src = obj;
4185 Node* dest = alloc_obj;
4186 Node* size = _gvn.transform(obj_size);
4188 // Exclude the header but include array length to copy by 8 bytes words.
4189 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4190 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4191 instanceOopDesc::base_offset_in_bytes();
4192 // base_off:
4193 // 8 - 32-bit VM
4194 // 12 - 64-bit VM, compressed klass
4195 // 16 - 64-bit VM, normal klass
4196 if (base_off % BytesPerLong != 0) {
4197 assert(UseCompressedKlassPointers, "");
4198 if (is_array) {
4199 // Exclude length to copy by 8 bytes words.
4200 base_off += sizeof(int);
4201 } else {
4202 // Include klass to copy by 8 bytes words.
4203 base_off = instanceOopDesc::klass_offset_in_bytes();
4204 }
4205 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4206 }
4207 src = basic_plus_adr(src, base_off);
4208 dest = basic_plus_adr(dest, base_off);
4210 // Compute the length also, if needed:
4211 Node* countx = size;
4212 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4213 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4215 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4216 bool disjoint_bases = true;
4217 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4218 src, NULL, dest, NULL, countx,
4219 /*dest_uninitialized*/true);
4221 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4222 if (card_mark) {
4223 assert(!is_array, "");
4224 // Put in store barrier for any and all oops we are sticking
4225 // into this object. (We could avoid this if we could prove
4226 // that the object type contains no oop fields at all.)
4227 Node* no_particular_value = NULL;
4228 Node* no_particular_field = NULL;
4229 int raw_adr_idx = Compile::AliasIdxRaw;
4230 post_barrier(control(),
4231 memory(raw_adr_type),
4232 alloc_obj,
4233 no_particular_field,
4234 raw_adr_idx,
4235 no_particular_value,
4236 T_OBJECT,
4237 false);
4238 }
4240 // Do not let reads from the cloned object float above the arraycopy.
4241 if (alloc != NULL) {
4242 // Do not let stores that initialize this object be reordered with
4243 // a subsequent store that would make this object accessible by
4244 // other threads.
4245 // Record what AllocateNode this StoreStore protects so that
4246 // escape analysis can go from the MemBarStoreStoreNode to the
4247 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4248 // based on the escape status of the AllocateNode.
4249 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4250 } else {
4251 insert_mem_bar(Op_MemBarCPUOrder);
4252 }
4253 }
4255 //------------------------inline_native_clone----------------------------
4256 // protected native Object java.lang.Object.clone();
4257 //
4258 // Here are the simple edge cases:
4259 // null receiver => normal trap
4260 // virtual and clone was overridden => slow path to out-of-line clone
4261 // not cloneable or finalizer => slow path to out-of-line Object.clone
4262 //
4263 // The general case has two steps, allocation and copying.
4264 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4265 //
4266 // Copying also has two cases, oop arrays and everything else.
4267 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4268 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4269 //
4270 // These steps fold up nicely if and when the cloned object's klass
4271 // can be sharply typed as an object array, a type array, or an instance.
4272 //
4273 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4274 PhiNode* result_val;
4276 // Set the reexecute bit for the interpreter to reexecute
4277 // the bytecode that invokes Object.clone if deoptimization happens.
4278 { PreserveReexecuteState preexecs(this);
4279 jvms()->set_should_reexecute(true);
4281 Node* obj = null_check_receiver();
4282 if (stopped()) return true;
4284 Node* obj_klass = load_object_klass(obj);
4285 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4286 const TypeOopPtr* toop = ((tklass != NULL)
4287 ? tklass->as_instance_type()
4288 : TypeInstPtr::NOTNULL);
4290 // Conservatively insert a memory barrier on all memory slices.
4291 // Do not let writes into the original float below the clone.
4292 insert_mem_bar(Op_MemBarCPUOrder);
4294 // paths into result_reg:
4295 enum {
4296 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4297 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4298 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4299 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4300 PATH_LIMIT
4301 };
4302 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4303 result_val = new(C) PhiNode(result_reg,
4304 TypeInstPtr::NOTNULL);
4305 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4306 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4307 TypePtr::BOTTOM);
4308 record_for_igvn(result_reg);
4310 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4311 int raw_adr_idx = Compile::AliasIdxRaw;
4313 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4314 if (array_ctl != NULL) {
4315 // It's an array.
4316 PreserveJVMState pjvms(this);
4317 set_control(array_ctl);
4318 Node* obj_length = load_array_length(obj);
4319 Node* obj_size = NULL;
4320 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4322 if (!use_ReduceInitialCardMarks()) {
4323 // If it is an oop array, it requires very special treatment,
4324 // because card marking is required on each card of the array.
4325 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4326 if (is_obja != NULL) {
4327 PreserveJVMState pjvms2(this);
4328 set_control(is_obja);
4329 // Generate a direct call to the right arraycopy function(s).
4330 bool disjoint_bases = true;
4331 bool length_never_negative = true;
4332 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4333 obj, intcon(0), alloc_obj, intcon(0),
4334 obj_length,
4335 disjoint_bases, length_never_negative);
4336 result_reg->init_req(_objArray_path, control());
4337 result_val->init_req(_objArray_path, alloc_obj);
4338 result_i_o ->set_req(_objArray_path, i_o());
4339 result_mem ->set_req(_objArray_path, reset_memory());
4340 }
4341 }
4342 // Otherwise, there are no card marks to worry about.
4343 // (We can dispense with card marks if we know the allocation
4344 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4345 // causes the non-eden paths to take compensating steps to
4346 // simulate a fresh allocation, so that no further
4347 // card marks are required in compiled code to initialize
4348 // the object.)
4350 if (!stopped()) {
4351 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4353 // Present the results of the copy.
4354 result_reg->init_req(_array_path, control());
4355 result_val->init_req(_array_path, alloc_obj);
4356 result_i_o ->set_req(_array_path, i_o());
4357 result_mem ->set_req(_array_path, reset_memory());
4358 }
4359 }
4361 // We only go to the instance fast case code if we pass a number of guards.
4362 // The paths which do not pass are accumulated in the slow_region.
4363 RegionNode* slow_region = new (C) RegionNode(1);
4364 record_for_igvn(slow_region);
4365 if (!stopped()) {
4366 // It's an instance (we did array above). Make the slow-path tests.
4367 // If this is a virtual call, we generate a funny guard. We grab
4368 // the vtable entry corresponding to clone() from the target object.
4369 // If the target method which we are calling happens to be the
4370 // Object clone() method, we pass the guard. We do not need this
4371 // guard for non-virtual calls; the caller is known to be the native
4372 // Object clone().
4373 if (is_virtual) {
4374 generate_virtual_guard(obj_klass, slow_region);
4375 }
4377 // The object must be cloneable and must not have a finalizer.
4378 // Both of these conditions may be checked in a single test.
4379 // We could optimize the cloneable test further, but we don't care.
4380 generate_access_flags_guard(obj_klass,
4381 // Test both conditions:
4382 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4383 // Must be cloneable but not finalizer:
4384 JVM_ACC_IS_CLONEABLE,
4385 slow_region);
4386 }
4388 if (!stopped()) {
4389 // It's an instance, and it passed the slow-path tests.
4390 PreserveJVMState pjvms(this);
4391 Node* obj_size = NULL;
4392 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4394 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4396 // Present the results of the slow call.
4397 result_reg->init_req(_instance_path, control());
4398 result_val->init_req(_instance_path, alloc_obj);
4399 result_i_o ->set_req(_instance_path, i_o());
4400 result_mem ->set_req(_instance_path, reset_memory());
4401 }
4403 // Generate code for the slow case. We make a call to clone().
4404 set_control(_gvn.transform(slow_region));
4405 if (!stopped()) {
4406 PreserveJVMState pjvms(this);
4407 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4408 Node* slow_result = set_results_for_java_call(slow_call);
4409 // this->control() comes from set_results_for_java_call
4410 result_reg->init_req(_slow_path, control());
4411 result_val->init_req(_slow_path, slow_result);
4412 result_i_o ->set_req(_slow_path, i_o());
4413 result_mem ->set_req(_slow_path, reset_memory());
4414 }
4416 // Return the combined state.
4417 set_control( _gvn.transform(result_reg) );
4418 set_i_o( _gvn.transform(result_i_o) );
4419 set_all_memory( _gvn.transform(result_mem) );
4420 } // original reexecute is set back here
4422 set_result(_gvn.transform(result_val));
4423 return true;
4424 }
4426 //------------------------------basictype2arraycopy----------------------------
4427 address LibraryCallKit::basictype2arraycopy(BasicType t,
4428 Node* src_offset,
4429 Node* dest_offset,
4430 bool disjoint_bases,
4431 const char* &name,
4432 bool dest_uninitialized) {
4433 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4434 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4436 bool aligned = false;
4437 bool disjoint = disjoint_bases;
4439 // if the offsets are the same, we can treat the memory regions as
4440 // disjoint, because either the memory regions are in different arrays,
4441 // or they are identical (which we can treat as disjoint.) We can also
4442 // treat a copy with a destination index less that the source index
4443 // as disjoint since a low->high copy will work correctly in this case.
4444 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4445 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4446 // both indices are constants
4447 int s_offs = src_offset_inttype->get_con();
4448 int d_offs = dest_offset_inttype->get_con();
4449 int element_size = type2aelembytes(t);
4450 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4451 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4452 if (s_offs >= d_offs) disjoint = true;
4453 } else if (src_offset == dest_offset && src_offset != NULL) {
4454 // This can occur if the offsets are identical non-constants.
4455 disjoint = true;
4456 }
4458 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4459 }
4462 //------------------------------inline_arraycopy-----------------------
4463 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4464 // Object dest, int destPos,
4465 // int length);
4466 bool LibraryCallKit::inline_arraycopy() {
4467 // Get the arguments.
4468 Node* src = argument(0); // type: oop
4469 Node* src_offset = argument(1); // type: int
4470 Node* dest = argument(2); // type: oop
4471 Node* dest_offset = argument(3); // type: int
4472 Node* length = argument(4); // type: int
4474 // Compile time checks. If any of these checks cannot be verified at compile time,
4475 // we do not make a fast path for this call. Instead, we let the call remain as it
4476 // is. The checks we choose to mandate at compile time are:
4477 //
4478 // (1) src and dest are arrays.
4479 const Type* src_type = src->Value(&_gvn);
4480 const Type* dest_type = dest->Value(&_gvn);
4481 const TypeAryPtr* top_src = src_type->isa_aryptr();
4482 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4483 if (top_src == NULL || top_src->klass() == NULL ||
4484 top_dest == NULL || top_dest->klass() == NULL) {
4485 // Conservatively insert a memory barrier on all memory slices.
4486 // Do not let writes into the source float below the arraycopy.
4487 insert_mem_bar(Op_MemBarCPUOrder);
4489 // Call StubRoutines::generic_arraycopy stub.
4490 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4491 src, src_offset, dest, dest_offset, length);
4493 // Do not let reads from the destination float above the arraycopy.
4494 // Since we cannot type the arrays, we don't know which slices
4495 // might be affected. We could restrict this barrier only to those
4496 // memory slices which pertain to array elements--but don't bother.
4497 if (!InsertMemBarAfterArraycopy)
4498 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4499 insert_mem_bar(Op_MemBarCPUOrder);
4500 return true;
4501 }
4503 // (2) src and dest arrays must have elements of the same BasicType
4504 // Figure out the size and type of the elements we will be copying.
4505 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4506 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4507 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4508 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4510 if (src_elem != dest_elem || dest_elem == T_VOID) {
4511 // The component types are not the same or are not recognized. Punt.
4512 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4513 generate_slow_arraycopy(TypePtr::BOTTOM,
4514 src, src_offset, dest, dest_offset, length,
4515 /*dest_uninitialized*/false);
4516 return true;
4517 }
4519 //---------------------------------------------------------------------------
4520 // We will make a fast path for this call to arraycopy.
4522 // We have the following tests left to perform:
4523 //
4524 // (3) src and dest must not be null.
4525 // (4) src_offset must not be negative.
4526 // (5) dest_offset must not be negative.
4527 // (6) length must not be negative.
4528 // (7) src_offset + length must not exceed length of src.
4529 // (8) dest_offset + length must not exceed length of dest.
4530 // (9) each element of an oop array must be assignable
4532 RegionNode* slow_region = new (C) RegionNode(1);
4533 record_for_igvn(slow_region);
4535 // (3) operands must not be null
4536 // We currently perform our null checks with the null_check routine.
4537 // This means that the null exceptions will be reported in the caller
4538 // rather than (correctly) reported inside of the native arraycopy call.
4539 // This should be corrected, given time. We do our null check with the
4540 // stack pointer restored.
4541 src = null_check(src, T_ARRAY);
4542 dest = null_check(dest, T_ARRAY);
4544 // (4) src_offset must not be negative.
4545 generate_negative_guard(src_offset, slow_region);
4547 // (5) dest_offset must not be negative.
4548 generate_negative_guard(dest_offset, slow_region);
4550 // (6) length must not be negative (moved to generate_arraycopy()).
4551 // generate_negative_guard(length, slow_region);
4553 // (7) src_offset + length must not exceed length of src.
4554 generate_limit_guard(src_offset, length,
4555 load_array_length(src),
4556 slow_region);
4558 // (8) dest_offset + length must not exceed length of dest.
4559 generate_limit_guard(dest_offset, length,
4560 load_array_length(dest),
4561 slow_region);
4563 // (9) each element of an oop array must be assignable
4564 // The generate_arraycopy subroutine checks this.
4566 // This is where the memory effects are placed:
4567 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4568 generate_arraycopy(adr_type, dest_elem,
4569 src, src_offset, dest, dest_offset, length,
4570 false, false, slow_region);
4572 return true;
4573 }
4575 //-----------------------------generate_arraycopy----------------------
4576 // Generate an optimized call to arraycopy.
4577 // Caller must guard against non-arrays.
4578 // Caller must determine a common array basic-type for both arrays.
4579 // Caller must validate offsets against array bounds.
4580 // The slow_region has already collected guard failure paths
4581 // (such as out of bounds length or non-conformable array types).
4582 // The generated code has this shape, in general:
4583 //
4584 // if (length == 0) return // via zero_path
4585 // slowval = -1
4586 // if (types unknown) {
4587 // slowval = call generic copy loop
4588 // if (slowval == 0) return // via checked_path
4589 // } else if (indexes in bounds) {
4590 // if ((is object array) && !(array type check)) {
4591 // slowval = call checked copy loop
4592 // if (slowval == 0) return // via checked_path
4593 // } else {
4594 // call bulk copy loop
4595 // return // via fast_path
4596 // }
4597 // }
4598 // // adjust params for remaining work:
4599 // if (slowval != -1) {
4600 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4601 // }
4602 // slow_region:
4603 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4604 // return // via slow_call_path
4605 //
4606 // This routine is used from several intrinsics: System.arraycopy,
4607 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4608 //
4609 void
4610 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4611 BasicType basic_elem_type,
4612 Node* src, Node* src_offset,
4613 Node* dest, Node* dest_offset,
4614 Node* copy_length,
4615 bool disjoint_bases,
4616 bool length_never_negative,
4617 RegionNode* slow_region) {
4619 if (slow_region == NULL) {
4620 slow_region = new(C) RegionNode(1);
4621 record_for_igvn(slow_region);
4622 }
4624 Node* original_dest = dest;
4625 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4626 bool dest_uninitialized = false;
4628 // See if this is the initialization of a newly-allocated array.
4629 // If so, we will take responsibility here for initializing it to zero.
4630 // (Note: Because tightly_coupled_allocation performs checks on the
4631 // out-edges of the dest, we need to avoid making derived pointers
4632 // from it until we have checked its uses.)
4633 if (ReduceBulkZeroing
4634 && !ZeroTLAB // pointless if already zeroed
4635 && basic_elem_type != T_CONFLICT // avoid corner case
4636 && !src->eqv_uncast(dest)
4637 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4638 != NULL)
4639 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4640 && alloc->maybe_set_complete(&_gvn)) {
4641 // "You break it, you buy it."
4642 InitializeNode* init = alloc->initialization();
4643 assert(init->is_complete(), "we just did this");
4644 init->set_complete_with_arraycopy();
4645 assert(dest->is_CheckCastPP(), "sanity");
4646 assert(dest->in(0)->in(0) == init, "dest pinned");
4647 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4648 // From this point on, every exit path is responsible for
4649 // initializing any non-copied parts of the object to zero.
4650 // Also, if this flag is set we make sure that arraycopy interacts properly
4651 // with G1, eliding pre-barriers. See CR 6627983.
4652 dest_uninitialized = true;
4653 } else {
4654 // No zeroing elimination here.
4655 alloc = NULL;
4656 //original_dest = dest;
4657 //dest_uninitialized = false;
4658 }
4660 // Results are placed here:
4661 enum { fast_path = 1, // normal void-returning assembly stub
4662 checked_path = 2, // special assembly stub with cleanup
4663 slow_call_path = 3, // something went wrong; call the VM
4664 zero_path = 4, // bypass when length of copy is zero
4665 bcopy_path = 5, // copy primitive array by 64-bit blocks
4666 PATH_LIMIT = 6
4667 };
4668 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4669 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4670 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4671 record_for_igvn(result_region);
4672 _gvn.set_type_bottom(result_i_o);
4673 _gvn.set_type_bottom(result_memory);
4674 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4676 // The slow_control path:
4677 Node* slow_control;
4678 Node* slow_i_o = i_o();
4679 Node* slow_mem = memory(adr_type);
4680 debug_only(slow_control = (Node*) badAddress);
4682 // Checked control path:
4683 Node* checked_control = top();
4684 Node* checked_mem = NULL;
4685 Node* checked_i_o = NULL;
4686 Node* checked_value = NULL;
4688 if (basic_elem_type == T_CONFLICT) {
4689 assert(!dest_uninitialized, "");
4690 Node* cv = generate_generic_arraycopy(adr_type,
4691 src, src_offset, dest, dest_offset,
4692 copy_length, dest_uninitialized);
4693 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4694 checked_control = control();
4695 checked_i_o = i_o();
4696 checked_mem = memory(adr_type);
4697 checked_value = cv;
4698 set_control(top()); // no fast path
4699 }
4701 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4702 if (not_pos != NULL) {
4703 PreserveJVMState pjvms(this);
4704 set_control(not_pos);
4706 // (6) length must not be negative.
4707 if (!length_never_negative) {
4708 generate_negative_guard(copy_length, slow_region);
4709 }
4711 // copy_length is 0.
4712 if (!stopped() && dest_uninitialized) {
4713 Node* dest_length = alloc->in(AllocateNode::ALength);
4714 if (copy_length->eqv_uncast(dest_length)
4715 || _gvn.find_int_con(dest_length, 1) <= 0) {
4716 // There is no zeroing to do. No need for a secondary raw memory barrier.
4717 } else {
4718 // Clear the whole thing since there are no source elements to copy.
4719 generate_clear_array(adr_type, dest, basic_elem_type,
4720 intcon(0), NULL,
4721 alloc->in(AllocateNode::AllocSize));
4722 // Use a secondary InitializeNode as raw memory barrier.
4723 // Currently it is needed only on this path since other
4724 // paths have stub or runtime calls as raw memory barriers.
4725 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4726 Compile::AliasIdxRaw,
4727 top())->as_Initialize();
4728 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4729 }
4730 }
4732 // Present the results of the fast call.
4733 result_region->init_req(zero_path, control());
4734 result_i_o ->init_req(zero_path, i_o());
4735 result_memory->init_req(zero_path, memory(adr_type));
4736 }
4738 if (!stopped() && dest_uninitialized) {
4739 // We have to initialize the *uncopied* part of the array to zero.
4740 // The copy destination is the slice dest[off..off+len]. The other slices
4741 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4742 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4743 Node* dest_length = alloc->in(AllocateNode::ALength);
4744 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
4745 copy_length) );
4747 // If there is a head section that needs zeroing, do it now.
4748 if (find_int_con(dest_offset, -1) != 0) {
4749 generate_clear_array(adr_type, dest, basic_elem_type,
4750 intcon(0), dest_offset,
4751 NULL);
4752 }
4754 // Next, perform a dynamic check on the tail length.
4755 // It is often zero, and we can win big if we prove this.
4756 // There are two wins: Avoid generating the ClearArray
4757 // with its attendant messy index arithmetic, and upgrade
4758 // the copy to a more hardware-friendly word size of 64 bits.
4759 Node* tail_ctl = NULL;
4760 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4761 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
4762 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
4763 tail_ctl = generate_slow_guard(bol_lt, NULL);
4764 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4765 }
4767 // At this point, let's assume there is no tail.
4768 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4769 // There is no tail. Try an upgrade to a 64-bit copy.
4770 bool didit = false;
4771 { PreserveJVMState pjvms(this);
4772 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4773 src, src_offset, dest, dest_offset,
4774 dest_size, dest_uninitialized);
4775 if (didit) {
4776 // Present the results of the block-copying fast call.
4777 result_region->init_req(bcopy_path, control());
4778 result_i_o ->init_req(bcopy_path, i_o());
4779 result_memory->init_req(bcopy_path, memory(adr_type));
4780 }
4781 }
4782 if (didit)
4783 set_control(top()); // no regular fast path
4784 }
4786 // Clear the tail, if any.
4787 if (tail_ctl != NULL) {
4788 Node* notail_ctl = stopped() ? NULL : control();
4789 set_control(tail_ctl);
4790 if (notail_ctl == NULL) {
4791 generate_clear_array(adr_type, dest, basic_elem_type,
4792 dest_tail, NULL,
4793 dest_size);
4794 } else {
4795 // Make a local merge.
4796 Node* done_ctl = new(C) RegionNode(3);
4797 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4798 done_ctl->init_req(1, notail_ctl);
4799 done_mem->init_req(1, memory(adr_type));
4800 generate_clear_array(adr_type, dest, basic_elem_type,
4801 dest_tail, NULL,
4802 dest_size);
4803 done_ctl->init_req(2, control());
4804 done_mem->init_req(2, memory(adr_type));
4805 set_control( _gvn.transform(done_ctl) );
4806 set_memory( _gvn.transform(done_mem), adr_type );
4807 }
4808 }
4809 }
4811 BasicType copy_type = basic_elem_type;
4812 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4813 if (!stopped() && copy_type == T_OBJECT) {
4814 // If src and dest have compatible element types, we can copy bits.
4815 // Types S[] and D[] are compatible if D is a supertype of S.
4816 //
4817 // If they are not, we will use checked_oop_disjoint_arraycopy,
4818 // which performs a fast optimistic per-oop check, and backs off
4819 // further to JVM_ArrayCopy on the first per-oop check that fails.
4820 // (Actually, we don't move raw bits only; the GC requires card marks.)
4822 // Get the Klass* for both src and dest
4823 Node* src_klass = load_object_klass(src);
4824 Node* dest_klass = load_object_klass(dest);
4826 // Generate the subtype check.
4827 // This might fold up statically, or then again it might not.
4828 //
4829 // Non-static example: Copying List<String>.elements to a new String[].
4830 // The backing store for a List<String> is always an Object[],
4831 // but its elements are always type String, if the generic types
4832 // are correct at the source level.
4833 //
4834 // Test S[] against D[], not S against D, because (probably)
4835 // the secondary supertype cache is less busy for S[] than S.
4836 // This usually only matters when D is an interface.
4837 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4838 // Plug failing path into checked_oop_disjoint_arraycopy
4839 if (not_subtype_ctrl != top()) {
4840 PreserveJVMState pjvms(this);
4841 set_control(not_subtype_ctrl);
4842 // (At this point we can assume disjoint_bases, since types differ.)
4843 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
4844 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4845 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4846 Node* dest_elem_klass = _gvn.transform(n1);
4847 Node* cv = generate_checkcast_arraycopy(adr_type,
4848 dest_elem_klass,
4849 src, src_offset, dest, dest_offset,
4850 ConvI2X(copy_length), dest_uninitialized);
4851 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4852 checked_control = control();
4853 checked_i_o = i_o();
4854 checked_mem = memory(adr_type);
4855 checked_value = cv;
4856 }
4857 // At this point we know we do not need type checks on oop stores.
4859 // Let's see if we need card marks:
4860 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4861 // If we do not need card marks, copy using the jint or jlong stub.
4862 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4863 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4864 "sizes agree");
4865 }
4866 }
4868 if (!stopped()) {
4869 // Generate the fast path, if possible.
4870 PreserveJVMState pjvms(this);
4871 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4872 src, src_offset, dest, dest_offset,
4873 ConvI2X(copy_length), dest_uninitialized);
4875 // Present the results of the fast call.
4876 result_region->init_req(fast_path, control());
4877 result_i_o ->init_req(fast_path, i_o());
4878 result_memory->init_req(fast_path, memory(adr_type));
4879 }
4881 // Here are all the slow paths up to this point, in one bundle:
4882 slow_control = top();
4883 if (slow_region != NULL)
4884 slow_control = _gvn.transform(slow_region);
4885 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
4887 set_control(checked_control);
4888 if (!stopped()) {
4889 // Clean up after the checked call.
4890 // The returned value is either 0 or -1^K,
4891 // where K = number of partially transferred array elements.
4892 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
4893 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
4894 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4896 // If it is 0, we are done, so transfer to the end.
4897 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
4898 result_region->init_req(checked_path, checks_done);
4899 result_i_o ->init_req(checked_path, checked_i_o);
4900 result_memory->init_req(checked_path, checked_mem);
4902 // If it is not zero, merge into the slow call.
4903 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
4904 RegionNode* slow_reg2 = new(C) RegionNode(3);
4905 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
4906 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4907 record_for_igvn(slow_reg2);
4908 slow_reg2 ->init_req(1, slow_control);
4909 slow_i_o2 ->init_req(1, slow_i_o);
4910 slow_mem2 ->init_req(1, slow_mem);
4911 slow_reg2 ->init_req(2, control());
4912 slow_i_o2 ->init_req(2, checked_i_o);
4913 slow_mem2 ->init_req(2, checked_mem);
4915 slow_control = _gvn.transform(slow_reg2);
4916 slow_i_o = _gvn.transform(slow_i_o2);
4917 slow_mem = _gvn.transform(slow_mem2);
4919 if (alloc != NULL) {
4920 // We'll restart from the very beginning, after zeroing the whole thing.
4921 // This can cause double writes, but that's OK since dest is brand new.
4922 // So we ignore the low 31 bits of the value returned from the stub.
4923 } else {
4924 // We must continue the copy exactly where it failed, or else
4925 // another thread might see the wrong number of writes to dest.
4926 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
4927 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
4928 slow_offset->init_req(1, intcon(0));
4929 slow_offset->init_req(2, checked_offset);
4930 slow_offset = _gvn.transform(slow_offset);
4932 // Adjust the arguments by the conditionally incoming offset.
4933 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
4934 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
4935 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
4937 // Tweak the node variables to adjust the code produced below:
4938 src_offset = src_off_plus;
4939 dest_offset = dest_off_plus;
4940 copy_length = length_minus;
4941 }
4942 }
4944 set_control(slow_control);
4945 if (!stopped()) {
4946 // Generate the slow path, if needed.
4947 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4949 set_memory(slow_mem, adr_type);
4950 set_i_o(slow_i_o);
4952 if (dest_uninitialized) {
4953 generate_clear_array(adr_type, dest, basic_elem_type,
4954 intcon(0), NULL,
4955 alloc->in(AllocateNode::AllocSize));
4956 }
4958 generate_slow_arraycopy(adr_type,
4959 src, src_offset, dest, dest_offset,
4960 copy_length, /*dest_uninitialized*/false);
4962 result_region->init_req(slow_call_path, control());
4963 result_i_o ->init_req(slow_call_path, i_o());
4964 result_memory->init_req(slow_call_path, memory(adr_type));
4965 }
4967 // Remove unused edges.
4968 for (uint i = 1; i < result_region->req(); i++) {
4969 if (result_region->in(i) == NULL)
4970 result_region->init_req(i, top());
4971 }
4973 // Finished; return the combined state.
4974 set_control( _gvn.transform(result_region) );
4975 set_i_o( _gvn.transform(result_i_o) );
4976 set_memory( _gvn.transform(result_memory), adr_type );
4978 // The memory edges above are precise in order to model effects around
4979 // array copies accurately to allow value numbering of field loads around
4980 // arraycopy. Such field loads, both before and after, are common in Java
4981 // collections and similar classes involving header/array data structures.
4982 //
4983 // But with low number of register or when some registers are used or killed
4984 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4985 // The next memory barrier is added to avoid it. If the arraycopy can be
4986 // optimized away (which it can, sometimes) then we can manually remove
4987 // the membar also.
4988 //
4989 // Do not let reads from the cloned object float above the arraycopy.
4990 if (alloc != NULL) {
4991 // Do not let stores that initialize this object be reordered with
4992 // a subsequent store that would make this object accessible by
4993 // other threads.
4994 // Record what AllocateNode this StoreStore protects so that
4995 // escape analysis can go from the MemBarStoreStoreNode to the
4996 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4997 // based on the escape status of the AllocateNode.
4998 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4999 } else if (InsertMemBarAfterArraycopy)
5000 insert_mem_bar(Op_MemBarCPUOrder);
5001 }
5004 // Helper function which determines if an arraycopy immediately follows
5005 // an allocation, with no intervening tests or other escapes for the object.
5006 AllocateArrayNode*
5007 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5008 RegionNode* slow_region) {
5009 if (stopped()) return NULL; // no fast path
5010 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5012 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5013 if (alloc == NULL) return NULL;
5015 Node* rawmem = memory(Compile::AliasIdxRaw);
5016 // Is the allocation's memory state untouched?
5017 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5018 // Bail out if there have been raw-memory effects since the allocation.
5019 // (Example: There might have been a call or safepoint.)
5020 return NULL;
5021 }
5022 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5023 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5024 return NULL;
5025 }
5027 // There must be no unexpected observers of this allocation.
5028 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5029 Node* obs = ptr->fast_out(i);
5030 if (obs != this->map()) {
5031 return NULL;
5032 }
5033 }
5035 // This arraycopy must unconditionally follow the allocation of the ptr.
5036 Node* alloc_ctl = ptr->in(0);
5037 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5039 Node* ctl = control();
5040 while (ctl != alloc_ctl) {
5041 // There may be guards which feed into the slow_region.
5042 // Any other control flow means that we might not get a chance
5043 // to finish initializing the allocated object.
5044 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5045 IfNode* iff = ctl->in(0)->as_If();
5046 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5047 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5048 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5049 ctl = iff->in(0); // This test feeds the known slow_region.
5050 continue;
5051 }
5052 // One more try: Various low-level checks bottom out in
5053 // uncommon traps. If the debug-info of the trap omits
5054 // any reference to the allocation, as we've already
5055 // observed, then there can be no objection to the trap.
5056 bool found_trap = false;
5057 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5058 Node* obs = not_ctl->fast_out(j);
5059 if (obs->in(0) == not_ctl && obs->is_Call() &&
5060 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5061 found_trap = true; break;
5062 }
5063 }
5064 if (found_trap) {
5065 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5066 continue;
5067 }
5068 }
5069 return NULL;
5070 }
5072 // If we get this far, we have an allocation which immediately
5073 // precedes the arraycopy, and we can take over zeroing the new object.
5074 // The arraycopy will finish the initialization, and provide
5075 // a new control state to which we will anchor the destination pointer.
5077 return alloc;
5078 }
5080 // Helper for initialization of arrays, creating a ClearArray.
5081 // It writes zero bits in [start..end), within the body of an array object.
5082 // The memory effects are all chained onto the 'adr_type' alias category.
5083 //
5084 // Since the object is otherwise uninitialized, we are free
5085 // to put a little "slop" around the edges of the cleared area,
5086 // as long as it does not go back into the array's header,
5087 // or beyond the array end within the heap.
5088 //
5089 // The lower edge can be rounded down to the nearest jint and the
5090 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5091 //
5092 // Arguments:
5093 // adr_type memory slice where writes are generated
5094 // dest oop of the destination array
5095 // basic_elem_type element type of the destination
5096 // slice_idx array index of first element to store
5097 // slice_len number of elements to store (or NULL)
5098 // dest_size total size in bytes of the array object
5099 //
5100 // Exactly one of slice_len or dest_size must be non-NULL.
5101 // If dest_size is non-NULL, zeroing extends to the end of the object.
5102 // If slice_len is non-NULL, the slice_idx value must be a constant.
5103 void
5104 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5105 Node* dest,
5106 BasicType basic_elem_type,
5107 Node* slice_idx,
5108 Node* slice_len,
5109 Node* dest_size) {
5110 // one or the other but not both of slice_len and dest_size:
5111 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5112 if (slice_len == NULL) slice_len = top();
5113 if (dest_size == NULL) dest_size = top();
5115 // operate on this memory slice:
5116 Node* mem = memory(adr_type); // memory slice to operate on
5118 // scaling and rounding of indexes:
5119 int scale = exact_log2(type2aelembytes(basic_elem_type));
5120 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5121 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5122 int bump_bit = (-1 << scale) & BytesPerInt;
5124 // determine constant starts and ends
5125 const intptr_t BIG_NEG = -128;
5126 assert(BIG_NEG + 2*abase < 0, "neg enough");
5127 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5128 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5129 if (slice_len_con == 0) {
5130 return; // nothing to do here
5131 }
5132 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5133 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5134 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5135 assert(end_con < 0, "not two cons");
5136 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5137 BytesPerLong);
5138 }
5140 if (start_con >= 0 && end_con >= 0) {
5141 // Constant start and end. Simple.
5142 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5143 start_con, end_con, &_gvn);
5144 } else if (start_con >= 0 && dest_size != top()) {
5145 // Constant start, pre-rounded end after the tail of the array.
5146 Node* end = dest_size;
5147 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5148 start_con, end, &_gvn);
5149 } else if (start_con >= 0 && slice_len != top()) {
5150 // Constant start, non-constant end. End needs rounding up.
5151 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5152 intptr_t end_base = abase + (slice_idx_con << scale);
5153 int end_round = (-1 << scale) & (BytesPerLong - 1);
5154 Node* end = ConvI2X(slice_len);
5155 if (scale != 0)
5156 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5157 end_base += end_round;
5158 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5159 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5160 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5161 start_con, end, &_gvn);
5162 } else if (start_con < 0 && dest_size != top()) {
5163 // Non-constant start, pre-rounded end after the tail of the array.
5164 // This is almost certainly a "round-to-end" operation.
5165 Node* start = slice_idx;
5166 start = ConvI2X(start);
5167 if (scale != 0)
5168 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5169 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5170 if ((bump_bit | clear_low) != 0) {
5171 int to_clear = (bump_bit | clear_low);
5172 // Align up mod 8, then store a jint zero unconditionally
5173 // just before the mod-8 boundary.
5174 if (((abase + bump_bit) & ~to_clear) - bump_bit
5175 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5176 bump_bit = 0;
5177 assert((abase & to_clear) == 0, "array base must be long-aligned");
5178 } else {
5179 // Bump 'start' up to (or past) the next jint boundary:
5180 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5181 assert((abase & clear_low) == 0, "array base must be int-aligned");
5182 }
5183 // Round bumped 'start' down to jlong boundary in body of array.
5184 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5185 if (bump_bit != 0) {
5186 // Store a zero to the immediately preceding jint:
5187 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5188 Node* p1 = basic_plus_adr(dest, x1);
5189 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5190 mem = _gvn.transform(mem);
5191 }
5192 }
5193 Node* end = dest_size; // pre-rounded
5194 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5195 start, end, &_gvn);
5196 } else {
5197 // Non-constant start, unrounded non-constant end.
5198 // (Nobody zeroes a random midsection of an array using this routine.)
5199 ShouldNotReachHere(); // fix caller
5200 }
5202 // Done.
5203 set_memory(mem, adr_type);
5204 }
5207 bool
5208 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5209 BasicType basic_elem_type,
5210 AllocateNode* alloc,
5211 Node* src, Node* src_offset,
5212 Node* dest, Node* dest_offset,
5213 Node* dest_size, bool dest_uninitialized) {
5214 // See if there is an advantage from block transfer.
5215 int scale = exact_log2(type2aelembytes(basic_elem_type));
5216 if (scale >= LogBytesPerLong)
5217 return false; // it is already a block transfer
5219 // Look at the alignment of the starting offsets.
5220 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5222 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5223 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5224 if (src_off_con < 0 || dest_off_con < 0)
5225 // At present, we can only understand constants.
5226 return false;
5228 intptr_t src_off = abase + (src_off_con << scale);
5229 intptr_t dest_off = abase + (dest_off_con << scale);
5231 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5232 // Non-aligned; too bad.
5233 // One more chance: Pick off an initial 32-bit word.
5234 // This is a common case, since abase can be odd mod 8.
5235 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5236 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5237 Node* sptr = basic_plus_adr(src, src_off);
5238 Node* dptr = basic_plus_adr(dest, dest_off);
5239 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5240 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5241 src_off += BytesPerInt;
5242 dest_off += BytesPerInt;
5243 } else {
5244 return false;
5245 }
5246 }
5247 assert(src_off % BytesPerLong == 0, "");
5248 assert(dest_off % BytesPerLong == 0, "");
5250 // Do this copy by giant steps.
5251 Node* sptr = basic_plus_adr(src, src_off);
5252 Node* dptr = basic_plus_adr(dest, dest_off);
5253 Node* countx = dest_size;
5254 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5255 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5257 bool disjoint_bases = true; // since alloc != NULL
5258 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5259 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5261 return true;
5262 }
5265 // Helper function; generates code for the slow case.
5266 // We make a call to a runtime method which emulates the native method,
5267 // but without the native wrapper overhead.
5268 void
5269 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5270 Node* src, Node* src_offset,
5271 Node* dest, Node* dest_offset,
5272 Node* copy_length, bool dest_uninitialized) {
5273 assert(!dest_uninitialized, "Invariant");
5274 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5275 OptoRuntime::slow_arraycopy_Type(),
5276 OptoRuntime::slow_arraycopy_Java(),
5277 "slow_arraycopy", adr_type,
5278 src, src_offset, dest, dest_offset,
5279 copy_length);
5281 // Handle exceptions thrown by this fellow:
5282 make_slow_call_ex(call, env()->Throwable_klass(), false);
5283 }
5285 // Helper function; generates code for cases requiring runtime checks.
5286 Node*
5287 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5288 Node* dest_elem_klass,
5289 Node* src, Node* src_offset,
5290 Node* dest, Node* dest_offset,
5291 Node* copy_length, bool dest_uninitialized) {
5292 if (stopped()) return NULL;
5294 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5295 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5296 return NULL;
5297 }
5299 // Pick out the parameters required to perform a store-check
5300 // for the target array. This is an optimistic check. It will
5301 // look in each non-null element's class, at the desired klass's
5302 // super_check_offset, for the desired klass.
5303 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5304 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5305 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5306 Node* check_offset = ConvI2X(_gvn.transform(n3));
5307 Node* check_value = dest_elem_klass;
5309 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5310 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5312 // (We know the arrays are never conjoint, because their types differ.)
5313 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5314 OptoRuntime::checkcast_arraycopy_Type(),
5315 copyfunc_addr, "checkcast_arraycopy", adr_type,
5316 // five arguments, of which two are
5317 // intptr_t (jlong in LP64)
5318 src_start, dest_start,
5319 copy_length XTOP,
5320 check_offset XTOP,
5321 check_value);
5323 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5324 }
5327 // Helper function; generates code for cases requiring runtime checks.
5328 Node*
5329 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5330 Node* src, Node* src_offset,
5331 Node* dest, Node* dest_offset,
5332 Node* copy_length, bool dest_uninitialized) {
5333 assert(!dest_uninitialized, "Invariant");
5334 if (stopped()) return NULL;
5335 address copyfunc_addr = StubRoutines::generic_arraycopy();
5336 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5337 return NULL;
5338 }
5340 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5341 OptoRuntime::generic_arraycopy_Type(),
5342 copyfunc_addr, "generic_arraycopy", adr_type,
5343 src, src_offset, dest, dest_offset, copy_length);
5345 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5346 }
5348 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5349 void
5350 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5351 BasicType basic_elem_type,
5352 bool disjoint_bases,
5353 Node* src, Node* src_offset,
5354 Node* dest, Node* dest_offset,
5355 Node* copy_length, bool dest_uninitialized) {
5356 if (stopped()) return; // nothing to do
5358 Node* src_start = src;
5359 Node* dest_start = dest;
5360 if (src_offset != NULL || dest_offset != NULL) {
5361 assert(src_offset != NULL && dest_offset != NULL, "");
5362 src_start = array_element_address(src, src_offset, basic_elem_type);
5363 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5364 }
5366 // Figure out which arraycopy runtime method to call.
5367 const char* copyfunc_name = "arraycopy";
5368 address copyfunc_addr =
5369 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5370 disjoint_bases, copyfunc_name, dest_uninitialized);
5372 // Call it. Note that the count_ix value is not scaled to a byte-size.
5373 make_runtime_call(RC_LEAF|RC_NO_FP,
5374 OptoRuntime::fast_arraycopy_Type(),
5375 copyfunc_addr, copyfunc_name, adr_type,
5376 src_start, dest_start, copy_length XTOP);
5377 }
5379 //-------------inline_encodeISOArray-----------------------------------
5380 // encode char[] to byte[] in ISO_8859_1
5381 bool LibraryCallKit::inline_encodeISOArray() {
5382 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5383 // no receiver since it is static method
5384 Node *src = argument(0);
5385 Node *src_offset = argument(1);
5386 Node *dst = argument(2);
5387 Node *dst_offset = argument(3);
5388 Node *length = argument(4);
5390 const Type* src_type = src->Value(&_gvn);
5391 const Type* dst_type = dst->Value(&_gvn);
5392 const TypeAryPtr* top_src = src_type->isa_aryptr();
5393 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5394 if (top_src == NULL || top_src->klass() == NULL ||
5395 top_dest == NULL || top_dest->klass() == NULL) {
5396 // failed array check
5397 return false;
5398 }
5400 // Figure out the size and type of the elements we will be copying.
5401 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5402 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5403 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5404 return false;
5405 }
5406 Node* src_start = array_element_address(src, src_offset, src_elem);
5407 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5408 // 'src_start' points to src array + scaled offset
5409 // 'dst_start' points to dst array + scaled offset
5411 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5412 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5413 enc = _gvn.transform(enc);
5414 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5415 set_memory(res_mem, mtype);
5416 set_result(enc);
5417 return true;
5418 }
5420 //----------------------------inline_reference_get----------------------------
5421 // public T java.lang.ref.Reference.get();
5422 bool LibraryCallKit::inline_reference_get() {
5423 const int referent_offset = java_lang_ref_Reference::referent_offset;
5424 guarantee(referent_offset > 0, "should have already been set");
5426 // Get the argument:
5427 Node* reference_obj = null_check_receiver();
5428 if (stopped()) return true;
5430 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5432 ciInstanceKlass* klass = env()->Object_klass();
5433 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5435 Node* no_ctrl = NULL;
5436 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5438 // Use the pre-barrier to record the value in the referent field
5439 pre_barrier(false /* do_load */,
5440 control(),
5441 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5442 result /* pre_val */,
5443 T_OBJECT);
5445 // Add memory barrier to prevent commoning reads from this field
5446 // across safepoint since GC can change its value.
5447 insert_mem_bar(Op_MemBarCPUOrder);
5449 set_result(result);
5450 return true;
5451 }
5454 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5455 bool is_exact=true, bool is_static=false) {
5457 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5458 assert(tinst != NULL, "obj is null");
5459 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5460 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5462 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5463 ciSymbol::make(fieldTypeString),
5464 is_static);
5465 if (field == NULL) return (Node *) NULL;
5466 assert (field != NULL, "undefined field");
5468 // Next code copied from Parse::do_get_xxx():
5470 // Compute address and memory type.
5471 int offset = field->offset_in_bytes();
5472 bool is_vol = field->is_volatile();
5473 ciType* field_klass = field->type();
5474 assert(field_klass->is_loaded(), "should be loaded");
5475 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5476 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5477 BasicType bt = field->layout_type();
5479 // Build the resultant type of the load
5480 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5482 // Build the load.
5483 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5484 return loadedField;
5485 }
5488 //------------------------------inline_aescrypt_Block-----------------------
5489 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5490 address stubAddr;
5491 const char *stubName;
5492 assert(UseAES, "need AES instruction support");
5494 switch(id) {
5495 case vmIntrinsics::_aescrypt_encryptBlock:
5496 stubAddr = StubRoutines::aescrypt_encryptBlock();
5497 stubName = "aescrypt_encryptBlock";
5498 break;
5499 case vmIntrinsics::_aescrypt_decryptBlock:
5500 stubAddr = StubRoutines::aescrypt_decryptBlock();
5501 stubName = "aescrypt_decryptBlock";
5502 break;
5503 }
5504 if (stubAddr == NULL) return false;
5506 Node* aescrypt_object = argument(0);
5507 Node* src = argument(1);
5508 Node* src_offset = argument(2);
5509 Node* dest = argument(3);
5510 Node* dest_offset = argument(4);
5512 // (1) src and dest are arrays.
5513 const Type* src_type = src->Value(&_gvn);
5514 const Type* dest_type = dest->Value(&_gvn);
5515 const TypeAryPtr* top_src = src_type->isa_aryptr();
5516 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5517 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5519 // for the quick and dirty code we will skip all the checks.
5520 // we are just trying to get the call to be generated.
5521 Node* src_start = src;
5522 Node* dest_start = dest;
5523 if (src_offset != NULL || dest_offset != NULL) {
5524 assert(src_offset != NULL && dest_offset != NULL, "");
5525 src_start = array_element_address(src, src_offset, T_BYTE);
5526 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5527 }
5529 // now need to get the start of its expanded key array
5530 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5531 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5532 if (k_start == NULL) return false;
5534 // Call the stub.
5535 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5536 stubAddr, stubName, TypePtr::BOTTOM,
5537 src_start, dest_start, k_start);
5539 return true;
5540 }
5542 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5543 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5544 address stubAddr;
5545 const char *stubName;
5547 assert(UseAES, "need AES instruction support");
5549 switch(id) {
5550 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5551 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5552 stubName = "cipherBlockChaining_encryptAESCrypt";
5553 break;
5554 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5555 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5556 stubName = "cipherBlockChaining_decryptAESCrypt";
5557 break;
5558 }
5559 if (stubAddr == NULL) return false;
5561 Node* cipherBlockChaining_object = argument(0);
5562 Node* src = argument(1);
5563 Node* src_offset = argument(2);
5564 Node* len = argument(3);
5565 Node* dest = argument(4);
5566 Node* dest_offset = argument(5);
5568 // (1) src and dest are arrays.
5569 const Type* src_type = src->Value(&_gvn);
5570 const Type* dest_type = dest->Value(&_gvn);
5571 const TypeAryPtr* top_src = src_type->isa_aryptr();
5572 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5573 assert (top_src != NULL && top_src->klass() != NULL
5574 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5576 // checks are the responsibility of the caller
5577 Node* src_start = src;
5578 Node* dest_start = dest;
5579 if (src_offset != NULL || dest_offset != NULL) {
5580 assert(src_offset != NULL && dest_offset != NULL, "");
5581 src_start = array_element_address(src, src_offset, T_BYTE);
5582 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5583 }
5585 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5586 // (because of the predicated logic executed earlier).
5587 // so we cast it here safely.
5588 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5590 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5591 if (embeddedCipherObj == NULL) return false;
5593 // cast it to what we know it will be at runtime
5594 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5595 assert(tinst != NULL, "CBC obj is null");
5596 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5597 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5598 if (!klass_AESCrypt->is_loaded()) return false;
5600 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5601 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5602 const TypeOopPtr* xtype = aklass->as_instance_type();
5603 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5604 aescrypt_object = _gvn.transform(aescrypt_object);
5606 // we need to get the start of the aescrypt_object's expanded key array
5607 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5608 if (k_start == NULL) return false;
5610 // similarly, get the start address of the r vector
5611 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5612 if (objRvec == NULL) return false;
5613 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5615 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5616 make_runtime_call(RC_LEAF|RC_NO_FP,
5617 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5618 stubAddr, stubName, TypePtr::BOTTOM,
5619 src_start, dest_start, k_start, r_start, len);
5621 // return is void so no result needs to be pushed
5623 return true;
5624 }
5626 //------------------------------get_key_start_from_aescrypt_object-----------------------
5627 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5628 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5629 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5630 if (objAESCryptKey == NULL) return (Node *) NULL;
5632 // now have the array, need to get the start address of the K array
5633 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5634 return k_start;
5635 }
5637 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5638 // Return node representing slow path of predicate check.
5639 // the pseudo code we want to emulate with this predicate is:
5640 // for encryption:
5641 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5642 // for decryption:
5643 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5644 // note cipher==plain is more conservative than the original java code but that's OK
5645 //
5646 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5647 // First, check receiver for NULL since it is virtual method.
5648 Node* objCBC = argument(0);
5649 objCBC = null_check(objCBC);
5651 if (stopped()) return NULL; // Always NULL
5653 // Load embeddedCipher field of CipherBlockChaining object.
5654 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5656 // get AESCrypt klass for instanceOf check
5657 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5658 // will have same classloader as CipherBlockChaining object
5659 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5660 assert(tinst != NULL, "CBCobj is null");
5661 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5663 // we want to do an instanceof comparison against the AESCrypt class
5664 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5665 if (!klass_AESCrypt->is_loaded()) {
5666 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5667 Node* ctrl = control();
5668 set_control(top()); // no regular fast path
5669 return ctrl;
5670 }
5671 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5673 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5674 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5675 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5677 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5679 // for encryption, we are done
5680 if (!decrypting)
5681 return instof_false; // even if it is NULL
5683 // for decryption, we need to add a further check to avoid
5684 // taking the intrinsic path when cipher and plain are the same
5685 // see the original java code for why.
5686 RegionNode* region = new(C) RegionNode(3);
5687 region->init_req(1, instof_false);
5688 Node* src = argument(1);
5689 Node* dest = argument(4);
5690 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5691 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5692 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5693 region->init_req(2, src_dest_conjoint);
5695 record_for_igvn(region);
5696 return _gvn.transform(region);
5697 }