1 /*
2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "ci/ciUtilities.inline.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "classfile/vmSymbols.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/c2compiler.hpp"
37 #include "opto/callGenerator.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/convertnode.hpp"
41 #include "opto/countbitsnode.hpp"
42 #include "opto/intrinsicnode.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/mathexactnode.hpp"
45 #include "opto/movenode.hpp"
46 #include "opto/mulnode.hpp"
47 #include "opto/narrowptrnode.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/parse.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/rootnode.hpp"
52 #include "opto/subnode.hpp"
53 #include "prims/nativeLookup.hpp"
54 #include "prims/unsafe.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #ifdef TRACE_HAVE_INTRINSICS
57 #include "trace/traceMacros.hpp"
58 #endif
59
60 class LibraryIntrinsic : public InlineCallGenerator {
61 // Extend the set of intrinsics known to the runtime:
62 public:
63 private:
64 bool _is_virtual;
65 bool _does_virtual_dispatch;
66 int8_t _predicates_count; // Intrinsic is predicated by several conditions
67 int8_t _last_predicate; // Last generated predicate
68 vmIntrinsics::ID _intrinsic_id;
69
70 public:
71 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
72 : InlineCallGenerator(m),
73 _is_virtual(is_virtual),
74 _does_virtual_dispatch(does_virtual_dispatch),
75 _predicates_count((int8_t)predicates_count),
76 _last_predicate((int8_t)-1),
77 _intrinsic_id(id)
78 {
79 }
80 virtual bool is_intrinsic() const { return true; }
81 virtual bool is_virtual() const { return _is_virtual; }
82 virtual bool is_predicated() const { return _predicates_count > 0; }
83 virtual int predicates_count() const { return _predicates_count; }
84 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
85 virtual JVMState* generate(JVMState* jvms);
86 virtual Node* generate_predicate(JVMState* jvms, int predicate);
87 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
88 };
89
90
91 // Local helper class for LibraryIntrinsic:
92 class LibraryCallKit : public GraphKit {
93 private:
94 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
95 Node* _result; // the result node, if any
96 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
97
98 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
99
100 public:
101 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
102 : GraphKit(jvms),
103 _intrinsic(intrinsic),
104 _result(NULL)
105 {
106 // Check if this is a root compile. In that case we don't have a caller.
107 if (!jvms->has_method()) {
108 _reexecute_sp = sp();
109 } else {
110 // Find out how many arguments the interpreter needs when deoptimizing
111 // and save the stack pointer value so it can used by uncommon_trap.
112 // We find the argument count by looking at the declared signature.
113 bool ignored_will_link;
114 ciSignature* declared_signature = NULL;
115 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
116 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
117 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
118 }
119 }
120
121 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
122
123 ciMethod* caller() const { return jvms()->method(); }
124 int bci() const { return jvms()->bci(); }
125 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
126 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
127 ciMethod* callee() const { return _intrinsic->method(); }
128
129 bool try_to_inline(int predicate);
130 Node* try_to_predicate(int predicate);
131
132 void push_result() {
133 // Push the result onto the stack.
134 if (!stopped() && result() != NULL) {
135 BasicType bt = result()->bottom_type()->basic_type();
136 push_node(bt, result());
137 }
138 }
139
140 private:
141 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
142 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
143 }
144
145 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
146 void set_result(RegionNode* region, PhiNode* value);
147 Node* result() { return _result; }
148
149 virtual int reexecute_sp() { return _reexecute_sp; }
150
151 // Helper functions to inline natives
152 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
153 Node* generate_slow_guard(Node* test, RegionNode* region);
154 Node* generate_fair_guard(Node* test, RegionNode* region);
155 Node* generate_negative_guard(Node* index, RegionNode* region,
156 // resulting CastII of index:
157 Node* *pos_index = NULL);
158 Node* generate_limit_guard(Node* offset, Node* subseq_length,
159 Node* array_length,
160 RegionNode* region);
161 void generate_string_range_check(Node* array, Node* offset,
162 Node* length, bool char_count);
163 Node* generate_current_thread(Node* &tls_output);
164 Node* load_mirror_from_klass(Node* klass);
165 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
166 RegionNode* region, int null_path,
167 int offset);
168 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
169 RegionNode* region, int null_path) {
170 int offset = java_lang_Class::klass_offset_in_bytes();
171 return load_klass_from_mirror_common(mirror, never_see_null,
172 region, null_path,
173 offset);
174 }
175 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
176 RegionNode* region, int null_path) {
177 int offset = java_lang_Class::array_klass_offset_in_bytes();
178 return load_klass_from_mirror_common(mirror, never_see_null,
179 region, null_path,
180 offset);
181 }
182 Node* generate_access_flags_guard(Node* kls,
183 int modifier_mask, int modifier_bits,
184 RegionNode* region);
185 Node* generate_interface_guard(Node* kls, RegionNode* region);
186 Node* generate_array_guard(Node* kls, RegionNode* region) {
187 return generate_array_guard_common(kls, region, false, false);
188 }
189 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
190 return generate_array_guard_common(kls, region, false, true);
191 }
192 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
193 return generate_array_guard_common(kls, region, true, false);
194 }
195 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
196 return generate_array_guard_common(kls, region, true, true);
197 }
198 Node* generate_array_guard_common(Node* kls, RegionNode* region,
199 bool obj_array, bool not_array);
200 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
201 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
202 bool is_virtual = false, bool is_static = false);
203 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
204 return generate_method_call(method_id, false, true);
205 }
206 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
207 return generate_method_call(method_id, true, false);
208 }
209 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
210 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
211
212 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
215 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
216 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
217 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_indexOfChar();
219 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
220 bool inline_string_toBytesU();
221 bool inline_string_getCharsU();
222 bool inline_string_copy(bool compress);
223 bool inline_string_char_access(bool is_store);
224 Node* round_double_node(Node* n);
225 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
226 bool inline_math_native(vmIntrinsics::ID id);
227 bool inline_math(vmIntrinsics::ID id);
228 template <typename OverflowOp>
229 bool inline_math_overflow(Node* arg1, Node* arg2);
230 void inline_math_mathExact(Node* math, Node* test);
231 bool inline_math_addExactI(bool is_increment);
232 bool inline_math_addExactL(bool is_increment);
233 bool inline_math_multiplyExactI();
234 bool inline_math_multiplyExactL();
235 bool inline_math_multiplyHigh();
236 bool inline_math_negateExactI();
237 bool inline_math_negateExactL();
238 bool inline_math_subtractExactI(bool is_decrement);
239 bool inline_math_subtractExactL(bool is_decrement);
240 bool inline_min_max(vmIntrinsics::ID id);
241 bool inline_notify(vmIntrinsics::ID id);
242 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
243 // This returns Type::AnyPtr, RawPtr, or OopPtr.
244 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
245 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
246 // Helper for inline_unsafe_access.
247 // Generates the guards that check whether the result of
248 // Unsafe.getObject should be recorded in an SATB log buffer.
249 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
250
251 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
252 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
253 static bool klass_needs_init_guard(Node* kls);
254 bool inline_unsafe_allocate();
255 bool inline_unsafe_newArray(bool uninitialized);
256 bool inline_unsafe_copyMemory();
257 bool inline_native_currentThread();
258
259 bool inline_native_time_funcs(address method, const char* funcName);
260 #ifdef TRACE_HAVE_INTRINSICS
261 bool inline_native_classID();
262 bool inline_native_getBufferWriter();
263 #endif
264 bool inline_native_isInterrupted();
265 bool inline_native_Class_query(vmIntrinsics::ID id);
266 bool inline_native_subtype_check();
267 bool inline_native_getLength();
268 bool inline_array_copyOf(bool is_copyOfRange);
269 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
270 bool inline_preconditions_checkIndex();
271 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
272 bool inline_native_clone(bool is_virtual);
273 bool inline_native_Reflection_getCallerClass();
274 // Helper function for inlining native object hash method
275 bool inline_native_hashcode(bool is_virtual, bool is_static);
276 bool inline_native_getClass();
277
278 // Helper functions for inlining arraycopy
279 bool inline_arraycopy();
280 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
281 RegionNode* slow_region);
282 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
283 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
284 uint new_idx);
285
286 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
287 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
288 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
289 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
290 bool inline_unsafe_fence(vmIntrinsics::ID id);
291 bool inline_onspinwait();
292 bool inline_fp_conversions(vmIntrinsics::ID id);
293 bool inline_number_methods(vmIntrinsics::ID id);
294 bool inline_reference_get();
295 bool inline_Class_cast();
296 bool inline_aescrypt_Block(vmIntrinsics::ID id);
297 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
298 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
299 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
300 Node* inline_counterMode_AESCrypt_predicate();
301 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
302 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
303 bool inline_ghash_processBlocks();
304 bool inline_sha_implCompress(vmIntrinsics::ID id);
305 bool inline_digestBase_implCompressMB(int predicate);
306 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
307 bool long_state, address stubAddr, const char *stubName,
308 Node* src_start, Node* ofs, Node* limit);
309 Node* get_state_from_sha_object(Node *sha_object);
310 Node* get_state_from_sha5_object(Node *sha_object);
311 Node* inline_digestBase_implCompressMB_predicate(int predicate);
312 bool inline_encodeISOArray();
313 bool inline_updateCRC32();
314 bool inline_updateBytesCRC32();
315 bool inline_updateByteBufferCRC32();
316 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
317 bool inline_updateBytesCRC32C();
318 bool inline_updateDirectByteBufferCRC32C();
319 bool inline_updateBytesAdler32();
320 bool inline_updateByteBufferAdler32();
321 bool inline_multiplyToLen();
322 bool inline_hasNegatives();
323 bool inline_squareToLen();
324 bool inline_mulAdd();
325 bool inline_montgomeryMultiply();
326 bool inline_montgomerySquare();
327 bool inline_vectorizedMismatch();
328 bool inline_fma(vmIntrinsics::ID id);
329
330 bool inline_profileBoolean();
331 bool inline_isCompileConstant();
332 void clear_upper_avx() {
333 #ifdef X86
334 if (UseAVX >= 2) {
335 C->set_clear_upper_avx(true);
336 }
337 #endif
338 }
339 };
340
341 //---------------------------make_vm_intrinsic----------------------------
342 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
343 vmIntrinsics::ID id = m->intrinsic_id();
344 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
345
346 if (!m->is_loaded()) {
347 // Do not attempt to inline unloaded methods.
348 return NULL;
349 }
350
351 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
352 bool is_available = false;
353
354 {
355 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
356 // the compiler must transition to '_thread_in_vm' state because both
357 // methods access VM-internal data.
358 VM_ENTRY_MARK;
359 methodHandle mh(THREAD, m->get_Method());
360 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
361 !C->directive()->is_intrinsic_disabled(mh) &&
362 !vmIntrinsics::is_disabled_by_flags(mh);
363
364 }
365
366 if (is_available) {
367 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
368 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
369 return new LibraryIntrinsic(m, is_virtual,
370 vmIntrinsics::predicates_needed(id),
371 vmIntrinsics::does_virtual_dispatch(id),
372 (vmIntrinsics::ID) id);
373 } else {
374 return NULL;
375 }
376 }
377
378 //----------------------register_library_intrinsics-----------------------
379 // Initialize this file's data structures, for each Compile instance.
380 void Compile::register_library_intrinsics() {
381 // Nothing to do here.
382 }
383
384 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
385 LibraryCallKit kit(jvms, this);
386 Compile* C = kit.C;
387 int nodes = C->unique();
388 #ifndef PRODUCT
389 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
390 char buf[1000];
391 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
392 tty->print_cr("Intrinsic %s", str);
393 }
394 #endif
395 ciMethod* callee = kit.callee();
396 const int bci = kit.bci();
397
398 // Try to inline the intrinsic.
399 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
400 kit.try_to_inline(_last_predicate)) {
401 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
402 : "(intrinsic)";
403 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
404 if (C->print_intrinsics() || C->print_inlining()) {
405 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
406 }
407 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
408 if (C->log()) {
409 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
410 vmIntrinsics::name_at(intrinsic_id()),
411 (is_virtual() ? " virtual='1'" : ""),
412 C->unique() - nodes);
413 }
414 // Push the result from the inlined method onto the stack.
415 kit.push_result();
416 C->print_inlining_update(this);
417 return kit.transfer_exceptions_into_jvms();
418 }
419
420 // The intrinsic bailed out
421 if (jvms->has_method()) {
422 // Not a root compile.
423 const char* msg;
424 if (callee->intrinsic_candidate()) {
425 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
426 } else {
427 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
428 : "failed to inline (intrinsic), method not annotated";
429 }
430 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
431 if (C->print_intrinsics() || C->print_inlining()) {
432 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
433 }
434 } else {
435 // Root compile
436 ResourceMark rm;
437 stringStream msg_stream;
438 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
439 vmIntrinsics::name_at(intrinsic_id()),
440 is_virtual() ? " (virtual)" : "", bci);
441 const char *msg = msg_stream.as_string();
442 log_debug(jit, inlining)("%s", msg);
443 if (C->print_intrinsics() || C->print_inlining()) {
444 tty->print("%s", msg);
445 }
446 }
447 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
448 C->print_inlining_update(this);
449 return NULL;
450 }
451
452 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
453 LibraryCallKit kit(jvms, this);
454 Compile* C = kit.C;
455 int nodes = C->unique();
456 _last_predicate = predicate;
457 #ifndef PRODUCT
458 assert(is_predicated() && predicate < predicates_count(), "sanity");
459 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
460 char buf[1000];
461 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
462 tty->print_cr("Predicate for intrinsic %s", str);
463 }
464 #endif
465 ciMethod* callee = kit.callee();
466 const int bci = kit.bci();
467
468 Node* slow_ctl = kit.try_to_predicate(predicate);
469 if (!kit.failing()) {
470 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
471 : "(intrinsic, predicate)";
472 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
473 if (C->print_intrinsics() || C->print_inlining()) {
474 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
475 }
476 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
477 if (C->log()) {
478 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
479 vmIntrinsics::name_at(intrinsic_id()),
480 (is_virtual() ? " virtual='1'" : ""),
481 C->unique() - nodes);
482 }
483 return slow_ctl; // Could be NULL if the check folds.
484 }
485
486 // The intrinsic bailed out
487 if (jvms->has_method()) {
488 // Not a root compile.
489 const char* msg = "failed to generate predicate for intrinsic";
490 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
491 if (C->print_intrinsics() || C->print_inlining()) {
492 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
493 }
494 } else {
495 // Root compile
496 ResourceMark rm;
497 stringStream msg_stream;
498 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
499 vmIntrinsics::name_at(intrinsic_id()),
500 is_virtual() ? " (virtual)" : "", bci);
501 const char *msg = msg_stream.as_string();
502 log_debug(jit, inlining)("%s", msg);
503 if (C->print_intrinsics() || C->print_inlining()) {
504 C->print_inlining_stream()->print("%s", msg);
505 }
506 }
507 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
508 return NULL;
509 }
510
511 bool LibraryCallKit::try_to_inline(int predicate) {
512 // Handle symbolic names for otherwise undistinguished boolean switches:
513 const bool is_store = true;
514 const bool is_compress = true;
515 const bool is_static = true;
516 const bool is_volatile = true;
517
518 if (!jvms()->has_method()) {
519 // Root JVMState has a null method.
520 assert(map()->memory()->Opcode() == Op_Parm, "");
521 // Insert the memory aliasing node
522 set_all_memory(reset_memory());
523 }
524 assert(merged_memory(), "");
525
526
527 switch (intrinsic_id()) {
528 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
529 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
530 case vmIntrinsics::_getClass: return inline_native_getClass();
531
532 case vmIntrinsics::_dsin:
533 case vmIntrinsics::_dcos:
534 case vmIntrinsics::_dtan:
535 case vmIntrinsics::_dabs:
536 case vmIntrinsics::_datan2:
537 case vmIntrinsics::_dsqrt:
538 case vmIntrinsics::_dexp:
539 case vmIntrinsics::_dlog:
540 case vmIntrinsics::_dlog10:
541 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
542
543 case vmIntrinsics::_min:
544 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
545
546 case vmIntrinsics::_notify:
547 case vmIntrinsics::_notifyAll:
548 if (ObjectMonitor::Knob_InlineNotify) {
549 return inline_notify(intrinsic_id());
550 }
551 return false;
552
553 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
554 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
555 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
556 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
557 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
558 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
559 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
560 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
561 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
562 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
563 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
564 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
565 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
566
567 case vmIntrinsics::_arraycopy: return inline_arraycopy();
568
569 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
570 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
571 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
572 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
573
574 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
575 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
576 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
577 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
578 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
579 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
580 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
581
582 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
583 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
584
585 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
586 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
587 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
588 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
589
590 case vmIntrinsics::_compressStringC:
591 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
592 case vmIntrinsics::_inflateStringC:
593 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
594
595 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
596 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
597 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
598 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
599 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
600 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
601 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
602 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
603 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
604
605 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
606 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
607 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
608 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
609 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
610 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
611 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
612 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
613 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
614
615 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
616 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
617 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
618 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
619 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
620 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
621 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
622 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
623 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
624
625 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
626 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
627 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
628 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
629 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
630 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
631 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
632 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
633 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
634
635 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
636 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
637 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
638 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
639
640 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
641 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
642 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
643 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
644
645 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
646 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
647 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
648 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
649 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
650 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
651 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
652 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
653 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
654
655 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
656 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
657 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
658 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
659 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
660 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
661 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
662 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
663 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
664
665 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
666 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
667 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
668 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
669 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
670 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
671 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
672 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
673 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
674
675 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
676 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
677 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
678 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
679 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
680 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
681 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
682 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
683 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
684
685 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
686 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
687 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
688 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
689 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
690
691 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
692 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
693 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
694 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
695 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
696 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
697 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
698 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
699 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
700 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
701 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
702 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
703 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
704 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
705 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
706 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
707 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
708 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
709 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
710 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
711
712 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
713 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
714 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
715 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
716 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
717 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
718 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
719 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
720 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
721 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
722 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
723 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
724 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
725 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
726 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
727
728 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
729 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
730 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
731 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
732
733 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
734 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
735 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
736 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
737 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
738
739 case vmIntrinsics::_loadFence:
740 case vmIntrinsics::_storeFence:
741 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
742
743 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
744
745 case vmIntrinsics::_currentThread: return inline_native_currentThread();
746 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
747
748 #ifdef TRACE_HAVE_INTRINSICS
749 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
750 case vmIntrinsics::_getClassId: return inline_native_classID();
751 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
752 #endif
753 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
754 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
755 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
756 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
757 case vmIntrinsics::_getLength: return inline_native_getLength();
758 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
759 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
760 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
761 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
762 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
763 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
764
765 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
766 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
767
768 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
769
770 case vmIntrinsics::_isInstance:
771 case vmIntrinsics::_getModifiers:
772 case vmIntrinsics::_isInterface:
773 case vmIntrinsics::_isArray:
774 case vmIntrinsics::_isPrimitive:
775 case vmIntrinsics::_getSuperclass:
776 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
777
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());
784
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());
795
796 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
797
798 case vmIntrinsics::_Reference_get: return inline_reference_get();
799
800 case vmIntrinsics::_Class_cast: return inline_Class_cast();
801
802 case vmIntrinsics::_aescrypt_encryptBlock:
803 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
804
805 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
806 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
807 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
808
809 case vmIntrinsics::_counterMode_AESCrypt:
810 return inline_counterMode_AESCrypt(intrinsic_id());
811
812 case vmIntrinsics::_sha_implCompress:
813 case vmIntrinsics::_sha2_implCompress:
814 case vmIntrinsics::_sha5_implCompress:
815 return inline_sha_implCompress(intrinsic_id());
816
817 case vmIntrinsics::_digestBase_implCompressMB:
818 return inline_digestBase_implCompressMB(predicate);
819
820 case vmIntrinsics::_multiplyToLen:
821 return inline_multiplyToLen();
822
823 case vmIntrinsics::_squareToLen:
824 return inline_squareToLen();
825
826 case vmIntrinsics::_mulAdd:
827 return inline_mulAdd();
828
829 case vmIntrinsics::_montgomeryMultiply:
830 return inline_montgomeryMultiply();
831 case vmIntrinsics::_montgomerySquare:
832 return inline_montgomerySquare();
833
834 case vmIntrinsics::_vectorizedMismatch:
835 return inline_vectorizedMismatch();
836
837 case vmIntrinsics::_ghash_processBlocks:
838 return inline_ghash_processBlocks();
839
840 case vmIntrinsics::_encodeISOArray:
841 case vmIntrinsics::_encodeByteISOArray:
842 return inline_encodeISOArray();
843
844 case vmIntrinsics::_updateCRC32:
845 return inline_updateCRC32();
846 case vmIntrinsics::_updateBytesCRC32:
847 return inline_updateBytesCRC32();
848 case vmIntrinsics::_updateByteBufferCRC32:
849 return inline_updateByteBufferCRC32();
850
851 case vmIntrinsics::_updateBytesCRC32C:
852 return inline_updateBytesCRC32C();
853 case vmIntrinsics::_updateDirectByteBufferCRC32C:
854 return inline_updateDirectByteBufferCRC32C();
855
856 case vmIntrinsics::_updateBytesAdler32:
857 return inline_updateBytesAdler32();
858 case vmIntrinsics::_updateByteBufferAdler32:
859 return inline_updateByteBufferAdler32();
860
861 case vmIntrinsics::_profileBoolean:
862 return inline_profileBoolean();
863 case vmIntrinsics::_isCompileConstant:
864 return inline_isCompileConstant();
865
866 case vmIntrinsics::_hasNegatives:
867 return inline_hasNegatives();
868
869 case vmIntrinsics::_fmaD:
870 case vmIntrinsics::_fmaF:
871 return inline_fma(intrinsic_id());
872
873 default:
874 // If you get here, it may be that someone has added a new intrinsic
875 // to the list in vmSymbols.hpp without implementing it here.
876 #ifndef PRODUCT
877 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
878 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
879 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
880 }
881 #endif
882 return false;
883 }
884 }
885
886 Node* LibraryCallKit::try_to_predicate(int predicate) {
887 if (!jvms()->has_method()) {
888 // Root JVMState has a null method.
889 assert(map()->memory()->Opcode() == Op_Parm, "");
890 // Insert the memory aliasing node
891 set_all_memory(reset_memory());
892 }
893 assert(merged_memory(), "");
894
895 switch (intrinsic_id()) {
896 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
897 return inline_cipherBlockChaining_AESCrypt_predicate(false);
898 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
899 return inline_cipherBlockChaining_AESCrypt_predicate(true);
900 case vmIntrinsics::_counterMode_AESCrypt:
901 return inline_counterMode_AESCrypt_predicate();
902 case vmIntrinsics::_digestBase_implCompressMB:
903 return inline_digestBase_implCompressMB_predicate(predicate);
904
905 default:
906 // If you get here, it may be that someone has added a new intrinsic
907 // to the list in vmSymbols.hpp without implementing it here.
908 #ifndef PRODUCT
909 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
910 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
911 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
912 }
913 #endif
914 Node* slow_ctl = control();
915 set_control(top()); // No fast path instrinsic
916 return slow_ctl;
917 }
918 }
919
920 //------------------------------set_result-------------------------------
921 // Helper function for finishing intrinsics.
922 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
923 record_for_igvn(region);
924 set_control(_gvn.transform(region));
925 set_result( _gvn.transform(value));
926 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
927 }
928
929 //------------------------------generate_guard---------------------------
930 // Helper function for generating guarded fast-slow graph structures.
931 // The given 'test', if true, guards a slow path. If the test fails
932 // then a fast path can be taken. (We generally hope it fails.)
933 // In all cases, GraphKit::control() is updated to the fast path.
934 // The returned value represents the control for the slow path.
935 // The return value is never 'top'; it is either a valid control
936 // or NULL if it is obvious that the slow path can never be taken.
937 // Also, if region and the slow control are not NULL, the slow edge
938 // is appended to the region.
939 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
940 if (stopped()) {
941 // Already short circuited.
942 return NULL;
943 }
944
945 // Build an if node and its projections.
946 // If test is true we take the slow path, which we assume is uncommon.
947 if (_gvn.type(test) == TypeInt::ZERO) {
948 // The slow branch is never taken. No need to build this guard.
949 return NULL;
950 }
951
952 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
953
954 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
955 if (if_slow == top()) {
956 // The slow branch is never taken. No need to build this guard.
957 return NULL;
958 }
959
960 if (region != NULL)
961 region->add_req(if_slow);
962
963 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
964 set_control(if_fast);
965
966 return if_slow;
967 }
968
969 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
970 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
971 }
972 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
973 return generate_guard(test, region, PROB_FAIR);
974 }
975
976 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
977 Node* *pos_index) {
978 if (stopped())
979 return NULL; // already stopped
980 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
981 return NULL; // index is already adequately typed
982 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
983 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
984 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
985 if (is_neg != NULL && pos_index != NULL) {
986 // Emulate effect of Parse::adjust_map_after_if.
987 Node* ccast = new CastIINode(index, TypeInt::POS);
988 ccast->set_req(0, control());
989 (*pos_index) = _gvn.transform(ccast);
990 }
991 return is_neg;
992 }
993
994 // Make sure that 'position' is a valid limit index, in [0..length].
995 // There are two equivalent plans for checking this:
996 // A. (offset + copyLength) unsigned<= arrayLength
997 // B. offset <= (arrayLength - copyLength)
998 // We require that all of the values above, except for the sum and
999 // difference, are already known to be non-negative.
1000 // Plan A is robust in the face of overflow, if offset and copyLength
1001 // are both hugely positive.
1002 //
1003 // Plan B is less direct and intuitive, but it does not overflow at
1004 // all, since the difference of two non-negatives is always
1005 // representable. Whenever Java methods must perform the equivalent
1006 // check they generally use Plan B instead of Plan A.
1007 // For the moment we use Plan A.
1008 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1009 Node* subseq_length,
1010 Node* array_length,
1011 RegionNode* region) {
1012 if (stopped())
1013 return NULL; // already stopped
1014 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1015 if (zero_offset && subseq_length->eqv_uncast(array_length))
1016 return NULL; // common case of whole-array copy
1017 Node* last = subseq_length;
1018 if (!zero_offset) // last += offset
1019 last = _gvn.transform(new AddINode(last, offset));
1020 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1021 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1022 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1023 return is_over;
1024 }
1025
1026 // Emit range checks for the given String.value byte array
1027 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1028 if (stopped()) {
1029 return; // already stopped
1030 }
1031 RegionNode* bailout = new RegionNode(1);
1032 record_for_igvn(bailout);
1033 if (char_count) {
1034 // Convert char count to byte count
1035 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1036 }
1037
1038 // Offset and count must not be negative
1039 generate_negative_guard(offset, bailout);
1040 generate_negative_guard(count, bailout);
1041 // Offset + count must not exceed length of array
1042 generate_limit_guard(offset, count, load_array_length(array), bailout);
1043
1044 if (bailout->req() > 1) {
1045 PreserveJVMState pjvms(this);
1046 set_control(_gvn.transform(bailout));
1047 uncommon_trap(Deoptimization::Reason_intrinsic,
1048 Deoptimization::Action_maybe_recompile);
1049 }
1050 }
1051
1052 //--------------------------generate_current_thread--------------------
1053 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1054 ciKlass* thread_klass = env()->Thread_klass();
1055 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1056 Node* thread = _gvn.transform(new ThreadLocalNode());
1057 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1058 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1059 tls_output = thread;
1060 return threadObj;
1061 }
1062
1063
1064 //------------------------------make_string_method_node------------------------
1065 // Helper method for String intrinsic functions. This version is called with
1066 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1067 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1068 // containing the lengths of str1 and str2.
1069 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1070 Node* result = NULL;
1071 switch (opcode) {
1072 case Op_StrIndexOf:
1073 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1074 str1_start, cnt1, str2_start, cnt2, ae);
1075 break;
1076 case Op_StrComp:
1077 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1078 str1_start, cnt1, str2_start, cnt2, ae);
1079 break;
1080 case Op_StrEquals:
1081 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1082 // Use the constant length if there is one because optimized match rule may exist.
1083 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1084 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1085 break;
1086 default:
1087 ShouldNotReachHere();
1088 return NULL;
1089 }
1090
1091 // All these intrinsics have checks.
1092 C->set_has_split_ifs(true); // Has chance for split-if optimization
1093 clear_upper_avx();
1094
1095 return _gvn.transform(result);
1096 }
1097
1098 //------------------------------inline_string_compareTo------------------------
1099 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1100 Node* arg1 = argument(0);
1101 Node* arg2 = argument(1);
1102
1103 // Get start addr and length of first argument
1104 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1105 Node* arg1_cnt = load_array_length(arg1);
1106
1107 // Get start addr and length of second argument
1108 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1109 Node* arg2_cnt = load_array_length(arg2);
1110
1111 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1112 set_result(result);
1113 return true;
1114 }
1115
1116 //------------------------------inline_string_equals------------------------
1117 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1118 Node* arg1 = argument(0);
1119 Node* arg2 = argument(1);
1120
1121 // paths (plus control) merge
1122 RegionNode* region = new RegionNode(3);
1123 Node* phi = new PhiNode(region, TypeInt::BOOL);
1124
1125 if (!stopped()) {
1126 // Get start addr and length of first argument
1127 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1128 Node* arg1_cnt = load_array_length(arg1);
1129
1130 // Get start addr and length of second argument
1131 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1132 Node* arg2_cnt = load_array_length(arg2);
1133
1134 // Check for arg1_cnt != arg2_cnt
1135 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1136 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1137 Node* if_ne = generate_slow_guard(bol, NULL);
1138 if (if_ne != NULL) {
1139 phi->init_req(2, intcon(0));
1140 region->init_req(2, if_ne);
1141 }
1142
1143 // Check for count == 0 is done by assembler code for StrEquals.
1144
1145 if (!stopped()) {
1146 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1147 phi->init_req(1, equals);
1148 region->init_req(1, control());
1149 }
1150 }
1151
1152 // post merge
1153 set_control(_gvn.transform(region));
1154 record_for_igvn(region);
1155
1156 set_result(_gvn.transform(phi));
1157 return true;
1158 }
1159
1160 //------------------------------inline_array_equals----------------------------
1161 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1162 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1163 Node* arg1 = argument(0);
1164 Node* arg2 = argument(1);
1165
1166 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1167 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1168 clear_upper_avx();
1169
1170 return true;
1171 }
1172
1173 //------------------------------inline_hasNegatives------------------------------
1174 bool LibraryCallKit::inline_hasNegatives() {
1175 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1176 return false;
1177 }
1178
1179 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1180 // no receiver since it is static method
1181 Node* ba = argument(0);
1182 Node* offset = argument(1);
1183 Node* len = argument(2);
1184
1185 // Range checks
1186 generate_string_range_check(ba, offset, len, false);
1187 if (stopped()) {
1188 return true;
1189 }
1190 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1191 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1192 set_result(_gvn.transform(result));
1193 return true;
1194 }
1195
1196 bool LibraryCallKit::inline_preconditions_checkIndex() {
1197 Node* index = argument(0);
1198 Node* length = argument(1);
1199 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1200 return false;
1201 }
1202
1203 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1204 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1205
1206 {
1207 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1208 uncommon_trap(Deoptimization::Reason_intrinsic,
1209 Deoptimization::Action_make_not_entrant);
1210 }
1211
1212 if (stopped()) {
1213 return false;
1214 }
1215
1216 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1217 BoolTest::mask btest = BoolTest::lt;
1218 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1219 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1220 _gvn.set_type(rc, rc->Value(&_gvn));
1221 if (!rc_bool->is_Con()) {
1222 record_for_igvn(rc);
1223 }
1224 set_control(_gvn.transform(new IfTrueNode(rc)));
1225 {
1226 PreserveJVMState pjvms(this);
1227 set_control(_gvn.transform(new IfFalseNode(rc)));
1228 uncommon_trap(Deoptimization::Reason_range_check,
1229 Deoptimization::Action_make_not_entrant);
1230 }
1231
1232 if (stopped()) {
1233 return false;
1234 }
1235
1236 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1237 result->set_req(0, control());
1238 result = _gvn.transform(result);
1239 set_result(result);
1240 replace_in_map(index, result);
1241 clear_upper_avx();
1242 return true;
1243 }
1244
1245 //------------------------------inline_string_indexOf------------------------
1246 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1247 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1248 return false;
1249 }
1250 Node* src = argument(0);
1251 Node* tgt = argument(1);
1252
1253 // Make the merge point
1254 RegionNode* result_rgn = new RegionNode(4);
1255 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1256
1257 // Get start addr and length of source string
1258 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1259 Node* src_count = load_array_length(src);
1260
1261 // Get start addr and length of substring
1262 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1263 Node* tgt_count = load_array_length(tgt);
1264
1265 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1266 // Divide src size by 2 if String is UTF16 encoded
1267 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1268 }
1269 if (ae == StrIntrinsicNode::UU) {
1270 // Divide substring size by 2 if String is UTF16 encoded
1271 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1272 }
1273
1274 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1275 if (result != NULL) {
1276 result_phi->init_req(3, result);
1277 result_rgn->init_req(3, control());
1278 }
1279 set_control(_gvn.transform(result_rgn));
1280 record_for_igvn(result_rgn);
1281 set_result(_gvn.transform(result_phi));
1282
1283 return true;
1284 }
1285
1286 //-----------------------------inline_string_indexOf-----------------------
1287 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1288 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1289 return false;
1290 }
1291 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1292 return false;
1293 }
1294 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1295 Node* src = argument(0); // byte[]
1296 Node* src_count = argument(1); // char count
1297 Node* tgt = argument(2); // byte[]
1298 Node* tgt_count = argument(3); // char count
1299 Node* from_index = argument(4); // char index
1300
1301 // Multiply byte array index by 2 if String is UTF16 encoded
1302 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1303 src_count = _gvn.transform(new SubINode(src_count, from_index));
1304 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1305 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1306
1307 // Range checks
1308 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1309 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1310 if (stopped()) {
1311 return true;
1312 }
1313
1314 RegionNode* region = new RegionNode(5);
1315 Node* phi = new PhiNode(region, TypeInt::INT);
1316
1317 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1318 if (result != NULL) {
1319 // The result is index relative to from_index if substring was found, -1 otherwise.
1320 // Generate code which will fold into cmove.
1321 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1322 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1323
1324 Node* if_lt = generate_slow_guard(bol, NULL);
1325 if (if_lt != NULL) {
1326 // result == -1
1327 phi->init_req(3, result);
1328 region->init_req(3, if_lt);
1329 }
1330 if (!stopped()) {
1331 result = _gvn.transform(new AddINode(result, from_index));
1332 phi->init_req(4, result);
1333 region->init_req(4, control());
1334 }
1335 }
1336
1337 set_control(_gvn.transform(region));
1338 record_for_igvn(region);
1339 set_result(_gvn.transform(phi));
1340 clear_upper_avx();
1341
1342 return true;
1343 }
1344
1345 // Create StrIndexOfNode with fast path checks
1346 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1347 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1348 // Check for substr count > string count
1349 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1350 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1351 Node* if_gt = generate_slow_guard(bol, NULL);
1352 if (if_gt != NULL) {
1353 phi->init_req(1, intcon(-1));
1354 region->init_req(1, if_gt);
1355 }
1356 if (!stopped()) {
1357 // Check for substr count == 0
1358 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1359 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1360 Node* if_zero = generate_slow_guard(bol, NULL);
1361 if (if_zero != NULL) {
1362 phi->init_req(2, intcon(0));
1363 region->init_req(2, if_zero);
1364 }
1365 }
1366 if (!stopped()) {
1367 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1368 }
1369 return NULL;
1370 }
1371
1372 //-----------------------------inline_string_indexOfChar-----------------------
1373 bool LibraryCallKit::inline_string_indexOfChar() {
1374 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1375 return false;
1376 }
1377 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1378 return false;
1379 }
1380 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1381 Node* src = argument(0); // byte[]
1382 Node* tgt = argument(1); // tgt is int ch
1383 Node* from_index = argument(2);
1384 Node* max = argument(3);
1385
1386 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1387 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1388 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1389
1390 // Range checks
1391 generate_string_range_check(src, src_offset, src_count, true);
1392 if (stopped()) {
1393 return true;
1394 }
1395
1396 RegionNode* region = new RegionNode(3);
1397 Node* phi = new PhiNode(region, TypeInt::INT);
1398
1399 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1400 C->set_has_split_ifs(true); // Has chance for split-if optimization
1401 _gvn.transform(result);
1402
1403 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1404 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1405
1406 Node* if_lt = generate_slow_guard(bol, NULL);
1407 if (if_lt != NULL) {
1408 // result == -1
1409 phi->init_req(2, result);
1410 region->init_req(2, if_lt);
1411 }
1412 if (!stopped()) {
1413 result = _gvn.transform(new AddINode(result, from_index));
1414 phi->init_req(1, result);
1415 region->init_req(1, control());
1416 }
1417 set_control(_gvn.transform(region));
1418 record_for_igvn(region);
1419 set_result(_gvn.transform(phi));
1420
1421 return true;
1422 }
1423 //---------------------------inline_string_copy---------------------
1424 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1425 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1426 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1427 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1428 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1429 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1430 bool LibraryCallKit::inline_string_copy(bool compress) {
1431 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1432 return false;
1433 }
1434 int nargs = 5; // 2 oops, 3 ints
1435 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1436
1437 Node* src = argument(0);
1438 Node* src_offset = argument(1);
1439 Node* dst = argument(2);
1440 Node* dst_offset = argument(3);
1441 Node* length = argument(4);
1442
1443 // Check for allocation before we add nodes that would confuse
1444 // tightly_coupled_allocation()
1445 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1446
1447 // Figure out the size and type of the elements we will be copying.
1448 const Type* src_type = src->Value(&_gvn);
1449 const Type* dst_type = dst->Value(&_gvn);
1450 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1451 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1452 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1453 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1454 "Unsupported array types for inline_string_copy");
1455
1456 // Convert char[] offsets to byte[] offsets
1457 bool convert_src = (compress && src_elem == T_BYTE);
1458 bool convert_dst = (!compress && dst_elem == T_BYTE);
1459 if (convert_src) {
1460 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1461 } else if (convert_dst) {
1462 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1463 }
1464
1465 // Range checks
1466 generate_string_range_check(src, src_offset, length, convert_src);
1467 generate_string_range_check(dst, dst_offset, length, convert_dst);
1468 if (stopped()) {
1469 return true;
1470 }
1471
1472 Node* src_start = array_element_address(src, src_offset, src_elem);
1473 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1474 // 'src_start' points to src array + scaled offset
1475 // 'dst_start' points to dst array + scaled offset
1476 Node* count = NULL;
1477 if (compress) {
1478 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1479 } else {
1480 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1481 }
1482
1483 if (alloc != NULL) {
1484 if (alloc->maybe_set_complete(&_gvn)) {
1485 // "You break it, you buy it."
1486 InitializeNode* init = alloc->initialization();
1487 assert(init->is_complete(), "we just did this");
1488 init->set_complete_with_arraycopy();
1489 assert(dst->is_CheckCastPP(), "sanity");
1490 assert(dst->in(0)->in(0) == init, "dest pinned");
1491 }
1492 // Do not let stores that initialize this object be reordered with
1493 // a subsequent store that would make this object accessible by
1494 // other threads.
1495 // Record what AllocateNode this StoreStore protects so that
1496 // escape analysis can go from the MemBarStoreStoreNode to the
1497 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1498 // based on the escape status of the AllocateNode.
1499 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1500 }
1501 if (compress) {
1502 set_result(_gvn.transform(count));
1503 }
1504 clear_upper_avx();
1505
1506 return true;
1507 }
1508
1509 #ifdef _LP64
1510 #define XTOP ,top() /*additional argument*/
1511 #else //_LP64
1512 #define XTOP /*no additional argument*/
1513 #endif //_LP64
1514
1515 //------------------------inline_string_toBytesU--------------------------
1516 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1517 bool LibraryCallKit::inline_string_toBytesU() {
1518 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1519 return false;
1520 }
1521 // Get the arguments.
1522 Node* value = argument(0);
1523 Node* offset = argument(1);
1524 Node* length = argument(2);
1525
1526 Node* newcopy = NULL;
1527
1528 // Set the original stack and the reexecute bit for the interpreter to reexecute
1529 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1530 { PreserveReexecuteState preexecs(this);
1531 jvms()->set_should_reexecute(true);
1532
1533 // Check if a null path was taken unconditionally.
1534 value = null_check(value);
1535
1536 RegionNode* bailout = new RegionNode(1);
1537 record_for_igvn(bailout);
1538
1539 // Range checks
1540 generate_negative_guard(offset, bailout);
1541 generate_negative_guard(length, bailout);
1542 generate_limit_guard(offset, length, load_array_length(value), bailout);
1543 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1544 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1545
1546 if (bailout->req() > 1) {
1547 PreserveJVMState pjvms(this);
1548 set_control(_gvn.transform(bailout));
1549 uncommon_trap(Deoptimization::Reason_intrinsic,
1550 Deoptimization::Action_maybe_recompile);
1551 }
1552 if (stopped()) {
1553 return true;
1554 }
1555
1556 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1557 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1558 newcopy = new_array(klass_node, size, 0); // no arguments to push
1559 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1560
1561 // Calculate starting addresses.
1562 Node* src_start = array_element_address(value, offset, T_CHAR);
1563 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1564
1565 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1566 const TypeInt* toffset = gvn().type(offset)->is_int();
1567 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1568
1569 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1570 const char* copyfunc_name = "arraycopy";
1571 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1572 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1573 OptoRuntime::fast_arraycopy_Type(),
1574 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1575 src_start, dst_start, ConvI2X(length) XTOP);
1576 // Do not let reads from the cloned object float above the arraycopy.
1577 if (alloc != NULL) {
1578 if (alloc->maybe_set_complete(&_gvn)) {
1579 // "You break it, you buy it."
1580 InitializeNode* init = alloc->initialization();
1581 assert(init->is_complete(), "we just did this");
1582 init->set_complete_with_arraycopy();
1583 assert(newcopy->is_CheckCastPP(), "sanity");
1584 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1585 }
1586 // Do not let stores that initialize this object be reordered with
1587 // a subsequent store that would make this object accessible by
1588 // other threads.
1589 // Record what AllocateNode this StoreStore protects so that
1590 // escape analysis can go from the MemBarStoreStoreNode to the
1591 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1592 // based on the escape status of the AllocateNode.
1593 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1594 } else {
1595 insert_mem_bar(Op_MemBarCPUOrder);
1596 }
1597 } // original reexecute is set back here
1598
1599 C->set_has_split_ifs(true); // Has chance for split-if optimization
1600 if (!stopped()) {
1601 set_result(newcopy);
1602 }
1603 clear_upper_avx();
1604
1605 return true;
1606 }
1607
1608 //------------------------inline_string_getCharsU--------------------------
1609 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1610 bool LibraryCallKit::inline_string_getCharsU() {
1611 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1612 return false;
1613 }
1614
1615 // Get the arguments.
1616 Node* src = argument(0);
1617 Node* src_begin = argument(1);
1618 Node* src_end = argument(2); // exclusive offset (i < src_end)
1619 Node* dst = argument(3);
1620 Node* dst_begin = argument(4);
1621
1622 // Check for allocation before we add nodes that would confuse
1623 // tightly_coupled_allocation()
1624 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1625
1626 // Check if a null path was taken unconditionally.
1627 src = null_check(src);
1628 dst = null_check(dst);
1629 if (stopped()) {
1630 return true;
1631 }
1632
1633 // Get length and convert char[] offset to byte[] offset
1634 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1635 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1636
1637 // Range checks
1638 generate_string_range_check(src, src_begin, length, true);
1639 generate_string_range_check(dst, dst_begin, length, false);
1640 if (stopped()) {
1641 return true;
1642 }
1643
1644 if (!stopped()) {
1645 // Calculate starting addresses.
1646 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1647 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1648
1649 // Check if array addresses are aligned to HeapWordSize
1650 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1651 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1652 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1653 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1654
1655 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1656 const char* copyfunc_name = "arraycopy";
1657 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1658 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1659 OptoRuntime::fast_arraycopy_Type(),
1660 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1661 src_start, dst_start, ConvI2X(length) XTOP);
1662 // Do not let reads from the cloned object float above the arraycopy.
1663 if (alloc != NULL) {
1664 if (alloc->maybe_set_complete(&_gvn)) {
1665 // "You break it, you buy it."
1666 InitializeNode* init = alloc->initialization();
1667 assert(init->is_complete(), "we just did this");
1668 init->set_complete_with_arraycopy();
1669 assert(dst->is_CheckCastPP(), "sanity");
1670 assert(dst->in(0)->in(0) == init, "dest pinned");
1671 }
1672 // Do not let stores that initialize this object be reordered with
1673 // a subsequent store that would make this object accessible by
1674 // other threads.
1675 // Record what AllocateNode this StoreStore protects so that
1676 // escape analysis can go from the MemBarStoreStoreNode to the
1677 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1678 // based on the escape status of the AllocateNode.
1679 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1680 } else {
1681 insert_mem_bar(Op_MemBarCPUOrder);
1682 }
1683 }
1684
1685 C->set_has_split_ifs(true); // Has chance for split-if optimization
1686 return true;
1687 }
1688
1689 //----------------------inline_string_char_access----------------------------
1690 // Store/Load char to/from byte[] array.
1691 // static void StringUTF16.putChar(byte[] val, int index, int c)
1692 // static char StringUTF16.getChar(byte[] val, int index)
1693 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1694 Node* value = argument(0);
1695 Node* index = argument(1);
1696 Node* ch = is_store ? argument(2) : NULL;
1697
1698 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1699 // correctly requires matched array shapes.
1700 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1701 "sanity: byte[] and char[] bases agree");
1702 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1703 "sanity: byte[] and char[] scales agree");
1704
1705 // Bail when getChar over constants is requested: constant folding would
1706 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1707 // Java method would constant fold nicely instead.
1708 if (!is_store && value->is_Con() && index->is_Con()) {
1709 return false;
1710 }
1711
1712 Node* adr = array_element_address(value, index, T_CHAR);
1713 if (adr->is_top()) {
1714 return false;
1715 }
1716 if (is_store) {
1717 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1718 false, false, true /* mismatched */);
1719 } else {
1720 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1721 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1722 set_result(ch);
1723 }
1724 return true;
1725 }
1726
1727 //--------------------------round_double_node--------------------------------
1728 // Round a double node if necessary.
1729 Node* LibraryCallKit::round_double_node(Node* n) {
1730 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1731 n = _gvn.transform(new RoundDoubleNode(0, n));
1732 return n;
1733 }
1734
1735 //------------------------------inline_math-----------------------------------
1736 // public static double Math.abs(double)
1737 // public static double Math.sqrt(double)
1738 // public static double Math.log(double)
1739 // public static double Math.log10(double)
1740 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1741 Node* arg = round_double_node(argument(0));
1742 Node* n = NULL;
1743 switch (id) {
1744 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1745 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1746 default: fatal_unexpected_iid(id); break;
1747 }
1748 set_result(_gvn.transform(n));
1749 return true;
1750 }
1751
1752 //------------------------------runtime_math-----------------------------
1753 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1754 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1755 "must be (DD)D or (D)D type");
1756
1757 // Inputs
1758 Node* a = round_double_node(argument(0));
1759 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1760
1761 const TypePtr* no_memory_effects = NULL;
1762 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1763 no_memory_effects,
1764 a, top(), b, b ? top() : NULL);
1765 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1766 #ifdef ASSERT
1767 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1768 assert(value_top == top(), "second value must be top");
1769 #endif
1770
1771 set_result(value);
1772 return true;
1773 }
1774
1775 //------------------------------inline_math_native-----------------------------
1776 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1777 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1778 switch (id) {
1779 // These intrinsics are not properly supported on all hardware
1780 case vmIntrinsics::_dsin:
1781 return StubRoutines::dsin() != NULL ?
1782 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1783 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1784 case vmIntrinsics::_dcos:
1785 return StubRoutines::dcos() != NULL ?
1786 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1787 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1788 case vmIntrinsics::_dtan:
1789 return StubRoutines::dtan() != NULL ?
1790 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1791 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1792 case vmIntrinsics::_dlog:
1793 return StubRoutines::dlog() != NULL ?
1794 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1795 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1796 case vmIntrinsics::_dlog10:
1797 return StubRoutines::dlog10() != NULL ?
1798 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1799 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1800
1801 // These intrinsics are supported on all hardware
1802 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1803 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1804
1805 case vmIntrinsics::_dexp:
1806 return StubRoutines::dexp() != NULL ?
1807 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1808 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1809 case vmIntrinsics::_dpow: {
1810 Node* exp = round_double_node(argument(2));
1811 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1812 if (d != NULL && d->getd() == 2.0) {
1813 // Special case: pow(x, 2.0) => x * x
1814 Node* base = round_double_node(argument(0));
1815 set_result(_gvn.transform(new MulDNode(base, base)));
1816 return true;
1817 }
1818 return StubRoutines::dexp() != NULL ?
1819 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1820 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1821 }
1822 #undef FN_PTR
1823
1824 // These intrinsics are not yet correctly implemented
1825 case vmIntrinsics::_datan2:
1826 return false;
1827
1828 default:
1829 fatal_unexpected_iid(id);
1830 return false;
1831 }
1832 }
1833
1834 static bool is_simple_name(Node* n) {
1835 return (n->req() == 1 // constant
1836 || (n->is_Type() && n->as_Type()->type()->singleton())
1837 || n->is_Proj() // parameter or return value
1838 || n->is_Phi() // local of some sort
1839 );
1840 }
1841
1842 //----------------------------inline_notify-----------------------------------*
1843 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1844 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1845 address func;
1846 if (id == vmIntrinsics::_notify) {
1847 func = OptoRuntime::monitor_notify_Java();
1848 } else {
1849 func = OptoRuntime::monitor_notifyAll_Java();
1850 }
1851 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1852 make_slow_call_ex(call, env()->Throwable_klass(), false);
1853 return true;
1854 }
1855
1856
1857 //----------------------------inline_min_max-----------------------------------
1858 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1859 set_result(generate_min_max(id, argument(0), argument(1)));
1860 return true;
1861 }
1862
1863 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1864 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1865 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1866 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1867 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1868
1869 {
1870 PreserveJVMState pjvms(this);
1871 PreserveReexecuteState preexecs(this);
1872 jvms()->set_should_reexecute(true);
1873
1874 set_control(slow_path);
1875 set_i_o(i_o());
1876
1877 uncommon_trap(Deoptimization::Reason_intrinsic,
1878 Deoptimization::Action_none);
1879 }
1880
1881 set_control(fast_path);
1882 set_result(math);
1883 }
1884
1885 template <typename OverflowOp>
1886 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1887 typedef typename OverflowOp::MathOp MathOp;
1888
1889 MathOp* mathOp = new MathOp(arg1, arg2);
1890 Node* operation = _gvn.transform( mathOp );
1891 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1892 inline_math_mathExact(operation, ofcheck);
1893 return true;
1894 }
1895
1896 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1897 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1898 }
1899
1900 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1901 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1902 }
1903
1904 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1905 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1906 }
1907
1908 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1909 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1910 }
1911
1912 bool LibraryCallKit::inline_math_negateExactI() {
1913 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1914 }
1915
1916 bool LibraryCallKit::inline_math_negateExactL() {
1917 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1918 }
1919
1920 bool LibraryCallKit::inline_math_multiplyExactI() {
1921 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1922 }
1923
1924 bool LibraryCallKit::inline_math_multiplyExactL() {
1925 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1926 }
1927
1928 bool LibraryCallKit::inline_math_multiplyHigh() {
1929 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1930 return true;
1931 }
1932
1933 Node*
1934 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1935 // These are the candidate return value:
1936 Node* xvalue = x0;
1937 Node* yvalue = y0;
1938
1939 if (xvalue == yvalue) {
1940 return xvalue;
1941 }
1942
1943 bool want_max = (id == vmIntrinsics::_max);
1944
1945 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1946 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1947 if (txvalue == NULL || tyvalue == NULL) return top();
1948 // This is not really necessary, but it is consistent with a
1949 // hypothetical MaxINode::Value method:
1950 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1951
1952 // %%% This folding logic should (ideally) be in a different place.
1953 // Some should be inside IfNode, and there to be a more reliable
1954 // transformation of ?: style patterns into cmoves. We also want
1955 // more powerful optimizations around cmove and min/max.
1956
1957 // Try to find a dominating comparison of these guys.
1958 // It can simplify the index computation for Arrays.copyOf
1959 // and similar uses of System.arraycopy.
1960 // First, compute the normalized version of CmpI(x, y).
1961 int cmp_op = Op_CmpI;
1962 Node* xkey = xvalue;
1963 Node* ykey = yvalue;
1964 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1965 if (ideal_cmpxy->is_Cmp()) {
1966 // E.g., if we have CmpI(length - offset, count),
1967 // it might idealize to CmpI(length, count + offset)
1968 cmp_op = ideal_cmpxy->Opcode();
1969 xkey = ideal_cmpxy->in(1);
1970 ykey = ideal_cmpxy->in(2);
1971 }
1972
1973 // Start by locating any relevant comparisons.
1974 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1975 Node* cmpxy = NULL;
1976 Node* cmpyx = NULL;
1977 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1978 Node* cmp = start_from->fast_out(k);
1979 if (cmp->outcnt() > 0 && // must have prior uses
1980 cmp->in(0) == NULL && // must be context-independent
1981 cmp->Opcode() == cmp_op) { // right kind of compare
1982 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1983 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1984 }
1985 }
1986
1987 const int NCMPS = 2;
1988 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1989 int cmpn;
1990 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1991 if (cmps[cmpn] != NULL) break; // find a result
1992 }
1993 if (cmpn < NCMPS) {
1994 // Look for a dominating test that tells us the min and max.
1995 int depth = 0; // Limit search depth for speed
1996 Node* dom = control();
1997 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1998 if (++depth >= 100) break;
1999 Node* ifproj = dom;
2000 if (!ifproj->is_Proj()) continue;
2001 Node* iff = ifproj->in(0);
2002 if (!iff->is_If()) continue;
2003 Node* bol = iff->in(1);
2004 if (!bol->is_Bool()) continue;
2005 Node* cmp = bol->in(1);
2006 if (cmp == NULL) continue;
2007 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2008 if (cmps[cmpn] == cmp) break;
2009 if (cmpn == NCMPS) continue;
2010 BoolTest::mask btest = bol->as_Bool()->_test._test;
2011 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2012 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2013 // At this point, we know that 'x btest y' is true.
2014 switch (btest) {
2015 case BoolTest::eq:
2016 // They are proven equal, so we can collapse the min/max.
2017 // Either value is the answer. Choose the simpler.
2018 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2019 return yvalue;
2020 return xvalue;
2021 case BoolTest::lt: // x < y
2022 case BoolTest::le: // x <= y
2023 return (want_max ? yvalue : xvalue);
2024 case BoolTest::gt: // x > y
2025 case BoolTest::ge: // x >= y
2026 return (want_max ? xvalue : yvalue);
2027 default:
2028 break;
2029 }
2030 }
2031 }
2032
2033 // We failed to find a dominating test.
2034 // Let's pick a test that might GVN with prior tests.
2035 Node* best_bol = NULL;
2036 BoolTest::mask best_btest = BoolTest::illegal;
2037 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2038 Node* cmp = cmps[cmpn];
2039 if (cmp == NULL) continue;
2040 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2041 Node* bol = cmp->fast_out(j);
2042 if (!bol->is_Bool()) continue;
2043 BoolTest::mask btest = bol->as_Bool()->_test._test;
2044 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2045 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2046 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2047 best_bol = bol->as_Bool();
2048 best_btest = btest;
2049 }
2050 }
2051 }
2052
2053 Node* answer_if_true = NULL;
2054 Node* answer_if_false = NULL;
2055 switch (best_btest) {
2056 default:
2057 if (cmpxy == NULL)
2058 cmpxy = ideal_cmpxy;
2059 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2060 // and fall through:
2061 case BoolTest::lt: // x < y
2062 case BoolTest::le: // x <= y
2063 answer_if_true = (want_max ? yvalue : xvalue);
2064 answer_if_false = (want_max ? xvalue : yvalue);
2065 break;
2066 case BoolTest::gt: // x > y
2067 case BoolTest::ge: // x >= y
2068 answer_if_true = (want_max ? xvalue : yvalue);
2069 answer_if_false = (want_max ? yvalue : xvalue);
2070 break;
2071 }
2072
2073 jint hi, lo;
2074 if (want_max) {
2075 // We can sharpen the minimum.
2076 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2077 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2078 } else {
2079 // We can sharpen the maximum.
2080 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2081 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2082 }
2083
2084 // Use a flow-free graph structure, to avoid creating excess control edges
2085 // which could hinder other optimizations.
2086 // Since Math.min/max is often used with arraycopy, we want
2087 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2088 Node* cmov = CMoveNode::make(NULL, best_bol,
2089 answer_if_false, answer_if_true,
2090 TypeInt::make(lo, hi, widen));
2091
2092 return _gvn.transform(cmov);
2093
2094 /*
2095 // This is not as desirable as it may seem, since Min and Max
2096 // nodes do not have a full set of optimizations.
2097 // And they would interfere, anyway, with 'if' optimizations
2098 // and with CMoveI canonical forms.
2099 switch (id) {
2100 case vmIntrinsics::_min:
2101 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2102 case vmIntrinsics::_max:
2103 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2104 default:
2105 ShouldNotReachHere();
2106 }
2107 */
2108 }
2109
2110 inline int
2111 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2112 const TypePtr* base_type = TypePtr::NULL_PTR;
2113 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2114 if (base_type == NULL) {
2115 // Unknown type.
2116 return Type::AnyPtr;
2117 } else if (base_type == TypePtr::NULL_PTR) {
2118 // Since this is a NULL+long form, we have to switch to a rawptr.
2119 base = _gvn.transform(new CastX2PNode(offset));
2120 offset = MakeConX(0);
2121 return Type::RawPtr;
2122 } else if (base_type->base() == Type::RawPtr) {
2123 return Type::RawPtr;
2124 } else if (base_type->isa_oopptr()) {
2125 // Base is never null => always a heap address.
2126 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2127 return Type::OopPtr;
2128 }
2129 // Offset is small => always a heap address.
2130 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2131 if (offset_type != NULL &&
2132 base_type->offset() == 0 && // (should always be?)
2133 offset_type->_lo >= 0 &&
2134 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2135 return Type::OopPtr;
2136 } else if (type == T_OBJECT) {
2137 // off heap access to an oop doesn't make any sense. Has to be on
2138 // heap.
2139 return Type::OopPtr;
2140 }
2141 // Otherwise, it might either be oop+off or NULL+addr.
2142 return Type::AnyPtr;
2143 } else {
2144 // No information:
2145 return Type::AnyPtr;
2146 }
2147 }
2148
2149 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2150 Node* uncasted_base = base;
2151 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2152 if (kind == Type::RawPtr) {
2153 return basic_plus_adr(top(), uncasted_base, offset);
2154 } else if (kind == Type::AnyPtr) {
2155 assert(base == uncasted_base, "unexpected base change");
2156 if (can_cast) {
2157 if (!_gvn.type(base)->speculative_maybe_null() &&
2158 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2159 // According to profiling, this access is always on
2160 // heap. Casting the base to not null and thus avoiding membars
2161 // around the access should allow better optimizations
2162 Node* null_ctl = top();
2163 base = null_check_oop(base, &null_ctl, true, true, true);
2164 assert(null_ctl->is_top(), "no null control here");
2165 return basic_plus_adr(base, offset);
2166 } else if (_gvn.type(base)->speculative_always_null() &&
2167 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2168 // According to profiling, this access is always off
2169 // heap.
2170 base = null_assert(base);
2171 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2172 offset = MakeConX(0);
2173 return basic_plus_adr(top(), raw_base, offset);
2174 }
2175 }
2176 // We don't know if it's an on heap or off heap access. Fall back
2177 // to raw memory access.
2178 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2179 return basic_plus_adr(top(), raw, offset);
2180 } else {
2181 assert(base == uncasted_base, "unexpected base change");
2182 // We know it's an on heap access so base can't be null
2183 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2184 base = must_be_not_null(base, true);
2185 }
2186 return basic_plus_adr(base, offset);
2187 }
2188 }
2189
2190 //--------------------------inline_number_methods-----------------------------
2191 // inline int Integer.numberOfLeadingZeros(int)
2192 // inline int Long.numberOfLeadingZeros(long)
2193 //
2194 // inline int Integer.numberOfTrailingZeros(int)
2195 // inline int Long.numberOfTrailingZeros(long)
2196 //
2197 // inline int Integer.bitCount(int)
2198 // inline int Long.bitCount(long)
2199 //
2200 // inline char Character.reverseBytes(char)
2201 // inline short Short.reverseBytes(short)
2202 // inline int Integer.reverseBytes(int)
2203 // inline long Long.reverseBytes(long)
2204 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2205 Node* arg = argument(0);
2206 Node* n = NULL;
2207 switch (id) {
2208 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2209 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2210 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2211 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2212 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2213 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2214 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2215 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2216 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2217 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2218 default: fatal_unexpected_iid(id); break;
2219 }
2220 set_result(_gvn.transform(n));
2221 return true;
2222 }
2223
2224 //----------------------------inline_unsafe_access----------------------------
2225
2226 // Helper that guards and inserts a pre-barrier.
2227 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2228 Node* pre_val, bool need_mem_bar) {
2229 // We could be accessing the referent field of a reference object. If so, when G1
2230 // is enabled, we need to log the value in the referent field in an SATB buffer.
2231 // This routine performs some compile time filters and generates suitable
2232 // runtime filters that guard the pre-barrier code.
2233 // Also add memory barrier for non volatile load from the referent field
2234 // to prevent commoning of loads across safepoint.
2235 if (!UseG1GC && !need_mem_bar)
2236 return;
2237
2238 // Some compile time checks.
2239
2240 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2241 const TypeX* otype = offset->find_intptr_t_type();
2242 if (otype != NULL && otype->is_con() &&
2243 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2244 // Constant offset but not the reference_offset so just return
2245 return;
2246 }
2247
2248 // We only need to generate the runtime guards for instances.
2249 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2250 if (btype != NULL) {
2251 if (btype->isa_aryptr()) {
2252 // Array type so nothing to do
2253 return;
2254 }
2255
2256 const TypeInstPtr* itype = btype->isa_instptr();
2257 if (itype != NULL) {
2258 // Can the klass of base_oop be statically determined to be
2259 // _not_ a sub-class of Reference and _not_ Object?
2260 ciKlass* klass = itype->klass();
2261 if ( klass->is_loaded() &&
2262 !klass->is_subtype_of(env()->Reference_klass()) &&
2263 !env()->Object_klass()->is_subtype_of(klass)) {
2264 return;
2265 }
2266 }
2267 }
2268
2269 // The compile time filters did not reject base_oop/offset so
2270 // we need to generate the following runtime filters
2271 //
2272 // if (offset == java_lang_ref_Reference::_reference_offset) {
2273 // if (instance_of(base, java.lang.ref.Reference)) {
2274 // pre_barrier(_, pre_val, ...);
2275 // }
2276 // }
2277
2278 float likely = PROB_LIKELY( 0.999);
2279 float unlikely = PROB_UNLIKELY(0.999);
2280
2281 IdealKit ideal(this);
2282 #define __ ideal.
2283
2284 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2285
2286 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2287 // Update graphKit memory and control from IdealKit.
2288 sync_kit(ideal);
2289
2290 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2291 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2292
2293 // Update IdealKit memory and control from graphKit.
2294 __ sync_kit(this);
2295
2296 Node* one = __ ConI(1);
2297 // is_instof == 0 if base_oop == NULL
2298 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2299
2300 // Update graphKit from IdeakKit.
2301 sync_kit(ideal);
2302
2303 // Use the pre-barrier to record the value in the referent field
2304 pre_barrier(false /* do_load */,
2305 __ ctrl(),
2306 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2307 pre_val /* pre_val */,
2308 T_OBJECT);
2309 if (need_mem_bar) {
2310 // Add memory barrier to prevent commoning reads from this field
2311 // across safepoint since GC can change its value.
2312 insert_mem_bar(Op_MemBarCPUOrder);
2313 }
2314 // Update IdealKit from graphKit.
2315 __ sync_kit(this);
2316
2317 } __ end_if(); // _ref_type != ref_none
2318 } __ end_if(); // offset == referent_offset
2319
2320 // Final sync IdealKit and GraphKit.
2321 final_sync(ideal);
2322 #undef __
2323 }
2324
2325
2326 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2327 // Attempt to infer a sharper value type from the offset and base type.
2328 ciKlass* sharpened_klass = NULL;
2329
2330 // See if it is an instance field, with an object type.
2331 if (alias_type->field() != NULL) {
2332 if (alias_type->field()->type()->is_klass()) {
2333 sharpened_klass = alias_type->field()->type()->as_klass();
2334 }
2335 }
2336
2337 // See if it is a narrow oop array.
2338 if (adr_type->isa_aryptr()) {
2339 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2340 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2341 if (elem_type != NULL) {
2342 sharpened_klass = elem_type->klass();
2343 }
2344 }
2345 }
2346
2347 // The sharpened class might be unloaded if there is no class loader
2348 // contraint in place.
2349 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2350 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2351
2352 #ifndef PRODUCT
2353 if (C->print_intrinsics() || C->print_inlining()) {
2354 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2355 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2356 }
2357 #endif
2358 // Sharpen the value type.
2359 return tjp;
2360 }
2361 return NULL;
2362 }
2363
2364 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2365 if (callee()->is_static()) return false; // caller must have the capability!
2366 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2367 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2368 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2369
2370 #ifndef PRODUCT
2371 {
2372 ResourceMark rm;
2373 // Check the signatures.
2374 ciSignature* sig = callee()->signature();
2375 #ifdef ASSERT
2376 if (!is_store) {
2377 // Object getObject(Object base, int/long offset), etc.
2378 BasicType rtype = sig->return_type()->basic_type();
2379 assert(rtype == type, "getter must return the expected value");
2380 assert(sig->count() == 2, "oop getter has 2 arguments");
2381 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2382 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2383 } else {
2384 // void putObject(Object base, int/long offset, Object x), etc.
2385 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2386 assert(sig->count() == 3, "oop putter has 3 arguments");
2387 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2388 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2389 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2390 assert(vtype == type, "putter must accept the expected value");
2391 }
2392 #endif // ASSERT
2393 }
2394 #endif //PRODUCT
2395
2396 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2397
2398 Node* receiver = argument(0); // type: oop
2399
2400 // Build address expression.
2401 Node* adr;
2402 Node* heap_base_oop = top();
2403 Node* offset = top();
2404 Node* val;
2405
2406 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2407 Node* base = argument(1); // type: oop
2408 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2409 offset = argument(2); // type: long
2410 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2411 // to be plain byte offsets, which are also the same as those accepted
2412 // by oopDesc::field_addr.
2413 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2414 "fieldOffset must be byte-scaled");
2415 // 32-bit machines ignore the high half!
2416 offset = ConvL2X(offset);
2417 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2418
2419 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2420 heap_base_oop = base;
2421 } else if (type == T_OBJECT) {
2422 return false; // off-heap oop accesses are not supported
2423 }
2424
2425 // Can base be NULL? Otherwise, always on-heap access.
2426 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2427
2428 val = is_store ? argument(4) : NULL;
2429
2430 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2431
2432 // Try to categorize the address.
2433 Compile::AliasType* alias_type = C->alias_type(adr_type);
2434 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2435
2436 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2437 alias_type->adr_type() == TypeAryPtr::RANGE) {
2438 return false; // not supported
2439 }
2440
2441 bool mismatched = false;
2442 BasicType bt = alias_type->basic_type();
2443 if (bt != T_ILLEGAL) {
2444 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2445 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2446 // Alias type doesn't differentiate between byte[] and boolean[]).
2447 // Use address type to get the element type.
2448 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2449 }
2450 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2451 // accessing an array field with getObject is not a mismatch
2452 bt = T_OBJECT;
2453 }
2454 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2455 // Don't intrinsify mismatched object accesses
2456 return false;
2457 }
2458 mismatched = (bt != type);
2459 } else if (alias_type->adr_type()->isa_oopptr()) {
2460 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2461 }
2462
2463 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2464
2465 // First guess at the value type.
2466 const Type *value_type = Type::get_const_basic_type(type);
2467
2468 // We will need memory barriers unless we can determine a unique
2469 // alias category for this reference. (Note: If for some reason
2470 // the barriers get omitted and the unsafe reference begins to "pollute"
2471 // the alias analysis of the rest of the graph, either Compile::can_alias
2472 // or Compile::must_alias will throw a diagnostic assert.)
2473 bool need_mem_bar = false;
2474 switch (kind) {
2475 case Relaxed:
2476 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2477 break;
2478 case Opaque:
2479 // Opaque uses CPUOrder membars for protection against code movement.
2480 case Acquire:
2481 case Release:
2482 case Volatile:
2483 need_mem_bar = true;
2484 break;
2485 default:
2486 ShouldNotReachHere();
2487 }
2488
2489 // Some accesses require access atomicity for all types, notably longs and doubles.
2490 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2491 bool requires_atomic_access = false;
2492 switch (kind) {
2493 case Relaxed:
2494 requires_atomic_access = AlwaysAtomicAccesses;
2495 break;
2496 case Opaque:
2497 // Opaque accesses are atomic.
2498 case Acquire:
2499 case Release:
2500 case Volatile:
2501 requires_atomic_access = true;
2502 break;
2503 default:
2504 ShouldNotReachHere();
2505 }
2506
2507 // Figure out the memory ordering.
2508 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2509 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2510
2511 // If we are reading the value of the referent field of a Reference
2512 // object (either by using Unsafe directly or through reflection)
2513 // then, if G1 is enabled, we need to record the referent in an
2514 // SATB log buffer using the pre-barrier mechanism.
2515 // Also we need to add memory barrier to prevent commoning reads
2516 // from this field across safepoint since GC can change its value.
2517 bool need_read_barrier = !is_store &&
2518 offset != top() && heap_base_oop != top();
2519
2520 if (!is_store && type == T_OBJECT) {
2521 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2522 if (tjp != NULL) {
2523 value_type = tjp;
2524 }
2525 }
2526
2527 receiver = null_check(receiver);
2528 if (stopped()) {
2529 return true;
2530 }
2531 // Heap pointers get a null-check from the interpreter,
2532 // as a courtesy. However, this is not guaranteed by Unsafe,
2533 // and it is not possible to fully distinguish unintended nulls
2534 // from intended ones in this API.
2535
2536 // We need to emit leading and trailing CPU membars (see below) in
2537 // addition to memory membars for special access modes. This is a little
2538 // too strong, but avoids the need to insert per-alias-type
2539 // volatile membars (for stores; compare Parse::do_put_xxx), which
2540 // we cannot do effectively here because we probably only have a
2541 // rough approximation of type.
2542
2543 switch(kind) {
2544 case Relaxed:
2545 case Opaque:
2546 case Acquire:
2547 break;
2548 case Release:
2549 case Volatile:
2550 if (is_store) {
2551 insert_mem_bar(Op_MemBarRelease);
2552 } else {
2553 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2554 insert_mem_bar(Op_MemBarVolatile);
2555 }
2556 }
2557 break;
2558 default:
2559 ShouldNotReachHere();
2560 }
2561
2562 // Memory barrier to prevent normal and 'unsafe' accesses from
2563 // bypassing each other. Happens after null checks, so the
2564 // exception paths do not take memory state from the memory barrier,
2565 // so there's no problems making a strong assert about mixing users
2566 // of safe & unsafe memory.
2567 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2568
2569 if (!is_store) {
2570 Node* p = NULL;
2571 // Try to constant fold a load from a constant field
2572 ciField* field = alias_type->field();
2573 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2574 // final or stable field
2575 p = make_constant_from_field(field, heap_base_oop);
2576 }
2577 if (p == NULL) {
2578 // To be valid, unsafe loads may depend on other conditions than
2579 // the one that guards them: pin the Load node
2580 LoadNode::ControlDependency dep = LoadNode::Pinned;
2581 Node* ctrl = control();
2582 // non volatile loads may be able to float
2583 if (!need_mem_bar && adr_type->isa_instptr()) {
2584 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2585 intptr_t offset = Type::OffsetBot;
2586 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2587 if (offset >= 0) {
2588 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2589 if (offset < s) {
2590 // Guaranteed to be a valid access, no need to pin it
2591 dep = LoadNode::DependsOnlyOnTest;
2592 ctrl = NULL;
2593 }
2594 }
2595 }
2596 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2597 // load value
2598 switch (type) {
2599 case T_BOOLEAN:
2600 {
2601 // Normalize the value returned by getBoolean in the following cases
2602 if (mismatched ||
2603 heap_base_oop == top() || // - heap_base_oop is NULL or
2604 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2605 // and the unsafe access is made to large offset
2606 // (i.e., larger than the maximum offset necessary for any
2607 // field access)
2608 ) {
2609 IdealKit ideal = IdealKit(this);
2610 #define __ ideal.
2611 IdealVariable normalized_result(ideal);
2612 __ declarations_done();
2613 __ set(normalized_result, p);
2614 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2615 __ set(normalized_result, ideal.ConI(1));
2616 ideal.end_if();
2617 final_sync(ideal);
2618 p = __ value(normalized_result);
2619 #undef __
2620 }
2621 }
2622 case T_CHAR:
2623 case T_BYTE:
2624 case T_SHORT:
2625 case T_INT:
2626 case T_LONG:
2627 case T_FLOAT:
2628 case T_DOUBLE:
2629 break;
2630 case T_OBJECT:
2631 if (need_read_barrier) {
2632 // We do not require a mem bar inside pre_barrier if need_mem_bar
2633 // is set: the barriers would be emitted by us.
2634 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2635 }
2636 break;
2637 case T_ADDRESS:
2638 // Cast to an int type.
2639 p = _gvn.transform(new CastP2XNode(NULL, p));
2640 p = ConvX2UL(p);
2641 break;
2642 default:
2643 fatal("unexpected type %d: %s", type, type2name(type));
2644 break;
2645 }
2646 }
2647 // The load node has the control of the preceding MemBarCPUOrder. All
2648 // following nodes will have the control of the MemBarCPUOrder inserted at
2649 // the end of this method. So, pushing the load onto the stack at a later
2650 // point is fine.
2651 set_result(p);
2652 } else {
2653 // place effect of store into memory
2654 switch (type) {
2655 case T_DOUBLE:
2656 val = dstore_rounding(val);
2657 break;
2658 case T_ADDRESS:
2659 // Repackage the long as a pointer.
2660 val = ConvL2X(val);
2661 val = _gvn.transform(new CastX2PNode(val));
2662 break;
2663 default:
2664 break;
2665 }
2666
2667 if (type == T_OBJECT) {
2668 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2669 } else {
2670 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2671 }
2672 }
2673
2674 switch(kind) {
2675 case Relaxed:
2676 case Opaque:
2677 case Release:
2678 break;
2679 case Acquire:
2680 case Volatile:
2681 if (!is_store) {
2682 insert_mem_bar(Op_MemBarAcquire);
2683 } else {
2684 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2685 insert_mem_bar(Op_MemBarVolatile);
2686 }
2687 }
2688 break;
2689 default:
2690 ShouldNotReachHere();
2691 }
2692
2693 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2694
2695 return true;
2696 }
2697
2698 //----------------------------inline_unsafe_load_store----------------------------
2699 // This method serves a couple of different customers (depending on LoadStoreKind):
2700 //
2701 // LS_cmp_swap:
2702 //
2703 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2704 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2705 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2706 //
2707 // LS_cmp_swap_weak:
2708 //
2709 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2710 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2711 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2712 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2713 //
2714 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2715 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2716 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2717 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2718 //
2719 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2720 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2721 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2722 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2723 //
2724 // LS_cmp_exchange:
2725 //
2726 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2727 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2728 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2729 //
2730 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2731 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2732 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2733 //
2734 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2735 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2736 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2737 //
2738 // LS_get_add:
2739 //
2740 // int getAndAddInt( Object o, long offset, int delta)
2741 // long getAndAddLong(Object o, long offset, long delta)
2742 //
2743 // LS_get_set:
2744 //
2745 // int getAndSet(Object o, long offset, int newValue)
2746 // long getAndSet(Object o, long offset, long newValue)
2747 // Object getAndSet(Object o, long offset, Object newValue)
2748 //
2749 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2750 // This basic scheme here is the same as inline_unsafe_access, but
2751 // differs in enough details that combining them would make the code
2752 // overly confusing. (This is a true fact! I originally combined
2753 // them, but even I was confused by it!) As much code/comments as
2754 // possible are retained from inline_unsafe_access though to make
2755 // the correspondences clearer. - dl
2756
2757 if (callee()->is_static()) return false; // caller must have the capability!
2758
2759 #ifndef PRODUCT
2760 BasicType rtype;
2761 {
2762 ResourceMark rm;
2763 // Check the signatures.
2764 ciSignature* sig = callee()->signature();
2765 rtype = sig->return_type()->basic_type();
2766 switch(kind) {
2767 case LS_get_add:
2768 case LS_get_set: {
2769 // Check the signatures.
2770 #ifdef ASSERT
2771 assert(rtype == type, "get and set must return the expected type");
2772 assert(sig->count() == 3, "get and set has 3 arguments");
2773 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2774 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2775 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2776 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2777 #endif // ASSERT
2778 break;
2779 }
2780 case LS_cmp_swap:
2781 case LS_cmp_swap_weak: {
2782 // Check the signatures.
2783 #ifdef ASSERT
2784 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2785 assert(sig->count() == 4, "CAS has 4 arguments");
2786 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2787 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2788 #endif // ASSERT
2789 break;
2790 }
2791 case LS_cmp_exchange: {
2792 // Check the signatures.
2793 #ifdef ASSERT
2794 assert(rtype == type, "CAS must return the expected type");
2795 assert(sig->count() == 4, "CAS has 4 arguments");
2796 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2797 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2798 #endif // ASSERT
2799 break;
2800 }
2801 default:
2802 ShouldNotReachHere();
2803 }
2804 }
2805 #endif //PRODUCT
2806
2807 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2808
2809 // Get arguments:
2810 Node* receiver = NULL;
2811 Node* base = NULL;
2812 Node* offset = NULL;
2813 Node* oldval = NULL;
2814 Node* newval = NULL;
2815 switch(kind) {
2816 case LS_cmp_swap:
2817 case LS_cmp_swap_weak:
2818 case LS_cmp_exchange: {
2819 const bool two_slot_type = type2size[type] == 2;
2820 receiver = argument(0); // type: oop
2821 base = argument(1); // type: oop
2822 offset = argument(2); // type: long
2823 oldval = argument(4); // type: oop, int, or long
2824 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2825 break;
2826 }
2827 case LS_get_add:
2828 case LS_get_set: {
2829 receiver = argument(0); // type: oop
2830 base = argument(1); // type: oop
2831 offset = argument(2); // type: long
2832 oldval = NULL;
2833 newval = argument(4); // type: oop, int, or long
2834 break;
2835 }
2836 default:
2837 ShouldNotReachHere();
2838 }
2839
2840 // Build field offset expression.
2841 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2842 // to be plain byte offsets, which are also the same as those accepted
2843 // by oopDesc::field_addr.
2844 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2845 // 32-bit machines ignore the high half of long offsets
2846 offset = ConvL2X(offset);
2847 Node* adr = make_unsafe_address(base, offset, type, false);
2848 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2849
2850 Compile::AliasType* alias_type = C->alias_type(adr_type);
2851 BasicType bt = alias_type->basic_type();
2852 if (bt != T_ILLEGAL &&
2853 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2854 // Don't intrinsify mismatched object accesses.
2855 return false;
2856 }
2857
2858 // For CAS, unlike inline_unsafe_access, there seems no point in
2859 // trying to refine types. Just use the coarse types here.
2860 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2861 const Type *value_type = Type::get_const_basic_type(type);
2862
2863 switch (kind) {
2864 case LS_get_set:
2865 case LS_cmp_exchange: {
2866 if (type == T_OBJECT) {
2867 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2868 if (tjp != NULL) {
2869 value_type = tjp;
2870 }
2871 }
2872 break;
2873 }
2874 case LS_cmp_swap:
2875 case LS_cmp_swap_weak:
2876 case LS_get_add:
2877 break;
2878 default:
2879 ShouldNotReachHere();
2880 }
2881
2882 // Null check receiver.
2883 receiver = null_check(receiver);
2884 if (stopped()) {
2885 return true;
2886 }
2887
2888 int alias_idx = C->get_alias_index(adr_type);
2889
2890 // Memory-model-wise, a LoadStore acts like a little synchronized
2891 // block, so needs barriers on each side. These don't translate
2892 // into actual barriers on most machines, but we still need rest of
2893 // compiler to respect ordering.
2894
2895 switch (access_kind) {
2896 case Relaxed:
2897 case Acquire:
2898 break;
2899 case Release:
2900 insert_mem_bar(Op_MemBarRelease);
2901 break;
2902 case Volatile:
2903 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2904 insert_mem_bar(Op_MemBarVolatile);
2905 } else {
2906 insert_mem_bar(Op_MemBarRelease);
2907 }
2908 break;
2909 default:
2910 ShouldNotReachHere();
2911 }
2912 insert_mem_bar(Op_MemBarCPUOrder);
2913
2914 // Figure out the memory ordering.
2915 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2916
2917 // 4984716: MemBars must be inserted before this
2918 // memory node in order to avoid a false
2919 // dependency which will confuse the scheduler.
2920 Node *mem = memory(alias_idx);
2921
2922 // For now, we handle only those cases that actually exist: ints,
2923 // longs, and Object. Adding others should be straightforward.
2924 Node* load_store = NULL;
2925 switch(type) {
2926 case T_BYTE:
2927 switch(kind) {
2928 case LS_get_add:
2929 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2930 break;
2931 case LS_get_set:
2932 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2933 break;
2934 case LS_cmp_swap_weak:
2935 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2936 break;
2937 case LS_cmp_swap:
2938 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2939 break;
2940 case LS_cmp_exchange:
2941 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2942 break;
2943 default:
2944 ShouldNotReachHere();
2945 }
2946 break;
2947 case T_SHORT:
2948 switch(kind) {
2949 case LS_get_add:
2950 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2951 break;
2952 case LS_get_set:
2953 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2954 break;
2955 case LS_cmp_swap_weak:
2956 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2957 break;
2958 case LS_cmp_swap:
2959 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2960 break;
2961 case LS_cmp_exchange:
2962 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2963 break;
2964 default:
2965 ShouldNotReachHere();
2966 }
2967 break;
2968 case T_INT:
2969 switch(kind) {
2970 case LS_get_add:
2971 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2972 break;
2973 case LS_get_set:
2974 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2975 break;
2976 case LS_cmp_swap_weak:
2977 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2978 break;
2979 case LS_cmp_swap:
2980 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2981 break;
2982 case LS_cmp_exchange:
2983 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2984 break;
2985 default:
2986 ShouldNotReachHere();
2987 }
2988 break;
2989 case T_LONG:
2990 switch(kind) {
2991 case LS_get_add:
2992 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2993 break;
2994 case LS_get_set:
2995 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2996 break;
2997 case LS_cmp_swap_weak:
2998 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2999 break;
3000 case LS_cmp_swap:
3001 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3002 break;
3003 case LS_cmp_exchange:
3004 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
3005 break;
3006 default:
3007 ShouldNotReachHere();
3008 }
3009 break;
3010 case T_OBJECT:
3011 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3012 // could be delayed during Parse (for example, in adjust_map_after_if()).
3013 // Execute transformation here to avoid barrier generation in such case.
3014 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3015 newval = _gvn.makecon(TypePtr::NULL_PTR);
3016
3017 // Reference stores need a store barrier.
3018 switch(kind) {
3019 case LS_get_set: {
3020 // If pre-barrier must execute before the oop store, old value will require do_load here.
3021 if (!can_move_pre_barrier()) {
3022 pre_barrier(true /* do_load*/,
3023 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3024 NULL /* pre_val*/,
3025 T_OBJECT);
3026 } // Else move pre_barrier to use load_store value, see below.
3027 break;
3028 }
3029 case LS_cmp_swap_weak:
3030 case LS_cmp_swap:
3031 case LS_cmp_exchange: {
3032 // Same as for newval above:
3033 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3034 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3035 }
3036 // The only known value which might get overwritten is oldval.
3037 pre_barrier(false /* do_load */,
3038 control(), NULL, NULL, max_juint, NULL, NULL,
3039 oldval /* pre_val */,
3040 T_OBJECT);
3041 break;
3042 }
3043 default:
3044 ShouldNotReachHere();
3045 }
3046
3047 #ifdef _LP64
3048 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3049 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3050
3051 switch(kind) {
3052 case LS_get_set:
3053 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3054 break;
3055 case LS_cmp_swap_weak: {
3056 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3057 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3058 break;
3059 }
3060 case LS_cmp_swap: {
3061 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3062 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3063 break;
3064 }
3065 case LS_cmp_exchange: {
3066 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3067 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3068 break;
3069 }
3070 default:
3071 ShouldNotReachHere();
3072 }
3073 } else
3074 #endif
3075 switch (kind) {
3076 case LS_get_set:
3077 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3078 break;
3079 case LS_cmp_swap_weak:
3080 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3081 break;
3082 case LS_cmp_swap:
3083 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3084 break;
3085 case LS_cmp_exchange:
3086 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3087 break;
3088 default:
3089 ShouldNotReachHere();
3090 }
3091
3092 // Emit the post barrier only when the actual store happened. This makes sense
3093 // to check only for LS_cmp_* that can fail to set the value.
3094 // LS_cmp_exchange does not produce any branches by default, so there is no
3095 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3096 // CompareAndExchange and move branches here, it would make sense to conditionalize
3097 // post_barriers for LS_cmp_exchange as well.
3098 //
3099 // CAS success path is marked more likely since we anticipate this is a performance
3100 // critical path, while CAS failure path can use the penalty for going through unlikely
3101 // path as backoff. Which is still better than doing a store barrier there.
3102 switch (kind) {
3103 case LS_get_set:
3104 case LS_cmp_exchange: {
3105 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3106 break;
3107 }
3108 case LS_cmp_swap_weak:
3109 case LS_cmp_swap: {
3110 IdealKit ideal(this);
3111 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3112 sync_kit(ideal);
3113 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3114 ideal.sync_kit(this);
3115 } ideal.end_if();
3116 final_sync(ideal);
3117 break;
3118 }
3119 default:
3120 ShouldNotReachHere();
3121 }
3122 break;
3123 default:
3124 fatal("unexpected type %d: %s", type, type2name(type));
3125 break;
3126 }
3127
3128 // SCMemProjNodes represent the memory state of a LoadStore. Their
3129 // main role is to prevent LoadStore nodes from being optimized away
3130 // when their results aren't used.
3131 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3132 set_memory(proj, alias_idx);
3133
3134 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3135 #ifdef _LP64
3136 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3137 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3138 }
3139 #endif
3140 if (can_move_pre_barrier() && kind == LS_get_set) {
3141 // Don't need to load pre_val. The old value is returned by load_store.
3142 // The pre_barrier can execute after the xchg as long as no safepoint
3143 // gets inserted between them.
3144 pre_barrier(false /* do_load */,
3145 control(), NULL, NULL, max_juint, NULL, NULL,
3146 load_store /* pre_val */,
3147 T_OBJECT);
3148 }
3149 }
3150
3151 // Add the trailing membar surrounding the access
3152 insert_mem_bar(Op_MemBarCPUOrder);
3153
3154 switch (access_kind) {
3155 case Relaxed:
3156 case Release:
3157 break; // do nothing
3158 case Acquire:
3159 case Volatile:
3160 insert_mem_bar(Op_MemBarAcquire);
3161 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3162 break;
3163 default:
3164 ShouldNotReachHere();
3165 }
3166
3167 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3168 set_result(load_store);
3169 return true;
3170 }
3171
3172 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3173 MemNode::MemOrd mo = MemNode::unset;
3174 switch(kind) {
3175 case Opaque:
3176 case Relaxed: mo = MemNode::unordered; break;
3177 case Acquire: mo = MemNode::acquire; break;
3178 case Release: mo = MemNode::release; break;
3179 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3180 default:
3181 ShouldNotReachHere();
3182 }
3183 guarantee(mo != MemNode::unset, "Should select memory ordering");
3184 return mo;
3185 }
3186
3187 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3188 MemNode::MemOrd mo = MemNode::unset;
3189 switch(kind) {
3190 case Opaque:
3191 case Relaxed: mo = MemNode::unordered; break;
3192 case Acquire: mo = MemNode::acquire; break;
3193 case Release: mo = MemNode::release; break;
3194 case Volatile: mo = MemNode::seqcst; break;
3195 default:
3196 ShouldNotReachHere();
3197 }
3198 guarantee(mo != MemNode::unset, "Should select memory ordering");
3199 return mo;
3200 }
3201
3202 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3203 // Regardless of form, don't allow previous ld/st to move down,
3204 // then issue acquire, release, or volatile mem_bar.
3205 insert_mem_bar(Op_MemBarCPUOrder);
3206 switch(id) {
3207 case vmIntrinsics::_loadFence:
3208 insert_mem_bar(Op_LoadFence);
3209 return true;
3210 case vmIntrinsics::_storeFence:
3211 insert_mem_bar(Op_StoreFence);
3212 return true;
3213 case vmIntrinsics::_fullFence:
3214 insert_mem_bar(Op_MemBarVolatile);
3215 return true;
3216 default:
3217 fatal_unexpected_iid(id);
3218 return false;
3219 }
3220 }
3221
3222 bool LibraryCallKit::inline_onspinwait() {
3223 insert_mem_bar(Op_OnSpinWait);
3224 return true;
3225 }
3226
3227 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3228 if (!kls->is_Con()) {
3229 return true;
3230 }
3231 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3232 if (klsptr == NULL) {
3233 return true;
3234 }
3235 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3236 // don't need a guard for a klass that is already initialized
3237 return !ik->is_initialized();
3238 }
3239
3240 //----------------------------inline_unsafe_allocate---------------------------
3241 // public native Object Unsafe.allocateInstance(Class<?> cls);
3242 bool LibraryCallKit::inline_unsafe_allocate() {
3243 if (callee()->is_static()) return false; // caller must have the capability!
3244
3245 null_check_receiver(); // null-check, then ignore
3246 Node* cls = null_check(argument(1));
3247 if (stopped()) return true;
3248
3249 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3250 kls = null_check(kls);
3251 if (stopped()) return true; // argument was like int.class
3252
3253 Node* test = NULL;
3254 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3255 // Note: The argument might still be an illegal value like
3256 // Serializable.class or Object[].class. The runtime will handle it.
3257 // But we must make an explicit check for initialization.
3258 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3259 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3260 // can generate code to load it as unsigned byte.
3261 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3262 Node* bits = intcon(InstanceKlass::fully_initialized);
3263 test = _gvn.transform(new SubINode(inst, bits));
3264 // The 'test' is non-zero if we need to take a slow path.
3265 }
3266
3267 Node* obj = new_instance(kls, test);
3268 set_result(obj);
3269 return true;
3270 }
3271
3272 //------------------------inline_native_time_funcs--------------
3273 // inline code for System.currentTimeMillis() and System.nanoTime()
3274 // these have the same type and signature
3275 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3276 const TypeFunc* tf = OptoRuntime::void_long_Type();
3277 const TypePtr* no_memory_effects = NULL;
3278 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3279 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3280 #ifdef ASSERT
3281 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3282 assert(value_top == top(), "second value must be top");
3283 #endif
3284 set_result(value);
3285 return true;
3286 }
3287
3288 #ifdef TRACE_HAVE_INTRINSICS
3289
3290 /*
3291 * oop -> myklass
3292 * myklass->trace_id |= USED
3293 * return myklass->trace_id & ~0x3
3294 */
3295 bool LibraryCallKit::inline_native_classID() {
3296 Node* cls = null_check(argument(0), T_OBJECT);
3297 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3298 kls = null_check(kls, T_OBJECT);
3299
3300 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3301 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3302 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3303
3304 Node* clsused = longcon(0x01l); // set the class bit
3305 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3306 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3307 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3308
3309 #ifdef TRACE_ID_META_BITS
3310 Node* mbits = longcon(~TRACE_ID_META_BITS);
3311 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3312 #endif
3313 #ifdef TRACE_ID_CLASS_SHIFT
3314 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3315 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3316 #endif
3317
3318 set_result(tvalue);
3319 return true;
3320
3321 }
3322
3323 bool LibraryCallKit::inline_native_getBufferWriter() {
3324 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3325
3326 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3327 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3328 );
3329
3330 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3331
3332 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3333 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3334
3335 IfNode* iff_jobj_null =
3336 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3337
3338 enum { _normal_path = 1,
3339 _null_path = 2,
3340 PATH_LIMIT };
3341
3342 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3343 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3344
3345 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3346 result_rgn->init_req(_null_path, jobj_is_null);
3347 result_val->init_req(_null_path, null());
3348
3349 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3350 result_rgn->init_req(_normal_path, jobj_is_not_null);
3351
3352 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3353 result_val->init_req(_normal_path, res);
3354
3355 set_result(result_rgn, result_val);
3356
3357 return true;
3358 }
3359
3360 #endif
3361
3362 //------------------------inline_native_currentThread------------------
3363 bool LibraryCallKit::inline_native_currentThread() {
3364 Node* junk = NULL;
3365 set_result(generate_current_thread(junk));
3366 return true;
3367 }
3368
3369 //------------------------inline_native_isInterrupted------------------
3370 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3371 bool LibraryCallKit::inline_native_isInterrupted() {
3372 // Add a fast path to t.isInterrupted(clear_int):
3373 // (t == Thread.current() &&
3374 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3375 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3376 // So, in the common case that the interrupt bit is false,
3377 // we avoid making a call into the VM. Even if the interrupt bit
3378 // is true, if the clear_int argument is false, we avoid the VM call.
3379 // However, if the receiver is not currentThread, we must call the VM,
3380 // because there must be some locking done around the operation.
3381
3382 // We only go to the fast case code if we pass two guards.
3383 // Paths which do not pass are accumulated in the slow_region.
3384
3385 enum {
3386 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3387 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3388 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3389 PATH_LIMIT
3390 };
3391
3392 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3393 // out of the function.
3394 insert_mem_bar(Op_MemBarCPUOrder);
3395
3396 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3397 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3398
3399 RegionNode* slow_region = new RegionNode(1);
3400 record_for_igvn(slow_region);
3401
3402 // (a) Receiving thread must be the current thread.
3403 Node* rec_thr = argument(0);
3404 Node* tls_ptr = NULL;
3405 Node* cur_thr = generate_current_thread(tls_ptr);
3406 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3407 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3408
3409 generate_slow_guard(bol_thr, slow_region);
3410
3411 // (b) Interrupt bit on TLS must be false.
3412 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3413 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3414 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3415
3416 // Set the control input on the field _interrupted read to prevent it floating up.
3417 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3418 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3419 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3420
3421 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3422
3423 // First fast path: if (!TLS._interrupted) return false;
3424 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3425 result_rgn->init_req(no_int_result_path, false_bit);
3426 result_val->init_req(no_int_result_path, intcon(0));
3427
3428 // drop through to next case
3429 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3430
3431 #ifndef _WINDOWS
3432 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3433 Node* clr_arg = argument(1);
3434 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3435 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3436 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3437
3438 // Second fast path: ... else if (!clear_int) return true;
3439 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3440 result_rgn->init_req(no_clear_result_path, false_arg);
3441 result_val->init_req(no_clear_result_path, intcon(1));
3442
3443 // drop through to next case
3444 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3445 #else
3446 // To return true on Windows you must read the _interrupted field
3447 // and check the event state i.e. take the slow path.
3448 #endif // _WINDOWS
3449
3450 // (d) Otherwise, go to the slow path.
3451 slow_region->add_req(control());
3452 set_control( _gvn.transform(slow_region));
3453
3454 if (stopped()) {
3455 // There is no slow path.
3456 result_rgn->init_req(slow_result_path, top());
3457 result_val->init_req(slow_result_path, top());
3458 } else {
3459 // non-virtual because it is a private non-static
3460 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3461
3462 Node* slow_val = set_results_for_java_call(slow_call);
3463 // this->control() comes from set_results_for_java_call
3464
3465 Node* fast_io = slow_call->in(TypeFunc::I_O);
3466 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3467
3468 // These two phis are pre-filled with copies of of the fast IO and Memory
3469 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3470 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3471
3472 result_rgn->init_req(slow_result_path, control());
3473 result_io ->init_req(slow_result_path, i_o());
3474 result_mem->init_req(slow_result_path, reset_memory());
3475 result_val->init_req(slow_result_path, slow_val);
3476
3477 set_all_memory(_gvn.transform(result_mem));
3478 set_i_o( _gvn.transform(result_io));
3479 }
3480
3481 C->set_has_split_ifs(true); // Has chance for split-if optimization
3482 set_result(result_rgn, result_val);
3483 return true;
3484 }
3485
3486 //---------------------------load_mirror_from_klass----------------------------
3487 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3488 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3489 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3490 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3491 // mirror = ((OopHandle)mirror)->resolve();
3492 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3493 }
3494
3495 //-----------------------load_klass_from_mirror_common-------------------------
3496 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3497 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3498 // and branch to the given path on the region.
3499 // If never_see_null, take an uncommon trap on null, so we can optimistically
3500 // compile for the non-null case.
3501 // If the region is NULL, force never_see_null = true.
3502 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3503 bool never_see_null,
3504 RegionNode* region,
3505 int null_path,
3506 int offset) {
3507 if (region == NULL) never_see_null = true;
3508 Node* p = basic_plus_adr(mirror, offset);
3509 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3510 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3511 Node* null_ctl = top();
3512 kls = null_check_oop(kls, &null_ctl, never_see_null);
3513 if (region != NULL) {
3514 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3515 region->init_req(null_path, null_ctl);
3516 } else {
3517 assert(null_ctl == top(), "no loose ends");
3518 }
3519 return kls;
3520 }
3521
3522 //--------------------(inline_native_Class_query helpers)---------------------
3523 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3524 // Fall through if (mods & mask) == bits, take the guard otherwise.
3525 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3526 // Branch around if the given klass has the given modifier bit set.
3527 // Like generate_guard, adds a new path onto the region.
3528 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3529 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3530 Node* mask = intcon(modifier_mask);
3531 Node* bits = intcon(modifier_bits);
3532 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3533 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3534 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3535 return generate_fair_guard(bol, region);
3536 }
3537 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3538 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3539 }
3540
3541 //-------------------------inline_native_Class_query-------------------
3542 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3543 const Type* return_type = TypeInt::BOOL;
3544 Node* prim_return_value = top(); // what happens if it's a primitive class?
3545 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3546 bool expect_prim = false; // most of these guys expect to work on refs
3547
3548 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3549
3550 Node* mirror = argument(0);
3551 Node* obj = top();
3552
3553 switch (id) {
3554 case vmIntrinsics::_isInstance:
3555 // nothing is an instance of a primitive type
3556 prim_return_value = intcon(0);
3557 obj = argument(1);
3558 break;
3559 case vmIntrinsics::_getModifiers:
3560 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3561 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3562 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3563 break;
3564 case vmIntrinsics::_isInterface:
3565 prim_return_value = intcon(0);
3566 break;
3567 case vmIntrinsics::_isArray:
3568 prim_return_value = intcon(0);
3569 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3570 break;
3571 case vmIntrinsics::_isPrimitive:
3572 prim_return_value = intcon(1);
3573 expect_prim = true; // obviously
3574 break;
3575 case vmIntrinsics::_getSuperclass:
3576 prim_return_value = null();
3577 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3578 break;
3579 case vmIntrinsics::_getClassAccessFlags:
3580 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3581 return_type = TypeInt::INT; // not bool! 6297094
3582 break;
3583 default:
3584 fatal_unexpected_iid(id);
3585 break;
3586 }
3587
3588 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3589 if (mirror_con == NULL) return false; // cannot happen?
3590
3591 #ifndef PRODUCT
3592 if (C->print_intrinsics() || C->print_inlining()) {
3593 ciType* k = mirror_con->java_mirror_type();
3594 if (k) {
3595 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3596 k->print_name();
3597 tty->cr();
3598 }
3599 }
3600 #endif
3601
3602 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3603 RegionNode* region = new RegionNode(PATH_LIMIT);
3604 record_for_igvn(region);
3605 PhiNode* phi = new PhiNode(region, return_type);
3606
3607 // The mirror will never be null of Reflection.getClassAccessFlags, however
3608 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3609 // if it is. See bug 4774291.
3610
3611 // For Reflection.getClassAccessFlags(), the null check occurs in
3612 // the wrong place; see inline_unsafe_access(), above, for a similar
3613 // situation.
3614 mirror = null_check(mirror);
3615 // If mirror or obj is dead, only null-path is taken.
3616 if (stopped()) return true;
3617
3618 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3619
3620 // Now load the mirror's klass metaobject, and null-check it.
3621 // Side-effects region with the control path if the klass is null.
3622 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3623 // If kls is null, we have a primitive mirror.
3624 phi->init_req(_prim_path, prim_return_value);
3625 if (stopped()) { set_result(region, phi); return true; }
3626 bool safe_for_replace = (region->in(_prim_path) == top());
3627
3628 Node* p; // handy temp
3629 Node* null_ctl;
3630
3631 // Now that we have the non-null klass, we can perform the real query.
3632 // For constant classes, the query will constant-fold in LoadNode::Value.
3633 Node* query_value = top();
3634 switch (id) {
3635 case vmIntrinsics::_isInstance:
3636 // nothing is an instance of a primitive type
3637 query_value = gen_instanceof(obj, kls, safe_for_replace);
3638 break;
3639
3640 case vmIntrinsics::_getModifiers:
3641 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3642 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3643 break;
3644
3645 case vmIntrinsics::_isInterface:
3646 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3647 if (generate_interface_guard(kls, region) != NULL)
3648 // A guard was added. If the guard is taken, it was an interface.
3649 phi->add_req(intcon(1));
3650 // If we fall through, it's a plain class.
3651 query_value = intcon(0);
3652 break;
3653
3654 case vmIntrinsics::_isArray:
3655 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3656 if (generate_array_guard(kls, region) != NULL)
3657 // A guard was added. If the guard is taken, it was an array.
3658 phi->add_req(intcon(1));
3659 // If we fall through, it's a plain class.
3660 query_value = intcon(0);
3661 break;
3662
3663 case vmIntrinsics::_isPrimitive:
3664 query_value = intcon(0); // "normal" path produces false
3665 break;
3666
3667 case vmIntrinsics::_getSuperclass:
3668 // The rules here are somewhat unfortunate, but we can still do better
3669 // with random logic than with a JNI call.
3670 // Interfaces store null or Object as _super, but must report null.
3671 // Arrays store an intermediate super as _super, but must report Object.
3672 // Other types can report the actual _super.
3673 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3674 if (generate_interface_guard(kls, region) != NULL)
3675 // A guard was added. If the guard is taken, it was an interface.
3676 phi->add_req(null());
3677 if (generate_array_guard(kls, region) != NULL)
3678 // A guard was added. If the guard is taken, it was an array.
3679 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3680 // If we fall through, it's a plain class. Get its _super.
3681 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3682 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3683 null_ctl = top();
3684 kls = null_check_oop(kls, &null_ctl);
3685 if (null_ctl != top()) {
3686 // If the guard is taken, Object.superClass is null (both klass and mirror).
3687 region->add_req(null_ctl);
3688 phi ->add_req(null());
3689 }
3690 if (!stopped()) {
3691 query_value = load_mirror_from_klass(kls);
3692 }
3693 break;
3694
3695 case vmIntrinsics::_getClassAccessFlags:
3696 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3697 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3698 break;
3699
3700 default:
3701 fatal_unexpected_iid(id);
3702 break;
3703 }
3704
3705 // Fall-through is the normal case of a query to a real class.
3706 phi->init_req(1, query_value);
3707 region->init_req(1, control());
3708
3709 C->set_has_split_ifs(true); // Has chance for split-if optimization
3710 set_result(region, phi);
3711 return true;
3712 }
3713
3714 //-------------------------inline_Class_cast-------------------
3715 bool LibraryCallKit::inline_Class_cast() {
3716 Node* mirror = argument(0); // Class
3717 Node* obj = argument(1);
3718 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3719 if (mirror_con == NULL) {
3720 return false; // dead path (mirror->is_top()).
3721 }
3722 if (obj == NULL || obj->is_top()) {
3723 return false; // dead path
3724 }
3725 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3726
3727 // First, see if Class.cast() can be folded statically.
3728 // java_mirror_type() returns non-null for compile-time Class constants.
3729 ciType* tm = mirror_con->java_mirror_type();
3730 if (tm != NULL && tm->is_klass() &&
3731 tp != NULL && tp->klass() != NULL) {
3732 if (!tp->klass()->is_loaded()) {
3733 // Don't use intrinsic when class is not loaded.
3734 return false;
3735 } else {
3736 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3737 if (static_res == Compile::SSC_always_true) {
3738 // isInstance() is true - fold the code.
3739 set_result(obj);
3740 return true;
3741 } else if (static_res == Compile::SSC_always_false) {
3742 // Don't use intrinsic, have to throw ClassCastException.
3743 // If the reference is null, the non-intrinsic bytecode will
3744 // be optimized appropriately.
3745 return false;
3746 }
3747 }
3748 }
3749
3750 // Bailout intrinsic and do normal inlining if exception path is frequent.
3751 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3752 return false;
3753 }
3754
3755 // Generate dynamic checks.
3756 // Class.cast() is java implementation of _checkcast bytecode.
3757 // Do checkcast (Parse::do_checkcast()) optimizations here.
3758
3759 mirror = null_check(mirror);
3760 // If mirror is dead, only null-path is taken.
3761 if (stopped()) {
3762 return true;
3763 }
3764
3765 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3766 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3767 RegionNode* region = new RegionNode(PATH_LIMIT);
3768 record_for_igvn(region);
3769
3770 // Now load the mirror's klass metaobject, and null-check it.
3771 // If kls is null, we have a primitive mirror and
3772 // nothing is an instance of a primitive type.
3773 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3774
3775 Node* res = top();
3776 if (!stopped()) {
3777 Node* bad_type_ctrl = top();
3778 // Do checkcast optimizations.
3779 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3780 region->init_req(_bad_type_path, bad_type_ctrl);
3781 }
3782 if (region->in(_prim_path) != top() ||
3783 region->in(_bad_type_path) != top()) {
3784 // Let Interpreter throw ClassCastException.
3785 PreserveJVMState pjvms(this);
3786 set_control(_gvn.transform(region));
3787 uncommon_trap(Deoptimization::Reason_intrinsic,
3788 Deoptimization::Action_maybe_recompile);
3789 }
3790 if (!stopped()) {
3791 set_result(res);
3792 }
3793 return true;
3794 }
3795
3796
3797 //--------------------------inline_native_subtype_check------------------------
3798 // This intrinsic takes the JNI calls out of the heart of
3799 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3800 bool LibraryCallKit::inline_native_subtype_check() {
3801 // Pull both arguments off the stack.
3802 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3803 args[0] = argument(0);
3804 args[1] = argument(1);
3805 Node* klasses[2]; // corresponding Klasses: superk, subk
3806 klasses[0] = klasses[1] = top();
3807
3808 enum {
3809 // A full decision tree on {superc is prim, subc is prim}:
3810 _prim_0_path = 1, // {P,N} => false
3811 // {P,P} & superc!=subc => false
3812 _prim_same_path, // {P,P} & superc==subc => true
3813 _prim_1_path, // {N,P} => false
3814 _ref_subtype_path, // {N,N} & subtype check wins => true
3815 _both_ref_path, // {N,N} & subtype check loses => false
3816 PATH_LIMIT
3817 };
3818
3819 RegionNode* region = new RegionNode(PATH_LIMIT);
3820 Node* phi = new PhiNode(region, TypeInt::BOOL);
3821 record_for_igvn(region);
3822
3823 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3824 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3825 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3826
3827 // First null-check both mirrors and load each mirror's klass metaobject.
3828 int which_arg;
3829 for (which_arg = 0; which_arg <= 1; which_arg++) {
3830 Node* arg = args[which_arg];
3831 arg = null_check(arg);
3832 if (stopped()) break;
3833 args[which_arg] = arg;
3834
3835 Node* p = basic_plus_adr(arg, class_klass_offset);
3836 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3837 klasses[which_arg] = _gvn.transform(kls);
3838 }
3839
3840 // Having loaded both klasses, test each for null.
3841 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3842 for (which_arg = 0; which_arg <= 1; which_arg++) {
3843 Node* kls = klasses[which_arg];
3844 Node* null_ctl = top();
3845 kls = null_check_oop(kls, &null_ctl, never_see_null);
3846 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3847 region->init_req(prim_path, null_ctl);
3848 if (stopped()) break;
3849 klasses[which_arg] = kls;
3850 }
3851
3852 if (!stopped()) {
3853 // now we have two reference types, in klasses[0..1]
3854 Node* subk = klasses[1]; // the argument to isAssignableFrom
3855 Node* superk = klasses[0]; // the receiver
3856 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3857 // now we have a successful reference subtype check
3858 region->set_req(_ref_subtype_path, control());
3859 }
3860
3861 // If both operands are primitive (both klasses null), then
3862 // we must return true when they are identical primitives.
3863 // It is convenient to test this after the first null klass check.
3864 set_control(region->in(_prim_0_path)); // go back to first null check
3865 if (!stopped()) {
3866 // Since superc is primitive, make a guard for the superc==subc case.
3867 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3868 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3869 generate_guard(bol_eq, region, PROB_FAIR);
3870 if (region->req() == PATH_LIMIT+1) {
3871 // A guard was added. If the added guard is taken, superc==subc.
3872 region->swap_edges(PATH_LIMIT, _prim_same_path);
3873 region->del_req(PATH_LIMIT);
3874 }
3875 region->set_req(_prim_0_path, control()); // Not equal after all.
3876 }
3877
3878 // these are the only paths that produce 'true':
3879 phi->set_req(_prim_same_path, intcon(1));
3880 phi->set_req(_ref_subtype_path, intcon(1));
3881
3882 // pull together the cases:
3883 assert(region->req() == PATH_LIMIT, "sane region");
3884 for (uint i = 1; i < region->req(); i++) {
3885 Node* ctl = region->in(i);
3886 if (ctl == NULL || ctl == top()) {
3887 region->set_req(i, top());
3888 phi ->set_req(i, top());
3889 } else if (phi->in(i) == NULL) {
3890 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3891 }
3892 }
3893
3894 set_control(_gvn.transform(region));
3895 set_result(_gvn.transform(phi));
3896 return true;
3897 }
3898
3899 //---------------------generate_array_guard_common------------------------
3900 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3901 bool obj_array, bool not_array) {
3902
3903 if (stopped()) {
3904 return NULL;
3905 }
3906
3907 // If obj_array/non_array==false/false:
3908 // Branch around if the given klass is in fact an array (either obj or prim).
3909 // If obj_array/non_array==false/true:
3910 // Branch around if the given klass is not an array klass of any kind.
3911 // If obj_array/non_array==true/true:
3912 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3913 // If obj_array/non_array==true/false:
3914 // Branch around if the kls is an oop array (Object[] or subtype)
3915 //
3916 // Like generate_guard, adds a new path onto the region.
3917 jint layout_con = 0;
3918 Node* layout_val = get_layout_helper(kls, layout_con);
3919 if (layout_val == NULL) {
3920 bool query = (obj_array
3921 ? Klass::layout_helper_is_objArray(layout_con)
3922 : Klass::layout_helper_is_array(layout_con));
3923 if (query == not_array) {
3924 return NULL; // never a branch
3925 } else { // always a branch
3926 Node* always_branch = control();
3927 if (region != NULL)
3928 region->add_req(always_branch);
3929 set_control(top());
3930 return always_branch;
3931 }
3932 }
3933 // Now test the correct condition.
3934 jint nval = (obj_array
3935 ? (jint)(Klass::_lh_array_tag_type_value
3936 << Klass::_lh_array_tag_shift)
3937 : Klass::_lh_neutral_value);
3938 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3939 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3940 // invert the test if we are looking for a non-array
3941 if (not_array) btest = BoolTest(btest).negate();
3942 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3943 return generate_fair_guard(bol, region);
3944 }
3945
3946
3947 //-----------------------inline_native_newArray--------------------------
3948 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3949 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3950 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3951 Node* mirror;
3952 Node* count_val;
3953 if (uninitialized) {
3954 mirror = argument(1);
3955 count_val = argument(2);
3956 } else {
3957 mirror = argument(0);
3958 count_val = argument(1);
3959 }
3960
3961 mirror = null_check(mirror);
3962 // If mirror or obj is dead, only null-path is taken.
3963 if (stopped()) return true;
3964
3965 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3966 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3967 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3968 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3969 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3970
3971 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3972 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3973 result_reg, _slow_path);
3974 Node* normal_ctl = control();
3975 Node* no_array_ctl = result_reg->in(_slow_path);
3976
3977 // Generate code for the slow case. We make a call to newArray().
3978 set_control(no_array_ctl);
3979 if (!stopped()) {
3980 // Either the input type is void.class, or else the
3981 // array klass has not yet been cached. Either the
3982 // ensuing call will throw an exception, or else it
3983 // will cache the array klass for next time.
3984 PreserveJVMState pjvms(this);
3985 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3986 Node* slow_result = set_results_for_java_call(slow_call);
3987 // this->control() comes from set_results_for_java_call
3988 result_reg->set_req(_slow_path, control());
3989 result_val->set_req(_slow_path, slow_result);
3990 result_io ->set_req(_slow_path, i_o());
3991 result_mem->set_req(_slow_path, reset_memory());
3992 }
3993
3994 set_control(normal_ctl);
3995 if (!stopped()) {
3996 // Normal case: The array type has been cached in the java.lang.Class.
3997 // The following call works fine even if the array type is polymorphic.
3998 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3999 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
4000 result_reg->init_req(_normal_path, control());
4001 result_val->init_req(_normal_path, obj);
4002 result_io ->init_req(_normal_path, i_o());
4003 result_mem->init_req(_normal_path, reset_memory());
4004
4005 if (uninitialized) {
4006 // Mark the allocation so that zeroing is skipped
4007 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4008 alloc->maybe_set_complete(&_gvn);
4009 }
4010 }
4011
4012 // Return the combined state.
4013 set_i_o( _gvn.transform(result_io) );
4014 set_all_memory( _gvn.transform(result_mem));
4015
4016 C->set_has_split_ifs(true); // Has chance for split-if optimization
4017 set_result(result_reg, result_val);
4018 return true;
4019 }
4020
4021 //----------------------inline_native_getLength--------------------------
4022 // public static native int java.lang.reflect.Array.getLength(Object array);
4023 bool LibraryCallKit::inline_native_getLength() {
4024 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4025
4026 Node* array = null_check(argument(0));
4027 // If array is dead, only null-path is taken.
4028 if (stopped()) return true;
4029
4030 // Deoptimize if it is a non-array.
4031 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4032
4033 if (non_array != NULL) {
4034 PreserveJVMState pjvms(this);
4035 set_control(non_array);
4036 uncommon_trap(Deoptimization::Reason_intrinsic,
4037 Deoptimization::Action_maybe_recompile);
4038 }
4039
4040 // If control is dead, only non-array-path is taken.
4041 if (stopped()) return true;
4042
4043 // The works fine even if the array type is polymorphic.
4044 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4045 Node* result = load_array_length(array);
4046
4047 C->set_has_split_ifs(true); // Has chance for split-if optimization
4048 set_result(result);
4049 return true;
4050 }
4051
4052 //------------------------inline_array_copyOf----------------------------
4053 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4054 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4055 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4056 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4057
4058 // Get the arguments.
4059 Node* original = argument(0);
4060 Node* start = is_copyOfRange? argument(1): intcon(0);
4061 Node* end = is_copyOfRange? argument(2): argument(1);
4062 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4063
4064 Node* newcopy = NULL;
4065
4066 // Set the original stack and the reexecute bit for the interpreter to reexecute
4067 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4068 { PreserveReexecuteState preexecs(this);
4069 jvms()->set_should_reexecute(true);
4070
4071 array_type_mirror = null_check(array_type_mirror);
4072 original = null_check(original);
4073
4074 // Check if a null path was taken unconditionally.
4075 if (stopped()) return true;
4076
4077 Node* orig_length = load_array_length(original);
4078
4079 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4080 klass_node = null_check(klass_node);
4081
4082 RegionNode* bailout = new RegionNode(1);
4083 record_for_igvn(bailout);
4084
4085 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4086 // Bail out if that is so.
4087 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4088 if (not_objArray != NULL) {
4089 // Improve the klass node's type from the new optimistic assumption:
4090 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4091 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4092 Node* cast = new CastPPNode(klass_node, akls);
4093 cast->init_req(0, control());
4094 klass_node = _gvn.transform(cast);
4095 }
4096
4097 // Bail out if either start or end is negative.
4098 generate_negative_guard(start, bailout, &start);
4099 generate_negative_guard(end, bailout, &end);
4100
4101 Node* length = end;
4102 if (_gvn.type(start) != TypeInt::ZERO) {
4103 length = _gvn.transform(new SubINode(end, start));
4104 }
4105
4106 // Bail out if length is negative.
4107 // Without this the new_array would throw
4108 // NegativeArraySizeException but IllegalArgumentException is what
4109 // should be thrown
4110 generate_negative_guard(length, bailout, &length);
4111
4112 if (bailout->req() > 1) {
4113 PreserveJVMState pjvms(this);
4114 set_control(_gvn.transform(bailout));
4115 uncommon_trap(Deoptimization::Reason_intrinsic,
4116 Deoptimization::Action_maybe_recompile);
4117 }
4118
4119 if (!stopped()) {
4120 // How many elements will we copy from the original?
4121 // The answer is MinI(orig_length - start, length).
4122 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4123 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4124
4125 // Generate a direct call to the right arraycopy function(s).
4126 // We know the copy is disjoint but we might not know if the
4127 // oop stores need checking.
4128 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4129 // This will fail a store-check if x contains any non-nulls.
4130
4131 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4132 // loads/stores but it is legal only if we're sure the
4133 // Arrays.copyOf would succeed. So we need all input arguments
4134 // to the copyOf to be validated, including that the copy to the
4135 // new array won't trigger an ArrayStoreException. That subtype
4136 // check can be optimized if we know something on the type of
4137 // the input array from type speculation.
4138 if (_gvn.type(klass_node)->singleton()) {
4139 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4140 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4141
4142 int test = C->static_subtype_check(superk, subk);
4143 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4144 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4145 if (t_original->speculative_type() != NULL) {
4146 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4147 }
4148 }
4149 }
4150
4151 bool validated = false;
4152 // Reason_class_check rather than Reason_intrinsic because we
4153 // want to intrinsify even if this traps.
4154 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4155 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4156 klass_node);
4157
4158 if (not_subtype_ctrl != top()) {
4159 PreserveJVMState pjvms(this);
4160 set_control(not_subtype_ctrl);
4161 uncommon_trap(Deoptimization::Reason_class_check,
4162 Deoptimization::Action_make_not_entrant);
4163 assert(stopped(), "Should be stopped");
4164 }
4165 validated = true;
4166 }
4167
4168 if (!stopped()) {
4169 newcopy = new_array(klass_node, length, 0); // no arguments to push
4170
4171 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4172 load_object_klass(original), klass_node);
4173 if (!is_copyOfRange) {
4174 ac->set_copyof(validated);
4175 } else {
4176 ac->set_copyofrange(validated);
4177 }
4178 Node* n = _gvn.transform(ac);
4179 if (n == ac) {
4180 ac->connect_outputs(this);
4181 } else {
4182 assert(validated, "shouldn't transform if all arguments not validated");
4183 set_all_memory(n);
4184 }
4185 }
4186 }
4187 } // original reexecute is set back here
4188
4189 C->set_has_split_ifs(true); // Has chance for split-if optimization
4190 if (!stopped()) {
4191 set_result(newcopy);
4192 }
4193 return true;
4194 }
4195
4196
4197 //----------------------generate_virtual_guard---------------------------
4198 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4199 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4200 RegionNode* slow_region) {
4201 ciMethod* method = callee();
4202 int vtable_index = method->vtable_index();
4203 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4204 "bad index %d", vtable_index);
4205 // Get the Method* out of the appropriate vtable entry.
4206 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4207 vtable_index*vtableEntry::size_in_bytes() +
4208 vtableEntry::method_offset_in_bytes();
4209 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4210 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4211
4212 // Compare the target method with the expected method (e.g., Object.hashCode).
4213 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4214
4215 Node* native_call = makecon(native_call_addr);
4216 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4217 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4218
4219 return generate_slow_guard(test_native, slow_region);
4220 }
4221
4222 //-----------------------generate_method_call----------------------------
4223 // Use generate_method_call to make a slow-call to the real
4224 // method if the fast path fails. An alternative would be to
4225 // use a stub like OptoRuntime::slow_arraycopy_Java.
4226 // This only works for expanding the current library call,
4227 // not another intrinsic. (E.g., don't use this for making an
4228 // arraycopy call inside of the copyOf intrinsic.)
4229 CallJavaNode*
4230 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4231 // When compiling the intrinsic method itself, do not use this technique.
4232 guarantee(callee() != C->method(), "cannot make slow-call to self");
4233
4234 ciMethod* method = callee();
4235 // ensure the JVMS we have will be correct for this call
4236 guarantee(method_id == method->intrinsic_id(), "must match");
4237
4238 const TypeFunc* tf = TypeFunc::make(method);
4239 CallJavaNode* slow_call;
4240 if (is_static) {
4241 assert(!is_virtual, "");
4242 slow_call = new CallStaticJavaNode(C, tf,
4243 SharedRuntime::get_resolve_static_call_stub(),
4244 method, bci());
4245 } else if (is_virtual) {
4246 null_check_receiver();
4247 int vtable_index = Method::invalid_vtable_index;
4248 if (UseInlineCaches) {
4249 // Suppress the vtable call
4250 } else {
4251 // hashCode and clone are not a miranda methods,
4252 // so the vtable index is fixed.
4253 // No need to use the linkResolver to get it.
4254 vtable_index = method->vtable_index();
4255 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4256 "bad index %d", vtable_index);
4257 }
4258 slow_call = new CallDynamicJavaNode(tf,
4259 SharedRuntime::get_resolve_virtual_call_stub(),
4260 method, vtable_index, bci());
4261 } else { // neither virtual nor static: opt_virtual
4262 null_check_receiver();
4263 slow_call = new CallStaticJavaNode(C, tf,
4264 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4265 method, bci());
4266 slow_call->set_optimized_virtual(true);
4267 }
4268 set_arguments_for_java_call(slow_call);
4269 set_edges_for_java_call(slow_call);
4270 return slow_call;
4271 }
4272
4273
4274 /**
4275 * Build special case code for calls to hashCode on an object. This call may
4276 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4277 * slightly different code.
4278 */
4279 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4280 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4281 assert(!(is_virtual && is_static), "either virtual, special, or static");
4282
4283 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4284
4285 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4286 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4287 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4288 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4289 Node* obj = NULL;
4290 if (!is_static) {
4291 // Check for hashing null object
4292 obj = null_check_receiver();
4293 if (stopped()) return true; // unconditionally null
4294 result_reg->init_req(_null_path, top());
4295 result_val->init_req(_null_path, top());
4296 } else {
4297 // Do a null check, and return zero if null.
4298 // System.identityHashCode(null) == 0
4299 obj = argument(0);
4300 Node* null_ctl = top();
4301 obj = null_check_oop(obj, &null_ctl);
4302 result_reg->init_req(_null_path, null_ctl);
4303 result_val->init_req(_null_path, _gvn.intcon(0));
4304 }
4305
4306 // Unconditionally null? Then return right away.
4307 if (stopped()) {
4308 set_control( result_reg->in(_null_path));
4309 if (!stopped())
4310 set_result(result_val->in(_null_path));
4311 return true;
4312 }
4313
4314 // We only go to the fast case code if we pass a number of guards. The
4315 // paths which do not pass are accumulated in the slow_region.
4316 RegionNode* slow_region = new RegionNode(1);
4317 record_for_igvn(slow_region);
4318
4319 // If this is a virtual call, we generate a funny guard. We pull out
4320 // the vtable entry corresponding to hashCode() from the target object.
4321 // If the target method which we are calling happens to be the native
4322 // Object hashCode() method, we pass the guard. We do not need this
4323 // guard for non-virtual calls -- the caller is known to be the native
4324 // Object hashCode().
4325 if (is_virtual) {
4326 // After null check, get the object's klass.
4327 Node* obj_klass = load_object_klass(obj);
4328 generate_virtual_guard(obj_klass, slow_region);
4329 }
4330
4331 // Get the header out of the object, use LoadMarkNode when available
4332 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4333 // The control of the load must be NULL. Otherwise, the load can move before
4334 // the null check after castPP removal.
4335 Node* no_ctrl = NULL;
4336 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4337
4338 // Test the header to see if it is unlocked.
4339 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4340 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4341 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4342 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4343 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4344
4345 generate_slow_guard(test_unlocked, slow_region);
4346
4347 // Get the hash value and check to see that it has been properly assigned.
4348 // We depend on hash_mask being at most 32 bits and avoid the use of
4349 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4350 // vm: see markOop.hpp.
4351 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4352 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4353 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4354 // This hack lets the hash bits live anywhere in the mark object now, as long
4355 // as the shift drops the relevant bits into the low 32 bits. Note that
4356 // Java spec says that HashCode is an int so there's no point in capturing
4357 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4358 hshifted_header = ConvX2I(hshifted_header);
4359 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4360
4361 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4362 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4363 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4364
4365 generate_slow_guard(test_assigned, slow_region);
4366
4367 Node* init_mem = reset_memory();
4368 // fill in the rest of the null path:
4369 result_io ->init_req(_null_path, i_o());
4370 result_mem->init_req(_null_path, init_mem);
4371
4372 result_val->init_req(_fast_path, hash_val);
4373 result_reg->init_req(_fast_path, control());
4374 result_io ->init_req(_fast_path, i_o());
4375 result_mem->init_req(_fast_path, init_mem);
4376
4377 // Generate code for the slow case. We make a call to hashCode().
4378 set_control(_gvn.transform(slow_region));
4379 if (!stopped()) {
4380 // No need for PreserveJVMState, because we're using up the present state.
4381 set_all_memory(init_mem);
4382 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4383 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4384 Node* slow_result = set_results_for_java_call(slow_call);
4385 // this->control() comes from set_results_for_java_call
4386 result_reg->init_req(_slow_path, control());
4387 result_val->init_req(_slow_path, slow_result);
4388 result_io ->set_req(_slow_path, i_o());
4389 result_mem ->set_req(_slow_path, reset_memory());
4390 }
4391
4392 // Return the combined state.
4393 set_i_o( _gvn.transform(result_io) );
4394 set_all_memory( _gvn.transform(result_mem));
4395
4396 set_result(result_reg, result_val);
4397 return true;
4398 }
4399
4400 //---------------------------inline_native_getClass----------------------------
4401 // public final native Class<?> java.lang.Object.getClass();
4402 //
4403 // Build special case code for calls to getClass on an object.
4404 bool LibraryCallKit::inline_native_getClass() {
4405 Node* obj = null_check_receiver();
4406 if (stopped()) return true;
4407 set_result(load_mirror_from_klass(load_object_klass(obj)));
4408 return true;
4409 }
4410
4411 //-----------------inline_native_Reflection_getCallerClass---------------------
4412 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4413 //
4414 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4415 //
4416 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4417 // in that it must skip particular security frames and checks for
4418 // caller sensitive methods.
4419 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4420 #ifndef PRODUCT
4421 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4422 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4423 }
4424 #endif
4425
4426 if (!jvms()->has_method()) {
4427 #ifndef PRODUCT
4428 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4429 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4430 }
4431 #endif
4432 return false;
4433 }
4434
4435 // Walk back up the JVM state to find the caller at the required
4436 // depth.
4437 JVMState* caller_jvms = jvms();
4438
4439 // Cf. JVM_GetCallerClass
4440 // NOTE: Start the loop at depth 1 because the current JVM state does
4441 // not include the Reflection.getCallerClass() frame.
4442 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4443 ciMethod* m = caller_jvms->method();
4444 switch (n) {
4445 case 0:
4446 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4447 break;
4448 case 1:
4449 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4450 if (!m->caller_sensitive()) {
4451 #ifndef PRODUCT
4452 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4453 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4454 }
4455 #endif
4456 return false; // bail-out; let JVM_GetCallerClass do the work
4457 }
4458 break;
4459 default:
4460 if (!m->is_ignored_by_security_stack_walk()) {
4461 // We have reached the desired frame; return the holder class.
4462 // Acquire method holder as java.lang.Class and push as constant.
4463 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4464 ciInstance* caller_mirror = caller_klass->java_mirror();
4465 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4466
4467 #ifndef PRODUCT
4468 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4469 tty->print_cr(" Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth());
4470 tty->print_cr(" JVM state at this point:");
4471 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4472 ciMethod* m = jvms()->of_depth(i)->method();
4473 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4474 }
4475 }
4476 #endif
4477 return true;
4478 }
4479 break;
4480 }
4481 }
4482
4483 #ifndef PRODUCT
4484 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4485 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4486 tty->print_cr(" JVM state at this point:");
4487 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4488 ciMethod* m = jvms()->of_depth(i)->method();
4489 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4490 }
4491 }
4492 #endif
4493
4494 return false; // bail-out; let JVM_GetCallerClass do the work
4495 }
4496
4497 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4498 Node* arg = argument(0);
4499 Node* result = NULL;
4500
4501 switch (id) {
4502 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4503 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4504 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4505 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4506
4507 case vmIntrinsics::_doubleToLongBits: {
4508 // two paths (plus control) merge in a wood
4509 RegionNode *r = new RegionNode(3);
4510 Node *phi = new PhiNode(r, TypeLong::LONG);
4511
4512 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4513 // Build the boolean node
4514 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4515
4516 // Branch either way.
4517 // NaN case is less traveled, which makes all the difference.
4518 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4519 Node *opt_isnan = _gvn.transform(ifisnan);
4520 assert( opt_isnan->is_If(), "Expect an IfNode");
4521 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4522 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4523
4524 set_control(iftrue);
4525
4526 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4527 Node *slow_result = longcon(nan_bits); // return NaN
4528 phi->init_req(1, _gvn.transform( slow_result ));
4529 r->init_req(1, iftrue);
4530
4531 // Else fall through
4532 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4533 set_control(iffalse);
4534
4535 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4536 r->init_req(2, iffalse);
4537
4538 // Post merge
4539 set_control(_gvn.transform(r));
4540 record_for_igvn(r);
4541
4542 C->set_has_split_ifs(true); // Has chance for split-if optimization
4543 result = phi;
4544 assert(result->bottom_type()->isa_long(), "must be");
4545 break;
4546 }
4547
4548 case vmIntrinsics::_floatToIntBits: {
4549 // two paths (plus control) merge in a wood
4550 RegionNode *r = new RegionNode(3);
4551 Node *phi = new PhiNode(r, TypeInt::INT);
4552
4553 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4554 // Build the boolean node
4555 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4556
4557 // Branch either way.
4558 // NaN case is less traveled, which makes all the difference.
4559 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4560 Node *opt_isnan = _gvn.transform(ifisnan);
4561 assert( opt_isnan->is_If(), "Expect an IfNode");
4562 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4563 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4564
4565 set_control(iftrue);
4566
4567 static const jint nan_bits = 0x7fc00000;
4568 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4569 phi->init_req(1, _gvn.transform( slow_result ));
4570 r->init_req(1, iftrue);
4571
4572 // Else fall through
4573 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4574 set_control(iffalse);
4575
4576 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4577 r->init_req(2, iffalse);
4578
4579 // Post merge
4580 set_control(_gvn.transform(r));
4581 record_for_igvn(r);
4582
4583 C->set_has_split_ifs(true); // Has chance for split-if optimization
4584 result = phi;
4585 assert(result->bottom_type()->isa_int(), "must be");
4586 break;
4587 }
4588
4589 default:
4590 fatal_unexpected_iid(id);
4591 break;
4592 }
4593 set_result(_gvn.transform(result));
4594 return true;
4595 }
4596
4597 //----------------------inline_unsafe_copyMemory-------------------------
4598 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4599 bool LibraryCallKit::inline_unsafe_copyMemory() {
4600 if (callee()->is_static()) return false; // caller must have the capability!
4601 null_check_receiver(); // null-check receiver
4602 if (stopped()) return true;
4603
4604 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4605
4606 Node* src_ptr = argument(1); // type: oop
4607 Node* src_off = ConvL2X(argument(2)); // type: long
4608 Node* dst_ptr = argument(4); // type: oop
4609 Node* dst_off = ConvL2X(argument(5)); // type: long
4610 Node* size = ConvL2X(argument(7)); // type: long
4611
4612 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4613 "fieldOffset must be byte-scaled");
4614
4615 Node* src = make_unsafe_address(src_ptr, src_off);
4616 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4617
4618 // Conservatively insert a memory barrier on all memory slices.
4619 // Do not let writes of the copy source or destination float below the copy.
4620 insert_mem_bar(Op_MemBarCPUOrder);
4621
4622 // Call it. Note that the length argument is not scaled.
4623 make_runtime_call(RC_LEAF|RC_NO_FP,
4624 OptoRuntime::fast_arraycopy_Type(),
4625 StubRoutines::unsafe_arraycopy(),
4626 "unsafe_arraycopy",
4627 TypeRawPtr::BOTTOM,
4628 src, dst, size XTOP);
4629
4630 // Do not let reads of the copy destination float above the copy.
4631 insert_mem_bar(Op_MemBarCPUOrder);
4632
4633 return true;
4634 }
4635
4636 //------------------------clone_coping-----------------------------------
4637 // Helper function for inline_native_clone.
4638 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4639 assert(obj_size != NULL, "");
4640 Node* raw_obj = alloc_obj->in(1);
4641 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4642
4643 AllocateNode* alloc = NULL;
4644 if (ReduceBulkZeroing) {
4645 // We will be completely responsible for initializing this object -
4646 // mark Initialize node as complete.
4647 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4648 // The object was just allocated - there should be no any stores!
4649 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4650 // Mark as complete_with_arraycopy so that on AllocateNode
4651 // expansion, we know this AllocateNode is initialized by an array
4652 // copy and a StoreStore barrier exists after the array copy.
4653 alloc->initialization()->set_complete_with_arraycopy();
4654 }
4655
4656 // Copy the fastest available way.
4657 // TODO: generate fields copies for small objects instead.
4658 Node* src = obj;
4659 Node* dest = alloc_obj;
4660 Node* size = _gvn.transform(obj_size);
4661
4662 // Exclude the header but include array length to copy by 8 bytes words.
4663 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4664 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4665 instanceOopDesc::base_offset_in_bytes();
4666 // base_off:
4667 // 8 - 32-bit VM
4668 // 12 - 64-bit VM, compressed klass
4669 // 16 - 64-bit VM, normal klass
4670 if (base_off % BytesPerLong != 0) {
4671 assert(UseCompressedClassPointers, "");
4672 if (is_array) {
4673 // Exclude length to copy by 8 bytes words.
4674 base_off += sizeof(int);
4675 } else {
4676 // Include klass to copy by 8 bytes words.
4677 base_off = instanceOopDesc::klass_offset_in_bytes();
4678 }
4679 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4680 }
4681 src = basic_plus_adr(src, base_off);
4682 dest = basic_plus_adr(dest, base_off);
4683
4684 // Compute the length also, if needed:
4685 Node* countx = size;
4686 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4687 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4688
4689 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4690
4691 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4692 ac->set_clonebasic();
4693 Node* n = _gvn.transform(ac);
4694 if (n == ac) {
4695 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4696 } else {
4697 set_all_memory(n);
4698 }
4699
4700 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4701 if (card_mark) {
4702 assert(!is_array, "");
4703 // Put in store barrier for any and all oops we are sticking
4704 // into this object. (We could avoid this if we could prove
4705 // that the object type contains no oop fields at all.)
4706 Node* no_particular_value = NULL;
4707 Node* no_particular_field = NULL;
4708 int raw_adr_idx = Compile::AliasIdxRaw;
4709 post_barrier(control(),
4710 memory(raw_adr_type),
4711 alloc_obj,
4712 no_particular_field,
4713 raw_adr_idx,
4714 no_particular_value,
4715 T_OBJECT,
4716 false);
4717 }
4718
4719 // Do not let reads from the cloned object float above the arraycopy.
4720 if (alloc != NULL) {
4721 // Do not let stores that initialize this object be reordered with
4722 // a subsequent store that would make this object accessible by
4723 // other threads.
4724 // Record what AllocateNode this StoreStore protects so that
4725 // escape analysis can go from the MemBarStoreStoreNode to the
4726 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4727 // based on the escape status of the AllocateNode.
4728 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4729 } else {
4730 insert_mem_bar(Op_MemBarCPUOrder);
4731 }
4732 }
4733
4734 //------------------------inline_native_clone----------------------------
4735 // protected native Object java.lang.Object.clone();
4736 //
4737 // Here are the simple edge cases:
4738 // null receiver => normal trap
4739 // virtual and clone was overridden => slow path to out-of-line clone
4740 // not cloneable or finalizer => slow path to out-of-line Object.clone
4741 //
4742 // The general case has two steps, allocation and copying.
4743 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4744 //
4745 // Copying also has two cases, oop arrays and everything else.
4746 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4747 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4748 //
4749 // These steps fold up nicely if and when the cloned object's klass
4750 // can be sharply typed as an object array, a type array, or an instance.
4751 //
4752 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4753 PhiNode* result_val;
4754
4755 // Set the reexecute bit for the interpreter to reexecute
4756 // the bytecode that invokes Object.clone if deoptimization happens.
4757 { PreserveReexecuteState preexecs(this);
4758 jvms()->set_should_reexecute(true);
4759
4760 Node* obj = null_check_receiver();
4761 if (stopped()) return true;
4762
4763 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4764
4765 // If we are going to clone an instance, we need its exact type to
4766 // know the number and types of fields to convert the clone to
4767 // loads/stores. Maybe a speculative type can help us.
4768 if (!obj_type->klass_is_exact() &&
4769 obj_type->speculative_type() != NULL &&
4770 obj_type->speculative_type()->is_instance_klass()) {
4771 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4772 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4773 !spec_ik->has_injected_fields()) {
4774 ciKlass* k = obj_type->klass();
4775 if (!k->is_instance_klass() ||
4776 k->as_instance_klass()->is_interface() ||
4777 k->as_instance_klass()->has_subklass()) {
4778 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4779 }
4780 }
4781 }
4782
4783 Node* obj_klass = load_object_klass(obj);
4784 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4785 const TypeOopPtr* toop = ((tklass != NULL)
4786 ? tklass->as_instance_type()
4787 : TypeInstPtr::NOTNULL);
4788
4789 // Conservatively insert a memory barrier on all memory slices.
4790 // Do not let writes into the original float below the clone.
4791 insert_mem_bar(Op_MemBarCPUOrder);
4792
4793 // paths into result_reg:
4794 enum {
4795 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4796 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4797 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4798 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4799 PATH_LIMIT
4800 };
4801 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4802 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4803 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4804 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4805 record_for_igvn(result_reg);
4806
4807 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4808 int raw_adr_idx = Compile::AliasIdxRaw;
4809
4810 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4811 if (array_ctl != NULL) {
4812 // It's an array.
4813 PreserveJVMState pjvms(this);
4814 set_control(array_ctl);
4815 Node* obj_length = load_array_length(obj);
4816 Node* obj_size = NULL;
4817 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4818
4819 if (!use_ReduceInitialCardMarks()) {
4820 // If it is an oop array, it requires very special treatment,
4821 // because card marking is required on each card of the array.
4822 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4823 if (is_obja != NULL) {
4824 PreserveJVMState pjvms2(this);
4825 set_control(is_obja);
4826 // Generate a direct call to the right arraycopy function(s).
4827 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4828 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4829 ac->set_cloneoop();
4830 Node* n = _gvn.transform(ac);
4831 assert(n == ac, "cannot disappear");
4832 ac->connect_outputs(this);
4833
4834 result_reg->init_req(_objArray_path, control());
4835 result_val->init_req(_objArray_path, alloc_obj);
4836 result_i_o ->set_req(_objArray_path, i_o());
4837 result_mem ->set_req(_objArray_path, reset_memory());
4838 }
4839 }
4840 // Otherwise, there are no card marks to worry about.
4841 // (We can dispense with card marks if we know the allocation
4842 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4843 // causes the non-eden paths to take compensating steps to
4844 // simulate a fresh allocation, so that no further
4845 // card marks are required in compiled code to initialize
4846 // the object.)
4847
4848 if (!stopped()) {
4849 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4850
4851 // Present the results of the copy.
4852 result_reg->init_req(_array_path, control());
4853 result_val->init_req(_array_path, alloc_obj);
4854 result_i_o ->set_req(_array_path, i_o());
4855 result_mem ->set_req(_array_path, reset_memory());
4856 }
4857 }
4858
4859 // We only go to the instance fast case code if we pass a number of guards.
4860 // The paths which do not pass are accumulated in the slow_region.
4861 RegionNode* slow_region = new RegionNode(1);
4862 record_for_igvn(slow_region);
4863 if (!stopped()) {
4864 // It's an instance (we did array above). Make the slow-path tests.
4865 // If this is a virtual call, we generate a funny guard. We grab
4866 // the vtable entry corresponding to clone() from the target object.
4867 // If the target method which we are calling happens to be the
4868 // Object clone() method, we pass the guard. We do not need this
4869 // guard for non-virtual calls; the caller is known to be the native
4870 // Object clone().
4871 if (is_virtual) {
4872 generate_virtual_guard(obj_klass, slow_region);
4873 }
4874
4875 // The object must be easily cloneable and must not have a finalizer.
4876 // Both of these conditions may be checked in a single test.
4877 // We could optimize the test further, but we don't care.
4878 generate_access_flags_guard(obj_klass,
4879 // Test both conditions:
4880 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4881 // Must be cloneable but not finalizer:
4882 JVM_ACC_IS_CLONEABLE_FAST,
4883 slow_region);
4884 }
4885
4886 if (!stopped()) {
4887 // It's an instance, and it passed the slow-path tests.
4888 PreserveJVMState pjvms(this);
4889 Node* obj_size = NULL;
4890 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4891 // is reexecuted if deoptimization occurs and there could be problems when merging
4892 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4893 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4894
4895 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4896
4897 // Present the results of the slow call.
4898 result_reg->init_req(_instance_path, control());
4899 result_val->init_req(_instance_path, alloc_obj);
4900 result_i_o ->set_req(_instance_path, i_o());
4901 result_mem ->set_req(_instance_path, reset_memory());
4902 }
4903
4904 // Generate code for the slow case. We make a call to clone().
4905 set_control(_gvn.transform(slow_region));
4906 if (!stopped()) {
4907 PreserveJVMState pjvms(this);
4908 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4909 Node* slow_result = set_results_for_java_call(slow_call);
4910 // this->control() comes from set_results_for_java_call
4911 result_reg->init_req(_slow_path, control());
4912 result_val->init_req(_slow_path, slow_result);
4913 result_i_o ->set_req(_slow_path, i_o());
4914 result_mem ->set_req(_slow_path, reset_memory());
4915 }
4916
4917 // Return the combined state.
4918 set_control( _gvn.transform(result_reg));
4919 set_i_o( _gvn.transform(result_i_o));
4920 set_all_memory( _gvn.transform(result_mem));
4921 } // original reexecute is set back here
4922
4923 set_result(_gvn.transform(result_val));
4924 return true;
4925 }
4926
4927 // If we have a tighly coupled allocation, the arraycopy may take care
4928 // of the array initialization. If one of the guards we insert between
4929 // the allocation and the arraycopy causes a deoptimization, an
4930 // unitialized array will escape the compiled method. To prevent that
4931 // we set the JVM state for uncommon traps between the allocation and
4932 // the arraycopy to the state before the allocation so, in case of
4933 // deoptimization, we'll reexecute the allocation and the
4934 // initialization.
4935 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4936 if (alloc != NULL) {
4937 ciMethod* trap_method = alloc->jvms()->method();
4938 int trap_bci = alloc->jvms()->bci();
4939
4940 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4941 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4942 // Make sure there's no store between the allocation and the
4943 // arraycopy otherwise visible side effects could be rexecuted
4944 // in case of deoptimization and cause incorrect execution.
4945 bool no_interfering_store = true;
4946 Node* mem = alloc->in(TypeFunc::Memory);
4947 if (mem->is_MergeMem()) {
4948 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4949 Node* n = mms.memory();
4950 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4951 assert(n->is_Store(), "what else?");
4952 no_interfering_store = false;
4953 break;
4954 }
4955 }
4956 } else {
4957 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4958 Node* n = mms.memory();
4959 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4960 assert(n->is_Store(), "what else?");
4961 no_interfering_store = false;
4962 break;
4963 }
4964 }
4965 }
4966
4967 if (no_interfering_store) {
4968 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4969 uint size = alloc->req();
4970 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4971 old_jvms->set_map(sfpt);
4972 for (uint i = 0; i < size; i++) {
4973 sfpt->init_req(i, alloc->in(i));
4974 }
4975 // re-push array length for deoptimization
4976 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4977 old_jvms->set_sp(old_jvms->sp()+1);
4978 old_jvms->set_monoff(old_jvms->monoff()+1);
4979 old_jvms->set_scloff(old_jvms->scloff()+1);
4980 old_jvms->set_endoff(old_jvms->endoff()+1);
4981 old_jvms->set_should_reexecute(true);
4982
4983 sfpt->set_i_o(map()->i_o());
4984 sfpt->set_memory(map()->memory());
4985 sfpt->set_control(map()->control());
4986
4987 JVMState* saved_jvms = jvms();
4988 saved_reexecute_sp = _reexecute_sp;
4989
4990 set_jvms(sfpt->jvms());
4991 _reexecute_sp = jvms()->sp();
4992
4993 return saved_jvms;
4994 }
4995 }
4996 }
4997 return NULL;
4998 }
4999
5000 // In case of a deoptimization, we restart execution at the
5001 // allocation, allocating a new array. We would leave an uninitialized
5002 // array in the heap that GCs wouldn't expect. Move the allocation
5003 // after the traps so we don't allocate the array if we
5004 // deoptimize. This is possible because tightly_coupled_allocation()
5005 // guarantees there's no observer of the allocated array at this point
5006 // and the control flow is simple enough.
5007 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
5008 int saved_reexecute_sp, uint new_idx) {
5009 if (saved_jvms != NULL && !stopped()) {
5010 assert(alloc != NULL, "only with a tightly coupled allocation");
5011 // restore JVM state to the state at the arraycopy
5012 saved_jvms->map()->set_control(map()->control());
5013 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5014 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5015 // If we've improved the types of some nodes (null check) while
5016 // emitting the guards, propagate them to the current state
5017 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
5018 set_jvms(saved_jvms);
5019 _reexecute_sp = saved_reexecute_sp;
5020
5021 // Remove the allocation from above the guards
5022 CallProjections callprojs;
5023 alloc->extract_projections(&callprojs, true);
5024 InitializeNode* init = alloc->initialization();
5025 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5026 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5027 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5028 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5029
5030 // move the allocation here (after the guards)
5031 _gvn.hash_delete(alloc);
5032 alloc->set_req(TypeFunc::Control, control());
5033 alloc->set_req(TypeFunc::I_O, i_o());
5034 Node *mem = reset_memory();
5035 set_all_memory(mem);
5036 alloc->set_req(TypeFunc::Memory, mem);
5037 set_control(init->proj_out_or_null(TypeFunc::Control));
5038 set_i_o(callprojs.fallthrough_ioproj);
5039
5040 // Update memory as done in GraphKit::set_output_for_allocation()
5041 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5042 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5043 if (ary_type->isa_aryptr() && length_type != NULL) {
5044 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5045 }
5046 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5047 int elemidx = C->get_alias_index(telemref);
5048 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
5049 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
5050
5051 Node* allocx = _gvn.transform(alloc);
5052 assert(allocx == alloc, "where has the allocation gone?");
5053 assert(dest->is_CheckCastPP(), "not an allocation result?");
5054
5055 _gvn.hash_delete(dest);
5056 dest->set_req(0, control());
5057 Node* destx = _gvn.transform(dest);
5058 assert(destx == dest, "where has the allocation result gone?");
5059 }
5060 }
5061
5062
5063 //------------------------------inline_arraycopy-----------------------
5064 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5065 // Object dest, int destPos,
5066 // int length);
5067 bool LibraryCallKit::inline_arraycopy() {
5068 // Get the arguments.
5069 Node* src = argument(0); // type: oop
5070 Node* src_offset = argument(1); // type: int
5071 Node* dest = argument(2); // type: oop
5072 Node* dest_offset = argument(3); // type: int
5073 Node* length = argument(4); // type: int
5074
5075 uint new_idx = C->unique();
5076
5077 // Check for allocation before we add nodes that would confuse
5078 // tightly_coupled_allocation()
5079 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5080
5081 int saved_reexecute_sp = -1;
5082 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5083 // See arraycopy_restore_alloc_state() comment
5084 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5085 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5086 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5087 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5088
5089 // The following tests must be performed
5090 // (1) src and dest are arrays.
5091 // (2) src and dest arrays must have elements of the same BasicType
5092 // (3) src and dest must not be null.
5093 // (4) src_offset must not be negative.
5094 // (5) dest_offset must not be negative.
5095 // (6) length must not be negative.
5096 // (7) src_offset + length must not exceed length of src.
5097 // (8) dest_offset + length must not exceed length of dest.
5098 // (9) each element of an oop array must be assignable
5099
5100 // (3) src and dest must not be null.
5101 // always do this here because we need the JVM state for uncommon traps
5102 Node* null_ctl = top();
5103 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5104 assert(null_ctl->is_top(), "no null control here");
5105 dest = null_check(dest, T_ARRAY);
5106
5107 if (!can_emit_guards) {
5108 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5109 // guards but the arraycopy node could still take advantage of a
5110 // tightly allocated allocation. tightly_coupled_allocation() is
5111 // called again to make sure it takes the null check above into
5112 // account: the null check is mandatory and if it caused an
5113 // uncommon trap to be emitted then the allocation can't be
5114 // considered tightly coupled in this context.
5115 alloc = tightly_coupled_allocation(dest, NULL);
5116 }
5117
5118 bool validated = false;
5119
5120 const Type* src_type = _gvn.type(src);
5121 const Type* dest_type = _gvn.type(dest);
5122 const TypeAryPtr* top_src = src_type->isa_aryptr();
5123 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5124
5125 // Do we have the type of src?
5126 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5127 // Do we have the type of dest?
5128 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5129 // Is the type for src from speculation?
5130 bool src_spec = false;
5131 // Is the type for dest from speculation?
5132 bool dest_spec = false;
5133
5134 if ((!has_src || !has_dest) && can_emit_guards) {
5135 // We don't have sufficient type information, let's see if
5136 // speculative types can help. We need to have types for both src
5137 // and dest so that it pays off.
5138
5139 // Do we already have or could we have type information for src
5140 bool could_have_src = has_src;
5141 // Do we already have or could we have type information for dest
5142 bool could_have_dest = has_dest;
5143
5144 ciKlass* src_k = NULL;
5145 if (!has_src) {
5146 src_k = src_type->speculative_type_not_null();
5147 if (src_k != NULL && src_k->is_array_klass()) {
5148 could_have_src = true;
5149 }
5150 }
5151
5152 ciKlass* dest_k = NULL;
5153 if (!has_dest) {
5154 dest_k = dest_type->speculative_type_not_null();
5155 if (dest_k != NULL && dest_k->is_array_klass()) {
5156 could_have_dest = true;
5157 }
5158 }
5159
5160 if (could_have_src && could_have_dest) {
5161 // This is going to pay off so emit the required guards
5162 if (!has_src) {
5163 src = maybe_cast_profiled_obj(src, src_k, true);
5164 src_type = _gvn.type(src);
5165 top_src = src_type->isa_aryptr();
5166 has_src = (top_src != NULL && top_src->klass() != NULL);
5167 src_spec = true;
5168 }
5169 if (!has_dest) {
5170 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5171 dest_type = _gvn.type(dest);
5172 top_dest = dest_type->isa_aryptr();
5173 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5174 dest_spec = true;
5175 }
5176 }
5177 }
5178
5179 if (has_src && has_dest && can_emit_guards) {
5180 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5181 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5182 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5183 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5184
5185 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5186 // If both arrays are object arrays then having the exact types
5187 // for both will remove the need for a subtype check at runtime
5188 // before the call and may make it possible to pick a faster copy
5189 // routine (without a subtype check on every element)
5190 // Do we have the exact type of src?
5191 bool could_have_src = src_spec;
5192 // Do we have the exact type of dest?
5193 bool could_have_dest = dest_spec;
5194 ciKlass* src_k = top_src->klass();
5195 ciKlass* dest_k = top_dest->klass();
5196 if (!src_spec) {
5197 src_k = src_type->speculative_type_not_null();
5198 if (src_k != NULL && src_k->is_array_klass()) {
5199 could_have_src = true;
5200 }
5201 }
5202 if (!dest_spec) {
5203 dest_k = dest_type->speculative_type_not_null();
5204 if (dest_k != NULL && dest_k->is_array_klass()) {
5205 could_have_dest = true;
5206 }
5207 }
5208 if (could_have_src && could_have_dest) {
5209 // If we can have both exact types, emit the missing guards
5210 if (could_have_src && !src_spec) {
5211 src = maybe_cast_profiled_obj(src, src_k, true);
5212 }
5213 if (could_have_dest && !dest_spec) {
5214 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5215 }
5216 }
5217 }
5218 }
5219
5220 ciMethod* trap_method = method();
5221 int trap_bci = bci();
5222 if (saved_jvms != NULL) {
5223 trap_method = alloc->jvms()->method();
5224 trap_bci = alloc->jvms()->bci();
5225 }
5226
5227 bool negative_length_guard_generated = false;
5228
5229 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5230 can_emit_guards &&
5231 !src->is_top() && !dest->is_top()) {
5232 // validate arguments: enables transformation the ArrayCopyNode
5233 validated = true;
5234
5235 RegionNode* slow_region = new RegionNode(1);
5236 record_for_igvn(slow_region);
5237
5238 // (1) src and dest are arrays.
5239 generate_non_array_guard(load_object_klass(src), slow_region);
5240 generate_non_array_guard(load_object_klass(dest), slow_region);
5241
5242 // (2) src and dest arrays must have elements of the same BasicType
5243 // done at macro expansion or at Ideal transformation time
5244
5245 // (4) src_offset must not be negative.
5246 generate_negative_guard(src_offset, slow_region);
5247
5248 // (5) dest_offset must not be negative.
5249 generate_negative_guard(dest_offset, slow_region);
5250
5251 // (7) src_offset + length must not exceed length of src.
5252 generate_limit_guard(src_offset, length,
5253 load_array_length(src),
5254 slow_region);
5255
5256 // (8) dest_offset + length must not exceed length of dest.
5257 generate_limit_guard(dest_offset, length,
5258 load_array_length(dest),
5259 slow_region);
5260
5261 // (6) length must not be negative.
5262 // This is also checked in generate_arraycopy() during macro expansion, but
5263 // we also have to check it here for the case where the ArrayCopyNode will
5264 // be eliminated by Escape Analysis.
5265 if (EliminateAllocations) {
5266 generate_negative_guard(length, slow_region);
5267 negative_length_guard_generated = true;
5268 }
5269
5270 // (9) each element of an oop array must be assignable
5271 Node* src_klass = load_object_klass(src);
5272 Node* dest_klass = load_object_klass(dest);
5273 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5274
5275 if (not_subtype_ctrl != top()) {
5276 PreserveJVMState pjvms(this);
5277 set_control(not_subtype_ctrl);
5278 uncommon_trap(Deoptimization::Reason_intrinsic,
5279 Deoptimization::Action_make_not_entrant);
5280 assert(stopped(), "Should be stopped");
5281 }
5282 {
5283 PreserveJVMState pjvms(this);
5284 set_control(_gvn.transform(slow_region));
5285 uncommon_trap(Deoptimization::Reason_intrinsic,
5286 Deoptimization::Action_make_not_entrant);
5287 assert(stopped(), "Should be stopped");
5288 }
5289
5290 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
5291 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
5292 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
5293 }
5294
5295 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5296
5297 if (stopped()) {
5298 return true;
5299 }
5300
5301 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5302 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5303 // so the compiler has a chance to eliminate them: during macro expansion,
5304 // we have to set their control (CastPP nodes are eliminated).
5305 load_object_klass(src), load_object_klass(dest),
5306 load_array_length(src), load_array_length(dest));
5307
5308 ac->set_arraycopy(validated);
5309
5310 Node* n = _gvn.transform(ac);
5311 if (n == ac) {
5312 ac->connect_outputs(this);
5313 } else {
5314 assert(validated, "shouldn't transform if all arguments not validated");
5315 set_all_memory(n);
5316 }
5317 clear_upper_avx();
5318
5319
5320 return true;
5321 }
5322
5323
5324 // Helper function which determines if an arraycopy immediately follows
5325 // an allocation, with no intervening tests or other escapes for the object.
5326 AllocateArrayNode*
5327 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5328 RegionNode* slow_region) {
5329 if (stopped()) return NULL; // no fast path
5330 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5331
5332 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5333 if (alloc == NULL) return NULL;
5334
5335 Node* rawmem = memory(Compile::AliasIdxRaw);
5336 // Is the allocation's memory state untouched?
5337 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5338 // Bail out if there have been raw-memory effects since the allocation.
5339 // (Example: There might have been a call or safepoint.)
5340 return NULL;
5341 }
5342 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5343 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5344 return NULL;
5345 }
5346
5347 // There must be no unexpected observers of this allocation.
5348 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5349 Node* obs = ptr->fast_out(i);
5350 if (obs != this->map()) {
5351 return NULL;
5352 }
5353 }
5354
5355 // This arraycopy must unconditionally follow the allocation of the ptr.
5356 Node* alloc_ctl = ptr->in(0);
5357 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5358
5359 Node* ctl = control();
5360 while (ctl != alloc_ctl) {
5361 // There may be guards which feed into the slow_region.
5362 // Any other control flow means that we might not get a chance
5363 // to finish initializing the allocated object.
5364 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5365 IfNode* iff = ctl->in(0)->as_If();
5366 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
5367 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5368 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5369 ctl = iff->in(0); // This test feeds the known slow_region.
5370 continue;
5371 }
5372 // One more try: Various low-level checks bottom out in
5373 // uncommon traps. If the debug-info of the trap omits
5374 // any reference to the allocation, as we've already
5375 // observed, then there can be no objection to the trap.
5376 bool found_trap = false;
5377 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5378 Node* obs = not_ctl->fast_out(j);
5379 if (obs->in(0) == not_ctl && obs->is_Call() &&
5380 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5381 found_trap = true; break;
5382 }
5383 }
5384 if (found_trap) {
5385 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5386 continue;
5387 }
5388 }
5389 return NULL;
5390 }
5391
5392 // If we get this far, we have an allocation which immediately
5393 // precedes the arraycopy, and we can take over zeroing the new object.
5394 // The arraycopy will finish the initialization, and provide
5395 // a new control state to which we will anchor the destination pointer.
5396
5397 return alloc;
5398 }
5399
5400 //-------------inline_encodeISOArray-----------------------------------
5401 // encode char[] to byte[] in ISO_8859_1
5402 bool LibraryCallKit::inline_encodeISOArray() {
5403 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5404 // no receiver since it is static method
5405 Node *src = argument(0);
5406 Node *src_offset = argument(1);
5407 Node *dst = argument(2);
5408 Node *dst_offset = argument(3);
5409 Node *length = argument(4);
5410
5411 const Type* src_type = src->Value(&_gvn);
5412 const Type* dst_type = dst->Value(&_gvn);
5413 const TypeAryPtr* top_src = src_type->isa_aryptr();
5414 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5415 if (top_src == NULL || top_src->klass() == NULL ||
5416 top_dest == NULL || top_dest->klass() == NULL) {
5417 // failed array check
5418 return false;
5419 }
5420
5421 // Figure out the size and type of the elements we will be copying.
5422 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5423 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5424 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5425 return false;
5426 }
5427
5428 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5429 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5430 // 'src_start' points to src array + scaled offset
5431 // 'dst_start' points to dst array + scaled offset
5432
5433 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5434 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5435 enc = _gvn.transform(enc);
5436 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5437 set_memory(res_mem, mtype);
5438 set_result(enc);
5439 clear_upper_avx();
5440
5441 return true;
5442 }
5443
5444 //-------------inline_multiplyToLen-----------------------------------
5445 bool LibraryCallKit::inline_multiplyToLen() {
5446 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5447
5448 address stubAddr = StubRoutines::multiplyToLen();
5449 if (stubAddr == NULL) {
5450 return false; // Intrinsic's stub is not implemented on this platform
5451 }
5452 const char* stubName = "multiplyToLen";
5453
5454 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5455
5456 // no receiver because it is a static method
5457 Node* x = argument(0);
5458 Node* xlen = argument(1);
5459 Node* y = argument(2);
5460 Node* ylen = argument(3);
5461 Node* z = argument(4);
5462
5463 const Type* x_type = x->Value(&_gvn);
5464 const Type* y_type = y->Value(&_gvn);
5465 const TypeAryPtr* top_x = x_type->isa_aryptr();
5466 const TypeAryPtr* top_y = y_type->isa_aryptr();
5467 if (top_x == NULL || top_x->klass() == NULL ||
5468 top_y == NULL || top_y->klass() == NULL) {
5469 // failed array check
5470 return false;
5471 }
5472
5473 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5474 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5475 if (x_elem != T_INT || y_elem != T_INT) {
5476 return false;
5477 }
5478
5479 // Set the original stack and the reexecute bit for the interpreter to reexecute
5480 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5481 // on the return from z array allocation in runtime.
5482 { PreserveReexecuteState preexecs(this);
5483 jvms()->set_should_reexecute(true);
5484
5485 Node* x_start = array_element_address(x, intcon(0), x_elem);
5486 Node* y_start = array_element_address(y, intcon(0), y_elem);
5487 // 'x_start' points to x array + scaled xlen
5488 // 'y_start' points to y array + scaled ylen
5489
5490 // Allocate the result array
5491 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5492 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5493 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5494
5495 IdealKit ideal(this);
5496
5497 #define __ ideal.
5498 Node* one = __ ConI(1);
5499 Node* zero = __ ConI(0);
5500 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5501 __ set(need_alloc, zero);
5502 __ set(z_alloc, z);
5503 __ if_then(z, BoolTest::eq, null()); {
5504 __ increment (need_alloc, one);
5505 } __ else_(); {
5506 // Update graphKit memory and control from IdealKit.
5507 sync_kit(ideal);
5508 Node* zlen_arg = load_array_length(z);
5509 // Update IdealKit memory and control from graphKit.
5510 __ sync_kit(this);
5511 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5512 __ increment (need_alloc, one);
5513 } __ end_if();
5514 } __ end_if();
5515
5516 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5517 // Update graphKit memory and control from IdealKit.
5518 sync_kit(ideal);
5519 Node * narr = new_array(klass_node, zlen, 1);
5520 // Update IdealKit memory and control from graphKit.
5521 __ sync_kit(this);
5522 __ set(z_alloc, narr);
5523 } __ end_if();
5524
5525 sync_kit(ideal);
5526 z = __ value(z_alloc);
5527 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5528 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5529 // Final sync IdealKit and GraphKit.
5530 final_sync(ideal);
5531 #undef __
5532
5533 Node* z_start = array_element_address(z, intcon(0), T_INT);
5534
5535 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5536 OptoRuntime::multiplyToLen_Type(),
5537 stubAddr, stubName, TypePtr::BOTTOM,
5538 x_start, xlen, y_start, ylen, z_start, zlen);
5539 } // original reexecute is set back here
5540
5541 C->set_has_split_ifs(true); // Has chance for split-if optimization
5542 set_result(z);
5543 return true;
5544 }
5545
5546 //-------------inline_squareToLen------------------------------------
5547 bool LibraryCallKit::inline_squareToLen() {
5548 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5549
5550 address stubAddr = StubRoutines::squareToLen();
5551 if (stubAddr == NULL) {
5552 return false; // Intrinsic's stub is not implemented on this platform
5553 }
5554 const char* stubName = "squareToLen";
5555
5556 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5557
5558 Node* x = argument(0);
5559 Node* len = argument(1);
5560 Node* z = argument(2);
5561 Node* zlen = argument(3);
5562
5563 const Type* x_type = x->Value(&_gvn);
5564 const Type* z_type = z->Value(&_gvn);
5565 const TypeAryPtr* top_x = x_type->isa_aryptr();
5566 const TypeAryPtr* top_z = z_type->isa_aryptr();
5567 if (top_x == NULL || top_x->klass() == NULL ||
5568 top_z == NULL || top_z->klass() == NULL) {
5569 // failed array check
5570 return false;
5571 }
5572
5573 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5574 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5575 if (x_elem != T_INT || z_elem != T_INT) {
5576 return false;
5577 }
5578
5579
5580 Node* x_start = array_element_address(x, intcon(0), x_elem);
5581 Node* z_start = array_element_address(z, intcon(0), z_elem);
5582
5583 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5584 OptoRuntime::squareToLen_Type(),
5585 stubAddr, stubName, TypePtr::BOTTOM,
5586 x_start, len, z_start, zlen);
5587
5588 set_result(z);
5589 return true;
5590 }
5591
5592 //-------------inline_mulAdd------------------------------------------
5593 bool LibraryCallKit::inline_mulAdd() {
5594 assert(UseMulAddIntrinsic, "not implemented on this platform");
5595
5596 address stubAddr = StubRoutines::mulAdd();
5597 if (stubAddr == NULL) {
5598 return false; // Intrinsic's stub is not implemented on this platform
5599 }
5600 const char* stubName = "mulAdd";
5601
5602 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5603
5604 Node* out = argument(0);
5605 Node* in = argument(1);
5606 Node* offset = argument(2);
5607 Node* len = argument(3);
5608 Node* k = argument(4);
5609
5610 const Type* out_type = out->Value(&_gvn);
5611 const Type* in_type = in->Value(&_gvn);
5612 const TypeAryPtr* top_out = out_type->isa_aryptr();
5613 const TypeAryPtr* top_in = in_type->isa_aryptr();
5614 if (top_out == NULL || top_out->klass() == NULL ||
5615 top_in == NULL || top_in->klass() == NULL) {
5616 // failed array check
5617 return false;
5618 }
5619
5620 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5621 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5622 if (out_elem != T_INT || in_elem != T_INT) {
5623 return false;
5624 }
5625
5626 Node* outlen = load_array_length(out);
5627 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5628 Node* out_start = array_element_address(out, intcon(0), out_elem);
5629 Node* in_start = array_element_address(in, intcon(0), in_elem);
5630
5631 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5632 OptoRuntime::mulAdd_Type(),
5633 stubAddr, stubName, TypePtr::BOTTOM,
5634 out_start,in_start, new_offset, len, k);
5635 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5636 set_result(result);
5637 return true;
5638 }
5639
5640 //-------------inline_montgomeryMultiply-----------------------------------
5641 bool LibraryCallKit::inline_montgomeryMultiply() {
5642 address stubAddr = StubRoutines::montgomeryMultiply();
5643 if (stubAddr == NULL) {
5644 return false; // Intrinsic's stub is not implemented on this platform
5645 }
5646
5647 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5648 const char* stubName = "montgomery_multiply";
5649
5650 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5651
5652 Node* a = argument(0);
5653 Node* b = argument(1);
5654 Node* n = argument(2);
5655 Node* len = argument(3);
5656 Node* inv = argument(4);
5657 Node* m = argument(6);
5658
5659 const Type* a_type = a->Value(&_gvn);
5660 const TypeAryPtr* top_a = a_type->isa_aryptr();
5661 const Type* b_type = b->Value(&_gvn);
5662 const TypeAryPtr* top_b = b_type->isa_aryptr();
5663 const Type* n_type = a->Value(&_gvn);
5664 const TypeAryPtr* top_n = n_type->isa_aryptr();
5665 const Type* m_type = a->Value(&_gvn);
5666 const TypeAryPtr* top_m = m_type->isa_aryptr();
5667 if (top_a == NULL || top_a->klass() == NULL ||
5668 top_b == NULL || top_b->klass() == NULL ||
5669 top_n == NULL || top_n->klass() == NULL ||
5670 top_m == NULL || top_m->klass() == NULL) {
5671 // failed array check
5672 return false;
5673 }
5674
5675 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5676 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5677 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5678 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5679 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5680 return false;
5681 }
5682
5683 // Make the call
5684 {
5685 Node* a_start = array_element_address(a, intcon(0), a_elem);
5686 Node* b_start = array_element_address(b, intcon(0), b_elem);
5687 Node* n_start = array_element_address(n, intcon(0), n_elem);
5688 Node* m_start = array_element_address(m, intcon(0), m_elem);
5689
5690 Node* call = make_runtime_call(RC_LEAF,
5691 OptoRuntime::montgomeryMultiply_Type(),
5692 stubAddr, stubName, TypePtr::BOTTOM,
5693 a_start, b_start, n_start, len, inv, top(),
5694 m_start);
5695 set_result(m);
5696 }
5697
5698 return true;
5699 }
5700
5701 bool LibraryCallKit::inline_montgomerySquare() {
5702 address stubAddr = StubRoutines::montgomerySquare();
5703 if (stubAddr == NULL) {
5704 return false; // Intrinsic's stub is not implemented on this platform
5705 }
5706
5707 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5708 const char* stubName = "montgomery_square";
5709
5710 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5711
5712 Node* a = argument(0);
5713 Node* n = argument(1);
5714 Node* len = argument(2);
5715 Node* inv = argument(3);
5716 Node* m = argument(5);
5717
5718 const Type* a_type = a->Value(&_gvn);
5719 const TypeAryPtr* top_a = a_type->isa_aryptr();
5720 const Type* n_type = a->Value(&_gvn);
5721 const TypeAryPtr* top_n = n_type->isa_aryptr();
5722 const Type* m_type = a->Value(&_gvn);
5723 const TypeAryPtr* top_m = m_type->isa_aryptr();
5724 if (top_a == NULL || top_a->klass() == NULL ||
5725 top_n == NULL || top_n->klass() == NULL ||
5726 top_m == NULL || top_m->klass() == NULL) {
5727 // failed array check
5728 return false;
5729 }
5730
5731 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5732 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5733 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5734 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5735 return false;
5736 }
5737
5738 // Make the call
5739 {
5740 Node* a_start = array_element_address(a, intcon(0), a_elem);
5741 Node* n_start = array_element_address(n, intcon(0), n_elem);
5742 Node* m_start = array_element_address(m, intcon(0), m_elem);
5743
5744 Node* call = make_runtime_call(RC_LEAF,
5745 OptoRuntime::montgomerySquare_Type(),
5746 stubAddr, stubName, TypePtr::BOTTOM,
5747 a_start, n_start, len, inv, top(),
5748 m_start);
5749 set_result(m);
5750 }
5751
5752 return true;
5753 }
5754
5755 //-------------inline_vectorizedMismatch------------------------------
5756 bool LibraryCallKit::inline_vectorizedMismatch() {
5757 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5758
5759 address stubAddr = StubRoutines::vectorizedMismatch();
5760 if (stubAddr == NULL) {
5761 return false; // Intrinsic's stub is not implemented on this platform
5762 }
5763 const char* stubName = "vectorizedMismatch";
5764 int size_l = callee()->signature()->size();
5765 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5766
5767 Node* obja = argument(0);
5768 Node* aoffset = argument(1);
5769 Node* objb = argument(3);
5770 Node* boffset = argument(4);
5771 Node* length = argument(6);
5772 Node* scale = argument(7);
5773
5774 const Type* a_type = obja->Value(&_gvn);
5775 const Type* b_type = objb->Value(&_gvn);
5776 const TypeAryPtr* top_a = a_type->isa_aryptr();
5777 const TypeAryPtr* top_b = b_type->isa_aryptr();
5778 if (top_a == NULL || top_a->klass() == NULL ||
5779 top_b == NULL || top_b->klass() == NULL) {
5780 // failed array check
5781 return false;
5782 }
5783
5784 Node* call;
5785 jvms()->set_should_reexecute(true);
5786
5787 Node* obja_adr = make_unsafe_address(obja, aoffset);
5788 Node* objb_adr = make_unsafe_address(objb, boffset);
5789
5790 call = make_runtime_call(RC_LEAF,
5791 OptoRuntime::vectorizedMismatch_Type(),
5792 stubAddr, stubName, TypePtr::BOTTOM,
5793 obja_adr, objb_adr, length, scale);
5794
5795 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5796 set_result(result);
5797 return true;
5798 }
5799
5800 /**
5801 * Calculate CRC32 for byte.
5802 * int java.util.zip.CRC32.update(int crc, int b)
5803 */
5804 bool LibraryCallKit::inline_updateCRC32() {
5805 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5806 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5807 // no receiver since it is static method
5808 Node* crc = argument(0); // type: int
5809 Node* b = argument(1); // type: int
5810
5811 /*
5812 * int c = ~ crc;
5813 * b = timesXtoThe32[(b ^ c) & 0xFF];
5814 * b = b ^ (c >>> 8);
5815 * crc = ~b;
5816 */
5817
5818 Node* M1 = intcon(-1);
5819 crc = _gvn.transform(new XorINode(crc, M1));
5820 Node* result = _gvn.transform(new XorINode(crc, b));
5821 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5822
5823 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5824 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5825 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5826 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5827
5828 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5829 result = _gvn.transform(new XorINode(crc, result));
5830 result = _gvn.transform(new XorINode(result, M1));
5831 set_result(result);
5832 return true;
5833 }
5834
5835 /**
5836 * Calculate CRC32 for byte[] array.
5837 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5838 */
5839 bool LibraryCallKit::inline_updateBytesCRC32() {
5840 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5841 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5842 // no receiver since it is static method
5843 Node* crc = argument(0); // type: int
5844 Node* src = argument(1); // type: oop
5845 Node* offset = argument(2); // type: int
5846 Node* length = argument(3); // type: int
5847
5848 const Type* src_type = src->Value(&_gvn);
5849 const TypeAryPtr* top_src = src_type->isa_aryptr();
5850 if (top_src == NULL || top_src->klass() == NULL) {
5851 // failed array check
5852 return false;
5853 }
5854
5855 // Figure out the size and type of the elements we will be copying.
5856 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5857 if (src_elem != T_BYTE) {
5858 return false;
5859 }
5860
5861 // 'src_start' points to src array + scaled offset
5862 Node* src_start = array_element_address(src, offset, src_elem);
5863
5864 // We assume that range check is done by caller.
5865 // TODO: generate range check (offset+length < src.length) in debug VM.
5866
5867 // Call the stub.
5868 address stubAddr = StubRoutines::updateBytesCRC32();
5869 const char *stubName = "updateBytesCRC32";
5870
5871 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5872 stubAddr, stubName, TypePtr::BOTTOM,
5873 crc, src_start, length);
5874 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5875 set_result(result);
5876 return true;
5877 }
5878
5879 /**
5880 * Calculate CRC32 for ByteBuffer.
5881 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5882 */
5883 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5884 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5885 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5886 // no receiver since it is static method
5887 Node* crc = argument(0); // type: int
5888 Node* src = argument(1); // type: long
5889 Node* offset = argument(3); // type: int
5890 Node* length = argument(4); // type: int
5891
5892 src = ConvL2X(src); // adjust Java long to machine word
5893 Node* base = _gvn.transform(new CastX2PNode(src));
5894 offset = ConvI2X(offset);
5895
5896 // 'src_start' points to src array + scaled offset
5897 Node* src_start = basic_plus_adr(top(), base, offset);
5898
5899 // Call the stub.
5900 address stubAddr = StubRoutines::updateBytesCRC32();
5901 const char *stubName = "updateBytesCRC32";
5902
5903 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5904 stubAddr, stubName, TypePtr::BOTTOM,
5905 crc, src_start, length);
5906 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5907 set_result(result);
5908 return true;
5909 }
5910
5911 //------------------------------get_table_from_crc32c_class-----------------------
5912 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5913 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5914 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5915
5916 return table;
5917 }
5918
5919 //------------------------------inline_updateBytesCRC32C-----------------------
5920 //
5921 // Calculate CRC32C for byte[] array.
5922 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5923 //
5924 bool LibraryCallKit::inline_updateBytesCRC32C() {
5925 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5926 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5927 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5928 // no receiver since it is a static method
5929 Node* crc = argument(0); // type: int
5930 Node* src = argument(1); // type: oop
5931 Node* offset = argument(2); // type: int
5932 Node* end = argument(3); // type: int
5933
5934 Node* length = _gvn.transform(new SubINode(end, offset));
5935
5936 const Type* src_type = src->Value(&_gvn);
5937 const TypeAryPtr* top_src = src_type->isa_aryptr();
5938 if (top_src == NULL || top_src->klass() == NULL) {
5939 // failed array check
5940 return false;
5941 }
5942
5943 // Figure out the size and type of the elements we will be copying.
5944 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5945 if (src_elem != T_BYTE) {
5946 return false;
5947 }
5948
5949 // 'src_start' points to src array + scaled offset
5950 Node* src_start = array_element_address(src, offset, src_elem);
5951
5952 // static final int[] byteTable in class CRC32C
5953 Node* table = get_table_from_crc32c_class(callee()->holder());
5954 Node* table_start = array_element_address(table, intcon(0), T_INT);
5955
5956 // We assume that range check is done by caller.
5957 // TODO: generate range check (offset+length < src.length) in debug VM.
5958
5959 // Call the stub.
5960 address stubAddr = StubRoutines::updateBytesCRC32C();
5961 const char *stubName = "updateBytesCRC32C";
5962
5963 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5964 stubAddr, stubName, TypePtr::BOTTOM,
5965 crc, src_start, length, table_start);
5966 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5967 set_result(result);
5968 return true;
5969 }
5970
5971 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5972 //
5973 // Calculate CRC32C for DirectByteBuffer.
5974 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5975 //
5976 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5977 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5978 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5979 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5980 // no receiver since it is a static method
5981 Node* crc = argument(0); // type: int
5982 Node* src = argument(1); // type: long
5983 Node* offset = argument(3); // type: int
5984 Node* end = argument(4); // type: int
5985
5986 Node* length = _gvn.transform(new SubINode(end, offset));
5987
5988 src = ConvL2X(src); // adjust Java long to machine word
5989 Node* base = _gvn.transform(new CastX2PNode(src));
5990 offset = ConvI2X(offset);
5991
5992 // 'src_start' points to src array + scaled offset
5993 Node* src_start = basic_plus_adr(top(), base, offset);
5994
5995 // static final int[] byteTable in class CRC32C
5996 Node* table = get_table_from_crc32c_class(callee()->holder());
5997 Node* table_start = array_element_address(table, intcon(0), T_INT);
5998
5999 // Call the stub.
6000 address stubAddr = StubRoutines::updateBytesCRC32C();
6001 const char *stubName = "updateBytesCRC32C";
6002
6003 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6004 stubAddr, stubName, TypePtr::BOTTOM,
6005 crc, src_start, length, table_start);
6006 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6007 set_result(result);
6008 return true;
6009 }
6010
6011 //------------------------------inline_updateBytesAdler32----------------------
6012 //
6013 // Calculate Adler32 checksum for byte[] array.
6014 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6015 //
6016 bool LibraryCallKit::inline_updateBytesAdler32() {
6017 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6018 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6019 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6020 // no receiver since it is static method
6021 Node* crc = argument(0); // type: int
6022 Node* src = argument(1); // type: oop
6023 Node* offset = argument(2); // type: int
6024 Node* length = argument(3); // type: int
6025
6026 const Type* src_type = src->Value(&_gvn);
6027 const TypeAryPtr* top_src = src_type->isa_aryptr();
6028 if (top_src == NULL || top_src->klass() == NULL) {
6029 // failed array check
6030 return false;
6031 }
6032
6033 // Figure out the size and type of the elements we will be copying.
6034 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6035 if (src_elem != T_BYTE) {
6036 return false;
6037 }
6038
6039 // 'src_start' points to src array + scaled offset
6040 Node* src_start = array_element_address(src, offset, src_elem);
6041
6042 // We assume that range check is done by caller.
6043 // TODO: generate range check (offset+length < src.length) in debug VM.
6044
6045 // Call the stub.
6046 address stubAddr = StubRoutines::updateBytesAdler32();
6047 const char *stubName = "updateBytesAdler32";
6048
6049 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6050 stubAddr, stubName, TypePtr::BOTTOM,
6051 crc, src_start, length);
6052 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6053 set_result(result);
6054 return true;
6055 }
6056
6057 //------------------------------inline_updateByteBufferAdler32---------------
6058 //
6059 // Calculate Adler32 checksum for DirectByteBuffer.
6060 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6061 //
6062 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6063 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6064 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6065 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6066 // no receiver since it is static method
6067 Node* crc = argument(0); // type: int
6068 Node* src = argument(1); // type: long
6069 Node* offset = argument(3); // type: int
6070 Node* length = argument(4); // type: int
6071
6072 src = ConvL2X(src); // adjust Java long to machine word
6073 Node* base = _gvn.transform(new CastX2PNode(src));
6074 offset = ConvI2X(offset);
6075
6076 // 'src_start' points to src array + scaled offset
6077 Node* src_start = basic_plus_adr(top(), base, offset);
6078
6079 // Call the stub.
6080 address stubAddr = StubRoutines::updateBytesAdler32();
6081 const char *stubName = "updateBytesAdler32";
6082
6083 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6084 stubAddr, stubName, TypePtr::BOTTOM,
6085 crc, src_start, length);
6086
6087 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6088 set_result(result);
6089 return true;
6090 }
6091
6092 //----------------------------inline_reference_get----------------------------
6093 // public T java.lang.ref.Reference.get();
6094 bool LibraryCallKit::inline_reference_get() {
6095 const int referent_offset = java_lang_ref_Reference::referent_offset;
6096 guarantee(referent_offset > 0, "should have already been set");
6097
6098 // Get the argument:
6099 Node* reference_obj = null_check_receiver();
6100 if (stopped()) return true;
6101
6102 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6103
6104 ciInstanceKlass* klass = env()->Object_klass();
6105 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6106
6107 Node* no_ctrl = NULL;
6108 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6109
6110 // Use the pre-barrier to record the value in the referent field
6111 pre_barrier(false /* do_load */,
6112 control(),
6113 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6114 result /* pre_val */,
6115 T_OBJECT);
6116
6117 // Add memory barrier to prevent commoning reads from this field
6118 // across safepoint since GC can change its value.
6119 insert_mem_bar(Op_MemBarCPUOrder);
6120
6121 set_result(result);
6122 return true;
6123 }
6124
6125
6126 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6127 bool is_exact=true, bool is_static=false,
6128 ciInstanceKlass * fromKls=NULL) {
6129 if (fromKls == NULL) {
6130 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6131 assert(tinst != NULL, "obj is null");
6132 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6133 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6134 fromKls = tinst->klass()->as_instance_klass();
6135 } else {
6136 assert(is_static, "only for static field access");
6137 }
6138 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6139 ciSymbol::make(fieldTypeString),
6140 is_static);
6141
6142 assert (field != NULL, "undefined field");
6143 if (field == NULL) return (Node *) NULL;
6144
6145 if (is_static) {
6146 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6147 fromObj = makecon(tip);
6148 }
6149
6150 // Next code copied from Parse::do_get_xxx():
6151
6152 // Compute address and memory type.
6153 int offset = field->offset_in_bytes();
6154 bool is_vol = field->is_volatile();
6155 ciType* field_klass = field->type();
6156 assert(field_klass->is_loaded(), "should be loaded");
6157 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6158 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6159 BasicType bt = field->layout_type();
6160
6161 // Build the resultant type of the load
6162 const Type *type;
6163 if (bt == T_OBJECT) {
6164 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6165 } else {
6166 type = Type::get_const_basic_type(bt);
6167 }
6168
6169 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6170 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6171 }
6172 // Build the load.
6173 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6174 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6175 // If reference is volatile, prevent following memory ops from
6176 // floating up past the volatile read. Also prevents commoning
6177 // another volatile read.
6178 if (is_vol) {
6179 // Memory barrier includes bogus read of value to force load BEFORE membar
6180 insert_mem_bar(Op_MemBarAcquire, loadedField);
6181 }
6182 return loadedField;
6183 }
6184
6185 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6186 bool is_exact = true, bool is_static = false,
6187 ciInstanceKlass * fromKls = NULL) {
6188 if (fromKls == NULL) {
6189 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6190 assert(tinst != NULL, "obj is null");
6191 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6192 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6193 fromKls = tinst->klass()->as_instance_klass();
6194 }
6195 else {
6196 assert(is_static, "only for static field access");
6197 }
6198 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6199 ciSymbol::make(fieldTypeString),
6200 is_static);
6201
6202 assert(field != NULL, "undefined field");
6203 assert(!field->is_volatile(), "not defined for volatile fields");
6204
6205 if (is_static) {
6206 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6207 fromObj = makecon(tip);
6208 }
6209
6210 // Next code copied from Parse::do_get_xxx():
6211
6212 // Compute address and memory type.
6213 int offset = field->offset_in_bytes();
6214 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6215
6216 return adr;
6217 }
6218
6219 //------------------------------inline_aescrypt_Block-----------------------
6220 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6221 address stubAddr = NULL;
6222 const char *stubName;
6223 assert(UseAES, "need AES instruction support");
6224
6225 switch(id) {
6226 case vmIntrinsics::_aescrypt_encryptBlock:
6227 stubAddr = StubRoutines::aescrypt_encryptBlock();
6228 stubName = "aescrypt_encryptBlock";
6229 break;
6230 case vmIntrinsics::_aescrypt_decryptBlock:
6231 stubAddr = StubRoutines::aescrypt_decryptBlock();
6232 stubName = "aescrypt_decryptBlock";
6233 break;
6234 default:
6235 break;
6236 }
6237 if (stubAddr == NULL) return false;
6238
6239 Node* aescrypt_object = argument(0);
6240 Node* src = argument(1);
6241 Node* src_offset = argument(2);
6242 Node* dest = argument(3);
6243 Node* dest_offset = argument(4);
6244
6245 // (1) src and dest are arrays.
6246 const Type* src_type = src->Value(&_gvn);
6247 const Type* dest_type = dest->Value(&_gvn);
6248 const TypeAryPtr* top_src = src_type->isa_aryptr();
6249 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6250 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6251
6252 // for the quick and dirty code we will skip all the checks.
6253 // we are just trying to get the call to be generated.
6254 Node* src_start = src;
6255 Node* dest_start = dest;
6256 if (src_offset != NULL || dest_offset != NULL) {
6257 assert(src_offset != NULL && dest_offset != NULL, "");
6258 src_start = array_element_address(src, src_offset, T_BYTE);
6259 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6260 }
6261
6262 // now need to get the start of its expanded key array
6263 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6264 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6265 if (k_start == NULL) return false;
6266
6267 if (Matcher::pass_original_key_for_aes()) {
6268 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6269 // compatibility issues between Java key expansion and SPARC crypto instructions
6270 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6271 if (original_k_start == NULL) return false;
6272
6273 // Call the stub.
6274 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6275 stubAddr, stubName, TypePtr::BOTTOM,
6276 src_start, dest_start, k_start, original_k_start);
6277 } else {
6278 // Call the stub.
6279 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6280 stubAddr, stubName, TypePtr::BOTTOM,
6281 src_start, dest_start, k_start);
6282 }
6283
6284 return true;
6285 }
6286
6287 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6288 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6289 address stubAddr = NULL;
6290 const char *stubName = NULL;
6291
6292 assert(UseAES, "need AES instruction support");
6293
6294 switch(id) {
6295 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6296 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6297 stubName = "cipherBlockChaining_encryptAESCrypt";
6298 break;
6299 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6300 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6301 stubName = "cipherBlockChaining_decryptAESCrypt";
6302 break;
6303 default:
6304 break;
6305 }
6306 if (stubAddr == NULL) return false;
6307
6308 Node* cipherBlockChaining_object = argument(0);
6309 Node* src = argument(1);
6310 Node* src_offset = argument(2);
6311 Node* len = argument(3);
6312 Node* dest = argument(4);
6313 Node* dest_offset = argument(5);
6314
6315 // (1) src and dest are arrays.
6316 const Type* src_type = src->Value(&_gvn);
6317 const Type* dest_type = dest->Value(&_gvn);
6318 const TypeAryPtr* top_src = src_type->isa_aryptr();
6319 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6320 assert (top_src != NULL && top_src->klass() != NULL
6321 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6322
6323 // checks are the responsibility of the caller
6324 Node* src_start = src;
6325 Node* dest_start = dest;
6326 if (src_offset != NULL || dest_offset != NULL) {
6327 assert(src_offset != NULL && dest_offset != NULL, "");
6328 src_start = array_element_address(src, src_offset, T_BYTE);
6329 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6330 }
6331
6332 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6333 // (because of the predicated logic executed earlier).
6334 // so we cast it here safely.
6335 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6336
6337 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6338 if (embeddedCipherObj == NULL) return false;
6339
6340 // cast it to what we know it will be at runtime
6341 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6342 assert(tinst != NULL, "CBC obj is null");
6343 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6344 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6345 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6346
6347 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6348 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6349 const TypeOopPtr* xtype = aklass->as_instance_type();
6350 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6351 aescrypt_object = _gvn.transform(aescrypt_object);
6352
6353 // we need to get the start of the aescrypt_object's expanded key array
6354 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6355 if (k_start == NULL) return false;
6356
6357 // similarly, get the start address of the r vector
6358 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6359 if (objRvec == NULL) return false;
6360 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6361
6362 Node* cbcCrypt;
6363 if (Matcher::pass_original_key_for_aes()) {
6364 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6365 // compatibility issues between Java key expansion and SPARC crypto instructions
6366 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6367 if (original_k_start == NULL) return false;
6368
6369 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6370 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6371 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6372 stubAddr, stubName, TypePtr::BOTTOM,
6373 src_start, dest_start, k_start, r_start, len, original_k_start);
6374 } else {
6375 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6376 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6377 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6378 stubAddr, stubName, TypePtr::BOTTOM,
6379 src_start, dest_start, k_start, r_start, len);
6380 }
6381
6382 // return cipher length (int)
6383 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6384 set_result(retvalue);
6385 return true;
6386 }
6387
6388 //------------------------------inline_counterMode_AESCrypt-----------------------
6389 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6390 assert(UseAES, "need AES instruction support");
6391 if (!UseAESCTRIntrinsics) return false;
6392
6393 address stubAddr = NULL;
6394 const char *stubName = NULL;
6395 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6396 stubAddr = StubRoutines::counterMode_AESCrypt();
6397 stubName = "counterMode_AESCrypt";
6398 }
6399 if (stubAddr == NULL) return false;
6400
6401 Node* counterMode_object = argument(0);
6402 Node* src = argument(1);
6403 Node* src_offset = argument(2);
6404 Node* len = argument(3);
6405 Node* dest = argument(4);
6406 Node* dest_offset = argument(5);
6407
6408 // (1) src and dest are arrays.
6409 const Type* src_type = src->Value(&_gvn);
6410 const Type* dest_type = dest->Value(&_gvn);
6411 const TypeAryPtr* top_src = src_type->isa_aryptr();
6412 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6413 assert(top_src != NULL && top_src->klass() != NULL &&
6414 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6415
6416 // checks are the responsibility of the caller
6417 Node* src_start = src;
6418 Node* dest_start = dest;
6419 if (src_offset != NULL || dest_offset != NULL) {
6420 assert(src_offset != NULL && dest_offset != NULL, "");
6421 src_start = array_element_address(src, src_offset, T_BYTE);
6422 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6423 }
6424
6425 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6426 // (because of the predicated logic executed earlier).
6427 // so we cast it here safely.
6428 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6429 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6430 if (embeddedCipherObj == NULL) return false;
6431 // cast it to what we know it will be at runtime
6432 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6433 assert(tinst != NULL, "CTR obj is null");
6434 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6435 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6436 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6437 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6438 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6439 const TypeOopPtr* xtype = aklass->as_instance_type();
6440 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6441 aescrypt_object = _gvn.transform(aescrypt_object);
6442 // we need to get the start of the aescrypt_object's expanded key array
6443 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6444 if (k_start == NULL) return false;
6445 // similarly, get the start address of the r vector
6446 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6447 if (obj_counter == NULL) return false;
6448 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6449
6450 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6451 if (saved_encCounter == NULL) return false;
6452 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6453 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6454
6455 Node* ctrCrypt;
6456 if (Matcher::pass_original_key_for_aes()) {
6457 // no SPARC version for AES/CTR intrinsics now.
6458 return false;
6459 }
6460 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6461 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6462 OptoRuntime::counterMode_aescrypt_Type(),
6463 stubAddr, stubName, TypePtr::BOTTOM,
6464 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6465
6466 // return cipher length (int)
6467 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6468 set_result(retvalue);
6469 return true;
6470 }
6471
6472 //------------------------------get_key_start_from_aescrypt_object-----------------------
6473 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6474 #if defined(PPC64) || defined(S390)
6475 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6476 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6477 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6478 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6479 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6480 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6481 if (objSessionK == NULL) {
6482 return (Node *) NULL;
6483 }
6484 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6485 #else
6486 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6487 #endif // PPC64
6488 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6489 if (objAESCryptKey == NULL) return (Node *) NULL;
6490
6491 // now have the array, need to get the start address of the K array
6492 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6493 return k_start;
6494 }
6495
6496 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6497 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6498 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6499 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6500 if (objAESCryptKey == NULL) return (Node *) NULL;
6501
6502 // now have the array, need to get the start address of the lastKey array
6503 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6504 return original_k_start;
6505 }
6506
6507 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6508 // Return node representing slow path of predicate check.
6509 // the pseudo code we want to emulate with this predicate is:
6510 // for encryption:
6511 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6512 // for decryption:
6513 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6514 // note cipher==plain is more conservative than the original java code but that's OK
6515 //
6516 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6517 // The receiver was checked for NULL already.
6518 Node* objCBC = argument(0);
6519
6520 // Load embeddedCipher field of CipherBlockChaining object.
6521 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6522
6523 // get AESCrypt klass for instanceOf check
6524 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6525 // will have same classloader as CipherBlockChaining object
6526 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6527 assert(tinst != NULL, "CBCobj is null");
6528 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6529
6530 // we want to do an instanceof comparison against the AESCrypt class
6531 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6532 if (!klass_AESCrypt->is_loaded()) {
6533 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6534 Node* ctrl = control();
6535 set_control(top()); // no regular fast path
6536 return ctrl;
6537 }
6538 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6539
6540 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6541 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6542 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6543
6544 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6545
6546 // for encryption, we are done
6547 if (!decrypting)
6548 return instof_false; // even if it is NULL
6549
6550 // for decryption, we need to add a further check to avoid
6551 // taking the intrinsic path when cipher and plain are the same
6552 // see the original java code for why.
6553 RegionNode* region = new RegionNode(3);
6554 region->init_req(1, instof_false);
6555 Node* src = argument(1);
6556 Node* dest = argument(4);
6557 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6558 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6559 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6560 region->init_req(2, src_dest_conjoint);
6561
6562 record_for_igvn(region);
6563 return _gvn.transform(region);
6564 }
6565
6566 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6567 // Return node representing slow path of predicate check.
6568 // the pseudo code we want to emulate with this predicate is:
6569 // for encryption:
6570 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6571 // for decryption:
6572 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6573 // note cipher==plain is more conservative than the original java code but that's OK
6574 //
6575
6576 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6577 // The receiver was checked for NULL already.
6578 Node* objCTR = argument(0);
6579
6580 // Load embeddedCipher field of CipherBlockChaining object.
6581 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6582
6583 // get AESCrypt klass for instanceOf check
6584 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6585 // will have same classloader as CipherBlockChaining object
6586 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6587 assert(tinst != NULL, "CTRobj is null");
6588 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6589
6590 // we want to do an instanceof comparison against the AESCrypt class
6591 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6592 if (!klass_AESCrypt->is_loaded()) {
6593 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6594 Node* ctrl = control();
6595 set_control(top()); // no regular fast path
6596 return ctrl;
6597 }
6598
6599 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6600 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6601 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6602 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6603 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6604
6605 return instof_false; // even if it is NULL
6606 }
6607
6608 //------------------------------inline_ghash_processBlocks
6609 bool LibraryCallKit::inline_ghash_processBlocks() {
6610 address stubAddr;
6611 const char *stubName;
6612 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6613
6614 stubAddr = StubRoutines::ghash_processBlocks();
6615 stubName = "ghash_processBlocks";
6616
6617 Node* data = argument(0);
6618 Node* offset = argument(1);
6619 Node* len = argument(2);
6620 Node* state = argument(3);
6621 Node* subkeyH = argument(4);
6622
6623 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6624 assert(state_start, "state is NULL");
6625 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6626 assert(subkeyH_start, "subkeyH is NULL");
6627 Node* data_start = array_element_address(data, offset, T_BYTE);
6628 assert(data_start, "data is NULL");
6629
6630 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6631 OptoRuntime::ghash_processBlocks_Type(),
6632 stubAddr, stubName, TypePtr::BOTTOM,
6633 state_start, subkeyH_start, data_start, len);
6634 return true;
6635 }
6636
6637 //------------------------------inline_sha_implCompress-----------------------
6638 //
6639 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6640 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6641 //
6642 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6643 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6644 //
6645 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6646 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6647 //
6648 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6649 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6650
6651 Node* sha_obj = argument(0);
6652 Node* src = argument(1); // type oop
6653 Node* ofs = argument(2); // type int
6654
6655 const Type* src_type = src->Value(&_gvn);
6656 const TypeAryPtr* top_src = src_type->isa_aryptr();
6657 if (top_src == NULL || top_src->klass() == NULL) {
6658 // failed array check
6659 return false;
6660 }
6661 // Figure out the size and type of the elements we will be copying.
6662 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6663 if (src_elem != T_BYTE) {
6664 return false;
6665 }
6666 // 'src_start' points to src array + offset
6667 Node* src_start = array_element_address(src, ofs, src_elem);
6668 Node* state = NULL;
6669 address stubAddr;
6670 const char *stubName;
6671
6672 switch(id) {
6673 case vmIntrinsics::_sha_implCompress:
6674 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6675 state = get_state_from_sha_object(sha_obj);
6676 stubAddr = StubRoutines::sha1_implCompress();
6677 stubName = "sha1_implCompress";
6678 break;
6679 case vmIntrinsics::_sha2_implCompress:
6680 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6681 state = get_state_from_sha_object(sha_obj);
6682 stubAddr = StubRoutines::sha256_implCompress();
6683 stubName = "sha256_implCompress";
6684 break;
6685 case vmIntrinsics::_sha5_implCompress:
6686 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6687 state = get_state_from_sha5_object(sha_obj);
6688 stubAddr = StubRoutines::sha512_implCompress();
6689 stubName = "sha512_implCompress";
6690 break;
6691 default:
6692 fatal_unexpected_iid(id);
6693 return false;
6694 }
6695 if (state == NULL) return false;
6696
6697 // Call the stub.
6698 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6699 stubAddr, stubName, TypePtr::BOTTOM,
6700 src_start, state);
6701
6702 return true;
6703 }
6704
6705 //------------------------------inline_digestBase_implCompressMB-----------------------
6706 //
6707 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6708 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6709 //
6710 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6711 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6712 "need SHA1/SHA256/SHA512 instruction support");
6713 assert((uint)predicate < 3, "sanity");
6714 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6715
6716 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6717 Node* src = argument(1); // byte[] array
6718 Node* ofs = argument(2); // type int
6719 Node* limit = argument(3); // type int
6720
6721 const Type* src_type = src->Value(&_gvn);
6722 const TypeAryPtr* top_src = src_type->isa_aryptr();
6723 if (top_src == NULL || top_src->klass() == NULL) {
6724 // failed array check
6725 return false;
6726 }
6727 // Figure out the size and type of the elements we will be copying.
6728 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6729 if (src_elem != T_BYTE) {
6730 return false;
6731 }
6732 // 'src_start' points to src array + offset
6733 Node* src_start = array_element_address(src, ofs, src_elem);
6734
6735 const char* klass_SHA_name = NULL;
6736 const char* stub_name = NULL;
6737 address stub_addr = NULL;
6738 bool long_state = false;
6739
6740 switch (predicate) {
6741 case 0:
6742 if (UseSHA1Intrinsics) {
6743 klass_SHA_name = "sun/security/provider/SHA";
6744 stub_name = "sha1_implCompressMB";
6745 stub_addr = StubRoutines::sha1_implCompressMB();
6746 }
6747 break;
6748 case 1:
6749 if (UseSHA256Intrinsics) {
6750 klass_SHA_name = "sun/security/provider/SHA2";
6751 stub_name = "sha256_implCompressMB";
6752 stub_addr = StubRoutines::sha256_implCompressMB();
6753 }
6754 break;
6755 case 2:
6756 if (UseSHA512Intrinsics) {
6757 klass_SHA_name = "sun/security/provider/SHA5";
6758 stub_name = "sha512_implCompressMB";
6759 stub_addr = StubRoutines::sha512_implCompressMB();
6760 long_state = true;
6761 }
6762 break;
6763 default:
6764 fatal("unknown SHA intrinsic predicate: %d", predicate);
6765 }
6766 if (klass_SHA_name != NULL) {
6767 // get DigestBase klass to lookup for SHA klass
6768 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6769 assert(tinst != NULL, "digestBase_obj is not instance???");
6770 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6771
6772 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6773 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6774 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6775 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6776 }
6777 return false;
6778 }
6779 //------------------------------inline_sha_implCompressMB-----------------------
6780 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6781 bool long_state, address stubAddr, const char *stubName,
6782 Node* src_start, Node* ofs, Node* limit) {
6783 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6784 const TypeOopPtr* xtype = aklass->as_instance_type();
6785 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6786 sha_obj = _gvn.transform(sha_obj);
6787
6788 Node* state;
6789 if (long_state) {
6790 state = get_state_from_sha5_object(sha_obj);
6791 } else {
6792 state = get_state_from_sha_object(sha_obj);
6793 }
6794 if (state == NULL) return false;
6795
6796 // Call the stub.
6797 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6798 OptoRuntime::digestBase_implCompressMB_Type(),
6799 stubAddr, stubName, TypePtr::BOTTOM,
6800 src_start, state, ofs, limit);
6801 // return ofs (int)
6802 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6803 set_result(result);
6804
6805 return true;
6806 }
6807
6808 //------------------------------get_state_from_sha_object-----------------------
6809 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6810 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6811 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6812 if (sha_state == NULL) return (Node *) NULL;
6813
6814 // now have the array, need to get the start address of the state array
6815 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6816 return state;
6817 }
6818
6819 //------------------------------get_state_from_sha5_object-----------------------
6820 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6821 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6822 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6823 if (sha_state == NULL) return (Node *) NULL;
6824
6825 // now have the array, need to get the start address of the state array
6826 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6827 return state;
6828 }
6829
6830 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6831 // Return node representing slow path of predicate check.
6832 // the pseudo code we want to emulate with this predicate is:
6833 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6834 //
6835 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6836 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6837 "need SHA1/SHA256/SHA512 instruction support");
6838 assert((uint)predicate < 3, "sanity");
6839
6840 // The receiver was checked for NULL already.
6841 Node* digestBaseObj = argument(0);
6842
6843 // get DigestBase klass for instanceOf check
6844 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6845 assert(tinst != NULL, "digestBaseObj is null");
6846 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6847
6848 const char* klass_SHA_name = NULL;
6849 switch (predicate) {
6850 case 0:
6851 if (UseSHA1Intrinsics) {
6852 // we want to do an instanceof comparison against the SHA class
6853 klass_SHA_name = "sun/security/provider/SHA";
6854 }
6855 break;
6856 case 1:
6857 if (UseSHA256Intrinsics) {
6858 // we want to do an instanceof comparison against the SHA2 class
6859 klass_SHA_name = "sun/security/provider/SHA2";
6860 }
6861 break;
6862 case 2:
6863 if (UseSHA512Intrinsics) {
6864 // we want to do an instanceof comparison against the SHA5 class
6865 klass_SHA_name = "sun/security/provider/SHA5";
6866 }
6867 break;
6868 default:
6869 fatal("unknown SHA intrinsic predicate: %d", predicate);
6870 }
6871
6872 ciKlass* klass_SHA = NULL;
6873 if (klass_SHA_name != NULL) {
6874 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6875 }
6876 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6877 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6878 Node* ctrl = control();
6879 set_control(top()); // no intrinsic path
6880 return ctrl;
6881 }
6882 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6883
6884 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6885 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6886 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6887 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6888
6889 return instof_false; // even if it is NULL
6890 }
6891
6892 //-------------inline_fma-----------------------------------
6893 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6894 Node *a = NULL;
6895 Node *b = NULL;
6896 Node *c = NULL;
6897 Node* result = NULL;
6898 switch (id) {
6899 case vmIntrinsics::_fmaD:
6900 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6901 // no receiver since it is static method
6902 a = round_double_node(argument(0));
6903 b = round_double_node(argument(2));
6904 c = round_double_node(argument(4));
6905 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6906 break;
6907 case vmIntrinsics::_fmaF:
6908 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6909 a = argument(0);
6910 b = argument(1);
6911 c = argument(2);
6912 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6913 break;
6914 default:
6915 fatal_unexpected_iid(id); break;
6916 }
6917 set_result(result);
6918 return true;
6919 }
6920
6921 bool LibraryCallKit::inline_profileBoolean() {
6922 Node* counts = argument(1);
6923 const TypeAryPtr* ary = NULL;
6924 ciArray* aobj = NULL;
6925 if (counts->is_Con()
6926 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6927 && (aobj = ary->const_oop()->as_array()) != NULL
6928 && (aobj->length() == 2)) {
6929 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6930 jint false_cnt = aobj->element_value(0).as_int();
6931 jint true_cnt = aobj->element_value(1).as_int();
6932
6933 if (C->log() != NULL) {
6934 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6935 false_cnt, true_cnt);
6936 }
6937
6938 if (false_cnt + true_cnt == 0) {
6939 // According to profile, never executed.
6940 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6941 Deoptimization::Action_reinterpret);
6942 return true;
6943 }
6944
6945 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6946 // is a number of each value occurrences.
6947 Node* result = argument(0);
6948 if (false_cnt == 0 || true_cnt == 0) {
6949 // According to profile, one value has been never seen.
6950 int expected_val = (false_cnt == 0) ? 1 : 0;
6951
6952 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6953 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6954
6955 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6956 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6957 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6958
6959 { // Slow path: uncommon trap for never seen value and then reexecute
6960 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6961 // the value has been seen at least once.
6962 PreserveJVMState pjvms(this);
6963 PreserveReexecuteState preexecs(this);
6964 jvms()->set_should_reexecute(true);
6965
6966 set_control(slow_path);
6967 set_i_o(i_o());
6968
6969 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6970 Deoptimization::Action_reinterpret);
6971 }
6972 // The guard for never seen value enables sharpening of the result and
6973 // returning a constant. It allows to eliminate branches on the same value
6974 // later on.
6975 set_control(fast_path);
6976 result = intcon(expected_val);
6977 }
6978 // Stop profiling.
6979 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6980 // By replacing method body with profile data (represented as ProfileBooleanNode
6981 // on IR level) we effectively disable profiling.
6982 // It enables full speed execution once optimized code is generated.
6983 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6984 C->record_for_igvn(profile);
6985 set_result(profile);
6986 return true;
6987 } else {
6988 // Continue profiling.
6989 // Profile data isn't available at the moment. So, execute method's bytecode version.
6990 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6991 // is compiled and counters aren't available since corresponding MethodHandle
6992 // isn't a compile-time constant.
6993 return false;
6994 }
6995 }
6996
6997 bool LibraryCallKit::inline_isCompileConstant() {
6998 Node* n = argument(0);
6999 set_result(n->is_Con() ? intcon(1) : intcon(0));
7000 return true;
7001 }