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_base64_encodeBlock();
305 bool inline_sha_implCompress(vmIntrinsics::ID id);
306 bool inline_digestBase_implCompressMB(int predicate);
307 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
308 bool long_state, address stubAddr, const char *stubName,
309 Node* src_start, Node* ofs, Node* limit);
310 Node* get_state_from_sha_object(Node *sha_object);
311 Node* get_state_from_sha5_object(Node *sha_object);
312 Node* inline_digestBase_implCompressMB_predicate(int predicate);
313 bool inline_encodeISOArray();
314 bool inline_updateCRC32();
315 bool inline_updateBytesCRC32();
316 bool inline_updateByteBufferCRC32();
317 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
318 bool inline_updateBytesCRC32C();
319 bool inline_updateDirectByteBufferCRC32C();
320 bool inline_updateBytesAdler32();
321 bool inline_updateByteBufferAdler32();
322 bool inline_multiplyToLen();
323 bool inline_hasNegatives();
324 bool inline_squareToLen();
325 bool inline_mulAdd();
326 bool inline_montgomeryMultiply();
327 bool inline_montgomerySquare();
328 bool inline_vectorizedMismatch();
329 bool inline_fma(vmIntrinsics::ID id);
330
331 bool inline_profileBoolean();
332 bool inline_isCompileConstant();
333 void clear_upper_avx() {
334 #ifdef X86
335 if (UseAVX >= 2) {
336 C->set_clear_upper_avx(true);
337 }
338 #endif
339 }
340 };
341
342 //---------------------------make_vm_intrinsic----------------------------
343 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
344 vmIntrinsics::ID id = m->intrinsic_id();
345 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
346
347 if (!m->is_loaded()) {
348 // Do not attempt to inline unloaded methods.
349 return NULL;
350 }
351
352 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
353 bool is_available = false;
354
355 {
356 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
357 // the compiler must transition to '_thread_in_vm' state because both
358 // methods access VM-internal data.
359 VM_ENTRY_MARK;
360 methodHandle mh(THREAD, m->get_Method());
361 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
362 !C->directive()->is_intrinsic_disabled(mh) &&
363 !vmIntrinsics::is_disabled_by_flags(mh);
364
365 }
366
367 if (is_available) {
368 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
369 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
370 return new LibraryIntrinsic(m, is_virtual,
371 vmIntrinsics::predicates_needed(id),
372 vmIntrinsics::does_virtual_dispatch(id),
373 (vmIntrinsics::ID) id);
374 } else {
375 return NULL;
376 }
377 }
378
379 //----------------------register_library_intrinsics-----------------------
380 // Initialize this file's data structures, for each Compile instance.
381 void Compile::register_library_intrinsics() {
382 // Nothing to do here.
383 }
384
385 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
386 LibraryCallKit kit(jvms, this);
387 Compile* C = kit.C;
388 int nodes = C->unique();
389 #ifndef PRODUCT
390 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
391 char buf[1000];
392 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
393 tty->print_cr("Intrinsic %s", str);
394 }
395 #endif
396 ciMethod* callee = kit.callee();
397 const int bci = kit.bci();
398
399 // Try to inline the intrinsic.
400 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
401 kit.try_to_inline(_last_predicate)) {
402 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
403 : "(intrinsic)";
404 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
405 if (C->print_intrinsics() || C->print_inlining()) {
406 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
407 }
408 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
409 if (C->log()) {
410 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
411 vmIntrinsics::name_at(intrinsic_id()),
412 (is_virtual() ? " virtual='1'" : ""),
413 C->unique() - nodes);
414 }
415 // Push the result from the inlined method onto the stack.
416 kit.push_result();
417 C->print_inlining_update(this);
418 return kit.transfer_exceptions_into_jvms();
419 }
420
421 // The intrinsic bailed out
422 if (jvms->has_method()) {
423 // Not a root compile.
424 const char* msg;
425 if (callee->intrinsic_candidate()) {
426 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
427 } else {
428 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
429 : "failed to inline (intrinsic), method not annotated";
430 }
431 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
432 if (C->print_intrinsics() || C->print_inlining()) {
433 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
434 }
435 } else {
436 // Root compile
437 ResourceMark rm;
438 stringStream msg_stream;
439 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
440 vmIntrinsics::name_at(intrinsic_id()),
441 is_virtual() ? " (virtual)" : "", bci);
442 const char *msg = msg_stream.as_string();
443 log_debug(jit, inlining)("%s", msg);
444 if (C->print_intrinsics() || C->print_inlining()) {
445 tty->print("%s", msg);
446 }
447 }
448 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
449 C->print_inlining_update(this);
450 return NULL;
451 }
452
453 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
454 LibraryCallKit kit(jvms, this);
455 Compile* C = kit.C;
456 int nodes = C->unique();
457 _last_predicate = predicate;
458 #ifndef PRODUCT
459 assert(is_predicated() && predicate < predicates_count(), "sanity");
460 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
461 char buf[1000];
462 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
463 tty->print_cr("Predicate for intrinsic %s", str);
464 }
465 #endif
466 ciMethod* callee = kit.callee();
467 const int bci = kit.bci();
468
469 Node* slow_ctl = kit.try_to_predicate(predicate);
470 if (!kit.failing()) {
471 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
472 : "(intrinsic, predicate)";
473 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
474 if (C->print_intrinsics() || C->print_inlining()) {
475 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
476 }
477 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
478 if (C->log()) {
479 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
480 vmIntrinsics::name_at(intrinsic_id()),
481 (is_virtual() ? " virtual='1'" : ""),
482 C->unique() - nodes);
483 }
484 return slow_ctl; // Could be NULL if the check folds.
485 }
486
487 // The intrinsic bailed out
488 if (jvms->has_method()) {
489 // Not a root compile.
490 const char* msg = "failed to generate predicate for intrinsic";
491 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
492 if (C->print_intrinsics() || C->print_inlining()) {
493 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
494 }
495 } else {
496 // Root compile
497 ResourceMark rm;
498 stringStream msg_stream;
499 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
500 vmIntrinsics::name_at(intrinsic_id()),
501 is_virtual() ? " (virtual)" : "", bci);
502 const char *msg = msg_stream.as_string();
503 log_debug(jit, inlining)("%s", msg);
504 if (C->print_intrinsics() || C->print_inlining()) {
505 C->print_inlining_stream()->print("%s", msg);
506 }
507 }
508 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
509 return NULL;
510 }
511
512 bool LibraryCallKit::try_to_inline(int predicate) {
513 // Handle symbolic names for otherwise undistinguished boolean switches:
514 const bool is_store = true;
515 const bool is_compress = true;
516 const bool is_static = true;
517 const bool is_volatile = true;
518
519 if (!jvms()->has_method()) {
520 // Root JVMState has a null method.
521 assert(map()->memory()->Opcode() == Op_Parm, "");
522 // Insert the memory aliasing node
523 set_all_memory(reset_memory());
524 }
525 assert(merged_memory(), "");
526
527
528 switch (intrinsic_id()) {
529 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
530 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
531 case vmIntrinsics::_getClass: return inline_native_getClass();
532
533 case vmIntrinsics::_dsin:
534 case vmIntrinsics::_dcos:
535 case vmIntrinsics::_dtan:
536 case vmIntrinsics::_dabs:
537 case vmIntrinsics::_datan2:
538 case vmIntrinsics::_dsqrt:
539 case vmIntrinsics::_dexp:
540 case vmIntrinsics::_dlog:
541 case vmIntrinsics::_dlog10:
542 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
543
544 case vmIntrinsics::_min:
545 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
546
547 case vmIntrinsics::_notify:
548 case vmIntrinsics::_notifyAll:
549 if (InlineNotify) {
550 return inline_notify(intrinsic_id());
551 }
552 return false;
553
554 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
555 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
556 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
557 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
558 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
559 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
560 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
561 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
562 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
563 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
564 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
565 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
566 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
567
568 case vmIntrinsics::_arraycopy: return inline_arraycopy();
569
570 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
571 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
572 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
573 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
574
575 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
576 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
577 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
578 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
579 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
580 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
581 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
582
583 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
584 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
585
586 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
587 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
588 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
589 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
590
591 case vmIntrinsics::_compressStringC:
592 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
593 case vmIntrinsics::_inflateStringC:
594 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
595
596 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
597 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
598 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
599 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
600 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
601 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
602 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
603 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
604 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
605
606 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
607 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
608 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
609 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
610 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
611 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
612 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
613 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
614 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
615
616 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
617 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
618 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
619 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
620 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
621 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
622 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
623 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
624 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
625
626 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
627 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
628 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
629 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
630 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
631 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
632 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
633 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
634 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
635
636 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
637 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
638 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
639 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
640
641 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
642 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
643 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
644 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
645
646 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
647 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
648 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
649 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
650 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
651 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
652 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
653 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
654 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
655
656 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
657 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
658 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
659 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
660 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
661 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
662 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
663 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
664 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
665
666 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
667 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
668 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
669 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
670 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
671 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
672 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
673 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
674 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
675
676 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
677 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
678 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
679 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
680 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
681 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
682 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
683 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
684 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
685
686 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
687 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
688 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
689 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
690 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
691
692 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
693 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
694 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
695 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
696 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
697 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
698 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
699 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
700 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
701 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
702 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
703 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
704 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
705 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
706 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
707 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
708 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
709 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
710 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
711 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
712
713 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
714 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
715 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
716 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
717 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
718 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
719 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
720 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
721 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
722 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
723 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
724 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
725 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
726 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
727 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
728
729 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
730 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
731 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
732 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
733
734 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
735 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
736 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
737 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
738 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
739
740 case vmIntrinsics::_loadFence:
741 case vmIntrinsics::_storeFence:
742 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
743
744 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
745
746 case vmIntrinsics::_currentThread: return inline_native_currentThread();
747 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
748
749 #ifdef TRACE_HAVE_INTRINSICS
750 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
751 case vmIntrinsics::_getClassId: return inline_native_classID();
752 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
753 #endif
754 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
755 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
756 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
757 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
758 case vmIntrinsics::_getLength: return inline_native_getLength();
759 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
760 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
761 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
762 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
763 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
764 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
765
766 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
767 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
768
769 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
770
771 case vmIntrinsics::_isInstance:
772 case vmIntrinsics::_getModifiers:
773 case vmIntrinsics::_isInterface:
774 case vmIntrinsics::_isArray:
775 case vmIntrinsics::_isPrimitive:
776 case vmIntrinsics::_getSuperclass:
777 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
778
779 case vmIntrinsics::_floatToRawIntBits:
780 case vmIntrinsics::_floatToIntBits:
781 case vmIntrinsics::_intBitsToFloat:
782 case vmIntrinsics::_doubleToRawLongBits:
783 case vmIntrinsics::_doubleToLongBits:
784 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
785
786 case vmIntrinsics::_numberOfLeadingZeros_i:
787 case vmIntrinsics::_numberOfLeadingZeros_l:
788 case vmIntrinsics::_numberOfTrailingZeros_i:
789 case vmIntrinsics::_numberOfTrailingZeros_l:
790 case vmIntrinsics::_bitCount_i:
791 case vmIntrinsics::_bitCount_l:
792 case vmIntrinsics::_reverseBytes_i:
793 case vmIntrinsics::_reverseBytes_l:
794 case vmIntrinsics::_reverseBytes_s:
795 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
796
797 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
798
799 case vmIntrinsics::_Reference_get: return inline_reference_get();
800
801 case vmIntrinsics::_Class_cast: return inline_Class_cast();
802
803 case vmIntrinsics::_aescrypt_encryptBlock:
804 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
805
806 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
807 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
808 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
809
810 case vmIntrinsics::_counterMode_AESCrypt:
811 return inline_counterMode_AESCrypt(intrinsic_id());
812
813 case vmIntrinsics::_sha_implCompress:
814 case vmIntrinsics::_sha2_implCompress:
815 case vmIntrinsics::_sha5_implCompress:
816 return inline_sha_implCompress(intrinsic_id());
817
818 case vmIntrinsics::_digestBase_implCompressMB:
819 return inline_digestBase_implCompressMB(predicate);
820
821 case vmIntrinsics::_multiplyToLen:
822 return inline_multiplyToLen();
823
824 case vmIntrinsics::_squareToLen:
825 return inline_squareToLen();
826
827 case vmIntrinsics::_mulAdd:
828 return inline_mulAdd();
829
830 case vmIntrinsics::_montgomeryMultiply:
831 return inline_montgomeryMultiply();
832 case vmIntrinsics::_montgomerySquare:
833 return inline_montgomerySquare();
834
835 case vmIntrinsics::_vectorizedMismatch:
836 return inline_vectorizedMismatch();
837
838 case vmIntrinsics::_ghash_processBlocks:
839 return inline_ghash_processBlocks();
840 case vmIntrinsics::_base64_encodeBlock:
841 return inline_base64_encodeBlock();
842
843 case vmIntrinsics::_encodeISOArray:
844 case vmIntrinsics::_encodeByteISOArray:
845 return inline_encodeISOArray();
846
847 case vmIntrinsics::_updateCRC32:
848 return inline_updateCRC32();
849 case vmIntrinsics::_updateBytesCRC32:
850 return inline_updateBytesCRC32();
851 case vmIntrinsics::_updateByteBufferCRC32:
852 return inline_updateByteBufferCRC32();
853
854 case vmIntrinsics::_updateBytesCRC32C:
855 return inline_updateBytesCRC32C();
856 case vmIntrinsics::_updateDirectByteBufferCRC32C:
857 return inline_updateDirectByteBufferCRC32C();
858
859 case vmIntrinsics::_updateBytesAdler32:
860 return inline_updateBytesAdler32();
861 case vmIntrinsics::_updateByteBufferAdler32:
862 return inline_updateByteBufferAdler32();
863
864 case vmIntrinsics::_profileBoolean:
865 return inline_profileBoolean();
866 case vmIntrinsics::_isCompileConstant:
867 return inline_isCompileConstant();
868
869 case vmIntrinsics::_hasNegatives:
870 return inline_hasNegatives();
871
872 case vmIntrinsics::_fmaD:
873 case vmIntrinsics::_fmaF:
874 return inline_fma(intrinsic_id());
875
876 default:
877 // If you get here, it may be that someone has added a new intrinsic
878 // to the list in vmSymbols.hpp without implementing it here.
879 #ifndef PRODUCT
880 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
881 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
882 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
883 }
884 #endif
885 return false;
886 }
887 }
888
889 Node* LibraryCallKit::try_to_predicate(int predicate) {
890 if (!jvms()->has_method()) {
891 // Root JVMState has a null method.
892 assert(map()->memory()->Opcode() == Op_Parm, "");
893 // Insert the memory aliasing node
894 set_all_memory(reset_memory());
895 }
896 assert(merged_memory(), "");
897
898 switch (intrinsic_id()) {
899 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
900 return inline_cipherBlockChaining_AESCrypt_predicate(false);
901 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
902 return inline_cipherBlockChaining_AESCrypt_predicate(true);
903 case vmIntrinsics::_counterMode_AESCrypt:
904 return inline_counterMode_AESCrypt_predicate();
905 case vmIntrinsics::_digestBase_implCompressMB:
906 return inline_digestBase_implCompressMB_predicate(predicate);
907
908 default:
909 // If you get here, it may be that someone has added a new intrinsic
910 // to the list in vmSymbols.hpp without implementing it here.
911 #ifndef PRODUCT
912 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
913 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
914 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
915 }
916 #endif
917 Node* slow_ctl = control();
918 set_control(top()); // No fast path instrinsic
919 return slow_ctl;
920 }
921 }
922
923 //------------------------------set_result-------------------------------
924 // Helper function for finishing intrinsics.
925 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
926 record_for_igvn(region);
927 set_control(_gvn.transform(region));
928 set_result( _gvn.transform(value));
929 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
930 }
931
932 //------------------------------generate_guard---------------------------
933 // Helper function for generating guarded fast-slow graph structures.
934 // The given 'test', if true, guards a slow path. If the test fails
935 // then a fast path can be taken. (We generally hope it fails.)
936 // In all cases, GraphKit::control() is updated to the fast path.
937 // The returned value represents the control for the slow path.
938 // The return value is never 'top'; it is either a valid control
939 // or NULL if it is obvious that the slow path can never be taken.
940 // Also, if region and the slow control are not NULL, the slow edge
941 // is appended to the region.
942 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
943 if (stopped()) {
944 // Already short circuited.
945 return NULL;
946 }
947
948 // Build an if node and its projections.
949 // If test is true we take the slow path, which we assume is uncommon.
950 if (_gvn.type(test) == TypeInt::ZERO) {
951 // The slow branch is never taken. No need to build this guard.
952 return NULL;
953 }
954
955 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
956
957 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
958 if (if_slow == top()) {
959 // The slow branch is never taken. No need to build this guard.
960 return NULL;
961 }
962
963 if (region != NULL)
964 region->add_req(if_slow);
965
966 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
967 set_control(if_fast);
968
969 return if_slow;
970 }
971
972 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
973 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
974 }
975 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
976 return generate_guard(test, region, PROB_FAIR);
977 }
978
979 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
980 Node* *pos_index) {
981 if (stopped())
982 return NULL; // already stopped
983 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
984 return NULL; // index is already adequately typed
985 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
986 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
987 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
988 if (is_neg != NULL && pos_index != NULL) {
989 // Emulate effect of Parse::adjust_map_after_if.
990 Node* ccast = new CastIINode(index, TypeInt::POS);
991 ccast->set_req(0, control());
992 (*pos_index) = _gvn.transform(ccast);
993 }
994 return is_neg;
995 }
996
997 // Make sure that 'position' is a valid limit index, in [0..length].
998 // There are two equivalent plans for checking this:
999 // A. (offset + copyLength) unsigned<= arrayLength
1000 // B. offset <= (arrayLength - copyLength)
1001 // We require that all of the values above, except for the sum and
1002 // difference, are already known to be non-negative.
1003 // Plan A is robust in the face of overflow, if offset and copyLength
1004 // are both hugely positive.
1005 //
1006 // Plan B is less direct and intuitive, but it does not overflow at
1007 // all, since the difference of two non-negatives is always
1008 // representable. Whenever Java methods must perform the equivalent
1009 // check they generally use Plan B instead of Plan A.
1010 // For the moment we use Plan A.
1011 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1012 Node* subseq_length,
1013 Node* array_length,
1014 RegionNode* region) {
1015 if (stopped())
1016 return NULL; // already stopped
1017 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1018 if (zero_offset && subseq_length->eqv_uncast(array_length))
1019 return NULL; // common case of whole-array copy
1020 Node* last = subseq_length;
1021 if (!zero_offset) // last += offset
1022 last = _gvn.transform(new AddINode(last, offset));
1023 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1024 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1025 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1026 return is_over;
1027 }
1028
1029 // Emit range checks for the given String.value byte array
1030 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1031 if (stopped()) {
1032 return; // already stopped
1033 }
1034 RegionNode* bailout = new RegionNode(1);
1035 record_for_igvn(bailout);
1036 if (char_count) {
1037 // Convert char count to byte count
1038 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1039 }
1040
1041 // Offset and count must not be negative
1042 generate_negative_guard(offset, bailout);
1043 generate_negative_guard(count, bailout);
1044 // Offset + count must not exceed length of array
1045 generate_limit_guard(offset, count, load_array_length(array), bailout);
1046
1047 if (bailout->req() > 1) {
1048 PreserveJVMState pjvms(this);
1049 set_control(_gvn.transform(bailout));
1050 uncommon_trap(Deoptimization::Reason_intrinsic,
1051 Deoptimization::Action_maybe_recompile);
1052 }
1053 }
1054
1055 //--------------------------generate_current_thread--------------------
1056 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1057 ciKlass* thread_klass = env()->Thread_klass();
1058 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1059 Node* thread = _gvn.transform(new ThreadLocalNode());
1060 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1061 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1062 tls_output = thread;
1063 return threadObj;
1064 }
1065
1066
1067 //------------------------------make_string_method_node------------------------
1068 // Helper method for String intrinsic functions. This version is called with
1069 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1070 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1071 // containing the lengths of str1 and str2.
1072 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1073 Node* result = NULL;
1074 switch (opcode) {
1075 case Op_StrIndexOf:
1076 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1077 str1_start, cnt1, str2_start, cnt2, ae);
1078 break;
1079 case Op_StrComp:
1080 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1081 str1_start, cnt1, str2_start, cnt2, ae);
1082 break;
1083 case Op_StrEquals:
1084 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1085 // Use the constant length if there is one because optimized match rule may exist.
1086 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1087 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1088 break;
1089 default:
1090 ShouldNotReachHere();
1091 return NULL;
1092 }
1093
1094 // All these intrinsics have checks.
1095 C->set_has_split_ifs(true); // Has chance for split-if optimization
1096 clear_upper_avx();
1097
1098 return _gvn.transform(result);
1099 }
1100
1101 //------------------------------inline_string_compareTo------------------------
1102 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1103 Node* arg1 = argument(0);
1104 Node* arg2 = argument(1);
1105
1106 // Get start addr and length of first argument
1107 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1108 Node* arg1_cnt = load_array_length(arg1);
1109
1110 // Get start addr and length of second argument
1111 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1112 Node* arg2_cnt = load_array_length(arg2);
1113
1114 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1115 set_result(result);
1116 return true;
1117 }
1118
1119 //------------------------------inline_string_equals------------------------
1120 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1121 Node* arg1 = argument(0);
1122 Node* arg2 = argument(1);
1123
1124 // paths (plus control) merge
1125 RegionNode* region = new RegionNode(3);
1126 Node* phi = new PhiNode(region, TypeInt::BOOL);
1127
1128 if (!stopped()) {
1129 // Get start addr and length of first argument
1130 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1131 Node* arg1_cnt = load_array_length(arg1);
1132
1133 // Get start addr and length of second argument
1134 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1135 Node* arg2_cnt = load_array_length(arg2);
1136
1137 // Check for arg1_cnt != arg2_cnt
1138 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1139 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1140 Node* if_ne = generate_slow_guard(bol, NULL);
1141 if (if_ne != NULL) {
1142 phi->init_req(2, intcon(0));
1143 region->init_req(2, if_ne);
1144 }
1145
1146 // Check for count == 0 is done by assembler code for StrEquals.
1147
1148 if (!stopped()) {
1149 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1150 phi->init_req(1, equals);
1151 region->init_req(1, control());
1152 }
1153 }
1154
1155 // post merge
1156 set_control(_gvn.transform(region));
1157 record_for_igvn(region);
1158
1159 set_result(_gvn.transform(phi));
1160 return true;
1161 }
1162
1163 //------------------------------inline_array_equals----------------------------
1164 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1165 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1166 Node* arg1 = argument(0);
1167 Node* arg2 = argument(1);
1168
1169 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1170 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1171 clear_upper_avx();
1172
1173 return true;
1174 }
1175
1176 //------------------------------inline_hasNegatives------------------------------
1177 bool LibraryCallKit::inline_hasNegatives() {
1178 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1179 return false;
1180 }
1181
1182 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1183 // no receiver since it is static method
1184 Node* ba = argument(0);
1185 Node* offset = argument(1);
1186 Node* len = argument(2);
1187
1188 // Range checks
1189 generate_string_range_check(ba, offset, len, false);
1190 if (stopped()) {
1191 return true;
1192 }
1193 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1194 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1195 set_result(_gvn.transform(result));
1196 return true;
1197 }
1198
1199 bool LibraryCallKit::inline_preconditions_checkIndex() {
1200 Node* index = argument(0);
1201 Node* length = argument(1);
1202 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1203 return false;
1204 }
1205
1206 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1207 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1208
1209 {
1210 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1211 uncommon_trap(Deoptimization::Reason_intrinsic,
1212 Deoptimization::Action_make_not_entrant);
1213 }
1214
1215 if (stopped()) {
1216 return false;
1217 }
1218
1219 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1220 BoolTest::mask btest = BoolTest::lt;
1221 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1222 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1223 _gvn.set_type(rc, rc->Value(&_gvn));
1224 if (!rc_bool->is_Con()) {
1225 record_for_igvn(rc);
1226 }
1227 set_control(_gvn.transform(new IfTrueNode(rc)));
1228 {
1229 PreserveJVMState pjvms(this);
1230 set_control(_gvn.transform(new IfFalseNode(rc)));
1231 uncommon_trap(Deoptimization::Reason_range_check,
1232 Deoptimization::Action_make_not_entrant);
1233 }
1234
1235 if (stopped()) {
1236 return false;
1237 }
1238
1239 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1240 result->set_req(0, control());
1241 result = _gvn.transform(result);
1242 set_result(result);
1243 replace_in_map(index, result);
1244 clear_upper_avx();
1245 return true;
1246 }
1247
1248 //------------------------------inline_string_indexOf------------------------
1249 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1250 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1251 return false;
1252 }
1253 Node* src = argument(0);
1254 Node* tgt = argument(1);
1255
1256 // Make the merge point
1257 RegionNode* result_rgn = new RegionNode(4);
1258 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1259
1260 // Get start addr and length of source string
1261 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1262 Node* src_count = load_array_length(src);
1263
1264 // Get start addr and length of substring
1265 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1266 Node* tgt_count = load_array_length(tgt);
1267
1268 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1269 // Divide src size by 2 if String is UTF16 encoded
1270 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1271 }
1272 if (ae == StrIntrinsicNode::UU) {
1273 // Divide substring size by 2 if String is UTF16 encoded
1274 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1275 }
1276
1277 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1278 if (result != NULL) {
1279 result_phi->init_req(3, result);
1280 result_rgn->init_req(3, control());
1281 }
1282 set_control(_gvn.transform(result_rgn));
1283 record_for_igvn(result_rgn);
1284 set_result(_gvn.transform(result_phi));
1285
1286 return true;
1287 }
1288
1289 //-----------------------------inline_string_indexOf-----------------------
1290 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1291 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1292 return false;
1293 }
1294 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1295 return false;
1296 }
1297 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1298 Node* src = argument(0); // byte[]
1299 Node* src_count = argument(1); // char count
1300 Node* tgt = argument(2); // byte[]
1301 Node* tgt_count = argument(3); // char count
1302 Node* from_index = argument(4); // char index
1303
1304 // Multiply byte array index by 2 if String is UTF16 encoded
1305 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1306 src_count = _gvn.transform(new SubINode(src_count, from_index));
1307 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1308 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1309
1310 // Range checks
1311 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1312 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1313 if (stopped()) {
1314 return true;
1315 }
1316
1317 RegionNode* region = new RegionNode(5);
1318 Node* phi = new PhiNode(region, TypeInt::INT);
1319
1320 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1321 if (result != NULL) {
1322 // The result is index relative to from_index if substring was found, -1 otherwise.
1323 // Generate code which will fold into cmove.
1324 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1325 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1326
1327 Node* if_lt = generate_slow_guard(bol, NULL);
1328 if (if_lt != NULL) {
1329 // result == -1
1330 phi->init_req(3, result);
1331 region->init_req(3, if_lt);
1332 }
1333 if (!stopped()) {
1334 result = _gvn.transform(new AddINode(result, from_index));
1335 phi->init_req(4, result);
1336 region->init_req(4, control());
1337 }
1338 }
1339
1340 set_control(_gvn.transform(region));
1341 record_for_igvn(region);
1342 set_result(_gvn.transform(phi));
1343 clear_upper_avx();
1344
1345 return true;
1346 }
1347
1348 // Create StrIndexOfNode with fast path checks
1349 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1350 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1351 // Check for substr count > string count
1352 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1353 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1354 Node* if_gt = generate_slow_guard(bol, NULL);
1355 if (if_gt != NULL) {
1356 phi->init_req(1, intcon(-1));
1357 region->init_req(1, if_gt);
1358 }
1359 if (!stopped()) {
1360 // Check for substr count == 0
1361 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1362 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1363 Node* if_zero = generate_slow_guard(bol, NULL);
1364 if (if_zero != NULL) {
1365 phi->init_req(2, intcon(0));
1366 region->init_req(2, if_zero);
1367 }
1368 }
1369 if (!stopped()) {
1370 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1371 }
1372 return NULL;
1373 }
1374
1375 //-----------------------------inline_string_indexOfChar-----------------------
1376 bool LibraryCallKit::inline_string_indexOfChar() {
1377 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1378 return false;
1379 }
1380 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1381 return false;
1382 }
1383 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1384 Node* src = argument(0); // byte[]
1385 Node* tgt = argument(1); // tgt is int ch
1386 Node* from_index = argument(2);
1387 Node* max = argument(3);
1388
1389 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1390 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1391 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1392
1393 // Range checks
1394 generate_string_range_check(src, src_offset, src_count, true);
1395 if (stopped()) {
1396 return true;
1397 }
1398
1399 RegionNode* region = new RegionNode(3);
1400 Node* phi = new PhiNode(region, TypeInt::INT);
1401
1402 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1403 C->set_has_split_ifs(true); // Has chance for split-if optimization
1404 _gvn.transform(result);
1405
1406 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1407 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1408
1409 Node* if_lt = generate_slow_guard(bol, NULL);
1410 if (if_lt != NULL) {
1411 // result == -1
1412 phi->init_req(2, result);
1413 region->init_req(2, if_lt);
1414 }
1415 if (!stopped()) {
1416 result = _gvn.transform(new AddINode(result, from_index));
1417 phi->init_req(1, result);
1418 region->init_req(1, control());
1419 }
1420 set_control(_gvn.transform(region));
1421 record_for_igvn(region);
1422 set_result(_gvn.transform(phi));
1423
1424 return true;
1425 }
1426 //---------------------------inline_string_copy---------------------
1427 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1428 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1429 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1430 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1431 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1432 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1433 bool LibraryCallKit::inline_string_copy(bool compress) {
1434 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1435 return false;
1436 }
1437 int nargs = 5; // 2 oops, 3 ints
1438 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1439
1440 Node* src = argument(0);
1441 Node* src_offset = argument(1);
1442 Node* dst = argument(2);
1443 Node* dst_offset = argument(3);
1444 Node* length = argument(4);
1445
1446 // Check for allocation before we add nodes that would confuse
1447 // tightly_coupled_allocation()
1448 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1449
1450 // Figure out the size and type of the elements we will be copying.
1451 const Type* src_type = src->Value(&_gvn);
1452 const Type* dst_type = dst->Value(&_gvn);
1453 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1454 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1455 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1456 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1457 "Unsupported array types for inline_string_copy");
1458
1459 // Convert char[] offsets to byte[] offsets
1460 bool convert_src = (compress && src_elem == T_BYTE);
1461 bool convert_dst = (!compress && dst_elem == T_BYTE);
1462 if (convert_src) {
1463 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1464 } else if (convert_dst) {
1465 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1466 }
1467
1468 // Range checks
1469 generate_string_range_check(src, src_offset, length, convert_src);
1470 generate_string_range_check(dst, dst_offset, length, convert_dst);
1471 if (stopped()) {
1472 return true;
1473 }
1474
1475 Node* src_start = array_element_address(src, src_offset, src_elem);
1476 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1477 // 'src_start' points to src array + scaled offset
1478 // 'dst_start' points to dst array + scaled offset
1479 Node* count = NULL;
1480 if (compress) {
1481 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1482 } else {
1483 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1484 }
1485
1486 if (alloc != NULL) {
1487 if (alloc->maybe_set_complete(&_gvn)) {
1488 // "You break it, you buy it."
1489 InitializeNode* init = alloc->initialization();
1490 assert(init->is_complete(), "we just did this");
1491 init->set_complete_with_arraycopy();
1492 assert(dst->is_CheckCastPP(), "sanity");
1493 assert(dst->in(0)->in(0) == init, "dest pinned");
1494 }
1495 // Do not let stores that initialize this object be reordered with
1496 // a subsequent store that would make this object accessible by
1497 // other threads.
1498 // Record what AllocateNode this StoreStore protects so that
1499 // escape analysis can go from the MemBarStoreStoreNode to the
1500 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1501 // based on the escape status of the AllocateNode.
1502 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1503 }
1504 if (compress) {
1505 set_result(_gvn.transform(count));
1506 }
1507 clear_upper_avx();
1508
1509 return true;
1510 }
1511
1512 #ifdef _LP64
1513 #define XTOP ,top() /*additional argument*/
1514 #else //_LP64
1515 #define XTOP /*no additional argument*/
1516 #endif //_LP64
1517
1518 //------------------------inline_string_toBytesU--------------------------
1519 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1520 bool LibraryCallKit::inline_string_toBytesU() {
1521 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1522 return false;
1523 }
1524 // Get the arguments.
1525 Node* value = argument(0);
1526 Node* offset = argument(1);
1527 Node* length = argument(2);
1528
1529 Node* newcopy = NULL;
1530
1531 // Set the original stack and the reexecute bit for the interpreter to reexecute
1532 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1533 { PreserveReexecuteState preexecs(this);
1534 jvms()->set_should_reexecute(true);
1535
1536 // Check if a null path was taken unconditionally.
1537 value = null_check(value);
1538
1539 RegionNode* bailout = new RegionNode(1);
1540 record_for_igvn(bailout);
1541
1542 // Range checks
1543 generate_negative_guard(offset, bailout);
1544 generate_negative_guard(length, bailout);
1545 generate_limit_guard(offset, length, load_array_length(value), bailout);
1546 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1547 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1548
1549 if (bailout->req() > 1) {
1550 PreserveJVMState pjvms(this);
1551 set_control(_gvn.transform(bailout));
1552 uncommon_trap(Deoptimization::Reason_intrinsic,
1553 Deoptimization::Action_maybe_recompile);
1554 }
1555 if (stopped()) {
1556 return true;
1557 }
1558
1559 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1560 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1561 newcopy = new_array(klass_node, size, 0); // no arguments to push
1562 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1563
1564 // Calculate starting addresses.
1565 Node* src_start = array_element_address(value, offset, T_CHAR);
1566 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1567
1568 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1569 const TypeInt* toffset = gvn().type(offset)->is_int();
1570 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1571
1572 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1573 const char* copyfunc_name = "arraycopy";
1574 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1575 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1576 OptoRuntime::fast_arraycopy_Type(),
1577 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1578 src_start, dst_start, ConvI2X(length) XTOP);
1579 // Do not let reads from the cloned object float above the arraycopy.
1580 if (alloc != NULL) {
1581 if (alloc->maybe_set_complete(&_gvn)) {
1582 // "You break it, you buy it."
1583 InitializeNode* init = alloc->initialization();
1584 assert(init->is_complete(), "we just did this");
1585 init->set_complete_with_arraycopy();
1586 assert(newcopy->is_CheckCastPP(), "sanity");
1587 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1588 }
1589 // Do not let stores that initialize this object be reordered with
1590 // a subsequent store that would make this object accessible by
1591 // other threads.
1592 // Record what AllocateNode this StoreStore protects so that
1593 // escape analysis can go from the MemBarStoreStoreNode to the
1594 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1595 // based on the escape status of the AllocateNode.
1596 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1597 } else {
1598 insert_mem_bar(Op_MemBarCPUOrder);
1599 }
1600 } // original reexecute is set back here
1601
1602 C->set_has_split_ifs(true); // Has chance for split-if optimization
1603 if (!stopped()) {
1604 set_result(newcopy);
1605 }
1606 clear_upper_avx();
1607
1608 return true;
1609 }
1610
1611 //------------------------inline_string_getCharsU--------------------------
1612 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1613 bool LibraryCallKit::inline_string_getCharsU() {
1614 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1615 return false;
1616 }
1617
1618 // Get the arguments.
1619 Node* src = argument(0);
1620 Node* src_begin = argument(1);
1621 Node* src_end = argument(2); // exclusive offset (i < src_end)
1622 Node* dst = argument(3);
1623 Node* dst_begin = argument(4);
1624
1625 // Check for allocation before we add nodes that would confuse
1626 // tightly_coupled_allocation()
1627 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1628
1629 // Check if a null path was taken unconditionally.
1630 src = null_check(src);
1631 dst = null_check(dst);
1632 if (stopped()) {
1633 return true;
1634 }
1635
1636 // Get length and convert char[] offset to byte[] offset
1637 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1638 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1639
1640 // Range checks
1641 generate_string_range_check(src, src_begin, length, true);
1642 generate_string_range_check(dst, dst_begin, length, false);
1643 if (stopped()) {
1644 return true;
1645 }
1646
1647 if (!stopped()) {
1648 // Calculate starting addresses.
1649 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1650 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1651
1652 // Check if array addresses are aligned to HeapWordSize
1653 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1654 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1655 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1656 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1657
1658 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1659 const char* copyfunc_name = "arraycopy";
1660 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1661 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1662 OptoRuntime::fast_arraycopy_Type(),
1663 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1664 src_start, dst_start, ConvI2X(length) XTOP);
1665 // Do not let reads from the cloned object float above the arraycopy.
1666 if (alloc != NULL) {
1667 if (alloc->maybe_set_complete(&_gvn)) {
1668 // "You break it, you buy it."
1669 InitializeNode* init = alloc->initialization();
1670 assert(init->is_complete(), "we just did this");
1671 init->set_complete_with_arraycopy();
1672 assert(dst->is_CheckCastPP(), "sanity");
1673 assert(dst->in(0)->in(0) == init, "dest pinned");
1674 }
1675 // Do not let stores that initialize this object be reordered with
1676 // a subsequent store that would make this object accessible by
1677 // other threads.
1678 // Record what AllocateNode this StoreStore protects so that
1679 // escape analysis can go from the MemBarStoreStoreNode to the
1680 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1681 // based on the escape status of the AllocateNode.
1682 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1683 } else {
1684 insert_mem_bar(Op_MemBarCPUOrder);
1685 }
1686 }
1687
1688 C->set_has_split_ifs(true); // Has chance for split-if optimization
1689 return true;
1690 }
1691
1692 //----------------------inline_string_char_access----------------------------
1693 // Store/Load char to/from byte[] array.
1694 // static void StringUTF16.putChar(byte[] val, int index, int c)
1695 // static char StringUTF16.getChar(byte[] val, int index)
1696 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1697 Node* value = argument(0);
1698 Node* index = argument(1);
1699 Node* ch = is_store ? argument(2) : NULL;
1700
1701 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1702 // correctly requires matched array shapes.
1703 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1704 "sanity: byte[] and char[] bases agree");
1705 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1706 "sanity: byte[] and char[] scales agree");
1707
1708 // Bail when getChar over constants is requested: constant folding would
1709 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1710 // Java method would constant fold nicely instead.
1711 if (!is_store && value->is_Con() && index->is_Con()) {
1712 return false;
1713 }
1714
1715 Node* adr = array_element_address(value, index, T_CHAR);
1716 if (adr->is_top()) {
1717 return false;
1718 }
1719 if (is_store) {
1720 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1721 false, false, true /* mismatched */);
1722 } else {
1723 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1724 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1725 set_result(ch);
1726 }
1727 return true;
1728 }
1729
1730 //--------------------------round_double_node--------------------------------
1731 // Round a double node if necessary.
1732 Node* LibraryCallKit::round_double_node(Node* n) {
1733 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1734 n = _gvn.transform(new RoundDoubleNode(0, n));
1735 return n;
1736 }
1737
1738 //------------------------------inline_math-----------------------------------
1739 // public static double Math.abs(double)
1740 // public static double Math.sqrt(double)
1741 // public static double Math.log(double)
1742 // public static double Math.log10(double)
1743 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1744 Node* arg = round_double_node(argument(0));
1745 Node* n = NULL;
1746 switch (id) {
1747 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1748 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1749 default: fatal_unexpected_iid(id); break;
1750 }
1751 set_result(_gvn.transform(n));
1752 return true;
1753 }
1754
1755 //------------------------------runtime_math-----------------------------
1756 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1757 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1758 "must be (DD)D or (D)D type");
1759
1760 // Inputs
1761 Node* a = round_double_node(argument(0));
1762 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1763
1764 const TypePtr* no_memory_effects = NULL;
1765 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1766 no_memory_effects,
1767 a, top(), b, b ? top() : NULL);
1768 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1769 #ifdef ASSERT
1770 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1771 assert(value_top == top(), "second value must be top");
1772 #endif
1773
1774 set_result(value);
1775 return true;
1776 }
1777
1778 //------------------------------inline_math_native-----------------------------
1779 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1780 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1781 switch (id) {
1782 // These intrinsics are not properly supported on all hardware
1783 case vmIntrinsics::_dsin:
1784 return StubRoutines::dsin() != NULL ?
1785 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1786 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1787 case vmIntrinsics::_dcos:
1788 return StubRoutines::dcos() != NULL ?
1789 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1790 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1791 case vmIntrinsics::_dtan:
1792 return StubRoutines::dtan() != NULL ?
1793 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1794 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1795 case vmIntrinsics::_dlog:
1796 return StubRoutines::dlog() != NULL ?
1797 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1798 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1799 case vmIntrinsics::_dlog10:
1800 return StubRoutines::dlog10() != NULL ?
1801 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1802 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1803
1804 // These intrinsics are supported on all hardware
1805 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1806 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1807
1808 case vmIntrinsics::_dexp:
1809 return StubRoutines::dexp() != NULL ?
1810 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1811 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1812 case vmIntrinsics::_dpow: {
1813 Node* exp = round_double_node(argument(2));
1814 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1815 if (d != NULL && d->getd() == 2.0) {
1816 // Special case: pow(x, 2.0) => x * x
1817 Node* base = round_double_node(argument(0));
1818 set_result(_gvn.transform(new MulDNode(base, base)));
1819 return true;
1820 }
1821 return StubRoutines::dexp() != NULL ?
1822 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1823 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1824 }
1825 #undef FN_PTR
1826
1827 // These intrinsics are not yet correctly implemented
1828 case vmIntrinsics::_datan2:
1829 return false;
1830
1831 default:
1832 fatal_unexpected_iid(id);
1833 return false;
1834 }
1835 }
1836
1837 static bool is_simple_name(Node* n) {
1838 return (n->req() == 1 // constant
1839 || (n->is_Type() && n->as_Type()->type()->singleton())
1840 || n->is_Proj() // parameter or return value
1841 || n->is_Phi() // local of some sort
1842 );
1843 }
1844
1845 //----------------------------inline_notify-----------------------------------*
1846 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1847 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1848 address func;
1849 if (id == vmIntrinsics::_notify) {
1850 func = OptoRuntime::monitor_notify_Java();
1851 } else {
1852 func = OptoRuntime::monitor_notifyAll_Java();
1853 }
1854 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1855 make_slow_call_ex(call, env()->Throwable_klass(), false);
1856 return true;
1857 }
1858
1859
1860 //----------------------------inline_min_max-----------------------------------
1861 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1862 set_result(generate_min_max(id, argument(0), argument(1)));
1863 return true;
1864 }
1865
1866 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1867 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1868 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1869 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1870 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1871
1872 {
1873 PreserveJVMState pjvms(this);
1874 PreserveReexecuteState preexecs(this);
1875 jvms()->set_should_reexecute(true);
1876
1877 set_control(slow_path);
1878 set_i_o(i_o());
1879
1880 uncommon_trap(Deoptimization::Reason_intrinsic,
1881 Deoptimization::Action_none);
1882 }
1883
1884 set_control(fast_path);
1885 set_result(math);
1886 }
1887
1888 template <typename OverflowOp>
1889 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1890 typedef typename OverflowOp::MathOp MathOp;
1891
1892 MathOp* mathOp = new MathOp(arg1, arg2);
1893 Node* operation = _gvn.transform( mathOp );
1894 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1895 inline_math_mathExact(operation, ofcheck);
1896 return true;
1897 }
1898
1899 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1900 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1901 }
1902
1903 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1904 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1905 }
1906
1907 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1908 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1909 }
1910
1911 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1912 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1913 }
1914
1915 bool LibraryCallKit::inline_math_negateExactI() {
1916 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1917 }
1918
1919 bool LibraryCallKit::inline_math_negateExactL() {
1920 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1921 }
1922
1923 bool LibraryCallKit::inline_math_multiplyExactI() {
1924 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1925 }
1926
1927 bool LibraryCallKit::inline_math_multiplyExactL() {
1928 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1929 }
1930
1931 bool LibraryCallKit::inline_math_multiplyHigh() {
1932 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1933 return true;
1934 }
1935
1936 Node*
1937 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1938 // These are the candidate return value:
1939 Node* xvalue = x0;
1940 Node* yvalue = y0;
1941
1942 if (xvalue == yvalue) {
1943 return xvalue;
1944 }
1945
1946 bool want_max = (id == vmIntrinsics::_max);
1947
1948 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1949 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1950 if (txvalue == NULL || tyvalue == NULL) return top();
1951 // This is not really necessary, but it is consistent with a
1952 // hypothetical MaxINode::Value method:
1953 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1954
1955 // %%% This folding logic should (ideally) be in a different place.
1956 // Some should be inside IfNode, and there to be a more reliable
1957 // transformation of ?: style patterns into cmoves. We also want
1958 // more powerful optimizations around cmove and min/max.
1959
1960 // Try to find a dominating comparison of these guys.
1961 // It can simplify the index computation for Arrays.copyOf
1962 // and similar uses of System.arraycopy.
1963 // First, compute the normalized version of CmpI(x, y).
1964 int cmp_op = Op_CmpI;
1965 Node* xkey = xvalue;
1966 Node* ykey = yvalue;
1967 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1968 if (ideal_cmpxy->is_Cmp()) {
1969 // E.g., if we have CmpI(length - offset, count),
1970 // it might idealize to CmpI(length, count + offset)
1971 cmp_op = ideal_cmpxy->Opcode();
1972 xkey = ideal_cmpxy->in(1);
1973 ykey = ideal_cmpxy->in(2);
1974 }
1975
1976 // Start by locating any relevant comparisons.
1977 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1978 Node* cmpxy = NULL;
1979 Node* cmpyx = NULL;
1980 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1981 Node* cmp = start_from->fast_out(k);
1982 if (cmp->outcnt() > 0 && // must have prior uses
1983 cmp->in(0) == NULL && // must be context-independent
1984 cmp->Opcode() == cmp_op) { // right kind of compare
1985 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1986 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1987 }
1988 }
1989
1990 const int NCMPS = 2;
1991 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1992 int cmpn;
1993 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1994 if (cmps[cmpn] != NULL) break; // find a result
1995 }
1996 if (cmpn < NCMPS) {
1997 // Look for a dominating test that tells us the min and max.
1998 int depth = 0; // Limit search depth for speed
1999 Node* dom = control();
2000 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2001 if (++depth >= 100) break;
2002 Node* ifproj = dom;
2003 if (!ifproj->is_Proj()) continue;
2004 Node* iff = ifproj->in(0);
2005 if (!iff->is_If()) continue;
2006 Node* bol = iff->in(1);
2007 if (!bol->is_Bool()) continue;
2008 Node* cmp = bol->in(1);
2009 if (cmp == NULL) continue;
2010 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2011 if (cmps[cmpn] == cmp) break;
2012 if (cmpn == NCMPS) continue;
2013 BoolTest::mask btest = bol->as_Bool()->_test._test;
2014 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2015 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2016 // At this point, we know that 'x btest y' is true.
2017 switch (btest) {
2018 case BoolTest::eq:
2019 // They are proven equal, so we can collapse the min/max.
2020 // Either value is the answer. Choose the simpler.
2021 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2022 return yvalue;
2023 return xvalue;
2024 case BoolTest::lt: // x < y
2025 case BoolTest::le: // x <= y
2026 return (want_max ? yvalue : xvalue);
2027 case BoolTest::gt: // x > y
2028 case BoolTest::ge: // x >= y
2029 return (want_max ? xvalue : yvalue);
2030 default:
2031 break;
2032 }
2033 }
2034 }
2035
2036 // We failed to find a dominating test.
2037 // Let's pick a test that might GVN with prior tests.
2038 Node* best_bol = NULL;
2039 BoolTest::mask best_btest = BoolTest::illegal;
2040 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2041 Node* cmp = cmps[cmpn];
2042 if (cmp == NULL) continue;
2043 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2044 Node* bol = cmp->fast_out(j);
2045 if (!bol->is_Bool()) continue;
2046 BoolTest::mask btest = bol->as_Bool()->_test._test;
2047 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2048 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2049 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2050 best_bol = bol->as_Bool();
2051 best_btest = btest;
2052 }
2053 }
2054 }
2055
2056 Node* answer_if_true = NULL;
2057 Node* answer_if_false = NULL;
2058 switch (best_btest) {
2059 default:
2060 if (cmpxy == NULL)
2061 cmpxy = ideal_cmpxy;
2062 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2063 // and fall through:
2064 case BoolTest::lt: // x < y
2065 case BoolTest::le: // x <= y
2066 answer_if_true = (want_max ? yvalue : xvalue);
2067 answer_if_false = (want_max ? xvalue : yvalue);
2068 break;
2069 case BoolTest::gt: // x > y
2070 case BoolTest::ge: // x >= y
2071 answer_if_true = (want_max ? xvalue : yvalue);
2072 answer_if_false = (want_max ? yvalue : xvalue);
2073 break;
2074 }
2075
2076 jint hi, lo;
2077 if (want_max) {
2078 // We can sharpen the minimum.
2079 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2080 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2081 } else {
2082 // We can sharpen the maximum.
2083 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2084 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2085 }
2086
2087 // Use a flow-free graph structure, to avoid creating excess control edges
2088 // which could hinder other optimizations.
2089 // Since Math.min/max is often used with arraycopy, we want
2090 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2091 Node* cmov = CMoveNode::make(NULL, best_bol,
2092 answer_if_false, answer_if_true,
2093 TypeInt::make(lo, hi, widen));
2094
2095 return _gvn.transform(cmov);
2096
2097 /*
2098 // This is not as desirable as it may seem, since Min and Max
2099 // nodes do not have a full set of optimizations.
2100 // And they would interfere, anyway, with 'if' optimizations
2101 // and with CMoveI canonical forms.
2102 switch (id) {
2103 case vmIntrinsics::_min:
2104 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2105 case vmIntrinsics::_max:
2106 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2107 default:
2108 ShouldNotReachHere();
2109 }
2110 */
2111 }
2112
2113 inline int
2114 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2115 const TypePtr* base_type = TypePtr::NULL_PTR;
2116 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2117 if (base_type == NULL) {
2118 // Unknown type.
2119 return Type::AnyPtr;
2120 } else if (base_type == TypePtr::NULL_PTR) {
2121 // Since this is a NULL+long form, we have to switch to a rawptr.
2122 base = _gvn.transform(new CastX2PNode(offset));
2123 offset = MakeConX(0);
2124 return Type::RawPtr;
2125 } else if (base_type->base() == Type::RawPtr) {
2126 return Type::RawPtr;
2127 } else if (base_type->isa_oopptr()) {
2128 // Base is never null => always a heap address.
2129 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2130 return Type::OopPtr;
2131 }
2132 // Offset is small => always a heap address.
2133 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2134 if (offset_type != NULL &&
2135 base_type->offset() == 0 && // (should always be?)
2136 offset_type->_lo >= 0 &&
2137 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2138 return Type::OopPtr;
2139 } else if (type == T_OBJECT) {
2140 // off heap access to an oop doesn't make any sense. Has to be on
2141 // heap.
2142 return Type::OopPtr;
2143 }
2144 // Otherwise, it might either be oop+off or NULL+addr.
2145 return Type::AnyPtr;
2146 } else {
2147 // No information:
2148 return Type::AnyPtr;
2149 }
2150 }
2151
2152 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2153 Node* uncasted_base = base;
2154 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2155 if (kind == Type::RawPtr) {
2156 return basic_plus_adr(top(), uncasted_base, offset);
2157 } else if (kind == Type::AnyPtr) {
2158 assert(base == uncasted_base, "unexpected base change");
2159 if (can_cast) {
2160 if (!_gvn.type(base)->speculative_maybe_null() &&
2161 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2162 // According to profiling, this access is always on
2163 // heap. Casting the base to not null and thus avoiding membars
2164 // around the access should allow better optimizations
2165 Node* null_ctl = top();
2166 base = null_check_oop(base, &null_ctl, true, true, true);
2167 assert(null_ctl->is_top(), "no null control here");
2168 return basic_plus_adr(base, offset);
2169 } else if (_gvn.type(base)->speculative_always_null() &&
2170 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2171 // According to profiling, this access is always off
2172 // heap.
2173 base = null_assert(base);
2174 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2175 offset = MakeConX(0);
2176 return basic_plus_adr(top(), raw_base, offset);
2177 }
2178 }
2179 // We don't know if it's an on heap or off heap access. Fall back
2180 // to raw memory access.
2181 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2182 return basic_plus_adr(top(), raw, offset);
2183 } else {
2184 assert(base == uncasted_base, "unexpected base change");
2185 // We know it's an on heap access so base can't be null
2186 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2187 base = must_be_not_null(base, true);
2188 }
2189 return basic_plus_adr(base, offset);
2190 }
2191 }
2192
2193 //--------------------------inline_number_methods-----------------------------
2194 // inline int Integer.numberOfLeadingZeros(int)
2195 // inline int Long.numberOfLeadingZeros(long)
2196 //
2197 // inline int Integer.numberOfTrailingZeros(int)
2198 // inline int Long.numberOfTrailingZeros(long)
2199 //
2200 // inline int Integer.bitCount(int)
2201 // inline int Long.bitCount(long)
2202 //
2203 // inline char Character.reverseBytes(char)
2204 // inline short Short.reverseBytes(short)
2205 // inline int Integer.reverseBytes(int)
2206 // inline long Long.reverseBytes(long)
2207 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2208 Node* arg = argument(0);
2209 Node* n = NULL;
2210 switch (id) {
2211 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2212 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2213 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2214 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2215 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2216 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2217 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2218 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2219 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2220 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2221 default: fatal_unexpected_iid(id); break;
2222 }
2223 set_result(_gvn.transform(n));
2224 return true;
2225 }
2226
2227 //----------------------------inline_unsafe_access----------------------------
2228
2229 // Helper that guards and inserts a pre-barrier.
2230 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2231 Node* pre_val, bool need_mem_bar) {
2232 // We could be accessing the referent field of a reference object. If so, when G1
2233 // is enabled, we need to log the value in the referent field in an SATB buffer.
2234 // This routine performs some compile time filters and generates suitable
2235 // runtime filters that guard the pre-barrier code.
2236 // Also add memory barrier for non volatile load from the referent field
2237 // to prevent commoning of loads across safepoint.
2238 if (!UseG1GC && !need_mem_bar)
2239 return;
2240
2241 // Some compile time checks.
2242
2243 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2244 const TypeX* otype = offset->find_intptr_t_type();
2245 if (otype != NULL && otype->is_con() &&
2246 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2247 // Constant offset but not the reference_offset so just return
2248 return;
2249 }
2250
2251 // We only need to generate the runtime guards for instances.
2252 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2253 if (btype != NULL) {
2254 if (btype->isa_aryptr()) {
2255 // Array type so nothing to do
2256 return;
2257 }
2258
2259 const TypeInstPtr* itype = btype->isa_instptr();
2260 if (itype != NULL) {
2261 // Can the klass of base_oop be statically determined to be
2262 // _not_ a sub-class of Reference and _not_ Object?
2263 ciKlass* klass = itype->klass();
2264 if ( klass->is_loaded() &&
2265 !klass->is_subtype_of(env()->Reference_klass()) &&
2266 !env()->Object_klass()->is_subtype_of(klass)) {
2267 return;
2268 }
2269 }
2270 }
2271
2272 // The compile time filters did not reject base_oop/offset so
2273 // we need to generate the following runtime filters
2274 //
2275 // if (offset == java_lang_ref_Reference::_reference_offset) {
2276 // if (instance_of(base, java.lang.ref.Reference)) {
2277 // pre_barrier(_, pre_val, ...);
2278 // }
2279 // }
2280
2281 float likely = PROB_LIKELY( 0.999);
2282 float unlikely = PROB_UNLIKELY(0.999);
2283
2284 IdealKit ideal(this);
2285 #define __ ideal.
2286
2287 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2288
2289 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2290 // Update graphKit memory and control from IdealKit.
2291 sync_kit(ideal);
2292
2293 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2294 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2295
2296 // Update IdealKit memory and control from graphKit.
2297 __ sync_kit(this);
2298
2299 Node* one = __ ConI(1);
2300 // is_instof == 0 if base_oop == NULL
2301 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2302
2303 // Update graphKit from IdeakKit.
2304 sync_kit(ideal);
2305
2306 // Use the pre-barrier to record the value in the referent field
2307 pre_barrier(false /* do_load */,
2308 __ ctrl(),
2309 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2310 pre_val /* pre_val */,
2311 T_OBJECT);
2312 if (need_mem_bar) {
2313 // Add memory barrier to prevent commoning reads from this field
2314 // across safepoint since GC can change its value.
2315 insert_mem_bar(Op_MemBarCPUOrder);
2316 }
2317 // Update IdealKit from graphKit.
2318 __ sync_kit(this);
2319
2320 } __ end_if(); // _ref_type != ref_none
2321 } __ end_if(); // offset == referent_offset
2322
2323 // Final sync IdealKit and GraphKit.
2324 final_sync(ideal);
2325 #undef __
2326 }
2327
2328
2329 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2330 // Attempt to infer a sharper value type from the offset and base type.
2331 ciKlass* sharpened_klass = NULL;
2332
2333 // See if it is an instance field, with an object type.
2334 if (alias_type->field() != NULL) {
2335 if (alias_type->field()->type()->is_klass()) {
2336 sharpened_klass = alias_type->field()->type()->as_klass();
2337 }
2338 }
2339
2340 // See if it is a narrow oop array.
2341 if (adr_type->isa_aryptr()) {
2342 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2343 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2344 if (elem_type != NULL) {
2345 sharpened_klass = elem_type->klass();
2346 }
2347 }
2348 }
2349
2350 // The sharpened class might be unloaded if there is no class loader
2351 // contraint in place.
2352 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2353 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2354
2355 #ifndef PRODUCT
2356 if (C->print_intrinsics() || C->print_inlining()) {
2357 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2358 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2359 }
2360 #endif
2361 // Sharpen the value type.
2362 return tjp;
2363 }
2364 return NULL;
2365 }
2366
2367 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2368 if (callee()->is_static()) return false; // caller must have the capability!
2369 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2370 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2371 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2372
2373 #ifndef PRODUCT
2374 {
2375 ResourceMark rm;
2376 // Check the signatures.
2377 ciSignature* sig = callee()->signature();
2378 #ifdef ASSERT
2379 if (!is_store) {
2380 // Object getObject(Object base, int/long offset), etc.
2381 BasicType rtype = sig->return_type()->basic_type();
2382 assert(rtype == type, "getter must return the expected value");
2383 assert(sig->count() == 2, "oop getter has 2 arguments");
2384 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2385 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2386 } else {
2387 // void putObject(Object base, int/long offset, Object x), etc.
2388 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2389 assert(sig->count() == 3, "oop putter has 3 arguments");
2390 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2391 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2392 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2393 assert(vtype == type, "putter must accept the expected value");
2394 }
2395 #endif // ASSERT
2396 }
2397 #endif //PRODUCT
2398
2399 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2400
2401 Node* receiver = argument(0); // type: oop
2402
2403 // Build address expression.
2404 Node* adr;
2405 Node* heap_base_oop = top();
2406 Node* offset = top();
2407 Node* val;
2408
2409 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2410 Node* base = argument(1); // type: oop
2411 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2412 offset = argument(2); // type: long
2413 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2414 // to be plain byte offsets, which are also the same as those accepted
2415 // by oopDesc::field_addr.
2416 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2417 "fieldOffset must be byte-scaled");
2418 // 32-bit machines ignore the high half!
2419 offset = ConvL2X(offset);
2420 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2421
2422 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2423 heap_base_oop = base;
2424 } else if (type == T_OBJECT) {
2425 return false; // off-heap oop accesses are not supported
2426 }
2427
2428 // Can base be NULL? Otherwise, always on-heap access.
2429 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2430
2431 val = is_store ? argument(4) : NULL;
2432
2433 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2434
2435 // Try to categorize the address.
2436 Compile::AliasType* alias_type = C->alias_type(adr_type);
2437 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2438
2439 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2440 alias_type->adr_type() == TypeAryPtr::RANGE) {
2441 return false; // not supported
2442 }
2443
2444 bool mismatched = false;
2445 BasicType bt = alias_type->basic_type();
2446 if (bt != T_ILLEGAL) {
2447 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2448 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2449 // Alias type doesn't differentiate between byte[] and boolean[]).
2450 // Use address type to get the element type.
2451 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2452 }
2453 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2454 // accessing an array field with getObject is not a mismatch
2455 bt = T_OBJECT;
2456 }
2457 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2458 // Don't intrinsify mismatched object accesses
2459 return false;
2460 }
2461 mismatched = (bt != type);
2462 } else if (alias_type->adr_type()->isa_oopptr()) {
2463 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2464 }
2465
2466 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2467
2468 // First guess at the value type.
2469 const Type *value_type = Type::get_const_basic_type(type);
2470
2471 // We will need memory barriers unless we can determine a unique
2472 // alias category for this reference. (Note: If for some reason
2473 // the barriers get omitted and the unsafe reference begins to "pollute"
2474 // the alias analysis of the rest of the graph, either Compile::can_alias
2475 // or Compile::must_alias will throw a diagnostic assert.)
2476 bool need_mem_bar = false;
2477 switch (kind) {
2478 case Relaxed:
2479 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2480 break;
2481 case Opaque:
2482 // Opaque uses CPUOrder membars for protection against code movement.
2483 case Acquire:
2484 case Release:
2485 case Volatile:
2486 need_mem_bar = true;
2487 break;
2488 default:
2489 ShouldNotReachHere();
2490 }
2491
2492 // Some accesses require access atomicity for all types, notably longs and doubles.
2493 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2494 bool requires_atomic_access = false;
2495 switch (kind) {
2496 case Relaxed:
2497 requires_atomic_access = AlwaysAtomicAccesses;
2498 break;
2499 case Opaque:
2500 // Opaque accesses are atomic.
2501 case Acquire:
2502 case Release:
2503 case Volatile:
2504 requires_atomic_access = true;
2505 break;
2506 default:
2507 ShouldNotReachHere();
2508 }
2509
2510 // Figure out the memory ordering.
2511 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2512 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2513
2514 // If we are reading the value of the referent field of a Reference
2515 // object (either by using Unsafe directly or through reflection)
2516 // then, if G1 is enabled, we need to record the referent in an
2517 // SATB log buffer using the pre-barrier mechanism.
2518 // Also we need to add memory barrier to prevent commoning reads
2519 // from this field across safepoint since GC can change its value.
2520 bool need_read_barrier = !is_store &&
2521 offset != top() && heap_base_oop != top();
2522
2523 if (!is_store && type == T_OBJECT) {
2524 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2525 if (tjp != NULL) {
2526 value_type = tjp;
2527 }
2528 }
2529
2530 receiver = null_check(receiver);
2531 if (stopped()) {
2532 return true;
2533 }
2534 // Heap pointers get a null-check from the interpreter,
2535 // as a courtesy. However, this is not guaranteed by Unsafe,
2536 // and it is not possible to fully distinguish unintended nulls
2537 // from intended ones in this API.
2538
2539 // We need to emit leading and trailing CPU membars (see below) in
2540 // addition to memory membars for special access modes. This is a little
2541 // too strong, but avoids the need to insert per-alias-type
2542 // volatile membars (for stores; compare Parse::do_put_xxx), which
2543 // we cannot do effectively here because we probably only have a
2544 // rough approximation of type.
2545
2546 switch(kind) {
2547 case Relaxed:
2548 case Opaque:
2549 case Acquire:
2550 break;
2551 case Release:
2552 case Volatile:
2553 if (is_store) {
2554 insert_mem_bar(Op_MemBarRelease);
2555 } else {
2556 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2557 insert_mem_bar(Op_MemBarVolatile);
2558 }
2559 }
2560 break;
2561 default:
2562 ShouldNotReachHere();
2563 }
2564
2565 // Memory barrier to prevent normal and 'unsafe' accesses from
2566 // bypassing each other. Happens after null checks, so the
2567 // exception paths do not take memory state from the memory barrier,
2568 // so there's no problems making a strong assert about mixing users
2569 // of safe & unsafe memory.
2570 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2571
2572 if (!is_store) {
2573 Node* p = NULL;
2574 // Try to constant fold a load from a constant field
2575 ciField* field = alias_type->field();
2576 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2577 // final or stable field
2578 p = make_constant_from_field(field, heap_base_oop);
2579 }
2580 if (p == NULL) {
2581 // To be valid, unsafe loads may depend on other conditions than
2582 // the one that guards them: pin the Load node
2583 LoadNode::ControlDependency dep = LoadNode::Pinned;
2584 Node* ctrl = control();
2585 // non volatile loads may be able to float
2586 if (!need_mem_bar && adr_type->isa_instptr()) {
2587 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2588 intptr_t offset = Type::OffsetBot;
2589 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2590 if (offset >= 0) {
2591 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2592 if (offset < s) {
2593 // Guaranteed to be a valid access, no need to pin it
2594 dep = LoadNode::DependsOnlyOnTest;
2595 ctrl = NULL;
2596 }
2597 }
2598 }
2599 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2600 // load value
2601 switch (type) {
2602 case T_BOOLEAN:
2603 {
2604 // Normalize the value returned by getBoolean in the following cases
2605 if (mismatched ||
2606 heap_base_oop == top() || // - heap_base_oop is NULL or
2607 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2608 // and the unsafe access is made to large offset
2609 // (i.e., larger than the maximum offset necessary for any
2610 // field access)
2611 ) {
2612 IdealKit ideal = IdealKit(this);
2613 #define __ ideal.
2614 IdealVariable normalized_result(ideal);
2615 __ declarations_done();
2616 __ set(normalized_result, p);
2617 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2618 __ set(normalized_result, ideal.ConI(1));
2619 ideal.end_if();
2620 final_sync(ideal);
2621 p = __ value(normalized_result);
2622 #undef __
2623 }
2624 }
2625 case T_CHAR:
2626 case T_BYTE:
2627 case T_SHORT:
2628 case T_INT:
2629 case T_LONG:
2630 case T_FLOAT:
2631 case T_DOUBLE:
2632 break;
2633 case T_OBJECT:
2634 if (need_read_barrier) {
2635 // We do not require a mem bar inside pre_barrier if need_mem_bar
2636 // is set: the barriers would be emitted by us.
2637 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2638 }
2639 break;
2640 case T_ADDRESS:
2641 // Cast to an int type.
2642 p = _gvn.transform(new CastP2XNode(NULL, p));
2643 p = ConvX2UL(p);
2644 break;
2645 default:
2646 fatal("unexpected type %d: %s", type, type2name(type));
2647 break;
2648 }
2649 }
2650 // The load node has the control of the preceding MemBarCPUOrder. All
2651 // following nodes will have the control of the MemBarCPUOrder inserted at
2652 // the end of this method. So, pushing the load onto the stack at a later
2653 // point is fine.
2654 set_result(p);
2655 } else {
2656 // place effect of store into memory
2657 switch (type) {
2658 case T_DOUBLE:
2659 val = dstore_rounding(val);
2660 break;
2661 case T_ADDRESS:
2662 // Repackage the long as a pointer.
2663 val = ConvL2X(val);
2664 val = _gvn.transform(new CastX2PNode(val));
2665 break;
2666 default:
2667 break;
2668 }
2669
2670 if (type == T_OBJECT) {
2671 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2672 } else {
2673 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2674 }
2675 }
2676
2677 switch(kind) {
2678 case Relaxed:
2679 case Opaque:
2680 case Release:
2681 break;
2682 case Acquire:
2683 case Volatile:
2684 if (!is_store) {
2685 insert_mem_bar(Op_MemBarAcquire);
2686 } else {
2687 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2688 insert_mem_bar(Op_MemBarVolatile);
2689 }
2690 }
2691 break;
2692 default:
2693 ShouldNotReachHere();
2694 }
2695
2696 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2697
2698 return true;
2699 }
2700
2701 //----------------------------inline_unsafe_load_store----------------------------
2702 // This method serves a couple of different customers (depending on LoadStoreKind):
2703 //
2704 // LS_cmp_swap:
2705 //
2706 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2707 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2708 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2709 //
2710 // LS_cmp_swap_weak:
2711 //
2712 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2713 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2714 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2715 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2716 //
2717 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2718 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2719 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2720 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2721 //
2722 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2723 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2724 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2725 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2726 //
2727 // LS_cmp_exchange:
2728 //
2729 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2730 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2731 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2732 //
2733 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2734 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2735 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2736 //
2737 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2738 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2739 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2740 //
2741 // LS_get_add:
2742 //
2743 // int getAndAddInt( Object o, long offset, int delta)
2744 // long getAndAddLong(Object o, long offset, long delta)
2745 //
2746 // LS_get_set:
2747 //
2748 // int getAndSet(Object o, long offset, int newValue)
2749 // long getAndSet(Object o, long offset, long newValue)
2750 // Object getAndSet(Object o, long offset, Object newValue)
2751 //
2752 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2753 // This basic scheme here is the same as inline_unsafe_access, but
2754 // differs in enough details that combining them would make the code
2755 // overly confusing. (This is a true fact! I originally combined
2756 // them, but even I was confused by it!) As much code/comments as
2757 // possible are retained from inline_unsafe_access though to make
2758 // the correspondences clearer. - dl
2759
2760 if (callee()->is_static()) return false; // caller must have the capability!
2761
2762 #ifndef PRODUCT
2763 BasicType rtype;
2764 {
2765 ResourceMark rm;
2766 // Check the signatures.
2767 ciSignature* sig = callee()->signature();
2768 rtype = sig->return_type()->basic_type();
2769 switch(kind) {
2770 case LS_get_add:
2771 case LS_get_set: {
2772 // Check the signatures.
2773 #ifdef ASSERT
2774 assert(rtype == type, "get and set must return the expected type");
2775 assert(sig->count() == 3, "get and set has 3 arguments");
2776 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2777 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2778 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2779 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2780 #endif // ASSERT
2781 break;
2782 }
2783 case LS_cmp_swap:
2784 case LS_cmp_swap_weak: {
2785 // Check the signatures.
2786 #ifdef ASSERT
2787 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2788 assert(sig->count() == 4, "CAS has 4 arguments");
2789 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2790 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2791 #endif // ASSERT
2792 break;
2793 }
2794 case LS_cmp_exchange: {
2795 // Check the signatures.
2796 #ifdef ASSERT
2797 assert(rtype == type, "CAS must return the expected type");
2798 assert(sig->count() == 4, "CAS has 4 arguments");
2799 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2800 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2801 #endif // ASSERT
2802 break;
2803 }
2804 default:
2805 ShouldNotReachHere();
2806 }
2807 }
2808 #endif //PRODUCT
2809
2810 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2811
2812 // Get arguments:
2813 Node* receiver = NULL;
2814 Node* base = NULL;
2815 Node* offset = NULL;
2816 Node* oldval = NULL;
2817 Node* newval = NULL;
2818 switch(kind) {
2819 case LS_cmp_swap:
2820 case LS_cmp_swap_weak:
2821 case LS_cmp_exchange: {
2822 const bool two_slot_type = type2size[type] == 2;
2823 receiver = argument(0); // type: oop
2824 base = argument(1); // type: oop
2825 offset = argument(2); // type: long
2826 oldval = argument(4); // type: oop, int, or long
2827 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2828 break;
2829 }
2830 case LS_get_add:
2831 case LS_get_set: {
2832 receiver = argument(0); // type: oop
2833 base = argument(1); // type: oop
2834 offset = argument(2); // type: long
2835 oldval = NULL;
2836 newval = argument(4); // type: oop, int, or long
2837 break;
2838 }
2839 default:
2840 ShouldNotReachHere();
2841 }
2842
2843 // Build field offset expression.
2844 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2845 // to be plain byte offsets, which are also the same as those accepted
2846 // by oopDesc::field_addr.
2847 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2848 // 32-bit machines ignore the high half of long offsets
2849 offset = ConvL2X(offset);
2850 Node* adr = make_unsafe_address(base, offset, type, false);
2851 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2852
2853 Compile::AliasType* alias_type = C->alias_type(adr_type);
2854 BasicType bt = alias_type->basic_type();
2855 if (bt != T_ILLEGAL &&
2856 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2857 // Don't intrinsify mismatched object accesses.
2858 return false;
2859 }
2860
2861 // For CAS, unlike inline_unsafe_access, there seems no point in
2862 // trying to refine types. Just use the coarse types here.
2863 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2864 const Type *value_type = Type::get_const_basic_type(type);
2865
2866 switch (kind) {
2867 case LS_get_set:
2868 case LS_cmp_exchange: {
2869 if (type == T_OBJECT) {
2870 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2871 if (tjp != NULL) {
2872 value_type = tjp;
2873 }
2874 }
2875 break;
2876 }
2877 case LS_cmp_swap:
2878 case LS_cmp_swap_weak:
2879 case LS_get_add:
2880 break;
2881 default:
2882 ShouldNotReachHere();
2883 }
2884
2885 // Null check receiver.
2886 receiver = null_check(receiver);
2887 if (stopped()) {
2888 return true;
2889 }
2890
2891 int alias_idx = C->get_alias_index(adr_type);
2892
2893 // Memory-model-wise, a LoadStore acts like a little synchronized
2894 // block, so needs barriers on each side. These don't translate
2895 // into actual barriers on most machines, but we still need rest of
2896 // compiler to respect ordering.
2897
2898 switch (access_kind) {
2899 case Relaxed:
2900 case Acquire:
2901 break;
2902 case Release:
2903 insert_mem_bar(Op_MemBarRelease);
2904 break;
2905 case Volatile:
2906 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2907 insert_mem_bar(Op_MemBarVolatile);
2908 } else {
2909 insert_mem_bar(Op_MemBarRelease);
2910 }
2911 break;
2912 default:
2913 ShouldNotReachHere();
2914 }
2915 insert_mem_bar(Op_MemBarCPUOrder);
2916
2917 // Figure out the memory ordering.
2918 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2919
2920 // 4984716: MemBars must be inserted before this
2921 // memory node in order to avoid a false
2922 // dependency which will confuse the scheduler.
2923 Node *mem = memory(alias_idx);
2924
2925 // For now, we handle only those cases that actually exist: ints,
2926 // longs, and Object. Adding others should be straightforward.
2927 Node* load_store = NULL;
2928 switch(type) {
2929 case T_BYTE:
2930 switch(kind) {
2931 case LS_get_add:
2932 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2933 break;
2934 case LS_get_set:
2935 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2936 break;
2937 case LS_cmp_swap_weak:
2938 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2939 break;
2940 case LS_cmp_swap:
2941 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2942 break;
2943 case LS_cmp_exchange:
2944 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2945 break;
2946 default:
2947 ShouldNotReachHere();
2948 }
2949 break;
2950 case T_SHORT:
2951 switch(kind) {
2952 case LS_get_add:
2953 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2954 break;
2955 case LS_get_set:
2956 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2957 break;
2958 case LS_cmp_swap_weak:
2959 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2960 break;
2961 case LS_cmp_swap:
2962 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2963 break;
2964 case LS_cmp_exchange:
2965 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2966 break;
2967 default:
2968 ShouldNotReachHere();
2969 }
2970 break;
2971 case T_INT:
2972 switch(kind) {
2973 case LS_get_add:
2974 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2975 break;
2976 case LS_get_set:
2977 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2978 break;
2979 case LS_cmp_swap_weak:
2980 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2981 break;
2982 case LS_cmp_swap:
2983 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2984 break;
2985 case LS_cmp_exchange:
2986 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2987 break;
2988 default:
2989 ShouldNotReachHere();
2990 }
2991 break;
2992 case T_LONG:
2993 switch(kind) {
2994 case LS_get_add:
2995 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2996 break;
2997 case LS_get_set:
2998 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2999 break;
3000 case LS_cmp_swap_weak:
3001 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3002 break;
3003 case LS_cmp_swap:
3004 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3005 break;
3006 case LS_cmp_exchange:
3007 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
3008 break;
3009 default:
3010 ShouldNotReachHere();
3011 }
3012 break;
3013 case T_OBJECT:
3014 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3015 // could be delayed during Parse (for example, in adjust_map_after_if()).
3016 // Execute transformation here to avoid barrier generation in such case.
3017 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3018 newval = _gvn.makecon(TypePtr::NULL_PTR);
3019
3020 // Reference stores need a store barrier.
3021 switch(kind) {
3022 case LS_get_set: {
3023 // If pre-barrier must execute before the oop store, old value will require do_load here.
3024 if (!can_move_pre_barrier()) {
3025 pre_barrier(true /* do_load*/,
3026 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3027 NULL /* pre_val*/,
3028 T_OBJECT);
3029 } // Else move pre_barrier to use load_store value, see below.
3030 break;
3031 }
3032 case LS_cmp_swap_weak:
3033 case LS_cmp_swap:
3034 case LS_cmp_exchange: {
3035 // Same as for newval above:
3036 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3037 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3038 }
3039 // The only known value which might get overwritten is oldval.
3040 pre_barrier(false /* do_load */,
3041 control(), NULL, NULL, max_juint, NULL, NULL,
3042 oldval /* pre_val */,
3043 T_OBJECT);
3044 break;
3045 }
3046 default:
3047 ShouldNotReachHere();
3048 }
3049
3050 #ifdef _LP64
3051 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3052 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3053
3054 switch(kind) {
3055 case LS_get_set:
3056 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3057 break;
3058 case LS_cmp_swap_weak: {
3059 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3060 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3061 break;
3062 }
3063 case LS_cmp_swap: {
3064 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3065 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3066 break;
3067 }
3068 case LS_cmp_exchange: {
3069 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3070 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3071 break;
3072 }
3073 default:
3074 ShouldNotReachHere();
3075 }
3076 } else
3077 #endif
3078 switch (kind) {
3079 case LS_get_set:
3080 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3081 break;
3082 case LS_cmp_swap_weak:
3083 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3084 break;
3085 case LS_cmp_swap:
3086 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3087 break;
3088 case LS_cmp_exchange:
3089 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3090 break;
3091 default:
3092 ShouldNotReachHere();
3093 }
3094
3095 // Emit the post barrier only when the actual store happened. This makes sense
3096 // to check only for LS_cmp_* that can fail to set the value.
3097 // LS_cmp_exchange does not produce any branches by default, so there is no
3098 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3099 // CompareAndExchange and move branches here, it would make sense to conditionalize
3100 // post_barriers for LS_cmp_exchange as well.
3101 //
3102 // CAS success path is marked more likely since we anticipate this is a performance
3103 // critical path, while CAS failure path can use the penalty for going through unlikely
3104 // path as backoff. Which is still better than doing a store barrier there.
3105 switch (kind) {
3106 case LS_get_set:
3107 case LS_cmp_exchange: {
3108 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3109 break;
3110 }
3111 case LS_cmp_swap_weak:
3112 case LS_cmp_swap: {
3113 IdealKit ideal(this);
3114 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3115 sync_kit(ideal);
3116 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3117 ideal.sync_kit(this);
3118 } ideal.end_if();
3119 final_sync(ideal);
3120 break;
3121 }
3122 default:
3123 ShouldNotReachHere();
3124 }
3125 break;
3126 default:
3127 fatal("unexpected type %d: %s", type, type2name(type));
3128 break;
3129 }
3130
3131 // SCMemProjNodes represent the memory state of a LoadStore. Their
3132 // main role is to prevent LoadStore nodes from being optimized away
3133 // when their results aren't used.
3134 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3135 set_memory(proj, alias_idx);
3136
3137 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3138 #ifdef _LP64
3139 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3140 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3141 }
3142 #endif
3143 if (can_move_pre_barrier() && kind == LS_get_set) {
3144 // Don't need to load pre_val. The old value is returned by load_store.
3145 // The pre_barrier can execute after the xchg as long as no safepoint
3146 // gets inserted between them.
3147 pre_barrier(false /* do_load */,
3148 control(), NULL, NULL, max_juint, NULL, NULL,
3149 load_store /* pre_val */,
3150 T_OBJECT);
3151 }
3152 }
3153
3154 // Add the trailing membar surrounding the access
3155 insert_mem_bar(Op_MemBarCPUOrder);
3156
3157 switch (access_kind) {
3158 case Relaxed:
3159 case Release:
3160 break; // do nothing
3161 case Acquire:
3162 case Volatile:
3163 insert_mem_bar(Op_MemBarAcquire);
3164 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3165 break;
3166 default:
3167 ShouldNotReachHere();
3168 }
3169
3170 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3171 set_result(load_store);
3172 return true;
3173 }
3174
3175 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3176 MemNode::MemOrd mo = MemNode::unset;
3177 switch(kind) {
3178 case Opaque:
3179 case Relaxed: mo = MemNode::unordered; break;
3180 case Acquire: mo = MemNode::acquire; break;
3181 case Release: mo = MemNode::release; break;
3182 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3183 default:
3184 ShouldNotReachHere();
3185 }
3186 guarantee(mo != MemNode::unset, "Should select memory ordering");
3187 return mo;
3188 }
3189
3190 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3191 MemNode::MemOrd mo = MemNode::unset;
3192 switch(kind) {
3193 case Opaque:
3194 case Relaxed: mo = MemNode::unordered; break;
3195 case Acquire: mo = MemNode::acquire; break;
3196 case Release: mo = MemNode::release; break;
3197 case Volatile: mo = MemNode::seqcst; break;
3198 default:
3199 ShouldNotReachHere();
3200 }
3201 guarantee(mo != MemNode::unset, "Should select memory ordering");
3202 return mo;
3203 }
3204
3205 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3206 // Regardless of form, don't allow previous ld/st to move down,
3207 // then issue acquire, release, or volatile mem_bar.
3208 insert_mem_bar(Op_MemBarCPUOrder);
3209 switch(id) {
3210 case vmIntrinsics::_loadFence:
3211 insert_mem_bar(Op_LoadFence);
3212 return true;
3213 case vmIntrinsics::_storeFence:
3214 insert_mem_bar(Op_StoreFence);
3215 return true;
3216 case vmIntrinsics::_fullFence:
3217 insert_mem_bar(Op_MemBarVolatile);
3218 return true;
3219 default:
3220 fatal_unexpected_iid(id);
3221 return false;
3222 }
3223 }
3224
3225 bool LibraryCallKit::inline_onspinwait() {
3226 insert_mem_bar(Op_OnSpinWait);
3227 return true;
3228 }
3229
3230 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3231 if (!kls->is_Con()) {
3232 return true;
3233 }
3234 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3235 if (klsptr == NULL) {
3236 return true;
3237 }
3238 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3239 // don't need a guard for a klass that is already initialized
3240 return !ik->is_initialized();
3241 }
3242
3243 //----------------------------inline_unsafe_allocate---------------------------
3244 // public native Object Unsafe.allocateInstance(Class<?> cls);
3245 bool LibraryCallKit::inline_unsafe_allocate() {
3246 if (callee()->is_static()) return false; // caller must have the capability!
3247
3248 null_check_receiver(); // null-check, then ignore
3249 Node* cls = null_check(argument(1));
3250 if (stopped()) return true;
3251
3252 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3253 kls = null_check(kls);
3254 if (stopped()) return true; // argument was like int.class
3255
3256 Node* test = NULL;
3257 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3258 // Note: The argument might still be an illegal value like
3259 // Serializable.class or Object[].class. The runtime will handle it.
3260 // But we must make an explicit check for initialization.
3261 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3262 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3263 // can generate code to load it as unsigned byte.
3264 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3265 Node* bits = intcon(InstanceKlass::fully_initialized);
3266 test = _gvn.transform(new SubINode(inst, bits));
3267 // The 'test' is non-zero if we need to take a slow path.
3268 }
3269
3270 Node* obj = new_instance(kls, test);
3271 set_result(obj);
3272 return true;
3273 }
3274
3275 //------------------------inline_native_time_funcs--------------
3276 // inline code for System.currentTimeMillis() and System.nanoTime()
3277 // these have the same type and signature
3278 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3279 const TypeFunc* tf = OptoRuntime::void_long_Type();
3280 const TypePtr* no_memory_effects = NULL;
3281 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3282 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3283 #ifdef ASSERT
3284 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3285 assert(value_top == top(), "second value must be top");
3286 #endif
3287 set_result(value);
3288 return true;
3289 }
3290
3291 #ifdef TRACE_HAVE_INTRINSICS
3292
3293 /*
3294 * oop -> myklass
3295 * myklass->trace_id |= USED
3296 * return myklass->trace_id & ~0x3
3297 */
3298 bool LibraryCallKit::inline_native_classID() {
3299 Node* cls = null_check(argument(0), T_OBJECT);
3300 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3301 kls = null_check(kls, T_OBJECT);
3302
3303 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3304 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3305 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3306
3307 Node* clsused = longcon(0x01l); // set the class bit
3308 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3309 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3310 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3311
3312 #ifdef TRACE_ID_META_BITS
3313 Node* mbits = longcon(~TRACE_ID_META_BITS);
3314 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3315 #endif
3316 #ifdef TRACE_ID_CLASS_SHIFT
3317 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3318 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3319 #endif
3320
3321 set_result(tvalue);
3322 return true;
3323
3324 }
3325
3326 bool LibraryCallKit::inline_native_getBufferWriter() {
3327 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3328
3329 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3330 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3331 );
3332
3333 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3334
3335 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3336 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3337
3338 IfNode* iff_jobj_null =
3339 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3340
3341 enum { _normal_path = 1,
3342 _null_path = 2,
3343 PATH_LIMIT };
3344
3345 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3346 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3347
3348 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3349 result_rgn->init_req(_null_path, jobj_is_null);
3350 result_val->init_req(_null_path, null());
3351
3352 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3353 result_rgn->init_req(_normal_path, jobj_is_not_null);
3354
3355 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3356 result_val->init_req(_normal_path, res);
3357
3358 set_result(result_rgn, result_val);
3359
3360 return true;
3361 }
3362
3363 #endif
3364
3365 //------------------------inline_native_currentThread------------------
3366 bool LibraryCallKit::inline_native_currentThread() {
3367 Node* junk = NULL;
3368 set_result(generate_current_thread(junk));
3369 return true;
3370 }
3371
3372 //------------------------inline_native_isInterrupted------------------
3373 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3374 bool LibraryCallKit::inline_native_isInterrupted() {
3375 // Add a fast path to t.isInterrupted(clear_int):
3376 // (t == Thread.current() &&
3377 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3378 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3379 // So, in the common case that the interrupt bit is false,
3380 // we avoid making a call into the VM. Even if the interrupt bit
3381 // is true, if the clear_int argument is false, we avoid the VM call.
3382 // However, if the receiver is not currentThread, we must call the VM,
3383 // because there must be some locking done around the operation.
3384
3385 // We only go to the fast case code if we pass two guards.
3386 // Paths which do not pass are accumulated in the slow_region.
3387
3388 enum {
3389 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3390 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3391 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3392 PATH_LIMIT
3393 };
3394
3395 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3396 // out of the function.
3397 insert_mem_bar(Op_MemBarCPUOrder);
3398
3399 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3400 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3401
3402 RegionNode* slow_region = new RegionNode(1);
3403 record_for_igvn(slow_region);
3404
3405 // (a) Receiving thread must be the current thread.
3406 Node* rec_thr = argument(0);
3407 Node* tls_ptr = NULL;
3408 Node* cur_thr = generate_current_thread(tls_ptr);
3409 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3410 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3411
3412 generate_slow_guard(bol_thr, slow_region);
3413
3414 // (b) Interrupt bit on TLS must be false.
3415 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3416 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3417 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3418
3419 // Set the control input on the field _interrupted read to prevent it floating up.
3420 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3421 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3422 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3423
3424 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3425
3426 // First fast path: if (!TLS._interrupted) return false;
3427 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3428 result_rgn->init_req(no_int_result_path, false_bit);
3429 result_val->init_req(no_int_result_path, intcon(0));
3430
3431 // drop through to next case
3432 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3433
3434 #ifndef _WINDOWS
3435 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3436 Node* clr_arg = argument(1);
3437 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3438 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3439 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3440
3441 // Second fast path: ... else if (!clear_int) return true;
3442 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3443 result_rgn->init_req(no_clear_result_path, false_arg);
3444 result_val->init_req(no_clear_result_path, intcon(1));
3445
3446 // drop through to next case
3447 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3448 #else
3449 // To return true on Windows you must read the _interrupted field
3450 // and check the event state i.e. take the slow path.
3451 #endif // _WINDOWS
3452
3453 // (d) Otherwise, go to the slow path.
3454 slow_region->add_req(control());
3455 set_control( _gvn.transform(slow_region));
3456
3457 if (stopped()) {
3458 // There is no slow path.
3459 result_rgn->init_req(slow_result_path, top());
3460 result_val->init_req(slow_result_path, top());
3461 } else {
3462 // non-virtual because it is a private non-static
3463 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3464
3465 Node* slow_val = set_results_for_java_call(slow_call);
3466 // this->control() comes from set_results_for_java_call
3467
3468 Node* fast_io = slow_call->in(TypeFunc::I_O);
3469 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3470
3471 // These two phis are pre-filled with copies of of the fast IO and Memory
3472 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3473 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3474
3475 result_rgn->init_req(slow_result_path, control());
3476 result_io ->init_req(slow_result_path, i_o());
3477 result_mem->init_req(slow_result_path, reset_memory());
3478 result_val->init_req(slow_result_path, slow_val);
3479
3480 set_all_memory(_gvn.transform(result_mem));
3481 set_i_o( _gvn.transform(result_io));
3482 }
3483
3484 C->set_has_split_ifs(true); // Has chance for split-if optimization
3485 set_result(result_rgn, result_val);
3486 return true;
3487 }
3488
3489 //---------------------------load_mirror_from_klass----------------------------
3490 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3491 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3492 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3493 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3494 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3495 }
3496
3497 //-----------------------load_klass_from_mirror_common-------------------------
3498 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3499 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3500 // and branch to the given path on the region.
3501 // If never_see_null, take an uncommon trap on null, so we can optimistically
3502 // compile for the non-null case.
3503 // If the region is NULL, force never_see_null = true.
3504 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3505 bool never_see_null,
3506 RegionNode* region,
3507 int null_path,
3508 int offset) {
3509 if (region == NULL) never_see_null = true;
3510 Node* p = basic_plus_adr(mirror, offset);
3511 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3512 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3513 Node* null_ctl = top();
3514 kls = null_check_oop(kls, &null_ctl, never_see_null);
3515 if (region != NULL) {
3516 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3517 region->init_req(null_path, null_ctl);
3518 } else {
3519 assert(null_ctl == top(), "no loose ends");
3520 }
3521 return kls;
3522 }
3523
3524 //--------------------(inline_native_Class_query helpers)---------------------
3525 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3526 // Fall through if (mods & mask) == bits, take the guard otherwise.
3527 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3528 // Branch around if the given klass has the given modifier bit set.
3529 // Like generate_guard, adds a new path onto the region.
3530 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3531 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3532 Node* mask = intcon(modifier_mask);
3533 Node* bits = intcon(modifier_bits);
3534 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3535 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3536 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3537 return generate_fair_guard(bol, region);
3538 }
3539 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3540 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3541 }
3542
3543 //-------------------------inline_native_Class_query-------------------
3544 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3545 const Type* return_type = TypeInt::BOOL;
3546 Node* prim_return_value = top(); // what happens if it's a primitive class?
3547 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3548 bool expect_prim = false; // most of these guys expect to work on refs
3549
3550 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3551
3552 Node* mirror = argument(0);
3553 Node* obj = top();
3554
3555 switch (id) {
3556 case vmIntrinsics::_isInstance:
3557 // nothing is an instance of a primitive type
3558 prim_return_value = intcon(0);
3559 obj = argument(1);
3560 break;
3561 case vmIntrinsics::_getModifiers:
3562 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3563 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3564 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3565 break;
3566 case vmIntrinsics::_isInterface:
3567 prim_return_value = intcon(0);
3568 break;
3569 case vmIntrinsics::_isArray:
3570 prim_return_value = intcon(0);
3571 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3572 break;
3573 case vmIntrinsics::_isPrimitive:
3574 prim_return_value = intcon(1);
3575 expect_prim = true; // obviously
3576 break;
3577 case vmIntrinsics::_getSuperclass:
3578 prim_return_value = null();
3579 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3580 break;
3581 case vmIntrinsics::_getClassAccessFlags:
3582 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3583 return_type = TypeInt::INT; // not bool! 6297094
3584 break;
3585 default:
3586 fatal_unexpected_iid(id);
3587 break;
3588 }
3589
3590 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3591 if (mirror_con == NULL) return false; // cannot happen?
3592
3593 #ifndef PRODUCT
3594 if (C->print_intrinsics() || C->print_inlining()) {
3595 ciType* k = mirror_con->java_mirror_type();
3596 if (k) {
3597 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3598 k->print_name();
3599 tty->cr();
3600 }
3601 }
3602 #endif
3603
3604 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3605 RegionNode* region = new RegionNode(PATH_LIMIT);
3606 record_for_igvn(region);
3607 PhiNode* phi = new PhiNode(region, return_type);
3608
3609 // The mirror will never be null of Reflection.getClassAccessFlags, however
3610 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3611 // if it is. See bug 4774291.
3612
3613 // For Reflection.getClassAccessFlags(), the null check occurs in
3614 // the wrong place; see inline_unsafe_access(), above, for a similar
3615 // situation.
3616 mirror = null_check(mirror);
3617 // If mirror or obj is dead, only null-path is taken.
3618 if (stopped()) return true;
3619
3620 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3621
3622 // Now load the mirror's klass metaobject, and null-check it.
3623 // Side-effects region with the control path if the klass is null.
3624 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3625 // If kls is null, we have a primitive mirror.
3626 phi->init_req(_prim_path, prim_return_value);
3627 if (stopped()) { set_result(region, phi); return true; }
3628 bool safe_for_replace = (region->in(_prim_path) == top());
3629
3630 Node* p; // handy temp
3631 Node* null_ctl;
3632
3633 // Now that we have the non-null klass, we can perform the real query.
3634 // For constant classes, the query will constant-fold in LoadNode::Value.
3635 Node* query_value = top();
3636 switch (id) {
3637 case vmIntrinsics::_isInstance:
3638 // nothing is an instance of a primitive type
3639 query_value = gen_instanceof(obj, kls, safe_for_replace);
3640 break;
3641
3642 case vmIntrinsics::_getModifiers:
3643 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3644 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3645 break;
3646
3647 case vmIntrinsics::_isInterface:
3648 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3649 if (generate_interface_guard(kls, region) != NULL)
3650 // A guard was added. If the guard is taken, it was an interface.
3651 phi->add_req(intcon(1));
3652 // If we fall through, it's a plain class.
3653 query_value = intcon(0);
3654 break;
3655
3656 case vmIntrinsics::_isArray:
3657 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3658 if (generate_array_guard(kls, region) != NULL)
3659 // A guard was added. If the guard is taken, it was an array.
3660 phi->add_req(intcon(1));
3661 // If we fall through, it's a plain class.
3662 query_value = intcon(0);
3663 break;
3664
3665 case vmIntrinsics::_isPrimitive:
3666 query_value = intcon(0); // "normal" path produces false
3667 break;
3668
3669 case vmIntrinsics::_getSuperclass:
3670 // The rules here are somewhat unfortunate, but we can still do better
3671 // with random logic than with a JNI call.
3672 // Interfaces store null or Object as _super, but must report null.
3673 // Arrays store an intermediate super as _super, but must report Object.
3674 // Other types can report the actual _super.
3675 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3676 if (generate_interface_guard(kls, region) != NULL)
3677 // A guard was added. If the guard is taken, it was an interface.
3678 phi->add_req(null());
3679 if (generate_array_guard(kls, region) != NULL)
3680 // A guard was added. If the guard is taken, it was an array.
3681 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3682 // If we fall through, it's a plain class. Get its _super.
3683 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3684 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3685 null_ctl = top();
3686 kls = null_check_oop(kls, &null_ctl);
3687 if (null_ctl != top()) {
3688 // If the guard is taken, Object.superClass is null (both klass and mirror).
3689 region->add_req(null_ctl);
3690 phi ->add_req(null());
3691 }
3692 if (!stopped()) {
3693 query_value = load_mirror_from_klass(kls);
3694 }
3695 break;
3696
3697 case vmIntrinsics::_getClassAccessFlags:
3698 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3699 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3700 break;
3701
3702 default:
3703 fatal_unexpected_iid(id);
3704 break;
3705 }
3706
3707 // Fall-through is the normal case of a query to a real class.
3708 phi->init_req(1, query_value);
3709 region->init_req(1, control());
3710
3711 C->set_has_split_ifs(true); // Has chance for split-if optimization
3712 set_result(region, phi);
3713 return true;
3714 }
3715
3716 //-------------------------inline_Class_cast-------------------
3717 bool LibraryCallKit::inline_Class_cast() {
3718 Node* mirror = argument(0); // Class
3719 Node* obj = argument(1);
3720 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3721 if (mirror_con == NULL) {
3722 return false; // dead path (mirror->is_top()).
3723 }
3724 if (obj == NULL || obj->is_top()) {
3725 return false; // dead path
3726 }
3727 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3728
3729 // First, see if Class.cast() can be folded statically.
3730 // java_mirror_type() returns non-null for compile-time Class constants.
3731 ciType* tm = mirror_con->java_mirror_type();
3732 if (tm != NULL && tm->is_klass() &&
3733 tp != NULL && tp->klass() != NULL) {
3734 if (!tp->klass()->is_loaded()) {
3735 // Don't use intrinsic when class is not loaded.
3736 return false;
3737 } else {
3738 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3739 if (static_res == Compile::SSC_always_true) {
3740 // isInstance() is true - fold the code.
3741 set_result(obj);
3742 return true;
3743 } else if (static_res == Compile::SSC_always_false) {
3744 // Don't use intrinsic, have to throw ClassCastException.
3745 // If the reference is null, the non-intrinsic bytecode will
3746 // be optimized appropriately.
3747 return false;
3748 }
3749 }
3750 }
3751
3752 // Bailout intrinsic and do normal inlining if exception path is frequent.
3753 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3754 return false;
3755 }
3756
3757 // Generate dynamic checks.
3758 // Class.cast() is java implementation of _checkcast bytecode.
3759 // Do checkcast (Parse::do_checkcast()) optimizations here.
3760
3761 mirror = null_check(mirror);
3762 // If mirror is dead, only null-path is taken.
3763 if (stopped()) {
3764 return true;
3765 }
3766
3767 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3768 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3769 RegionNode* region = new RegionNode(PATH_LIMIT);
3770 record_for_igvn(region);
3771
3772 // Now load the mirror's klass metaobject, and null-check it.
3773 // If kls is null, we have a primitive mirror and
3774 // nothing is an instance of a primitive type.
3775 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3776
3777 Node* res = top();
3778 if (!stopped()) {
3779 Node* bad_type_ctrl = top();
3780 // Do checkcast optimizations.
3781 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3782 region->init_req(_bad_type_path, bad_type_ctrl);
3783 }
3784 if (region->in(_prim_path) != top() ||
3785 region->in(_bad_type_path) != top()) {
3786 // Let Interpreter throw ClassCastException.
3787 PreserveJVMState pjvms(this);
3788 set_control(_gvn.transform(region));
3789 uncommon_trap(Deoptimization::Reason_intrinsic,
3790 Deoptimization::Action_maybe_recompile);
3791 }
3792 if (!stopped()) {
3793 set_result(res);
3794 }
3795 return true;
3796 }
3797
3798
3799 //--------------------------inline_native_subtype_check------------------------
3800 // This intrinsic takes the JNI calls out of the heart of
3801 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3802 bool LibraryCallKit::inline_native_subtype_check() {
3803 // Pull both arguments off the stack.
3804 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3805 args[0] = argument(0);
3806 args[1] = argument(1);
3807 Node* klasses[2]; // corresponding Klasses: superk, subk
3808 klasses[0] = klasses[1] = top();
3809
3810 enum {
3811 // A full decision tree on {superc is prim, subc is prim}:
3812 _prim_0_path = 1, // {P,N} => false
3813 // {P,P} & superc!=subc => false
3814 _prim_same_path, // {P,P} & superc==subc => true
3815 _prim_1_path, // {N,P} => false
3816 _ref_subtype_path, // {N,N} & subtype check wins => true
3817 _both_ref_path, // {N,N} & subtype check loses => false
3818 PATH_LIMIT
3819 };
3820
3821 RegionNode* region = new RegionNode(PATH_LIMIT);
3822 Node* phi = new PhiNode(region, TypeInt::BOOL);
3823 record_for_igvn(region);
3824
3825 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3826 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3827 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3828
3829 // First null-check both mirrors and load each mirror's klass metaobject.
3830 int which_arg;
3831 for (which_arg = 0; which_arg <= 1; which_arg++) {
3832 Node* arg = args[which_arg];
3833 arg = null_check(arg);
3834 if (stopped()) break;
3835 args[which_arg] = arg;
3836
3837 Node* p = basic_plus_adr(arg, class_klass_offset);
3838 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3839 klasses[which_arg] = _gvn.transform(kls);
3840 }
3841
3842 // Having loaded both klasses, test each for null.
3843 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3844 for (which_arg = 0; which_arg <= 1; which_arg++) {
3845 Node* kls = klasses[which_arg];
3846 Node* null_ctl = top();
3847 kls = null_check_oop(kls, &null_ctl, never_see_null);
3848 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3849 region->init_req(prim_path, null_ctl);
3850 if (stopped()) break;
3851 klasses[which_arg] = kls;
3852 }
3853
3854 if (!stopped()) {
3855 // now we have two reference types, in klasses[0..1]
3856 Node* subk = klasses[1]; // the argument to isAssignableFrom
3857 Node* superk = klasses[0]; // the receiver
3858 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3859 // now we have a successful reference subtype check
3860 region->set_req(_ref_subtype_path, control());
3861 }
3862
3863 // If both operands are primitive (both klasses null), then
3864 // we must return true when they are identical primitives.
3865 // It is convenient to test this after the first null klass check.
3866 set_control(region->in(_prim_0_path)); // go back to first null check
3867 if (!stopped()) {
3868 // Since superc is primitive, make a guard for the superc==subc case.
3869 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3870 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3871 generate_guard(bol_eq, region, PROB_FAIR);
3872 if (region->req() == PATH_LIMIT+1) {
3873 // A guard was added. If the added guard is taken, superc==subc.
3874 region->swap_edges(PATH_LIMIT, _prim_same_path);
3875 region->del_req(PATH_LIMIT);
3876 }
3877 region->set_req(_prim_0_path, control()); // Not equal after all.
3878 }
3879
3880 // these are the only paths that produce 'true':
3881 phi->set_req(_prim_same_path, intcon(1));
3882 phi->set_req(_ref_subtype_path, intcon(1));
3883
3884 // pull together the cases:
3885 assert(region->req() == PATH_LIMIT, "sane region");
3886 for (uint i = 1; i < region->req(); i++) {
3887 Node* ctl = region->in(i);
3888 if (ctl == NULL || ctl == top()) {
3889 region->set_req(i, top());
3890 phi ->set_req(i, top());
3891 } else if (phi->in(i) == NULL) {
3892 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3893 }
3894 }
3895
3896 set_control(_gvn.transform(region));
3897 set_result(_gvn.transform(phi));
3898 return true;
3899 }
3900
3901 //---------------------generate_array_guard_common------------------------
3902 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3903 bool obj_array, bool not_array) {
3904
3905 if (stopped()) {
3906 return NULL;
3907 }
3908
3909 // If obj_array/non_array==false/false:
3910 // Branch around if the given klass is in fact an array (either obj or prim).
3911 // If obj_array/non_array==false/true:
3912 // Branch around if the given klass is not an array klass of any kind.
3913 // If obj_array/non_array==true/true:
3914 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3915 // If obj_array/non_array==true/false:
3916 // Branch around if the kls is an oop array (Object[] or subtype)
3917 //
3918 // Like generate_guard, adds a new path onto the region.
3919 jint layout_con = 0;
3920 Node* layout_val = get_layout_helper(kls, layout_con);
3921 if (layout_val == NULL) {
3922 bool query = (obj_array
3923 ? Klass::layout_helper_is_objArray(layout_con)
3924 : Klass::layout_helper_is_array(layout_con));
3925 if (query == not_array) {
3926 return NULL; // never a branch
3927 } else { // always a branch
3928 Node* always_branch = control();
3929 if (region != NULL)
3930 region->add_req(always_branch);
3931 set_control(top());
3932 return always_branch;
3933 }
3934 }
3935 // Now test the correct condition.
3936 jint nval = (obj_array
3937 ? (jint)(Klass::_lh_array_tag_type_value
3938 << Klass::_lh_array_tag_shift)
3939 : Klass::_lh_neutral_value);
3940 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3941 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3942 // invert the test if we are looking for a non-array
3943 if (not_array) btest = BoolTest(btest).negate();
3944 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3945 return generate_fair_guard(bol, region);
3946 }
3947
3948
3949 //-----------------------inline_native_newArray--------------------------
3950 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3951 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3952 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3953 Node* mirror;
3954 Node* count_val;
3955 if (uninitialized) {
3956 mirror = argument(1);
3957 count_val = argument(2);
3958 } else {
3959 mirror = argument(0);
3960 count_val = argument(1);
3961 }
3962
3963 mirror = null_check(mirror);
3964 // If mirror or obj is dead, only null-path is taken.
3965 if (stopped()) return true;
3966
3967 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3968 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3969 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3970 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3971 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3972
3973 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3974 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3975 result_reg, _slow_path);
3976 Node* normal_ctl = control();
3977 Node* no_array_ctl = result_reg->in(_slow_path);
3978
3979 // Generate code for the slow case. We make a call to newArray().
3980 set_control(no_array_ctl);
3981 if (!stopped()) {
3982 // Either the input type is void.class, or else the
3983 // array klass has not yet been cached. Either the
3984 // ensuing call will throw an exception, or else it
3985 // will cache the array klass for next time.
3986 PreserveJVMState pjvms(this);
3987 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3988 Node* slow_result = set_results_for_java_call(slow_call);
3989 // this->control() comes from set_results_for_java_call
3990 result_reg->set_req(_slow_path, control());
3991 result_val->set_req(_slow_path, slow_result);
3992 result_io ->set_req(_slow_path, i_o());
3993 result_mem->set_req(_slow_path, reset_memory());
3994 }
3995
3996 set_control(normal_ctl);
3997 if (!stopped()) {
3998 // Normal case: The array type has been cached in the java.lang.Class.
3999 // The following call works fine even if the array type is polymorphic.
4000 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4001 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
4002 result_reg->init_req(_normal_path, control());
4003 result_val->init_req(_normal_path, obj);
4004 result_io ->init_req(_normal_path, i_o());
4005 result_mem->init_req(_normal_path, reset_memory());
4006
4007 if (uninitialized) {
4008 // Mark the allocation so that zeroing is skipped
4009 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4010 alloc->maybe_set_complete(&_gvn);
4011 }
4012 }
4013
4014 // Return the combined state.
4015 set_i_o( _gvn.transform(result_io) );
4016 set_all_memory( _gvn.transform(result_mem));
4017
4018 C->set_has_split_ifs(true); // Has chance for split-if optimization
4019 set_result(result_reg, result_val);
4020 return true;
4021 }
4022
4023 //----------------------inline_native_getLength--------------------------
4024 // public static native int java.lang.reflect.Array.getLength(Object array);
4025 bool LibraryCallKit::inline_native_getLength() {
4026 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4027
4028 Node* array = null_check(argument(0));
4029 // If array is dead, only null-path is taken.
4030 if (stopped()) return true;
4031
4032 // Deoptimize if it is a non-array.
4033 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4034
4035 if (non_array != NULL) {
4036 PreserveJVMState pjvms(this);
4037 set_control(non_array);
4038 uncommon_trap(Deoptimization::Reason_intrinsic,
4039 Deoptimization::Action_maybe_recompile);
4040 }
4041
4042 // If control is dead, only non-array-path is taken.
4043 if (stopped()) return true;
4044
4045 // The works fine even if the array type is polymorphic.
4046 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4047 Node* result = load_array_length(array);
4048
4049 C->set_has_split_ifs(true); // Has chance for split-if optimization
4050 set_result(result);
4051 return true;
4052 }
4053
4054 //------------------------inline_array_copyOf----------------------------
4055 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4056 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4057 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4058 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4059
4060 // Get the arguments.
4061 Node* original = argument(0);
4062 Node* start = is_copyOfRange? argument(1): intcon(0);
4063 Node* end = is_copyOfRange? argument(2): argument(1);
4064 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4065
4066 Node* newcopy = NULL;
4067
4068 // Set the original stack and the reexecute bit for the interpreter to reexecute
4069 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4070 { PreserveReexecuteState preexecs(this);
4071 jvms()->set_should_reexecute(true);
4072
4073 array_type_mirror = null_check(array_type_mirror);
4074 original = null_check(original);
4075
4076 // Check if a null path was taken unconditionally.
4077 if (stopped()) return true;
4078
4079 Node* orig_length = load_array_length(original);
4080
4081 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4082 klass_node = null_check(klass_node);
4083
4084 RegionNode* bailout = new RegionNode(1);
4085 record_for_igvn(bailout);
4086
4087 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4088 // Bail out if that is so.
4089 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4090 if (not_objArray != NULL) {
4091 // Improve the klass node's type from the new optimistic assumption:
4092 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4093 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4094 Node* cast = new CastPPNode(klass_node, akls);
4095 cast->init_req(0, control());
4096 klass_node = _gvn.transform(cast);
4097 }
4098
4099 // Bail out if either start or end is negative.
4100 generate_negative_guard(start, bailout, &start);
4101 generate_negative_guard(end, bailout, &end);
4102
4103 Node* length = end;
4104 if (_gvn.type(start) != TypeInt::ZERO) {
4105 length = _gvn.transform(new SubINode(end, start));
4106 }
4107
4108 // Bail out if length is negative.
4109 // Without this the new_array would throw
4110 // NegativeArraySizeException but IllegalArgumentException is what
4111 // should be thrown
4112 generate_negative_guard(length, bailout, &length);
4113
4114 if (bailout->req() > 1) {
4115 PreserveJVMState pjvms(this);
4116 set_control(_gvn.transform(bailout));
4117 uncommon_trap(Deoptimization::Reason_intrinsic,
4118 Deoptimization::Action_maybe_recompile);
4119 }
4120
4121 if (!stopped()) {
4122 // How many elements will we copy from the original?
4123 // The answer is MinI(orig_length - start, length).
4124 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4125 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4126
4127 // Generate a direct call to the right arraycopy function(s).
4128 // We know the copy is disjoint but we might not know if the
4129 // oop stores need checking.
4130 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4131 // This will fail a store-check if x contains any non-nulls.
4132
4133 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4134 // loads/stores but it is legal only if we're sure the
4135 // Arrays.copyOf would succeed. So we need all input arguments
4136 // to the copyOf to be validated, including that the copy to the
4137 // new array won't trigger an ArrayStoreException. That subtype
4138 // check can be optimized if we know something on the type of
4139 // the input array from type speculation.
4140 if (_gvn.type(klass_node)->singleton()) {
4141 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4142 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4143
4144 int test = C->static_subtype_check(superk, subk);
4145 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4146 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4147 if (t_original->speculative_type() != NULL) {
4148 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4149 }
4150 }
4151 }
4152
4153 bool validated = false;
4154 // Reason_class_check rather than Reason_intrinsic because we
4155 // want to intrinsify even if this traps.
4156 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4157 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4158 klass_node);
4159
4160 if (not_subtype_ctrl != top()) {
4161 PreserveJVMState pjvms(this);
4162 set_control(not_subtype_ctrl);
4163 uncommon_trap(Deoptimization::Reason_class_check,
4164 Deoptimization::Action_make_not_entrant);
4165 assert(stopped(), "Should be stopped");
4166 }
4167 validated = true;
4168 }
4169
4170 if (!stopped()) {
4171 newcopy = new_array(klass_node, length, 0); // no arguments to push
4172
4173 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4174 load_object_klass(original), klass_node);
4175 if (!is_copyOfRange) {
4176 ac->set_copyof(validated);
4177 } else {
4178 ac->set_copyofrange(validated);
4179 }
4180 Node* n = _gvn.transform(ac);
4181 if (n == ac) {
4182 ac->connect_outputs(this);
4183 } else {
4184 assert(validated, "shouldn't transform if all arguments not validated");
4185 set_all_memory(n);
4186 }
4187 }
4188 }
4189 } // original reexecute is set back here
4190
4191 C->set_has_split_ifs(true); // Has chance for split-if optimization
4192 if (!stopped()) {
4193 set_result(newcopy);
4194 }
4195 return true;
4196 }
4197
4198
4199 //----------------------generate_virtual_guard---------------------------
4200 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4201 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4202 RegionNode* slow_region) {
4203 ciMethod* method = callee();
4204 int vtable_index = method->vtable_index();
4205 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4206 "bad index %d", vtable_index);
4207 // Get the Method* out of the appropriate vtable entry.
4208 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4209 vtable_index*vtableEntry::size_in_bytes() +
4210 vtableEntry::method_offset_in_bytes();
4211 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4212 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4213
4214 // Compare the target method with the expected method (e.g., Object.hashCode).
4215 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4216
4217 Node* native_call = makecon(native_call_addr);
4218 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4219 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4220
4221 return generate_slow_guard(test_native, slow_region);
4222 }
4223
4224 //-----------------------generate_method_call----------------------------
4225 // Use generate_method_call to make a slow-call to the real
4226 // method if the fast path fails. An alternative would be to
4227 // use a stub like OptoRuntime::slow_arraycopy_Java.
4228 // This only works for expanding the current library call,
4229 // not another intrinsic. (E.g., don't use this for making an
4230 // arraycopy call inside of the copyOf intrinsic.)
4231 CallJavaNode*
4232 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4233 // When compiling the intrinsic method itself, do not use this technique.
4234 guarantee(callee() != C->method(), "cannot make slow-call to self");
4235
4236 ciMethod* method = callee();
4237 // ensure the JVMS we have will be correct for this call
4238 guarantee(method_id == method->intrinsic_id(), "must match");
4239
4240 const TypeFunc* tf = TypeFunc::make(method);
4241 CallJavaNode* slow_call;
4242 if (is_static) {
4243 assert(!is_virtual, "");
4244 slow_call = new CallStaticJavaNode(C, tf,
4245 SharedRuntime::get_resolve_static_call_stub(),
4246 method, bci());
4247 } else if (is_virtual) {
4248 null_check_receiver();
4249 int vtable_index = Method::invalid_vtable_index;
4250 if (UseInlineCaches) {
4251 // Suppress the vtable call
4252 } else {
4253 // hashCode and clone are not a miranda methods,
4254 // so the vtable index is fixed.
4255 // No need to use the linkResolver to get it.
4256 vtable_index = method->vtable_index();
4257 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4258 "bad index %d", vtable_index);
4259 }
4260 slow_call = new CallDynamicJavaNode(tf,
4261 SharedRuntime::get_resolve_virtual_call_stub(),
4262 method, vtable_index, bci());
4263 } else { // neither virtual nor static: opt_virtual
4264 null_check_receiver();
4265 slow_call = new CallStaticJavaNode(C, tf,
4266 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4267 method, bci());
4268 slow_call->set_optimized_virtual(true);
4269 }
4270 set_arguments_for_java_call(slow_call);
4271 set_edges_for_java_call(slow_call);
4272 return slow_call;
4273 }
4274
4275
4276 /**
4277 * Build special case code for calls to hashCode on an object. This call may
4278 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4279 * slightly different code.
4280 */
4281 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4282 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4283 assert(!(is_virtual && is_static), "either virtual, special, or static");
4284
4285 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4286
4287 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4288 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4289 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4290 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4291 Node* obj = NULL;
4292 if (!is_static) {
4293 // Check for hashing null object
4294 obj = null_check_receiver();
4295 if (stopped()) return true; // unconditionally null
4296 result_reg->init_req(_null_path, top());
4297 result_val->init_req(_null_path, top());
4298 } else {
4299 // Do a null check, and return zero if null.
4300 // System.identityHashCode(null) == 0
4301 obj = argument(0);
4302 Node* null_ctl = top();
4303 obj = null_check_oop(obj, &null_ctl);
4304 result_reg->init_req(_null_path, null_ctl);
4305 result_val->init_req(_null_path, _gvn.intcon(0));
4306 }
4307
4308 // Unconditionally null? Then return right away.
4309 if (stopped()) {
4310 set_control( result_reg->in(_null_path));
4311 if (!stopped())
4312 set_result(result_val->in(_null_path));
4313 return true;
4314 }
4315
4316 // We only go to the fast case code if we pass a number of guards. The
4317 // paths which do not pass are accumulated in the slow_region.
4318 RegionNode* slow_region = new RegionNode(1);
4319 record_for_igvn(slow_region);
4320
4321 // If this is a virtual call, we generate a funny guard. We pull out
4322 // the vtable entry corresponding to hashCode() from the target object.
4323 // If the target method which we are calling happens to be the native
4324 // Object hashCode() method, we pass the guard. We do not need this
4325 // guard for non-virtual calls -- the caller is known to be the native
4326 // Object hashCode().
4327 if (is_virtual) {
4328 // After null check, get the object's klass.
4329 Node* obj_klass = load_object_klass(obj);
4330 generate_virtual_guard(obj_klass, slow_region);
4331 }
4332
4333 // Get the header out of the object, use LoadMarkNode when available
4334 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4335 // The control of the load must be NULL. Otherwise, the load can move before
4336 // the null check after castPP removal.
4337 Node* no_ctrl = NULL;
4338 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4339
4340 // Test the header to see if it is unlocked.
4341 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4342 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4343 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4344 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4345 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4346
4347 generate_slow_guard(test_unlocked, slow_region);
4348
4349 // Get the hash value and check to see that it has been properly assigned.
4350 // We depend on hash_mask being at most 32 bits and avoid the use of
4351 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4352 // vm: see markOop.hpp.
4353 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4354 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4355 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4356 // This hack lets the hash bits live anywhere in the mark object now, as long
4357 // as the shift drops the relevant bits into the low 32 bits. Note that
4358 // Java spec says that HashCode is an int so there's no point in capturing
4359 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4360 hshifted_header = ConvX2I(hshifted_header);
4361 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4362
4363 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4364 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4365 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4366
4367 generate_slow_guard(test_assigned, slow_region);
4368
4369 Node* init_mem = reset_memory();
4370 // fill in the rest of the null path:
4371 result_io ->init_req(_null_path, i_o());
4372 result_mem->init_req(_null_path, init_mem);
4373
4374 result_val->init_req(_fast_path, hash_val);
4375 result_reg->init_req(_fast_path, control());
4376 result_io ->init_req(_fast_path, i_o());
4377 result_mem->init_req(_fast_path, init_mem);
4378
4379 // Generate code for the slow case. We make a call to hashCode().
4380 set_control(_gvn.transform(slow_region));
4381 if (!stopped()) {
4382 // No need for PreserveJVMState, because we're using up the present state.
4383 set_all_memory(init_mem);
4384 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4385 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4386 Node* slow_result = set_results_for_java_call(slow_call);
4387 // this->control() comes from set_results_for_java_call
4388 result_reg->init_req(_slow_path, control());
4389 result_val->init_req(_slow_path, slow_result);
4390 result_io ->set_req(_slow_path, i_o());
4391 result_mem ->set_req(_slow_path, reset_memory());
4392 }
4393
4394 // Return the combined state.
4395 set_i_o( _gvn.transform(result_io) );
4396 set_all_memory( _gvn.transform(result_mem));
4397
4398 set_result(result_reg, result_val);
4399 return true;
4400 }
4401
4402 //---------------------------inline_native_getClass----------------------------
4403 // public final native Class<?> java.lang.Object.getClass();
4404 //
4405 // Build special case code for calls to getClass on an object.
4406 bool LibraryCallKit::inline_native_getClass() {
4407 Node* obj = null_check_receiver();
4408 if (stopped()) return true;
4409 set_result(load_mirror_from_klass(load_object_klass(obj)));
4410 return true;
4411 }
4412
4413 //-----------------inline_native_Reflection_getCallerClass---------------------
4414 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4415 //
4416 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4417 //
4418 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4419 // in that it must skip particular security frames and checks for
4420 // caller sensitive methods.
4421 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4422 #ifndef PRODUCT
4423 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4424 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4425 }
4426 #endif
4427
4428 if (!jvms()->has_method()) {
4429 #ifndef PRODUCT
4430 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4431 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4432 }
4433 #endif
4434 return false;
4435 }
4436
4437 // Walk back up the JVM state to find the caller at the required
4438 // depth.
4439 JVMState* caller_jvms = jvms();
4440
4441 // Cf. JVM_GetCallerClass
4442 // NOTE: Start the loop at depth 1 because the current JVM state does
4443 // not include the Reflection.getCallerClass() frame.
4444 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4445 ciMethod* m = caller_jvms->method();
4446 switch (n) {
4447 case 0:
4448 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4449 break;
4450 case 1:
4451 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4452 if (!m->caller_sensitive()) {
4453 #ifndef PRODUCT
4454 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4455 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4456 }
4457 #endif
4458 return false; // bail-out; let JVM_GetCallerClass do the work
4459 }
4460 break;
4461 default:
4462 if (!m->is_ignored_by_security_stack_walk()) {
4463 // We have reached the desired frame; return the holder class.
4464 // Acquire method holder as java.lang.Class and push as constant.
4465 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4466 ciInstance* caller_mirror = caller_klass->java_mirror();
4467 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4468
4469 #ifndef PRODUCT
4470 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4471 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());
4472 tty->print_cr(" JVM state at this point:");
4473 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4474 ciMethod* m = jvms()->of_depth(i)->method();
4475 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4476 }
4477 }
4478 #endif
4479 return true;
4480 }
4481 break;
4482 }
4483 }
4484
4485 #ifndef PRODUCT
4486 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4487 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4488 tty->print_cr(" JVM state at this point:");
4489 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4490 ciMethod* m = jvms()->of_depth(i)->method();
4491 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4492 }
4493 }
4494 #endif
4495
4496 return false; // bail-out; let JVM_GetCallerClass do the work
4497 }
4498
4499 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4500 Node* arg = argument(0);
4501 Node* result = NULL;
4502
4503 switch (id) {
4504 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4505 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4506 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4507 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4508
4509 case vmIntrinsics::_doubleToLongBits: {
4510 // two paths (plus control) merge in a wood
4511 RegionNode *r = new RegionNode(3);
4512 Node *phi = new PhiNode(r, TypeLong::LONG);
4513
4514 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4515 // Build the boolean node
4516 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4517
4518 // Branch either way.
4519 // NaN case is less traveled, which makes all the difference.
4520 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4521 Node *opt_isnan = _gvn.transform(ifisnan);
4522 assert( opt_isnan->is_If(), "Expect an IfNode");
4523 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4524 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4525
4526 set_control(iftrue);
4527
4528 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4529 Node *slow_result = longcon(nan_bits); // return NaN
4530 phi->init_req(1, _gvn.transform( slow_result ));
4531 r->init_req(1, iftrue);
4532
4533 // Else fall through
4534 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4535 set_control(iffalse);
4536
4537 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4538 r->init_req(2, iffalse);
4539
4540 // Post merge
4541 set_control(_gvn.transform(r));
4542 record_for_igvn(r);
4543
4544 C->set_has_split_ifs(true); // Has chance for split-if optimization
4545 result = phi;
4546 assert(result->bottom_type()->isa_long(), "must be");
4547 break;
4548 }
4549
4550 case vmIntrinsics::_floatToIntBits: {
4551 // two paths (plus control) merge in a wood
4552 RegionNode *r = new RegionNode(3);
4553 Node *phi = new PhiNode(r, TypeInt::INT);
4554
4555 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4556 // Build the boolean node
4557 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4558
4559 // Branch either way.
4560 // NaN case is less traveled, which makes all the difference.
4561 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4562 Node *opt_isnan = _gvn.transform(ifisnan);
4563 assert( opt_isnan->is_If(), "Expect an IfNode");
4564 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4565 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4566
4567 set_control(iftrue);
4568
4569 static const jint nan_bits = 0x7fc00000;
4570 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4571 phi->init_req(1, _gvn.transform( slow_result ));
4572 r->init_req(1, iftrue);
4573
4574 // Else fall through
4575 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4576 set_control(iffalse);
4577
4578 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4579 r->init_req(2, iffalse);
4580
4581 // Post merge
4582 set_control(_gvn.transform(r));
4583 record_for_igvn(r);
4584
4585 C->set_has_split_ifs(true); // Has chance for split-if optimization
4586 result = phi;
4587 assert(result->bottom_type()->isa_int(), "must be");
4588 break;
4589 }
4590
4591 default:
4592 fatal_unexpected_iid(id);
4593 break;
4594 }
4595 set_result(_gvn.transform(result));
4596 return true;
4597 }
4598
4599 //----------------------inline_unsafe_copyMemory-------------------------
4600 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4601 bool LibraryCallKit::inline_unsafe_copyMemory() {
4602 if (callee()->is_static()) return false; // caller must have the capability!
4603 null_check_receiver(); // null-check receiver
4604 if (stopped()) return true;
4605
4606 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4607
4608 Node* src_ptr = argument(1); // type: oop
4609 Node* src_off = ConvL2X(argument(2)); // type: long
4610 Node* dst_ptr = argument(4); // type: oop
4611 Node* dst_off = ConvL2X(argument(5)); // type: long
4612 Node* size = ConvL2X(argument(7)); // type: long
4613
4614 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4615 "fieldOffset must be byte-scaled");
4616
4617 Node* src = make_unsafe_address(src_ptr, src_off);
4618 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4619
4620 // Conservatively insert a memory barrier on all memory slices.
4621 // Do not let writes of the copy source or destination float below the copy.
4622 insert_mem_bar(Op_MemBarCPUOrder);
4623
4624 // Call it. Note that the length argument is not scaled.
4625 make_runtime_call(RC_LEAF|RC_NO_FP,
4626 OptoRuntime::fast_arraycopy_Type(),
4627 StubRoutines::unsafe_arraycopy(),
4628 "unsafe_arraycopy",
4629 TypeRawPtr::BOTTOM,
4630 src, dst, size XTOP);
4631
4632 // Do not let reads of the copy destination float above the copy.
4633 insert_mem_bar(Op_MemBarCPUOrder);
4634
4635 return true;
4636 }
4637
4638 //------------------------clone_coping-----------------------------------
4639 // Helper function for inline_native_clone.
4640 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4641 assert(obj_size != NULL, "");
4642 Node* raw_obj = alloc_obj->in(1);
4643 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4644
4645 AllocateNode* alloc = NULL;
4646 if (ReduceBulkZeroing) {
4647 // We will be completely responsible for initializing this object -
4648 // mark Initialize node as complete.
4649 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4650 // The object was just allocated - there should be no any stores!
4651 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4652 // Mark as complete_with_arraycopy so that on AllocateNode
4653 // expansion, we know this AllocateNode is initialized by an array
4654 // copy and a StoreStore barrier exists after the array copy.
4655 alloc->initialization()->set_complete_with_arraycopy();
4656 }
4657
4658 // Copy the fastest available way.
4659 // TODO: generate fields copies for small objects instead.
4660 Node* src = obj;
4661 Node* dest = alloc_obj;
4662 Node* size = _gvn.transform(obj_size);
4663
4664 // Exclude the header but include array length to copy by 8 bytes words.
4665 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4666 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4667 instanceOopDesc::base_offset_in_bytes();
4668 // base_off:
4669 // 8 - 32-bit VM
4670 // 12 - 64-bit VM, compressed klass
4671 // 16 - 64-bit VM, normal klass
4672 if (base_off % BytesPerLong != 0) {
4673 assert(UseCompressedClassPointers, "");
4674 if (is_array) {
4675 // Exclude length to copy by 8 bytes words.
4676 base_off += sizeof(int);
4677 } else {
4678 // Include klass to copy by 8 bytes words.
4679 base_off = instanceOopDesc::klass_offset_in_bytes();
4680 }
4681 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4682 }
4683 src = basic_plus_adr(src, base_off);
4684 dest = basic_plus_adr(dest, base_off);
4685
4686 // Compute the length also, if needed:
4687 Node* countx = size;
4688 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4689 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4690
4691 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4692
4693 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4694 ac->set_clonebasic();
4695 Node* n = _gvn.transform(ac);
4696 if (n == ac) {
4697 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4698 } else {
4699 set_all_memory(n);
4700 }
4701
4702 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4703 if (card_mark) {
4704 assert(!is_array, "");
4705 // Put in store barrier for any and all oops we are sticking
4706 // into this object. (We could avoid this if we could prove
4707 // that the object type contains no oop fields at all.)
4708 Node* no_particular_value = NULL;
4709 Node* no_particular_field = NULL;
4710 int raw_adr_idx = Compile::AliasIdxRaw;
4711 post_barrier(control(),
4712 memory(raw_adr_type),
4713 alloc_obj,
4714 no_particular_field,
4715 raw_adr_idx,
4716 no_particular_value,
4717 T_OBJECT,
4718 false);
4719 }
4720
4721 // Do not let reads from the cloned object float above the arraycopy.
4722 if (alloc != NULL) {
4723 // Do not let stores that initialize this object be reordered with
4724 // a subsequent store that would make this object accessible by
4725 // other threads.
4726 // Record what AllocateNode this StoreStore protects so that
4727 // escape analysis can go from the MemBarStoreStoreNode to the
4728 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4729 // based on the escape status of the AllocateNode.
4730 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4731 } else {
4732 insert_mem_bar(Op_MemBarCPUOrder);
4733 }
4734 }
4735
4736 //------------------------inline_native_clone----------------------------
4737 // protected native Object java.lang.Object.clone();
4738 //
4739 // Here are the simple edge cases:
4740 // null receiver => normal trap
4741 // virtual and clone was overridden => slow path to out-of-line clone
4742 // not cloneable or finalizer => slow path to out-of-line Object.clone
4743 //
4744 // The general case has two steps, allocation and copying.
4745 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4746 //
4747 // Copying also has two cases, oop arrays and everything else.
4748 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4749 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4750 //
4751 // These steps fold up nicely if and when the cloned object's klass
4752 // can be sharply typed as an object array, a type array, or an instance.
4753 //
4754 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4755 PhiNode* result_val;
4756
4757 // Set the reexecute bit for the interpreter to reexecute
4758 // the bytecode that invokes Object.clone if deoptimization happens.
4759 { PreserveReexecuteState preexecs(this);
4760 jvms()->set_should_reexecute(true);
4761
4762 Node* obj = null_check_receiver();
4763 if (stopped()) return true;
4764
4765 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4766
4767 // If we are going to clone an instance, we need its exact type to
4768 // know the number and types of fields to convert the clone to
4769 // loads/stores. Maybe a speculative type can help us.
4770 if (!obj_type->klass_is_exact() &&
4771 obj_type->speculative_type() != NULL &&
4772 obj_type->speculative_type()->is_instance_klass()) {
4773 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4774 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4775 !spec_ik->has_injected_fields()) {
4776 ciKlass* k = obj_type->klass();
4777 if (!k->is_instance_klass() ||
4778 k->as_instance_klass()->is_interface() ||
4779 k->as_instance_klass()->has_subklass()) {
4780 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4781 }
4782 }
4783 }
4784
4785 Node* obj_klass = load_object_klass(obj);
4786 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4787 const TypeOopPtr* toop = ((tklass != NULL)
4788 ? tklass->as_instance_type()
4789 : TypeInstPtr::NOTNULL);
4790
4791 // Conservatively insert a memory barrier on all memory slices.
4792 // Do not let writes into the original float below the clone.
4793 insert_mem_bar(Op_MemBarCPUOrder);
4794
4795 // paths into result_reg:
4796 enum {
4797 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4798 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4799 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4800 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4801 PATH_LIMIT
4802 };
4803 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4804 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4805 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4806 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4807 record_for_igvn(result_reg);
4808
4809 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4810 int raw_adr_idx = Compile::AliasIdxRaw;
4811
4812 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4813 if (array_ctl != NULL) {
4814 // It's an array.
4815 PreserveJVMState pjvms(this);
4816 set_control(array_ctl);
4817 Node* obj_length = load_array_length(obj);
4818 Node* obj_size = NULL;
4819 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4820
4821 if (!use_ReduceInitialCardMarks()) {
4822 // If it is an oop array, it requires very special treatment,
4823 // because card marking is required on each card of the array.
4824 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4825 if (is_obja != NULL) {
4826 PreserveJVMState pjvms2(this);
4827 set_control(is_obja);
4828 // Generate a direct call to the right arraycopy function(s).
4829 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4830 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4831 ac->set_cloneoop();
4832 Node* n = _gvn.transform(ac);
4833 assert(n == ac, "cannot disappear");
4834 ac->connect_outputs(this);
4835
4836 result_reg->init_req(_objArray_path, control());
4837 result_val->init_req(_objArray_path, alloc_obj);
4838 result_i_o ->set_req(_objArray_path, i_o());
4839 result_mem ->set_req(_objArray_path, reset_memory());
4840 }
4841 }
4842 // Otherwise, there are no card marks to worry about.
4843 // (We can dispense with card marks if we know the allocation
4844 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4845 // causes the non-eden paths to take compensating steps to
4846 // simulate a fresh allocation, so that no further
4847 // card marks are required in compiled code to initialize
4848 // the object.)
4849
4850 if (!stopped()) {
4851 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4852
4853 // Present the results of the copy.
4854 result_reg->init_req(_array_path, control());
4855 result_val->init_req(_array_path, alloc_obj);
4856 result_i_o ->set_req(_array_path, i_o());
4857 result_mem ->set_req(_array_path, reset_memory());
4858 }
4859 }
4860
4861 // We only go to the instance fast case code if we pass a number of guards.
4862 // The paths which do not pass are accumulated in the slow_region.
4863 RegionNode* slow_region = new RegionNode(1);
4864 record_for_igvn(slow_region);
4865 if (!stopped()) {
4866 // It's an instance (we did array above). Make the slow-path tests.
4867 // If this is a virtual call, we generate a funny guard. We grab
4868 // the vtable entry corresponding to clone() from the target object.
4869 // If the target method which we are calling happens to be the
4870 // Object clone() method, we pass the guard. We do not need this
4871 // guard for non-virtual calls; the caller is known to be the native
4872 // Object clone().
4873 if (is_virtual) {
4874 generate_virtual_guard(obj_klass, slow_region);
4875 }
4876
4877 // The object must be easily cloneable and must not have a finalizer.
4878 // Both of these conditions may be checked in a single test.
4879 // We could optimize the test further, but we don't care.
4880 generate_access_flags_guard(obj_klass,
4881 // Test both conditions:
4882 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4883 // Must be cloneable but not finalizer:
4884 JVM_ACC_IS_CLONEABLE_FAST,
4885 slow_region);
4886 }
4887
4888 if (!stopped()) {
4889 // It's an instance, and it passed the slow-path tests.
4890 PreserveJVMState pjvms(this);
4891 Node* obj_size = NULL;
4892 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4893 // is reexecuted if deoptimization occurs and there could be problems when merging
4894 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4895 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4896
4897 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4898
4899 // Present the results of the slow call.
4900 result_reg->init_req(_instance_path, control());
4901 result_val->init_req(_instance_path, alloc_obj);
4902 result_i_o ->set_req(_instance_path, i_o());
4903 result_mem ->set_req(_instance_path, reset_memory());
4904 }
4905
4906 // Generate code for the slow case. We make a call to clone().
4907 set_control(_gvn.transform(slow_region));
4908 if (!stopped()) {
4909 PreserveJVMState pjvms(this);
4910 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4911 Node* slow_result = set_results_for_java_call(slow_call);
4912 // this->control() comes from set_results_for_java_call
4913 result_reg->init_req(_slow_path, control());
4914 result_val->init_req(_slow_path, slow_result);
4915 result_i_o ->set_req(_slow_path, i_o());
4916 result_mem ->set_req(_slow_path, reset_memory());
4917 }
4918
4919 // Return the combined state.
4920 set_control( _gvn.transform(result_reg));
4921 set_i_o( _gvn.transform(result_i_o));
4922 set_all_memory( _gvn.transform(result_mem));
4923 } // original reexecute is set back here
4924
4925 set_result(_gvn.transform(result_val));
4926 return true;
4927 }
4928
4929 // If we have a tighly coupled allocation, the arraycopy may take care
4930 // of the array initialization. If one of the guards we insert between
4931 // the allocation and the arraycopy causes a deoptimization, an
4932 // unitialized array will escape the compiled method. To prevent that
4933 // we set the JVM state for uncommon traps between the allocation and
4934 // the arraycopy to the state before the allocation so, in case of
4935 // deoptimization, we'll reexecute the allocation and the
4936 // initialization.
4937 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4938 if (alloc != NULL) {
4939 ciMethod* trap_method = alloc->jvms()->method();
4940 int trap_bci = alloc->jvms()->bci();
4941
4942 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4943 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4944 // Make sure there's no store between the allocation and the
4945 // arraycopy otherwise visible side effects could be rexecuted
4946 // in case of deoptimization and cause incorrect execution.
4947 bool no_interfering_store = true;
4948 Node* mem = alloc->in(TypeFunc::Memory);
4949 if (mem->is_MergeMem()) {
4950 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4951 Node* n = mms.memory();
4952 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4953 assert(n->is_Store(), "what else?");
4954 no_interfering_store = false;
4955 break;
4956 }
4957 }
4958 } else {
4959 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4960 Node* n = mms.memory();
4961 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4962 assert(n->is_Store(), "what else?");
4963 no_interfering_store = false;
4964 break;
4965 }
4966 }
4967 }
4968
4969 if (no_interfering_store) {
4970 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4971 uint size = alloc->req();
4972 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4973 old_jvms->set_map(sfpt);
4974 for (uint i = 0; i < size; i++) {
4975 sfpt->init_req(i, alloc->in(i));
4976 }
4977 // re-push array length for deoptimization
4978 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4979 old_jvms->set_sp(old_jvms->sp()+1);
4980 old_jvms->set_monoff(old_jvms->monoff()+1);
4981 old_jvms->set_scloff(old_jvms->scloff()+1);
4982 old_jvms->set_endoff(old_jvms->endoff()+1);
4983 old_jvms->set_should_reexecute(true);
4984
4985 sfpt->set_i_o(map()->i_o());
4986 sfpt->set_memory(map()->memory());
4987 sfpt->set_control(map()->control());
4988
4989 JVMState* saved_jvms = jvms();
4990 saved_reexecute_sp = _reexecute_sp;
4991
4992 set_jvms(sfpt->jvms());
4993 _reexecute_sp = jvms()->sp();
4994
4995 return saved_jvms;
4996 }
4997 }
4998 }
4999 return NULL;
5000 }
5001
5002 // In case of a deoptimization, we restart execution at the
5003 // allocation, allocating a new array. We would leave an uninitialized
5004 // array in the heap that GCs wouldn't expect. Move the allocation
5005 // after the traps so we don't allocate the array if we
5006 // deoptimize. This is possible because tightly_coupled_allocation()
5007 // guarantees there's no observer of the allocated array at this point
5008 // and the control flow is simple enough.
5009 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
5010 int saved_reexecute_sp, uint new_idx) {
5011 if (saved_jvms != NULL && !stopped()) {
5012 assert(alloc != NULL, "only with a tightly coupled allocation");
5013 // restore JVM state to the state at the arraycopy
5014 saved_jvms->map()->set_control(map()->control());
5015 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5016 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5017 // If we've improved the types of some nodes (null check) while
5018 // emitting the guards, propagate them to the current state
5019 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
5020 set_jvms(saved_jvms);
5021 _reexecute_sp = saved_reexecute_sp;
5022
5023 // Remove the allocation from above the guards
5024 CallProjections callprojs;
5025 alloc->extract_projections(&callprojs, true);
5026 InitializeNode* init = alloc->initialization();
5027 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5028 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5029 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5030 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5031
5032 // move the allocation here (after the guards)
5033 _gvn.hash_delete(alloc);
5034 alloc->set_req(TypeFunc::Control, control());
5035 alloc->set_req(TypeFunc::I_O, i_o());
5036 Node *mem = reset_memory();
5037 set_all_memory(mem);
5038 alloc->set_req(TypeFunc::Memory, mem);
5039 set_control(init->proj_out_or_null(TypeFunc::Control));
5040 set_i_o(callprojs.fallthrough_ioproj);
5041
5042 // Update memory as done in GraphKit::set_output_for_allocation()
5043 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5044 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5045 if (ary_type->isa_aryptr() && length_type != NULL) {
5046 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5047 }
5048 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5049 int elemidx = C->get_alias_index(telemref);
5050 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
5051 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
5052
5053 Node* allocx = _gvn.transform(alloc);
5054 assert(allocx == alloc, "where has the allocation gone?");
5055 assert(dest->is_CheckCastPP(), "not an allocation result?");
5056
5057 _gvn.hash_delete(dest);
5058 dest->set_req(0, control());
5059 Node* destx = _gvn.transform(dest);
5060 assert(destx == dest, "where has the allocation result gone?");
5061 }
5062 }
5063
5064
5065 //------------------------------inline_arraycopy-----------------------
5066 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5067 // Object dest, int destPos,
5068 // int length);
5069 bool LibraryCallKit::inline_arraycopy() {
5070 // Get the arguments.
5071 Node* src = argument(0); // type: oop
5072 Node* src_offset = argument(1); // type: int
5073 Node* dest = argument(2); // type: oop
5074 Node* dest_offset = argument(3); // type: int
5075 Node* length = argument(4); // type: int
5076
5077 uint new_idx = C->unique();
5078
5079 // Check for allocation before we add nodes that would confuse
5080 // tightly_coupled_allocation()
5081 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5082
5083 int saved_reexecute_sp = -1;
5084 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5085 // See arraycopy_restore_alloc_state() comment
5086 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5087 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5088 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5089 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5090
5091 // The following tests must be performed
5092 // (1) src and dest are arrays.
5093 // (2) src and dest arrays must have elements of the same BasicType
5094 // (3) src and dest must not be null.
5095 // (4) src_offset must not be negative.
5096 // (5) dest_offset must not be negative.
5097 // (6) length must not be negative.
5098 // (7) src_offset + length must not exceed length of src.
5099 // (8) dest_offset + length must not exceed length of dest.
5100 // (9) each element of an oop array must be assignable
5101
5102 // (3) src and dest must not be null.
5103 // always do this here because we need the JVM state for uncommon traps
5104 Node* null_ctl = top();
5105 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5106 assert(null_ctl->is_top(), "no null control here");
5107 dest = null_check(dest, T_ARRAY);
5108
5109 if (!can_emit_guards) {
5110 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5111 // guards but the arraycopy node could still take advantage of a
5112 // tightly allocated allocation. tightly_coupled_allocation() is
5113 // called again to make sure it takes the null check above into
5114 // account: the null check is mandatory and if it caused an
5115 // uncommon trap to be emitted then the allocation can't be
5116 // considered tightly coupled in this context.
5117 alloc = tightly_coupled_allocation(dest, NULL);
5118 }
5119
5120 bool validated = false;
5121
5122 const Type* src_type = _gvn.type(src);
5123 const Type* dest_type = _gvn.type(dest);
5124 const TypeAryPtr* top_src = src_type->isa_aryptr();
5125 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5126
5127 // Do we have the type of src?
5128 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5129 // Do we have the type of dest?
5130 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5131 // Is the type for src from speculation?
5132 bool src_spec = false;
5133 // Is the type for dest from speculation?
5134 bool dest_spec = false;
5135
5136 if ((!has_src || !has_dest) && can_emit_guards) {
5137 // We don't have sufficient type information, let's see if
5138 // speculative types can help. We need to have types for both src
5139 // and dest so that it pays off.
5140
5141 // Do we already have or could we have type information for src
5142 bool could_have_src = has_src;
5143 // Do we already have or could we have type information for dest
5144 bool could_have_dest = has_dest;
5145
5146 ciKlass* src_k = NULL;
5147 if (!has_src) {
5148 src_k = src_type->speculative_type_not_null();
5149 if (src_k != NULL && src_k->is_array_klass()) {
5150 could_have_src = true;
5151 }
5152 }
5153
5154 ciKlass* dest_k = NULL;
5155 if (!has_dest) {
5156 dest_k = dest_type->speculative_type_not_null();
5157 if (dest_k != NULL && dest_k->is_array_klass()) {
5158 could_have_dest = true;
5159 }
5160 }
5161
5162 if (could_have_src && could_have_dest) {
5163 // This is going to pay off so emit the required guards
5164 if (!has_src) {
5165 src = maybe_cast_profiled_obj(src, src_k, true);
5166 src_type = _gvn.type(src);
5167 top_src = src_type->isa_aryptr();
5168 has_src = (top_src != NULL && top_src->klass() != NULL);
5169 src_spec = true;
5170 }
5171 if (!has_dest) {
5172 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5173 dest_type = _gvn.type(dest);
5174 top_dest = dest_type->isa_aryptr();
5175 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5176 dest_spec = true;
5177 }
5178 }
5179 }
5180
5181 if (has_src && has_dest && can_emit_guards) {
5182 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5183 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5184 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5185 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5186
5187 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5188 // If both arrays are object arrays then having the exact types
5189 // for both will remove the need for a subtype check at runtime
5190 // before the call and may make it possible to pick a faster copy
5191 // routine (without a subtype check on every element)
5192 // Do we have the exact type of src?
5193 bool could_have_src = src_spec;
5194 // Do we have the exact type of dest?
5195 bool could_have_dest = dest_spec;
5196 ciKlass* src_k = top_src->klass();
5197 ciKlass* dest_k = top_dest->klass();
5198 if (!src_spec) {
5199 src_k = src_type->speculative_type_not_null();
5200 if (src_k != NULL && src_k->is_array_klass()) {
5201 could_have_src = true;
5202 }
5203 }
5204 if (!dest_spec) {
5205 dest_k = dest_type->speculative_type_not_null();
5206 if (dest_k != NULL && dest_k->is_array_klass()) {
5207 could_have_dest = true;
5208 }
5209 }
5210 if (could_have_src && could_have_dest) {
5211 // If we can have both exact types, emit the missing guards
5212 if (could_have_src && !src_spec) {
5213 src = maybe_cast_profiled_obj(src, src_k, true);
5214 }
5215 if (could_have_dest && !dest_spec) {
5216 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5217 }
5218 }
5219 }
5220 }
5221
5222 ciMethod* trap_method = method();
5223 int trap_bci = bci();
5224 if (saved_jvms != NULL) {
5225 trap_method = alloc->jvms()->method();
5226 trap_bci = alloc->jvms()->bci();
5227 }
5228
5229 bool negative_length_guard_generated = false;
5230
5231 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5232 can_emit_guards &&
5233 !src->is_top() && !dest->is_top()) {
5234 // validate arguments: enables transformation the ArrayCopyNode
5235 validated = true;
5236
5237 RegionNode* slow_region = new RegionNode(1);
5238 record_for_igvn(slow_region);
5239
5240 // (1) src and dest are arrays.
5241 generate_non_array_guard(load_object_klass(src), slow_region);
5242 generate_non_array_guard(load_object_klass(dest), slow_region);
5243
5244 // (2) src and dest arrays must have elements of the same BasicType
5245 // done at macro expansion or at Ideal transformation time
5246
5247 // (4) src_offset must not be negative.
5248 generate_negative_guard(src_offset, slow_region);
5249
5250 // (5) dest_offset must not be negative.
5251 generate_negative_guard(dest_offset, slow_region);
5252
5253 // (7) src_offset + length must not exceed length of src.
5254 generate_limit_guard(src_offset, length,
5255 load_array_length(src),
5256 slow_region);
5257
5258 // (8) dest_offset + length must not exceed length of dest.
5259 generate_limit_guard(dest_offset, length,
5260 load_array_length(dest),
5261 slow_region);
5262
5263 // (6) length must not be negative.
5264 // This is also checked in generate_arraycopy() during macro expansion, but
5265 // we also have to check it here for the case where the ArrayCopyNode will
5266 // be eliminated by Escape Analysis.
5267 if (EliminateAllocations) {
5268 generate_negative_guard(length, slow_region);
5269 negative_length_guard_generated = true;
5270 }
5271
5272 // (9) each element of an oop array must be assignable
5273 Node* src_klass = load_object_klass(src);
5274 Node* dest_klass = load_object_klass(dest);
5275 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5276
5277 if (not_subtype_ctrl != top()) {
5278 PreserveJVMState pjvms(this);
5279 set_control(not_subtype_ctrl);
5280 uncommon_trap(Deoptimization::Reason_intrinsic,
5281 Deoptimization::Action_make_not_entrant);
5282 assert(stopped(), "Should be stopped");
5283 }
5284 {
5285 PreserveJVMState pjvms(this);
5286 set_control(_gvn.transform(slow_region));
5287 uncommon_trap(Deoptimization::Reason_intrinsic,
5288 Deoptimization::Action_make_not_entrant);
5289 assert(stopped(), "Should be stopped");
5290 }
5291
5292 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
5293 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
5294 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
5295 }
5296
5297 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5298
5299 if (stopped()) {
5300 return true;
5301 }
5302
5303 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5304 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5305 // so the compiler has a chance to eliminate them: during macro expansion,
5306 // we have to set their control (CastPP nodes are eliminated).
5307 load_object_klass(src), load_object_klass(dest),
5308 load_array_length(src), load_array_length(dest));
5309
5310 ac->set_arraycopy(validated);
5311
5312 Node* n = _gvn.transform(ac);
5313 if (n == ac) {
5314 ac->connect_outputs(this);
5315 } else {
5316 assert(validated, "shouldn't transform if all arguments not validated");
5317 set_all_memory(n);
5318 }
5319 clear_upper_avx();
5320
5321
5322 return true;
5323 }
5324
5325
5326 // Helper function which determines if an arraycopy immediately follows
5327 // an allocation, with no intervening tests or other escapes for the object.
5328 AllocateArrayNode*
5329 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5330 RegionNode* slow_region) {
5331 if (stopped()) return NULL; // no fast path
5332 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5333
5334 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5335 if (alloc == NULL) return NULL;
5336
5337 Node* rawmem = memory(Compile::AliasIdxRaw);
5338 // Is the allocation's memory state untouched?
5339 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5340 // Bail out if there have been raw-memory effects since the allocation.
5341 // (Example: There might have been a call or safepoint.)
5342 return NULL;
5343 }
5344 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5345 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5346 return NULL;
5347 }
5348
5349 // There must be no unexpected observers of this allocation.
5350 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5351 Node* obs = ptr->fast_out(i);
5352 if (obs != this->map()) {
5353 return NULL;
5354 }
5355 }
5356
5357 // This arraycopy must unconditionally follow the allocation of the ptr.
5358 Node* alloc_ctl = ptr->in(0);
5359 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5360
5361 Node* ctl = control();
5362 while (ctl != alloc_ctl) {
5363 // There may be guards which feed into the slow_region.
5364 // Any other control flow means that we might not get a chance
5365 // to finish initializing the allocated object.
5366 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5367 IfNode* iff = ctl->in(0)->as_If();
5368 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
5369 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5370 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5371 ctl = iff->in(0); // This test feeds the known slow_region.
5372 continue;
5373 }
5374 // One more try: Various low-level checks bottom out in
5375 // uncommon traps. If the debug-info of the trap omits
5376 // any reference to the allocation, as we've already
5377 // observed, then there can be no objection to the trap.
5378 bool found_trap = false;
5379 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5380 Node* obs = not_ctl->fast_out(j);
5381 if (obs->in(0) == not_ctl && obs->is_Call() &&
5382 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5383 found_trap = true; break;
5384 }
5385 }
5386 if (found_trap) {
5387 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5388 continue;
5389 }
5390 }
5391 return NULL;
5392 }
5393
5394 // If we get this far, we have an allocation which immediately
5395 // precedes the arraycopy, and we can take over zeroing the new object.
5396 // The arraycopy will finish the initialization, and provide
5397 // a new control state to which we will anchor the destination pointer.
5398
5399 return alloc;
5400 }
5401
5402 //-------------inline_encodeISOArray-----------------------------------
5403 // encode char[] to byte[] in ISO_8859_1
5404 bool LibraryCallKit::inline_encodeISOArray() {
5405 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5406 // no receiver since it is static method
5407 Node *src = argument(0);
5408 Node *src_offset = argument(1);
5409 Node *dst = argument(2);
5410 Node *dst_offset = argument(3);
5411 Node *length = argument(4);
5412
5413 const Type* src_type = src->Value(&_gvn);
5414 const Type* dst_type = dst->Value(&_gvn);
5415 const TypeAryPtr* top_src = src_type->isa_aryptr();
5416 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5417 if (top_src == NULL || top_src->klass() == NULL ||
5418 top_dest == NULL || top_dest->klass() == NULL) {
5419 // failed array check
5420 return false;
5421 }
5422
5423 // Figure out the size and type of the elements we will be copying.
5424 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5425 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5426 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5427 return false;
5428 }
5429
5430 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5431 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5432 // 'src_start' points to src array + scaled offset
5433 // 'dst_start' points to dst array + scaled offset
5434
5435 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5436 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5437 enc = _gvn.transform(enc);
5438 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5439 set_memory(res_mem, mtype);
5440 set_result(enc);
5441 clear_upper_avx();
5442
5443 return true;
5444 }
5445
5446 //-------------inline_multiplyToLen-----------------------------------
5447 bool LibraryCallKit::inline_multiplyToLen() {
5448 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5449
5450 address stubAddr = StubRoutines::multiplyToLen();
5451 if (stubAddr == NULL) {
5452 return false; // Intrinsic's stub is not implemented on this platform
5453 }
5454 const char* stubName = "multiplyToLen";
5455
5456 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5457
5458 // no receiver because it is a static method
5459 Node* x = argument(0);
5460 Node* xlen = argument(1);
5461 Node* y = argument(2);
5462 Node* ylen = argument(3);
5463 Node* z = argument(4);
5464
5465 const Type* x_type = x->Value(&_gvn);
5466 const Type* y_type = y->Value(&_gvn);
5467 const TypeAryPtr* top_x = x_type->isa_aryptr();
5468 const TypeAryPtr* top_y = y_type->isa_aryptr();
5469 if (top_x == NULL || top_x->klass() == NULL ||
5470 top_y == NULL || top_y->klass() == NULL) {
5471 // failed array check
5472 return false;
5473 }
5474
5475 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5476 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5477 if (x_elem != T_INT || y_elem != T_INT) {
5478 return false;
5479 }
5480
5481 // Set the original stack and the reexecute bit for the interpreter to reexecute
5482 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5483 // on the return from z array allocation in runtime.
5484 { PreserveReexecuteState preexecs(this);
5485 jvms()->set_should_reexecute(true);
5486
5487 Node* x_start = array_element_address(x, intcon(0), x_elem);
5488 Node* y_start = array_element_address(y, intcon(0), y_elem);
5489 // 'x_start' points to x array + scaled xlen
5490 // 'y_start' points to y array + scaled ylen
5491
5492 // Allocate the result array
5493 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5494 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5495 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5496
5497 IdealKit ideal(this);
5498
5499 #define __ ideal.
5500 Node* one = __ ConI(1);
5501 Node* zero = __ ConI(0);
5502 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5503 __ set(need_alloc, zero);
5504 __ set(z_alloc, z);
5505 __ if_then(z, BoolTest::eq, null()); {
5506 __ increment (need_alloc, one);
5507 } __ else_(); {
5508 // Update graphKit memory and control from IdealKit.
5509 sync_kit(ideal);
5510 Node* zlen_arg = load_array_length(z);
5511 // Update IdealKit memory and control from graphKit.
5512 __ sync_kit(this);
5513 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5514 __ increment (need_alloc, one);
5515 } __ end_if();
5516 } __ end_if();
5517
5518 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5519 // Update graphKit memory and control from IdealKit.
5520 sync_kit(ideal);
5521 Node * narr = new_array(klass_node, zlen, 1);
5522 // Update IdealKit memory and control from graphKit.
5523 __ sync_kit(this);
5524 __ set(z_alloc, narr);
5525 } __ end_if();
5526
5527 sync_kit(ideal);
5528 z = __ value(z_alloc);
5529 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5530 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5531 // Final sync IdealKit and GraphKit.
5532 final_sync(ideal);
5533 #undef __
5534
5535 Node* z_start = array_element_address(z, intcon(0), T_INT);
5536
5537 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5538 OptoRuntime::multiplyToLen_Type(),
5539 stubAddr, stubName, TypePtr::BOTTOM,
5540 x_start, xlen, y_start, ylen, z_start, zlen);
5541 } // original reexecute is set back here
5542
5543 C->set_has_split_ifs(true); // Has chance for split-if optimization
5544 set_result(z);
5545 return true;
5546 }
5547
5548 //-------------inline_squareToLen------------------------------------
5549 bool LibraryCallKit::inline_squareToLen() {
5550 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5551
5552 address stubAddr = StubRoutines::squareToLen();
5553 if (stubAddr == NULL) {
5554 return false; // Intrinsic's stub is not implemented on this platform
5555 }
5556 const char* stubName = "squareToLen";
5557
5558 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5559
5560 Node* x = argument(0);
5561 Node* len = argument(1);
5562 Node* z = argument(2);
5563 Node* zlen = argument(3);
5564
5565 const Type* x_type = x->Value(&_gvn);
5566 const Type* z_type = z->Value(&_gvn);
5567 const TypeAryPtr* top_x = x_type->isa_aryptr();
5568 const TypeAryPtr* top_z = z_type->isa_aryptr();
5569 if (top_x == NULL || top_x->klass() == NULL ||
5570 top_z == NULL || top_z->klass() == NULL) {
5571 // failed array check
5572 return false;
5573 }
5574
5575 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5576 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5577 if (x_elem != T_INT || z_elem != T_INT) {
5578 return false;
5579 }
5580
5581
5582 Node* x_start = array_element_address(x, intcon(0), x_elem);
5583 Node* z_start = array_element_address(z, intcon(0), z_elem);
5584
5585 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5586 OptoRuntime::squareToLen_Type(),
5587 stubAddr, stubName, TypePtr::BOTTOM,
5588 x_start, len, z_start, zlen);
5589
5590 set_result(z);
5591 return true;
5592 }
5593
5594 //-------------inline_mulAdd------------------------------------------
5595 bool LibraryCallKit::inline_mulAdd() {
5596 assert(UseMulAddIntrinsic, "not implemented on this platform");
5597
5598 address stubAddr = StubRoutines::mulAdd();
5599 if (stubAddr == NULL) {
5600 return false; // Intrinsic's stub is not implemented on this platform
5601 }
5602 const char* stubName = "mulAdd";
5603
5604 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5605
5606 Node* out = argument(0);
5607 Node* in = argument(1);
5608 Node* offset = argument(2);
5609 Node* len = argument(3);
5610 Node* k = argument(4);
5611
5612 const Type* out_type = out->Value(&_gvn);
5613 const Type* in_type = in->Value(&_gvn);
5614 const TypeAryPtr* top_out = out_type->isa_aryptr();
5615 const TypeAryPtr* top_in = in_type->isa_aryptr();
5616 if (top_out == NULL || top_out->klass() == NULL ||
5617 top_in == NULL || top_in->klass() == NULL) {
5618 // failed array check
5619 return false;
5620 }
5621
5622 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5623 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5624 if (out_elem != T_INT || in_elem != T_INT) {
5625 return false;
5626 }
5627
5628 Node* outlen = load_array_length(out);
5629 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5630 Node* out_start = array_element_address(out, intcon(0), out_elem);
5631 Node* in_start = array_element_address(in, intcon(0), in_elem);
5632
5633 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5634 OptoRuntime::mulAdd_Type(),
5635 stubAddr, stubName, TypePtr::BOTTOM,
5636 out_start,in_start, new_offset, len, k);
5637 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5638 set_result(result);
5639 return true;
5640 }
5641
5642 //-------------inline_montgomeryMultiply-----------------------------------
5643 bool LibraryCallKit::inline_montgomeryMultiply() {
5644 address stubAddr = StubRoutines::montgomeryMultiply();
5645 if (stubAddr == NULL) {
5646 return false; // Intrinsic's stub is not implemented on this platform
5647 }
5648
5649 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5650 const char* stubName = "montgomery_multiply";
5651
5652 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5653
5654 Node* a = argument(0);
5655 Node* b = argument(1);
5656 Node* n = argument(2);
5657 Node* len = argument(3);
5658 Node* inv = argument(4);
5659 Node* m = argument(6);
5660
5661 const Type* a_type = a->Value(&_gvn);
5662 const TypeAryPtr* top_a = a_type->isa_aryptr();
5663 const Type* b_type = b->Value(&_gvn);
5664 const TypeAryPtr* top_b = b_type->isa_aryptr();
5665 const Type* n_type = a->Value(&_gvn);
5666 const TypeAryPtr* top_n = n_type->isa_aryptr();
5667 const Type* m_type = a->Value(&_gvn);
5668 const TypeAryPtr* top_m = m_type->isa_aryptr();
5669 if (top_a == NULL || top_a->klass() == NULL ||
5670 top_b == NULL || top_b->klass() == NULL ||
5671 top_n == NULL || top_n->klass() == NULL ||
5672 top_m == NULL || top_m->klass() == NULL) {
5673 // failed array check
5674 return false;
5675 }
5676
5677 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5678 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5679 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5680 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5681 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5682 return false;
5683 }
5684
5685 // Make the call
5686 {
5687 Node* a_start = array_element_address(a, intcon(0), a_elem);
5688 Node* b_start = array_element_address(b, intcon(0), b_elem);
5689 Node* n_start = array_element_address(n, intcon(0), n_elem);
5690 Node* m_start = array_element_address(m, intcon(0), m_elem);
5691
5692 Node* call = make_runtime_call(RC_LEAF,
5693 OptoRuntime::montgomeryMultiply_Type(),
5694 stubAddr, stubName, TypePtr::BOTTOM,
5695 a_start, b_start, n_start, len, inv, top(),
5696 m_start);
5697 set_result(m);
5698 }
5699
5700 return true;
5701 }
5702
5703 bool LibraryCallKit::inline_montgomerySquare() {
5704 address stubAddr = StubRoutines::montgomerySquare();
5705 if (stubAddr == NULL) {
5706 return false; // Intrinsic's stub is not implemented on this platform
5707 }
5708
5709 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5710 const char* stubName = "montgomery_square";
5711
5712 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5713
5714 Node* a = argument(0);
5715 Node* n = argument(1);
5716 Node* len = argument(2);
5717 Node* inv = argument(3);
5718 Node* m = argument(5);
5719
5720 const Type* a_type = a->Value(&_gvn);
5721 const TypeAryPtr* top_a = a_type->isa_aryptr();
5722 const Type* n_type = a->Value(&_gvn);
5723 const TypeAryPtr* top_n = n_type->isa_aryptr();
5724 const Type* m_type = a->Value(&_gvn);
5725 const TypeAryPtr* top_m = m_type->isa_aryptr();
5726 if (top_a == NULL || top_a->klass() == NULL ||
5727 top_n == NULL || top_n->klass() == NULL ||
5728 top_m == NULL || top_m->klass() == NULL) {
5729 // failed array check
5730 return false;
5731 }
5732
5733 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5734 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5735 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5736 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5737 return false;
5738 }
5739
5740 // Make the call
5741 {
5742 Node* a_start = array_element_address(a, intcon(0), a_elem);
5743 Node* n_start = array_element_address(n, intcon(0), n_elem);
5744 Node* m_start = array_element_address(m, intcon(0), m_elem);
5745
5746 Node* call = make_runtime_call(RC_LEAF,
5747 OptoRuntime::montgomerySquare_Type(),
5748 stubAddr, stubName, TypePtr::BOTTOM,
5749 a_start, n_start, len, inv, top(),
5750 m_start);
5751 set_result(m);
5752 }
5753
5754 return true;
5755 }
5756
5757 //-------------inline_vectorizedMismatch------------------------------
5758 bool LibraryCallKit::inline_vectorizedMismatch() {
5759 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5760
5761 address stubAddr = StubRoutines::vectorizedMismatch();
5762 if (stubAddr == NULL) {
5763 return false; // Intrinsic's stub is not implemented on this platform
5764 }
5765 const char* stubName = "vectorizedMismatch";
5766 int size_l = callee()->signature()->size();
5767 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5768
5769 Node* obja = argument(0);
5770 Node* aoffset = argument(1);
5771 Node* objb = argument(3);
5772 Node* boffset = argument(4);
5773 Node* length = argument(6);
5774 Node* scale = argument(7);
5775
5776 const Type* a_type = obja->Value(&_gvn);
5777 const Type* b_type = objb->Value(&_gvn);
5778 const TypeAryPtr* top_a = a_type->isa_aryptr();
5779 const TypeAryPtr* top_b = b_type->isa_aryptr();
5780 if (top_a == NULL || top_a->klass() == NULL ||
5781 top_b == NULL || top_b->klass() == NULL) {
5782 // failed array check
5783 return false;
5784 }
5785
5786 Node* call;
5787 jvms()->set_should_reexecute(true);
5788
5789 Node* obja_adr = make_unsafe_address(obja, aoffset);
5790 Node* objb_adr = make_unsafe_address(objb, boffset);
5791
5792 call = make_runtime_call(RC_LEAF,
5793 OptoRuntime::vectorizedMismatch_Type(),
5794 stubAddr, stubName, TypePtr::BOTTOM,
5795 obja_adr, objb_adr, length, scale);
5796
5797 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5798 set_result(result);
5799 return true;
5800 }
5801
5802 /**
5803 * Calculate CRC32 for byte.
5804 * int java.util.zip.CRC32.update(int crc, int b)
5805 */
5806 bool LibraryCallKit::inline_updateCRC32() {
5807 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5808 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5809 // no receiver since it is static method
5810 Node* crc = argument(0); // type: int
5811 Node* b = argument(1); // type: int
5812
5813 /*
5814 * int c = ~ crc;
5815 * b = timesXtoThe32[(b ^ c) & 0xFF];
5816 * b = b ^ (c >>> 8);
5817 * crc = ~b;
5818 */
5819
5820 Node* M1 = intcon(-1);
5821 crc = _gvn.transform(new XorINode(crc, M1));
5822 Node* result = _gvn.transform(new XorINode(crc, b));
5823 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5824
5825 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5826 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5827 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5828 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5829
5830 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5831 result = _gvn.transform(new XorINode(crc, result));
5832 result = _gvn.transform(new XorINode(result, M1));
5833 set_result(result);
5834 return true;
5835 }
5836
5837 /**
5838 * Calculate CRC32 for byte[] array.
5839 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5840 */
5841 bool LibraryCallKit::inline_updateBytesCRC32() {
5842 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5843 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5844 // no receiver since it is static method
5845 Node* crc = argument(0); // type: int
5846 Node* src = argument(1); // type: oop
5847 Node* offset = argument(2); // type: int
5848 Node* length = argument(3); // type: int
5849
5850 const Type* src_type = src->Value(&_gvn);
5851 const TypeAryPtr* top_src = src_type->isa_aryptr();
5852 if (top_src == NULL || top_src->klass() == NULL) {
5853 // failed array check
5854 return false;
5855 }
5856
5857 // Figure out the size and type of the elements we will be copying.
5858 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5859 if (src_elem != T_BYTE) {
5860 return false;
5861 }
5862
5863 // 'src_start' points to src array + scaled offset
5864 Node* src_start = array_element_address(src, offset, src_elem);
5865
5866 // We assume that range check is done by caller.
5867 // TODO: generate range check (offset+length < src.length) in debug VM.
5868
5869 // Call the stub.
5870 address stubAddr = StubRoutines::updateBytesCRC32();
5871 const char *stubName = "updateBytesCRC32";
5872
5873 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5874 stubAddr, stubName, TypePtr::BOTTOM,
5875 crc, src_start, length);
5876 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5877 set_result(result);
5878 return true;
5879 }
5880
5881 /**
5882 * Calculate CRC32 for ByteBuffer.
5883 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5884 */
5885 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5886 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5887 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5888 // no receiver since it is static method
5889 Node* crc = argument(0); // type: int
5890 Node* src = argument(1); // type: long
5891 Node* offset = argument(3); // type: int
5892 Node* length = argument(4); // type: int
5893
5894 src = ConvL2X(src); // adjust Java long to machine word
5895 Node* base = _gvn.transform(new CastX2PNode(src));
5896 offset = ConvI2X(offset);
5897
5898 // 'src_start' points to src array + scaled offset
5899 Node* src_start = basic_plus_adr(top(), base, offset);
5900
5901 // Call the stub.
5902 address stubAddr = StubRoutines::updateBytesCRC32();
5903 const char *stubName = "updateBytesCRC32";
5904
5905 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5906 stubAddr, stubName, TypePtr::BOTTOM,
5907 crc, src_start, length);
5908 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5909 set_result(result);
5910 return true;
5911 }
5912
5913 //------------------------------get_table_from_crc32c_class-----------------------
5914 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5915 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5916 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5917
5918 return table;
5919 }
5920
5921 //------------------------------inline_updateBytesCRC32C-----------------------
5922 //
5923 // Calculate CRC32C for byte[] array.
5924 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5925 //
5926 bool LibraryCallKit::inline_updateBytesCRC32C() {
5927 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5928 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5929 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5930 // no receiver since it is a static method
5931 Node* crc = argument(0); // type: int
5932 Node* src = argument(1); // type: oop
5933 Node* offset = argument(2); // type: int
5934 Node* end = argument(3); // type: int
5935
5936 Node* length = _gvn.transform(new SubINode(end, offset));
5937
5938 const Type* src_type = src->Value(&_gvn);
5939 const TypeAryPtr* top_src = src_type->isa_aryptr();
5940 if (top_src == NULL || top_src->klass() == NULL) {
5941 // failed array check
5942 return false;
5943 }
5944
5945 // Figure out the size and type of the elements we will be copying.
5946 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5947 if (src_elem != T_BYTE) {
5948 return false;
5949 }
5950
5951 // 'src_start' points to src array + scaled offset
5952 Node* src_start = array_element_address(src, offset, src_elem);
5953
5954 // static final int[] byteTable in class CRC32C
5955 Node* table = get_table_from_crc32c_class(callee()->holder());
5956 Node* table_start = array_element_address(table, intcon(0), T_INT);
5957
5958 // We assume that range check is done by caller.
5959 // TODO: generate range check (offset+length < src.length) in debug VM.
5960
5961 // Call the stub.
5962 address stubAddr = StubRoutines::updateBytesCRC32C();
5963 const char *stubName = "updateBytesCRC32C";
5964
5965 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5966 stubAddr, stubName, TypePtr::BOTTOM,
5967 crc, src_start, length, table_start);
5968 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5969 set_result(result);
5970 return true;
5971 }
5972
5973 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5974 //
5975 // Calculate CRC32C for DirectByteBuffer.
5976 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5977 //
5978 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5979 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5980 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5981 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5982 // no receiver since it is a static method
5983 Node* crc = argument(0); // type: int
5984 Node* src = argument(1); // type: long
5985 Node* offset = argument(3); // type: int
5986 Node* end = argument(4); // type: int
5987
5988 Node* length = _gvn.transform(new SubINode(end, offset));
5989
5990 src = ConvL2X(src); // adjust Java long to machine word
5991 Node* base = _gvn.transform(new CastX2PNode(src));
5992 offset = ConvI2X(offset);
5993
5994 // 'src_start' points to src array + scaled offset
5995 Node* src_start = basic_plus_adr(top(), base, offset);
5996
5997 // static final int[] byteTable in class CRC32C
5998 Node* table = get_table_from_crc32c_class(callee()->holder());
5999 Node* table_start = array_element_address(table, intcon(0), T_INT);
6000
6001 // Call the stub.
6002 address stubAddr = StubRoutines::updateBytesCRC32C();
6003 const char *stubName = "updateBytesCRC32C";
6004
6005 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6006 stubAddr, stubName, TypePtr::BOTTOM,
6007 crc, src_start, length, table_start);
6008 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6009 set_result(result);
6010 return true;
6011 }
6012
6013 //------------------------------inline_updateBytesAdler32----------------------
6014 //
6015 // Calculate Adler32 checksum for byte[] array.
6016 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6017 //
6018 bool LibraryCallKit::inline_updateBytesAdler32() {
6019 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6020 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6021 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6022 // no receiver since it is static method
6023 Node* crc = argument(0); // type: int
6024 Node* src = argument(1); // type: oop
6025 Node* offset = argument(2); // type: int
6026 Node* length = argument(3); // type: int
6027
6028 const Type* src_type = src->Value(&_gvn);
6029 const TypeAryPtr* top_src = src_type->isa_aryptr();
6030 if (top_src == NULL || top_src->klass() == NULL) {
6031 // failed array check
6032 return false;
6033 }
6034
6035 // Figure out the size and type of the elements we will be copying.
6036 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6037 if (src_elem != T_BYTE) {
6038 return false;
6039 }
6040
6041 // 'src_start' points to src array + scaled offset
6042 Node* src_start = array_element_address(src, offset, src_elem);
6043
6044 // We assume that range check is done by caller.
6045 // TODO: generate range check (offset+length < src.length) in debug VM.
6046
6047 // Call the stub.
6048 address stubAddr = StubRoutines::updateBytesAdler32();
6049 const char *stubName = "updateBytesAdler32";
6050
6051 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6052 stubAddr, stubName, TypePtr::BOTTOM,
6053 crc, src_start, length);
6054 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6055 set_result(result);
6056 return true;
6057 }
6058
6059 //------------------------------inline_updateByteBufferAdler32---------------
6060 //
6061 // Calculate Adler32 checksum for DirectByteBuffer.
6062 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6063 //
6064 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6065 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6066 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6067 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6068 // no receiver since it is static method
6069 Node* crc = argument(0); // type: int
6070 Node* src = argument(1); // type: long
6071 Node* offset = argument(3); // type: int
6072 Node* length = argument(4); // type: int
6073
6074 src = ConvL2X(src); // adjust Java long to machine word
6075 Node* base = _gvn.transform(new CastX2PNode(src));
6076 offset = ConvI2X(offset);
6077
6078 // 'src_start' points to src array + scaled offset
6079 Node* src_start = basic_plus_adr(top(), base, offset);
6080
6081 // Call the stub.
6082 address stubAddr = StubRoutines::updateBytesAdler32();
6083 const char *stubName = "updateBytesAdler32";
6084
6085 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6086 stubAddr, stubName, TypePtr::BOTTOM,
6087 crc, src_start, length);
6088
6089 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6090 set_result(result);
6091 return true;
6092 }
6093
6094 //----------------------------inline_reference_get----------------------------
6095 // public T java.lang.ref.Reference.get();
6096 bool LibraryCallKit::inline_reference_get() {
6097 const int referent_offset = java_lang_ref_Reference::referent_offset;
6098 guarantee(referent_offset > 0, "should have already been set");
6099
6100 // Get the argument:
6101 Node* reference_obj = null_check_receiver();
6102 if (stopped()) return true;
6103
6104 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6105
6106 ciInstanceKlass* klass = env()->Object_klass();
6107 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6108
6109 Node* no_ctrl = NULL;
6110 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6111
6112 // Use the pre-barrier to record the value in the referent field
6113 pre_barrier(false /* do_load */,
6114 control(),
6115 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6116 result /* pre_val */,
6117 T_OBJECT);
6118
6119 // Add memory barrier to prevent commoning reads from this field
6120 // across safepoint since GC can change its value.
6121 insert_mem_bar(Op_MemBarCPUOrder);
6122
6123 set_result(result);
6124 return true;
6125 }
6126
6127
6128 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6129 bool is_exact=true, bool is_static=false,
6130 ciInstanceKlass * fromKls=NULL) {
6131 if (fromKls == NULL) {
6132 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6133 assert(tinst != NULL, "obj is null");
6134 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6135 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6136 fromKls = tinst->klass()->as_instance_klass();
6137 } else {
6138 assert(is_static, "only for static field access");
6139 }
6140 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6141 ciSymbol::make(fieldTypeString),
6142 is_static);
6143
6144 assert (field != NULL, "undefined field");
6145 if (field == NULL) return (Node *) NULL;
6146
6147 if (is_static) {
6148 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6149 fromObj = makecon(tip);
6150 }
6151
6152 // Next code copied from Parse::do_get_xxx():
6153
6154 // Compute address and memory type.
6155 int offset = field->offset_in_bytes();
6156 bool is_vol = field->is_volatile();
6157 ciType* field_klass = field->type();
6158 assert(field_klass->is_loaded(), "should be loaded");
6159 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6160 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6161 BasicType bt = field->layout_type();
6162
6163 // Build the resultant type of the load
6164 const Type *type;
6165 if (bt == T_OBJECT) {
6166 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6167 } else {
6168 type = Type::get_const_basic_type(bt);
6169 }
6170
6171 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6172 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6173 }
6174 // Build the load.
6175 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6176 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6177 // If reference is volatile, prevent following memory ops from
6178 // floating up past the volatile read. Also prevents commoning
6179 // another volatile read.
6180 if (is_vol) {
6181 // Memory barrier includes bogus read of value to force load BEFORE membar
6182 insert_mem_bar(Op_MemBarAcquire, loadedField);
6183 }
6184 return loadedField;
6185 }
6186
6187 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6188 bool is_exact = true, bool is_static = false,
6189 ciInstanceKlass * fromKls = NULL) {
6190 if (fromKls == NULL) {
6191 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6192 assert(tinst != NULL, "obj is null");
6193 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6194 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6195 fromKls = tinst->klass()->as_instance_klass();
6196 }
6197 else {
6198 assert(is_static, "only for static field access");
6199 }
6200 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6201 ciSymbol::make(fieldTypeString),
6202 is_static);
6203
6204 assert(field != NULL, "undefined field");
6205 assert(!field->is_volatile(), "not defined for volatile fields");
6206
6207 if (is_static) {
6208 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6209 fromObj = makecon(tip);
6210 }
6211
6212 // Next code copied from Parse::do_get_xxx():
6213
6214 // Compute address and memory type.
6215 int offset = field->offset_in_bytes();
6216 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6217
6218 return adr;
6219 }
6220
6221 //------------------------------inline_aescrypt_Block-----------------------
6222 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6223 address stubAddr = NULL;
6224 const char *stubName;
6225 assert(UseAES, "need AES instruction support");
6226
6227 switch(id) {
6228 case vmIntrinsics::_aescrypt_encryptBlock:
6229 stubAddr = StubRoutines::aescrypt_encryptBlock();
6230 stubName = "aescrypt_encryptBlock";
6231 break;
6232 case vmIntrinsics::_aescrypt_decryptBlock:
6233 stubAddr = StubRoutines::aescrypt_decryptBlock();
6234 stubName = "aescrypt_decryptBlock";
6235 break;
6236 default:
6237 break;
6238 }
6239 if (stubAddr == NULL) return false;
6240
6241 Node* aescrypt_object = argument(0);
6242 Node* src = argument(1);
6243 Node* src_offset = argument(2);
6244 Node* dest = argument(3);
6245 Node* dest_offset = argument(4);
6246
6247 // (1) src and dest are arrays.
6248 const Type* src_type = src->Value(&_gvn);
6249 const Type* dest_type = dest->Value(&_gvn);
6250 const TypeAryPtr* top_src = src_type->isa_aryptr();
6251 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6252 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6253
6254 // for the quick and dirty code we will skip all the checks.
6255 // we are just trying to get the call to be generated.
6256 Node* src_start = src;
6257 Node* dest_start = dest;
6258 if (src_offset != NULL || dest_offset != NULL) {
6259 assert(src_offset != NULL && dest_offset != NULL, "");
6260 src_start = array_element_address(src, src_offset, T_BYTE);
6261 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6262 }
6263
6264 // now need to get the start of its expanded key array
6265 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6266 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6267 if (k_start == NULL) return false;
6268
6269 if (Matcher::pass_original_key_for_aes()) {
6270 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6271 // compatibility issues between Java key expansion and SPARC crypto instructions
6272 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6273 if (original_k_start == NULL) return false;
6274
6275 // Call the stub.
6276 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6277 stubAddr, stubName, TypePtr::BOTTOM,
6278 src_start, dest_start, k_start, original_k_start);
6279 } else {
6280 // Call the stub.
6281 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6282 stubAddr, stubName, TypePtr::BOTTOM,
6283 src_start, dest_start, k_start);
6284 }
6285
6286 return true;
6287 }
6288
6289 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6290 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6291 address stubAddr = NULL;
6292 const char *stubName = NULL;
6293
6294 assert(UseAES, "need AES instruction support");
6295
6296 switch(id) {
6297 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6298 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6299 stubName = "cipherBlockChaining_encryptAESCrypt";
6300 break;
6301 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6302 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6303 stubName = "cipherBlockChaining_decryptAESCrypt";
6304 break;
6305 default:
6306 break;
6307 }
6308 if (stubAddr == NULL) return false;
6309
6310 Node* cipherBlockChaining_object = argument(0);
6311 Node* src = argument(1);
6312 Node* src_offset = argument(2);
6313 Node* len = argument(3);
6314 Node* dest = argument(4);
6315 Node* dest_offset = argument(5);
6316
6317 // (1) src and dest are arrays.
6318 const Type* src_type = src->Value(&_gvn);
6319 const Type* dest_type = dest->Value(&_gvn);
6320 const TypeAryPtr* top_src = src_type->isa_aryptr();
6321 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6322 assert (top_src != NULL && top_src->klass() != NULL
6323 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6324
6325 // checks are the responsibility of the caller
6326 Node* src_start = src;
6327 Node* dest_start = dest;
6328 if (src_offset != NULL || dest_offset != NULL) {
6329 assert(src_offset != NULL && dest_offset != NULL, "");
6330 src_start = array_element_address(src, src_offset, T_BYTE);
6331 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6332 }
6333
6334 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6335 // (because of the predicated logic executed earlier).
6336 // so we cast it here safely.
6337 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6338
6339 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6340 if (embeddedCipherObj == NULL) return false;
6341
6342 // cast it to what we know it will be at runtime
6343 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6344 assert(tinst != NULL, "CBC obj is null");
6345 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6346 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6347 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6348
6349 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6350 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6351 const TypeOopPtr* xtype = aklass->as_instance_type();
6352 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6353 aescrypt_object = _gvn.transform(aescrypt_object);
6354
6355 // we need to get the start of the aescrypt_object's expanded key array
6356 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6357 if (k_start == NULL) return false;
6358
6359 // similarly, get the start address of the r vector
6360 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6361 if (objRvec == NULL) return false;
6362 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6363
6364 Node* cbcCrypt;
6365 if (Matcher::pass_original_key_for_aes()) {
6366 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6367 // compatibility issues between Java key expansion and SPARC crypto instructions
6368 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6369 if (original_k_start == NULL) return false;
6370
6371 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6372 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6373 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6374 stubAddr, stubName, TypePtr::BOTTOM,
6375 src_start, dest_start, k_start, r_start, len, original_k_start);
6376 } else {
6377 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6378 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6379 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6380 stubAddr, stubName, TypePtr::BOTTOM,
6381 src_start, dest_start, k_start, r_start, len);
6382 }
6383
6384 // return cipher length (int)
6385 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6386 set_result(retvalue);
6387 return true;
6388 }
6389
6390 //------------------------------inline_counterMode_AESCrypt-----------------------
6391 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6392 assert(UseAES, "need AES instruction support");
6393 if (!UseAESCTRIntrinsics) return false;
6394
6395 address stubAddr = NULL;
6396 const char *stubName = NULL;
6397 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6398 stubAddr = StubRoutines::counterMode_AESCrypt();
6399 stubName = "counterMode_AESCrypt";
6400 }
6401 if (stubAddr == NULL) return false;
6402
6403 Node* counterMode_object = argument(0);
6404 Node* src = argument(1);
6405 Node* src_offset = argument(2);
6406 Node* len = argument(3);
6407 Node* dest = argument(4);
6408 Node* dest_offset = argument(5);
6409
6410 // (1) src and dest are arrays.
6411 const Type* src_type = src->Value(&_gvn);
6412 const Type* dest_type = dest->Value(&_gvn);
6413 const TypeAryPtr* top_src = src_type->isa_aryptr();
6414 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6415 assert(top_src != NULL && top_src->klass() != NULL &&
6416 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6417
6418 // checks are the responsibility of the caller
6419 Node* src_start = src;
6420 Node* dest_start = dest;
6421 if (src_offset != NULL || dest_offset != NULL) {
6422 assert(src_offset != NULL && dest_offset != NULL, "");
6423 src_start = array_element_address(src, src_offset, T_BYTE);
6424 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6425 }
6426
6427 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6428 // (because of the predicated logic executed earlier).
6429 // so we cast it here safely.
6430 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6431 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6432 if (embeddedCipherObj == NULL) return false;
6433 // cast it to what we know it will be at runtime
6434 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6435 assert(tinst != NULL, "CTR obj is null");
6436 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6437 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6438 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6439 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6440 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6441 const TypeOopPtr* xtype = aklass->as_instance_type();
6442 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6443 aescrypt_object = _gvn.transform(aescrypt_object);
6444 // we need to get the start of the aescrypt_object's expanded key array
6445 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6446 if (k_start == NULL) return false;
6447 // similarly, get the start address of the r vector
6448 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6449 if (obj_counter == NULL) return false;
6450 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6451
6452 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6453 if (saved_encCounter == NULL) return false;
6454 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6455 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6456
6457 Node* ctrCrypt;
6458 if (Matcher::pass_original_key_for_aes()) {
6459 // no SPARC version for AES/CTR intrinsics now.
6460 return false;
6461 }
6462 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6463 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6464 OptoRuntime::counterMode_aescrypt_Type(),
6465 stubAddr, stubName, TypePtr::BOTTOM,
6466 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6467
6468 // return cipher length (int)
6469 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6470 set_result(retvalue);
6471 return true;
6472 }
6473
6474 //------------------------------get_key_start_from_aescrypt_object-----------------------
6475 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6476 #if defined(PPC64) || defined(S390)
6477 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6478 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6479 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6480 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6481 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6482 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6483 if (objSessionK == NULL) {
6484 return (Node *) NULL;
6485 }
6486 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6487 #else
6488 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6489 #endif // PPC64
6490 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6491 if (objAESCryptKey == NULL) return (Node *) NULL;
6492
6493 // now have the array, need to get the start address of the K array
6494 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6495 return k_start;
6496 }
6497
6498 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6499 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6500 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6501 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6502 if (objAESCryptKey == NULL) return (Node *) NULL;
6503
6504 // now have the array, need to get the start address of the lastKey array
6505 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6506 return original_k_start;
6507 }
6508
6509 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6510 // Return node representing slow path of predicate check.
6511 // the pseudo code we want to emulate with this predicate is:
6512 // for encryption:
6513 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6514 // for decryption:
6515 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6516 // note cipher==plain is more conservative than the original java code but that's OK
6517 //
6518 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6519 // The receiver was checked for NULL already.
6520 Node* objCBC = argument(0);
6521
6522 // Load embeddedCipher field of CipherBlockChaining object.
6523 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6524
6525 // get AESCrypt klass for instanceOf check
6526 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6527 // will have same classloader as CipherBlockChaining object
6528 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6529 assert(tinst != NULL, "CBCobj is null");
6530 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6531
6532 // we want to do an instanceof comparison against the AESCrypt class
6533 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6534 if (!klass_AESCrypt->is_loaded()) {
6535 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6536 Node* ctrl = control();
6537 set_control(top()); // no regular fast path
6538 return ctrl;
6539 }
6540 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6541
6542 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6543 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6544 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6545
6546 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6547
6548 // for encryption, we are done
6549 if (!decrypting)
6550 return instof_false; // even if it is NULL
6551
6552 // for decryption, we need to add a further check to avoid
6553 // taking the intrinsic path when cipher and plain are the same
6554 // see the original java code for why.
6555 RegionNode* region = new RegionNode(3);
6556 region->init_req(1, instof_false);
6557 Node* src = argument(1);
6558 Node* dest = argument(4);
6559 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6560 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6561 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6562 region->init_req(2, src_dest_conjoint);
6563
6564 record_for_igvn(region);
6565 return _gvn.transform(region);
6566 }
6567
6568 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6569 // Return node representing slow path of predicate check.
6570 // the pseudo code we want to emulate with this predicate is:
6571 // for encryption:
6572 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6573 // for decryption:
6574 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6575 // note cipher==plain is more conservative than the original java code but that's OK
6576 //
6577
6578 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6579 // The receiver was checked for NULL already.
6580 Node* objCTR = argument(0);
6581
6582 // Load embeddedCipher field of CipherBlockChaining object.
6583 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6584
6585 // get AESCrypt klass for instanceOf check
6586 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6587 // will have same classloader as CipherBlockChaining object
6588 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6589 assert(tinst != NULL, "CTRobj is null");
6590 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6591
6592 // we want to do an instanceof comparison against the AESCrypt class
6593 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6594 if (!klass_AESCrypt->is_loaded()) {
6595 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6596 Node* ctrl = control();
6597 set_control(top()); // no regular fast path
6598 return ctrl;
6599 }
6600
6601 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6602 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6603 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6604 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6605 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6606
6607 return instof_false; // even if it is NULL
6608 }
6609
6610 //------------------------------inline_ghash_processBlocks
6611 bool LibraryCallKit::inline_ghash_processBlocks() {
6612 address stubAddr;
6613 const char *stubName;
6614 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6615
6616 stubAddr = StubRoutines::ghash_processBlocks();
6617 stubName = "ghash_processBlocks";
6618
6619 Node* data = argument(0);
6620 Node* offset = argument(1);
6621 Node* len = argument(2);
6622 Node* state = argument(3);
6623 Node* subkeyH = argument(4);
6624
6625 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6626 assert(state_start, "state is NULL");
6627 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6628 assert(subkeyH_start, "subkeyH is NULL");
6629 Node* data_start = array_element_address(data, offset, T_BYTE);
6630 assert(data_start, "data is NULL");
6631
6632 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6633 OptoRuntime::ghash_processBlocks_Type(),
6634 stubAddr, stubName, TypePtr::BOTTOM,
6635 state_start, subkeyH_start, data_start, len);
6636 return true;
6637 }
6638
6639 bool LibraryCallKit::inline_base64_encodeBlock() {
6640 address stubAddr;
6641 const char *stubName;
6642 assert(UseBASE64Intrinsics, "need Base64 intrinsics support");
6643 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters");
6644 stubAddr = StubRoutines::base64_encodeBlock();
6645 stubName = "encodeBlock";
6646
6647 if (!stubAddr) return false;
6648 Node* base64obj = argument(0);
6649 Node* src = argument(1);
6650 Node* offset = argument(2);
6651 Node* len = argument(3);
6652 Node* dest = argument(4);
6653 Node* dp = argument(5);
6654 Node* isURL = argument(6);
6655
6656 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
6657 assert(src_start, "source array is NULL");
6658 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE);
6659 assert(dest_start, "destination array is NULL");
6660
6661 Node* base64 = make_runtime_call(RC_LEAF,
6662 OptoRuntime::base64_encodeBlock_Type(),
6663 stubAddr, stubName, TypePtr::BOTTOM,
6664 src_start, offset, len, dest_start, dp, isURL);
6665 return true;
6666 }
6667
6668 //------------------------------inline_sha_implCompress-----------------------
6669 //
6670 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6671 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6672 //
6673 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6674 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6675 //
6676 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6677 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6678 //
6679 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6680 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6681
6682 Node* sha_obj = argument(0);
6683 Node* src = argument(1); // type oop
6684 Node* ofs = argument(2); // type int
6685
6686 const Type* src_type = src->Value(&_gvn);
6687 const TypeAryPtr* top_src = src_type->isa_aryptr();
6688 if (top_src == NULL || top_src->klass() == NULL) {
6689 // failed array check
6690 return false;
6691 }
6692 // Figure out the size and type of the elements we will be copying.
6693 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6694 if (src_elem != T_BYTE) {
6695 return false;
6696 }
6697 // 'src_start' points to src array + offset
6698 Node* src_start = array_element_address(src, ofs, src_elem);
6699 Node* state = NULL;
6700 address stubAddr;
6701 const char *stubName;
6702
6703 switch(id) {
6704 case vmIntrinsics::_sha_implCompress:
6705 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6706 state = get_state_from_sha_object(sha_obj);
6707 stubAddr = StubRoutines::sha1_implCompress();
6708 stubName = "sha1_implCompress";
6709 break;
6710 case vmIntrinsics::_sha2_implCompress:
6711 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6712 state = get_state_from_sha_object(sha_obj);
6713 stubAddr = StubRoutines::sha256_implCompress();
6714 stubName = "sha256_implCompress";
6715 break;
6716 case vmIntrinsics::_sha5_implCompress:
6717 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6718 state = get_state_from_sha5_object(sha_obj);
6719 stubAddr = StubRoutines::sha512_implCompress();
6720 stubName = "sha512_implCompress";
6721 break;
6722 default:
6723 fatal_unexpected_iid(id);
6724 return false;
6725 }
6726 if (state == NULL) return false;
6727
6728 // Call the stub.
6729 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6730 stubAddr, stubName, TypePtr::BOTTOM,
6731 src_start, state);
6732
6733 return true;
6734 }
6735
6736 //------------------------------inline_digestBase_implCompressMB-----------------------
6737 //
6738 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6739 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6740 //
6741 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6742 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6743 "need SHA1/SHA256/SHA512 instruction support");
6744 assert((uint)predicate < 3, "sanity");
6745 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6746
6747 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6748 Node* src = argument(1); // byte[] array
6749 Node* ofs = argument(2); // type int
6750 Node* limit = argument(3); // type int
6751
6752 const Type* src_type = src->Value(&_gvn);
6753 const TypeAryPtr* top_src = src_type->isa_aryptr();
6754 if (top_src == NULL || top_src->klass() == NULL) {
6755 // failed array check
6756 return false;
6757 }
6758 // Figure out the size and type of the elements we will be copying.
6759 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6760 if (src_elem != T_BYTE) {
6761 return false;
6762 }
6763 // 'src_start' points to src array + offset
6764 Node* src_start = array_element_address(src, ofs, src_elem);
6765
6766 const char* klass_SHA_name = NULL;
6767 const char* stub_name = NULL;
6768 address stub_addr = NULL;
6769 bool long_state = false;
6770
6771 switch (predicate) {
6772 case 0:
6773 if (UseSHA1Intrinsics) {
6774 klass_SHA_name = "sun/security/provider/SHA";
6775 stub_name = "sha1_implCompressMB";
6776 stub_addr = StubRoutines::sha1_implCompressMB();
6777 }
6778 break;
6779 case 1:
6780 if (UseSHA256Intrinsics) {
6781 klass_SHA_name = "sun/security/provider/SHA2";
6782 stub_name = "sha256_implCompressMB";
6783 stub_addr = StubRoutines::sha256_implCompressMB();
6784 }
6785 break;
6786 case 2:
6787 if (UseSHA512Intrinsics) {
6788 klass_SHA_name = "sun/security/provider/SHA5";
6789 stub_name = "sha512_implCompressMB";
6790 stub_addr = StubRoutines::sha512_implCompressMB();
6791 long_state = true;
6792 }
6793 break;
6794 default:
6795 fatal("unknown SHA intrinsic predicate: %d", predicate);
6796 }
6797 if (klass_SHA_name != NULL) {
6798 // get DigestBase klass to lookup for SHA klass
6799 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6800 assert(tinst != NULL, "digestBase_obj is not instance???");
6801 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6802
6803 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6804 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6805 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6806 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6807 }
6808 return false;
6809 }
6810 //------------------------------inline_sha_implCompressMB-----------------------
6811 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6812 bool long_state, address stubAddr, const char *stubName,
6813 Node* src_start, Node* ofs, Node* limit) {
6814 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6815 const TypeOopPtr* xtype = aklass->as_instance_type();
6816 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6817 sha_obj = _gvn.transform(sha_obj);
6818
6819 Node* state;
6820 if (long_state) {
6821 state = get_state_from_sha5_object(sha_obj);
6822 } else {
6823 state = get_state_from_sha_object(sha_obj);
6824 }
6825 if (state == NULL) return false;
6826
6827 // Call the stub.
6828 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6829 OptoRuntime::digestBase_implCompressMB_Type(),
6830 stubAddr, stubName, TypePtr::BOTTOM,
6831 src_start, state, ofs, limit);
6832 // return ofs (int)
6833 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6834 set_result(result);
6835
6836 return true;
6837 }
6838
6839 //------------------------------get_state_from_sha_object-----------------------
6840 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6841 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6842 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6843 if (sha_state == NULL) return (Node *) NULL;
6844
6845 // now have the array, need to get the start address of the state array
6846 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6847 return state;
6848 }
6849
6850 //------------------------------get_state_from_sha5_object-----------------------
6851 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6852 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6853 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6854 if (sha_state == NULL) return (Node *) NULL;
6855
6856 // now have the array, need to get the start address of the state array
6857 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6858 return state;
6859 }
6860
6861 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6862 // Return node representing slow path of predicate check.
6863 // the pseudo code we want to emulate with this predicate is:
6864 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6865 //
6866 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6867 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6868 "need SHA1/SHA256/SHA512 instruction support");
6869 assert((uint)predicate < 3, "sanity");
6870
6871 // The receiver was checked for NULL already.
6872 Node* digestBaseObj = argument(0);
6873
6874 // get DigestBase klass for instanceOf check
6875 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6876 assert(tinst != NULL, "digestBaseObj is null");
6877 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6878
6879 const char* klass_SHA_name = NULL;
6880 switch (predicate) {
6881 case 0:
6882 if (UseSHA1Intrinsics) {
6883 // we want to do an instanceof comparison against the SHA class
6884 klass_SHA_name = "sun/security/provider/SHA";
6885 }
6886 break;
6887 case 1:
6888 if (UseSHA256Intrinsics) {
6889 // we want to do an instanceof comparison against the SHA2 class
6890 klass_SHA_name = "sun/security/provider/SHA2";
6891 }
6892 break;
6893 case 2:
6894 if (UseSHA512Intrinsics) {
6895 // we want to do an instanceof comparison against the SHA5 class
6896 klass_SHA_name = "sun/security/provider/SHA5";
6897 }
6898 break;
6899 default:
6900 fatal("unknown SHA intrinsic predicate: %d", predicate);
6901 }
6902
6903 ciKlass* klass_SHA = NULL;
6904 if (klass_SHA_name != NULL) {
6905 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6906 }
6907 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6908 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6909 Node* ctrl = control();
6910 set_control(top()); // no intrinsic path
6911 return ctrl;
6912 }
6913 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6914
6915 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6916 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6917 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6918 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6919
6920 return instof_false; // even if it is NULL
6921 }
6922
6923 //-------------inline_fma-----------------------------------
6924 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6925 Node *a = NULL;
6926 Node *b = NULL;
6927 Node *c = NULL;
6928 Node* result = NULL;
6929 switch (id) {
6930 case vmIntrinsics::_fmaD:
6931 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6932 // no receiver since it is static method
6933 a = round_double_node(argument(0));
6934 b = round_double_node(argument(2));
6935 c = round_double_node(argument(4));
6936 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6937 break;
6938 case vmIntrinsics::_fmaF:
6939 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6940 a = argument(0);
6941 b = argument(1);
6942 c = argument(2);
6943 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6944 break;
6945 default:
6946 fatal_unexpected_iid(id); break;
6947 }
6948 set_result(result);
6949 return true;
6950 }
6951
6952 bool LibraryCallKit::inline_profileBoolean() {
6953 Node* counts = argument(1);
6954 const TypeAryPtr* ary = NULL;
6955 ciArray* aobj = NULL;
6956 if (counts->is_Con()
6957 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6958 && (aobj = ary->const_oop()->as_array()) != NULL
6959 && (aobj->length() == 2)) {
6960 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6961 jint false_cnt = aobj->element_value(0).as_int();
6962 jint true_cnt = aobj->element_value(1).as_int();
6963
6964 if (C->log() != NULL) {
6965 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6966 false_cnt, true_cnt);
6967 }
6968
6969 if (false_cnt + true_cnt == 0) {
6970 // According to profile, never executed.
6971 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6972 Deoptimization::Action_reinterpret);
6973 return true;
6974 }
6975
6976 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6977 // is a number of each value occurrences.
6978 Node* result = argument(0);
6979 if (false_cnt == 0 || true_cnt == 0) {
6980 // According to profile, one value has been never seen.
6981 int expected_val = (false_cnt == 0) ? 1 : 0;
6982
6983 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6984 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6985
6986 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6987 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6988 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6989
6990 { // Slow path: uncommon trap for never seen value and then reexecute
6991 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6992 // the value has been seen at least once.
6993 PreserveJVMState pjvms(this);
6994 PreserveReexecuteState preexecs(this);
6995 jvms()->set_should_reexecute(true);
6996
6997 set_control(slow_path);
6998 set_i_o(i_o());
6999
7000 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7001 Deoptimization::Action_reinterpret);
7002 }
7003 // The guard for never seen value enables sharpening of the result and
7004 // returning a constant. It allows to eliminate branches on the same value
7005 // later on.
7006 set_control(fast_path);
7007 result = intcon(expected_val);
7008 }
7009 // Stop profiling.
7010 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
7011 // By replacing method body with profile data (represented as ProfileBooleanNode
7012 // on IR level) we effectively disable profiling.
7013 // It enables full speed execution once optimized code is generated.
7014 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
7015 C->record_for_igvn(profile);
7016 set_result(profile);
7017 return true;
7018 } else {
7019 // Continue profiling.
7020 // Profile data isn't available at the moment. So, execute method's bytecode version.
7021 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
7022 // is compiled and counters aren't available since corresponding MethodHandle
7023 // isn't a compile-time constant.
7024 return false;
7025 }
7026 }
7027
7028 bool LibraryCallKit::inline_isCompileConstant() {
7029 Node* n = argument(0);
7030 set_result(n->is_Con() ? intcon(1) : intcon(0));
7031 return true;
7032 }