1 /*
2 * Copyright (c) 1999, 2020, 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 "gc/shared/barrierSet.hpp"
33 #include "jfr/support/jfrIntrinsics.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/klass.inline.hpp"
36 #include "oops/objArrayKlass.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/arraycopynode.hpp"
39 #include "opto/c2compiler.hpp"
40 #include "opto/callGenerator.hpp"
41 #include "opto/castnode.hpp"
42 #include "opto/cfgnode.hpp"
43 #include "opto/convertnode.hpp"
44 #include "opto/countbitsnode.hpp"
45 #include "opto/intrinsicnode.hpp"
46 #include "opto/idealKit.hpp"
47 #include "opto/mathexactnode.hpp"
48 #include "opto/movenode.hpp"
49 #include "opto/mulnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/opaquenode.hpp"
52 #include "opto/parse.hpp"
53 #include "opto/runtime.hpp"
54 #include "opto/rootnode.hpp"
55 #include "opto/subnode.hpp"
56 #include "prims/nativeLookup.hpp"
57 #include "prims/unsafe.hpp"
58 #include "runtime/objectMonitor.hpp"
59 #include "runtime/sharedRuntime.hpp"
60 #include "utilities/macros.hpp"
61 #include "utilities/powerOfTwo.hpp"
62
63 class LibraryIntrinsic : public InlineCallGenerator {
64 // Extend the set of intrinsics known to the runtime:
65 public:
66 private:
67 bool _is_virtual;
68 bool _does_virtual_dispatch;
69 int8_t _predicates_count; // Intrinsic is predicated by several conditions
70 int8_t _last_predicate; // Last generated predicate
71 vmIntrinsics::ID _intrinsic_id;
72
73 public:
74 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
75 : InlineCallGenerator(m),
76 _is_virtual(is_virtual),
77 _does_virtual_dispatch(does_virtual_dispatch),
78 _predicates_count((int8_t)predicates_count),
79 _last_predicate((int8_t)-1),
80 _intrinsic_id(id)
81 {
82 }
83 virtual bool is_intrinsic() const { return true; }
84 virtual bool is_virtual() const { return _is_virtual; }
85 virtual bool is_predicated() const { return _predicates_count > 0; }
86 virtual int predicates_count() const { return _predicates_count; }
87 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
88 virtual JVMState* generate(JVMState* jvms);
89 virtual Node* generate_predicate(JVMState* jvms, int predicate);
90 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
91 };
92
93
94 // Local helper class for LibraryIntrinsic:
95 class LibraryCallKit : public GraphKit {
96 private:
97 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
98 Node* _result; // the result node, if any
99 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
100
101 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
102
103 public:
104 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
105 : GraphKit(jvms),
106 _intrinsic(intrinsic),
107 _result(NULL)
108 {
109 // Check if this is a root compile. In that case we don't have a caller.
110 if (!jvms->has_method()) {
111 _reexecute_sp = sp();
112 } else {
113 // Find out how many arguments the interpreter needs when deoptimizing
114 // and save the stack pointer value so it can used by uncommon_trap.
115 // We find the argument count by looking at the declared signature.
116 bool ignored_will_link;
117 ciSignature* declared_signature = NULL;
118 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
119 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
120 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
121 }
122 }
123
124 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
125
126 ciMethod* caller() const { return jvms()->method(); }
127 int bci() const { return jvms()->bci(); }
128 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
129 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
130 ciMethod* callee() const { return _intrinsic->method(); }
131
132 bool try_to_inline(int predicate);
133 Node* try_to_predicate(int predicate);
134
135 void push_result() {
136 // Push the result onto the stack.
137 if (!stopped() && result() != NULL) {
138 BasicType bt = result()->bottom_type()->basic_type();
139 push_node(bt, result());
140 }
141 }
142
143 private:
144 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
145 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
146 }
147
148 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
149 void set_result(RegionNode* region, PhiNode* value);
150 Node* result() { return _result; }
151
152 virtual int reexecute_sp() { return _reexecute_sp; }
153
154 // Helper functions to inline natives
155 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
156 Node* generate_slow_guard(Node* test, RegionNode* region);
157 Node* generate_fair_guard(Node* test, RegionNode* region);
158 Node* generate_negative_guard(Node* index, RegionNode* region,
159 // resulting CastII of index:
160 Node* *pos_index = NULL);
161 Node* generate_limit_guard(Node* offset, Node* subseq_length,
162 Node* array_length,
163 RegionNode* region);
164 void generate_string_range_check(Node* array, Node* offset,
165 Node* length, bool char_count);
166 Node* generate_current_thread(Node* &tls_output);
167 Node* load_mirror_from_klass(Node* klass);
168 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
169 RegionNode* region, int null_path,
170 int offset);
171 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
172 RegionNode* region, int null_path) {
173 int offset = java_lang_Class::klass_offset();
174 return load_klass_from_mirror_common(mirror, never_see_null,
175 region, null_path,
176 offset);
177 }
178 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
179 RegionNode* region, int null_path) {
180 int offset = java_lang_Class::array_klass_offset();
181 return load_klass_from_mirror_common(mirror, never_see_null,
182 region, null_path,
183 offset);
184 }
185 Node* generate_access_flags_guard(Node* kls,
186 int modifier_mask, int modifier_bits,
187 RegionNode* region);
188 Node* generate_interface_guard(Node* kls, RegionNode* region);
189 Node* generate_hidden_class_guard(Node* kls, RegionNode* region);
190 Node* generate_array_guard(Node* kls, RegionNode* region) {
191 return generate_array_guard_common(kls, region, false, false);
192 }
193 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
194 return generate_array_guard_common(kls, region, false, true);
195 }
196 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
197 return generate_array_guard_common(kls, region, true, false);
198 }
199 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
200 return generate_array_guard_common(kls, region, true, true);
201 }
202 Node* generate_array_guard_common(Node* kls, RegionNode* region,
203 bool obj_array, bool not_array);
204 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
205 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
206 bool is_virtual = false, bool is_static = false);
207 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
208 return generate_method_call(method_id, false, true);
209 }
210 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
211 return generate_method_call(method_id, true, false);
212 }
213 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
214 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
215
216 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
217 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
219 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
220 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
221 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
222 bool inline_string_indexOfChar();
223 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
224 bool inline_string_toBytesU();
225 bool inline_string_getCharsU();
226 bool inline_string_copy(bool compress);
227 bool inline_string_char_access(bool is_store);
228 Node* round_double_node(Node* n);
229 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
230 bool inline_math_native(vmIntrinsics::ID id);
231 bool inline_math(vmIntrinsics::ID id);
232 bool inline_double_math(vmIntrinsics::ID id);
233 template <typename OverflowOp>
234 bool inline_math_overflow(Node* arg1, Node* arg2);
235 void inline_math_mathExact(Node* math, Node* test);
236 bool inline_math_addExactI(bool is_increment);
237 bool inline_math_addExactL(bool is_increment);
238 bool inline_math_multiplyExactI();
239 bool inline_math_multiplyExactL();
240 bool inline_math_multiplyHigh();
241 bool inline_math_negateExactI();
242 bool inline_math_negateExactL();
243 bool inline_math_subtractExactI(bool is_decrement);
244 bool inline_math_subtractExactL(bool is_decrement);
245 bool inline_min_max(vmIntrinsics::ID id);
246 bool inline_notify(vmIntrinsics::ID id);
247 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
248 // This returns Type::AnyPtr, RawPtr, or OopPtr.
249 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
250 Node* make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type = T_ILLEGAL, bool can_cast = false);
251
252 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
253 DecoratorSet mo_decorator_for_access_kind(AccessKind kind);
254 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
255 static bool klass_needs_init_guard(Node* kls);
256 bool inline_unsafe_allocate();
257 bool inline_unsafe_newArray(bool uninitialized);
258 bool inline_unsafe_writeback0();
259 bool inline_unsafe_writebackSync0(bool is_pre);
260 bool inline_unsafe_copyMemory();
261 bool inline_native_currentThread();
262
263 bool inline_native_time_funcs(address method, const char* funcName);
264 #ifdef JFR_HAVE_INTRINSICS
265 bool inline_native_classID();
266 bool inline_native_getEventWriter();
267 #endif
268 bool inline_native_Class_query(vmIntrinsics::ID id);
269 bool inline_native_subtype_check();
270 bool inline_native_getLength();
271 bool inline_array_copyOf(bool is_copyOfRange);
272 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
273 bool inline_preconditions_checkIndex();
274 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array);
275 bool inline_native_clone(bool is_virtual);
276 bool inline_native_Reflection_getCallerClass();
277 // Helper function for inlining native object hash method
278 bool inline_native_hashcode(bool is_virtual, bool is_static);
279 bool inline_native_getClass();
280
281 // Helper functions for inlining arraycopy
282 bool inline_arraycopy();
283 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
284 RegionNode* slow_region);
285 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
286 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
287 uint new_idx);
288
289 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
290 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
291 bool inline_unsafe_fence(vmIntrinsics::ID id);
292 bool inline_onspinwait();
293 bool inline_fp_conversions(vmIntrinsics::ID id);
294 bool inline_number_methods(vmIntrinsics::ID id);
295 bool inline_reference_get();
296 bool inline_Class_cast();
297 bool inline_aescrypt_Block(vmIntrinsics::ID id);
298 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
299 bool inline_electronicCodeBook_AESCrypt(vmIntrinsics::ID id);
300 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
301 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
302 Node* inline_electronicCodeBook_AESCrypt_predicate(bool decrypting);
303 Node* inline_counterMode_AESCrypt_predicate();
304 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
305 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
306 bool inline_ghash_processBlocks();
307 bool inline_base64_encodeBlock();
308 bool inline_digestBase_implCompress(vmIntrinsics::ID id);
309 bool inline_digestBase_implCompressMB(int predicate);
310 bool inline_digestBase_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass,
311 bool long_state, address stubAddr, const char *stubName,
312 Node* src_start, Node* ofs, Node* limit);
313 Node* get_state_from_digest_object(Node *digestBase_object);
314 Node* get_long_state_from_digest_object(Node *digestBase_object);
315 Node* inline_digestBase_implCompressMB_predicate(int predicate);
316 bool inline_encodeISOArray();
317 bool inline_updateCRC32();
318 bool inline_updateBytesCRC32();
319 bool inline_updateByteBufferCRC32();
320 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
321 bool inline_updateBytesCRC32C();
322 bool inline_updateDirectByteBufferCRC32C();
323 bool inline_updateBytesAdler32();
324 bool inline_updateByteBufferAdler32();
325 bool inline_multiplyToLen();
326 bool inline_hasNegatives();
327 bool inline_squareToLen();
328 bool inline_mulAdd();
329 bool inline_montgomeryMultiply();
330 bool inline_montgomerySquare();
331 bool inline_bigIntegerShift(bool isRightShift);
332 bool inline_vectorizedMismatch();
333 bool inline_fma(vmIntrinsics::ID id);
334 bool inline_character_compare(vmIntrinsics::ID id);
335 bool inline_fp_min_max(vmIntrinsics::ID id);
336
337 bool inline_addressOf();
338 bool inline_sizeOf();
339 bool inline_getReferencedObjects();
340
341 bool inline_profileBoolean();
342 bool inline_isCompileConstant();
343 void clear_upper_avx() {
344 #ifdef X86
345 if (UseAVX >= 2) {
346 C->set_clear_upper_avx(true);
347 }
348 #endif
349 }
350 };
351
352 //---------------------------make_vm_intrinsic----------------------------
353 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
354 vmIntrinsics::ID id = m->intrinsic_id();
355 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
356
357 if (!m->is_loaded()) {
358 // Do not attempt to inline unloaded methods.
359 return NULL;
360 }
361
362 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
363 bool is_available = false;
364
365 {
366 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
367 // the compiler must transition to '_thread_in_vm' state because both
368 // methods access VM-internal data.
369 VM_ENTRY_MARK;
370 methodHandle mh(THREAD, m->get_Method());
371 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
372 !C->directive()->is_intrinsic_disabled(mh) &&
373 !vmIntrinsics::is_disabled_by_flags(mh);
374
375 }
376
377 if (is_available) {
378 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
379 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
380 return new LibraryIntrinsic(m, is_virtual,
381 vmIntrinsics::predicates_needed(id),
382 vmIntrinsics::does_virtual_dispatch(id),
383 (vmIntrinsics::ID) id);
384 } else {
385 return NULL;
386 }
387 }
388
389 //----------------------register_library_intrinsics-----------------------
390 // Initialize this file's data structures, for each Compile instance.
391 void Compile::register_library_intrinsics() {
392 // Nothing to do here.
393 }
394
395 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
396 LibraryCallKit kit(jvms, this);
397 Compile* C = kit.C;
398 int nodes = C->unique();
399 #ifndef PRODUCT
400 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
401 char buf[1000];
402 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
403 tty->print_cr("Intrinsic %s", str);
404 }
405 #endif
406 ciMethod* callee = kit.callee();
407 const int bci = kit.bci();
408
409 // Try to inline the intrinsic.
410 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
411 kit.try_to_inline(_last_predicate)) {
412 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
413 : "(intrinsic)";
414 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
415 if (C->print_intrinsics() || C->print_inlining()) {
416 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
417 }
418 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
419 if (C->log()) {
420 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
421 vmIntrinsics::name_at(intrinsic_id()),
422 (is_virtual() ? " virtual='1'" : ""),
423 C->unique() - nodes);
424 }
425 // Push the result from the inlined method onto the stack.
426 kit.push_result();
427 C->print_inlining_update(this);
428 return kit.transfer_exceptions_into_jvms();
429 }
430
431 // The intrinsic bailed out
432 if (jvms->has_method()) {
433 // Not a root compile.
434 const char* msg;
435 if (callee->intrinsic_candidate()) {
436 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
437 } else {
438 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
439 : "failed to inline (intrinsic), method not annotated";
440 }
441 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
442 if (C->print_intrinsics() || C->print_inlining()) {
443 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
444 }
445 } else {
446 // Root compile
447 ResourceMark rm;
448 stringStream msg_stream;
449 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
450 vmIntrinsics::name_at(intrinsic_id()),
451 is_virtual() ? " (virtual)" : "", bci);
452 const char *msg = msg_stream.as_string();
453 log_debug(jit, inlining)("%s", msg);
454 if (C->print_intrinsics() || C->print_inlining()) {
455 tty->print("%s", msg);
456 }
457 }
458 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
459 C->print_inlining_update(this);
460 return NULL;
461 }
462
463 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
464 LibraryCallKit kit(jvms, this);
465 Compile* C = kit.C;
466 int nodes = C->unique();
467 _last_predicate = predicate;
468 #ifndef PRODUCT
469 assert(is_predicated() && predicate < predicates_count(), "sanity");
470 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
471 char buf[1000];
472 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
473 tty->print_cr("Predicate for intrinsic %s", str);
474 }
475 #endif
476 ciMethod* callee = kit.callee();
477 const int bci = kit.bci();
478
479 Node* slow_ctl = kit.try_to_predicate(predicate);
480 if (!kit.failing()) {
481 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
482 : "(intrinsic, predicate)";
483 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
484 if (C->print_intrinsics() || C->print_inlining()) {
485 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
486 }
487 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
488 if (C->log()) {
489 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
490 vmIntrinsics::name_at(intrinsic_id()),
491 (is_virtual() ? " virtual='1'" : ""),
492 C->unique() - nodes);
493 }
494 return slow_ctl; // Could be NULL if the check folds.
495 }
496
497 // The intrinsic bailed out
498 if (jvms->has_method()) {
499 // Not a root compile.
500 const char* msg = "failed to generate predicate for intrinsic";
501 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
502 if (C->print_intrinsics() || C->print_inlining()) {
503 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
504 }
505 } else {
506 // Root compile
507 ResourceMark rm;
508 stringStream msg_stream;
509 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
510 vmIntrinsics::name_at(intrinsic_id()),
511 is_virtual() ? " (virtual)" : "", bci);
512 const char *msg = msg_stream.as_string();
513 log_debug(jit, inlining)("%s", msg);
514 if (C->print_intrinsics() || C->print_inlining()) {
515 C->print_inlining_stream()->print("%s", msg);
516 }
517 }
518 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
519 return NULL;
520 }
521
522 bool LibraryCallKit::try_to_inline(int predicate) {
523 // Handle symbolic names for otherwise undistinguished boolean switches:
524 const bool is_store = true;
525 const bool is_compress = true;
526 const bool is_static = true;
527 const bool is_volatile = true;
528
529 if (!jvms()->has_method()) {
530 // Root JVMState has a null method.
531 assert(map()->memory()->Opcode() == Op_Parm, "");
532 // Insert the memory aliasing node
533 set_all_memory(reset_memory());
534 }
535 assert(merged_memory(), "");
536
537
538 switch (intrinsic_id()) {
539 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
540 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
541 case vmIntrinsics::_getClass: return inline_native_getClass();
542
543 case vmIntrinsics::_ceil:
544 case vmIntrinsics::_floor:
545 case vmIntrinsics::_rint:
546 case vmIntrinsics::_dsin:
547 case vmIntrinsics::_dcos:
548 case vmIntrinsics::_dtan:
549 case vmIntrinsics::_dabs:
550 case vmIntrinsics::_fabs:
551 case vmIntrinsics::_iabs:
552 case vmIntrinsics::_labs:
553 case vmIntrinsics::_datan2:
554 case vmIntrinsics::_dsqrt:
555 case vmIntrinsics::_dexp:
556 case vmIntrinsics::_dlog:
557 case vmIntrinsics::_dlog10:
558 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
559
560 case vmIntrinsics::_min:
561 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
562
563 case vmIntrinsics::_notify:
564 case vmIntrinsics::_notifyAll:
565 return inline_notify(intrinsic_id());
566
567 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
568 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
569 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
570 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
571 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
572 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
573 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
574 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
575 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
576 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
577 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
578 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
579 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
580
581 case vmIntrinsics::_arraycopy: return inline_arraycopy();
582
583 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
584 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
585 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
586 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
587
588 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
589 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
590 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
591 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
592 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
593 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
594 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
595
596 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
597 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
598
599 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
600 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
601 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
602 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
603
604 case vmIntrinsics::_compressStringC:
605 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
606 case vmIntrinsics::_inflateStringC:
607 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
608
609 case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
610 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
611 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
612 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
613 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
614 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
615 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
616 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
617 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
618
619 case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
620 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
621 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
622 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
623 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
624 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
625 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
626 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
627 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
628
629 case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
630 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
631 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
632 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
633 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
634 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
635 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
636 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
637 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
638
639 case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
640 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
641 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
642 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
643 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
644 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
645 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
646 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
647 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
648
649 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
650 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
651 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
652 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
653
654 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
655 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
656 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
657 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
658
659 case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
660 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
661 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
662 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
663 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
664 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
665 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
666 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
667 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
668
669 case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
670 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
671 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
672 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
673 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
674 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
675 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
676 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
677 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
678
679 case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
680 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
681 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
682 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
683 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
684 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
685 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
686 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
687 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
688
689 case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
690 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
691 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
692 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
693 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
694 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
695 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
696 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
697 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
698
699 case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
700 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
701 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
702 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
703 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
704
705 case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
706 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
707 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
708 case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
709 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
710 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
711 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
712 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
713 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
714 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
715 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
716 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
717 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
718 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
719 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
720 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
721 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
722 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
723 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
724 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
725
726 case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
727 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
728 case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
729 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
730 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
731 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
732 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
733 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
734 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
735 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
736 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
737 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
738 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
739 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
740 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
741
742 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
743 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
744 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
745 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
746
747 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
748 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
749 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
750 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
751 case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
752
753 case vmIntrinsics::_loadFence:
754 case vmIntrinsics::_storeFence:
755 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
756
757 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
758
759 case vmIntrinsics::_currentThread: return inline_native_currentThread();
760
761 #ifdef JFR_HAVE_INTRINSICS
762 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime");
763 case vmIntrinsics::_getClassId: return inline_native_classID();
764 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter();
765 #endif
766 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
767 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
768 case vmIntrinsics::_writeback0: return inline_unsafe_writeback0();
769 case vmIntrinsics::_writebackPreSync0: return inline_unsafe_writebackSync0(true);
770 case vmIntrinsics::_writebackPostSync0: return inline_unsafe_writebackSync0(false);
771 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
772 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
773 case vmIntrinsics::_getLength: return inline_native_getLength();
774 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
775 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
776 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
777 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
778 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
779 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
780
781 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
782 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
783
784 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
785
786 case vmIntrinsics::_isInstance:
787 case vmIntrinsics::_getModifiers:
788 case vmIntrinsics::_isInterface:
789 case vmIntrinsics::_isArray:
790 case vmIntrinsics::_isPrimitive:
791 case vmIntrinsics::_isHidden:
792 case vmIntrinsics::_getSuperclass:
793 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
794
795 case vmIntrinsics::_floatToRawIntBits:
796 case vmIntrinsics::_floatToIntBits:
797 case vmIntrinsics::_intBitsToFloat:
798 case vmIntrinsics::_doubleToRawLongBits:
799 case vmIntrinsics::_doubleToLongBits:
800 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
801
802 case vmIntrinsics::_numberOfLeadingZeros_i:
803 case vmIntrinsics::_numberOfLeadingZeros_l:
804 case vmIntrinsics::_numberOfTrailingZeros_i:
805 case vmIntrinsics::_numberOfTrailingZeros_l:
806 case vmIntrinsics::_bitCount_i:
807 case vmIntrinsics::_bitCount_l:
808 case vmIntrinsics::_reverseBytes_i:
809 case vmIntrinsics::_reverseBytes_l:
810 case vmIntrinsics::_reverseBytes_s:
811 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
812
813 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
814
815 case vmIntrinsics::_Reference_get: return inline_reference_get();
816
817 case vmIntrinsics::_Class_cast: return inline_Class_cast();
818
819 case vmIntrinsics::_aescrypt_encryptBlock:
820 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
821
822 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
823 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
824 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
825
826 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt:
827 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt:
828 return inline_electronicCodeBook_AESCrypt(intrinsic_id());
829
830 case vmIntrinsics::_counterMode_AESCrypt:
831 return inline_counterMode_AESCrypt(intrinsic_id());
832
833 case vmIntrinsics::_md5_implCompress:
834 case vmIntrinsics::_sha_implCompress:
835 case vmIntrinsics::_sha2_implCompress:
836 case vmIntrinsics::_sha5_implCompress:
837 return inline_digestBase_implCompress(intrinsic_id());
838
839 case vmIntrinsics::_digestBase_implCompressMB:
840 return inline_digestBase_implCompressMB(predicate);
841
842 case vmIntrinsics::_multiplyToLen:
843 return inline_multiplyToLen();
844
845 case vmIntrinsics::_squareToLen:
846 return inline_squareToLen();
847
848 case vmIntrinsics::_mulAdd:
849 return inline_mulAdd();
850
851 case vmIntrinsics::_montgomeryMultiply:
852 return inline_montgomeryMultiply();
853 case vmIntrinsics::_montgomerySquare:
854 return inline_montgomerySquare();
855
856 case vmIntrinsics::_bigIntegerRightShiftWorker:
857 return inline_bigIntegerShift(true);
858 case vmIntrinsics::_bigIntegerLeftShiftWorker:
859 return inline_bigIntegerShift(false);
860
861 case vmIntrinsics::_vectorizedMismatch:
862 return inline_vectorizedMismatch();
863
864 case vmIntrinsics::_ghash_processBlocks:
865 return inline_ghash_processBlocks();
866 case vmIntrinsics::_base64_encodeBlock:
867 return inline_base64_encodeBlock();
868
869 case vmIntrinsics::_encodeISOArray:
870 case vmIntrinsics::_encodeByteISOArray:
871 return inline_encodeISOArray();
872
873 case vmIntrinsics::_updateCRC32:
874 return inline_updateCRC32();
875 case vmIntrinsics::_updateBytesCRC32:
876 return inline_updateBytesCRC32();
877 case vmIntrinsics::_updateByteBufferCRC32:
878 return inline_updateByteBufferCRC32();
879
880 case vmIntrinsics::_updateBytesCRC32C:
881 return inline_updateBytesCRC32C();
882 case vmIntrinsics::_updateDirectByteBufferCRC32C:
883 return inline_updateDirectByteBufferCRC32C();
884
885 case vmIntrinsics::_updateBytesAdler32:
886 return inline_updateBytesAdler32();
887 case vmIntrinsics::_updateByteBufferAdler32:
888 return inline_updateByteBufferAdler32();
889
890 case vmIntrinsics::_profileBoolean:
891 return inline_profileBoolean();
892 case vmIntrinsics::_isCompileConstant:
893 return inline_isCompileConstant();
894
895 case vmIntrinsics::_hasNegatives:
896 return inline_hasNegatives();
897
898 case vmIntrinsics::_fmaD:
899 case vmIntrinsics::_fmaF:
900 return inline_fma(intrinsic_id());
901
902 case vmIntrinsics::_isDigit:
903 case vmIntrinsics::_isLowerCase:
904 case vmIntrinsics::_isUpperCase:
905 case vmIntrinsics::_isWhitespace:
906 return inline_character_compare(intrinsic_id());
907
908 case vmIntrinsics::_maxF:
909 case vmIntrinsics::_minF:
910 case vmIntrinsics::_maxD:
911 case vmIntrinsics::_minD:
912 return inline_fp_min_max(intrinsic_id());
913
914 case vmIntrinsics::_sizeOf:
915 return inline_sizeOf();
916
917 case vmIntrinsics::_addressOf:
918 return inline_addressOf();
919
920 case vmIntrinsics::_getReferencedObjects:
921 return inline_getReferencedObjects();
922
923 default:
924 // If you get here, it may be that someone has added a new intrinsic
925 // to the list in vmSymbols.hpp without implementing it here.
926 #ifndef PRODUCT
927 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
928 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
929 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
930 }
931 #endif
932 return false;
933 }
934 }
935
936 Node* LibraryCallKit::try_to_predicate(int predicate) {
937 if (!jvms()->has_method()) {
938 // Root JVMState has a null method.
939 assert(map()->memory()->Opcode() == Op_Parm, "");
940 // Insert the memory aliasing node
941 set_all_memory(reset_memory());
942 }
943 assert(merged_memory(), "");
944
945 switch (intrinsic_id()) {
946 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
947 return inline_cipherBlockChaining_AESCrypt_predicate(false);
948 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
949 return inline_cipherBlockChaining_AESCrypt_predicate(true);
950 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt:
951 return inline_electronicCodeBook_AESCrypt_predicate(false);
952 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt:
953 return inline_electronicCodeBook_AESCrypt_predicate(true);
954 case vmIntrinsics::_counterMode_AESCrypt:
955 return inline_counterMode_AESCrypt_predicate();
956 case vmIntrinsics::_digestBase_implCompressMB:
957 return inline_digestBase_implCompressMB_predicate(predicate);
958
959 default:
960 // If you get here, it may be that someone has added a new intrinsic
961 // to the list in vmSymbols.hpp without implementing it here.
962 #ifndef PRODUCT
963 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
964 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
965 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
966 }
967 #endif
968 Node* slow_ctl = control();
969 set_control(top()); // No fast path instrinsic
970 return slow_ctl;
971 }
972 }
973
974 //------------------------------set_result-------------------------------
975 // Helper function for finishing intrinsics.
976 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
977 record_for_igvn(region);
978 set_control(_gvn.transform(region));
979 set_result( _gvn.transform(value));
980 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
981 }
982
983 //------------------------------generate_guard---------------------------
984 // Helper function for generating guarded fast-slow graph structures.
985 // The given 'test', if true, guards a slow path. If the test fails
986 // then a fast path can be taken. (We generally hope it fails.)
987 // In all cases, GraphKit::control() is updated to the fast path.
988 // The returned value represents the control for the slow path.
989 // The return value is never 'top'; it is either a valid control
990 // or NULL if it is obvious that the slow path can never be taken.
991 // Also, if region and the slow control are not NULL, the slow edge
992 // is appended to the region.
993 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
994 if (stopped()) {
995 // Already short circuited.
996 return NULL;
997 }
998
999 // Build an if node and its projections.
1000 // If test is true we take the slow path, which we assume is uncommon.
1001 if (_gvn.type(test) == TypeInt::ZERO) {
1002 // The slow branch is never taken. No need to build this guard.
1003 return NULL;
1004 }
1005
1006 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1007
1008 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1009 if (if_slow == top()) {
1010 // The slow branch is never taken. No need to build this guard.
1011 return NULL;
1012 }
1013
1014 if (region != NULL)
1015 region->add_req(if_slow);
1016
1017 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1018 set_control(if_fast);
1019
1020 return if_slow;
1021 }
1022
1023 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1024 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1025 }
1026 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1027 return generate_guard(test, region, PROB_FAIR);
1028 }
1029
1030 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1031 Node* *pos_index) {
1032 if (stopped())
1033 return NULL; // already stopped
1034 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1035 return NULL; // index is already adequately typed
1036 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1037 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1038 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1039 if (is_neg != NULL && pos_index != NULL) {
1040 // Emulate effect of Parse::adjust_map_after_if.
1041 Node* ccast = new CastIINode(index, TypeInt::POS);
1042 ccast->set_req(0, control());
1043 (*pos_index) = _gvn.transform(ccast);
1044 }
1045 return is_neg;
1046 }
1047
1048 // Make sure that 'position' is a valid limit index, in [0..length].
1049 // There are two equivalent plans for checking this:
1050 // A. (offset + copyLength) unsigned<= arrayLength
1051 // B. offset <= (arrayLength - copyLength)
1052 // We require that all of the values above, except for the sum and
1053 // difference, are already known to be non-negative.
1054 // Plan A is robust in the face of overflow, if offset and copyLength
1055 // are both hugely positive.
1056 //
1057 // Plan B is less direct and intuitive, but it does not overflow at
1058 // all, since the difference of two non-negatives is always
1059 // representable. Whenever Java methods must perform the equivalent
1060 // check they generally use Plan B instead of Plan A.
1061 // For the moment we use Plan A.
1062 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1063 Node* subseq_length,
1064 Node* array_length,
1065 RegionNode* region) {
1066 if (stopped())
1067 return NULL; // already stopped
1068 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1069 if (zero_offset && subseq_length->eqv_uncast(array_length))
1070 return NULL; // common case of whole-array copy
1071 Node* last = subseq_length;
1072 if (!zero_offset) // last += offset
1073 last = _gvn.transform(new AddINode(last, offset));
1074 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1075 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1076 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1077 return is_over;
1078 }
1079
1080 // Emit range checks for the given String.value byte array
1081 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1082 if (stopped()) {
1083 return; // already stopped
1084 }
1085 RegionNode* bailout = new RegionNode(1);
1086 record_for_igvn(bailout);
1087 if (char_count) {
1088 // Convert char count to byte count
1089 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1090 }
1091
1092 // Offset and count must not be negative
1093 generate_negative_guard(offset, bailout);
1094 generate_negative_guard(count, bailout);
1095 // Offset + count must not exceed length of array
1096 generate_limit_guard(offset, count, load_array_length(array), bailout);
1097
1098 if (bailout->req() > 1) {
1099 PreserveJVMState pjvms(this);
1100 set_control(_gvn.transform(bailout));
1101 uncommon_trap(Deoptimization::Reason_intrinsic,
1102 Deoptimization::Action_maybe_recompile);
1103 }
1104 }
1105
1106 //--------------------------generate_current_thread--------------------
1107 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1108 ciKlass* thread_klass = env()->Thread_klass();
1109 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1110 Node* thread = _gvn.transform(new ThreadLocalNode());
1111 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1112 tls_output = thread;
1113 Node* thread_obj_handle = LoadNode::make(_gvn, NULL, immutable_memory(), p, p->bottom_type()->is_ptr(), TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
1114 thread_obj_handle = _gvn.transform(thread_obj_handle);
1115 return access_load(thread_obj_handle, thread_type, T_OBJECT, IN_NATIVE | C2_IMMUTABLE_MEMORY);
1116 }
1117
1118
1119 //------------------------------make_string_method_node------------------------
1120 // Helper method for String intrinsic functions. This version is called with
1121 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1122 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1123 // containing the lengths of str1 and str2.
1124 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1125 Node* result = NULL;
1126 switch (opcode) {
1127 case Op_StrIndexOf:
1128 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1129 str1_start, cnt1, str2_start, cnt2, ae);
1130 break;
1131 case Op_StrComp:
1132 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1133 str1_start, cnt1, str2_start, cnt2, ae);
1134 break;
1135 case Op_StrEquals:
1136 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1137 // Use the constant length if there is one because optimized match rule may exist.
1138 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1139 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1140 break;
1141 default:
1142 ShouldNotReachHere();
1143 return NULL;
1144 }
1145
1146 // All these intrinsics have checks.
1147 C->set_has_split_ifs(true); // Has chance for split-if optimization
1148 clear_upper_avx();
1149
1150 return _gvn.transform(result);
1151 }
1152
1153 //------------------------------inline_string_compareTo------------------------
1154 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1155 Node* arg1 = argument(0);
1156 Node* arg2 = argument(1);
1157
1158 arg1 = must_be_not_null(arg1, true);
1159 arg2 = must_be_not_null(arg2, true);
1160
1161 // Get start addr and length of first argument
1162 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1163 Node* arg1_cnt = load_array_length(arg1);
1164
1165 // Get start addr and length of second argument
1166 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1167 Node* arg2_cnt = load_array_length(arg2);
1168
1169 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1170 set_result(result);
1171 return true;
1172 }
1173
1174 //------------------------------inline_string_equals------------------------
1175 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1176 Node* arg1 = argument(0);
1177 Node* arg2 = argument(1);
1178
1179 // paths (plus control) merge
1180 RegionNode* region = new RegionNode(3);
1181 Node* phi = new PhiNode(region, TypeInt::BOOL);
1182
1183 if (!stopped()) {
1184
1185 arg1 = must_be_not_null(arg1, true);
1186 arg2 = must_be_not_null(arg2, true);
1187
1188 // Get start addr and length of first argument
1189 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1190 Node* arg1_cnt = load_array_length(arg1);
1191
1192 // Get start addr and length of second argument
1193 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1194 Node* arg2_cnt = load_array_length(arg2);
1195
1196 // Check for arg1_cnt != arg2_cnt
1197 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1198 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1199 Node* if_ne = generate_slow_guard(bol, NULL);
1200 if (if_ne != NULL) {
1201 phi->init_req(2, intcon(0));
1202 region->init_req(2, if_ne);
1203 }
1204
1205 // Check for count == 0 is done by assembler code for StrEquals.
1206
1207 if (!stopped()) {
1208 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1209 phi->init_req(1, equals);
1210 region->init_req(1, control());
1211 }
1212 }
1213
1214 // post merge
1215 set_control(_gvn.transform(region));
1216 record_for_igvn(region);
1217
1218 set_result(_gvn.transform(phi));
1219 return true;
1220 }
1221
1222 //------------------------------inline_array_equals----------------------------
1223 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1224 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1225 Node* arg1 = argument(0);
1226 Node* arg2 = argument(1);
1227
1228 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1229 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1230 clear_upper_avx();
1231
1232 return true;
1233 }
1234
1235 //------------------------------inline_hasNegatives------------------------------
1236 bool LibraryCallKit::inline_hasNegatives() {
1237 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1238 return false;
1239 }
1240
1241 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1242 // no receiver since it is static method
1243 Node* ba = argument(0);
1244 Node* offset = argument(1);
1245 Node* len = argument(2);
1246
1247 ba = must_be_not_null(ba, true);
1248
1249 // Range checks
1250 generate_string_range_check(ba, offset, len, false);
1251 if (stopped()) {
1252 return true;
1253 }
1254 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1255 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1256 set_result(_gvn.transform(result));
1257 return true;
1258 }
1259
1260 bool LibraryCallKit::inline_preconditions_checkIndex() {
1261 Node* index = argument(0);
1262 Node* length = argument(1);
1263 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1264 return false;
1265 }
1266
1267 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1268 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1269
1270 {
1271 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1272 uncommon_trap(Deoptimization::Reason_intrinsic,
1273 Deoptimization::Action_make_not_entrant);
1274 }
1275
1276 if (stopped()) {
1277 return false;
1278 }
1279
1280 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1281 BoolTest::mask btest = BoolTest::lt;
1282 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1283 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1284 _gvn.set_type(rc, rc->Value(&_gvn));
1285 if (!rc_bool->is_Con()) {
1286 record_for_igvn(rc);
1287 }
1288 set_control(_gvn.transform(new IfTrueNode(rc)));
1289 {
1290 PreserveJVMState pjvms(this);
1291 set_control(_gvn.transform(new IfFalseNode(rc)));
1292 uncommon_trap(Deoptimization::Reason_range_check,
1293 Deoptimization::Action_make_not_entrant);
1294 }
1295
1296 if (stopped()) {
1297 return false;
1298 }
1299
1300 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1301 result->set_req(0, control());
1302 result = _gvn.transform(result);
1303 set_result(result);
1304 replace_in_map(index, result);
1305 clear_upper_avx();
1306 return true;
1307 }
1308
1309 //------------------------------inline_string_indexOf------------------------
1310 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1311 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1312 return false;
1313 }
1314 Node* src = argument(0);
1315 Node* tgt = argument(1);
1316
1317 // Make the merge point
1318 RegionNode* result_rgn = new RegionNode(4);
1319 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1320
1321 src = must_be_not_null(src, true);
1322 tgt = must_be_not_null(tgt, true);
1323
1324 // Get start addr and length of source string
1325 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1326 Node* src_count = load_array_length(src);
1327
1328 // Get start addr and length of substring
1329 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1330 Node* tgt_count = load_array_length(tgt);
1331
1332 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1333 // Divide src size by 2 if String is UTF16 encoded
1334 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1335 }
1336 if (ae == StrIntrinsicNode::UU) {
1337 // Divide substring size by 2 if String is UTF16 encoded
1338 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1339 }
1340
1341 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1342 if (result != NULL) {
1343 result_phi->init_req(3, result);
1344 result_rgn->init_req(3, control());
1345 }
1346 set_control(_gvn.transform(result_rgn));
1347 record_for_igvn(result_rgn);
1348 set_result(_gvn.transform(result_phi));
1349
1350 return true;
1351 }
1352
1353 //-----------------------------inline_string_indexOf-----------------------
1354 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1355 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1356 return false;
1357 }
1358 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1359 return false;
1360 }
1361 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1362 Node* src = argument(0); // byte[]
1363 Node* src_count = argument(1); // char count
1364 Node* tgt = argument(2); // byte[]
1365 Node* tgt_count = argument(3); // char count
1366 Node* from_index = argument(4); // char index
1367
1368 src = must_be_not_null(src, true);
1369 tgt = must_be_not_null(tgt, true);
1370
1371 // Multiply byte array index by 2 if String is UTF16 encoded
1372 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1373 src_count = _gvn.transform(new SubINode(src_count, from_index));
1374 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1375 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1376
1377 // Range checks
1378 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1379 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1380 if (stopped()) {
1381 return true;
1382 }
1383
1384 RegionNode* region = new RegionNode(5);
1385 Node* phi = new PhiNode(region, TypeInt::INT);
1386
1387 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1388 if (result != NULL) {
1389 // The result is index relative to from_index if substring was found, -1 otherwise.
1390 // Generate code which will fold into cmove.
1391 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1392 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1393
1394 Node* if_lt = generate_slow_guard(bol, NULL);
1395 if (if_lt != NULL) {
1396 // result == -1
1397 phi->init_req(3, result);
1398 region->init_req(3, if_lt);
1399 }
1400 if (!stopped()) {
1401 result = _gvn.transform(new AddINode(result, from_index));
1402 phi->init_req(4, result);
1403 region->init_req(4, control());
1404 }
1405 }
1406
1407 set_control(_gvn.transform(region));
1408 record_for_igvn(region);
1409 set_result(_gvn.transform(phi));
1410 clear_upper_avx();
1411
1412 return true;
1413 }
1414
1415 // Create StrIndexOfNode with fast path checks
1416 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1417 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1418 // Check for substr count > string count
1419 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1420 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1421 Node* if_gt = generate_slow_guard(bol, NULL);
1422 if (if_gt != NULL) {
1423 phi->init_req(1, intcon(-1));
1424 region->init_req(1, if_gt);
1425 }
1426 if (!stopped()) {
1427 // Check for substr count == 0
1428 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1429 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1430 Node* if_zero = generate_slow_guard(bol, NULL);
1431 if (if_zero != NULL) {
1432 phi->init_req(2, intcon(0));
1433 region->init_req(2, if_zero);
1434 }
1435 }
1436 if (!stopped()) {
1437 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1438 }
1439 return NULL;
1440 }
1441
1442 //-----------------------------inline_string_indexOfChar-----------------------
1443 bool LibraryCallKit::inline_string_indexOfChar() {
1444 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1445 return false;
1446 }
1447 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1448 return false;
1449 }
1450 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1451 Node* src = argument(0); // byte[]
1452 Node* tgt = argument(1); // tgt is int ch
1453 Node* from_index = argument(2);
1454 Node* max = argument(3);
1455
1456 src = must_be_not_null(src, true);
1457
1458 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1459 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1460 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1461
1462 // Range checks
1463 generate_string_range_check(src, src_offset, src_count, true);
1464 if (stopped()) {
1465 return true;
1466 }
1467
1468 RegionNode* region = new RegionNode(3);
1469 Node* phi = new PhiNode(region, TypeInt::INT);
1470
1471 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1472 C->set_has_split_ifs(true); // Has chance for split-if optimization
1473 _gvn.transform(result);
1474
1475 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1476 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1477
1478 Node* if_lt = generate_slow_guard(bol, NULL);
1479 if (if_lt != NULL) {
1480 // result == -1
1481 phi->init_req(2, result);
1482 region->init_req(2, if_lt);
1483 }
1484 if (!stopped()) {
1485 result = _gvn.transform(new AddINode(result, from_index));
1486 phi->init_req(1, result);
1487 region->init_req(1, control());
1488 }
1489 set_control(_gvn.transform(region));
1490 record_for_igvn(region);
1491 set_result(_gvn.transform(phi));
1492
1493 return true;
1494 }
1495 //---------------------------inline_string_copy---------------------
1496 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1497 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1498 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1499 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1500 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1501 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1502 bool LibraryCallKit::inline_string_copy(bool compress) {
1503 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1504 return false;
1505 }
1506 int nargs = 5; // 2 oops, 3 ints
1507 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1508
1509 Node* src = argument(0);
1510 Node* src_offset = argument(1);
1511 Node* dst = argument(2);
1512 Node* dst_offset = argument(3);
1513 Node* length = argument(4);
1514
1515 // Check for allocation before we add nodes that would confuse
1516 // tightly_coupled_allocation()
1517 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1518
1519 // Figure out the size and type of the elements we will be copying.
1520 const Type* src_type = src->Value(&_gvn);
1521 const Type* dst_type = dst->Value(&_gvn);
1522 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1523 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1524 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1525 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1526 "Unsupported array types for inline_string_copy");
1527
1528 src = must_be_not_null(src, true);
1529 dst = must_be_not_null(dst, true);
1530
1531 // Convert char[] offsets to byte[] offsets
1532 bool convert_src = (compress && src_elem == T_BYTE);
1533 bool convert_dst = (!compress && dst_elem == T_BYTE);
1534 if (convert_src) {
1535 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1536 } else if (convert_dst) {
1537 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1538 }
1539
1540 // Range checks
1541 generate_string_range_check(src, src_offset, length, convert_src);
1542 generate_string_range_check(dst, dst_offset, length, convert_dst);
1543 if (stopped()) {
1544 return true;
1545 }
1546
1547 Node* src_start = array_element_address(src, src_offset, src_elem);
1548 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1549 // 'src_start' points to src array + scaled offset
1550 // 'dst_start' points to dst array + scaled offset
1551 Node* count = NULL;
1552 if (compress) {
1553 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1554 } else {
1555 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1556 }
1557
1558 if (alloc != NULL) {
1559 if (alloc->maybe_set_complete(&_gvn)) {
1560 // "You break it, you buy it."
1561 InitializeNode* init = alloc->initialization();
1562 assert(init->is_complete(), "we just did this");
1563 init->set_complete_with_arraycopy();
1564 assert(dst->is_CheckCastPP(), "sanity");
1565 assert(dst->in(0)->in(0) == init, "dest pinned");
1566 }
1567 // Do not let stores that initialize this object be reordered with
1568 // a subsequent store that would make this object accessible by
1569 // other threads.
1570 // Record what AllocateNode this StoreStore protects so that
1571 // escape analysis can go from the MemBarStoreStoreNode to the
1572 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1573 // based on the escape status of the AllocateNode.
1574 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1575 }
1576 if (compress) {
1577 set_result(_gvn.transform(count));
1578 }
1579 clear_upper_avx();
1580
1581 return true;
1582 }
1583
1584 #ifdef _LP64
1585 #define XTOP ,top() /*additional argument*/
1586 #else //_LP64
1587 #define XTOP /*no additional argument*/
1588 #endif //_LP64
1589
1590 //------------------------inline_string_toBytesU--------------------------
1591 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1592 bool LibraryCallKit::inline_string_toBytesU() {
1593 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1594 return false;
1595 }
1596 // Get the arguments.
1597 Node* value = argument(0);
1598 Node* offset = argument(1);
1599 Node* length = argument(2);
1600
1601 Node* newcopy = NULL;
1602
1603 // Set the original stack and the reexecute bit for the interpreter to reexecute
1604 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1605 { PreserveReexecuteState preexecs(this);
1606 jvms()->set_should_reexecute(true);
1607
1608 // Check if a null path was taken unconditionally.
1609 value = null_check(value);
1610
1611 RegionNode* bailout = new RegionNode(1);
1612 record_for_igvn(bailout);
1613
1614 // Range checks
1615 generate_negative_guard(offset, bailout);
1616 generate_negative_guard(length, bailout);
1617 generate_limit_guard(offset, length, load_array_length(value), bailout);
1618 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1619 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1620
1621 if (bailout->req() > 1) {
1622 PreserveJVMState pjvms(this);
1623 set_control(_gvn.transform(bailout));
1624 uncommon_trap(Deoptimization::Reason_intrinsic,
1625 Deoptimization::Action_maybe_recompile);
1626 }
1627 if (stopped()) {
1628 return true;
1629 }
1630
1631 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1632 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1633 newcopy = new_array(klass_node, size, 0); // no arguments to push
1634 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1635
1636 // Calculate starting addresses.
1637 Node* src_start = array_element_address(value, offset, T_CHAR);
1638 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1639
1640 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1641 const TypeInt* toffset = gvn().type(offset)->is_int();
1642 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1643
1644 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1645 const char* copyfunc_name = "arraycopy";
1646 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1647 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1648 OptoRuntime::fast_arraycopy_Type(),
1649 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1650 src_start, dst_start, ConvI2X(length) XTOP);
1651 // Do not let reads from the cloned object float above the arraycopy.
1652 if (alloc != NULL) {
1653 if (alloc->maybe_set_complete(&_gvn)) {
1654 // "You break it, you buy it."
1655 InitializeNode* init = alloc->initialization();
1656 assert(init->is_complete(), "we just did this");
1657 init->set_complete_with_arraycopy();
1658 assert(newcopy->is_CheckCastPP(), "sanity");
1659 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1660 }
1661 // Do not let stores that initialize this object be reordered with
1662 // a subsequent store that would make this object accessible by
1663 // other threads.
1664 // Record what AllocateNode this StoreStore protects so that
1665 // escape analysis can go from the MemBarStoreStoreNode to the
1666 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1667 // based on the escape status of the AllocateNode.
1668 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1669 } else {
1670 insert_mem_bar(Op_MemBarCPUOrder);
1671 }
1672 } // original reexecute is set back here
1673
1674 C->set_has_split_ifs(true); // Has chance for split-if optimization
1675 if (!stopped()) {
1676 set_result(newcopy);
1677 }
1678 clear_upper_avx();
1679
1680 return true;
1681 }
1682
1683 //------------------------inline_string_getCharsU--------------------------
1684 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1685 bool LibraryCallKit::inline_string_getCharsU() {
1686 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1687 return false;
1688 }
1689
1690 // Get the arguments.
1691 Node* src = argument(0);
1692 Node* src_begin = argument(1);
1693 Node* src_end = argument(2); // exclusive offset (i < src_end)
1694 Node* dst = argument(3);
1695 Node* dst_begin = argument(4);
1696
1697 // Check for allocation before we add nodes that would confuse
1698 // tightly_coupled_allocation()
1699 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1700
1701 // Check if a null path was taken unconditionally.
1702 src = null_check(src);
1703 dst = null_check(dst);
1704 if (stopped()) {
1705 return true;
1706 }
1707
1708 // Get length and convert char[] offset to byte[] offset
1709 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1710 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1711
1712 // Range checks
1713 generate_string_range_check(src, src_begin, length, true);
1714 generate_string_range_check(dst, dst_begin, length, false);
1715 if (stopped()) {
1716 return true;
1717 }
1718
1719 if (!stopped()) {
1720 // Calculate starting addresses.
1721 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1722 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1723
1724 // Check if array addresses are aligned to HeapWordSize
1725 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1726 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1727 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1728 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1729
1730 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1731 const char* copyfunc_name = "arraycopy";
1732 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1733 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1734 OptoRuntime::fast_arraycopy_Type(),
1735 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1736 src_start, dst_start, ConvI2X(length) XTOP);
1737 // Do not let reads from the cloned object float above the arraycopy.
1738 if (alloc != NULL) {
1739 if (alloc->maybe_set_complete(&_gvn)) {
1740 // "You break it, you buy it."
1741 InitializeNode* init = alloc->initialization();
1742 assert(init->is_complete(), "we just did this");
1743 init->set_complete_with_arraycopy();
1744 assert(dst->is_CheckCastPP(), "sanity");
1745 assert(dst->in(0)->in(0) == init, "dest pinned");
1746 }
1747 // Do not let stores that initialize this object be reordered with
1748 // a subsequent store that would make this object accessible by
1749 // other threads.
1750 // Record what AllocateNode this StoreStore protects so that
1751 // escape analysis can go from the MemBarStoreStoreNode to the
1752 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1753 // based on the escape status of the AllocateNode.
1754 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1755 } else {
1756 insert_mem_bar(Op_MemBarCPUOrder);
1757 }
1758 }
1759
1760 C->set_has_split_ifs(true); // Has chance for split-if optimization
1761 return true;
1762 }
1763
1764 //----------------------inline_string_char_access----------------------------
1765 // Store/Load char to/from byte[] array.
1766 // static void StringUTF16.putChar(byte[] val, int index, int c)
1767 // static char StringUTF16.getChar(byte[] val, int index)
1768 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1769 Node* value = argument(0);
1770 Node* index = argument(1);
1771 Node* ch = is_store ? argument(2) : NULL;
1772
1773 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1774 // correctly requires matched array shapes.
1775 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1776 "sanity: byte[] and char[] bases agree");
1777 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1778 "sanity: byte[] and char[] scales agree");
1779
1780 // Bail when getChar over constants is requested: constant folding would
1781 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1782 // Java method would constant fold nicely instead.
1783 if (!is_store && value->is_Con() && index->is_Con()) {
1784 return false;
1785 }
1786
1787 value = must_be_not_null(value, true);
1788
1789 Node* adr = array_element_address(value, index, T_CHAR);
1790 if (adr->is_top()) {
1791 return false;
1792 }
1793 if (is_store) {
1794 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED);
1795 } else {
1796 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD);
1797 set_result(ch);
1798 }
1799 return true;
1800 }
1801
1802 //--------------------------round_double_node--------------------------------
1803 // Round a double node if necessary.
1804 Node* LibraryCallKit::round_double_node(Node* n) {
1805 if (Matcher::strict_fp_requires_explicit_rounding) {
1806 #ifdef IA32
1807 if (UseSSE < 2) {
1808 n = _gvn.transform(new RoundDoubleNode(NULL, n));
1809 }
1810 #else
1811 Unimplemented();
1812 #endif // IA32
1813 }
1814 return n;
1815 }
1816
1817 //------------------------------inline_math-----------------------------------
1818 // public static double Math.abs(double)
1819 // public static double Math.sqrt(double)
1820 // public static double Math.log(double)
1821 // public static double Math.log10(double)
1822 bool LibraryCallKit::inline_double_math(vmIntrinsics::ID id) {
1823 Node* arg = round_double_node(argument(0));
1824 Node* n = NULL;
1825 switch (id) {
1826 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1827 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1828 case vmIntrinsics::_ceil: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_ceil); break;
1829 case vmIntrinsics::_floor: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_floor); break;
1830 case vmIntrinsics::_rint: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_rint); break;
1831 default: fatal_unexpected_iid(id); break;
1832 }
1833 set_result(_gvn.transform(n));
1834 return true;
1835 }
1836
1837 //------------------------------inline_math-----------------------------------
1838 // public static float Math.abs(float)
1839 // public static int Math.abs(int)
1840 // public static long Math.abs(long)
1841 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1842 Node* arg = argument(0);
1843 Node* n = NULL;
1844 switch (id) {
1845 case vmIntrinsics::_fabs: n = new AbsFNode( arg); break;
1846 case vmIntrinsics::_iabs: n = new AbsINode( arg); break;
1847 case vmIntrinsics::_labs: n = new AbsLNode( arg); break;
1848 default: fatal_unexpected_iid(id); break;
1849 }
1850 set_result(_gvn.transform(n));
1851 return true;
1852 }
1853
1854 //------------------------------runtime_math-----------------------------
1855 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1856 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1857 "must be (DD)D or (D)D type");
1858
1859 // Inputs
1860 Node* a = round_double_node(argument(0));
1861 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1862
1863 const TypePtr* no_memory_effects = NULL;
1864 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1865 no_memory_effects,
1866 a, top(), b, b ? top() : NULL);
1867 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1868 #ifdef ASSERT
1869 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1870 assert(value_top == top(), "second value must be top");
1871 #endif
1872
1873 set_result(value);
1874 return true;
1875 }
1876
1877 //------------------------------inline_math_native-----------------------------
1878 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1879 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1880 switch (id) {
1881 // These intrinsics are not properly supported on all hardware
1882 case vmIntrinsics::_dsin:
1883 return StubRoutines::dsin() != NULL ?
1884 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1885 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1886 case vmIntrinsics::_dcos:
1887 return StubRoutines::dcos() != NULL ?
1888 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1889 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1890 case vmIntrinsics::_dtan:
1891 return StubRoutines::dtan() != NULL ?
1892 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1893 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1894 case vmIntrinsics::_dlog:
1895 return StubRoutines::dlog() != NULL ?
1896 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1897 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1898 case vmIntrinsics::_dlog10:
1899 return StubRoutines::dlog10() != NULL ?
1900 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1901 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1902
1903 // These intrinsics are supported on all hardware
1904 case vmIntrinsics::_ceil:
1905 case vmIntrinsics::_floor:
1906 case vmIntrinsics::_rint: return Matcher::match_rule_supported(Op_RoundDoubleMode) ? inline_double_math(id) : false;
1907 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_double_math(id) : false;
1908 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_double_math(id) : false;
1909 case vmIntrinsics::_fabs: return Matcher::match_rule_supported(Op_AbsF) ? inline_math(id) : false;
1910 case vmIntrinsics::_iabs: return Matcher::match_rule_supported(Op_AbsI) ? inline_math(id) : false;
1911 case vmIntrinsics::_labs: return Matcher::match_rule_supported(Op_AbsL) ? inline_math(id) : false;
1912
1913 case vmIntrinsics::_dexp:
1914 return StubRoutines::dexp() != NULL ?
1915 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1916 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1917 case vmIntrinsics::_dpow: {
1918 Node* exp = round_double_node(argument(2));
1919 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1920 if (d != NULL && d->getd() == 2.0) {
1921 // Special case: pow(x, 2.0) => x * x
1922 Node* base = round_double_node(argument(0));
1923 set_result(_gvn.transform(new MulDNode(base, base)));
1924 return true;
1925 }
1926 return StubRoutines::dpow() != NULL ?
1927 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1928 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1929 }
1930 #undef FN_PTR
1931
1932 // These intrinsics are not yet correctly implemented
1933 case vmIntrinsics::_datan2:
1934 return false;
1935
1936 default:
1937 fatal_unexpected_iid(id);
1938 return false;
1939 }
1940 }
1941
1942 static bool is_simple_name(Node* n) {
1943 return (n->req() == 1 // constant
1944 || (n->is_Type() && n->as_Type()->type()->singleton())
1945 || n->is_Proj() // parameter or return value
1946 || n->is_Phi() // local of some sort
1947 );
1948 }
1949
1950 //----------------------------inline_notify-----------------------------------*
1951 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1952 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1953 address func;
1954 if (id == vmIntrinsics::_notify) {
1955 func = OptoRuntime::monitor_notify_Java();
1956 } else {
1957 func = OptoRuntime::monitor_notifyAll_Java();
1958 }
1959 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1960 make_slow_call_ex(call, env()->Throwable_klass(), false);
1961 return true;
1962 }
1963
1964
1965 //----------------------------inline_min_max-----------------------------------
1966 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1967 set_result(generate_min_max(id, argument(0), argument(1)));
1968 return true;
1969 }
1970
1971 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1972 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1973 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1974 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1975 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1976
1977 {
1978 PreserveJVMState pjvms(this);
1979 PreserveReexecuteState preexecs(this);
1980 jvms()->set_should_reexecute(true);
1981
1982 set_control(slow_path);
1983 set_i_o(i_o());
1984
1985 uncommon_trap(Deoptimization::Reason_intrinsic,
1986 Deoptimization::Action_none);
1987 }
1988
1989 set_control(fast_path);
1990 set_result(math);
1991 }
1992
1993 template <typename OverflowOp>
1994 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1995 typedef typename OverflowOp::MathOp MathOp;
1996
1997 MathOp* mathOp = new MathOp(arg1, arg2);
1998 Node* operation = _gvn.transform( mathOp );
1999 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2000 inline_math_mathExact(operation, ofcheck);
2001 return true;
2002 }
2003
2004 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2005 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2006 }
2007
2008 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2009 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2010 }
2011
2012 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2013 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2014 }
2015
2016 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2017 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2018 }
2019
2020 bool LibraryCallKit::inline_math_negateExactI() {
2021 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2022 }
2023
2024 bool LibraryCallKit::inline_math_negateExactL() {
2025 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2026 }
2027
2028 bool LibraryCallKit::inline_math_multiplyExactI() {
2029 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2030 }
2031
2032 bool LibraryCallKit::inline_math_multiplyExactL() {
2033 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2034 }
2035
2036 bool LibraryCallKit::inline_math_multiplyHigh() {
2037 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
2038 return true;
2039 }
2040
2041 Node*
2042 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2043 // These are the candidate return value:
2044 Node* xvalue = x0;
2045 Node* yvalue = y0;
2046
2047 if (xvalue == yvalue) {
2048 return xvalue;
2049 }
2050
2051 bool want_max = (id == vmIntrinsics::_max);
2052
2053 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2054 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2055 if (txvalue == NULL || tyvalue == NULL) return top();
2056 // This is not really necessary, but it is consistent with a
2057 // hypothetical MaxINode::Value method:
2058 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2059
2060 // %%% This folding logic should (ideally) be in a different place.
2061 // Some should be inside IfNode, and there to be a more reliable
2062 // transformation of ?: style patterns into cmoves. We also want
2063 // more powerful optimizations around cmove and min/max.
2064
2065 // Try to find a dominating comparison of these guys.
2066 // It can simplify the index computation for Arrays.copyOf
2067 // and similar uses of System.arraycopy.
2068 // First, compute the normalized version of CmpI(x, y).
2069 int cmp_op = Op_CmpI;
2070 Node* xkey = xvalue;
2071 Node* ykey = yvalue;
2072 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2073 if (ideal_cmpxy->is_Cmp()) {
2074 // E.g., if we have CmpI(length - offset, count),
2075 // it might idealize to CmpI(length, count + offset)
2076 cmp_op = ideal_cmpxy->Opcode();
2077 xkey = ideal_cmpxy->in(1);
2078 ykey = ideal_cmpxy->in(2);
2079 }
2080
2081 // Start by locating any relevant comparisons.
2082 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2083 Node* cmpxy = NULL;
2084 Node* cmpyx = NULL;
2085 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2086 Node* cmp = start_from->fast_out(k);
2087 if (cmp->outcnt() > 0 && // must have prior uses
2088 cmp->in(0) == NULL && // must be context-independent
2089 cmp->Opcode() == cmp_op) { // right kind of compare
2090 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2091 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2092 }
2093 }
2094
2095 const int NCMPS = 2;
2096 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2097 int cmpn;
2098 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2099 if (cmps[cmpn] != NULL) break; // find a result
2100 }
2101 if (cmpn < NCMPS) {
2102 // Look for a dominating test that tells us the min and max.
2103 int depth = 0; // Limit search depth for speed
2104 Node* dom = control();
2105 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2106 if (++depth >= 100) break;
2107 Node* ifproj = dom;
2108 if (!ifproj->is_Proj()) continue;
2109 Node* iff = ifproj->in(0);
2110 if (!iff->is_If()) continue;
2111 Node* bol = iff->in(1);
2112 if (!bol->is_Bool()) continue;
2113 Node* cmp = bol->in(1);
2114 if (cmp == NULL) continue;
2115 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2116 if (cmps[cmpn] == cmp) break;
2117 if (cmpn == NCMPS) continue;
2118 BoolTest::mask btest = bol->as_Bool()->_test._test;
2119 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2120 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2121 // At this point, we know that 'x btest y' is true.
2122 switch (btest) {
2123 case BoolTest::eq:
2124 // They are proven equal, so we can collapse the min/max.
2125 // Either value is the answer. Choose the simpler.
2126 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2127 return yvalue;
2128 return xvalue;
2129 case BoolTest::lt: // x < y
2130 case BoolTest::le: // x <= y
2131 return (want_max ? yvalue : xvalue);
2132 case BoolTest::gt: // x > y
2133 case BoolTest::ge: // x >= y
2134 return (want_max ? xvalue : yvalue);
2135 default:
2136 break;
2137 }
2138 }
2139 }
2140
2141 // We failed to find a dominating test.
2142 // Let's pick a test that might GVN with prior tests.
2143 Node* best_bol = NULL;
2144 BoolTest::mask best_btest = BoolTest::illegal;
2145 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2146 Node* cmp = cmps[cmpn];
2147 if (cmp == NULL) continue;
2148 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2149 Node* bol = cmp->fast_out(j);
2150 if (!bol->is_Bool()) continue;
2151 BoolTest::mask btest = bol->as_Bool()->_test._test;
2152 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2153 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2154 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2155 best_bol = bol->as_Bool();
2156 best_btest = btest;
2157 }
2158 }
2159 }
2160
2161 Node* answer_if_true = NULL;
2162 Node* answer_if_false = NULL;
2163 switch (best_btest) {
2164 default:
2165 if (cmpxy == NULL)
2166 cmpxy = ideal_cmpxy;
2167 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2168 // and fall through:
2169 case BoolTest::lt: // x < y
2170 case BoolTest::le: // x <= y
2171 answer_if_true = (want_max ? yvalue : xvalue);
2172 answer_if_false = (want_max ? xvalue : yvalue);
2173 break;
2174 case BoolTest::gt: // x > y
2175 case BoolTest::ge: // x >= y
2176 answer_if_true = (want_max ? xvalue : yvalue);
2177 answer_if_false = (want_max ? yvalue : xvalue);
2178 break;
2179 }
2180
2181 jint hi, lo;
2182 if (want_max) {
2183 // We can sharpen the minimum.
2184 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2185 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2186 } else {
2187 // We can sharpen the maximum.
2188 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2189 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2190 }
2191
2192 // Use a flow-free graph structure, to avoid creating excess control edges
2193 // which could hinder other optimizations.
2194 // Since Math.min/max is often used with arraycopy, we want
2195 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2196 Node* cmov = CMoveNode::make(NULL, best_bol,
2197 answer_if_false, answer_if_true,
2198 TypeInt::make(lo, hi, widen));
2199
2200 return _gvn.transform(cmov);
2201
2202 /*
2203 // This is not as desirable as it may seem, since Min and Max
2204 // nodes do not have a full set of optimizations.
2205 // And they would interfere, anyway, with 'if' optimizations
2206 // and with CMoveI canonical forms.
2207 switch (id) {
2208 case vmIntrinsics::_min:
2209 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2210 case vmIntrinsics::_max:
2211 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2212 default:
2213 ShouldNotReachHere();
2214 }
2215 */
2216 }
2217
2218 inline int
2219 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2220 const TypePtr* base_type = TypePtr::NULL_PTR;
2221 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2222 if (base_type == NULL) {
2223 // Unknown type.
2224 return Type::AnyPtr;
2225 } else if (base_type == TypePtr::NULL_PTR) {
2226 // Since this is a NULL+long form, we have to switch to a rawptr.
2227 base = _gvn.transform(new CastX2PNode(offset));
2228 offset = MakeConX(0);
2229 return Type::RawPtr;
2230 } else if (base_type->base() == Type::RawPtr) {
2231 return Type::RawPtr;
2232 } else if (base_type->isa_oopptr()) {
2233 // Base is never null => always a heap address.
2234 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2235 return Type::OopPtr;
2236 }
2237 // Offset is small => always a heap address.
2238 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2239 if (offset_type != NULL &&
2240 base_type->offset() == 0 && // (should always be?)
2241 offset_type->_lo >= 0 &&
2242 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2243 return Type::OopPtr;
2244 } else if (type == T_OBJECT) {
2245 // off heap access to an oop doesn't make any sense. Has to be on
2246 // heap.
2247 return Type::OopPtr;
2248 }
2249 // Otherwise, it might either be oop+off or NULL+addr.
2250 return Type::AnyPtr;
2251 } else {
2252 // No information:
2253 return Type::AnyPtr;
2254 }
2255 }
2256
2257 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) {
2258 Node* uncasted_base = base;
2259 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2260 if (kind == Type::RawPtr) {
2261 return basic_plus_adr(top(), uncasted_base, offset);
2262 } else if (kind == Type::AnyPtr) {
2263 assert(base == uncasted_base, "unexpected base change");
2264 if (can_cast) {
2265 if (!_gvn.type(base)->speculative_maybe_null() &&
2266 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2267 // According to profiling, this access is always on
2268 // heap. Casting the base to not null and thus avoiding membars
2269 // around the access should allow better optimizations
2270 Node* null_ctl = top();
2271 base = null_check_oop(base, &null_ctl, true, true, true);
2272 assert(null_ctl->is_top(), "no null control here");
2273 return basic_plus_adr(base, offset);
2274 } else if (_gvn.type(base)->speculative_always_null() &&
2275 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2276 // According to profiling, this access is always off
2277 // heap.
2278 base = null_assert(base);
2279 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2280 offset = MakeConX(0);
2281 return basic_plus_adr(top(), raw_base, offset);
2282 }
2283 }
2284 // We don't know if it's an on heap or off heap access. Fall back
2285 // to raw memory access.
2286 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2287 return basic_plus_adr(top(), raw, offset);
2288 } else {
2289 assert(base == uncasted_base, "unexpected base change");
2290 // We know it's an on heap access so base can't be null
2291 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2292 base = must_be_not_null(base, true);
2293 }
2294 return basic_plus_adr(base, offset);
2295 }
2296 }
2297
2298 //--------------------------inline_number_methods-----------------------------
2299 // inline int Integer.numberOfLeadingZeros(int)
2300 // inline int Long.numberOfLeadingZeros(long)
2301 //
2302 // inline int Integer.numberOfTrailingZeros(int)
2303 // inline int Long.numberOfTrailingZeros(long)
2304 //
2305 // inline int Integer.bitCount(int)
2306 // inline int Long.bitCount(long)
2307 //
2308 // inline char Character.reverseBytes(char)
2309 // inline short Short.reverseBytes(short)
2310 // inline int Integer.reverseBytes(int)
2311 // inline long Long.reverseBytes(long)
2312 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2313 Node* arg = argument(0);
2314 Node* n = NULL;
2315 switch (id) {
2316 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2317 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2318 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2319 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2320 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2321 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2322 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2323 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2324 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2325 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2326 default: fatal_unexpected_iid(id); break;
2327 }
2328 set_result(_gvn.transform(n));
2329 return true;
2330 }
2331
2332 //----------------------------inline_unsafe_access----------------------------
2333
2334 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2335 // Attempt to infer a sharper value type from the offset and base type.
2336 ciKlass* sharpened_klass = NULL;
2337
2338 // See if it is an instance field, with an object type.
2339 if (alias_type->field() != NULL) {
2340 if (alias_type->field()->type()->is_klass()) {
2341 sharpened_klass = alias_type->field()->type()->as_klass();
2342 }
2343 }
2344
2345 // See if it is a narrow oop array.
2346 if (adr_type->isa_aryptr()) {
2347 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2348 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2349 if (elem_type != NULL) {
2350 sharpened_klass = elem_type->klass();
2351 }
2352 }
2353 }
2354
2355 // The sharpened class might be unloaded if there is no class loader
2356 // contraint in place.
2357 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2358 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2359
2360 #ifndef PRODUCT
2361 if (C->print_intrinsics() || C->print_inlining()) {
2362 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2363 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2364 }
2365 #endif
2366 // Sharpen the value type.
2367 return tjp;
2368 }
2369 return NULL;
2370 }
2371
2372 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) {
2373 switch (kind) {
2374 case Relaxed:
2375 return MO_UNORDERED;
2376 case Opaque:
2377 return MO_RELAXED;
2378 case Acquire:
2379 return MO_ACQUIRE;
2380 case Release:
2381 return MO_RELEASE;
2382 case Volatile:
2383 return MO_SEQ_CST;
2384 default:
2385 ShouldNotReachHere();
2386 return 0;
2387 }
2388 }
2389
2390 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2391 if (callee()->is_static()) return false; // caller must have the capability!
2392 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2393 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2394 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2395 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2396
2397 if (is_reference_type(type)) {
2398 decorators |= ON_UNKNOWN_OOP_REF;
2399 }
2400
2401 if (unaligned) {
2402 decorators |= C2_UNALIGNED;
2403 }
2404
2405 #ifndef PRODUCT
2406 {
2407 ResourceMark rm;
2408 // Check the signatures.
2409 ciSignature* sig = callee()->signature();
2410 #ifdef ASSERT
2411 if (!is_store) {
2412 // Object getReference(Object base, int/long offset), etc.
2413 BasicType rtype = sig->return_type()->basic_type();
2414 assert(rtype == type, "getter must return the expected value");
2415 assert(sig->count() == 2, "oop getter has 2 arguments");
2416 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2417 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2418 } else {
2419 // void putReference(Object base, int/long offset, Object x), etc.
2420 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2421 assert(sig->count() == 3, "oop putter has 3 arguments");
2422 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2423 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2424 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2425 assert(vtype == type, "putter must accept the expected value");
2426 }
2427 #endif // ASSERT
2428 }
2429 #endif //PRODUCT
2430
2431 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2432
2433 Node* receiver = argument(0); // type: oop
2434
2435 // Build address expression.
2436 Node* adr;
2437 Node* heap_base_oop = top();
2438 Node* offset = top();
2439 Node* val;
2440
2441 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2442 Node* base = argument(1); // type: oop
2443 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2444 offset = argument(2); // type: long
2445 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2446 // to be plain byte offsets, which are also the same as those accepted
2447 // by oopDesc::field_addr.
2448 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2449 "fieldOffset must be byte-scaled");
2450 // 32-bit machines ignore the high half!
2451 offset = ConvL2X(offset);
2452 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed);
2453
2454 if (_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) {
2455 if (type != T_OBJECT) {
2456 decorators |= IN_NATIVE; // off-heap primitive access
2457 } else {
2458 return false; // off-heap oop accesses are not supported
2459 }
2460 } else {
2461 heap_base_oop = base; // on-heap or mixed access
2462 }
2463
2464 // Can base be NULL? Otherwise, always on-heap access.
2465 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base));
2466
2467 if (!can_access_non_heap) {
2468 decorators |= IN_HEAP;
2469 }
2470
2471 val = is_store ? argument(4) : NULL;
2472
2473 const TypePtr* adr_type = _gvn.type(adr)->isa_ptr();
2474 if (adr_type == TypePtr::NULL_PTR) {
2475 return false; // off-heap access with zero address
2476 }
2477
2478 // Try to categorize the address.
2479 Compile::AliasType* alias_type = C->alias_type(adr_type);
2480 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2481
2482 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2483 alias_type->adr_type() == TypeAryPtr::RANGE) {
2484 return false; // not supported
2485 }
2486
2487 bool mismatched = false;
2488 BasicType bt = alias_type->basic_type();
2489 if (bt != T_ILLEGAL) {
2490 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2491 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2492 // Alias type doesn't differentiate between byte[] and boolean[]).
2493 // Use address type to get the element type.
2494 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2495 }
2496 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2497 // accessing an array field with getReference is not a mismatch
2498 bt = T_OBJECT;
2499 }
2500 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2501 // Don't intrinsify mismatched object accesses
2502 return false;
2503 }
2504 mismatched = (bt != type);
2505 } else if (alias_type->adr_type()->isa_oopptr()) {
2506 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2507 }
2508
2509 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2510
2511 if (mismatched) {
2512 decorators |= C2_MISMATCHED;
2513 }
2514
2515 // First guess at the value type.
2516 const Type *value_type = Type::get_const_basic_type(type);
2517
2518 // Figure out the memory ordering.
2519 decorators |= mo_decorator_for_access_kind(kind);
2520
2521 if (!is_store && type == T_OBJECT) {
2522 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2523 if (tjp != NULL) {
2524 value_type = tjp;
2525 }
2526 }
2527
2528 receiver = null_check(receiver);
2529 if (stopped()) {
2530 return true;
2531 }
2532 // Heap pointers get a null-check from the interpreter,
2533 // as a courtesy. However, this is not guaranteed by Unsafe,
2534 // and it is not possible to fully distinguish unintended nulls
2535 // from intended ones in this API.
2536
2537 if (!is_store) {
2538 Node* p = NULL;
2539 // Try to constant fold a load from a constant field
2540 ciField* field = alias_type->field();
2541 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2542 // final or stable field
2543 p = make_constant_from_field(field, heap_base_oop);
2544 }
2545
2546 if (p == NULL) { // Could not constant fold the load
2547 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators);
2548 // Normalize the value returned by getBoolean in the following cases
2549 if (type == T_BOOLEAN &&
2550 (mismatched ||
2551 heap_base_oop == top() || // - heap_base_oop is NULL or
2552 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL
2553 // and the unsafe access is made to large offset
2554 // (i.e., larger than the maximum offset necessary for any
2555 // field access)
2556 ) {
2557 IdealKit ideal = IdealKit(this);
2558 #define __ ideal.
2559 IdealVariable normalized_result(ideal);
2560 __ declarations_done();
2561 __ set(normalized_result, p);
2562 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2563 __ set(normalized_result, ideal.ConI(1));
2564 ideal.end_if();
2565 final_sync(ideal);
2566 p = __ value(normalized_result);
2567 #undef __
2568 }
2569 }
2570 if (type == T_ADDRESS) {
2571 p = gvn().transform(new CastP2XNode(NULL, p));
2572 p = ConvX2UL(p);
2573 }
2574 // The load node has the control of the preceding MemBarCPUOrder. All
2575 // following nodes will have the control of the MemBarCPUOrder inserted at
2576 // the end of this method. So, pushing the load onto the stack at a later
2577 // point is fine.
2578 set_result(p);
2579 } else {
2580 if (bt == T_ADDRESS) {
2581 // Repackage the long as a pointer.
2582 val = ConvL2X(val);
2583 val = gvn().transform(new CastX2PNode(val));
2584 }
2585 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators);
2586 }
2587
2588 return true;
2589 }
2590
2591 //----------------------------inline_unsafe_load_store----------------------------
2592 // This method serves a couple of different customers (depending on LoadStoreKind):
2593 //
2594 // LS_cmp_swap:
2595 //
2596 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x);
2597 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2598 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2599 //
2600 // LS_cmp_swap_weak:
2601 //
2602 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x);
2603 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x);
2604 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x);
2605 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x);
2606 //
2607 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2608 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2609 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2610 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2611 //
2612 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2613 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2614 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2615 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2616 //
2617 // LS_cmp_exchange:
2618 //
2619 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x);
2620 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x);
2621 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x);
2622 //
2623 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2624 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2625 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2626 //
2627 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2628 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2629 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2630 //
2631 // LS_get_add:
2632 //
2633 // int getAndAddInt( Object o, long offset, int delta)
2634 // long getAndAddLong(Object o, long offset, long delta)
2635 //
2636 // LS_get_set:
2637 //
2638 // int getAndSet(Object o, long offset, int newValue)
2639 // long getAndSet(Object o, long offset, long newValue)
2640 // Object getAndSet(Object o, long offset, Object newValue)
2641 //
2642 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2643 // This basic scheme here is the same as inline_unsafe_access, but
2644 // differs in enough details that combining them would make the code
2645 // overly confusing. (This is a true fact! I originally combined
2646 // them, but even I was confused by it!) As much code/comments as
2647 // possible are retained from inline_unsafe_access though to make
2648 // the correspondences clearer. - dl
2649
2650 if (callee()->is_static()) return false; // caller must have the capability!
2651
2652 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2653 decorators |= mo_decorator_for_access_kind(access_kind);
2654
2655 #ifndef PRODUCT
2656 BasicType rtype;
2657 {
2658 ResourceMark rm;
2659 // Check the signatures.
2660 ciSignature* sig = callee()->signature();
2661 rtype = sig->return_type()->basic_type();
2662 switch(kind) {
2663 case LS_get_add:
2664 case LS_get_set: {
2665 // Check the signatures.
2666 #ifdef ASSERT
2667 assert(rtype == type, "get and set must return the expected type");
2668 assert(sig->count() == 3, "get and set has 3 arguments");
2669 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2670 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2671 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2672 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2673 #endif // ASSERT
2674 break;
2675 }
2676 case LS_cmp_swap:
2677 case LS_cmp_swap_weak: {
2678 // Check the signatures.
2679 #ifdef ASSERT
2680 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2681 assert(sig->count() == 4, "CAS has 4 arguments");
2682 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2683 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2684 #endif // ASSERT
2685 break;
2686 }
2687 case LS_cmp_exchange: {
2688 // Check the signatures.
2689 #ifdef ASSERT
2690 assert(rtype == type, "CAS must return the expected type");
2691 assert(sig->count() == 4, "CAS has 4 arguments");
2692 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2693 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2694 #endif // ASSERT
2695 break;
2696 }
2697 default:
2698 ShouldNotReachHere();
2699 }
2700 }
2701 #endif //PRODUCT
2702
2703 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2704
2705 // Get arguments:
2706 Node* receiver = NULL;
2707 Node* base = NULL;
2708 Node* offset = NULL;
2709 Node* oldval = NULL;
2710 Node* newval = NULL;
2711 switch(kind) {
2712 case LS_cmp_swap:
2713 case LS_cmp_swap_weak:
2714 case LS_cmp_exchange: {
2715 const bool two_slot_type = type2size[type] == 2;
2716 receiver = argument(0); // type: oop
2717 base = argument(1); // type: oop
2718 offset = argument(2); // type: long
2719 oldval = argument(4); // type: oop, int, or long
2720 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2721 break;
2722 }
2723 case LS_get_add:
2724 case LS_get_set: {
2725 receiver = argument(0); // type: oop
2726 base = argument(1); // type: oop
2727 offset = argument(2); // type: long
2728 oldval = NULL;
2729 newval = argument(4); // type: oop, int, or long
2730 break;
2731 }
2732 default:
2733 ShouldNotReachHere();
2734 }
2735
2736 // Build field offset expression.
2737 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2738 // to be plain byte offsets, which are also the same as those accepted
2739 // by oopDesc::field_addr.
2740 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2741 // 32-bit machines ignore the high half of long offsets
2742 offset = ConvL2X(offset);
2743 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false);
2744 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2745
2746 Compile::AliasType* alias_type = C->alias_type(adr_type);
2747 BasicType bt = alias_type->basic_type();
2748 if (bt != T_ILLEGAL &&
2749 (is_reference_type(bt) != (type == T_OBJECT))) {
2750 // Don't intrinsify mismatched object accesses.
2751 return false;
2752 }
2753
2754 // For CAS, unlike inline_unsafe_access, there seems no point in
2755 // trying to refine types. Just use the coarse types here.
2756 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2757 const Type *value_type = Type::get_const_basic_type(type);
2758
2759 switch (kind) {
2760 case LS_get_set:
2761 case LS_cmp_exchange: {
2762 if (type == T_OBJECT) {
2763 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2764 if (tjp != NULL) {
2765 value_type = tjp;
2766 }
2767 }
2768 break;
2769 }
2770 case LS_cmp_swap:
2771 case LS_cmp_swap_weak:
2772 case LS_get_add:
2773 break;
2774 default:
2775 ShouldNotReachHere();
2776 }
2777
2778 // Null check receiver.
2779 receiver = null_check(receiver);
2780 if (stopped()) {
2781 return true;
2782 }
2783
2784 int alias_idx = C->get_alias_index(adr_type);
2785
2786 if (is_reference_type(type)) {
2787 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF;
2788
2789 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2790 // could be delayed during Parse (for example, in adjust_map_after_if()).
2791 // Execute transformation here to avoid barrier generation in such case.
2792 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2793 newval = _gvn.makecon(TypePtr::NULL_PTR);
2794
2795 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) {
2796 // Refine the value to a null constant, when it is known to be null
2797 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2798 }
2799 }
2800
2801 Node* result = NULL;
2802 switch (kind) {
2803 case LS_cmp_exchange: {
2804 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx,
2805 oldval, newval, value_type, type, decorators);
2806 break;
2807 }
2808 case LS_cmp_swap_weak:
2809 decorators |= C2_WEAK_CMPXCHG;
2810 case LS_cmp_swap: {
2811 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx,
2812 oldval, newval, value_type, type, decorators);
2813 break;
2814 }
2815 case LS_get_set: {
2816 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx,
2817 newval, value_type, type, decorators);
2818 break;
2819 }
2820 case LS_get_add: {
2821 result = access_atomic_add_at(base, adr, adr_type, alias_idx,
2822 newval, value_type, type, decorators);
2823 break;
2824 }
2825 default:
2826 ShouldNotReachHere();
2827 }
2828
2829 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2830 set_result(result);
2831 return true;
2832 }
2833
2834 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2835 // Regardless of form, don't allow previous ld/st to move down,
2836 // then issue acquire, release, or volatile mem_bar.
2837 insert_mem_bar(Op_MemBarCPUOrder);
2838 switch(id) {
2839 case vmIntrinsics::_loadFence:
2840 insert_mem_bar(Op_LoadFence);
2841 return true;
2842 case vmIntrinsics::_storeFence:
2843 insert_mem_bar(Op_StoreFence);
2844 return true;
2845 case vmIntrinsics::_fullFence:
2846 insert_mem_bar(Op_MemBarVolatile);
2847 return true;
2848 default:
2849 fatal_unexpected_iid(id);
2850 return false;
2851 }
2852 }
2853
2854 bool LibraryCallKit::inline_onspinwait() {
2855 insert_mem_bar(Op_OnSpinWait);
2856 return true;
2857 }
2858
2859 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2860 if (!kls->is_Con()) {
2861 return true;
2862 }
2863 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2864 if (klsptr == NULL) {
2865 return true;
2866 }
2867 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2868 // don't need a guard for a klass that is already initialized
2869 return !ik->is_initialized();
2870 }
2871
2872 //----------------------------inline_unsafe_writeback0-------------------------
2873 // public native void Unsafe.writeback0(long address)
2874 bool LibraryCallKit::inline_unsafe_writeback0() {
2875 if (!Matcher::has_match_rule(Op_CacheWB)) {
2876 return false;
2877 }
2878 #ifndef PRODUCT
2879 assert(Matcher::has_match_rule(Op_CacheWBPreSync), "found match rule for CacheWB but not CacheWBPreSync");
2880 assert(Matcher::has_match_rule(Op_CacheWBPostSync), "found match rule for CacheWB but not CacheWBPostSync");
2881 ciSignature* sig = callee()->signature();
2882 assert(sig->type_at(0)->basic_type() == T_LONG, "Unsafe_writeback0 address is long!");
2883 #endif
2884 null_check_receiver(); // null-check, then ignore
2885 Node *addr = argument(1);
2886 addr = new CastX2PNode(addr);
2887 addr = _gvn.transform(addr);
2888 Node *flush = new CacheWBNode(control(), memory(TypeRawPtr::BOTTOM), addr);
2889 flush = _gvn.transform(flush);
2890 set_memory(flush, TypeRawPtr::BOTTOM);
2891 return true;
2892 }
2893
2894 //----------------------------inline_unsafe_writeback0-------------------------
2895 // public native void Unsafe.writeback0(long address)
2896 bool LibraryCallKit::inline_unsafe_writebackSync0(bool is_pre) {
2897 if (is_pre && !Matcher::has_match_rule(Op_CacheWBPreSync)) {
2898 return false;
2899 }
2900 if (!is_pre && !Matcher::has_match_rule(Op_CacheWBPostSync)) {
2901 return false;
2902 }
2903 #ifndef PRODUCT
2904 assert(Matcher::has_match_rule(Op_CacheWB),
2905 (is_pre ? "found match rule for CacheWBPreSync but not CacheWB"
2906 : "found match rule for CacheWBPostSync but not CacheWB"));
2907
2908 #endif
2909 null_check_receiver(); // null-check, then ignore
2910 Node *sync;
2911 if (is_pre) {
2912 sync = new CacheWBPreSyncNode(control(), memory(TypeRawPtr::BOTTOM));
2913 } else {
2914 sync = new CacheWBPostSyncNode(control(), memory(TypeRawPtr::BOTTOM));
2915 }
2916 sync = _gvn.transform(sync);
2917 set_memory(sync, TypeRawPtr::BOTTOM);
2918 return true;
2919 }
2920
2921 //----------------------------inline_unsafe_allocate---------------------------
2922 // public native Object Unsafe.allocateInstance(Class<?> cls);
2923 bool LibraryCallKit::inline_unsafe_allocate() {
2924 if (callee()->is_static()) return false; // caller must have the capability!
2925
2926 null_check_receiver(); // null-check, then ignore
2927 Node* cls = null_check(argument(1));
2928 if (stopped()) return true;
2929
2930 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2931 kls = null_check(kls);
2932 if (stopped()) return true; // argument was like int.class
2933
2934 Node* test = NULL;
2935 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2936 // Note: The argument might still be an illegal value like
2937 // Serializable.class or Object[].class. The runtime will handle it.
2938 // But we must make an explicit check for initialization.
2939 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2940 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2941 // can generate code to load it as unsigned byte.
2942 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2943 Node* bits = intcon(InstanceKlass::fully_initialized);
2944 test = _gvn.transform(new SubINode(inst, bits));
2945 // The 'test' is non-zero if we need to take a slow path.
2946 }
2947
2948 Node* obj = new_instance(kls, test);
2949 set_result(obj);
2950 return true;
2951 }
2952
2953 //------------------------inline_native_time_funcs--------------
2954 // inline code for System.currentTimeMillis() and System.nanoTime()
2955 // these have the same type and signature
2956 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2957 const TypeFunc* tf = OptoRuntime::void_long_Type();
2958 const TypePtr* no_memory_effects = NULL;
2959 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2960 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2961 #ifdef ASSERT
2962 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2963 assert(value_top == top(), "second value must be top");
2964 #endif
2965 set_result(value);
2966 return true;
2967 }
2968
2969 #ifdef JFR_HAVE_INTRINSICS
2970
2971 /*
2972 * oop -> myklass
2973 * myklass->trace_id |= USED
2974 * return myklass->trace_id & ~0x3
2975 */
2976 bool LibraryCallKit::inline_native_classID() {
2977 Node* cls = null_check(argument(0), T_OBJECT);
2978 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2979 kls = null_check(kls, T_OBJECT);
2980
2981 ByteSize offset = KLASS_TRACE_ID_OFFSET;
2982 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2983 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2984
2985 Node* clsused = longcon(0x01l); // set the class bit
2986 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2987 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2988 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2989
2990 #ifdef TRACE_ID_META_BITS
2991 Node* mbits = longcon(~TRACE_ID_META_BITS);
2992 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
2993 #endif
2994 #ifdef TRACE_ID_SHIFT
2995 Node* cbits = intcon(TRACE_ID_SHIFT);
2996 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
2997 #endif
2998
2999 set_result(tvalue);
3000 return true;
3001
3002 }
3003
3004 bool LibraryCallKit::inline_native_getEventWriter() {
3005 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3006
3007 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3008 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR));
3009
3010 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3011
3012 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3013 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3014
3015 IfNode* iff_jobj_null =
3016 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3017
3018 enum { _normal_path = 1,
3019 _null_path = 2,
3020 PATH_LIMIT };
3021
3022 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3023 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM);
3024
3025 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3026 result_rgn->init_req(_null_path, jobj_is_null);
3027 result_val->init_req(_null_path, null());
3028
3029 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3030 set_control(jobj_is_not_null);
3031 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT,
3032 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD);
3033 result_rgn->init_req(_normal_path, control());
3034 result_val->init_req(_normal_path, res);
3035
3036 set_result(result_rgn, result_val);
3037
3038 return true;
3039 }
3040
3041 #endif // JFR_HAVE_INTRINSICS
3042
3043 //------------------------inline_native_currentThread------------------
3044 bool LibraryCallKit::inline_native_currentThread() {
3045 Node* junk = NULL;
3046 set_result(generate_current_thread(junk));
3047 return true;
3048 }
3049
3050 //---------------------------load_mirror_from_klass----------------------------
3051 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3052 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3053 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3054 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3055 // mirror = ((OopHandle)mirror)->resolve();
3056 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
3057 }
3058
3059 //-----------------------load_klass_from_mirror_common-------------------------
3060 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3061 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3062 // and branch to the given path on the region.
3063 // If never_see_null, take an uncommon trap on null, so we can optimistically
3064 // compile for the non-null case.
3065 // If the region is NULL, force never_see_null = true.
3066 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3067 bool never_see_null,
3068 RegionNode* region,
3069 int null_path,
3070 int offset) {
3071 if (region == NULL) never_see_null = true;
3072 Node* p = basic_plus_adr(mirror, offset);
3073 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3074 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3075 Node* null_ctl = top();
3076 kls = null_check_oop(kls, &null_ctl, never_see_null);
3077 if (region != NULL) {
3078 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3079 region->init_req(null_path, null_ctl);
3080 } else {
3081 assert(null_ctl == top(), "no loose ends");
3082 }
3083 return kls;
3084 }
3085
3086 //--------------------(inline_native_Class_query helpers)---------------------
3087 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3088 // Fall through if (mods & mask) == bits, take the guard otherwise.
3089 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3090 // Branch around if the given klass has the given modifier bit set.
3091 // Like generate_guard, adds a new path onto the region.
3092 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3093 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3094 Node* mask = intcon(modifier_mask);
3095 Node* bits = intcon(modifier_bits);
3096 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3097 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3098 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3099 return generate_fair_guard(bol, region);
3100 }
3101 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3102 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3103 }
3104 Node* LibraryCallKit::generate_hidden_class_guard(Node* kls, RegionNode* region) {
3105 return generate_access_flags_guard(kls, JVM_ACC_IS_HIDDEN_CLASS, 0, region);
3106 }
3107
3108 //-------------------------inline_native_Class_query-------------------
3109 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3110 const Type* return_type = TypeInt::BOOL;
3111 Node* prim_return_value = top(); // what happens if it's a primitive class?
3112 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3113 bool expect_prim = false; // most of these guys expect to work on refs
3114
3115 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3116
3117 Node* mirror = argument(0);
3118 Node* obj = top();
3119
3120 switch (id) {
3121 case vmIntrinsics::_isInstance:
3122 // nothing is an instance of a primitive type
3123 prim_return_value = intcon(0);
3124 obj = argument(1);
3125 break;
3126 case vmIntrinsics::_getModifiers:
3127 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3128 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3129 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3130 break;
3131 case vmIntrinsics::_isInterface:
3132 prim_return_value = intcon(0);
3133 break;
3134 case vmIntrinsics::_isArray:
3135 prim_return_value = intcon(0);
3136 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3137 break;
3138 case vmIntrinsics::_isPrimitive:
3139 prim_return_value = intcon(1);
3140 expect_prim = true; // obviously
3141 break;
3142 case vmIntrinsics::_isHidden:
3143 prim_return_value = intcon(0);
3144 break;
3145 case vmIntrinsics::_getSuperclass:
3146 prim_return_value = null();
3147 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3148 break;
3149 case vmIntrinsics::_getClassAccessFlags:
3150 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3151 return_type = TypeInt::INT; // not bool! 6297094
3152 break;
3153 default:
3154 fatal_unexpected_iid(id);
3155 break;
3156 }
3157
3158 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3159 if (mirror_con == NULL) return false; // cannot happen?
3160
3161 #ifndef PRODUCT
3162 if (C->print_intrinsics() || C->print_inlining()) {
3163 ciType* k = mirror_con->java_mirror_type();
3164 if (k) {
3165 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3166 k->print_name();
3167 tty->cr();
3168 }
3169 }
3170 #endif
3171
3172 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3173 RegionNode* region = new RegionNode(PATH_LIMIT);
3174 record_for_igvn(region);
3175 PhiNode* phi = new PhiNode(region, return_type);
3176
3177 // The mirror will never be null of Reflection.getClassAccessFlags, however
3178 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3179 // if it is. See bug 4774291.
3180
3181 // For Reflection.getClassAccessFlags(), the null check occurs in
3182 // the wrong place; see inline_unsafe_access(), above, for a similar
3183 // situation.
3184 mirror = null_check(mirror);
3185 // If mirror or obj is dead, only null-path is taken.
3186 if (stopped()) return true;
3187
3188 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3189
3190 // Now load the mirror's klass metaobject, and null-check it.
3191 // Side-effects region with the control path if the klass is null.
3192 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3193 // If kls is null, we have a primitive mirror.
3194 phi->init_req(_prim_path, prim_return_value);
3195 if (stopped()) { set_result(region, phi); return true; }
3196 bool safe_for_replace = (region->in(_prim_path) == top());
3197
3198 Node* p; // handy temp
3199 Node* null_ctl;
3200
3201 // Now that we have the non-null klass, we can perform the real query.
3202 // For constant classes, the query will constant-fold in LoadNode::Value.
3203 Node* query_value = top();
3204 switch (id) {
3205 case vmIntrinsics::_isInstance:
3206 // nothing is an instance of a primitive type
3207 query_value = gen_instanceof(obj, kls, safe_for_replace);
3208 break;
3209
3210 case vmIntrinsics::_getModifiers:
3211 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3212 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3213 break;
3214
3215 case vmIntrinsics::_isInterface:
3216 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3217 if (generate_interface_guard(kls, region) != NULL)
3218 // A guard was added. If the guard is taken, it was an interface.
3219 phi->add_req(intcon(1));
3220 // If we fall through, it's a plain class.
3221 query_value = intcon(0);
3222 break;
3223
3224 case vmIntrinsics::_isArray:
3225 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3226 if (generate_array_guard(kls, region) != NULL)
3227 // A guard was added. If the guard is taken, it was an array.
3228 phi->add_req(intcon(1));
3229 // If we fall through, it's a plain class.
3230 query_value = intcon(0);
3231 break;
3232
3233 case vmIntrinsics::_isPrimitive:
3234 query_value = intcon(0); // "normal" path produces false
3235 break;
3236
3237 case vmIntrinsics::_isHidden:
3238 // (To verify this code sequence, check the asserts in JVM_IsHiddenClass.)
3239 if (generate_hidden_class_guard(kls, region) != NULL)
3240 // A guard was added. If the guard is taken, it was an hidden class.
3241 phi->add_req(intcon(1));
3242 // If we fall through, it's a plain class.
3243 query_value = intcon(0);
3244 break;
3245
3246
3247 case vmIntrinsics::_getSuperclass:
3248 // The rules here are somewhat unfortunate, but we can still do better
3249 // with random logic than with a JNI call.
3250 // Interfaces store null or Object as _super, but must report null.
3251 // Arrays store an intermediate super as _super, but must report Object.
3252 // Other types can report the actual _super.
3253 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3254 if (generate_interface_guard(kls, region) != NULL)
3255 // A guard was added. If the guard is taken, it was an interface.
3256 phi->add_req(null());
3257 if (generate_array_guard(kls, region) != NULL)
3258 // A guard was added. If the guard is taken, it was an array.
3259 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3260 // If we fall through, it's a plain class. Get its _super.
3261 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3262 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3263 null_ctl = top();
3264 kls = null_check_oop(kls, &null_ctl);
3265 if (null_ctl != top()) {
3266 // If the guard is taken, Object.superClass is null (both klass and mirror).
3267 region->add_req(null_ctl);
3268 phi ->add_req(null());
3269 }
3270 if (!stopped()) {
3271 query_value = load_mirror_from_klass(kls);
3272 }
3273 break;
3274
3275 case vmIntrinsics::_getClassAccessFlags:
3276 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3277 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3278 break;
3279
3280 default:
3281 fatal_unexpected_iid(id);
3282 break;
3283 }
3284
3285 // Fall-through is the normal case of a query to a real class.
3286 phi->init_req(1, query_value);
3287 region->init_req(1, control());
3288
3289 C->set_has_split_ifs(true); // Has chance for split-if optimization
3290 set_result(region, phi);
3291 return true;
3292 }
3293
3294 //-------------------------inline_Class_cast-------------------
3295 bool LibraryCallKit::inline_Class_cast() {
3296 Node* mirror = argument(0); // Class
3297 Node* obj = argument(1);
3298 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3299 if (mirror_con == NULL) {
3300 return false; // dead path (mirror->is_top()).
3301 }
3302 if (obj == NULL || obj->is_top()) {
3303 return false; // dead path
3304 }
3305 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3306
3307 // First, see if Class.cast() can be folded statically.
3308 // java_mirror_type() returns non-null for compile-time Class constants.
3309 ciType* tm = mirror_con->java_mirror_type();
3310 if (tm != NULL && tm->is_klass() &&
3311 tp != NULL && tp->klass() != NULL) {
3312 if (!tp->klass()->is_loaded()) {
3313 // Don't use intrinsic when class is not loaded.
3314 return false;
3315 } else {
3316 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3317 if (static_res == Compile::SSC_always_true) {
3318 // isInstance() is true - fold the code.
3319 set_result(obj);
3320 return true;
3321 } else if (static_res == Compile::SSC_always_false) {
3322 // Don't use intrinsic, have to throw ClassCastException.
3323 // If the reference is null, the non-intrinsic bytecode will
3324 // be optimized appropriately.
3325 return false;
3326 }
3327 }
3328 }
3329
3330 // Bailout intrinsic and do normal inlining if exception path is frequent.
3331 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3332 return false;
3333 }
3334
3335 // Generate dynamic checks.
3336 // Class.cast() is java implementation of _checkcast bytecode.
3337 // Do checkcast (Parse::do_checkcast()) optimizations here.
3338
3339 mirror = null_check(mirror);
3340 // If mirror is dead, only null-path is taken.
3341 if (stopped()) {
3342 return true;
3343 }
3344
3345 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3346 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3347 RegionNode* region = new RegionNode(PATH_LIMIT);
3348 record_for_igvn(region);
3349
3350 // Now load the mirror's klass metaobject, and null-check it.
3351 // If kls is null, we have a primitive mirror and
3352 // nothing is an instance of a primitive type.
3353 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3354
3355 Node* res = top();
3356 if (!stopped()) {
3357 Node* bad_type_ctrl = top();
3358 // Do checkcast optimizations.
3359 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3360 region->init_req(_bad_type_path, bad_type_ctrl);
3361 }
3362 if (region->in(_prim_path) != top() ||
3363 region->in(_bad_type_path) != top()) {
3364 // Let Interpreter throw ClassCastException.
3365 PreserveJVMState pjvms(this);
3366 set_control(_gvn.transform(region));
3367 uncommon_trap(Deoptimization::Reason_intrinsic,
3368 Deoptimization::Action_maybe_recompile);
3369 }
3370 if (!stopped()) {
3371 set_result(res);
3372 }
3373 return true;
3374 }
3375
3376
3377 //--------------------------inline_native_subtype_check------------------------
3378 // This intrinsic takes the JNI calls out of the heart of
3379 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3380 bool LibraryCallKit::inline_native_subtype_check() {
3381 // Pull both arguments off the stack.
3382 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3383 args[0] = argument(0);
3384 args[1] = argument(1);
3385 Node* klasses[2]; // corresponding Klasses: superk, subk
3386 klasses[0] = klasses[1] = top();
3387
3388 enum {
3389 // A full decision tree on {superc is prim, subc is prim}:
3390 _prim_0_path = 1, // {P,N} => false
3391 // {P,P} & superc!=subc => false
3392 _prim_same_path, // {P,P} & superc==subc => true
3393 _prim_1_path, // {N,P} => false
3394 _ref_subtype_path, // {N,N} & subtype check wins => true
3395 _both_ref_path, // {N,N} & subtype check loses => false
3396 PATH_LIMIT
3397 };
3398
3399 RegionNode* region = new RegionNode(PATH_LIMIT);
3400 Node* phi = new PhiNode(region, TypeInt::BOOL);
3401 record_for_igvn(region);
3402
3403 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3404 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3405 int class_klass_offset = java_lang_Class::klass_offset();
3406
3407 // First null-check both mirrors and load each mirror's klass metaobject.
3408 int which_arg;
3409 for (which_arg = 0; which_arg <= 1; which_arg++) {
3410 Node* arg = args[which_arg];
3411 arg = null_check(arg);
3412 if (stopped()) break;
3413 args[which_arg] = arg;
3414
3415 Node* p = basic_plus_adr(arg, class_klass_offset);
3416 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3417 klasses[which_arg] = _gvn.transform(kls);
3418 }
3419
3420 // Having loaded both klasses, test each for null.
3421 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3422 for (which_arg = 0; which_arg <= 1; which_arg++) {
3423 Node* kls = klasses[which_arg];
3424 Node* null_ctl = top();
3425 kls = null_check_oop(kls, &null_ctl, never_see_null);
3426 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3427 region->init_req(prim_path, null_ctl);
3428 if (stopped()) break;
3429 klasses[which_arg] = kls;
3430 }
3431
3432 if (!stopped()) {
3433 // now we have two reference types, in klasses[0..1]
3434 Node* subk = klasses[1]; // the argument to isAssignableFrom
3435 Node* superk = klasses[0]; // the receiver
3436 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3437 // now we have a successful reference subtype check
3438 region->set_req(_ref_subtype_path, control());
3439 }
3440
3441 // If both operands are primitive (both klasses null), then
3442 // we must return true when they are identical primitives.
3443 // It is convenient to test this after the first null klass check.
3444 set_control(region->in(_prim_0_path)); // go back to first null check
3445 if (!stopped()) {
3446 // Since superc is primitive, make a guard for the superc==subc case.
3447 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3448 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3449 generate_guard(bol_eq, region, PROB_FAIR);
3450 if (region->req() == PATH_LIMIT+1) {
3451 // A guard was added. If the added guard is taken, superc==subc.
3452 region->swap_edges(PATH_LIMIT, _prim_same_path);
3453 region->del_req(PATH_LIMIT);
3454 }
3455 region->set_req(_prim_0_path, control()); // Not equal after all.
3456 }
3457
3458 // these are the only paths that produce 'true':
3459 phi->set_req(_prim_same_path, intcon(1));
3460 phi->set_req(_ref_subtype_path, intcon(1));
3461
3462 // pull together the cases:
3463 assert(region->req() == PATH_LIMIT, "sane region");
3464 for (uint i = 1; i < region->req(); i++) {
3465 Node* ctl = region->in(i);
3466 if (ctl == NULL || ctl == top()) {
3467 region->set_req(i, top());
3468 phi ->set_req(i, top());
3469 } else if (phi->in(i) == NULL) {
3470 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3471 }
3472 }
3473
3474 set_control(_gvn.transform(region));
3475 set_result(_gvn.transform(phi));
3476 return true;
3477 }
3478
3479 //---------------------generate_array_guard_common------------------------
3480 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3481 bool obj_array, bool not_array) {
3482
3483 if (stopped()) {
3484 return NULL;
3485 }
3486
3487 // If obj_array/non_array==false/false:
3488 // Branch around if the given klass is in fact an array (either obj or prim).
3489 // If obj_array/non_array==false/true:
3490 // Branch around if the given klass is not an array klass of any kind.
3491 // If obj_array/non_array==true/true:
3492 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3493 // If obj_array/non_array==true/false:
3494 // Branch around if the kls is an oop array (Object[] or subtype)
3495 //
3496 // Like generate_guard, adds a new path onto the region.
3497 jint layout_con = 0;
3498 Node* layout_val = get_layout_helper(kls, layout_con);
3499 if (layout_val == NULL) {
3500 bool query = (obj_array
3501 ? Klass::layout_helper_is_objArray(layout_con)
3502 : Klass::layout_helper_is_array(layout_con));
3503 if (query == not_array) {
3504 return NULL; // never a branch
3505 } else { // always a branch
3506 Node* always_branch = control();
3507 if (region != NULL)
3508 region->add_req(always_branch);
3509 set_control(top());
3510 return always_branch;
3511 }
3512 }
3513 // Now test the correct condition.
3514 jint nval = (obj_array
3515 ? (jint)(Klass::_lh_array_tag_type_value
3516 << Klass::_lh_array_tag_shift)
3517 : Klass::_lh_neutral_value);
3518 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3519 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3520 // invert the test if we are looking for a non-array
3521 if (not_array) btest = BoolTest(btest).negate();
3522 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3523 return generate_fair_guard(bol, region);
3524 }
3525
3526
3527 //-----------------------inline_native_newArray--------------------------
3528 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3529 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3530 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3531 Node* mirror;
3532 Node* count_val;
3533 if (uninitialized) {
3534 mirror = argument(1);
3535 count_val = argument(2);
3536 } else {
3537 mirror = argument(0);
3538 count_val = argument(1);
3539 }
3540
3541 mirror = null_check(mirror);
3542 // If mirror or obj is dead, only null-path is taken.
3543 if (stopped()) return true;
3544
3545 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3546 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3547 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3548 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3549 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3550
3551 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3552 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3553 result_reg, _slow_path);
3554 Node* normal_ctl = control();
3555 Node* no_array_ctl = result_reg->in(_slow_path);
3556
3557 // Generate code for the slow case. We make a call to newArray().
3558 set_control(no_array_ctl);
3559 if (!stopped()) {
3560 // Either the input type is void.class, or else the
3561 // array klass has not yet been cached. Either the
3562 // ensuing call will throw an exception, or else it
3563 // will cache the array klass for next time.
3564 PreserveJVMState pjvms(this);
3565 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3566 Node* slow_result = set_results_for_java_call(slow_call);
3567 // this->control() comes from set_results_for_java_call
3568 result_reg->set_req(_slow_path, control());
3569 result_val->set_req(_slow_path, slow_result);
3570 result_io ->set_req(_slow_path, i_o());
3571 result_mem->set_req(_slow_path, reset_memory());
3572 }
3573
3574 set_control(normal_ctl);
3575 if (!stopped()) {
3576 // Normal case: The array type has been cached in the java.lang.Class.
3577 // The following call works fine even if the array type is polymorphic.
3578 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3579 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3580 result_reg->init_req(_normal_path, control());
3581 result_val->init_req(_normal_path, obj);
3582 result_io ->init_req(_normal_path, i_o());
3583 result_mem->init_req(_normal_path, reset_memory());
3584
3585 if (uninitialized) {
3586 // Mark the allocation so that zeroing is skipped
3587 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3588 alloc->maybe_set_complete(&_gvn);
3589 }
3590 }
3591
3592 // Return the combined state.
3593 set_i_o( _gvn.transform(result_io) );
3594 set_all_memory( _gvn.transform(result_mem));
3595
3596 C->set_has_split_ifs(true); // Has chance for split-if optimization
3597 set_result(result_reg, result_val);
3598 return true;
3599 }
3600
3601 //----------------------inline_native_getLength--------------------------
3602 // public static native int java.lang.reflect.Array.getLength(Object array);
3603 bool LibraryCallKit::inline_native_getLength() {
3604 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3605
3606 Node* array = null_check(argument(0));
3607 // If array is dead, only null-path is taken.
3608 if (stopped()) return true;
3609
3610 // Deoptimize if it is a non-array.
3611 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3612
3613 if (non_array != NULL) {
3614 PreserveJVMState pjvms(this);
3615 set_control(non_array);
3616 uncommon_trap(Deoptimization::Reason_intrinsic,
3617 Deoptimization::Action_maybe_recompile);
3618 }
3619
3620 // If control is dead, only non-array-path is taken.
3621 if (stopped()) return true;
3622
3623 // The works fine even if the array type is polymorphic.
3624 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3625 Node* result = load_array_length(array);
3626
3627 C->set_has_split_ifs(true); // Has chance for split-if optimization
3628 set_result(result);
3629 return true;
3630 }
3631
3632 //------------------------inline_array_copyOf----------------------------
3633 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3634 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3635 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3636 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3637
3638 // Get the arguments.
3639 Node* original = argument(0);
3640 Node* start = is_copyOfRange? argument(1): intcon(0);
3641 Node* end = is_copyOfRange? argument(2): argument(1);
3642 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3643
3644 Node* newcopy = NULL;
3645
3646 // Set the original stack and the reexecute bit for the interpreter to reexecute
3647 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3648 { PreserveReexecuteState preexecs(this);
3649 jvms()->set_should_reexecute(true);
3650
3651 array_type_mirror = null_check(array_type_mirror);
3652 original = null_check(original);
3653
3654 // Check if a null path was taken unconditionally.
3655 if (stopped()) return true;
3656
3657 Node* orig_length = load_array_length(original);
3658
3659 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3660 klass_node = null_check(klass_node);
3661
3662 RegionNode* bailout = new RegionNode(1);
3663 record_for_igvn(bailout);
3664
3665 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3666 // Bail out if that is so.
3667 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3668 if (not_objArray != NULL) {
3669 // Improve the klass node's type from the new optimistic assumption:
3670 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3671 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3672 Node* cast = new CastPPNode(klass_node, akls);
3673 cast->init_req(0, control());
3674 klass_node = _gvn.transform(cast);
3675 }
3676
3677 // Bail out if either start or end is negative.
3678 generate_negative_guard(start, bailout, &start);
3679 generate_negative_guard(end, bailout, &end);
3680
3681 Node* length = end;
3682 if (_gvn.type(start) != TypeInt::ZERO) {
3683 length = _gvn.transform(new SubINode(end, start));
3684 }
3685
3686 // Bail out if length is negative.
3687 // Without this the new_array would throw
3688 // NegativeArraySizeException but IllegalArgumentException is what
3689 // should be thrown
3690 generate_negative_guard(length, bailout, &length);
3691
3692 if (bailout->req() > 1) {
3693 PreserveJVMState pjvms(this);
3694 set_control(_gvn.transform(bailout));
3695 uncommon_trap(Deoptimization::Reason_intrinsic,
3696 Deoptimization::Action_maybe_recompile);
3697 }
3698
3699 if (!stopped()) {
3700 // How many elements will we copy from the original?
3701 // The answer is MinI(orig_length - start, length).
3702 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3703 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3704
3705 // Generate a direct call to the right arraycopy function(s).
3706 // We know the copy is disjoint but we might not know if the
3707 // oop stores need checking.
3708 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3709 // This will fail a store-check if x contains any non-nulls.
3710
3711 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3712 // loads/stores but it is legal only if we're sure the
3713 // Arrays.copyOf would succeed. So we need all input arguments
3714 // to the copyOf to be validated, including that the copy to the
3715 // new array won't trigger an ArrayStoreException. That subtype
3716 // check can be optimized if we know something on the type of
3717 // the input array from type speculation.
3718 if (_gvn.type(klass_node)->singleton()) {
3719 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3720 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3721
3722 int test = C->static_subtype_check(superk, subk);
3723 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3724 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3725 if (t_original->speculative_type() != NULL) {
3726 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3727 }
3728 }
3729 }
3730
3731 bool validated = false;
3732 // Reason_class_check rather than Reason_intrinsic because we
3733 // want to intrinsify even if this traps.
3734 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3735 Node* not_subtype_ctrl = gen_subtype_check(original, klass_node);
3736
3737 if (not_subtype_ctrl != top()) {
3738 PreserveJVMState pjvms(this);
3739 set_control(not_subtype_ctrl);
3740 uncommon_trap(Deoptimization::Reason_class_check,
3741 Deoptimization::Action_make_not_entrant);
3742 assert(stopped(), "Should be stopped");
3743 }
3744 validated = true;
3745 }
3746
3747 if (!stopped()) {
3748 newcopy = new_array(klass_node, length, 0); // no arguments to push
3749
3750 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
3751 load_object_klass(original), klass_node);
3752 if (!is_copyOfRange) {
3753 ac->set_copyof(validated);
3754 } else {
3755 ac->set_copyofrange(validated);
3756 }
3757 Node* n = _gvn.transform(ac);
3758 if (n == ac) {
3759 ac->connect_outputs(this);
3760 } else {
3761 assert(validated, "shouldn't transform if all arguments not validated");
3762 set_all_memory(n);
3763 }
3764 }
3765 }
3766 } // original reexecute is set back here
3767
3768 C->set_has_split_ifs(true); // Has chance for split-if optimization
3769 if (!stopped()) {
3770 set_result(newcopy);
3771 }
3772 return true;
3773 }
3774
3775
3776 //----------------------generate_virtual_guard---------------------------
3777 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3778 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3779 RegionNode* slow_region) {
3780 ciMethod* method = callee();
3781 int vtable_index = method->vtable_index();
3782 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3783 "bad index %d", vtable_index);
3784 // Get the Method* out of the appropriate vtable entry.
3785 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
3786 vtable_index*vtableEntry::size_in_bytes() +
3787 vtableEntry::method_offset_in_bytes();
3788 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3789 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3790
3791 // Compare the target method with the expected method (e.g., Object.hashCode).
3792 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3793
3794 Node* native_call = makecon(native_call_addr);
3795 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3796 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3797
3798 return generate_slow_guard(test_native, slow_region);
3799 }
3800
3801 //-----------------------generate_method_call----------------------------
3802 // Use generate_method_call to make a slow-call to the real
3803 // method if the fast path fails. An alternative would be to
3804 // use a stub like OptoRuntime::slow_arraycopy_Java.
3805 // This only works for expanding the current library call,
3806 // not another intrinsic. (E.g., don't use this for making an
3807 // arraycopy call inside of the copyOf intrinsic.)
3808 CallJavaNode*
3809 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3810 // When compiling the intrinsic method itself, do not use this technique.
3811 guarantee(callee() != C->method(), "cannot make slow-call to self");
3812
3813 ciMethod* method = callee();
3814 // ensure the JVMS we have will be correct for this call
3815 guarantee(method_id == method->intrinsic_id(), "must match");
3816
3817 const TypeFunc* tf = TypeFunc::make(method);
3818 CallJavaNode* slow_call;
3819 if (is_static) {
3820 assert(!is_virtual, "");
3821 slow_call = new CallStaticJavaNode(C, tf,
3822 SharedRuntime::get_resolve_static_call_stub(),
3823 method, bci());
3824 } else if (is_virtual) {
3825 null_check_receiver();
3826 int vtable_index = Method::invalid_vtable_index;
3827 if (UseInlineCaches) {
3828 // Suppress the vtable call
3829 } else {
3830 // hashCode and clone are not a miranda methods,
3831 // so the vtable index is fixed.
3832 // No need to use the linkResolver to get it.
3833 vtable_index = method->vtable_index();
3834 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3835 "bad index %d", vtable_index);
3836 }
3837 slow_call = new CallDynamicJavaNode(tf,
3838 SharedRuntime::get_resolve_virtual_call_stub(),
3839 method, vtable_index, bci());
3840 } else { // neither virtual nor static: opt_virtual
3841 null_check_receiver();
3842 slow_call = new CallStaticJavaNode(C, tf,
3843 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3844 method, bci());
3845 slow_call->set_optimized_virtual(true);
3846 }
3847 if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) {
3848 // To be able to issue a direct call (optimized virtual or virtual)
3849 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information
3850 // about the method being invoked should be attached to the call site to
3851 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C).
3852 slow_call->set_override_symbolic_info(true);
3853 }
3854 set_arguments_for_java_call(slow_call);
3855 set_edges_for_java_call(slow_call);
3856 return slow_call;
3857 }
3858
3859
3860 /**
3861 * Build special case code for calls to hashCode on an object. This call may
3862 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3863 * slightly different code.
3864 */
3865 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3866 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3867 assert(!(is_virtual && is_static), "either virtual, special, or static");
3868
3869 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3870
3871 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3872 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3873 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3874 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3875 Node* obj = NULL;
3876 if (!is_static) {
3877 // Check for hashing null object
3878 obj = null_check_receiver();
3879 if (stopped()) return true; // unconditionally null
3880 result_reg->init_req(_null_path, top());
3881 result_val->init_req(_null_path, top());
3882 } else {
3883 // Do a null check, and return zero if null.
3884 // System.identityHashCode(null) == 0
3885 obj = argument(0);
3886 Node* null_ctl = top();
3887 obj = null_check_oop(obj, &null_ctl);
3888 result_reg->init_req(_null_path, null_ctl);
3889 result_val->init_req(_null_path, _gvn.intcon(0));
3890 }
3891
3892 // Unconditionally null? Then return right away.
3893 if (stopped()) {
3894 set_control( result_reg->in(_null_path));
3895 if (!stopped())
3896 set_result(result_val->in(_null_path));
3897 return true;
3898 }
3899
3900 // We only go to the fast case code if we pass a number of guards. The
3901 // paths which do not pass are accumulated in the slow_region.
3902 RegionNode* slow_region = new RegionNode(1);
3903 record_for_igvn(slow_region);
3904
3905 // If this is a virtual call, we generate a funny guard. We pull out
3906 // the vtable entry corresponding to hashCode() from the target object.
3907 // If the target method which we are calling happens to be the native
3908 // Object hashCode() method, we pass the guard. We do not need this
3909 // guard for non-virtual calls -- the caller is known to be the native
3910 // Object hashCode().
3911 if (is_virtual) {
3912 // After null check, get the object's klass.
3913 Node* obj_klass = load_object_klass(obj);
3914 generate_virtual_guard(obj_klass, slow_region);
3915 }
3916
3917 // Get the header out of the object, use LoadMarkNode when available
3918 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3919 // The control of the load must be NULL. Otherwise, the load can move before
3920 // the null check after castPP removal.
3921 Node* no_ctrl = NULL;
3922 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3923
3924 // Test the header to see if it is unlocked.
3925 Node *lock_mask = _gvn.MakeConX(markWord::biased_lock_mask_in_place);
3926 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3927 Node *unlocked_val = _gvn.MakeConX(markWord::unlocked_value);
3928 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3929 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3930
3931 generate_slow_guard(test_unlocked, slow_region);
3932
3933 // Get the hash value and check to see that it has been properly assigned.
3934 // We depend on hash_mask being at most 32 bits and avoid the use of
3935 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3936 // vm: see markWord.hpp.
3937 Node *hash_mask = _gvn.intcon(markWord::hash_mask);
3938 Node *hash_shift = _gvn.intcon(markWord::hash_shift);
3939 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3940 // This hack lets the hash bits live anywhere in the mark object now, as long
3941 // as the shift drops the relevant bits into the low 32 bits. Note that
3942 // Java spec says that HashCode is an int so there's no point in capturing
3943 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3944 hshifted_header = ConvX2I(hshifted_header);
3945 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3946
3947 Node *no_hash_val = _gvn.intcon(markWord::no_hash);
3948 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3949 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3950
3951 generate_slow_guard(test_assigned, slow_region);
3952
3953 Node* init_mem = reset_memory();
3954 // fill in the rest of the null path:
3955 result_io ->init_req(_null_path, i_o());
3956 result_mem->init_req(_null_path, init_mem);
3957
3958 result_val->init_req(_fast_path, hash_val);
3959 result_reg->init_req(_fast_path, control());
3960 result_io ->init_req(_fast_path, i_o());
3961 result_mem->init_req(_fast_path, init_mem);
3962
3963 // Generate code for the slow case. We make a call to hashCode().
3964 set_control(_gvn.transform(slow_region));
3965 if (!stopped()) {
3966 // No need for PreserveJVMState, because we're using up the present state.
3967 set_all_memory(init_mem);
3968 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3969 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3970 Node* slow_result = set_results_for_java_call(slow_call);
3971 // this->control() comes from set_results_for_java_call
3972 result_reg->init_req(_slow_path, control());
3973 result_val->init_req(_slow_path, slow_result);
3974 result_io ->set_req(_slow_path, i_o());
3975 result_mem ->set_req(_slow_path, reset_memory());
3976 }
3977
3978 // Return the combined state.
3979 set_i_o( _gvn.transform(result_io) );
3980 set_all_memory( _gvn.transform(result_mem));
3981
3982 set_result(result_reg, result_val);
3983 return true;
3984 }
3985
3986 //---------------------------inline_native_getClass----------------------------
3987 // public final native Class<?> java.lang.Object.getClass();
3988 //
3989 // Build special case code for calls to getClass on an object.
3990 bool LibraryCallKit::inline_native_getClass() {
3991 Node* obj = null_check_receiver();
3992 if (stopped()) return true;
3993 set_result(load_mirror_from_klass(load_object_klass(obj)));
3994 return true;
3995 }
3996
3997 //-----------------inline_native_Reflection_getCallerClass---------------------
3998 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
3999 //
4000 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4001 //
4002 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4003 // in that it must skip particular security frames and checks for
4004 // caller sensitive methods.
4005 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4006 #ifndef PRODUCT
4007 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4008 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4009 }
4010 #endif
4011
4012 if (!jvms()->has_method()) {
4013 #ifndef PRODUCT
4014 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4015 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4016 }
4017 #endif
4018 return false;
4019 }
4020
4021 // Walk back up the JVM state to find the caller at the required
4022 // depth.
4023 JVMState* caller_jvms = jvms();
4024
4025 // Cf. JVM_GetCallerClass
4026 // NOTE: Start the loop at depth 1 because the current JVM state does
4027 // not include the Reflection.getCallerClass() frame.
4028 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4029 ciMethod* m = caller_jvms->method();
4030 switch (n) {
4031 case 0:
4032 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4033 break;
4034 case 1:
4035 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4036 if (!m->caller_sensitive()) {
4037 #ifndef PRODUCT
4038 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4039 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4040 }
4041 #endif
4042 return false; // bail-out; let JVM_GetCallerClass do the work
4043 }
4044 break;
4045 default:
4046 if (!m->is_ignored_by_security_stack_walk()) {
4047 // We have reached the desired frame; return the holder class.
4048 // Acquire method holder as java.lang.Class and push as constant.
4049 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4050 ciInstance* caller_mirror = caller_klass->java_mirror();
4051 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4052
4053 #ifndef PRODUCT
4054 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4055 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());
4056 tty->print_cr(" JVM state at this point:");
4057 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4058 ciMethod* m = jvms()->of_depth(i)->method();
4059 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4060 }
4061 }
4062 #endif
4063 return true;
4064 }
4065 break;
4066 }
4067 }
4068
4069 #ifndef PRODUCT
4070 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4071 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4072 tty->print_cr(" JVM state at this point:");
4073 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4074 ciMethod* m = jvms()->of_depth(i)->method();
4075 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4076 }
4077 }
4078 #endif
4079
4080 return false; // bail-out; let JVM_GetCallerClass do the work
4081 }
4082
4083 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4084 Node* arg = argument(0);
4085 Node* result = NULL;
4086
4087 switch (id) {
4088 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4089 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4090 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4091 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4092
4093 case vmIntrinsics::_doubleToLongBits: {
4094 // two paths (plus control) merge in a wood
4095 RegionNode *r = new RegionNode(3);
4096 Node *phi = new PhiNode(r, TypeLong::LONG);
4097
4098 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4099 // Build the boolean node
4100 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4101
4102 // Branch either way.
4103 // NaN case is less traveled, which makes all the difference.
4104 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4105 Node *opt_isnan = _gvn.transform(ifisnan);
4106 assert( opt_isnan->is_If(), "Expect an IfNode");
4107 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4108 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4109
4110 set_control(iftrue);
4111
4112 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4113 Node *slow_result = longcon(nan_bits); // return NaN
4114 phi->init_req(1, _gvn.transform( slow_result ));
4115 r->init_req(1, iftrue);
4116
4117 // Else fall through
4118 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4119 set_control(iffalse);
4120
4121 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4122 r->init_req(2, iffalse);
4123
4124 // Post merge
4125 set_control(_gvn.transform(r));
4126 record_for_igvn(r);
4127
4128 C->set_has_split_ifs(true); // Has chance for split-if optimization
4129 result = phi;
4130 assert(result->bottom_type()->isa_long(), "must be");
4131 break;
4132 }
4133
4134 case vmIntrinsics::_floatToIntBits: {
4135 // two paths (plus control) merge in a wood
4136 RegionNode *r = new RegionNode(3);
4137 Node *phi = new PhiNode(r, TypeInt::INT);
4138
4139 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4140 // Build the boolean node
4141 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4142
4143 // Branch either way.
4144 // NaN case is less traveled, which makes all the difference.
4145 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4146 Node *opt_isnan = _gvn.transform(ifisnan);
4147 assert( opt_isnan->is_If(), "Expect an IfNode");
4148 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4149 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4150
4151 set_control(iftrue);
4152
4153 static const jint nan_bits = 0x7fc00000;
4154 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4155 phi->init_req(1, _gvn.transform( slow_result ));
4156 r->init_req(1, iftrue);
4157
4158 // Else fall through
4159 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4160 set_control(iffalse);
4161
4162 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4163 r->init_req(2, iffalse);
4164
4165 // Post merge
4166 set_control(_gvn.transform(r));
4167 record_for_igvn(r);
4168
4169 C->set_has_split_ifs(true); // Has chance for split-if optimization
4170 result = phi;
4171 assert(result->bottom_type()->isa_int(), "must be");
4172 break;
4173 }
4174
4175 default:
4176 fatal_unexpected_iid(id);
4177 break;
4178 }
4179 set_result(_gvn.transform(result));
4180 return true;
4181 }
4182
4183 //----------------------inline_unsafe_copyMemory-------------------------
4184 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4185 bool LibraryCallKit::inline_unsafe_copyMemory() {
4186 if (callee()->is_static()) return false; // caller must have the capability!
4187 null_check_receiver(); // null-check receiver
4188 if (stopped()) return true;
4189
4190 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4191
4192 Node* src_ptr = argument(1); // type: oop
4193 Node* src_off = ConvL2X(argument(2)); // type: long
4194 Node* dst_ptr = argument(4); // type: oop
4195 Node* dst_off = ConvL2X(argument(5)); // type: long
4196 Node* size = ConvL2X(argument(7)); // type: long
4197
4198 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4199 "fieldOffset must be byte-scaled");
4200
4201 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ);
4202 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE);
4203
4204 // Conservatively insert a memory barrier on all memory slices.
4205 // Do not let writes of the copy source or destination float below the copy.
4206 insert_mem_bar(Op_MemBarCPUOrder);
4207
4208 Node* thread = _gvn.transform(new ThreadLocalNode());
4209 Node* doing_unsafe_access_addr = basic_plus_adr(top(), thread, in_bytes(JavaThread::doing_unsafe_access_offset()));
4210 BasicType doing_unsafe_access_bt = T_BYTE;
4211 assert((sizeof(bool) * CHAR_BIT) == 8, "not implemented");
4212
4213 // update volatile field
4214 store_to_memory(control(), doing_unsafe_access_addr, intcon(1), doing_unsafe_access_bt, Compile::AliasIdxRaw, MemNode::unordered);
4215
4216 // Call it. Note that the length argument is not scaled.
4217 make_runtime_call(RC_LEAF|RC_NO_FP,
4218 OptoRuntime::fast_arraycopy_Type(),
4219 StubRoutines::unsafe_arraycopy(),
4220 "unsafe_arraycopy",
4221 TypeRawPtr::BOTTOM,
4222 src, dst, size XTOP);
4223
4224 store_to_memory(control(), doing_unsafe_access_addr, intcon(0), doing_unsafe_access_bt, Compile::AliasIdxRaw, MemNode::unordered);
4225
4226 // Do not let reads of the copy destination float above the copy.
4227 insert_mem_bar(Op_MemBarCPUOrder);
4228
4229 return true;
4230 }
4231
4232 //------------------------clone_coping-----------------------------------
4233 // Helper function for inline_native_clone.
4234 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) {
4235 assert(obj_size != NULL, "");
4236 Node* raw_obj = alloc_obj->in(1);
4237 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4238
4239 AllocateNode* alloc = NULL;
4240 if (ReduceBulkZeroing) {
4241 // We will be completely responsible for initializing this object -
4242 // mark Initialize node as complete.
4243 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4244 // The object was just allocated - there should be no any stores!
4245 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4246 // Mark as complete_with_arraycopy so that on AllocateNode
4247 // expansion, we know this AllocateNode is initialized by an array
4248 // copy and a StoreStore barrier exists after the array copy.
4249 alloc->initialization()->set_complete_with_arraycopy();
4250 }
4251
4252 Node* size = _gvn.transform(obj_size);
4253 access_clone(obj, alloc_obj, size, is_array);
4254
4255 // Do not let reads from the cloned object float above the arraycopy.
4256 if (alloc != NULL) {
4257 // Do not let stores that initialize this object be reordered with
4258 // a subsequent store that would make this object accessible by
4259 // other threads.
4260 // Record what AllocateNode this StoreStore protects so that
4261 // escape analysis can go from the MemBarStoreStoreNode to the
4262 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4263 // based on the escape status of the AllocateNode.
4264 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4265 } else {
4266 insert_mem_bar(Op_MemBarCPUOrder);
4267 }
4268 }
4269
4270 //------------------------inline_native_clone----------------------------
4271 // protected native Object java.lang.Object.clone();
4272 //
4273 // Here are the simple edge cases:
4274 // null receiver => normal trap
4275 // virtual and clone was overridden => slow path to out-of-line clone
4276 // not cloneable or finalizer => slow path to out-of-line Object.clone
4277 //
4278 // The general case has two steps, allocation and copying.
4279 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4280 //
4281 // Copying also has two cases, oop arrays and everything else.
4282 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4283 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4284 //
4285 // These steps fold up nicely if and when the cloned object's klass
4286 // can be sharply typed as an object array, a type array, or an instance.
4287 //
4288 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4289 PhiNode* result_val;
4290
4291 // Set the reexecute bit for the interpreter to reexecute
4292 // the bytecode that invokes Object.clone if deoptimization happens.
4293 { PreserveReexecuteState preexecs(this);
4294 jvms()->set_should_reexecute(true);
4295
4296 Node* obj = null_check_receiver();
4297 if (stopped()) return true;
4298
4299 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4300
4301 // If we are going to clone an instance, we need its exact type to
4302 // know the number and types of fields to convert the clone to
4303 // loads/stores. Maybe a speculative type can help us.
4304 if (!obj_type->klass_is_exact() &&
4305 obj_type->speculative_type() != NULL &&
4306 obj_type->speculative_type()->is_instance_klass()) {
4307 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4308 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4309 !spec_ik->has_injected_fields()) {
4310 ciKlass* k = obj_type->klass();
4311 if (!k->is_instance_klass() ||
4312 k->as_instance_klass()->is_interface() ||
4313 k->as_instance_klass()->has_subklass()) {
4314 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4315 }
4316 }
4317 }
4318
4319 // Conservatively insert a memory barrier on all memory slices.
4320 // Do not let writes into the original float below the clone.
4321 insert_mem_bar(Op_MemBarCPUOrder);
4322
4323 // paths into result_reg:
4324 enum {
4325 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4326 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4327 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4328 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4329 PATH_LIMIT
4330 };
4331 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4332 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4333 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4334 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4335 record_for_igvn(result_reg);
4336
4337 Node* obj_klass = load_object_klass(obj);
4338 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4339 if (array_ctl != NULL) {
4340 // It's an array.
4341 PreserveJVMState pjvms(this);
4342 set_control(array_ctl);
4343 Node* obj_length = load_array_length(obj);
4344 Node* obj_size = NULL;
4345 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size, /*deoptimize_on_exception=*/true);
4346
4347 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4348 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) {
4349 // If it is an oop array, it requires very special treatment,
4350 // because gc barriers are required when accessing the array.
4351 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4352 if (is_obja != NULL) {
4353 PreserveJVMState pjvms2(this);
4354 set_control(is_obja);
4355 // Generate a direct call to the right arraycopy function(s).
4356 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4357 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4358 ac->set_clone_oop_array();
4359 Node* n = _gvn.transform(ac);
4360 assert(n == ac, "cannot disappear");
4361 ac->connect_outputs(this, /*deoptimize_on_exception=*/true);
4362
4363 result_reg->init_req(_objArray_path, control());
4364 result_val->init_req(_objArray_path, alloc_obj);
4365 result_i_o ->set_req(_objArray_path, i_o());
4366 result_mem ->set_req(_objArray_path, reset_memory());
4367 }
4368 }
4369 // Otherwise, there are no barriers to worry about.
4370 // (We can dispense with card marks if we know the allocation
4371 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4372 // causes the non-eden paths to take compensating steps to
4373 // simulate a fresh allocation, so that no further
4374 // card marks are required in compiled code to initialize
4375 // the object.)
4376
4377 if (!stopped()) {
4378 copy_to_clone(obj, alloc_obj, obj_size, true);
4379
4380 // Present the results of the copy.
4381 result_reg->init_req(_array_path, control());
4382 result_val->init_req(_array_path, alloc_obj);
4383 result_i_o ->set_req(_array_path, i_o());
4384 result_mem ->set_req(_array_path, reset_memory());
4385 }
4386 }
4387
4388 // We only go to the instance fast case code if we pass a number of guards.
4389 // The paths which do not pass are accumulated in the slow_region.
4390 RegionNode* slow_region = new RegionNode(1);
4391 record_for_igvn(slow_region);
4392 if (!stopped()) {
4393 // It's an instance (we did array above). Make the slow-path tests.
4394 // If this is a virtual call, we generate a funny guard. We grab
4395 // the vtable entry corresponding to clone() from the target object.
4396 // If the target method which we are calling happens to be the
4397 // Object clone() method, we pass the guard. We do not need this
4398 // guard for non-virtual calls; the caller is known to be the native
4399 // Object clone().
4400 if (is_virtual) {
4401 generate_virtual_guard(obj_klass, slow_region);
4402 }
4403
4404 // The object must be easily cloneable and must not have a finalizer.
4405 // Both of these conditions may be checked in a single test.
4406 // We could optimize the test further, but we don't care.
4407 generate_access_flags_guard(obj_klass,
4408 // Test both conditions:
4409 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4410 // Must be cloneable but not finalizer:
4411 JVM_ACC_IS_CLONEABLE_FAST,
4412 slow_region);
4413 }
4414
4415 if (!stopped()) {
4416 // It's an instance, and it passed the slow-path tests.
4417 PreserveJVMState pjvms(this);
4418 Node* obj_size = NULL;
4419 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4420 // is reexecuted if deoptimization occurs and there could be problems when merging
4421 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4422 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4423
4424 copy_to_clone(obj, alloc_obj, obj_size, false);
4425
4426 // Present the results of the slow call.
4427 result_reg->init_req(_instance_path, control());
4428 result_val->init_req(_instance_path, alloc_obj);
4429 result_i_o ->set_req(_instance_path, i_o());
4430 result_mem ->set_req(_instance_path, reset_memory());
4431 }
4432
4433 // Generate code for the slow case. We make a call to clone().
4434 set_control(_gvn.transform(slow_region));
4435 if (!stopped()) {
4436 PreserveJVMState pjvms(this);
4437 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4438 // We need to deoptimize on exception (see comment above)
4439 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true);
4440 // this->control() comes from set_results_for_java_call
4441 result_reg->init_req(_slow_path, control());
4442 result_val->init_req(_slow_path, slow_result);
4443 result_i_o ->set_req(_slow_path, i_o());
4444 result_mem ->set_req(_slow_path, reset_memory());
4445 }
4446
4447 // Return the combined state.
4448 set_control( _gvn.transform(result_reg));
4449 set_i_o( _gvn.transform(result_i_o));
4450 set_all_memory( _gvn.transform(result_mem));
4451 } // original reexecute is set back here
4452
4453 set_result(_gvn.transform(result_val));
4454 return true;
4455 }
4456
4457 // If we have a tightly coupled allocation, the arraycopy may take care
4458 // of the array initialization. If one of the guards we insert between
4459 // the allocation and the arraycopy causes a deoptimization, an
4460 // unitialized array will escape the compiled method. To prevent that
4461 // we set the JVM state for uncommon traps between the allocation and
4462 // the arraycopy to the state before the allocation so, in case of
4463 // deoptimization, we'll reexecute the allocation and the
4464 // initialization.
4465 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4466 if (alloc != NULL) {
4467 ciMethod* trap_method = alloc->jvms()->method();
4468 int trap_bci = alloc->jvms()->bci();
4469
4470 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4471 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4472 // Make sure there's no store between the allocation and the
4473 // arraycopy otherwise visible side effects could be rexecuted
4474 // in case of deoptimization and cause incorrect execution.
4475 bool no_interfering_store = true;
4476 Node* mem = alloc->in(TypeFunc::Memory);
4477 if (mem->is_MergeMem()) {
4478 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4479 Node* n = mms.memory();
4480 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4481 assert(n->is_Store(), "what else?");
4482 no_interfering_store = false;
4483 break;
4484 }
4485 }
4486 } else {
4487 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4488 Node* n = mms.memory();
4489 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4490 assert(n->is_Store(), "what else?");
4491 no_interfering_store = false;
4492 break;
4493 }
4494 }
4495 }
4496
4497 if (no_interfering_store) {
4498 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4499 uint size = alloc->req();
4500 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4501 old_jvms->set_map(sfpt);
4502 for (uint i = 0; i < size; i++) {
4503 sfpt->init_req(i, alloc->in(i));
4504 }
4505 // re-push array length for deoptimization
4506 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4507 old_jvms->set_sp(old_jvms->sp()+1);
4508 old_jvms->set_monoff(old_jvms->monoff()+1);
4509 old_jvms->set_scloff(old_jvms->scloff()+1);
4510 old_jvms->set_endoff(old_jvms->endoff()+1);
4511 old_jvms->set_should_reexecute(true);
4512
4513 sfpt->set_i_o(map()->i_o());
4514 sfpt->set_memory(map()->memory());
4515 sfpt->set_control(map()->control());
4516
4517 JVMState* saved_jvms = jvms();
4518 saved_reexecute_sp = _reexecute_sp;
4519
4520 set_jvms(sfpt->jvms());
4521 _reexecute_sp = jvms()->sp();
4522
4523 return saved_jvms;
4524 }
4525 }
4526 }
4527 return NULL;
4528 }
4529
4530 // In case of a deoptimization, we restart execution at the
4531 // allocation, allocating a new array. We would leave an uninitialized
4532 // array in the heap that GCs wouldn't expect. Move the allocation
4533 // after the traps so we don't allocate the array if we
4534 // deoptimize. This is possible because tightly_coupled_allocation()
4535 // guarantees there's no observer of the allocated array at this point
4536 // and the control flow is simple enough.
4537 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4538 int saved_reexecute_sp, uint new_idx) {
4539 if (saved_jvms != NULL && !stopped()) {
4540 assert(alloc != NULL, "only with a tightly coupled allocation");
4541 // restore JVM state to the state at the arraycopy
4542 saved_jvms->map()->set_control(map()->control());
4543 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4544 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4545 // If we've improved the types of some nodes (null check) while
4546 // emitting the guards, propagate them to the current state
4547 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4548 set_jvms(saved_jvms);
4549 _reexecute_sp = saved_reexecute_sp;
4550
4551 // Remove the allocation from above the guards
4552 CallProjections callprojs;
4553 alloc->extract_projections(&callprojs, true);
4554 InitializeNode* init = alloc->initialization();
4555 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4556 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4557 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4558 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4559
4560 // move the allocation here (after the guards)
4561 _gvn.hash_delete(alloc);
4562 alloc->set_req(TypeFunc::Control, control());
4563 alloc->set_req(TypeFunc::I_O, i_o());
4564 Node *mem = reset_memory();
4565 set_all_memory(mem);
4566 alloc->set_req(TypeFunc::Memory, mem);
4567 set_control(init->proj_out_or_null(TypeFunc::Control));
4568 set_i_o(callprojs.fallthrough_ioproj);
4569
4570 // Update memory as done in GraphKit::set_output_for_allocation()
4571 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4572 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4573 if (ary_type->isa_aryptr() && length_type != NULL) {
4574 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4575 }
4576 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4577 int elemidx = C->get_alias_index(telemref);
4578 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
4579 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
4580
4581 Node* allocx = _gvn.transform(alloc);
4582 assert(allocx == alloc, "where has the allocation gone?");
4583 assert(dest->is_CheckCastPP(), "not an allocation result?");
4584
4585 _gvn.hash_delete(dest);
4586 dest->set_req(0, control());
4587 Node* destx = _gvn.transform(dest);
4588 assert(destx == dest, "where has the allocation result gone?");
4589 }
4590 }
4591
4592
4593 //------------------------------inline_arraycopy-----------------------
4594 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4595 // Object dest, int destPos,
4596 // int length);
4597 bool LibraryCallKit::inline_arraycopy() {
4598 // Get the arguments.
4599 Node* src = argument(0); // type: oop
4600 Node* src_offset = argument(1); // type: int
4601 Node* dest = argument(2); // type: oop
4602 Node* dest_offset = argument(3); // type: int
4603 Node* length = argument(4); // type: int
4604
4605 uint new_idx = C->unique();
4606
4607 // Check for allocation before we add nodes that would confuse
4608 // tightly_coupled_allocation()
4609 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4610
4611 int saved_reexecute_sp = -1;
4612 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4613 // See arraycopy_restore_alloc_state() comment
4614 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4615 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4616 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
4617 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4618
4619 // The following tests must be performed
4620 // (1) src and dest are arrays.
4621 // (2) src and dest arrays must have elements of the same BasicType
4622 // (3) src and dest must not be null.
4623 // (4) src_offset must not be negative.
4624 // (5) dest_offset must not be negative.
4625 // (6) length must not be negative.
4626 // (7) src_offset + length must not exceed length of src.
4627 // (8) dest_offset + length must not exceed length of dest.
4628 // (9) each element of an oop array must be assignable
4629
4630 // (3) src and dest must not be null.
4631 // always do this here because we need the JVM state for uncommon traps
4632 Node* null_ctl = top();
4633 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4634 assert(null_ctl->is_top(), "no null control here");
4635 dest = null_check(dest, T_ARRAY);
4636
4637 if (!can_emit_guards) {
4638 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4639 // guards but the arraycopy node could still take advantage of a
4640 // tightly allocated allocation. tightly_coupled_allocation() is
4641 // called again to make sure it takes the null check above into
4642 // account: the null check is mandatory and if it caused an
4643 // uncommon trap to be emitted then the allocation can't be
4644 // considered tightly coupled in this context.
4645 alloc = tightly_coupled_allocation(dest, NULL);
4646 }
4647
4648 bool validated = false;
4649
4650 const Type* src_type = _gvn.type(src);
4651 const Type* dest_type = _gvn.type(dest);
4652 const TypeAryPtr* top_src = src_type->isa_aryptr();
4653 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4654
4655 // Do we have the type of src?
4656 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4657 // Do we have the type of dest?
4658 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4659 // Is the type for src from speculation?
4660 bool src_spec = false;
4661 // Is the type for dest from speculation?
4662 bool dest_spec = false;
4663
4664 if ((!has_src || !has_dest) && can_emit_guards) {
4665 // We don't have sufficient type information, let's see if
4666 // speculative types can help. We need to have types for both src
4667 // and dest so that it pays off.
4668
4669 // Do we already have or could we have type information for src
4670 bool could_have_src = has_src;
4671 // Do we already have or could we have type information for dest
4672 bool could_have_dest = has_dest;
4673
4674 ciKlass* src_k = NULL;
4675 if (!has_src) {
4676 src_k = src_type->speculative_type_not_null();
4677 if (src_k != NULL && src_k->is_array_klass()) {
4678 could_have_src = true;
4679 }
4680 }
4681
4682 ciKlass* dest_k = NULL;
4683 if (!has_dest) {
4684 dest_k = dest_type->speculative_type_not_null();
4685 if (dest_k != NULL && dest_k->is_array_klass()) {
4686 could_have_dest = true;
4687 }
4688 }
4689
4690 if (could_have_src && could_have_dest) {
4691 // This is going to pay off so emit the required guards
4692 if (!has_src) {
4693 src = maybe_cast_profiled_obj(src, src_k, true);
4694 src_type = _gvn.type(src);
4695 top_src = src_type->isa_aryptr();
4696 has_src = (top_src != NULL && top_src->klass() != NULL);
4697 src_spec = true;
4698 }
4699 if (!has_dest) {
4700 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4701 dest_type = _gvn.type(dest);
4702 top_dest = dest_type->isa_aryptr();
4703 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4704 dest_spec = true;
4705 }
4706 }
4707 }
4708
4709 if (has_src && has_dest && can_emit_guards) {
4710 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4711 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4712 if (is_reference_type(src_elem)) src_elem = T_OBJECT;
4713 if (is_reference_type(dest_elem)) dest_elem = T_OBJECT;
4714
4715 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4716 // If both arrays are object arrays then having the exact types
4717 // for both will remove the need for a subtype check at runtime
4718 // before the call and may make it possible to pick a faster copy
4719 // routine (without a subtype check on every element)
4720 // Do we have the exact type of src?
4721 bool could_have_src = src_spec;
4722 // Do we have the exact type of dest?
4723 bool could_have_dest = dest_spec;
4724 ciKlass* src_k = top_src->klass();
4725 ciKlass* dest_k = top_dest->klass();
4726 if (!src_spec) {
4727 src_k = src_type->speculative_type_not_null();
4728 if (src_k != NULL && src_k->is_array_klass()) {
4729 could_have_src = true;
4730 }
4731 }
4732 if (!dest_spec) {
4733 dest_k = dest_type->speculative_type_not_null();
4734 if (dest_k != NULL && dest_k->is_array_klass()) {
4735 could_have_dest = true;
4736 }
4737 }
4738 if (could_have_src && could_have_dest) {
4739 // If we can have both exact types, emit the missing guards
4740 if (could_have_src && !src_spec) {
4741 src = maybe_cast_profiled_obj(src, src_k, true);
4742 }
4743 if (could_have_dest && !dest_spec) {
4744 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4745 }
4746 }
4747 }
4748 }
4749
4750 ciMethod* trap_method = method();
4751 int trap_bci = bci();
4752 if (saved_jvms != NULL) {
4753 trap_method = alloc->jvms()->method();
4754 trap_bci = alloc->jvms()->bci();
4755 }
4756
4757 bool negative_length_guard_generated = false;
4758
4759 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4760 can_emit_guards &&
4761 !src->is_top() && !dest->is_top()) {
4762 // validate arguments: enables transformation the ArrayCopyNode
4763 validated = true;
4764
4765 RegionNode* slow_region = new RegionNode(1);
4766 record_for_igvn(slow_region);
4767
4768 // (1) src and dest are arrays.
4769 generate_non_array_guard(load_object_klass(src), slow_region);
4770 generate_non_array_guard(load_object_klass(dest), slow_region);
4771
4772 // (2) src and dest arrays must have elements of the same BasicType
4773 // done at macro expansion or at Ideal transformation time
4774
4775 // (4) src_offset must not be negative.
4776 generate_negative_guard(src_offset, slow_region);
4777
4778 // (5) dest_offset must not be negative.
4779 generate_negative_guard(dest_offset, slow_region);
4780
4781 // (7) src_offset + length must not exceed length of src.
4782 generate_limit_guard(src_offset, length,
4783 load_array_length(src),
4784 slow_region);
4785
4786 // (8) dest_offset + length must not exceed length of dest.
4787 generate_limit_guard(dest_offset, length,
4788 load_array_length(dest),
4789 slow_region);
4790
4791 // (6) length must not be negative.
4792 // This is also checked in generate_arraycopy() during macro expansion, but
4793 // we also have to check it here for the case where the ArrayCopyNode will
4794 // be eliminated by Escape Analysis.
4795 if (EliminateAllocations) {
4796 generate_negative_guard(length, slow_region);
4797 negative_length_guard_generated = true;
4798 }
4799
4800 // (9) each element of an oop array must be assignable
4801 Node* dest_klass = load_object_klass(dest);
4802 if (src != dest) {
4803 Node* not_subtype_ctrl = gen_subtype_check(src, dest_klass);
4804
4805 if (not_subtype_ctrl != top()) {
4806 PreserveJVMState pjvms(this);
4807 set_control(not_subtype_ctrl);
4808 uncommon_trap(Deoptimization::Reason_intrinsic,
4809 Deoptimization::Action_make_not_entrant);
4810 assert(stopped(), "Should be stopped");
4811 }
4812 }
4813 {
4814 PreserveJVMState pjvms(this);
4815 set_control(_gvn.transform(slow_region));
4816 uncommon_trap(Deoptimization::Reason_intrinsic,
4817 Deoptimization::Action_make_not_entrant);
4818 assert(stopped(), "Should be stopped");
4819 }
4820
4821 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
4822 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
4823 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
4824 }
4825
4826 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
4827
4828 if (stopped()) {
4829 return true;
4830 }
4831
4832 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
4833 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4834 // so the compiler has a chance to eliminate them: during macro expansion,
4835 // we have to set their control (CastPP nodes are eliminated).
4836 load_object_klass(src), load_object_klass(dest),
4837 load_array_length(src), load_array_length(dest));
4838
4839 ac->set_arraycopy(validated);
4840
4841 Node* n = _gvn.transform(ac);
4842 if (n == ac) {
4843 ac->connect_outputs(this);
4844 } else {
4845 assert(validated, "shouldn't transform if all arguments not validated");
4846 set_all_memory(n);
4847 }
4848 clear_upper_avx();
4849
4850
4851 return true;
4852 }
4853
4854
4855 // Helper function which determines if an arraycopy immediately follows
4856 // an allocation, with no intervening tests or other escapes for the object.
4857 AllocateArrayNode*
4858 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4859 RegionNode* slow_region) {
4860 if (stopped()) return NULL; // no fast path
4861 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4862
4863 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4864 if (alloc == NULL) return NULL;
4865
4866 Node* rawmem = memory(Compile::AliasIdxRaw);
4867 // Is the allocation's memory state untouched?
4868 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4869 // Bail out if there have been raw-memory effects since the allocation.
4870 // (Example: There might have been a call or safepoint.)
4871 return NULL;
4872 }
4873 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4874 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4875 return NULL;
4876 }
4877
4878 // There must be no unexpected observers of this allocation.
4879 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4880 Node* obs = ptr->fast_out(i);
4881 if (obs != this->map()) {
4882 return NULL;
4883 }
4884 }
4885
4886 // This arraycopy must unconditionally follow the allocation of the ptr.
4887 Node* alloc_ctl = ptr->in(0);
4888 Node* ctl = control();
4889 while (ctl != alloc_ctl) {
4890 // There may be guards which feed into the slow_region.
4891 // Any other control flow means that we might not get a chance
4892 // to finish initializing the allocated object.
4893 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4894 IfNode* iff = ctl->in(0)->as_If();
4895 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
4896 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4897 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4898 ctl = iff->in(0); // This test feeds the known slow_region.
4899 continue;
4900 }
4901 // One more try: Various low-level checks bottom out in
4902 // uncommon traps. If the debug-info of the trap omits
4903 // any reference to the allocation, as we've already
4904 // observed, then there can be no objection to the trap.
4905 bool found_trap = false;
4906 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4907 Node* obs = not_ctl->fast_out(j);
4908 if (obs->in(0) == not_ctl && obs->is_Call() &&
4909 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4910 found_trap = true; break;
4911 }
4912 }
4913 if (found_trap) {
4914 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4915 continue;
4916 }
4917 }
4918 return NULL;
4919 }
4920
4921 // If we get this far, we have an allocation which immediately
4922 // precedes the arraycopy, and we can take over zeroing the new object.
4923 // The arraycopy will finish the initialization, and provide
4924 // a new control state to which we will anchor the destination pointer.
4925
4926 return alloc;
4927 }
4928
4929 //-------------inline_encodeISOArray-----------------------------------
4930 // encode char[] to byte[] in ISO_8859_1
4931 bool LibraryCallKit::inline_encodeISOArray() {
4932 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4933 // no receiver since it is static method
4934 Node *src = argument(0);
4935 Node *src_offset = argument(1);
4936 Node *dst = argument(2);
4937 Node *dst_offset = argument(3);
4938 Node *length = argument(4);
4939
4940 src = must_be_not_null(src, true);
4941 dst = must_be_not_null(dst, true);
4942
4943 const Type* src_type = src->Value(&_gvn);
4944 const Type* dst_type = dst->Value(&_gvn);
4945 const TypeAryPtr* top_src = src_type->isa_aryptr();
4946 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4947 if (top_src == NULL || top_src->klass() == NULL ||
4948 top_dest == NULL || top_dest->klass() == NULL) {
4949 // failed array check
4950 return false;
4951 }
4952
4953 // Figure out the size and type of the elements we will be copying.
4954 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4955 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4956 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
4957 return false;
4958 }
4959
4960 Node* src_start = array_element_address(src, src_offset, T_CHAR);
4961 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
4962 // 'src_start' points to src array + scaled offset
4963 // 'dst_start' points to dst array + scaled offset
4964
4965 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
4966 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
4967 enc = _gvn.transform(enc);
4968 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
4969 set_memory(res_mem, mtype);
4970 set_result(enc);
4971 clear_upper_avx();
4972
4973 return true;
4974 }
4975
4976 //-------------inline_multiplyToLen-----------------------------------
4977 bool LibraryCallKit::inline_multiplyToLen() {
4978 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
4979
4980 address stubAddr = StubRoutines::multiplyToLen();
4981 if (stubAddr == NULL) {
4982 return false; // Intrinsic's stub is not implemented on this platform
4983 }
4984 const char* stubName = "multiplyToLen";
4985
4986 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
4987
4988 // no receiver because it is a static method
4989 Node* x = argument(0);
4990 Node* xlen = argument(1);
4991 Node* y = argument(2);
4992 Node* ylen = argument(3);
4993 Node* z = argument(4);
4994
4995 x = must_be_not_null(x, true);
4996 y = must_be_not_null(y, true);
4997
4998 const Type* x_type = x->Value(&_gvn);
4999 const Type* y_type = y->Value(&_gvn);
5000 const TypeAryPtr* top_x = x_type->isa_aryptr();
5001 const TypeAryPtr* top_y = y_type->isa_aryptr();
5002 if (top_x == NULL || top_x->klass() == NULL ||
5003 top_y == NULL || top_y->klass() == NULL) {
5004 // failed array check
5005 return false;
5006 }
5007
5008 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5009 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5010 if (x_elem != T_INT || y_elem != T_INT) {
5011 return false;
5012 }
5013
5014 // Set the original stack and the reexecute bit for the interpreter to reexecute
5015 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5016 // on the return from z array allocation in runtime.
5017 { PreserveReexecuteState preexecs(this);
5018 jvms()->set_should_reexecute(true);
5019
5020 Node* x_start = array_element_address(x, intcon(0), x_elem);
5021 Node* y_start = array_element_address(y, intcon(0), y_elem);
5022 // 'x_start' points to x array + scaled xlen
5023 // 'y_start' points to y array + scaled ylen
5024
5025 // Allocate the result array
5026 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5027 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5028 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5029
5030 IdealKit ideal(this);
5031
5032 #define __ ideal.
5033 Node* one = __ ConI(1);
5034 Node* zero = __ ConI(0);
5035 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5036 __ set(need_alloc, zero);
5037 __ set(z_alloc, z);
5038 __ if_then(z, BoolTest::eq, null()); {
5039 __ increment (need_alloc, one);
5040 } __ else_(); {
5041 // Update graphKit memory and control from IdealKit.
5042 sync_kit(ideal);
5043 Node *cast = new CastPPNode(z, TypePtr::NOTNULL);
5044 cast->init_req(0, control());
5045 _gvn.set_type(cast, cast->bottom_type());
5046 C->record_for_igvn(cast);
5047
5048 Node* zlen_arg = load_array_length(cast);
5049 // Update IdealKit memory and control from graphKit.
5050 __ sync_kit(this);
5051 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5052 __ increment (need_alloc, one);
5053 } __ end_if();
5054 } __ end_if();
5055
5056 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5057 // Update graphKit memory and control from IdealKit.
5058 sync_kit(ideal);
5059 Node * narr = new_array(klass_node, zlen, 1);
5060 // Update IdealKit memory and control from graphKit.
5061 __ sync_kit(this);
5062 __ set(z_alloc, narr);
5063 } __ end_if();
5064
5065 sync_kit(ideal);
5066 z = __ value(z_alloc);
5067 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5068 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5069 // Final sync IdealKit and GraphKit.
5070 final_sync(ideal);
5071 #undef __
5072
5073 Node* z_start = array_element_address(z, intcon(0), T_INT);
5074
5075 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5076 OptoRuntime::multiplyToLen_Type(),
5077 stubAddr, stubName, TypePtr::BOTTOM,
5078 x_start, xlen, y_start, ylen, z_start, zlen);
5079 } // original reexecute is set back here
5080
5081 C->set_has_split_ifs(true); // Has chance for split-if optimization
5082 set_result(z);
5083 return true;
5084 }
5085
5086 //-------------inline_squareToLen------------------------------------
5087 bool LibraryCallKit::inline_squareToLen() {
5088 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5089
5090 address stubAddr = StubRoutines::squareToLen();
5091 if (stubAddr == NULL) {
5092 return false; // Intrinsic's stub is not implemented on this platform
5093 }
5094 const char* stubName = "squareToLen";
5095
5096 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5097
5098 Node* x = argument(0);
5099 Node* len = argument(1);
5100 Node* z = argument(2);
5101 Node* zlen = argument(3);
5102
5103 x = must_be_not_null(x, true);
5104 z = must_be_not_null(z, true);
5105
5106 const Type* x_type = x->Value(&_gvn);
5107 const Type* z_type = z->Value(&_gvn);
5108 const TypeAryPtr* top_x = x_type->isa_aryptr();
5109 const TypeAryPtr* top_z = z_type->isa_aryptr();
5110 if (top_x == NULL || top_x->klass() == NULL ||
5111 top_z == NULL || top_z->klass() == NULL) {
5112 // failed array check
5113 return false;
5114 }
5115
5116 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5117 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5118 if (x_elem != T_INT || z_elem != T_INT) {
5119 return false;
5120 }
5121
5122
5123 Node* x_start = array_element_address(x, intcon(0), x_elem);
5124 Node* z_start = array_element_address(z, intcon(0), z_elem);
5125
5126 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5127 OptoRuntime::squareToLen_Type(),
5128 stubAddr, stubName, TypePtr::BOTTOM,
5129 x_start, len, z_start, zlen);
5130
5131 set_result(z);
5132 return true;
5133 }
5134
5135 //-------------inline_mulAdd------------------------------------------
5136 bool LibraryCallKit::inline_mulAdd() {
5137 assert(UseMulAddIntrinsic, "not implemented on this platform");
5138
5139 address stubAddr = StubRoutines::mulAdd();
5140 if (stubAddr == NULL) {
5141 return false; // Intrinsic's stub is not implemented on this platform
5142 }
5143 const char* stubName = "mulAdd";
5144
5145 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5146
5147 Node* out = argument(0);
5148 Node* in = argument(1);
5149 Node* offset = argument(2);
5150 Node* len = argument(3);
5151 Node* k = argument(4);
5152
5153 out = must_be_not_null(out, true);
5154
5155 const Type* out_type = out->Value(&_gvn);
5156 const Type* in_type = in->Value(&_gvn);
5157 const TypeAryPtr* top_out = out_type->isa_aryptr();
5158 const TypeAryPtr* top_in = in_type->isa_aryptr();
5159 if (top_out == NULL || top_out->klass() == NULL ||
5160 top_in == NULL || top_in->klass() == NULL) {
5161 // failed array check
5162 return false;
5163 }
5164
5165 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5166 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5167 if (out_elem != T_INT || in_elem != T_INT) {
5168 return false;
5169 }
5170
5171 Node* outlen = load_array_length(out);
5172 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5173 Node* out_start = array_element_address(out, intcon(0), out_elem);
5174 Node* in_start = array_element_address(in, intcon(0), in_elem);
5175
5176 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5177 OptoRuntime::mulAdd_Type(),
5178 stubAddr, stubName, TypePtr::BOTTOM,
5179 out_start,in_start, new_offset, len, k);
5180 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5181 set_result(result);
5182 return true;
5183 }
5184
5185 //-------------inline_montgomeryMultiply-----------------------------------
5186 bool LibraryCallKit::inline_montgomeryMultiply() {
5187 address stubAddr = StubRoutines::montgomeryMultiply();
5188 if (stubAddr == NULL) {
5189 return false; // Intrinsic's stub is not implemented on this platform
5190 }
5191
5192 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5193 const char* stubName = "montgomery_multiply";
5194
5195 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5196
5197 Node* a = argument(0);
5198 Node* b = argument(1);
5199 Node* n = argument(2);
5200 Node* len = argument(3);
5201 Node* inv = argument(4);
5202 Node* m = argument(6);
5203
5204 const Type* a_type = a->Value(&_gvn);
5205 const TypeAryPtr* top_a = a_type->isa_aryptr();
5206 const Type* b_type = b->Value(&_gvn);
5207 const TypeAryPtr* top_b = b_type->isa_aryptr();
5208 const Type* n_type = a->Value(&_gvn);
5209 const TypeAryPtr* top_n = n_type->isa_aryptr();
5210 const Type* m_type = a->Value(&_gvn);
5211 const TypeAryPtr* top_m = m_type->isa_aryptr();
5212 if (top_a == NULL || top_a->klass() == NULL ||
5213 top_b == NULL || top_b->klass() == NULL ||
5214 top_n == NULL || top_n->klass() == NULL ||
5215 top_m == NULL || top_m->klass() == NULL) {
5216 // failed array check
5217 return false;
5218 }
5219
5220 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5221 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5222 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5223 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5224 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5225 return false;
5226 }
5227
5228 // Make the call
5229 {
5230 Node* a_start = array_element_address(a, intcon(0), a_elem);
5231 Node* b_start = array_element_address(b, intcon(0), b_elem);
5232 Node* n_start = array_element_address(n, intcon(0), n_elem);
5233 Node* m_start = array_element_address(m, intcon(0), m_elem);
5234
5235 Node* call = make_runtime_call(RC_LEAF,
5236 OptoRuntime::montgomeryMultiply_Type(),
5237 stubAddr, stubName, TypePtr::BOTTOM,
5238 a_start, b_start, n_start, len, inv, top(),
5239 m_start);
5240 set_result(m);
5241 }
5242
5243 return true;
5244 }
5245
5246 bool LibraryCallKit::inline_montgomerySquare() {
5247 address stubAddr = StubRoutines::montgomerySquare();
5248 if (stubAddr == NULL) {
5249 return false; // Intrinsic's stub is not implemented on this platform
5250 }
5251
5252 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5253 const char* stubName = "montgomery_square";
5254
5255 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5256
5257 Node* a = argument(0);
5258 Node* n = argument(1);
5259 Node* len = argument(2);
5260 Node* inv = argument(3);
5261 Node* m = argument(5);
5262
5263 const Type* a_type = a->Value(&_gvn);
5264 const TypeAryPtr* top_a = a_type->isa_aryptr();
5265 const Type* n_type = a->Value(&_gvn);
5266 const TypeAryPtr* top_n = n_type->isa_aryptr();
5267 const Type* m_type = a->Value(&_gvn);
5268 const TypeAryPtr* top_m = m_type->isa_aryptr();
5269 if (top_a == NULL || top_a->klass() == NULL ||
5270 top_n == NULL || top_n->klass() == NULL ||
5271 top_m == NULL || top_m->klass() == NULL) {
5272 // failed array check
5273 return false;
5274 }
5275
5276 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5277 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5278 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5279 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5280 return false;
5281 }
5282
5283 // Make the call
5284 {
5285 Node* a_start = array_element_address(a, intcon(0), a_elem);
5286 Node* n_start = array_element_address(n, intcon(0), n_elem);
5287 Node* m_start = array_element_address(m, intcon(0), m_elem);
5288
5289 Node* call = make_runtime_call(RC_LEAF,
5290 OptoRuntime::montgomerySquare_Type(),
5291 stubAddr, stubName, TypePtr::BOTTOM,
5292 a_start, n_start, len, inv, top(),
5293 m_start);
5294 set_result(m);
5295 }
5296
5297 return true;
5298 }
5299
5300 bool LibraryCallKit::inline_bigIntegerShift(bool isRightShift) {
5301 address stubAddr = NULL;
5302 const char* stubName = NULL;
5303
5304 stubAddr = isRightShift? StubRoutines::bigIntegerRightShift(): StubRoutines::bigIntegerLeftShift();
5305 if (stubAddr == NULL) {
5306 return false; // Intrinsic's stub is not implemented on this platform
5307 }
5308
5309 stubName = isRightShift? "bigIntegerRightShiftWorker" : "bigIntegerLeftShiftWorker";
5310
5311 assert(callee()->signature()->size() == 5, "expected 5 arguments");
5312
5313 Node* newArr = argument(0);
5314 Node* oldArr = argument(1);
5315 Node* newIdx = argument(2);
5316 Node* shiftCount = argument(3);
5317 Node* numIter = argument(4);
5318
5319 const Type* newArr_type = newArr->Value(&_gvn);
5320 const TypeAryPtr* top_newArr = newArr_type->isa_aryptr();
5321 const Type* oldArr_type = oldArr->Value(&_gvn);
5322 const TypeAryPtr* top_oldArr = oldArr_type->isa_aryptr();
5323 if (top_newArr == NULL || top_newArr->klass() == NULL || top_oldArr == NULL
5324 || top_oldArr->klass() == NULL) {
5325 return false;
5326 }
5327
5328 BasicType newArr_elem = newArr_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5329 BasicType oldArr_elem = oldArr_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5330 if (newArr_elem != T_INT || oldArr_elem != T_INT) {
5331 return false;
5332 }
5333
5334 // Make the call
5335 {
5336 Node* newArr_start = array_element_address(newArr, intcon(0), newArr_elem);
5337 Node* oldArr_start = array_element_address(oldArr, intcon(0), oldArr_elem);
5338
5339 Node* call = make_runtime_call(RC_LEAF,
5340 OptoRuntime::bigIntegerShift_Type(),
5341 stubAddr,
5342 stubName,
5343 TypePtr::BOTTOM,
5344 newArr_start,
5345 oldArr_start,
5346 newIdx,
5347 shiftCount,
5348 numIter);
5349 }
5350
5351 return true;
5352 }
5353
5354 //-------------inline_vectorizedMismatch------------------------------
5355 bool LibraryCallKit::inline_vectorizedMismatch() {
5356 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5357
5358 address stubAddr = StubRoutines::vectorizedMismatch();
5359 if (stubAddr == NULL) {
5360 return false; // Intrinsic's stub is not implemented on this platform
5361 }
5362 const char* stubName = "vectorizedMismatch";
5363 int size_l = callee()->signature()->size();
5364 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5365
5366 Node* obja = argument(0);
5367 Node* aoffset = argument(1);
5368 Node* objb = argument(3);
5369 Node* boffset = argument(4);
5370 Node* length = argument(6);
5371 Node* scale = argument(7);
5372
5373 const Type* a_type = obja->Value(&_gvn);
5374 const Type* b_type = objb->Value(&_gvn);
5375 const TypeAryPtr* top_a = a_type->isa_aryptr();
5376 const TypeAryPtr* top_b = b_type->isa_aryptr();
5377 if (top_a == NULL || top_a->klass() == NULL ||
5378 top_b == NULL || top_b->klass() == NULL) {
5379 // failed array check
5380 return false;
5381 }
5382
5383 Node* call;
5384 jvms()->set_should_reexecute(true);
5385
5386 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ);
5387 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ);
5388
5389 call = make_runtime_call(RC_LEAF,
5390 OptoRuntime::vectorizedMismatch_Type(),
5391 stubAddr, stubName, TypePtr::BOTTOM,
5392 obja_adr, objb_adr, length, scale);
5393
5394 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5395 set_result(result);
5396 return true;
5397 }
5398
5399 /**
5400 * Calculate CRC32 for byte.
5401 * int java.util.zip.CRC32.update(int crc, int b)
5402 */
5403 bool LibraryCallKit::inline_updateCRC32() {
5404 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5405 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5406 // no receiver since it is static method
5407 Node* crc = argument(0); // type: int
5408 Node* b = argument(1); // type: int
5409
5410 /*
5411 * int c = ~ crc;
5412 * b = timesXtoThe32[(b ^ c) & 0xFF];
5413 * b = b ^ (c >>> 8);
5414 * crc = ~b;
5415 */
5416
5417 Node* M1 = intcon(-1);
5418 crc = _gvn.transform(new XorINode(crc, M1));
5419 Node* result = _gvn.transform(new XorINode(crc, b));
5420 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5421
5422 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5423 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5424 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5425 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5426
5427 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5428 result = _gvn.transform(new XorINode(crc, result));
5429 result = _gvn.transform(new XorINode(result, M1));
5430 set_result(result);
5431 return true;
5432 }
5433
5434 /**
5435 * Calculate CRC32 for byte[] array.
5436 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5437 */
5438 bool LibraryCallKit::inline_updateBytesCRC32() {
5439 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5440 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5441 // no receiver since it is static method
5442 Node* crc = argument(0); // type: int
5443 Node* src = argument(1); // type: oop
5444 Node* offset = argument(2); // type: int
5445 Node* length = argument(3); // type: int
5446
5447 const Type* src_type = src->Value(&_gvn);
5448 const TypeAryPtr* top_src = src_type->isa_aryptr();
5449 if (top_src == NULL || top_src->klass() == NULL) {
5450 // failed array check
5451 return false;
5452 }
5453
5454 // Figure out the size and type of the elements we will be copying.
5455 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5456 if (src_elem != T_BYTE) {
5457 return false;
5458 }
5459
5460 // 'src_start' points to src array + scaled offset
5461 src = must_be_not_null(src, true);
5462 Node* src_start = array_element_address(src, offset, src_elem);
5463
5464 // We assume that range check is done by caller.
5465 // TODO: generate range check (offset+length < src.length) in debug VM.
5466
5467 // Call the stub.
5468 address stubAddr = StubRoutines::updateBytesCRC32();
5469 const char *stubName = "updateBytesCRC32";
5470
5471 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5472 stubAddr, stubName, TypePtr::BOTTOM,
5473 crc, src_start, length);
5474 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5475 set_result(result);
5476 return true;
5477 }
5478
5479 /**
5480 * Calculate CRC32 for ByteBuffer.
5481 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5482 */
5483 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5484 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5485 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5486 // no receiver since it is static method
5487 Node* crc = argument(0); // type: int
5488 Node* src = argument(1); // type: long
5489 Node* offset = argument(3); // type: int
5490 Node* length = argument(4); // type: int
5491
5492 src = ConvL2X(src); // adjust Java long to machine word
5493 Node* base = _gvn.transform(new CastX2PNode(src));
5494 offset = ConvI2X(offset);
5495
5496 // 'src_start' points to src array + scaled offset
5497 Node* src_start = basic_plus_adr(top(), base, offset);
5498
5499 // Call the stub.
5500 address stubAddr = StubRoutines::updateBytesCRC32();
5501 const char *stubName = "updateBytesCRC32";
5502
5503 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5504 stubAddr, stubName, TypePtr::BOTTOM,
5505 crc, src_start, length);
5506 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5507 set_result(result);
5508 return true;
5509 }
5510
5511 //------------------------------get_table_from_crc32c_class-----------------------
5512 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5513 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5514 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5515
5516 return table;
5517 }
5518
5519 //------------------------------inline_updateBytesCRC32C-----------------------
5520 //
5521 // Calculate CRC32C for byte[] array.
5522 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5523 //
5524 bool LibraryCallKit::inline_updateBytesCRC32C() {
5525 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5526 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5527 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5528 // no receiver since it is a static method
5529 Node* crc = argument(0); // type: int
5530 Node* src = argument(1); // type: oop
5531 Node* offset = argument(2); // type: int
5532 Node* end = argument(3); // type: int
5533
5534 Node* length = _gvn.transform(new SubINode(end, offset));
5535
5536 const Type* src_type = src->Value(&_gvn);
5537 const TypeAryPtr* top_src = src_type->isa_aryptr();
5538 if (top_src == NULL || top_src->klass() == NULL) {
5539 // failed array check
5540 return false;
5541 }
5542
5543 // Figure out the size and type of the elements we will be copying.
5544 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5545 if (src_elem != T_BYTE) {
5546 return false;
5547 }
5548
5549 // 'src_start' points to src array + scaled offset
5550 src = must_be_not_null(src, true);
5551 Node* src_start = array_element_address(src, offset, src_elem);
5552
5553 // static final int[] byteTable in class CRC32C
5554 Node* table = get_table_from_crc32c_class(callee()->holder());
5555 table = must_be_not_null(table, true);
5556 Node* table_start = array_element_address(table, intcon(0), T_INT);
5557
5558 // We assume that range check is done by caller.
5559 // TODO: generate range check (offset+length < src.length) in debug VM.
5560
5561 // Call the stub.
5562 address stubAddr = StubRoutines::updateBytesCRC32C();
5563 const char *stubName = "updateBytesCRC32C";
5564
5565 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5566 stubAddr, stubName, TypePtr::BOTTOM,
5567 crc, src_start, length, table_start);
5568 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5569 set_result(result);
5570 return true;
5571 }
5572
5573 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5574 //
5575 // Calculate CRC32C for DirectByteBuffer.
5576 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5577 //
5578 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5579 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5580 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5581 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5582 // no receiver since it is a static method
5583 Node* crc = argument(0); // type: int
5584 Node* src = argument(1); // type: long
5585 Node* offset = argument(3); // type: int
5586 Node* end = argument(4); // type: int
5587
5588 Node* length = _gvn.transform(new SubINode(end, offset));
5589
5590 src = ConvL2X(src); // adjust Java long to machine word
5591 Node* base = _gvn.transform(new CastX2PNode(src));
5592 offset = ConvI2X(offset);
5593
5594 // 'src_start' points to src array + scaled offset
5595 Node* src_start = basic_plus_adr(top(), base, offset);
5596
5597 // static final int[] byteTable in class CRC32C
5598 Node* table = get_table_from_crc32c_class(callee()->holder());
5599 table = must_be_not_null(table, true);
5600 Node* table_start = array_element_address(table, intcon(0), T_INT);
5601
5602 // Call the stub.
5603 address stubAddr = StubRoutines::updateBytesCRC32C();
5604 const char *stubName = "updateBytesCRC32C";
5605
5606 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5607 stubAddr, stubName, TypePtr::BOTTOM,
5608 crc, src_start, length, table_start);
5609 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5610 set_result(result);
5611 return true;
5612 }
5613
5614 //------------------------------inline_updateBytesAdler32----------------------
5615 //
5616 // Calculate Adler32 checksum for byte[] array.
5617 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5618 //
5619 bool LibraryCallKit::inline_updateBytesAdler32() {
5620 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5621 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5622 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5623 // no receiver since it is static method
5624 Node* crc = argument(0); // type: int
5625 Node* src = argument(1); // type: oop
5626 Node* offset = argument(2); // type: int
5627 Node* length = argument(3); // type: int
5628
5629 const Type* src_type = src->Value(&_gvn);
5630 const TypeAryPtr* top_src = src_type->isa_aryptr();
5631 if (top_src == NULL || top_src->klass() == NULL) {
5632 // failed array check
5633 return false;
5634 }
5635
5636 // Figure out the size and type of the elements we will be copying.
5637 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5638 if (src_elem != T_BYTE) {
5639 return false;
5640 }
5641
5642 // 'src_start' points to src array + scaled offset
5643 Node* src_start = array_element_address(src, offset, src_elem);
5644
5645 // We assume that range check is done by caller.
5646 // TODO: generate range check (offset+length < src.length) in debug VM.
5647
5648 // Call the stub.
5649 address stubAddr = StubRoutines::updateBytesAdler32();
5650 const char *stubName = "updateBytesAdler32";
5651
5652 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5653 stubAddr, stubName, TypePtr::BOTTOM,
5654 crc, src_start, length);
5655 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5656 set_result(result);
5657 return true;
5658 }
5659
5660 //------------------------------inline_updateByteBufferAdler32---------------
5661 //
5662 // Calculate Adler32 checksum for DirectByteBuffer.
5663 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5664 //
5665 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5666 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5667 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5668 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5669 // no receiver since it is static method
5670 Node* crc = argument(0); // type: int
5671 Node* src = argument(1); // type: long
5672 Node* offset = argument(3); // type: int
5673 Node* length = argument(4); // type: int
5674
5675 src = ConvL2X(src); // adjust Java long to machine word
5676 Node* base = _gvn.transform(new CastX2PNode(src));
5677 offset = ConvI2X(offset);
5678
5679 // 'src_start' points to src array + scaled offset
5680 Node* src_start = basic_plus_adr(top(), base, offset);
5681
5682 // Call the stub.
5683 address stubAddr = StubRoutines::updateBytesAdler32();
5684 const char *stubName = "updateBytesAdler32";
5685
5686 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5687 stubAddr, stubName, TypePtr::BOTTOM,
5688 crc, src_start, length);
5689
5690 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5691 set_result(result);
5692 return true;
5693 }
5694
5695 //----------------------------inline_reference_get----------------------------
5696 // public T java.lang.ref.Reference.get();
5697 bool LibraryCallKit::inline_reference_get() {
5698 const int referent_offset = java_lang_ref_Reference::referent_offset();
5699
5700 // Get the argument:
5701 Node* reference_obj = null_check_receiver();
5702 if (stopped()) return true;
5703
5704 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr();
5705 assert(tinst != NULL, "obj is null");
5706 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5707 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass();
5708 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"),
5709 ciSymbol::make("Ljava/lang/Object;"),
5710 false);
5711 assert (field != NULL, "undefined field");
5712
5713 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5714 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5715
5716 ciInstanceKlass* klass = env()->Object_klass();
5717 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5718
5719 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF;
5720 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators);
5721 // Add memory barrier to prevent commoning reads from this field
5722 // across safepoint since GC can change its value.
5723 insert_mem_bar(Op_MemBarCPUOrder);
5724
5725 set_result(result);
5726 return true;
5727 }
5728
5729
5730 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5731 bool is_exact=true, bool is_static=false,
5732 ciInstanceKlass * fromKls=NULL) {
5733 if (fromKls == NULL) {
5734 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5735 assert(tinst != NULL, "obj is null");
5736 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5737 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5738 fromKls = tinst->klass()->as_instance_klass();
5739 } else {
5740 assert(is_static, "only for static field access");
5741 }
5742 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5743 ciSymbol::make(fieldTypeString),
5744 is_static);
5745
5746 assert (field != NULL, "undefined field");
5747 if (field == NULL) return (Node *) NULL;
5748
5749 if (is_static) {
5750 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5751 fromObj = makecon(tip);
5752 }
5753
5754 // Next code copied from Parse::do_get_xxx():
5755
5756 // Compute address and memory type.
5757 int offset = field->offset_in_bytes();
5758 bool is_vol = field->is_volatile();
5759 ciType* field_klass = field->type();
5760 assert(field_klass->is_loaded(), "should be loaded");
5761 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5762 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5763 BasicType bt = field->layout_type();
5764
5765 // Build the resultant type of the load
5766 const Type *type;
5767 if (bt == T_OBJECT) {
5768 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5769 } else {
5770 type = Type::get_const_basic_type(bt);
5771 }
5772
5773 DecoratorSet decorators = IN_HEAP;
5774
5775 if (is_vol) {
5776 decorators |= MO_SEQ_CST;
5777 }
5778
5779 return access_load_at(fromObj, adr, adr_type, type, bt, decorators);
5780 }
5781
5782 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5783 bool is_exact = true, bool is_static = false,
5784 ciInstanceKlass * fromKls = NULL) {
5785 if (fromKls == NULL) {
5786 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5787 assert(tinst != NULL, "obj is null");
5788 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5789 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5790 fromKls = tinst->klass()->as_instance_klass();
5791 }
5792 else {
5793 assert(is_static, "only for static field access");
5794 }
5795 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5796 ciSymbol::make(fieldTypeString),
5797 is_static);
5798
5799 assert(field != NULL, "undefined field");
5800 assert(!field->is_volatile(), "not defined for volatile fields");
5801
5802 if (is_static) {
5803 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5804 fromObj = makecon(tip);
5805 }
5806
5807 // Next code copied from Parse::do_get_xxx():
5808
5809 // Compute address and memory type.
5810 int offset = field->offset_in_bytes();
5811 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5812
5813 return adr;
5814 }
5815
5816 //------------------------------inline_aescrypt_Block-----------------------
5817 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5818 address stubAddr = NULL;
5819 const char *stubName;
5820 assert(UseAES, "need AES instruction support");
5821
5822 switch(id) {
5823 case vmIntrinsics::_aescrypt_encryptBlock:
5824 stubAddr = StubRoutines::aescrypt_encryptBlock();
5825 stubName = "aescrypt_encryptBlock";
5826 break;
5827 case vmIntrinsics::_aescrypt_decryptBlock:
5828 stubAddr = StubRoutines::aescrypt_decryptBlock();
5829 stubName = "aescrypt_decryptBlock";
5830 break;
5831 default:
5832 break;
5833 }
5834 if (stubAddr == NULL) return false;
5835
5836 Node* aescrypt_object = argument(0);
5837 Node* src = argument(1);
5838 Node* src_offset = argument(2);
5839 Node* dest = argument(3);
5840 Node* dest_offset = argument(4);
5841
5842 src = must_be_not_null(src, true);
5843 dest = must_be_not_null(dest, true);
5844
5845 // (1) src and dest are arrays.
5846 const Type* src_type = src->Value(&_gvn);
5847 const Type* dest_type = dest->Value(&_gvn);
5848 const TypeAryPtr* top_src = src_type->isa_aryptr();
5849 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5850 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5851
5852 // for the quick and dirty code we will skip all the checks.
5853 // we are just trying to get the call to be generated.
5854 Node* src_start = src;
5855 Node* dest_start = dest;
5856 if (src_offset != NULL || dest_offset != NULL) {
5857 assert(src_offset != NULL && dest_offset != NULL, "");
5858 src_start = array_element_address(src, src_offset, T_BYTE);
5859 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5860 }
5861
5862 // now need to get the start of its expanded key array
5863 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5864 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5865 if (k_start == NULL) return false;
5866
5867 if (Matcher::pass_original_key_for_aes()) {
5868 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5869 // compatibility issues between Java key expansion and SPARC crypto instructions
5870 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5871 if (original_k_start == NULL) return false;
5872
5873 // Call the stub.
5874 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5875 stubAddr, stubName, TypePtr::BOTTOM,
5876 src_start, dest_start, k_start, original_k_start);
5877 } else {
5878 // Call the stub.
5879 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5880 stubAddr, stubName, TypePtr::BOTTOM,
5881 src_start, dest_start, k_start);
5882 }
5883
5884 return true;
5885 }
5886
5887 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5888 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5889 address stubAddr = NULL;
5890 const char *stubName = NULL;
5891
5892 assert(UseAES, "need AES instruction support");
5893
5894 switch(id) {
5895 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5896 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5897 stubName = "cipherBlockChaining_encryptAESCrypt";
5898 break;
5899 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5900 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5901 stubName = "cipherBlockChaining_decryptAESCrypt";
5902 break;
5903 default:
5904 break;
5905 }
5906 if (stubAddr == NULL) return false;
5907
5908 Node* cipherBlockChaining_object = argument(0);
5909 Node* src = argument(1);
5910 Node* src_offset = argument(2);
5911 Node* len = argument(3);
5912 Node* dest = argument(4);
5913 Node* dest_offset = argument(5);
5914
5915 src = must_be_not_null(src, false);
5916 dest = must_be_not_null(dest, false);
5917
5918 // (1) src and dest are arrays.
5919 const Type* src_type = src->Value(&_gvn);
5920 const Type* dest_type = dest->Value(&_gvn);
5921 const TypeAryPtr* top_src = src_type->isa_aryptr();
5922 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5923 assert (top_src != NULL && top_src->klass() != NULL
5924 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5925
5926 // checks are the responsibility of the caller
5927 Node* src_start = src;
5928 Node* dest_start = dest;
5929 if (src_offset != NULL || dest_offset != NULL) {
5930 assert(src_offset != NULL && dest_offset != NULL, "");
5931 src_start = array_element_address(src, src_offset, T_BYTE);
5932 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5933 }
5934
5935 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5936 // (because of the predicated logic executed earlier).
5937 // so we cast it here safely.
5938 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5939
5940 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5941 if (embeddedCipherObj == NULL) return false;
5942
5943 // cast it to what we know it will be at runtime
5944 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5945 assert(tinst != NULL, "CBC obj is null");
5946 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5947 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5948 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5949
5950 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5951 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5952 const TypeOopPtr* xtype = aklass->as_instance_type();
5953 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5954 aescrypt_object = _gvn.transform(aescrypt_object);
5955
5956 // we need to get the start of the aescrypt_object's expanded key array
5957 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5958 if (k_start == NULL) return false;
5959
5960 // similarly, get the start address of the r vector
5961 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5962 if (objRvec == NULL) return false;
5963 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5964
5965 Node* cbcCrypt;
5966 if (Matcher::pass_original_key_for_aes()) {
5967 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5968 // compatibility issues between Java key expansion and SPARC crypto instructions
5969 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5970 if (original_k_start == NULL) return false;
5971
5972 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5973 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5974 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5975 stubAddr, stubName, TypePtr::BOTTOM,
5976 src_start, dest_start, k_start, r_start, len, original_k_start);
5977 } else {
5978 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5979 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5980 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5981 stubAddr, stubName, TypePtr::BOTTOM,
5982 src_start, dest_start, k_start, r_start, len);
5983 }
5984
5985 // return cipher length (int)
5986 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5987 set_result(retvalue);
5988 return true;
5989 }
5990
5991 //------------------------------inline_electronicCodeBook_AESCrypt-----------------------
5992 bool LibraryCallKit::inline_electronicCodeBook_AESCrypt(vmIntrinsics::ID id) {
5993 address stubAddr = NULL;
5994 const char *stubName = NULL;
5995
5996 assert(UseAES, "need AES instruction support");
5997
5998 switch (id) {
5999 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt:
6000 stubAddr = StubRoutines::electronicCodeBook_encryptAESCrypt();
6001 stubName = "electronicCodeBook_encryptAESCrypt";
6002 break;
6003 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt:
6004 stubAddr = StubRoutines::electronicCodeBook_decryptAESCrypt();
6005 stubName = "electronicCodeBook_decryptAESCrypt";
6006 break;
6007 default:
6008 break;
6009 }
6010
6011 if (stubAddr == NULL) return false;
6012
6013 Node* electronicCodeBook_object = argument(0);
6014 Node* src = argument(1);
6015 Node* src_offset = argument(2);
6016 Node* len = argument(3);
6017 Node* dest = argument(4);
6018 Node* dest_offset = argument(5);
6019
6020 // (1) src and dest are arrays.
6021 const Type* src_type = src->Value(&_gvn);
6022 const Type* dest_type = dest->Value(&_gvn);
6023 const TypeAryPtr* top_src = src_type->isa_aryptr();
6024 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6025 assert(top_src != NULL && top_src->klass() != NULL
6026 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6027
6028 // checks are the responsibility of the caller
6029 Node* src_start = src;
6030 Node* dest_start = dest;
6031 if (src_offset != NULL || dest_offset != NULL) {
6032 assert(src_offset != NULL && dest_offset != NULL, "");
6033 src_start = array_element_address(src, src_offset, T_BYTE);
6034 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6035 }
6036
6037 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6038 // (because of the predicated logic executed earlier).
6039 // so we cast it here safely.
6040 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6041
6042 Node* embeddedCipherObj = load_field_from_object(electronicCodeBook_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6043 if (embeddedCipherObj == NULL) return false;
6044
6045 // cast it to what we know it will be at runtime
6046 const TypeInstPtr* tinst = _gvn.type(electronicCodeBook_object)->isa_instptr();
6047 assert(tinst != NULL, "ECB obj is null");
6048 assert(tinst->klass()->is_loaded(), "ECB obj is not loaded");
6049 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6050 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6051
6052 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6053 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6054 const TypeOopPtr* xtype = aklass->as_instance_type();
6055 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6056 aescrypt_object = _gvn.transform(aescrypt_object);
6057
6058 // we need to get the start of the aescrypt_object's expanded key array
6059 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6060 if (k_start == NULL) return false;
6061
6062 Node* ecbCrypt;
6063 if (Matcher::pass_original_key_for_aes()) {
6064 // no SPARC version for AES/ECB intrinsics now.
6065 return false;
6066 }
6067 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6068 ecbCrypt = make_runtime_call(RC_LEAF | RC_NO_FP,
6069 OptoRuntime::electronicCodeBook_aescrypt_Type(),
6070 stubAddr, stubName, TypePtr::BOTTOM,
6071 src_start, dest_start, k_start, len);
6072
6073 // return cipher length (int)
6074 Node* retvalue = _gvn.transform(new ProjNode(ecbCrypt, TypeFunc::Parms));
6075 set_result(retvalue);
6076 return true;
6077 }
6078
6079 //------------------------------inline_counterMode_AESCrypt-----------------------
6080 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6081 assert(UseAES, "need AES instruction support");
6082 if (!UseAESCTRIntrinsics) return false;
6083
6084 address stubAddr = NULL;
6085 const char *stubName = NULL;
6086 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6087 stubAddr = StubRoutines::counterMode_AESCrypt();
6088 stubName = "counterMode_AESCrypt";
6089 }
6090 if (stubAddr == NULL) return false;
6091
6092 Node* counterMode_object = argument(0);
6093 Node* src = argument(1);
6094 Node* src_offset = argument(2);
6095 Node* len = argument(3);
6096 Node* dest = argument(4);
6097 Node* dest_offset = argument(5);
6098
6099 // (1) src and dest are arrays.
6100 const Type* src_type = src->Value(&_gvn);
6101 const Type* dest_type = dest->Value(&_gvn);
6102 const TypeAryPtr* top_src = src_type->isa_aryptr();
6103 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6104 assert(top_src != NULL && top_src->klass() != NULL &&
6105 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6106
6107 // checks are the responsibility of the caller
6108 Node* src_start = src;
6109 Node* dest_start = dest;
6110 if (src_offset != NULL || dest_offset != NULL) {
6111 assert(src_offset != NULL && dest_offset != NULL, "");
6112 src_start = array_element_address(src, src_offset, T_BYTE);
6113 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6114 }
6115
6116 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6117 // (because of the predicated logic executed earlier).
6118 // so we cast it here safely.
6119 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6120 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6121 if (embeddedCipherObj == NULL) return false;
6122 // cast it to what we know it will be at runtime
6123 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6124 assert(tinst != NULL, "CTR obj is null");
6125 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6126 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6127 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6128 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6129 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6130 const TypeOopPtr* xtype = aklass->as_instance_type();
6131 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6132 aescrypt_object = _gvn.transform(aescrypt_object);
6133 // we need to get the start of the aescrypt_object's expanded key array
6134 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6135 if (k_start == NULL) return false;
6136 // similarly, get the start address of the r vector
6137 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6138 if (obj_counter == NULL) return false;
6139 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6140
6141 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6142 if (saved_encCounter == NULL) return false;
6143 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6144 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6145
6146 Node* ctrCrypt;
6147 if (Matcher::pass_original_key_for_aes()) {
6148 // no SPARC version for AES/CTR intrinsics now.
6149 return false;
6150 }
6151 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6152 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6153 OptoRuntime::counterMode_aescrypt_Type(),
6154 stubAddr, stubName, TypePtr::BOTTOM,
6155 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6156
6157 // return cipher length (int)
6158 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6159 set_result(retvalue);
6160 return true;
6161 }
6162
6163 //------------------------------get_key_start_from_aescrypt_object-----------------------
6164 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6165 #if defined(PPC64) || defined(S390)
6166 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6167 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6168 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6169 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6170 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6171 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6172 if (objSessionK == NULL) {
6173 return (Node *) NULL;
6174 }
6175 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6176 #else
6177 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6178 #endif // PPC64
6179 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6180 if (objAESCryptKey == NULL) return (Node *) NULL;
6181
6182 // now have the array, need to get the start address of the K array
6183 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6184 return k_start;
6185 }
6186
6187 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6188 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6189 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6190 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6191 if (objAESCryptKey == NULL) return (Node *) NULL;
6192
6193 // now have the array, need to get the start address of the lastKey array
6194 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6195 return original_k_start;
6196 }
6197
6198 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6199 // Return node representing slow path of predicate check.
6200 // the pseudo code we want to emulate with this predicate is:
6201 // for encryption:
6202 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6203 // for decryption:
6204 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6205 // note cipher==plain is more conservative than the original java code but that's OK
6206 //
6207 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6208 // The receiver was checked for NULL already.
6209 Node* objCBC = argument(0);
6210
6211 Node* src = argument(1);
6212 Node* dest = argument(4);
6213
6214 // Load embeddedCipher field of CipherBlockChaining object.
6215 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6216
6217 // get AESCrypt klass for instanceOf check
6218 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6219 // will have same classloader as CipherBlockChaining object
6220 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6221 assert(tinst != NULL, "CBCobj is null");
6222 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6223
6224 // we want to do an instanceof comparison against the AESCrypt class
6225 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6226 if (!klass_AESCrypt->is_loaded()) {
6227 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6228 Node* ctrl = control();
6229 set_control(top()); // no regular fast path
6230 return ctrl;
6231 }
6232
6233 src = must_be_not_null(src, true);
6234 dest = must_be_not_null(dest, true);
6235
6236 // Resolve oops to stable for CmpP below.
6237 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6238
6239 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6240 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6241 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6242
6243 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6244
6245 // for encryption, we are done
6246 if (!decrypting)
6247 return instof_false; // even if it is NULL
6248
6249 // for decryption, we need to add a further check to avoid
6250 // taking the intrinsic path when cipher and plain are the same
6251 // see the original java code for why.
6252 RegionNode* region = new RegionNode(3);
6253 region->init_req(1, instof_false);
6254
6255 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6256 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6257 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6258 region->init_req(2, src_dest_conjoint);
6259
6260 record_for_igvn(region);
6261 return _gvn.transform(region);
6262 }
6263
6264 //----------------------------inline_electronicCodeBook_AESCrypt_predicate----------------------------
6265 // Return node representing slow path of predicate check.
6266 // the pseudo code we want to emulate with this predicate is:
6267 // for encryption:
6268 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6269 // for decryption:
6270 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6271 // note cipher==plain is more conservative than the original java code but that's OK
6272 //
6273 Node* LibraryCallKit::inline_electronicCodeBook_AESCrypt_predicate(bool decrypting) {
6274 // The receiver was checked for NULL already.
6275 Node* objECB = argument(0);
6276
6277 // Load embeddedCipher field of ElectronicCodeBook object.
6278 Node* embeddedCipherObj = load_field_from_object(objECB, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6279
6280 // get AESCrypt klass for instanceOf check
6281 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6282 // will have same classloader as ElectronicCodeBook object
6283 const TypeInstPtr* tinst = _gvn.type(objECB)->isa_instptr();
6284 assert(tinst != NULL, "ECBobj is null");
6285 assert(tinst->klass()->is_loaded(), "ECBobj is not loaded");
6286
6287 // we want to do an instanceof comparison against the AESCrypt class
6288 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6289 if (!klass_AESCrypt->is_loaded()) {
6290 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6291 Node* ctrl = control();
6292 set_control(top()); // no regular fast path
6293 return ctrl;
6294 }
6295 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6296
6297 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6298 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6299 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6300
6301 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6302
6303 // for encryption, we are done
6304 if (!decrypting)
6305 return instof_false; // even if it is NULL
6306
6307 // for decryption, we need to add a further check to avoid
6308 // taking the intrinsic path when cipher and plain are the same
6309 // see the original java code for why.
6310 RegionNode* region = new RegionNode(3);
6311 region->init_req(1, instof_false);
6312 Node* src = argument(1);
6313 Node* dest = argument(4);
6314 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6315 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6316 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6317 region->init_req(2, src_dest_conjoint);
6318
6319 record_for_igvn(region);
6320 return _gvn.transform(region);
6321 }
6322
6323 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6324 // Return node representing slow path of predicate check.
6325 // the pseudo code we want to emulate with this predicate is:
6326 // for encryption:
6327 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6328 // for decryption:
6329 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6330 // note cipher==plain is more conservative than the original java code but that's OK
6331 //
6332
6333 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6334 // The receiver was checked for NULL already.
6335 Node* objCTR = argument(0);
6336
6337 // Load embeddedCipher field of CipherBlockChaining object.
6338 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6339
6340 // get AESCrypt klass for instanceOf check
6341 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6342 // will have same classloader as CipherBlockChaining object
6343 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6344 assert(tinst != NULL, "CTRobj is null");
6345 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6346
6347 // we want to do an instanceof comparison against the AESCrypt class
6348 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6349 if (!klass_AESCrypt->is_loaded()) {
6350 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6351 Node* ctrl = control();
6352 set_control(top()); // no regular fast path
6353 return ctrl;
6354 }
6355
6356 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6357 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6358 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6359 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6360 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6361
6362 return instof_false; // even if it is NULL
6363 }
6364
6365 //------------------------------inline_ghash_processBlocks
6366 bool LibraryCallKit::inline_ghash_processBlocks() {
6367 address stubAddr;
6368 const char *stubName;
6369 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6370
6371 stubAddr = StubRoutines::ghash_processBlocks();
6372 stubName = "ghash_processBlocks";
6373
6374 Node* data = argument(0);
6375 Node* offset = argument(1);
6376 Node* len = argument(2);
6377 Node* state = argument(3);
6378 Node* subkeyH = argument(4);
6379
6380 state = must_be_not_null(state, true);
6381 subkeyH = must_be_not_null(subkeyH, true);
6382 data = must_be_not_null(data, true);
6383
6384 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6385 assert(state_start, "state is NULL");
6386 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6387 assert(subkeyH_start, "subkeyH is NULL");
6388 Node* data_start = array_element_address(data, offset, T_BYTE);
6389 assert(data_start, "data is NULL");
6390
6391 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6392 OptoRuntime::ghash_processBlocks_Type(),
6393 stubAddr, stubName, TypePtr::BOTTOM,
6394 state_start, subkeyH_start, data_start, len);
6395 return true;
6396 }
6397
6398 bool LibraryCallKit::inline_base64_encodeBlock() {
6399 address stubAddr;
6400 const char *stubName;
6401 assert(UseBASE64Intrinsics, "need Base64 intrinsics support");
6402 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters");
6403 stubAddr = StubRoutines::base64_encodeBlock();
6404 stubName = "encodeBlock";
6405
6406 if (!stubAddr) return false;
6407 Node* base64obj = argument(0);
6408 Node* src = argument(1);
6409 Node* offset = argument(2);
6410 Node* len = argument(3);
6411 Node* dest = argument(4);
6412 Node* dp = argument(5);
6413 Node* isURL = argument(6);
6414
6415 src = must_be_not_null(src, true);
6416 dest = must_be_not_null(dest, true);
6417
6418 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
6419 assert(src_start, "source array is NULL");
6420 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE);
6421 assert(dest_start, "destination array is NULL");
6422
6423 Node* base64 = make_runtime_call(RC_LEAF,
6424 OptoRuntime::base64_encodeBlock_Type(),
6425 stubAddr, stubName, TypePtr::BOTTOM,
6426 src_start, offset, len, dest_start, dp, isURL);
6427 return true;
6428 }
6429
6430 //------------------------------inline_digestBase_implCompress-----------------------
6431 //
6432 // Calculate MD5 for single-block byte[] array.
6433 // void com.sun.security.provider.MD5.implCompress(byte[] buf, int ofs)
6434 //
6435 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6436 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6437 //
6438 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6439 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6440 //
6441 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6442 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6443 //
6444 bool LibraryCallKit::inline_digestBase_implCompress(vmIntrinsics::ID id) {
6445 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6446
6447 Node* digestBase_obj = argument(0);
6448 Node* src = argument(1); // type oop
6449 Node* ofs = argument(2); // type int
6450
6451 const Type* src_type = src->Value(&_gvn);
6452 const TypeAryPtr* top_src = src_type->isa_aryptr();
6453 if (top_src == NULL || top_src->klass() == NULL) {
6454 // failed array check
6455 return false;
6456 }
6457 // Figure out the size and type of the elements we will be copying.
6458 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6459 if (src_elem != T_BYTE) {
6460 return false;
6461 }
6462 // 'src_start' points to src array + offset
6463 src = must_be_not_null(src, true);
6464 Node* src_start = array_element_address(src, ofs, src_elem);
6465 Node* state = NULL;
6466 address stubAddr;
6467 const char *stubName;
6468
6469 switch(id) {
6470 case vmIntrinsics::_md5_implCompress:
6471 assert(UseMD5Intrinsics, "need MD5 instruction support");
6472 state = get_state_from_digest_object(digestBase_obj);
6473 stubAddr = StubRoutines::md5_implCompress();
6474 stubName = "md5_implCompress";
6475 break;
6476 case vmIntrinsics::_sha_implCompress:
6477 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6478 state = get_state_from_digest_object(digestBase_obj);
6479 stubAddr = StubRoutines::sha1_implCompress();
6480 stubName = "sha1_implCompress";
6481 break;
6482 case vmIntrinsics::_sha2_implCompress:
6483 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6484 state = get_state_from_digest_object(digestBase_obj);
6485 stubAddr = StubRoutines::sha256_implCompress();
6486 stubName = "sha256_implCompress";
6487 break;
6488 case vmIntrinsics::_sha5_implCompress:
6489 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6490 state = get_long_state_from_digest_object(digestBase_obj);
6491 stubAddr = StubRoutines::sha512_implCompress();
6492 stubName = "sha512_implCompress";
6493 break;
6494 default:
6495 fatal_unexpected_iid(id);
6496 return false;
6497 }
6498 if (state == NULL) return false;
6499
6500 assert(stubAddr != NULL, "Stub is generated");
6501 if (stubAddr == NULL) return false;
6502
6503 // Call the stub.
6504 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::digestBase_implCompress_Type(),
6505 stubAddr, stubName, TypePtr::BOTTOM,
6506 src_start, state);
6507
6508 return true;
6509 }
6510
6511 //------------------------------inline_digestBase_implCompressMB-----------------------
6512 //
6513 // Calculate MD5/SHA/SHA2/SHA5 for multi-block byte[] array.
6514 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6515 //
6516 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6517 assert(UseMD5Intrinsics || UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6518 "need MD5/SHA1/SHA256/SHA512 instruction support");
6519 assert((uint)predicate < 4, "sanity");
6520 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6521
6522 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6523 Node* src = argument(1); // byte[] array
6524 Node* ofs = argument(2); // type int
6525 Node* limit = argument(3); // type int
6526
6527 const Type* src_type = src->Value(&_gvn);
6528 const TypeAryPtr* top_src = src_type->isa_aryptr();
6529 if (top_src == NULL || top_src->klass() == NULL) {
6530 // failed array check
6531 return false;
6532 }
6533 // Figure out the size and type of the elements we will be copying.
6534 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6535 if (src_elem != T_BYTE) {
6536 return false;
6537 }
6538 // 'src_start' points to src array + offset
6539 src = must_be_not_null(src, false);
6540 Node* src_start = array_element_address(src, ofs, src_elem);
6541
6542 const char* klass_digestBase_name = NULL;
6543 const char* stub_name = NULL;
6544 address stub_addr = NULL;
6545 bool long_state = false;
6546
6547 switch (predicate) {
6548 case 0:
6549 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_md5_implCompress)) {
6550 klass_digestBase_name = "sun/security/provider/MD5";
6551 stub_name = "md5_implCompressMB";
6552 stub_addr = StubRoutines::md5_implCompressMB();
6553 }
6554 break;
6555 case 1:
6556 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha_implCompress)) {
6557 klass_digestBase_name = "sun/security/provider/SHA";
6558 stub_name = "sha1_implCompressMB";
6559 stub_addr = StubRoutines::sha1_implCompressMB();
6560 }
6561 break;
6562 case 2:
6563 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha2_implCompress)) {
6564 klass_digestBase_name = "sun/security/provider/SHA2";
6565 stub_name = "sha256_implCompressMB";
6566 stub_addr = StubRoutines::sha256_implCompressMB();
6567 }
6568 break;
6569 case 3:
6570 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha5_implCompress)) {
6571 klass_digestBase_name = "sun/security/provider/SHA5";
6572 stub_name = "sha512_implCompressMB";
6573 stub_addr = StubRoutines::sha512_implCompressMB();
6574 long_state = true;
6575 }
6576 break;
6577 default:
6578 fatal("unknown DigestBase intrinsic predicate: %d", predicate);
6579 }
6580 if (klass_digestBase_name != NULL) {
6581 assert(stub_addr != NULL, "Stub is generated");
6582 if (stub_addr == NULL) return false;
6583
6584 // get DigestBase klass to lookup for SHA klass
6585 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6586 assert(tinst != NULL, "digestBase_obj is not instance???");
6587 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6588
6589 ciKlass* klass_digestBase = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_digestBase_name));
6590 assert(klass_digestBase->is_loaded(), "predicate checks that this class is loaded");
6591 ciInstanceKlass* instklass_digestBase = klass_digestBase->as_instance_klass();
6592 return inline_digestBase_implCompressMB(digestBase_obj, instklass_digestBase, long_state, stub_addr, stub_name, src_start, ofs, limit);
6593 }
6594 return false;
6595 }
6596
6597 //------------------------------inline_digestBase_implCompressMB-----------------------
6598 bool LibraryCallKit::inline_digestBase_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_digestBase,
6599 bool long_state, address stubAddr, const char *stubName,
6600 Node* src_start, Node* ofs, Node* limit) {
6601 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_digestBase);
6602 const TypeOopPtr* xtype = aklass->as_instance_type();
6603 Node* digest_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6604 digest_obj = _gvn.transform(digest_obj);
6605
6606 Node* state;
6607 if (long_state) {
6608 state = get_long_state_from_digest_object(digest_obj);
6609 } else {
6610 state = get_state_from_digest_object(digest_obj);
6611 }
6612 if (state == NULL) return false;
6613
6614 // Call the stub.
6615 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6616 OptoRuntime::digestBase_implCompressMB_Type(),
6617 stubAddr, stubName, TypePtr::BOTTOM,
6618 src_start, state, ofs, limit);
6619 // return ofs (int)
6620 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6621 set_result(result);
6622
6623 return true;
6624 }
6625
6626 //------------------------------get_state_from_digest_object-----------------------
6627 Node * LibraryCallKit::get_state_from_digest_object(Node *digest_object) {
6628 Node* digest_state = load_field_from_object(digest_object, "state", "[I", /*is_exact*/ false);
6629 assert (digest_state != NULL, "wrong version of sun.security.provider.MD5/SHA/SHA2");
6630 if (digest_state == NULL) return (Node *) NULL;
6631
6632 // now have the array, need to get the start address of the state array
6633 Node* state = array_element_address(digest_state, intcon(0), T_INT);
6634 return state;
6635 }
6636
6637 //------------------------------get_long_state_from_digest_object-----------------------
6638 Node * LibraryCallKit::get_long_state_from_digest_object(Node *digest_object) {
6639 Node* digest_state = load_field_from_object(digest_object, "state", "[J", /*is_exact*/ false);
6640 assert (digest_state != NULL, "wrong version of sun.security.provider.SHA5");
6641 if (digest_state == NULL) return (Node *) NULL;
6642
6643 // now have the array, need to get the start address of the state array
6644 Node* state = array_element_address(digest_state, intcon(0), T_LONG);
6645 return state;
6646 }
6647
6648 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6649 // Return node representing slow path of predicate check.
6650 // the pseudo code we want to emulate with this predicate is:
6651 // if (digestBaseObj instanceof MD5/SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6652 //
6653 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6654 assert(UseMD5Intrinsics || UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6655 "need MD5/SHA1/SHA256/SHA512 instruction support");
6656 assert((uint)predicate < 4, "sanity");
6657
6658 // The receiver was checked for NULL already.
6659 Node* digestBaseObj = argument(0);
6660
6661 // get DigestBase klass for instanceOf check
6662 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6663 assert(tinst != NULL, "digestBaseObj is null");
6664 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6665
6666 const char* klass_name = NULL;
6667 switch (predicate) {
6668 case 0:
6669 if (UseMD5Intrinsics) {
6670 // we want to do an instanceof comparison against the MD5 class
6671 klass_name = "sun/security/provider/MD5";
6672 }
6673 break;
6674 case 1:
6675 if (UseSHA1Intrinsics) {
6676 // we want to do an instanceof comparison against the SHA class
6677 klass_name = "sun/security/provider/SHA";
6678 }
6679 break;
6680 case 2:
6681 if (UseSHA256Intrinsics) {
6682 // we want to do an instanceof comparison against the SHA2 class
6683 klass_name = "sun/security/provider/SHA2";
6684 }
6685 break;
6686 case 3:
6687 if (UseSHA512Intrinsics) {
6688 // we want to do an instanceof comparison against the SHA5 class
6689 klass_name = "sun/security/provider/SHA5";
6690 }
6691 break;
6692 default:
6693 fatal("unknown SHA intrinsic predicate: %d", predicate);
6694 }
6695
6696 ciKlass* klass = NULL;
6697 if (klass_name != NULL) {
6698 klass = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_name));
6699 }
6700 if ((klass == NULL) || !klass->is_loaded()) {
6701 // if none of MD5/SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6702 Node* ctrl = control();
6703 set_control(top()); // no intrinsic path
6704 return ctrl;
6705 }
6706 ciInstanceKlass* instklass = klass->as_instance_klass();
6707
6708 Node* instof = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass)));
6709 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6710 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6711 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6712
6713 return instof_false; // even if it is NULL
6714 }
6715
6716 //-------------inline_fma-----------------------------------
6717 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6718 Node *a = NULL;
6719 Node *b = NULL;
6720 Node *c = NULL;
6721 Node* result = NULL;
6722 switch (id) {
6723 case vmIntrinsics::_fmaD:
6724 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6725 // no receiver since it is static method
6726 a = round_double_node(argument(0));
6727 b = round_double_node(argument(2));
6728 c = round_double_node(argument(4));
6729 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6730 break;
6731 case vmIntrinsics::_fmaF:
6732 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6733 a = argument(0);
6734 b = argument(1);
6735 c = argument(2);
6736 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6737 break;
6738 default:
6739 fatal_unexpected_iid(id); break;
6740 }
6741 set_result(result);
6742 return true;
6743 }
6744
6745 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) {
6746 // argument(0) is receiver
6747 Node* codePoint = argument(1);
6748 Node* n = NULL;
6749
6750 switch (id) {
6751 case vmIntrinsics::_isDigit :
6752 n = new DigitNode(control(), codePoint);
6753 break;
6754 case vmIntrinsics::_isLowerCase :
6755 n = new LowerCaseNode(control(), codePoint);
6756 break;
6757 case vmIntrinsics::_isUpperCase :
6758 n = new UpperCaseNode(control(), codePoint);
6759 break;
6760 case vmIntrinsics::_isWhitespace :
6761 n = new WhitespaceNode(control(), codePoint);
6762 break;
6763 default:
6764 fatal_unexpected_iid(id);
6765 }
6766
6767 set_result(_gvn.transform(n));
6768 return true;
6769 }
6770
6771 //------------------------------inline_fp_min_max------------------------------
6772 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) {
6773 /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416.
6774
6775 // The intrinsic should be used only when the API branches aren't predictable,
6776 // the last one performing the most important comparison. The following heuristic
6777 // uses the branch statistics to eventually bail out if necessary.
6778
6779 ciMethodData *md = callee()->method_data();
6780
6781 if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) {
6782 ciCallProfile cp = caller()->call_profile_at_bci(bci());
6783
6784 if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) {
6785 // Bail out if the call-site didn't contribute enough to the statistics.
6786 return false;
6787 }
6788
6789 uint taken = 0, not_taken = 0;
6790
6791 for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) {
6792 if (p->is_BranchData()) {
6793 taken = ((ciBranchData*)p)->taken();
6794 not_taken = ((ciBranchData*)p)->not_taken();
6795 }
6796 }
6797
6798 double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count());
6799 balance = balance < 0 ? -balance : balance;
6800 if ( balance > 0.2 ) {
6801 // Bail out if the most important branch is predictable enough.
6802 return false;
6803 }
6804 }
6805 */
6806
6807 Node *a = NULL;
6808 Node *b = NULL;
6809 Node *n = NULL;
6810 switch (id) {
6811 case vmIntrinsics::_maxF:
6812 case vmIntrinsics::_minF:
6813 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each.");
6814 a = argument(0);
6815 b = argument(1);
6816 break;
6817 case vmIntrinsics::_maxD:
6818 case vmIntrinsics::_minD:
6819 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each.");
6820 a = round_double_node(argument(0));
6821 b = round_double_node(argument(2));
6822 break;
6823 default:
6824 fatal_unexpected_iid(id);
6825 break;
6826 }
6827 switch (id) {
6828 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break;
6829 case vmIntrinsics::_minF: n = new MinFNode(a, b); break;
6830 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break;
6831 case vmIntrinsics::_minD: n = new MinDNode(a, b); break;
6832 default: fatal_unexpected_iid(id); break;
6833 }
6834 set_result(_gvn.transform(n));
6835 return true;
6836 }
6837
6838 bool LibraryCallKit::inline_profileBoolean() {
6839 Node* counts = argument(1);
6840 const TypeAryPtr* ary = NULL;
6841 ciArray* aobj = NULL;
6842 if (counts->is_Con()
6843 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6844 && (aobj = ary->const_oop()->as_array()) != NULL
6845 && (aobj->length() == 2)) {
6846 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6847 jint false_cnt = aobj->element_value(0).as_int();
6848 jint true_cnt = aobj->element_value(1).as_int();
6849
6850 if (C->log() != NULL) {
6851 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6852 false_cnt, true_cnt);
6853 }
6854
6855 if (false_cnt + true_cnt == 0) {
6856 // According to profile, never executed.
6857 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6858 Deoptimization::Action_reinterpret);
6859 return true;
6860 }
6861
6862 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6863 // is a number of each value occurrences.
6864 Node* result = argument(0);
6865 if (false_cnt == 0 || true_cnt == 0) {
6866 // According to profile, one value has been never seen.
6867 int expected_val = (false_cnt == 0) ? 1 : 0;
6868
6869 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6870 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6871
6872 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6873 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6874 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6875
6876 { // Slow path: uncommon trap for never seen value and then reexecute
6877 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6878 // the value has been seen at least once.
6879 PreserveJVMState pjvms(this);
6880 PreserveReexecuteState preexecs(this);
6881 jvms()->set_should_reexecute(true);
6882
6883 set_control(slow_path);
6884 set_i_o(i_o());
6885
6886 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6887 Deoptimization::Action_reinterpret);
6888 }
6889 // The guard for never seen value enables sharpening of the result and
6890 // returning a constant. It allows to eliminate branches on the same value
6891 // later on.
6892 set_control(fast_path);
6893 result = intcon(expected_val);
6894 }
6895 // Stop profiling.
6896 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6897 // By replacing method body with profile data (represented as ProfileBooleanNode
6898 // on IR level) we effectively disable profiling.
6899 // It enables full speed execution once optimized code is generated.
6900 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6901 C->record_for_igvn(profile);
6902 set_result(profile);
6903 return true;
6904 } else {
6905 // Continue profiling.
6906 // Profile data isn't available at the moment. So, execute method's bytecode version.
6907 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6908 // is compiled and counters aren't available since corresponding MethodHandle
6909 // isn't a compile-time constant.
6910 return false;
6911 }
6912 }
6913
6914 bool LibraryCallKit::inline_isCompileConstant() {
6915 Node* n = argument(0);
6916 set_result(n->is_Con() ? intcon(1) : intcon(0));
6917 return true;
6918 }
6919
6920 bool LibraryCallKit::inline_sizeOf() {
6921 if (!RuntimeSizeOf) {
6922 set_result(longcon(-1));
6923 return true;
6924 }
6925
6926 Node* obj = argument(0);
6927 Node* klass_node = load_object_klass(obj);
6928
6929 jint layout_con = Klass::_lh_neutral_value;
6930 Node* layout_val = get_layout_helper(klass_node, layout_con);
6931 int layout_is_con = (layout_val == NULL);
6932
6933 if (layout_is_con) {
6934 // Layout helper is constant, can figure out things at compile time.
6935 assert(false, "Since sizeOf is @DontInline, this path should be unvisited");
6936
6937 if (Klass::layout_helper_is_instance(layout_con)) {
6938 // Instance case: layout_con contains the size itself.
6939 Node *size = longcon(Klass::layout_helper_size_in_bytes(layout_con));
6940 set_result(size);
6941 } else {
6942 // Array case: size is round(header + element_size*arraylength).
6943 // Since arraylength is different for every array instance, we have to
6944 // compute the whole thing at runtime.
6945
6946 Node* arr_length = load_array_length(obj);
6947
6948 int round_mask = MinObjAlignmentInBytes - 1;
6949 int hsize = Klass::layout_helper_header_size(layout_con);
6950 int eshift = Klass::layout_helper_log2_element_size(layout_con);
6951
6952 if ((round_mask & ~right_n_bits(eshift)) == 0) {
6953 round_mask = 0; // strength-reduce it if it goes away completely
6954 }
6955 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
6956 Node* header_size = intcon(hsize + round_mask);
6957
6958 Node* lengthx = ConvI2X(arr_length);
6959 Node* headerx = ConvI2X(header_size);
6960
6961 Node* abody = lengthx;
6962 if (eshift != 0) {
6963 abody = _gvn.transform(new LShiftXNode(lengthx, intcon(eshift)));
6964 }
6965 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
6966 if (round_mask != 0) {
6967 size = _gvn.transform( new AndXNode(size, MakeConX(~round_mask)) );
6968 }
6969 size = ConvX2L(size);
6970 set_result(size);
6971 }
6972 } else {
6973 // Layout helper is not constant, need to test for array-ness at runtime.
6974
6975 enum { _instance_path = 1, _array_path, PATH_LIMIT };
6976 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
6977 PhiNode* result_val = new PhiNode(result_reg, TypeLong::LONG);
6978 record_for_igvn(result_reg);
6979
6980 Node* array_ctl = generate_array_guard(klass_node, NULL);
6981 if (array_ctl != NULL) {
6982 // Array case: size is round(header + element_size*arraylength).
6983 // Since arraylength is different for every array instance, we have to
6984 // compute the whole thing at runtime.
6985
6986 PreserveJVMState pjvms(this);
6987 set_control(array_ctl);
6988 Node* arr_length = load_array_length(obj);
6989
6990 int round_mask = MinObjAlignmentInBytes - 1;
6991 Node* mask = intcon(round_mask);
6992
6993 Node* hss = intcon(Klass::_lh_header_size_shift);
6994 Node* hsm = intcon(Klass::_lh_header_size_mask);
6995 Node* header_size = _gvn.transform(new URShiftINode(layout_val, hss));
6996 header_size = _gvn.transform(new AndINode(header_size, hsm));
6997 header_size = _gvn.transform(new AddINode(header_size, mask));
6998
6999 // There is no need to mask or shift this value.
7000 // The semantics of LShiftINode include an implicit mask to 0x1F.
7001 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
7002 Node* elem_shift = layout_val;
7003
7004 Node* lengthx = ConvI2X(arr_length);
7005 Node* headerx = ConvI2X(header_size);
7006
7007 Node* abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
7008 Node* size = _gvn.transform(new AddXNode(headerx, abody));
7009 if (round_mask != 0) {
7010 size = _gvn.transform(new AndXNode(size, MakeConX(~round_mask)));
7011 }
7012 size = ConvX2L(size);
7013
7014 result_reg->init_req(_array_path, control());
7015 result_val->init_req(_array_path, size);
7016 }
7017
7018 if (!stopped()) {
7019 // Instance case: the layout helper gives us instance size almost directly,
7020 // but we need to mask out the _lh_instance_slow_path_bit.
7021 Node* size = ConvI2X(layout_val);
7022 assert((int) Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
7023 Node* mask = MakeConX(~(intptr_t) right_n_bits(LogBytesPerLong));
7024 size = _gvn.transform(new AndXNode(size, mask));
7025 size = ConvX2L(size);
7026
7027 result_reg->init_req(_instance_path, control());
7028 result_val->init_req(_instance_path, size);
7029 }
7030
7031 set_result(result_reg, result_val);
7032 }
7033
7034 return true;
7035 }
7036
7037 bool LibraryCallKit::inline_addressOf() {
7038 if (!RuntimeAddressOf) {
7039 set_result(longcon(-1));
7040 return true;
7041 }
7042
7043 Node* obj = argument(0);
7044 Node* raw_val = _gvn.transform(new CastP2XNode(NULL, obj));
7045 Node* off_val = _gvn.transform(new AddXNode(raw_val, MakeConX(Universe::non_heap_offset())));
7046 Node* long_val = ConvX2L(off_val);
7047
7048 set_result(long_val);
7049 return true;
7050 }
7051
7052 bool LibraryCallKit::inline_getReferencedObjects() {
7053 Node* a1 = argument(0);
7054 Node* a2 = argument(1);
7055
7056 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
7057 OptoRuntime::get_referenced_objects_Type(),
7058 CAST_FROM_FN_PTR(address, SharedRuntime::get_referenced_objects),
7059 "get_referenced_objects",
7060 TypePtr::BOTTOM,
7061 a1, a2);
7062
7063 Node* value = _gvn.transform(new ProjNode(call, TypeFunc::Parms+0));
7064 set_result(value);
7065 return true;
7066 }