1 /*
2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/subnode.hpp"
52 #include "opto/vectornode.hpp"
53 #include "prims/nativeLookup.hpp"
54 #include "prims/unsafe.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #ifdef TRACE_HAVE_INTRINSICS
57 #include "trace/traceMacros.hpp"
58 #endif
59
60 class LibraryIntrinsic : public InlineCallGenerator {
61 // Extend the set of intrinsics known to the runtime:
62 public:
63 private:
64 bool _is_virtual;
65 bool _does_virtual_dispatch;
66 int8_t _predicates_count; // Intrinsic is predicated by several conditions
67 int8_t _last_predicate; // Last generated predicate
68 vmIntrinsics::ID _intrinsic_id;
69
70 public:
71 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
72 : InlineCallGenerator(m),
73 _is_virtual(is_virtual),
74 _does_virtual_dispatch(does_virtual_dispatch),
75 _predicates_count((int8_t)predicates_count),
76 _last_predicate((int8_t)-1),
77 _intrinsic_id(id)
78 {
79 }
80 virtual bool is_intrinsic() const { return true; }
81 virtual bool is_virtual() const { return _is_virtual; }
82 virtual bool is_predicated() const { return _predicates_count > 0; }
83 virtual int predicates_count() const { return _predicates_count; }
84 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
85 virtual JVMState* generate(JVMState* jvms);
86 virtual Node* generate_predicate(JVMState* jvms, int predicate);
87 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
88 };
89
90
91 // Local helper class for LibraryIntrinsic:
92 class LibraryCallKit : public GraphKit {
93 private:
94 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
95 Node* _result; // the result node, if any
96 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
97
98 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
99
100 public:
101 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
102 : GraphKit(jvms),
103 _intrinsic(intrinsic),
104 _result(NULL)
105 {
106 // Check if this is a root compile. In that case we don't have a caller.
107 if (!jvms->has_method()) {
108 _reexecute_sp = sp();
109 } else {
110 // Find out how many arguments the interpreter needs when deoptimizing
111 // and save the stack pointer value so it can used by uncommon_trap.
112 // We find the argument count by looking at the declared signature.
113 bool ignored_will_link;
114 ciSignature* declared_signature = NULL;
115 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
116 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
117 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
118 }
119 }
120
121 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
122
123 ciMethod* caller() const { return jvms()->method(); }
124 int bci() const { return jvms()->bci(); }
125 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
126 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
127 ciMethod* callee() const { return _intrinsic->method(); }
128
129 bool try_to_inline(int predicate);
130 Node* try_to_predicate(int predicate);
131
132 void push_result() {
133 // Push the result onto the stack.
134 if (!stopped() && result() != NULL) {
135 BasicType bt = result()->bottom_type()->basic_type();
136 push_node(bt, result());
137 }
138 }
139
140 private:
141 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
142 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
143 }
144
145 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
146 void set_result(RegionNode* region, PhiNode* value);
147 Node* result() { return _result; }
148
149 virtual int reexecute_sp() { return _reexecute_sp; }
150
151 // Helper functions to inline natives
152 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
153 Node* generate_slow_guard(Node* test, RegionNode* region);
154 Node* generate_fair_guard(Node* test, RegionNode* region);
155 Node* generate_negative_guard(Node* index, RegionNode* region,
156 // resulting CastII of index:
157 Node* *pos_index = NULL);
158 Node* generate_limit_guard(Node* offset, Node* subseq_length,
159 Node* array_length,
160 RegionNode* region);
161 void generate_string_range_check(Node* array, Node* offset,
162 Node* length, bool char_count);
163 Node* generate_current_thread(Node* &tls_output);
164 Node* load_mirror_from_klass(Node* klass);
165 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
166 RegionNode* region, int null_path,
167 int offset);
168 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
169 RegionNode* region, int null_path) {
170 int offset = java_lang_Class::klass_offset_in_bytes();
171 return load_klass_from_mirror_common(mirror, never_see_null,
172 region, null_path,
173 offset);
174 }
175 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
176 RegionNode* region, int null_path) {
177 int offset = java_lang_Class::array_klass_offset_in_bytes();
178 return load_klass_from_mirror_common(mirror, never_see_null,
179 region, null_path,
180 offset);
181 }
182 Node* generate_access_flags_guard(Node* kls,
183 int modifier_mask, int modifier_bits,
184 RegionNode* region);
185 Node* generate_interface_guard(Node* kls, RegionNode* region);
186 Node* generate_array_guard(Node* kls, RegionNode* region) {
187 return generate_array_guard_common(kls, region, false, false);
188 }
189 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
190 return generate_array_guard_common(kls, region, false, true);
191 }
192 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
193 return generate_array_guard_common(kls, region, true, false);
194 }
195 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
196 return generate_array_guard_common(kls, region, true, true);
197 }
198 Node* generate_array_guard_common(Node* kls, RegionNode* region,
199 bool obj_array, bool not_array);
200 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
201 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
202 bool is_virtual = false, bool is_static = false);
203 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
204 return generate_method_call(method_id, false, true);
205 }
206 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
207 return generate_method_call(method_id, true, false);
208 }
209 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
210 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
211
212 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
215 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
216 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
217 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_indexOfChar();
219 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
220 bool inline_string_toBytesU();
221 bool inline_string_getCharsU();
222 bool inline_string_copy(bool compress);
223 bool inline_string_char_access(bool is_store);
224 Node* round_double_node(Node* n);
225 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
226 bool inline_math_native(vmIntrinsics::ID id);
227 bool inline_math(vmIntrinsics::ID id);
228 template <typename OverflowOp>
229 bool inline_math_overflow(Node* arg1, Node* arg2);
230 void inline_math_mathExact(Node* math, Node* test);
231 bool inline_math_addExactI(bool is_increment);
232 bool inline_math_addExactL(bool is_increment);
233 bool inline_math_multiplyExactI();
234 bool inline_math_multiplyExactL();
235 bool inline_math_multiplyHigh();
236 bool inline_math_negateExactI();
237 bool inline_math_negateExactL();
238 bool inline_math_subtractExactI(bool is_decrement);
239 bool inline_math_subtractExactL(bool is_decrement);
240 bool inline_min_max(vmIntrinsics::ID id);
241 bool inline_notify(vmIntrinsics::ID id);
242 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
243 // This returns Type::AnyPtr, RawPtr, or OopPtr.
244 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
245 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
246 // Helper for inline_unsafe_access.
247 // Generates the guards that check whether the result of
248 // Unsafe.getObject should be recorded in an SATB log buffer.
249 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
250
251 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
252 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
253 static bool klass_needs_init_guard(Node* kls);
254 bool inline_unsafe_allocate();
255 bool inline_unsafe_newArray(bool uninitialized);
256 bool inline_unsafe_copyMemory();
257 bool inline_native_currentThread();
258
259 bool inline_native_time_funcs(address method, const char* funcName);
260 #ifdef TRACE_HAVE_INTRINSICS
261 bool inline_native_classID();
262 bool inline_native_getBufferWriter();
263 #endif
264 bool inline_native_isInterrupted();
265 bool inline_native_Class_query(vmIntrinsics::ID id);
266 bool inline_native_subtype_check();
267 bool inline_native_getLength();
268 bool inline_array_copyOf(bool is_copyOfRange);
269 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
270 bool inline_preconditions_checkIndex();
271 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
272 bool inline_native_clone(bool is_virtual);
273 bool inline_native_Reflection_getCallerClass();
274 // Helper function for inlining native object hash method
275 bool inline_native_hashcode(bool is_virtual, bool is_static);
276 bool inline_native_getClass();
277
278 // Helper functions for inlining arraycopy
279 bool inline_arraycopy();
280 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
281 RegionNode* slow_region);
282 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
283 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
284 uint new_idx);
285
286 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
287 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
288 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
289 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
290 bool inline_unsafe_fence(vmIntrinsics::ID id);
291 bool inline_onspinwait();
292 bool inline_fp_conversions(vmIntrinsics::ID id);
293 bool inline_number_methods(vmIntrinsics::ID id);
294 bool inline_reference_get();
295 bool inline_Class_cast();
296 bool inline_aescrypt_Block(vmIntrinsics::ID id);
297 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
298 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
299 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
300 Node* inline_counterMode_AESCrypt_predicate();
301 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
302 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
303 bool inline_ghash_processBlocks();
304 bool inline_sha_implCompress(vmIntrinsics::ID id);
305 bool inline_digestBase_implCompressMB(int predicate);
306 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
307 bool long_state, address stubAddr, const char *stubName,
308 Node* src_start, Node* ofs, Node* limit);
309 Node* get_state_from_sha_object(Node *sha_object);
310 Node* get_state_from_sha5_object(Node *sha_object);
311 Node* inline_digestBase_implCompressMB_predicate(int predicate);
312 bool inline_encodeISOArray();
313 bool inline_updateCRC32();
314 bool inline_updateBytesCRC32();
315 bool inline_updateByteBufferCRC32();
316 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
317 bool inline_updateBytesCRC32C();
318 bool inline_updateDirectByteBufferCRC32C();
319 bool inline_updateBytesAdler32();
320 bool inline_updateByteBufferAdler32();
321 bool inline_multiplyToLen();
322 bool inline_hasNegatives();
323 bool inline_squareToLen();
324 bool inline_mulAdd();
325 bool inline_montgomeryMultiply();
326 bool inline_montgomerySquare();
327 bool inline_vectorizedMismatch();
328 bool inline_fma(vmIntrinsics::ID id);
329
330 bool inline_profileBoolean();
331 bool inline_isCompileConstant();
332
333 // Vector API support
334 bool inline_vector_nary_operation(int n);
335 bool inline_vector_broadcast_coerced();
336 bool inline_vector_mem_operation(bool is_store);
337 bool inline_vector_reduction();
338 bool inline_vector_test();
339 bool inline_vector_blend();
340 bool inline_vector_compare();
341 bool inline_vector_broadcast_int();
342 bool inline_vector_rebracket();
343 Node* box_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
344 Node* unbox_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
345 Node* shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem);
346
347 void clear_upper_avx() {
348 #ifdef X86
349 if (UseAVX >= 2) {
350 C->set_clear_upper_avx(true);
351 }
352 #endif
353 }
354 };
355
356 //---------------------------make_vm_intrinsic----------------------------
357 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
358 vmIntrinsics::ID id = m->intrinsic_id();
359 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
360
361 if (!m->is_loaded()) {
362 // Do not attempt to inline unloaded methods.
363 return NULL;
364 }
365
366 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
367 bool is_available = false;
368
369 {
370 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
371 // the compiler must transition to '_thread_in_vm' state because both
372 // methods access VM-internal data.
373 VM_ENTRY_MARK;
374 methodHandle mh(THREAD, m->get_Method());
375 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
376 !C->directive()->is_intrinsic_disabled(mh) &&
377 !vmIntrinsics::is_disabled_by_flags(mh);
378
379 }
380
381 if (is_available) {
382 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
383 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
384 return new LibraryIntrinsic(m, is_virtual,
385 vmIntrinsics::predicates_needed(id),
386 vmIntrinsics::does_virtual_dispatch(id),
387 (vmIntrinsics::ID) id);
388 } else {
389 return NULL;
390 }
391 }
392
393 //----------------------register_library_intrinsics-----------------------
394 // Initialize this file's data structures, for each Compile instance.
395 void Compile::register_library_intrinsics() {
396 // Nothing to do here.
397 }
398
399 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
400 LibraryCallKit kit(jvms, this);
401 Compile* C = kit.C;
402 int nodes = C->unique();
403 #ifndef PRODUCT
404 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
405 char buf[1000];
406 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
407 tty->print_cr("Intrinsic %s", str);
408 }
409 #endif
410 ciMethod* callee = kit.callee();
411 const int bci = kit.bci();
412
413 // Try to inline the intrinsic.
414 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
415 kit.try_to_inline(_last_predicate)) {
416 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
417 : "(intrinsic)";
418 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
419 if (C->print_intrinsics() || C->print_inlining()) {
420 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
421 }
422 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
423 if (C->log()) {
424 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
425 vmIntrinsics::name_at(intrinsic_id()),
426 (is_virtual() ? " virtual='1'" : ""),
427 C->unique() - nodes);
428 }
429 // Push the result from the inlined method onto the stack.
430 kit.push_result();
431 C->print_inlining_update(this);
432 return kit.transfer_exceptions_into_jvms();
433 }
434
435 // The intrinsic bailed out
436 if (jvms->has_method()) {
437 // Not a root compile.
438 const char* msg;
439 if (callee->intrinsic_candidate()) {
440 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
441 } else {
442 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
443 : "failed to inline (intrinsic), method not annotated";
444 }
445 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
446 if (C->print_intrinsics() || C->print_inlining()) {
447 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
448 }
449 } else {
450 // Root compile
451 ResourceMark rm;
452 stringStream msg_stream;
453 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
454 vmIntrinsics::name_at(intrinsic_id()),
455 is_virtual() ? " (virtual)" : "", bci);
456 const char *msg = msg_stream.as_string();
457 log_debug(jit, inlining)("%s", msg);
458 if (C->print_intrinsics() || C->print_inlining()) {
459 tty->print("%s", msg);
460 }
461 }
462 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
463 C->print_inlining_update(this);
464 return NULL;
465 }
466
467 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
468 LibraryCallKit kit(jvms, this);
469 Compile* C = kit.C;
470 int nodes = C->unique();
471 _last_predicate = predicate;
472 #ifndef PRODUCT
473 assert(is_predicated() && predicate < predicates_count(), "sanity");
474 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
475 char buf[1000];
476 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
477 tty->print_cr("Predicate for intrinsic %s", str);
478 }
479 #endif
480 ciMethod* callee = kit.callee();
481 const int bci = kit.bci();
482
483 Node* slow_ctl = kit.try_to_predicate(predicate);
484 if (!kit.failing()) {
485 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
486 : "(intrinsic, predicate)";
487 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
488 if (C->print_intrinsics() || C->print_inlining()) {
489 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
490 }
491 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
492 if (C->log()) {
493 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
494 vmIntrinsics::name_at(intrinsic_id()),
495 (is_virtual() ? " virtual='1'" : ""),
496 C->unique() - nodes);
497 }
498 return slow_ctl; // Could be NULL if the check folds.
499 }
500
501 // The intrinsic bailed out
502 if (jvms->has_method()) {
503 // Not a root compile.
504 const char* msg = "failed to generate predicate for intrinsic";
505 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
506 if (C->print_intrinsics() || C->print_inlining()) {
507 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
508 }
509 } else {
510 // Root compile
511 ResourceMark rm;
512 stringStream msg_stream;
513 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
514 vmIntrinsics::name_at(intrinsic_id()),
515 is_virtual() ? " (virtual)" : "", bci);
516 const char *msg = msg_stream.as_string();
517 log_debug(jit, inlining)("%s", msg);
518 if (C->print_intrinsics() || C->print_inlining()) {
519 C->print_inlining_stream()->print("%s", msg);
520 }
521 }
522 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
523 return NULL;
524 }
525
526 bool LibraryCallKit::try_to_inline(int predicate) {
527 // Handle symbolic names for otherwise undistinguished boolean switches:
528 const bool is_store = true;
529 const bool is_compress = true;
530 const bool is_static = true;
531 const bool is_volatile = true;
532
533 if (!jvms()->has_method()) {
534 // Root JVMState has a null method.
535 assert(map()->memory()->Opcode() == Op_Parm, "");
536 // Insert the memory aliasing node
537 set_all_memory(reset_memory());
538 }
539 assert(merged_memory(), "");
540
541 switch (intrinsic_id()) {
542 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
543 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
544 case vmIntrinsics::_getClass: return inline_native_getClass();
545
546 case vmIntrinsics::_dsin:
547 case vmIntrinsics::_dcos:
548 case vmIntrinsics::_dtan:
549 case vmIntrinsics::_dabs:
550 case vmIntrinsics::_datan2:
551 case vmIntrinsics::_dsqrt:
552 case vmIntrinsics::_dexp:
553 case vmIntrinsics::_dlog:
554 case vmIntrinsics::_dlog10:
555 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
556
557 case vmIntrinsics::_min:
558 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
559
560 case vmIntrinsics::_notify:
561 case vmIntrinsics::_notifyAll:
562 if (InlineNotify) {
563 return inline_notify(intrinsic_id());
564 }
565 return false;
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::_getObject: 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::_putObject: 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::_getObjectVolatile: 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::_putObjectVolatile: 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::_getObjectAcquire: 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::_putObjectRelease: 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::_getObjectOpaque: 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::_putObjectOpaque: 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::_compareAndSetObject: 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::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
706 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
707 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
708 case vmIntrinsics::_weakCompareAndSetObject: 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::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
727 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
728 case vmIntrinsics::_compareAndExchangeObjectRelease: 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::_getAndSetObject: 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 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
761
762 #ifdef TRACE_HAVE_INTRINSICS
763 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
764 case vmIntrinsics::_getClassId: return inline_native_classID();
765 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
766 #endif
767 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
768 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
769 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
770 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
771 case vmIntrinsics::_getLength: return inline_native_getLength();
772 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
773 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
774 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
775 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
776 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
777 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
778
779 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
780 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
781
782 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
783
784 case vmIntrinsics::_isInstance:
785 case vmIntrinsics::_getModifiers:
786 case vmIntrinsics::_isInterface:
787 case vmIntrinsics::_isArray:
788 case vmIntrinsics::_isPrimitive:
789 case vmIntrinsics::_getSuperclass:
790 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
791
792 case vmIntrinsics::_floatToRawIntBits:
793 case vmIntrinsics::_floatToIntBits:
794 case vmIntrinsics::_intBitsToFloat:
795 case vmIntrinsics::_doubleToRawLongBits:
796 case vmIntrinsics::_doubleToLongBits:
797 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
798
799 case vmIntrinsics::_numberOfLeadingZeros_i:
800 case vmIntrinsics::_numberOfLeadingZeros_l:
801 case vmIntrinsics::_numberOfTrailingZeros_i:
802 case vmIntrinsics::_numberOfTrailingZeros_l:
803 case vmIntrinsics::_bitCount_i:
804 case vmIntrinsics::_bitCount_l:
805 case vmIntrinsics::_reverseBytes_i:
806 case vmIntrinsics::_reverseBytes_l:
807 case vmIntrinsics::_reverseBytes_s:
808 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
809
810 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
811
812 case vmIntrinsics::_Reference_get: return inline_reference_get();
813
814 case vmIntrinsics::_Class_cast: return inline_Class_cast();
815
816 case vmIntrinsics::_aescrypt_encryptBlock:
817 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
818
819 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
820 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
821 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
822
823 case vmIntrinsics::_counterMode_AESCrypt:
824 return inline_counterMode_AESCrypt(intrinsic_id());
825
826 case vmIntrinsics::_sha_implCompress:
827 case vmIntrinsics::_sha2_implCompress:
828 case vmIntrinsics::_sha5_implCompress:
829 return inline_sha_implCompress(intrinsic_id());
830
831 case vmIntrinsics::_digestBase_implCompressMB:
832 return inline_digestBase_implCompressMB(predicate);
833
834 case vmIntrinsics::_multiplyToLen:
835 return inline_multiplyToLen();
836
837 case vmIntrinsics::_squareToLen:
838 return inline_squareToLen();
839
840 case vmIntrinsics::_mulAdd:
841 return inline_mulAdd();
842
843 case vmIntrinsics::_montgomeryMultiply:
844 return inline_montgomeryMultiply();
845 case vmIntrinsics::_montgomerySquare:
846 return inline_montgomerySquare();
847
848 case vmIntrinsics::_vectorizedMismatch:
849 return inline_vectorizedMismatch();
850
851 case vmIntrinsics::_ghash_processBlocks:
852 return inline_ghash_processBlocks();
853
854 case vmIntrinsics::_encodeISOArray:
855 case vmIntrinsics::_encodeByteISOArray:
856 return inline_encodeISOArray();
857
858 case vmIntrinsics::_updateCRC32:
859 return inline_updateCRC32();
860 case vmIntrinsics::_updateBytesCRC32:
861 return inline_updateBytesCRC32();
862 case vmIntrinsics::_updateByteBufferCRC32:
863 return inline_updateByteBufferCRC32();
864
865 case vmIntrinsics::_updateBytesCRC32C:
866 return inline_updateBytesCRC32C();
867 case vmIntrinsics::_updateDirectByteBufferCRC32C:
868 return inline_updateDirectByteBufferCRC32C();
869
870 case vmIntrinsics::_updateBytesAdler32:
871 return inline_updateBytesAdler32();
872 case vmIntrinsics::_updateByteBufferAdler32:
873 return inline_updateByteBufferAdler32();
874
875 case vmIntrinsics::_profileBoolean:
876 return inline_profileBoolean();
877 case vmIntrinsics::_isCompileConstant:
878 return inline_isCompileConstant();
879
880 case vmIntrinsics::_hasNegatives:
881 return inline_hasNegatives();
882
883 case vmIntrinsics::_fmaD:
884 case vmIntrinsics::_fmaF:
885 return inline_fma(intrinsic_id());
886
887 case vmIntrinsics::_VectorUnaryOp:
888 return inline_vector_nary_operation(1);
889 case vmIntrinsics::_VectorBinaryOp:
890 return inline_vector_nary_operation(2);
891 case vmIntrinsics::_VectorTernaryOp:
892 return inline_vector_nary_operation(3);
893
894 case vmIntrinsics::_VectorBroadcastCoerced:
895 return inline_vector_broadcast_coerced();
896 case vmIntrinsics::_VectorLoadOp:
897 return inline_vector_mem_operation(/*is_store=*/false);
898 case vmIntrinsics::_VectorStoreOp:
899 return inline_vector_mem_operation(/*is_store=*/true);
900 case vmIntrinsics::_VectorReductionCoerced:
901 return inline_vector_reduction();
902 case vmIntrinsics::_VectorTest:
903 return inline_vector_test();
904 case vmIntrinsics::_VectorBlend:
905 return inline_vector_blend();
906 case vmIntrinsics::_VectorCompare:
907 return inline_vector_compare();
908 case vmIntrinsics::_VectorBroadcastInt:
909 return inline_vector_broadcast_int();
910 case vmIntrinsics::_VectorRebracket:
911 return inline_vector_rebracket();
912
913 default:
914 // If you get here, it may be that someone has added a new intrinsic
915 // to the list in vmSymbols.hpp without implementing it here.
916 #ifndef PRODUCT
917 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
918 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
919 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
920 }
921 #endif
922 return false;
923 }
924 }
925
926 Node* LibraryCallKit::try_to_predicate(int predicate) {
927 if (!jvms()->has_method()) {
928 // Root JVMState has a null method.
929 assert(map()->memory()->Opcode() == Op_Parm, "");
930 // Insert the memory aliasing node
931 set_all_memory(reset_memory());
932 }
933 assert(merged_memory(), "");
934
935 switch (intrinsic_id()) {
936 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
937 return inline_cipherBlockChaining_AESCrypt_predicate(false);
938 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
939 return inline_cipherBlockChaining_AESCrypt_predicate(true);
940 case vmIntrinsics::_counterMode_AESCrypt:
941 return inline_counterMode_AESCrypt_predicate();
942 case vmIntrinsics::_digestBase_implCompressMB:
943 return inline_digestBase_implCompressMB_predicate(predicate);
944
945 default:
946 // If you get here, it may be that someone has added a new intrinsic
947 // to the list in vmSymbols.hpp without implementing it here.
948 #ifndef PRODUCT
949 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
950 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
951 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
952 }
953 #endif
954 Node* slow_ctl = control();
955 set_control(top()); // No fast path instrinsic
956 return slow_ctl;
957 }
958 }
959
960 //------------------------------set_result-------------------------------
961 // Helper function for finishing intrinsics.
962 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
963 record_for_igvn(region);
964 set_control(_gvn.transform(region));
965 set_result( _gvn.transform(value));
966 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
967 }
968
969 //------------------------------generate_guard---------------------------
970 // Helper function for generating guarded fast-slow graph structures.
971 // The given 'test', if true, guards a slow path. If the test fails
972 // then a fast path can be taken. (We generally hope it fails.)
973 // In all cases, GraphKit::control() is updated to the fast path.
974 // The returned value represents the control for the slow path.
975 // The return value is never 'top'; it is either a valid control
976 // or NULL if it is obvious that the slow path can never be taken.
977 // Also, if region and the slow control are not NULL, the slow edge
978 // is appended to the region.
979 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
980 if (stopped()) {
981 // Already short circuited.
982 return NULL;
983 }
984
985 // Build an if node and its projections.
986 // If test is true we take the slow path, which we assume is uncommon.
987 if (_gvn.type(test) == TypeInt::ZERO) {
988 // The slow branch is never taken. No need to build this guard.
989 return NULL;
990 }
991
992 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
993
994 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
995 if (if_slow == top()) {
996 // The slow branch is never taken. No need to build this guard.
997 return NULL;
998 }
999
1000 if (region != NULL)
1001 region->add_req(if_slow);
1002
1003 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1004 set_control(if_fast);
1005
1006 return if_slow;
1007 }
1008
1009 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1010 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1011 }
1012 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1013 return generate_guard(test, region, PROB_FAIR);
1014 }
1015
1016 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1017 Node* *pos_index) {
1018 if (stopped())
1019 return NULL; // already stopped
1020 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1021 return NULL; // index is already adequately typed
1022 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1023 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1024 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1025 if (is_neg != NULL && pos_index != NULL) {
1026 // Emulate effect of Parse::adjust_map_after_if.
1027 Node* ccast = new CastIINode(index, TypeInt::POS);
1028 ccast->set_req(0, control());
1029 (*pos_index) = _gvn.transform(ccast);
1030 }
1031 return is_neg;
1032 }
1033
1034 // Make sure that 'position' is a valid limit index, in [0..length].
1035 // There are two equivalent plans for checking this:
1036 // A. (offset + copyLength) unsigned<= arrayLength
1037 // B. offset <= (arrayLength - copyLength)
1038 // We require that all of the values above, except for the sum and
1039 // difference, are already known to be non-negative.
1040 // Plan A is robust in the face of overflow, if offset and copyLength
1041 // are both hugely positive.
1042 //
1043 // Plan B is less direct and intuitive, but it does not overflow at
1044 // all, since the difference of two non-negatives is always
1045 // representable. Whenever Java methods must perform the equivalent
1046 // check they generally use Plan B instead of Plan A.
1047 // For the moment we use Plan A.
1048 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1049 Node* subseq_length,
1050 Node* array_length,
1051 RegionNode* region) {
1052 if (stopped())
1053 return NULL; // already stopped
1054 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1055 if (zero_offset && subseq_length->eqv_uncast(array_length))
1056 return NULL; // common case of whole-array copy
1057 Node* last = subseq_length;
1058 if (!zero_offset) // last += offset
1059 last = _gvn.transform(new AddINode(last, offset));
1060 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1061 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1062 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1063 return is_over;
1064 }
1065
1066 // Emit range checks for the given String.value byte array
1067 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1068 if (stopped()) {
1069 return; // already stopped
1070 }
1071 RegionNode* bailout = new RegionNode(1);
1072 record_for_igvn(bailout);
1073 if (char_count) {
1074 // Convert char count to byte count
1075 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1076 }
1077
1078 // Offset and count must not be negative
1079 generate_negative_guard(offset, bailout);
1080 generate_negative_guard(count, bailout);
1081 // Offset + count must not exceed length of array
1082 generate_limit_guard(offset, count, load_array_length(array), bailout);
1083
1084 if (bailout->req() > 1) {
1085 PreserveJVMState pjvms(this);
1086 set_control(_gvn.transform(bailout));
1087 uncommon_trap(Deoptimization::Reason_intrinsic,
1088 Deoptimization::Action_maybe_recompile);
1089 }
1090 }
1091
1092 //--------------------------generate_current_thread--------------------
1093 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1094 ciKlass* thread_klass = env()->Thread_klass();
1095 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1096 Node* thread = _gvn.transform(new ThreadLocalNode());
1097 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1098 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1099 tls_output = thread;
1100 return threadObj;
1101 }
1102
1103
1104 //------------------------------make_string_method_node------------------------
1105 // Helper method for String intrinsic functions. This version is called with
1106 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1107 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1108 // containing the lengths of str1 and str2.
1109 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1110 Node* result = NULL;
1111 switch (opcode) {
1112 case Op_StrIndexOf:
1113 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1114 str1_start, cnt1, str2_start, cnt2, ae);
1115 break;
1116 case Op_StrComp:
1117 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1118 str1_start, cnt1, str2_start, cnt2, ae);
1119 break;
1120 case Op_StrEquals:
1121 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1122 // Use the constant length if there is one because optimized match rule may exist.
1123 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1124 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1125 break;
1126 default:
1127 ShouldNotReachHere();
1128 return NULL;
1129 }
1130
1131 // All these intrinsics have checks.
1132 C->set_has_split_ifs(true); // Has chance for split-if optimization
1133 clear_upper_avx();
1134
1135 return _gvn.transform(result);
1136 }
1137
1138 //------------------------------inline_string_compareTo------------------------
1139 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1140 Node* arg1 = argument(0);
1141 Node* arg2 = argument(1);
1142
1143 // Get start addr and length of first argument
1144 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1145 Node* arg1_cnt = load_array_length(arg1);
1146
1147 // Get start addr and length of second argument
1148 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1149 Node* arg2_cnt = load_array_length(arg2);
1150
1151 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1152 set_result(result);
1153 return true;
1154 }
1155
1156 //------------------------------inline_string_equals------------------------
1157 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1158 Node* arg1 = argument(0);
1159 Node* arg2 = argument(1);
1160
1161 // paths (plus control) merge
1162 RegionNode* region = new RegionNode(3);
1163 Node* phi = new PhiNode(region, TypeInt::BOOL);
1164
1165 if (!stopped()) {
1166 // Get start addr and length of first argument
1167 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1168 Node* arg1_cnt = load_array_length(arg1);
1169
1170 // Get start addr and length of second argument
1171 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1172 Node* arg2_cnt = load_array_length(arg2);
1173
1174 // Check for arg1_cnt != arg2_cnt
1175 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1176 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1177 Node* if_ne = generate_slow_guard(bol, NULL);
1178 if (if_ne != NULL) {
1179 phi->init_req(2, intcon(0));
1180 region->init_req(2, if_ne);
1181 }
1182
1183 // Check for count == 0 is done by assembler code for StrEquals.
1184
1185 if (!stopped()) {
1186 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1187 phi->init_req(1, equals);
1188 region->init_req(1, control());
1189 }
1190 }
1191
1192 // post merge
1193 set_control(_gvn.transform(region));
1194 record_for_igvn(region);
1195
1196 set_result(_gvn.transform(phi));
1197 return true;
1198 }
1199
1200 //------------------------------inline_array_equals----------------------------
1201 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1202 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1203 Node* arg1 = argument(0);
1204 Node* arg2 = argument(1);
1205
1206 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1207 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1208 clear_upper_avx();
1209
1210 return true;
1211 }
1212
1213 //------------------------------inline_hasNegatives------------------------------
1214 bool LibraryCallKit::inline_hasNegatives() {
1215 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1216 return false;
1217 }
1218
1219 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1220 // no receiver since it is static method
1221 Node* ba = argument(0);
1222 Node* offset = argument(1);
1223 Node* len = argument(2);
1224
1225 // Range checks
1226 generate_string_range_check(ba, offset, len, false);
1227 if (stopped()) {
1228 return true;
1229 }
1230 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1231 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1232 set_result(_gvn.transform(result));
1233 return true;
1234 }
1235
1236 bool LibraryCallKit::inline_preconditions_checkIndex() {
1237 Node* index = argument(0);
1238 Node* length = argument(1);
1239 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1240 return false;
1241 }
1242
1243 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1244 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1245
1246 {
1247 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1248 uncommon_trap(Deoptimization::Reason_intrinsic,
1249 Deoptimization::Action_make_not_entrant);
1250 }
1251
1252 if (stopped()) {
1253 return false;
1254 }
1255
1256 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1257 BoolTest::mask btest = BoolTest::lt;
1258 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1259 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1260 _gvn.set_type(rc, rc->Value(&_gvn));
1261 if (!rc_bool->is_Con()) {
1262 record_for_igvn(rc);
1263 }
1264 set_control(_gvn.transform(new IfTrueNode(rc)));
1265 {
1266 PreserveJVMState pjvms(this);
1267 set_control(_gvn.transform(new IfFalseNode(rc)));
1268 uncommon_trap(Deoptimization::Reason_range_check,
1269 Deoptimization::Action_make_not_entrant);
1270 }
1271
1272 if (stopped()) {
1273 return false;
1274 }
1275
1276 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1277 result->set_req(0, control());
1278 result = _gvn.transform(result);
1279 set_result(result);
1280 replace_in_map(index, result);
1281 clear_upper_avx();
1282 return true;
1283 }
1284
1285 //------------------------------inline_string_indexOf------------------------
1286 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1287 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1288 return false;
1289 }
1290 Node* src = argument(0);
1291 Node* tgt = argument(1);
1292
1293 // Make the merge point
1294 RegionNode* result_rgn = new RegionNode(4);
1295 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1296
1297 // Get start addr and length of source string
1298 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1299 Node* src_count = load_array_length(src);
1300
1301 // Get start addr and length of substring
1302 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1303 Node* tgt_count = load_array_length(tgt);
1304
1305 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1306 // Divide src size by 2 if String is UTF16 encoded
1307 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1308 }
1309 if (ae == StrIntrinsicNode::UU) {
1310 // Divide substring size by 2 if String is UTF16 encoded
1311 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1312 }
1313
1314 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1315 if (result != NULL) {
1316 result_phi->init_req(3, result);
1317 result_rgn->init_req(3, control());
1318 }
1319 set_control(_gvn.transform(result_rgn));
1320 record_for_igvn(result_rgn);
1321 set_result(_gvn.transform(result_phi));
1322
1323 return true;
1324 }
1325
1326 //-----------------------------inline_string_indexOf-----------------------
1327 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1328 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1329 return false;
1330 }
1331 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1332 return false;
1333 }
1334 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1335 Node* src = argument(0); // byte[]
1336 Node* src_count = argument(1); // char count
1337 Node* tgt = argument(2); // byte[]
1338 Node* tgt_count = argument(3); // char count
1339 Node* from_index = argument(4); // char index
1340
1341 // Multiply byte array index by 2 if String is UTF16 encoded
1342 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1343 src_count = _gvn.transform(new SubINode(src_count, from_index));
1344 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1345 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1346
1347 // Range checks
1348 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1349 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1350 if (stopped()) {
1351 return true;
1352 }
1353
1354 RegionNode* region = new RegionNode(5);
1355 Node* phi = new PhiNode(region, TypeInt::INT);
1356
1357 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1358 if (result != NULL) {
1359 // The result is index relative to from_index if substring was found, -1 otherwise.
1360 // Generate code which will fold into cmove.
1361 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1362 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1363
1364 Node* if_lt = generate_slow_guard(bol, NULL);
1365 if (if_lt != NULL) {
1366 // result == -1
1367 phi->init_req(3, result);
1368 region->init_req(3, if_lt);
1369 }
1370 if (!stopped()) {
1371 result = _gvn.transform(new AddINode(result, from_index));
1372 phi->init_req(4, result);
1373 region->init_req(4, control());
1374 }
1375 }
1376
1377 set_control(_gvn.transform(region));
1378 record_for_igvn(region);
1379 set_result(_gvn.transform(phi));
1380 clear_upper_avx();
1381
1382 return true;
1383 }
1384
1385 // Create StrIndexOfNode with fast path checks
1386 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1387 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1388 // Check for substr count > string count
1389 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1390 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1391 Node* if_gt = generate_slow_guard(bol, NULL);
1392 if (if_gt != NULL) {
1393 phi->init_req(1, intcon(-1));
1394 region->init_req(1, if_gt);
1395 }
1396 if (!stopped()) {
1397 // Check for substr count == 0
1398 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1399 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1400 Node* if_zero = generate_slow_guard(bol, NULL);
1401 if (if_zero != NULL) {
1402 phi->init_req(2, intcon(0));
1403 region->init_req(2, if_zero);
1404 }
1405 }
1406 if (!stopped()) {
1407 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1408 }
1409 return NULL;
1410 }
1411
1412 //-----------------------------inline_string_indexOfChar-----------------------
1413 bool LibraryCallKit::inline_string_indexOfChar() {
1414 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1415 return false;
1416 }
1417 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1418 return false;
1419 }
1420 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1421 Node* src = argument(0); // byte[]
1422 Node* tgt = argument(1); // tgt is int ch
1423 Node* from_index = argument(2);
1424 Node* max = argument(3);
1425
1426 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1427 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1428 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1429
1430 // Range checks
1431 generate_string_range_check(src, src_offset, src_count, true);
1432 if (stopped()) {
1433 return true;
1434 }
1435
1436 RegionNode* region = new RegionNode(3);
1437 Node* phi = new PhiNode(region, TypeInt::INT);
1438
1439 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1440 C->set_has_split_ifs(true); // Has chance for split-if optimization
1441 _gvn.transform(result);
1442
1443 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1444 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1445
1446 Node* if_lt = generate_slow_guard(bol, NULL);
1447 if (if_lt != NULL) {
1448 // result == -1
1449 phi->init_req(2, result);
1450 region->init_req(2, if_lt);
1451 }
1452 if (!stopped()) {
1453 result = _gvn.transform(new AddINode(result, from_index));
1454 phi->init_req(1, result);
1455 region->init_req(1, control());
1456 }
1457 set_control(_gvn.transform(region));
1458 record_for_igvn(region);
1459 set_result(_gvn.transform(phi));
1460
1461 return true;
1462 }
1463 //---------------------------inline_string_copy---------------------
1464 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1465 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1466 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1467 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1468 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1469 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1470 bool LibraryCallKit::inline_string_copy(bool compress) {
1471 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1472 return false;
1473 }
1474 int nargs = 5; // 2 oops, 3 ints
1475 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1476
1477 Node* src = argument(0);
1478 Node* src_offset = argument(1);
1479 Node* dst = argument(2);
1480 Node* dst_offset = argument(3);
1481 Node* length = argument(4);
1482
1483 // Check for allocation before we add nodes that would confuse
1484 // tightly_coupled_allocation()
1485 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1486
1487 // Figure out the size and type of the elements we will be copying.
1488 const Type* src_type = src->Value(&_gvn);
1489 const Type* dst_type = dst->Value(&_gvn);
1490 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1491 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1492 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1493 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1494 "Unsupported array types for inline_string_copy");
1495
1496 // Convert char[] offsets to byte[] offsets
1497 bool convert_src = (compress && src_elem == T_BYTE);
1498 bool convert_dst = (!compress && dst_elem == T_BYTE);
1499 if (convert_src) {
1500 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1501 } else if (convert_dst) {
1502 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1503 }
1504
1505 // Range checks
1506 generate_string_range_check(src, src_offset, length, convert_src);
1507 generate_string_range_check(dst, dst_offset, length, convert_dst);
1508 if (stopped()) {
1509 return true;
1510 }
1511
1512 Node* src_start = array_element_address(src, src_offset, src_elem);
1513 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1514 // 'src_start' points to src array + scaled offset
1515 // 'dst_start' points to dst array + scaled offset
1516 Node* count = NULL;
1517 if (compress) {
1518 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1519 } else {
1520 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1521 }
1522
1523 if (alloc != NULL) {
1524 if (alloc->maybe_set_complete(&_gvn)) {
1525 // "You break it, you buy it."
1526 InitializeNode* init = alloc->initialization();
1527 assert(init->is_complete(), "we just did this");
1528 init->set_complete_with_arraycopy();
1529 assert(dst->is_CheckCastPP(), "sanity");
1530 assert(dst->in(0)->in(0) == init, "dest pinned");
1531 }
1532 // Do not let stores that initialize this object be reordered with
1533 // a subsequent store that would make this object accessible by
1534 // other threads.
1535 // Record what AllocateNode this StoreStore protects so that
1536 // escape analysis can go from the MemBarStoreStoreNode to the
1537 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1538 // based on the escape status of the AllocateNode.
1539 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1540 }
1541 if (compress) {
1542 set_result(_gvn.transform(count));
1543 }
1544 clear_upper_avx();
1545
1546 return true;
1547 }
1548
1549 #ifdef _LP64
1550 #define XTOP ,top() /*additional argument*/
1551 #else //_LP64
1552 #define XTOP /*no additional argument*/
1553 #endif //_LP64
1554
1555 //------------------------inline_string_toBytesU--------------------------
1556 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1557 bool LibraryCallKit::inline_string_toBytesU() {
1558 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1559 return false;
1560 }
1561 // Get the arguments.
1562 Node* value = argument(0);
1563 Node* offset = argument(1);
1564 Node* length = argument(2);
1565
1566 Node* newcopy = NULL;
1567
1568 // Set the original stack and the reexecute bit for the interpreter to reexecute
1569 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1570 { PreserveReexecuteState preexecs(this);
1571 jvms()->set_should_reexecute(true);
1572
1573 // Check if a null path was taken unconditionally.
1574 value = null_check(value);
1575
1576 RegionNode* bailout = new RegionNode(1);
1577 record_for_igvn(bailout);
1578
1579 // Range checks
1580 generate_negative_guard(offset, bailout);
1581 generate_negative_guard(length, bailout);
1582 generate_limit_guard(offset, length, load_array_length(value), bailout);
1583 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1584 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1585
1586 if (bailout->req() > 1) {
1587 PreserveJVMState pjvms(this);
1588 set_control(_gvn.transform(bailout));
1589 uncommon_trap(Deoptimization::Reason_intrinsic,
1590 Deoptimization::Action_maybe_recompile);
1591 }
1592 if (stopped()) {
1593 return true;
1594 }
1595
1596 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1597 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1598 newcopy = new_array(klass_node, size, 0); // no arguments to push
1599 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1600
1601 // Calculate starting addresses.
1602 Node* src_start = array_element_address(value, offset, T_CHAR);
1603 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1604
1605 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1606 const TypeInt* toffset = gvn().type(offset)->is_int();
1607 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1608
1609 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1610 const char* copyfunc_name = "arraycopy";
1611 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1612 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1613 OptoRuntime::fast_arraycopy_Type(),
1614 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1615 src_start, dst_start, ConvI2X(length) XTOP);
1616 // Do not let reads from the cloned object float above the arraycopy.
1617 if (alloc != NULL) {
1618 if (alloc->maybe_set_complete(&_gvn)) {
1619 // "You break it, you buy it."
1620 InitializeNode* init = alloc->initialization();
1621 assert(init->is_complete(), "we just did this");
1622 init->set_complete_with_arraycopy();
1623 assert(newcopy->is_CheckCastPP(), "sanity");
1624 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1625 }
1626 // Do not let stores that initialize this object be reordered with
1627 // a subsequent store that would make this object accessible by
1628 // other threads.
1629 // Record what AllocateNode this StoreStore protects so that
1630 // escape analysis can go from the MemBarStoreStoreNode to the
1631 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1632 // based on the escape status of the AllocateNode.
1633 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1634 } else {
1635 insert_mem_bar(Op_MemBarCPUOrder);
1636 }
1637 } // original reexecute is set back here
1638
1639 C->set_has_split_ifs(true); // Has chance for split-if optimization
1640 if (!stopped()) {
1641 set_result(newcopy);
1642 }
1643 clear_upper_avx();
1644
1645 return true;
1646 }
1647
1648 //------------------------inline_string_getCharsU--------------------------
1649 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1650 bool LibraryCallKit::inline_string_getCharsU() {
1651 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1652 return false;
1653 }
1654
1655 // Get the arguments.
1656 Node* src = argument(0);
1657 Node* src_begin = argument(1);
1658 Node* src_end = argument(2); // exclusive offset (i < src_end)
1659 Node* dst = argument(3);
1660 Node* dst_begin = argument(4);
1661
1662 // Check for allocation before we add nodes that would confuse
1663 // tightly_coupled_allocation()
1664 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1665
1666 // Check if a null path was taken unconditionally.
1667 src = null_check(src);
1668 dst = null_check(dst);
1669 if (stopped()) {
1670 return true;
1671 }
1672
1673 // Get length and convert char[] offset to byte[] offset
1674 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1675 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1676
1677 // Range checks
1678 generate_string_range_check(src, src_begin, length, true);
1679 generate_string_range_check(dst, dst_begin, length, false);
1680 if (stopped()) {
1681 return true;
1682 }
1683
1684 if (!stopped()) {
1685 // Calculate starting addresses.
1686 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1687 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1688
1689 // Check if array addresses are aligned to HeapWordSize
1690 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1691 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1692 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1693 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1694
1695 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1696 const char* copyfunc_name = "arraycopy";
1697 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1698 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1699 OptoRuntime::fast_arraycopy_Type(),
1700 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1701 src_start, dst_start, ConvI2X(length) XTOP);
1702 // Do not let reads from the cloned object float above the arraycopy.
1703 if (alloc != NULL) {
1704 if (alloc->maybe_set_complete(&_gvn)) {
1705 // "You break it, you buy it."
1706 InitializeNode* init = alloc->initialization();
1707 assert(init->is_complete(), "we just did this");
1708 init->set_complete_with_arraycopy();
1709 assert(dst->is_CheckCastPP(), "sanity");
1710 assert(dst->in(0)->in(0) == init, "dest pinned");
1711 }
1712 // Do not let stores that initialize this object be reordered with
1713 // a subsequent store that would make this object accessible by
1714 // other threads.
1715 // Record what AllocateNode this StoreStore protects so that
1716 // escape analysis can go from the MemBarStoreStoreNode to the
1717 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1718 // based on the escape status of the AllocateNode.
1719 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1720 } else {
1721 insert_mem_bar(Op_MemBarCPUOrder);
1722 }
1723 }
1724
1725 C->set_has_split_ifs(true); // Has chance for split-if optimization
1726 return true;
1727 }
1728
1729 //----------------------inline_string_char_access----------------------------
1730 // Store/Load char to/from byte[] array.
1731 // static void StringUTF16.putChar(byte[] val, int index, int c)
1732 // static char StringUTF16.getChar(byte[] val, int index)
1733 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1734 Node* value = argument(0);
1735 Node* index = argument(1);
1736 Node* ch = is_store ? argument(2) : NULL;
1737
1738 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1739 // correctly requires matched array shapes.
1740 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1741 "sanity: byte[] and char[] bases agree");
1742 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1743 "sanity: byte[] and char[] scales agree");
1744
1745 // Bail when getChar over constants is requested: constant folding would
1746 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1747 // Java method would constant fold nicely instead.
1748 if (!is_store && value->is_Con() && index->is_Con()) {
1749 return false;
1750 }
1751
1752 Node* adr = array_element_address(value, index, T_CHAR);
1753 if (adr->is_top()) {
1754 return false;
1755 }
1756 if (is_store) {
1757 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1758 false, false, true /* mismatched */);
1759 } else {
1760 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1761 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1762 set_result(ch);
1763 }
1764 return true;
1765 }
1766
1767 //--------------------------round_double_node--------------------------------
1768 // Round a double node if necessary.
1769 Node* LibraryCallKit::round_double_node(Node* n) {
1770 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1771 n = _gvn.transform(new RoundDoubleNode(0, n));
1772 return n;
1773 }
1774
1775 //------------------------------inline_math-----------------------------------
1776 // public static double Math.abs(double)
1777 // public static double Math.sqrt(double)
1778 // public static double Math.log(double)
1779 // public static double Math.log10(double)
1780 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1781 Node* arg = round_double_node(argument(0));
1782 Node* n = NULL;
1783 switch (id) {
1784 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1785 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1786 default: fatal_unexpected_iid(id); break;
1787 }
1788 set_result(_gvn.transform(n));
1789 return true;
1790 }
1791
1792 //------------------------------runtime_math-----------------------------
1793 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1794 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1795 "must be (DD)D or (D)D type");
1796
1797 // Inputs
1798 Node* a = round_double_node(argument(0));
1799 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1800
1801 const TypePtr* no_memory_effects = NULL;
1802 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1803 no_memory_effects,
1804 a, top(), b, b ? top() : NULL);
1805 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1806 #ifdef ASSERT
1807 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1808 assert(value_top == top(), "second value must be top");
1809 #endif
1810
1811 set_result(value);
1812 return true;
1813 }
1814
1815 //------------------------------inline_math_native-----------------------------
1816 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1817 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1818 switch (id) {
1819 // These intrinsics are not properly supported on all hardware
1820 case vmIntrinsics::_dsin:
1821 return StubRoutines::dsin() != NULL ?
1822 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1823 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1824 case vmIntrinsics::_dcos:
1825 return StubRoutines::dcos() != NULL ?
1826 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1827 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1828 case vmIntrinsics::_dtan:
1829 return StubRoutines::dtan() != NULL ?
1830 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1831 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1832 case vmIntrinsics::_dlog:
1833 return StubRoutines::dlog() != NULL ?
1834 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1835 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1836 case vmIntrinsics::_dlog10:
1837 return StubRoutines::dlog10() != NULL ?
1838 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1839 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1840
1841 // These intrinsics are supported on all hardware
1842 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1843 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1844
1845 case vmIntrinsics::_dexp:
1846 return StubRoutines::dexp() != NULL ?
1847 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1848 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1849 case vmIntrinsics::_dpow: {
1850 Node* exp = round_double_node(argument(2));
1851 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1852 if (d != NULL && d->getd() == 2.0) {
1853 // Special case: pow(x, 2.0) => x * x
1854 Node* base = round_double_node(argument(0));
1855 set_result(_gvn.transform(new MulDNode(base, base)));
1856 return true;
1857 }
1858 return StubRoutines::dexp() != NULL ?
1859 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1860 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1861 }
1862 #undef FN_PTR
1863
1864 // These intrinsics are not yet correctly implemented
1865 case vmIntrinsics::_datan2:
1866 return false;
1867
1868 default:
1869 fatal_unexpected_iid(id);
1870 return false;
1871 }
1872 }
1873
1874 static bool is_simple_name(Node* n) {
1875 return (n->req() == 1 // constant
1876 || (n->is_Type() && n->as_Type()->type()->singleton())
1877 || n->is_Proj() // parameter or return value
1878 || n->is_Phi() // local of some sort
1879 );
1880 }
1881
1882 //----------------------------inline_notify-----------------------------------*
1883 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1884 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1885 address func;
1886 if (id == vmIntrinsics::_notify) {
1887 func = OptoRuntime::monitor_notify_Java();
1888 } else {
1889 func = OptoRuntime::monitor_notifyAll_Java();
1890 }
1891 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1892 make_slow_call_ex(call, env()->Throwable_klass(), false);
1893 return true;
1894 }
1895
1896
1897 //----------------------------inline_min_max-----------------------------------
1898 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1899 set_result(generate_min_max(id, argument(0), argument(1)));
1900 return true;
1901 }
1902
1903 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1904 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1905 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1906 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1907 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1908
1909 {
1910 PreserveJVMState pjvms(this);
1911 PreserveReexecuteState preexecs(this);
1912 jvms()->set_should_reexecute(true);
1913
1914 set_control(slow_path);
1915 set_i_o(i_o());
1916
1917 uncommon_trap(Deoptimization::Reason_intrinsic,
1918 Deoptimization::Action_none);
1919 }
1920
1921 set_control(fast_path);
1922 set_result(math);
1923 }
1924
1925 template <typename OverflowOp>
1926 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1927 typedef typename OverflowOp::MathOp MathOp;
1928
1929 MathOp* mathOp = new MathOp(arg1, arg2);
1930 Node* operation = _gvn.transform( mathOp );
1931 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1932 inline_math_mathExact(operation, ofcheck);
1933 return true;
1934 }
1935
1936 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1937 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1938 }
1939
1940 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1941 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1942 }
1943
1944 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1945 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1946 }
1947
1948 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1949 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1950 }
1951
1952 bool LibraryCallKit::inline_math_negateExactI() {
1953 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1954 }
1955
1956 bool LibraryCallKit::inline_math_negateExactL() {
1957 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1958 }
1959
1960 bool LibraryCallKit::inline_math_multiplyExactI() {
1961 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1962 }
1963
1964 bool LibraryCallKit::inline_math_multiplyExactL() {
1965 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1966 }
1967
1968 bool LibraryCallKit::inline_math_multiplyHigh() {
1969 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1970 return true;
1971 }
1972
1973 Node*
1974 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1975 // These are the candidate return value:
1976 Node* xvalue = x0;
1977 Node* yvalue = y0;
1978
1979 if (xvalue == yvalue) {
1980 return xvalue;
1981 }
1982
1983 bool want_max = (id == vmIntrinsics::_max);
1984
1985 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1986 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1987 if (txvalue == NULL || tyvalue == NULL) return top();
1988 // This is not really necessary, but it is consistent with a
1989 // hypothetical MaxINode::Value method:
1990 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1991
1992 // %%% This folding logic should (ideally) be in a different place.
1993 // Some should be inside IfNode, and there to be a more reliable
1994 // transformation of ?: style patterns into cmoves. We also want
1995 // more powerful optimizations around cmove and min/max.
1996
1997 // Try to find a dominating comparison of these guys.
1998 // It can simplify the index computation for Arrays.copyOf
1999 // and similar uses of System.arraycopy.
2000 // First, compute the normalized version of CmpI(x, y).
2001 int cmp_op = Op_CmpI;
2002 Node* xkey = xvalue;
2003 Node* ykey = yvalue;
2004 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2005 if (ideal_cmpxy->is_Cmp()) {
2006 // E.g., if we have CmpI(length - offset, count),
2007 // it might idealize to CmpI(length, count + offset)
2008 cmp_op = ideal_cmpxy->Opcode();
2009 xkey = ideal_cmpxy->in(1);
2010 ykey = ideal_cmpxy->in(2);
2011 }
2012
2013 // Start by locating any relevant comparisons.
2014 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2015 Node* cmpxy = NULL;
2016 Node* cmpyx = NULL;
2017 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2018 Node* cmp = start_from->fast_out(k);
2019 if (cmp->outcnt() > 0 && // must have prior uses
2020 cmp->in(0) == NULL && // must be context-independent
2021 cmp->Opcode() == cmp_op) { // right kind of compare
2022 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2023 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2024 }
2025 }
2026
2027 const int NCMPS = 2;
2028 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2029 int cmpn;
2030 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2031 if (cmps[cmpn] != NULL) break; // find a result
2032 }
2033 if (cmpn < NCMPS) {
2034 // Look for a dominating test that tells us the min and max.
2035 int depth = 0; // Limit search depth for speed
2036 Node* dom = control();
2037 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2038 if (++depth >= 100) break;
2039 Node* ifproj = dom;
2040 if (!ifproj->is_Proj()) continue;
2041 Node* iff = ifproj->in(0);
2042 if (!iff->is_If()) continue;
2043 Node* bol = iff->in(1);
2044 if (!bol->is_Bool()) continue;
2045 Node* cmp = bol->in(1);
2046 if (cmp == NULL) continue;
2047 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2048 if (cmps[cmpn] == cmp) break;
2049 if (cmpn == NCMPS) continue;
2050 BoolTest::mask btest = bol->as_Bool()->_test._test;
2051 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2052 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2053 // At this point, we know that 'x btest y' is true.
2054 switch (btest) {
2055 case BoolTest::eq:
2056 // They are proven equal, so we can collapse the min/max.
2057 // Either value is the answer. Choose the simpler.
2058 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2059 return yvalue;
2060 return xvalue;
2061 case BoolTest::lt: // x < y
2062 case BoolTest::le: // x <= y
2063 return (want_max ? yvalue : xvalue);
2064 case BoolTest::gt: // x > y
2065 case BoolTest::ge: // x >= y
2066 return (want_max ? xvalue : yvalue);
2067 default:
2068 break;
2069 }
2070 }
2071 }
2072
2073 // We failed to find a dominating test.
2074 // Let's pick a test that might GVN with prior tests.
2075 Node* best_bol = NULL;
2076 BoolTest::mask best_btest = BoolTest::illegal;
2077 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2078 Node* cmp = cmps[cmpn];
2079 if (cmp == NULL) continue;
2080 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2081 Node* bol = cmp->fast_out(j);
2082 if (!bol->is_Bool()) continue;
2083 BoolTest::mask btest = bol->as_Bool()->_test._test;
2084 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2085 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2086 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2087 best_bol = bol->as_Bool();
2088 best_btest = btest;
2089 }
2090 }
2091 }
2092
2093 Node* answer_if_true = NULL;
2094 Node* answer_if_false = NULL;
2095 switch (best_btest) {
2096 default:
2097 if (cmpxy == NULL)
2098 cmpxy = ideal_cmpxy;
2099 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2100 // and fall through:
2101 case BoolTest::lt: // x < y
2102 case BoolTest::le: // x <= y
2103 answer_if_true = (want_max ? yvalue : xvalue);
2104 answer_if_false = (want_max ? xvalue : yvalue);
2105 break;
2106 case BoolTest::gt: // x > y
2107 case BoolTest::ge: // x >= y
2108 answer_if_true = (want_max ? xvalue : yvalue);
2109 answer_if_false = (want_max ? yvalue : xvalue);
2110 break;
2111 }
2112
2113 jint hi, lo;
2114 if (want_max) {
2115 // We can sharpen the minimum.
2116 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2117 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2118 } else {
2119 // We can sharpen the maximum.
2120 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2121 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2122 }
2123
2124 // Use a flow-free graph structure, to avoid creating excess control edges
2125 // which could hinder other optimizations.
2126 // Since Math.min/max is often used with arraycopy, we want
2127 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2128 Node* cmov = CMoveNode::make(NULL, best_bol,
2129 answer_if_false, answer_if_true,
2130 TypeInt::make(lo, hi, widen));
2131
2132 return _gvn.transform(cmov);
2133
2134 /*
2135 // This is not as desirable as it may seem, since Min and Max
2136 // nodes do not have a full set of optimizations.
2137 // And they would interfere, anyway, with 'if' optimizations
2138 // and with CMoveI canonical forms.
2139 switch (id) {
2140 case vmIntrinsics::_min:
2141 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2142 case vmIntrinsics::_max:
2143 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2144 default:
2145 ShouldNotReachHere();
2146 }
2147 */
2148 }
2149
2150 inline int
2151 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2152 const TypePtr* base_type = TypePtr::NULL_PTR;
2153 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2154 if (base_type == NULL) {
2155 // Unknown type.
2156 return Type::AnyPtr;
2157 } else if (base_type == TypePtr::NULL_PTR) {
2158 // Since this is a NULL+long form, we have to switch to a rawptr.
2159 base = _gvn.transform(new CastX2PNode(offset));
2160 offset = MakeConX(0);
2161 return Type::RawPtr;
2162 } else if (base_type->base() == Type::RawPtr) {
2163 return Type::RawPtr;
2164 } else if (base_type->isa_oopptr()) {
2165 // Base is never null => always a heap address.
2166 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2167 return Type::OopPtr;
2168 }
2169 // Offset is small => always a heap address.
2170 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2171 if (offset_type != NULL &&
2172 base_type->offset() == 0 && // (should always be?)
2173 offset_type->_lo >= 0 &&
2174 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2175 return Type::OopPtr;
2176 } else if (type == T_OBJECT) {
2177 // off heap access to an oop doesn't make any sense. Has to be on
2178 // heap.
2179 return Type::OopPtr;
2180 }
2181 // Otherwise, it might either be oop+off or NULL+addr.
2182 return Type::AnyPtr;
2183 } else {
2184 // No information:
2185 return Type::AnyPtr;
2186 }
2187 }
2188
2189 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2190 Node* uncasted_base = base;
2191 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2192 if (kind == Type::RawPtr) {
2193 return basic_plus_adr(top(), uncasted_base, offset);
2194 } else if (kind == Type::AnyPtr) {
2195 assert(base == uncasted_base, "unexpected base change");
2196 if (can_cast) {
2197 if (!_gvn.type(base)->speculative_maybe_null() &&
2198 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2199 // According to profiling, this access is always on
2200 // heap. Casting the base to not null and thus avoiding membars
2201 // around the access should allow better optimizations
2202 Node* null_ctl = top();
2203 base = null_check_oop(base, &null_ctl, true, true, true);
2204 assert(null_ctl->is_top(), "no null control here");
2205 return basic_plus_adr(base, offset);
2206 } else if (_gvn.type(base)->speculative_always_null() &&
2207 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2208 // According to profiling, this access is always off
2209 // heap.
2210 base = null_assert(base);
2211 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2212 offset = MakeConX(0);
2213 return basic_plus_adr(top(), raw_base, offset);
2214 }
2215 }
2216 // We don't know if it's an on heap or off heap access. Fall back
2217 // to raw memory access.
2218 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2219 return basic_plus_adr(top(), raw, offset);
2220 } else {
2221 assert(base == uncasted_base, "unexpected base change");
2222 // We know it's an on heap access so base can't be null
2223 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2224 base = must_be_not_null(base, true);
2225 }
2226 return basic_plus_adr(base, offset);
2227 }
2228 }
2229
2230 //--------------------------inline_number_methods-----------------------------
2231 // inline int Integer.numberOfLeadingZeros(int)
2232 // inline int Long.numberOfLeadingZeros(long)
2233 //
2234 // inline int Integer.numberOfTrailingZeros(int)
2235 // inline int Long.numberOfTrailingZeros(long)
2236 //
2237 // inline int Integer.bitCount(int)
2238 // inline int Long.bitCount(long)
2239 //
2240 // inline char Character.reverseBytes(char)
2241 // inline short Short.reverseBytes(short)
2242 // inline int Integer.reverseBytes(int)
2243 // inline long Long.reverseBytes(long)
2244 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2245 Node* arg = argument(0);
2246 Node* n = NULL;
2247 switch (id) {
2248 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2249 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2250 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2251 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2252 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2253 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2254 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2255 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2256 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2257 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2258 default: fatal_unexpected_iid(id); break;
2259 }
2260 set_result(_gvn.transform(n));
2261 return true;
2262 }
2263
2264 //----------------------------inline_unsafe_access----------------------------
2265
2266 // Helper that guards and inserts a pre-barrier.
2267 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2268 Node* pre_val, bool need_mem_bar) {
2269 // We could be accessing the referent field of a reference object. If so, when G1
2270 // is enabled, we need to log the value in the referent field in an SATB buffer.
2271 // This routine performs some compile time filters and generates suitable
2272 // runtime filters that guard the pre-barrier code.
2273 // Also add memory barrier for non volatile load from the referent field
2274 // to prevent commoning of loads across safepoint.
2275 if (!UseG1GC && !need_mem_bar)
2276 return;
2277
2278 // Some compile time checks.
2279
2280 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2281 const TypeX* otype = offset->find_intptr_t_type();
2282 if (otype != NULL && otype->is_con() &&
2283 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2284 // Constant offset but not the reference_offset so just return
2285 return;
2286 }
2287
2288 // We only need to generate the runtime guards for instances.
2289 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2290 if (btype != NULL) {
2291 if (btype->isa_aryptr()) {
2292 // Array type so nothing to do
2293 return;
2294 }
2295
2296 const TypeInstPtr* itype = btype->isa_instptr();
2297 if (itype != NULL) {
2298 // Can the klass of base_oop be statically determined to be
2299 // _not_ a sub-class of Reference and _not_ Object?
2300 ciKlass* klass = itype->klass();
2301 if ( klass->is_loaded() &&
2302 !klass->is_subtype_of(env()->Reference_klass()) &&
2303 !env()->Object_klass()->is_subtype_of(klass)) {
2304 return;
2305 }
2306 }
2307 }
2308
2309 // The compile time filters did not reject base_oop/offset so
2310 // we need to generate the following runtime filters
2311 //
2312 // if (offset == java_lang_ref_Reference::_reference_offset) {
2313 // if (instance_of(base, java.lang.ref.Reference)) {
2314 // pre_barrier(_, pre_val, ...);
2315 // }
2316 // }
2317
2318 float likely = PROB_LIKELY( 0.999);
2319 float unlikely = PROB_UNLIKELY(0.999);
2320
2321 IdealKit ideal(this);
2322 #define __ ideal.
2323
2324 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2325
2326 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2327 // Update graphKit memory and control from IdealKit.
2328 sync_kit(ideal);
2329
2330 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2331 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2332
2333 // Update IdealKit memory and control from graphKit.
2334 __ sync_kit(this);
2335
2336 Node* one = __ ConI(1);
2337 // is_instof == 0 if base_oop == NULL
2338 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2339
2340 // Update graphKit from IdeakKit.
2341 sync_kit(ideal);
2342
2343 // Use the pre-barrier to record the value in the referent field
2344 pre_barrier(false /* do_load */,
2345 __ ctrl(),
2346 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2347 pre_val /* pre_val */,
2348 T_OBJECT);
2349 if (need_mem_bar) {
2350 // Add memory barrier to prevent commoning reads from this field
2351 // across safepoint since GC can change its value.
2352 insert_mem_bar(Op_MemBarCPUOrder);
2353 }
2354 // Update IdealKit from graphKit.
2355 __ sync_kit(this);
2356
2357 } __ end_if(); // _ref_type != ref_none
2358 } __ end_if(); // offset == referent_offset
2359
2360 // Final sync IdealKit and GraphKit.
2361 final_sync(ideal);
2362 #undef __
2363 }
2364
2365
2366 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2367 // Attempt to infer a sharper value type from the offset and base type.
2368 ciKlass* sharpened_klass = NULL;
2369
2370 // See if it is an instance field, with an object type.
2371 if (alias_type->field() != NULL) {
2372 if (alias_type->field()->type()->is_klass()) {
2373 sharpened_klass = alias_type->field()->type()->as_klass();
2374 }
2375 }
2376
2377 // See if it is a narrow oop array.
2378 if (adr_type->isa_aryptr()) {
2379 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2380 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2381 if (elem_type != NULL) {
2382 sharpened_klass = elem_type->klass();
2383 }
2384 }
2385 }
2386
2387 // The sharpened class might be unloaded if there is no class loader
2388 // contraint in place.
2389 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2390 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2391
2392 #ifndef PRODUCT
2393 if (C->print_intrinsics() || C->print_inlining()) {
2394 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2395 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2396 }
2397 #endif
2398 // Sharpen the value type.
2399 return tjp;
2400 }
2401 return NULL;
2402 }
2403
2404 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2405 if (callee()->is_static()) return false; // caller must have the capability!
2406 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2407 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2408 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2409
2410 #ifndef PRODUCT
2411 {
2412 ResourceMark rm;
2413 // Check the signatures.
2414 ciSignature* sig = callee()->signature();
2415 #ifdef ASSERT
2416 if (!is_store) {
2417 // Object getObject(Object base, int/long offset), etc.
2418 BasicType rtype = sig->return_type()->basic_type();
2419 assert(rtype == type, "getter must return the expected value");
2420 assert(sig->count() == 2, "oop getter has 2 arguments");
2421 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2422 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2423 } else {
2424 // void putObject(Object base, int/long offset, Object x), etc.
2425 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2426 assert(sig->count() == 3, "oop putter has 3 arguments");
2427 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2428 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2429 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2430 assert(vtype == type, "putter must accept the expected value");
2431 }
2432 #endif // ASSERT
2433 }
2434 #endif //PRODUCT
2435
2436 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2437
2438 Node* receiver = argument(0); // type: oop
2439
2440 // Build address expression.
2441 Node* adr;
2442 Node* heap_base_oop = top();
2443 Node* offset = top();
2444 Node* val;
2445
2446 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2447 Node* base = argument(1); // type: oop
2448 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2449 offset = argument(2); // type: long
2450 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2451 // to be plain byte offsets, which are also the same as those accepted
2452 // by oopDesc::field_addr.
2453 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2454 "fieldOffset must be byte-scaled");
2455 // 32-bit machines ignore the high half!
2456 offset = ConvL2X(offset);
2457 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2458
2459 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2460 heap_base_oop = base;
2461 } else if (type == T_OBJECT) {
2462 return false; // off-heap oop accesses are not supported
2463 }
2464
2465 // Can base be NULL? Otherwise, always on-heap access.
2466 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2467
2468 val = is_store ? argument(4) : NULL;
2469
2470 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2471
2472 // Try to categorize the address.
2473 Compile::AliasType* alias_type = C->alias_type(adr_type);
2474 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2475
2476 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2477 alias_type->adr_type() == TypeAryPtr::RANGE) {
2478 return false; // not supported
2479 }
2480
2481 bool mismatched = false;
2482 BasicType bt = alias_type->basic_type();
2483 if (bt != T_ILLEGAL) {
2484 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2485 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2486 // Alias type doesn't differentiate between byte[] and boolean[]).
2487 // Use address type to get the element type.
2488 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2489 }
2490 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2491 // accessing an array field with getObject is not a mismatch
2492 bt = T_OBJECT;
2493 }
2494 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2495 // Don't intrinsify mismatched object accesses
2496 return false;
2497 }
2498 mismatched = (bt != type);
2499 } else if (alias_type->adr_type()->isa_oopptr()) {
2500 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2501 }
2502
2503 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2504
2505 // First guess at the value type.
2506 const Type *value_type = Type::get_const_basic_type(type);
2507
2508 // We will need memory barriers unless we can determine a unique
2509 // alias category for this reference. (Note: If for some reason
2510 // the barriers get omitted and the unsafe reference begins to "pollute"
2511 // the alias analysis of the rest of the graph, either Compile::can_alias
2512 // or Compile::must_alias will throw a diagnostic assert.)
2513 bool need_mem_bar = false;
2514 switch (kind) {
2515 case Relaxed:
2516 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2517 break;
2518 case Opaque:
2519 // Opaque uses CPUOrder membars for protection against code movement.
2520 case Acquire:
2521 case Release:
2522 case Volatile:
2523 need_mem_bar = true;
2524 break;
2525 default:
2526 ShouldNotReachHere();
2527 }
2528
2529 // Some accesses require access atomicity for all types, notably longs and doubles.
2530 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2531 bool requires_atomic_access = false;
2532 switch (kind) {
2533 case Relaxed:
2534 requires_atomic_access = AlwaysAtomicAccesses;
2535 break;
2536 case Opaque:
2537 // Opaque accesses are atomic.
2538 case Acquire:
2539 case Release:
2540 case Volatile:
2541 requires_atomic_access = true;
2542 break;
2543 default:
2544 ShouldNotReachHere();
2545 }
2546
2547 // Figure out the memory ordering.
2548 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2549 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2550
2551 // If we are reading the value of the referent field of a Reference
2552 // object (either by using Unsafe directly or through reflection)
2553 // then, if G1 is enabled, we need to record the referent in an
2554 // SATB log buffer using the pre-barrier mechanism.
2555 // Also we need to add memory barrier to prevent commoning reads
2556 // from this field across safepoint since GC can change its value.
2557 bool need_read_barrier = !is_store &&
2558 offset != top() && heap_base_oop != top();
2559
2560 if (!is_store && type == T_OBJECT) {
2561 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2562 if (tjp != NULL) {
2563 value_type = tjp;
2564 }
2565 }
2566
2567 receiver = null_check(receiver);
2568 if (stopped()) {
2569 return true;
2570 }
2571 // Heap pointers get a null-check from the interpreter,
2572 // as a courtesy. However, this is not guaranteed by Unsafe,
2573 // and it is not possible to fully distinguish unintended nulls
2574 // from intended ones in this API.
2575
2576 // We need to emit leading and trailing CPU membars (see below) in
2577 // addition to memory membars for special access modes. This is a little
2578 // too strong, but avoids the need to insert per-alias-type
2579 // volatile membars (for stores; compare Parse::do_put_xxx), which
2580 // we cannot do effectively here because we probably only have a
2581 // rough approximation of type.
2582
2583 switch(kind) {
2584 case Relaxed:
2585 case Opaque:
2586 case Acquire:
2587 break;
2588 case Release:
2589 case Volatile:
2590 if (is_store) {
2591 insert_mem_bar(Op_MemBarRelease);
2592 } else {
2593 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2594 insert_mem_bar(Op_MemBarVolatile);
2595 }
2596 }
2597 break;
2598 default:
2599 ShouldNotReachHere();
2600 }
2601
2602 // Memory barrier to prevent normal and 'unsafe' accesses from
2603 // bypassing each other. Happens after null checks, so the
2604 // exception paths do not take memory state from the memory barrier,
2605 // so there's no problems making a strong assert about mixing users
2606 // of safe & unsafe memory.
2607 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2608
2609 if (!is_store) {
2610 Node* p = NULL;
2611 // Try to constant fold a load from a constant field
2612 ciField* field = alias_type->field();
2613 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2614 // final or stable field
2615 p = make_constant_from_field(field, heap_base_oop);
2616 }
2617 if (p == NULL) {
2618 // To be valid, unsafe loads may depend on other conditions than
2619 // the one that guards them: pin the Load node
2620 LoadNode::ControlDependency dep = LoadNode::Pinned;
2621 Node* ctrl = control();
2622 if (adr_type->isa_instptr()) {
2623 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2624 intptr_t offset = Type::OffsetBot;
2625 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2626 if (offset >= 0) {
2627 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2628 if (offset < s) {
2629 // Guaranteed to be a valid access, no need to pin it
2630 dep = LoadNode::DependsOnlyOnTest;
2631 ctrl = NULL;
2632 }
2633 }
2634 }
2635 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2636 // load value
2637 switch (type) {
2638 case T_BOOLEAN:
2639 {
2640 // Normalize the value returned by getBoolean in the following cases
2641 if (mismatched ||
2642 heap_base_oop == top() || // - heap_base_oop is NULL or
2643 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2644 // and the unsafe access is made to large offset
2645 // (i.e., larger than the maximum offset necessary for any
2646 // field access)
2647 ) {
2648 IdealKit ideal = IdealKit(this);
2649 #define __ ideal.
2650 IdealVariable normalized_result(ideal);
2651 __ declarations_done();
2652 __ set(normalized_result, p);
2653 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2654 __ set(normalized_result, ideal.ConI(1));
2655 ideal.end_if();
2656 final_sync(ideal);
2657 p = __ value(normalized_result);
2658 #undef __
2659 }
2660 }
2661 case T_CHAR:
2662 case T_BYTE:
2663 case T_SHORT:
2664 case T_INT:
2665 case T_LONG:
2666 case T_FLOAT:
2667 case T_DOUBLE:
2668 break;
2669 case T_OBJECT:
2670 if (need_read_barrier) {
2671 // We do not require a mem bar inside pre_barrier if need_mem_bar
2672 // is set: the barriers would be emitted by us.
2673 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2674 }
2675 break;
2676 case T_ADDRESS:
2677 // Cast to an int type.
2678 p = _gvn.transform(new CastP2XNode(NULL, p));
2679 p = ConvX2UL(p);
2680 break;
2681 default:
2682 fatal("unexpected type %d: %s", type, type2name(type));
2683 break;
2684 }
2685 }
2686 // The load node has the control of the preceding MemBarCPUOrder. All
2687 // following nodes will have the control of the MemBarCPUOrder inserted at
2688 // the end of this method. So, pushing the load onto the stack at a later
2689 // point is fine.
2690 set_result(p);
2691 } else {
2692 // place effect of store into memory
2693 switch (type) {
2694 case T_DOUBLE:
2695 val = dstore_rounding(val);
2696 break;
2697 case T_ADDRESS:
2698 // Repackage the long as a pointer.
2699 val = ConvL2X(val);
2700 val = _gvn.transform(new CastX2PNode(val));
2701 break;
2702 default:
2703 break;
2704 }
2705
2706 if (type == T_OBJECT) {
2707 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2708 } else {
2709 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2710 }
2711 }
2712
2713 switch(kind) {
2714 case Relaxed:
2715 case Opaque:
2716 case Release:
2717 break;
2718 case Acquire:
2719 case Volatile:
2720 if (!is_store) {
2721 insert_mem_bar(Op_MemBarAcquire);
2722 } else {
2723 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2724 insert_mem_bar(Op_MemBarVolatile);
2725 }
2726 }
2727 break;
2728 default:
2729 ShouldNotReachHere();
2730 }
2731
2732 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2733
2734 return true;
2735 }
2736
2737 //----------------------------inline_unsafe_load_store----------------------------
2738 // This method serves a couple of different customers (depending on LoadStoreKind):
2739 //
2740 // LS_cmp_swap:
2741 //
2742 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2743 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2744 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2745 //
2746 // LS_cmp_swap_weak:
2747 //
2748 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2749 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2750 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2751 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2752 //
2753 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2754 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2755 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2756 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2757 //
2758 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2759 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2760 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2761 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2762 //
2763 // LS_cmp_exchange:
2764 //
2765 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2766 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2767 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2768 //
2769 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2770 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2771 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2772 //
2773 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2774 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2775 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2776 //
2777 // LS_get_add:
2778 //
2779 // int getAndAddInt( Object o, long offset, int delta)
2780 // long getAndAddLong(Object o, long offset, long delta)
2781 //
2782 // LS_get_set:
2783 //
2784 // int getAndSet(Object o, long offset, int newValue)
2785 // long getAndSet(Object o, long offset, long newValue)
2786 // Object getAndSet(Object o, long offset, Object newValue)
2787 //
2788 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2789 // This basic scheme here is the same as inline_unsafe_access, but
2790 // differs in enough details that combining them would make the code
2791 // overly confusing. (This is a true fact! I originally combined
2792 // them, but even I was confused by it!) As much code/comments as
2793 // possible are retained from inline_unsafe_access though to make
2794 // the correspondences clearer. - dl
2795
2796 if (callee()->is_static()) return false; // caller must have the capability!
2797
2798 #ifndef PRODUCT
2799 BasicType rtype;
2800 {
2801 ResourceMark rm;
2802 // Check the signatures.
2803 ciSignature* sig = callee()->signature();
2804 rtype = sig->return_type()->basic_type();
2805 switch(kind) {
2806 case LS_get_add:
2807 case LS_get_set: {
2808 // Check the signatures.
2809 #ifdef ASSERT
2810 assert(rtype == type, "get and set must return the expected type");
2811 assert(sig->count() == 3, "get and set has 3 arguments");
2812 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2813 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2814 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2815 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2816 #endif // ASSERT
2817 break;
2818 }
2819 case LS_cmp_swap:
2820 case LS_cmp_swap_weak: {
2821 // Check the signatures.
2822 #ifdef ASSERT
2823 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2824 assert(sig->count() == 4, "CAS has 4 arguments");
2825 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2826 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2827 #endif // ASSERT
2828 break;
2829 }
2830 case LS_cmp_exchange: {
2831 // Check the signatures.
2832 #ifdef ASSERT
2833 assert(rtype == type, "CAS must return the expected type");
2834 assert(sig->count() == 4, "CAS has 4 arguments");
2835 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2836 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2837 #endif // ASSERT
2838 break;
2839 }
2840 default:
2841 ShouldNotReachHere();
2842 }
2843 }
2844 #endif //PRODUCT
2845
2846 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2847
2848 // Get arguments:
2849 Node* receiver = NULL;
2850 Node* base = NULL;
2851 Node* offset = NULL;
2852 Node* oldval = NULL;
2853 Node* newval = NULL;
2854 switch(kind) {
2855 case LS_cmp_swap:
2856 case LS_cmp_swap_weak:
2857 case LS_cmp_exchange: {
2858 const bool two_slot_type = type2size[type] == 2;
2859 receiver = argument(0); // type: oop
2860 base = argument(1); // type: oop
2861 offset = argument(2); // type: long
2862 oldval = argument(4); // type: oop, int, or long
2863 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2864 break;
2865 }
2866 case LS_get_add:
2867 case LS_get_set: {
2868 receiver = argument(0); // type: oop
2869 base = argument(1); // type: oop
2870 offset = argument(2); // type: long
2871 oldval = NULL;
2872 newval = argument(4); // type: oop, int, or long
2873 break;
2874 }
2875 default:
2876 ShouldNotReachHere();
2877 }
2878
2879 // Build field offset expression.
2880 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2881 // to be plain byte offsets, which are also the same as those accepted
2882 // by oopDesc::field_addr.
2883 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2884 // 32-bit machines ignore the high half of long offsets
2885 offset = ConvL2X(offset);
2886 Node* adr = make_unsafe_address(base, offset, type, false);
2887 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2888
2889 Compile::AliasType* alias_type = C->alias_type(adr_type);
2890 BasicType bt = alias_type->basic_type();
2891 if (bt != T_ILLEGAL &&
2892 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2893 // Don't intrinsify mismatched object accesses.
2894 return false;
2895 }
2896
2897 // For CAS, unlike inline_unsafe_access, there seems no point in
2898 // trying to refine types. Just use the coarse types here.
2899 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2900 const Type *value_type = Type::get_const_basic_type(type);
2901
2902 switch (kind) {
2903 case LS_get_set:
2904 case LS_cmp_exchange: {
2905 if (type == T_OBJECT) {
2906 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2907 if (tjp != NULL) {
2908 value_type = tjp;
2909 }
2910 }
2911 break;
2912 }
2913 case LS_cmp_swap:
2914 case LS_cmp_swap_weak:
2915 case LS_get_add:
2916 break;
2917 default:
2918 ShouldNotReachHere();
2919 }
2920
2921 // Null check receiver.
2922 receiver = null_check(receiver);
2923 if (stopped()) {
2924 return true;
2925 }
2926
2927 int alias_idx = C->get_alias_index(adr_type);
2928
2929 // Memory-model-wise, a LoadStore acts like a little synchronized
2930 // block, so needs barriers on each side. These don't translate
2931 // into actual barriers on most machines, but we still need rest of
2932 // compiler to respect ordering.
2933
2934 switch (access_kind) {
2935 case Relaxed:
2936 case Acquire:
2937 break;
2938 case Release:
2939 insert_mem_bar(Op_MemBarRelease);
2940 break;
2941 case Volatile:
2942 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2943 insert_mem_bar(Op_MemBarVolatile);
2944 } else {
2945 insert_mem_bar(Op_MemBarRelease);
2946 }
2947 break;
2948 default:
2949 ShouldNotReachHere();
2950 }
2951 insert_mem_bar(Op_MemBarCPUOrder);
2952
2953 // Figure out the memory ordering.
2954 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2955
2956 // 4984716: MemBars must be inserted before this
2957 // memory node in order to avoid a false
2958 // dependency which will confuse the scheduler.
2959 Node *mem = memory(alias_idx);
2960
2961 // For now, we handle only those cases that actually exist: ints,
2962 // longs, and Object. Adding others should be straightforward.
2963 Node* load_store = NULL;
2964 switch(type) {
2965 case T_BYTE:
2966 switch(kind) {
2967 case LS_get_add:
2968 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2969 break;
2970 case LS_get_set:
2971 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2972 break;
2973 case LS_cmp_swap_weak:
2974 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2975 break;
2976 case LS_cmp_swap:
2977 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2978 break;
2979 case LS_cmp_exchange:
2980 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2981 break;
2982 default:
2983 ShouldNotReachHere();
2984 }
2985 break;
2986 case T_SHORT:
2987 switch(kind) {
2988 case LS_get_add:
2989 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2990 break;
2991 case LS_get_set:
2992 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2993 break;
2994 case LS_cmp_swap_weak:
2995 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2996 break;
2997 case LS_cmp_swap:
2998 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2999 break;
3000 case LS_cmp_exchange:
3001 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
3002 break;
3003 default:
3004 ShouldNotReachHere();
3005 }
3006 break;
3007 case T_INT:
3008 switch(kind) {
3009 case LS_get_add:
3010 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
3011 break;
3012 case LS_get_set:
3013 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
3014 break;
3015 case LS_cmp_swap_weak:
3016 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
3017 break;
3018 case LS_cmp_swap:
3019 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
3020 break;
3021 case LS_cmp_exchange:
3022 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
3023 break;
3024 default:
3025 ShouldNotReachHere();
3026 }
3027 break;
3028 case T_LONG:
3029 switch(kind) {
3030 case LS_get_add:
3031 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
3032 break;
3033 case LS_get_set:
3034 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
3035 break;
3036 case LS_cmp_swap_weak:
3037 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3038 break;
3039 case LS_cmp_swap:
3040 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3041 break;
3042 case LS_cmp_exchange:
3043 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
3044 break;
3045 default:
3046 ShouldNotReachHere();
3047 }
3048 break;
3049 case T_OBJECT:
3050 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3051 // could be delayed during Parse (for example, in adjust_map_after_if()).
3052 // Execute transformation here to avoid barrier generation in such case.
3053 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3054 newval = _gvn.makecon(TypePtr::NULL_PTR);
3055
3056 // Reference stores need a store barrier.
3057 switch(kind) {
3058 case LS_get_set: {
3059 // If pre-barrier must execute before the oop store, old value will require do_load here.
3060 if (!can_move_pre_barrier()) {
3061 pre_barrier(true /* do_load*/,
3062 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3063 NULL /* pre_val*/,
3064 T_OBJECT);
3065 } // Else move pre_barrier to use load_store value, see below.
3066 break;
3067 }
3068 case LS_cmp_swap_weak:
3069 case LS_cmp_swap:
3070 case LS_cmp_exchange: {
3071 // Same as for newval above:
3072 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3073 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3074 }
3075 // The only known value which might get overwritten is oldval.
3076 pre_barrier(false /* do_load */,
3077 control(), NULL, NULL, max_juint, NULL, NULL,
3078 oldval /* pre_val */,
3079 T_OBJECT);
3080 break;
3081 }
3082 default:
3083 ShouldNotReachHere();
3084 }
3085
3086 #ifdef _LP64
3087 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3088 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3089
3090 switch(kind) {
3091 case LS_get_set:
3092 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3093 break;
3094 case LS_cmp_swap_weak: {
3095 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3096 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3097 break;
3098 }
3099 case LS_cmp_swap: {
3100 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3101 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3102 break;
3103 }
3104 case LS_cmp_exchange: {
3105 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3106 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3107 break;
3108 }
3109 default:
3110 ShouldNotReachHere();
3111 }
3112 } else
3113 #endif
3114 switch (kind) {
3115 case LS_get_set:
3116 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3117 break;
3118 case LS_cmp_swap_weak:
3119 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3120 break;
3121 case LS_cmp_swap:
3122 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3123 break;
3124 case LS_cmp_exchange:
3125 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3126 break;
3127 default:
3128 ShouldNotReachHere();
3129 }
3130
3131 // Emit the post barrier only when the actual store happened. This makes sense
3132 // to check only for LS_cmp_* that can fail to set the value.
3133 // LS_cmp_exchange does not produce any branches by default, so there is no
3134 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3135 // CompareAndExchange and move branches here, it would make sense to conditionalize
3136 // post_barriers for LS_cmp_exchange as well.
3137 //
3138 // CAS success path is marked more likely since we anticipate this is a performance
3139 // critical path, while CAS failure path can use the penalty for going through unlikely
3140 // path as backoff. Which is still better than doing a store barrier there.
3141 switch (kind) {
3142 case LS_get_set:
3143 case LS_cmp_exchange: {
3144 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3145 break;
3146 }
3147 case LS_cmp_swap_weak:
3148 case LS_cmp_swap: {
3149 IdealKit ideal(this);
3150 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3151 sync_kit(ideal);
3152 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3153 ideal.sync_kit(this);
3154 } ideal.end_if();
3155 final_sync(ideal);
3156 break;
3157 }
3158 default:
3159 ShouldNotReachHere();
3160 }
3161 break;
3162 default:
3163 fatal("unexpected type %d: %s", type, type2name(type));
3164 break;
3165 }
3166
3167 // SCMemProjNodes represent the memory state of a LoadStore. Their
3168 // main role is to prevent LoadStore nodes from being optimized away
3169 // when their results aren't used.
3170 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3171 set_memory(proj, alias_idx);
3172
3173 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3174 #ifdef _LP64
3175 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3176 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3177 }
3178 #endif
3179 if (can_move_pre_barrier() && kind == LS_get_set) {
3180 // Don't need to load pre_val. The old value is returned by load_store.
3181 // The pre_barrier can execute after the xchg as long as no safepoint
3182 // gets inserted between them.
3183 pre_barrier(false /* do_load */,
3184 control(), NULL, NULL, max_juint, NULL, NULL,
3185 load_store /* pre_val */,
3186 T_OBJECT);
3187 }
3188 }
3189
3190 // Add the trailing membar surrounding the access
3191 insert_mem_bar(Op_MemBarCPUOrder);
3192
3193 switch (access_kind) {
3194 case Relaxed:
3195 case Release:
3196 break; // do nothing
3197 case Acquire:
3198 case Volatile:
3199 insert_mem_bar(Op_MemBarAcquire);
3200 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3201 break;
3202 default:
3203 ShouldNotReachHere();
3204 }
3205
3206 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3207 set_result(load_store);
3208 return true;
3209 }
3210
3211 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3212 MemNode::MemOrd mo = MemNode::unset;
3213 switch(kind) {
3214 case Opaque:
3215 case Relaxed: mo = MemNode::unordered; break;
3216 case Acquire: mo = MemNode::acquire; break;
3217 case Release: mo = MemNode::release; break;
3218 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3219 default:
3220 ShouldNotReachHere();
3221 }
3222 guarantee(mo != MemNode::unset, "Should select memory ordering");
3223 return mo;
3224 }
3225
3226 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3227 MemNode::MemOrd mo = MemNode::unset;
3228 switch(kind) {
3229 case Opaque:
3230 case Relaxed: mo = MemNode::unordered; break;
3231 case Acquire: mo = MemNode::acquire; break;
3232 case Release: mo = MemNode::release; break;
3233 case Volatile: mo = MemNode::seqcst; break;
3234 default:
3235 ShouldNotReachHere();
3236 }
3237 guarantee(mo != MemNode::unset, "Should select memory ordering");
3238 return mo;
3239 }
3240
3241 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3242 // Regardless of form, don't allow previous ld/st to move down,
3243 // then issue acquire, release, or volatile mem_bar.
3244 insert_mem_bar(Op_MemBarCPUOrder);
3245 switch(id) {
3246 case vmIntrinsics::_loadFence:
3247 insert_mem_bar(Op_LoadFence);
3248 return true;
3249 case vmIntrinsics::_storeFence:
3250 insert_mem_bar(Op_StoreFence);
3251 return true;
3252 case vmIntrinsics::_fullFence:
3253 insert_mem_bar(Op_MemBarVolatile);
3254 return true;
3255 default:
3256 fatal_unexpected_iid(id);
3257 return false;
3258 }
3259 }
3260
3261 bool LibraryCallKit::inline_onspinwait() {
3262 insert_mem_bar(Op_OnSpinWait);
3263 return true;
3264 }
3265
3266 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3267 if (!kls->is_Con()) {
3268 return true;
3269 }
3270 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3271 if (klsptr == NULL) {
3272 return true;
3273 }
3274 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3275 // don't need a guard for a klass that is already initialized
3276 return !ik->is_initialized();
3277 }
3278
3279 //----------------------------inline_unsafe_allocate---------------------------
3280 // public native Object Unsafe.allocateInstance(Class<?> cls);
3281 bool LibraryCallKit::inline_unsafe_allocate() {
3282 if (callee()->is_static()) return false; // caller must have the capability!
3283
3284 null_check_receiver(); // null-check, then ignore
3285 Node* cls = null_check(argument(1));
3286 if (stopped()) return true;
3287
3288 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3289 kls = null_check(kls);
3290 if (stopped()) return true; // argument was like int.class
3291
3292 Node* test = NULL;
3293 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3294 // Note: The argument might still be an illegal value like
3295 // Serializable.class or Object[].class. The runtime will handle it.
3296 // But we must make an explicit check for initialization.
3297 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3298 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3299 // can generate code to load it as unsigned byte.
3300 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3301 Node* bits = intcon(InstanceKlass::fully_initialized);
3302 test = _gvn.transform(new SubINode(inst, bits));
3303 // The 'test' is non-zero if we need to take a slow path.
3304 }
3305
3306 Node* obj = new_instance(kls, test);
3307 set_result(obj);
3308 return true;
3309 }
3310
3311 //------------------------inline_native_time_funcs--------------
3312 // inline code for System.currentTimeMillis() and System.nanoTime()
3313 // these have the same type and signature
3314 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3315 const TypeFunc* tf = OptoRuntime::void_long_Type();
3316 const TypePtr* no_memory_effects = NULL;
3317 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3318 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3319 #ifdef ASSERT
3320 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3321 assert(value_top == top(), "second value must be top");
3322 #endif
3323 set_result(value);
3324 return true;
3325 }
3326
3327 #ifdef TRACE_HAVE_INTRINSICS
3328
3329 /*
3330 * oop -> myklass
3331 * myklass->trace_id |= USED
3332 * return myklass->trace_id & ~0x3
3333 */
3334 bool LibraryCallKit::inline_native_classID() {
3335 Node* cls = null_check(argument(0), T_OBJECT);
3336 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3337 kls = null_check(kls, T_OBJECT);
3338
3339 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3340 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3341 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3342
3343 Node* clsused = longcon(0x01l); // set the class bit
3344 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3345 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3346 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3347
3348 #ifdef TRACE_ID_META_BITS
3349 Node* mbits = longcon(~TRACE_ID_META_BITS);
3350 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3351 #endif
3352 #ifdef TRACE_ID_CLASS_SHIFT
3353 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3354 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3355 #endif
3356
3357 set_result(tvalue);
3358 return true;
3359
3360 }
3361
3362 bool LibraryCallKit::inline_native_getBufferWriter() {
3363 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3364
3365 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3366 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3367 );
3368
3369 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3370
3371 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3372 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3373
3374 IfNode* iff_jobj_null =
3375 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3376
3377 enum { _normal_path = 1,
3378 _null_path = 2,
3379 PATH_LIMIT };
3380
3381 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3382 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3383
3384 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3385 result_rgn->init_req(_null_path, jobj_is_null);
3386 result_val->init_req(_null_path, null());
3387
3388 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3389 result_rgn->init_req(_normal_path, jobj_is_not_null);
3390
3391 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3392 result_val->init_req(_normal_path, res);
3393
3394 set_result(result_rgn, result_val);
3395
3396 return true;
3397 }
3398
3399 #endif
3400
3401 //------------------------inline_native_currentThread------------------
3402 bool LibraryCallKit::inline_native_currentThread() {
3403 Node* junk = NULL;
3404 set_result(generate_current_thread(junk));
3405 return true;
3406 }
3407
3408 //------------------------inline_native_isInterrupted------------------
3409 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3410 bool LibraryCallKit::inline_native_isInterrupted() {
3411 // Add a fast path to t.isInterrupted(clear_int):
3412 // (t == Thread.current() &&
3413 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3414 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3415 // So, in the common case that the interrupt bit is false,
3416 // we avoid making a call into the VM. Even if the interrupt bit
3417 // is true, if the clear_int argument is false, we avoid the VM call.
3418 // However, if the receiver is not currentThread, we must call the VM,
3419 // because there must be some locking done around the operation.
3420
3421 // We only go to the fast case code if we pass two guards.
3422 // Paths which do not pass are accumulated in the slow_region.
3423
3424 enum {
3425 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3426 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3427 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3428 PATH_LIMIT
3429 };
3430
3431 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3432 // out of the function.
3433 insert_mem_bar(Op_MemBarCPUOrder);
3434
3435 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3436 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3437
3438 RegionNode* slow_region = new RegionNode(1);
3439 record_for_igvn(slow_region);
3440
3441 // (a) Receiving thread must be the current thread.
3442 Node* rec_thr = argument(0);
3443 Node* tls_ptr = NULL;
3444 Node* cur_thr = generate_current_thread(tls_ptr);
3445 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3446 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3447
3448 generate_slow_guard(bol_thr, slow_region);
3449
3450 // (b) Interrupt bit on TLS must be false.
3451 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3452 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3453 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3454
3455 // Set the control input on the field _interrupted read to prevent it floating up.
3456 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3457 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3458 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3459
3460 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3461
3462 // First fast path: if (!TLS._interrupted) return false;
3463 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3464 result_rgn->init_req(no_int_result_path, false_bit);
3465 result_val->init_req(no_int_result_path, intcon(0));
3466
3467 // drop through to next case
3468 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3469
3470 #ifndef _WINDOWS
3471 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3472 Node* clr_arg = argument(1);
3473 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3474 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3475 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3476
3477 // Second fast path: ... else if (!clear_int) return true;
3478 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3479 result_rgn->init_req(no_clear_result_path, false_arg);
3480 result_val->init_req(no_clear_result_path, intcon(1));
3481
3482 // drop through to next case
3483 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3484 #else
3485 // To return true on Windows you must read the _interrupted field
3486 // and check the event state i.e. take the slow path.
3487 #endif // _WINDOWS
3488
3489 // (d) Otherwise, go to the slow path.
3490 slow_region->add_req(control());
3491 set_control( _gvn.transform(slow_region));
3492
3493 if (stopped()) {
3494 // There is no slow path.
3495 result_rgn->init_req(slow_result_path, top());
3496 result_val->init_req(slow_result_path, top());
3497 } else {
3498 // non-virtual because it is a private non-static
3499 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3500
3501 Node* slow_val = set_results_for_java_call(slow_call);
3502 // this->control() comes from set_results_for_java_call
3503
3504 Node* fast_io = slow_call->in(TypeFunc::I_O);
3505 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3506
3507 // These two phis are pre-filled with copies of of the fast IO and Memory
3508 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3509 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3510
3511 result_rgn->init_req(slow_result_path, control());
3512 result_io ->init_req(slow_result_path, i_o());
3513 result_mem->init_req(slow_result_path, reset_memory());
3514 result_val->init_req(slow_result_path, slow_val);
3515
3516 set_all_memory(_gvn.transform(result_mem));
3517 set_i_o( _gvn.transform(result_io));
3518 }
3519
3520 C->set_has_split_ifs(true); // Has chance for split-if optimization
3521 set_result(result_rgn, result_val);
3522 return true;
3523 }
3524
3525 //---------------------------load_mirror_from_klass----------------------------
3526 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3527 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3528 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3529 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3530 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3531 }
3532
3533 //-----------------------load_klass_from_mirror_common-------------------------
3534 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3535 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3536 // and branch to the given path on the region.
3537 // If never_see_null, take an uncommon trap on null, so we can optimistically
3538 // compile for the non-null case.
3539 // If the region is NULL, force never_see_null = true.
3540 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3541 bool never_see_null,
3542 RegionNode* region,
3543 int null_path,
3544 int offset) {
3545 if (region == NULL) never_see_null = true;
3546 Node* p = basic_plus_adr(mirror, offset);
3547 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3548 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3549 Node* null_ctl = top();
3550 kls = null_check_oop(kls, &null_ctl, never_see_null);
3551 if (region != NULL) {
3552 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3553 region->init_req(null_path, null_ctl);
3554 } else {
3555 assert(null_ctl == top(), "no loose ends");
3556 }
3557 return kls;
3558 }
3559
3560 //--------------------(inline_native_Class_query helpers)---------------------
3561 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3562 // Fall through if (mods & mask) == bits, take the guard otherwise.
3563 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3564 // Branch around if the given klass has the given modifier bit set.
3565 // Like generate_guard, adds a new path onto the region.
3566 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3567 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3568 Node* mask = intcon(modifier_mask);
3569 Node* bits = intcon(modifier_bits);
3570 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3571 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3572 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3573 return generate_fair_guard(bol, region);
3574 }
3575 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3576 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3577 }
3578
3579 //-------------------------inline_native_Class_query-------------------
3580 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3581 const Type* return_type = TypeInt::BOOL;
3582 Node* prim_return_value = top(); // what happens if it's a primitive class?
3583 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3584 bool expect_prim = false; // most of these guys expect to work on refs
3585
3586 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3587
3588 Node* mirror = argument(0);
3589 Node* obj = top();
3590
3591 switch (id) {
3592 case vmIntrinsics::_isInstance:
3593 // nothing is an instance of a primitive type
3594 prim_return_value = intcon(0);
3595 obj = argument(1);
3596 break;
3597 case vmIntrinsics::_getModifiers:
3598 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3599 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3600 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3601 break;
3602 case vmIntrinsics::_isInterface:
3603 prim_return_value = intcon(0);
3604 break;
3605 case vmIntrinsics::_isArray:
3606 prim_return_value = intcon(0);
3607 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3608 break;
3609 case vmIntrinsics::_isPrimitive:
3610 prim_return_value = intcon(1);
3611 expect_prim = true; // obviously
3612 break;
3613 case vmIntrinsics::_getSuperclass:
3614 prim_return_value = null();
3615 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3616 break;
3617 case vmIntrinsics::_getClassAccessFlags:
3618 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3619 return_type = TypeInt::INT; // not bool! 6297094
3620 break;
3621 default:
3622 fatal_unexpected_iid(id);
3623 break;
3624 }
3625
3626 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3627 if (mirror_con == NULL) return false; // cannot happen?
3628
3629 #ifndef PRODUCT
3630 if (C->print_intrinsics() || C->print_inlining()) {
3631 ciType* k = mirror_con->java_mirror_type();
3632 if (k) {
3633 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3634 k->print_name();
3635 tty->cr();
3636 }
3637 }
3638 #endif
3639
3640 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3641 RegionNode* region = new RegionNode(PATH_LIMIT);
3642 record_for_igvn(region);
3643 PhiNode* phi = new PhiNode(region, return_type);
3644
3645 // The mirror will never be null of Reflection.getClassAccessFlags, however
3646 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3647 // if it is. See bug 4774291.
3648
3649 // For Reflection.getClassAccessFlags(), the null check occurs in
3650 // the wrong place; see inline_unsafe_access(), above, for a similar
3651 // situation.
3652 mirror = null_check(mirror);
3653 // If mirror or obj is dead, only null-path is taken.
3654 if (stopped()) return true;
3655
3656 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3657
3658 // Now load the mirror's klass metaobject, and null-check it.
3659 // Side-effects region with the control path if the klass is null.
3660 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3661 // If kls is null, we have a primitive mirror.
3662 phi->init_req(_prim_path, prim_return_value);
3663 if (stopped()) { set_result(region, phi); return true; }
3664 bool safe_for_replace = (region->in(_prim_path) == top());
3665
3666 Node* p; // handy temp
3667 Node* null_ctl;
3668
3669 // Now that we have the non-null klass, we can perform the real query.
3670 // For constant classes, the query will constant-fold in LoadNode::Value.
3671 Node* query_value = top();
3672 switch (id) {
3673 case vmIntrinsics::_isInstance:
3674 // nothing is an instance of a primitive type
3675 query_value = gen_instanceof(obj, kls, safe_for_replace);
3676 break;
3677
3678 case vmIntrinsics::_getModifiers:
3679 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3680 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3681 break;
3682
3683 case vmIntrinsics::_isInterface:
3684 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3685 if (generate_interface_guard(kls, region) != NULL)
3686 // A guard was added. If the guard is taken, it was an interface.
3687 phi->add_req(intcon(1));
3688 // If we fall through, it's a plain class.
3689 query_value = intcon(0);
3690 break;
3691
3692 case vmIntrinsics::_isArray:
3693 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3694 if (generate_array_guard(kls, region) != NULL)
3695 // A guard was added. If the guard is taken, it was an array.
3696 phi->add_req(intcon(1));
3697 // If we fall through, it's a plain class.
3698 query_value = intcon(0);
3699 break;
3700
3701 case vmIntrinsics::_isPrimitive:
3702 query_value = intcon(0); // "normal" path produces false
3703 break;
3704
3705 case vmIntrinsics::_getSuperclass:
3706 // The rules here are somewhat unfortunate, but we can still do better
3707 // with random logic than with a JNI call.
3708 // Interfaces store null or Object as _super, but must report null.
3709 // Arrays store an intermediate super as _super, but must report Object.
3710 // Other types can report the actual _super.
3711 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3712 if (generate_interface_guard(kls, region) != NULL)
3713 // A guard was added. If the guard is taken, it was an interface.
3714 phi->add_req(null());
3715 if (generate_array_guard(kls, region) != NULL)
3716 // A guard was added. If the guard is taken, it was an array.
3717 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3718 // If we fall through, it's a plain class. Get its _super.
3719 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3720 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3721 null_ctl = top();
3722 kls = null_check_oop(kls, &null_ctl);
3723 if (null_ctl != top()) {
3724 // If the guard is taken, Object.superClass is null (both klass and mirror).
3725 region->add_req(null_ctl);
3726 phi ->add_req(null());
3727 }
3728 if (!stopped()) {
3729 query_value = load_mirror_from_klass(kls);
3730 }
3731 break;
3732
3733 case vmIntrinsics::_getClassAccessFlags:
3734 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3735 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3736 break;
3737
3738 default:
3739 fatal_unexpected_iid(id);
3740 break;
3741 }
3742
3743 // Fall-through is the normal case of a query to a real class.
3744 phi->init_req(1, query_value);
3745 region->init_req(1, control());
3746
3747 C->set_has_split_ifs(true); // Has chance for split-if optimization
3748 set_result(region, phi);
3749 return true;
3750 }
3751
3752 //-------------------------inline_Class_cast-------------------
3753 bool LibraryCallKit::inline_Class_cast() {
3754 Node* mirror = argument(0); // Class
3755 Node* obj = argument(1);
3756 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3757 if (mirror_con == NULL) {
3758 return false; // dead path (mirror->is_top()).
3759 }
3760 if (obj == NULL || obj->is_top()) {
3761 return false; // dead path
3762 }
3763 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3764
3765 // First, see if Class.cast() can be folded statically.
3766 // java_mirror_type() returns non-null for compile-time Class constants.
3767 ciType* tm = mirror_con->java_mirror_type();
3768 if (tm != NULL && tm->is_klass() &&
3769 tp != NULL && tp->klass() != NULL) {
3770 if (!tp->klass()->is_loaded()) {
3771 // Don't use intrinsic when class is not loaded.
3772 return false;
3773 } else {
3774 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3775 if (static_res == Compile::SSC_always_true) {
3776 // isInstance() is true - fold the code.
3777 set_result(obj);
3778 return true;
3779 } else if (static_res == Compile::SSC_always_false) {
3780 // Don't use intrinsic, have to throw ClassCastException.
3781 // If the reference is null, the non-intrinsic bytecode will
3782 // be optimized appropriately.
3783 return false;
3784 }
3785 }
3786 }
3787
3788 // Bailout intrinsic and do normal inlining if exception path is frequent.
3789 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3790 return false;
3791 }
3792
3793 // Generate dynamic checks.
3794 // Class.cast() is java implementation of _checkcast bytecode.
3795 // Do checkcast (Parse::do_checkcast()) optimizations here.
3796
3797 mirror = null_check(mirror);
3798 // If mirror is dead, only null-path is taken.
3799 if (stopped()) {
3800 return true;
3801 }
3802
3803 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3804 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3805 RegionNode* region = new RegionNode(PATH_LIMIT);
3806 record_for_igvn(region);
3807
3808 // Now load the mirror's klass metaobject, and null-check it.
3809 // If kls is null, we have a primitive mirror and
3810 // nothing is an instance of a primitive type.
3811 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3812
3813 Node* res = top();
3814 if (!stopped()) {
3815 Node* bad_type_ctrl = top();
3816 // Do checkcast optimizations.
3817 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3818 region->init_req(_bad_type_path, bad_type_ctrl);
3819 }
3820 if (region->in(_prim_path) != top() ||
3821 region->in(_bad_type_path) != top()) {
3822 // Let Interpreter throw ClassCastException.
3823 PreserveJVMState pjvms(this);
3824 set_control(_gvn.transform(region));
3825 uncommon_trap(Deoptimization::Reason_intrinsic,
3826 Deoptimization::Action_maybe_recompile);
3827 }
3828 if (!stopped()) {
3829 set_result(res);
3830 }
3831 return true;
3832 }
3833
3834
3835 //--------------------------inline_native_subtype_check------------------------
3836 // This intrinsic takes the JNI calls out of the heart of
3837 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3838 bool LibraryCallKit::inline_native_subtype_check() {
3839 // Pull both arguments off the stack.
3840 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3841 args[0] = argument(0);
3842 args[1] = argument(1);
3843 Node* klasses[2]; // corresponding Klasses: superk, subk
3844 klasses[0] = klasses[1] = top();
3845
3846 enum {
3847 // A full decision tree on {superc is prim, subc is prim}:
3848 _prim_0_path = 1, // {P,N} => false
3849 // {P,P} & superc!=subc => false
3850 _prim_same_path, // {P,P} & superc==subc => true
3851 _prim_1_path, // {N,P} => false
3852 _ref_subtype_path, // {N,N} & subtype check wins => true
3853 _both_ref_path, // {N,N} & subtype check loses => false
3854 PATH_LIMIT
3855 };
3856
3857 RegionNode* region = new RegionNode(PATH_LIMIT);
3858 Node* phi = new PhiNode(region, TypeInt::BOOL);
3859 record_for_igvn(region);
3860
3861 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3862 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3863 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3864
3865 // First null-check both mirrors and load each mirror's klass metaobject.
3866 int which_arg;
3867 for (which_arg = 0; which_arg <= 1; which_arg++) {
3868 Node* arg = args[which_arg];
3869 arg = null_check(arg);
3870 if (stopped()) break;
3871 args[which_arg] = arg;
3872
3873 Node* p = basic_plus_adr(arg, class_klass_offset);
3874 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3875 klasses[which_arg] = _gvn.transform(kls);
3876 }
3877
3878 // Having loaded both klasses, test each for null.
3879 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3880 for (which_arg = 0; which_arg <= 1; which_arg++) {
3881 Node* kls = klasses[which_arg];
3882 Node* null_ctl = top();
3883 kls = null_check_oop(kls, &null_ctl, never_see_null);
3884 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3885 region->init_req(prim_path, null_ctl);
3886 if (stopped()) break;
3887 klasses[which_arg] = kls;
3888 }
3889
3890 if (!stopped()) {
3891 // now we have two reference types, in klasses[0..1]
3892 Node* subk = klasses[1]; // the argument to isAssignableFrom
3893 Node* superk = klasses[0]; // the receiver
3894 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3895 // now we have a successful reference subtype check
3896 region->set_req(_ref_subtype_path, control());
3897 }
3898
3899 // If both operands are primitive (both klasses null), then
3900 // we must return true when they are identical primitives.
3901 // It is convenient to test this after the first null klass check.
3902 set_control(region->in(_prim_0_path)); // go back to first null check
3903 if (!stopped()) {
3904 // Since superc is primitive, make a guard for the superc==subc case.
3905 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3906 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3907 generate_guard(bol_eq, region, PROB_FAIR);
3908 if (region->req() == PATH_LIMIT+1) {
3909 // A guard was added. If the added guard is taken, superc==subc.
3910 region->swap_edges(PATH_LIMIT, _prim_same_path);
3911 region->del_req(PATH_LIMIT);
3912 }
3913 region->set_req(_prim_0_path, control()); // Not equal after all.
3914 }
3915
3916 // these are the only paths that produce 'true':
3917 phi->set_req(_prim_same_path, intcon(1));
3918 phi->set_req(_ref_subtype_path, intcon(1));
3919
3920 // pull together the cases:
3921 assert(region->req() == PATH_LIMIT, "sane region");
3922 for (uint i = 1; i < region->req(); i++) {
3923 Node* ctl = region->in(i);
3924 if (ctl == NULL || ctl == top()) {
3925 region->set_req(i, top());
3926 phi ->set_req(i, top());
3927 } else if (phi->in(i) == NULL) {
3928 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3929 }
3930 }
3931
3932 set_control(_gvn.transform(region));
3933 set_result(_gvn.transform(phi));
3934 return true;
3935 }
3936
3937 //---------------------generate_array_guard_common------------------------
3938 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3939 bool obj_array, bool not_array) {
3940
3941 if (stopped()) {
3942 return NULL;
3943 }
3944
3945 // If obj_array/non_array==false/false:
3946 // Branch around if the given klass is in fact an array (either obj or prim).
3947 // If obj_array/non_array==false/true:
3948 // Branch around if the given klass is not an array klass of any kind.
3949 // If obj_array/non_array==true/true:
3950 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3951 // If obj_array/non_array==true/false:
3952 // Branch around if the kls is an oop array (Object[] or subtype)
3953 //
3954 // Like generate_guard, adds a new path onto the region.
3955 jint layout_con = 0;
3956 Node* layout_val = get_layout_helper(kls, layout_con);
3957 if (layout_val == NULL) {
3958 bool query = (obj_array
3959 ? Klass::layout_helper_is_objArray(layout_con)
3960 : Klass::layout_helper_is_array(layout_con));
3961 if (query == not_array) {
3962 return NULL; // never a branch
3963 } else { // always a branch
3964 Node* always_branch = control();
3965 if (region != NULL)
3966 region->add_req(always_branch);
3967 set_control(top());
3968 return always_branch;
3969 }
3970 }
3971 // Now test the correct condition.
3972 jint nval = (obj_array
3973 ? (jint)(Klass::_lh_array_tag_type_value
3974 << Klass::_lh_array_tag_shift)
3975 : Klass::_lh_neutral_value);
3976 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3977 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3978 // invert the test if we are looking for a non-array
3979 if (not_array) btest = BoolTest(btest).negate();
3980 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3981 return generate_fair_guard(bol, region);
3982 }
3983
3984
3985 //-----------------------inline_native_newArray--------------------------
3986 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3987 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3988 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3989 Node* mirror;
3990 Node* count_val;
3991 if (uninitialized) {
3992 mirror = argument(1);
3993 count_val = argument(2);
3994 } else {
3995 mirror = argument(0);
3996 count_val = argument(1);
3997 }
3998
3999 mirror = null_check(mirror);
4000 // If mirror or obj is dead, only null-path is taken.
4001 if (stopped()) return true;
4002
4003 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
4004 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4005 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4006 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4007 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4008
4009 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
4010 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
4011 result_reg, _slow_path);
4012 Node* normal_ctl = control();
4013 Node* no_array_ctl = result_reg->in(_slow_path);
4014
4015 // Generate code for the slow case. We make a call to newArray().
4016 set_control(no_array_ctl);
4017 if (!stopped()) {
4018 // Either the input type is void.class, or else the
4019 // array klass has not yet been cached. Either the
4020 // ensuing call will throw an exception, or else it
4021 // will cache the array klass for next time.
4022 PreserveJVMState pjvms(this);
4023 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
4024 Node* slow_result = set_results_for_java_call(slow_call);
4025 // this->control() comes from set_results_for_java_call
4026 result_reg->set_req(_slow_path, control());
4027 result_val->set_req(_slow_path, slow_result);
4028 result_io ->set_req(_slow_path, i_o());
4029 result_mem->set_req(_slow_path, reset_memory());
4030 }
4031
4032 set_control(normal_ctl);
4033 if (!stopped()) {
4034 // Normal case: The array type has been cached in the java.lang.Class.
4035 // The following call works fine even if the array type is polymorphic.
4036 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4037 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
4038 result_reg->init_req(_normal_path, control());
4039 result_val->init_req(_normal_path, obj);
4040 result_io ->init_req(_normal_path, i_o());
4041 result_mem->init_req(_normal_path, reset_memory());
4042
4043 if (uninitialized) {
4044 // Mark the allocation so that zeroing is skipped
4045 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4046 alloc->maybe_set_complete(&_gvn);
4047 }
4048 }
4049
4050 // Return the combined state.
4051 set_i_o( _gvn.transform(result_io) );
4052 set_all_memory( _gvn.transform(result_mem));
4053
4054 C->set_has_split_ifs(true); // Has chance for split-if optimization
4055 set_result(result_reg, result_val);
4056 return true;
4057 }
4058
4059 //----------------------inline_native_getLength--------------------------
4060 // public static native int java.lang.reflect.Array.getLength(Object array);
4061 bool LibraryCallKit::inline_native_getLength() {
4062 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4063
4064 Node* array = null_check(argument(0));
4065 // If array is dead, only null-path is taken.
4066 if (stopped()) return true;
4067
4068 // Deoptimize if it is a non-array.
4069 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4070
4071 if (non_array != NULL) {
4072 PreserveJVMState pjvms(this);
4073 set_control(non_array);
4074 uncommon_trap(Deoptimization::Reason_intrinsic,
4075 Deoptimization::Action_maybe_recompile);
4076 }
4077
4078 // If control is dead, only non-array-path is taken.
4079 if (stopped()) return true;
4080
4081 // The works fine even if the array type is polymorphic.
4082 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4083 Node* result = load_array_length(array);
4084
4085 C->set_has_split_ifs(true); // Has chance for split-if optimization
4086 set_result(result);
4087 return true;
4088 }
4089
4090 //------------------------inline_array_copyOf----------------------------
4091 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4092 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4093 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4094 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4095
4096 // Get the arguments.
4097 Node* original = argument(0);
4098 Node* start = is_copyOfRange? argument(1): intcon(0);
4099 Node* end = is_copyOfRange? argument(2): argument(1);
4100 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4101
4102 Node* newcopy = NULL;
4103
4104 // Set the original stack and the reexecute bit for the interpreter to reexecute
4105 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4106 { PreserveReexecuteState preexecs(this);
4107 jvms()->set_should_reexecute(true);
4108
4109 array_type_mirror = null_check(array_type_mirror);
4110 original = null_check(original);
4111
4112 // Check if a null path was taken unconditionally.
4113 if (stopped()) return true;
4114
4115 Node* orig_length = load_array_length(original);
4116
4117 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4118 klass_node = null_check(klass_node);
4119
4120 RegionNode* bailout = new RegionNode(1);
4121 record_for_igvn(bailout);
4122
4123 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4124 // Bail out if that is so.
4125 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4126 if (not_objArray != NULL) {
4127 // Improve the klass node's type from the new optimistic assumption:
4128 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4129 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4130 Node* cast = new CastPPNode(klass_node, akls);
4131 cast->init_req(0, control());
4132 klass_node = _gvn.transform(cast);
4133 }
4134
4135 // Bail out if either start or end is negative.
4136 generate_negative_guard(start, bailout, &start);
4137 generate_negative_guard(end, bailout, &end);
4138
4139 Node* length = end;
4140 if (_gvn.type(start) != TypeInt::ZERO) {
4141 length = _gvn.transform(new SubINode(end, start));
4142 }
4143
4144 // Bail out if length is negative.
4145 // Without this the new_array would throw
4146 // NegativeArraySizeException but IllegalArgumentException is what
4147 // should be thrown
4148 generate_negative_guard(length, bailout, &length);
4149
4150 if (bailout->req() > 1) {
4151 PreserveJVMState pjvms(this);
4152 set_control(_gvn.transform(bailout));
4153 uncommon_trap(Deoptimization::Reason_intrinsic,
4154 Deoptimization::Action_maybe_recompile);
4155 }
4156
4157 if (!stopped()) {
4158 // How many elements will we copy from the original?
4159 // The answer is MinI(orig_length - start, length).
4160 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4161 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4162
4163 // Generate a direct call to the right arraycopy function(s).
4164 // We know the copy is disjoint but we might not know if the
4165 // oop stores need checking.
4166 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4167 // This will fail a store-check if x contains any non-nulls.
4168
4169 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4170 // loads/stores but it is legal only if we're sure the
4171 // Arrays.copyOf would succeed. So we need all input arguments
4172 // to the copyOf to be validated, including that the copy to the
4173 // new array won't trigger an ArrayStoreException. That subtype
4174 // check can be optimized if we know something on the type of
4175 // the input array from type speculation.
4176 if (_gvn.type(klass_node)->singleton()) {
4177 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4178 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4179
4180 int test = C->static_subtype_check(superk, subk);
4181 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4182 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4183 if (t_original->speculative_type() != NULL) {
4184 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4185 }
4186 }
4187 }
4188
4189 bool validated = false;
4190 // Reason_class_check rather than Reason_intrinsic because we
4191 // want to intrinsify even if this traps.
4192 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4193 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4194 klass_node);
4195
4196 if (not_subtype_ctrl != top()) {
4197 PreserveJVMState pjvms(this);
4198 set_control(not_subtype_ctrl);
4199 uncommon_trap(Deoptimization::Reason_class_check,
4200 Deoptimization::Action_make_not_entrant);
4201 assert(stopped(), "Should be stopped");
4202 }
4203 validated = true;
4204 }
4205
4206 if (!stopped()) {
4207 newcopy = new_array(klass_node, length, 0); // no arguments to push
4208
4209 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4210 load_object_klass(original), klass_node);
4211 if (!is_copyOfRange) {
4212 ac->set_copyof(validated);
4213 } else {
4214 ac->set_copyofrange(validated);
4215 }
4216 Node* n = _gvn.transform(ac);
4217 if (n == ac) {
4218 ac->connect_outputs(this);
4219 } else {
4220 assert(validated, "shouldn't transform if all arguments not validated");
4221 set_all_memory(n);
4222 }
4223 }
4224 }
4225 } // original reexecute is set back here
4226
4227 C->set_has_split_ifs(true); // Has chance for split-if optimization
4228 if (!stopped()) {
4229 set_result(newcopy);
4230 }
4231 return true;
4232 }
4233
4234
4235 //----------------------generate_virtual_guard---------------------------
4236 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4237 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4238 RegionNode* slow_region) {
4239 ciMethod* method = callee();
4240 int vtable_index = method->vtable_index();
4241 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4242 "bad index %d", vtable_index);
4243 // Get the Method* out of the appropriate vtable entry.
4244 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4245 vtable_index*vtableEntry::size_in_bytes() +
4246 vtableEntry::method_offset_in_bytes();
4247 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4248 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4249
4250 // Compare the target method with the expected method (e.g., Object.hashCode).
4251 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4252
4253 Node* native_call = makecon(native_call_addr);
4254 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4255 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4256
4257 return generate_slow_guard(test_native, slow_region);
4258 }
4259
4260 //-----------------------generate_method_call----------------------------
4261 // Use generate_method_call to make a slow-call to the real
4262 // method if the fast path fails. An alternative would be to
4263 // use a stub like OptoRuntime::slow_arraycopy_Java.
4264 // This only works for expanding the current library call,
4265 // not another intrinsic. (E.g., don't use this for making an
4266 // arraycopy call inside of the copyOf intrinsic.)
4267 CallJavaNode*
4268 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4269 // When compiling the intrinsic method itself, do not use this technique.
4270 guarantee(callee() != C->method(), "cannot make slow-call to self");
4271
4272 ciMethod* method = callee();
4273 // ensure the JVMS we have will be correct for this call
4274 guarantee(method_id == method->intrinsic_id(), "must match");
4275
4276 const TypeFunc* tf = TypeFunc::make(method);
4277 CallJavaNode* slow_call;
4278 if (is_static) {
4279 assert(!is_virtual, "");
4280 slow_call = new CallStaticJavaNode(C, tf,
4281 SharedRuntime::get_resolve_static_call_stub(),
4282 method, bci());
4283 } else if (is_virtual) {
4284 null_check_receiver();
4285 int vtable_index = Method::invalid_vtable_index;
4286 if (UseInlineCaches) {
4287 // Suppress the vtable call
4288 } else {
4289 // hashCode and clone are not a miranda methods,
4290 // so the vtable index is fixed.
4291 // No need to use the linkResolver to get it.
4292 vtable_index = method->vtable_index();
4293 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4294 "bad index %d", vtable_index);
4295 }
4296 slow_call = new CallDynamicJavaNode(tf,
4297 SharedRuntime::get_resolve_virtual_call_stub(),
4298 method, vtable_index, bci());
4299 } else { // neither virtual nor static: opt_virtual
4300 null_check_receiver();
4301 slow_call = new CallStaticJavaNode(C, tf,
4302 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4303 method, bci());
4304 slow_call->set_optimized_virtual(true);
4305 }
4306 set_arguments_for_java_call(slow_call);
4307 set_edges_for_java_call(slow_call);
4308 return slow_call;
4309 }
4310
4311
4312 /**
4313 * Build special case code for calls to hashCode on an object. This call may
4314 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4315 * slightly different code.
4316 */
4317 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4318 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4319 assert(!(is_virtual && is_static), "either virtual, special, or static");
4320
4321 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4322
4323 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4324 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4325 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4326 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4327 Node* obj = NULL;
4328 if (!is_static) {
4329 // Check for hashing null object
4330 obj = null_check_receiver();
4331 if (stopped()) return true; // unconditionally null
4332 result_reg->init_req(_null_path, top());
4333 result_val->init_req(_null_path, top());
4334 } else {
4335 // Do a null check, and return zero if null.
4336 // System.identityHashCode(null) == 0
4337 obj = argument(0);
4338 Node* null_ctl = top();
4339 obj = null_check_oop(obj, &null_ctl);
4340 result_reg->init_req(_null_path, null_ctl);
4341 result_val->init_req(_null_path, _gvn.intcon(0));
4342 }
4343
4344 // Unconditionally null? Then return right away.
4345 if (stopped()) {
4346 set_control( result_reg->in(_null_path));
4347 if (!stopped())
4348 set_result(result_val->in(_null_path));
4349 return true;
4350 }
4351
4352 // We only go to the fast case code if we pass a number of guards. The
4353 // paths which do not pass are accumulated in the slow_region.
4354 RegionNode* slow_region = new RegionNode(1);
4355 record_for_igvn(slow_region);
4356
4357 // If this is a virtual call, we generate a funny guard. We pull out
4358 // the vtable entry corresponding to hashCode() from the target object.
4359 // If the target method which we are calling happens to be the native
4360 // Object hashCode() method, we pass the guard. We do not need this
4361 // guard for non-virtual calls -- the caller is known to be the native
4362 // Object hashCode().
4363 if (is_virtual) {
4364 // After null check, get the object's klass.
4365 Node* obj_klass = load_object_klass(obj);
4366 generate_virtual_guard(obj_klass, slow_region);
4367 }
4368
4369 // Get the header out of the object, use LoadMarkNode when available
4370 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4371 // The control of the load must be NULL. Otherwise, the load can move before
4372 // the null check after castPP removal.
4373 Node* no_ctrl = NULL;
4374 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4375
4376 // Test the header to see if it is unlocked.
4377 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4378 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4379 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4380 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4381 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4382
4383 generate_slow_guard(test_unlocked, slow_region);
4384
4385 // Get the hash value and check to see that it has been properly assigned.
4386 // We depend on hash_mask being at most 32 bits and avoid the use of
4387 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4388 // vm: see markOop.hpp.
4389 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4390 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4391 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4392 // This hack lets the hash bits live anywhere in the mark object now, as long
4393 // as the shift drops the relevant bits into the low 32 bits. Note that
4394 // Java spec says that HashCode is an int so there's no point in capturing
4395 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4396 hshifted_header = ConvX2I(hshifted_header);
4397 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4398
4399 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4400 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4401 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4402
4403 generate_slow_guard(test_assigned, slow_region);
4404
4405 Node* init_mem = reset_memory();
4406 // fill in the rest of the null path:
4407 result_io ->init_req(_null_path, i_o());
4408 result_mem->init_req(_null_path, init_mem);
4409
4410 result_val->init_req(_fast_path, hash_val);
4411 result_reg->init_req(_fast_path, control());
4412 result_io ->init_req(_fast_path, i_o());
4413 result_mem->init_req(_fast_path, init_mem);
4414
4415 // Generate code for the slow case. We make a call to hashCode().
4416 set_control(_gvn.transform(slow_region));
4417 if (!stopped()) {
4418 // No need for PreserveJVMState, because we're using up the present state.
4419 set_all_memory(init_mem);
4420 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4421 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4422 Node* slow_result = set_results_for_java_call(slow_call);
4423 // this->control() comes from set_results_for_java_call
4424 result_reg->init_req(_slow_path, control());
4425 result_val->init_req(_slow_path, slow_result);
4426 result_io ->set_req(_slow_path, i_o());
4427 result_mem ->set_req(_slow_path, reset_memory());
4428 }
4429
4430 // Return the combined state.
4431 set_i_o( _gvn.transform(result_io) );
4432 set_all_memory( _gvn.transform(result_mem));
4433
4434 set_result(result_reg, result_val);
4435 return true;
4436 }
4437
4438 //---------------------------inline_native_getClass----------------------------
4439 // public final native Class<?> java.lang.Object.getClass();
4440 //
4441 // Build special case code for calls to getClass on an object.
4442 bool LibraryCallKit::inline_native_getClass() {
4443 Node* obj = null_check_receiver();
4444 if (stopped()) return true;
4445 set_result(load_mirror_from_klass(load_object_klass(obj)));
4446 return true;
4447 }
4448
4449 //-----------------inline_native_Reflection_getCallerClass---------------------
4450 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4451 //
4452 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4453 //
4454 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4455 // in that it must skip particular security frames and checks for
4456 // caller sensitive methods.
4457 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4458 #ifndef PRODUCT
4459 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4460 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4461 }
4462 #endif
4463
4464 if (!jvms()->has_method()) {
4465 #ifndef PRODUCT
4466 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4467 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4468 }
4469 #endif
4470 return false;
4471 }
4472
4473 // Walk back up the JVM state to find the caller at the required
4474 // depth.
4475 JVMState* caller_jvms = jvms();
4476
4477 // Cf. JVM_GetCallerClass
4478 // NOTE: Start the loop at depth 1 because the current JVM state does
4479 // not include the Reflection.getCallerClass() frame.
4480 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4481 ciMethod* m = caller_jvms->method();
4482 switch (n) {
4483 case 0:
4484 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4485 break;
4486 case 1:
4487 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4488 if (!m->caller_sensitive()) {
4489 #ifndef PRODUCT
4490 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4491 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4492 }
4493 #endif
4494 return false; // bail-out; let JVM_GetCallerClass do the work
4495 }
4496 break;
4497 default:
4498 if (!m->is_ignored_by_security_stack_walk()) {
4499 // We have reached the desired frame; return the holder class.
4500 // Acquire method holder as java.lang.Class and push as constant.
4501 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4502 ciInstance* caller_mirror = caller_klass->java_mirror();
4503 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4504
4505 #ifndef PRODUCT
4506 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4507 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());
4508 tty->print_cr(" JVM state at this point:");
4509 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4510 ciMethod* m = jvms()->of_depth(i)->method();
4511 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4512 }
4513 }
4514 #endif
4515 return true;
4516 }
4517 break;
4518 }
4519 }
4520
4521 #ifndef PRODUCT
4522 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4523 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4524 tty->print_cr(" JVM state at this point:");
4525 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4526 ciMethod* m = jvms()->of_depth(i)->method();
4527 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4528 }
4529 }
4530 #endif
4531
4532 return false; // bail-out; let JVM_GetCallerClass do the work
4533 }
4534
4535 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4536 Node* arg = argument(0);
4537 Node* result = NULL;
4538
4539 switch (id) {
4540 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4541 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4542 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4543 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4544
4545 case vmIntrinsics::_doubleToLongBits: {
4546 // two paths (plus control) merge in a wood
4547 RegionNode *r = new RegionNode(3);
4548 Node *phi = new PhiNode(r, TypeLong::LONG);
4549
4550 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4551 // Build the boolean node
4552 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4553
4554 // Branch either way.
4555 // NaN case is less traveled, which makes all the difference.
4556 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4557 Node *opt_isnan = _gvn.transform(ifisnan);
4558 assert( opt_isnan->is_If(), "Expect an IfNode");
4559 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4560 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4561
4562 set_control(iftrue);
4563
4564 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4565 Node *slow_result = longcon(nan_bits); // return NaN
4566 phi->init_req(1, _gvn.transform( slow_result ));
4567 r->init_req(1, iftrue);
4568
4569 // Else fall through
4570 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4571 set_control(iffalse);
4572
4573 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4574 r->init_req(2, iffalse);
4575
4576 // Post merge
4577 set_control(_gvn.transform(r));
4578 record_for_igvn(r);
4579
4580 C->set_has_split_ifs(true); // Has chance for split-if optimization
4581 result = phi;
4582 assert(result->bottom_type()->isa_long(), "must be");
4583 break;
4584 }
4585
4586 case vmIntrinsics::_floatToIntBits: {
4587 // two paths (plus control) merge in a wood
4588 RegionNode *r = new RegionNode(3);
4589 Node *phi = new PhiNode(r, TypeInt::INT);
4590
4591 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4592 // Build the boolean node
4593 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4594
4595 // Branch either way.
4596 // NaN case is less traveled, which makes all the difference.
4597 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4598 Node *opt_isnan = _gvn.transform(ifisnan);
4599 assert( opt_isnan->is_If(), "Expect an IfNode");
4600 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4601 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4602
4603 set_control(iftrue);
4604
4605 static const jint nan_bits = 0x7fc00000;
4606 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4607 phi->init_req(1, _gvn.transform( slow_result ));
4608 r->init_req(1, iftrue);
4609
4610 // Else fall through
4611 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4612 set_control(iffalse);
4613
4614 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4615 r->init_req(2, iffalse);
4616
4617 // Post merge
4618 set_control(_gvn.transform(r));
4619 record_for_igvn(r);
4620
4621 C->set_has_split_ifs(true); // Has chance for split-if optimization
4622 result = phi;
4623 assert(result->bottom_type()->isa_int(), "must be");
4624 break;
4625 }
4626
4627 default:
4628 fatal_unexpected_iid(id);
4629 break;
4630 }
4631 set_result(_gvn.transform(result));
4632 return true;
4633 }
4634
4635 //----------------------inline_unsafe_copyMemory-------------------------
4636 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4637 bool LibraryCallKit::inline_unsafe_copyMemory() {
4638 if (callee()->is_static()) return false; // caller must have the capability!
4639 null_check_receiver(); // null-check receiver
4640 if (stopped()) return true;
4641
4642 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4643
4644 Node* src_ptr = argument(1); // type: oop
4645 Node* src_off = ConvL2X(argument(2)); // type: long
4646 Node* dst_ptr = argument(4); // type: oop
4647 Node* dst_off = ConvL2X(argument(5)); // type: long
4648 Node* size = ConvL2X(argument(7)); // type: long
4649
4650 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4651 "fieldOffset must be byte-scaled");
4652
4653 Node* src = make_unsafe_address(src_ptr, src_off);
4654 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4655
4656 // Conservatively insert a memory barrier on all memory slices.
4657 // Do not let writes of the copy source or destination float below the copy.
4658 insert_mem_bar(Op_MemBarCPUOrder);
4659
4660 // Call it. Note that the length argument is not scaled.
4661 make_runtime_call(RC_LEAF|RC_NO_FP,
4662 OptoRuntime::fast_arraycopy_Type(),
4663 StubRoutines::unsafe_arraycopy(),
4664 "unsafe_arraycopy",
4665 TypeRawPtr::BOTTOM,
4666 src, dst, size XTOP);
4667
4668 // Do not let reads of the copy destination float above the copy.
4669 insert_mem_bar(Op_MemBarCPUOrder);
4670
4671 return true;
4672 }
4673
4674 //------------------------clone_coping-----------------------------------
4675 // Helper function for inline_native_clone.
4676 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4677 assert(obj_size != NULL, "");
4678 Node* raw_obj = alloc_obj->in(1);
4679 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4680
4681 AllocateNode* alloc = NULL;
4682 if (ReduceBulkZeroing) {
4683 // We will be completely responsible for initializing this object -
4684 // mark Initialize node as complete.
4685 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4686 // The object was just allocated - there should be no any stores!
4687 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4688 // Mark as complete_with_arraycopy so that on AllocateNode
4689 // expansion, we know this AllocateNode is initialized by an array
4690 // copy and a StoreStore barrier exists after the array copy.
4691 alloc->initialization()->set_complete_with_arraycopy();
4692 }
4693
4694 // Copy the fastest available way.
4695 // TODO: generate fields copies for small objects instead.
4696 Node* src = obj;
4697 Node* dest = alloc_obj;
4698 Node* size = _gvn.transform(obj_size);
4699
4700 // Exclude the header but include array length to copy by 8 bytes words.
4701 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4702 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4703 instanceOopDesc::base_offset_in_bytes();
4704 // base_off:
4705 // 8 - 32-bit VM
4706 // 12 - 64-bit VM, compressed klass
4707 // 16 - 64-bit VM, normal klass
4708 if (base_off % BytesPerLong != 0) {
4709 assert(UseCompressedClassPointers, "");
4710 if (is_array) {
4711 // Exclude length to copy by 8 bytes words.
4712 base_off += sizeof(int);
4713 } else {
4714 // Include klass to copy by 8 bytes words.
4715 base_off = instanceOopDesc::klass_offset_in_bytes();
4716 }
4717 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4718 }
4719 src = basic_plus_adr(src, base_off);
4720 dest = basic_plus_adr(dest, base_off);
4721
4722 // Compute the length also, if needed:
4723 Node* countx = size;
4724 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4725 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4726
4727 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4728
4729 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4730 ac->set_clonebasic();
4731 Node* n = _gvn.transform(ac);
4732 if (n == ac) {
4733 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4734 } else {
4735 set_all_memory(n);
4736 }
4737
4738 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4739 if (card_mark) {
4740 assert(!is_array, "");
4741 // Put in store barrier for any and all oops we are sticking
4742 // into this object. (We could avoid this if we could prove
4743 // that the object type contains no oop fields at all.)
4744 Node* no_particular_value = NULL;
4745 Node* no_particular_field = NULL;
4746 int raw_adr_idx = Compile::AliasIdxRaw;
4747 post_barrier(control(),
4748 memory(raw_adr_type),
4749 alloc_obj,
4750 no_particular_field,
4751 raw_adr_idx,
4752 no_particular_value,
4753 T_OBJECT,
4754 false);
4755 }
4756
4757 // Do not let reads from the cloned object float above the arraycopy.
4758 if (alloc != NULL) {
4759 // Do not let stores that initialize this object be reordered with
4760 // a subsequent store that would make this object accessible by
4761 // other threads.
4762 // Record what AllocateNode this StoreStore protects so that
4763 // escape analysis can go from the MemBarStoreStoreNode to the
4764 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4765 // based on the escape status of the AllocateNode.
4766 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4767 } else {
4768 insert_mem_bar(Op_MemBarCPUOrder);
4769 }
4770 }
4771
4772 //------------------------inline_native_clone----------------------------
4773 // protected native Object java.lang.Object.clone();
4774 //
4775 // Here are the simple edge cases:
4776 // null receiver => normal trap
4777 // virtual and clone was overridden => slow path to out-of-line clone
4778 // not cloneable or finalizer => slow path to out-of-line Object.clone
4779 //
4780 // The general case has two steps, allocation and copying.
4781 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4782 //
4783 // Copying also has two cases, oop arrays and everything else.
4784 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4785 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4786 //
4787 // These steps fold up nicely if and when the cloned object's klass
4788 // can be sharply typed as an object array, a type array, or an instance.
4789 //
4790 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4791 PhiNode* result_val;
4792
4793 // Set the reexecute bit for the interpreter to reexecute
4794 // the bytecode that invokes Object.clone if deoptimization happens.
4795 { PreserveReexecuteState preexecs(this);
4796 jvms()->set_should_reexecute(true);
4797
4798 Node* obj = null_check_receiver();
4799 if (stopped()) return true;
4800
4801 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4802
4803 // If we are going to clone an instance, we need its exact type to
4804 // know the number and types of fields to convert the clone to
4805 // loads/stores. Maybe a speculative type can help us.
4806 if (!obj_type->klass_is_exact() &&
4807 obj_type->speculative_type() != NULL &&
4808 obj_type->speculative_type()->is_instance_klass()) {
4809 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4810 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4811 !spec_ik->has_injected_fields()) {
4812 ciKlass* k = obj_type->klass();
4813 if (!k->is_instance_klass() ||
4814 k->as_instance_klass()->is_interface() ||
4815 k->as_instance_klass()->has_subklass()) {
4816 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4817 }
4818 }
4819 }
4820
4821 Node* obj_klass = load_object_klass(obj);
4822 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4823 const TypeOopPtr* toop = ((tklass != NULL)
4824 ? tklass->as_instance_type()
4825 : TypeInstPtr::NOTNULL);
4826
4827 // Conservatively insert a memory barrier on all memory slices.
4828 // Do not let writes into the original float below the clone.
4829 insert_mem_bar(Op_MemBarCPUOrder);
4830
4831 // paths into result_reg:
4832 enum {
4833 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4834 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4835 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4836 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4837 PATH_LIMIT
4838 };
4839 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4840 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4841 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4842 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4843 record_for_igvn(result_reg);
4844
4845 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4846 int raw_adr_idx = Compile::AliasIdxRaw;
4847
4848 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4849 if (array_ctl != NULL) {
4850 // It's an array.
4851 PreserveJVMState pjvms(this);
4852 set_control(array_ctl);
4853 Node* obj_length = load_array_length(obj);
4854 Node* obj_size = NULL;
4855 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4856
4857 if (!use_ReduceInitialCardMarks()) {
4858 // If it is an oop array, it requires very special treatment,
4859 // because card marking is required on each card of the array.
4860 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4861 if (is_obja != NULL) {
4862 PreserveJVMState pjvms2(this);
4863 set_control(is_obja);
4864 // Generate a direct call to the right arraycopy function(s).
4865 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4866 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4867 ac->set_cloneoop();
4868 Node* n = _gvn.transform(ac);
4869 assert(n == ac, "cannot disappear");
4870 ac->connect_outputs(this);
4871
4872 result_reg->init_req(_objArray_path, control());
4873 result_val->init_req(_objArray_path, alloc_obj);
4874 result_i_o ->set_req(_objArray_path, i_o());
4875 result_mem ->set_req(_objArray_path, reset_memory());
4876 }
4877 }
4878 // Otherwise, there are no card marks to worry about.
4879 // (We can dispense with card marks if we know the allocation
4880 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4881 // causes the non-eden paths to take compensating steps to
4882 // simulate a fresh allocation, so that no further
4883 // card marks are required in compiled code to initialize
4884 // the object.)
4885
4886 if (!stopped()) {
4887 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4888
4889 // Present the results of the copy.
4890 result_reg->init_req(_array_path, control());
4891 result_val->init_req(_array_path, alloc_obj);
4892 result_i_o ->set_req(_array_path, i_o());
4893 result_mem ->set_req(_array_path, reset_memory());
4894 }
4895 }
4896
4897 // We only go to the instance fast case code if we pass a number of guards.
4898 // The paths which do not pass are accumulated in the slow_region.
4899 RegionNode* slow_region = new RegionNode(1);
4900 record_for_igvn(slow_region);
4901 if (!stopped()) {
4902 // It's an instance (we did array above). Make the slow-path tests.
4903 // If this is a virtual call, we generate a funny guard. We grab
4904 // the vtable entry corresponding to clone() from the target object.
4905 // If the target method which we are calling happens to be the
4906 // Object clone() method, we pass the guard. We do not need this
4907 // guard for non-virtual calls; the caller is known to be the native
4908 // Object clone().
4909 if (is_virtual) {
4910 generate_virtual_guard(obj_klass, slow_region);
4911 }
4912
4913 // The object must be easily cloneable and must not have a finalizer.
4914 // Both of these conditions may be checked in a single test.
4915 // We could optimize the test further, but we don't care.
4916 generate_access_flags_guard(obj_klass,
4917 // Test both conditions:
4918 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4919 // Must be cloneable but not finalizer:
4920 JVM_ACC_IS_CLONEABLE_FAST,
4921 slow_region);
4922 }
4923
4924 if (!stopped()) {
4925 // It's an instance, and it passed the slow-path tests.
4926 PreserveJVMState pjvms(this);
4927 Node* obj_size = NULL;
4928 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4929 // is reexecuted if deoptimization occurs and there could be problems when merging
4930 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4931 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4932
4933 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4934
4935 // Present the results of the slow call.
4936 result_reg->init_req(_instance_path, control());
4937 result_val->init_req(_instance_path, alloc_obj);
4938 result_i_o ->set_req(_instance_path, i_o());
4939 result_mem ->set_req(_instance_path, reset_memory());
4940 }
4941
4942 // Generate code for the slow case. We make a call to clone().
4943 set_control(_gvn.transform(slow_region));
4944 if (!stopped()) {
4945 PreserveJVMState pjvms(this);
4946 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4947 Node* slow_result = set_results_for_java_call(slow_call);
4948 // this->control() comes from set_results_for_java_call
4949 result_reg->init_req(_slow_path, control());
4950 result_val->init_req(_slow_path, slow_result);
4951 result_i_o ->set_req(_slow_path, i_o());
4952 result_mem ->set_req(_slow_path, reset_memory());
4953 }
4954
4955 // Return the combined state.
4956 set_control( _gvn.transform(result_reg));
4957 set_i_o( _gvn.transform(result_i_o));
4958 set_all_memory( _gvn.transform(result_mem));
4959 } // original reexecute is set back here
4960
4961 set_result(_gvn.transform(result_val));
4962 return true;
4963 }
4964
4965 // If we have a tighly coupled allocation, the arraycopy may take care
4966 // of the array initialization. If one of the guards we insert between
4967 // the allocation and the arraycopy causes a deoptimization, an
4968 // unitialized array will escape the compiled method. To prevent that
4969 // we set the JVM state for uncommon traps between the allocation and
4970 // the arraycopy to the state before the allocation so, in case of
4971 // deoptimization, we'll reexecute the allocation and the
4972 // initialization.
4973 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4974 if (alloc != NULL) {
4975 ciMethod* trap_method = alloc->jvms()->method();
4976 int trap_bci = alloc->jvms()->bci();
4977
4978 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4979 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4980 // Make sure there's no store between the allocation and the
4981 // arraycopy otherwise visible side effects could be rexecuted
4982 // in case of deoptimization and cause incorrect execution.
4983 bool no_interfering_store = true;
4984 Node* mem = alloc->in(TypeFunc::Memory);
4985 if (mem->is_MergeMem()) {
4986 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4987 Node* n = mms.memory();
4988 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4989 assert(n->is_Store(), "what else?");
4990 no_interfering_store = false;
4991 break;
4992 }
4993 }
4994 } else {
4995 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4996 Node* n = mms.memory();
4997 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4998 assert(n->is_Store(), "what else?");
4999 no_interfering_store = false;
5000 break;
5001 }
5002 }
5003 }
5004
5005 if (no_interfering_store) {
5006 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
5007 uint size = alloc->req();
5008 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
5009 old_jvms->set_map(sfpt);
5010 for (uint i = 0; i < size; i++) {
5011 sfpt->init_req(i, alloc->in(i));
5012 }
5013 // re-push array length for deoptimization
5014 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
5015 old_jvms->set_sp(old_jvms->sp()+1);
5016 old_jvms->set_monoff(old_jvms->monoff()+1);
5017 old_jvms->set_scloff(old_jvms->scloff()+1);
5018 old_jvms->set_endoff(old_jvms->endoff()+1);
5019 old_jvms->set_should_reexecute(true);
5020
5021 sfpt->set_i_o(map()->i_o());
5022 sfpt->set_memory(map()->memory());
5023 sfpt->set_control(map()->control());
5024
5025 JVMState* saved_jvms = jvms();
5026 saved_reexecute_sp = _reexecute_sp;
5027
5028 set_jvms(sfpt->jvms());
5029 _reexecute_sp = jvms()->sp();
5030
5031 return saved_jvms;
5032 }
5033 }
5034 }
5035 return NULL;
5036 }
5037
5038 // In case of a deoptimization, we restart execution at the
5039 // allocation, allocating a new array. We would leave an uninitialized
5040 // array in the heap that GCs wouldn't expect. Move the allocation
5041 // after the traps so we don't allocate the array if we
5042 // deoptimize. This is possible because tightly_coupled_allocation()
5043 // guarantees there's no observer of the allocated array at this point
5044 // and the control flow is simple enough.
5045 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
5046 int saved_reexecute_sp, uint new_idx) {
5047 if (saved_jvms != NULL && !stopped()) {
5048 assert(alloc != NULL, "only with a tightly coupled allocation");
5049 // restore JVM state to the state at the arraycopy
5050 saved_jvms->map()->set_control(map()->control());
5051 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5052 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5053 // If we've improved the types of some nodes (null check) while
5054 // emitting the guards, propagate them to the current state
5055 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
5056 set_jvms(saved_jvms);
5057 _reexecute_sp = saved_reexecute_sp;
5058
5059 // Remove the allocation from above the guards
5060 CallProjections callprojs;
5061 alloc->extract_projections(&callprojs, true);
5062 InitializeNode* init = alloc->initialization();
5063 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5064 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5065 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5066 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5067
5068 // move the allocation here (after the guards)
5069 _gvn.hash_delete(alloc);
5070 alloc->set_req(TypeFunc::Control, control());
5071 alloc->set_req(TypeFunc::I_O, i_o());
5072 Node *mem = reset_memory();
5073 set_all_memory(mem);
5074 alloc->set_req(TypeFunc::Memory, mem);
5075 set_control(init->proj_out_or_null(TypeFunc::Control));
5076 set_i_o(callprojs.fallthrough_ioproj);
5077
5078 // Update memory as done in GraphKit::set_output_for_allocation()
5079 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5080 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5081 if (ary_type->isa_aryptr() && length_type != NULL) {
5082 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5083 }
5084 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5085 int elemidx = C->get_alias_index(telemref);
5086 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
5087 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
5088
5089 Node* allocx = _gvn.transform(alloc);
5090 assert(allocx == alloc, "where has the allocation gone?");
5091 assert(dest->is_CheckCastPP(), "not an allocation result?");
5092
5093 _gvn.hash_delete(dest);
5094 dest->set_req(0, control());
5095 Node* destx = _gvn.transform(dest);
5096 assert(destx == dest, "where has the allocation result gone?");
5097 }
5098 }
5099
5100
5101 //------------------------------inline_arraycopy-----------------------
5102 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5103 // Object dest, int destPos,
5104 // int length);
5105 bool LibraryCallKit::inline_arraycopy() {
5106 // Get the arguments.
5107 Node* src = argument(0); // type: oop
5108 Node* src_offset = argument(1); // type: int
5109 Node* dest = argument(2); // type: oop
5110 Node* dest_offset = argument(3); // type: int
5111 Node* length = argument(4); // type: int
5112
5113 uint new_idx = C->unique();
5114
5115 // Check for allocation before we add nodes that would confuse
5116 // tightly_coupled_allocation()
5117 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5118
5119 int saved_reexecute_sp = -1;
5120 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5121 // See arraycopy_restore_alloc_state() comment
5122 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5123 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5124 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5125 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5126
5127 // The following tests must be performed
5128 // (1) src and dest are arrays.
5129 // (2) src and dest arrays must have elements of the same BasicType
5130 // (3) src and dest must not be null.
5131 // (4) src_offset must not be negative.
5132 // (5) dest_offset must not be negative.
5133 // (6) length must not be negative.
5134 // (7) src_offset + length must not exceed length of src.
5135 // (8) dest_offset + length must not exceed length of dest.
5136 // (9) each element of an oop array must be assignable
5137
5138 // (3) src and dest must not be null.
5139 // always do this here because we need the JVM state for uncommon traps
5140 Node* null_ctl = top();
5141 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5142 assert(null_ctl->is_top(), "no null control here");
5143 dest = null_check(dest, T_ARRAY);
5144
5145 if (!can_emit_guards) {
5146 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5147 // guards but the arraycopy node could still take advantage of a
5148 // tightly allocated allocation. tightly_coupled_allocation() is
5149 // called again to make sure it takes the null check above into
5150 // account: the null check is mandatory and if it caused an
5151 // uncommon trap to be emitted then the allocation can't be
5152 // considered tightly coupled in this context.
5153 alloc = tightly_coupled_allocation(dest, NULL);
5154 }
5155
5156 bool validated = false;
5157
5158 const Type* src_type = _gvn.type(src);
5159 const Type* dest_type = _gvn.type(dest);
5160 const TypeAryPtr* top_src = src_type->isa_aryptr();
5161 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5162
5163 // Do we have the type of src?
5164 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5165 // Do we have the type of dest?
5166 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5167 // Is the type for src from speculation?
5168 bool src_spec = false;
5169 // Is the type for dest from speculation?
5170 bool dest_spec = false;
5171
5172 if ((!has_src || !has_dest) && can_emit_guards) {
5173 // We don't have sufficient type information, let's see if
5174 // speculative types can help. We need to have types for both src
5175 // and dest so that it pays off.
5176
5177 // Do we already have or could we have type information for src
5178 bool could_have_src = has_src;
5179 // Do we already have or could we have type information for dest
5180 bool could_have_dest = has_dest;
5181
5182 ciKlass* src_k = NULL;
5183 if (!has_src) {
5184 src_k = src_type->speculative_type_not_null();
5185 if (src_k != NULL && src_k->is_array_klass()) {
5186 could_have_src = true;
5187 }
5188 }
5189
5190 ciKlass* dest_k = NULL;
5191 if (!has_dest) {
5192 dest_k = dest_type->speculative_type_not_null();
5193 if (dest_k != NULL && dest_k->is_array_klass()) {
5194 could_have_dest = true;
5195 }
5196 }
5197
5198 if (could_have_src && could_have_dest) {
5199 // This is going to pay off so emit the required guards
5200 if (!has_src) {
5201 src = maybe_cast_profiled_obj(src, src_k, true);
5202 src_type = _gvn.type(src);
5203 top_src = src_type->isa_aryptr();
5204 has_src = (top_src != NULL && top_src->klass() != NULL);
5205 src_spec = true;
5206 }
5207 if (!has_dest) {
5208 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5209 dest_type = _gvn.type(dest);
5210 top_dest = dest_type->isa_aryptr();
5211 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5212 dest_spec = true;
5213 }
5214 }
5215 }
5216
5217 if (has_src && has_dest && can_emit_guards) {
5218 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5219 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5220 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5221 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5222
5223 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5224 // If both arrays are object arrays then having the exact types
5225 // for both will remove the need for a subtype check at runtime
5226 // before the call and may make it possible to pick a faster copy
5227 // routine (without a subtype check on every element)
5228 // Do we have the exact type of src?
5229 bool could_have_src = src_spec;
5230 // Do we have the exact type of dest?
5231 bool could_have_dest = dest_spec;
5232 ciKlass* src_k = top_src->klass();
5233 ciKlass* dest_k = top_dest->klass();
5234 if (!src_spec) {
5235 src_k = src_type->speculative_type_not_null();
5236 if (src_k != NULL && src_k->is_array_klass()) {
5237 could_have_src = true;
5238 }
5239 }
5240 if (!dest_spec) {
5241 dest_k = dest_type->speculative_type_not_null();
5242 if (dest_k != NULL && dest_k->is_array_klass()) {
5243 could_have_dest = true;
5244 }
5245 }
5246 if (could_have_src && could_have_dest) {
5247 // If we can have both exact types, emit the missing guards
5248 if (could_have_src && !src_spec) {
5249 src = maybe_cast_profiled_obj(src, src_k, true);
5250 }
5251 if (could_have_dest && !dest_spec) {
5252 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5253 }
5254 }
5255 }
5256 }
5257
5258 ciMethod* trap_method = method();
5259 int trap_bci = bci();
5260 if (saved_jvms != NULL) {
5261 trap_method = alloc->jvms()->method();
5262 trap_bci = alloc->jvms()->bci();
5263 }
5264
5265 bool negative_length_guard_generated = false;
5266
5267 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5268 can_emit_guards &&
5269 !src->is_top() && !dest->is_top()) {
5270 // validate arguments: enables transformation the ArrayCopyNode
5271 validated = true;
5272
5273 RegionNode* slow_region = new RegionNode(1);
5274 record_for_igvn(slow_region);
5275
5276 // (1) src and dest are arrays.
5277 generate_non_array_guard(load_object_klass(src), slow_region);
5278 generate_non_array_guard(load_object_klass(dest), slow_region);
5279
5280 // (2) src and dest arrays must have elements of the same BasicType
5281 // done at macro expansion or at Ideal transformation time
5282
5283 // (4) src_offset must not be negative.
5284 generate_negative_guard(src_offset, slow_region);
5285
5286 // (5) dest_offset must not be negative.
5287 generate_negative_guard(dest_offset, slow_region);
5288
5289 // (7) src_offset + length must not exceed length of src.
5290 generate_limit_guard(src_offset, length,
5291 load_array_length(src),
5292 slow_region);
5293
5294 // (8) dest_offset + length must not exceed length of dest.
5295 generate_limit_guard(dest_offset, length,
5296 load_array_length(dest),
5297 slow_region);
5298
5299 // (6) length must not be negative.
5300 // This is also checked in generate_arraycopy() during macro expansion, but
5301 // we also have to check it here for the case where the ArrayCopyNode will
5302 // be eliminated by Escape Analysis.
5303 if (EliminateAllocations) {
5304 generate_negative_guard(length, slow_region);
5305 negative_length_guard_generated = true;
5306 }
5307
5308 // (9) each element of an oop array must be assignable
5309 Node* src_klass = load_object_klass(src);
5310 Node* dest_klass = load_object_klass(dest);
5311 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5312
5313 if (not_subtype_ctrl != top()) {
5314 PreserveJVMState pjvms(this);
5315 set_control(not_subtype_ctrl);
5316 uncommon_trap(Deoptimization::Reason_intrinsic,
5317 Deoptimization::Action_make_not_entrant);
5318 assert(stopped(), "Should be stopped");
5319 }
5320 {
5321 PreserveJVMState pjvms(this);
5322 set_control(_gvn.transform(slow_region));
5323 uncommon_trap(Deoptimization::Reason_intrinsic,
5324 Deoptimization::Action_make_not_entrant);
5325 assert(stopped(), "Should be stopped");
5326 }
5327
5328 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
5329 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
5330 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
5331 }
5332
5333 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5334
5335 if (stopped()) {
5336 return true;
5337 }
5338
5339 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5340 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5341 // so the compiler has a chance to eliminate them: during macro expansion,
5342 // we have to set their control (CastPP nodes are eliminated).
5343 load_object_klass(src), load_object_klass(dest),
5344 load_array_length(src), load_array_length(dest));
5345
5346 ac->set_arraycopy(validated);
5347
5348 Node* n = _gvn.transform(ac);
5349 if (n == ac) {
5350 ac->connect_outputs(this);
5351 } else {
5352 assert(validated, "shouldn't transform if all arguments not validated");
5353 set_all_memory(n);
5354 }
5355 clear_upper_avx();
5356
5357
5358 return true;
5359 }
5360
5361
5362 // Helper function which determines if an arraycopy immediately follows
5363 // an allocation, with no intervening tests or other escapes for the object.
5364 AllocateArrayNode*
5365 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5366 RegionNode* slow_region) {
5367 if (stopped()) return NULL; // no fast path
5368 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5369
5370 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5371 if (alloc == NULL) return NULL;
5372
5373 Node* rawmem = memory(Compile::AliasIdxRaw);
5374 // Is the allocation's memory state untouched?
5375 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5376 // Bail out if there have been raw-memory effects since the allocation.
5377 // (Example: There might have been a call or safepoint.)
5378 return NULL;
5379 }
5380 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5381 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5382 return NULL;
5383 }
5384
5385 // There must be no unexpected observers of this allocation.
5386 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5387 Node* obs = ptr->fast_out(i);
5388 if (obs != this->map()) {
5389 return NULL;
5390 }
5391 }
5392
5393 // This arraycopy must unconditionally follow the allocation of the ptr.
5394 Node* alloc_ctl = ptr->in(0);
5395 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5396
5397 Node* ctl = control();
5398 while (ctl != alloc_ctl) {
5399 // There may be guards which feed into the slow_region.
5400 // Any other control flow means that we might not get a chance
5401 // to finish initializing the allocated object.
5402 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5403 IfNode* iff = ctl->in(0)->as_If();
5404 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
5405 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5406 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5407 ctl = iff->in(0); // This test feeds the known slow_region.
5408 continue;
5409 }
5410 // One more try: Various low-level checks bottom out in
5411 // uncommon traps. If the debug-info of the trap omits
5412 // any reference to the allocation, as we've already
5413 // observed, then there can be no objection to the trap.
5414 bool found_trap = false;
5415 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5416 Node* obs = not_ctl->fast_out(j);
5417 if (obs->in(0) == not_ctl && obs->is_Call() &&
5418 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5419 found_trap = true; break;
5420 }
5421 }
5422 if (found_trap) {
5423 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5424 continue;
5425 }
5426 }
5427 return NULL;
5428 }
5429
5430 // If we get this far, we have an allocation which immediately
5431 // precedes the arraycopy, and we can take over zeroing the new object.
5432 // The arraycopy will finish the initialization, and provide
5433 // a new control state to which we will anchor the destination pointer.
5434
5435 return alloc;
5436 }
5437
5438 //-------------inline_encodeISOArray-----------------------------------
5439 // encode char[] to byte[] in ISO_8859_1
5440 bool LibraryCallKit::inline_encodeISOArray() {
5441 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5442 // no receiver since it is static method
5443 Node *src = argument(0);
5444 Node *src_offset = argument(1);
5445 Node *dst = argument(2);
5446 Node *dst_offset = argument(3);
5447 Node *length = argument(4);
5448
5449 const Type* src_type = src->Value(&_gvn);
5450 const Type* dst_type = dst->Value(&_gvn);
5451 const TypeAryPtr* top_src = src_type->isa_aryptr();
5452 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5453 if (top_src == NULL || top_src->klass() == NULL ||
5454 top_dest == NULL || top_dest->klass() == NULL) {
5455 // failed array check
5456 return false;
5457 }
5458
5459 // Figure out the size and type of the elements we will be copying.
5460 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5461 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5462 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5463 return false;
5464 }
5465
5466 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5467 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5468 // 'src_start' points to src array + scaled offset
5469 // 'dst_start' points to dst array + scaled offset
5470
5471 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5472 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5473 enc = _gvn.transform(enc);
5474 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5475 set_memory(res_mem, mtype);
5476 set_result(enc);
5477 clear_upper_avx();
5478
5479 return true;
5480 }
5481
5482 //-------------inline_multiplyToLen-----------------------------------
5483 bool LibraryCallKit::inline_multiplyToLen() {
5484 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5485
5486 address stubAddr = StubRoutines::multiplyToLen();
5487 if (stubAddr == NULL) {
5488 return false; // Intrinsic's stub is not implemented on this platform
5489 }
5490 const char* stubName = "multiplyToLen";
5491
5492 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5493
5494 // no receiver because it is a static method
5495 Node* x = argument(0);
5496 Node* xlen = argument(1);
5497 Node* y = argument(2);
5498 Node* ylen = argument(3);
5499 Node* z = argument(4);
5500
5501 const Type* x_type = x->Value(&_gvn);
5502 const Type* y_type = y->Value(&_gvn);
5503 const TypeAryPtr* top_x = x_type->isa_aryptr();
5504 const TypeAryPtr* top_y = y_type->isa_aryptr();
5505 if (top_x == NULL || top_x->klass() == NULL ||
5506 top_y == NULL || top_y->klass() == NULL) {
5507 // failed array check
5508 return false;
5509 }
5510
5511 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5512 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5513 if (x_elem != T_INT || y_elem != T_INT) {
5514 return false;
5515 }
5516
5517 // Set the original stack and the reexecute bit for the interpreter to reexecute
5518 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5519 // on the return from z array allocation in runtime.
5520 { PreserveReexecuteState preexecs(this);
5521 jvms()->set_should_reexecute(true);
5522
5523 Node* x_start = array_element_address(x, intcon(0), x_elem);
5524 Node* y_start = array_element_address(y, intcon(0), y_elem);
5525 // 'x_start' points to x array + scaled xlen
5526 // 'y_start' points to y array + scaled ylen
5527
5528 // Allocate the result array
5529 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5530 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5531 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5532
5533 IdealKit ideal(this);
5534
5535 #define __ ideal.
5536 Node* one = __ ConI(1);
5537 Node* zero = __ ConI(0);
5538 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5539 __ set(need_alloc, zero);
5540 __ set(z_alloc, z);
5541 __ if_then(z, BoolTest::eq, null()); {
5542 __ increment (need_alloc, one);
5543 } __ else_(); {
5544 // Update graphKit memory and control from IdealKit.
5545 sync_kit(ideal);
5546 Node* zlen_arg = load_array_length(z);
5547 // Update IdealKit memory and control from graphKit.
5548 __ sync_kit(this);
5549 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5550 __ increment (need_alloc, one);
5551 } __ end_if();
5552 } __ end_if();
5553
5554 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5555 // Update graphKit memory and control from IdealKit.
5556 sync_kit(ideal);
5557 Node * narr = new_array(klass_node, zlen, 1);
5558 // Update IdealKit memory and control from graphKit.
5559 __ sync_kit(this);
5560 __ set(z_alloc, narr);
5561 } __ end_if();
5562
5563 sync_kit(ideal);
5564 z = __ value(z_alloc);
5565 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5566 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5567 // Final sync IdealKit and GraphKit.
5568 final_sync(ideal);
5569 #undef __
5570
5571 Node* z_start = array_element_address(z, intcon(0), T_INT);
5572
5573 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5574 OptoRuntime::multiplyToLen_Type(),
5575 stubAddr, stubName, TypePtr::BOTTOM,
5576 x_start, xlen, y_start, ylen, z_start, zlen);
5577 } // original reexecute is set back here
5578
5579 C->set_has_split_ifs(true); // Has chance for split-if optimization
5580 set_result(z);
5581 return true;
5582 }
5583
5584 //-------------inline_squareToLen------------------------------------
5585 bool LibraryCallKit::inline_squareToLen() {
5586 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5587
5588 address stubAddr = StubRoutines::squareToLen();
5589 if (stubAddr == NULL) {
5590 return false; // Intrinsic's stub is not implemented on this platform
5591 }
5592 const char* stubName = "squareToLen";
5593
5594 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5595
5596 Node* x = argument(0);
5597 Node* len = argument(1);
5598 Node* z = argument(2);
5599 Node* zlen = argument(3);
5600
5601 const Type* x_type = x->Value(&_gvn);
5602 const Type* z_type = z->Value(&_gvn);
5603 const TypeAryPtr* top_x = x_type->isa_aryptr();
5604 const TypeAryPtr* top_z = z_type->isa_aryptr();
5605 if (top_x == NULL || top_x->klass() == NULL ||
5606 top_z == NULL || top_z->klass() == NULL) {
5607 // failed array check
5608 return false;
5609 }
5610
5611 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5612 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5613 if (x_elem != T_INT || z_elem != T_INT) {
5614 return false;
5615 }
5616
5617
5618 Node* x_start = array_element_address(x, intcon(0), x_elem);
5619 Node* z_start = array_element_address(z, intcon(0), z_elem);
5620
5621 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5622 OptoRuntime::squareToLen_Type(),
5623 stubAddr, stubName, TypePtr::BOTTOM,
5624 x_start, len, z_start, zlen);
5625
5626 set_result(z);
5627 return true;
5628 }
5629
5630 //-------------inline_mulAdd------------------------------------------
5631 bool LibraryCallKit::inline_mulAdd() {
5632 assert(UseMulAddIntrinsic, "not implemented on this platform");
5633
5634 address stubAddr = StubRoutines::mulAdd();
5635 if (stubAddr == NULL) {
5636 return false; // Intrinsic's stub is not implemented on this platform
5637 }
5638 const char* stubName = "mulAdd";
5639
5640 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5641
5642 Node* out = argument(0);
5643 Node* in = argument(1);
5644 Node* offset = argument(2);
5645 Node* len = argument(3);
5646 Node* k = argument(4);
5647
5648 const Type* out_type = out->Value(&_gvn);
5649 const Type* in_type = in->Value(&_gvn);
5650 const TypeAryPtr* top_out = out_type->isa_aryptr();
5651 const TypeAryPtr* top_in = in_type->isa_aryptr();
5652 if (top_out == NULL || top_out->klass() == NULL ||
5653 top_in == NULL || top_in->klass() == NULL) {
5654 // failed array check
5655 return false;
5656 }
5657
5658 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5659 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5660 if (out_elem != T_INT || in_elem != T_INT) {
5661 return false;
5662 }
5663
5664 Node* outlen = load_array_length(out);
5665 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5666 Node* out_start = array_element_address(out, intcon(0), out_elem);
5667 Node* in_start = array_element_address(in, intcon(0), in_elem);
5668
5669 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5670 OptoRuntime::mulAdd_Type(),
5671 stubAddr, stubName, TypePtr::BOTTOM,
5672 out_start,in_start, new_offset, len, k);
5673 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5674 set_result(result);
5675 return true;
5676 }
5677
5678 //-------------inline_montgomeryMultiply-----------------------------------
5679 bool LibraryCallKit::inline_montgomeryMultiply() {
5680 address stubAddr = StubRoutines::montgomeryMultiply();
5681 if (stubAddr == NULL) {
5682 return false; // Intrinsic's stub is not implemented on this platform
5683 }
5684
5685 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5686 const char* stubName = "montgomery_multiply";
5687
5688 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5689
5690 Node* a = argument(0);
5691 Node* b = argument(1);
5692 Node* n = argument(2);
5693 Node* len = argument(3);
5694 Node* inv = argument(4);
5695 Node* m = argument(6);
5696
5697 const Type* a_type = a->Value(&_gvn);
5698 const TypeAryPtr* top_a = a_type->isa_aryptr();
5699 const Type* b_type = b->Value(&_gvn);
5700 const TypeAryPtr* top_b = b_type->isa_aryptr();
5701 const Type* n_type = a->Value(&_gvn);
5702 const TypeAryPtr* top_n = n_type->isa_aryptr();
5703 const Type* m_type = a->Value(&_gvn);
5704 const TypeAryPtr* top_m = m_type->isa_aryptr();
5705 if (top_a == NULL || top_a->klass() == NULL ||
5706 top_b == NULL || top_b->klass() == NULL ||
5707 top_n == NULL || top_n->klass() == NULL ||
5708 top_m == NULL || top_m->klass() == NULL) {
5709 // failed array check
5710 return false;
5711 }
5712
5713 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5714 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5715 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5716 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5717 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5718 return false;
5719 }
5720
5721 // Make the call
5722 {
5723 Node* a_start = array_element_address(a, intcon(0), a_elem);
5724 Node* b_start = array_element_address(b, intcon(0), b_elem);
5725 Node* n_start = array_element_address(n, intcon(0), n_elem);
5726 Node* m_start = array_element_address(m, intcon(0), m_elem);
5727
5728 Node* call = make_runtime_call(RC_LEAF,
5729 OptoRuntime::montgomeryMultiply_Type(),
5730 stubAddr, stubName, TypePtr::BOTTOM,
5731 a_start, b_start, n_start, len, inv, top(),
5732 m_start);
5733 set_result(m);
5734 }
5735
5736 return true;
5737 }
5738
5739 bool LibraryCallKit::inline_montgomerySquare() {
5740 address stubAddr = StubRoutines::montgomerySquare();
5741 if (stubAddr == NULL) {
5742 return false; // Intrinsic's stub is not implemented on this platform
5743 }
5744
5745 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5746 const char* stubName = "montgomery_square";
5747
5748 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5749
5750 Node* a = argument(0);
5751 Node* n = argument(1);
5752 Node* len = argument(2);
5753 Node* inv = argument(3);
5754 Node* m = argument(5);
5755
5756 const Type* a_type = a->Value(&_gvn);
5757 const TypeAryPtr* top_a = a_type->isa_aryptr();
5758 const Type* n_type = a->Value(&_gvn);
5759 const TypeAryPtr* top_n = n_type->isa_aryptr();
5760 const Type* m_type = a->Value(&_gvn);
5761 const TypeAryPtr* top_m = m_type->isa_aryptr();
5762 if (top_a == NULL || top_a->klass() == NULL ||
5763 top_n == NULL || top_n->klass() == NULL ||
5764 top_m == NULL || top_m->klass() == NULL) {
5765 // failed array check
5766 return false;
5767 }
5768
5769 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5770 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5771 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5772 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5773 return false;
5774 }
5775
5776 // Make the call
5777 {
5778 Node* a_start = array_element_address(a, intcon(0), a_elem);
5779 Node* n_start = array_element_address(n, intcon(0), n_elem);
5780 Node* m_start = array_element_address(m, intcon(0), m_elem);
5781
5782 Node* call = make_runtime_call(RC_LEAF,
5783 OptoRuntime::montgomerySquare_Type(),
5784 stubAddr, stubName, TypePtr::BOTTOM,
5785 a_start, n_start, len, inv, top(),
5786 m_start);
5787 set_result(m);
5788 }
5789
5790 return true;
5791 }
5792
5793 //-------------inline_vectorizedMismatch------------------------------
5794 bool LibraryCallKit::inline_vectorizedMismatch() {
5795 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5796
5797 address stubAddr = StubRoutines::vectorizedMismatch();
5798 if (stubAddr == NULL) {
5799 return false; // Intrinsic's stub is not implemented on this platform
5800 }
5801 const char* stubName = "vectorizedMismatch";
5802 int size_l = callee()->signature()->size();
5803 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5804
5805 Node* obja = argument(0);
5806 Node* aoffset = argument(1);
5807 Node* objb = argument(3);
5808 Node* boffset = argument(4);
5809 Node* length = argument(6);
5810 Node* scale = argument(7);
5811
5812 const Type* a_type = obja->Value(&_gvn);
5813 const Type* b_type = objb->Value(&_gvn);
5814 const TypeAryPtr* top_a = a_type->isa_aryptr();
5815 const TypeAryPtr* top_b = b_type->isa_aryptr();
5816 if (top_a == NULL || top_a->klass() == NULL ||
5817 top_b == NULL || top_b->klass() == NULL) {
5818 // failed array check
5819 return false;
5820 }
5821
5822 Node* call;
5823 jvms()->set_should_reexecute(true);
5824
5825 Node* obja_adr = make_unsafe_address(obja, aoffset);
5826 Node* objb_adr = make_unsafe_address(objb, boffset);
5827
5828 call = make_runtime_call(RC_LEAF,
5829 OptoRuntime::vectorizedMismatch_Type(),
5830 stubAddr, stubName, TypePtr::BOTTOM,
5831 obja_adr, objb_adr, length, scale);
5832
5833 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5834 set_result(result);
5835 return true;
5836 }
5837
5838 /**
5839 * Calculate CRC32 for byte.
5840 * int java.util.zip.CRC32.update(int crc, int b)
5841 */
5842 bool LibraryCallKit::inline_updateCRC32() {
5843 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5844 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5845 // no receiver since it is static method
5846 Node* crc = argument(0); // type: int
5847 Node* b = argument(1); // type: int
5848
5849 /*
5850 * int c = ~ crc;
5851 * b = timesXtoThe32[(b ^ c) & 0xFF];
5852 * b = b ^ (c >>> 8);
5853 * crc = ~b;
5854 */
5855
5856 Node* M1 = intcon(-1);
5857 crc = _gvn.transform(new XorINode(crc, M1));
5858 Node* result = _gvn.transform(new XorINode(crc, b));
5859 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5860
5861 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5862 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5863 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5864 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5865
5866 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5867 result = _gvn.transform(new XorINode(crc, result));
5868 result = _gvn.transform(new XorINode(result, M1));
5869 set_result(result);
5870 return true;
5871 }
5872
5873 /**
5874 * Calculate CRC32 for byte[] array.
5875 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5876 */
5877 bool LibraryCallKit::inline_updateBytesCRC32() {
5878 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5879 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5880 // no receiver since it is static method
5881 Node* crc = argument(0); // type: int
5882 Node* src = argument(1); // type: oop
5883 Node* offset = argument(2); // type: int
5884 Node* length = argument(3); // type: int
5885
5886 const Type* src_type = src->Value(&_gvn);
5887 const TypeAryPtr* top_src = src_type->isa_aryptr();
5888 if (top_src == NULL || top_src->klass() == NULL) {
5889 // failed array check
5890 return false;
5891 }
5892
5893 // Figure out the size and type of the elements we will be copying.
5894 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5895 if (src_elem != T_BYTE) {
5896 return false;
5897 }
5898
5899 // 'src_start' points to src array + scaled offset
5900 Node* src_start = array_element_address(src, offset, src_elem);
5901
5902 // We assume that range check is done by caller.
5903 // TODO: generate range check (offset+length < src.length) in debug VM.
5904
5905 // Call the stub.
5906 address stubAddr = StubRoutines::updateBytesCRC32();
5907 const char *stubName = "updateBytesCRC32";
5908
5909 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5910 stubAddr, stubName, TypePtr::BOTTOM,
5911 crc, src_start, length);
5912 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5913 set_result(result);
5914 return true;
5915 }
5916
5917 /**
5918 * Calculate CRC32 for ByteBuffer.
5919 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5920 */
5921 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5922 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5923 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5924 // no receiver since it is static method
5925 Node* crc = argument(0); // type: int
5926 Node* src = argument(1); // type: long
5927 Node* offset = argument(3); // type: int
5928 Node* length = argument(4); // type: int
5929
5930 src = ConvL2X(src); // adjust Java long to machine word
5931 Node* base = _gvn.transform(new CastX2PNode(src));
5932 offset = ConvI2X(offset);
5933
5934 // 'src_start' points to src array + scaled offset
5935 Node* src_start = basic_plus_adr(top(), base, offset);
5936
5937 // Call the stub.
5938 address stubAddr = StubRoutines::updateBytesCRC32();
5939 const char *stubName = "updateBytesCRC32";
5940
5941 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5942 stubAddr, stubName, TypePtr::BOTTOM,
5943 crc, src_start, length);
5944 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5945 set_result(result);
5946 return true;
5947 }
5948
5949 //------------------------------get_table_from_crc32c_class-----------------------
5950 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5951 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5952 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5953
5954 return table;
5955 }
5956
5957 //------------------------------inline_updateBytesCRC32C-----------------------
5958 //
5959 // Calculate CRC32C for byte[] array.
5960 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5961 //
5962 bool LibraryCallKit::inline_updateBytesCRC32C() {
5963 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5964 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5965 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5966 // no receiver since it is a static method
5967 Node* crc = argument(0); // type: int
5968 Node* src = argument(1); // type: oop
5969 Node* offset = argument(2); // type: int
5970 Node* end = argument(3); // type: int
5971
5972 Node* length = _gvn.transform(new SubINode(end, offset));
5973
5974 const Type* src_type = src->Value(&_gvn);
5975 const TypeAryPtr* top_src = src_type->isa_aryptr();
5976 if (top_src == NULL || top_src->klass() == NULL) {
5977 // failed array check
5978 return false;
5979 }
5980
5981 // Figure out the size and type of the elements we will be copying.
5982 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5983 if (src_elem != T_BYTE) {
5984 return false;
5985 }
5986
5987 // 'src_start' points to src array + scaled offset
5988 Node* src_start = array_element_address(src, offset, src_elem);
5989
5990 // static final int[] byteTable in class CRC32C
5991 Node* table = get_table_from_crc32c_class(callee()->holder());
5992 Node* table_start = array_element_address(table, intcon(0), T_INT);
5993
5994 // We assume that range check is done by caller.
5995 // TODO: generate range check (offset+length < src.length) in debug VM.
5996
5997 // Call the stub.
5998 address stubAddr = StubRoutines::updateBytesCRC32C();
5999 const char *stubName = "updateBytesCRC32C";
6000
6001 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6002 stubAddr, stubName, TypePtr::BOTTOM,
6003 crc, src_start, length, table_start);
6004 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6005 set_result(result);
6006 return true;
6007 }
6008
6009 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
6010 //
6011 // Calculate CRC32C for DirectByteBuffer.
6012 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
6013 //
6014 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
6015 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
6016 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
6017 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
6018 // no receiver since it is a static method
6019 Node* crc = argument(0); // type: int
6020 Node* src = argument(1); // type: long
6021 Node* offset = argument(3); // type: int
6022 Node* end = argument(4); // type: int
6023
6024 Node* length = _gvn.transform(new SubINode(end, offset));
6025
6026 src = ConvL2X(src); // adjust Java long to machine word
6027 Node* base = _gvn.transform(new CastX2PNode(src));
6028 offset = ConvI2X(offset);
6029
6030 // 'src_start' points to src array + scaled offset
6031 Node* src_start = basic_plus_adr(top(), base, offset);
6032
6033 // static final int[] byteTable in class CRC32C
6034 Node* table = get_table_from_crc32c_class(callee()->holder());
6035 Node* table_start = array_element_address(table, intcon(0), T_INT);
6036
6037 // Call the stub.
6038 address stubAddr = StubRoutines::updateBytesCRC32C();
6039 const char *stubName = "updateBytesCRC32C";
6040
6041 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6042 stubAddr, stubName, TypePtr::BOTTOM,
6043 crc, src_start, length, table_start);
6044 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6045 set_result(result);
6046 return true;
6047 }
6048
6049 //------------------------------inline_updateBytesAdler32----------------------
6050 //
6051 // Calculate Adler32 checksum for byte[] array.
6052 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6053 //
6054 bool LibraryCallKit::inline_updateBytesAdler32() {
6055 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6056 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6057 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6058 // no receiver since it is static method
6059 Node* crc = argument(0); // type: int
6060 Node* src = argument(1); // type: oop
6061 Node* offset = argument(2); // type: int
6062 Node* length = argument(3); // type: int
6063
6064 const Type* src_type = src->Value(&_gvn);
6065 const TypeAryPtr* top_src = src_type->isa_aryptr();
6066 if (top_src == NULL || top_src->klass() == NULL) {
6067 // failed array check
6068 return false;
6069 }
6070
6071 // Figure out the size and type of the elements we will be copying.
6072 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6073 if (src_elem != T_BYTE) {
6074 return false;
6075 }
6076
6077 // 'src_start' points to src array + scaled offset
6078 Node* src_start = array_element_address(src, offset, src_elem);
6079
6080 // We assume that range check is done by caller.
6081 // TODO: generate range check (offset+length < src.length) in debug VM.
6082
6083 // Call the stub.
6084 address stubAddr = StubRoutines::updateBytesAdler32();
6085 const char *stubName = "updateBytesAdler32";
6086
6087 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6088 stubAddr, stubName, TypePtr::BOTTOM,
6089 crc, src_start, length);
6090 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6091 set_result(result);
6092 return true;
6093 }
6094
6095 //------------------------------inline_updateByteBufferAdler32---------------
6096 //
6097 // Calculate Adler32 checksum for DirectByteBuffer.
6098 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6099 //
6100 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6101 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6102 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6103 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6104 // no receiver since it is static method
6105 Node* crc = argument(0); // type: int
6106 Node* src = argument(1); // type: long
6107 Node* offset = argument(3); // type: int
6108 Node* length = argument(4); // type: int
6109
6110 src = ConvL2X(src); // adjust Java long to machine word
6111 Node* base = _gvn.transform(new CastX2PNode(src));
6112 offset = ConvI2X(offset);
6113
6114 // 'src_start' points to src array + scaled offset
6115 Node* src_start = basic_plus_adr(top(), base, offset);
6116
6117 // Call the stub.
6118 address stubAddr = StubRoutines::updateBytesAdler32();
6119 const char *stubName = "updateBytesAdler32";
6120
6121 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6122 stubAddr, stubName, TypePtr::BOTTOM,
6123 crc, src_start, length);
6124
6125 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6126 set_result(result);
6127 return true;
6128 }
6129
6130 //----------------------------inline_reference_get----------------------------
6131 // public T java.lang.ref.Reference.get();
6132 bool LibraryCallKit::inline_reference_get() {
6133 const int referent_offset = java_lang_ref_Reference::referent_offset;
6134 guarantee(referent_offset > 0, "should have already been set");
6135
6136 // Get the argument:
6137 Node* reference_obj = null_check_receiver();
6138 if (stopped()) return true;
6139
6140 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6141
6142 ciInstanceKlass* klass = env()->Object_klass();
6143 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6144
6145 Node* no_ctrl = NULL;
6146 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6147
6148 // Use the pre-barrier to record the value in the referent field
6149 pre_barrier(false /* do_load */,
6150 control(),
6151 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6152 result /* pre_val */,
6153 T_OBJECT);
6154
6155 // Add memory barrier to prevent commoning reads from this field
6156 // across safepoint since GC can change its value.
6157 insert_mem_bar(Op_MemBarCPUOrder);
6158
6159 set_result(result);
6160 return true;
6161 }
6162
6163
6164 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6165 bool is_exact=true, bool is_static=false,
6166 ciInstanceKlass * fromKls=NULL) {
6167 if (fromKls == NULL) {
6168 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6169 assert(tinst != NULL, "obj is null");
6170 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6171 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6172 fromKls = tinst->klass()->as_instance_klass();
6173 } else {
6174 assert(is_static, "only for static field access");
6175 }
6176 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6177 ciSymbol::make(fieldTypeString),
6178 is_static);
6179
6180 assert (field != NULL, "undefined field");
6181 if (field == NULL) return (Node *) NULL;
6182
6183 if (is_static) {
6184 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6185 fromObj = makecon(tip);
6186 }
6187
6188 // Next code copied from Parse::do_get_xxx():
6189
6190 // Compute address and memory type.
6191 int offset = field->offset_in_bytes();
6192 bool is_vol = field->is_volatile();
6193 ciType* field_klass = field->type();
6194 assert(field_klass->is_loaded(), "should be loaded");
6195 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6196 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6197 BasicType bt = field->layout_type();
6198
6199 // Build the resultant type of the load
6200 const Type *type;
6201 if (bt == T_OBJECT) {
6202 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6203 } else {
6204 type = Type::get_const_basic_type(bt);
6205 }
6206
6207 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6208 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6209 }
6210 // Build the load.
6211 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6212 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6213 // If reference is volatile, prevent following memory ops from
6214 // floating up past the volatile read. Also prevents commoning
6215 // another volatile read.
6216 if (is_vol) {
6217 // Memory barrier includes bogus read of value to force load BEFORE membar
6218 insert_mem_bar(Op_MemBarAcquire, loadedField);
6219 }
6220 return loadedField;
6221 }
6222
6223 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6224 bool is_exact = true, bool is_static = false,
6225 ciInstanceKlass * fromKls = NULL) {
6226 if (fromKls == NULL) {
6227 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6228 assert(tinst != NULL, "obj is null");
6229 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6230 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6231 fromKls = tinst->klass()->as_instance_klass();
6232 }
6233 else {
6234 assert(is_static, "only for static field access");
6235 }
6236 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6237 ciSymbol::make(fieldTypeString),
6238 is_static);
6239
6240 assert(field != NULL, "undefined field");
6241 assert(!field->is_volatile(), "not defined for volatile fields");
6242
6243 if (is_static) {
6244 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6245 fromObj = makecon(tip);
6246 }
6247
6248 // Next code copied from Parse::do_get_xxx():
6249
6250 // Compute address and memory type.
6251 int offset = field->offset_in_bytes();
6252 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6253
6254 return adr;
6255 }
6256
6257 //------------------------------inline_aescrypt_Block-----------------------
6258 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6259 address stubAddr = NULL;
6260 const char *stubName;
6261 assert(UseAES, "need AES instruction support");
6262
6263 switch(id) {
6264 case vmIntrinsics::_aescrypt_encryptBlock:
6265 stubAddr = StubRoutines::aescrypt_encryptBlock();
6266 stubName = "aescrypt_encryptBlock";
6267 break;
6268 case vmIntrinsics::_aescrypt_decryptBlock:
6269 stubAddr = StubRoutines::aescrypt_decryptBlock();
6270 stubName = "aescrypt_decryptBlock";
6271 break;
6272 default:
6273 break;
6274 }
6275 if (stubAddr == NULL) return false;
6276
6277 Node* aescrypt_object = argument(0);
6278 Node* src = argument(1);
6279 Node* src_offset = argument(2);
6280 Node* dest = argument(3);
6281 Node* dest_offset = argument(4);
6282
6283 // (1) src and dest are arrays.
6284 const Type* src_type = src->Value(&_gvn);
6285 const Type* dest_type = dest->Value(&_gvn);
6286 const TypeAryPtr* top_src = src_type->isa_aryptr();
6287 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6288 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6289
6290 // for the quick and dirty code we will skip all the checks.
6291 // we are just trying to get the call to be generated.
6292 Node* src_start = src;
6293 Node* dest_start = dest;
6294 if (src_offset != NULL || dest_offset != NULL) {
6295 assert(src_offset != NULL && dest_offset != NULL, "");
6296 src_start = array_element_address(src, src_offset, T_BYTE);
6297 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6298 }
6299
6300 // now need to get the start of its expanded key array
6301 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6302 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6303 if (k_start == NULL) return false;
6304
6305 if (Matcher::pass_original_key_for_aes()) {
6306 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6307 // compatibility issues between Java key expansion and SPARC crypto instructions
6308 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6309 if (original_k_start == NULL) return false;
6310
6311 // Call the stub.
6312 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6313 stubAddr, stubName, TypePtr::BOTTOM,
6314 src_start, dest_start, k_start, original_k_start);
6315 } else {
6316 // Call the stub.
6317 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6318 stubAddr, stubName, TypePtr::BOTTOM,
6319 src_start, dest_start, k_start);
6320 }
6321
6322 return true;
6323 }
6324
6325 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6326 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6327 address stubAddr = NULL;
6328 const char *stubName = NULL;
6329
6330 assert(UseAES, "need AES instruction support");
6331
6332 switch(id) {
6333 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6334 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6335 stubName = "cipherBlockChaining_encryptAESCrypt";
6336 break;
6337 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6338 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6339 stubName = "cipherBlockChaining_decryptAESCrypt";
6340 break;
6341 default:
6342 break;
6343 }
6344 if (stubAddr == NULL) return false;
6345
6346 Node* cipherBlockChaining_object = argument(0);
6347 Node* src = argument(1);
6348 Node* src_offset = argument(2);
6349 Node* len = argument(3);
6350 Node* dest = argument(4);
6351 Node* dest_offset = argument(5);
6352
6353 // (1) src and dest are arrays.
6354 const Type* src_type = src->Value(&_gvn);
6355 const Type* dest_type = dest->Value(&_gvn);
6356 const TypeAryPtr* top_src = src_type->isa_aryptr();
6357 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6358 assert (top_src != NULL && top_src->klass() != NULL
6359 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6360
6361 // checks are the responsibility of the caller
6362 Node* src_start = src;
6363 Node* dest_start = dest;
6364 if (src_offset != NULL || dest_offset != NULL) {
6365 assert(src_offset != NULL && dest_offset != NULL, "");
6366 src_start = array_element_address(src, src_offset, T_BYTE);
6367 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6368 }
6369
6370 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6371 // (because of the predicated logic executed earlier).
6372 // so we cast it here safely.
6373 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6374
6375 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6376 if (embeddedCipherObj == NULL) return false;
6377
6378 // cast it to what we know it will be at runtime
6379 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6380 assert(tinst != NULL, "CBC obj is null");
6381 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6382 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6383 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6384
6385 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6386 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6387 const TypeOopPtr* xtype = aklass->as_instance_type();
6388 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6389 aescrypt_object = _gvn.transform(aescrypt_object);
6390
6391 // we need to get the start of the aescrypt_object's expanded key array
6392 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6393 if (k_start == NULL) return false;
6394
6395 // similarly, get the start address of the r vector
6396 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6397 if (objRvec == NULL) return false;
6398 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6399
6400 Node* cbcCrypt;
6401 if (Matcher::pass_original_key_for_aes()) {
6402 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6403 // compatibility issues between Java key expansion and SPARC crypto instructions
6404 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6405 if (original_k_start == NULL) return false;
6406
6407 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6408 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6409 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6410 stubAddr, stubName, TypePtr::BOTTOM,
6411 src_start, dest_start, k_start, r_start, len, original_k_start);
6412 } else {
6413 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6414 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6415 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6416 stubAddr, stubName, TypePtr::BOTTOM,
6417 src_start, dest_start, k_start, r_start, len);
6418 }
6419
6420 // return cipher length (int)
6421 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6422 set_result(retvalue);
6423 return true;
6424 }
6425
6426 //------------------------------inline_counterMode_AESCrypt-----------------------
6427 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6428 assert(UseAES, "need AES instruction support");
6429 if (!UseAESCTRIntrinsics) return false;
6430
6431 address stubAddr = NULL;
6432 const char *stubName = NULL;
6433 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6434 stubAddr = StubRoutines::counterMode_AESCrypt();
6435 stubName = "counterMode_AESCrypt";
6436 }
6437 if (stubAddr == NULL) return false;
6438
6439 Node* counterMode_object = argument(0);
6440 Node* src = argument(1);
6441 Node* src_offset = argument(2);
6442 Node* len = argument(3);
6443 Node* dest = argument(4);
6444 Node* dest_offset = argument(5);
6445
6446 // (1) src and dest are arrays.
6447 const Type* src_type = src->Value(&_gvn);
6448 const Type* dest_type = dest->Value(&_gvn);
6449 const TypeAryPtr* top_src = src_type->isa_aryptr();
6450 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6451 assert(top_src != NULL && top_src->klass() != NULL &&
6452 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6453
6454 // checks are the responsibility of the caller
6455 Node* src_start = src;
6456 Node* dest_start = dest;
6457 if (src_offset != NULL || dest_offset != NULL) {
6458 assert(src_offset != NULL && dest_offset != NULL, "");
6459 src_start = array_element_address(src, src_offset, T_BYTE);
6460 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6461 }
6462
6463 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6464 // (because of the predicated logic executed earlier).
6465 // so we cast it here safely.
6466 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6467 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6468 if (embeddedCipherObj == NULL) return false;
6469 // cast it to what we know it will be at runtime
6470 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6471 assert(tinst != NULL, "CTR obj is null");
6472 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6473 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6474 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6475 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6476 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6477 const TypeOopPtr* xtype = aklass->as_instance_type();
6478 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6479 aescrypt_object = _gvn.transform(aescrypt_object);
6480 // we need to get the start of the aescrypt_object's expanded key array
6481 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6482 if (k_start == NULL) return false;
6483 // similarly, get the start address of the r vector
6484 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6485 if (obj_counter == NULL) return false;
6486 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6487
6488 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6489 if (saved_encCounter == NULL) return false;
6490 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6491 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6492
6493 Node* ctrCrypt;
6494 if (Matcher::pass_original_key_for_aes()) {
6495 // no SPARC version for AES/CTR intrinsics now.
6496 return false;
6497 }
6498 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6499 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6500 OptoRuntime::counterMode_aescrypt_Type(),
6501 stubAddr, stubName, TypePtr::BOTTOM,
6502 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6503
6504 // return cipher length (int)
6505 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6506 set_result(retvalue);
6507 return true;
6508 }
6509
6510 //------------------------------get_key_start_from_aescrypt_object-----------------------
6511 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6512 #if defined(PPC64) || defined(S390)
6513 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6514 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6515 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6516 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6517 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6518 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6519 if (objSessionK == NULL) {
6520 return (Node *) NULL;
6521 }
6522 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6523 #else
6524 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6525 #endif // PPC64
6526 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6527 if (objAESCryptKey == NULL) return (Node *) NULL;
6528
6529 // now have the array, need to get the start address of the K array
6530 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6531 return k_start;
6532 }
6533
6534 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6535 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6536 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6537 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6538 if (objAESCryptKey == NULL) return (Node *) NULL;
6539
6540 // now have the array, need to get the start address of the lastKey array
6541 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6542 return original_k_start;
6543 }
6544
6545 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6546 // Return node representing slow path of predicate check.
6547 // the pseudo code we want to emulate with this predicate is:
6548 // for encryption:
6549 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6550 // for decryption:
6551 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6552 // note cipher==plain is more conservative than the original java code but that's OK
6553 //
6554 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6555 // The receiver was checked for NULL already.
6556 Node* objCBC = argument(0);
6557
6558 // Load embeddedCipher field of CipherBlockChaining object.
6559 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6560
6561 // get AESCrypt klass for instanceOf check
6562 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6563 // will have same classloader as CipherBlockChaining object
6564 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6565 assert(tinst != NULL, "CBCobj is null");
6566 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6567
6568 // we want to do an instanceof comparison against the AESCrypt class
6569 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6570 if (!klass_AESCrypt->is_loaded()) {
6571 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6572 Node* ctrl = control();
6573 set_control(top()); // no regular fast path
6574 return ctrl;
6575 }
6576 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6577
6578 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6579 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6580 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6581
6582 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6583
6584 // for encryption, we are done
6585 if (!decrypting)
6586 return instof_false; // even if it is NULL
6587
6588 // for decryption, we need to add a further check to avoid
6589 // taking the intrinsic path when cipher and plain are the same
6590 // see the original java code for why.
6591 RegionNode* region = new RegionNode(3);
6592 region->init_req(1, instof_false);
6593 Node* src = argument(1);
6594 Node* dest = argument(4);
6595 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6596 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6597 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6598 region->init_req(2, src_dest_conjoint);
6599
6600 record_for_igvn(region);
6601 return _gvn.transform(region);
6602 }
6603
6604 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6605 // Return node representing slow path of predicate check.
6606 // the pseudo code we want to emulate with this predicate is:
6607 // for encryption:
6608 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6609 // for decryption:
6610 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6611 // note cipher==plain is more conservative than the original java code but that's OK
6612 //
6613
6614 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6615 // The receiver was checked for NULL already.
6616 Node* objCTR = argument(0);
6617
6618 // Load embeddedCipher field of CipherBlockChaining object.
6619 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6620
6621 // get AESCrypt klass for instanceOf check
6622 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6623 // will have same classloader as CipherBlockChaining object
6624 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6625 assert(tinst != NULL, "CTRobj is null");
6626 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6627
6628 // we want to do an instanceof comparison against the AESCrypt class
6629 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6630 if (!klass_AESCrypt->is_loaded()) {
6631 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6632 Node* ctrl = control();
6633 set_control(top()); // no regular fast path
6634 return ctrl;
6635 }
6636
6637 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6638 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6639 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6640 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6641 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6642
6643 return instof_false; // even if it is NULL
6644 }
6645
6646 //------------------------------inline_ghash_processBlocks
6647 bool LibraryCallKit::inline_ghash_processBlocks() {
6648 address stubAddr;
6649 const char *stubName;
6650 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6651
6652 stubAddr = StubRoutines::ghash_processBlocks();
6653 stubName = "ghash_processBlocks";
6654
6655 Node* data = argument(0);
6656 Node* offset = argument(1);
6657 Node* len = argument(2);
6658 Node* state = argument(3);
6659 Node* subkeyH = argument(4);
6660
6661 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6662 assert(state_start, "state is NULL");
6663 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6664 assert(subkeyH_start, "subkeyH is NULL");
6665 Node* data_start = array_element_address(data, offset, T_BYTE);
6666 assert(data_start, "data is NULL");
6667
6668 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6669 OptoRuntime::ghash_processBlocks_Type(),
6670 stubAddr, stubName, TypePtr::BOTTOM,
6671 state_start, subkeyH_start, data_start, len);
6672 return true;
6673 }
6674
6675 //------------------------------inline_sha_implCompress-----------------------
6676 //
6677 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6678 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6679 //
6680 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6681 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6682 //
6683 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6684 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6685 //
6686 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6687 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6688
6689 Node* sha_obj = argument(0);
6690 Node* src = argument(1); // type oop
6691 Node* ofs = argument(2); // type int
6692
6693 const Type* src_type = src->Value(&_gvn);
6694 const TypeAryPtr* top_src = src_type->isa_aryptr();
6695 if (top_src == NULL || top_src->klass() == NULL) {
6696 // failed array check
6697 return false;
6698 }
6699 // Figure out the size and type of the elements we will be copying.
6700 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6701 if (src_elem != T_BYTE) {
6702 return false;
6703 }
6704 // 'src_start' points to src array + offset
6705 Node* src_start = array_element_address(src, ofs, src_elem);
6706 Node* state = NULL;
6707 address stubAddr;
6708 const char *stubName;
6709
6710 switch(id) {
6711 case vmIntrinsics::_sha_implCompress:
6712 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6713 state = get_state_from_sha_object(sha_obj);
6714 stubAddr = StubRoutines::sha1_implCompress();
6715 stubName = "sha1_implCompress";
6716 break;
6717 case vmIntrinsics::_sha2_implCompress:
6718 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6719 state = get_state_from_sha_object(sha_obj);
6720 stubAddr = StubRoutines::sha256_implCompress();
6721 stubName = "sha256_implCompress";
6722 break;
6723 case vmIntrinsics::_sha5_implCompress:
6724 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6725 state = get_state_from_sha5_object(sha_obj);
6726 stubAddr = StubRoutines::sha512_implCompress();
6727 stubName = "sha512_implCompress";
6728 break;
6729 default:
6730 fatal_unexpected_iid(id);
6731 return false;
6732 }
6733 if (state == NULL) return false;
6734
6735 // Call the stub.
6736 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6737 stubAddr, stubName, TypePtr::BOTTOM,
6738 src_start, state);
6739
6740 return true;
6741 }
6742
6743 //------------------------------inline_digestBase_implCompressMB-----------------------
6744 //
6745 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6746 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6747 //
6748 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6749 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6750 "need SHA1/SHA256/SHA512 instruction support");
6751 assert((uint)predicate < 3, "sanity");
6752 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6753
6754 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6755 Node* src = argument(1); // byte[] array
6756 Node* ofs = argument(2); // type int
6757 Node* limit = argument(3); // type int
6758
6759 const Type* src_type = src->Value(&_gvn);
6760 const TypeAryPtr* top_src = src_type->isa_aryptr();
6761 if (top_src == NULL || top_src->klass() == NULL) {
6762 // failed array check
6763 return false;
6764 }
6765 // Figure out the size and type of the elements we will be copying.
6766 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6767 if (src_elem != T_BYTE) {
6768 return false;
6769 }
6770 // 'src_start' points to src array + offset
6771 Node* src_start = array_element_address(src, ofs, src_elem);
6772
6773 const char* klass_SHA_name = NULL;
6774 const char* stub_name = NULL;
6775 address stub_addr = NULL;
6776 bool long_state = false;
6777
6778 switch (predicate) {
6779 case 0:
6780 if (UseSHA1Intrinsics) {
6781 klass_SHA_name = "sun/security/provider/SHA";
6782 stub_name = "sha1_implCompressMB";
6783 stub_addr = StubRoutines::sha1_implCompressMB();
6784 }
6785 break;
6786 case 1:
6787 if (UseSHA256Intrinsics) {
6788 klass_SHA_name = "sun/security/provider/SHA2";
6789 stub_name = "sha256_implCompressMB";
6790 stub_addr = StubRoutines::sha256_implCompressMB();
6791 }
6792 break;
6793 case 2:
6794 if (UseSHA512Intrinsics) {
6795 klass_SHA_name = "sun/security/provider/SHA5";
6796 stub_name = "sha512_implCompressMB";
6797 stub_addr = StubRoutines::sha512_implCompressMB();
6798 long_state = true;
6799 }
6800 break;
6801 default:
6802 fatal("unknown SHA intrinsic predicate: %d", predicate);
6803 }
6804 if (klass_SHA_name != NULL) {
6805 // get DigestBase klass to lookup for SHA klass
6806 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6807 assert(tinst != NULL, "digestBase_obj is not instance???");
6808 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6809
6810 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6811 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6812 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6813 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6814 }
6815 return false;
6816 }
6817 //------------------------------inline_sha_implCompressMB-----------------------
6818 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6819 bool long_state, address stubAddr, const char *stubName,
6820 Node* src_start, Node* ofs, Node* limit) {
6821 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6822 const TypeOopPtr* xtype = aklass->as_instance_type();
6823 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6824 sha_obj = _gvn.transform(sha_obj);
6825
6826 Node* state;
6827 if (long_state) {
6828 state = get_state_from_sha5_object(sha_obj);
6829 } else {
6830 state = get_state_from_sha_object(sha_obj);
6831 }
6832 if (state == NULL) return false;
6833
6834 // Call the stub.
6835 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6836 OptoRuntime::digestBase_implCompressMB_Type(),
6837 stubAddr, stubName, TypePtr::BOTTOM,
6838 src_start, state, ofs, limit);
6839 // return ofs (int)
6840 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6841 set_result(result);
6842
6843 return true;
6844 }
6845
6846 enum VectorApiObjectType {
6847 VECAPI_VECTOR,
6848 VECAPI_MASK,
6849 VECAPI_SPECIES,
6850 VECAPI_SHUFFLE,
6851 };
6852
6853 static ciKlass* get_exact_klass_for_vector_box(ciKlass* ctx, BasicType bt, int num_elem, VectorApiObjectType obj_type) {
6854 // FIXME: use VM symbol table instead
6855 ResourceMark rm;
6856 stringStream ss;
6857 ss.print_raw("jdk/incubator/vector/");
6858 switch (bt) {
6859 case T_BYTE: ss.print_raw("Byte"); break;
6860 case T_SHORT: ss.print_raw("Short"); break;
6861 case T_INT: ss.print_raw("Int"); break;
6862 case T_LONG: ss.print_raw("Long"); break;
6863 case T_FLOAT: ss.print_raw("Float"); break;
6864 case T_DOUBLE: ss.print_raw("Double"); break;
6865 default: fatal("unknown element type: %d", bt);
6866 }
6867 int bits = num_elem * BitsPerByte * type2aelembytes(bt);
6868 switch (bits) {
6869 case 64: // fall through
6870 case 128: // fall through
6871 case 256: // fall through
6872 case 512: ss.print("%d", bits); break;
6873 default: fatal("unknown vector size: %d", bits);
6874 }
6875 ss.print_raw("Vector");
6876 if (obj_type != VECAPI_VECTOR) {
6877 ss.print_raw("$");
6878 switch (bt) {
6879 case T_BYTE: ss.print_raw("Byte"); break;
6880 case T_SHORT: ss.print_raw("Short"); break;
6881 case T_INT: ss.print_raw("Int"); break;
6882 case T_LONG: ss.print_raw("Long"); break;
6883 case T_FLOAT: ss.print_raw("Float"); break;
6884 case T_DOUBLE: ss.print_raw("Double"); break;
6885 default: fatal("unknown element type: %d", bt);
6886 }
6887 switch (bits) {
6888 case 64: // fall through
6889 case 128: // fall through
6890 case 256: // fall through
6891 case 512: ss.print("%d", bits); break;
6892 default: fatal("unknown vector size: %d", bits);
6893 }
6894 switch (obj_type) {
6895 case VECAPI_MASK: ss.print_raw("Mask"); break;
6896 case VECAPI_SPECIES: ss.print_raw("Species"); break;
6897 case VECAPI_SHUFFLE: ss.print_raw("Shuffle"); break;
6898 default: fatal("unknown vector object type");
6899 }
6900 }
6901
6902 ciSymbol* vector_klass_name = ciSymbol::make(ss.as_string());
6903 ciKlass* vector_klass = ctx->find_klass(vector_klass_name);
6904 return vector_klass;
6905 }
6906
6907 static BasicType getMaskBasicType(BasicType use_type) {
6908 if (use_type == T_FLOAT) {
6909 return T_INT;
6910 }
6911 if (use_type == T_DOUBLE) {
6912 return T_LONG;
6913 }
6914 assert(is_java_primitive(use_type), "must be java primitive");
6915 return use_type;
6916 }
6917
6918 enum VectorMaskUseType {
6919 VecMaskUseLoad,
6920 VecMaskUseStore,
6921 VecMaskUseAll,
6922 VecMaskNotUsed
6923 };
6924
6925 static bool arch_supports_vector(int op, int num_elem, BasicType type, VectorMaskUseType mask_use_type) {
6926 // Check that the operation is valid.
6927 if (op <= 0) {
6928 #ifndef PRODUCT
6929 if (DebugVectorApi) {
6930 tty->print_cr("Rejected intrinsification because no valid vector op could be extracted");
6931 }
6932 #endif
6933 return false;
6934 }
6935
6936 // Check that architecture supports this op-size-type combination.
6937 if (!Matcher::match_rule_supported_vector(op, num_elem, type)) {
6938 #ifndef PRODUCT
6939 if (DebugVectorApi) {
6940 tty->print_cr("Rejected vector op (%s,%s,%d) because architecture does not support it",
6941 NodeClassNames[op], type2name(type), num_elem);
6942 }
6943 #endif
6944 return false;
6945 } else {
6946 assert(Matcher::match_rule_supported(op), "must be supported");
6947 }
6948
6949 // Check whether mask unboxing is supported.
6950 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseLoad) {
6951 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type)) {
6952 #ifndef PRODUCT
6953 if (DebugVectorApi) {
6954 tty->print_cr("Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
6955 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
6956 }
6957 #endif
6958 return false;
6959 }
6960 }
6961
6962 // Check whether mask boxing is supported.
6963 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseStore) {
6964 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type)) {
6965 #ifndef PRODUCT
6966 if (DebugVectorApi) {
6967 tty->print_cr("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
6968 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
6969 }
6970 #endif
6971 return false;
6972 }
6973 }
6974
6975 return true;
6976 }
6977
6978 // Should be aligned with constants in jdk.incubator.vector.VectorIntrinsics.
6979 enum {
6980 // Unary
6981 OP_ABS = 0,
6982 OP_NEG = 1,
6983 OP_SQRT = 2,
6984 OP_NOT = 3,
6985 // Binary
6986 OP_ADD = 4,
6987 OP_SUB = 5,
6988 OP_MUL = 6,
6989 OP_DIV = 7,
6990 OP_AND = 8,
6991 OP_OR = 9,
6992 OP_XOR = 10,
6993 // Ternary
6994 OP_FMA = 11,
6995 // Broadcast int
6996 OP_LSHIFT = 12,
6997 OP_RSHIFT = 13,
6998 OP_URSHIFT = 14
6999 };
7000
7001 static int get_opc(jint op, BasicType bt) {
7002 switch (op) {
7003 case OP_ADD: {
7004 switch (bt) {
7005 case T_BYTE: // fall-through
7006 case T_SHORT: // fall-through
7007 case T_INT: return Op_AddI;
7008 case T_LONG: return Op_AddL;
7009 case T_FLOAT: return Op_AddF;
7010 case T_DOUBLE: return Op_AddD;
7011 default: fatal("ADD: %s", type2name(bt));
7012 }
7013 break;
7014 }
7015 case OP_SUB: {
7016 switch (bt) {
7017 case T_BYTE: // fall-through
7018 case T_SHORT: // fall-through
7019 case T_INT: return Op_SubI;
7020 case T_LONG: return Op_SubL;
7021 case T_FLOAT: return Op_SubF;
7022 case T_DOUBLE: return Op_SubD;
7023 default: fatal("SUB: %s", type2name(bt));
7024 }
7025 break;
7026 }
7027 case OP_MUL: {
7028 switch (bt) {
7029 case T_BYTE: // fall-through
7030 case T_SHORT: // fall-through
7031 case T_INT: return Op_MulI;
7032 case T_LONG: return Op_MulL;
7033 case T_FLOAT: return Op_MulF;
7034 case T_DOUBLE: return Op_MulD;
7035 default: fatal("MUL: %s", type2name(bt));
7036 }
7037 break;
7038 }
7039 case OP_DIV: {
7040 switch (bt) {
7041 case T_BYTE: // fall-through
7042 case T_SHORT: // fall-through
7043 case T_INT: return Op_DivI;
7044 case T_LONG: return Op_DivL;
7045 case T_FLOAT: return Op_DivF;
7046 case T_DOUBLE: return Op_DivD;
7047 default: fatal("DIV: %s", type2name(bt));
7048 }
7049 break;
7050 }
7051 case OP_ABS: {
7052 switch (bt) {
7053 case T_BYTE: // fall-through
7054 case T_SHORT: // fall-through
7055 case T_INT: return Op_AbsI;
7056 case T_FLOAT: return Op_AbsF;
7057 case T_DOUBLE: return Op_AbsD;
7058 default: fatal("ABS: %s", type2name(bt));
7059 }
7060 break;
7061 }
7062 case OP_NEG: {
7063 switch (bt) {
7064 case T_BYTE: // fall-through
7065 case T_SHORT: // fall-through
7066 case T_INT: return Op_NegI;
7067 case T_FLOAT: return Op_NegF;
7068 case T_DOUBLE: return Op_NegD;
7069 default: fatal("NEG: %s", type2name(bt));
7070 }
7071 break;
7072 }
7073 case OP_AND: {
7074 switch (bt) {
7075 case T_BYTE: // fall-through
7076 case T_SHORT: // fall-through
7077 case T_INT: return Op_AndI;
7078 case T_LONG: return Op_AndL;
7079 default: fatal("AND: %s", type2name(bt));
7080 }
7081 break;
7082 }
7083 case OP_OR: {
7084 switch (bt) {
7085 case T_BYTE: // fall-through
7086 case T_SHORT: // fall-through
7087 case T_INT: return Op_OrI;
7088 case T_LONG: return Op_OrL;
7089 default: fatal("OR: %s", type2name(bt));
7090 }
7091 break;
7092 }
7093 case OP_XOR: {
7094 switch (bt) {
7095 case T_BYTE: // fall-through
7096 case T_SHORT: // fall-through
7097 case T_INT: return Op_XorI;
7098 case T_LONG: return Op_XorL;
7099 default: fatal("XOR: %s", type2name(bt));
7100 }
7101 break;
7102 }
7103 case OP_SQRT: {
7104 switch (bt) {
7105 case T_FLOAT: return Op_SqrtF;
7106 case T_DOUBLE: return Op_SqrtD;
7107 default: fatal("SQRT: %s", type2name(bt));
7108 }
7109 break;
7110 }
7111 case OP_NOT: {
7112 switch (bt) {
7113 case T_BYTE: // fall-through
7114 case T_SHORT: // fall-through
7115 case T_INT: // fall-through
7116 case T_LONG: return Op_Not;
7117 default: fatal("NOT: %s", type2name(bt));
7118 }
7119 break;
7120 }
7121 case OP_FMA: {
7122 switch (bt) {
7123 case T_FLOAT: return Op_FmaF;
7124 case T_DOUBLE: return Op_FmaD;
7125 default: fatal("FMA: %s", type2name(bt));
7126 }
7127 break;
7128 }
7129 case OP_LSHIFT: {
7130 switch (bt) {
7131 case T_BYTE: // fall-through
7132 case T_SHORT: // fall-through
7133 case T_INT: return Op_LShiftI;
7134 case T_LONG: return Op_LShiftL;
7135 default: fatal("LSHIFT: %s", type2name(bt));
7136 }
7137 }
7138 case OP_RSHIFT: {
7139 switch (bt) {
7140 case T_BYTE: // fall-through
7141 case T_SHORT: // fall-through
7142 case T_INT: return Op_RShiftI;
7143 case T_LONG: return Op_RShiftL;
7144 default: fatal("RSHIFT: %s", type2name(bt));
7145 }
7146 }
7147 case OP_URSHIFT: {
7148 switch (bt) {
7149 case T_BYTE: // fall-through
7150 case T_SHORT: // fall-through
7151 case T_INT: return Op_URShiftI;
7152 case T_LONG: return Op_URShiftL;
7153 default: fatal("URSHIFT: %s", type2name(bt));
7154 }
7155 }
7156 default: fatal("unknown op: %d", op);
7157 }
7158 return 0; // Unimplemented
7159 }
7160
7161 Node* LibraryCallKit::box_vector(Node* vector, const TypeInstPtr* vbox_type,
7162 BasicType elem_bt, int num_elem) {
7163
7164 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
7165
7166 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
7167 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
7168 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true);
7169 set_i_o(_gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
7170 set_all_memory(_gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
7171 Node* ret = _gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
7172
7173 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vec_type);
7174 return _gvn.transform(vbox);
7175 }
7176
7177 Node* LibraryCallKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) {
7178 const TypeInstPtr* vbox_type_v = gvn().type(v)->is_instptr();
7179 if (vbox_type->klass() != vbox_type_v->klass()) {
7180 return NULL; // arguments don't agree on vector shapes
7181 }
7182 if (vbox_type_v->maybe_null()) {
7183 return NULL; // no nulls are allowed
7184 }
7185 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
7186 Node* unbox = gvn().transform(new VectorUnboxNode(C, vec_type, v, merged_memory()));
7187 return unbox;
7188 }
7189
7190 bool LibraryCallKit::inline_vector_nary_operation(int n) {
7191 const TypeInt* opr = gvn().type(argument(0))->is_int();
7192 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7193 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7194 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7195
7196 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7197 return false; // not enough info for intrinsification
7198 }
7199 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7200 if (!elem_type->is_primitive_type()) {
7201 return false; // should be primitive type
7202 }
7203 BasicType elem_bt = elem_type->basic_type();
7204 int num_elem = vlen->get_con();
7205 int opc = get_opc(opr->get_con(), elem_bt);
7206 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
7207 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7208 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7209
7210 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7211 if (!arch_supports_vector(sopc, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseAll : VecMaskNotUsed)) {
7212 return false; // not supported
7213 }
7214
7215 Node* opd1 = NULL; Node* opd2 = NULL; Node* opd3 = NULL;
7216 switch (n) {
7217 case 3: {
7218 opd1 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7219 if (opd1 == NULL) {
7220 return false;
7221 }
7222 // fall-through
7223 }
7224 case 2: {
7225 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7226 if (opd2 == NULL) {
7227 return false;
7228 }
7229 // fall-through
7230 }
7231 case 1: {
7232 opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7233 if (opd1 == NULL) {
7234 return false;
7235 }
7236 break;
7237 }
7238 default: fatal("unsupported arity: %d", n);
7239 }
7240
7241 Node* operation = NULL;
7242 switch (n) {
7243 case 1:
7244 case 2: {
7245 operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, num_elem, elem_bt));
7246 break;
7247 }
7248 case 3: {
7249 operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, opd3, num_elem, elem_bt));
7250 break;
7251 }
7252 default: fatal("unsupported arity: %d", n);
7253 }
7254 // Wrap it up in VectorBox to keep object type information.
7255 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
7256 set_result(operation);
7257
7258 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7259 return true;
7260 }
7261
7262 // <V extends Vector<?,?>>
7263 // V broadcastCoerced(Class<?> vectorClass, Class<?> elementType, int vlen,
7264 // long bits,
7265 // LongFunction<V> defaultImpl)
7266 bool LibraryCallKit::inline_vector_broadcast_coerced() {
7267 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7268 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7269 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7270
7271 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7272 return false; // not enough info for intrinsification
7273 }
7274
7275 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7276 if (!elem_type->is_primitive_type()) {
7277 return false; // should be primitive type
7278 }
7279 BasicType elem_bt = elem_type->basic_type();
7280 int num_elem = vlen->get_con();
7281 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7282 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7283
7284 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7285 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt,
7286 vbox_klass->is_vectormask() ? VecMaskUseStore : VecMaskNotUsed)) {
7287 return false; // not supported
7288 }
7289
7290 Node* bits = argument(3); // long
7291
7292 Node* elem = NULL;
7293 switch (elem_bt) {
7294 case T_BOOLEAN: // fall-through
7295 case T_BYTE: // fall-through
7296 case T_SHORT: // fall-through
7297 case T_CHAR: // fall-through
7298 case T_INT: {
7299 elem = gvn().transform(new ConvL2INode(bits));
7300 break;
7301 }
7302 case T_DOUBLE: {
7303 elem = gvn().transform(new MoveL2DNode(bits));
7304 break;
7305 }
7306 case T_FLOAT: {
7307 bits = gvn().transform(new ConvL2INode(bits));
7308 elem = gvn().transform(new MoveI2FNode(bits));
7309 break;
7310 }
7311 case T_LONG: {
7312 elem = bits; // no conversion needed
7313 break;
7314 }
7315 default: fatal("%s", type2name(elem_bt));
7316 }
7317
7318 Node* broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt));
7319 broadcast = gvn().transform(broadcast);
7320 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
7321 set_result(box);
7322
7323 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7324 return true;
7325 }
7326
7327 // <V extends Vector<?,?>>
7328 // V load(Class<?> vectorClass, Class<?> elementType, int vlen,
7329 // Object array, int index, /* Vector.Mask<E,S> m*/
7330 // BiFunction<Object, Integer, V> defaultImpl) {
7331 //
7332 // <V extends Vector<?,?>>
7333 // void store(Class<?> vectorClass, Class<?> elementType, int vlen,
7334 // Object array, int index, V v, /*Vector.Mask<E,S> m*/
7335 // StoreVectorOperation<V> defaultImpl) {
7336
7337 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
7338 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7339 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7340 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7341
7342 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7343 return false; // not enough info for intrinsification
7344 }
7345
7346 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7347 if (!elem_type->is_primitive_type()) {
7348 return false; // should be primitive type
7349 }
7350 BasicType elem_bt = elem_type->basic_type();
7351 int num_elem = vlen->get_con();
7352
7353 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7354 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
7355 return false; // not supported
7356 }
7357
7358 Node* arr = argument(3);
7359 Node* idx = argument(4);
7360
7361 const TypeAryPtr* arr_type = gvn().type(arr)->isa_aryptr();
7362 if (arr_type == NULL) {
7363 return false; // should be an array
7364 }
7365 if (elem_bt != arr_type->elem()->array_element_basic_type()) {
7366 return false; // array & vector element types should be the same
7367 }
7368 Node* adr = array_element_address(arr, idx, elem_bt);
7369 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
7370
7371 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7372 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7373
7374 if (is_store) {
7375 Node* val = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7376 if (val == NULL) {
7377 return false; // operand unboxing failed
7378 }
7379 set_all_memory(reset_memory());
7380 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(adr), adr, adr_type, val, num_elem));
7381 set_memory(vstore, adr_type);
7382 } else {
7383 Node* vload = gvn().transform(LoadVectorNode::make(0, control(), memory(adr), adr, adr_type, num_elem, elem_bt));
7384 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7385 set_result(box);
7386 }
7387
7388 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7389 return true;
7390 }
7391
7392 // <V extends Vector<?,?>>
7393 // long reductionCoerced(int oprId, Class<?> vectorClass, Class<?> elementType, int vlen,
7394 // V v,
7395 // Function<V,Long> defaultImpl)
7396
7397 bool LibraryCallKit::inline_vector_reduction() {
7398 const TypeInt* opr = gvn().type(argument(0))->is_int();
7399 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7400 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7401 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7402
7403 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7404 return false; // not enough info for intrinsification
7405 }
7406 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7407 if (!elem_type->is_primitive_type()) {
7408 return false; // should be primitive type
7409 }
7410 BasicType elem_bt = elem_type->basic_type();
7411 int num_elem = vlen->get_con();
7412
7413 int opc = get_opc(opr->get_con(), elem_bt);
7414 int sopc = ReductionNode::opcode(opc, elem_bt);
7415
7416 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7417 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7418 return false;
7419 }
7420
7421 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7422 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7423
7424 Node* opd = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7425 if (opd == NULL) {
7426 return false; // operand unboxing failed
7427 }
7428
7429 Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt);
7430 Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, elem_bt));
7431
7432 Node* bits = NULL;
7433 switch (elem_bt) {
7434 case T_INT: {
7435 bits = gvn().transform(new ConvI2LNode(rn));
7436 break;
7437 }
7438 case T_FLOAT: {
7439 rn = gvn().transform(new MoveF2INode(rn));
7440 bits = gvn().transform(new ConvI2LNode(rn));
7441 break;
7442 }
7443 case T_DOUBLE: {
7444 bits = gvn().transform(new MoveD2LNode(rn));
7445 break;
7446 }
7447 case T_LONG: {
7448 bits = rn; // no conversion needed
7449 break;
7450 }
7451 default: fatal("%s", type2name(elem_bt));
7452 }
7453 set_result(bits);
7454 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7455 return true;
7456 }
7457
7458 // static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
7459 // V v1, V v2,
7460 // BiFunction<V, V, Boolean> defaultImpl) {
7461
7462 bool LibraryCallKit::inline_vector_test() {
7463 const TypeInt* cond = gvn().type(argument(0))->is_int();
7464 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7465 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7466 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7467
7468 if (!cond->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7469 return false; // not enough info for intrinsification
7470 }
7471 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7472 if (!elem_type->is_primitive_type()) {
7473 return false; // should be primitive type
7474 }
7475 BasicType elem_bt = elem_type->basic_type();
7476 int num_elem = vlen->get_con();
7477 Assembler::Condition booltest = (Assembler::Condition)cond->get_con();
7478 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7479 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7480
7481 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseLoad : VecMaskNotUsed)) {
7482 return false;
7483 }
7484
7485 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7486 Node* opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7487 if (opd1 == NULL || opd2 == NULL) {
7488 return false; // operand unboxing failed
7489 }
7490 Node* test = new VectorTestNode(opd1, opd2, booltest);
7491 test = _gvn.transform(test);
7492 set_result(test);
7493 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7494 return true;
7495 }
7496
7497 // static
7498 // <V extends Vector, M extends Mask>
7499 // V blend(Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7500 // V v1, V v2, M m,
7501 // VectorBlendOp<V,M> defaultImpl) { ...
7502 //
7503 bool LibraryCallKit::inline_vector_blend() {
7504 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7505 const TypeInstPtr* mask_klass = gvn().type(argument(1))->is_instptr();
7506 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7507 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7508
7509 if (mask_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7510 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7511 return false; // not enough info for intrinsification
7512 }
7513 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7514 if (!elem_type->is_primitive_type()) {
7515 return false; // should be primitive type
7516 }
7517 BasicType elem_bt = elem_type->basic_type();
7518 BasicType mask_bt = getMaskBasicType(elem_bt);
7519 int num_elem = vlen->get_con();
7520
7521 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
7522 return false; // not supported
7523 }
7524 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7525 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7526
7527 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7528 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7529
7530 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7531 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7532 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
7533
7534 if (v1 == NULL || v2 == NULL || mask == NULL) {
7535 return false; // operand unboxing failed
7536 }
7537
7538 Node* blend = _gvn.transform(new VectorBlendNode(v1, v2, mask));
7539 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
7540 set_result(box);
7541
7542 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7543 return true;
7544 }
7545
7546 // static <V extends Vector<E,S>,
7547 // M extends Vector.Mask<E,S>,
7548 // S extends Vector.Shape, E>
7549 // M compare(int cond, Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7550 // V v1, V v2,
7551 // VectorCompareOp<V,M> defaultImpl) { ...
7552 //
7553 bool LibraryCallKit::inline_vector_compare() {
7554 const TypeInt* cond = gvn().type(argument(0))->is_int();
7555 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7556 const TypeInstPtr* mask_klass = gvn().type(argument(2))->is_instptr();
7557 const TypeInstPtr* elem_klass = gvn().type(argument(3))->is_instptr();
7558 const TypeInt* vlen = gvn().type(argument(4))->is_int();
7559
7560 if (!cond->is_con() || vector_klass->const_oop() == NULL || mask_klass->const_oop() == NULL ||
7561 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7562 return false; // not enough info for intrinsification
7563 }
7564 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7565 if (!elem_type->is_primitive_type()) {
7566 return false; // should be primitive type
7567 }
7568
7569 int num_elem = vlen->get_con();
7570 BasicType elem_bt = elem_type->basic_type();
7571 BasicType mask_bt = getMaskBasicType(elem_bt);
7572
7573 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
7574 return false;
7575 }
7576
7577 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7578 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7579
7580 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7581 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7582
7583 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7584 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7585
7586 if (v1 == NULL || v2 == NULL) {
7587 return false; // operand unboxing failed
7588 }
7589 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
7590 const TypeVect* vt = TypeVect::make(mask_bt, num_elem);
7591 Node* operation = _gvn.transform(new VectorMaskCmpNode(pred, v1, v2, vt));
7592
7593 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
7594 set_result(box);
7595
7596 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7597 return true;
7598 }
7599
7600 Node* LibraryCallKit::shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
7601 assert(bt == T_INT || bt == T_LONG, "Only long and int are supported");
7602 juint mask = (bt == T_INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
7603 const TypeInt* t = cnt->find_int_type();
7604 if (t != NULL && t->is_con()) {
7605 juint shift = t->get_con();
7606 if (shift > mask) {
7607 return _gvn.transform(ConNode::make(TypeInt::make(shift & mask)));
7608 } else {
7609 return cnt;
7610 }
7611 } else {
7612 Node* maskedcnt = cnt;
7613 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
7614 Node* nmask = _gvn.transform(ConNode::make(TypeInt::make(mask)));
7615 maskedcnt = _gvn.transform(new AndINode(cnt, nmask));
7616 }
7617 return VectorNode::shift_count(shift_op, maskedcnt, num_elem, bt);
7618 }
7619 }
7620
7621 // static
7622 // <V extends Vector<?,?>>
7623 // V broadcastInt(int opr, Class<V> vectorClass, Class<?> elementType, int vlen,
7624 // V v, int i,
7625 // VectorBroadcastIntOp<V> defaultImpl) {
7626 //
7627 bool LibraryCallKit::inline_vector_broadcast_int() {
7628 const TypeInt* opr = gvn().type(argument(0))->is_int();
7629 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7630 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7631 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7632
7633 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7634 return false; // not enough info for intrinsification
7635 }
7636 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7637 if (!elem_type->is_primitive_type()) {
7638 return false; // should be primitive type
7639 }
7640 BasicType elem_bt = elem_type->basic_type();
7641 int num_elem = vlen->get_con();
7642 int opc = get_opc(opr->get_con(), elem_bt);
7643 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
7644 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7645 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7646
7647 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7648 return false; // not supported
7649 }
7650 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7651 Node* opd2 = shift_count(argument(5), opc, elem_bt, num_elem);
7652 if (opd1 == NULL || opd2 == NULL) {
7653 return false;
7654 }
7655 Node* operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, num_elem, elem_bt));
7656
7657 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
7658 set_result(vbox);
7659 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7660 return true;
7661 }
7662
7663 static BasicType extract_bt_from_box_class(ciType* t) {
7664 if (t->basic_type() == T_OBJECT) {
7665 switch(t->as_klass()->name()->sid()) {
7666 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Boolean): return T_BOOLEAN;
7667 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Byte): return T_BYTE;
7668 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Short): return T_SHORT;
7669 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Character): return T_CHAR;
7670 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Integer): return T_INT;
7671 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Long): return T_LONG;
7672 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Float): return T_FLOAT;
7673 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Double): return T_DOUBLE;
7674 default: fatal("unknown type: %s", t->as_klass()->name()->as_utf8());
7675 }
7676 }
7677 return T_ILLEGAL;
7678 }
7679
7680 bool LibraryCallKit::inline_vector_rebracket() {
7681 const TypeInstPtr* vector_klass_from = gvn().type(argument(0))->is_instptr();
7682 const TypeInstPtr* elem_klass_from = gvn().type(argument(1))->is_instptr();
7683 const TypeInt* vlen_from = gvn().type(argument(2))->is_int();
7684
7685 const TypeInstPtr* elem_klass_to = gvn().type(argument(3))->is_instptr();
7686
7687 if (vector_klass_from->const_oop() == NULL || elem_klass_from->const_oop() == NULL || !vlen_from->is_con() ||
7688 elem_klass_to->const_oop() == NULL ) {
7689 return false; // not enough info for intrinsification
7690 }
7691
7692 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
7693 if (!vbox_klass_from->is_vectorapi_vector()) {
7694 return false; // only vector & mask are supported
7695 }
7696 bool is_mask = vbox_klass_from->is_vectormask();
7697
7698 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
7699 if (!elem_type_from->is_primitive_type()) {
7700 return false; // should be primitive type
7701 }
7702 BasicType elem_bt_from = elem_type_from->basic_type();
7703 if (is_mask) {
7704 elem_bt_from = getMaskBasicType(elem_bt_from);
7705 }
7706 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
7707 BasicType elem_bt_to = extract_bt_from_box_class(elem_type_to);
7708 if (elem_bt_to == T_ILLEGAL) {
7709 return false;
7710 }
7711 if (is_mask) {
7712 elem_bt_to = getMaskBasicType(elem_bt_to);
7713 }
7714 if (is_mask && elem_bt_from != elem_bt_to) {
7715 return false; // type mismatch
7716 }
7717 int num_elem_from = vlen_from->get_con();
7718
7719 int num_elem_to = (type2aelembytes(elem_bt_from) * num_elem_from) / type2aelembytes(elem_bt_to);
7720 if (num_elem_to == 0) {
7721 return false;
7722 }
7723 if (is_mask) {
7724 elem_bt_to = getMaskBasicType(elem_bt_to);
7725 }
7726
7727 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
7728
7729 Node* opd1 = unbox_vector(argument(4), vbox_type_from, elem_bt_from, num_elem_from);
7730 if (opd1 == NULL) {
7731 return false;
7732 }
7733
7734 // Can assert when Phi merges vectors of different types:
7735 // # Internal Error (/Users/vlivanov/ws/jdk/panama-dev/open/src/hotspot/share/opto/type.cpp:2291), pid=67536, tid=14083
7736 // # Error: assert(length() == v->length()) failed
7737 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from);
7738 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to);
7739 Node* op = opd1;
7740 if (src_type != dst_type) {
7741 op = _gvn.transform(new VectorReinterpretNode(op, src_type, dst_type));
7742 }
7743 ciKlass* vbox_klass_to = get_exact_klass_for_vector_box(vbox_klass_from, extract_bt_from_box_class(elem_type_to),
7744 num_elem_to, is_mask ? VECAPI_MASK : VECAPI_VECTOR);
7745 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
7746 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
7747 set_result(vbox);
7748 return true;
7749 }
7750
7751 //------------------------------get_state_from_sha_object-----------------------
7752 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
7753 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
7754 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
7755 if (sha_state == NULL) return (Node *) NULL;
7756
7757 // now have the array, need to get the start address of the state array
7758 Node* state = array_element_address(sha_state, intcon(0), T_INT);
7759 return state;
7760 }
7761
7762 //------------------------------get_state_from_sha5_object-----------------------
7763 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
7764 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
7765 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
7766 if (sha_state == NULL) return (Node *) NULL;
7767
7768 // now have the array, need to get the start address of the state array
7769 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
7770 return state;
7771 }
7772
7773 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
7774 // Return node representing slow path of predicate check.
7775 // the pseudo code we want to emulate with this predicate is:
7776 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
7777 //
7778 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
7779 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
7780 "need SHA1/SHA256/SHA512 instruction support");
7781 assert((uint)predicate < 3, "sanity");
7782
7783 // The receiver was checked for NULL already.
7784 Node* digestBaseObj = argument(0);
7785
7786 // get DigestBase klass for instanceOf check
7787 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
7788 assert(tinst != NULL, "digestBaseObj is null");
7789 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
7790
7791 const char* klass_SHA_name = NULL;
7792 switch (predicate) {
7793 case 0:
7794 if (UseSHA1Intrinsics) {
7795 // we want to do an instanceof comparison against the SHA class
7796 klass_SHA_name = "sun/security/provider/SHA";
7797 }
7798 break;
7799 case 1:
7800 if (UseSHA256Intrinsics) {
7801 // we want to do an instanceof comparison against the SHA2 class
7802 klass_SHA_name = "sun/security/provider/SHA2";
7803 }
7804 break;
7805 case 2:
7806 if (UseSHA512Intrinsics) {
7807 // we want to do an instanceof comparison against the SHA5 class
7808 klass_SHA_name = "sun/security/provider/SHA5";
7809 }
7810 break;
7811 default:
7812 fatal("unknown SHA intrinsic predicate: %d", predicate);
7813 }
7814
7815 ciKlass* klass_SHA = NULL;
7816 if (klass_SHA_name != NULL) {
7817 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
7818 }
7819 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
7820 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
7821 Node* ctrl = control();
7822 set_control(top()); // no intrinsic path
7823 return ctrl;
7824 }
7825 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
7826
7827 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
7828 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
7829 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
7830 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
7831
7832 return instof_false; // even if it is NULL
7833 }
7834
7835 //-------------inline_fma-----------------------------------
7836 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
7837 Node *a = NULL;
7838 Node *b = NULL;
7839 Node *c = NULL;
7840 Node* result = NULL;
7841 switch (id) {
7842 case vmIntrinsics::_fmaD:
7843 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
7844 // no receiver since it is static method
7845 a = round_double_node(argument(0));
7846 b = round_double_node(argument(2));
7847 c = round_double_node(argument(4));
7848 result = _gvn.transform(new FmaDNode(control(), a, b, c));
7849 break;
7850 case vmIntrinsics::_fmaF:
7851 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
7852 a = argument(0);
7853 b = argument(1);
7854 c = argument(2);
7855 result = _gvn.transform(new FmaFNode(control(), a, b, c));
7856 break;
7857 default:
7858 fatal_unexpected_iid(id); break;
7859 }
7860 set_result(result);
7861 return true;
7862 }
7863
7864 bool LibraryCallKit::inline_profileBoolean() {
7865 Node* counts = argument(1);
7866 const TypeAryPtr* ary = NULL;
7867 ciArray* aobj = NULL;
7868 if (counts->is_Con()
7869 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
7870 && (aobj = ary->const_oop()->as_array()) != NULL
7871 && (aobj->length() == 2)) {
7872 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
7873 jint false_cnt = aobj->element_value(0).as_int();
7874 jint true_cnt = aobj->element_value(1).as_int();
7875
7876 if (C->log() != NULL) {
7877 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
7878 false_cnt, true_cnt);
7879 }
7880
7881 if (false_cnt + true_cnt == 0) {
7882 // According to profile, never executed.
7883 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7884 Deoptimization::Action_reinterpret);
7885 return true;
7886 }
7887
7888 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
7889 // is a number of each value occurrences.
7890 Node* result = argument(0);
7891 if (false_cnt == 0 || true_cnt == 0) {
7892 // According to profile, one value has been never seen.
7893 int expected_val = (false_cnt == 0) ? 1 : 0;
7894
7895 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
7896 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
7897
7898 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
7899 Node* fast_path = _gvn.transform(new IfTrueNode(check));
7900 Node* slow_path = _gvn.transform(new IfFalseNode(check));
7901
7902 { // Slow path: uncommon trap for never seen value and then reexecute
7903 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
7904 // the value has been seen at least once.
7905 PreserveJVMState pjvms(this);
7906 PreserveReexecuteState preexecs(this);
7907 jvms()->set_should_reexecute(true);
7908
7909 set_control(slow_path);
7910 set_i_o(i_o());
7911
7912 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7913 Deoptimization::Action_reinterpret);
7914 }
7915 // The guard for never seen value enables sharpening of the result and
7916 // returning a constant. It allows to eliminate branches on the same value
7917 // later on.
7918 set_control(fast_path);
7919 result = intcon(expected_val);
7920 }
7921 // Stop profiling.
7922 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
7923 // By replacing method body with profile data (represented as ProfileBooleanNode
7924 // on IR level) we effectively disable profiling.
7925 // It enables full speed execution once optimized code is generated.
7926 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
7927 C->record_for_igvn(profile);
7928 set_result(profile);
7929 return true;
7930 } else {
7931 // Continue profiling.
7932 // Profile data isn't available at the moment. So, execute method's bytecode version.
7933 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
7934 // is compiled and counters aren't available since corresponding MethodHandle
7935 // isn't a compile-time constant.
7936 return false;
7937 }
7938 }
7939
7940 bool LibraryCallKit::inline_isCompileConstant() {
7941 Node* n = argument(0);
7942 set_result(n->is_Con() ? intcon(1) : intcon(0));
7943 return true;
7944 }