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_reinterpret();
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 void set_vector_result(Node* result, bool set_res = true);
347
348 void clear_upper_avx() {
349 #ifdef X86
350 if (UseAVX >= 2) {
351 C->set_clear_upper_avx(true);
352 }
353 #endif
354 }
355 };
356
357 //---------------------------make_vm_intrinsic----------------------------
358 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
359 vmIntrinsics::ID id = m->intrinsic_id();
360 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
361
362 if (!m->is_loaded()) {
363 // Do not attempt to inline unloaded methods.
364 return NULL;
365 }
366
367 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
368 bool is_available = false;
369
370 {
371 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
372 // the compiler must transition to '_thread_in_vm' state because both
373 // methods access VM-internal data.
374 VM_ENTRY_MARK;
375 methodHandle mh(THREAD, m->get_Method());
376 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
377 !C->directive()->is_intrinsic_disabled(mh) &&
378 !vmIntrinsics::is_disabled_by_flags(mh);
379
380 }
381
382 if (is_available) {
383 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
384 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
385 return new LibraryIntrinsic(m, is_virtual,
386 vmIntrinsics::predicates_needed(id),
387 vmIntrinsics::does_virtual_dispatch(id),
388 (vmIntrinsics::ID) id);
389 } else {
390 return NULL;
391 }
392 }
393
394 //----------------------register_library_intrinsics-----------------------
395 // Initialize this file's data structures, for each Compile instance.
396 void Compile::register_library_intrinsics() {
397 // Nothing to do here.
398 }
399
400 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
401 LibraryCallKit kit(jvms, this);
402 Compile* C = kit.C;
403 int nodes = C->unique();
404 #ifndef PRODUCT
405 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
406 char buf[1000];
407 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
408 tty->print_cr("Intrinsic %s", str);
409 }
410 #endif
411 ciMethod* callee = kit.callee();
412 const int bci = kit.bci();
413
414 // Try to inline the intrinsic.
415 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
416 kit.try_to_inline(_last_predicate)) {
417 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
418 : "(intrinsic)";
419 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
420 if (C->print_intrinsics() || C->print_inlining()) {
421 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
422 }
423 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
424 if (C->log()) {
425 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
426 vmIntrinsics::name_at(intrinsic_id()),
427 (is_virtual() ? " virtual='1'" : ""),
428 C->unique() - nodes);
429 }
430 // Push the result from the inlined method onto the stack.
431 kit.push_result();
432 C->print_inlining_update(this);
433 return kit.transfer_exceptions_into_jvms();
434 }
435
436 // The intrinsic bailed out
437 if (jvms->has_method()) {
438 // Not a root compile.
439 const char* msg;
440 if (callee->intrinsic_candidate()) {
441 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
442 } else {
443 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
444 : "failed to inline (intrinsic), method not annotated";
445 }
446 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
447 if (C->print_intrinsics() || C->print_inlining()) {
448 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
449 }
450 } else {
451 // Root compile
452 ResourceMark rm;
453 stringStream msg_stream;
454 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
455 vmIntrinsics::name_at(intrinsic_id()),
456 is_virtual() ? " (virtual)" : "", bci);
457 const char *msg = msg_stream.as_string();
458 log_debug(jit, inlining)("%s", msg);
459 if (C->print_intrinsics() || C->print_inlining()) {
460 tty->print("%s", msg);
461 }
462 }
463 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
464 C->print_inlining_update(this);
465 return NULL;
466 }
467
468 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
469 LibraryCallKit kit(jvms, this);
470 Compile* C = kit.C;
471 int nodes = C->unique();
472 _last_predicate = predicate;
473 #ifndef PRODUCT
474 assert(is_predicated() && predicate < predicates_count(), "sanity");
475 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
476 char buf[1000];
477 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
478 tty->print_cr("Predicate for intrinsic %s", str);
479 }
480 #endif
481 ciMethod* callee = kit.callee();
482 const int bci = kit.bci();
483
484 Node* slow_ctl = kit.try_to_predicate(predicate);
485 if (!kit.failing()) {
486 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
487 : "(intrinsic, predicate)";
488 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
489 if (C->print_intrinsics() || C->print_inlining()) {
490 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
491 }
492 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
493 if (C->log()) {
494 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
495 vmIntrinsics::name_at(intrinsic_id()),
496 (is_virtual() ? " virtual='1'" : ""),
497 C->unique() - nodes);
498 }
499 return slow_ctl; // Could be NULL if the check folds.
500 }
501
502 // The intrinsic bailed out
503 if (jvms->has_method()) {
504 // Not a root compile.
505 const char* msg = "failed to generate predicate for intrinsic";
506 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
507 if (C->print_intrinsics() || C->print_inlining()) {
508 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
509 }
510 } else {
511 // Root compile
512 ResourceMark rm;
513 stringStream msg_stream;
514 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
515 vmIntrinsics::name_at(intrinsic_id()),
516 is_virtual() ? " (virtual)" : "", bci);
517 const char *msg = msg_stream.as_string();
518 log_debug(jit, inlining)("%s", msg);
519 if (C->print_intrinsics() || C->print_inlining()) {
520 C->print_inlining_stream()->print("%s", msg);
521 }
522 }
523 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
524 return NULL;
525 }
526
527 bool LibraryCallKit::try_to_inline(int predicate) {
528 // Handle symbolic names for otherwise undistinguished boolean switches:
529 const bool is_store = true;
530 const bool is_compress = true;
531 const bool is_static = true;
532 const bool is_volatile = true;
533
534 if (!jvms()->has_method()) {
535 // Root JVMState has a null method.
536 assert(map()->memory()->Opcode() == Op_Parm, "");
537 // Insert the memory aliasing node
538 set_all_memory(reset_memory());
539 }
540 assert(merged_memory(), "");
541
542 switch (intrinsic_id()) {
543 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
544 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
545 case vmIntrinsics::_getClass: return inline_native_getClass();
546
547 case vmIntrinsics::_dsin:
548 case vmIntrinsics::_dcos:
549 case vmIntrinsics::_dtan:
550 case vmIntrinsics::_dabs:
551 case vmIntrinsics::_datan2:
552 case vmIntrinsics::_dsqrt:
553 case vmIntrinsics::_dexp:
554 case vmIntrinsics::_dlog:
555 case vmIntrinsics::_dlog10:
556 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
557
558 case vmIntrinsics::_min:
559 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
560
561 case vmIntrinsics::_notify:
562 case vmIntrinsics::_notifyAll:
563 if (InlineNotify) {
564 return inline_notify(intrinsic_id());
565 }
566 return false;
567
568 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
569 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
570 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
571 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
572 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
573 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
574 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
575 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
576 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
577 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
578 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
579 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
580 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
581
582 case vmIntrinsics::_arraycopy: return inline_arraycopy();
583
584 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
585 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
586 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
587 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
588
589 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
590 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
591 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
592 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
593 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
594 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
595 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
596
597 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
598 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
599
600 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
601 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
602 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
603 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
604
605 case vmIntrinsics::_compressStringC:
606 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
607 case vmIntrinsics::_inflateStringC:
608 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
609
610 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
611 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
612 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
613 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
614 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
615 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
616 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
617 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
618 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
619
620 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
621 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
622 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
623 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
624 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
625 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
626 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
627 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
628 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
629
630 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
631 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
632 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
633 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
634 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
635 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
636 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
637 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
638 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
639
640 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
641 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
642 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
643 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
644 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
645 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
646 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
647 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
648 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
649
650 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
651 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
652 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
653 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
654
655 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
656 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
657 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
658 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
659
660 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
661 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
662 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
663 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
664 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
665 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
666 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
667 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
668 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
669
670 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
671 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
672 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
673 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
674 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
675 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
676 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
677 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
678 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
679
680 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
681 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
682 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
683 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
684 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
685 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
686 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
687 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
688 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
689
690 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
691 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
692 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
693 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
694 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
695 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
696 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
697 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
698 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
699
700 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
701 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
702 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
703 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
704 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
705
706 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
707 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
708 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
709 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
710 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
711 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
712 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
713 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
714 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
715 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
716 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
717 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
718 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
719 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
720 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
721 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
722 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
723 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
724 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
725 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
726
727 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
728 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
729 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
730 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
731 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
732 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
733 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
734 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
735 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
736 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
737 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
738 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
739 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
740 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
741 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
742
743 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
744 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
745 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
746 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
747
748 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
749 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
750 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
751 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
752 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
753
754 case vmIntrinsics::_loadFence:
755 case vmIntrinsics::_storeFence:
756 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
757
758 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
759
760 case vmIntrinsics::_currentThread: return inline_native_currentThread();
761 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
762
763 #ifdef TRACE_HAVE_INTRINSICS
764 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
765 case vmIntrinsics::_getClassId: return inline_native_classID();
766 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
767 #endif
768 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
769 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
770 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
771 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
772 case vmIntrinsics::_getLength: return inline_native_getLength();
773 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
774 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
775 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
776 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
777 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
778 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
779
780 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
781 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
782
783 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
784
785 case vmIntrinsics::_isInstance:
786 case vmIntrinsics::_getModifiers:
787 case vmIntrinsics::_isInterface:
788 case vmIntrinsics::_isArray:
789 case vmIntrinsics::_isPrimitive:
790 case vmIntrinsics::_getSuperclass:
791 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
792
793 case vmIntrinsics::_floatToRawIntBits:
794 case vmIntrinsics::_floatToIntBits:
795 case vmIntrinsics::_intBitsToFloat:
796 case vmIntrinsics::_doubleToRawLongBits:
797 case vmIntrinsics::_doubleToLongBits:
798 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
799
800 case vmIntrinsics::_numberOfLeadingZeros_i:
801 case vmIntrinsics::_numberOfLeadingZeros_l:
802 case vmIntrinsics::_numberOfTrailingZeros_i:
803 case vmIntrinsics::_numberOfTrailingZeros_l:
804 case vmIntrinsics::_bitCount_i:
805 case vmIntrinsics::_bitCount_l:
806 case vmIntrinsics::_reverseBytes_i:
807 case vmIntrinsics::_reverseBytes_l:
808 case vmIntrinsics::_reverseBytes_s:
809 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
810
811 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
812
813 case vmIntrinsics::_Reference_get: return inline_reference_get();
814
815 case vmIntrinsics::_Class_cast: return inline_Class_cast();
816
817 case vmIntrinsics::_aescrypt_encryptBlock:
818 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
819
820 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
821 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
822 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
823
824 case vmIntrinsics::_counterMode_AESCrypt:
825 return inline_counterMode_AESCrypt(intrinsic_id());
826
827 case vmIntrinsics::_sha_implCompress:
828 case vmIntrinsics::_sha2_implCompress:
829 case vmIntrinsics::_sha5_implCompress:
830 return inline_sha_implCompress(intrinsic_id());
831
832 case vmIntrinsics::_digestBase_implCompressMB:
833 return inline_digestBase_implCompressMB(predicate);
834
835 case vmIntrinsics::_multiplyToLen:
836 return inline_multiplyToLen();
837
838 case vmIntrinsics::_squareToLen:
839 return inline_squareToLen();
840
841 case vmIntrinsics::_mulAdd:
842 return inline_mulAdd();
843
844 case vmIntrinsics::_montgomeryMultiply:
845 return inline_montgomeryMultiply();
846 case vmIntrinsics::_montgomerySquare:
847 return inline_montgomerySquare();
848
849 case vmIntrinsics::_vectorizedMismatch:
850 return inline_vectorizedMismatch();
851
852 case vmIntrinsics::_ghash_processBlocks:
853 return inline_ghash_processBlocks();
854
855 case vmIntrinsics::_encodeISOArray:
856 case vmIntrinsics::_encodeByteISOArray:
857 return inline_encodeISOArray();
858
859 case vmIntrinsics::_updateCRC32:
860 return inline_updateCRC32();
861 case vmIntrinsics::_updateBytesCRC32:
862 return inline_updateBytesCRC32();
863 case vmIntrinsics::_updateByteBufferCRC32:
864 return inline_updateByteBufferCRC32();
865
866 case vmIntrinsics::_updateBytesCRC32C:
867 return inline_updateBytesCRC32C();
868 case vmIntrinsics::_updateDirectByteBufferCRC32C:
869 return inline_updateDirectByteBufferCRC32C();
870
871 case vmIntrinsics::_updateBytesAdler32:
872 return inline_updateBytesAdler32();
873 case vmIntrinsics::_updateByteBufferAdler32:
874 return inline_updateByteBufferAdler32();
875
876 case vmIntrinsics::_profileBoolean:
877 return inline_profileBoolean();
878 case vmIntrinsics::_isCompileConstant:
879 return inline_isCompileConstant();
880
881 case vmIntrinsics::_hasNegatives:
882 return inline_hasNegatives();
883
884 case vmIntrinsics::_fmaD:
885 case vmIntrinsics::_fmaF:
886 return inline_fma(intrinsic_id());
887
888 case vmIntrinsics::_VectorUnaryOp:
889 return inline_vector_nary_operation(1);
890 case vmIntrinsics::_VectorBinaryOp:
891 return inline_vector_nary_operation(2);
892 case vmIntrinsics::_VectorTernaryOp:
893 return inline_vector_nary_operation(3);
894
895 case vmIntrinsics::_VectorBroadcastCoerced:
896 return inline_vector_broadcast_coerced();
897 case vmIntrinsics::_VectorLoadOp:
898 return inline_vector_mem_operation(/*is_store=*/false);
899 case vmIntrinsics::_VectorStoreOp:
900 return inline_vector_mem_operation(/*is_store=*/true);
901 case vmIntrinsics::_VectorReductionCoerced:
902 return inline_vector_reduction();
903 case vmIntrinsics::_VectorTest:
904 return inline_vector_test();
905 case vmIntrinsics::_VectorBlend:
906 return inline_vector_blend();
907 case vmIntrinsics::_VectorCompare:
908 return inline_vector_compare();
909 case vmIntrinsics::_VectorBroadcastInt:
910 return inline_vector_broadcast_int();
911 case vmIntrinsics::_VectorReinterpret:
912 return inline_vector_reinterpret();
913
914 default:
915 // If you get here, it may be that someone has added a new intrinsic
916 // to the list in vmSymbols.hpp without implementing it here.
917 #ifndef PRODUCT
918 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
919 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
920 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
921 }
922 #endif
923 return false;
924 }
925 }
926
927 Node* LibraryCallKit::try_to_predicate(int predicate) {
928 if (!jvms()->has_method()) {
929 // Root JVMState has a null method.
930 assert(map()->memory()->Opcode() == Op_Parm, "");
931 // Insert the memory aliasing node
932 set_all_memory(reset_memory());
933 }
934 assert(merged_memory(), "");
935
936 switch (intrinsic_id()) {
937 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
938 return inline_cipherBlockChaining_AESCrypt_predicate(false);
939 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
940 return inline_cipherBlockChaining_AESCrypt_predicate(true);
941 case vmIntrinsics::_counterMode_AESCrypt:
942 return inline_counterMode_AESCrypt_predicate();
943 case vmIntrinsics::_digestBase_implCompressMB:
944 return inline_digestBase_implCompressMB_predicate(predicate);
945
946 default:
947 // If you get here, it may be that someone has added a new intrinsic
948 // to the list in vmSymbols.hpp without implementing it here.
949 #ifndef PRODUCT
950 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
951 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
952 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
953 }
954 #endif
955 Node* slow_ctl = control();
956 set_control(top()); // No fast path instrinsic
957 return slow_ctl;
958 }
959 }
960
961 //------------------------------set_result-------------------------------
962 // Helper function for finishing intrinsics.
963 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
964 record_for_igvn(region);
965 set_control(_gvn.transform(region));
966 set_result( _gvn.transform(value));
967 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
968 }
969
970 //------------------------------generate_guard---------------------------
971 // Helper function for generating guarded fast-slow graph structures.
972 // The given 'test', if true, guards a slow path. If the test fails
973 // then a fast path can be taken. (We generally hope it fails.)
974 // In all cases, GraphKit::control() is updated to the fast path.
975 // The returned value represents the control for the slow path.
976 // The return value is never 'top'; it is either a valid control
977 // or NULL if it is obvious that the slow path can never be taken.
978 // Also, if region and the slow control are not NULL, the slow edge
979 // is appended to the region.
980 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
981 if (stopped()) {
982 // Already short circuited.
983 return NULL;
984 }
985
986 // Build an if node and its projections.
987 // If test is true we take the slow path, which we assume is uncommon.
988 if (_gvn.type(test) == TypeInt::ZERO) {
989 // The slow branch is never taken. No need to build this guard.
990 return NULL;
991 }
992
993 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
994
995 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
996 if (if_slow == top()) {
997 // The slow branch is never taken. No need to build this guard.
998 return NULL;
999 }
1000
1001 if (region != NULL)
1002 region->add_req(if_slow);
1003
1004 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1005 set_control(if_fast);
1006
1007 return if_slow;
1008 }
1009
1010 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1011 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1012 }
1013 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1014 return generate_guard(test, region, PROB_FAIR);
1015 }
1016
1017 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1018 Node* *pos_index) {
1019 if (stopped())
1020 return NULL; // already stopped
1021 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1022 return NULL; // index is already adequately typed
1023 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1024 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1025 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1026 if (is_neg != NULL && pos_index != NULL) {
1027 // Emulate effect of Parse::adjust_map_after_if.
1028 Node* ccast = new CastIINode(index, TypeInt::POS);
1029 ccast->set_req(0, control());
1030 (*pos_index) = _gvn.transform(ccast);
1031 }
1032 return is_neg;
1033 }
1034
1035 // Make sure that 'position' is a valid limit index, in [0..length].
1036 // There are two equivalent plans for checking this:
1037 // A. (offset + copyLength) unsigned<= arrayLength
1038 // B. offset <= (arrayLength - copyLength)
1039 // We require that all of the values above, except for the sum and
1040 // difference, are already known to be non-negative.
1041 // Plan A is robust in the face of overflow, if offset and copyLength
1042 // are both hugely positive.
1043 //
1044 // Plan B is less direct and intuitive, but it does not overflow at
1045 // all, since the difference of two non-negatives is always
1046 // representable. Whenever Java methods must perform the equivalent
1047 // check they generally use Plan B instead of Plan A.
1048 // For the moment we use Plan A.
1049 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1050 Node* subseq_length,
1051 Node* array_length,
1052 RegionNode* region) {
1053 if (stopped())
1054 return NULL; // already stopped
1055 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1056 if (zero_offset && subseq_length->eqv_uncast(array_length))
1057 return NULL; // common case of whole-array copy
1058 Node* last = subseq_length;
1059 if (!zero_offset) // last += offset
1060 last = _gvn.transform(new AddINode(last, offset));
1061 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1062 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1063 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1064 return is_over;
1065 }
1066
1067 // Emit range checks for the given String.value byte array
1068 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1069 if (stopped()) {
1070 return; // already stopped
1071 }
1072 RegionNode* bailout = new RegionNode(1);
1073 record_for_igvn(bailout);
1074 if (char_count) {
1075 // Convert char count to byte count
1076 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1077 }
1078
1079 // Offset and count must not be negative
1080 generate_negative_guard(offset, bailout);
1081 generate_negative_guard(count, bailout);
1082 // Offset + count must not exceed length of array
1083 generate_limit_guard(offset, count, load_array_length(array), bailout);
1084
1085 if (bailout->req() > 1) {
1086 PreserveJVMState pjvms(this);
1087 set_control(_gvn.transform(bailout));
1088 uncommon_trap(Deoptimization::Reason_intrinsic,
1089 Deoptimization::Action_maybe_recompile);
1090 }
1091 }
1092
1093 //--------------------------generate_current_thread--------------------
1094 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1095 ciKlass* thread_klass = env()->Thread_klass();
1096 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1097 Node* thread = _gvn.transform(new ThreadLocalNode());
1098 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1099 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1100 tls_output = thread;
1101 return threadObj;
1102 }
1103
1104
1105 //------------------------------make_string_method_node------------------------
1106 // Helper method for String intrinsic functions. This version is called with
1107 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1108 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1109 // containing the lengths of str1 and str2.
1110 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1111 Node* result = NULL;
1112 switch (opcode) {
1113 case Op_StrIndexOf:
1114 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1115 str1_start, cnt1, str2_start, cnt2, ae);
1116 break;
1117 case Op_StrComp:
1118 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1119 str1_start, cnt1, str2_start, cnt2, ae);
1120 break;
1121 case Op_StrEquals:
1122 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1123 // Use the constant length if there is one because optimized match rule may exist.
1124 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1125 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1126 break;
1127 default:
1128 ShouldNotReachHere();
1129 return NULL;
1130 }
1131
1132 // All these intrinsics have checks.
1133 C->set_has_split_ifs(true); // Has chance for split-if optimization
1134 clear_upper_avx();
1135
1136 return _gvn.transform(result);
1137 }
1138
1139 //------------------------------inline_string_compareTo------------------------
1140 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1141 Node* arg1 = argument(0);
1142 Node* arg2 = argument(1);
1143
1144 // Get start addr and length of first argument
1145 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1146 Node* arg1_cnt = load_array_length(arg1);
1147
1148 // Get start addr and length of second argument
1149 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1150 Node* arg2_cnt = load_array_length(arg2);
1151
1152 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1153 set_result(result);
1154 return true;
1155 }
1156
1157 //------------------------------inline_string_equals------------------------
1158 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1159 Node* arg1 = argument(0);
1160 Node* arg2 = argument(1);
1161
1162 // paths (plus control) merge
1163 RegionNode* region = new RegionNode(3);
1164 Node* phi = new PhiNode(region, TypeInt::BOOL);
1165
1166 if (!stopped()) {
1167 // Get start addr and length of first argument
1168 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1169 Node* arg1_cnt = load_array_length(arg1);
1170
1171 // Get start addr and length of second argument
1172 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1173 Node* arg2_cnt = load_array_length(arg2);
1174
1175 // Check for arg1_cnt != arg2_cnt
1176 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1177 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1178 Node* if_ne = generate_slow_guard(bol, NULL);
1179 if (if_ne != NULL) {
1180 phi->init_req(2, intcon(0));
1181 region->init_req(2, if_ne);
1182 }
1183
1184 // Check for count == 0 is done by assembler code for StrEquals.
1185
1186 if (!stopped()) {
1187 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1188 phi->init_req(1, equals);
1189 region->init_req(1, control());
1190 }
1191 }
1192
1193 // post merge
1194 set_control(_gvn.transform(region));
1195 record_for_igvn(region);
1196
1197 set_result(_gvn.transform(phi));
1198 return true;
1199 }
1200
1201 //------------------------------inline_array_equals----------------------------
1202 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1203 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1204 Node* arg1 = argument(0);
1205 Node* arg2 = argument(1);
1206
1207 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1208 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1209 clear_upper_avx();
1210
1211 return true;
1212 }
1213
1214 //------------------------------inline_hasNegatives------------------------------
1215 bool LibraryCallKit::inline_hasNegatives() {
1216 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1217 return false;
1218 }
1219
1220 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1221 // no receiver since it is static method
1222 Node* ba = argument(0);
1223 Node* offset = argument(1);
1224 Node* len = argument(2);
1225
1226 // Range checks
1227 generate_string_range_check(ba, offset, len, false);
1228 if (stopped()) {
1229 return true;
1230 }
1231 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1232 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1233 set_result(_gvn.transform(result));
1234 return true;
1235 }
1236
1237 bool LibraryCallKit::inline_preconditions_checkIndex() {
1238 Node* index = argument(0);
1239 Node* length = argument(1);
1240 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1241 return false;
1242 }
1243
1244 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1245 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1246
1247 {
1248 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1249 uncommon_trap(Deoptimization::Reason_intrinsic,
1250 Deoptimization::Action_make_not_entrant);
1251 }
1252
1253 if (stopped()) {
1254 return false;
1255 }
1256
1257 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1258 BoolTest::mask btest = BoolTest::lt;
1259 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1260 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1261 _gvn.set_type(rc, rc->Value(&_gvn));
1262 if (!rc_bool->is_Con()) {
1263 record_for_igvn(rc);
1264 }
1265 set_control(_gvn.transform(new IfTrueNode(rc)));
1266 {
1267 PreserveJVMState pjvms(this);
1268 set_control(_gvn.transform(new IfFalseNode(rc)));
1269 uncommon_trap(Deoptimization::Reason_range_check,
1270 Deoptimization::Action_make_not_entrant);
1271 }
1272
1273 if (stopped()) {
1274 return false;
1275 }
1276
1277 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1278 result->set_req(0, control());
1279 result = _gvn.transform(result);
1280 set_result(result);
1281 replace_in_map(index, result);
1282 clear_upper_avx();
1283 return true;
1284 }
1285
1286 //------------------------------inline_string_indexOf------------------------
1287 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1288 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1289 return false;
1290 }
1291 Node* src = argument(0);
1292 Node* tgt = argument(1);
1293
1294 // Make the merge point
1295 RegionNode* result_rgn = new RegionNode(4);
1296 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1297
1298 // Get start addr and length of source string
1299 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1300 Node* src_count = load_array_length(src);
1301
1302 // Get start addr and length of substring
1303 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1304 Node* tgt_count = load_array_length(tgt);
1305
1306 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1307 // Divide src size by 2 if String is UTF16 encoded
1308 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1309 }
1310 if (ae == StrIntrinsicNode::UU) {
1311 // Divide substring size by 2 if String is UTF16 encoded
1312 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1313 }
1314
1315 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1316 if (result != NULL) {
1317 result_phi->init_req(3, result);
1318 result_rgn->init_req(3, control());
1319 }
1320 set_control(_gvn.transform(result_rgn));
1321 record_for_igvn(result_rgn);
1322 set_result(_gvn.transform(result_phi));
1323
1324 return true;
1325 }
1326
1327 //-----------------------------inline_string_indexOf-----------------------
1328 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1329 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1330 return false;
1331 }
1332 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1333 return false;
1334 }
1335 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1336 Node* src = argument(0); // byte[]
1337 Node* src_count = argument(1); // char count
1338 Node* tgt = argument(2); // byte[]
1339 Node* tgt_count = argument(3); // char count
1340 Node* from_index = argument(4); // char index
1341
1342 // Multiply byte array index by 2 if String is UTF16 encoded
1343 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1344 src_count = _gvn.transform(new SubINode(src_count, from_index));
1345 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1346 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1347
1348 // Range checks
1349 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1350 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1351 if (stopped()) {
1352 return true;
1353 }
1354
1355 RegionNode* region = new RegionNode(5);
1356 Node* phi = new PhiNode(region, TypeInt::INT);
1357
1358 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1359 if (result != NULL) {
1360 // The result is index relative to from_index if substring was found, -1 otherwise.
1361 // Generate code which will fold into cmove.
1362 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1363 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1364
1365 Node* if_lt = generate_slow_guard(bol, NULL);
1366 if (if_lt != NULL) {
1367 // result == -1
1368 phi->init_req(3, result);
1369 region->init_req(3, if_lt);
1370 }
1371 if (!stopped()) {
1372 result = _gvn.transform(new AddINode(result, from_index));
1373 phi->init_req(4, result);
1374 region->init_req(4, control());
1375 }
1376 }
1377
1378 set_control(_gvn.transform(region));
1379 record_for_igvn(region);
1380 set_result(_gvn.transform(phi));
1381 clear_upper_avx();
1382
1383 return true;
1384 }
1385
1386 // Create StrIndexOfNode with fast path checks
1387 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1388 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1389 // Check for substr count > string count
1390 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1391 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1392 Node* if_gt = generate_slow_guard(bol, NULL);
1393 if (if_gt != NULL) {
1394 phi->init_req(1, intcon(-1));
1395 region->init_req(1, if_gt);
1396 }
1397 if (!stopped()) {
1398 // Check for substr count == 0
1399 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1400 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1401 Node* if_zero = generate_slow_guard(bol, NULL);
1402 if (if_zero != NULL) {
1403 phi->init_req(2, intcon(0));
1404 region->init_req(2, if_zero);
1405 }
1406 }
1407 if (!stopped()) {
1408 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1409 }
1410 return NULL;
1411 }
1412
1413 //-----------------------------inline_string_indexOfChar-----------------------
1414 bool LibraryCallKit::inline_string_indexOfChar() {
1415 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1416 return false;
1417 }
1418 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1419 return false;
1420 }
1421 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1422 Node* src = argument(0); // byte[]
1423 Node* tgt = argument(1); // tgt is int ch
1424 Node* from_index = argument(2);
1425 Node* max = argument(3);
1426
1427 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1428 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1429 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1430
1431 // Range checks
1432 generate_string_range_check(src, src_offset, src_count, true);
1433 if (stopped()) {
1434 return true;
1435 }
1436
1437 RegionNode* region = new RegionNode(3);
1438 Node* phi = new PhiNode(region, TypeInt::INT);
1439
1440 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1441 C->set_has_split_ifs(true); // Has chance for split-if optimization
1442 _gvn.transform(result);
1443
1444 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1445 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1446
1447 Node* if_lt = generate_slow_guard(bol, NULL);
1448 if (if_lt != NULL) {
1449 // result == -1
1450 phi->init_req(2, result);
1451 region->init_req(2, if_lt);
1452 }
1453 if (!stopped()) {
1454 result = _gvn.transform(new AddINode(result, from_index));
1455 phi->init_req(1, result);
1456 region->init_req(1, control());
1457 }
1458 set_control(_gvn.transform(region));
1459 record_for_igvn(region);
1460 set_result(_gvn.transform(phi));
1461
1462 return true;
1463 }
1464 //---------------------------inline_string_copy---------------------
1465 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1466 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1467 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1468 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1469 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1470 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1471 bool LibraryCallKit::inline_string_copy(bool compress) {
1472 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1473 return false;
1474 }
1475 int nargs = 5; // 2 oops, 3 ints
1476 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1477
1478 Node* src = argument(0);
1479 Node* src_offset = argument(1);
1480 Node* dst = argument(2);
1481 Node* dst_offset = argument(3);
1482 Node* length = argument(4);
1483
1484 // Check for allocation before we add nodes that would confuse
1485 // tightly_coupled_allocation()
1486 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1487
1488 // Figure out the size and type of the elements we will be copying.
1489 const Type* src_type = src->Value(&_gvn);
1490 const Type* dst_type = dst->Value(&_gvn);
1491 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1492 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1493 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1494 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1495 "Unsupported array types for inline_string_copy");
1496
1497 // Convert char[] offsets to byte[] offsets
1498 bool convert_src = (compress && src_elem == T_BYTE);
1499 bool convert_dst = (!compress && dst_elem == T_BYTE);
1500 if (convert_src) {
1501 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1502 } else if (convert_dst) {
1503 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1504 }
1505
1506 // Range checks
1507 generate_string_range_check(src, src_offset, length, convert_src);
1508 generate_string_range_check(dst, dst_offset, length, convert_dst);
1509 if (stopped()) {
1510 return true;
1511 }
1512
1513 Node* src_start = array_element_address(src, src_offset, src_elem);
1514 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1515 // 'src_start' points to src array + scaled offset
1516 // 'dst_start' points to dst array + scaled offset
1517 Node* count = NULL;
1518 if (compress) {
1519 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1520 } else {
1521 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1522 }
1523
1524 if (alloc != NULL) {
1525 if (alloc->maybe_set_complete(&_gvn)) {
1526 // "You break it, you buy it."
1527 InitializeNode* init = alloc->initialization();
1528 assert(init->is_complete(), "we just did this");
1529 init->set_complete_with_arraycopy();
1530 assert(dst->is_CheckCastPP(), "sanity");
1531 assert(dst->in(0)->in(0) == init, "dest pinned");
1532 }
1533 // Do not let stores that initialize this object be reordered with
1534 // a subsequent store that would make this object accessible by
1535 // other threads.
1536 // Record what AllocateNode this StoreStore protects so that
1537 // escape analysis can go from the MemBarStoreStoreNode to the
1538 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1539 // based on the escape status of the AllocateNode.
1540 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1541 }
1542 if (compress) {
1543 set_result(_gvn.transform(count));
1544 }
1545 clear_upper_avx();
1546
1547 return true;
1548 }
1549
1550 #ifdef _LP64
1551 #define XTOP ,top() /*additional argument*/
1552 #else //_LP64
1553 #define XTOP /*no additional argument*/
1554 #endif //_LP64
1555
1556 //------------------------inline_string_toBytesU--------------------------
1557 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1558 bool LibraryCallKit::inline_string_toBytesU() {
1559 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1560 return false;
1561 }
1562 // Get the arguments.
1563 Node* value = argument(0);
1564 Node* offset = argument(1);
1565 Node* length = argument(2);
1566
1567 Node* newcopy = NULL;
1568
1569 // Set the original stack and the reexecute bit for the interpreter to reexecute
1570 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1571 { PreserveReexecuteState preexecs(this);
1572 jvms()->set_should_reexecute(true);
1573
1574 // Check if a null path was taken unconditionally.
1575 value = null_check(value);
1576
1577 RegionNode* bailout = new RegionNode(1);
1578 record_for_igvn(bailout);
1579
1580 // Range checks
1581 generate_negative_guard(offset, bailout);
1582 generate_negative_guard(length, bailout);
1583 generate_limit_guard(offset, length, load_array_length(value), bailout);
1584 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1585 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1586
1587 if (bailout->req() > 1) {
1588 PreserveJVMState pjvms(this);
1589 set_control(_gvn.transform(bailout));
1590 uncommon_trap(Deoptimization::Reason_intrinsic,
1591 Deoptimization::Action_maybe_recompile);
1592 }
1593 if (stopped()) {
1594 return true;
1595 }
1596
1597 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1598 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1599 newcopy = new_array(klass_node, size, 0); // no arguments to push
1600 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1601
1602 // Calculate starting addresses.
1603 Node* src_start = array_element_address(value, offset, T_CHAR);
1604 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1605
1606 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1607 const TypeInt* toffset = gvn().type(offset)->is_int();
1608 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1609
1610 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1611 const char* copyfunc_name = "arraycopy";
1612 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1613 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1614 OptoRuntime::fast_arraycopy_Type(),
1615 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1616 src_start, dst_start, ConvI2X(length) XTOP);
1617 // Do not let reads from the cloned object float above the arraycopy.
1618 if (alloc != NULL) {
1619 if (alloc->maybe_set_complete(&_gvn)) {
1620 // "You break it, you buy it."
1621 InitializeNode* init = alloc->initialization();
1622 assert(init->is_complete(), "we just did this");
1623 init->set_complete_with_arraycopy();
1624 assert(newcopy->is_CheckCastPP(), "sanity");
1625 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1626 }
1627 // Do not let stores that initialize this object be reordered with
1628 // a subsequent store that would make this object accessible by
1629 // other threads.
1630 // Record what AllocateNode this StoreStore protects so that
1631 // escape analysis can go from the MemBarStoreStoreNode to the
1632 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1633 // based on the escape status of the AllocateNode.
1634 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1635 } else {
1636 insert_mem_bar(Op_MemBarCPUOrder);
1637 }
1638 } // original reexecute is set back here
1639
1640 C->set_has_split_ifs(true); // Has chance for split-if optimization
1641 if (!stopped()) {
1642 set_result(newcopy);
1643 }
1644 clear_upper_avx();
1645
1646 return true;
1647 }
1648
1649 //------------------------inline_string_getCharsU--------------------------
1650 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1651 bool LibraryCallKit::inline_string_getCharsU() {
1652 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1653 return false;
1654 }
1655
1656 // Get the arguments.
1657 Node* src = argument(0);
1658 Node* src_begin = argument(1);
1659 Node* src_end = argument(2); // exclusive offset (i < src_end)
1660 Node* dst = argument(3);
1661 Node* dst_begin = argument(4);
1662
1663 // Check for allocation before we add nodes that would confuse
1664 // tightly_coupled_allocation()
1665 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1666
1667 // Check if a null path was taken unconditionally.
1668 src = null_check(src);
1669 dst = null_check(dst);
1670 if (stopped()) {
1671 return true;
1672 }
1673
1674 // Get length and convert char[] offset to byte[] offset
1675 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1676 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1677
1678 // Range checks
1679 generate_string_range_check(src, src_begin, length, true);
1680 generate_string_range_check(dst, dst_begin, length, false);
1681 if (stopped()) {
1682 return true;
1683 }
1684
1685 if (!stopped()) {
1686 // Calculate starting addresses.
1687 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1688 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1689
1690 // Check if array addresses are aligned to HeapWordSize
1691 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1692 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1693 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1694 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1695
1696 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1697 const char* copyfunc_name = "arraycopy";
1698 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1699 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1700 OptoRuntime::fast_arraycopy_Type(),
1701 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1702 src_start, dst_start, ConvI2X(length) XTOP);
1703 // Do not let reads from the cloned object float above the arraycopy.
1704 if (alloc != NULL) {
1705 if (alloc->maybe_set_complete(&_gvn)) {
1706 // "You break it, you buy it."
1707 InitializeNode* init = alloc->initialization();
1708 assert(init->is_complete(), "we just did this");
1709 init->set_complete_with_arraycopy();
1710 assert(dst->is_CheckCastPP(), "sanity");
1711 assert(dst->in(0)->in(0) == init, "dest pinned");
1712 }
1713 // Do not let stores that initialize this object be reordered with
1714 // a subsequent store that would make this object accessible by
1715 // other threads.
1716 // Record what AllocateNode this StoreStore protects so that
1717 // escape analysis can go from the MemBarStoreStoreNode to the
1718 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1719 // based on the escape status of the AllocateNode.
1720 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1721 } else {
1722 insert_mem_bar(Op_MemBarCPUOrder);
1723 }
1724 }
1725
1726 C->set_has_split_ifs(true); // Has chance for split-if optimization
1727 return true;
1728 }
1729
1730 //----------------------inline_string_char_access----------------------------
1731 // Store/Load char to/from byte[] array.
1732 // static void StringUTF16.putChar(byte[] val, int index, int c)
1733 // static char StringUTF16.getChar(byte[] val, int index)
1734 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1735 Node* value = argument(0);
1736 Node* index = argument(1);
1737 Node* ch = is_store ? argument(2) : NULL;
1738
1739 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1740 // correctly requires matched array shapes.
1741 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1742 "sanity: byte[] and char[] bases agree");
1743 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1744 "sanity: byte[] and char[] scales agree");
1745
1746 // Bail when getChar over constants is requested: constant folding would
1747 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1748 // Java method would constant fold nicely instead.
1749 if (!is_store && value->is_Con() && index->is_Con()) {
1750 return false;
1751 }
1752
1753 Node* adr = array_element_address(value, index, T_CHAR);
1754 if (adr->is_top()) {
1755 return false;
1756 }
1757 if (is_store) {
1758 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1759 false, false, true /* mismatched */);
1760 } else {
1761 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1762 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1763 set_result(ch);
1764 }
1765 return true;
1766 }
1767
1768 //--------------------------round_double_node--------------------------------
1769 // Round a double node if necessary.
1770 Node* LibraryCallKit::round_double_node(Node* n) {
1771 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1772 n = _gvn.transform(new RoundDoubleNode(0, n));
1773 return n;
1774 }
1775
1776 //------------------------------inline_math-----------------------------------
1777 // public static double Math.abs(double)
1778 // public static double Math.sqrt(double)
1779 // public static double Math.log(double)
1780 // public static double Math.log10(double)
1781 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1782 Node* arg = round_double_node(argument(0));
1783 Node* n = NULL;
1784 switch (id) {
1785 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1786 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1787 default: fatal_unexpected_iid(id); break;
1788 }
1789 set_result(_gvn.transform(n));
1790 return true;
1791 }
1792
1793 //------------------------------runtime_math-----------------------------
1794 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1795 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1796 "must be (DD)D or (D)D type");
1797
1798 // Inputs
1799 Node* a = round_double_node(argument(0));
1800 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1801
1802 const TypePtr* no_memory_effects = NULL;
1803 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1804 no_memory_effects,
1805 a, top(), b, b ? top() : NULL);
1806 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1807 #ifdef ASSERT
1808 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1809 assert(value_top == top(), "second value must be top");
1810 #endif
1811
1812 set_result(value);
1813 return true;
1814 }
1815
1816 //------------------------------inline_math_native-----------------------------
1817 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1818 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1819 switch (id) {
1820 // These intrinsics are not properly supported on all hardware
1821 case vmIntrinsics::_dsin:
1822 return StubRoutines::dsin() != NULL ?
1823 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1824 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1825 case vmIntrinsics::_dcos:
1826 return StubRoutines::dcos() != NULL ?
1827 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1828 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1829 case vmIntrinsics::_dtan:
1830 return StubRoutines::dtan() != NULL ?
1831 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1832 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1833 case vmIntrinsics::_dlog:
1834 return StubRoutines::dlog() != NULL ?
1835 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1836 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1837 case vmIntrinsics::_dlog10:
1838 return StubRoutines::dlog10() != NULL ?
1839 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1840 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1841
1842 // These intrinsics are supported on all hardware
1843 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1844 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1845
1846 case vmIntrinsics::_dexp:
1847 return StubRoutines::dexp() != NULL ?
1848 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1849 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1850 case vmIntrinsics::_dpow: {
1851 Node* exp = round_double_node(argument(2));
1852 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1853 if (d != NULL && d->getd() == 2.0) {
1854 // Special case: pow(x, 2.0) => x * x
1855 Node* base = round_double_node(argument(0));
1856 set_result(_gvn.transform(new MulDNode(base, base)));
1857 return true;
1858 }
1859 return StubRoutines::dexp() != NULL ?
1860 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1861 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1862 }
1863 #undef FN_PTR
1864
1865 // These intrinsics are not yet correctly implemented
1866 case vmIntrinsics::_datan2:
1867 return false;
1868
1869 default:
1870 fatal_unexpected_iid(id);
1871 return false;
1872 }
1873 }
1874
1875 static bool is_simple_name(Node* n) {
1876 return (n->req() == 1 // constant
1877 || (n->is_Type() && n->as_Type()->type()->singleton())
1878 || n->is_Proj() // parameter or return value
1879 || n->is_Phi() // local of some sort
1880 );
1881 }
1882
1883 //----------------------------inline_notify-----------------------------------*
1884 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1885 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1886 address func;
1887 if (id == vmIntrinsics::_notify) {
1888 func = OptoRuntime::monitor_notify_Java();
1889 } else {
1890 func = OptoRuntime::monitor_notifyAll_Java();
1891 }
1892 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1893 make_slow_call_ex(call, env()->Throwable_klass(), false);
1894 return true;
1895 }
1896
1897
1898 //----------------------------inline_min_max-----------------------------------
1899 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1900 set_result(generate_min_max(id, argument(0), argument(1)));
1901 return true;
1902 }
1903
1904 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1905 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1906 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1907 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1908 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1909
1910 {
1911 PreserveJVMState pjvms(this);
1912 PreserveReexecuteState preexecs(this);
1913 jvms()->set_should_reexecute(true);
1914
1915 set_control(slow_path);
1916 set_i_o(i_o());
1917
1918 uncommon_trap(Deoptimization::Reason_intrinsic,
1919 Deoptimization::Action_none);
1920 }
1921
1922 set_control(fast_path);
1923 set_result(math);
1924 }
1925
1926 template <typename OverflowOp>
1927 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1928 typedef typename OverflowOp::MathOp MathOp;
1929
1930 MathOp* mathOp = new MathOp(arg1, arg2);
1931 Node* operation = _gvn.transform( mathOp );
1932 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1933 inline_math_mathExact(operation, ofcheck);
1934 return true;
1935 }
1936
1937 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1938 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1939 }
1940
1941 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1942 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1943 }
1944
1945 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1946 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1947 }
1948
1949 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1950 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1951 }
1952
1953 bool LibraryCallKit::inline_math_negateExactI() {
1954 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1955 }
1956
1957 bool LibraryCallKit::inline_math_negateExactL() {
1958 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1959 }
1960
1961 bool LibraryCallKit::inline_math_multiplyExactI() {
1962 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1963 }
1964
1965 bool LibraryCallKit::inline_math_multiplyExactL() {
1966 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1967 }
1968
1969 bool LibraryCallKit::inline_math_multiplyHigh() {
1970 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1971 return true;
1972 }
1973
1974 Node*
1975 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1976 // These are the candidate return value:
1977 Node* xvalue = x0;
1978 Node* yvalue = y0;
1979
1980 if (xvalue == yvalue) {
1981 return xvalue;
1982 }
1983
1984 bool want_max = (id == vmIntrinsics::_max);
1985
1986 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1987 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1988 if (txvalue == NULL || tyvalue == NULL) return top();
1989 // This is not really necessary, but it is consistent with a
1990 // hypothetical MaxINode::Value method:
1991 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1992
1993 // %%% This folding logic should (ideally) be in a different place.
1994 // Some should be inside IfNode, and there to be a more reliable
1995 // transformation of ?: style patterns into cmoves. We also want
1996 // more powerful optimizations around cmove and min/max.
1997
1998 // Try to find a dominating comparison of these guys.
1999 // It can simplify the index computation for Arrays.copyOf
2000 // and similar uses of System.arraycopy.
2001 // First, compute the normalized version of CmpI(x, y).
2002 int cmp_op = Op_CmpI;
2003 Node* xkey = xvalue;
2004 Node* ykey = yvalue;
2005 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2006 if (ideal_cmpxy->is_Cmp()) {
2007 // E.g., if we have CmpI(length - offset, count),
2008 // it might idealize to CmpI(length, count + offset)
2009 cmp_op = ideal_cmpxy->Opcode();
2010 xkey = ideal_cmpxy->in(1);
2011 ykey = ideal_cmpxy->in(2);
2012 }
2013
2014 // Start by locating any relevant comparisons.
2015 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2016 Node* cmpxy = NULL;
2017 Node* cmpyx = NULL;
2018 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2019 Node* cmp = start_from->fast_out(k);
2020 if (cmp->outcnt() > 0 && // must have prior uses
2021 cmp->in(0) == NULL && // must be context-independent
2022 cmp->Opcode() == cmp_op) { // right kind of compare
2023 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2024 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2025 }
2026 }
2027
2028 const int NCMPS = 2;
2029 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2030 int cmpn;
2031 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2032 if (cmps[cmpn] != NULL) break; // find a result
2033 }
2034 if (cmpn < NCMPS) {
2035 // Look for a dominating test that tells us the min and max.
2036 int depth = 0; // Limit search depth for speed
2037 Node* dom = control();
2038 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2039 if (++depth >= 100) break;
2040 Node* ifproj = dom;
2041 if (!ifproj->is_Proj()) continue;
2042 Node* iff = ifproj->in(0);
2043 if (!iff->is_If()) continue;
2044 Node* bol = iff->in(1);
2045 if (!bol->is_Bool()) continue;
2046 Node* cmp = bol->in(1);
2047 if (cmp == NULL) continue;
2048 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2049 if (cmps[cmpn] == cmp) break;
2050 if (cmpn == NCMPS) continue;
2051 BoolTest::mask btest = bol->as_Bool()->_test._test;
2052 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2053 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2054 // At this point, we know that 'x btest y' is true.
2055 switch (btest) {
2056 case BoolTest::eq:
2057 // They are proven equal, so we can collapse the min/max.
2058 // Either value is the answer. Choose the simpler.
2059 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2060 return yvalue;
2061 return xvalue;
2062 case BoolTest::lt: // x < y
2063 case BoolTest::le: // x <= y
2064 return (want_max ? yvalue : xvalue);
2065 case BoolTest::gt: // x > y
2066 case BoolTest::ge: // x >= y
2067 return (want_max ? xvalue : yvalue);
2068 default:
2069 break;
2070 }
2071 }
2072 }
2073
2074 // We failed to find a dominating test.
2075 // Let's pick a test that might GVN with prior tests.
2076 Node* best_bol = NULL;
2077 BoolTest::mask best_btest = BoolTest::illegal;
2078 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2079 Node* cmp = cmps[cmpn];
2080 if (cmp == NULL) continue;
2081 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2082 Node* bol = cmp->fast_out(j);
2083 if (!bol->is_Bool()) continue;
2084 BoolTest::mask btest = bol->as_Bool()->_test._test;
2085 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2086 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2087 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2088 best_bol = bol->as_Bool();
2089 best_btest = btest;
2090 }
2091 }
2092 }
2093
2094 Node* answer_if_true = NULL;
2095 Node* answer_if_false = NULL;
2096 switch (best_btest) {
2097 default:
2098 if (cmpxy == NULL)
2099 cmpxy = ideal_cmpxy;
2100 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2101 // and fall through:
2102 case BoolTest::lt: // x < y
2103 case BoolTest::le: // x <= y
2104 answer_if_true = (want_max ? yvalue : xvalue);
2105 answer_if_false = (want_max ? xvalue : yvalue);
2106 break;
2107 case BoolTest::gt: // x > y
2108 case BoolTest::ge: // x >= y
2109 answer_if_true = (want_max ? xvalue : yvalue);
2110 answer_if_false = (want_max ? yvalue : xvalue);
2111 break;
2112 }
2113
2114 jint hi, lo;
2115 if (want_max) {
2116 // We can sharpen the minimum.
2117 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2118 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2119 } else {
2120 // We can sharpen the maximum.
2121 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2122 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2123 }
2124
2125 // Use a flow-free graph structure, to avoid creating excess control edges
2126 // which could hinder other optimizations.
2127 // Since Math.min/max is often used with arraycopy, we want
2128 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2129 Node* cmov = CMoveNode::make(NULL, best_bol,
2130 answer_if_false, answer_if_true,
2131 TypeInt::make(lo, hi, widen));
2132
2133 return _gvn.transform(cmov);
2134
2135 /*
2136 // This is not as desirable as it may seem, since Min and Max
2137 // nodes do not have a full set of optimizations.
2138 // And they would interfere, anyway, with 'if' optimizations
2139 // and with CMoveI canonical forms.
2140 switch (id) {
2141 case vmIntrinsics::_min:
2142 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2143 case vmIntrinsics::_max:
2144 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2145 default:
2146 ShouldNotReachHere();
2147 }
2148 */
2149 }
2150
2151 inline int
2152 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2153 const TypePtr* base_type = TypePtr::NULL_PTR;
2154 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2155 if (base_type == NULL) {
2156 // Unknown type.
2157 return Type::AnyPtr;
2158 } else if (base_type == TypePtr::NULL_PTR) {
2159 // Since this is a NULL+long form, we have to switch to a rawptr.
2160 base = _gvn.transform(new CastX2PNode(offset));
2161 offset = MakeConX(0);
2162 return Type::RawPtr;
2163 } else if (base_type->base() == Type::RawPtr) {
2164 return Type::RawPtr;
2165 } else if (base_type->isa_oopptr()) {
2166 // Base is never null => always a heap address.
2167 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2168 return Type::OopPtr;
2169 }
2170 // Offset is small => always a heap address.
2171 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2172 if (offset_type != NULL &&
2173 base_type->offset() == 0 && // (should always be?)
2174 offset_type->_lo >= 0 &&
2175 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2176 return Type::OopPtr;
2177 } else if (type == T_OBJECT) {
2178 // off heap access to an oop doesn't make any sense. Has to be on
2179 // heap.
2180 return Type::OopPtr;
2181 }
2182 // Otherwise, it might either be oop+off or NULL+addr.
2183 return Type::AnyPtr;
2184 } else {
2185 // No information:
2186 return Type::AnyPtr;
2187 }
2188 }
2189
2190 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2191 Node* uncasted_base = base;
2192 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2193 if (kind == Type::RawPtr) {
2194 return basic_plus_adr(top(), uncasted_base, offset);
2195 } else if (kind == Type::AnyPtr) {
2196 assert(base == uncasted_base, "unexpected base change");
2197 if (can_cast) {
2198 if (!_gvn.type(base)->speculative_maybe_null() &&
2199 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2200 // According to profiling, this access is always on
2201 // heap. Casting the base to not null and thus avoiding membars
2202 // around the access should allow better optimizations
2203 Node* null_ctl = top();
2204 base = null_check_oop(base, &null_ctl, true, true, true);
2205 assert(null_ctl->is_top(), "no null control here");
2206 return basic_plus_adr(base, offset);
2207 } else if (_gvn.type(base)->speculative_always_null() &&
2208 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2209 // According to profiling, this access is always off
2210 // heap.
2211 base = null_assert(base);
2212 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2213 offset = MakeConX(0);
2214 return basic_plus_adr(top(), raw_base, offset);
2215 }
2216 }
2217 // We don't know if it's an on heap or off heap access. Fall back
2218 // to raw memory access.
2219 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2220 return basic_plus_adr(top(), raw, offset);
2221 } else {
2222 assert(base == uncasted_base, "unexpected base change");
2223 // We know it's an on heap access so base can't be null
2224 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2225 base = must_be_not_null(base, true);
2226 }
2227 return basic_plus_adr(base, offset);
2228 }
2229 }
2230
2231 //--------------------------inline_number_methods-----------------------------
2232 // inline int Integer.numberOfLeadingZeros(int)
2233 // inline int Long.numberOfLeadingZeros(long)
2234 //
2235 // inline int Integer.numberOfTrailingZeros(int)
2236 // inline int Long.numberOfTrailingZeros(long)
2237 //
2238 // inline int Integer.bitCount(int)
2239 // inline int Long.bitCount(long)
2240 //
2241 // inline char Character.reverseBytes(char)
2242 // inline short Short.reverseBytes(short)
2243 // inline int Integer.reverseBytes(int)
2244 // inline long Long.reverseBytes(long)
2245 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2246 Node* arg = argument(0);
2247 Node* n = NULL;
2248 switch (id) {
2249 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2250 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2251 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2252 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2253 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2254 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2255 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2256 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2257 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2258 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2259 default: fatal_unexpected_iid(id); break;
2260 }
2261 set_result(_gvn.transform(n));
2262 return true;
2263 }
2264
2265 //----------------------------inline_unsafe_access----------------------------
2266
2267 // Helper that guards and inserts a pre-barrier.
2268 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2269 Node* pre_val, bool need_mem_bar) {
2270 // We could be accessing the referent field of a reference object. If so, when G1
2271 // is enabled, we need to log the value in the referent field in an SATB buffer.
2272 // This routine performs some compile time filters and generates suitable
2273 // runtime filters that guard the pre-barrier code.
2274 // Also add memory barrier for non volatile load from the referent field
2275 // to prevent commoning of loads across safepoint.
2276 if (!UseG1GC && !need_mem_bar)
2277 return;
2278
2279 // Some compile time checks.
2280
2281 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2282 const TypeX* otype = offset->find_intptr_t_type();
2283 if (otype != NULL && otype->is_con() &&
2284 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2285 // Constant offset but not the reference_offset so just return
2286 return;
2287 }
2288
2289 // We only need to generate the runtime guards for instances.
2290 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2291 if (btype != NULL) {
2292 if (btype->isa_aryptr()) {
2293 // Array type so nothing to do
2294 return;
2295 }
2296
2297 const TypeInstPtr* itype = btype->isa_instptr();
2298 if (itype != NULL) {
2299 // Can the klass of base_oop be statically determined to be
2300 // _not_ a sub-class of Reference and _not_ Object?
2301 ciKlass* klass = itype->klass();
2302 if ( klass->is_loaded() &&
2303 !klass->is_subtype_of(env()->Reference_klass()) &&
2304 !env()->Object_klass()->is_subtype_of(klass)) {
2305 return;
2306 }
2307 }
2308 }
2309
2310 // The compile time filters did not reject base_oop/offset so
2311 // we need to generate the following runtime filters
2312 //
2313 // if (offset == java_lang_ref_Reference::_reference_offset) {
2314 // if (instance_of(base, java.lang.ref.Reference)) {
2315 // pre_barrier(_, pre_val, ...);
2316 // }
2317 // }
2318
2319 float likely = PROB_LIKELY( 0.999);
2320 float unlikely = PROB_UNLIKELY(0.999);
2321
2322 IdealKit ideal(this);
2323 #define __ ideal.
2324
2325 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2326
2327 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2328 // Update graphKit memory and control from IdealKit.
2329 sync_kit(ideal);
2330
2331 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2332 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2333
2334 // Update IdealKit memory and control from graphKit.
2335 __ sync_kit(this);
2336
2337 Node* one = __ ConI(1);
2338 // is_instof == 0 if base_oop == NULL
2339 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2340
2341 // Update graphKit from IdeakKit.
2342 sync_kit(ideal);
2343
2344 // Use the pre-barrier to record the value in the referent field
2345 pre_barrier(false /* do_load */,
2346 __ ctrl(),
2347 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2348 pre_val /* pre_val */,
2349 T_OBJECT);
2350 if (need_mem_bar) {
2351 // Add memory barrier to prevent commoning reads from this field
2352 // across safepoint since GC can change its value.
2353 insert_mem_bar(Op_MemBarCPUOrder);
2354 }
2355 // Update IdealKit from graphKit.
2356 __ sync_kit(this);
2357
2358 } __ end_if(); // _ref_type != ref_none
2359 } __ end_if(); // offset == referent_offset
2360
2361 // Final sync IdealKit and GraphKit.
2362 final_sync(ideal);
2363 #undef __
2364 }
2365
2366
2367 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2368 // Attempt to infer a sharper value type from the offset and base type.
2369 ciKlass* sharpened_klass = NULL;
2370
2371 // See if it is an instance field, with an object type.
2372 if (alias_type->field() != NULL) {
2373 if (alias_type->field()->type()->is_klass()) {
2374 sharpened_klass = alias_type->field()->type()->as_klass();
2375 }
2376 }
2377
2378 // See if it is a narrow oop array.
2379 if (adr_type->isa_aryptr()) {
2380 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2381 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2382 if (elem_type != NULL) {
2383 sharpened_klass = elem_type->klass();
2384 }
2385 }
2386 }
2387
2388 // The sharpened class might be unloaded if there is no class loader
2389 // contraint in place.
2390 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2391 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2392
2393 #ifndef PRODUCT
2394 if (C->print_intrinsics() || C->print_inlining()) {
2395 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2396 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2397 }
2398 #endif
2399 // Sharpen the value type.
2400 return tjp;
2401 }
2402 return NULL;
2403 }
2404
2405 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2406 if (callee()->is_static()) return false; // caller must have the capability!
2407 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2408 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2409 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2410
2411 #ifndef PRODUCT
2412 {
2413 ResourceMark rm;
2414 // Check the signatures.
2415 ciSignature* sig = callee()->signature();
2416 #ifdef ASSERT
2417 if (!is_store) {
2418 // Object getObject(Object base, int/long offset), etc.
2419 BasicType rtype = sig->return_type()->basic_type();
2420 assert(rtype == type, "getter must return the expected value");
2421 assert(sig->count() == 2, "oop getter has 2 arguments");
2422 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2423 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2424 } else {
2425 // void putObject(Object base, int/long offset, Object x), etc.
2426 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2427 assert(sig->count() == 3, "oop putter has 3 arguments");
2428 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2429 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2430 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2431 assert(vtype == type, "putter must accept the expected value");
2432 }
2433 #endif // ASSERT
2434 }
2435 #endif //PRODUCT
2436
2437 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2438
2439 Node* receiver = argument(0); // type: oop
2440
2441 // Build address expression.
2442 Node* adr;
2443 Node* heap_base_oop = top();
2444 Node* offset = top();
2445 Node* val;
2446
2447 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2448 Node* base = argument(1); // type: oop
2449 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2450 offset = argument(2); // type: long
2451 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2452 // to be plain byte offsets, which are also the same as those accepted
2453 // by oopDesc::field_addr.
2454 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2455 "fieldOffset must be byte-scaled");
2456 // 32-bit machines ignore the high half!
2457 offset = ConvL2X(offset);
2458 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2459
2460 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2461 heap_base_oop = base;
2462 } else if (type == T_OBJECT) {
2463 return false; // off-heap oop accesses are not supported
2464 }
2465
2466 // Can base be NULL? Otherwise, always on-heap access.
2467 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2468
2469 val = is_store ? argument(4) : NULL;
2470
2471 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2472
2473 // Try to categorize the address.
2474 Compile::AliasType* alias_type = C->alias_type(adr_type);
2475 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2476
2477 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2478 alias_type->adr_type() == TypeAryPtr::RANGE) {
2479 return false; // not supported
2480 }
2481
2482 bool mismatched = false;
2483 BasicType bt = alias_type->basic_type();
2484 if (bt != T_ILLEGAL) {
2485 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2486 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2487 // Alias type doesn't differentiate between byte[] and boolean[]).
2488 // Use address type to get the element type.
2489 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2490 }
2491 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2492 // accessing an array field with getObject is not a mismatch
2493 bt = T_OBJECT;
2494 }
2495 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2496 // Don't intrinsify mismatched object accesses
2497 return false;
2498 }
2499 mismatched = (bt != type);
2500 } else if (alias_type->adr_type()->isa_oopptr()) {
2501 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2502 }
2503
2504 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2505
2506 // First guess at the value type.
2507 const Type *value_type = Type::get_const_basic_type(type);
2508
2509 // We will need memory barriers unless we can determine a unique
2510 // alias category for this reference. (Note: If for some reason
2511 // the barriers get omitted and the unsafe reference begins to "pollute"
2512 // the alias analysis of the rest of the graph, either Compile::can_alias
2513 // or Compile::must_alias will throw a diagnostic assert.)
2514 bool need_mem_bar = false;
2515 switch (kind) {
2516 case Relaxed:
2517 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2518 break;
2519 case Opaque:
2520 // Opaque uses CPUOrder membars for protection against code movement.
2521 case Acquire:
2522 case Release:
2523 case Volatile:
2524 need_mem_bar = true;
2525 break;
2526 default:
2527 ShouldNotReachHere();
2528 }
2529
2530 // Some accesses require access atomicity for all types, notably longs and doubles.
2531 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2532 bool requires_atomic_access = false;
2533 switch (kind) {
2534 case Relaxed:
2535 requires_atomic_access = AlwaysAtomicAccesses;
2536 break;
2537 case Opaque:
2538 // Opaque accesses are atomic.
2539 case Acquire:
2540 case Release:
2541 case Volatile:
2542 requires_atomic_access = true;
2543 break;
2544 default:
2545 ShouldNotReachHere();
2546 }
2547
2548 // Figure out the memory ordering.
2549 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2550 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2551
2552 // If we are reading the value of the referent field of a Reference
2553 // object (either by using Unsafe directly or through reflection)
2554 // then, if G1 is enabled, we need to record the referent in an
2555 // SATB log buffer using the pre-barrier mechanism.
2556 // Also we need to add memory barrier to prevent commoning reads
2557 // from this field across safepoint since GC can change its value.
2558 bool need_read_barrier = !is_store &&
2559 offset != top() && heap_base_oop != top();
2560
2561 if (!is_store && type == T_OBJECT) {
2562 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2563 if (tjp != NULL) {
2564 value_type = tjp;
2565 }
2566 }
2567
2568 receiver = null_check(receiver);
2569 if (stopped()) {
2570 return true;
2571 }
2572 // Heap pointers get a null-check from the interpreter,
2573 // as a courtesy. However, this is not guaranteed by Unsafe,
2574 // and it is not possible to fully distinguish unintended nulls
2575 // from intended ones in this API.
2576
2577 // We need to emit leading and trailing CPU membars (see below) in
2578 // addition to memory membars for special access modes. This is a little
2579 // too strong, but avoids the need to insert per-alias-type
2580 // volatile membars (for stores; compare Parse::do_put_xxx), which
2581 // we cannot do effectively here because we probably only have a
2582 // rough approximation of type.
2583
2584 switch(kind) {
2585 case Relaxed:
2586 case Opaque:
2587 case Acquire:
2588 break;
2589 case Release:
2590 case Volatile:
2591 if (is_store) {
2592 insert_mem_bar(Op_MemBarRelease);
2593 } else {
2594 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2595 insert_mem_bar(Op_MemBarVolatile);
2596 }
2597 }
2598 break;
2599 default:
2600 ShouldNotReachHere();
2601 }
2602
2603 // Memory barrier to prevent normal and 'unsafe' accesses from
2604 // bypassing each other. Happens after null checks, so the
2605 // exception paths do not take memory state from the memory barrier,
2606 // so there's no problems making a strong assert about mixing users
2607 // of safe & unsafe memory.
2608 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2609
2610 if (!is_store) {
2611 Node* p = NULL;
2612 // Try to constant fold a load from a constant field
2613 ciField* field = alias_type->field();
2614 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2615 // final or stable field
2616 p = make_constant_from_field(field, heap_base_oop);
2617 }
2618 if (p == NULL) {
2619 // To be valid, unsafe loads may depend on other conditions than
2620 // the one that guards them: pin the Load node
2621 LoadNode::ControlDependency dep = LoadNode::Pinned;
2622 Node* ctrl = control();
2623 if (adr_type->isa_instptr()) {
2624 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2625 intptr_t offset = Type::OffsetBot;
2626 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2627 if (offset >= 0) {
2628 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2629 if (offset < s) {
2630 // Guaranteed to be a valid access, no need to pin it
2631 dep = LoadNode::DependsOnlyOnTest;
2632 ctrl = NULL;
2633 }
2634 }
2635 }
2636 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2637 // load value
2638 switch (type) {
2639 case T_BOOLEAN:
2640 {
2641 // Normalize the value returned by getBoolean in the following cases
2642 if (mismatched ||
2643 heap_base_oop == top() || // - heap_base_oop is NULL or
2644 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2645 // and the unsafe access is made to large offset
2646 // (i.e., larger than the maximum offset necessary for any
2647 // field access)
2648 ) {
2649 IdealKit ideal = IdealKit(this);
2650 #define __ ideal.
2651 IdealVariable normalized_result(ideal);
2652 __ declarations_done();
2653 __ set(normalized_result, p);
2654 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2655 __ set(normalized_result, ideal.ConI(1));
2656 ideal.end_if();
2657 final_sync(ideal);
2658 p = __ value(normalized_result);
2659 #undef __
2660 }
2661 }
2662 case T_CHAR:
2663 case T_BYTE:
2664 case T_SHORT:
2665 case T_INT:
2666 case T_LONG:
2667 case T_FLOAT:
2668 case T_DOUBLE:
2669 break;
2670 case T_OBJECT:
2671 if (need_read_barrier) {
2672 // We do not require a mem bar inside pre_barrier if need_mem_bar
2673 // is set: the barriers would be emitted by us.
2674 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2675 }
2676 break;
2677 case T_ADDRESS:
2678 // Cast to an int type.
2679 p = _gvn.transform(new CastP2XNode(NULL, p));
2680 p = ConvX2UL(p);
2681 break;
2682 default:
2683 fatal("unexpected type %d: %s", type, type2name(type));
2684 break;
2685 }
2686 }
2687 // The load node has the control of the preceding MemBarCPUOrder. All
2688 // following nodes will have the control of the MemBarCPUOrder inserted at
2689 // the end of this method. So, pushing the load onto the stack at a later
2690 // point is fine.
2691 set_result(p);
2692 } else {
2693 // place effect of store into memory
2694 switch (type) {
2695 case T_DOUBLE:
2696 val = dstore_rounding(val);
2697 break;
2698 case T_ADDRESS:
2699 // Repackage the long as a pointer.
2700 val = ConvL2X(val);
2701 val = _gvn.transform(new CastX2PNode(val));
2702 break;
2703 default:
2704 break;
2705 }
2706
2707 if (type == T_OBJECT) {
2708 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2709 } else {
2710 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2711 }
2712 }
2713
2714 switch(kind) {
2715 case Relaxed:
2716 case Opaque:
2717 case Release:
2718 break;
2719 case Acquire:
2720 case Volatile:
2721 if (!is_store) {
2722 insert_mem_bar(Op_MemBarAcquire);
2723 } else {
2724 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2725 insert_mem_bar(Op_MemBarVolatile);
2726 }
2727 }
2728 break;
2729 default:
2730 ShouldNotReachHere();
2731 }
2732
2733 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2734
2735 return true;
2736 }
2737
2738 //----------------------------inline_unsafe_load_store----------------------------
2739 // This method serves a couple of different customers (depending on LoadStoreKind):
2740 //
2741 // LS_cmp_swap:
2742 //
2743 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2744 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2745 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2746 //
2747 // LS_cmp_swap_weak:
2748 //
2749 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2750 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2751 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2752 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2753 //
2754 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2755 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2756 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2757 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2758 //
2759 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2760 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2761 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2762 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2763 //
2764 // LS_cmp_exchange:
2765 //
2766 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2767 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2768 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2769 //
2770 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2771 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2772 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2773 //
2774 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2775 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2776 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2777 //
2778 // LS_get_add:
2779 //
2780 // int getAndAddInt( Object o, long offset, int delta)
2781 // long getAndAddLong(Object o, long offset, long delta)
2782 //
2783 // LS_get_set:
2784 //
2785 // int getAndSet(Object o, long offset, int newValue)
2786 // long getAndSet(Object o, long offset, long newValue)
2787 // Object getAndSet(Object o, long offset, Object newValue)
2788 //
2789 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2790 // This basic scheme here is the same as inline_unsafe_access, but
2791 // differs in enough details that combining them would make the code
2792 // overly confusing. (This is a true fact! I originally combined
2793 // them, but even I was confused by it!) As much code/comments as
2794 // possible are retained from inline_unsafe_access though to make
2795 // the correspondences clearer. - dl
2796
2797 if (callee()->is_static()) return false; // caller must have the capability!
2798
2799 #ifndef PRODUCT
2800 BasicType rtype;
2801 {
2802 ResourceMark rm;
2803 // Check the signatures.
2804 ciSignature* sig = callee()->signature();
2805 rtype = sig->return_type()->basic_type();
2806 switch(kind) {
2807 case LS_get_add:
2808 case LS_get_set: {
2809 // Check the signatures.
2810 #ifdef ASSERT
2811 assert(rtype == type, "get and set must return the expected type");
2812 assert(sig->count() == 3, "get and set has 3 arguments");
2813 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2814 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2815 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2816 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2817 #endif // ASSERT
2818 break;
2819 }
2820 case LS_cmp_swap:
2821 case LS_cmp_swap_weak: {
2822 // Check the signatures.
2823 #ifdef ASSERT
2824 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2825 assert(sig->count() == 4, "CAS has 4 arguments");
2826 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2827 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2828 #endif // ASSERT
2829 break;
2830 }
2831 case LS_cmp_exchange: {
2832 // Check the signatures.
2833 #ifdef ASSERT
2834 assert(rtype == type, "CAS must return the expected type");
2835 assert(sig->count() == 4, "CAS has 4 arguments");
2836 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2837 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2838 #endif // ASSERT
2839 break;
2840 }
2841 default:
2842 ShouldNotReachHere();
2843 }
2844 }
2845 #endif //PRODUCT
2846
2847 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2848
2849 // Get arguments:
2850 Node* receiver = NULL;
2851 Node* base = NULL;
2852 Node* offset = NULL;
2853 Node* oldval = NULL;
2854 Node* newval = NULL;
2855 switch(kind) {
2856 case LS_cmp_swap:
2857 case LS_cmp_swap_weak:
2858 case LS_cmp_exchange: {
2859 const bool two_slot_type = type2size[type] == 2;
2860 receiver = argument(0); // type: oop
2861 base = argument(1); // type: oop
2862 offset = argument(2); // type: long
2863 oldval = argument(4); // type: oop, int, or long
2864 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2865 break;
2866 }
2867 case LS_get_add:
2868 case LS_get_set: {
2869 receiver = argument(0); // type: oop
2870 base = argument(1); // type: oop
2871 offset = argument(2); // type: long
2872 oldval = NULL;
2873 newval = argument(4); // type: oop, int, or long
2874 break;
2875 }
2876 default:
2877 ShouldNotReachHere();
2878 }
2879
2880 // Build field offset expression.
2881 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2882 // to be plain byte offsets, which are also the same as those accepted
2883 // by oopDesc::field_addr.
2884 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2885 // 32-bit machines ignore the high half of long offsets
2886 offset = ConvL2X(offset);
2887 Node* adr = make_unsafe_address(base, offset, type, false);
2888 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2889
2890 Compile::AliasType* alias_type = C->alias_type(adr_type);
2891 BasicType bt = alias_type->basic_type();
2892 if (bt != T_ILLEGAL &&
2893 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2894 // Don't intrinsify mismatched object accesses.
2895 return false;
2896 }
2897
2898 // For CAS, unlike inline_unsafe_access, there seems no point in
2899 // trying to refine types. Just use the coarse types here.
2900 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2901 const Type *value_type = Type::get_const_basic_type(type);
2902
2903 switch (kind) {
2904 case LS_get_set:
2905 case LS_cmp_exchange: {
2906 if (type == T_OBJECT) {
2907 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2908 if (tjp != NULL) {
2909 value_type = tjp;
2910 }
2911 }
2912 break;
2913 }
2914 case LS_cmp_swap:
2915 case LS_cmp_swap_weak:
2916 case LS_get_add:
2917 break;
2918 default:
2919 ShouldNotReachHere();
2920 }
2921
2922 // Null check receiver.
2923 receiver = null_check(receiver);
2924 if (stopped()) {
2925 return true;
2926 }
2927
2928 int alias_idx = C->get_alias_index(adr_type);
2929
2930 // Memory-model-wise, a LoadStore acts like a little synchronized
2931 // block, so needs barriers on each side. These don't translate
2932 // into actual barriers on most machines, but we still need rest of
2933 // compiler to respect ordering.
2934
2935 switch (access_kind) {
2936 case Relaxed:
2937 case Acquire:
2938 break;
2939 case Release:
2940 insert_mem_bar(Op_MemBarRelease);
2941 break;
2942 case Volatile:
2943 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2944 insert_mem_bar(Op_MemBarVolatile);
2945 } else {
2946 insert_mem_bar(Op_MemBarRelease);
2947 }
2948 break;
2949 default:
2950 ShouldNotReachHere();
2951 }
2952 insert_mem_bar(Op_MemBarCPUOrder);
2953
2954 // Figure out the memory ordering.
2955 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2956
2957 // 4984716: MemBars must be inserted before this
2958 // memory node in order to avoid a false
2959 // dependency which will confuse the scheduler.
2960 Node *mem = memory(alias_idx);
2961
2962 // For now, we handle only those cases that actually exist: ints,
2963 // longs, and Object. Adding others should be straightforward.
2964 Node* load_store = NULL;
2965 switch(type) {
2966 case T_BYTE:
2967 switch(kind) {
2968 case LS_get_add:
2969 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2970 break;
2971 case LS_get_set:
2972 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2973 break;
2974 case LS_cmp_swap_weak:
2975 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2976 break;
2977 case LS_cmp_swap:
2978 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2979 break;
2980 case LS_cmp_exchange:
2981 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2982 break;
2983 default:
2984 ShouldNotReachHere();
2985 }
2986 break;
2987 case T_SHORT:
2988 switch(kind) {
2989 case LS_get_add:
2990 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2991 break;
2992 case LS_get_set:
2993 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2994 break;
2995 case LS_cmp_swap_weak:
2996 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2997 break;
2998 case LS_cmp_swap:
2999 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
3000 break;
3001 case LS_cmp_exchange:
3002 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
3003 break;
3004 default:
3005 ShouldNotReachHere();
3006 }
3007 break;
3008 case T_INT:
3009 switch(kind) {
3010 case LS_get_add:
3011 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
3012 break;
3013 case LS_get_set:
3014 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
3015 break;
3016 case LS_cmp_swap_weak:
3017 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
3018 break;
3019 case LS_cmp_swap:
3020 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
3021 break;
3022 case LS_cmp_exchange:
3023 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
3024 break;
3025 default:
3026 ShouldNotReachHere();
3027 }
3028 break;
3029 case T_LONG:
3030 switch(kind) {
3031 case LS_get_add:
3032 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
3033 break;
3034 case LS_get_set:
3035 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
3036 break;
3037 case LS_cmp_swap_weak:
3038 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3039 break;
3040 case LS_cmp_swap:
3041 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3042 break;
3043 case LS_cmp_exchange:
3044 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
3045 break;
3046 default:
3047 ShouldNotReachHere();
3048 }
3049 break;
3050 case T_OBJECT:
3051 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3052 // could be delayed during Parse (for example, in adjust_map_after_if()).
3053 // Execute transformation here to avoid barrier generation in such case.
3054 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3055 newval = _gvn.makecon(TypePtr::NULL_PTR);
3056
3057 // Reference stores need a store barrier.
3058 switch(kind) {
3059 case LS_get_set: {
3060 // If pre-barrier must execute before the oop store, old value will require do_load here.
3061 if (!can_move_pre_barrier()) {
3062 pre_barrier(true /* do_load*/,
3063 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3064 NULL /* pre_val*/,
3065 T_OBJECT);
3066 } // Else move pre_barrier to use load_store value, see below.
3067 break;
3068 }
3069 case LS_cmp_swap_weak:
3070 case LS_cmp_swap:
3071 case LS_cmp_exchange: {
3072 // Same as for newval above:
3073 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3074 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3075 }
3076 // The only known value which might get overwritten is oldval.
3077 pre_barrier(false /* do_load */,
3078 control(), NULL, NULL, max_juint, NULL, NULL,
3079 oldval /* pre_val */,
3080 T_OBJECT);
3081 break;
3082 }
3083 default:
3084 ShouldNotReachHere();
3085 }
3086
3087 #ifdef _LP64
3088 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3089 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3090
3091 switch(kind) {
3092 case LS_get_set:
3093 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3094 break;
3095 case LS_cmp_swap_weak: {
3096 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3097 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3098 break;
3099 }
3100 case LS_cmp_swap: {
3101 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3102 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3103 break;
3104 }
3105 case LS_cmp_exchange: {
3106 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3107 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3108 break;
3109 }
3110 default:
3111 ShouldNotReachHere();
3112 }
3113 } else
3114 #endif
3115 switch (kind) {
3116 case LS_get_set:
3117 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3118 break;
3119 case LS_cmp_swap_weak:
3120 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3121 break;
3122 case LS_cmp_swap:
3123 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3124 break;
3125 case LS_cmp_exchange:
3126 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3127 break;
3128 default:
3129 ShouldNotReachHere();
3130 }
3131
3132 // Emit the post barrier only when the actual store happened. This makes sense
3133 // to check only for LS_cmp_* that can fail to set the value.
3134 // LS_cmp_exchange does not produce any branches by default, so there is no
3135 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3136 // CompareAndExchange and move branches here, it would make sense to conditionalize
3137 // post_barriers for LS_cmp_exchange as well.
3138 //
3139 // CAS success path is marked more likely since we anticipate this is a performance
3140 // critical path, while CAS failure path can use the penalty for going through unlikely
3141 // path as backoff. Which is still better than doing a store barrier there.
3142 switch (kind) {
3143 case LS_get_set:
3144 case LS_cmp_exchange: {
3145 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3146 break;
3147 }
3148 case LS_cmp_swap_weak:
3149 case LS_cmp_swap: {
3150 IdealKit ideal(this);
3151 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3152 sync_kit(ideal);
3153 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3154 ideal.sync_kit(this);
3155 } ideal.end_if();
3156 final_sync(ideal);
3157 break;
3158 }
3159 default:
3160 ShouldNotReachHere();
3161 }
3162 break;
3163 default:
3164 fatal("unexpected type %d: %s", type, type2name(type));
3165 break;
3166 }
3167
3168 // SCMemProjNodes represent the memory state of a LoadStore. Their
3169 // main role is to prevent LoadStore nodes from being optimized away
3170 // when their results aren't used.
3171 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3172 set_memory(proj, alias_idx);
3173
3174 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3175 #ifdef _LP64
3176 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3177 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3178 }
3179 #endif
3180 if (can_move_pre_barrier() && kind == LS_get_set) {
3181 // Don't need to load pre_val. The old value is returned by load_store.
3182 // The pre_barrier can execute after the xchg as long as no safepoint
3183 // gets inserted between them.
3184 pre_barrier(false /* do_load */,
3185 control(), NULL, NULL, max_juint, NULL, NULL,
3186 load_store /* pre_val */,
3187 T_OBJECT);
3188 }
3189 }
3190
3191 // Add the trailing membar surrounding the access
3192 insert_mem_bar(Op_MemBarCPUOrder);
3193
3194 switch (access_kind) {
3195 case Relaxed:
3196 case Release:
3197 break; // do nothing
3198 case Acquire:
3199 case Volatile:
3200 insert_mem_bar(Op_MemBarAcquire);
3201 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3202 break;
3203 default:
3204 ShouldNotReachHere();
3205 }
3206
3207 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3208 set_result(load_store);
3209 return true;
3210 }
3211
3212 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3213 MemNode::MemOrd mo = MemNode::unset;
3214 switch(kind) {
3215 case Opaque:
3216 case Relaxed: mo = MemNode::unordered; break;
3217 case Acquire: mo = MemNode::acquire; break;
3218 case Release: mo = MemNode::release; break;
3219 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3220 default:
3221 ShouldNotReachHere();
3222 }
3223 guarantee(mo != MemNode::unset, "Should select memory ordering");
3224 return mo;
3225 }
3226
3227 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3228 MemNode::MemOrd mo = MemNode::unset;
3229 switch(kind) {
3230 case Opaque:
3231 case Relaxed: mo = MemNode::unordered; break;
3232 case Acquire: mo = MemNode::acquire; break;
3233 case Release: mo = MemNode::release; break;
3234 case Volatile: mo = MemNode::seqcst; break;
3235 default:
3236 ShouldNotReachHere();
3237 }
3238 guarantee(mo != MemNode::unset, "Should select memory ordering");
3239 return mo;
3240 }
3241
3242 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3243 // Regardless of form, don't allow previous ld/st to move down,
3244 // then issue acquire, release, or volatile mem_bar.
3245 insert_mem_bar(Op_MemBarCPUOrder);
3246 switch(id) {
3247 case vmIntrinsics::_loadFence:
3248 insert_mem_bar(Op_LoadFence);
3249 return true;
3250 case vmIntrinsics::_storeFence:
3251 insert_mem_bar(Op_StoreFence);
3252 return true;
3253 case vmIntrinsics::_fullFence:
3254 insert_mem_bar(Op_MemBarVolatile);
3255 return true;
3256 default:
3257 fatal_unexpected_iid(id);
3258 return false;
3259 }
3260 }
3261
3262 bool LibraryCallKit::inline_onspinwait() {
3263 insert_mem_bar(Op_OnSpinWait);
3264 return true;
3265 }
3266
3267 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3268 if (!kls->is_Con()) {
3269 return true;
3270 }
3271 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3272 if (klsptr == NULL) {
3273 return true;
3274 }
3275 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3276 // don't need a guard for a klass that is already initialized
3277 return !ik->is_initialized();
3278 }
3279
3280 //----------------------------inline_unsafe_allocate---------------------------
3281 // public native Object Unsafe.allocateInstance(Class<?> cls);
3282 bool LibraryCallKit::inline_unsafe_allocate() {
3283 if (callee()->is_static()) return false; // caller must have the capability!
3284
3285 null_check_receiver(); // null-check, then ignore
3286 Node* cls = null_check(argument(1));
3287 if (stopped()) return true;
3288
3289 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3290 kls = null_check(kls);
3291 if (stopped()) return true; // argument was like int.class
3292
3293 Node* test = NULL;
3294 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3295 // Note: The argument might still be an illegal value like
3296 // Serializable.class or Object[].class. The runtime will handle it.
3297 // But we must make an explicit check for initialization.
3298 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3299 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3300 // can generate code to load it as unsigned byte.
3301 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3302 Node* bits = intcon(InstanceKlass::fully_initialized);
3303 test = _gvn.transform(new SubINode(inst, bits));
3304 // The 'test' is non-zero if we need to take a slow path.
3305 }
3306
3307 Node* obj = new_instance(kls, test);
3308 set_result(obj);
3309 return true;
3310 }
3311
3312 //------------------------inline_native_time_funcs--------------
3313 // inline code for System.currentTimeMillis() and System.nanoTime()
3314 // these have the same type and signature
3315 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3316 const TypeFunc* tf = OptoRuntime::void_long_Type();
3317 const TypePtr* no_memory_effects = NULL;
3318 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3319 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3320 #ifdef ASSERT
3321 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3322 assert(value_top == top(), "second value must be top");
3323 #endif
3324 set_result(value);
3325 return true;
3326 }
3327
3328 #ifdef TRACE_HAVE_INTRINSICS
3329
3330 /*
3331 * oop -> myklass
3332 * myklass->trace_id |= USED
3333 * return myklass->trace_id & ~0x3
3334 */
3335 bool LibraryCallKit::inline_native_classID() {
3336 Node* cls = null_check(argument(0), T_OBJECT);
3337 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3338 kls = null_check(kls, T_OBJECT);
3339
3340 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3341 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3342 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3343
3344 Node* clsused = longcon(0x01l); // set the class bit
3345 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3346 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3347 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3348
3349 #ifdef TRACE_ID_META_BITS
3350 Node* mbits = longcon(~TRACE_ID_META_BITS);
3351 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3352 #endif
3353 #ifdef TRACE_ID_CLASS_SHIFT
3354 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3355 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3356 #endif
3357
3358 set_result(tvalue);
3359 return true;
3360
3361 }
3362
3363 bool LibraryCallKit::inline_native_getBufferWriter() {
3364 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3365
3366 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3367 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3368 );
3369
3370 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3371
3372 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3373 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3374
3375 IfNode* iff_jobj_null =
3376 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3377
3378 enum { _normal_path = 1,
3379 _null_path = 2,
3380 PATH_LIMIT };
3381
3382 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3383 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3384
3385 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3386 result_rgn->init_req(_null_path, jobj_is_null);
3387 result_val->init_req(_null_path, null());
3388
3389 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3390 result_rgn->init_req(_normal_path, jobj_is_not_null);
3391
3392 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3393 result_val->init_req(_normal_path, res);
3394
3395 set_result(result_rgn, result_val);
3396
3397 return true;
3398 }
3399
3400 #endif
3401
3402 //------------------------inline_native_currentThread------------------
3403 bool LibraryCallKit::inline_native_currentThread() {
3404 Node* junk = NULL;
3405 set_result(generate_current_thread(junk));
3406 return true;
3407 }
3408
3409 //------------------------inline_native_isInterrupted------------------
3410 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3411 bool LibraryCallKit::inline_native_isInterrupted() {
3412 // Add a fast path to t.isInterrupted(clear_int):
3413 // (t == Thread.current() &&
3414 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3415 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3416 // So, in the common case that the interrupt bit is false,
3417 // we avoid making a call into the VM. Even if the interrupt bit
3418 // is true, if the clear_int argument is false, we avoid the VM call.
3419 // However, if the receiver is not currentThread, we must call the VM,
3420 // because there must be some locking done around the operation.
3421
3422 // We only go to the fast case code if we pass two guards.
3423 // Paths which do not pass are accumulated in the slow_region.
3424
3425 enum {
3426 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3427 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3428 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3429 PATH_LIMIT
3430 };
3431
3432 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3433 // out of the function.
3434 insert_mem_bar(Op_MemBarCPUOrder);
3435
3436 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3437 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3438
3439 RegionNode* slow_region = new RegionNode(1);
3440 record_for_igvn(slow_region);
3441
3442 // (a) Receiving thread must be the current thread.
3443 Node* rec_thr = argument(0);
3444 Node* tls_ptr = NULL;
3445 Node* cur_thr = generate_current_thread(tls_ptr);
3446 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3447 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3448
3449 generate_slow_guard(bol_thr, slow_region);
3450
3451 // (b) Interrupt bit on TLS must be false.
3452 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3453 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3454 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3455
3456 // Set the control input on the field _interrupted read to prevent it floating up.
3457 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3458 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3459 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3460
3461 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3462
3463 // First fast path: if (!TLS._interrupted) return false;
3464 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3465 result_rgn->init_req(no_int_result_path, false_bit);
3466 result_val->init_req(no_int_result_path, intcon(0));
3467
3468 // drop through to next case
3469 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3470
3471 #ifndef _WINDOWS
3472 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3473 Node* clr_arg = argument(1);
3474 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3475 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3476 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3477
3478 // Second fast path: ... else if (!clear_int) return true;
3479 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3480 result_rgn->init_req(no_clear_result_path, false_arg);
3481 result_val->init_req(no_clear_result_path, intcon(1));
3482
3483 // drop through to next case
3484 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3485 #else
3486 // To return true on Windows you must read the _interrupted field
3487 // and check the event state i.e. take the slow path.
3488 #endif // _WINDOWS
3489
3490 // (d) Otherwise, go to the slow path.
3491 slow_region->add_req(control());
3492 set_control( _gvn.transform(slow_region));
3493
3494 if (stopped()) {
3495 // There is no slow path.
3496 result_rgn->init_req(slow_result_path, top());
3497 result_val->init_req(slow_result_path, top());
3498 } else {
3499 // non-virtual because it is a private non-static
3500 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3501
3502 Node* slow_val = set_results_for_java_call(slow_call);
3503 // this->control() comes from set_results_for_java_call
3504
3505 Node* fast_io = slow_call->in(TypeFunc::I_O);
3506 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3507
3508 // These two phis are pre-filled with copies of of the fast IO and Memory
3509 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3510 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3511
3512 result_rgn->init_req(slow_result_path, control());
3513 result_io ->init_req(slow_result_path, i_o());
3514 result_mem->init_req(slow_result_path, reset_memory());
3515 result_val->init_req(slow_result_path, slow_val);
3516
3517 set_all_memory(_gvn.transform(result_mem));
3518 set_i_o( _gvn.transform(result_io));
3519 }
3520
3521 C->set_has_split_ifs(true); // Has chance for split-if optimization
3522 set_result(result_rgn, result_val);
3523 return true;
3524 }
3525
3526 //---------------------------load_mirror_from_klass----------------------------
3527 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3528 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3529 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3530 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3531 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3532 }
3533
3534 //-----------------------load_klass_from_mirror_common-------------------------
3535 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3536 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3537 // and branch to the given path on the region.
3538 // If never_see_null, take an uncommon trap on null, so we can optimistically
3539 // compile for the non-null case.
3540 // If the region is NULL, force never_see_null = true.
3541 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3542 bool never_see_null,
3543 RegionNode* region,
3544 int null_path,
3545 int offset) {
3546 if (region == NULL) never_see_null = true;
3547 Node* p = basic_plus_adr(mirror, offset);
3548 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3549 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3550 Node* null_ctl = top();
3551 kls = null_check_oop(kls, &null_ctl, never_see_null);
3552 if (region != NULL) {
3553 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3554 region->init_req(null_path, null_ctl);
3555 } else {
3556 assert(null_ctl == top(), "no loose ends");
3557 }
3558 return kls;
3559 }
3560
3561 //--------------------(inline_native_Class_query helpers)---------------------
3562 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3563 // Fall through if (mods & mask) == bits, take the guard otherwise.
3564 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3565 // Branch around if the given klass has the given modifier bit set.
3566 // Like generate_guard, adds a new path onto the region.
3567 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3568 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3569 Node* mask = intcon(modifier_mask);
3570 Node* bits = intcon(modifier_bits);
3571 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3572 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3573 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3574 return generate_fair_guard(bol, region);
3575 }
3576 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3577 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3578 }
3579
3580 //-------------------------inline_native_Class_query-------------------
3581 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3582 const Type* return_type = TypeInt::BOOL;
3583 Node* prim_return_value = top(); // what happens if it's a primitive class?
3584 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3585 bool expect_prim = false; // most of these guys expect to work on refs
3586
3587 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3588
3589 Node* mirror = argument(0);
3590 Node* obj = top();
3591
3592 switch (id) {
3593 case vmIntrinsics::_isInstance:
3594 // nothing is an instance of a primitive type
3595 prim_return_value = intcon(0);
3596 obj = argument(1);
3597 break;
3598 case vmIntrinsics::_getModifiers:
3599 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3600 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3601 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3602 break;
3603 case vmIntrinsics::_isInterface:
3604 prim_return_value = intcon(0);
3605 break;
3606 case vmIntrinsics::_isArray:
3607 prim_return_value = intcon(0);
3608 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3609 break;
3610 case vmIntrinsics::_isPrimitive:
3611 prim_return_value = intcon(1);
3612 expect_prim = true; // obviously
3613 break;
3614 case vmIntrinsics::_getSuperclass:
3615 prim_return_value = null();
3616 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3617 break;
3618 case vmIntrinsics::_getClassAccessFlags:
3619 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3620 return_type = TypeInt::INT; // not bool! 6297094
3621 break;
3622 default:
3623 fatal_unexpected_iid(id);
3624 break;
3625 }
3626
3627 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3628 if (mirror_con == NULL) return false; // cannot happen?
3629
3630 #ifndef PRODUCT
3631 if (C->print_intrinsics() || C->print_inlining()) {
3632 ciType* k = mirror_con->java_mirror_type();
3633 if (k) {
3634 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3635 k->print_name();
3636 tty->cr();
3637 }
3638 }
3639 #endif
3640
3641 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3642 RegionNode* region = new RegionNode(PATH_LIMIT);
3643 record_for_igvn(region);
3644 PhiNode* phi = new PhiNode(region, return_type);
3645
3646 // The mirror will never be null of Reflection.getClassAccessFlags, however
3647 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3648 // if it is. See bug 4774291.
3649
3650 // For Reflection.getClassAccessFlags(), the null check occurs in
3651 // the wrong place; see inline_unsafe_access(), above, for a similar
3652 // situation.
3653 mirror = null_check(mirror);
3654 // If mirror or obj is dead, only null-path is taken.
3655 if (stopped()) return true;
3656
3657 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3658
3659 // Now load the mirror's klass metaobject, and null-check it.
3660 // Side-effects region with the control path if the klass is null.
3661 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3662 // If kls is null, we have a primitive mirror.
3663 phi->init_req(_prim_path, prim_return_value);
3664 if (stopped()) { set_result(region, phi); return true; }
3665 bool safe_for_replace = (region->in(_prim_path) == top());
3666
3667 Node* p; // handy temp
3668 Node* null_ctl;
3669
3670 // Now that we have the non-null klass, we can perform the real query.
3671 // For constant classes, the query will constant-fold in LoadNode::Value.
3672 Node* query_value = top();
3673 switch (id) {
3674 case vmIntrinsics::_isInstance:
3675 // nothing is an instance of a primitive type
3676 query_value = gen_instanceof(obj, kls, safe_for_replace);
3677 break;
3678
3679 case vmIntrinsics::_getModifiers:
3680 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3681 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3682 break;
3683
3684 case vmIntrinsics::_isInterface:
3685 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3686 if (generate_interface_guard(kls, region) != NULL)
3687 // A guard was added. If the guard is taken, it was an interface.
3688 phi->add_req(intcon(1));
3689 // If we fall through, it's a plain class.
3690 query_value = intcon(0);
3691 break;
3692
3693 case vmIntrinsics::_isArray:
3694 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3695 if (generate_array_guard(kls, region) != NULL)
3696 // A guard was added. If the guard is taken, it was an array.
3697 phi->add_req(intcon(1));
3698 // If we fall through, it's a plain class.
3699 query_value = intcon(0);
3700 break;
3701
3702 case vmIntrinsics::_isPrimitive:
3703 query_value = intcon(0); // "normal" path produces false
3704 break;
3705
3706 case vmIntrinsics::_getSuperclass:
3707 // The rules here are somewhat unfortunate, but we can still do better
3708 // with random logic than with a JNI call.
3709 // Interfaces store null or Object as _super, but must report null.
3710 // Arrays store an intermediate super as _super, but must report Object.
3711 // Other types can report the actual _super.
3712 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3713 if (generate_interface_guard(kls, region) != NULL)
3714 // A guard was added. If the guard is taken, it was an interface.
3715 phi->add_req(null());
3716 if (generate_array_guard(kls, region) != NULL)
3717 // A guard was added. If the guard is taken, it was an array.
3718 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3719 // If we fall through, it's a plain class. Get its _super.
3720 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3721 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3722 null_ctl = top();
3723 kls = null_check_oop(kls, &null_ctl);
3724 if (null_ctl != top()) {
3725 // If the guard is taken, Object.superClass is null (both klass and mirror).
3726 region->add_req(null_ctl);
3727 phi ->add_req(null());
3728 }
3729 if (!stopped()) {
3730 query_value = load_mirror_from_klass(kls);
3731 }
3732 break;
3733
3734 case vmIntrinsics::_getClassAccessFlags:
3735 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3736 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3737 break;
3738
3739 default:
3740 fatal_unexpected_iid(id);
3741 break;
3742 }
3743
3744 // Fall-through is the normal case of a query to a real class.
3745 phi->init_req(1, query_value);
3746 region->init_req(1, control());
3747
3748 C->set_has_split_ifs(true); // Has chance for split-if optimization
3749 set_result(region, phi);
3750 return true;
3751 }
3752
3753 //-------------------------inline_Class_cast-------------------
3754 bool LibraryCallKit::inline_Class_cast() {
3755 Node* mirror = argument(0); // Class
3756 Node* obj = argument(1);
3757 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3758 if (mirror_con == NULL) {
3759 return false; // dead path (mirror->is_top()).
3760 }
3761 if (obj == NULL || obj->is_top()) {
3762 return false; // dead path
3763 }
3764 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3765
3766 // First, see if Class.cast() can be folded statically.
3767 // java_mirror_type() returns non-null for compile-time Class constants.
3768 ciType* tm = mirror_con->java_mirror_type();
3769 if (tm != NULL && tm->is_klass() &&
3770 tp != NULL && tp->klass() != NULL) {
3771 if (!tp->klass()->is_loaded()) {
3772 // Don't use intrinsic when class is not loaded.
3773 return false;
3774 } else {
3775 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3776 if (static_res == Compile::SSC_always_true) {
3777 // isInstance() is true - fold the code.
3778 set_result(obj);
3779 return true;
3780 } else if (static_res == Compile::SSC_always_false) {
3781 // Don't use intrinsic, have to throw ClassCastException.
3782 // If the reference is null, the non-intrinsic bytecode will
3783 // be optimized appropriately.
3784 return false;
3785 }
3786 }
3787 }
3788
3789 // Bailout intrinsic and do normal inlining if exception path is frequent.
3790 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3791 return false;
3792 }
3793
3794 // Generate dynamic checks.
3795 // Class.cast() is java implementation of _checkcast bytecode.
3796 // Do checkcast (Parse::do_checkcast()) optimizations here.
3797
3798 mirror = null_check(mirror);
3799 // If mirror is dead, only null-path is taken.
3800 if (stopped()) {
3801 return true;
3802 }
3803
3804 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3805 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3806 RegionNode* region = new RegionNode(PATH_LIMIT);
3807 record_for_igvn(region);
3808
3809 // Now load the mirror's klass metaobject, and null-check it.
3810 // If kls is null, we have a primitive mirror and
3811 // nothing is an instance of a primitive type.
3812 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3813
3814 Node* res = top();
3815 if (!stopped()) {
3816 Node* bad_type_ctrl = top();
3817 // Do checkcast optimizations.
3818 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3819 region->init_req(_bad_type_path, bad_type_ctrl);
3820 }
3821 if (region->in(_prim_path) != top() ||
3822 region->in(_bad_type_path) != top()) {
3823 // Let Interpreter throw ClassCastException.
3824 PreserveJVMState pjvms(this);
3825 set_control(_gvn.transform(region));
3826 uncommon_trap(Deoptimization::Reason_intrinsic,
3827 Deoptimization::Action_maybe_recompile);
3828 }
3829 if (!stopped()) {
3830 set_result(res);
3831 }
3832 return true;
3833 }
3834
3835
3836 //--------------------------inline_native_subtype_check------------------------
3837 // This intrinsic takes the JNI calls out of the heart of
3838 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3839 bool LibraryCallKit::inline_native_subtype_check() {
3840 // Pull both arguments off the stack.
3841 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3842 args[0] = argument(0);
3843 args[1] = argument(1);
3844 Node* klasses[2]; // corresponding Klasses: superk, subk
3845 klasses[0] = klasses[1] = top();
3846
3847 enum {
3848 // A full decision tree on {superc is prim, subc is prim}:
3849 _prim_0_path = 1, // {P,N} => false
3850 // {P,P} & superc!=subc => false
3851 _prim_same_path, // {P,P} & superc==subc => true
3852 _prim_1_path, // {N,P} => false
3853 _ref_subtype_path, // {N,N} & subtype check wins => true
3854 _both_ref_path, // {N,N} & subtype check loses => false
3855 PATH_LIMIT
3856 };
3857
3858 RegionNode* region = new RegionNode(PATH_LIMIT);
3859 Node* phi = new PhiNode(region, TypeInt::BOOL);
3860 record_for_igvn(region);
3861
3862 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3863 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3864 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3865
3866 // First null-check both mirrors and load each mirror's klass metaobject.
3867 int which_arg;
3868 for (which_arg = 0; which_arg <= 1; which_arg++) {
3869 Node* arg = args[which_arg];
3870 arg = null_check(arg);
3871 if (stopped()) break;
3872 args[which_arg] = arg;
3873
3874 Node* p = basic_plus_adr(arg, class_klass_offset);
3875 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3876 klasses[which_arg] = _gvn.transform(kls);
3877 }
3878
3879 // Having loaded both klasses, test each for null.
3880 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3881 for (which_arg = 0; which_arg <= 1; which_arg++) {
3882 Node* kls = klasses[which_arg];
3883 Node* null_ctl = top();
3884 kls = null_check_oop(kls, &null_ctl, never_see_null);
3885 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3886 region->init_req(prim_path, null_ctl);
3887 if (stopped()) break;
3888 klasses[which_arg] = kls;
3889 }
3890
3891 if (!stopped()) {
3892 // now we have two reference types, in klasses[0..1]
3893 Node* subk = klasses[1]; // the argument to isAssignableFrom
3894 Node* superk = klasses[0]; // the receiver
3895 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3896 // now we have a successful reference subtype check
3897 region->set_req(_ref_subtype_path, control());
3898 }
3899
3900 // If both operands are primitive (both klasses null), then
3901 // we must return true when they are identical primitives.
3902 // It is convenient to test this after the first null klass check.
3903 set_control(region->in(_prim_0_path)); // go back to first null check
3904 if (!stopped()) {
3905 // Since superc is primitive, make a guard for the superc==subc case.
3906 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3907 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3908 generate_guard(bol_eq, region, PROB_FAIR);
3909 if (region->req() == PATH_LIMIT+1) {
3910 // A guard was added. If the added guard is taken, superc==subc.
3911 region->swap_edges(PATH_LIMIT, _prim_same_path);
3912 region->del_req(PATH_LIMIT);
3913 }
3914 region->set_req(_prim_0_path, control()); // Not equal after all.
3915 }
3916
3917 // these are the only paths that produce 'true':
3918 phi->set_req(_prim_same_path, intcon(1));
3919 phi->set_req(_ref_subtype_path, intcon(1));
3920
3921 // pull together the cases:
3922 assert(region->req() == PATH_LIMIT, "sane region");
3923 for (uint i = 1; i < region->req(); i++) {
3924 Node* ctl = region->in(i);
3925 if (ctl == NULL || ctl == top()) {
3926 region->set_req(i, top());
3927 phi ->set_req(i, top());
3928 } else if (phi->in(i) == NULL) {
3929 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3930 }
3931 }
3932
3933 set_control(_gvn.transform(region));
3934 set_result(_gvn.transform(phi));
3935 return true;
3936 }
3937
3938 //---------------------generate_array_guard_common------------------------
3939 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3940 bool obj_array, bool not_array) {
3941
3942 if (stopped()) {
3943 return NULL;
3944 }
3945
3946 // If obj_array/non_array==false/false:
3947 // Branch around if the given klass is in fact an array (either obj or prim).
3948 // If obj_array/non_array==false/true:
3949 // Branch around if the given klass is not an array klass of any kind.
3950 // If obj_array/non_array==true/true:
3951 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3952 // If obj_array/non_array==true/false:
3953 // Branch around if the kls is an oop array (Object[] or subtype)
3954 //
3955 // Like generate_guard, adds a new path onto the region.
3956 jint layout_con = 0;
3957 Node* layout_val = get_layout_helper(kls, layout_con);
3958 if (layout_val == NULL) {
3959 bool query = (obj_array
3960 ? Klass::layout_helper_is_objArray(layout_con)
3961 : Klass::layout_helper_is_array(layout_con));
3962 if (query == not_array) {
3963 return NULL; // never a branch
3964 } else { // always a branch
3965 Node* always_branch = control();
3966 if (region != NULL)
3967 region->add_req(always_branch);
3968 set_control(top());
3969 return always_branch;
3970 }
3971 }
3972 // Now test the correct condition.
3973 jint nval = (obj_array
3974 ? (jint)(Klass::_lh_array_tag_type_value
3975 << Klass::_lh_array_tag_shift)
3976 : Klass::_lh_neutral_value);
3977 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3978 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3979 // invert the test if we are looking for a non-array
3980 if (not_array) btest = BoolTest(btest).negate();
3981 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3982 return generate_fair_guard(bol, region);
3983 }
3984
3985
3986 //-----------------------inline_native_newArray--------------------------
3987 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3988 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3989 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3990 Node* mirror;
3991 Node* count_val;
3992 if (uninitialized) {
3993 mirror = argument(1);
3994 count_val = argument(2);
3995 } else {
3996 mirror = argument(0);
3997 count_val = argument(1);
3998 }
3999
4000 mirror = null_check(mirror);
4001 // If mirror or obj is dead, only null-path is taken.
4002 if (stopped()) return true;
4003
4004 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
4005 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4006 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4007 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4008 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4009
4010 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
4011 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
4012 result_reg, _slow_path);
4013 Node* normal_ctl = control();
4014 Node* no_array_ctl = result_reg->in(_slow_path);
4015
4016 // Generate code for the slow case. We make a call to newArray().
4017 set_control(no_array_ctl);
4018 if (!stopped()) {
4019 // Either the input type is void.class, or else the
4020 // array klass has not yet been cached. Either the
4021 // ensuing call will throw an exception, or else it
4022 // will cache the array klass for next time.
4023 PreserveJVMState pjvms(this);
4024 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
4025 Node* slow_result = set_results_for_java_call(slow_call);
4026 // this->control() comes from set_results_for_java_call
4027 result_reg->set_req(_slow_path, control());
4028 result_val->set_req(_slow_path, slow_result);
4029 result_io ->set_req(_slow_path, i_o());
4030 result_mem->set_req(_slow_path, reset_memory());
4031 }
4032
4033 set_control(normal_ctl);
4034 if (!stopped()) {
4035 // Normal case: The array type has been cached in the java.lang.Class.
4036 // The following call works fine even if the array type is polymorphic.
4037 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4038 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
4039 result_reg->init_req(_normal_path, control());
4040 result_val->init_req(_normal_path, obj);
4041 result_io ->init_req(_normal_path, i_o());
4042 result_mem->init_req(_normal_path, reset_memory());
4043
4044 if (uninitialized) {
4045 // Mark the allocation so that zeroing is skipped
4046 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4047 alloc->maybe_set_complete(&_gvn);
4048 }
4049 }
4050
4051 // Return the combined state.
4052 set_i_o( _gvn.transform(result_io) );
4053 set_all_memory( _gvn.transform(result_mem));
4054
4055 C->set_has_split_ifs(true); // Has chance for split-if optimization
4056 set_result(result_reg, result_val);
4057 return true;
4058 }
4059
4060 //----------------------inline_native_getLength--------------------------
4061 // public static native int java.lang.reflect.Array.getLength(Object array);
4062 bool LibraryCallKit::inline_native_getLength() {
4063 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4064
4065 Node* array = null_check(argument(0));
4066 // If array is dead, only null-path is taken.
4067 if (stopped()) return true;
4068
4069 // Deoptimize if it is a non-array.
4070 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4071
4072 if (non_array != NULL) {
4073 PreserveJVMState pjvms(this);
4074 set_control(non_array);
4075 uncommon_trap(Deoptimization::Reason_intrinsic,
4076 Deoptimization::Action_maybe_recompile);
4077 }
4078
4079 // If control is dead, only non-array-path is taken.
4080 if (stopped()) return true;
4081
4082 // The works fine even if the array type is polymorphic.
4083 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4084 Node* result = load_array_length(array);
4085
4086 C->set_has_split_ifs(true); // Has chance for split-if optimization
4087 set_result(result);
4088 return true;
4089 }
4090
4091 //------------------------inline_array_copyOf----------------------------
4092 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4093 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4094 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4095 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4096
4097 // Get the arguments.
4098 Node* original = argument(0);
4099 Node* start = is_copyOfRange? argument(1): intcon(0);
4100 Node* end = is_copyOfRange? argument(2): argument(1);
4101 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4102
4103 Node* newcopy = NULL;
4104
4105 // Set the original stack and the reexecute bit for the interpreter to reexecute
4106 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4107 { PreserveReexecuteState preexecs(this);
4108 jvms()->set_should_reexecute(true);
4109
4110 array_type_mirror = null_check(array_type_mirror);
4111 original = null_check(original);
4112
4113 // Check if a null path was taken unconditionally.
4114 if (stopped()) return true;
4115
4116 Node* orig_length = load_array_length(original);
4117
4118 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4119 klass_node = null_check(klass_node);
4120
4121 RegionNode* bailout = new RegionNode(1);
4122 record_for_igvn(bailout);
4123
4124 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4125 // Bail out if that is so.
4126 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4127 if (not_objArray != NULL) {
4128 // Improve the klass node's type from the new optimistic assumption:
4129 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4130 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4131 Node* cast = new CastPPNode(klass_node, akls);
4132 cast->init_req(0, control());
4133 klass_node = _gvn.transform(cast);
4134 }
4135
4136 // Bail out if either start or end is negative.
4137 generate_negative_guard(start, bailout, &start);
4138 generate_negative_guard(end, bailout, &end);
4139
4140 Node* length = end;
4141 if (_gvn.type(start) != TypeInt::ZERO) {
4142 length = _gvn.transform(new SubINode(end, start));
4143 }
4144
4145 // Bail out if length is negative.
4146 // Without this the new_array would throw
4147 // NegativeArraySizeException but IllegalArgumentException is what
4148 // should be thrown
4149 generate_negative_guard(length, bailout, &length);
4150
4151 if (bailout->req() > 1) {
4152 PreserveJVMState pjvms(this);
4153 set_control(_gvn.transform(bailout));
4154 uncommon_trap(Deoptimization::Reason_intrinsic,
4155 Deoptimization::Action_maybe_recompile);
4156 }
4157
4158 if (!stopped()) {
4159 // How many elements will we copy from the original?
4160 // The answer is MinI(orig_length - start, length).
4161 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4162 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4163
4164 // Generate a direct call to the right arraycopy function(s).
4165 // We know the copy is disjoint but we might not know if the
4166 // oop stores need checking.
4167 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4168 // This will fail a store-check if x contains any non-nulls.
4169
4170 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4171 // loads/stores but it is legal only if we're sure the
4172 // Arrays.copyOf would succeed. So we need all input arguments
4173 // to the copyOf to be validated, including that the copy to the
4174 // new array won't trigger an ArrayStoreException. That subtype
4175 // check can be optimized if we know something on the type of
4176 // the input array from type speculation.
4177 if (_gvn.type(klass_node)->singleton()) {
4178 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4179 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4180
4181 int test = C->static_subtype_check(superk, subk);
4182 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4183 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4184 if (t_original->speculative_type() != NULL) {
4185 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4186 }
4187 }
4188 }
4189
4190 bool validated = false;
4191 // Reason_class_check rather than Reason_intrinsic because we
4192 // want to intrinsify even if this traps.
4193 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4194 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4195 klass_node);
4196
4197 if (not_subtype_ctrl != top()) {
4198 PreserveJVMState pjvms(this);
4199 set_control(not_subtype_ctrl);
4200 uncommon_trap(Deoptimization::Reason_class_check,
4201 Deoptimization::Action_make_not_entrant);
4202 assert(stopped(), "Should be stopped");
4203 }
4204 validated = true;
4205 }
4206
4207 if (!stopped()) {
4208 newcopy = new_array(klass_node, length, 0); // no arguments to push
4209
4210 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4211 load_object_klass(original), klass_node);
4212 if (!is_copyOfRange) {
4213 ac->set_copyof(validated);
4214 } else {
4215 ac->set_copyofrange(validated);
4216 }
4217 Node* n = _gvn.transform(ac);
4218 if (n == ac) {
4219 ac->connect_outputs(this);
4220 } else {
4221 assert(validated, "shouldn't transform if all arguments not validated");
4222 set_all_memory(n);
4223 }
4224 }
4225 }
4226 } // original reexecute is set back here
4227
4228 C->set_has_split_ifs(true); // Has chance for split-if optimization
4229 if (!stopped()) {
4230 set_result(newcopy);
4231 }
4232 return true;
4233 }
4234
4235
4236 //----------------------generate_virtual_guard---------------------------
4237 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4238 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4239 RegionNode* slow_region) {
4240 ciMethod* method = callee();
4241 int vtable_index = method->vtable_index();
4242 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4243 "bad index %d", vtable_index);
4244 // Get the Method* out of the appropriate vtable entry.
4245 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4246 vtable_index*vtableEntry::size_in_bytes() +
4247 vtableEntry::method_offset_in_bytes();
4248 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4249 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4250
4251 // Compare the target method with the expected method (e.g., Object.hashCode).
4252 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4253
4254 Node* native_call = makecon(native_call_addr);
4255 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4256 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4257
4258 return generate_slow_guard(test_native, slow_region);
4259 }
4260
4261 //-----------------------generate_method_call----------------------------
4262 // Use generate_method_call to make a slow-call to the real
4263 // method if the fast path fails. An alternative would be to
4264 // use a stub like OptoRuntime::slow_arraycopy_Java.
4265 // This only works for expanding the current library call,
4266 // not another intrinsic. (E.g., don't use this for making an
4267 // arraycopy call inside of the copyOf intrinsic.)
4268 CallJavaNode*
4269 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4270 // When compiling the intrinsic method itself, do not use this technique.
4271 guarantee(callee() != C->method(), "cannot make slow-call to self");
4272
4273 ciMethod* method = callee();
4274 // ensure the JVMS we have will be correct for this call
4275 guarantee(method_id == method->intrinsic_id(), "must match");
4276
4277 const TypeFunc* tf = TypeFunc::make(method);
4278 CallJavaNode* slow_call;
4279 if (is_static) {
4280 assert(!is_virtual, "");
4281 slow_call = new CallStaticJavaNode(C, tf,
4282 SharedRuntime::get_resolve_static_call_stub(),
4283 method, bci());
4284 } else if (is_virtual) {
4285 null_check_receiver();
4286 int vtable_index = Method::invalid_vtable_index;
4287 if (UseInlineCaches) {
4288 // Suppress the vtable call
4289 } else {
4290 // hashCode and clone are not a miranda methods,
4291 // so the vtable index is fixed.
4292 // No need to use the linkResolver to get it.
4293 vtable_index = method->vtable_index();
4294 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4295 "bad index %d", vtable_index);
4296 }
4297 slow_call = new CallDynamicJavaNode(tf,
4298 SharedRuntime::get_resolve_virtual_call_stub(),
4299 method, vtable_index, bci());
4300 } else { // neither virtual nor static: opt_virtual
4301 null_check_receiver();
4302 slow_call = new CallStaticJavaNode(C, tf,
4303 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4304 method, bci());
4305 slow_call->set_optimized_virtual(true);
4306 }
4307 set_arguments_for_java_call(slow_call);
4308 set_edges_for_java_call(slow_call);
4309 return slow_call;
4310 }
4311
4312
4313 /**
4314 * Build special case code for calls to hashCode on an object. This call may
4315 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4316 * slightly different code.
4317 */
4318 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4319 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4320 assert(!(is_virtual && is_static), "either virtual, special, or static");
4321
4322 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4323
4324 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4325 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4326 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4327 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4328 Node* obj = NULL;
4329 if (!is_static) {
4330 // Check for hashing null object
4331 obj = null_check_receiver();
4332 if (stopped()) return true; // unconditionally null
4333 result_reg->init_req(_null_path, top());
4334 result_val->init_req(_null_path, top());
4335 } else {
4336 // Do a null check, and return zero if null.
4337 // System.identityHashCode(null) == 0
4338 obj = argument(0);
4339 Node* null_ctl = top();
4340 obj = null_check_oop(obj, &null_ctl);
4341 result_reg->init_req(_null_path, null_ctl);
4342 result_val->init_req(_null_path, _gvn.intcon(0));
4343 }
4344
4345 // Unconditionally null? Then return right away.
4346 if (stopped()) {
4347 set_control( result_reg->in(_null_path));
4348 if (!stopped())
4349 set_result(result_val->in(_null_path));
4350 return true;
4351 }
4352
4353 // We only go to the fast case code if we pass a number of guards. The
4354 // paths which do not pass are accumulated in the slow_region.
4355 RegionNode* slow_region = new RegionNode(1);
4356 record_for_igvn(slow_region);
4357
4358 // If this is a virtual call, we generate a funny guard. We pull out
4359 // the vtable entry corresponding to hashCode() from the target object.
4360 // If the target method which we are calling happens to be the native
4361 // Object hashCode() method, we pass the guard. We do not need this
4362 // guard for non-virtual calls -- the caller is known to be the native
4363 // Object hashCode().
4364 if (is_virtual) {
4365 // After null check, get the object's klass.
4366 Node* obj_klass = load_object_klass(obj);
4367 generate_virtual_guard(obj_klass, slow_region);
4368 }
4369
4370 // Get the header out of the object, use LoadMarkNode when available
4371 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4372 // The control of the load must be NULL. Otherwise, the load can move before
4373 // the null check after castPP removal.
4374 Node* no_ctrl = NULL;
4375 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4376
4377 // Test the header to see if it is unlocked.
4378 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4379 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4380 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4381 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4382 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4383
4384 generate_slow_guard(test_unlocked, slow_region);
4385
4386 // Get the hash value and check to see that it has been properly assigned.
4387 // We depend on hash_mask being at most 32 bits and avoid the use of
4388 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4389 // vm: see markOop.hpp.
4390 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4391 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4392 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4393 // This hack lets the hash bits live anywhere in the mark object now, as long
4394 // as the shift drops the relevant bits into the low 32 bits. Note that
4395 // Java spec says that HashCode is an int so there's no point in capturing
4396 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4397 hshifted_header = ConvX2I(hshifted_header);
4398 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4399
4400 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4401 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4402 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4403
4404 generate_slow_guard(test_assigned, slow_region);
4405
4406 Node* init_mem = reset_memory();
4407 // fill in the rest of the null path:
4408 result_io ->init_req(_null_path, i_o());
4409 result_mem->init_req(_null_path, init_mem);
4410
4411 result_val->init_req(_fast_path, hash_val);
4412 result_reg->init_req(_fast_path, control());
4413 result_io ->init_req(_fast_path, i_o());
4414 result_mem->init_req(_fast_path, init_mem);
4415
4416 // Generate code for the slow case. We make a call to hashCode().
4417 set_control(_gvn.transform(slow_region));
4418 if (!stopped()) {
4419 // No need for PreserveJVMState, because we're using up the present state.
4420 set_all_memory(init_mem);
4421 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4422 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4423 Node* slow_result = set_results_for_java_call(slow_call);
4424 // this->control() comes from set_results_for_java_call
4425 result_reg->init_req(_slow_path, control());
4426 result_val->init_req(_slow_path, slow_result);
4427 result_io ->set_req(_slow_path, i_o());
4428 result_mem ->set_req(_slow_path, reset_memory());
4429 }
4430
4431 // Return the combined state.
4432 set_i_o( _gvn.transform(result_io) );
4433 set_all_memory( _gvn.transform(result_mem));
4434
4435 set_result(result_reg, result_val);
4436 return true;
4437 }
4438
4439 //---------------------------inline_native_getClass----------------------------
4440 // public final native Class<?> java.lang.Object.getClass();
4441 //
4442 // Build special case code for calls to getClass on an object.
4443 bool LibraryCallKit::inline_native_getClass() {
4444 Node* obj = null_check_receiver();
4445 if (stopped()) return true;
4446 set_result(load_mirror_from_klass(load_object_klass(obj)));
4447 return true;
4448 }
4449
4450 //-----------------inline_native_Reflection_getCallerClass---------------------
4451 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4452 //
4453 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4454 //
4455 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4456 // in that it must skip particular security frames and checks for
4457 // caller sensitive methods.
4458 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4459 #ifndef PRODUCT
4460 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4461 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4462 }
4463 #endif
4464
4465 if (!jvms()->has_method()) {
4466 #ifndef PRODUCT
4467 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4468 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4469 }
4470 #endif
4471 return false;
4472 }
4473
4474 // Walk back up the JVM state to find the caller at the required
4475 // depth.
4476 JVMState* caller_jvms = jvms();
4477
4478 // Cf. JVM_GetCallerClass
4479 // NOTE: Start the loop at depth 1 because the current JVM state does
4480 // not include the Reflection.getCallerClass() frame.
4481 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4482 ciMethod* m = caller_jvms->method();
4483 switch (n) {
4484 case 0:
4485 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4486 break;
4487 case 1:
4488 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4489 if (!m->caller_sensitive()) {
4490 #ifndef PRODUCT
4491 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4492 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4493 }
4494 #endif
4495 return false; // bail-out; let JVM_GetCallerClass do the work
4496 }
4497 break;
4498 default:
4499 if (!m->is_ignored_by_security_stack_walk()) {
4500 // We have reached the desired frame; return the holder class.
4501 // Acquire method holder as java.lang.Class and push as constant.
4502 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4503 ciInstance* caller_mirror = caller_klass->java_mirror();
4504 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4505
4506 #ifndef PRODUCT
4507 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4508 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());
4509 tty->print_cr(" JVM state at this point:");
4510 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4511 ciMethod* m = jvms()->of_depth(i)->method();
4512 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4513 }
4514 }
4515 #endif
4516 return true;
4517 }
4518 break;
4519 }
4520 }
4521
4522 #ifndef PRODUCT
4523 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4524 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4525 tty->print_cr(" JVM state at this point:");
4526 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4527 ciMethod* m = jvms()->of_depth(i)->method();
4528 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4529 }
4530 }
4531 #endif
4532
4533 return false; // bail-out; let JVM_GetCallerClass do the work
4534 }
4535
4536 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4537 Node* arg = argument(0);
4538 Node* result = NULL;
4539
4540 switch (id) {
4541 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4542 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4543 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4544 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4545
4546 case vmIntrinsics::_doubleToLongBits: {
4547 // two paths (plus control) merge in a wood
4548 RegionNode *r = new RegionNode(3);
4549 Node *phi = new PhiNode(r, TypeLong::LONG);
4550
4551 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4552 // Build the boolean node
4553 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4554
4555 // Branch either way.
4556 // NaN case is less traveled, which makes all the difference.
4557 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4558 Node *opt_isnan = _gvn.transform(ifisnan);
4559 assert( opt_isnan->is_If(), "Expect an IfNode");
4560 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4561 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4562
4563 set_control(iftrue);
4564
4565 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4566 Node *slow_result = longcon(nan_bits); // return NaN
4567 phi->init_req(1, _gvn.transform( slow_result ));
4568 r->init_req(1, iftrue);
4569
4570 // Else fall through
4571 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4572 set_control(iffalse);
4573
4574 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4575 r->init_req(2, iffalse);
4576
4577 // Post merge
4578 set_control(_gvn.transform(r));
4579 record_for_igvn(r);
4580
4581 C->set_has_split_ifs(true); // Has chance for split-if optimization
4582 result = phi;
4583 assert(result->bottom_type()->isa_long(), "must be");
4584 break;
4585 }
4586
4587 case vmIntrinsics::_floatToIntBits: {
4588 // two paths (plus control) merge in a wood
4589 RegionNode *r = new RegionNode(3);
4590 Node *phi = new PhiNode(r, TypeInt::INT);
4591
4592 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4593 // Build the boolean node
4594 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4595
4596 // Branch either way.
4597 // NaN case is less traveled, which makes all the difference.
4598 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4599 Node *opt_isnan = _gvn.transform(ifisnan);
4600 assert( opt_isnan->is_If(), "Expect an IfNode");
4601 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4602 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4603
4604 set_control(iftrue);
4605
4606 static const jint nan_bits = 0x7fc00000;
4607 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4608 phi->init_req(1, _gvn.transform( slow_result ));
4609 r->init_req(1, iftrue);
4610
4611 // Else fall through
4612 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4613 set_control(iffalse);
4614
4615 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4616 r->init_req(2, iffalse);
4617
4618 // Post merge
4619 set_control(_gvn.transform(r));
4620 record_for_igvn(r);
4621
4622 C->set_has_split_ifs(true); // Has chance for split-if optimization
4623 result = phi;
4624 assert(result->bottom_type()->isa_int(), "must be");
4625 break;
4626 }
4627
4628 default:
4629 fatal_unexpected_iid(id);
4630 break;
4631 }
4632 set_result(_gvn.transform(result));
4633 return true;
4634 }
4635
4636 //----------------------inline_unsafe_copyMemory-------------------------
4637 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4638 bool LibraryCallKit::inline_unsafe_copyMemory() {
4639 if (callee()->is_static()) return false; // caller must have the capability!
4640 null_check_receiver(); // null-check receiver
4641 if (stopped()) return true;
4642
4643 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4644
4645 Node* src_ptr = argument(1); // type: oop
4646 Node* src_off = ConvL2X(argument(2)); // type: long
4647 Node* dst_ptr = argument(4); // type: oop
4648 Node* dst_off = ConvL2X(argument(5)); // type: long
4649 Node* size = ConvL2X(argument(7)); // type: long
4650
4651 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4652 "fieldOffset must be byte-scaled");
4653
4654 Node* src = make_unsafe_address(src_ptr, src_off);
4655 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4656
4657 // Conservatively insert a memory barrier on all memory slices.
4658 // Do not let writes of the copy source or destination float below the copy.
4659 insert_mem_bar(Op_MemBarCPUOrder);
4660
4661 // Call it. Note that the length argument is not scaled.
4662 make_runtime_call(RC_LEAF|RC_NO_FP,
4663 OptoRuntime::fast_arraycopy_Type(),
4664 StubRoutines::unsafe_arraycopy(),
4665 "unsafe_arraycopy",
4666 TypeRawPtr::BOTTOM,
4667 src, dst, size XTOP);
4668
4669 // Do not let reads of the copy destination float above the copy.
4670 insert_mem_bar(Op_MemBarCPUOrder);
4671
4672 return true;
4673 }
4674
4675 //------------------------clone_coping-----------------------------------
4676 // Helper function for inline_native_clone.
4677 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4678 assert(obj_size != NULL, "");
4679 Node* raw_obj = alloc_obj->in(1);
4680 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4681
4682 AllocateNode* alloc = NULL;
4683 if (ReduceBulkZeroing) {
4684 // We will be completely responsible for initializing this object -
4685 // mark Initialize node as complete.
4686 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4687 // The object was just allocated - there should be no any stores!
4688 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4689 // Mark as complete_with_arraycopy so that on AllocateNode
4690 // expansion, we know this AllocateNode is initialized by an array
4691 // copy and a StoreStore barrier exists after the array copy.
4692 alloc->initialization()->set_complete_with_arraycopy();
4693 }
4694
4695 // Copy the fastest available way.
4696 // TODO: generate fields copies for small objects instead.
4697 Node* src = obj;
4698 Node* dest = alloc_obj;
4699 Node* size = _gvn.transform(obj_size);
4700
4701 // Exclude the header but include array length to copy by 8 bytes words.
4702 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4703 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4704 instanceOopDesc::base_offset_in_bytes();
4705 // base_off:
4706 // 8 - 32-bit VM
4707 // 12 - 64-bit VM, compressed klass
4708 // 16 - 64-bit VM, normal klass
4709 if (base_off % BytesPerLong != 0) {
4710 assert(UseCompressedClassPointers, "");
4711 if (is_array) {
4712 // Exclude length to copy by 8 bytes words.
4713 base_off += sizeof(int);
4714 } else {
4715 // Include klass to copy by 8 bytes words.
4716 base_off = instanceOopDesc::klass_offset_in_bytes();
4717 }
4718 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4719 }
4720 src = basic_plus_adr(src, base_off);
4721 dest = basic_plus_adr(dest, base_off);
4722
4723 // Compute the length also, if needed:
4724 Node* countx = size;
4725 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4726 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4727
4728 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4729
4730 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4731 ac->set_clonebasic();
4732 Node* n = _gvn.transform(ac);
4733 if (n == ac) {
4734 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4735 } else {
4736 set_all_memory(n);
4737 }
4738
4739 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4740 if (card_mark) {
4741 assert(!is_array, "");
4742 // Put in store barrier for any and all oops we are sticking
4743 // into this object. (We could avoid this if we could prove
4744 // that the object type contains no oop fields at all.)
4745 Node* no_particular_value = NULL;
4746 Node* no_particular_field = NULL;
4747 int raw_adr_idx = Compile::AliasIdxRaw;
4748 post_barrier(control(),
4749 memory(raw_adr_type),
4750 alloc_obj,
4751 no_particular_field,
4752 raw_adr_idx,
4753 no_particular_value,
4754 T_OBJECT,
4755 false);
4756 }
4757
4758 // Do not let reads from the cloned object float above the arraycopy.
4759 if (alloc != NULL) {
4760 // Do not let stores that initialize this object be reordered with
4761 // a subsequent store that would make this object accessible by
4762 // other threads.
4763 // Record what AllocateNode this StoreStore protects so that
4764 // escape analysis can go from the MemBarStoreStoreNode to the
4765 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4766 // based on the escape status of the AllocateNode.
4767 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4768 } else {
4769 insert_mem_bar(Op_MemBarCPUOrder);
4770 }
4771 }
4772
4773 //------------------------inline_native_clone----------------------------
4774 // protected native Object java.lang.Object.clone();
4775 //
4776 // Here are the simple edge cases:
4777 // null receiver => normal trap
4778 // virtual and clone was overridden => slow path to out-of-line clone
4779 // not cloneable or finalizer => slow path to out-of-line Object.clone
4780 //
4781 // The general case has two steps, allocation and copying.
4782 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4783 //
4784 // Copying also has two cases, oop arrays and everything else.
4785 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4786 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4787 //
4788 // These steps fold up nicely if and when the cloned object's klass
4789 // can be sharply typed as an object array, a type array, or an instance.
4790 //
4791 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4792 PhiNode* result_val;
4793
4794 // Set the reexecute bit for the interpreter to reexecute
4795 // the bytecode that invokes Object.clone if deoptimization happens.
4796 { PreserveReexecuteState preexecs(this);
4797 jvms()->set_should_reexecute(true);
4798
4799 Node* obj = null_check_receiver();
4800 if (stopped()) return true;
4801
4802 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4803
4804 // If we are going to clone an instance, we need its exact type to
4805 // know the number and types of fields to convert the clone to
4806 // loads/stores. Maybe a speculative type can help us.
4807 if (!obj_type->klass_is_exact() &&
4808 obj_type->speculative_type() != NULL &&
4809 obj_type->speculative_type()->is_instance_klass()) {
4810 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4811 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4812 !spec_ik->has_injected_fields()) {
4813 ciKlass* k = obj_type->klass();
4814 if (!k->is_instance_klass() ||
4815 k->as_instance_klass()->is_interface() ||
4816 k->as_instance_klass()->has_subklass()) {
4817 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4818 }
4819 }
4820 }
4821
4822 Node* obj_klass = load_object_klass(obj);
4823 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4824 const TypeOopPtr* toop = ((tklass != NULL)
4825 ? tklass->as_instance_type()
4826 : TypeInstPtr::NOTNULL);
4827
4828 // Conservatively insert a memory barrier on all memory slices.
4829 // Do not let writes into the original float below the clone.
4830 insert_mem_bar(Op_MemBarCPUOrder);
4831
4832 // paths into result_reg:
4833 enum {
4834 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4835 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4836 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4837 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4838 PATH_LIMIT
4839 };
4840 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4841 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4842 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4843 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4844 record_for_igvn(result_reg);
4845
4846 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4847 int raw_adr_idx = Compile::AliasIdxRaw;
4848
4849 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4850 if (array_ctl != NULL) {
4851 // It's an array.
4852 PreserveJVMState pjvms(this);
4853 set_control(array_ctl);
4854 Node* obj_length = load_array_length(obj);
4855 Node* obj_size = NULL;
4856 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4857
4858 if (!use_ReduceInitialCardMarks()) {
4859 // If it is an oop array, it requires very special treatment,
4860 // because card marking is required on each card of the array.
4861 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4862 if (is_obja != NULL) {
4863 PreserveJVMState pjvms2(this);
4864 set_control(is_obja);
4865 // Generate a direct call to the right arraycopy function(s).
4866 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4867 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4868 ac->set_cloneoop();
4869 Node* n = _gvn.transform(ac);
4870 assert(n == ac, "cannot disappear");
4871 ac->connect_outputs(this);
4872
4873 result_reg->init_req(_objArray_path, control());
4874 result_val->init_req(_objArray_path, alloc_obj);
4875 result_i_o ->set_req(_objArray_path, i_o());
4876 result_mem ->set_req(_objArray_path, reset_memory());
4877 }
4878 }
4879 // Otherwise, there are no card marks to worry about.
4880 // (We can dispense with card marks if we know the allocation
4881 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4882 // causes the non-eden paths to take compensating steps to
4883 // simulate a fresh allocation, so that no further
4884 // card marks are required in compiled code to initialize
4885 // the object.)
4886
4887 if (!stopped()) {
4888 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4889
4890 // Present the results of the copy.
4891 result_reg->init_req(_array_path, control());
4892 result_val->init_req(_array_path, alloc_obj);
4893 result_i_o ->set_req(_array_path, i_o());
4894 result_mem ->set_req(_array_path, reset_memory());
4895 }
4896 }
4897
4898 // We only go to the instance fast case code if we pass a number of guards.
4899 // The paths which do not pass are accumulated in the slow_region.
4900 RegionNode* slow_region = new RegionNode(1);
4901 record_for_igvn(slow_region);
4902 if (!stopped()) {
4903 // It's an instance (we did array above). Make the slow-path tests.
4904 // If this is a virtual call, we generate a funny guard. We grab
4905 // the vtable entry corresponding to clone() from the target object.
4906 // If the target method which we are calling happens to be the
4907 // Object clone() method, we pass the guard. We do not need this
4908 // guard for non-virtual calls; the caller is known to be the native
4909 // Object clone().
4910 if (is_virtual) {
4911 generate_virtual_guard(obj_klass, slow_region);
4912 }
4913
4914 // The object must be easily cloneable and must not have a finalizer.
4915 // Both of these conditions may be checked in a single test.
4916 // We could optimize the test further, but we don't care.
4917 generate_access_flags_guard(obj_klass,
4918 // Test both conditions:
4919 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4920 // Must be cloneable but not finalizer:
4921 JVM_ACC_IS_CLONEABLE_FAST,
4922 slow_region);
4923 }
4924
4925 if (!stopped()) {
4926 // It's an instance, and it passed the slow-path tests.
4927 PreserveJVMState pjvms(this);
4928 Node* obj_size = NULL;
4929 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4930 // is reexecuted if deoptimization occurs and there could be problems when merging
4931 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4932 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4933
4934 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4935
4936 // Present the results of the slow call.
4937 result_reg->init_req(_instance_path, control());
4938 result_val->init_req(_instance_path, alloc_obj);
4939 result_i_o ->set_req(_instance_path, i_o());
4940 result_mem ->set_req(_instance_path, reset_memory());
4941 }
4942
4943 // Generate code for the slow case. We make a call to clone().
4944 set_control(_gvn.transform(slow_region));
4945 if (!stopped()) {
4946 PreserveJVMState pjvms(this);
4947 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4948 Node* slow_result = set_results_for_java_call(slow_call);
4949 // this->control() comes from set_results_for_java_call
4950 result_reg->init_req(_slow_path, control());
4951 result_val->init_req(_slow_path, slow_result);
4952 result_i_o ->set_req(_slow_path, i_o());
4953 result_mem ->set_req(_slow_path, reset_memory());
4954 }
4955
4956 // Return the combined state.
4957 set_control( _gvn.transform(result_reg));
4958 set_i_o( _gvn.transform(result_i_o));
4959 set_all_memory( _gvn.transform(result_mem));
4960 } // original reexecute is set back here
4961
4962 set_result(_gvn.transform(result_val));
4963 return true;
4964 }
4965
4966 // If we have a tighly coupled allocation, the arraycopy may take care
4967 // of the array initialization. If one of the guards we insert between
4968 // the allocation and the arraycopy causes a deoptimization, an
4969 // unitialized array will escape the compiled method. To prevent that
4970 // we set the JVM state for uncommon traps between the allocation and
4971 // the arraycopy to the state before the allocation so, in case of
4972 // deoptimization, we'll reexecute the allocation and the
4973 // initialization.
4974 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4975 if (alloc != NULL) {
4976 ciMethod* trap_method = alloc->jvms()->method();
4977 int trap_bci = alloc->jvms()->bci();
4978
4979 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4980 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4981 // Make sure there's no store between the allocation and the
4982 // arraycopy otherwise visible side effects could be rexecuted
4983 // in case of deoptimization and cause incorrect execution.
4984 bool no_interfering_store = true;
4985 Node* mem = alloc->in(TypeFunc::Memory);
4986 if (mem->is_MergeMem()) {
4987 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4988 Node* n = mms.memory();
4989 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4990 assert(n->is_Store(), "what else?");
4991 no_interfering_store = false;
4992 break;
4993 }
4994 }
4995 } else {
4996 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4997 Node* n = mms.memory();
4998 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4999 assert(n->is_Store(), "what else?");
5000 no_interfering_store = false;
5001 break;
5002 }
5003 }
5004 }
5005
5006 if (no_interfering_store) {
5007 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
5008 uint size = alloc->req();
5009 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
5010 old_jvms->set_map(sfpt);
5011 for (uint i = 0; i < size; i++) {
5012 sfpt->init_req(i, alloc->in(i));
5013 }
5014 // re-push array length for deoptimization
5015 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
5016 old_jvms->set_sp(old_jvms->sp()+1);
5017 old_jvms->set_monoff(old_jvms->monoff()+1);
5018 old_jvms->set_scloff(old_jvms->scloff()+1);
5019 old_jvms->set_endoff(old_jvms->endoff()+1);
5020 old_jvms->set_should_reexecute(true);
5021
5022 sfpt->set_i_o(map()->i_o());
5023 sfpt->set_memory(map()->memory());
5024 sfpt->set_control(map()->control());
5025
5026 JVMState* saved_jvms = jvms();
5027 saved_reexecute_sp = _reexecute_sp;
5028
5029 set_jvms(sfpt->jvms());
5030 _reexecute_sp = jvms()->sp();
5031
5032 return saved_jvms;
5033 }
5034 }
5035 }
5036 return NULL;
5037 }
5038
5039 // In case of a deoptimization, we restart execution at the
5040 // allocation, allocating a new array. We would leave an uninitialized
5041 // array in the heap that GCs wouldn't expect. Move the allocation
5042 // after the traps so we don't allocate the array if we
5043 // deoptimize. This is possible because tightly_coupled_allocation()
5044 // guarantees there's no observer of the allocated array at this point
5045 // and the control flow is simple enough.
5046 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
5047 int saved_reexecute_sp, uint new_idx) {
5048 if (saved_jvms != NULL && !stopped()) {
5049 assert(alloc != NULL, "only with a tightly coupled allocation");
5050 // restore JVM state to the state at the arraycopy
5051 saved_jvms->map()->set_control(map()->control());
5052 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5053 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5054 // If we've improved the types of some nodes (null check) while
5055 // emitting the guards, propagate them to the current state
5056 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
5057 set_jvms(saved_jvms);
5058 _reexecute_sp = saved_reexecute_sp;
5059
5060 // Remove the allocation from above the guards
5061 CallProjections callprojs;
5062 alloc->extract_projections(&callprojs, true);
5063 InitializeNode* init = alloc->initialization();
5064 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5065 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5066 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5067 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5068
5069 // move the allocation here (after the guards)
5070 _gvn.hash_delete(alloc);
5071 alloc->set_req(TypeFunc::Control, control());
5072 alloc->set_req(TypeFunc::I_O, i_o());
5073 Node *mem = reset_memory();
5074 set_all_memory(mem);
5075 alloc->set_req(TypeFunc::Memory, mem);
5076 set_control(init->proj_out_or_null(TypeFunc::Control));
5077 set_i_o(callprojs.fallthrough_ioproj);
5078
5079 // Update memory as done in GraphKit::set_output_for_allocation()
5080 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5081 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5082 if (ary_type->isa_aryptr() && length_type != NULL) {
5083 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5084 }
5085 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5086 int elemidx = C->get_alias_index(telemref);
5087 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
5088 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
5089
5090 Node* allocx = _gvn.transform(alloc);
5091 assert(allocx == alloc, "where has the allocation gone?");
5092 assert(dest->is_CheckCastPP(), "not an allocation result?");
5093
5094 _gvn.hash_delete(dest);
5095 dest->set_req(0, control());
5096 Node* destx = _gvn.transform(dest);
5097 assert(destx == dest, "where has the allocation result gone?");
5098 }
5099 }
5100
5101
5102 //------------------------------inline_arraycopy-----------------------
5103 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5104 // Object dest, int destPos,
5105 // int length);
5106 bool LibraryCallKit::inline_arraycopy() {
5107 // Get the arguments.
5108 Node* src = argument(0); // type: oop
5109 Node* src_offset = argument(1); // type: int
5110 Node* dest = argument(2); // type: oop
5111 Node* dest_offset = argument(3); // type: int
5112 Node* length = argument(4); // type: int
5113
5114 uint new_idx = C->unique();
5115
5116 // Check for allocation before we add nodes that would confuse
5117 // tightly_coupled_allocation()
5118 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5119
5120 int saved_reexecute_sp = -1;
5121 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5122 // See arraycopy_restore_alloc_state() comment
5123 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5124 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5125 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5126 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5127
5128 // The following tests must be performed
5129 // (1) src and dest are arrays.
5130 // (2) src and dest arrays must have elements of the same BasicType
5131 // (3) src and dest must not be null.
5132 // (4) src_offset must not be negative.
5133 // (5) dest_offset must not be negative.
5134 // (6) length must not be negative.
5135 // (7) src_offset + length must not exceed length of src.
5136 // (8) dest_offset + length must not exceed length of dest.
5137 // (9) each element of an oop array must be assignable
5138
5139 // (3) src and dest must not be null.
5140 // always do this here because we need the JVM state for uncommon traps
5141 Node* null_ctl = top();
5142 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5143 assert(null_ctl->is_top(), "no null control here");
5144 dest = null_check(dest, T_ARRAY);
5145
5146 if (!can_emit_guards) {
5147 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5148 // guards but the arraycopy node could still take advantage of a
5149 // tightly allocated allocation. tightly_coupled_allocation() is
5150 // called again to make sure it takes the null check above into
5151 // account: the null check is mandatory and if it caused an
5152 // uncommon trap to be emitted then the allocation can't be
5153 // considered tightly coupled in this context.
5154 alloc = tightly_coupled_allocation(dest, NULL);
5155 }
5156
5157 bool validated = false;
5158
5159 const Type* src_type = _gvn.type(src);
5160 const Type* dest_type = _gvn.type(dest);
5161 const TypeAryPtr* top_src = src_type->isa_aryptr();
5162 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5163
5164 // Do we have the type of src?
5165 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5166 // Do we have the type of dest?
5167 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5168 // Is the type for src from speculation?
5169 bool src_spec = false;
5170 // Is the type for dest from speculation?
5171 bool dest_spec = false;
5172
5173 if ((!has_src || !has_dest) && can_emit_guards) {
5174 // We don't have sufficient type information, let's see if
5175 // speculative types can help. We need to have types for both src
5176 // and dest so that it pays off.
5177
5178 // Do we already have or could we have type information for src
5179 bool could_have_src = has_src;
5180 // Do we already have or could we have type information for dest
5181 bool could_have_dest = has_dest;
5182
5183 ciKlass* src_k = NULL;
5184 if (!has_src) {
5185 src_k = src_type->speculative_type_not_null();
5186 if (src_k != NULL && src_k->is_array_klass()) {
5187 could_have_src = true;
5188 }
5189 }
5190
5191 ciKlass* dest_k = NULL;
5192 if (!has_dest) {
5193 dest_k = dest_type->speculative_type_not_null();
5194 if (dest_k != NULL && dest_k->is_array_klass()) {
5195 could_have_dest = true;
5196 }
5197 }
5198
5199 if (could_have_src && could_have_dest) {
5200 // This is going to pay off so emit the required guards
5201 if (!has_src) {
5202 src = maybe_cast_profiled_obj(src, src_k, true);
5203 src_type = _gvn.type(src);
5204 top_src = src_type->isa_aryptr();
5205 has_src = (top_src != NULL && top_src->klass() != NULL);
5206 src_spec = true;
5207 }
5208 if (!has_dest) {
5209 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5210 dest_type = _gvn.type(dest);
5211 top_dest = dest_type->isa_aryptr();
5212 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5213 dest_spec = true;
5214 }
5215 }
5216 }
5217
5218 if (has_src && has_dest && can_emit_guards) {
5219 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5220 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5221 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5222 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5223
5224 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5225 // If both arrays are object arrays then having the exact types
5226 // for both will remove the need for a subtype check at runtime
5227 // before the call and may make it possible to pick a faster copy
5228 // routine (without a subtype check on every element)
5229 // Do we have the exact type of src?
5230 bool could_have_src = src_spec;
5231 // Do we have the exact type of dest?
5232 bool could_have_dest = dest_spec;
5233 ciKlass* src_k = top_src->klass();
5234 ciKlass* dest_k = top_dest->klass();
5235 if (!src_spec) {
5236 src_k = src_type->speculative_type_not_null();
5237 if (src_k != NULL && src_k->is_array_klass()) {
5238 could_have_src = true;
5239 }
5240 }
5241 if (!dest_spec) {
5242 dest_k = dest_type->speculative_type_not_null();
5243 if (dest_k != NULL && dest_k->is_array_klass()) {
5244 could_have_dest = true;
5245 }
5246 }
5247 if (could_have_src && could_have_dest) {
5248 // If we can have both exact types, emit the missing guards
5249 if (could_have_src && !src_spec) {
5250 src = maybe_cast_profiled_obj(src, src_k, true);
5251 }
5252 if (could_have_dest && !dest_spec) {
5253 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5254 }
5255 }
5256 }
5257 }
5258
5259 ciMethod* trap_method = method();
5260 int trap_bci = bci();
5261 if (saved_jvms != NULL) {
5262 trap_method = alloc->jvms()->method();
5263 trap_bci = alloc->jvms()->bci();
5264 }
5265
5266 bool negative_length_guard_generated = false;
5267
5268 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5269 can_emit_guards &&
5270 !src->is_top() && !dest->is_top()) {
5271 // validate arguments: enables transformation the ArrayCopyNode
5272 validated = true;
5273
5274 RegionNode* slow_region = new RegionNode(1);
5275 record_for_igvn(slow_region);
5276
5277 // (1) src and dest are arrays.
5278 generate_non_array_guard(load_object_klass(src), slow_region);
5279 generate_non_array_guard(load_object_klass(dest), slow_region);
5280
5281 // (2) src and dest arrays must have elements of the same BasicType
5282 // done at macro expansion or at Ideal transformation time
5283
5284 // (4) src_offset must not be negative.
5285 generate_negative_guard(src_offset, slow_region);
5286
5287 // (5) dest_offset must not be negative.
5288 generate_negative_guard(dest_offset, slow_region);
5289
5290 // (7) src_offset + length must not exceed length of src.
5291 generate_limit_guard(src_offset, length,
5292 load_array_length(src),
5293 slow_region);
5294
5295 // (8) dest_offset + length must not exceed length of dest.
5296 generate_limit_guard(dest_offset, length,
5297 load_array_length(dest),
5298 slow_region);
5299
5300 // (6) length must not be negative.
5301 // This is also checked in generate_arraycopy() during macro expansion, but
5302 // we also have to check it here for the case where the ArrayCopyNode will
5303 // be eliminated by Escape Analysis.
5304 if (EliminateAllocations) {
5305 generate_negative_guard(length, slow_region);
5306 negative_length_guard_generated = true;
5307 }
5308
5309 // (9) each element of an oop array must be assignable
5310 Node* src_klass = load_object_klass(src);
5311 Node* dest_klass = load_object_klass(dest);
5312 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5313
5314 if (not_subtype_ctrl != top()) {
5315 PreserveJVMState pjvms(this);
5316 set_control(not_subtype_ctrl);
5317 uncommon_trap(Deoptimization::Reason_intrinsic,
5318 Deoptimization::Action_make_not_entrant);
5319 assert(stopped(), "Should be stopped");
5320 }
5321 {
5322 PreserveJVMState pjvms(this);
5323 set_control(_gvn.transform(slow_region));
5324 uncommon_trap(Deoptimization::Reason_intrinsic,
5325 Deoptimization::Action_make_not_entrant);
5326 assert(stopped(), "Should be stopped");
5327 }
5328
5329 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
5330 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
5331 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
5332 }
5333
5334 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5335
5336 if (stopped()) {
5337 return true;
5338 }
5339
5340 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5341 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5342 // so the compiler has a chance to eliminate them: during macro expansion,
5343 // we have to set their control (CastPP nodes are eliminated).
5344 load_object_klass(src), load_object_klass(dest),
5345 load_array_length(src), load_array_length(dest));
5346
5347 ac->set_arraycopy(validated);
5348
5349 Node* n = _gvn.transform(ac);
5350 if (n == ac) {
5351 ac->connect_outputs(this);
5352 } else {
5353 assert(validated, "shouldn't transform if all arguments not validated");
5354 set_all_memory(n);
5355 }
5356 clear_upper_avx();
5357
5358
5359 return true;
5360 }
5361
5362
5363 // Helper function which determines if an arraycopy immediately follows
5364 // an allocation, with no intervening tests or other escapes for the object.
5365 AllocateArrayNode*
5366 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5367 RegionNode* slow_region) {
5368 if (stopped()) return NULL; // no fast path
5369 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5370
5371 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5372 if (alloc == NULL) return NULL;
5373
5374 Node* rawmem = memory(Compile::AliasIdxRaw);
5375 // Is the allocation's memory state untouched?
5376 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5377 // Bail out if there have been raw-memory effects since the allocation.
5378 // (Example: There might have been a call or safepoint.)
5379 return NULL;
5380 }
5381 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5382 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5383 return NULL;
5384 }
5385
5386 // There must be no unexpected observers of this allocation.
5387 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5388 Node* obs = ptr->fast_out(i);
5389 if (obs != this->map()) {
5390 return NULL;
5391 }
5392 }
5393
5394 // This arraycopy must unconditionally follow the allocation of the ptr.
5395 Node* alloc_ctl = ptr->in(0);
5396 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5397
5398 Node* ctl = control();
5399 while (ctl != alloc_ctl) {
5400 // There may be guards which feed into the slow_region.
5401 // Any other control flow means that we might not get a chance
5402 // to finish initializing the allocated object.
5403 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5404 IfNode* iff = ctl->in(0)->as_If();
5405 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
5406 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5407 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5408 ctl = iff->in(0); // This test feeds the known slow_region.
5409 continue;
5410 }
5411 // One more try: Various low-level checks bottom out in
5412 // uncommon traps. If the debug-info of the trap omits
5413 // any reference to the allocation, as we've already
5414 // observed, then there can be no objection to the trap.
5415 bool found_trap = false;
5416 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5417 Node* obs = not_ctl->fast_out(j);
5418 if (obs->in(0) == not_ctl && obs->is_Call() &&
5419 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5420 found_trap = true; break;
5421 }
5422 }
5423 if (found_trap) {
5424 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5425 continue;
5426 }
5427 }
5428 return NULL;
5429 }
5430
5431 // If we get this far, we have an allocation which immediately
5432 // precedes the arraycopy, and we can take over zeroing the new object.
5433 // The arraycopy will finish the initialization, and provide
5434 // a new control state to which we will anchor the destination pointer.
5435
5436 return alloc;
5437 }
5438
5439 //-------------inline_encodeISOArray-----------------------------------
5440 // encode char[] to byte[] in ISO_8859_1
5441 bool LibraryCallKit::inline_encodeISOArray() {
5442 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5443 // no receiver since it is static method
5444 Node *src = argument(0);
5445 Node *src_offset = argument(1);
5446 Node *dst = argument(2);
5447 Node *dst_offset = argument(3);
5448 Node *length = argument(4);
5449
5450 const Type* src_type = src->Value(&_gvn);
5451 const Type* dst_type = dst->Value(&_gvn);
5452 const TypeAryPtr* top_src = src_type->isa_aryptr();
5453 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5454 if (top_src == NULL || top_src->klass() == NULL ||
5455 top_dest == NULL || top_dest->klass() == NULL) {
5456 // failed array check
5457 return false;
5458 }
5459
5460 // Figure out the size and type of the elements we will be copying.
5461 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5462 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5463 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5464 return false;
5465 }
5466
5467 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5468 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5469 // 'src_start' points to src array + scaled offset
5470 // 'dst_start' points to dst array + scaled offset
5471
5472 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5473 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5474 enc = _gvn.transform(enc);
5475 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5476 set_memory(res_mem, mtype);
5477 set_result(enc);
5478 clear_upper_avx();
5479
5480 return true;
5481 }
5482
5483 //-------------inline_multiplyToLen-----------------------------------
5484 bool LibraryCallKit::inline_multiplyToLen() {
5485 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5486
5487 address stubAddr = StubRoutines::multiplyToLen();
5488 if (stubAddr == NULL) {
5489 return false; // Intrinsic's stub is not implemented on this platform
5490 }
5491 const char* stubName = "multiplyToLen";
5492
5493 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5494
5495 // no receiver because it is a static method
5496 Node* x = argument(0);
5497 Node* xlen = argument(1);
5498 Node* y = argument(2);
5499 Node* ylen = argument(3);
5500 Node* z = argument(4);
5501
5502 const Type* x_type = x->Value(&_gvn);
5503 const Type* y_type = y->Value(&_gvn);
5504 const TypeAryPtr* top_x = x_type->isa_aryptr();
5505 const TypeAryPtr* top_y = y_type->isa_aryptr();
5506 if (top_x == NULL || top_x->klass() == NULL ||
5507 top_y == NULL || top_y->klass() == NULL) {
5508 // failed array check
5509 return false;
5510 }
5511
5512 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5513 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5514 if (x_elem != T_INT || y_elem != T_INT) {
5515 return false;
5516 }
5517
5518 // Set the original stack and the reexecute bit for the interpreter to reexecute
5519 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5520 // on the return from z array allocation in runtime.
5521 { PreserveReexecuteState preexecs(this);
5522 jvms()->set_should_reexecute(true);
5523
5524 Node* x_start = array_element_address(x, intcon(0), x_elem);
5525 Node* y_start = array_element_address(y, intcon(0), y_elem);
5526 // 'x_start' points to x array + scaled xlen
5527 // 'y_start' points to y array + scaled ylen
5528
5529 // Allocate the result array
5530 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5531 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5532 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5533
5534 IdealKit ideal(this);
5535
5536 #define __ ideal.
5537 Node* one = __ ConI(1);
5538 Node* zero = __ ConI(0);
5539 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5540 __ set(need_alloc, zero);
5541 __ set(z_alloc, z);
5542 __ if_then(z, BoolTest::eq, null()); {
5543 __ increment (need_alloc, one);
5544 } __ else_(); {
5545 // Update graphKit memory and control from IdealKit.
5546 sync_kit(ideal);
5547 Node* zlen_arg = load_array_length(z);
5548 // Update IdealKit memory and control from graphKit.
5549 __ sync_kit(this);
5550 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5551 __ increment (need_alloc, one);
5552 } __ end_if();
5553 } __ end_if();
5554
5555 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5556 // Update graphKit memory and control from IdealKit.
5557 sync_kit(ideal);
5558 Node * narr = new_array(klass_node, zlen, 1);
5559 // Update IdealKit memory and control from graphKit.
5560 __ sync_kit(this);
5561 __ set(z_alloc, narr);
5562 } __ end_if();
5563
5564 sync_kit(ideal);
5565 z = __ value(z_alloc);
5566 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5567 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5568 // Final sync IdealKit and GraphKit.
5569 final_sync(ideal);
5570 #undef __
5571
5572 Node* z_start = array_element_address(z, intcon(0), T_INT);
5573
5574 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5575 OptoRuntime::multiplyToLen_Type(),
5576 stubAddr, stubName, TypePtr::BOTTOM,
5577 x_start, xlen, y_start, ylen, z_start, zlen);
5578 } // original reexecute is set back here
5579
5580 C->set_has_split_ifs(true); // Has chance for split-if optimization
5581 set_result(z);
5582 return true;
5583 }
5584
5585 //-------------inline_squareToLen------------------------------------
5586 bool LibraryCallKit::inline_squareToLen() {
5587 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5588
5589 address stubAddr = StubRoutines::squareToLen();
5590 if (stubAddr == NULL) {
5591 return false; // Intrinsic's stub is not implemented on this platform
5592 }
5593 const char* stubName = "squareToLen";
5594
5595 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5596
5597 Node* x = argument(0);
5598 Node* len = argument(1);
5599 Node* z = argument(2);
5600 Node* zlen = argument(3);
5601
5602 const Type* x_type = x->Value(&_gvn);
5603 const Type* z_type = z->Value(&_gvn);
5604 const TypeAryPtr* top_x = x_type->isa_aryptr();
5605 const TypeAryPtr* top_z = z_type->isa_aryptr();
5606 if (top_x == NULL || top_x->klass() == NULL ||
5607 top_z == NULL || top_z->klass() == NULL) {
5608 // failed array check
5609 return false;
5610 }
5611
5612 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5613 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5614 if (x_elem != T_INT || z_elem != T_INT) {
5615 return false;
5616 }
5617
5618
5619 Node* x_start = array_element_address(x, intcon(0), x_elem);
5620 Node* z_start = array_element_address(z, intcon(0), z_elem);
5621
5622 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5623 OptoRuntime::squareToLen_Type(),
5624 stubAddr, stubName, TypePtr::BOTTOM,
5625 x_start, len, z_start, zlen);
5626
5627 set_result(z);
5628 return true;
5629 }
5630
5631 //-------------inline_mulAdd------------------------------------------
5632 bool LibraryCallKit::inline_mulAdd() {
5633 assert(UseMulAddIntrinsic, "not implemented on this platform");
5634
5635 address stubAddr = StubRoutines::mulAdd();
5636 if (stubAddr == NULL) {
5637 return false; // Intrinsic's stub is not implemented on this platform
5638 }
5639 const char* stubName = "mulAdd";
5640
5641 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5642
5643 Node* out = argument(0);
5644 Node* in = argument(1);
5645 Node* offset = argument(2);
5646 Node* len = argument(3);
5647 Node* k = argument(4);
5648
5649 const Type* out_type = out->Value(&_gvn);
5650 const Type* in_type = in->Value(&_gvn);
5651 const TypeAryPtr* top_out = out_type->isa_aryptr();
5652 const TypeAryPtr* top_in = in_type->isa_aryptr();
5653 if (top_out == NULL || top_out->klass() == NULL ||
5654 top_in == NULL || top_in->klass() == NULL) {
5655 // failed array check
5656 return false;
5657 }
5658
5659 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5660 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5661 if (out_elem != T_INT || in_elem != T_INT) {
5662 return false;
5663 }
5664
5665 Node* outlen = load_array_length(out);
5666 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5667 Node* out_start = array_element_address(out, intcon(0), out_elem);
5668 Node* in_start = array_element_address(in, intcon(0), in_elem);
5669
5670 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5671 OptoRuntime::mulAdd_Type(),
5672 stubAddr, stubName, TypePtr::BOTTOM,
5673 out_start,in_start, new_offset, len, k);
5674 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5675 set_result(result);
5676 return true;
5677 }
5678
5679 //-------------inline_montgomeryMultiply-----------------------------------
5680 bool LibraryCallKit::inline_montgomeryMultiply() {
5681 address stubAddr = StubRoutines::montgomeryMultiply();
5682 if (stubAddr == NULL) {
5683 return false; // Intrinsic's stub is not implemented on this platform
5684 }
5685
5686 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5687 const char* stubName = "montgomery_multiply";
5688
5689 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5690
5691 Node* a = argument(0);
5692 Node* b = argument(1);
5693 Node* n = argument(2);
5694 Node* len = argument(3);
5695 Node* inv = argument(4);
5696 Node* m = argument(6);
5697
5698 const Type* a_type = a->Value(&_gvn);
5699 const TypeAryPtr* top_a = a_type->isa_aryptr();
5700 const Type* b_type = b->Value(&_gvn);
5701 const TypeAryPtr* top_b = b_type->isa_aryptr();
5702 const Type* n_type = a->Value(&_gvn);
5703 const TypeAryPtr* top_n = n_type->isa_aryptr();
5704 const Type* m_type = a->Value(&_gvn);
5705 const TypeAryPtr* top_m = m_type->isa_aryptr();
5706 if (top_a == NULL || top_a->klass() == NULL ||
5707 top_b == NULL || top_b->klass() == NULL ||
5708 top_n == NULL || top_n->klass() == NULL ||
5709 top_m == NULL || top_m->klass() == NULL) {
5710 // failed array check
5711 return false;
5712 }
5713
5714 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5715 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5716 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5717 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5718 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5719 return false;
5720 }
5721
5722 // Make the call
5723 {
5724 Node* a_start = array_element_address(a, intcon(0), a_elem);
5725 Node* b_start = array_element_address(b, intcon(0), b_elem);
5726 Node* n_start = array_element_address(n, intcon(0), n_elem);
5727 Node* m_start = array_element_address(m, intcon(0), m_elem);
5728
5729 Node* call = make_runtime_call(RC_LEAF,
5730 OptoRuntime::montgomeryMultiply_Type(),
5731 stubAddr, stubName, TypePtr::BOTTOM,
5732 a_start, b_start, n_start, len, inv, top(),
5733 m_start);
5734 set_result(m);
5735 }
5736
5737 return true;
5738 }
5739
5740 bool LibraryCallKit::inline_montgomerySquare() {
5741 address stubAddr = StubRoutines::montgomerySquare();
5742 if (stubAddr == NULL) {
5743 return false; // Intrinsic's stub is not implemented on this platform
5744 }
5745
5746 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5747 const char* stubName = "montgomery_square";
5748
5749 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5750
5751 Node* a = argument(0);
5752 Node* n = argument(1);
5753 Node* len = argument(2);
5754 Node* inv = argument(3);
5755 Node* m = argument(5);
5756
5757 const Type* a_type = a->Value(&_gvn);
5758 const TypeAryPtr* top_a = a_type->isa_aryptr();
5759 const Type* n_type = a->Value(&_gvn);
5760 const TypeAryPtr* top_n = n_type->isa_aryptr();
5761 const Type* m_type = a->Value(&_gvn);
5762 const TypeAryPtr* top_m = m_type->isa_aryptr();
5763 if (top_a == NULL || top_a->klass() == NULL ||
5764 top_n == NULL || top_n->klass() == NULL ||
5765 top_m == NULL || top_m->klass() == NULL) {
5766 // failed array check
5767 return false;
5768 }
5769
5770 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5771 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5772 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5773 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5774 return false;
5775 }
5776
5777 // Make the call
5778 {
5779 Node* a_start = array_element_address(a, intcon(0), a_elem);
5780 Node* n_start = array_element_address(n, intcon(0), n_elem);
5781 Node* m_start = array_element_address(m, intcon(0), m_elem);
5782
5783 Node* call = make_runtime_call(RC_LEAF,
5784 OptoRuntime::montgomerySquare_Type(),
5785 stubAddr, stubName, TypePtr::BOTTOM,
5786 a_start, n_start, len, inv, top(),
5787 m_start);
5788 set_result(m);
5789 }
5790
5791 return true;
5792 }
5793
5794 //-------------inline_vectorizedMismatch------------------------------
5795 bool LibraryCallKit::inline_vectorizedMismatch() {
5796 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5797
5798 address stubAddr = StubRoutines::vectorizedMismatch();
5799 if (stubAddr == NULL) {
5800 return false; // Intrinsic's stub is not implemented on this platform
5801 }
5802 const char* stubName = "vectorizedMismatch";
5803 int size_l = callee()->signature()->size();
5804 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5805
5806 Node* obja = argument(0);
5807 Node* aoffset = argument(1);
5808 Node* objb = argument(3);
5809 Node* boffset = argument(4);
5810 Node* length = argument(6);
5811 Node* scale = argument(7);
5812
5813 const Type* a_type = obja->Value(&_gvn);
5814 const Type* b_type = objb->Value(&_gvn);
5815 const TypeAryPtr* top_a = a_type->isa_aryptr();
5816 const TypeAryPtr* top_b = b_type->isa_aryptr();
5817 if (top_a == NULL || top_a->klass() == NULL ||
5818 top_b == NULL || top_b->klass() == NULL) {
5819 // failed array check
5820 return false;
5821 }
5822
5823 Node* call;
5824 jvms()->set_should_reexecute(true);
5825
5826 Node* obja_adr = make_unsafe_address(obja, aoffset);
5827 Node* objb_adr = make_unsafe_address(objb, boffset);
5828
5829 call = make_runtime_call(RC_LEAF,
5830 OptoRuntime::vectorizedMismatch_Type(),
5831 stubAddr, stubName, TypePtr::BOTTOM,
5832 obja_adr, objb_adr, length, scale);
5833
5834 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5835 set_result(result);
5836 return true;
5837 }
5838
5839 /**
5840 * Calculate CRC32 for byte.
5841 * int java.util.zip.CRC32.update(int crc, int b)
5842 */
5843 bool LibraryCallKit::inline_updateCRC32() {
5844 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5845 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5846 // no receiver since it is static method
5847 Node* crc = argument(0); // type: int
5848 Node* b = argument(1); // type: int
5849
5850 /*
5851 * int c = ~ crc;
5852 * b = timesXtoThe32[(b ^ c) & 0xFF];
5853 * b = b ^ (c >>> 8);
5854 * crc = ~b;
5855 */
5856
5857 Node* M1 = intcon(-1);
5858 crc = _gvn.transform(new XorINode(crc, M1));
5859 Node* result = _gvn.transform(new XorINode(crc, b));
5860 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5861
5862 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5863 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5864 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5865 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5866
5867 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5868 result = _gvn.transform(new XorINode(crc, result));
5869 result = _gvn.transform(new XorINode(result, M1));
5870 set_result(result);
5871 return true;
5872 }
5873
5874 /**
5875 * Calculate CRC32 for byte[] array.
5876 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5877 */
5878 bool LibraryCallKit::inline_updateBytesCRC32() {
5879 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5880 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5881 // no receiver since it is static method
5882 Node* crc = argument(0); // type: int
5883 Node* src = argument(1); // type: oop
5884 Node* offset = argument(2); // type: int
5885 Node* length = argument(3); // type: int
5886
5887 const Type* src_type = src->Value(&_gvn);
5888 const TypeAryPtr* top_src = src_type->isa_aryptr();
5889 if (top_src == NULL || top_src->klass() == NULL) {
5890 // failed array check
5891 return false;
5892 }
5893
5894 // Figure out the size and type of the elements we will be copying.
5895 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5896 if (src_elem != T_BYTE) {
5897 return false;
5898 }
5899
5900 // 'src_start' points to src array + scaled offset
5901 Node* src_start = array_element_address(src, offset, src_elem);
5902
5903 // We assume that range check is done by caller.
5904 // TODO: generate range check (offset+length < src.length) in debug VM.
5905
5906 // Call the stub.
5907 address stubAddr = StubRoutines::updateBytesCRC32();
5908 const char *stubName = "updateBytesCRC32";
5909
5910 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5911 stubAddr, stubName, TypePtr::BOTTOM,
5912 crc, src_start, length);
5913 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5914 set_result(result);
5915 return true;
5916 }
5917
5918 /**
5919 * Calculate CRC32 for ByteBuffer.
5920 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5921 */
5922 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5923 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5924 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5925 // no receiver since it is static method
5926 Node* crc = argument(0); // type: int
5927 Node* src = argument(1); // type: long
5928 Node* offset = argument(3); // type: int
5929 Node* length = argument(4); // type: int
5930
5931 src = ConvL2X(src); // adjust Java long to machine word
5932 Node* base = _gvn.transform(new CastX2PNode(src));
5933 offset = ConvI2X(offset);
5934
5935 // 'src_start' points to src array + scaled offset
5936 Node* src_start = basic_plus_adr(top(), base, offset);
5937
5938 // Call the stub.
5939 address stubAddr = StubRoutines::updateBytesCRC32();
5940 const char *stubName = "updateBytesCRC32";
5941
5942 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5943 stubAddr, stubName, TypePtr::BOTTOM,
5944 crc, src_start, length);
5945 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5946 set_result(result);
5947 return true;
5948 }
5949
5950 //------------------------------get_table_from_crc32c_class-----------------------
5951 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5952 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5953 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5954
5955 return table;
5956 }
5957
5958 //------------------------------inline_updateBytesCRC32C-----------------------
5959 //
5960 // Calculate CRC32C for byte[] array.
5961 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5962 //
5963 bool LibraryCallKit::inline_updateBytesCRC32C() {
5964 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5965 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5966 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5967 // no receiver since it is a static method
5968 Node* crc = argument(0); // type: int
5969 Node* src = argument(1); // type: oop
5970 Node* offset = argument(2); // type: int
5971 Node* end = argument(3); // type: int
5972
5973 Node* length = _gvn.transform(new SubINode(end, offset));
5974
5975 const Type* src_type = src->Value(&_gvn);
5976 const TypeAryPtr* top_src = src_type->isa_aryptr();
5977 if (top_src == NULL || top_src->klass() == NULL) {
5978 // failed array check
5979 return false;
5980 }
5981
5982 // Figure out the size and type of the elements we will be copying.
5983 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5984 if (src_elem != T_BYTE) {
5985 return false;
5986 }
5987
5988 // 'src_start' points to src array + scaled offset
5989 Node* src_start = array_element_address(src, offset, src_elem);
5990
5991 // static final int[] byteTable in class CRC32C
5992 Node* table = get_table_from_crc32c_class(callee()->holder());
5993 Node* table_start = array_element_address(table, intcon(0), T_INT);
5994
5995 // We assume that range check is done by caller.
5996 // TODO: generate range check (offset+length < src.length) in debug VM.
5997
5998 // Call the stub.
5999 address stubAddr = StubRoutines::updateBytesCRC32C();
6000 const char *stubName = "updateBytesCRC32C";
6001
6002 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6003 stubAddr, stubName, TypePtr::BOTTOM,
6004 crc, src_start, length, table_start);
6005 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6006 set_result(result);
6007 return true;
6008 }
6009
6010 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
6011 //
6012 // Calculate CRC32C for DirectByteBuffer.
6013 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
6014 //
6015 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
6016 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
6017 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
6018 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
6019 // no receiver since it is a static method
6020 Node* crc = argument(0); // type: int
6021 Node* src = argument(1); // type: long
6022 Node* offset = argument(3); // type: int
6023 Node* end = argument(4); // type: int
6024
6025 Node* length = _gvn.transform(new SubINode(end, offset));
6026
6027 src = ConvL2X(src); // adjust Java long to machine word
6028 Node* base = _gvn.transform(new CastX2PNode(src));
6029 offset = ConvI2X(offset);
6030
6031 // 'src_start' points to src array + scaled offset
6032 Node* src_start = basic_plus_adr(top(), base, offset);
6033
6034 // static final int[] byteTable in class CRC32C
6035 Node* table = get_table_from_crc32c_class(callee()->holder());
6036 Node* table_start = array_element_address(table, intcon(0), T_INT);
6037
6038 // Call the stub.
6039 address stubAddr = StubRoutines::updateBytesCRC32C();
6040 const char *stubName = "updateBytesCRC32C";
6041
6042 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6043 stubAddr, stubName, TypePtr::BOTTOM,
6044 crc, src_start, length, table_start);
6045 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6046 set_result(result);
6047 return true;
6048 }
6049
6050 //------------------------------inline_updateBytesAdler32----------------------
6051 //
6052 // Calculate Adler32 checksum for byte[] array.
6053 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6054 //
6055 bool LibraryCallKit::inline_updateBytesAdler32() {
6056 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6057 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6058 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6059 // no receiver since it is static method
6060 Node* crc = argument(0); // type: int
6061 Node* src = argument(1); // type: oop
6062 Node* offset = argument(2); // type: int
6063 Node* length = argument(3); // type: int
6064
6065 const Type* src_type = src->Value(&_gvn);
6066 const TypeAryPtr* top_src = src_type->isa_aryptr();
6067 if (top_src == NULL || top_src->klass() == NULL) {
6068 // failed array check
6069 return false;
6070 }
6071
6072 // Figure out the size and type of the elements we will be copying.
6073 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6074 if (src_elem != T_BYTE) {
6075 return false;
6076 }
6077
6078 // 'src_start' points to src array + scaled offset
6079 Node* src_start = array_element_address(src, offset, src_elem);
6080
6081 // We assume that range check is done by caller.
6082 // TODO: generate range check (offset+length < src.length) in debug VM.
6083
6084 // Call the stub.
6085 address stubAddr = StubRoutines::updateBytesAdler32();
6086 const char *stubName = "updateBytesAdler32";
6087
6088 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6089 stubAddr, stubName, TypePtr::BOTTOM,
6090 crc, src_start, length);
6091 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6092 set_result(result);
6093 return true;
6094 }
6095
6096 //------------------------------inline_updateByteBufferAdler32---------------
6097 //
6098 // Calculate Adler32 checksum for DirectByteBuffer.
6099 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6100 //
6101 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6102 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6103 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6104 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6105 // no receiver since it is static method
6106 Node* crc = argument(0); // type: int
6107 Node* src = argument(1); // type: long
6108 Node* offset = argument(3); // type: int
6109 Node* length = argument(4); // type: int
6110
6111 src = ConvL2X(src); // adjust Java long to machine word
6112 Node* base = _gvn.transform(new CastX2PNode(src));
6113 offset = ConvI2X(offset);
6114
6115 // 'src_start' points to src array + scaled offset
6116 Node* src_start = basic_plus_adr(top(), base, offset);
6117
6118 // Call the stub.
6119 address stubAddr = StubRoutines::updateBytesAdler32();
6120 const char *stubName = "updateBytesAdler32";
6121
6122 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6123 stubAddr, stubName, TypePtr::BOTTOM,
6124 crc, src_start, length);
6125
6126 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6127 set_result(result);
6128 return true;
6129 }
6130
6131 //----------------------------inline_reference_get----------------------------
6132 // public T java.lang.ref.Reference.get();
6133 bool LibraryCallKit::inline_reference_get() {
6134 const int referent_offset = java_lang_ref_Reference::referent_offset;
6135 guarantee(referent_offset > 0, "should have already been set");
6136
6137 // Get the argument:
6138 Node* reference_obj = null_check_receiver();
6139 if (stopped()) return true;
6140
6141 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6142
6143 ciInstanceKlass* klass = env()->Object_klass();
6144 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6145
6146 Node* no_ctrl = NULL;
6147 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6148
6149 // Use the pre-barrier to record the value in the referent field
6150 pre_barrier(false /* do_load */,
6151 control(),
6152 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6153 result /* pre_val */,
6154 T_OBJECT);
6155
6156 // Add memory barrier to prevent commoning reads from this field
6157 // across safepoint since GC can change its value.
6158 insert_mem_bar(Op_MemBarCPUOrder);
6159
6160 set_result(result);
6161 return true;
6162 }
6163
6164
6165 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6166 bool is_exact=true, bool is_static=false,
6167 ciInstanceKlass * fromKls=NULL) {
6168 if (fromKls == NULL) {
6169 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6170 assert(tinst != NULL, "obj is null");
6171 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6172 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6173 fromKls = tinst->klass()->as_instance_klass();
6174 } else {
6175 assert(is_static, "only for static field access");
6176 }
6177 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6178 ciSymbol::make(fieldTypeString),
6179 is_static);
6180
6181 assert (field != NULL, "undefined field");
6182 if (field == NULL) return (Node *) NULL;
6183
6184 if (is_static) {
6185 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6186 fromObj = makecon(tip);
6187 }
6188
6189 // Next code copied from Parse::do_get_xxx():
6190
6191 // Compute address and memory type.
6192 int offset = field->offset_in_bytes();
6193 bool is_vol = field->is_volatile();
6194 ciType* field_klass = field->type();
6195 assert(field_klass->is_loaded(), "should be loaded");
6196 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6197 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6198 BasicType bt = field->layout_type();
6199
6200 // Build the resultant type of the load
6201 const Type *type;
6202 if (bt == T_OBJECT) {
6203 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6204 } else {
6205 type = Type::get_const_basic_type(bt);
6206 }
6207
6208 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6209 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6210 }
6211 // Build the load.
6212 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6213 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6214 // If reference is volatile, prevent following memory ops from
6215 // floating up past the volatile read. Also prevents commoning
6216 // another volatile read.
6217 if (is_vol) {
6218 // Memory barrier includes bogus read of value to force load BEFORE membar
6219 insert_mem_bar(Op_MemBarAcquire, loadedField);
6220 }
6221 return loadedField;
6222 }
6223
6224 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6225 bool is_exact = true, bool is_static = false,
6226 ciInstanceKlass * fromKls = NULL) {
6227 if (fromKls == NULL) {
6228 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6229 assert(tinst != NULL, "obj is null");
6230 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6231 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6232 fromKls = tinst->klass()->as_instance_klass();
6233 }
6234 else {
6235 assert(is_static, "only for static field access");
6236 }
6237 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6238 ciSymbol::make(fieldTypeString),
6239 is_static);
6240
6241 assert(field != NULL, "undefined field");
6242 assert(!field->is_volatile(), "not defined for volatile fields");
6243
6244 if (is_static) {
6245 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6246 fromObj = makecon(tip);
6247 }
6248
6249 // Next code copied from Parse::do_get_xxx():
6250
6251 // Compute address and memory type.
6252 int offset = field->offset_in_bytes();
6253 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6254
6255 return adr;
6256 }
6257
6258 //------------------------------inline_aescrypt_Block-----------------------
6259 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6260 address stubAddr = NULL;
6261 const char *stubName;
6262 assert(UseAES, "need AES instruction support");
6263
6264 switch(id) {
6265 case vmIntrinsics::_aescrypt_encryptBlock:
6266 stubAddr = StubRoutines::aescrypt_encryptBlock();
6267 stubName = "aescrypt_encryptBlock";
6268 break;
6269 case vmIntrinsics::_aescrypt_decryptBlock:
6270 stubAddr = StubRoutines::aescrypt_decryptBlock();
6271 stubName = "aescrypt_decryptBlock";
6272 break;
6273 default:
6274 break;
6275 }
6276 if (stubAddr == NULL) return false;
6277
6278 Node* aescrypt_object = argument(0);
6279 Node* src = argument(1);
6280 Node* src_offset = argument(2);
6281 Node* dest = argument(3);
6282 Node* dest_offset = argument(4);
6283
6284 // (1) src and dest are arrays.
6285 const Type* src_type = src->Value(&_gvn);
6286 const Type* dest_type = dest->Value(&_gvn);
6287 const TypeAryPtr* top_src = src_type->isa_aryptr();
6288 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6289 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6290
6291 // for the quick and dirty code we will skip all the checks.
6292 // we are just trying to get the call to be generated.
6293 Node* src_start = src;
6294 Node* dest_start = dest;
6295 if (src_offset != NULL || dest_offset != NULL) {
6296 assert(src_offset != NULL && dest_offset != NULL, "");
6297 src_start = array_element_address(src, src_offset, T_BYTE);
6298 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6299 }
6300
6301 // now need to get the start of its expanded key array
6302 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6303 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6304 if (k_start == NULL) return false;
6305
6306 if (Matcher::pass_original_key_for_aes()) {
6307 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6308 // compatibility issues between Java key expansion and SPARC crypto instructions
6309 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6310 if (original_k_start == NULL) return false;
6311
6312 // Call the stub.
6313 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6314 stubAddr, stubName, TypePtr::BOTTOM,
6315 src_start, dest_start, k_start, original_k_start);
6316 } else {
6317 // Call the stub.
6318 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6319 stubAddr, stubName, TypePtr::BOTTOM,
6320 src_start, dest_start, k_start);
6321 }
6322
6323 return true;
6324 }
6325
6326 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6327 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6328 address stubAddr = NULL;
6329 const char *stubName = NULL;
6330
6331 assert(UseAES, "need AES instruction support");
6332
6333 switch(id) {
6334 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6335 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6336 stubName = "cipherBlockChaining_encryptAESCrypt";
6337 break;
6338 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6339 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6340 stubName = "cipherBlockChaining_decryptAESCrypt";
6341 break;
6342 default:
6343 break;
6344 }
6345 if (stubAddr == NULL) return false;
6346
6347 Node* cipherBlockChaining_object = argument(0);
6348 Node* src = argument(1);
6349 Node* src_offset = argument(2);
6350 Node* len = argument(3);
6351 Node* dest = argument(4);
6352 Node* dest_offset = argument(5);
6353
6354 // (1) src and dest are arrays.
6355 const Type* src_type = src->Value(&_gvn);
6356 const Type* dest_type = dest->Value(&_gvn);
6357 const TypeAryPtr* top_src = src_type->isa_aryptr();
6358 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6359 assert (top_src != NULL && top_src->klass() != NULL
6360 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6361
6362 // checks are the responsibility of the caller
6363 Node* src_start = src;
6364 Node* dest_start = dest;
6365 if (src_offset != NULL || dest_offset != NULL) {
6366 assert(src_offset != NULL && dest_offset != NULL, "");
6367 src_start = array_element_address(src, src_offset, T_BYTE);
6368 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6369 }
6370
6371 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6372 // (because of the predicated logic executed earlier).
6373 // so we cast it here safely.
6374 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6375
6376 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6377 if (embeddedCipherObj == NULL) return false;
6378
6379 // cast it to what we know it will be at runtime
6380 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6381 assert(tinst != NULL, "CBC obj is null");
6382 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6383 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6384 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6385
6386 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6387 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6388 const TypeOopPtr* xtype = aklass->as_instance_type();
6389 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6390 aescrypt_object = _gvn.transform(aescrypt_object);
6391
6392 // we need to get the start of the aescrypt_object's expanded key array
6393 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6394 if (k_start == NULL) return false;
6395
6396 // similarly, get the start address of the r vector
6397 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6398 if (objRvec == NULL) return false;
6399 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6400
6401 Node* cbcCrypt;
6402 if (Matcher::pass_original_key_for_aes()) {
6403 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6404 // compatibility issues between Java key expansion and SPARC crypto instructions
6405 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6406 if (original_k_start == NULL) return false;
6407
6408 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6409 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6410 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6411 stubAddr, stubName, TypePtr::BOTTOM,
6412 src_start, dest_start, k_start, r_start, len, original_k_start);
6413 } else {
6414 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6415 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6416 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6417 stubAddr, stubName, TypePtr::BOTTOM,
6418 src_start, dest_start, k_start, r_start, len);
6419 }
6420
6421 // return cipher length (int)
6422 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6423 set_result(retvalue);
6424 return true;
6425 }
6426
6427 //------------------------------inline_counterMode_AESCrypt-----------------------
6428 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6429 assert(UseAES, "need AES instruction support");
6430 if (!UseAESCTRIntrinsics) return false;
6431
6432 address stubAddr = NULL;
6433 const char *stubName = NULL;
6434 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6435 stubAddr = StubRoutines::counterMode_AESCrypt();
6436 stubName = "counterMode_AESCrypt";
6437 }
6438 if (stubAddr == NULL) return false;
6439
6440 Node* counterMode_object = argument(0);
6441 Node* src = argument(1);
6442 Node* src_offset = argument(2);
6443 Node* len = argument(3);
6444 Node* dest = argument(4);
6445 Node* dest_offset = argument(5);
6446
6447 // (1) src and dest are arrays.
6448 const Type* src_type = src->Value(&_gvn);
6449 const Type* dest_type = dest->Value(&_gvn);
6450 const TypeAryPtr* top_src = src_type->isa_aryptr();
6451 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6452 assert(top_src != NULL && top_src->klass() != NULL &&
6453 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6454
6455 // checks are the responsibility of the caller
6456 Node* src_start = src;
6457 Node* dest_start = dest;
6458 if (src_offset != NULL || dest_offset != NULL) {
6459 assert(src_offset != NULL && dest_offset != NULL, "");
6460 src_start = array_element_address(src, src_offset, T_BYTE);
6461 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6462 }
6463
6464 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6465 // (because of the predicated logic executed earlier).
6466 // so we cast it here safely.
6467 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6468 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6469 if (embeddedCipherObj == NULL) return false;
6470 // cast it to what we know it will be at runtime
6471 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6472 assert(tinst != NULL, "CTR obj is null");
6473 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6474 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6475 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6476 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6477 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6478 const TypeOopPtr* xtype = aklass->as_instance_type();
6479 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6480 aescrypt_object = _gvn.transform(aescrypt_object);
6481 // we need to get the start of the aescrypt_object's expanded key array
6482 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6483 if (k_start == NULL) return false;
6484 // similarly, get the start address of the r vector
6485 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6486 if (obj_counter == NULL) return false;
6487 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6488
6489 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6490 if (saved_encCounter == NULL) return false;
6491 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6492 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6493
6494 Node* ctrCrypt;
6495 if (Matcher::pass_original_key_for_aes()) {
6496 // no SPARC version for AES/CTR intrinsics now.
6497 return false;
6498 }
6499 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6500 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6501 OptoRuntime::counterMode_aescrypt_Type(),
6502 stubAddr, stubName, TypePtr::BOTTOM,
6503 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6504
6505 // return cipher length (int)
6506 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6507 set_result(retvalue);
6508 return true;
6509 }
6510
6511 //------------------------------get_key_start_from_aescrypt_object-----------------------
6512 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6513 #if defined(PPC64) || defined(S390)
6514 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6515 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6516 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6517 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6518 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6519 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6520 if (objSessionK == NULL) {
6521 return (Node *) NULL;
6522 }
6523 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6524 #else
6525 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6526 #endif // PPC64
6527 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6528 if (objAESCryptKey == NULL) return (Node *) NULL;
6529
6530 // now have the array, need to get the start address of the K array
6531 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6532 return k_start;
6533 }
6534
6535 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6536 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6537 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6538 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6539 if (objAESCryptKey == NULL) return (Node *) NULL;
6540
6541 // now have the array, need to get the start address of the lastKey array
6542 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6543 return original_k_start;
6544 }
6545
6546 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6547 // Return node representing slow path of predicate check.
6548 // the pseudo code we want to emulate with this predicate is:
6549 // for encryption:
6550 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6551 // for decryption:
6552 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6553 // note cipher==plain is more conservative than the original java code but that's OK
6554 //
6555 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6556 // The receiver was checked for NULL already.
6557 Node* objCBC = argument(0);
6558
6559 // Load embeddedCipher field of CipherBlockChaining object.
6560 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6561
6562 // get AESCrypt klass for instanceOf check
6563 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6564 // will have same classloader as CipherBlockChaining object
6565 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6566 assert(tinst != NULL, "CBCobj is null");
6567 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6568
6569 // we want to do an instanceof comparison against the AESCrypt class
6570 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6571 if (!klass_AESCrypt->is_loaded()) {
6572 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6573 Node* ctrl = control();
6574 set_control(top()); // no regular fast path
6575 return ctrl;
6576 }
6577 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6578
6579 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6580 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6581 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6582
6583 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6584
6585 // for encryption, we are done
6586 if (!decrypting)
6587 return instof_false; // even if it is NULL
6588
6589 // for decryption, we need to add a further check to avoid
6590 // taking the intrinsic path when cipher and plain are the same
6591 // see the original java code for why.
6592 RegionNode* region = new RegionNode(3);
6593 region->init_req(1, instof_false);
6594 Node* src = argument(1);
6595 Node* dest = argument(4);
6596 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6597 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6598 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6599 region->init_req(2, src_dest_conjoint);
6600
6601 record_for_igvn(region);
6602 return _gvn.transform(region);
6603 }
6604
6605 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6606 // Return node representing slow path of predicate check.
6607 // the pseudo code we want to emulate with this predicate is:
6608 // for encryption:
6609 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6610 // for decryption:
6611 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6612 // note cipher==plain is more conservative than the original java code but that's OK
6613 //
6614
6615 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6616 // The receiver was checked for NULL already.
6617 Node* objCTR = argument(0);
6618
6619 // Load embeddedCipher field of CipherBlockChaining object.
6620 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6621
6622 // get AESCrypt klass for instanceOf check
6623 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6624 // will have same classloader as CipherBlockChaining object
6625 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6626 assert(tinst != NULL, "CTRobj is null");
6627 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6628
6629 // we want to do an instanceof comparison against the AESCrypt class
6630 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6631 if (!klass_AESCrypt->is_loaded()) {
6632 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6633 Node* ctrl = control();
6634 set_control(top()); // no regular fast path
6635 return ctrl;
6636 }
6637
6638 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6639 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6640 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6641 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6642 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6643
6644 return instof_false; // even if it is NULL
6645 }
6646
6647 //------------------------------inline_ghash_processBlocks
6648 bool LibraryCallKit::inline_ghash_processBlocks() {
6649 address stubAddr;
6650 const char *stubName;
6651 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6652
6653 stubAddr = StubRoutines::ghash_processBlocks();
6654 stubName = "ghash_processBlocks";
6655
6656 Node* data = argument(0);
6657 Node* offset = argument(1);
6658 Node* len = argument(2);
6659 Node* state = argument(3);
6660 Node* subkeyH = argument(4);
6661
6662 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6663 assert(state_start, "state is NULL");
6664 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6665 assert(subkeyH_start, "subkeyH is NULL");
6666 Node* data_start = array_element_address(data, offset, T_BYTE);
6667 assert(data_start, "data is NULL");
6668
6669 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6670 OptoRuntime::ghash_processBlocks_Type(),
6671 stubAddr, stubName, TypePtr::BOTTOM,
6672 state_start, subkeyH_start, data_start, len);
6673 return true;
6674 }
6675
6676 //------------------------------inline_sha_implCompress-----------------------
6677 //
6678 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6679 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6680 //
6681 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6682 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6683 //
6684 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6685 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6686 //
6687 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6688 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6689
6690 Node* sha_obj = argument(0);
6691 Node* src = argument(1); // type oop
6692 Node* ofs = argument(2); // type int
6693
6694 const Type* src_type = src->Value(&_gvn);
6695 const TypeAryPtr* top_src = src_type->isa_aryptr();
6696 if (top_src == NULL || top_src->klass() == NULL) {
6697 // failed array check
6698 return false;
6699 }
6700 // Figure out the size and type of the elements we will be copying.
6701 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6702 if (src_elem != T_BYTE) {
6703 return false;
6704 }
6705 // 'src_start' points to src array + offset
6706 Node* src_start = array_element_address(src, ofs, src_elem);
6707 Node* state = NULL;
6708 address stubAddr;
6709 const char *stubName;
6710
6711 switch(id) {
6712 case vmIntrinsics::_sha_implCompress:
6713 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6714 state = get_state_from_sha_object(sha_obj);
6715 stubAddr = StubRoutines::sha1_implCompress();
6716 stubName = "sha1_implCompress";
6717 break;
6718 case vmIntrinsics::_sha2_implCompress:
6719 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6720 state = get_state_from_sha_object(sha_obj);
6721 stubAddr = StubRoutines::sha256_implCompress();
6722 stubName = "sha256_implCompress";
6723 break;
6724 case vmIntrinsics::_sha5_implCompress:
6725 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6726 state = get_state_from_sha5_object(sha_obj);
6727 stubAddr = StubRoutines::sha512_implCompress();
6728 stubName = "sha512_implCompress";
6729 break;
6730 default:
6731 fatal_unexpected_iid(id);
6732 return false;
6733 }
6734 if (state == NULL) return false;
6735
6736 // Call the stub.
6737 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6738 stubAddr, stubName, TypePtr::BOTTOM,
6739 src_start, state);
6740
6741 return true;
6742 }
6743
6744 //------------------------------inline_digestBase_implCompressMB-----------------------
6745 //
6746 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6747 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6748 //
6749 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6750 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6751 "need SHA1/SHA256/SHA512 instruction support");
6752 assert((uint)predicate < 3, "sanity");
6753 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6754
6755 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6756 Node* src = argument(1); // byte[] array
6757 Node* ofs = argument(2); // type int
6758 Node* limit = argument(3); // type int
6759
6760 const Type* src_type = src->Value(&_gvn);
6761 const TypeAryPtr* top_src = src_type->isa_aryptr();
6762 if (top_src == NULL || top_src->klass() == NULL) {
6763 // failed array check
6764 return false;
6765 }
6766 // Figure out the size and type of the elements we will be copying.
6767 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6768 if (src_elem != T_BYTE) {
6769 return false;
6770 }
6771 // 'src_start' points to src array + offset
6772 Node* src_start = array_element_address(src, ofs, src_elem);
6773
6774 const char* klass_SHA_name = NULL;
6775 const char* stub_name = NULL;
6776 address stub_addr = NULL;
6777 bool long_state = false;
6778
6779 switch (predicate) {
6780 case 0:
6781 if (UseSHA1Intrinsics) {
6782 klass_SHA_name = "sun/security/provider/SHA";
6783 stub_name = "sha1_implCompressMB";
6784 stub_addr = StubRoutines::sha1_implCompressMB();
6785 }
6786 break;
6787 case 1:
6788 if (UseSHA256Intrinsics) {
6789 klass_SHA_name = "sun/security/provider/SHA2";
6790 stub_name = "sha256_implCompressMB";
6791 stub_addr = StubRoutines::sha256_implCompressMB();
6792 }
6793 break;
6794 case 2:
6795 if (UseSHA512Intrinsics) {
6796 klass_SHA_name = "sun/security/provider/SHA5";
6797 stub_name = "sha512_implCompressMB";
6798 stub_addr = StubRoutines::sha512_implCompressMB();
6799 long_state = true;
6800 }
6801 break;
6802 default:
6803 fatal("unknown SHA intrinsic predicate: %d", predicate);
6804 }
6805 if (klass_SHA_name != NULL) {
6806 // get DigestBase klass to lookup for SHA klass
6807 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6808 assert(tinst != NULL, "digestBase_obj is not instance???");
6809 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6810
6811 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6812 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6813 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6814 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6815 }
6816 return false;
6817 }
6818 //------------------------------inline_sha_implCompressMB-----------------------
6819 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6820 bool long_state, address stubAddr, const char *stubName,
6821 Node* src_start, Node* ofs, Node* limit) {
6822 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6823 const TypeOopPtr* xtype = aklass->as_instance_type();
6824 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6825 sha_obj = _gvn.transform(sha_obj);
6826
6827 Node* state;
6828 if (long_state) {
6829 state = get_state_from_sha5_object(sha_obj);
6830 } else {
6831 state = get_state_from_sha_object(sha_obj);
6832 }
6833 if (state == NULL) return false;
6834
6835 // Call the stub.
6836 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6837 OptoRuntime::digestBase_implCompressMB_Type(),
6838 stubAddr, stubName, TypePtr::BOTTOM,
6839 src_start, state, ofs, limit);
6840 // return ofs (int)
6841 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6842 set_result(result);
6843
6844 return true;
6845 }
6846
6847 enum VectorApiObjectType {
6848 VECAPI_VECTOR,
6849 VECAPI_MASK,
6850 VECAPI_SPECIES,
6851 VECAPI_SHUFFLE,
6852 };
6853
6854 static ciKlass* get_exact_klass_for_vector_box(ciKlass* ctx, BasicType bt, int num_elem, VectorApiObjectType obj_type) {
6855 // FIXME: use VM symbol table instead
6856 ResourceMark rm;
6857 stringStream ss;
6858 ss.print_raw("jdk/incubator/vector/");
6859 switch (bt) {
6860 case T_BYTE: ss.print_raw("Byte"); break;
6861 case T_SHORT: ss.print_raw("Short"); break;
6862 case T_INT: ss.print_raw("Int"); break;
6863 case T_LONG: ss.print_raw("Long"); break;
6864 case T_FLOAT: ss.print_raw("Float"); break;
6865 case T_DOUBLE: ss.print_raw("Double"); break;
6866 default: fatal("unknown element type: %d", bt);
6867 }
6868 int bits = num_elem * BitsPerByte * type2aelembytes(bt);
6869 switch (bits) {
6870 case 64: // fall through
6871 case 128: // fall through
6872 case 256: // fall through
6873 case 512: ss.print("%d", bits); break;
6874 default: fatal("unknown vector size: %d", bits);
6875 }
6876 ss.print_raw("Vector");
6877 if (obj_type != VECAPI_VECTOR) {
6878 ss.print_raw("$");
6879 switch (bt) {
6880 case T_BYTE: ss.print_raw("Byte"); break;
6881 case T_SHORT: ss.print_raw("Short"); break;
6882 case T_INT: ss.print_raw("Int"); break;
6883 case T_LONG: ss.print_raw("Long"); break;
6884 case T_FLOAT: ss.print_raw("Float"); break;
6885 case T_DOUBLE: ss.print_raw("Double"); break;
6886 default: fatal("unknown element type: %d", bt);
6887 }
6888 switch (bits) {
6889 case 64: // fall through
6890 case 128: // fall through
6891 case 256: // fall through
6892 case 512: ss.print("%d", bits); break;
6893 default: fatal("unknown vector size: %d", bits);
6894 }
6895 switch (obj_type) {
6896 case VECAPI_MASK: ss.print_raw("Mask"); break;
6897 case VECAPI_SPECIES: ss.print_raw("Species"); break;
6898 case VECAPI_SHUFFLE: ss.print_raw("Shuffle"); break;
6899 default: fatal("unknown vector object type");
6900 }
6901 }
6902
6903 ciSymbol* vector_klass_name = ciSymbol::make(ss.as_string());
6904 ciKlass* vector_klass = ctx->find_klass(vector_klass_name);
6905 return vector_klass;
6906 }
6907
6908 static BasicType getMaskBasicType(BasicType use_type) {
6909 if (use_type == T_FLOAT) {
6910 return T_INT;
6911 }
6912 if (use_type == T_DOUBLE) {
6913 return T_LONG;
6914 }
6915 assert(is_java_primitive(use_type), "must be java primitive");
6916 return use_type;
6917 }
6918
6919 enum VectorMaskUseType {
6920 VecMaskUseLoad,
6921 VecMaskUseStore,
6922 VecMaskUseAll,
6923 VecMaskNotUsed
6924 };
6925
6926 static bool arch_supports_vector(int op, int num_elem, BasicType type, VectorMaskUseType mask_use_type) {
6927 // Check that the operation is valid.
6928 if (op <= 0) {
6929 #ifndef PRODUCT
6930 if (DebugVectorApi) {
6931 tty->print_cr("Rejected intrinsification because no valid vector op could be extracted");
6932 }
6933 #endif
6934 return false;
6935 }
6936
6937 // Check that architecture supports this op-size-type combination.
6938 if (!Matcher::match_rule_supported_vector(op, num_elem, type)) {
6939 #ifndef PRODUCT
6940 if (DebugVectorApi) {
6941 tty->print_cr("Rejected vector op (%s,%s,%d) because architecture does not support it",
6942 NodeClassNames[op], type2name(type), num_elem);
6943 }
6944 #endif
6945 return false;
6946 } else {
6947 assert(Matcher::match_rule_supported(op), "must be supported");
6948 }
6949
6950 // Check whether mask unboxing is supported.
6951 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseLoad) {
6952 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type)) {
6953 #ifndef PRODUCT
6954 if (DebugVectorApi) {
6955 tty->print_cr("Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
6956 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
6957 }
6958 #endif
6959 return false;
6960 }
6961 }
6962
6963 // Check whether mask boxing is supported.
6964 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseStore) {
6965 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type)) {
6966 #ifndef PRODUCT
6967 if (DebugVectorApi) {
6968 tty->print_cr("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
6969 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
6970 }
6971 #endif
6972 return false;
6973 }
6974 }
6975
6976 return true;
6977 }
6978
6979 // Should be aligned with constants in jdk.incubator.vector.VectorIntrinsics.
6980 enum {
6981 // Unary
6982 OP_ABS = 0,
6983 OP_NEG = 1,
6984 OP_SQRT = 2,
6985 OP_NOT = 3,
6986 // Binary
6987 OP_ADD = 4,
6988 OP_SUB = 5,
6989 OP_MUL = 6,
6990 OP_DIV = 7,
6991 OP_AND = 8,
6992 OP_OR = 9,
6993 OP_XOR = 10,
6994 // Ternary
6995 OP_FMA = 11,
6996 // Broadcast int
6997 OP_LSHIFT = 12,
6998 OP_RSHIFT = 13,
6999 OP_URSHIFT = 14
7000 };
7001
7002 static int get_opc(jint op, BasicType bt) {
7003 switch (op) {
7004 case OP_ADD: {
7005 switch (bt) {
7006 case T_BYTE: // fall-through
7007 case T_SHORT: // fall-through
7008 case T_INT: return Op_AddI;
7009 case T_LONG: return Op_AddL;
7010 case T_FLOAT: return Op_AddF;
7011 case T_DOUBLE: return Op_AddD;
7012 default: fatal("ADD: %s", type2name(bt));
7013 }
7014 break;
7015 }
7016 case OP_SUB: {
7017 switch (bt) {
7018 case T_BYTE: // fall-through
7019 case T_SHORT: // fall-through
7020 case T_INT: return Op_SubI;
7021 case T_LONG: return Op_SubL;
7022 case T_FLOAT: return Op_SubF;
7023 case T_DOUBLE: return Op_SubD;
7024 default: fatal("SUB: %s", type2name(bt));
7025 }
7026 break;
7027 }
7028 case OP_MUL: {
7029 switch (bt) {
7030 case T_BYTE: // fall-through
7031 case T_SHORT: // fall-through
7032 case T_INT: return Op_MulI;
7033 case T_LONG: return Op_MulL;
7034 case T_FLOAT: return Op_MulF;
7035 case T_DOUBLE: return Op_MulD;
7036 default: fatal("MUL: %s", type2name(bt));
7037 }
7038 break;
7039 }
7040 case OP_DIV: {
7041 switch (bt) {
7042 case T_BYTE: // fall-through
7043 case T_SHORT: // fall-through
7044 case T_INT: return Op_DivI;
7045 case T_LONG: return Op_DivL;
7046 case T_FLOAT: return Op_DivF;
7047 case T_DOUBLE: return Op_DivD;
7048 default: fatal("DIV: %s", type2name(bt));
7049 }
7050 break;
7051 }
7052 case OP_ABS: {
7053 switch (bt) {
7054 case T_BYTE: // fall-through
7055 case T_SHORT: // fall-through
7056 case T_INT: return Op_AbsI;
7057 case T_FLOAT: return Op_AbsF;
7058 case T_DOUBLE: return Op_AbsD;
7059 default: fatal("ABS: %s", type2name(bt));
7060 }
7061 break;
7062 }
7063 case OP_NEG: {
7064 switch (bt) {
7065 case T_BYTE: // fall-through
7066 case T_SHORT: // fall-through
7067 case T_INT: return Op_NegI;
7068 case T_FLOAT: return Op_NegF;
7069 case T_DOUBLE: return Op_NegD;
7070 default: fatal("NEG: %s", type2name(bt));
7071 }
7072 break;
7073 }
7074 case OP_AND: {
7075 switch (bt) {
7076 case T_BYTE: // fall-through
7077 case T_SHORT: // fall-through
7078 case T_INT: return Op_AndI;
7079 case T_LONG: return Op_AndL;
7080 default: fatal("AND: %s", type2name(bt));
7081 }
7082 break;
7083 }
7084 case OP_OR: {
7085 switch (bt) {
7086 case T_BYTE: // fall-through
7087 case T_SHORT: // fall-through
7088 case T_INT: return Op_OrI;
7089 case T_LONG: return Op_OrL;
7090 default: fatal("OR: %s", type2name(bt));
7091 }
7092 break;
7093 }
7094 case OP_XOR: {
7095 switch (bt) {
7096 case T_BYTE: // fall-through
7097 case T_SHORT: // fall-through
7098 case T_INT: return Op_XorI;
7099 case T_LONG: return Op_XorL;
7100 default: fatal("XOR: %s", type2name(bt));
7101 }
7102 break;
7103 }
7104 case OP_SQRT: {
7105 switch (bt) {
7106 case T_FLOAT: return Op_SqrtF;
7107 case T_DOUBLE: return Op_SqrtD;
7108 default: fatal("SQRT: %s", type2name(bt));
7109 }
7110 break;
7111 }
7112 case OP_NOT: {
7113 switch (bt) {
7114 case T_BYTE: // fall-through
7115 case T_SHORT: // fall-through
7116 case T_INT: // fall-through
7117 case T_LONG: return Op_Not;
7118 default: fatal("NOT: %s", type2name(bt));
7119 }
7120 break;
7121 }
7122 case OP_FMA: {
7123 switch (bt) {
7124 case T_FLOAT: return Op_FmaF;
7125 case T_DOUBLE: return Op_FmaD;
7126 default: fatal("FMA: %s", type2name(bt));
7127 }
7128 break;
7129 }
7130 case OP_LSHIFT: {
7131 switch (bt) {
7132 case T_BYTE: // fall-through
7133 case T_SHORT: // fall-through
7134 case T_INT: return Op_LShiftI;
7135 case T_LONG: return Op_LShiftL;
7136 default: fatal("LSHIFT: %s", type2name(bt));
7137 }
7138 }
7139 case OP_RSHIFT: {
7140 switch (bt) {
7141 case T_BYTE: // fall-through
7142 case T_SHORT: // fall-through
7143 case T_INT: return Op_RShiftI;
7144 case T_LONG: return Op_RShiftL;
7145 default: fatal("RSHIFT: %s", type2name(bt));
7146 }
7147 }
7148 case OP_URSHIFT: {
7149 switch (bt) {
7150 case T_BYTE: // fall-through
7151 case T_SHORT: // fall-through
7152 case T_INT: return Op_URShiftI;
7153 case T_LONG: return Op_URShiftL;
7154 default: fatal("URSHIFT: %s", type2name(bt));
7155 }
7156 }
7157 default: fatal("unknown op: %d", op);
7158 }
7159 return 0; // Unimplemented
7160 }
7161
7162 Node* LibraryCallKit::box_vector(Node* vector, const TypeInstPtr* vbox_type,
7163 BasicType elem_bt, int num_elem) {
7164
7165 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
7166
7167 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
7168 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
7169 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true);
7170 set_i_o(_gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
7171 set_all_memory(_gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
7172 Node* ret = _gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
7173
7174 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vec_type);
7175 return _gvn.transform(vbox);
7176 }
7177
7178 Node* LibraryCallKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) {
7179 const TypeInstPtr* vbox_type_v = gvn().type(v)->is_instptr();
7180 if (vbox_type->klass() != vbox_type_v->klass()) {
7181 return NULL; // arguments don't agree on vector shapes
7182 }
7183 if (vbox_type_v->maybe_null()) {
7184 return NULL; // no nulls are allowed
7185 }
7186 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
7187 Node* unbox = gvn().transform(new VectorUnboxNode(C, vec_type, v, merged_memory()));
7188 return unbox;
7189 }
7190
7191 void LibraryCallKit::set_vector_result(Node* result, bool set_res) {
7192 if (DebugVectorApi) {
7193 tty->print("============ ");
7194 callee()->print();
7195 tty->print_cr(" ============");
7196 result->dump(5);
7197 tty->print_cr("----------------------------------------------------");
7198 }
7199 if (set_res) {
7200 set_result(result);
7201 }
7202 }
7203
7204 bool LibraryCallKit::inline_vector_nary_operation(int n) {
7205 const TypeInt* opr = gvn().type(argument(0))->is_int();
7206 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7207 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7208 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7209
7210 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7211 return false; // not enough info for intrinsification
7212 }
7213 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7214 if (!elem_type->is_primitive_type()) {
7215 return false; // should be primitive type
7216 }
7217 BasicType elem_bt = elem_type->basic_type();
7218 int num_elem = vlen->get_con();
7219 int opc = get_opc(opr->get_con(), elem_bt);
7220 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
7221 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7222 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7223
7224 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7225 if (!arch_supports_vector(sopc, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseAll : VecMaskNotUsed)) {
7226 return false; // not supported
7227 }
7228
7229 Node* opd1 = NULL; Node* opd2 = NULL; Node* opd3 = NULL;
7230 switch (n) {
7231 case 3: {
7232 opd1 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7233 if (opd1 == NULL) {
7234 return false;
7235 }
7236 // fall-through
7237 }
7238 case 2: {
7239 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7240 if (opd2 == NULL) {
7241 return false;
7242 }
7243 // fall-through
7244 }
7245 case 1: {
7246 opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7247 if (opd1 == NULL) {
7248 return false;
7249 }
7250 break;
7251 }
7252 default: fatal("unsupported arity: %d", n);
7253 }
7254
7255 Node* operation = NULL;
7256 switch (n) {
7257 case 1:
7258 case 2: {
7259 operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, num_elem, elem_bt));
7260 break;
7261 }
7262 case 3: {
7263 operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, opd3, num_elem, elem_bt));
7264 break;
7265 }
7266 default: fatal("unsupported arity: %d", n);
7267 }
7268 // Wrap it up in VectorBox to keep object type information.
7269 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
7270 set_vector_result(operation);
7271
7272 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7273 return true;
7274 }
7275
7276 // <V extends Vector<?,?>>
7277 // V broadcastCoerced(Class<?> vectorClass, Class<?> elementType, int vlen,
7278 // long bits,
7279 // LongFunction<V> defaultImpl)
7280 bool LibraryCallKit::inline_vector_broadcast_coerced() {
7281 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7282 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7283 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7284
7285 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7286 return false; // not enough info for intrinsification
7287 }
7288
7289 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7290 if (!elem_type->is_primitive_type()) {
7291 return false; // should be primitive type
7292 }
7293 BasicType elem_bt = elem_type->basic_type();
7294 int num_elem = vlen->get_con();
7295 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7296 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7297
7298 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7299 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt,
7300 vbox_klass->is_vectormask() ? VecMaskUseStore : VecMaskNotUsed)) {
7301 return false; // not supported
7302 }
7303
7304 Node* bits = argument(3); // long
7305
7306 Node* elem = NULL;
7307 switch (elem_bt) {
7308 case T_BOOLEAN: // fall-through
7309 case T_BYTE: // fall-through
7310 case T_SHORT: // fall-through
7311 case T_CHAR: // fall-through
7312 case T_INT: {
7313 elem = gvn().transform(new ConvL2INode(bits));
7314 break;
7315 }
7316 case T_DOUBLE: {
7317 elem = gvn().transform(new MoveL2DNode(bits));
7318 break;
7319 }
7320 case T_FLOAT: {
7321 bits = gvn().transform(new ConvL2INode(bits));
7322 elem = gvn().transform(new MoveI2FNode(bits));
7323 break;
7324 }
7325 case T_LONG: {
7326 elem = bits; // no conversion needed
7327 break;
7328 }
7329 default: fatal("%s", type2name(elem_bt));
7330 }
7331
7332 Node* broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt));
7333 broadcast = gvn().transform(broadcast);
7334 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
7335 set_vector_result(box);
7336
7337 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7338 return true;
7339 }
7340
7341 // <V extends Vector<?,?>>
7342 // V load(Class<?> vectorClass, Class<?> elementType, int vlen,
7343 // Object array, int index, /* Vector.Mask<E,S> m*/
7344 // BiFunction<Object, Integer, V> defaultImpl) {
7345 //
7346 // <V extends Vector<?,?>>
7347 // void store(Class<?> vectorClass, Class<?> elementType, int vlen,
7348 // Object array, int index, V v, /*Vector.Mask<E,S> m*/
7349 // StoreVectorOperation<V> defaultImpl) {
7350
7351 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
7352 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7353 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7354 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7355
7356 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7357 return false; // not enough info for intrinsification
7358 }
7359
7360 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7361 if (!elem_type->is_primitive_type()) {
7362 return false; // should be primitive type
7363 }
7364 BasicType elem_bt = elem_type->basic_type();
7365 int num_elem = vlen->get_con();
7366
7367 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7368 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
7369 return false; // not supported
7370 }
7371
7372 Node* arr = argument(3);
7373 Node* idx = argument(4);
7374
7375 const TypeAryPtr* arr_type = gvn().type(arr)->isa_aryptr();
7376 if (arr_type == NULL) {
7377 return false; // should be an array
7378 }
7379 if (elem_bt != arr_type->elem()->array_element_basic_type()) {
7380 return false; // array & vector element types should be the same
7381 }
7382 Node* adr = array_element_address(arr, idx, elem_bt);
7383 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
7384
7385 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7386 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7387
7388 if (is_store) {
7389 Node* val = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7390 if (val == NULL) {
7391 return false; // operand unboxing failed
7392 }
7393 set_all_memory(reset_memory());
7394 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(adr), adr, adr_type, val, num_elem));
7395 set_memory(vstore, adr_type);
7396 set_vector_result(vstore, false);
7397 } else {
7398 Node* vload = gvn().transform(LoadVectorNode::make(0, control(), memory(adr), adr, adr_type, num_elem, elem_bt));
7399 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7400 set_vector_result(box);
7401 }
7402
7403 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7404 return true;
7405 }
7406
7407 // <V extends Vector<?,?>>
7408 // long reductionCoerced(int oprId, Class<?> vectorClass, Class<?> elementType, int vlen,
7409 // V v,
7410 // Function<V,Long> defaultImpl)
7411
7412 bool LibraryCallKit::inline_vector_reduction() {
7413 const TypeInt* opr = gvn().type(argument(0))->is_int();
7414 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7415 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7416 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7417
7418 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7419 return false; // not enough info for intrinsification
7420 }
7421 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7422 if (!elem_type->is_primitive_type()) {
7423 return false; // should be primitive type
7424 }
7425 BasicType elem_bt = elem_type->basic_type();
7426 int num_elem = vlen->get_con();
7427
7428 int opc = get_opc(opr->get_con(), elem_bt);
7429 int sopc = ReductionNode::opcode(opc, elem_bt);
7430
7431 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7432 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7433 return false;
7434 }
7435
7436 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7437 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7438
7439 Node* opd = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7440 if (opd == NULL) {
7441 return false; // operand unboxing failed
7442 }
7443
7444 Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt);
7445 Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, elem_bt));
7446
7447 Node* bits = NULL;
7448 switch (elem_bt) {
7449 case T_INT: {
7450 bits = gvn().transform(new ConvI2LNode(rn));
7451 break;
7452 }
7453 case T_FLOAT: {
7454 rn = gvn().transform(new MoveF2INode(rn));
7455 bits = gvn().transform(new ConvI2LNode(rn));
7456 break;
7457 }
7458 case T_DOUBLE: {
7459 bits = gvn().transform(new MoveD2LNode(rn));
7460 break;
7461 }
7462 case T_LONG: {
7463 bits = rn; // no conversion needed
7464 break;
7465 }
7466 default: fatal("%s", type2name(elem_bt));
7467 }
7468 set_vector_result(bits);
7469 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7470 return true;
7471 }
7472
7473 // static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
7474 // V v1, V v2,
7475 // BiFunction<V, V, Boolean> defaultImpl) {
7476
7477 bool LibraryCallKit::inline_vector_test() {
7478 const TypeInt* cond = gvn().type(argument(0))->is_int();
7479 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7480 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7481 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7482
7483 if (!cond->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7484 return false; // not enough info for intrinsification
7485 }
7486 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7487 if (!elem_type->is_primitive_type()) {
7488 return false; // should be primitive type
7489 }
7490 BasicType elem_bt = elem_type->basic_type();
7491 int num_elem = vlen->get_con();
7492 Assembler::Condition booltest = (Assembler::Condition)cond->get_con();
7493 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7494 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7495
7496 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseLoad : VecMaskNotUsed)) {
7497 return false;
7498 }
7499
7500 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7501 Node* opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7502 if (opd1 == NULL || opd2 == NULL) {
7503 return false; // operand unboxing failed
7504 }
7505 Node* test = new VectorTestNode(opd1, opd2, booltest);
7506 test = _gvn.transform(test);
7507 set_vector_result(test);
7508 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7509 return true;
7510 }
7511
7512 // static
7513 // <V extends Vector, M extends Mask>
7514 // V blend(Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7515 // V v1, V v2, M m,
7516 // VectorBlendOp<V,M> defaultImpl) { ...
7517 //
7518 bool LibraryCallKit::inline_vector_blend() {
7519 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7520 const TypeInstPtr* mask_klass = gvn().type(argument(1))->is_instptr();
7521 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7522 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7523
7524 if (mask_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7525 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7526 return false; // not enough info for intrinsification
7527 }
7528 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7529 if (!elem_type->is_primitive_type()) {
7530 return false; // should be primitive type
7531 }
7532 BasicType elem_bt = elem_type->basic_type();
7533 BasicType mask_bt = getMaskBasicType(elem_bt);
7534 int num_elem = vlen->get_con();
7535
7536 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
7537 return false; // not supported
7538 }
7539 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7540 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7541
7542 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7543 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7544
7545 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7546 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7547 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
7548
7549 if (v1 == NULL || v2 == NULL || mask == NULL) {
7550 return false; // operand unboxing failed
7551 }
7552
7553 Node* blend = _gvn.transform(new VectorBlendNode(v1, v2, mask));
7554 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
7555 set_vector_result(box);
7556
7557 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7558 return true;
7559 }
7560
7561 // static <V extends Vector<E,S>,
7562 // M extends Vector.Mask<E,S>,
7563 // S extends Vector.Shape, E>
7564 // M compare(int cond, Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7565 // V v1, V v2,
7566 // VectorCompareOp<V,M> defaultImpl) { ...
7567 //
7568 bool LibraryCallKit::inline_vector_compare() {
7569 const TypeInt* cond = gvn().type(argument(0))->is_int();
7570 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7571 const TypeInstPtr* mask_klass = gvn().type(argument(2))->is_instptr();
7572 const TypeInstPtr* elem_klass = gvn().type(argument(3))->is_instptr();
7573 const TypeInt* vlen = gvn().type(argument(4))->is_int();
7574
7575 if (!cond->is_con() || vector_klass->const_oop() == NULL || mask_klass->const_oop() == NULL ||
7576 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7577 return false; // not enough info for intrinsification
7578 }
7579 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7580 if (!elem_type->is_primitive_type()) {
7581 return false; // should be primitive type
7582 }
7583
7584 int num_elem = vlen->get_con();
7585 BasicType elem_bt = elem_type->basic_type();
7586 BasicType mask_bt = getMaskBasicType(elem_bt);
7587
7588 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
7589 return false;
7590 }
7591
7592 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7593 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7594
7595 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7596 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7597
7598 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7599 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7600
7601 if (v1 == NULL || v2 == NULL) {
7602 return false; // operand unboxing failed
7603 }
7604 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
7605 const TypeVect* vt = TypeVect::make(mask_bt, num_elem);
7606 Node* operation = _gvn.transform(new VectorMaskCmpNode(pred, v1, v2, vt));
7607
7608 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
7609 set_vector_result(box);
7610
7611 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7612 return true;
7613 }
7614
7615 Node* LibraryCallKit::shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
7616 assert(bt == T_INT || bt == T_LONG, "Only long and int are supported");
7617 juint mask = (bt == T_INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
7618 const TypeInt* t = cnt->find_int_type();
7619 if (t != NULL && t->is_con()) {
7620 juint shift = t->get_con();
7621 if (shift > mask) {
7622 return _gvn.transform(ConNode::make(TypeInt::make(shift & mask)));
7623 } else {
7624 return cnt;
7625 }
7626 } else {
7627 Node* maskedcnt = cnt;
7628 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
7629 Node* nmask = _gvn.transform(ConNode::make(TypeInt::make(mask)));
7630 maskedcnt = _gvn.transform(new AndINode(cnt, nmask));
7631 }
7632 return VectorNode::shift_count(shift_op, maskedcnt, num_elem, bt);
7633 }
7634 }
7635
7636 // static
7637 // <V extends Vector<?,?>>
7638 // V broadcastInt(int opr, Class<V> vectorClass, Class<?> elementType, int vlen,
7639 // V v, int i,
7640 // VectorBroadcastIntOp<V> defaultImpl) {
7641 //
7642 bool LibraryCallKit::inline_vector_broadcast_int() {
7643 const TypeInt* opr = gvn().type(argument(0))->is_int();
7644 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7645 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7646 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7647
7648 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7649 return false; // not enough info for intrinsification
7650 }
7651 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7652 if (!elem_type->is_primitive_type()) {
7653 return false; // should be primitive type
7654 }
7655 BasicType elem_bt = elem_type->basic_type();
7656 int num_elem = vlen->get_con();
7657 int opc = get_opc(opr->get_con(), elem_bt);
7658 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
7659 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7660 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7661
7662 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7663 return false; // not supported
7664 }
7665 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7666 Node* opd2 = shift_count(argument(5), opc, elem_bt, num_elem);
7667 if (opd1 == NULL || opd2 == NULL) {
7668 return false;
7669 }
7670 Node* operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, num_elem, elem_bt));
7671
7672 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
7673 set_vector_result(vbox);
7674 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7675 return true;
7676 }
7677
7678 static BasicType extract_bt_from_box_class(ciType* t) {
7679 if (t->basic_type() == T_OBJECT) {
7680 switch(t->as_klass()->name()->sid()) {
7681 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Boolean): return T_BOOLEAN;
7682 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Byte): return T_BYTE;
7683 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Short): return T_SHORT;
7684 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Character): return T_CHAR;
7685 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Integer): return T_INT;
7686 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Long): return T_LONG;
7687 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Float): return T_FLOAT;
7688 case vmSymbols::VM_SYMBOL_ENUM_NAME(java_lang_Double): return T_DOUBLE;
7689 default: fatal("unknown type: %s", t->as_klass()->name()->as_utf8());
7690 }
7691 }
7692 return T_ILLEGAL;
7693 }
7694
7695 bool LibraryCallKit::inline_vector_reinterpret() {
7696 const TypeInstPtr* vector_klass_from = gvn().type(argument(0))->is_instptr();
7697 const TypeInstPtr* elem_klass_from = gvn().type(argument(1))->is_instptr();
7698 const TypeInt* vlen_from = gvn().type(argument(2))->is_int();
7699
7700 const TypeInstPtr* elem_klass_to = gvn().type(argument(3))->is_instptr();
7701 const TypeInt* vlen_to = gvn().type(argument(4))->is_int();
7702
7703 if (vector_klass_from->const_oop() == NULL || elem_klass_from->const_oop() == NULL || !vlen_from->is_con() ||
7704 elem_klass_to->const_oop() == NULL || !vlen_to->is_con()) {
7705 return false; // not enough info for intrinsification
7706 }
7707
7708 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
7709 if (!vbox_klass_from->is_vectorapi_vector()) {
7710 return false; // only vector & mask are supported
7711 }
7712 bool is_mask = vbox_klass_from->is_vectormask();
7713
7714 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
7715 if (!elem_type_from->is_primitive_type()) {
7716 return false; // should be primitive type
7717 }
7718 BasicType elem_bt_from = elem_type_from->basic_type();
7719 if (is_mask) {
7720 elem_bt_from = getMaskBasicType(elem_bt_from);
7721 }
7722 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
7723 BasicType elem_bt_to = extract_bt_from_box_class(elem_type_to);
7724 if (elem_bt_to == T_ILLEGAL) {
7725 return false;
7726 }
7727 if (is_mask) {
7728 elem_bt_to = getMaskBasicType(elem_bt_to);
7729 }
7730 if (is_mask && elem_bt_from != elem_bt_to) {
7731 return false; // type mismatch
7732 }
7733 int num_elem_from = vlen_from->get_con();
7734 int num_elem_to = vlen_to->get_con();
7735 if (is_mask) {
7736 elem_bt_to = getMaskBasicType(elem_bt_to);
7737 }
7738
7739 // Check whether we can unbox to appropriate size.
7740 if (!arch_supports_vector(Op_VectorReinterpret,
7741 num_elem_from,
7742 elem_bt_from,
7743 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
7744 return false;
7745 }
7746
7747 // Check whether we can support resizing/reinterpreting to the new size.
7748 if (!arch_supports_vector(Op_VectorReinterpret,
7749 num_elem_to,
7750 elem_bt_to,
7751 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
7752 return false;
7753 }
7754
7755 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
7756
7757 Node* opd1 = unbox_vector(argument(5), vbox_type_from, elem_bt_from, num_elem_from);
7758 if (opd1 == NULL) {
7759 return false;
7760 }
7761
7762 // Can assert when Phi merges vectors of different types:
7763 // # Internal Error (/Users/vlivanov/ws/jdk/panama-dev/open/src/hotspot/share/opto/type.cpp:2291), pid=67536, tid=14083
7764 // # Error: assert(length() == v->length()) failed
7765 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from);
7766 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to);
7767 Node* op = opd1;
7768 if (src_type != dst_type) {
7769 op = _gvn.transform(new VectorReinterpretNode(op, src_type, dst_type));
7770 }
7771 ciKlass* vbox_klass_to = get_exact_klass_for_vector_box(vbox_klass_from, extract_bt_from_box_class(elem_type_to),
7772 num_elem_to, is_mask ? VECAPI_MASK : VECAPI_VECTOR);
7773 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
7774 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
7775 set_vector_result(vbox);
7776 return true;
7777 }
7778
7779 //------------------------------get_state_from_sha_object-----------------------
7780 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
7781 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
7782 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
7783 if (sha_state == NULL) return (Node *) NULL;
7784
7785 // now have the array, need to get the start address of the state array
7786 Node* state = array_element_address(sha_state, intcon(0), T_INT);
7787 return state;
7788 }
7789
7790 //------------------------------get_state_from_sha5_object-----------------------
7791 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
7792 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
7793 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
7794 if (sha_state == NULL) return (Node *) NULL;
7795
7796 // now have the array, need to get the start address of the state array
7797 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
7798 return state;
7799 }
7800
7801 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
7802 // Return node representing slow path of predicate check.
7803 // the pseudo code we want to emulate with this predicate is:
7804 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
7805 //
7806 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
7807 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
7808 "need SHA1/SHA256/SHA512 instruction support");
7809 assert((uint)predicate < 3, "sanity");
7810
7811 // The receiver was checked for NULL already.
7812 Node* digestBaseObj = argument(0);
7813
7814 // get DigestBase klass for instanceOf check
7815 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
7816 assert(tinst != NULL, "digestBaseObj is null");
7817 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
7818
7819 const char* klass_SHA_name = NULL;
7820 switch (predicate) {
7821 case 0:
7822 if (UseSHA1Intrinsics) {
7823 // we want to do an instanceof comparison against the SHA class
7824 klass_SHA_name = "sun/security/provider/SHA";
7825 }
7826 break;
7827 case 1:
7828 if (UseSHA256Intrinsics) {
7829 // we want to do an instanceof comparison against the SHA2 class
7830 klass_SHA_name = "sun/security/provider/SHA2";
7831 }
7832 break;
7833 case 2:
7834 if (UseSHA512Intrinsics) {
7835 // we want to do an instanceof comparison against the SHA5 class
7836 klass_SHA_name = "sun/security/provider/SHA5";
7837 }
7838 break;
7839 default:
7840 fatal("unknown SHA intrinsic predicate: %d", predicate);
7841 }
7842
7843 ciKlass* klass_SHA = NULL;
7844 if (klass_SHA_name != NULL) {
7845 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
7846 }
7847 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
7848 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
7849 Node* ctrl = control();
7850 set_control(top()); // no intrinsic path
7851 return ctrl;
7852 }
7853 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
7854
7855 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
7856 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
7857 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
7858 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
7859
7860 return instof_false; // even if it is NULL
7861 }
7862
7863 //-------------inline_fma-----------------------------------
7864 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
7865 Node *a = NULL;
7866 Node *b = NULL;
7867 Node *c = NULL;
7868 Node* result = NULL;
7869 switch (id) {
7870 case vmIntrinsics::_fmaD:
7871 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
7872 // no receiver since it is static method
7873 a = round_double_node(argument(0));
7874 b = round_double_node(argument(2));
7875 c = round_double_node(argument(4));
7876 result = _gvn.transform(new FmaDNode(control(), a, b, c));
7877 break;
7878 case vmIntrinsics::_fmaF:
7879 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
7880 a = argument(0);
7881 b = argument(1);
7882 c = argument(2);
7883 result = _gvn.transform(new FmaFNode(control(), a, b, c));
7884 break;
7885 default:
7886 fatal_unexpected_iid(id); break;
7887 }
7888 set_result(result);
7889 return true;
7890 }
7891
7892 bool LibraryCallKit::inline_profileBoolean() {
7893 Node* counts = argument(1);
7894 const TypeAryPtr* ary = NULL;
7895 ciArray* aobj = NULL;
7896 if (counts->is_Con()
7897 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
7898 && (aobj = ary->const_oop()->as_array()) != NULL
7899 && (aobj->length() == 2)) {
7900 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
7901 jint false_cnt = aobj->element_value(0).as_int();
7902 jint true_cnt = aobj->element_value(1).as_int();
7903
7904 if (C->log() != NULL) {
7905 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
7906 false_cnt, true_cnt);
7907 }
7908
7909 if (false_cnt + true_cnt == 0) {
7910 // According to profile, never executed.
7911 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7912 Deoptimization::Action_reinterpret);
7913 return true;
7914 }
7915
7916 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
7917 // is a number of each value occurrences.
7918 Node* result = argument(0);
7919 if (false_cnt == 0 || true_cnt == 0) {
7920 // According to profile, one value has been never seen.
7921 int expected_val = (false_cnt == 0) ? 1 : 0;
7922
7923 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
7924 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
7925
7926 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
7927 Node* fast_path = _gvn.transform(new IfTrueNode(check));
7928 Node* slow_path = _gvn.transform(new IfFalseNode(check));
7929
7930 { // Slow path: uncommon trap for never seen value and then reexecute
7931 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
7932 // the value has been seen at least once.
7933 PreserveJVMState pjvms(this);
7934 PreserveReexecuteState preexecs(this);
7935 jvms()->set_should_reexecute(true);
7936
7937 set_control(slow_path);
7938 set_i_o(i_o());
7939
7940 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7941 Deoptimization::Action_reinterpret);
7942 }
7943 // The guard for never seen value enables sharpening of the result and
7944 // returning a constant. It allows to eliminate branches on the same value
7945 // later on.
7946 set_control(fast_path);
7947 result = intcon(expected_val);
7948 }
7949 // Stop profiling.
7950 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
7951 // By replacing method body with profile data (represented as ProfileBooleanNode
7952 // on IR level) we effectively disable profiling.
7953 // It enables full speed execution once optimized code is generated.
7954 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
7955 C->record_for_igvn(profile);
7956 set_result(profile);
7957 return true;
7958 } else {
7959 // Continue profiling.
7960 // Profile data isn't available at the moment. So, execute method's bytecode version.
7961 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
7962 // is compiled and counters aren't available since corresponding MethodHandle
7963 // isn't a compile-time constant.
7964 return false;
7965 }
7966 }
7967
7968 bool LibraryCallKit::inline_isCompileConstant() {
7969 Node* n = argument(0);
7970 set_result(n->is_Con() ? intcon(1) : intcon(0));
7971 return true;
7972 }