1 /*
2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "ci/ciUtilities.inline.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "classfile/vmSymbols.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/c2compiler.hpp"
37 #include "opto/callGenerator.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/convertnode.hpp"
41 #include "opto/countbitsnode.hpp"
42 #include "opto/intrinsicnode.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/mathexactnode.hpp"
45 #include "opto/movenode.hpp"
46 #include "opto/mulnode.hpp"
47 #include "opto/narrowptrnode.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/parse.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/rootnode.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/vectornode.hpp"
54 #include "prims/nativeLookup.hpp"
55 #include "prims/unsafe.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #ifdef TRACE_HAVE_INTRINSICS
58 #include "trace/traceMacros.hpp"
59 #endif
60
61 class LibraryIntrinsic : public InlineCallGenerator {
62 // Extend the set of intrinsics known to the runtime:
63 public:
64 private:
65 bool _is_virtual;
66 bool _does_virtual_dispatch;
67 int8_t _predicates_count; // Intrinsic is predicated by several conditions
68 int8_t _last_predicate; // Last generated predicate
69 vmIntrinsics::ID _intrinsic_id;
70
71 public:
72 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
73 : InlineCallGenerator(m),
74 _is_virtual(is_virtual),
75 _does_virtual_dispatch(does_virtual_dispatch),
76 _predicates_count((int8_t)predicates_count),
77 _last_predicate((int8_t)-1),
78 _intrinsic_id(id)
79 {
80 }
81 virtual bool is_intrinsic() const { return true; }
82 virtual bool is_virtual() const { return _is_virtual; }
83 virtual bool is_predicated() const { return _predicates_count > 0; }
84 virtual int predicates_count() const { return _predicates_count; }
85 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
86 virtual JVMState* generate(JVMState* jvms);
87 virtual Node* generate_predicate(JVMState* jvms, int predicate);
88 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
89 };
90
91
92 // Local helper class for LibraryIntrinsic:
93 class LibraryCallKit : public GraphKit {
94 private:
95 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
96 Node* _result; // the result node, if any
97 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
98
99 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
100
101 public:
102 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
103 : GraphKit(jvms),
104 _intrinsic(intrinsic),
105 _result(NULL)
106 {
107 // Check if this is a root compile. In that case we don't have a caller.
108 if (!jvms->has_method()) {
109 _reexecute_sp = sp();
110 } else {
111 // Find out how many arguments the interpreter needs when deoptimizing
112 // and save the stack pointer value so it can used by uncommon_trap.
113 // We find the argument count by looking at the declared signature.
114 bool ignored_will_link;
115 ciSignature* declared_signature = NULL;
116 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
117 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
118 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
119 }
120 }
121
122 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
123
124 ciMethod* caller() const { return jvms()->method(); }
125 int bci() const { return jvms()->bci(); }
126 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
127 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
128 ciMethod* callee() const { return _intrinsic->method(); }
129
130 bool try_to_inline(int predicate);
131 Node* try_to_predicate(int predicate);
132
133 void push_result() {
134 // Push the result onto the stack.
135 if (!stopped() && result() != NULL) {
136 BasicType bt = result()->bottom_type()->basic_type();
137 push_node(bt, result());
138 }
139 }
140
141 private:
142 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
143 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
144 }
145
146 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
147 void set_result(RegionNode* region, PhiNode* value);
148 Node* result() { return _result; }
149
150 virtual int reexecute_sp() { return _reexecute_sp; }
151
152 // Helper functions to inline natives
153 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
154 Node* generate_slow_guard(Node* test, RegionNode* region);
155 Node* generate_fair_guard(Node* test, RegionNode* region);
156 Node* generate_negative_guard(Node* index, RegionNode* region,
157 // resulting CastII of index:
158 Node* *pos_index = NULL);
159 Node* generate_limit_guard(Node* offset, Node* subseq_length,
160 Node* array_length,
161 RegionNode* region);
162 void generate_string_range_check(Node* array, Node* offset,
163 Node* length, bool char_count);
164 Node* generate_current_thread(Node* &tls_output);
165 Node* load_mirror_from_klass(Node* klass);
166 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
167 RegionNode* region, int null_path,
168 int offset);
169 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
170 RegionNode* region, int null_path) {
171 int offset = java_lang_Class::klass_offset_in_bytes();
172 return load_klass_from_mirror_common(mirror, never_see_null,
173 region, null_path,
174 offset);
175 }
176 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
177 RegionNode* region, int null_path) {
178 int offset = java_lang_Class::array_klass_offset_in_bytes();
179 return load_klass_from_mirror_common(mirror, never_see_null,
180 region, null_path,
181 offset);
182 }
183 Node* generate_access_flags_guard(Node* kls,
184 int modifier_mask, int modifier_bits,
185 RegionNode* region);
186 Node* generate_interface_guard(Node* kls, RegionNode* region);
187 Node* generate_array_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, false, false);
189 }
190 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
191 return generate_array_guard_common(kls, region, false, true);
192 }
193 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
194 return generate_array_guard_common(kls, region, true, false);
195 }
196 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
197 return generate_array_guard_common(kls, region, true, true);
198 }
199 Node* generate_array_guard_common(Node* kls, RegionNode* region,
200 bool obj_array, bool not_array);
201 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
202 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
203 bool is_virtual = false, bool is_static = false);
204 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
205 return generate_method_call(method_id, false, true);
206 }
207 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
208 return generate_method_call(method_id, true, false);
209 }
210 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
211 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
212
213 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
215 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
216 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
217 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
218 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
219 bool inline_string_indexOfChar();
220 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
221 bool inline_string_toBytesU();
222 bool inline_string_getCharsU();
223 bool inline_string_copy(bool compress);
224 bool inline_string_char_access(bool is_store);
225 Node* round_double_node(Node* n);
226 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
227 bool inline_math_native(vmIntrinsics::ID id);
228 bool inline_math(vmIntrinsics::ID id);
229 template <typename OverflowOp>
230 bool inline_math_overflow(Node* arg1, Node* arg2);
231 void inline_math_mathExact(Node* math, Node* test);
232 bool inline_math_addExactI(bool is_increment);
233 bool inline_math_addExactL(bool is_increment);
234 bool inline_math_multiplyExactI();
235 bool inline_math_multiplyExactL();
236 bool inline_math_multiplyHigh();
237 bool inline_math_negateExactI();
238 bool inline_math_negateExactL();
239 bool inline_math_subtractExactI(bool is_decrement);
240 bool inline_math_subtractExactL(bool is_decrement);
241 bool inline_min_max(vmIntrinsics::ID id);
242 bool inline_notify(vmIntrinsics::ID id);
243 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
244 // This returns Type::AnyPtr, RawPtr, or OopPtr.
245 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
246 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
247 // Helper for inline_unsafe_access.
248 // Generates the guards that check whether the result of
249 // Unsafe.getObject should be recorded in an SATB log buffer.
250 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
251
252 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
253 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
254 static bool klass_needs_init_guard(Node* kls);
255 bool inline_unsafe_allocate();
256 bool inline_unsafe_newArray(bool uninitialized);
257 bool inline_unsafe_copyMemory();
258 bool inline_native_currentThread();
259
260 bool inline_native_time_funcs(address method, const char* funcName);
261 #ifdef TRACE_HAVE_INTRINSICS
262 bool inline_native_classID();
263 bool inline_native_getBufferWriter();
264 #endif
265 bool inline_native_isInterrupted();
266 bool inline_native_Class_query(vmIntrinsics::ID id);
267 bool inline_native_subtype_check();
268 bool inline_native_getLength();
269 bool inline_array_copyOf(bool is_copyOfRange);
270 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
271 bool inline_preconditions_checkIndex();
272 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
273 bool inline_native_clone(bool is_virtual);
274 bool inline_native_Reflection_getCallerClass();
275 // Helper function for inlining native object hash method
276 bool inline_native_hashcode(bool is_virtual, bool is_static);
277 bool inline_native_getClass();
278
279 // Helper functions for inlining arraycopy
280 bool inline_arraycopy();
281 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
282 RegionNode* slow_region);
283 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
284 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
285 uint new_idx);
286
287 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
288 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
289 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
290 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
291 bool inline_unsafe_fence(vmIntrinsics::ID id);
292 bool inline_onspinwait();
293 bool inline_fp_conversions(vmIntrinsics::ID id);
294 bool inline_number_methods(vmIntrinsics::ID id);
295 bool inline_reference_get();
296 bool inline_Class_cast();
297 bool inline_aescrypt_Block(vmIntrinsics::ID id);
298 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
299 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
300 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
301 Node* inline_counterMode_AESCrypt_predicate();
302 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
303 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
304 bool inline_ghash_processBlocks();
305 bool inline_sha_implCompress(vmIntrinsics::ID id);
306 bool inline_digestBase_implCompressMB(int predicate);
307 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
308 bool long_state, address stubAddr, const char *stubName,
309 Node* src_start, Node* ofs, Node* limit);
310 Node* get_state_from_sha_object(Node *sha_object);
311 Node* get_state_from_sha5_object(Node *sha_object);
312 Node* inline_digestBase_implCompressMB_predicate(int predicate);
313 bool inline_encodeISOArray();
314 bool inline_updateCRC32();
315 bool inline_updateBytesCRC32();
316 bool inline_updateByteBufferCRC32();
317 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
318 bool inline_updateBytesCRC32C();
319 bool inline_updateDirectByteBufferCRC32C();
320 bool inline_updateBytesAdler32();
321 bool inline_updateByteBufferAdler32();
322 bool inline_multiplyToLen();
323 bool inline_hasNegatives();
324 bool inline_squareToLen();
325 bool inline_mulAdd();
326 bool inline_montgomeryMultiply();
327 bool inline_montgomerySquare();
328 bool inline_vectorizedMismatch();
329 bool inline_fma(vmIntrinsics::ID id);
330
331 bool inline_profileBoolean();
332 bool inline_isCompileConstant();
333
334 // Vector API support
335 bool inline_vector_nary_operation(int n);
336 bool inline_vector_broadcast_coerced();
337 bool inline_vector_mem_operation(bool is_store);
338 bool inline_vector_reduction();
339 bool inline_vector_test();
340 bool inline_vector_blend();
341 bool inline_vector_compare();
342 bool inline_vector_broadcast_int();
343 bool inline_vector_reinterpret();
344 Node* box_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
345 Node* unbox_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
346 Node* shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem);
347 void set_vector_result(Node* result, bool set_res = true);
348
349 void clear_upper_avx() {
350 #ifdef X86
351 if (UseAVX >= 2) {
352 C->set_clear_upper_avx(true);
353 }
354 #endif
355 }
356 };
357
358 //---------------------------make_vm_intrinsic----------------------------
359 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
360 vmIntrinsics::ID id = m->intrinsic_id();
361 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
362
363 if (!m->is_loaded()) {
364 // Do not attempt to inline unloaded methods.
365 return NULL;
366 }
367
368 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
369 bool is_available = false;
370
371 {
372 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
373 // the compiler must transition to '_thread_in_vm' state because both
374 // methods access VM-internal data.
375 VM_ENTRY_MARK;
376 methodHandle mh(THREAD, m->get_Method());
377 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
378 !C->directive()->is_intrinsic_disabled(mh) &&
379 !vmIntrinsics::is_disabled_by_flags(mh);
380
381 }
382
383 if (is_available) {
384 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
385 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
386 return new LibraryIntrinsic(m, is_virtual,
387 vmIntrinsics::predicates_needed(id),
388 vmIntrinsics::does_virtual_dispatch(id),
389 (vmIntrinsics::ID) id);
390 } else {
391 return NULL;
392 }
393 }
394
395 //----------------------register_library_intrinsics-----------------------
396 // Initialize this file's data structures, for each Compile instance.
397 void Compile::register_library_intrinsics() {
398 // Nothing to do here.
399 }
400
401 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
402 LibraryCallKit kit(jvms, this);
403 Compile* C = kit.C;
404 int nodes = C->unique();
405 #ifndef PRODUCT
406 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
407 char buf[1000];
408 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
409 tty->print_cr("Intrinsic %s", str);
410 }
411 #endif
412 ciMethod* callee = kit.callee();
413 const int bci = kit.bci();
414
415 // Try to inline the intrinsic.
416 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
417 kit.try_to_inline(_last_predicate)) {
418 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
419 : "(intrinsic)";
420 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
421 if (C->print_intrinsics() || C->print_inlining()) {
422 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
423 }
424 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
425 if (C->log()) {
426 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
427 vmIntrinsics::name_at(intrinsic_id()),
428 (is_virtual() ? " virtual='1'" : ""),
429 C->unique() - nodes);
430 }
431 // Push the result from the inlined method onto the stack.
432 kit.push_result();
433 C->print_inlining_update(this);
434 return kit.transfer_exceptions_into_jvms();
435 }
436
437 // The intrinsic bailed out
438 if (jvms->has_method()) {
439 // Not a root compile.
440 const char* msg;
441 if (callee->intrinsic_candidate()) {
442 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
443 } else {
444 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
445 : "failed to inline (intrinsic), method not annotated";
446 }
447 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
448 if (C->print_intrinsics() || C->print_inlining()) {
449 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
450 }
451 } else {
452 // Root compile
453 ResourceMark rm;
454 stringStream msg_stream;
455 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
456 vmIntrinsics::name_at(intrinsic_id()),
457 is_virtual() ? " (virtual)" : "", bci);
458 const char *msg = msg_stream.as_string();
459 log_debug(jit, inlining)("%s", msg);
460 if (C->print_intrinsics() || C->print_inlining()) {
461 tty->print("%s", msg);
462 }
463 }
464 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
465 C->print_inlining_update(this);
466 return NULL;
467 }
468
469 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
470 LibraryCallKit kit(jvms, this);
471 Compile* C = kit.C;
472 int nodes = C->unique();
473 _last_predicate = predicate;
474 #ifndef PRODUCT
475 assert(is_predicated() && predicate < predicates_count(), "sanity");
476 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
477 char buf[1000];
478 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
479 tty->print_cr("Predicate for intrinsic %s", str);
480 }
481 #endif
482 ciMethod* callee = kit.callee();
483 const int bci = kit.bci();
484
485 Node* slow_ctl = kit.try_to_predicate(predicate);
486 if (!kit.failing()) {
487 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
488 : "(intrinsic, predicate)";
489 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
490 if (C->print_intrinsics() || C->print_inlining()) {
491 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
492 }
493 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
494 if (C->log()) {
495 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
496 vmIntrinsics::name_at(intrinsic_id()),
497 (is_virtual() ? " virtual='1'" : ""),
498 C->unique() - nodes);
499 }
500 return slow_ctl; // Could be NULL if the check folds.
501 }
502
503 // The intrinsic bailed out
504 if (jvms->has_method()) {
505 // Not a root compile.
506 const char* msg = "failed to generate predicate for intrinsic";
507 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
508 if (C->print_intrinsics() || C->print_inlining()) {
509 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
510 }
511 } else {
512 // Root compile
513 ResourceMark rm;
514 stringStream msg_stream;
515 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
516 vmIntrinsics::name_at(intrinsic_id()),
517 is_virtual() ? " (virtual)" : "", bci);
518 const char *msg = msg_stream.as_string();
519 log_debug(jit, inlining)("%s", msg);
520 if (C->print_intrinsics() || C->print_inlining()) {
521 C->print_inlining_stream()->print("%s", msg);
522 }
523 }
524 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
525 return NULL;
526 }
527
528 bool LibraryCallKit::try_to_inline(int predicate) {
529 // Handle symbolic names for otherwise undistinguished boolean switches:
530 const bool is_store = true;
531 const bool is_compress = true;
532 const bool is_static = true;
533 const bool is_volatile = true;
534
535 if (!jvms()->has_method()) {
536 // Root JVMState has a null method.
537 assert(map()->memory()->Opcode() == Op_Parm, "");
538 // Insert the memory aliasing node
539 set_all_memory(reset_memory());
540 }
541 assert(merged_memory(), "");
542
543 switch (intrinsic_id()) {
544 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
545 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
546 case vmIntrinsics::_getClass: return inline_native_getClass();
547
548 case vmIntrinsics::_dsin:
549 case vmIntrinsics::_dcos:
550 case vmIntrinsics::_dtan:
551 case vmIntrinsics::_dabs:
552 case vmIntrinsics::_datan2:
553 case vmIntrinsics::_dsqrt:
554 case vmIntrinsics::_dexp:
555 case vmIntrinsics::_dlog:
556 case vmIntrinsics::_dlog10:
557 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
558
559 case vmIntrinsics::_min:
560 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
561
562 case vmIntrinsics::_notify:
563 case vmIntrinsics::_notifyAll:
564 if (InlineNotify) {
565 return inline_notify(intrinsic_id());
566 }
567 return false;
568
569 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
570 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
571 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
572 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
573 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
574 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
575 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
576 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
577 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
578 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
579 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
580 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
581 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
582
583 case vmIntrinsics::_arraycopy: return inline_arraycopy();
584
585 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
586 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
587 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
588 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
589
590 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
591 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
592 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
593 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
594 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
595 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
596 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
597
598 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
599 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
600
601 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
602 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
603 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
604 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
605
606 case vmIntrinsics::_compressStringC:
607 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
608 case vmIntrinsics::_inflateStringC:
609 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
610
611 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
612 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
613 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
614 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
615 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
616 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
617 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
618 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
619 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
620
621 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
622 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
623 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
624 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
625 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
626 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
627 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
628 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
629 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
630
631 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
632 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
633 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
634 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
635 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
636 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
637 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
638 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
639 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
640
641 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
642 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
643 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
644 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
645 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
646 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
647 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
648 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
649 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
650
651 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
652 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
653 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
654 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
655
656 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
657 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
658 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
659 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
660
661 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
662 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
663 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
664 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
665 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
666 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
667 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
668 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
669 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
670
671 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
672 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
673 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
674 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
675 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
676 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
677 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
678 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
679 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
680
681 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
682 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
683 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
684 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
685 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
686 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
687 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
688 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
689 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
690
691 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
692 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
693 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
694 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
695 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
696 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
697 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
698 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
699 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
700
701 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
702 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
703 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
704 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
705 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
706
707 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
708 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
709 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
710 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
711 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
712 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
713 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
714 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
715 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
716 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
717 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
718 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
719 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
720 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
721 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
722 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
723 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
724 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
725 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
726 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
727
728 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
729 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
730 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
731 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
732 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
733 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
734 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
735 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
736 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
737 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
738 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
739 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
740 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
741 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
742 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
743
744 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
745 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
746 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
747 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
748
749 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
750 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
751 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
752 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
753 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
754
755 case vmIntrinsics::_loadFence:
756 case vmIntrinsics::_storeFence:
757 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
758
759 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
760
761 case vmIntrinsics::_currentThread: return inline_native_currentThread();
762 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
763
764 #ifdef TRACE_HAVE_INTRINSICS
765 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
766 case vmIntrinsics::_getClassId: return inline_native_classID();
767 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
768 #endif
769 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
770 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
771 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
772 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
773 case vmIntrinsics::_getLength: return inline_native_getLength();
774 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
775 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
776 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
777 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
778 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
779 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
780
781 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
782 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
783
784 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
785
786 case vmIntrinsics::_isInstance:
787 case vmIntrinsics::_getModifiers:
788 case vmIntrinsics::_isInterface:
789 case vmIntrinsics::_isArray:
790 case vmIntrinsics::_isPrimitive:
791 case vmIntrinsics::_getSuperclass:
792 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
793
794 case vmIntrinsics::_floatToRawIntBits:
795 case vmIntrinsics::_floatToIntBits:
796 case vmIntrinsics::_intBitsToFloat:
797 case vmIntrinsics::_doubleToRawLongBits:
798 case vmIntrinsics::_doubleToLongBits:
799 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
800
801 case vmIntrinsics::_numberOfLeadingZeros_i:
802 case vmIntrinsics::_numberOfLeadingZeros_l:
803 case vmIntrinsics::_numberOfTrailingZeros_i:
804 case vmIntrinsics::_numberOfTrailingZeros_l:
805 case vmIntrinsics::_bitCount_i:
806 case vmIntrinsics::_bitCount_l:
807 case vmIntrinsics::_reverseBytes_i:
808 case vmIntrinsics::_reverseBytes_l:
809 case vmIntrinsics::_reverseBytes_s:
810 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
811
812 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
813
814 case vmIntrinsics::_Reference_get: return inline_reference_get();
815
816 case vmIntrinsics::_Class_cast: return inline_Class_cast();
817
818 case vmIntrinsics::_aescrypt_encryptBlock:
819 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
820
821 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
822 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
823 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
824
825 case vmIntrinsics::_counterMode_AESCrypt:
826 return inline_counterMode_AESCrypt(intrinsic_id());
827
828 case vmIntrinsics::_sha_implCompress:
829 case vmIntrinsics::_sha2_implCompress:
830 case vmIntrinsics::_sha5_implCompress:
831 return inline_sha_implCompress(intrinsic_id());
832
833 case vmIntrinsics::_digestBase_implCompressMB:
834 return inline_digestBase_implCompressMB(predicate);
835
836 case vmIntrinsics::_multiplyToLen:
837 return inline_multiplyToLen();
838
839 case vmIntrinsics::_squareToLen:
840 return inline_squareToLen();
841
842 case vmIntrinsics::_mulAdd:
843 return inline_mulAdd();
844
845 case vmIntrinsics::_montgomeryMultiply:
846 return inline_montgomeryMultiply();
847 case vmIntrinsics::_montgomerySquare:
848 return inline_montgomerySquare();
849
850 case vmIntrinsics::_vectorizedMismatch:
851 return inline_vectorizedMismatch();
852
853 case vmIntrinsics::_ghash_processBlocks:
854 return inline_ghash_processBlocks();
855
856 case vmIntrinsics::_encodeISOArray:
857 case vmIntrinsics::_encodeByteISOArray:
858 return inline_encodeISOArray();
859
860 case vmIntrinsics::_updateCRC32:
861 return inline_updateCRC32();
862 case vmIntrinsics::_updateBytesCRC32:
863 return inline_updateBytesCRC32();
864 case vmIntrinsics::_updateByteBufferCRC32:
865 return inline_updateByteBufferCRC32();
866
867 case vmIntrinsics::_updateBytesCRC32C:
868 return inline_updateBytesCRC32C();
869 case vmIntrinsics::_updateDirectByteBufferCRC32C:
870 return inline_updateDirectByteBufferCRC32C();
871
872 case vmIntrinsics::_updateBytesAdler32:
873 return inline_updateBytesAdler32();
874 case vmIntrinsics::_updateByteBufferAdler32:
875 return inline_updateByteBufferAdler32();
876
877 case vmIntrinsics::_profileBoolean:
878 return inline_profileBoolean();
879 case vmIntrinsics::_isCompileConstant:
880 return inline_isCompileConstant();
881
882 case vmIntrinsics::_hasNegatives:
883 return inline_hasNegatives();
884
885 case vmIntrinsics::_fmaD:
886 case vmIntrinsics::_fmaF:
887 return inline_fma(intrinsic_id());
888
889 case vmIntrinsics::_VectorUnaryOp:
890 return inline_vector_nary_operation(1);
891 case vmIntrinsics::_VectorBinaryOp:
892 return inline_vector_nary_operation(2);
893 case vmIntrinsics::_VectorTernaryOp:
894 return inline_vector_nary_operation(3);
895
896 case vmIntrinsics::_VectorBroadcastCoerced:
897 return inline_vector_broadcast_coerced();
898 case vmIntrinsics::_VectorLoadOp:
899 return inline_vector_mem_operation(/*is_store=*/false);
900 case vmIntrinsics::_VectorStoreOp:
901 return inline_vector_mem_operation(/*is_store=*/true);
902 case vmIntrinsics::_VectorReductionCoerced:
903 return inline_vector_reduction();
904 case vmIntrinsics::_VectorTest:
905 return inline_vector_test();
906 case vmIntrinsics::_VectorBlend:
907 return inline_vector_blend();
908 case vmIntrinsics::_VectorCompare:
909 return inline_vector_compare();
910 case vmIntrinsics::_VectorBroadcastInt:
911 return inline_vector_broadcast_int();
912 case vmIntrinsics::_VectorReinterpret:
913 return inline_vector_reinterpret();
914
915 default:
916 // If you get here, it may be that someone has added a new intrinsic
917 // to the list in vmSymbols.hpp without implementing it here.
918 #ifndef PRODUCT
919 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
920 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
921 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
922 }
923 #endif
924 return false;
925 }
926 }
927
928 Node* LibraryCallKit::try_to_predicate(int predicate) {
929 if (!jvms()->has_method()) {
930 // Root JVMState has a null method.
931 assert(map()->memory()->Opcode() == Op_Parm, "");
932 // Insert the memory aliasing node
933 set_all_memory(reset_memory());
934 }
935 assert(merged_memory(), "");
936
937 switch (intrinsic_id()) {
938 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
939 return inline_cipherBlockChaining_AESCrypt_predicate(false);
940 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
941 return inline_cipherBlockChaining_AESCrypt_predicate(true);
942 case vmIntrinsics::_counterMode_AESCrypt:
943 return inline_counterMode_AESCrypt_predicate();
944 case vmIntrinsics::_digestBase_implCompressMB:
945 return inline_digestBase_implCompressMB_predicate(predicate);
946
947 default:
948 // If you get here, it may be that someone has added a new intrinsic
949 // to the list in vmSymbols.hpp without implementing it here.
950 #ifndef PRODUCT
951 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
952 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
953 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
954 }
955 #endif
956 Node* slow_ctl = control();
957 set_control(top()); // No fast path instrinsic
958 return slow_ctl;
959 }
960 }
961
962 //------------------------------set_result-------------------------------
963 // Helper function for finishing intrinsics.
964 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
965 record_for_igvn(region);
966 set_control(_gvn.transform(region));
967 set_result( _gvn.transform(value));
968 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
969 }
970
971 //------------------------------generate_guard---------------------------
972 // Helper function for generating guarded fast-slow graph structures.
973 // The given 'test', if true, guards a slow path. If the test fails
974 // then a fast path can be taken. (We generally hope it fails.)
975 // In all cases, GraphKit::control() is updated to the fast path.
976 // The returned value represents the control for the slow path.
977 // The return value is never 'top'; it is either a valid control
978 // or NULL if it is obvious that the slow path can never be taken.
979 // Also, if region and the slow control are not NULL, the slow edge
980 // is appended to the region.
981 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
982 if (stopped()) {
983 // Already short circuited.
984 return NULL;
985 }
986
987 // Build an if node and its projections.
988 // If test is true we take the slow path, which we assume is uncommon.
989 if (_gvn.type(test) == TypeInt::ZERO) {
990 // The slow branch is never taken. No need to build this guard.
991 return NULL;
992 }
993
994 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
995
996 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
997 if (if_slow == top()) {
998 // The slow branch is never taken. No need to build this guard.
999 return NULL;
1000 }
1001
1002 if (region != NULL)
1003 region->add_req(if_slow);
1004
1005 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1006 set_control(if_fast);
1007
1008 return if_slow;
1009 }
1010
1011 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1012 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1013 }
1014 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1015 return generate_guard(test, region, PROB_FAIR);
1016 }
1017
1018 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1019 Node* *pos_index) {
1020 if (stopped())
1021 return NULL; // already stopped
1022 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1023 return NULL; // index is already adequately typed
1024 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1025 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1026 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1027 if (is_neg != NULL && pos_index != NULL) {
1028 // Emulate effect of Parse::adjust_map_after_if.
1029 Node* ccast = new CastIINode(index, TypeInt::POS);
1030 ccast->set_req(0, control());
1031 (*pos_index) = _gvn.transform(ccast);
1032 }
1033 return is_neg;
1034 }
1035
1036 // Make sure that 'position' is a valid limit index, in [0..length].
1037 // There are two equivalent plans for checking this:
1038 // A. (offset + copyLength) unsigned<= arrayLength
1039 // B. offset <= (arrayLength - copyLength)
1040 // We require that all of the values above, except for the sum and
1041 // difference, are already known to be non-negative.
1042 // Plan A is robust in the face of overflow, if offset and copyLength
1043 // are both hugely positive.
1044 //
1045 // Plan B is less direct and intuitive, but it does not overflow at
1046 // all, since the difference of two non-negatives is always
1047 // representable. Whenever Java methods must perform the equivalent
1048 // check they generally use Plan B instead of Plan A.
1049 // For the moment we use Plan A.
1050 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1051 Node* subseq_length,
1052 Node* array_length,
1053 RegionNode* region) {
1054 if (stopped())
1055 return NULL; // already stopped
1056 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1057 if (zero_offset && subseq_length->eqv_uncast(array_length))
1058 return NULL; // common case of whole-array copy
1059 Node* last = subseq_length;
1060 if (!zero_offset) // last += offset
1061 last = _gvn.transform(new AddINode(last, offset));
1062 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1063 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1064 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1065 return is_over;
1066 }
1067
1068 // Emit range checks for the given String.value byte array
1069 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1070 if (stopped()) {
1071 return; // already stopped
1072 }
1073 RegionNode* bailout = new RegionNode(1);
1074 record_for_igvn(bailout);
1075 if (char_count) {
1076 // Convert char count to byte count
1077 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1078 }
1079
1080 // Offset and count must not be negative
1081 generate_negative_guard(offset, bailout);
1082 generate_negative_guard(count, bailout);
1083 // Offset + count must not exceed length of array
1084 generate_limit_guard(offset, count, load_array_length(array), bailout);
1085
1086 if (bailout->req() > 1) {
1087 PreserveJVMState pjvms(this);
1088 set_control(_gvn.transform(bailout));
1089 uncommon_trap(Deoptimization::Reason_intrinsic,
1090 Deoptimization::Action_maybe_recompile);
1091 }
1092 }
1093
1094 //--------------------------generate_current_thread--------------------
1095 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1096 ciKlass* thread_klass = env()->Thread_klass();
1097 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1098 Node* thread = _gvn.transform(new ThreadLocalNode());
1099 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1100 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1101 tls_output = thread;
1102 return threadObj;
1103 }
1104
1105
1106 //------------------------------make_string_method_node------------------------
1107 // Helper method for String intrinsic functions. This version is called with
1108 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1109 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1110 // containing the lengths of str1 and str2.
1111 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1112 Node* result = NULL;
1113 switch (opcode) {
1114 case Op_StrIndexOf:
1115 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1116 str1_start, cnt1, str2_start, cnt2, ae);
1117 break;
1118 case Op_StrComp:
1119 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1120 str1_start, cnt1, str2_start, cnt2, ae);
1121 break;
1122 case Op_StrEquals:
1123 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1124 // Use the constant length if there is one because optimized match rule may exist.
1125 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1126 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1127 break;
1128 default:
1129 ShouldNotReachHere();
1130 return NULL;
1131 }
1132
1133 // All these intrinsics have checks.
1134 C->set_has_split_ifs(true); // Has chance for split-if optimization
1135 clear_upper_avx();
1136
1137 return _gvn.transform(result);
1138 }
1139
1140 //------------------------------inline_string_compareTo------------------------
1141 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1142 Node* arg1 = argument(0);
1143 Node* arg2 = argument(1);
1144
1145 // Get start addr and length of first argument
1146 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1147 Node* arg1_cnt = load_array_length(arg1);
1148
1149 // Get start addr and length of second argument
1150 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1151 Node* arg2_cnt = load_array_length(arg2);
1152
1153 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1154 set_result(result);
1155 return true;
1156 }
1157
1158 //------------------------------inline_string_equals------------------------
1159 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1160 Node* arg1 = argument(0);
1161 Node* arg2 = argument(1);
1162
1163 // paths (plus control) merge
1164 RegionNode* region = new RegionNode(3);
1165 Node* phi = new PhiNode(region, TypeInt::BOOL);
1166
1167 if (!stopped()) {
1168 // Get start addr and length of first argument
1169 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1170 Node* arg1_cnt = load_array_length(arg1);
1171
1172 // Get start addr and length of second argument
1173 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1174 Node* arg2_cnt = load_array_length(arg2);
1175
1176 // Check for arg1_cnt != arg2_cnt
1177 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1178 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1179 Node* if_ne = generate_slow_guard(bol, NULL);
1180 if (if_ne != NULL) {
1181 phi->init_req(2, intcon(0));
1182 region->init_req(2, if_ne);
1183 }
1184
1185 // Check for count == 0 is done by assembler code for StrEquals.
1186
1187 if (!stopped()) {
1188 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1189 phi->init_req(1, equals);
1190 region->init_req(1, control());
1191 }
1192 }
1193
1194 // post merge
1195 set_control(_gvn.transform(region));
1196 record_for_igvn(region);
1197
1198 set_result(_gvn.transform(phi));
1199 return true;
1200 }
1201
1202 //------------------------------inline_array_equals----------------------------
1203 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1204 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1205 Node* arg1 = argument(0);
1206 Node* arg2 = argument(1);
1207
1208 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1209 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1210 clear_upper_avx();
1211
1212 return true;
1213 }
1214
1215 //------------------------------inline_hasNegatives------------------------------
1216 bool LibraryCallKit::inline_hasNegatives() {
1217 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1218 return false;
1219 }
1220
1221 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1222 // no receiver since it is static method
1223 Node* ba = argument(0);
1224 Node* offset = argument(1);
1225 Node* len = argument(2);
1226
1227 // Range checks
1228 generate_string_range_check(ba, offset, len, false);
1229 if (stopped()) {
1230 return true;
1231 }
1232 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1233 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1234 set_result(_gvn.transform(result));
1235 return true;
1236 }
1237
1238 bool LibraryCallKit::inline_preconditions_checkIndex() {
1239 Node* index = argument(0);
1240 Node* length = argument(1);
1241 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1242 return false;
1243 }
1244
1245 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1246 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1247
1248 {
1249 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1250 uncommon_trap(Deoptimization::Reason_intrinsic,
1251 Deoptimization::Action_make_not_entrant);
1252 }
1253
1254 if (stopped()) {
1255 return false;
1256 }
1257
1258 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1259 BoolTest::mask btest = BoolTest::lt;
1260 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1261 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1262 _gvn.set_type(rc, rc->Value(&_gvn));
1263 if (!rc_bool->is_Con()) {
1264 record_for_igvn(rc);
1265 }
1266 set_control(_gvn.transform(new IfTrueNode(rc)));
1267 {
1268 PreserveJVMState pjvms(this);
1269 set_control(_gvn.transform(new IfFalseNode(rc)));
1270 uncommon_trap(Deoptimization::Reason_range_check,
1271 Deoptimization::Action_make_not_entrant);
1272 }
1273
1274 if (stopped()) {
1275 return false;
1276 }
1277
1278 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1279 result->set_req(0, control());
1280 result = _gvn.transform(result);
1281 set_result(result);
1282 replace_in_map(index, result);
1283 clear_upper_avx();
1284 return true;
1285 }
1286
1287 //------------------------------inline_string_indexOf------------------------
1288 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1289 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1290 return false;
1291 }
1292 Node* src = argument(0);
1293 Node* tgt = argument(1);
1294
1295 // Make the merge point
1296 RegionNode* result_rgn = new RegionNode(4);
1297 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1298
1299 // Get start addr and length of source string
1300 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1301 Node* src_count = load_array_length(src);
1302
1303 // Get start addr and length of substring
1304 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1305 Node* tgt_count = load_array_length(tgt);
1306
1307 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1308 // Divide src size by 2 if String is UTF16 encoded
1309 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1310 }
1311 if (ae == StrIntrinsicNode::UU) {
1312 // Divide substring size by 2 if String is UTF16 encoded
1313 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1314 }
1315
1316 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1317 if (result != NULL) {
1318 result_phi->init_req(3, result);
1319 result_rgn->init_req(3, control());
1320 }
1321 set_control(_gvn.transform(result_rgn));
1322 record_for_igvn(result_rgn);
1323 set_result(_gvn.transform(result_phi));
1324
1325 return true;
1326 }
1327
1328 //-----------------------------inline_string_indexOf-----------------------
1329 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1330 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1331 return false;
1332 }
1333 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1334 return false;
1335 }
1336 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1337 Node* src = argument(0); // byte[]
1338 Node* src_count = argument(1); // char count
1339 Node* tgt = argument(2); // byte[]
1340 Node* tgt_count = argument(3); // char count
1341 Node* from_index = argument(4); // char index
1342
1343 // Multiply byte array index by 2 if String is UTF16 encoded
1344 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1345 src_count = _gvn.transform(new SubINode(src_count, from_index));
1346 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1347 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1348
1349 // Range checks
1350 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1351 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1352 if (stopped()) {
1353 return true;
1354 }
1355
1356 RegionNode* region = new RegionNode(5);
1357 Node* phi = new PhiNode(region, TypeInt::INT);
1358
1359 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1360 if (result != NULL) {
1361 // The result is index relative to from_index if substring was found, -1 otherwise.
1362 // Generate code which will fold into cmove.
1363 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1364 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1365
1366 Node* if_lt = generate_slow_guard(bol, NULL);
1367 if (if_lt != NULL) {
1368 // result == -1
1369 phi->init_req(3, result);
1370 region->init_req(3, if_lt);
1371 }
1372 if (!stopped()) {
1373 result = _gvn.transform(new AddINode(result, from_index));
1374 phi->init_req(4, result);
1375 region->init_req(4, control());
1376 }
1377 }
1378
1379 set_control(_gvn.transform(region));
1380 record_for_igvn(region);
1381 set_result(_gvn.transform(phi));
1382 clear_upper_avx();
1383
1384 return true;
1385 }
1386
1387 // Create StrIndexOfNode with fast path checks
1388 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1389 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1390 // Check for substr count > string count
1391 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1392 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1393 Node* if_gt = generate_slow_guard(bol, NULL);
1394 if (if_gt != NULL) {
1395 phi->init_req(1, intcon(-1));
1396 region->init_req(1, if_gt);
1397 }
1398 if (!stopped()) {
1399 // Check for substr count == 0
1400 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1401 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1402 Node* if_zero = generate_slow_guard(bol, NULL);
1403 if (if_zero != NULL) {
1404 phi->init_req(2, intcon(0));
1405 region->init_req(2, if_zero);
1406 }
1407 }
1408 if (!stopped()) {
1409 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1410 }
1411 return NULL;
1412 }
1413
1414 //-----------------------------inline_string_indexOfChar-----------------------
1415 bool LibraryCallKit::inline_string_indexOfChar() {
1416 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1417 return false;
1418 }
1419 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1420 return false;
1421 }
1422 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1423 Node* src = argument(0); // byte[]
1424 Node* tgt = argument(1); // tgt is int ch
1425 Node* from_index = argument(2);
1426 Node* max = argument(3);
1427
1428 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1429 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1430 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1431
1432 // Range checks
1433 generate_string_range_check(src, src_offset, src_count, true);
1434 if (stopped()) {
1435 return true;
1436 }
1437
1438 RegionNode* region = new RegionNode(3);
1439 Node* phi = new PhiNode(region, TypeInt::INT);
1440
1441 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1442 C->set_has_split_ifs(true); // Has chance for split-if optimization
1443 _gvn.transform(result);
1444
1445 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1446 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1447
1448 Node* if_lt = generate_slow_guard(bol, NULL);
1449 if (if_lt != NULL) {
1450 // result == -1
1451 phi->init_req(2, result);
1452 region->init_req(2, if_lt);
1453 }
1454 if (!stopped()) {
1455 result = _gvn.transform(new AddINode(result, from_index));
1456 phi->init_req(1, result);
1457 region->init_req(1, control());
1458 }
1459 set_control(_gvn.transform(region));
1460 record_for_igvn(region);
1461 set_result(_gvn.transform(phi));
1462
1463 return true;
1464 }
1465 //---------------------------inline_string_copy---------------------
1466 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1467 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1468 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1469 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1470 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1471 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1472 bool LibraryCallKit::inline_string_copy(bool compress) {
1473 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1474 return false;
1475 }
1476 int nargs = 5; // 2 oops, 3 ints
1477 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1478
1479 Node* src = argument(0);
1480 Node* src_offset = argument(1);
1481 Node* dst = argument(2);
1482 Node* dst_offset = argument(3);
1483 Node* length = argument(4);
1484
1485 // Check for allocation before we add nodes that would confuse
1486 // tightly_coupled_allocation()
1487 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1488
1489 // Figure out the size and type of the elements we will be copying.
1490 const Type* src_type = src->Value(&_gvn);
1491 const Type* dst_type = dst->Value(&_gvn);
1492 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1493 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1494 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1495 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1496 "Unsupported array types for inline_string_copy");
1497
1498 // Convert char[] offsets to byte[] offsets
1499 bool convert_src = (compress && src_elem == T_BYTE);
1500 bool convert_dst = (!compress && dst_elem == T_BYTE);
1501 if (convert_src) {
1502 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1503 } else if (convert_dst) {
1504 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1505 }
1506
1507 // Range checks
1508 generate_string_range_check(src, src_offset, length, convert_src);
1509 generate_string_range_check(dst, dst_offset, length, convert_dst);
1510 if (stopped()) {
1511 return true;
1512 }
1513
1514 Node* src_start = array_element_address(src, src_offset, src_elem);
1515 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1516 // 'src_start' points to src array + scaled offset
1517 // 'dst_start' points to dst array + scaled offset
1518 Node* count = NULL;
1519 if (compress) {
1520 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1521 } else {
1522 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1523 }
1524
1525 if (alloc != NULL) {
1526 if (alloc->maybe_set_complete(&_gvn)) {
1527 // "You break it, you buy it."
1528 InitializeNode* init = alloc->initialization();
1529 assert(init->is_complete(), "we just did this");
1530 init->set_complete_with_arraycopy();
1531 assert(dst->is_CheckCastPP(), "sanity");
1532 assert(dst->in(0)->in(0) == init, "dest pinned");
1533 }
1534 // Do not let stores that initialize this object be reordered with
1535 // a subsequent store that would make this object accessible by
1536 // other threads.
1537 // Record what AllocateNode this StoreStore protects so that
1538 // escape analysis can go from the MemBarStoreStoreNode to the
1539 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1540 // based on the escape status of the AllocateNode.
1541 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1542 }
1543 if (compress) {
1544 set_result(_gvn.transform(count));
1545 }
1546 clear_upper_avx();
1547
1548 return true;
1549 }
1550
1551 #ifdef _LP64
1552 #define XTOP ,top() /*additional argument*/
1553 #else //_LP64
1554 #define XTOP /*no additional argument*/
1555 #endif //_LP64
1556
1557 //------------------------inline_string_toBytesU--------------------------
1558 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1559 bool LibraryCallKit::inline_string_toBytesU() {
1560 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1561 return false;
1562 }
1563 // Get the arguments.
1564 Node* value = argument(0);
1565 Node* offset = argument(1);
1566 Node* length = argument(2);
1567
1568 Node* newcopy = NULL;
1569
1570 // Set the original stack and the reexecute bit for the interpreter to reexecute
1571 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1572 { PreserveReexecuteState preexecs(this);
1573 jvms()->set_should_reexecute(true);
1574
1575 // Check if a null path was taken unconditionally.
1576 value = null_check(value);
1577
1578 RegionNode* bailout = new RegionNode(1);
1579 record_for_igvn(bailout);
1580
1581 // Range checks
1582 generate_negative_guard(offset, bailout);
1583 generate_negative_guard(length, bailout);
1584 generate_limit_guard(offset, length, load_array_length(value), bailout);
1585 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1586 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1587
1588 if (bailout->req() > 1) {
1589 PreserveJVMState pjvms(this);
1590 set_control(_gvn.transform(bailout));
1591 uncommon_trap(Deoptimization::Reason_intrinsic,
1592 Deoptimization::Action_maybe_recompile);
1593 }
1594 if (stopped()) {
1595 return true;
1596 }
1597
1598 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1599 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1600 newcopy = new_array(klass_node, size, 0); // no arguments to push
1601 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1602
1603 // Calculate starting addresses.
1604 Node* src_start = array_element_address(value, offset, T_CHAR);
1605 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1606
1607 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1608 const TypeInt* toffset = gvn().type(offset)->is_int();
1609 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1610
1611 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1612 const char* copyfunc_name = "arraycopy";
1613 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1614 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1615 OptoRuntime::fast_arraycopy_Type(),
1616 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1617 src_start, dst_start, ConvI2X(length) XTOP);
1618 // Do not let reads from the cloned object float above the arraycopy.
1619 if (alloc != NULL) {
1620 if (alloc->maybe_set_complete(&_gvn)) {
1621 // "You break it, you buy it."
1622 InitializeNode* init = alloc->initialization();
1623 assert(init->is_complete(), "we just did this");
1624 init->set_complete_with_arraycopy();
1625 assert(newcopy->is_CheckCastPP(), "sanity");
1626 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1627 }
1628 // Do not let stores that initialize this object be reordered with
1629 // a subsequent store that would make this object accessible by
1630 // other threads.
1631 // Record what AllocateNode this StoreStore protects so that
1632 // escape analysis can go from the MemBarStoreStoreNode to the
1633 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1634 // based on the escape status of the AllocateNode.
1635 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1636 } else {
1637 insert_mem_bar(Op_MemBarCPUOrder);
1638 }
1639 } // original reexecute is set back here
1640
1641 C->set_has_split_ifs(true); // Has chance for split-if optimization
1642 if (!stopped()) {
1643 set_result(newcopy);
1644 }
1645 clear_upper_avx();
1646
1647 return true;
1648 }
1649
1650 //------------------------inline_string_getCharsU--------------------------
1651 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1652 bool LibraryCallKit::inline_string_getCharsU() {
1653 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1654 return false;
1655 }
1656
1657 // Get the arguments.
1658 Node* src = argument(0);
1659 Node* src_begin = argument(1);
1660 Node* src_end = argument(2); // exclusive offset (i < src_end)
1661 Node* dst = argument(3);
1662 Node* dst_begin = argument(4);
1663
1664 // Check for allocation before we add nodes that would confuse
1665 // tightly_coupled_allocation()
1666 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1667
1668 // Check if a null path was taken unconditionally.
1669 src = null_check(src);
1670 dst = null_check(dst);
1671 if (stopped()) {
1672 return true;
1673 }
1674
1675 // Get length and convert char[] offset to byte[] offset
1676 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1677 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1678
1679 // Range checks
1680 generate_string_range_check(src, src_begin, length, true);
1681 generate_string_range_check(dst, dst_begin, length, false);
1682 if (stopped()) {
1683 return true;
1684 }
1685
1686 if (!stopped()) {
1687 // Calculate starting addresses.
1688 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1689 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1690
1691 // Check if array addresses are aligned to HeapWordSize
1692 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1693 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1694 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1695 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1696
1697 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1698 const char* copyfunc_name = "arraycopy";
1699 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1700 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1701 OptoRuntime::fast_arraycopy_Type(),
1702 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1703 src_start, dst_start, ConvI2X(length) XTOP);
1704 // Do not let reads from the cloned object float above the arraycopy.
1705 if (alloc != NULL) {
1706 if (alloc->maybe_set_complete(&_gvn)) {
1707 // "You break it, you buy it."
1708 InitializeNode* init = alloc->initialization();
1709 assert(init->is_complete(), "we just did this");
1710 init->set_complete_with_arraycopy();
1711 assert(dst->is_CheckCastPP(), "sanity");
1712 assert(dst->in(0)->in(0) == init, "dest pinned");
1713 }
1714 // Do not let stores that initialize this object be reordered with
1715 // a subsequent store that would make this object accessible by
1716 // other threads.
1717 // Record what AllocateNode this StoreStore protects so that
1718 // escape analysis can go from the MemBarStoreStoreNode to the
1719 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1720 // based on the escape status of the AllocateNode.
1721 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1722 } else {
1723 insert_mem_bar(Op_MemBarCPUOrder);
1724 }
1725 }
1726
1727 C->set_has_split_ifs(true); // Has chance for split-if optimization
1728 return true;
1729 }
1730
1731 //----------------------inline_string_char_access----------------------------
1732 // Store/Load char to/from byte[] array.
1733 // static void StringUTF16.putChar(byte[] val, int index, int c)
1734 // static char StringUTF16.getChar(byte[] val, int index)
1735 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1736 Node* value = argument(0);
1737 Node* index = argument(1);
1738 Node* ch = is_store ? argument(2) : NULL;
1739
1740 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1741 // correctly requires matched array shapes.
1742 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1743 "sanity: byte[] and char[] bases agree");
1744 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1745 "sanity: byte[] and char[] scales agree");
1746
1747 // Bail when getChar over constants is requested: constant folding would
1748 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1749 // Java method would constant fold nicely instead.
1750 if (!is_store && value->is_Con() && index->is_Con()) {
1751 return false;
1752 }
1753
1754 Node* adr = array_element_address(value, index, T_CHAR);
1755 if (adr->is_top()) {
1756 return false;
1757 }
1758 if (is_store) {
1759 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1760 false, false, true /* mismatched */);
1761 } else {
1762 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1763 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1764 set_result(ch);
1765 }
1766 return true;
1767 }
1768
1769 //--------------------------round_double_node--------------------------------
1770 // Round a double node if necessary.
1771 Node* LibraryCallKit::round_double_node(Node* n) {
1772 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1773 n = _gvn.transform(new RoundDoubleNode(0, n));
1774 return n;
1775 }
1776
1777 //------------------------------inline_math-----------------------------------
1778 // public static double Math.abs(double)
1779 // public static double Math.sqrt(double)
1780 // public static double Math.log(double)
1781 // public static double Math.log10(double)
1782 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1783 Node* arg = round_double_node(argument(0));
1784 Node* n = NULL;
1785 switch (id) {
1786 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1787 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1788 default: fatal_unexpected_iid(id); break;
1789 }
1790 set_result(_gvn.transform(n));
1791 return true;
1792 }
1793
1794 //------------------------------runtime_math-----------------------------
1795 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1796 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1797 "must be (DD)D or (D)D type");
1798
1799 // Inputs
1800 Node* a = round_double_node(argument(0));
1801 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1802
1803 const TypePtr* no_memory_effects = NULL;
1804 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1805 no_memory_effects,
1806 a, top(), b, b ? top() : NULL);
1807 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1808 #ifdef ASSERT
1809 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1810 assert(value_top == top(), "second value must be top");
1811 #endif
1812
1813 set_result(value);
1814 return true;
1815 }
1816
1817 //------------------------------inline_math_native-----------------------------
1818 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1819 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1820 switch (id) {
1821 // These intrinsics are not properly supported on all hardware
1822 case vmIntrinsics::_dsin:
1823 return StubRoutines::dsin() != NULL ?
1824 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1825 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1826 case vmIntrinsics::_dcos:
1827 return StubRoutines::dcos() != NULL ?
1828 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1829 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1830 case vmIntrinsics::_dtan:
1831 return StubRoutines::dtan() != NULL ?
1832 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1833 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1834 case vmIntrinsics::_dlog:
1835 return StubRoutines::dlog() != NULL ?
1836 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1837 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1838 case vmIntrinsics::_dlog10:
1839 return StubRoutines::dlog10() != NULL ?
1840 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1841 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1842
1843 // These intrinsics are supported on all hardware
1844 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1845 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1846
1847 case vmIntrinsics::_dexp:
1848 return StubRoutines::dexp() != NULL ?
1849 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1850 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1851 case vmIntrinsics::_dpow: {
1852 Node* exp = round_double_node(argument(2));
1853 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1854 if (d != NULL && d->getd() == 2.0) {
1855 // Special case: pow(x, 2.0) => x * x
1856 Node* base = round_double_node(argument(0));
1857 set_result(_gvn.transform(new MulDNode(base, base)));
1858 return true;
1859 }
1860 return StubRoutines::dexp() != NULL ?
1861 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1862 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1863 }
1864 #undef FN_PTR
1865
1866 // These intrinsics are not yet correctly implemented
1867 case vmIntrinsics::_datan2:
1868 return false;
1869
1870 default:
1871 fatal_unexpected_iid(id);
1872 return false;
1873 }
1874 }
1875
1876 static bool is_simple_name(Node* n) {
1877 return (n->req() == 1 // constant
1878 || (n->is_Type() && n->as_Type()->type()->singleton())
1879 || n->is_Proj() // parameter or return value
1880 || n->is_Phi() // local of some sort
1881 );
1882 }
1883
1884 //----------------------------inline_notify-----------------------------------*
1885 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1886 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1887 address func;
1888 if (id == vmIntrinsics::_notify) {
1889 func = OptoRuntime::monitor_notify_Java();
1890 } else {
1891 func = OptoRuntime::monitor_notifyAll_Java();
1892 }
1893 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1894 make_slow_call_ex(call, env()->Throwable_klass(), false);
1895 return true;
1896 }
1897
1898
1899 //----------------------------inline_min_max-----------------------------------
1900 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1901 set_result(generate_min_max(id, argument(0), argument(1)));
1902 return true;
1903 }
1904
1905 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1906 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1907 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1908 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1909 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1910
1911 {
1912 PreserveJVMState pjvms(this);
1913 PreserveReexecuteState preexecs(this);
1914 jvms()->set_should_reexecute(true);
1915
1916 set_control(slow_path);
1917 set_i_o(i_o());
1918
1919 uncommon_trap(Deoptimization::Reason_intrinsic,
1920 Deoptimization::Action_none);
1921 }
1922
1923 set_control(fast_path);
1924 set_result(math);
1925 }
1926
1927 template <typename OverflowOp>
1928 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1929 typedef typename OverflowOp::MathOp MathOp;
1930
1931 MathOp* mathOp = new MathOp(arg1, arg2);
1932 Node* operation = _gvn.transform( mathOp );
1933 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1934 inline_math_mathExact(operation, ofcheck);
1935 return true;
1936 }
1937
1938 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1939 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1940 }
1941
1942 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1943 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1944 }
1945
1946 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1947 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1948 }
1949
1950 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1951 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1952 }
1953
1954 bool LibraryCallKit::inline_math_negateExactI() {
1955 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1956 }
1957
1958 bool LibraryCallKit::inline_math_negateExactL() {
1959 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1960 }
1961
1962 bool LibraryCallKit::inline_math_multiplyExactI() {
1963 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1964 }
1965
1966 bool LibraryCallKit::inline_math_multiplyExactL() {
1967 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1968 }
1969
1970 bool LibraryCallKit::inline_math_multiplyHigh() {
1971 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1972 return true;
1973 }
1974
1975 Node*
1976 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1977 // These are the candidate return value:
1978 Node* xvalue = x0;
1979 Node* yvalue = y0;
1980
1981 if (xvalue == yvalue) {
1982 return xvalue;
1983 }
1984
1985 bool want_max = (id == vmIntrinsics::_max);
1986
1987 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1988 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1989 if (txvalue == NULL || tyvalue == NULL) return top();
1990 // This is not really necessary, but it is consistent with a
1991 // hypothetical MaxINode::Value method:
1992 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1993
1994 // %%% This folding logic should (ideally) be in a different place.
1995 // Some should be inside IfNode, and there to be a more reliable
1996 // transformation of ?: style patterns into cmoves. We also want
1997 // more powerful optimizations around cmove and min/max.
1998
1999 // Try to find a dominating comparison of these guys.
2000 // It can simplify the index computation for Arrays.copyOf
2001 // and similar uses of System.arraycopy.
2002 // First, compute the normalized version of CmpI(x, y).
2003 int cmp_op = Op_CmpI;
2004 Node* xkey = xvalue;
2005 Node* ykey = yvalue;
2006 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2007 if (ideal_cmpxy->is_Cmp()) {
2008 // E.g., if we have CmpI(length - offset, count),
2009 // it might idealize to CmpI(length, count + offset)
2010 cmp_op = ideal_cmpxy->Opcode();
2011 xkey = ideal_cmpxy->in(1);
2012 ykey = ideal_cmpxy->in(2);
2013 }
2014
2015 // Start by locating any relevant comparisons.
2016 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2017 Node* cmpxy = NULL;
2018 Node* cmpyx = NULL;
2019 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2020 Node* cmp = start_from->fast_out(k);
2021 if (cmp->outcnt() > 0 && // must have prior uses
2022 cmp->in(0) == NULL && // must be context-independent
2023 cmp->Opcode() == cmp_op) { // right kind of compare
2024 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2025 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2026 }
2027 }
2028
2029 const int NCMPS = 2;
2030 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2031 int cmpn;
2032 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2033 if (cmps[cmpn] != NULL) break; // find a result
2034 }
2035 if (cmpn < NCMPS) {
2036 // Look for a dominating test that tells us the min and max.
2037 int depth = 0; // Limit search depth for speed
2038 Node* dom = control();
2039 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2040 if (++depth >= 100) break;
2041 Node* ifproj = dom;
2042 if (!ifproj->is_Proj()) continue;
2043 Node* iff = ifproj->in(0);
2044 if (!iff->is_If()) continue;
2045 Node* bol = iff->in(1);
2046 if (!bol->is_Bool()) continue;
2047 Node* cmp = bol->in(1);
2048 if (cmp == NULL) continue;
2049 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2050 if (cmps[cmpn] == cmp) break;
2051 if (cmpn == NCMPS) continue;
2052 BoolTest::mask btest = bol->as_Bool()->_test._test;
2053 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2054 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2055 // At this point, we know that 'x btest y' is true.
2056 switch (btest) {
2057 case BoolTest::eq:
2058 // They are proven equal, so we can collapse the min/max.
2059 // Either value is the answer. Choose the simpler.
2060 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2061 return yvalue;
2062 return xvalue;
2063 case BoolTest::lt: // x < y
2064 case BoolTest::le: // x <= y
2065 return (want_max ? yvalue : xvalue);
2066 case BoolTest::gt: // x > y
2067 case BoolTest::ge: // x >= y
2068 return (want_max ? xvalue : yvalue);
2069 default:
2070 break;
2071 }
2072 }
2073 }
2074
2075 // We failed to find a dominating test.
2076 // Let's pick a test that might GVN with prior tests.
2077 Node* best_bol = NULL;
2078 BoolTest::mask best_btest = BoolTest::illegal;
2079 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2080 Node* cmp = cmps[cmpn];
2081 if (cmp == NULL) continue;
2082 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2083 Node* bol = cmp->fast_out(j);
2084 if (!bol->is_Bool()) continue;
2085 BoolTest::mask btest = bol->as_Bool()->_test._test;
2086 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2087 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2088 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2089 best_bol = bol->as_Bool();
2090 best_btest = btest;
2091 }
2092 }
2093 }
2094
2095 Node* answer_if_true = NULL;
2096 Node* answer_if_false = NULL;
2097 switch (best_btest) {
2098 default:
2099 if (cmpxy == NULL)
2100 cmpxy = ideal_cmpxy;
2101 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2102 // and fall through:
2103 case BoolTest::lt: // x < y
2104 case BoolTest::le: // x <= y
2105 answer_if_true = (want_max ? yvalue : xvalue);
2106 answer_if_false = (want_max ? xvalue : yvalue);
2107 break;
2108 case BoolTest::gt: // x > y
2109 case BoolTest::ge: // x >= y
2110 answer_if_true = (want_max ? xvalue : yvalue);
2111 answer_if_false = (want_max ? yvalue : xvalue);
2112 break;
2113 }
2114
2115 jint hi, lo;
2116 if (want_max) {
2117 // We can sharpen the minimum.
2118 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2119 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2120 } else {
2121 // We can sharpen the maximum.
2122 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2123 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2124 }
2125
2126 // Use a flow-free graph structure, to avoid creating excess control edges
2127 // which could hinder other optimizations.
2128 // Since Math.min/max is often used with arraycopy, we want
2129 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2130 Node* cmov = CMoveNode::make(NULL, best_bol,
2131 answer_if_false, answer_if_true,
2132 TypeInt::make(lo, hi, widen));
2133
2134 return _gvn.transform(cmov);
2135
2136 /*
2137 // This is not as desirable as it may seem, since Min and Max
2138 // nodes do not have a full set of optimizations.
2139 // And they would interfere, anyway, with 'if' optimizations
2140 // and with CMoveI canonical forms.
2141 switch (id) {
2142 case vmIntrinsics::_min:
2143 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2144 case vmIntrinsics::_max:
2145 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2146 default:
2147 ShouldNotReachHere();
2148 }
2149 */
2150 }
2151
2152 inline int
2153 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2154 const TypePtr* base_type = TypePtr::NULL_PTR;
2155 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2156 if (base_type == NULL) {
2157 // Unknown type.
2158 return Type::AnyPtr;
2159 } else if (base_type == TypePtr::NULL_PTR) {
2160 // Since this is a NULL+long form, we have to switch to a rawptr.
2161 base = _gvn.transform(new CastX2PNode(offset));
2162 offset = MakeConX(0);
2163 return Type::RawPtr;
2164 } else if (base_type->base() == Type::RawPtr) {
2165 return Type::RawPtr;
2166 } else if (base_type->isa_oopptr()) {
2167 // Base is never null => always a heap address.
2168 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2169 return Type::OopPtr;
2170 }
2171 // Offset is small => always a heap address.
2172 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2173 if (offset_type != NULL &&
2174 base_type->offset() == 0 && // (should always be?)
2175 offset_type->_lo >= 0 &&
2176 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2177 return Type::OopPtr;
2178 } else if (type == T_OBJECT) {
2179 // off heap access to an oop doesn't make any sense. Has to be on
2180 // heap.
2181 return Type::OopPtr;
2182 }
2183 // Otherwise, it might either be oop+off or NULL+addr.
2184 return Type::AnyPtr;
2185 } else {
2186 // No information:
2187 return Type::AnyPtr;
2188 }
2189 }
2190
2191 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2192 Node* uncasted_base = base;
2193 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2194 if (kind == Type::RawPtr) {
2195 return basic_plus_adr(top(), uncasted_base, offset);
2196 } else if (kind == Type::AnyPtr) {
2197 assert(base == uncasted_base, "unexpected base change");
2198 if (can_cast) {
2199 if (!_gvn.type(base)->speculative_maybe_null() &&
2200 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2201 // According to profiling, this access is always on
2202 // heap. Casting the base to not null and thus avoiding membars
2203 // around the access should allow better optimizations
2204 Node* null_ctl = top();
2205 base = null_check_oop(base, &null_ctl, true, true, true);
2206 assert(null_ctl->is_top(), "no null control here");
2207 return basic_plus_adr(base, offset);
2208 } else if (_gvn.type(base)->speculative_always_null() &&
2209 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2210 // According to profiling, this access is always off
2211 // heap.
2212 base = null_assert(base);
2213 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2214 offset = MakeConX(0);
2215 return basic_plus_adr(top(), raw_base, offset);
2216 }
2217 }
2218 // We don't know if it's an on heap or off heap access. Fall back
2219 // to raw memory access.
2220 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2221 return basic_plus_adr(top(), raw, offset);
2222 } else {
2223 assert(base == uncasted_base, "unexpected base change");
2224 // We know it's an on heap access so base can't be null
2225 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2226 base = must_be_not_null(base, true);
2227 }
2228 return basic_plus_adr(base, offset);
2229 }
2230 }
2231
2232 //--------------------------inline_number_methods-----------------------------
2233 // inline int Integer.numberOfLeadingZeros(int)
2234 // inline int Long.numberOfLeadingZeros(long)
2235 //
2236 // inline int Integer.numberOfTrailingZeros(int)
2237 // inline int Long.numberOfTrailingZeros(long)
2238 //
2239 // inline int Integer.bitCount(int)
2240 // inline int Long.bitCount(long)
2241 //
2242 // inline char Character.reverseBytes(char)
2243 // inline short Short.reverseBytes(short)
2244 // inline int Integer.reverseBytes(int)
2245 // inline long Long.reverseBytes(long)
2246 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2247 Node* arg = argument(0);
2248 Node* n = NULL;
2249 switch (id) {
2250 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2251 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2252 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2253 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2254 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2255 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2256 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2257 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2258 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2259 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2260 default: fatal_unexpected_iid(id); break;
2261 }
2262 set_result(_gvn.transform(n));
2263 return true;
2264 }
2265
2266 //----------------------------inline_unsafe_access----------------------------
2267
2268 // Helper that guards and inserts a pre-barrier.
2269 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2270 Node* pre_val, bool need_mem_bar) {
2271 // We could be accessing the referent field of a reference object. If so, when G1
2272 // is enabled, we need to log the value in the referent field in an SATB buffer.
2273 // This routine performs some compile time filters and generates suitable
2274 // runtime filters that guard the pre-barrier code.
2275 // Also add memory barrier for non volatile load from the referent field
2276 // to prevent commoning of loads across safepoint.
2277 if (!UseG1GC && !need_mem_bar)
2278 return;
2279
2280 // Some compile time checks.
2281
2282 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2283 const TypeX* otype = offset->find_intptr_t_type();
2284 if (otype != NULL && otype->is_con() &&
2285 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2286 // Constant offset but not the reference_offset so just return
2287 return;
2288 }
2289
2290 // We only need to generate the runtime guards for instances.
2291 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2292 if (btype != NULL) {
2293 if (btype->isa_aryptr()) {
2294 // Array type so nothing to do
2295 return;
2296 }
2297
2298 const TypeInstPtr* itype = btype->isa_instptr();
2299 if (itype != NULL) {
2300 // Can the klass of base_oop be statically determined to be
2301 // _not_ a sub-class of Reference and _not_ Object?
2302 ciKlass* klass = itype->klass();
2303 if ( klass->is_loaded() &&
2304 !klass->is_subtype_of(env()->Reference_klass()) &&
2305 !env()->Object_klass()->is_subtype_of(klass)) {
2306 return;
2307 }
2308 }
2309 }
2310
2311 // The compile time filters did not reject base_oop/offset so
2312 // we need to generate the following runtime filters
2313 //
2314 // if (offset == java_lang_ref_Reference::_reference_offset) {
2315 // if (instance_of(base, java.lang.ref.Reference)) {
2316 // pre_barrier(_, pre_val, ...);
2317 // }
2318 // }
2319
2320 float likely = PROB_LIKELY( 0.999);
2321 float unlikely = PROB_UNLIKELY(0.999);
2322
2323 IdealKit ideal(this);
2324 #define __ ideal.
2325
2326 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2327
2328 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2329 // Update graphKit memory and control from IdealKit.
2330 sync_kit(ideal);
2331
2332 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2333 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2334
2335 // Update IdealKit memory and control from graphKit.
2336 __ sync_kit(this);
2337
2338 Node* one = __ ConI(1);
2339 // is_instof == 0 if base_oop == NULL
2340 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2341
2342 // Update graphKit from IdeakKit.
2343 sync_kit(ideal);
2344
2345 // Use the pre-barrier to record the value in the referent field
2346 pre_barrier(false /* do_load */,
2347 __ ctrl(),
2348 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2349 pre_val /* pre_val */,
2350 T_OBJECT);
2351 if (need_mem_bar) {
2352 // Add memory barrier to prevent commoning reads from this field
2353 // across safepoint since GC can change its value.
2354 insert_mem_bar(Op_MemBarCPUOrder);
2355 }
2356 // Update IdealKit from graphKit.
2357 __ sync_kit(this);
2358
2359 } __ end_if(); // _ref_type != ref_none
2360 } __ end_if(); // offset == referent_offset
2361
2362 // Final sync IdealKit and GraphKit.
2363 final_sync(ideal);
2364 #undef __
2365 }
2366
2367
2368 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2369 // Attempt to infer a sharper value type from the offset and base type.
2370 ciKlass* sharpened_klass = NULL;
2371
2372 // See if it is an instance field, with an object type.
2373 if (alias_type->field() != NULL) {
2374 if (alias_type->field()->type()->is_klass()) {
2375 sharpened_klass = alias_type->field()->type()->as_klass();
2376 }
2377 }
2378
2379 // See if it is a narrow oop array.
2380 if (adr_type->isa_aryptr()) {
2381 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2382 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2383 if (elem_type != NULL) {
2384 sharpened_klass = elem_type->klass();
2385 }
2386 }
2387 }
2388
2389 // The sharpened class might be unloaded if there is no class loader
2390 // contraint in place.
2391 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2392 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2393
2394 #ifndef PRODUCT
2395 if (C->print_intrinsics() || C->print_inlining()) {
2396 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2397 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2398 }
2399 #endif
2400 // Sharpen the value type.
2401 return tjp;
2402 }
2403 return NULL;
2404 }
2405
2406 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2407 if (callee()->is_static()) return false; // caller must have the capability!
2408 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2409 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2410 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2411
2412 #ifndef PRODUCT
2413 {
2414 ResourceMark rm;
2415 // Check the signatures.
2416 ciSignature* sig = callee()->signature();
2417 #ifdef ASSERT
2418 if (!is_store) {
2419 // Object getObject(Object base, int/long offset), etc.
2420 BasicType rtype = sig->return_type()->basic_type();
2421 assert(rtype == type, "getter must return the expected value");
2422 assert(sig->count() == 2, "oop getter has 2 arguments");
2423 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2424 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2425 } else {
2426 // void putObject(Object base, int/long offset, Object x), etc.
2427 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2428 assert(sig->count() == 3, "oop putter has 3 arguments");
2429 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2430 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2431 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2432 assert(vtype == type, "putter must accept the expected value");
2433 }
2434 #endif // ASSERT
2435 }
2436 #endif //PRODUCT
2437
2438 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2439
2440 Node* receiver = argument(0); // type: oop
2441
2442 // Build address expression.
2443 Node* adr;
2444 Node* heap_base_oop = top();
2445 Node* offset = top();
2446 Node* val;
2447
2448 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2449 Node* base = argument(1); // type: oop
2450 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2451 offset = argument(2); // type: long
2452 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2453 // to be plain byte offsets, which are also the same as those accepted
2454 // by oopDesc::field_addr.
2455 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2456 "fieldOffset must be byte-scaled");
2457 // 32-bit machines ignore the high half!
2458 offset = ConvL2X(offset);
2459 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2460
2461 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2462 heap_base_oop = base;
2463 } else if (type == T_OBJECT) {
2464 return false; // off-heap oop accesses are not supported
2465 }
2466
2467 // Can base be NULL? Otherwise, always on-heap access.
2468 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2469
2470 val = is_store ? argument(4) : NULL;
2471
2472 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2473
2474 // Try to categorize the address.
2475 Compile::AliasType* alias_type = C->alias_type(adr_type);
2476 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2477
2478 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2479 alias_type->adr_type() == TypeAryPtr::RANGE) {
2480 return false; // not supported
2481 }
2482
2483 bool mismatched = false;
2484 BasicType bt = alias_type->basic_type();
2485 if (bt != T_ILLEGAL) {
2486 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2487 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2488 // Alias type doesn't differentiate between byte[] and boolean[]).
2489 // Use address type to get the element type.
2490 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2491 }
2492 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2493 // accessing an array field with getObject is not a mismatch
2494 bt = T_OBJECT;
2495 }
2496 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2497 // Don't intrinsify mismatched object accesses
2498 return false;
2499 }
2500 mismatched = (bt != type);
2501 } else if (alias_type->adr_type()->isa_oopptr()) {
2502 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2503 }
2504
2505 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2506
2507 // First guess at the value type.
2508 const Type *value_type = Type::get_const_basic_type(type);
2509
2510 // We will need memory barriers unless we can determine a unique
2511 // alias category for this reference. (Note: If for some reason
2512 // the barriers get omitted and the unsafe reference begins to "pollute"
2513 // the alias analysis of the rest of the graph, either Compile::can_alias
2514 // or Compile::must_alias will throw a diagnostic assert.)
2515 bool need_mem_bar = false;
2516 switch (kind) {
2517 case Relaxed:
2518 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2519 break;
2520 case Opaque:
2521 // Opaque uses CPUOrder membars for protection against code movement.
2522 case Acquire:
2523 case Release:
2524 case Volatile:
2525 need_mem_bar = true;
2526 break;
2527 default:
2528 ShouldNotReachHere();
2529 }
2530
2531 // Some accesses require access atomicity for all types, notably longs and doubles.
2532 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2533 bool requires_atomic_access = false;
2534 switch (kind) {
2535 case Relaxed:
2536 requires_atomic_access = AlwaysAtomicAccesses;
2537 break;
2538 case Opaque:
2539 // Opaque accesses are atomic.
2540 case Acquire:
2541 case Release:
2542 case Volatile:
2543 requires_atomic_access = true;
2544 break;
2545 default:
2546 ShouldNotReachHere();
2547 }
2548
2549 // Figure out the memory ordering.
2550 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2551 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2552
2553 // If we are reading the value of the referent field of a Reference
2554 // object (either by using Unsafe directly or through reflection)
2555 // then, if G1 is enabled, we need to record the referent in an
2556 // SATB log buffer using the pre-barrier mechanism.
2557 // Also we need to add memory barrier to prevent commoning reads
2558 // from this field across safepoint since GC can change its value.
2559 bool need_read_barrier = !is_store &&
2560 offset != top() && heap_base_oop != top();
2561
2562 if (!is_store && type == T_OBJECT) {
2563 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2564 if (tjp != NULL) {
2565 value_type = tjp;
2566 }
2567 }
2568
2569 receiver = null_check(receiver);
2570 if (stopped()) {
2571 return true;
2572 }
2573 // Heap pointers get a null-check from the interpreter,
2574 // as a courtesy. However, this is not guaranteed by Unsafe,
2575 // and it is not possible to fully distinguish unintended nulls
2576 // from intended ones in this API.
2577
2578 // We need to emit leading and trailing CPU membars (see below) in
2579 // addition to memory membars for special access modes. This is a little
2580 // too strong, but avoids the need to insert per-alias-type
2581 // volatile membars (for stores; compare Parse::do_put_xxx), which
2582 // we cannot do effectively here because we probably only have a
2583 // rough approximation of type.
2584
2585 switch(kind) {
2586 case Relaxed:
2587 case Opaque:
2588 case Acquire:
2589 break;
2590 case Release:
2591 case Volatile:
2592 if (is_store) {
2593 insert_mem_bar(Op_MemBarRelease);
2594 } else {
2595 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2596 insert_mem_bar(Op_MemBarVolatile);
2597 }
2598 }
2599 break;
2600 default:
2601 ShouldNotReachHere();
2602 }
2603
2604 // Memory barrier to prevent normal and 'unsafe' accesses from
2605 // bypassing each other. Happens after null checks, so the
2606 // exception paths do not take memory state from the memory barrier,
2607 // so there's no problems making a strong assert about mixing users
2608 // of safe & unsafe memory.
2609 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2610
2611 if (!is_store) {
2612 Node* p = NULL;
2613 // Try to constant fold a load from a constant field
2614 ciField* field = alias_type->field();
2615 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2616 // final or stable field
2617 p = make_constant_from_field(field, heap_base_oop);
2618 }
2619 if (p == NULL) {
2620 // To be valid, unsafe loads may depend on other conditions than
2621 // the one that guards them: pin the Load node
2622 LoadNode::ControlDependency dep = LoadNode::Pinned;
2623 Node* ctrl = control();
2624 // non volatile loads may be able to float
2625 if (!need_mem_bar && adr_type->isa_instptr()) {
2626 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2627 intptr_t offset = Type::OffsetBot;
2628 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2629 if (offset >= 0) {
2630 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2631 if (offset < s) {
2632 // Guaranteed to be a valid access, no need to pin it
2633 dep = LoadNode::DependsOnlyOnTest;
2634 ctrl = NULL;
2635 }
2636 }
2637 }
2638 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2639 // load value
2640 switch (type) {
2641 case T_BOOLEAN:
2642 {
2643 // Normalize the value returned by getBoolean in the following cases
2644 if (mismatched ||
2645 heap_base_oop == top() || // - heap_base_oop is NULL or
2646 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2647 // and the unsafe access is made to large offset
2648 // (i.e., larger than the maximum offset necessary for any
2649 // field access)
2650 ) {
2651 IdealKit ideal = IdealKit(this);
2652 #define __ ideal.
2653 IdealVariable normalized_result(ideal);
2654 __ declarations_done();
2655 __ set(normalized_result, p);
2656 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2657 __ set(normalized_result, ideal.ConI(1));
2658 ideal.end_if();
2659 final_sync(ideal);
2660 p = __ value(normalized_result);
2661 #undef __
2662 }
2663 }
2664 case T_CHAR:
2665 case T_BYTE:
2666 case T_SHORT:
2667 case T_INT:
2668 case T_LONG:
2669 case T_FLOAT:
2670 case T_DOUBLE:
2671 break;
2672 case T_OBJECT:
2673 if (need_read_barrier) {
2674 // We do not require a mem bar inside pre_barrier if need_mem_bar
2675 // is set: the barriers would be emitted by us.
2676 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2677 }
2678 break;
2679 case T_ADDRESS:
2680 // Cast to an int type.
2681 p = _gvn.transform(new CastP2XNode(NULL, p));
2682 p = ConvX2UL(p);
2683 break;
2684 default:
2685 fatal("unexpected type %d: %s", type, type2name(type));
2686 break;
2687 }
2688 }
2689 // The load node has the control of the preceding MemBarCPUOrder. All
2690 // following nodes will have the control of the MemBarCPUOrder inserted at
2691 // the end of this method. So, pushing the load onto the stack at a later
2692 // point is fine.
2693 set_result(p);
2694 } else {
2695 // place effect of store into memory
2696 switch (type) {
2697 case T_DOUBLE:
2698 val = dstore_rounding(val);
2699 break;
2700 case T_ADDRESS:
2701 // Repackage the long as a pointer.
2702 val = ConvL2X(val);
2703 val = _gvn.transform(new CastX2PNode(val));
2704 break;
2705 default:
2706 break;
2707 }
2708
2709 if (type == T_OBJECT) {
2710 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2711 } else {
2712 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2713 }
2714 }
2715
2716 switch(kind) {
2717 case Relaxed:
2718 case Opaque:
2719 case Release:
2720 break;
2721 case Acquire:
2722 case Volatile:
2723 if (!is_store) {
2724 insert_mem_bar(Op_MemBarAcquire);
2725 } else {
2726 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2727 insert_mem_bar(Op_MemBarVolatile);
2728 }
2729 }
2730 break;
2731 default:
2732 ShouldNotReachHere();
2733 }
2734
2735 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2736
2737 return true;
2738 }
2739
2740 //----------------------------inline_unsafe_load_store----------------------------
2741 // This method serves a couple of different customers (depending on LoadStoreKind):
2742 //
2743 // LS_cmp_swap:
2744 //
2745 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2746 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2747 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2748 //
2749 // LS_cmp_swap_weak:
2750 //
2751 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2752 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2753 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2754 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2755 //
2756 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2757 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2758 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2759 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2760 //
2761 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2762 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2763 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2764 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2765 //
2766 // LS_cmp_exchange:
2767 //
2768 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2769 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2770 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2771 //
2772 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2773 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2774 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2775 //
2776 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2777 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2778 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2779 //
2780 // LS_get_add:
2781 //
2782 // int getAndAddInt( Object o, long offset, int delta)
2783 // long getAndAddLong(Object o, long offset, long delta)
2784 //
2785 // LS_get_set:
2786 //
2787 // int getAndSet(Object o, long offset, int newValue)
2788 // long getAndSet(Object o, long offset, long newValue)
2789 // Object getAndSet(Object o, long offset, Object newValue)
2790 //
2791 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2792 // This basic scheme here is the same as inline_unsafe_access, but
2793 // differs in enough details that combining them would make the code
2794 // overly confusing. (This is a true fact! I originally combined
2795 // them, but even I was confused by it!) As much code/comments as
2796 // possible are retained from inline_unsafe_access though to make
2797 // the correspondences clearer. - dl
2798
2799 if (callee()->is_static()) return false; // caller must have the capability!
2800
2801 #ifndef PRODUCT
2802 BasicType rtype;
2803 {
2804 ResourceMark rm;
2805 // Check the signatures.
2806 ciSignature* sig = callee()->signature();
2807 rtype = sig->return_type()->basic_type();
2808 switch(kind) {
2809 case LS_get_add:
2810 case LS_get_set: {
2811 // Check the signatures.
2812 #ifdef ASSERT
2813 assert(rtype == type, "get and set must return the expected type");
2814 assert(sig->count() == 3, "get and set has 3 arguments");
2815 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2816 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2817 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2818 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2819 #endif // ASSERT
2820 break;
2821 }
2822 case LS_cmp_swap:
2823 case LS_cmp_swap_weak: {
2824 // Check the signatures.
2825 #ifdef ASSERT
2826 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2827 assert(sig->count() == 4, "CAS has 4 arguments");
2828 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2829 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2830 #endif // ASSERT
2831 break;
2832 }
2833 case LS_cmp_exchange: {
2834 // Check the signatures.
2835 #ifdef ASSERT
2836 assert(rtype == type, "CAS must return the expected type");
2837 assert(sig->count() == 4, "CAS has 4 arguments");
2838 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2839 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2840 #endif // ASSERT
2841 break;
2842 }
2843 default:
2844 ShouldNotReachHere();
2845 }
2846 }
2847 #endif //PRODUCT
2848
2849 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2850
2851 // Get arguments:
2852 Node* receiver = NULL;
2853 Node* base = NULL;
2854 Node* offset = NULL;
2855 Node* oldval = NULL;
2856 Node* newval = NULL;
2857 switch(kind) {
2858 case LS_cmp_swap:
2859 case LS_cmp_swap_weak:
2860 case LS_cmp_exchange: {
2861 const bool two_slot_type = type2size[type] == 2;
2862 receiver = argument(0); // type: oop
2863 base = argument(1); // type: oop
2864 offset = argument(2); // type: long
2865 oldval = argument(4); // type: oop, int, or long
2866 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2867 break;
2868 }
2869 case LS_get_add:
2870 case LS_get_set: {
2871 receiver = argument(0); // type: oop
2872 base = argument(1); // type: oop
2873 offset = argument(2); // type: long
2874 oldval = NULL;
2875 newval = argument(4); // type: oop, int, or long
2876 break;
2877 }
2878 default:
2879 ShouldNotReachHere();
2880 }
2881
2882 // Build field offset expression.
2883 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2884 // to be plain byte offsets, which are also the same as those accepted
2885 // by oopDesc::field_addr.
2886 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2887 // 32-bit machines ignore the high half of long offsets
2888 offset = ConvL2X(offset);
2889 Node* adr = make_unsafe_address(base, offset, type, false);
2890 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2891
2892 Compile::AliasType* alias_type = C->alias_type(adr_type);
2893 BasicType bt = alias_type->basic_type();
2894 if (bt != T_ILLEGAL &&
2895 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2896 // Don't intrinsify mismatched object accesses.
2897 return false;
2898 }
2899
2900 // For CAS, unlike inline_unsafe_access, there seems no point in
2901 // trying to refine types. Just use the coarse types here.
2902 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2903 const Type *value_type = Type::get_const_basic_type(type);
2904
2905 switch (kind) {
2906 case LS_get_set:
2907 case LS_cmp_exchange: {
2908 if (type == T_OBJECT) {
2909 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2910 if (tjp != NULL) {
2911 value_type = tjp;
2912 }
2913 }
2914 break;
2915 }
2916 case LS_cmp_swap:
2917 case LS_cmp_swap_weak:
2918 case LS_get_add:
2919 break;
2920 default:
2921 ShouldNotReachHere();
2922 }
2923
2924 // Null check receiver.
2925 receiver = null_check(receiver);
2926 if (stopped()) {
2927 return true;
2928 }
2929
2930 int alias_idx = C->get_alias_index(adr_type);
2931
2932 // Memory-model-wise, a LoadStore acts like a little synchronized
2933 // block, so needs barriers on each side. These don't translate
2934 // into actual barriers on most machines, but we still need rest of
2935 // compiler to respect ordering.
2936
2937 switch (access_kind) {
2938 case Relaxed:
2939 case Acquire:
2940 break;
2941 case Release:
2942 insert_mem_bar(Op_MemBarRelease);
2943 break;
2944 case Volatile:
2945 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2946 insert_mem_bar(Op_MemBarVolatile);
2947 } else {
2948 insert_mem_bar(Op_MemBarRelease);
2949 }
2950 break;
2951 default:
2952 ShouldNotReachHere();
2953 }
2954 insert_mem_bar(Op_MemBarCPUOrder);
2955
2956 // Figure out the memory ordering.
2957 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2958
2959 // 4984716: MemBars must be inserted before this
2960 // memory node in order to avoid a false
2961 // dependency which will confuse the scheduler.
2962 Node *mem = memory(alias_idx);
2963
2964 // For now, we handle only those cases that actually exist: ints,
2965 // longs, and Object. Adding others should be straightforward.
2966 Node* load_store = NULL;
2967 switch(type) {
2968 case T_BYTE:
2969 switch(kind) {
2970 case LS_get_add:
2971 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2972 break;
2973 case LS_get_set:
2974 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2975 break;
2976 case LS_cmp_swap_weak:
2977 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2978 break;
2979 case LS_cmp_swap:
2980 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2981 break;
2982 case LS_cmp_exchange:
2983 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2984 break;
2985 default:
2986 ShouldNotReachHere();
2987 }
2988 break;
2989 case T_SHORT:
2990 switch(kind) {
2991 case LS_get_add:
2992 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2993 break;
2994 case LS_get_set:
2995 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2996 break;
2997 case LS_cmp_swap_weak:
2998 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2999 break;
3000 case LS_cmp_swap:
3001 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
3002 break;
3003 case LS_cmp_exchange:
3004 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
3005 break;
3006 default:
3007 ShouldNotReachHere();
3008 }
3009 break;
3010 case T_INT:
3011 switch(kind) {
3012 case LS_get_add:
3013 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
3014 break;
3015 case LS_get_set:
3016 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
3017 break;
3018 case LS_cmp_swap_weak:
3019 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
3020 break;
3021 case LS_cmp_swap:
3022 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
3023 break;
3024 case LS_cmp_exchange:
3025 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
3026 break;
3027 default:
3028 ShouldNotReachHere();
3029 }
3030 break;
3031 case T_LONG:
3032 switch(kind) {
3033 case LS_get_add:
3034 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
3035 break;
3036 case LS_get_set:
3037 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
3038 break;
3039 case LS_cmp_swap_weak:
3040 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3041 break;
3042 case LS_cmp_swap:
3043 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3044 break;
3045 case LS_cmp_exchange:
3046 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
3047 break;
3048 default:
3049 ShouldNotReachHere();
3050 }
3051 break;
3052 case T_OBJECT:
3053 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3054 // could be delayed during Parse (for example, in adjust_map_after_if()).
3055 // Execute transformation here to avoid barrier generation in such case.
3056 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3057 newval = _gvn.makecon(TypePtr::NULL_PTR);
3058
3059 // Reference stores need a store barrier.
3060 switch(kind) {
3061 case LS_get_set: {
3062 // If pre-barrier must execute before the oop store, old value will require do_load here.
3063 if (!can_move_pre_barrier()) {
3064 pre_barrier(true /* do_load*/,
3065 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3066 NULL /* pre_val*/,
3067 T_OBJECT);
3068 } // Else move pre_barrier to use load_store value, see below.
3069 break;
3070 }
3071 case LS_cmp_swap_weak:
3072 case LS_cmp_swap:
3073 case LS_cmp_exchange: {
3074 // Same as for newval above:
3075 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3076 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3077 }
3078 // The only known value which might get overwritten is oldval.
3079 pre_barrier(false /* do_load */,
3080 control(), NULL, NULL, max_juint, NULL, NULL,
3081 oldval /* pre_val */,
3082 T_OBJECT);
3083 break;
3084 }
3085 default:
3086 ShouldNotReachHere();
3087 }
3088
3089 #ifdef _LP64
3090 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3091 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3092
3093 switch(kind) {
3094 case LS_get_set:
3095 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3096 break;
3097 case LS_cmp_swap_weak: {
3098 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3099 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3100 break;
3101 }
3102 case LS_cmp_swap: {
3103 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3104 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3105 break;
3106 }
3107 case LS_cmp_exchange: {
3108 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3109 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3110 break;
3111 }
3112 default:
3113 ShouldNotReachHere();
3114 }
3115 } else
3116 #endif
3117 switch (kind) {
3118 case LS_get_set:
3119 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3120 break;
3121 case LS_cmp_swap_weak:
3122 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3123 break;
3124 case LS_cmp_swap:
3125 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3126 break;
3127 case LS_cmp_exchange:
3128 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3129 break;
3130 default:
3131 ShouldNotReachHere();
3132 }
3133
3134 // Emit the post barrier only when the actual store happened. This makes sense
3135 // to check only for LS_cmp_* that can fail to set the value.
3136 // LS_cmp_exchange does not produce any branches by default, so there is no
3137 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3138 // CompareAndExchange and move branches here, it would make sense to conditionalize
3139 // post_barriers for LS_cmp_exchange as well.
3140 //
3141 // CAS success path is marked more likely since we anticipate this is a performance
3142 // critical path, while CAS failure path can use the penalty for going through unlikely
3143 // path as backoff. Which is still better than doing a store barrier there.
3144 switch (kind) {
3145 case LS_get_set:
3146 case LS_cmp_exchange: {
3147 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3148 break;
3149 }
3150 case LS_cmp_swap_weak:
3151 case LS_cmp_swap: {
3152 IdealKit ideal(this);
3153 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3154 sync_kit(ideal);
3155 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3156 ideal.sync_kit(this);
3157 } ideal.end_if();
3158 final_sync(ideal);
3159 break;
3160 }
3161 default:
3162 ShouldNotReachHere();
3163 }
3164 break;
3165 default:
3166 fatal("unexpected type %d: %s", type, type2name(type));
3167 break;
3168 }
3169
3170 // SCMemProjNodes represent the memory state of a LoadStore. Their
3171 // main role is to prevent LoadStore nodes from being optimized away
3172 // when their results aren't used.
3173 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3174 set_memory(proj, alias_idx);
3175
3176 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3177 #ifdef _LP64
3178 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3179 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3180 }
3181 #endif
3182 if (can_move_pre_barrier() && kind == LS_get_set) {
3183 // Don't need to load pre_val. The old value is returned by load_store.
3184 // The pre_barrier can execute after the xchg as long as no safepoint
3185 // gets inserted between them.
3186 pre_barrier(false /* do_load */,
3187 control(), NULL, NULL, max_juint, NULL, NULL,
3188 load_store /* pre_val */,
3189 T_OBJECT);
3190 }
3191 }
3192
3193 // Add the trailing membar surrounding the access
3194 insert_mem_bar(Op_MemBarCPUOrder);
3195
3196 switch (access_kind) {
3197 case Relaxed:
3198 case Release:
3199 break; // do nothing
3200 case Acquire:
3201 case Volatile:
3202 insert_mem_bar(Op_MemBarAcquire);
3203 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3204 break;
3205 default:
3206 ShouldNotReachHere();
3207 }
3208
3209 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3210 set_result(load_store);
3211 return true;
3212 }
3213
3214 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3215 MemNode::MemOrd mo = MemNode::unset;
3216 switch(kind) {
3217 case Opaque:
3218 case Relaxed: mo = MemNode::unordered; break;
3219 case Acquire: mo = MemNode::acquire; break;
3220 case Release: mo = MemNode::release; break;
3221 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3222 default:
3223 ShouldNotReachHere();
3224 }
3225 guarantee(mo != MemNode::unset, "Should select memory ordering");
3226 return mo;
3227 }
3228
3229 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3230 MemNode::MemOrd mo = MemNode::unset;
3231 switch(kind) {
3232 case Opaque:
3233 case Relaxed: mo = MemNode::unordered; break;
3234 case Acquire: mo = MemNode::acquire; break;
3235 case Release: mo = MemNode::release; break;
3236 case Volatile: mo = MemNode::seqcst; break;
3237 default:
3238 ShouldNotReachHere();
3239 }
3240 guarantee(mo != MemNode::unset, "Should select memory ordering");
3241 return mo;
3242 }
3243
3244 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3245 // Regardless of form, don't allow previous ld/st to move down,
3246 // then issue acquire, release, or volatile mem_bar.
3247 insert_mem_bar(Op_MemBarCPUOrder);
3248 switch(id) {
3249 case vmIntrinsics::_loadFence:
3250 insert_mem_bar(Op_LoadFence);
3251 return true;
3252 case vmIntrinsics::_storeFence:
3253 insert_mem_bar(Op_StoreFence);
3254 return true;
3255 case vmIntrinsics::_fullFence:
3256 insert_mem_bar(Op_MemBarVolatile);
3257 return true;
3258 default:
3259 fatal_unexpected_iid(id);
3260 return false;
3261 }
3262 }
3263
3264 bool LibraryCallKit::inline_onspinwait() {
3265 insert_mem_bar(Op_OnSpinWait);
3266 return true;
3267 }
3268
3269 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3270 if (!kls->is_Con()) {
3271 return true;
3272 }
3273 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3274 if (klsptr == NULL) {
3275 return true;
3276 }
3277 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3278 // don't need a guard for a klass that is already initialized
3279 return !ik->is_initialized();
3280 }
3281
3282 //----------------------------inline_unsafe_allocate---------------------------
3283 // public native Object Unsafe.allocateInstance(Class<?> cls);
3284 bool LibraryCallKit::inline_unsafe_allocate() {
3285 if (callee()->is_static()) return false; // caller must have the capability!
3286
3287 null_check_receiver(); // null-check, then ignore
3288 Node* cls = null_check(argument(1));
3289 if (stopped()) return true;
3290
3291 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3292 kls = null_check(kls);
3293 if (stopped()) return true; // argument was like int.class
3294
3295 Node* test = NULL;
3296 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3297 // Note: The argument might still be an illegal value like
3298 // Serializable.class or Object[].class. The runtime will handle it.
3299 // But we must make an explicit check for initialization.
3300 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3301 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3302 // can generate code to load it as unsigned byte.
3303 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3304 Node* bits = intcon(InstanceKlass::fully_initialized);
3305 test = _gvn.transform(new SubINode(inst, bits));
3306 // The 'test' is non-zero if we need to take a slow path.
3307 }
3308
3309 Node* obj = new_instance(kls, test);
3310 set_result(obj);
3311 return true;
3312 }
3313
3314 //------------------------inline_native_time_funcs--------------
3315 // inline code for System.currentTimeMillis() and System.nanoTime()
3316 // these have the same type and signature
3317 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3318 const TypeFunc* tf = OptoRuntime::void_long_Type();
3319 const TypePtr* no_memory_effects = NULL;
3320 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3321 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3322 #ifdef ASSERT
3323 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3324 assert(value_top == top(), "second value must be top");
3325 #endif
3326 set_result(value);
3327 return true;
3328 }
3329
3330 #ifdef TRACE_HAVE_INTRINSICS
3331
3332 /*
3333 * oop -> myklass
3334 * myklass->trace_id |= USED
3335 * return myklass->trace_id & ~0x3
3336 */
3337 bool LibraryCallKit::inline_native_classID() {
3338 Node* cls = null_check(argument(0), T_OBJECT);
3339 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3340 kls = null_check(kls, T_OBJECT);
3341
3342 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3343 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3344 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3345
3346 Node* clsused = longcon(0x01l); // set the class bit
3347 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3348 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3349 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3350
3351 #ifdef TRACE_ID_META_BITS
3352 Node* mbits = longcon(~TRACE_ID_META_BITS);
3353 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3354 #endif
3355 #ifdef TRACE_ID_CLASS_SHIFT
3356 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3357 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3358 #endif
3359
3360 set_result(tvalue);
3361 return true;
3362
3363 }
3364
3365 bool LibraryCallKit::inline_native_getBufferWriter() {
3366 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3367
3368 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3369 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3370 );
3371
3372 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3373
3374 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3375 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3376
3377 IfNode* iff_jobj_null =
3378 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3379
3380 enum { _normal_path = 1,
3381 _null_path = 2,
3382 PATH_LIMIT };
3383
3384 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3385 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3386
3387 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3388 result_rgn->init_req(_null_path, jobj_is_null);
3389 result_val->init_req(_null_path, null());
3390
3391 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3392 result_rgn->init_req(_normal_path, jobj_is_not_null);
3393
3394 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3395 result_val->init_req(_normal_path, res);
3396
3397 set_result(result_rgn, result_val);
3398
3399 return true;
3400 }
3401
3402 #endif
3403
3404 //------------------------inline_native_currentThread------------------
3405 bool LibraryCallKit::inline_native_currentThread() {
3406 Node* junk = NULL;
3407 set_result(generate_current_thread(junk));
3408 return true;
3409 }
3410
3411 //------------------------inline_native_isInterrupted------------------
3412 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3413 bool LibraryCallKit::inline_native_isInterrupted() {
3414 // Add a fast path to t.isInterrupted(clear_int):
3415 // (t == Thread.current() &&
3416 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3417 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3418 // So, in the common case that the interrupt bit is false,
3419 // we avoid making a call into the VM. Even if the interrupt bit
3420 // is true, if the clear_int argument is false, we avoid the VM call.
3421 // However, if the receiver is not currentThread, we must call the VM,
3422 // because there must be some locking done around the operation.
3423
3424 // We only go to the fast case code if we pass two guards.
3425 // Paths which do not pass are accumulated in the slow_region.
3426
3427 enum {
3428 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3429 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3430 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3431 PATH_LIMIT
3432 };
3433
3434 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3435 // out of the function.
3436 insert_mem_bar(Op_MemBarCPUOrder);
3437
3438 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3439 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3440
3441 RegionNode* slow_region = new RegionNode(1);
3442 record_for_igvn(slow_region);
3443
3444 // (a) Receiving thread must be the current thread.
3445 Node* rec_thr = argument(0);
3446 Node* tls_ptr = NULL;
3447 Node* cur_thr = generate_current_thread(tls_ptr);
3448 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3449 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3450
3451 generate_slow_guard(bol_thr, slow_region);
3452
3453 // (b) Interrupt bit on TLS must be false.
3454 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3455 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3456 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3457
3458 // Set the control input on the field _interrupted read to prevent it floating up.
3459 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3460 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3461 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3462
3463 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3464
3465 // First fast path: if (!TLS._interrupted) return false;
3466 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3467 result_rgn->init_req(no_int_result_path, false_bit);
3468 result_val->init_req(no_int_result_path, intcon(0));
3469
3470 // drop through to next case
3471 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3472
3473 #ifndef _WINDOWS
3474 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3475 Node* clr_arg = argument(1);
3476 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3477 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3478 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3479
3480 // Second fast path: ... else if (!clear_int) return true;
3481 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3482 result_rgn->init_req(no_clear_result_path, false_arg);
3483 result_val->init_req(no_clear_result_path, intcon(1));
3484
3485 // drop through to next case
3486 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3487 #else
3488 // To return true on Windows you must read the _interrupted field
3489 // and check the event state i.e. take the slow path.
3490 #endif // _WINDOWS
3491
3492 // (d) Otherwise, go to the slow path.
3493 slow_region->add_req(control());
3494 set_control( _gvn.transform(slow_region));
3495
3496 if (stopped()) {
3497 // There is no slow path.
3498 result_rgn->init_req(slow_result_path, top());
3499 result_val->init_req(slow_result_path, top());
3500 } else {
3501 // non-virtual because it is a private non-static
3502 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3503
3504 Node* slow_val = set_results_for_java_call(slow_call);
3505 // this->control() comes from set_results_for_java_call
3506
3507 Node* fast_io = slow_call->in(TypeFunc::I_O);
3508 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3509
3510 // These two phis are pre-filled with copies of of the fast IO and Memory
3511 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3512 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3513
3514 result_rgn->init_req(slow_result_path, control());
3515 result_io ->init_req(slow_result_path, i_o());
3516 result_mem->init_req(slow_result_path, reset_memory());
3517 result_val->init_req(slow_result_path, slow_val);
3518
3519 set_all_memory(_gvn.transform(result_mem));
3520 set_i_o( _gvn.transform(result_io));
3521 }
3522
3523 C->set_has_split_ifs(true); // Has chance for split-if optimization
3524 set_result(result_rgn, result_val);
3525 return true;
3526 }
3527
3528 //---------------------------load_mirror_from_klass----------------------------
3529 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3530 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3531 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3532 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3533 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3534 }
3535
3536 //-----------------------load_klass_from_mirror_common-------------------------
3537 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3538 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3539 // and branch to the given path on the region.
3540 // If never_see_null, take an uncommon trap on null, so we can optimistically
3541 // compile for the non-null case.
3542 // If the region is NULL, force never_see_null = true.
3543 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3544 bool never_see_null,
3545 RegionNode* region,
3546 int null_path,
3547 int offset) {
3548 if (region == NULL) never_see_null = true;
3549 Node* p = basic_plus_adr(mirror, offset);
3550 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3551 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3552 Node* null_ctl = top();
3553 kls = null_check_oop(kls, &null_ctl, never_see_null);
3554 if (region != NULL) {
3555 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3556 region->init_req(null_path, null_ctl);
3557 } else {
3558 assert(null_ctl == top(), "no loose ends");
3559 }
3560 return kls;
3561 }
3562
3563 //--------------------(inline_native_Class_query helpers)---------------------
3564 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3565 // Fall through if (mods & mask) == bits, take the guard otherwise.
3566 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3567 // Branch around if the given klass has the given modifier bit set.
3568 // Like generate_guard, adds a new path onto the region.
3569 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3570 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3571 Node* mask = intcon(modifier_mask);
3572 Node* bits = intcon(modifier_bits);
3573 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3574 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3575 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3576 return generate_fair_guard(bol, region);
3577 }
3578 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3579 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3580 }
3581
3582 //-------------------------inline_native_Class_query-------------------
3583 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3584 const Type* return_type = TypeInt::BOOL;
3585 Node* prim_return_value = top(); // what happens if it's a primitive class?
3586 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3587 bool expect_prim = false; // most of these guys expect to work on refs
3588
3589 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3590
3591 Node* mirror = argument(0);
3592 Node* obj = top();
3593
3594 switch (id) {
3595 case vmIntrinsics::_isInstance:
3596 // nothing is an instance of a primitive type
3597 prim_return_value = intcon(0);
3598 obj = argument(1);
3599 break;
3600 case vmIntrinsics::_getModifiers:
3601 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3602 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3603 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3604 break;
3605 case vmIntrinsics::_isInterface:
3606 prim_return_value = intcon(0);
3607 break;
3608 case vmIntrinsics::_isArray:
3609 prim_return_value = intcon(0);
3610 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3611 break;
3612 case vmIntrinsics::_isPrimitive:
3613 prim_return_value = intcon(1);
3614 expect_prim = true; // obviously
3615 break;
3616 case vmIntrinsics::_getSuperclass:
3617 prim_return_value = null();
3618 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3619 break;
3620 case vmIntrinsics::_getClassAccessFlags:
3621 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3622 return_type = TypeInt::INT; // not bool! 6297094
3623 break;
3624 default:
3625 fatal_unexpected_iid(id);
3626 break;
3627 }
3628
3629 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3630 if (mirror_con == NULL) return false; // cannot happen?
3631
3632 #ifndef PRODUCT
3633 if (C->print_intrinsics() || C->print_inlining()) {
3634 ciType* k = mirror_con->java_mirror_type();
3635 if (k) {
3636 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3637 k->print_name();
3638 tty->cr();
3639 }
3640 }
3641 #endif
3642
3643 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3644 RegionNode* region = new RegionNode(PATH_LIMIT);
3645 record_for_igvn(region);
3646 PhiNode* phi = new PhiNode(region, return_type);
3647
3648 // The mirror will never be null of Reflection.getClassAccessFlags, however
3649 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3650 // if it is. See bug 4774291.
3651
3652 // For Reflection.getClassAccessFlags(), the null check occurs in
3653 // the wrong place; see inline_unsafe_access(), above, for a similar
3654 // situation.
3655 mirror = null_check(mirror);
3656 // If mirror or obj is dead, only null-path is taken.
3657 if (stopped()) return true;
3658
3659 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3660
3661 // Now load the mirror's klass metaobject, and null-check it.
3662 // Side-effects region with the control path if the klass is null.
3663 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3664 // If kls is null, we have a primitive mirror.
3665 phi->init_req(_prim_path, prim_return_value);
3666 if (stopped()) { set_result(region, phi); return true; }
3667 bool safe_for_replace = (region->in(_prim_path) == top());
3668
3669 Node* p; // handy temp
3670 Node* null_ctl;
3671
3672 // Now that we have the non-null klass, we can perform the real query.
3673 // For constant classes, the query will constant-fold in LoadNode::Value.
3674 Node* query_value = top();
3675 switch (id) {
3676 case vmIntrinsics::_isInstance:
3677 // nothing is an instance of a primitive type
3678 query_value = gen_instanceof(obj, kls, safe_for_replace);
3679 break;
3680
3681 case vmIntrinsics::_getModifiers:
3682 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3683 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3684 break;
3685
3686 case vmIntrinsics::_isInterface:
3687 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3688 if (generate_interface_guard(kls, region) != NULL)
3689 // A guard was added. If the guard is taken, it was an interface.
3690 phi->add_req(intcon(1));
3691 // If we fall through, it's a plain class.
3692 query_value = intcon(0);
3693 break;
3694
3695 case vmIntrinsics::_isArray:
3696 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3697 if (generate_array_guard(kls, region) != NULL)
3698 // A guard was added. If the guard is taken, it was an array.
3699 phi->add_req(intcon(1));
3700 // If we fall through, it's a plain class.
3701 query_value = intcon(0);
3702 break;
3703
3704 case vmIntrinsics::_isPrimitive:
3705 query_value = intcon(0); // "normal" path produces false
3706 break;
3707
3708 case vmIntrinsics::_getSuperclass:
3709 // The rules here are somewhat unfortunate, but we can still do better
3710 // with random logic than with a JNI call.
3711 // Interfaces store null or Object as _super, but must report null.
3712 // Arrays store an intermediate super as _super, but must report Object.
3713 // Other types can report the actual _super.
3714 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3715 if (generate_interface_guard(kls, region) != NULL)
3716 // A guard was added. If the guard is taken, it was an interface.
3717 phi->add_req(null());
3718 if (generate_array_guard(kls, region) != NULL)
3719 // A guard was added. If the guard is taken, it was an array.
3720 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3721 // If we fall through, it's a plain class. Get its _super.
3722 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3723 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3724 null_ctl = top();
3725 kls = null_check_oop(kls, &null_ctl);
3726 if (null_ctl != top()) {
3727 // If the guard is taken, Object.superClass is null (both klass and mirror).
3728 region->add_req(null_ctl);
3729 phi ->add_req(null());
3730 }
3731 if (!stopped()) {
3732 query_value = load_mirror_from_klass(kls);
3733 }
3734 break;
3735
3736 case vmIntrinsics::_getClassAccessFlags:
3737 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3738 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3739 break;
3740
3741 default:
3742 fatal_unexpected_iid(id);
3743 break;
3744 }
3745
3746 // Fall-through is the normal case of a query to a real class.
3747 phi->init_req(1, query_value);
3748 region->init_req(1, control());
3749
3750 C->set_has_split_ifs(true); // Has chance for split-if optimization
3751 set_result(region, phi);
3752 return true;
3753 }
3754
3755 //-------------------------inline_Class_cast-------------------
3756 bool LibraryCallKit::inline_Class_cast() {
3757 Node* mirror = argument(0); // Class
3758 Node* obj = argument(1);
3759 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3760 if (mirror_con == NULL) {
3761 return false; // dead path (mirror->is_top()).
3762 }
3763 if (obj == NULL || obj->is_top()) {
3764 return false; // dead path
3765 }
3766 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3767
3768 // First, see if Class.cast() can be folded statically.
3769 // java_mirror_type() returns non-null for compile-time Class constants.
3770 ciType* tm = mirror_con->java_mirror_type();
3771 if (tm != NULL && tm->is_klass() &&
3772 tp != NULL && tp->klass() != NULL) {
3773 if (!tp->klass()->is_loaded()) {
3774 // Don't use intrinsic when class is not loaded.
3775 return false;
3776 } else {
3777 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3778 if (static_res == Compile::SSC_always_true) {
3779 // isInstance() is true - fold the code.
3780 set_result(obj);
3781 return true;
3782 } else if (static_res == Compile::SSC_always_false) {
3783 // Don't use intrinsic, have to throw ClassCastException.
3784 // If the reference is null, the non-intrinsic bytecode will
3785 // be optimized appropriately.
3786 return false;
3787 }
3788 }
3789 }
3790
3791 // Bailout intrinsic and do normal inlining if exception path is frequent.
3792 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3793 return false;
3794 }
3795
3796 // Generate dynamic checks.
3797 // Class.cast() is java implementation of _checkcast bytecode.
3798 // Do checkcast (Parse::do_checkcast()) optimizations here.
3799
3800 mirror = null_check(mirror);
3801 // If mirror is dead, only null-path is taken.
3802 if (stopped()) {
3803 return true;
3804 }
3805
3806 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3807 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3808 RegionNode* region = new RegionNode(PATH_LIMIT);
3809 record_for_igvn(region);
3810
3811 // Now load the mirror's klass metaobject, and null-check it.
3812 // If kls is null, we have a primitive mirror and
3813 // nothing is an instance of a primitive type.
3814 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3815
3816 Node* res = top();
3817 if (!stopped()) {
3818 Node* bad_type_ctrl = top();
3819 // Do checkcast optimizations.
3820 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3821 region->init_req(_bad_type_path, bad_type_ctrl);
3822 }
3823 if (region->in(_prim_path) != top() ||
3824 region->in(_bad_type_path) != top()) {
3825 // Let Interpreter throw ClassCastException.
3826 PreserveJVMState pjvms(this);
3827 set_control(_gvn.transform(region));
3828 uncommon_trap(Deoptimization::Reason_intrinsic,
3829 Deoptimization::Action_maybe_recompile);
3830 }
3831 if (!stopped()) {
3832 set_result(res);
3833 }
3834 return true;
3835 }
3836
3837
3838 //--------------------------inline_native_subtype_check------------------------
3839 // This intrinsic takes the JNI calls out of the heart of
3840 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3841 bool LibraryCallKit::inline_native_subtype_check() {
3842 // Pull both arguments off the stack.
3843 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3844 args[0] = argument(0);
3845 args[1] = argument(1);
3846 Node* klasses[2]; // corresponding Klasses: superk, subk
3847 klasses[0] = klasses[1] = top();
3848
3849 enum {
3850 // A full decision tree on {superc is prim, subc is prim}:
3851 _prim_0_path = 1, // {P,N} => false
3852 // {P,P} & superc!=subc => false
3853 _prim_same_path, // {P,P} & superc==subc => true
3854 _prim_1_path, // {N,P} => false
3855 _ref_subtype_path, // {N,N} & subtype check wins => true
3856 _both_ref_path, // {N,N} & subtype check loses => false
3857 PATH_LIMIT
3858 };
3859
3860 RegionNode* region = new RegionNode(PATH_LIMIT);
3861 Node* phi = new PhiNode(region, TypeInt::BOOL);
3862 record_for_igvn(region);
3863
3864 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3865 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3866 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3867
3868 // First null-check both mirrors and load each mirror's klass metaobject.
3869 int which_arg;
3870 for (which_arg = 0; which_arg <= 1; which_arg++) {
3871 Node* arg = args[which_arg];
3872 arg = null_check(arg);
3873 if (stopped()) break;
3874 args[which_arg] = arg;
3875
3876 Node* p = basic_plus_adr(arg, class_klass_offset);
3877 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3878 klasses[which_arg] = _gvn.transform(kls);
3879 }
3880
3881 // Having loaded both klasses, test each for null.
3882 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3883 for (which_arg = 0; which_arg <= 1; which_arg++) {
3884 Node* kls = klasses[which_arg];
3885 Node* null_ctl = top();
3886 kls = null_check_oop(kls, &null_ctl, never_see_null);
3887 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3888 region->init_req(prim_path, null_ctl);
3889 if (stopped()) break;
3890 klasses[which_arg] = kls;
3891 }
3892
3893 if (!stopped()) {
3894 // now we have two reference types, in klasses[0..1]
3895 Node* subk = klasses[1]; // the argument to isAssignableFrom
3896 Node* superk = klasses[0]; // the receiver
3897 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3898 // now we have a successful reference subtype check
3899 region->set_req(_ref_subtype_path, control());
3900 }
3901
3902 // If both operands are primitive (both klasses null), then
3903 // we must return true when they are identical primitives.
3904 // It is convenient to test this after the first null klass check.
3905 set_control(region->in(_prim_0_path)); // go back to first null check
3906 if (!stopped()) {
3907 // Since superc is primitive, make a guard for the superc==subc case.
3908 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3909 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3910 generate_guard(bol_eq, region, PROB_FAIR);
3911 if (region->req() == PATH_LIMIT+1) {
3912 // A guard was added. If the added guard is taken, superc==subc.
3913 region->swap_edges(PATH_LIMIT, _prim_same_path);
3914 region->del_req(PATH_LIMIT);
3915 }
3916 region->set_req(_prim_0_path, control()); // Not equal after all.
3917 }
3918
3919 // these are the only paths that produce 'true':
3920 phi->set_req(_prim_same_path, intcon(1));
3921 phi->set_req(_ref_subtype_path, intcon(1));
3922
3923 // pull together the cases:
3924 assert(region->req() == PATH_LIMIT, "sane region");
3925 for (uint i = 1; i < region->req(); i++) {
3926 Node* ctl = region->in(i);
3927 if (ctl == NULL || ctl == top()) {
3928 region->set_req(i, top());
3929 phi ->set_req(i, top());
3930 } else if (phi->in(i) == NULL) {
3931 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3932 }
3933 }
3934
3935 set_control(_gvn.transform(region));
3936 set_result(_gvn.transform(phi));
3937 return true;
3938 }
3939
3940 //---------------------generate_array_guard_common------------------------
3941 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3942 bool obj_array, bool not_array) {
3943
3944 if (stopped()) {
3945 return NULL;
3946 }
3947
3948 // If obj_array/non_array==false/false:
3949 // Branch around if the given klass is in fact an array (either obj or prim).
3950 // If obj_array/non_array==false/true:
3951 // Branch around if the given klass is not an array klass of any kind.
3952 // If obj_array/non_array==true/true:
3953 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3954 // If obj_array/non_array==true/false:
3955 // Branch around if the kls is an oop array (Object[] or subtype)
3956 //
3957 // Like generate_guard, adds a new path onto the region.
3958 jint layout_con = 0;
3959 Node* layout_val = get_layout_helper(kls, layout_con);
3960 if (layout_val == NULL) {
3961 bool query = (obj_array
3962 ? Klass::layout_helper_is_objArray(layout_con)
3963 : Klass::layout_helper_is_array(layout_con));
3964 if (query == not_array) {
3965 return NULL; // never a branch
3966 } else { // always a branch
3967 Node* always_branch = control();
3968 if (region != NULL)
3969 region->add_req(always_branch);
3970 set_control(top());
3971 return always_branch;
3972 }
3973 }
3974 // Now test the correct condition.
3975 jint nval = (obj_array
3976 ? (jint)(Klass::_lh_array_tag_type_value
3977 << Klass::_lh_array_tag_shift)
3978 : Klass::_lh_neutral_value);
3979 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3980 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3981 // invert the test if we are looking for a non-array
3982 if (not_array) btest = BoolTest(btest).negate();
3983 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3984 return generate_fair_guard(bol, region);
3985 }
3986
3987
3988 //-----------------------inline_native_newArray--------------------------
3989 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3990 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3991 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3992 Node* mirror;
3993 Node* count_val;
3994 if (uninitialized) {
3995 mirror = argument(1);
3996 count_val = argument(2);
3997 } else {
3998 mirror = argument(0);
3999 count_val = argument(1);
4000 }
4001
4002 mirror = null_check(mirror);
4003 // If mirror or obj is dead, only null-path is taken.
4004 if (stopped()) return true;
4005
4006 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
4007 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4008 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4009 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4010 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4011
4012 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
4013 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
4014 result_reg, _slow_path);
4015 Node* normal_ctl = control();
4016 Node* no_array_ctl = result_reg->in(_slow_path);
4017
4018 // Generate code for the slow case. We make a call to newArray().
4019 set_control(no_array_ctl);
4020 if (!stopped()) {
4021 // Either the input type is void.class, or else the
4022 // array klass has not yet been cached. Either the
4023 // ensuing call will throw an exception, or else it
4024 // will cache the array klass for next time.
4025 PreserveJVMState pjvms(this);
4026 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
4027 Node* slow_result = set_results_for_java_call(slow_call);
4028 // this->control() comes from set_results_for_java_call
4029 result_reg->set_req(_slow_path, control());
4030 result_val->set_req(_slow_path, slow_result);
4031 result_io ->set_req(_slow_path, i_o());
4032 result_mem->set_req(_slow_path, reset_memory());
4033 }
4034
4035 set_control(normal_ctl);
4036 if (!stopped()) {
4037 // Normal case: The array type has been cached in the java.lang.Class.
4038 // The following call works fine even if the array type is polymorphic.
4039 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4040 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
4041 result_reg->init_req(_normal_path, control());
4042 result_val->init_req(_normal_path, obj);
4043 result_io ->init_req(_normal_path, i_o());
4044 result_mem->init_req(_normal_path, reset_memory());
4045
4046 if (uninitialized) {
4047 // Mark the allocation so that zeroing is skipped
4048 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4049 alloc->maybe_set_complete(&_gvn);
4050 }
4051 }
4052
4053 // Return the combined state.
4054 set_i_o( _gvn.transform(result_io) );
4055 set_all_memory( _gvn.transform(result_mem));
4056
4057 C->set_has_split_ifs(true); // Has chance for split-if optimization
4058 set_result(result_reg, result_val);
4059 return true;
4060 }
4061
4062 //----------------------inline_native_getLength--------------------------
4063 // public static native int java.lang.reflect.Array.getLength(Object array);
4064 bool LibraryCallKit::inline_native_getLength() {
4065 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4066
4067 Node* array = null_check(argument(0));
4068 // If array is dead, only null-path is taken.
4069 if (stopped()) return true;
4070
4071 // Deoptimize if it is a non-array.
4072 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4073
4074 if (non_array != NULL) {
4075 PreserveJVMState pjvms(this);
4076 set_control(non_array);
4077 uncommon_trap(Deoptimization::Reason_intrinsic,
4078 Deoptimization::Action_maybe_recompile);
4079 }
4080
4081 // If control is dead, only non-array-path is taken.
4082 if (stopped()) return true;
4083
4084 // The works fine even if the array type is polymorphic.
4085 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4086 Node* result = load_array_length(array);
4087
4088 C->set_has_split_ifs(true); // Has chance for split-if optimization
4089 set_result(result);
4090 return true;
4091 }
4092
4093 //------------------------inline_array_copyOf----------------------------
4094 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4095 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4096 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4097 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4098
4099 // Get the arguments.
4100 Node* original = argument(0);
4101 Node* start = is_copyOfRange? argument(1): intcon(0);
4102 Node* end = is_copyOfRange? argument(2): argument(1);
4103 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4104
4105 Node* newcopy = NULL;
4106
4107 // Set the original stack and the reexecute bit for the interpreter to reexecute
4108 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4109 { PreserveReexecuteState preexecs(this);
4110 jvms()->set_should_reexecute(true);
4111
4112 array_type_mirror = null_check(array_type_mirror);
4113 original = null_check(original);
4114
4115 // Check if a null path was taken unconditionally.
4116 if (stopped()) return true;
4117
4118 Node* orig_length = load_array_length(original);
4119
4120 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4121 klass_node = null_check(klass_node);
4122
4123 RegionNode* bailout = new RegionNode(1);
4124 record_for_igvn(bailout);
4125
4126 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4127 // Bail out if that is so.
4128 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4129 if (not_objArray != NULL) {
4130 // Improve the klass node's type from the new optimistic assumption:
4131 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4132 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4133 Node* cast = new CastPPNode(klass_node, akls);
4134 cast->init_req(0, control());
4135 klass_node = _gvn.transform(cast);
4136 }
4137
4138 // Bail out if either start or end is negative.
4139 generate_negative_guard(start, bailout, &start);
4140 generate_negative_guard(end, bailout, &end);
4141
4142 Node* length = end;
4143 if (_gvn.type(start) != TypeInt::ZERO) {
4144 length = _gvn.transform(new SubINode(end, start));
4145 }
4146
4147 // Bail out if length is negative.
4148 // Without this the new_array would throw
4149 // NegativeArraySizeException but IllegalArgumentException is what
4150 // should be thrown
4151 generate_negative_guard(length, bailout, &length);
4152
4153 if (bailout->req() > 1) {
4154 PreserveJVMState pjvms(this);
4155 set_control(_gvn.transform(bailout));
4156 uncommon_trap(Deoptimization::Reason_intrinsic,
4157 Deoptimization::Action_maybe_recompile);
4158 }
4159
4160 if (!stopped()) {
4161 // How many elements will we copy from the original?
4162 // The answer is MinI(orig_length - start, length).
4163 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4164 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4165
4166 // Generate a direct call to the right arraycopy function(s).
4167 // We know the copy is disjoint but we might not know if the
4168 // oop stores need checking.
4169 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4170 // This will fail a store-check if x contains any non-nulls.
4171
4172 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4173 // loads/stores but it is legal only if we're sure the
4174 // Arrays.copyOf would succeed. So we need all input arguments
4175 // to the copyOf to be validated, including that the copy to the
4176 // new array won't trigger an ArrayStoreException. That subtype
4177 // check can be optimized if we know something on the type of
4178 // the input array from type speculation.
4179 if (_gvn.type(klass_node)->singleton()) {
4180 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4181 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4182
4183 int test = C->static_subtype_check(superk, subk);
4184 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4185 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4186 if (t_original->speculative_type() != NULL) {
4187 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4188 }
4189 }
4190 }
4191
4192 bool validated = false;
4193 // Reason_class_check rather than Reason_intrinsic because we
4194 // want to intrinsify even if this traps.
4195 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4196 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4197 klass_node);
4198
4199 if (not_subtype_ctrl != top()) {
4200 PreserveJVMState pjvms(this);
4201 set_control(not_subtype_ctrl);
4202 uncommon_trap(Deoptimization::Reason_class_check,
4203 Deoptimization::Action_make_not_entrant);
4204 assert(stopped(), "Should be stopped");
4205 }
4206 validated = true;
4207 }
4208
4209 if (!stopped()) {
4210 newcopy = new_array(klass_node, length, 0); // no arguments to push
4211
4212 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4213 load_object_klass(original), klass_node);
4214 if (!is_copyOfRange) {
4215 ac->set_copyof(validated);
4216 } else {
4217 ac->set_copyofrange(validated);
4218 }
4219 Node* n = _gvn.transform(ac);
4220 if (n == ac) {
4221 ac->connect_outputs(this);
4222 } else {
4223 assert(validated, "shouldn't transform if all arguments not validated");
4224 set_all_memory(n);
4225 }
4226 }
4227 }
4228 } // original reexecute is set back here
4229
4230 C->set_has_split_ifs(true); // Has chance for split-if optimization
4231 if (!stopped()) {
4232 set_result(newcopy);
4233 }
4234 return true;
4235 }
4236
4237
4238 //----------------------generate_virtual_guard---------------------------
4239 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4240 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4241 RegionNode* slow_region) {
4242 ciMethod* method = callee();
4243 int vtable_index = method->vtable_index();
4244 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4245 "bad index %d", vtable_index);
4246 // Get the Method* out of the appropriate vtable entry.
4247 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4248 vtable_index*vtableEntry::size_in_bytes() +
4249 vtableEntry::method_offset_in_bytes();
4250 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4251 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4252
4253 // Compare the target method with the expected method (e.g., Object.hashCode).
4254 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4255
4256 Node* native_call = makecon(native_call_addr);
4257 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4258 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4259
4260 return generate_slow_guard(test_native, slow_region);
4261 }
4262
4263 //-----------------------generate_method_call----------------------------
4264 // Use generate_method_call to make a slow-call to the real
4265 // method if the fast path fails. An alternative would be to
4266 // use a stub like OptoRuntime::slow_arraycopy_Java.
4267 // This only works for expanding the current library call,
4268 // not another intrinsic. (E.g., don't use this for making an
4269 // arraycopy call inside of the copyOf intrinsic.)
4270 CallJavaNode*
4271 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4272 // When compiling the intrinsic method itself, do not use this technique.
4273 guarantee(callee() != C->method(), "cannot make slow-call to self");
4274
4275 ciMethod* method = callee();
4276 // ensure the JVMS we have will be correct for this call
4277 guarantee(method_id == method->intrinsic_id(), "must match");
4278
4279 const TypeFunc* tf = TypeFunc::make(method);
4280 CallJavaNode* slow_call;
4281 if (is_static) {
4282 assert(!is_virtual, "");
4283 slow_call = new CallStaticJavaNode(C, tf,
4284 SharedRuntime::get_resolve_static_call_stub(),
4285 method, bci());
4286 } else if (is_virtual) {
4287 null_check_receiver();
4288 int vtable_index = Method::invalid_vtable_index;
4289 if (UseInlineCaches) {
4290 // Suppress the vtable call
4291 } else {
4292 // hashCode and clone are not a miranda methods,
4293 // so the vtable index is fixed.
4294 // No need to use the linkResolver to get it.
4295 vtable_index = method->vtable_index();
4296 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4297 "bad index %d", vtable_index);
4298 }
4299 slow_call = new CallDynamicJavaNode(tf,
4300 SharedRuntime::get_resolve_virtual_call_stub(),
4301 method, vtable_index, bci());
4302 } else { // neither virtual nor static: opt_virtual
4303 null_check_receiver();
4304 slow_call = new CallStaticJavaNode(C, tf,
4305 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4306 method, bci());
4307 slow_call->set_optimized_virtual(true);
4308 }
4309 set_arguments_for_java_call(slow_call);
4310 set_edges_for_java_call(slow_call);
4311 return slow_call;
4312 }
4313
4314
4315 /**
4316 * Build special case code for calls to hashCode on an object. This call may
4317 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4318 * slightly different code.
4319 */
4320 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4321 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4322 assert(!(is_virtual && is_static), "either virtual, special, or static");
4323
4324 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4325
4326 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4327 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4328 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4329 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4330 Node* obj = NULL;
4331 if (!is_static) {
4332 // Check for hashing null object
4333 obj = null_check_receiver();
4334 if (stopped()) return true; // unconditionally null
4335 result_reg->init_req(_null_path, top());
4336 result_val->init_req(_null_path, top());
4337 } else {
4338 // Do a null check, and return zero if null.
4339 // System.identityHashCode(null) == 0
4340 obj = argument(0);
4341 Node* null_ctl = top();
4342 obj = null_check_oop(obj, &null_ctl);
4343 result_reg->init_req(_null_path, null_ctl);
4344 result_val->init_req(_null_path, _gvn.intcon(0));
4345 }
4346
4347 // Unconditionally null? Then return right away.
4348 if (stopped()) {
4349 set_control( result_reg->in(_null_path));
4350 if (!stopped())
4351 set_result(result_val->in(_null_path));
4352 return true;
4353 }
4354
4355 // We only go to the fast case code if we pass a number of guards. The
4356 // paths which do not pass are accumulated in the slow_region.
4357 RegionNode* slow_region = new RegionNode(1);
4358 record_for_igvn(slow_region);
4359
4360 // If this is a virtual call, we generate a funny guard. We pull out
4361 // the vtable entry corresponding to hashCode() from the target object.
4362 // If the target method which we are calling happens to be the native
4363 // Object hashCode() method, we pass the guard. We do not need this
4364 // guard for non-virtual calls -- the caller is known to be the native
4365 // Object hashCode().
4366 if (is_virtual) {
4367 // After null check, get the object's klass.
4368 Node* obj_klass = load_object_klass(obj);
4369 generate_virtual_guard(obj_klass, slow_region);
4370 }
4371
4372 // Get the header out of the object, use LoadMarkNode when available
4373 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4374 // The control of the load must be NULL. Otherwise, the load can move before
4375 // the null check after castPP removal.
4376 Node* no_ctrl = NULL;
4377 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4378
4379 // Test the header to see if it is unlocked.
4380 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4381 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4382 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4383 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4384 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4385
4386 generate_slow_guard(test_unlocked, slow_region);
4387
4388 // Get the hash value and check to see that it has been properly assigned.
4389 // We depend on hash_mask being at most 32 bits and avoid the use of
4390 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4391 // vm: see markOop.hpp.
4392 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4393 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4394 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4395 // This hack lets the hash bits live anywhere in the mark object now, as long
4396 // as the shift drops the relevant bits into the low 32 bits. Note that
4397 // Java spec says that HashCode is an int so there's no point in capturing
4398 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4399 hshifted_header = ConvX2I(hshifted_header);
4400 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4401
4402 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4403 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4404 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4405
4406 generate_slow_guard(test_assigned, slow_region);
4407
4408 Node* init_mem = reset_memory();
4409 // fill in the rest of the null path:
4410 result_io ->init_req(_null_path, i_o());
4411 result_mem->init_req(_null_path, init_mem);
4412
4413 result_val->init_req(_fast_path, hash_val);
4414 result_reg->init_req(_fast_path, control());
4415 result_io ->init_req(_fast_path, i_o());
4416 result_mem->init_req(_fast_path, init_mem);
4417
4418 // Generate code for the slow case. We make a call to hashCode().
4419 set_control(_gvn.transform(slow_region));
4420 if (!stopped()) {
4421 // No need for PreserveJVMState, because we're using up the present state.
4422 set_all_memory(init_mem);
4423 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4424 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4425 Node* slow_result = set_results_for_java_call(slow_call);
4426 // this->control() comes from set_results_for_java_call
4427 result_reg->init_req(_slow_path, control());
4428 result_val->init_req(_slow_path, slow_result);
4429 result_io ->set_req(_slow_path, i_o());
4430 result_mem ->set_req(_slow_path, reset_memory());
4431 }
4432
4433 // Return the combined state.
4434 set_i_o( _gvn.transform(result_io) );
4435 set_all_memory( _gvn.transform(result_mem));
4436
4437 set_result(result_reg, result_val);
4438 return true;
4439 }
4440
4441 //---------------------------inline_native_getClass----------------------------
4442 // public final native Class<?> java.lang.Object.getClass();
4443 //
4444 // Build special case code for calls to getClass on an object.
4445 bool LibraryCallKit::inline_native_getClass() {
4446 Node* obj = null_check_receiver();
4447 if (stopped()) return true;
4448 set_result(load_mirror_from_klass(load_object_klass(obj)));
4449 return true;
4450 }
4451
4452 //-----------------inline_native_Reflection_getCallerClass---------------------
4453 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4454 //
4455 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4456 //
4457 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4458 // in that it must skip particular security frames and checks for
4459 // caller sensitive methods.
4460 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4461 #ifndef PRODUCT
4462 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4463 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4464 }
4465 #endif
4466
4467 if (!jvms()->has_method()) {
4468 #ifndef PRODUCT
4469 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4470 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4471 }
4472 #endif
4473 return false;
4474 }
4475
4476 // Walk back up the JVM state to find the caller at the required
4477 // depth.
4478 JVMState* caller_jvms = jvms();
4479
4480 // Cf. JVM_GetCallerClass
4481 // NOTE: Start the loop at depth 1 because the current JVM state does
4482 // not include the Reflection.getCallerClass() frame.
4483 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4484 ciMethod* m = caller_jvms->method();
4485 switch (n) {
4486 case 0:
4487 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4488 break;
4489 case 1:
4490 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4491 if (!m->caller_sensitive()) {
4492 #ifndef PRODUCT
4493 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4494 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4495 }
4496 #endif
4497 return false; // bail-out; let JVM_GetCallerClass do the work
4498 }
4499 break;
4500 default:
4501 if (!m->is_ignored_by_security_stack_walk()) {
4502 // We have reached the desired frame; return the holder class.
4503 // Acquire method holder as java.lang.Class and push as constant.
4504 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4505 ciInstance* caller_mirror = caller_klass->java_mirror();
4506 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4507
4508 #ifndef PRODUCT
4509 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4510 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());
4511 tty->print_cr(" JVM state at this point:");
4512 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4513 ciMethod* m = jvms()->of_depth(i)->method();
4514 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4515 }
4516 }
4517 #endif
4518 return true;
4519 }
4520 break;
4521 }
4522 }
4523
4524 #ifndef PRODUCT
4525 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4526 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4527 tty->print_cr(" JVM state at this point:");
4528 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4529 ciMethod* m = jvms()->of_depth(i)->method();
4530 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4531 }
4532 }
4533 #endif
4534
4535 return false; // bail-out; let JVM_GetCallerClass do the work
4536 }
4537
4538 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4539 Node* arg = argument(0);
4540 Node* result = NULL;
4541
4542 switch (id) {
4543 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4544 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4545 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4546 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4547
4548 case vmIntrinsics::_doubleToLongBits: {
4549 // two paths (plus control) merge in a wood
4550 RegionNode *r = new RegionNode(3);
4551 Node *phi = new PhiNode(r, TypeLong::LONG);
4552
4553 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4554 // Build the boolean node
4555 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4556
4557 // Branch either way.
4558 // NaN case is less traveled, which makes all the difference.
4559 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4560 Node *opt_isnan = _gvn.transform(ifisnan);
4561 assert( opt_isnan->is_If(), "Expect an IfNode");
4562 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4563 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4564
4565 set_control(iftrue);
4566
4567 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4568 Node *slow_result = longcon(nan_bits); // return NaN
4569 phi->init_req(1, _gvn.transform( slow_result ));
4570 r->init_req(1, iftrue);
4571
4572 // Else fall through
4573 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4574 set_control(iffalse);
4575
4576 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4577 r->init_req(2, iffalse);
4578
4579 // Post merge
4580 set_control(_gvn.transform(r));
4581 record_for_igvn(r);
4582
4583 C->set_has_split_ifs(true); // Has chance for split-if optimization
4584 result = phi;
4585 assert(result->bottom_type()->isa_long(), "must be");
4586 break;
4587 }
4588
4589 case vmIntrinsics::_floatToIntBits: {
4590 // two paths (plus control) merge in a wood
4591 RegionNode *r = new RegionNode(3);
4592 Node *phi = new PhiNode(r, TypeInt::INT);
4593
4594 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4595 // Build the boolean node
4596 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4597
4598 // Branch either way.
4599 // NaN case is less traveled, which makes all the difference.
4600 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4601 Node *opt_isnan = _gvn.transform(ifisnan);
4602 assert( opt_isnan->is_If(), "Expect an IfNode");
4603 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4604 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4605
4606 set_control(iftrue);
4607
4608 static const jint nan_bits = 0x7fc00000;
4609 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4610 phi->init_req(1, _gvn.transform( slow_result ));
4611 r->init_req(1, iftrue);
4612
4613 // Else fall through
4614 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4615 set_control(iffalse);
4616
4617 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4618 r->init_req(2, iffalse);
4619
4620 // Post merge
4621 set_control(_gvn.transform(r));
4622 record_for_igvn(r);
4623
4624 C->set_has_split_ifs(true); // Has chance for split-if optimization
4625 result = phi;
4626 assert(result->bottom_type()->isa_int(), "must be");
4627 break;
4628 }
4629
4630 default:
4631 fatal_unexpected_iid(id);
4632 break;
4633 }
4634 set_result(_gvn.transform(result));
4635 return true;
4636 }
4637
4638 //----------------------inline_unsafe_copyMemory-------------------------
4639 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4640 bool LibraryCallKit::inline_unsafe_copyMemory() {
4641 if (callee()->is_static()) return false; // caller must have the capability!
4642 null_check_receiver(); // null-check receiver
4643 if (stopped()) return true;
4644
4645 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4646
4647 Node* src_ptr = argument(1); // type: oop
4648 Node* src_off = ConvL2X(argument(2)); // type: long
4649 Node* dst_ptr = argument(4); // type: oop
4650 Node* dst_off = ConvL2X(argument(5)); // type: long
4651 Node* size = ConvL2X(argument(7)); // type: long
4652
4653 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4654 "fieldOffset must be byte-scaled");
4655
4656 Node* src = make_unsafe_address(src_ptr, src_off);
4657 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4658
4659 // Conservatively insert a memory barrier on all memory slices.
4660 // Do not let writes of the copy source or destination float below the copy.
4661 insert_mem_bar(Op_MemBarCPUOrder);
4662
4663 // Call it. Note that the length argument is not scaled.
4664 make_runtime_call(RC_LEAF|RC_NO_FP,
4665 OptoRuntime::fast_arraycopy_Type(),
4666 StubRoutines::unsafe_arraycopy(),
4667 "unsafe_arraycopy",
4668 TypeRawPtr::BOTTOM,
4669 src, dst, size XTOP);
4670
4671 // Do not let reads of the copy destination float above the copy.
4672 insert_mem_bar(Op_MemBarCPUOrder);
4673
4674 return true;
4675 }
4676
4677 //------------------------clone_coping-----------------------------------
4678 // Helper function for inline_native_clone.
4679 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4680 assert(obj_size != NULL, "");
4681 Node* raw_obj = alloc_obj->in(1);
4682 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4683
4684 AllocateNode* alloc = NULL;
4685 if (ReduceBulkZeroing) {
4686 // We will be completely responsible for initializing this object -
4687 // mark Initialize node as complete.
4688 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4689 // The object was just allocated - there should be no any stores!
4690 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4691 // Mark as complete_with_arraycopy so that on AllocateNode
4692 // expansion, we know this AllocateNode is initialized by an array
4693 // copy and a StoreStore barrier exists after the array copy.
4694 alloc->initialization()->set_complete_with_arraycopy();
4695 }
4696
4697 // Copy the fastest available way.
4698 // TODO: generate fields copies for small objects instead.
4699 Node* src = obj;
4700 Node* dest = alloc_obj;
4701 Node* size = _gvn.transform(obj_size);
4702
4703 // Exclude the header but include array length to copy by 8 bytes words.
4704 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4705 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4706 instanceOopDesc::base_offset_in_bytes();
4707 // base_off:
4708 // 8 - 32-bit VM
4709 // 12 - 64-bit VM, compressed klass
4710 // 16 - 64-bit VM, normal klass
4711 if (base_off % BytesPerLong != 0) {
4712 assert(UseCompressedClassPointers, "");
4713 if (is_array) {
4714 // Exclude length to copy by 8 bytes words.
4715 base_off += sizeof(int);
4716 } else {
4717 // Include klass to copy by 8 bytes words.
4718 base_off = instanceOopDesc::klass_offset_in_bytes();
4719 }
4720 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4721 }
4722 src = basic_plus_adr(src, base_off);
4723 dest = basic_plus_adr(dest, base_off);
4724
4725 // Compute the length also, if needed:
4726 Node* countx = size;
4727 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4728 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4729
4730 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4731
4732 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4733 ac->set_clonebasic();
4734 Node* n = _gvn.transform(ac);
4735 if (n == ac) {
4736 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4737 } else {
4738 set_all_memory(n);
4739 }
4740
4741 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4742 if (card_mark) {
4743 assert(!is_array, "");
4744 // Put in store barrier for any and all oops we are sticking
4745 // into this object. (We could avoid this if we could prove
4746 // that the object type contains no oop fields at all.)
4747 Node* no_particular_value = NULL;
4748 Node* no_particular_field = NULL;
4749 int raw_adr_idx = Compile::AliasIdxRaw;
4750 post_barrier(control(),
4751 memory(raw_adr_type),
4752 alloc_obj,
4753 no_particular_field,
4754 raw_adr_idx,
4755 no_particular_value,
4756 T_OBJECT,
4757 false);
4758 }
4759
4760 // Do not let reads from the cloned object float above the arraycopy.
4761 if (alloc != NULL) {
4762 // Do not let stores that initialize this object be reordered with
4763 // a subsequent store that would make this object accessible by
4764 // other threads.
4765 // Record what AllocateNode this StoreStore protects so that
4766 // escape analysis can go from the MemBarStoreStoreNode to the
4767 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4768 // based on the escape status of the AllocateNode.
4769 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4770 } else {
4771 insert_mem_bar(Op_MemBarCPUOrder);
4772 }
4773 }
4774
4775 //------------------------inline_native_clone----------------------------
4776 // protected native Object java.lang.Object.clone();
4777 //
4778 // Here are the simple edge cases:
4779 // null receiver => normal trap
4780 // virtual and clone was overridden => slow path to out-of-line clone
4781 // not cloneable or finalizer => slow path to out-of-line Object.clone
4782 //
4783 // The general case has two steps, allocation and copying.
4784 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4785 //
4786 // Copying also has two cases, oop arrays and everything else.
4787 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4788 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4789 //
4790 // These steps fold up nicely if and when the cloned object's klass
4791 // can be sharply typed as an object array, a type array, or an instance.
4792 //
4793 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4794 PhiNode* result_val;
4795
4796 // Set the reexecute bit for the interpreter to reexecute
4797 // the bytecode that invokes Object.clone if deoptimization happens.
4798 { PreserveReexecuteState preexecs(this);
4799 jvms()->set_should_reexecute(true);
4800
4801 Node* obj = null_check_receiver();
4802 if (stopped()) return true;
4803
4804 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4805
4806 // If we are going to clone an instance, we need its exact type to
4807 // know the number and types of fields to convert the clone to
4808 // loads/stores. Maybe a speculative type can help us.
4809 if (!obj_type->klass_is_exact() &&
4810 obj_type->speculative_type() != NULL &&
4811 obj_type->speculative_type()->is_instance_klass()) {
4812 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4813 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4814 !spec_ik->has_injected_fields()) {
4815 ciKlass* k = obj_type->klass();
4816 if (!k->is_instance_klass() ||
4817 k->as_instance_klass()->is_interface() ||
4818 k->as_instance_klass()->has_subklass()) {
4819 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4820 }
4821 }
4822 }
4823
4824 Node* obj_klass = load_object_klass(obj);
4825 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4826 const TypeOopPtr* toop = ((tklass != NULL)
4827 ? tklass->as_instance_type()
4828 : TypeInstPtr::NOTNULL);
4829
4830 // Conservatively insert a memory barrier on all memory slices.
4831 // Do not let writes into the original float below the clone.
4832 insert_mem_bar(Op_MemBarCPUOrder);
4833
4834 // paths into result_reg:
4835 enum {
4836 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4837 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4838 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4839 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4840 PATH_LIMIT
4841 };
4842 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4843 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4844 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4845 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4846 record_for_igvn(result_reg);
4847
4848 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4849 int raw_adr_idx = Compile::AliasIdxRaw;
4850
4851 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4852 if (array_ctl != NULL) {
4853 // It's an array.
4854 PreserveJVMState pjvms(this);
4855 set_control(array_ctl);
4856 Node* obj_length = load_array_length(obj);
4857 Node* obj_size = NULL;
4858 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4859
4860 if (!use_ReduceInitialCardMarks()) {
4861 // If it is an oop array, it requires very special treatment,
4862 // because card marking is required on each card of the array.
4863 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4864 if (is_obja != NULL) {
4865 PreserveJVMState pjvms2(this);
4866 set_control(is_obja);
4867 // Generate a direct call to the right arraycopy function(s).
4868 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4869 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4870 ac->set_cloneoop();
4871 Node* n = _gvn.transform(ac);
4872 assert(n == ac, "cannot disappear");
4873 ac->connect_outputs(this);
4874
4875 result_reg->init_req(_objArray_path, control());
4876 result_val->init_req(_objArray_path, alloc_obj);
4877 result_i_o ->set_req(_objArray_path, i_o());
4878 result_mem ->set_req(_objArray_path, reset_memory());
4879 }
4880 }
4881 // Otherwise, there are no card marks to worry about.
4882 // (We can dispense with card marks if we know the allocation
4883 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4884 // causes the non-eden paths to take compensating steps to
4885 // simulate a fresh allocation, so that no further
4886 // card marks are required in compiled code to initialize
4887 // the object.)
4888
4889 if (!stopped()) {
4890 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4891
4892 // Present the results of the copy.
4893 result_reg->init_req(_array_path, control());
4894 result_val->init_req(_array_path, alloc_obj);
4895 result_i_o ->set_req(_array_path, i_o());
4896 result_mem ->set_req(_array_path, reset_memory());
4897 }
4898 }
4899
4900 // We only go to the instance fast case code if we pass a number of guards.
4901 // The paths which do not pass are accumulated in the slow_region.
4902 RegionNode* slow_region = new RegionNode(1);
4903 record_for_igvn(slow_region);
4904 if (!stopped()) {
4905 // It's an instance (we did array above). Make the slow-path tests.
4906 // If this is a virtual call, we generate a funny guard. We grab
4907 // the vtable entry corresponding to clone() from the target object.
4908 // If the target method which we are calling happens to be the
4909 // Object clone() method, we pass the guard. We do not need this
4910 // guard for non-virtual calls; the caller is known to be the native
4911 // Object clone().
4912 if (is_virtual) {
4913 generate_virtual_guard(obj_klass, slow_region);
4914 }
4915
4916 // The object must be easily cloneable and must not have a finalizer.
4917 // Both of these conditions may be checked in a single test.
4918 // We could optimize the test further, but we don't care.
4919 generate_access_flags_guard(obj_klass,
4920 // Test both conditions:
4921 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4922 // Must be cloneable but not finalizer:
4923 JVM_ACC_IS_CLONEABLE_FAST,
4924 slow_region);
4925 }
4926
4927 if (!stopped()) {
4928 // It's an instance, and it passed the slow-path tests.
4929 PreserveJVMState pjvms(this);
4930 Node* obj_size = NULL;
4931 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4932 // is reexecuted if deoptimization occurs and there could be problems when merging
4933 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4934 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4935
4936 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4937
4938 // Present the results of the slow call.
4939 result_reg->init_req(_instance_path, control());
4940 result_val->init_req(_instance_path, alloc_obj);
4941 result_i_o ->set_req(_instance_path, i_o());
4942 result_mem ->set_req(_instance_path, reset_memory());
4943 }
4944
4945 // Generate code for the slow case. We make a call to clone().
4946 set_control(_gvn.transform(slow_region));
4947 if (!stopped()) {
4948 PreserveJVMState pjvms(this);
4949 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4950 Node* slow_result = set_results_for_java_call(slow_call);
4951 // this->control() comes from set_results_for_java_call
4952 result_reg->init_req(_slow_path, control());
4953 result_val->init_req(_slow_path, slow_result);
4954 result_i_o ->set_req(_slow_path, i_o());
4955 result_mem ->set_req(_slow_path, reset_memory());
4956 }
4957
4958 // Return the combined state.
4959 set_control( _gvn.transform(result_reg));
4960 set_i_o( _gvn.transform(result_i_o));
4961 set_all_memory( _gvn.transform(result_mem));
4962 } // original reexecute is set back here
4963
4964 set_result(_gvn.transform(result_val));
4965 return true;
4966 }
4967
4968 // If we have a tighly coupled allocation, the arraycopy may take care
4969 // of the array initialization. If one of the guards we insert between
4970 // the allocation and the arraycopy causes a deoptimization, an
4971 // unitialized array will escape the compiled method. To prevent that
4972 // we set the JVM state for uncommon traps between the allocation and
4973 // the arraycopy to the state before the allocation so, in case of
4974 // deoptimization, we'll reexecute the allocation and the
4975 // initialization.
4976 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4977 if (alloc != NULL) {
4978 ciMethod* trap_method = alloc->jvms()->method();
4979 int trap_bci = alloc->jvms()->bci();
4980
4981 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4982 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4983 // Make sure there's no store between the allocation and the
4984 // arraycopy otherwise visible side effects could be rexecuted
4985 // in case of deoptimization and cause incorrect execution.
4986 bool no_interfering_store = true;
4987 Node* mem = alloc->in(TypeFunc::Memory);
4988 if (mem->is_MergeMem()) {
4989 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4990 Node* n = mms.memory();
4991 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4992 assert(n->is_Store(), "what else?");
4993 no_interfering_store = false;
4994 break;
4995 }
4996 }
4997 } else {
4998 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4999 Node* n = mms.memory();
5000 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
5001 assert(n->is_Store(), "what else?");
5002 no_interfering_store = false;
5003 break;
5004 }
5005 }
5006 }
5007
5008 if (no_interfering_store) {
5009 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
5010 uint size = alloc->req();
5011 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
5012 old_jvms->set_map(sfpt);
5013 for (uint i = 0; i < size; i++) {
5014 sfpt->init_req(i, alloc->in(i));
5015 }
5016 // re-push array length for deoptimization
5017 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
5018 old_jvms->set_sp(old_jvms->sp()+1);
5019 old_jvms->set_monoff(old_jvms->monoff()+1);
5020 old_jvms->set_scloff(old_jvms->scloff()+1);
5021 old_jvms->set_endoff(old_jvms->endoff()+1);
5022 old_jvms->set_should_reexecute(true);
5023
5024 sfpt->set_i_o(map()->i_o());
5025 sfpt->set_memory(map()->memory());
5026 sfpt->set_control(map()->control());
5027
5028 JVMState* saved_jvms = jvms();
5029 saved_reexecute_sp = _reexecute_sp;
5030
5031 set_jvms(sfpt->jvms());
5032 _reexecute_sp = jvms()->sp();
5033
5034 return saved_jvms;
5035 }
5036 }
5037 }
5038 return NULL;
5039 }
5040
5041 // In case of a deoptimization, we restart execution at the
5042 // allocation, allocating a new array. We would leave an uninitialized
5043 // array in the heap that GCs wouldn't expect. Move the allocation
5044 // after the traps so we don't allocate the array if we
5045 // deoptimize. This is possible because tightly_coupled_allocation()
5046 // guarantees there's no observer of the allocated array at this point
5047 // and the control flow is simple enough.
5048 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
5049 int saved_reexecute_sp, uint new_idx) {
5050 if (saved_jvms != NULL && !stopped()) {
5051 assert(alloc != NULL, "only with a tightly coupled allocation");
5052 // restore JVM state to the state at the arraycopy
5053 saved_jvms->map()->set_control(map()->control());
5054 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5055 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5056 // If we've improved the types of some nodes (null check) while
5057 // emitting the guards, propagate them to the current state
5058 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
5059 set_jvms(saved_jvms);
5060 _reexecute_sp = saved_reexecute_sp;
5061
5062 // Remove the allocation from above the guards
5063 CallProjections callprojs;
5064 alloc->extract_projections(&callprojs, true);
5065 InitializeNode* init = alloc->initialization();
5066 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5067 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5068 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5069 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5070
5071 // move the allocation here (after the guards)
5072 _gvn.hash_delete(alloc);
5073 alloc->set_req(TypeFunc::Control, control());
5074 alloc->set_req(TypeFunc::I_O, i_o());
5075 Node *mem = reset_memory();
5076 set_all_memory(mem);
5077 alloc->set_req(TypeFunc::Memory, mem);
5078 set_control(init->proj_out_or_null(TypeFunc::Control));
5079 set_i_o(callprojs.fallthrough_ioproj);
5080
5081 // Update memory as done in GraphKit::set_output_for_allocation()
5082 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5083 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5084 if (ary_type->isa_aryptr() && length_type != NULL) {
5085 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5086 }
5087 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5088 int elemidx = C->get_alias_index(telemref);
5089 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
5090 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
5091
5092 Node* allocx = _gvn.transform(alloc);
5093 assert(allocx == alloc, "where has the allocation gone?");
5094 assert(dest->is_CheckCastPP(), "not an allocation result?");
5095
5096 _gvn.hash_delete(dest);
5097 dest->set_req(0, control());
5098 Node* destx = _gvn.transform(dest);
5099 assert(destx == dest, "where has the allocation result gone?");
5100 }
5101 }
5102
5103
5104 //------------------------------inline_arraycopy-----------------------
5105 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5106 // Object dest, int destPos,
5107 // int length);
5108 bool LibraryCallKit::inline_arraycopy() {
5109 // Get the arguments.
5110 Node* src = argument(0); // type: oop
5111 Node* src_offset = argument(1); // type: int
5112 Node* dest = argument(2); // type: oop
5113 Node* dest_offset = argument(3); // type: int
5114 Node* length = argument(4); // type: int
5115
5116 uint new_idx = C->unique();
5117
5118 // Check for allocation before we add nodes that would confuse
5119 // tightly_coupled_allocation()
5120 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5121
5122 int saved_reexecute_sp = -1;
5123 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5124 // See arraycopy_restore_alloc_state() comment
5125 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5126 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5127 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5128 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5129
5130 // The following tests must be performed
5131 // (1) src and dest are arrays.
5132 // (2) src and dest arrays must have elements of the same BasicType
5133 // (3) src and dest must not be null.
5134 // (4) src_offset must not be negative.
5135 // (5) dest_offset must not be negative.
5136 // (6) length must not be negative.
5137 // (7) src_offset + length must not exceed length of src.
5138 // (8) dest_offset + length must not exceed length of dest.
5139 // (9) each element of an oop array must be assignable
5140
5141 // (3) src and dest must not be null.
5142 // always do this here because we need the JVM state for uncommon traps
5143 Node* null_ctl = top();
5144 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5145 assert(null_ctl->is_top(), "no null control here");
5146 dest = null_check(dest, T_ARRAY);
5147
5148 if (!can_emit_guards) {
5149 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5150 // guards but the arraycopy node could still take advantage of a
5151 // tightly allocated allocation. tightly_coupled_allocation() is
5152 // called again to make sure it takes the null check above into
5153 // account: the null check is mandatory and if it caused an
5154 // uncommon trap to be emitted then the allocation can't be
5155 // considered tightly coupled in this context.
5156 alloc = tightly_coupled_allocation(dest, NULL);
5157 }
5158
5159 bool validated = false;
5160
5161 const Type* src_type = _gvn.type(src);
5162 const Type* dest_type = _gvn.type(dest);
5163 const TypeAryPtr* top_src = src_type->isa_aryptr();
5164 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5165
5166 // Do we have the type of src?
5167 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5168 // Do we have the type of dest?
5169 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5170 // Is the type for src from speculation?
5171 bool src_spec = false;
5172 // Is the type for dest from speculation?
5173 bool dest_spec = false;
5174
5175 if ((!has_src || !has_dest) && can_emit_guards) {
5176 // We don't have sufficient type information, let's see if
5177 // speculative types can help. We need to have types for both src
5178 // and dest so that it pays off.
5179
5180 // Do we already have or could we have type information for src
5181 bool could_have_src = has_src;
5182 // Do we already have or could we have type information for dest
5183 bool could_have_dest = has_dest;
5184
5185 ciKlass* src_k = NULL;
5186 if (!has_src) {
5187 src_k = src_type->speculative_type_not_null();
5188 if (src_k != NULL && src_k->is_array_klass()) {
5189 could_have_src = true;
5190 }
5191 }
5192
5193 ciKlass* dest_k = NULL;
5194 if (!has_dest) {
5195 dest_k = dest_type->speculative_type_not_null();
5196 if (dest_k != NULL && dest_k->is_array_klass()) {
5197 could_have_dest = true;
5198 }
5199 }
5200
5201 if (could_have_src && could_have_dest) {
5202 // This is going to pay off so emit the required guards
5203 if (!has_src) {
5204 src = maybe_cast_profiled_obj(src, src_k, true);
5205 src_type = _gvn.type(src);
5206 top_src = src_type->isa_aryptr();
5207 has_src = (top_src != NULL && top_src->klass() != NULL);
5208 src_spec = true;
5209 }
5210 if (!has_dest) {
5211 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5212 dest_type = _gvn.type(dest);
5213 top_dest = dest_type->isa_aryptr();
5214 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5215 dest_spec = true;
5216 }
5217 }
5218 }
5219
5220 if (has_src && has_dest && can_emit_guards) {
5221 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5222 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5223 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5224 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5225
5226 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5227 // If both arrays are object arrays then having the exact types
5228 // for both will remove the need for a subtype check at runtime
5229 // before the call and may make it possible to pick a faster copy
5230 // routine (without a subtype check on every element)
5231 // Do we have the exact type of src?
5232 bool could_have_src = src_spec;
5233 // Do we have the exact type of dest?
5234 bool could_have_dest = dest_spec;
5235 ciKlass* src_k = top_src->klass();
5236 ciKlass* dest_k = top_dest->klass();
5237 if (!src_spec) {
5238 src_k = src_type->speculative_type_not_null();
5239 if (src_k != NULL && src_k->is_array_klass()) {
5240 could_have_src = true;
5241 }
5242 }
5243 if (!dest_spec) {
5244 dest_k = dest_type->speculative_type_not_null();
5245 if (dest_k != NULL && dest_k->is_array_klass()) {
5246 could_have_dest = true;
5247 }
5248 }
5249 if (could_have_src && could_have_dest) {
5250 // If we can have both exact types, emit the missing guards
5251 if (could_have_src && !src_spec) {
5252 src = maybe_cast_profiled_obj(src, src_k, true);
5253 }
5254 if (could_have_dest && !dest_spec) {
5255 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5256 }
5257 }
5258 }
5259 }
5260
5261 ciMethod* trap_method = method();
5262 int trap_bci = bci();
5263 if (saved_jvms != NULL) {
5264 trap_method = alloc->jvms()->method();
5265 trap_bci = alloc->jvms()->bci();
5266 }
5267
5268 bool negative_length_guard_generated = false;
5269
5270 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5271 can_emit_guards &&
5272 !src->is_top() && !dest->is_top()) {
5273 // validate arguments: enables transformation the ArrayCopyNode
5274 validated = true;
5275
5276 RegionNode* slow_region = new RegionNode(1);
5277 record_for_igvn(slow_region);
5278
5279 // (1) src and dest are arrays.
5280 generate_non_array_guard(load_object_klass(src), slow_region);
5281 generate_non_array_guard(load_object_klass(dest), slow_region);
5282
5283 // (2) src and dest arrays must have elements of the same BasicType
5284 // done at macro expansion or at Ideal transformation time
5285
5286 // (4) src_offset must not be negative.
5287 generate_negative_guard(src_offset, slow_region);
5288
5289 // (5) dest_offset must not be negative.
5290 generate_negative_guard(dest_offset, slow_region);
5291
5292 // (7) src_offset + length must not exceed length of src.
5293 generate_limit_guard(src_offset, length,
5294 load_array_length(src),
5295 slow_region);
5296
5297 // (8) dest_offset + length must not exceed length of dest.
5298 generate_limit_guard(dest_offset, length,
5299 load_array_length(dest),
5300 slow_region);
5301
5302 // (6) length must not be negative.
5303 // This is also checked in generate_arraycopy() during macro expansion, but
5304 // we also have to check it here for the case where the ArrayCopyNode will
5305 // be eliminated by Escape Analysis.
5306 if (EliminateAllocations) {
5307 generate_negative_guard(length, slow_region);
5308 negative_length_guard_generated = true;
5309 }
5310
5311 // (9) each element of an oop array must be assignable
5312 Node* src_klass = load_object_klass(src);
5313 Node* dest_klass = load_object_klass(dest);
5314 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5315
5316 if (not_subtype_ctrl != top()) {
5317 PreserveJVMState pjvms(this);
5318 set_control(not_subtype_ctrl);
5319 uncommon_trap(Deoptimization::Reason_intrinsic,
5320 Deoptimization::Action_make_not_entrant);
5321 assert(stopped(), "Should be stopped");
5322 }
5323 {
5324 PreserveJVMState pjvms(this);
5325 set_control(_gvn.transform(slow_region));
5326 uncommon_trap(Deoptimization::Reason_intrinsic,
5327 Deoptimization::Action_make_not_entrant);
5328 assert(stopped(), "Should be stopped");
5329 }
5330
5331 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
5332 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
5333 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
5334 }
5335
5336 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5337
5338 if (stopped()) {
5339 return true;
5340 }
5341
5342 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5343 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5344 // so the compiler has a chance to eliminate them: during macro expansion,
5345 // we have to set their control (CastPP nodes are eliminated).
5346 load_object_klass(src), load_object_klass(dest),
5347 load_array_length(src), load_array_length(dest));
5348
5349 ac->set_arraycopy(validated);
5350
5351 Node* n = _gvn.transform(ac);
5352 if (n == ac) {
5353 ac->connect_outputs(this);
5354 } else {
5355 assert(validated, "shouldn't transform if all arguments not validated");
5356 set_all_memory(n);
5357 }
5358 clear_upper_avx();
5359
5360
5361 return true;
5362 }
5363
5364
5365 // Helper function which determines if an arraycopy immediately follows
5366 // an allocation, with no intervening tests or other escapes for the object.
5367 AllocateArrayNode*
5368 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5369 RegionNode* slow_region) {
5370 if (stopped()) return NULL; // no fast path
5371 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5372
5373 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5374 if (alloc == NULL) return NULL;
5375
5376 Node* rawmem = memory(Compile::AliasIdxRaw);
5377 // Is the allocation's memory state untouched?
5378 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5379 // Bail out if there have been raw-memory effects since the allocation.
5380 // (Example: There might have been a call or safepoint.)
5381 return NULL;
5382 }
5383 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5384 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5385 return NULL;
5386 }
5387
5388 // There must be no unexpected observers of this allocation.
5389 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5390 Node* obs = ptr->fast_out(i);
5391 if (obs != this->map()) {
5392 return NULL;
5393 }
5394 }
5395
5396 // This arraycopy must unconditionally follow the allocation of the ptr.
5397 Node* alloc_ctl = ptr->in(0);
5398 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5399
5400 Node* ctl = control();
5401 while (ctl != alloc_ctl) {
5402 // There may be guards which feed into the slow_region.
5403 // Any other control flow means that we might not get a chance
5404 // to finish initializing the allocated object.
5405 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5406 IfNode* iff = ctl->in(0)->as_If();
5407 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
5408 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5409 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5410 ctl = iff->in(0); // This test feeds the known slow_region.
5411 continue;
5412 }
5413 // One more try: Various low-level checks bottom out in
5414 // uncommon traps. If the debug-info of the trap omits
5415 // any reference to the allocation, as we've already
5416 // observed, then there can be no objection to the trap.
5417 bool found_trap = false;
5418 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5419 Node* obs = not_ctl->fast_out(j);
5420 if (obs->in(0) == not_ctl && obs->is_Call() &&
5421 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5422 found_trap = true; break;
5423 }
5424 }
5425 if (found_trap) {
5426 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5427 continue;
5428 }
5429 }
5430 return NULL;
5431 }
5432
5433 // If we get this far, we have an allocation which immediately
5434 // precedes the arraycopy, and we can take over zeroing the new object.
5435 // The arraycopy will finish the initialization, and provide
5436 // a new control state to which we will anchor the destination pointer.
5437
5438 return alloc;
5439 }
5440
5441 //-------------inline_encodeISOArray-----------------------------------
5442 // encode char[] to byte[] in ISO_8859_1
5443 bool LibraryCallKit::inline_encodeISOArray() {
5444 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5445 // no receiver since it is static method
5446 Node *src = argument(0);
5447 Node *src_offset = argument(1);
5448 Node *dst = argument(2);
5449 Node *dst_offset = argument(3);
5450 Node *length = argument(4);
5451
5452 const Type* src_type = src->Value(&_gvn);
5453 const Type* dst_type = dst->Value(&_gvn);
5454 const TypeAryPtr* top_src = src_type->isa_aryptr();
5455 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5456 if (top_src == NULL || top_src->klass() == NULL ||
5457 top_dest == NULL || top_dest->klass() == NULL) {
5458 // failed array check
5459 return false;
5460 }
5461
5462 // Figure out the size and type of the elements we will be copying.
5463 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5464 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5465 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5466 return false;
5467 }
5468
5469 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5470 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5471 // 'src_start' points to src array + scaled offset
5472 // 'dst_start' points to dst array + scaled offset
5473
5474 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5475 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5476 enc = _gvn.transform(enc);
5477 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5478 set_memory(res_mem, mtype);
5479 set_result(enc);
5480 clear_upper_avx();
5481
5482 return true;
5483 }
5484
5485 //-------------inline_multiplyToLen-----------------------------------
5486 bool LibraryCallKit::inline_multiplyToLen() {
5487 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5488
5489 address stubAddr = StubRoutines::multiplyToLen();
5490 if (stubAddr == NULL) {
5491 return false; // Intrinsic's stub is not implemented on this platform
5492 }
5493 const char* stubName = "multiplyToLen";
5494
5495 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5496
5497 // no receiver because it is a static method
5498 Node* x = argument(0);
5499 Node* xlen = argument(1);
5500 Node* y = argument(2);
5501 Node* ylen = argument(3);
5502 Node* z = argument(4);
5503
5504 const Type* x_type = x->Value(&_gvn);
5505 const Type* y_type = y->Value(&_gvn);
5506 const TypeAryPtr* top_x = x_type->isa_aryptr();
5507 const TypeAryPtr* top_y = y_type->isa_aryptr();
5508 if (top_x == NULL || top_x->klass() == NULL ||
5509 top_y == NULL || top_y->klass() == NULL) {
5510 // failed array check
5511 return false;
5512 }
5513
5514 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5515 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5516 if (x_elem != T_INT || y_elem != T_INT) {
5517 return false;
5518 }
5519
5520 // Set the original stack and the reexecute bit for the interpreter to reexecute
5521 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5522 // on the return from z array allocation in runtime.
5523 { PreserveReexecuteState preexecs(this);
5524 jvms()->set_should_reexecute(true);
5525
5526 Node* x_start = array_element_address(x, intcon(0), x_elem);
5527 Node* y_start = array_element_address(y, intcon(0), y_elem);
5528 // 'x_start' points to x array + scaled xlen
5529 // 'y_start' points to y array + scaled ylen
5530
5531 // Allocate the result array
5532 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5533 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5534 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5535
5536 IdealKit ideal(this);
5537
5538 #define __ ideal.
5539 Node* one = __ ConI(1);
5540 Node* zero = __ ConI(0);
5541 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5542 __ set(need_alloc, zero);
5543 __ set(z_alloc, z);
5544 __ if_then(z, BoolTest::eq, null()); {
5545 __ increment (need_alloc, one);
5546 } __ else_(); {
5547 // Update graphKit memory and control from IdealKit.
5548 sync_kit(ideal);
5549 Node* zlen_arg = load_array_length(z);
5550 // Update IdealKit memory and control from graphKit.
5551 __ sync_kit(this);
5552 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5553 __ increment (need_alloc, one);
5554 } __ end_if();
5555 } __ end_if();
5556
5557 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5558 // Update graphKit memory and control from IdealKit.
5559 sync_kit(ideal);
5560 Node * narr = new_array(klass_node, zlen, 1);
5561 // Update IdealKit memory and control from graphKit.
5562 __ sync_kit(this);
5563 __ set(z_alloc, narr);
5564 } __ end_if();
5565
5566 sync_kit(ideal);
5567 z = __ value(z_alloc);
5568 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5569 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5570 // Final sync IdealKit and GraphKit.
5571 final_sync(ideal);
5572 #undef __
5573
5574 Node* z_start = array_element_address(z, intcon(0), T_INT);
5575
5576 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5577 OptoRuntime::multiplyToLen_Type(),
5578 stubAddr, stubName, TypePtr::BOTTOM,
5579 x_start, xlen, y_start, ylen, z_start, zlen);
5580 } // original reexecute is set back here
5581
5582 C->set_has_split_ifs(true); // Has chance for split-if optimization
5583 set_result(z);
5584 return true;
5585 }
5586
5587 //-------------inline_squareToLen------------------------------------
5588 bool LibraryCallKit::inline_squareToLen() {
5589 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5590
5591 address stubAddr = StubRoutines::squareToLen();
5592 if (stubAddr == NULL) {
5593 return false; // Intrinsic's stub is not implemented on this platform
5594 }
5595 const char* stubName = "squareToLen";
5596
5597 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5598
5599 Node* x = argument(0);
5600 Node* len = argument(1);
5601 Node* z = argument(2);
5602 Node* zlen = argument(3);
5603
5604 const Type* x_type = x->Value(&_gvn);
5605 const Type* z_type = z->Value(&_gvn);
5606 const TypeAryPtr* top_x = x_type->isa_aryptr();
5607 const TypeAryPtr* top_z = z_type->isa_aryptr();
5608 if (top_x == NULL || top_x->klass() == NULL ||
5609 top_z == NULL || top_z->klass() == NULL) {
5610 // failed array check
5611 return false;
5612 }
5613
5614 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5615 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5616 if (x_elem != T_INT || z_elem != T_INT) {
5617 return false;
5618 }
5619
5620
5621 Node* x_start = array_element_address(x, intcon(0), x_elem);
5622 Node* z_start = array_element_address(z, intcon(0), z_elem);
5623
5624 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5625 OptoRuntime::squareToLen_Type(),
5626 stubAddr, stubName, TypePtr::BOTTOM,
5627 x_start, len, z_start, zlen);
5628
5629 set_result(z);
5630 return true;
5631 }
5632
5633 //-------------inline_mulAdd------------------------------------------
5634 bool LibraryCallKit::inline_mulAdd() {
5635 assert(UseMulAddIntrinsic, "not implemented on this platform");
5636
5637 address stubAddr = StubRoutines::mulAdd();
5638 if (stubAddr == NULL) {
5639 return false; // Intrinsic's stub is not implemented on this platform
5640 }
5641 const char* stubName = "mulAdd";
5642
5643 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5644
5645 Node* out = argument(0);
5646 Node* in = argument(1);
5647 Node* offset = argument(2);
5648 Node* len = argument(3);
5649 Node* k = argument(4);
5650
5651 const Type* out_type = out->Value(&_gvn);
5652 const Type* in_type = in->Value(&_gvn);
5653 const TypeAryPtr* top_out = out_type->isa_aryptr();
5654 const TypeAryPtr* top_in = in_type->isa_aryptr();
5655 if (top_out == NULL || top_out->klass() == NULL ||
5656 top_in == NULL || top_in->klass() == NULL) {
5657 // failed array check
5658 return false;
5659 }
5660
5661 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5662 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5663 if (out_elem != T_INT || in_elem != T_INT) {
5664 return false;
5665 }
5666
5667 Node* outlen = load_array_length(out);
5668 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5669 Node* out_start = array_element_address(out, intcon(0), out_elem);
5670 Node* in_start = array_element_address(in, intcon(0), in_elem);
5671
5672 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5673 OptoRuntime::mulAdd_Type(),
5674 stubAddr, stubName, TypePtr::BOTTOM,
5675 out_start,in_start, new_offset, len, k);
5676 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5677 set_result(result);
5678 return true;
5679 }
5680
5681 //-------------inline_montgomeryMultiply-----------------------------------
5682 bool LibraryCallKit::inline_montgomeryMultiply() {
5683 address stubAddr = StubRoutines::montgomeryMultiply();
5684 if (stubAddr == NULL) {
5685 return false; // Intrinsic's stub is not implemented on this platform
5686 }
5687
5688 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5689 const char* stubName = "montgomery_multiply";
5690
5691 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5692
5693 Node* a = argument(0);
5694 Node* b = argument(1);
5695 Node* n = argument(2);
5696 Node* len = argument(3);
5697 Node* inv = argument(4);
5698 Node* m = argument(6);
5699
5700 const Type* a_type = a->Value(&_gvn);
5701 const TypeAryPtr* top_a = a_type->isa_aryptr();
5702 const Type* b_type = b->Value(&_gvn);
5703 const TypeAryPtr* top_b = b_type->isa_aryptr();
5704 const Type* n_type = a->Value(&_gvn);
5705 const TypeAryPtr* top_n = n_type->isa_aryptr();
5706 const Type* m_type = a->Value(&_gvn);
5707 const TypeAryPtr* top_m = m_type->isa_aryptr();
5708 if (top_a == NULL || top_a->klass() == NULL ||
5709 top_b == NULL || top_b->klass() == NULL ||
5710 top_n == NULL || top_n->klass() == NULL ||
5711 top_m == NULL || top_m->klass() == NULL) {
5712 // failed array check
5713 return false;
5714 }
5715
5716 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5717 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5718 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5719 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5720 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5721 return false;
5722 }
5723
5724 // Make the call
5725 {
5726 Node* a_start = array_element_address(a, intcon(0), a_elem);
5727 Node* b_start = array_element_address(b, intcon(0), b_elem);
5728 Node* n_start = array_element_address(n, intcon(0), n_elem);
5729 Node* m_start = array_element_address(m, intcon(0), m_elem);
5730
5731 Node* call = make_runtime_call(RC_LEAF,
5732 OptoRuntime::montgomeryMultiply_Type(),
5733 stubAddr, stubName, TypePtr::BOTTOM,
5734 a_start, b_start, n_start, len, inv, top(),
5735 m_start);
5736 set_result(m);
5737 }
5738
5739 return true;
5740 }
5741
5742 bool LibraryCallKit::inline_montgomerySquare() {
5743 address stubAddr = StubRoutines::montgomerySquare();
5744 if (stubAddr == NULL) {
5745 return false; // Intrinsic's stub is not implemented on this platform
5746 }
5747
5748 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5749 const char* stubName = "montgomery_square";
5750
5751 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5752
5753 Node* a = argument(0);
5754 Node* n = argument(1);
5755 Node* len = argument(2);
5756 Node* inv = argument(3);
5757 Node* m = argument(5);
5758
5759 const Type* a_type = a->Value(&_gvn);
5760 const TypeAryPtr* top_a = a_type->isa_aryptr();
5761 const Type* n_type = a->Value(&_gvn);
5762 const TypeAryPtr* top_n = n_type->isa_aryptr();
5763 const Type* m_type = a->Value(&_gvn);
5764 const TypeAryPtr* top_m = m_type->isa_aryptr();
5765 if (top_a == NULL || top_a->klass() == NULL ||
5766 top_n == NULL || top_n->klass() == NULL ||
5767 top_m == NULL || top_m->klass() == NULL) {
5768 // failed array check
5769 return false;
5770 }
5771
5772 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5773 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5774 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5775 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5776 return false;
5777 }
5778
5779 // Make the call
5780 {
5781 Node* a_start = array_element_address(a, intcon(0), a_elem);
5782 Node* n_start = array_element_address(n, intcon(0), n_elem);
5783 Node* m_start = array_element_address(m, intcon(0), m_elem);
5784
5785 Node* call = make_runtime_call(RC_LEAF,
5786 OptoRuntime::montgomerySquare_Type(),
5787 stubAddr, stubName, TypePtr::BOTTOM,
5788 a_start, n_start, len, inv, top(),
5789 m_start);
5790 set_result(m);
5791 }
5792
5793 return true;
5794 }
5795
5796 //-------------inline_vectorizedMismatch------------------------------
5797 bool LibraryCallKit::inline_vectorizedMismatch() {
5798 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5799
5800 address stubAddr = StubRoutines::vectorizedMismatch();
5801 if (stubAddr == NULL) {
5802 return false; // Intrinsic's stub is not implemented on this platform
5803 }
5804 const char* stubName = "vectorizedMismatch";
5805 int size_l = callee()->signature()->size();
5806 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5807
5808 Node* obja = argument(0);
5809 Node* aoffset = argument(1);
5810 Node* objb = argument(3);
5811 Node* boffset = argument(4);
5812 Node* length = argument(6);
5813 Node* scale = argument(7);
5814
5815 const Type* a_type = obja->Value(&_gvn);
5816 const Type* b_type = objb->Value(&_gvn);
5817 const TypeAryPtr* top_a = a_type->isa_aryptr();
5818 const TypeAryPtr* top_b = b_type->isa_aryptr();
5819 if (top_a == NULL || top_a->klass() == NULL ||
5820 top_b == NULL || top_b->klass() == NULL) {
5821 // failed array check
5822 return false;
5823 }
5824
5825 Node* call;
5826 jvms()->set_should_reexecute(true);
5827
5828 Node* obja_adr = make_unsafe_address(obja, aoffset);
5829 Node* objb_adr = make_unsafe_address(objb, boffset);
5830
5831 call = make_runtime_call(RC_LEAF,
5832 OptoRuntime::vectorizedMismatch_Type(),
5833 stubAddr, stubName, TypePtr::BOTTOM,
5834 obja_adr, objb_adr, length, scale);
5835
5836 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5837 set_result(result);
5838 return true;
5839 }
5840
5841 /**
5842 * Calculate CRC32 for byte.
5843 * int java.util.zip.CRC32.update(int crc, int b)
5844 */
5845 bool LibraryCallKit::inline_updateCRC32() {
5846 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5847 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5848 // no receiver since it is static method
5849 Node* crc = argument(0); // type: int
5850 Node* b = argument(1); // type: int
5851
5852 /*
5853 * int c = ~ crc;
5854 * b = timesXtoThe32[(b ^ c) & 0xFF];
5855 * b = b ^ (c >>> 8);
5856 * crc = ~b;
5857 */
5858
5859 Node* M1 = intcon(-1);
5860 crc = _gvn.transform(new XorINode(crc, M1));
5861 Node* result = _gvn.transform(new XorINode(crc, b));
5862 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5863
5864 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5865 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5866 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5867 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5868
5869 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5870 result = _gvn.transform(new XorINode(crc, result));
5871 result = _gvn.transform(new XorINode(result, M1));
5872 set_result(result);
5873 return true;
5874 }
5875
5876 /**
5877 * Calculate CRC32 for byte[] array.
5878 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5879 */
5880 bool LibraryCallKit::inline_updateBytesCRC32() {
5881 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5882 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5883 // no receiver since it is static method
5884 Node* crc = argument(0); // type: int
5885 Node* src = argument(1); // type: oop
5886 Node* offset = argument(2); // type: int
5887 Node* length = argument(3); // type: int
5888
5889 const Type* src_type = src->Value(&_gvn);
5890 const TypeAryPtr* top_src = src_type->isa_aryptr();
5891 if (top_src == NULL || top_src->klass() == NULL) {
5892 // failed array check
5893 return false;
5894 }
5895
5896 // Figure out the size and type of the elements we will be copying.
5897 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5898 if (src_elem != T_BYTE) {
5899 return false;
5900 }
5901
5902 // 'src_start' points to src array + scaled offset
5903 Node* src_start = array_element_address(src, offset, src_elem);
5904
5905 // We assume that range check is done by caller.
5906 // TODO: generate range check (offset+length < src.length) in debug VM.
5907
5908 // Call the stub.
5909 address stubAddr = StubRoutines::updateBytesCRC32();
5910 const char *stubName = "updateBytesCRC32";
5911
5912 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5913 stubAddr, stubName, TypePtr::BOTTOM,
5914 crc, src_start, length);
5915 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5916 set_result(result);
5917 return true;
5918 }
5919
5920 /**
5921 * Calculate CRC32 for ByteBuffer.
5922 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5923 */
5924 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5925 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5926 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5927 // no receiver since it is static method
5928 Node* crc = argument(0); // type: int
5929 Node* src = argument(1); // type: long
5930 Node* offset = argument(3); // type: int
5931 Node* length = argument(4); // type: int
5932
5933 src = ConvL2X(src); // adjust Java long to machine word
5934 Node* base = _gvn.transform(new CastX2PNode(src));
5935 offset = ConvI2X(offset);
5936
5937 // 'src_start' points to src array + scaled offset
5938 Node* src_start = basic_plus_adr(top(), base, offset);
5939
5940 // Call the stub.
5941 address stubAddr = StubRoutines::updateBytesCRC32();
5942 const char *stubName = "updateBytesCRC32";
5943
5944 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5945 stubAddr, stubName, TypePtr::BOTTOM,
5946 crc, src_start, length);
5947 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5948 set_result(result);
5949 return true;
5950 }
5951
5952 //------------------------------get_table_from_crc32c_class-----------------------
5953 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5954 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5955 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5956
5957 return table;
5958 }
5959
5960 //------------------------------inline_updateBytesCRC32C-----------------------
5961 //
5962 // Calculate CRC32C for byte[] array.
5963 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5964 //
5965 bool LibraryCallKit::inline_updateBytesCRC32C() {
5966 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5967 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5968 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5969 // no receiver since it is a static method
5970 Node* crc = argument(0); // type: int
5971 Node* src = argument(1); // type: oop
5972 Node* offset = argument(2); // type: int
5973 Node* end = argument(3); // type: int
5974
5975 Node* length = _gvn.transform(new SubINode(end, offset));
5976
5977 const Type* src_type = src->Value(&_gvn);
5978 const TypeAryPtr* top_src = src_type->isa_aryptr();
5979 if (top_src == NULL || top_src->klass() == NULL) {
5980 // failed array check
5981 return false;
5982 }
5983
5984 // Figure out the size and type of the elements we will be copying.
5985 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5986 if (src_elem != T_BYTE) {
5987 return false;
5988 }
5989
5990 // 'src_start' points to src array + scaled offset
5991 Node* src_start = array_element_address(src, offset, src_elem);
5992
5993 // static final int[] byteTable in class CRC32C
5994 Node* table = get_table_from_crc32c_class(callee()->holder());
5995 Node* table_start = array_element_address(table, intcon(0), T_INT);
5996
5997 // We assume that range check is done by caller.
5998 // TODO: generate range check (offset+length < src.length) in debug VM.
5999
6000 // Call the stub.
6001 address stubAddr = StubRoutines::updateBytesCRC32C();
6002 const char *stubName = "updateBytesCRC32C";
6003
6004 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6005 stubAddr, stubName, TypePtr::BOTTOM,
6006 crc, src_start, length, table_start);
6007 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6008 set_result(result);
6009 return true;
6010 }
6011
6012 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
6013 //
6014 // Calculate CRC32C for DirectByteBuffer.
6015 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
6016 //
6017 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
6018 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
6019 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
6020 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
6021 // no receiver since it is a static method
6022 Node* crc = argument(0); // type: int
6023 Node* src = argument(1); // type: long
6024 Node* offset = argument(3); // type: int
6025 Node* end = argument(4); // type: int
6026
6027 Node* length = _gvn.transform(new SubINode(end, offset));
6028
6029 src = ConvL2X(src); // adjust Java long to machine word
6030 Node* base = _gvn.transform(new CastX2PNode(src));
6031 offset = ConvI2X(offset);
6032
6033 // 'src_start' points to src array + scaled offset
6034 Node* src_start = basic_plus_adr(top(), base, offset);
6035
6036 // static final int[] byteTable in class CRC32C
6037 Node* table = get_table_from_crc32c_class(callee()->holder());
6038 Node* table_start = array_element_address(table, intcon(0), T_INT);
6039
6040 // Call the stub.
6041 address stubAddr = StubRoutines::updateBytesCRC32C();
6042 const char *stubName = "updateBytesCRC32C";
6043
6044 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6045 stubAddr, stubName, TypePtr::BOTTOM,
6046 crc, src_start, length, table_start);
6047 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6048 set_result(result);
6049 return true;
6050 }
6051
6052 //------------------------------inline_updateBytesAdler32----------------------
6053 //
6054 // Calculate Adler32 checksum for byte[] array.
6055 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6056 //
6057 bool LibraryCallKit::inline_updateBytesAdler32() {
6058 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6059 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6060 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6061 // no receiver since it is static method
6062 Node* crc = argument(0); // type: int
6063 Node* src = argument(1); // type: oop
6064 Node* offset = argument(2); // type: int
6065 Node* length = argument(3); // type: int
6066
6067 const Type* src_type = src->Value(&_gvn);
6068 const TypeAryPtr* top_src = src_type->isa_aryptr();
6069 if (top_src == NULL || top_src->klass() == NULL) {
6070 // failed array check
6071 return false;
6072 }
6073
6074 // Figure out the size and type of the elements we will be copying.
6075 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6076 if (src_elem != T_BYTE) {
6077 return false;
6078 }
6079
6080 // 'src_start' points to src array + scaled offset
6081 Node* src_start = array_element_address(src, offset, src_elem);
6082
6083 // We assume that range check is done by caller.
6084 // TODO: generate range check (offset+length < src.length) in debug VM.
6085
6086 // Call the stub.
6087 address stubAddr = StubRoutines::updateBytesAdler32();
6088 const char *stubName = "updateBytesAdler32";
6089
6090 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6091 stubAddr, stubName, TypePtr::BOTTOM,
6092 crc, src_start, length);
6093 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6094 set_result(result);
6095 return true;
6096 }
6097
6098 //------------------------------inline_updateByteBufferAdler32---------------
6099 //
6100 // Calculate Adler32 checksum for DirectByteBuffer.
6101 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6102 //
6103 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6104 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6105 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6106 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6107 // no receiver since it is static method
6108 Node* crc = argument(0); // type: int
6109 Node* src = argument(1); // type: long
6110 Node* offset = argument(3); // type: int
6111 Node* length = argument(4); // type: int
6112
6113 src = ConvL2X(src); // adjust Java long to machine word
6114 Node* base = _gvn.transform(new CastX2PNode(src));
6115 offset = ConvI2X(offset);
6116
6117 // 'src_start' points to src array + scaled offset
6118 Node* src_start = basic_plus_adr(top(), base, offset);
6119
6120 // Call the stub.
6121 address stubAddr = StubRoutines::updateBytesAdler32();
6122 const char *stubName = "updateBytesAdler32";
6123
6124 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6125 stubAddr, stubName, TypePtr::BOTTOM,
6126 crc, src_start, length);
6127
6128 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6129 set_result(result);
6130 return true;
6131 }
6132
6133 //----------------------------inline_reference_get----------------------------
6134 // public T java.lang.ref.Reference.get();
6135 bool LibraryCallKit::inline_reference_get() {
6136 const int referent_offset = java_lang_ref_Reference::referent_offset;
6137 guarantee(referent_offset > 0, "should have already been set");
6138
6139 // Get the argument:
6140 Node* reference_obj = null_check_receiver();
6141 if (stopped()) return true;
6142
6143 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6144
6145 ciInstanceKlass* klass = env()->Object_klass();
6146 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6147
6148 Node* no_ctrl = NULL;
6149 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6150
6151 // Use the pre-barrier to record the value in the referent field
6152 pre_barrier(false /* do_load */,
6153 control(),
6154 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6155 result /* pre_val */,
6156 T_OBJECT);
6157
6158 // Add memory barrier to prevent commoning reads from this field
6159 // across safepoint since GC can change its value.
6160 insert_mem_bar(Op_MemBarCPUOrder);
6161
6162 set_result(result);
6163 return true;
6164 }
6165
6166
6167 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6168 bool is_exact=true, bool is_static=false,
6169 ciInstanceKlass * fromKls=NULL) {
6170 if (fromKls == NULL) {
6171 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6172 assert(tinst != NULL, "obj is null");
6173 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6174 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6175 fromKls = tinst->klass()->as_instance_klass();
6176 } else {
6177 assert(is_static, "only for static field access");
6178 }
6179 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6180 ciSymbol::make(fieldTypeString),
6181 is_static);
6182
6183 assert (field != NULL, "undefined field");
6184 if (field == NULL) return (Node *) NULL;
6185
6186 if (is_static) {
6187 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6188 fromObj = makecon(tip);
6189 }
6190
6191 // Next code copied from Parse::do_get_xxx():
6192
6193 // Compute address and memory type.
6194 int offset = field->offset_in_bytes();
6195 bool is_vol = field->is_volatile();
6196 ciType* field_klass = field->type();
6197 assert(field_klass->is_loaded(), "should be loaded");
6198 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6199 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6200 BasicType bt = field->layout_type();
6201
6202 // Build the resultant type of the load
6203 const Type *type;
6204 if (bt == T_OBJECT) {
6205 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6206 } else {
6207 type = Type::get_const_basic_type(bt);
6208 }
6209
6210 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6211 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6212 }
6213 // Build the load.
6214 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6215 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6216 // If reference is volatile, prevent following memory ops from
6217 // floating up past the volatile read. Also prevents commoning
6218 // another volatile read.
6219 if (is_vol) {
6220 // Memory barrier includes bogus read of value to force load BEFORE membar
6221 insert_mem_bar(Op_MemBarAcquire, loadedField);
6222 }
6223 return loadedField;
6224 }
6225
6226 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6227 bool is_exact = true, bool is_static = false,
6228 ciInstanceKlass * fromKls = NULL) {
6229 if (fromKls == NULL) {
6230 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6231 assert(tinst != NULL, "obj is null");
6232 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6233 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6234 fromKls = tinst->klass()->as_instance_klass();
6235 }
6236 else {
6237 assert(is_static, "only for static field access");
6238 }
6239 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6240 ciSymbol::make(fieldTypeString),
6241 is_static);
6242
6243 assert(field != NULL, "undefined field");
6244 assert(!field->is_volatile(), "not defined for volatile fields");
6245
6246 if (is_static) {
6247 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6248 fromObj = makecon(tip);
6249 }
6250
6251 // Next code copied from Parse::do_get_xxx():
6252
6253 // Compute address and memory type.
6254 int offset = field->offset_in_bytes();
6255 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6256
6257 return adr;
6258 }
6259
6260 //------------------------------inline_aescrypt_Block-----------------------
6261 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6262 address stubAddr = NULL;
6263 const char *stubName;
6264 assert(UseAES, "need AES instruction support");
6265
6266 switch(id) {
6267 case vmIntrinsics::_aescrypt_encryptBlock:
6268 stubAddr = StubRoutines::aescrypt_encryptBlock();
6269 stubName = "aescrypt_encryptBlock";
6270 break;
6271 case vmIntrinsics::_aescrypt_decryptBlock:
6272 stubAddr = StubRoutines::aescrypt_decryptBlock();
6273 stubName = "aescrypt_decryptBlock";
6274 break;
6275 default:
6276 break;
6277 }
6278 if (stubAddr == NULL) return false;
6279
6280 Node* aescrypt_object = argument(0);
6281 Node* src = argument(1);
6282 Node* src_offset = argument(2);
6283 Node* dest = argument(3);
6284 Node* dest_offset = argument(4);
6285
6286 // (1) src and dest are arrays.
6287 const Type* src_type = src->Value(&_gvn);
6288 const Type* dest_type = dest->Value(&_gvn);
6289 const TypeAryPtr* top_src = src_type->isa_aryptr();
6290 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6291 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6292
6293 // for the quick and dirty code we will skip all the checks.
6294 // we are just trying to get the call to be generated.
6295 Node* src_start = src;
6296 Node* dest_start = dest;
6297 if (src_offset != NULL || dest_offset != NULL) {
6298 assert(src_offset != NULL && dest_offset != NULL, "");
6299 src_start = array_element_address(src, src_offset, T_BYTE);
6300 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6301 }
6302
6303 // now need to get the start of its expanded key array
6304 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6305 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6306 if (k_start == NULL) return false;
6307
6308 if (Matcher::pass_original_key_for_aes()) {
6309 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6310 // compatibility issues between Java key expansion and SPARC crypto instructions
6311 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6312 if (original_k_start == NULL) return false;
6313
6314 // Call the stub.
6315 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6316 stubAddr, stubName, TypePtr::BOTTOM,
6317 src_start, dest_start, k_start, original_k_start);
6318 } else {
6319 // Call the stub.
6320 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6321 stubAddr, stubName, TypePtr::BOTTOM,
6322 src_start, dest_start, k_start);
6323 }
6324
6325 return true;
6326 }
6327
6328 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6329 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6330 address stubAddr = NULL;
6331 const char *stubName = NULL;
6332
6333 assert(UseAES, "need AES instruction support");
6334
6335 switch(id) {
6336 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6337 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6338 stubName = "cipherBlockChaining_encryptAESCrypt";
6339 break;
6340 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6341 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6342 stubName = "cipherBlockChaining_decryptAESCrypt";
6343 break;
6344 default:
6345 break;
6346 }
6347 if (stubAddr == NULL) return false;
6348
6349 Node* cipherBlockChaining_object = argument(0);
6350 Node* src = argument(1);
6351 Node* src_offset = argument(2);
6352 Node* len = argument(3);
6353 Node* dest = argument(4);
6354 Node* dest_offset = argument(5);
6355
6356 // (1) src and dest are arrays.
6357 const Type* src_type = src->Value(&_gvn);
6358 const Type* dest_type = dest->Value(&_gvn);
6359 const TypeAryPtr* top_src = src_type->isa_aryptr();
6360 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6361 assert (top_src != NULL && top_src->klass() != NULL
6362 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6363
6364 // checks are the responsibility of the caller
6365 Node* src_start = src;
6366 Node* dest_start = dest;
6367 if (src_offset != NULL || dest_offset != NULL) {
6368 assert(src_offset != NULL && dest_offset != NULL, "");
6369 src_start = array_element_address(src, src_offset, T_BYTE);
6370 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6371 }
6372
6373 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6374 // (because of the predicated logic executed earlier).
6375 // so we cast it here safely.
6376 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6377
6378 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6379 if (embeddedCipherObj == NULL) return false;
6380
6381 // cast it to what we know it will be at runtime
6382 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6383 assert(tinst != NULL, "CBC obj is null");
6384 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6385 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6386 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6387
6388 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6389 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6390 const TypeOopPtr* xtype = aklass->as_instance_type();
6391 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6392 aescrypt_object = _gvn.transform(aescrypt_object);
6393
6394 // we need to get the start of the aescrypt_object's expanded key array
6395 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6396 if (k_start == NULL) return false;
6397
6398 // similarly, get the start address of the r vector
6399 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6400 if (objRvec == NULL) return false;
6401 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6402
6403 Node* cbcCrypt;
6404 if (Matcher::pass_original_key_for_aes()) {
6405 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6406 // compatibility issues between Java key expansion and SPARC crypto instructions
6407 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6408 if (original_k_start == NULL) return false;
6409
6410 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6411 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6412 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6413 stubAddr, stubName, TypePtr::BOTTOM,
6414 src_start, dest_start, k_start, r_start, len, original_k_start);
6415 } else {
6416 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6417 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6418 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6419 stubAddr, stubName, TypePtr::BOTTOM,
6420 src_start, dest_start, k_start, r_start, len);
6421 }
6422
6423 // return cipher length (int)
6424 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6425 set_result(retvalue);
6426 return true;
6427 }
6428
6429 //------------------------------inline_counterMode_AESCrypt-----------------------
6430 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6431 assert(UseAES, "need AES instruction support");
6432 if (!UseAESCTRIntrinsics) return false;
6433
6434 address stubAddr = NULL;
6435 const char *stubName = NULL;
6436 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6437 stubAddr = StubRoutines::counterMode_AESCrypt();
6438 stubName = "counterMode_AESCrypt";
6439 }
6440 if (stubAddr == NULL) return false;
6441
6442 Node* counterMode_object = argument(0);
6443 Node* src = argument(1);
6444 Node* src_offset = argument(2);
6445 Node* len = argument(3);
6446 Node* dest = argument(4);
6447 Node* dest_offset = argument(5);
6448
6449 // (1) src and dest are arrays.
6450 const Type* src_type = src->Value(&_gvn);
6451 const Type* dest_type = dest->Value(&_gvn);
6452 const TypeAryPtr* top_src = src_type->isa_aryptr();
6453 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6454 assert(top_src != NULL && top_src->klass() != NULL &&
6455 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6456
6457 // checks are the responsibility of the caller
6458 Node* src_start = src;
6459 Node* dest_start = dest;
6460 if (src_offset != NULL || dest_offset != NULL) {
6461 assert(src_offset != NULL && dest_offset != NULL, "");
6462 src_start = array_element_address(src, src_offset, T_BYTE);
6463 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6464 }
6465
6466 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6467 // (because of the predicated logic executed earlier).
6468 // so we cast it here safely.
6469 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6470 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6471 if (embeddedCipherObj == NULL) return false;
6472 // cast it to what we know it will be at runtime
6473 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6474 assert(tinst != NULL, "CTR obj is null");
6475 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6476 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6477 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6478 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6479 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6480 const TypeOopPtr* xtype = aklass->as_instance_type();
6481 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6482 aescrypt_object = _gvn.transform(aescrypt_object);
6483 // we need to get the start of the aescrypt_object's expanded key array
6484 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6485 if (k_start == NULL) return false;
6486 // similarly, get the start address of the r vector
6487 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6488 if (obj_counter == NULL) return false;
6489 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6490
6491 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6492 if (saved_encCounter == NULL) return false;
6493 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6494 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6495
6496 Node* ctrCrypt;
6497 if (Matcher::pass_original_key_for_aes()) {
6498 // no SPARC version for AES/CTR intrinsics now.
6499 return false;
6500 }
6501 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6502 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6503 OptoRuntime::counterMode_aescrypt_Type(),
6504 stubAddr, stubName, TypePtr::BOTTOM,
6505 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6506
6507 // return cipher length (int)
6508 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6509 set_result(retvalue);
6510 return true;
6511 }
6512
6513 //------------------------------get_key_start_from_aescrypt_object-----------------------
6514 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6515 #if defined(PPC64) || defined(S390)
6516 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6517 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6518 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6519 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6520 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6521 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6522 if (objSessionK == NULL) {
6523 return (Node *) NULL;
6524 }
6525 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6526 #else
6527 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6528 #endif // PPC64
6529 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6530 if (objAESCryptKey == NULL) return (Node *) NULL;
6531
6532 // now have the array, need to get the start address of the K array
6533 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6534 return k_start;
6535 }
6536
6537 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6538 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6539 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6540 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6541 if (objAESCryptKey == NULL) return (Node *) NULL;
6542
6543 // now have the array, need to get the start address of the lastKey array
6544 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6545 return original_k_start;
6546 }
6547
6548 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6549 // Return node representing slow path of predicate check.
6550 // the pseudo code we want to emulate with this predicate is:
6551 // for encryption:
6552 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6553 // for decryption:
6554 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6555 // note cipher==plain is more conservative than the original java code but that's OK
6556 //
6557 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6558 // The receiver was checked for NULL already.
6559 Node* objCBC = argument(0);
6560
6561 // Load embeddedCipher field of CipherBlockChaining object.
6562 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6563
6564 // get AESCrypt klass for instanceOf check
6565 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6566 // will have same classloader as CipherBlockChaining object
6567 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6568 assert(tinst != NULL, "CBCobj is null");
6569 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6570
6571 // we want to do an instanceof comparison against the AESCrypt class
6572 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6573 if (!klass_AESCrypt->is_loaded()) {
6574 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6575 Node* ctrl = control();
6576 set_control(top()); // no regular fast path
6577 return ctrl;
6578 }
6579 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6580
6581 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6582 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6583 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6584
6585 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6586
6587 // for encryption, we are done
6588 if (!decrypting)
6589 return instof_false; // even if it is NULL
6590
6591 // for decryption, we need to add a further check to avoid
6592 // taking the intrinsic path when cipher and plain are the same
6593 // see the original java code for why.
6594 RegionNode* region = new RegionNode(3);
6595 region->init_req(1, instof_false);
6596 Node* src = argument(1);
6597 Node* dest = argument(4);
6598 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6599 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6600 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6601 region->init_req(2, src_dest_conjoint);
6602
6603 record_for_igvn(region);
6604 return _gvn.transform(region);
6605 }
6606
6607 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6608 // Return node representing slow path of predicate check.
6609 // the pseudo code we want to emulate with this predicate is:
6610 // for encryption:
6611 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6612 // for decryption:
6613 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6614 // note cipher==plain is more conservative than the original java code but that's OK
6615 //
6616
6617 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6618 // The receiver was checked for NULL already.
6619 Node* objCTR = argument(0);
6620
6621 // Load embeddedCipher field of CipherBlockChaining object.
6622 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6623
6624 // get AESCrypt klass for instanceOf check
6625 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6626 // will have same classloader as CipherBlockChaining object
6627 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6628 assert(tinst != NULL, "CTRobj is null");
6629 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6630
6631 // we want to do an instanceof comparison against the AESCrypt class
6632 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6633 if (!klass_AESCrypt->is_loaded()) {
6634 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6635 Node* ctrl = control();
6636 set_control(top()); // no regular fast path
6637 return ctrl;
6638 }
6639
6640 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6641 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6642 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6643 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6644 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6645
6646 return instof_false; // even if it is NULL
6647 }
6648
6649 //------------------------------inline_ghash_processBlocks
6650 bool LibraryCallKit::inline_ghash_processBlocks() {
6651 address stubAddr;
6652 const char *stubName;
6653 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6654
6655 stubAddr = StubRoutines::ghash_processBlocks();
6656 stubName = "ghash_processBlocks";
6657
6658 Node* data = argument(0);
6659 Node* offset = argument(1);
6660 Node* len = argument(2);
6661 Node* state = argument(3);
6662 Node* subkeyH = argument(4);
6663
6664 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6665 assert(state_start, "state is NULL");
6666 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6667 assert(subkeyH_start, "subkeyH is NULL");
6668 Node* data_start = array_element_address(data, offset, T_BYTE);
6669 assert(data_start, "data is NULL");
6670
6671 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6672 OptoRuntime::ghash_processBlocks_Type(),
6673 stubAddr, stubName, TypePtr::BOTTOM,
6674 state_start, subkeyH_start, data_start, len);
6675 return true;
6676 }
6677
6678 //------------------------------inline_sha_implCompress-----------------------
6679 //
6680 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6681 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6682 //
6683 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6684 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6685 //
6686 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6687 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6688 //
6689 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6690 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6691
6692 Node* sha_obj = argument(0);
6693 Node* src = argument(1); // type oop
6694 Node* ofs = argument(2); // type int
6695
6696 const Type* src_type = src->Value(&_gvn);
6697 const TypeAryPtr* top_src = src_type->isa_aryptr();
6698 if (top_src == NULL || top_src->klass() == NULL) {
6699 // failed array check
6700 return false;
6701 }
6702 // Figure out the size and type of the elements we will be copying.
6703 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6704 if (src_elem != T_BYTE) {
6705 return false;
6706 }
6707 // 'src_start' points to src array + offset
6708 Node* src_start = array_element_address(src, ofs, src_elem);
6709 Node* state = NULL;
6710 address stubAddr;
6711 const char *stubName;
6712
6713 switch(id) {
6714 case vmIntrinsics::_sha_implCompress:
6715 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6716 state = get_state_from_sha_object(sha_obj);
6717 stubAddr = StubRoutines::sha1_implCompress();
6718 stubName = "sha1_implCompress";
6719 break;
6720 case vmIntrinsics::_sha2_implCompress:
6721 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6722 state = get_state_from_sha_object(sha_obj);
6723 stubAddr = StubRoutines::sha256_implCompress();
6724 stubName = "sha256_implCompress";
6725 break;
6726 case vmIntrinsics::_sha5_implCompress:
6727 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6728 state = get_state_from_sha5_object(sha_obj);
6729 stubAddr = StubRoutines::sha512_implCompress();
6730 stubName = "sha512_implCompress";
6731 break;
6732 default:
6733 fatal_unexpected_iid(id);
6734 return false;
6735 }
6736 if (state == NULL) return false;
6737
6738 // Call the stub.
6739 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6740 stubAddr, stubName, TypePtr::BOTTOM,
6741 src_start, state);
6742
6743 return true;
6744 }
6745
6746 //------------------------------inline_digestBase_implCompressMB-----------------------
6747 //
6748 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6749 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6750 //
6751 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6752 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6753 "need SHA1/SHA256/SHA512 instruction support");
6754 assert((uint)predicate < 3, "sanity");
6755 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6756
6757 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6758 Node* src = argument(1); // byte[] array
6759 Node* ofs = argument(2); // type int
6760 Node* limit = argument(3); // type int
6761
6762 const Type* src_type = src->Value(&_gvn);
6763 const TypeAryPtr* top_src = src_type->isa_aryptr();
6764 if (top_src == NULL || top_src->klass() == NULL) {
6765 // failed array check
6766 return false;
6767 }
6768 // Figure out the size and type of the elements we will be copying.
6769 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6770 if (src_elem != T_BYTE) {
6771 return false;
6772 }
6773 // 'src_start' points to src array + offset
6774 Node* src_start = array_element_address(src, ofs, src_elem);
6775
6776 const char* klass_SHA_name = NULL;
6777 const char* stub_name = NULL;
6778 address stub_addr = NULL;
6779 bool long_state = false;
6780
6781 switch (predicate) {
6782 case 0:
6783 if (UseSHA1Intrinsics) {
6784 klass_SHA_name = "sun/security/provider/SHA";
6785 stub_name = "sha1_implCompressMB";
6786 stub_addr = StubRoutines::sha1_implCompressMB();
6787 }
6788 break;
6789 case 1:
6790 if (UseSHA256Intrinsics) {
6791 klass_SHA_name = "sun/security/provider/SHA2";
6792 stub_name = "sha256_implCompressMB";
6793 stub_addr = StubRoutines::sha256_implCompressMB();
6794 }
6795 break;
6796 case 2:
6797 if (UseSHA512Intrinsics) {
6798 klass_SHA_name = "sun/security/provider/SHA5";
6799 stub_name = "sha512_implCompressMB";
6800 stub_addr = StubRoutines::sha512_implCompressMB();
6801 long_state = true;
6802 }
6803 break;
6804 default:
6805 fatal("unknown SHA intrinsic predicate: %d", predicate);
6806 }
6807 if (klass_SHA_name != NULL) {
6808 // get DigestBase klass to lookup for SHA klass
6809 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6810 assert(tinst != NULL, "digestBase_obj is not instance???");
6811 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6812
6813 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6814 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6815 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6816 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6817 }
6818 return false;
6819 }
6820 //------------------------------inline_sha_implCompressMB-----------------------
6821 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6822 bool long_state, address stubAddr, const char *stubName,
6823 Node* src_start, Node* ofs, Node* limit) {
6824 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6825 const TypeOopPtr* xtype = aklass->as_instance_type();
6826 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6827 sha_obj = _gvn.transform(sha_obj);
6828
6829 Node* state;
6830 if (long_state) {
6831 state = get_state_from_sha5_object(sha_obj);
6832 } else {
6833 state = get_state_from_sha_object(sha_obj);
6834 }
6835 if (state == NULL) return false;
6836
6837 // Call the stub.
6838 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6839 OptoRuntime::digestBase_implCompressMB_Type(),
6840 stubAddr, stubName, TypePtr::BOTTOM,
6841 src_start, state, ofs, limit);
6842 // return ofs (int)
6843 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6844 set_result(result);
6845
6846 return true;
6847 }
6848
6849 enum VectorApiObjectType {
6850 VECAPI_VECTOR,
6851 VECAPI_MASK,
6852 VECAPI_SPECIES,
6853 VECAPI_SHUFFLE,
6854 };
6855
6856 static ciKlass* get_exact_klass_for_vector_box(ciKlass* ctx, BasicType bt, int num_elem, VectorApiObjectType obj_type) {
6857 // FIXME: use VM symbol table instead
6858 ResourceMark rm;
6859 stringStream ss;
6860 ss.print_raw("jdk/incubator/vector/");
6861 switch (bt) {
6862 case T_BYTE: ss.print_raw("Byte"); break;
6863 case T_SHORT: ss.print_raw("Short"); break;
6864 case T_INT: ss.print_raw("Int"); break;
6865 case T_LONG: ss.print_raw("Long"); break;
6866 case T_FLOAT: ss.print_raw("Float"); break;
6867 case T_DOUBLE: ss.print_raw("Double"); break;
6868 default: fatal("unknown element type: %d", bt);
6869 }
6870 int bits = num_elem * BitsPerByte * type2aelembytes(bt);
6871 switch (bits) {
6872 case 64: // fall through
6873 case 128: // fall through
6874 case 256: // fall through
6875 case 512: ss.print("%d", bits); break;
6876 default: fatal("unknown vector size: %d", bits);
6877 }
6878 ss.print_raw("Vector");
6879 if (obj_type != VECAPI_VECTOR) {
6880 ss.print_raw("$");
6881 switch (bt) {
6882 case T_BYTE: ss.print_raw("Byte"); break;
6883 case T_SHORT: ss.print_raw("Short"); break;
6884 case T_INT: ss.print_raw("Int"); break;
6885 case T_LONG: ss.print_raw("Long"); break;
6886 case T_FLOAT: ss.print_raw("Float"); break;
6887 case T_DOUBLE: ss.print_raw("Double"); break;
6888 default: fatal("unknown element type: %d", bt);
6889 }
6890 switch (bits) {
6891 case 64: // fall through
6892 case 128: // fall through
6893 case 256: // fall through
6894 case 512: ss.print("%d", bits); break;
6895 default: fatal("unknown vector size: %d", bits);
6896 }
6897 switch (obj_type) {
6898 case VECAPI_MASK: ss.print_raw("Mask"); break;
6899 case VECAPI_SPECIES: ss.print_raw("Species"); break;
6900 case VECAPI_SHUFFLE: ss.print_raw("Shuffle"); break;
6901 default: fatal("unknown vector object type");
6902 }
6903 }
6904
6905 ciSymbol* vector_klass_name = ciSymbol::make(ss.as_string());
6906 ciKlass* vector_klass = ctx->find_klass(vector_klass_name);
6907 return vector_klass;
6908 }
6909
6910 static BasicType getMaskBasicType(BasicType use_type) {
6911 if (use_type == T_FLOAT) {
6912 return T_INT;
6913 }
6914 if (use_type == T_DOUBLE) {
6915 return T_LONG;
6916 }
6917 assert(is_java_primitive(use_type), "must be java primitive");
6918 return use_type;
6919 }
6920
6921 enum VectorMaskUseType {
6922 VecMaskUseLoad,
6923 VecMaskUseStore,
6924 VecMaskUseAll,
6925 VecMaskNotUsed
6926 };
6927
6928 static bool arch_supports_vector(int op, int num_elem, BasicType type, VectorMaskUseType mask_use_type, int op_arity = 0) {
6929 // Check that the operation is valid.
6930 if (op <= 0) {
6931 #ifndef PRODUCT
6932 if (DebugVectorApi) {
6933 tty->print_cr("Rejected intrinsification because no valid vector op could be extracted");
6934 }
6935 #endif
6936 return false;
6937 }
6938
6939 // Check that architecture supports this op-size-type combination.
6940 if (!Matcher::match_rule_supported_vector(op, num_elem, type, op_arity)) {
6941 #ifndef PRODUCT
6942 if (DebugVectorApi) {
6943 tty->print_cr("Rejected vector op (%s,%s,%d) because architecture does not support it",
6944 NodeClassNames[op], type2name(type), num_elem);
6945 }
6946 #endif
6947 return false;
6948 } else {
6949 assert(Matcher::match_rule_supported(op), "must be supported");
6950 }
6951
6952 // Check whether mask unboxing is supported.
6953 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseLoad) {
6954 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type)) {
6955 #ifndef PRODUCT
6956 if (DebugVectorApi) {
6957 tty->print_cr("Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
6958 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
6959 }
6960 #endif
6961 return false;
6962 }
6963 }
6964
6965 // Check whether mask boxing is supported.
6966 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseStore) {
6967 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type)) {
6968 #ifndef PRODUCT
6969 if (DebugVectorApi) {
6970 tty->print_cr("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
6971 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
6972 }
6973 #endif
6974 return false;
6975 }
6976 }
6977
6978 return true;
6979 }
6980
6981 // Should be aligned with constants in jdk.incubator.vector.VectorIntrinsics.
6982 enum {
6983 // Unary
6984 OP_ABS = 0,
6985 OP_NEG = 1,
6986 OP_SQRT = 2,
6987 OP_NOT = 3,
6988 // Binary
6989 OP_ADD = 4,
6990 OP_SUB = 5,
6991 OP_MUL = 6,
6992 OP_DIV = 7,
6993 OP_MIN = 8,
6994 OP_MAX = 9,
6995 OP_AND = 10,
6996 OP_OR = 11,
6997 OP_XOR = 12,
6998 // Ternary
6999 OP_FMA = 13,
7000 // Broadcast int
7001 OP_LSHIFT = 14,
7002 OP_RSHIFT = 15,
7003 OP_URSHIFT = 16
7004 };
7005
7006 static int get_opc(jint op, BasicType bt) {
7007 switch (op) {
7008 case OP_ADD: {
7009 switch (bt) {
7010 case T_BYTE: // fall-through
7011 case T_SHORT: // fall-through
7012 case T_INT: return Op_AddI;
7013 case T_LONG: return Op_AddL;
7014 case T_FLOAT: return Op_AddF;
7015 case T_DOUBLE: return Op_AddD;
7016 default: fatal("ADD: %s", type2name(bt));
7017 }
7018 break;
7019 }
7020 case OP_SUB: {
7021 switch (bt) {
7022 case T_BYTE: // fall-through
7023 case T_SHORT: // fall-through
7024 case T_INT: return Op_SubI;
7025 case T_LONG: return Op_SubL;
7026 case T_FLOAT: return Op_SubF;
7027 case T_DOUBLE: return Op_SubD;
7028 default: fatal("SUB: %s", type2name(bt));
7029 }
7030 break;
7031 }
7032 case OP_MUL: {
7033 switch (bt) {
7034 case T_BYTE: // fall-through
7035 case T_SHORT: // fall-through
7036 case T_INT: return Op_MulI;
7037 case T_LONG: return Op_MulL;
7038 case T_FLOAT: return Op_MulF;
7039 case T_DOUBLE: return Op_MulD;
7040 default: fatal("MUL: %s", type2name(bt));
7041 }
7042 break;
7043 }
7044 case OP_DIV: {
7045 switch (bt) {
7046 case T_BYTE: // fall-through
7047 case T_SHORT: // fall-through
7048 case T_INT: return Op_DivI;
7049 case T_LONG: return Op_DivL;
7050 case T_FLOAT: return Op_DivF;
7051 case T_DOUBLE: return Op_DivD;
7052 default: fatal("DIV: %s", type2name(bt));
7053 }
7054 break;
7055 }
7056 case OP_MIN: {
7057 switch (bt) {
7058 case T_BYTE:
7059 case T_SHORT:
7060 case T_INT: return Op_MinI;
7061 case T_LONG: return Op_MinL;
7062 case T_FLOAT: return Op_MinF;
7063 case T_DOUBLE: return Op_MinD;
7064 default: fatal("MIN: %s", type2name(bt));
7065 }
7066 break;
7067 }
7068 case OP_MAX: {
7069 switch (bt) {
7070 case T_BYTE:
7071 case T_SHORT:
7072 case T_INT: return Op_MaxI;
7073 case T_LONG: return Op_MaxL;
7074 case T_FLOAT: return Op_MaxF;
7075 case T_DOUBLE: return Op_MaxD;
7076 default: fatal("MAX: %s", type2name(bt));
7077 }
7078 break;
7079 }
7080 case OP_ABS: {
7081 switch (bt) {
7082 case T_BYTE: // fall-through
7083 case T_SHORT: // fall-through
7084 case T_LONG: // fall-through
7085 case T_INT: return Op_AbsI;
7086 case T_FLOAT: return Op_AbsF;
7087 case T_DOUBLE: return Op_AbsD;
7088 default: fatal("ABS: %s", type2name(bt));
7089 }
7090 break;
7091 }
7092 case OP_NEG: {
7093 switch (bt) {
7094 case T_BYTE: // fall-through
7095 case T_SHORT: // fall-through
7096 case T_INT: return Op_NegI;
7097 case T_FLOAT: return Op_NegF;
7098 case T_DOUBLE: return Op_NegD;
7099 default: fatal("NEG: %s", type2name(bt));
7100 }
7101 break;
7102 }
7103 case OP_AND: {
7104 switch (bt) {
7105 case T_BYTE: // fall-through
7106 case T_SHORT: // fall-through
7107 case T_INT: return Op_AndI;
7108 case T_LONG: return Op_AndL;
7109 default: fatal("AND: %s", type2name(bt));
7110 }
7111 break;
7112 }
7113 case OP_OR: {
7114 switch (bt) {
7115 case T_BYTE: // fall-through
7116 case T_SHORT: // fall-through
7117 case T_INT: return Op_OrI;
7118 case T_LONG: return Op_OrL;
7119 default: fatal("OR: %s", type2name(bt));
7120 }
7121 break;
7122 }
7123 case OP_XOR: {
7124 switch (bt) {
7125 case T_BYTE: // fall-through
7126 case T_SHORT: // fall-through
7127 case T_INT: return Op_XorI;
7128 case T_LONG: return Op_XorL;
7129 default: fatal("XOR: %s", type2name(bt));
7130 }
7131 break;
7132 }
7133 case OP_SQRT: {
7134 switch (bt) {
7135 case T_FLOAT: return Op_SqrtF;
7136 case T_DOUBLE: return Op_SqrtD;
7137 default: fatal("SQRT: %s", type2name(bt));
7138 }
7139 break;
7140 }
7141 case OP_NOT: {
7142 switch (bt) {
7143 case T_BYTE: // fall-through
7144 case T_SHORT: // fall-through
7145 case T_INT: // fall-through
7146 case T_LONG: return Op_Not;
7147 default: fatal("NOT: %s", type2name(bt));
7148 }
7149 break;
7150 }
7151 case OP_FMA: {
7152 switch (bt) {
7153 case T_FLOAT: return Op_FmaF;
7154 case T_DOUBLE: return Op_FmaD;
7155 default: fatal("FMA: %s", type2name(bt));
7156 }
7157 break;
7158 }
7159 case OP_LSHIFT: {
7160 switch (bt) {
7161 case T_BYTE: // fall-through
7162 case T_SHORT: // fall-through
7163 case T_INT: return Op_LShiftI;
7164 case T_LONG: return Op_LShiftL;
7165 default: fatal("LSHIFT: %s", type2name(bt));
7166 }
7167 }
7168 case OP_RSHIFT: {
7169 switch (bt) {
7170 case T_BYTE: // fall-through
7171 case T_SHORT: // fall-through
7172 case T_INT: return Op_RShiftI;
7173 case T_LONG: return Op_RShiftL;
7174 default: fatal("RSHIFT: %s", type2name(bt));
7175 }
7176 }
7177 case OP_URSHIFT: {
7178 switch (bt) {
7179 case T_BYTE: // fall-through
7180 case T_SHORT: // fall-through
7181 case T_INT: return Op_URShiftI;
7182 case T_LONG: return Op_URShiftL;
7183 default: fatal("URSHIFT: %s", type2name(bt));
7184 }
7185 }
7186 default: fatal("unknown op: %d", op);
7187 }
7188 return 0; // Unimplemented
7189 }
7190
7191 Node* LibraryCallKit::box_vector(Node* vector, const TypeInstPtr* vbox_type,
7192 BasicType elem_bt, int num_elem) {
7193
7194 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
7195
7196 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
7197 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
7198 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true);
7199 set_i_o(_gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
7200 set_all_memory(_gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
7201 Node* ret = _gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
7202
7203 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vec_type);
7204 return _gvn.transform(vbox);
7205 }
7206
7207 Node* LibraryCallKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) {
7208 const TypeInstPtr* vbox_type_v = gvn().type(v)->is_instptr();
7209 if (vbox_type->klass() != vbox_type_v->klass()) {
7210 return NULL; // arguments don't agree on vector shapes
7211 }
7212 if (vbox_type_v->maybe_null()) {
7213 return NULL; // no nulls are allowed
7214 }
7215 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
7216 Node* unbox = gvn().transform(new VectorUnboxNode(C, vec_type, v, merged_memory()));
7217 return unbox;
7218 }
7219
7220 void LibraryCallKit::set_vector_result(Node* result, bool set_res) {
7221 if (DebugVectorApi) {
7222 tty->print("============ ");
7223 callee()->print();
7224 tty->print_cr(" ============");
7225 result->dump(5);
7226 tty->print_cr("----------------------------------------------------");
7227 }
7228 if (set_res) {
7229 set_result(result);
7230 }
7231 }
7232
7233 bool LibraryCallKit::inline_vector_nary_operation(int n) {
7234 const TypeInt* opr = gvn().type(argument(0))->is_int();
7235 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7236 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7237 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7238
7239 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7240 return false; // not enough info for intrinsification
7241 }
7242 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7243 if (!elem_type->is_primitive_type()) {
7244 return false; // should be primitive type
7245 }
7246 BasicType elem_bt = elem_type->basic_type();
7247 int num_elem = vlen->get_con();
7248 int opc = get_opc(opr->get_con(), elem_bt);
7249 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
7250 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7251 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7252
7253 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7254 if (!arch_supports_vector(sopc, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseAll : VecMaskNotUsed, n)) {
7255 return false; // not supported
7256 }
7257
7258 Node* opd1 = NULL; Node* opd2 = NULL; Node* opd3 = NULL;
7259 switch (n) {
7260 case 3: {
7261 opd1 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7262 if (opd1 == NULL) {
7263 return false;
7264 }
7265 // fall-through
7266 }
7267 case 2: {
7268 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7269 if (opd2 == NULL) {
7270 return false;
7271 }
7272 // fall-through
7273 }
7274 case 1: {
7275 opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7276 if (opd1 == NULL) {
7277 return false;
7278 }
7279 break;
7280 }
7281 default: fatal("unsupported arity: %d", n);
7282 }
7283
7284 Node* operation = NULL;
7285 switch (n) {
7286 case 1:
7287 case 2: {
7288 operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, num_elem, elem_bt));
7289 break;
7290 }
7291 case 3: {
7292 operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, opd3, num_elem, elem_bt));
7293 break;
7294 }
7295 default: fatal("unsupported arity: %d", n);
7296 }
7297 // Wrap it up in VectorBox to keep object type information.
7298 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
7299 set_vector_result(operation);
7300
7301 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7302 return true;
7303 }
7304
7305 // <V extends Vector<?,?>>
7306 // V broadcastCoerced(Class<?> vectorClass, Class<?> elementType, int vlen,
7307 // long bits,
7308 // LongFunction<V> defaultImpl)
7309 bool LibraryCallKit::inline_vector_broadcast_coerced() {
7310 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7311 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7312 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7313
7314 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7315 return false; // not enough info for intrinsification
7316 }
7317
7318 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7319 if (!elem_type->is_primitive_type()) {
7320 return false; // should be primitive type
7321 }
7322 BasicType elem_bt = elem_type->basic_type();
7323 int num_elem = vlen->get_con();
7324 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7325 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7326
7327 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7328 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt,
7329 vbox_klass->is_vectormask() ? VecMaskUseStore : VecMaskNotUsed)) {
7330 return false; // not supported
7331 }
7332
7333 Node* bits = argument(3); // long
7334
7335 Node* elem = NULL;
7336 switch (elem_bt) {
7337 case T_BOOLEAN: // fall-through
7338 case T_BYTE: // fall-through
7339 case T_SHORT: // fall-through
7340 case T_CHAR: // fall-through
7341 case T_INT: {
7342 elem = gvn().transform(new ConvL2INode(bits));
7343 break;
7344 }
7345 case T_DOUBLE: {
7346 elem = gvn().transform(new MoveL2DNode(bits));
7347 break;
7348 }
7349 case T_FLOAT: {
7350 bits = gvn().transform(new ConvL2INode(bits));
7351 elem = gvn().transform(new MoveI2FNode(bits));
7352 break;
7353 }
7354 case T_LONG: {
7355 elem = bits; // no conversion needed
7356 break;
7357 }
7358 default: fatal("%s", type2name(elem_bt));
7359 }
7360
7361 Node* broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt));
7362 broadcast = gvn().transform(broadcast);
7363 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
7364 set_vector_result(box);
7365
7366 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7367 return true;
7368 }
7369
7370 // <V extends Vector<?,?>>
7371 // V load(Class<?> vectorClass, Class<?> elementType, int vlen,
7372 // Object array, int index, /* Vector.Mask<E,S> m*/
7373 // BiFunction<Object, Integer, V> defaultImpl) {
7374 //
7375 // <V extends Vector<?,?>>
7376 // void store(Class<?> vectorClass, Class<?> elementType, int vlen,
7377 // Object array, int index, V v, /*Vector.Mask<E,S> m*/
7378 // StoreVectorOperation<V> defaultImpl) {
7379
7380 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
7381 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7382 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7383 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7384
7385 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7386 return false; // not enough info for intrinsification
7387 }
7388
7389 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7390 if (!elem_type->is_primitive_type()) {
7391 return false; // should be primitive type
7392 }
7393 BasicType elem_bt = elem_type->basic_type();
7394 int num_elem = vlen->get_con();
7395
7396 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7397 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
7398 return false; // not supported
7399 }
7400
7401 Node* arr = argument(3);
7402 Node* idx = argument(4);
7403
7404 const TypeAryPtr* arr_type = gvn().type(arr)->isa_aryptr();
7405 if (arr_type == NULL) {
7406 return false; // should be an array
7407 }
7408 if (elem_bt != arr_type->elem()->array_element_basic_type()) {
7409 return false; // array & vector element types should be the same
7410 }
7411 Node* adr = array_element_address(arr, idx, elem_bt);
7412 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
7413
7414 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7415 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7416
7417 if (is_store) {
7418 Node* val = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7419 if (val == NULL) {
7420 return false; // operand unboxing failed
7421 }
7422 set_all_memory(reset_memory());
7423 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(adr), adr, adr_type, val, num_elem));
7424 set_memory(vstore, adr_type);
7425 set_vector_result(vstore, false);
7426 } else {
7427 Node* vload = gvn().transform(LoadVectorNode::make(0, control(), memory(adr), adr, adr_type, num_elem, elem_bt));
7428 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7429 set_vector_result(box);
7430 }
7431
7432 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7433 return true;
7434 }
7435
7436 // <V extends Vector<?,?>>
7437 // long reductionCoerced(int oprId, Class<?> vectorClass, Class<?> elementType, int vlen,
7438 // V v,
7439 // Function<V,Long> defaultImpl)
7440
7441 bool LibraryCallKit::inline_vector_reduction() {
7442 const TypeInt* opr = gvn().type(argument(0))->is_int();
7443 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7444 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7445 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7446
7447 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7448 return false; // not enough info for intrinsification
7449 }
7450 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7451 if (!elem_type->is_primitive_type()) {
7452 return false; // should be primitive type
7453 }
7454 BasicType elem_bt = elem_type->basic_type();
7455 int num_elem = vlen->get_con();
7456
7457 int opc = get_opc(opr->get_con(), elem_bt);
7458 int sopc = ReductionNode::opcode(opc, elem_bt);
7459
7460 // FIXME: When encountering a SubReduction, we want to check for support of
7461 // the corresponding AddReduction node.
7462 if (sopc == Op_SubReductionV) {
7463 if (gvn().type(argument(2))->isa_int()) {
7464 sopc = Op_AddReductionVI;
7465 } else if (gvn().type(argument(2))->isa_long()) {
7466 sopc = Op_AddReductionVL;
7467 }
7468 }
7469
7470 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7471 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7472 return false;
7473 }
7474
7475 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7476 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7477
7478 Node* opd = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7479 if (opd == NULL) {
7480 return false; // operand unboxing failed
7481 }
7482
7483 Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt);
7484 Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, elem_bt));
7485
7486 Node* bits = NULL;
7487 switch (elem_bt) {
7488 case T_BYTE:
7489 case T_SHORT:
7490 case T_INT: {
7491 bits = gvn().transform(new ConvI2LNode(rn));
7492 break;
7493 }
7494 case T_FLOAT: {
7495 rn = gvn().transform(new MoveF2INode(rn));
7496 bits = gvn().transform(new ConvI2LNode(rn));
7497 break;
7498 }
7499 case T_DOUBLE: {
7500 bits = gvn().transform(new MoveD2LNode(rn));
7501 break;
7502 }
7503 case T_LONG: {
7504 bits = rn; // no conversion needed
7505 break;
7506 }
7507 default: fatal("%s", type2name(elem_bt));
7508 }
7509 set_vector_result(bits);
7510 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7511 return true;
7512 }
7513
7514 // static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
7515 // V v1, V v2,
7516 // BiFunction<V, V, Boolean> defaultImpl) {
7517
7518 bool LibraryCallKit::inline_vector_test() {
7519 const TypeInt* cond = gvn().type(argument(0))->is_int();
7520 const TypeInstPtr* vector_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 (!cond->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7525 return false; // not enough info for intrinsification
7526 }
7527 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7528 if (!elem_type->is_primitive_type()) {
7529 return false; // should be primitive type
7530 }
7531 BasicType elem_bt = elem_type->basic_type();
7532 int num_elem = vlen->get_con();
7533 Assembler::Condition booltest = (Assembler::Condition)cond->get_con();
7534 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7535 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7536
7537 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseLoad : VecMaskNotUsed)) {
7538 return false;
7539 }
7540
7541 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7542 Node* opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7543 if (opd1 == NULL || opd2 == NULL) {
7544 return false; // operand unboxing failed
7545 }
7546 Node* test = new VectorTestNode(opd1, opd2, booltest);
7547 test = _gvn.transform(test);
7548 set_vector_result(test);
7549 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7550 return true;
7551 }
7552
7553 // static
7554 // <V extends Vector, M extends Mask>
7555 // V blend(Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7556 // V v1, V v2, M m,
7557 // VectorBlendOp<V,M> defaultImpl) { ...
7558 //
7559 bool LibraryCallKit::inline_vector_blend() {
7560 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7561 const TypeInstPtr* mask_klass = gvn().type(argument(1))->is_instptr();
7562 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7563 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7564
7565 if (mask_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7566 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7567 return false; // not enough info for intrinsification
7568 }
7569 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7570 if (!elem_type->is_primitive_type()) {
7571 return false; // should be primitive type
7572 }
7573 BasicType elem_bt = elem_type->basic_type();
7574 BasicType mask_bt = getMaskBasicType(elem_bt);
7575 int num_elem = vlen->get_con();
7576
7577 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
7578 return false; // not supported
7579 }
7580 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7581 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7582
7583 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7584 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7585
7586 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7587 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7588 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
7589
7590 if (v1 == NULL || v2 == NULL || mask == NULL) {
7591 return false; // operand unboxing failed
7592 }
7593
7594 Node* blend = _gvn.transform(new VectorBlendNode(v1, v2, mask));
7595 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
7596 set_vector_result(box);
7597
7598 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7599 return true;
7600 }
7601
7602 // static <V extends Vector<E,S>,
7603 // M extends Vector.Mask<E,S>,
7604 // S extends Vector.Shape, E>
7605 // M compare(int cond, Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7606 // V v1, V v2,
7607 // VectorCompareOp<V,M> defaultImpl) { ...
7608 //
7609 bool LibraryCallKit::inline_vector_compare() {
7610 const TypeInt* cond = gvn().type(argument(0))->is_int();
7611 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7612 const TypeInstPtr* mask_klass = gvn().type(argument(2))->is_instptr();
7613 const TypeInstPtr* elem_klass = gvn().type(argument(3))->is_instptr();
7614 const TypeInt* vlen = gvn().type(argument(4))->is_int();
7615
7616 if (!cond->is_con() || vector_klass->const_oop() == NULL || mask_klass->const_oop() == NULL ||
7617 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7618 return false; // not enough info for intrinsification
7619 }
7620 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7621 if (!elem_type->is_primitive_type()) {
7622 return false; // should be primitive type
7623 }
7624
7625 int num_elem = vlen->get_con();
7626 BasicType elem_bt = elem_type->basic_type();
7627 BasicType mask_bt = getMaskBasicType(elem_bt);
7628
7629 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
7630 return false;
7631 }
7632
7633 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7634 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7635
7636 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7637 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7638
7639 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7640 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7641
7642 if (v1 == NULL || v2 == NULL) {
7643 return false; // operand unboxing failed
7644 }
7645 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
7646 const TypeVect* vt = TypeVect::make(mask_bt, num_elem);
7647 Node* operation = _gvn.transform(new VectorMaskCmpNode(pred, v1, v2, vt));
7648
7649 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
7650 set_vector_result(box);
7651
7652 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7653 return true;
7654 }
7655
7656 Node* LibraryCallKit::shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
7657 assert(bt == T_INT || bt == T_LONG, "Only long and int are supported");
7658 juint mask = (bt == T_INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
7659 const TypeInt* t = cnt->find_int_type();
7660 if (t != NULL && t->is_con()) {
7661 juint shift = t->get_con();
7662 if (shift > mask) {
7663 return _gvn.transform(ConNode::make(TypeInt::make(shift & mask)));
7664 } else {
7665 return cnt;
7666 }
7667 } else {
7668 Node* maskedcnt = cnt;
7669 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
7670 Node* nmask = _gvn.transform(ConNode::make(TypeInt::make(mask)));
7671 maskedcnt = _gvn.transform(new AndINode(cnt, nmask));
7672 }
7673 return VectorNode::shift_count(shift_op, maskedcnt, num_elem, bt);
7674 }
7675 }
7676
7677 // static
7678 // <V extends Vector<?,?>>
7679 // V broadcastInt(int opr, Class<V> vectorClass, Class<?> elementType, int vlen,
7680 // V v, int i,
7681 // VectorBroadcastIntOp<V> defaultImpl) {
7682 //
7683 bool LibraryCallKit::inline_vector_broadcast_int() {
7684 const TypeInt* opr = gvn().type(argument(0))->is_int();
7685 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7686 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7687 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7688
7689 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7690 return false; // not enough info for intrinsification
7691 }
7692 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7693 if (!elem_type->is_primitive_type()) {
7694 return false; // should be primitive type
7695 }
7696 BasicType elem_bt = elem_type->basic_type();
7697 int num_elem = vlen->get_con();
7698 int opc = get_opc(opr->get_con(), elem_bt);
7699 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
7700 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7701 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7702
7703 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7704 return false; // not supported
7705 }
7706 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7707 Node* opd2 = shift_count(argument(5), opc, elem_bt, num_elem);
7708 if (opd1 == NULL || opd2 == NULL) {
7709 return false;
7710 }
7711 Node* operation = _gvn.transform(VectorNode::make(sopc, opd1, opd2, num_elem, elem_bt));
7712
7713 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
7714 set_vector_result(vbox);
7715 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7716 return true;
7717 }
7718
7719 bool LibraryCallKit::inline_vector_reinterpret() {
7720 const TypeInstPtr* vector_klass_from = gvn().type(argument(0))->is_instptr();
7721 const TypeInstPtr* elem_klass_from = gvn().type(argument(1))->is_instptr();
7722 const TypeInt* vlen_from = gvn().type(argument(2))->is_int();
7723
7724 const TypeInstPtr* elem_klass_to = gvn().type(argument(3))->is_instptr();
7725 const TypeInt* vlen_to = gvn().type(argument(4))->is_int();
7726
7727 if (vector_klass_from->const_oop() == NULL || elem_klass_from->const_oop() == NULL || !vlen_from->is_con() ||
7728 elem_klass_to->const_oop() == NULL || !vlen_to->is_con()) {
7729 return false; // not enough info for intrinsification
7730 }
7731
7732 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
7733 if (!vbox_klass_from->is_vectorapi_vector()) {
7734 return false; // only vector & mask are supported
7735 }
7736 bool is_mask = vbox_klass_from->is_vectormask();
7737
7738 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
7739 if (!elem_type_from->is_primitive_type()) {
7740 return false; // should be primitive type
7741 }
7742 BasicType elem_bt_from = elem_type_from->basic_type();
7743 if (is_mask) {
7744 elem_bt_from = getMaskBasicType(elem_bt_from);
7745 }
7746 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
7747 if (!elem_type_to->is_primitive_type()) {
7748 return false; // should be primitive type
7749 }
7750 BasicType elem_bt_to = elem_type_to->basic_type();
7751 if (is_mask) {
7752 elem_bt_to = getMaskBasicType(elem_bt_to);
7753 }
7754 if (is_mask && elem_bt_from != elem_bt_to) {
7755 return false; // type mismatch
7756 }
7757 int num_elem_from = vlen_from->get_con();
7758 int num_elem_to = vlen_to->get_con();
7759 if (is_mask) {
7760 elem_bt_to = getMaskBasicType(elem_bt_to);
7761 }
7762
7763 // Check whether we can unbox to appropriate size.
7764 if (!arch_supports_vector(Op_VectorReinterpret,
7765 num_elem_from,
7766 elem_bt_from,
7767 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
7768 return false;
7769 }
7770
7771 // Check whether we can support resizing/reinterpreting to the new size.
7772 if (!arch_supports_vector(Op_VectorReinterpret,
7773 num_elem_to,
7774 elem_bt_to,
7775 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
7776 return false;
7777 }
7778
7779 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
7780
7781 Node* opd1 = unbox_vector(argument(5), vbox_type_from, elem_bt_from, num_elem_from);
7782 if (opd1 == NULL) {
7783 return false;
7784 }
7785
7786 // Can assert when Phi merges vectors of different types:
7787 // # Internal Error (/Users/vlivanov/ws/jdk/panama-dev/open/src/hotspot/share/opto/type.cpp:2291), pid=67536, tid=14083
7788 // # Error: assert(length() == v->length()) failed
7789 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from);
7790 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to);
7791 Node* op = opd1;
7792 if (Type::cmp(src_type, dst_type) != 0) {
7793 op = _gvn.transform(new VectorReinterpretNode(op, src_type, dst_type));
7794 }
7795 ciKlass* vbox_klass_to = get_exact_klass_for_vector_box(vbox_klass_from, elem_type_to->basic_type(),
7796 num_elem_to, is_mask ? VECAPI_MASK : VECAPI_VECTOR);
7797 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
7798 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
7799 set_vector_result(vbox);
7800 return true;
7801 }
7802
7803 //------------------------------get_state_from_sha_object-----------------------
7804 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
7805 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
7806 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
7807 if (sha_state == NULL) return (Node *) NULL;
7808
7809 // now have the array, need to get the start address of the state array
7810 Node* state = array_element_address(sha_state, intcon(0), T_INT);
7811 return state;
7812 }
7813
7814 //------------------------------get_state_from_sha5_object-----------------------
7815 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
7816 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
7817 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
7818 if (sha_state == NULL) return (Node *) NULL;
7819
7820 // now have the array, need to get the start address of the state array
7821 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
7822 return state;
7823 }
7824
7825 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
7826 // Return node representing slow path of predicate check.
7827 // the pseudo code we want to emulate with this predicate is:
7828 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
7829 //
7830 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
7831 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
7832 "need SHA1/SHA256/SHA512 instruction support");
7833 assert((uint)predicate < 3, "sanity");
7834
7835 // The receiver was checked for NULL already.
7836 Node* digestBaseObj = argument(0);
7837
7838 // get DigestBase klass for instanceOf check
7839 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
7840 assert(tinst != NULL, "digestBaseObj is null");
7841 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
7842
7843 const char* klass_SHA_name = NULL;
7844 switch (predicate) {
7845 case 0:
7846 if (UseSHA1Intrinsics) {
7847 // we want to do an instanceof comparison against the SHA class
7848 klass_SHA_name = "sun/security/provider/SHA";
7849 }
7850 break;
7851 case 1:
7852 if (UseSHA256Intrinsics) {
7853 // we want to do an instanceof comparison against the SHA2 class
7854 klass_SHA_name = "sun/security/provider/SHA2";
7855 }
7856 break;
7857 case 2:
7858 if (UseSHA512Intrinsics) {
7859 // we want to do an instanceof comparison against the SHA5 class
7860 klass_SHA_name = "sun/security/provider/SHA5";
7861 }
7862 break;
7863 default:
7864 fatal("unknown SHA intrinsic predicate: %d", predicate);
7865 }
7866
7867 ciKlass* klass_SHA = NULL;
7868 if (klass_SHA_name != NULL) {
7869 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
7870 }
7871 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
7872 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
7873 Node* ctrl = control();
7874 set_control(top()); // no intrinsic path
7875 return ctrl;
7876 }
7877 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
7878
7879 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
7880 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
7881 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
7882 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
7883
7884 return instof_false; // even if it is NULL
7885 }
7886
7887 //-------------inline_fma-----------------------------------
7888 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
7889 Node *a = NULL;
7890 Node *b = NULL;
7891 Node *c = NULL;
7892 Node* result = NULL;
7893 switch (id) {
7894 case vmIntrinsics::_fmaD:
7895 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
7896 // no receiver since it is static method
7897 a = round_double_node(argument(0));
7898 b = round_double_node(argument(2));
7899 c = round_double_node(argument(4));
7900 result = _gvn.transform(new FmaDNode(control(), a, b, c));
7901 break;
7902 case vmIntrinsics::_fmaF:
7903 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
7904 a = argument(0);
7905 b = argument(1);
7906 c = argument(2);
7907 result = _gvn.transform(new FmaFNode(control(), a, b, c));
7908 break;
7909 default:
7910 fatal_unexpected_iid(id); break;
7911 }
7912 set_result(result);
7913 return true;
7914 }
7915
7916 bool LibraryCallKit::inline_profileBoolean() {
7917 Node* counts = argument(1);
7918 const TypeAryPtr* ary = NULL;
7919 ciArray* aobj = NULL;
7920 if (counts->is_Con()
7921 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
7922 && (aobj = ary->const_oop()->as_array()) != NULL
7923 && (aobj->length() == 2)) {
7924 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
7925 jint false_cnt = aobj->element_value(0).as_int();
7926 jint true_cnt = aobj->element_value(1).as_int();
7927
7928 if (C->log() != NULL) {
7929 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
7930 false_cnt, true_cnt);
7931 }
7932
7933 if (false_cnt + true_cnt == 0) {
7934 // According to profile, never executed.
7935 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7936 Deoptimization::Action_reinterpret);
7937 return true;
7938 }
7939
7940 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
7941 // is a number of each value occurrences.
7942 Node* result = argument(0);
7943 if (false_cnt == 0 || true_cnt == 0) {
7944 // According to profile, one value has been never seen.
7945 int expected_val = (false_cnt == 0) ? 1 : 0;
7946
7947 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
7948 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
7949
7950 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
7951 Node* fast_path = _gvn.transform(new IfTrueNode(check));
7952 Node* slow_path = _gvn.transform(new IfFalseNode(check));
7953
7954 { // Slow path: uncommon trap for never seen value and then reexecute
7955 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
7956 // the value has been seen at least once.
7957 PreserveJVMState pjvms(this);
7958 PreserveReexecuteState preexecs(this);
7959 jvms()->set_should_reexecute(true);
7960
7961 set_control(slow_path);
7962 set_i_o(i_o());
7963
7964 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7965 Deoptimization::Action_reinterpret);
7966 }
7967 // The guard for never seen value enables sharpening of the result and
7968 // returning a constant. It allows to eliminate branches on the same value
7969 // later on.
7970 set_control(fast_path);
7971 result = intcon(expected_val);
7972 }
7973 // Stop profiling.
7974 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
7975 // By replacing method body with profile data (represented as ProfileBooleanNode
7976 // on IR level) we effectively disable profiling.
7977 // It enables full speed execution once optimized code is generated.
7978 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
7979 C->record_for_igvn(profile);
7980 set_result(profile);
7981 return true;
7982 } else {
7983 // Continue profiling.
7984 // Profile data isn't available at the moment. So, execute method's bytecode version.
7985 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
7986 // is compiled and counters aren't available since corresponding MethodHandle
7987 // isn't a compile-time constant.
7988 return false;
7989 }
7990 }
7991
7992 bool LibraryCallKit::inline_isCompileConstant() {
7993 Node* n = argument(0);
7994 set_result(n->is_Con() ? intcon(1) : intcon(0));
7995 return true;
7996 }