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 "gc/shared/barrierSet.hpp"
33 #include "jfr/support/jfrIntrinsics.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/objArrayKlass.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/c2compiler.hpp"
39 #include "opto/callGenerator.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/cfgnode.hpp"
42 #include "opto/convertnode.hpp"
43 #include "opto/countbitsnode.hpp"
44 #include "opto/intrinsicnode.hpp"
45 #include "opto/idealKit.hpp"
46 #include "opto/mathexactnode.hpp"
47 #include "opto/movenode.hpp"
48 #include "opto/mulnode.hpp"
49 #include "opto/narrowptrnode.hpp"
50 #include "opto/opaquenode.hpp"
51 #include "opto/parse.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/subnode.hpp"
55 #include "opto/vectornode.hpp"
56 #include "prims/nativeLookup.hpp"
57 #include "prims/unsafe.hpp"
58 #include "runtime/objectMonitor.hpp"
59 #include "runtime/sharedRuntime.hpp"
60 #include "utilities/macros.hpp"
61
62
63 class LibraryIntrinsic : public InlineCallGenerator {
64 // Extend the set of intrinsics known to the runtime:
65 public:
66 private:
67 bool _is_virtual;
68 bool _does_virtual_dispatch;
69 int8_t _predicates_count; // Intrinsic is predicated by several conditions
70 int8_t _last_predicate; // Last generated predicate
71 vmIntrinsics::ID _intrinsic_id;
72
73 public:
74 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
75 : InlineCallGenerator(m),
76 _is_virtual(is_virtual),
77 _does_virtual_dispatch(does_virtual_dispatch),
78 _predicates_count((int8_t)predicates_count),
79 _last_predicate((int8_t)-1),
80 _intrinsic_id(id)
81 {
82 }
83 virtual bool is_intrinsic() const { return true; }
84 virtual bool is_virtual() const { return _is_virtual; }
85 virtual bool is_predicated() const { return _predicates_count > 0; }
86 virtual int predicates_count() const { return _predicates_count; }
87 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
88 virtual JVMState* generate(JVMState* jvms);
89 virtual Node* generate_predicate(JVMState* jvms, int predicate);
90 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
91 };
92
93
94 // Local helper class for LibraryIntrinsic:
95 class LibraryCallKit : public GraphKit {
96 private:
97 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
98 Node* _result; // the result node, if any
99 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
100
101 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
102
103 public:
104 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
105 : GraphKit(jvms),
106 _intrinsic(intrinsic),
107 _result(NULL)
108 {
109 // Check if this is a root compile. In that case we don't have a caller.
110 if (!jvms->has_method()) {
111 _reexecute_sp = sp();
112 } else {
113 // Find out how many arguments the interpreter needs when deoptimizing
114 // and save the stack pointer value so it can used by uncommon_trap.
115 // We find the argument count by looking at the declared signature.
116 bool ignored_will_link;
117 ciSignature* declared_signature = NULL;
118 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
119 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
120 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
121 }
122 }
123
124 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
125
126 ciMethod* caller() const { return jvms()->method(); }
127 int bci() const { return jvms()->bci(); }
128 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
129 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
130 ciMethod* callee() const { return _intrinsic->method(); }
131
132 bool try_to_inline(int predicate);
133 Node* try_to_predicate(int predicate);
134
135 void push_result() {
136 // Push the result onto the stack.
137 if (!stopped() && result() != NULL) {
138 BasicType bt = result()->bottom_type()->basic_type();
139 push_node(bt, result());
140 }
141 }
142
143 private:
144 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
145 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
146 }
147
148 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
149 void set_result(RegionNode* region, PhiNode* value);
150 Node* result() { return _result; }
151
152 virtual int reexecute_sp() { return _reexecute_sp; }
153
154 // Helper functions to inline natives
155 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
156 Node* generate_slow_guard(Node* test, RegionNode* region);
157 Node* generate_fair_guard(Node* test, RegionNode* region);
158 Node* generate_negative_guard(Node* index, RegionNode* region,
159 // resulting CastII of index:
160 Node* *pos_index = NULL);
161 Node* generate_limit_guard(Node* offset, Node* subseq_length,
162 Node* array_length,
163 RegionNode* region);
164 void generate_string_range_check(Node* array, Node* offset,
165 Node* length, bool char_count);
166 Node* generate_current_thread(Node* &tls_output);
167 Node* load_mirror_from_klass(Node* klass);
168 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
169 RegionNode* region, int null_path,
170 int offset);
171 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
172 RegionNode* region, int null_path) {
173 int offset = java_lang_Class::klass_offset_in_bytes();
174 return load_klass_from_mirror_common(mirror, never_see_null,
175 region, null_path,
176 offset);
177 }
178 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
179 RegionNode* region, int null_path) {
180 int offset = java_lang_Class::array_klass_offset_in_bytes();
181 return load_klass_from_mirror_common(mirror, never_see_null,
182 region, null_path,
183 offset);
184 }
185 Node* generate_access_flags_guard(Node* kls,
186 int modifier_mask, int modifier_bits,
187 RegionNode* region);
188 Node* generate_interface_guard(Node* kls, RegionNode* region);
189 Node* generate_array_guard(Node* kls, RegionNode* region) {
190 return generate_array_guard_common(kls, region, false, false);
191 }
192 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
193 return generate_array_guard_common(kls, region, false, true);
194 }
195 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
196 return generate_array_guard_common(kls, region, true, false);
197 }
198 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
199 return generate_array_guard_common(kls, region, true, true);
200 }
201 Node* generate_array_guard_common(Node* kls, RegionNode* region,
202 bool obj_array, bool not_array);
203 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
204 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
205 bool is_virtual = false, bool is_static = false);
206 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
207 return generate_method_call(method_id, false, true);
208 }
209 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
210 return generate_method_call(method_id, true, false);
211 }
212 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
213 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
214
215 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
216 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
217 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
219 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
220 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
221 bool inline_string_indexOfChar();
222 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
223 bool inline_string_toBytesU();
224 bool inline_string_getCharsU();
225 bool inline_string_copy(bool compress);
226 bool inline_string_char_access(bool is_store);
227 Node* round_double_node(Node* n);
228 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
229 bool inline_math_native(vmIntrinsics::ID id);
230 bool inline_math(vmIntrinsics::ID id);
231 template <typename OverflowOp>
232 bool inline_math_overflow(Node* arg1, Node* arg2);
233 void inline_math_mathExact(Node* math, Node* test);
234 bool inline_math_addExactI(bool is_increment);
235 bool inline_math_addExactL(bool is_increment);
236 bool inline_math_multiplyExactI();
237 bool inline_math_multiplyExactL();
238 bool inline_math_multiplyHigh();
239 bool inline_math_negateExactI();
240 bool inline_math_negateExactL();
241 bool inline_math_subtractExactI(bool is_decrement);
242 bool inline_math_subtractExactL(bool is_decrement);
243 bool inline_min_max(vmIntrinsics::ID id);
244 bool inline_notify(vmIntrinsics::ID id);
245 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
246 // This returns Type::AnyPtr, RawPtr, or OopPtr.
247 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
248 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
249
250 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
251 DecoratorSet mo_decorator_for_access_kind(AccessKind kind);
252 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
253 static bool klass_needs_init_guard(Node* kls);
254 bool inline_unsafe_allocate();
255 bool inline_unsafe_newArray(bool uninitialized);
256 bool inline_unsafe_copyMemory();
257 bool inline_native_currentThread();
258
259 bool inline_native_time_funcs(address method, const char* funcName);
260 #ifdef JFR_HAVE_INTRINSICS
261 bool inline_native_classID();
262 bool inline_native_getEventWriter();
263 #endif
264 bool inline_native_isInterrupted();
265 bool inline_native_Class_query(vmIntrinsics::ID id);
266 bool inline_native_subtype_check();
267 bool inline_native_getLength();
268 bool inline_array_copyOf(bool is_copyOfRange);
269 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
270 bool inline_preconditions_checkIndex();
271 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array);
272 bool inline_native_clone(bool is_virtual);
273 bool inline_native_Reflection_getCallerClass();
274 // Helper function for inlining native object hash method
275 bool inline_native_hashcode(bool is_virtual, bool is_static);
276 bool inline_native_getClass();
277
278 // Helper functions for inlining arraycopy
279 bool inline_arraycopy();
280 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
281 RegionNode* slow_region);
282 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
283 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
284 uint new_idx);
285
286 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
287 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
288 bool inline_unsafe_fence(vmIntrinsics::ID id);
289 bool inline_onspinwait();
290 bool inline_fp_conversions(vmIntrinsics::ID id);
291 bool inline_number_methods(vmIntrinsics::ID id);
292 bool inline_reference_get();
293 bool inline_Class_cast();
294 bool inline_aescrypt_Block(vmIntrinsics::ID id);
295 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
296 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
297 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
298 Node* inline_counterMode_AESCrypt_predicate();
299 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
300 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
301 bool inline_ghash_processBlocks();
302 bool inline_base64_encodeBlock();
303 bool inline_sha_implCompress(vmIntrinsics::ID id);
304 bool inline_digestBase_implCompressMB(int predicate);
305 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
306 bool long_state, address stubAddr, const char *stubName,
307 Node* src_start, Node* ofs, Node* limit);
308 Node* get_state_from_sha_object(Node *sha_object);
309 Node* get_state_from_sha5_object(Node *sha_object);
310 Node* inline_digestBase_implCompressMB_predicate(int predicate);
311 bool inline_encodeISOArray();
312 bool inline_updateCRC32();
313 bool inline_updateBytesCRC32();
314 bool inline_updateByteBufferCRC32();
315 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
316 bool inline_updateBytesCRC32C();
317 bool inline_updateDirectByteBufferCRC32C();
318 bool inline_updateBytesAdler32();
319 bool inline_updateByteBufferAdler32();
320 bool inline_multiplyToLen();
321 bool inline_hasNegatives();
322 bool inline_squareToLen();
323 bool inline_mulAdd();
324 bool inline_montgomeryMultiply();
325 bool inline_montgomerySquare();
326 bool inline_vectorizedMismatch();
327 bool inline_fma(vmIntrinsics::ID id);
328
329 bool inline_profileBoolean();
330 bool inline_isCompileConstant();
331
332 // Vector API support
333 bool inline_vector_nary_operation(int n);
334 bool inline_vector_broadcast_coerced();
335 bool inline_vector_mem_operation(bool is_store);
336 bool inline_vector_reduction();
337 bool inline_vector_test();
338 bool inline_vector_blend();
339 bool inline_vector_rearrange();
340 bool inline_vector_compare();
341 bool inline_vector_broadcast_int();
342 bool inline_vector_cast_reinterpret(bool is_cast);
343 bool inline_vector_extract();
344 bool inline_vector_insert();
345 Node* box_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
346 Node* unbox_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
347 Node* shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem);
348 Node* gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2);
349 void set_vector_result(Node* result, bool set_res = true);
350
351 void clear_upper_avx() {
352 #ifdef X86
353 if (UseAVX >= 2) {
354 C->set_clear_upper_avx(true);
355 }
356 #endif
357 }
358 };
359
360 //---------------------------make_vm_intrinsic----------------------------
361 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
362 vmIntrinsics::ID id = m->intrinsic_id();
363 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
364
365 if (!m->is_loaded()) {
366 // Do not attempt to inline unloaded methods.
367 return NULL;
368 }
369
370 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
371 bool is_available = false;
372
373 {
374 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
375 // the compiler must transition to '_thread_in_vm' state because both
376 // methods access VM-internal data.
377 VM_ENTRY_MARK;
378 methodHandle mh(THREAD, m->get_Method());
379 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
380 !C->directive()->is_intrinsic_disabled(mh) &&
381 !vmIntrinsics::is_disabled_by_flags(mh);
382
383 }
384
385 if (is_available) {
386 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
387 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
388 return new LibraryIntrinsic(m, is_virtual,
389 vmIntrinsics::predicates_needed(id),
390 vmIntrinsics::does_virtual_dispatch(id),
391 (vmIntrinsics::ID) id);
392 } else {
393 return NULL;
394 }
395 }
396
397 //----------------------register_library_intrinsics-----------------------
398 // Initialize this file's data structures, for each Compile instance.
399 void Compile::register_library_intrinsics() {
400 // Nothing to do here.
401 }
402
403 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
404 LibraryCallKit kit(jvms, this);
405 Compile* C = kit.C;
406 int nodes = C->unique();
407 #ifndef PRODUCT
408 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
409 char buf[1000];
410 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
411 tty->print_cr("Intrinsic %s", str);
412 }
413 #endif
414 ciMethod* callee = kit.callee();
415 const int bci = kit.bci();
416
417 // Try to inline the intrinsic.
418 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
419 kit.try_to_inline(_last_predicate)) {
420 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
421 : "(intrinsic)";
422 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
423 if (C->print_intrinsics() || C->print_inlining()) {
424 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
425 }
426 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
427 if (C->log()) {
428 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
429 vmIntrinsics::name_at(intrinsic_id()),
430 (is_virtual() ? " virtual='1'" : ""),
431 C->unique() - nodes);
432 }
433 // Push the result from the inlined method onto the stack.
434 kit.push_result();
435 C->print_inlining_update(this);
436 return kit.transfer_exceptions_into_jvms();
437 }
438
439 // The intrinsic bailed out
440 if (jvms->has_method()) {
441 // Not a root compile.
442 const char* msg;
443 if (callee->intrinsic_candidate()) {
444 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
445 } else {
446 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
447 : "failed to inline (intrinsic), method not annotated";
448 }
449 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
450 if (C->print_intrinsics() || C->print_inlining()) {
451 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
452 }
453 } else {
454 // Root compile
455 ResourceMark rm;
456 stringStream msg_stream;
457 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
458 vmIntrinsics::name_at(intrinsic_id()),
459 is_virtual() ? " (virtual)" : "", bci);
460 const char *msg = msg_stream.as_string();
461 log_debug(jit, inlining)("%s", msg);
462 if (C->print_intrinsics() || C->print_inlining()) {
463 tty->print("%s", msg);
464 }
465 }
466 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
467 C->print_inlining_update(this);
468 return NULL;
469 }
470
471 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
472 LibraryCallKit kit(jvms, this);
473 Compile* C = kit.C;
474 int nodes = C->unique();
475 _last_predicate = predicate;
476 #ifndef PRODUCT
477 assert(is_predicated() && predicate < predicates_count(), "sanity");
478 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
479 char buf[1000];
480 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
481 tty->print_cr("Predicate for intrinsic %s", str);
482 }
483 #endif
484 ciMethod* callee = kit.callee();
485 const int bci = kit.bci();
486
487 Node* slow_ctl = kit.try_to_predicate(predicate);
488 if (!kit.failing()) {
489 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
490 : "(intrinsic, predicate)";
491 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
492 if (C->print_intrinsics() || C->print_inlining()) {
493 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
494 }
495 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
496 if (C->log()) {
497 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
498 vmIntrinsics::name_at(intrinsic_id()),
499 (is_virtual() ? " virtual='1'" : ""),
500 C->unique() - nodes);
501 }
502 return slow_ctl; // Could be NULL if the check folds.
503 }
504
505 // The intrinsic bailed out
506 if (jvms->has_method()) {
507 // Not a root compile.
508 const char* msg = "failed to generate predicate for intrinsic";
509 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
510 if (C->print_intrinsics() || C->print_inlining()) {
511 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
512 }
513 } else {
514 // Root compile
515 ResourceMark rm;
516 stringStream msg_stream;
517 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
518 vmIntrinsics::name_at(intrinsic_id()),
519 is_virtual() ? " (virtual)" : "", bci);
520 const char *msg = msg_stream.as_string();
521 log_debug(jit, inlining)("%s", msg);
522 if (C->print_intrinsics() || C->print_inlining()) {
523 C->print_inlining_stream()->print("%s", msg);
524 }
525 }
526 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
527 return NULL;
528 }
529
530 bool LibraryCallKit::try_to_inline(int predicate) {
531 // Handle symbolic names for otherwise undistinguished boolean switches:
532 const bool is_store = true;
533 const bool is_compress = true;
534 const bool is_static = true;
535 const bool is_volatile = true;
536
537 if (!jvms()->has_method()) {
538 // Root JVMState has a null method.
539 assert(map()->memory()->Opcode() == Op_Parm, "");
540 // Insert the memory aliasing node
541 set_all_memory(reset_memory());
542 }
543 assert(merged_memory(), "");
544
545 switch (intrinsic_id()) {
546 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
547 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
548 case vmIntrinsics::_getClass: return inline_native_getClass();
549
550 case vmIntrinsics::_dsin:
551 case vmIntrinsics::_dcos:
552 case vmIntrinsics::_dtan:
553 case vmIntrinsics::_dabs:
554 case vmIntrinsics::_datan2:
555 case vmIntrinsics::_dsqrt:
556 case vmIntrinsics::_dexp:
557 case vmIntrinsics::_dlog:
558 case vmIntrinsics::_dlog10:
559 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
560
561 case vmIntrinsics::_min:
562 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
563
564 case vmIntrinsics::_notify:
565 case vmIntrinsics::_notifyAll:
566 if (ObjectMonitor::Knob_InlineNotify) {
567 return inline_notify(intrinsic_id());
568 }
569 return false;
570
571 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
572 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
573 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
574 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
575 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
576 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
577 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
578 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
579 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
580 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
581 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
582 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
583 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
584
585 case vmIntrinsics::_arraycopy: return inline_arraycopy();
586
587 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
588 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
589 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
590 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
591
592 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
593 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
594 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
595 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
596 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
597 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
598 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
599
600 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
601 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
602
603 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
604 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
605 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
606 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
607
608 case vmIntrinsics::_compressStringC:
609 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
610 case vmIntrinsics::_inflateStringC:
611 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
612
613 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
614 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
615 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
616 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
617 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
618 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
619 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
620 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
621 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
622
623 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
624 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
625 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
626 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
627 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
628 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
629 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
630 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
631 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
632
633 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
634 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
635 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
636 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
637 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
638 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
639 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
640 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
641 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
642
643 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
644 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
645 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
646 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
647 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
648 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
649 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
650 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
651 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
652
653 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
654 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
655 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
656 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
657
658 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
659 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
660 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
661 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
662
663 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
664 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
665 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
666 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
667 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
668 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
669 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
670 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
671 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
672
673 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
674 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
675 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
676 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
677 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
678 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
679 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
680 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
681 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
682
683 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
684 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
685 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
686 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
687 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
688 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
689 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
690 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
691 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
692
693 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
694 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
695 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
696 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
697 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
698 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
699 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
700 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
701 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
702
703 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
704 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
705 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
706 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
707 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
708
709 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
710 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
711 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
712 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
713 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
714 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
715 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
716 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
717 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
718 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
719 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
720 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
721 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
722 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
723 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
724 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
725 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
726 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
727 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
728 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
729
730 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
731 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
732 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
733 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
734 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
735 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
736 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
737 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
738 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
739 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
740 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
741 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
742 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
743 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
744 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
745
746 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
747 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
748 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
749 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
750
751 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
752 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
753 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
754 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
755 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
756
757 case vmIntrinsics::_loadFence:
758 case vmIntrinsics::_storeFence:
759 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
760
761 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
762
763 case vmIntrinsics::_currentThread: return inline_native_currentThread();
764 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
765
766 #ifdef JFR_HAVE_INTRINSICS
767 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime");
768 case vmIntrinsics::_getClassId: return inline_native_classID();
769 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter();
770 #endif
771 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
772 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
773 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
774 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
775 case vmIntrinsics::_getLength: return inline_native_getLength();
776 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
777 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
778 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
779 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
780 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
781 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
782
783 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
784 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
785
786 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
787
788 case vmIntrinsics::_isInstance:
789 case vmIntrinsics::_getModifiers:
790 case vmIntrinsics::_isInterface:
791 case vmIntrinsics::_isArray:
792 case vmIntrinsics::_isPrimitive:
793 case vmIntrinsics::_getSuperclass:
794 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
795
796 case vmIntrinsics::_floatToRawIntBits:
797 case vmIntrinsics::_floatToIntBits:
798 case vmIntrinsics::_intBitsToFloat:
799 case vmIntrinsics::_doubleToRawLongBits:
800 case vmIntrinsics::_doubleToLongBits:
801 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
802
803 case vmIntrinsics::_numberOfLeadingZeros_i:
804 case vmIntrinsics::_numberOfLeadingZeros_l:
805 case vmIntrinsics::_numberOfTrailingZeros_i:
806 case vmIntrinsics::_numberOfTrailingZeros_l:
807 case vmIntrinsics::_bitCount_i:
808 case vmIntrinsics::_bitCount_l:
809 case vmIntrinsics::_reverseBytes_i:
810 case vmIntrinsics::_reverseBytes_l:
811 case vmIntrinsics::_reverseBytes_s:
812 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
813
814 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
815
816 case vmIntrinsics::_Reference_get: return inline_reference_get();
817
818 case vmIntrinsics::_Class_cast: return inline_Class_cast();
819
820 case vmIntrinsics::_aescrypt_encryptBlock:
821 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
822
823 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
824 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
825 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
826
827 case vmIntrinsics::_counterMode_AESCrypt:
828 return inline_counterMode_AESCrypt(intrinsic_id());
829
830 case vmIntrinsics::_sha_implCompress:
831 case vmIntrinsics::_sha2_implCompress:
832 case vmIntrinsics::_sha5_implCompress:
833 return inline_sha_implCompress(intrinsic_id());
834
835 case vmIntrinsics::_digestBase_implCompressMB:
836 return inline_digestBase_implCompressMB(predicate);
837
838 case vmIntrinsics::_multiplyToLen:
839 return inline_multiplyToLen();
840
841 case vmIntrinsics::_squareToLen:
842 return inline_squareToLen();
843
844 case vmIntrinsics::_mulAdd:
845 return inline_mulAdd();
846
847 case vmIntrinsics::_montgomeryMultiply:
848 return inline_montgomeryMultiply();
849 case vmIntrinsics::_montgomerySquare:
850 return inline_montgomerySquare();
851
852 case vmIntrinsics::_vectorizedMismatch:
853 return inline_vectorizedMismatch();
854
855 case vmIntrinsics::_ghash_processBlocks:
856 return inline_ghash_processBlocks();
857 case vmIntrinsics::_base64_encodeBlock:
858 return inline_base64_encodeBlock();
859
860 case vmIntrinsics::_encodeISOArray:
861 case vmIntrinsics::_encodeByteISOArray:
862 return inline_encodeISOArray();
863
864 case vmIntrinsics::_updateCRC32:
865 return inline_updateCRC32();
866 case vmIntrinsics::_updateBytesCRC32:
867 return inline_updateBytesCRC32();
868 case vmIntrinsics::_updateByteBufferCRC32:
869 return inline_updateByteBufferCRC32();
870
871 case vmIntrinsics::_updateBytesCRC32C:
872 return inline_updateBytesCRC32C();
873 case vmIntrinsics::_updateDirectByteBufferCRC32C:
874 return inline_updateDirectByteBufferCRC32C();
875
876 case vmIntrinsics::_updateBytesAdler32:
877 return inline_updateBytesAdler32();
878 case vmIntrinsics::_updateByteBufferAdler32:
879 return inline_updateByteBufferAdler32();
880
881 case vmIntrinsics::_profileBoolean:
882 return inline_profileBoolean();
883 case vmIntrinsics::_isCompileConstant:
884 return inline_isCompileConstant();
885
886 case vmIntrinsics::_hasNegatives:
887 return inline_hasNegatives();
888
889 case vmIntrinsics::_fmaD:
890 case vmIntrinsics::_fmaF:
891 return inline_fma(intrinsic_id());
892
893 case vmIntrinsics::_VectorUnaryOp:
894 return inline_vector_nary_operation(1);
895 case vmIntrinsics::_VectorBinaryOp:
896 return inline_vector_nary_operation(2);
897 case vmIntrinsics::_VectorTernaryOp:
898 return inline_vector_nary_operation(3);
899
900 case vmIntrinsics::_VectorBroadcastCoerced:
901 return inline_vector_broadcast_coerced();
902 case vmIntrinsics::_VectorLoadOp:
903 return inline_vector_mem_operation(/*is_store=*/false);
904 case vmIntrinsics::_VectorStoreOp:
905 return inline_vector_mem_operation(/*is_store=*/true);
906 case vmIntrinsics::_VectorReductionCoerced:
907 return inline_vector_reduction();
908 case vmIntrinsics::_VectorTest:
909 return inline_vector_test();
910 case vmIntrinsics::_VectorBlend:
911 return inline_vector_blend();
912 case vmIntrinsics::_VectorRearrange:
913 return inline_vector_rearrange();
914 case vmIntrinsics::_VectorCompare:
915 return inline_vector_compare();
916 case vmIntrinsics::_VectorBroadcastInt:
917 return inline_vector_broadcast_int();
918 case vmIntrinsics::_VectorReinterpret:
919 return inline_vector_cast_reinterpret(/*is_cast*/ false);
920 case vmIntrinsics::_VectorCast:
921 return inline_vector_cast_reinterpret(/*is_cast*/ true);
922 case vmIntrinsics::_VectorInsert:
923 return inline_vector_insert();
924 case vmIntrinsics::_VectorExtract:
925 return inline_vector_extract();
926
927 default:
928 // If you get here, it may be that someone has added a new intrinsic
929 // to the list in vmSymbols.hpp without implementing it here.
930 #ifndef PRODUCT
931 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
932 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
933 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
934 }
935 #endif
936 return false;
937 }
938 }
939
940 Node* LibraryCallKit::try_to_predicate(int predicate) {
941 if (!jvms()->has_method()) {
942 // Root JVMState has a null method.
943 assert(map()->memory()->Opcode() == Op_Parm, "");
944 // Insert the memory aliasing node
945 set_all_memory(reset_memory());
946 }
947 assert(merged_memory(), "");
948
949 switch (intrinsic_id()) {
950 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
951 return inline_cipherBlockChaining_AESCrypt_predicate(false);
952 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
953 return inline_cipherBlockChaining_AESCrypt_predicate(true);
954 case vmIntrinsics::_counterMode_AESCrypt:
955 return inline_counterMode_AESCrypt_predicate();
956 case vmIntrinsics::_digestBase_implCompressMB:
957 return inline_digestBase_implCompressMB_predicate(predicate);
958
959 default:
960 // If you get here, it may be that someone has added a new intrinsic
961 // to the list in vmSymbols.hpp without implementing it here.
962 #ifndef PRODUCT
963 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
964 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
965 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
966 }
967 #endif
968 Node* slow_ctl = control();
969 set_control(top()); // No fast path instrinsic
970 return slow_ctl;
971 }
972 }
973
974 //------------------------------set_result-------------------------------
975 // Helper function for finishing intrinsics.
976 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
977 record_for_igvn(region);
978 set_control(_gvn.transform(region));
979 set_result( _gvn.transform(value));
980 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
981 }
982
983 //------------------------------generate_guard---------------------------
984 // Helper function for generating guarded fast-slow graph structures.
985 // The given 'test', if true, guards a slow path. If the test fails
986 // then a fast path can be taken. (We generally hope it fails.)
987 // In all cases, GraphKit::control() is updated to the fast path.
988 // The returned value represents the control for the slow path.
989 // The return value is never 'top'; it is either a valid control
990 // or NULL if it is obvious that the slow path can never be taken.
991 // Also, if region and the slow control are not NULL, the slow edge
992 // is appended to the region.
993 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
994 if (stopped()) {
995 // Already short circuited.
996 return NULL;
997 }
998
999 // Build an if node and its projections.
1000 // If test is true we take the slow path, which we assume is uncommon.
1001 if (_gvn.type(test) == TypeInt::ZERO) {
1002 // The slow branch is never taken. No need to build this guard.
1003 return NULL;
1004 }
1005
1006 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1007
1008 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1009 if (if_slow == top()) {
1010 // The slow branch is never taken. No need to build this guard.
1011 return NULL;
1012 }
1013
1014 if (region != NULL)
1015 region->add_req(if_slow);
1016
1017 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1018 set_control(if_fast);
1019
1020 return if_slow;
1021 }
1022
1023 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1024 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1025 }
1026 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1027 return generate_guard(test, region, PROB_FAIR);
1028 }
1029
1030 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1031 Node* *pos_index) {
1032 if (stopped())
1033 return NULL; // already stopped
1034 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1035 return NULL; // index is already adequately typed
1036 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1037 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1038 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1039 if (is_neg != NULL && pos_index != NULL) {
1040 // Emulate effect of Parse::adjust_map_after_if.
1041 Node* ccast = new CastIINode(index, TypeInt::POS);
1042 ccast->set_req(0, control());
1043 (*pos_index) = _gvn.transform(ccast);
1044 }
1045 return is_neg;
1046 }
1047
1048 // Make sure that 'position' is a valid limit index, in [0..length].
1049 // There are two equivalent plans for checking this:
1050 // A. (offset + copyLength) unsigned<= arrayLength
1051 // B. offset <= (arrayLength - copyLength)
1052 // We require that all of the values above, except for the sum and
1053 // difference, are already known to be non-negative.
1054 // Plan A is robust in the face of overflow, if offset and copyLength
1055 // are both hugely positive.
1056 //
1057 // Plan B is less direct and intuitive, but it does not overflow at
1058 // all, since the difference of two non-negatives is always
1059 // representable. Whenever Java methods must perform the equivalent
1060 // check they generally use Plan B instead of Plan A.
1061 // For the moment we use Plan A.
1062 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1063 Node* subseq_length,
1064 Node* array_length,
1065 RegionNode* region) {
1066 if (stopped())
1067 return NULL; // already stopped
1068 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1069 if (zero_offset && subseq_length->eqv_uncast(array_length))
1070 return NULL; // common case of whole-array copy
1071 Node* last = subseq_length;
1072 if (!zero_offset) // last += offset
1073 last = _gvn.transform(new AddINode(last, offset));
1074 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1075 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1076 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1077 return is_over;
1078 }
1079
1080 // Emit range checks for the given String.value byte array
1081 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1082 if (stopped()) {
1083 return; // already stopped
1084 }
1085 RegionNode* bailout = new RegionNode(1);
1086 record_for_igvn(bailout);
1087 if (char_count) {
1088 // Convert char count to byte count
1089 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1090 }
1091
1092 // Offset and count must not be negative
1093 generate_negative_guard(offset, bailout);
1094 generate_negative_guard(count, bailout);
1095 // Offset + count must not exceed length of array
1096 generate_limit_guard(offset, count, load_array_length(array), bailout);
1097
1098 if (bailout->req() > 1) {
1099 PreserveJVMState pjvms(this);
1100 set_control(_gvn.transform(bailout));
1101 uncommon_trap(Deoptimization::Reason_intrinsic,
1102 Deoptimization::Action_maybe_recompile);
1103 }
1104 }
1105
1106 //--------------------------generate_current_thread--------------------
1107 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1108 ciKlass* thread_klass = env()->Thread_klass();
1109 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1110 Node* thread = _gvn.transform(new ThreadLocalNode());
1111 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1112 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1113 tls_output = thread;
1114 return threadObj;
1115 }
1116
1117
1118 //------------------------------make_string_method_node------------------------
1119 // Helper method for String intrinsic functions. This version is called with
1120 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1121 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1122 // containing the lengths of str1 and str2.
1123 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1124 Node* result = NULL;
1125 switch (opcode) {
1126 case Op_StrIndexOf:
1127 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1128 str1_start, cnt1, str2_start, cnt2, ae);
1129 break;
1130 case Op_StrComp:
1131 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1132 str1_start, cnt1, str2_start, cnt2, ae);
1133 break;
1134 case Op_StrEquals:
1135 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1136 // Use the constant length if there is one because optimized match rule may exist.
1137 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1138 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1139 break;
1140 default:
1141 ShouldNotReachHere();
1142 return NULL;
1143 }
1144
1145 // All these intrinsics have checks.
1146 C->set_has_split_ifs(true); // Has chance for split-if optimization
1147 clear_upper_avx();
1148
1149 return _gvn.transform(result);
1150 }
1151
1152 //------------------------------inline_string_compareTo------------------------
1153 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1154 Node* arg1 = argument(0);
1155 Node* arg2 = argument(1);
1156
1157 arg1 = must_be_not_null(arg1, true);
1158 arg2 = must_be_not_null(arg2, true);
1159
1160 // Get start addr and length of first argument
1161 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1162 Node* arg1_cnt = load_array_length(arg1);
1163
1164 // Get start addr and length of second argument
1165 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1166 Node* arg2_cnt = load_array_length(arg2);
1167
1168 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1169 set_result(result);
1170 return true;
1171 }
1172
1173 //------------------------------inline_string_equals------------------------
1174 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1175 Node* arg1 = argument(0);
1176 Node* arg2 = argument(1);
1177
1178 // paths (plus control) merge
1179 RegionNode* region = new RegionNode(3);
1180 Node* phi = new PhiNode(region, TypeInt::BOOL);
1181
1182 if (!stopped()) {
1183
1184 arg1 = must_be_not_null(arg1, true);
1185 arg2 = must_be_not_null(arg2, true);
1186
1187 // Get start addr and length of first argument
1188 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1189 Node* arg1_cnt = load_array_length(arg1);
1190
1191 // Get start addr and length of second argument
1192 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1193 Node* arg2_cnt = load_array_length(arg2);
1194
1195 // Check for arg1_cnt != arg2_cnt
1196 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1197 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1198 Node* if_ne = generate_slow_guard(bol, NULL);
1199 if (if_ne != NULL) {
1200 phi->init_req(2, intcon(0));
1201 region->init_req(2, if_ne);
1202 }
1203
1204 // Check for count == 0 is done by assembler code for StrEquals.
1205
1206 if (!stopped()) {
1207 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1208 phi->init_req(1, equals);
1209 region->init_req(1, control());
1210 }
1211 }
1212
1213 // post merge
1214 set_control(_gvn.transform(region));
1215 record_for_igvn(region);
1216
1217 set_result(_gvn.transform(phi));
1218 return true;
1219 }
1220
1221 //------------------------------inline_array_equals----------------------------
1222 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1223 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1224 Node* arg1 = argument(0);
1225 Node* arg2 = argument(1);
1226
1227 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1228 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1229 clear_upper_avx();
1230
1231 return true;
1232 }
1233
1234 //------------------------------inline_hasNegatives------------------------------
1235 bool LibraryCallKit::inline_hasNegatives() {
1236 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1237 return false;
1238 }
1239
1240 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1241 // no receiver since it is static method
1242 Node* ba = argument(0);
1243 Node* offset = argument(1);
1244 Node* len = argument(2);
1245
1246 ba = must_be_not_null(ba, true);
1247
1248 // Range checks
1249 generate_string_range_check(ba, offset, len, false);
1250 if (stopped()) {
1251 return true;
1252 }
1253 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1254 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1255 set_result(_gvn.transform(result));
1256 return true;
1257 }
1258
1259 bool LibraryCallKit::inline_preconditions_checkIndex() {
1260 Node* index = argument(0);
1261 Node* length = argument(1);
1262 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1263 return false;
1264 }
1265
1266 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1267 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1268
1269 {
1270 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1271 uncommon_trap(Deoptimization::Reason_intrinsic,
1272 Deoptimization::Action_make_not_entrant);
1273 }
1274
1275 if (stopped()) {
1276 return false;
1277 }
1278
1279 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1280 BoolTest::mask btest = BoolTest::lt;
1281 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1282 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1283 _gvn.set_type(rc, rc->Value(&_gvn));
1284 if (!rc_bool->is_Con()) {
1285 record_for_igvn(rc);
1286 }
1287 set_control(_gvn.transform(new IfTrueNode(rc)));
1288 {
1289 PreserveJVMState pjvms(this);
1290 set_control(_gvn.transform(new IfFalseNode(rc)));
1291 uncommon_trap(Deoptimization::Reason_range_check,
1292 Deoptimization::Action_make_not_entrant);
1293 }
1294
1295 if (stopped()) {
1296 return false;
1297 }
1298
1299 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1300 result->set_req(0, control());
1301 result = _gvn.transform(result);
1302 set_result(result);
1303 replace_in_map(index, result);
1304 clear_upper_avx();
1305 return true;
1306 }
1307
1308 //------------------------------inline_string_indexOf------------------------
1309 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1310 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1311 return false;
1312 }
1313 Node* src = argument(0);
1314 Node* tgt = argument(1);
1315
1316 // Make the merge point
1317 RegionNode* result_rgn = new RegionNode(4);
1318 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1319
1320 src = must_be_not_null(src, true);
1321 tgt = must_be_not_null(tgt, true);
1322
1323 // Get start addr and length of source string
1324 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1325 Node* src_count = load_array_length(src);
1326
1327 // Get start addr and length of substring
1328 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1329 Node* tgt_count = load_array_length(tgt);
1330
1331 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1332 // Divide src size by 2 if String is UTF16 encoded
1333 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1334 }
1335 if (ae == StrIntrinsicNode::UU) {
1336 // Divide substring size by 2 if String is UTF16 encoded
1337 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1338 }
1339
1340 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1341 if (result != NULL) {
1342 result_phi->init_req(3, result);
1343 result_rgn->init_req(3, control());
1344 }
1345 set_control(_gvn.transform(result_rgn));
1346 record_for_igvn(result_rgn);
1347 set_result(_gvn.transform(result_phi));
1348
1349 return true;
1350 }
1351
1352 //-----------------------------inline_string_indexOf-----------------------
1353 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1354 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1355 return false;
1356 }
1357 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1358 return false;
1359 }
1360 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1361 Node* src = argument(0); // byte[]
1362 Node* src_count = argument(1); // char count
1363 Node* tgt = argument(2); // byte[]
1364 Node* tgt_count = argument(3); // char count
1365 Node* from_index = argument(4); // char index
1366
1367 src = must_be_not_null(src, true);
1368 tgt = must_be_not_null(tgt, true);
1369
1370 // Multiply byte array index by 2 if String is UTF16 encoded
1371 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1372 src_count = _gvn.transform(new SubINode(src_count, from_index));
1373 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1374 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1375
1376 // Range checks
1377 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1378 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1379 if (stopped()) {
1380 return true;
1381 }
1382
1383 RegionNode* region = new RegionNode(5);
1384 Node* phi = new PhiNode(region, TypeInt::INT);
1385
1386 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1387 if (result != NULL) {
1388 // The result is index relative to from_index if substring was found, -1 otherwise.
1389 // Generate code which will fold into cmove.
1390 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1391 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1392
1393 Node* if_lt = generate_slow_guard(bol, NULL);
1394 if (if_lt != NULL) {
1395 // result == -1
1396 phi->init_req(3, result);
1397 region->init_req(3, if_lt);
1398 }
1399 if (!stopped()) {
1400 result = _gvn.transform(new AddINode(result, from_index));
1401 phi->init_req(4, result);
1402 region->init_req(4, control());
1403 }
1404 }
1405
1406 set_control(_gvn.transform(region));
1407 record_for_igvn(region);
1408 set_result(_gvn.transform(phi));
1409 clear_upper_avx();
1410
1411 return true;
1412 }
1413
1414 // Create StrIndexOfNode with fast path checks
1415 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1416 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1417 // Check for substr count > string count
1418 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1419 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1420 Node* if_gt = generate_slow_guard(bol, NULL);
1421 if (if_gt != NULL) {
1422 phi->init_req(1, intcon(-1));
1423 region->init_req(1, if_gt);
1424 }
1425 if (!stopped()) {
1426 // Check for substr count == 0
1427 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1428 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1429 Node* if_zero = generate_slow_guard(bol, NULL);
1430 if (if_zero != NULL) {
1431 phi->init_req(2, intcon(0));
1432 region->init_req(2, if_zero);
1433 }
1434 }
1435 if (!stopped()) {
1436 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1437 }
1438 return NULL;
1439 }
1440
1441 //-----------------------------inline_string_indexOfChar-----------------------
1442 bool LibraryCallKit::inline_string_indexOfChar() {
1443 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1444 return false;
1445 }
1446 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1447 return false;
1448 }
1449 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1450 Node* src = argument(0); // byte[]
1451 Node* tgt = argument(1); // tgt is int ch
1452 Node* from_index = argument(2);
1453 Node* max = argument(3);
1454
1455 src = must_be_not_null(src, true);
1456
1457 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1458 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1459 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1460
1461 // Range checks
1462 generate_string_range_check(src, src_offset, src_count, true);
1463 if (stopped()) {
1464 return true;
1465 }
1466
1467 RegionNode* region = new RegionNode(3);
1468 Node* phi = new PhiNode(region, TypeInt::INT);
1469
1470 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1471 C->set_has_split_ifs(true); // Has chance for split-if optimization
1472 _gvn.transform(result);
1473
1474 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1475 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1476
1477 Node* if_lt = generate_slow_guard(bol, NULL);
1478 if (if_lt != NULL) {
1479 // result == -1
1480 phi->init_req(2, result);
1481 region->init_req(2, if_lt);
1482 }
1483 if (!stopped()) {
1484 result = _gvn.transform(new AddINode(result, from_index));
1485 phi->init_req(1, result);
1486 region->init_req(1, control());
1487 }
1488 set_control(_gvn.transform(region));
1489 record_for_igvn(region);
1490 set_result(_gvn.transform(phi));
1491
1492 return true;
1493 }
1494 //---------------------------inline_string_copy---------------------
1495 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1496 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1497 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1498 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1499 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1500 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1501 bool LibraryCallKit::inline_string_copy(bool compress) {
1502 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1503 return false;
1504 }
1505 int nargs = 5; // 2 oops, 3 ints
1506 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1507
1508 Node* src = argument(0);
1509 Node* src_offset = argument(1);
1510 Node* dst = argument(2);
1511 Node* dst_offset = argument(3);
1512 Node* length = argument(4);
1513
1514 // Check for allocation before we add nodes that would confuse
1515 // tightly_coupled_allocation()
1516 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1517
1518 // Figure out the size and type of the elements we will be copying.
1519 const Type* src_type = src->Value(&_gvn);
1520 const Type* dst_type = dst->Value(&_gvn);
1521 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1522 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1523 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1524 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1525 "Unsupported array types for inline_string_copy");
1526
1527 src = must_be_not_null(src, true);
1528 dst = must_be_not_null(dst, true);
1529
1530 // Convert char[] offsets to byte[] offsets
1531 bool convert_src = (compress && src_elem == T_BYTE);
1532 bool convert_dst = (!compress && dst_elem == T_BYTE);
1533 if (convert_src) {
1534 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1535 } else if (convert_dst) {
1536 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1537 }
1538
1539 // Range checks
1540 generate_string_range_check(src, src_offset, length, convert_src);
1541 generate_string_range_check(dst, dst_offset, length, convert_dst);
1542 if (stopped()) {
1543 return true;
1544 }
1545
1546 Node* src_start = array_element_address(src, src_offset, src_elem);
1547 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1548 // 'src_start' points to src array + scaled offset
1549 // 'dst_start' points to dst array + scaled offset
1550 Node* count = NULL;
1551 if (compress) {
1552 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1553 } else {
1554 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1555 }
1556
1557 if (alloc != NULL) {
1558 if (alloc->maybe_set_complete(&_gvn)) {
1559 // "You break it, you buy it."
1560 InitializeNode* init = alloc->initialization();
1561 assert(init->is_complete(), "we just did this");
1562 init->set_complete_with_arraycopy();
1563 assert(dst->is_CheckCastPP(), "sanity");
1564 assert(dst->in(0)->in(0) == init, "dest pinned");
1565 }
1566 // Do not let stores that initialize this object be reordered with
1567 // a subsequent store that would make this object accessible by
1568 // other threads.
1569 // Record what AllocateNode this StoreStore protects so that
1570 // escape analysis can go from the MemBarStoreStoreNode to the
1571 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1572 // based on the escape status of the AllocateNode.
1573 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1574 }
1575 if (compress) {
1576 set_result(_gvn.transform(count));
1577 }
1578 clear_upper_avx();
1579
1580 return true;
1581 }
1582
1583 #ifdef _LP64
1584 #define XTOP ,top() /*additional argument*/
1585 #else //_LP64
1586 #define XTOP /*no additional argument*/
1587 #endif //_LP64
1588
1589 //------------------------inline_string_toBytesU--------------------------
1590 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1591 bool LibraryCallKit::inline_string_toBytesU() {
1592 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1593 return false;
1594 }
1595 // Get the arguments.
1596 Node* value = argument(0);
1597 Node* offset = argument(1);
1598 Node* length = argument(2);
1599
1600 Node* newcopy = NULL;
1601
1602 // Set the original stack and the reexecute bit for the interpreter to reexecute
1603 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1604 { PreserveReexecuteState preexecs(this);
1605 jvms()->set_should_reexecute(true);
1606
1607 // Check if a null path was taken unconditionally.
1608 value = null_check(value);
1609
1610 RegionNode* bailout = new RegionNode(1);
1611 record_for_igvn(bailout);
1612
1613 // Range checks
1614 generate_negative_guard(offset, bailout);
1615 generate_negative_guard(length, bailout);
1616 generate_limit_guard(offset, length, load_array_length(value), bailout);
1617 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1618 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1619
1620 if (bailout->req() > 1) {
1621 PreserveJVMState pjvms(this);
1622 set_control(_gvn.transform(bailout));
1623 uncommon_trap(Deoptimization::Reason_intrinsic,
1624 Deoptimization::Action_maybe_recompile);
1625 }
1626 if (stopped()) {
1627 return true;
1628 }
1629
1630 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1631 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1632 newcopy = new_array(klass_node, size, 0); // no arguments to push
1633 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1634
1635 // Calculate starting addresses.
1636 Node* src_start = array_element_address(value, offset, T_CHAR);
1637 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1638
1639 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1640 const TypeInt* toffset = gvn().type(offset)->is_int();
1641 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1642
1643 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1644 const char* copyfunc_name = "arraycopy";
1645 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1646 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1647 OptoRuntime::fast_arraycopy_Type(),
1648 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1649 src_start, dst_start, ConvI2X(length) XTOP);
1650 // Do not let reads from the cloned object float above the arraycopy.
1651 if (alloc != NULL) {
1652 if (alloc->maybe_set_complete(&_gvn)) {
1653 // "You break it, you buy it."
1654 InitializeNode* init = alloc->initialization();
1655 assert(init->is_complete(), "we just did this");
1656 init->set_complete_with_arraycopy();
1657 assert(newcopy->is_CheckCastPP(), "sanity");
1658 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1659 }
1660 // Do not let stores that initialize this object be reordered with
1661 // a subsequent store that would make this object accessible by
1662 // other threads.
1663 // Record what AllocateNode this StoreStore protects so that
1664 // escape analysis can go from the MemBarStoreStoreNode to the
1665 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1666 // based on the escape status of the AllocateNode.
1667 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1668 } else {
1669 insert_mem_bar(Op_MemBarCPUOrder);
1670 }
1671 } // original reexecute is set back here
1672
1673 C->set_has_split_ifs(true); // Has chance for split-if optimization
1674 if (!stopped()) {
1675 set_result(newcopy);
1676 }
1677 clear_upper_avx();
1678
1679 return true;
1680 }
1681
1682 //------------------------inline_string_getCharsU--------------------------
1683 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1684 bool LibraryCallKit::inline_string_getCharsU() {
1685 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1686 return false;
1687 }
1688
1689 // Get the arguments.
1690 Node* src = argument(0);
1691 Node* src_begin = argument(1);
1692 Node* src_end = argument(2); // exclusive offset (i < src_end)
1693 Node* dst = argument(3);
1694 Node* dst_begin = argument(4);
1695
1696 // Check for allocation before we add nodes that would confuse
1697 // tightly_coupled_allocation()
1698 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1699
1700 // Check if a null path was taken unconditionally.
1701 src = null_check(src);
1702 dst = null_check(dst);
1703 if (stopped()) {
1704 return true;
1705 }
1706
1707 // Get length and convert char[] offset to byte[] offset
1708 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1709 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1710
1711 // Range checks
1712 generate_string_range_check(src, src_begin, length, true);
1713 generate_string_range_check(dst, dst_begin, length, false);
1714 if (stopped()) {
1715 return true;
1716 }
1717
1718 if (!stopped()) {
1719 // Calculate starting addresses.
1720 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1721 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1722
1723 // Check if array addresses are aligned to HeapWordSize
1724 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1725 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1726 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1727 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1728
1729 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1730 const char* copyfunc_name = "arraycopy";
1731 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1732 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1733 OptoRuntime::fast_arraycopy_Type(),
1734 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1735 src_start, dst_start, ConvI2X(length) XTOP);
1736 // Do not let reads from the cloned object float above the arraycopy.
1737 if (alloc != NULL) {
1738 if (alloc->maybe_set_complete(&_gvn)) {
1739 // "You break it, you buy it."
1740 InitializeNode* init = alloc->initialization();
1741 assert(init->is_complete(), "we just did this");
1742 init->set_complete_with_arraycopy();
1743 assert(dst->is_CheckCastPP(), "sanity");
1744 assert(dst->in(0)->in(0) == init, "dest pinned");
1745 }
1746 // Do not let stores that initialize this object be reordered with
1747 // a subsequent store that would make this object accessible by
1748 // other threads.
1749 // Record what AllocateNode this StoreStore protects so that
1750 // escape analysis can go from the MemBarStoreStoreNode to the
1751 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1752 // based on the escape status of the AllocateNode.
1753 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1754 } else {
1755 insert_mem_bar(Op_MemBarCPUOrder);
1756 }
1757 }
1758
1759 C->set_has_split_ifs(true); // Has chance for split-if optimization
1760 return true;
1761 }
1762
1763 //----------------------inline_string_char_access----------------------------
1764 // Store/Load char to/from byte[] array.
1765 // static void StringUTF16.putChar(byte[] val, int index, int c)
1766 // static char StringUTF16.getChar(byte[] val, int index)
1767 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1768 Node* value = argument(0);
1769 Node* index = argument(1);
1770 Node* ch = is_store ? argument(2) : NULL;
1771
1772 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1773 // correctly requires matched array shapes.
1774 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1775 "sanity: byte[] and char[] bases agree");
1776 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1777 "sanity: byte[] and char[] scales agree");
1778
1779 // Bail when getChar over constants is requested: constant folding would
1780 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1781 // Java method would constant fold nicely instead.
1782 if (!is_store && value->is_Con() && index->is_Con()) {
1783 return false;
1784 }
1785
1786 value = must_be_not_null(value, true);
1787
1788 Node* adr = array_element_address(value, index, T_CHAR);
1789 if (adr->is_top()) {
1790 return false;
1791 }
1792 if (is_store) {
1793 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1794 false, false, true /* mismatched */);
1795 } else {
1796 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1797 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1798 set_result(ch);
1799 }
1800 return true;
1801 }
1802
1803 //--------------------------round_double_node--------------------------------
1804 // Round a double node if necessary.
1805 Node* LibraryCallKit::round_double_node(Node* n) {
1806 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1807 n = _gvn.transform(new RoundDoubleNode(0, n));
1808 return n;
1809 }
1810
1811 //------------------------------inline_math-----------------------------------
1812 // public static double Math.abs(double)
1813 // public static double Math.sqrt(double)
1814 // public static double Math.log(double)
1815 // public static double Math.log10(double)
1816 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1817 Node* arg = round_double_node(argument(0));
1818 Node* n = NULL;
1819 switch (id) {
1820 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1821 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1822 default: fatal_unexpected_iid(id); break;
1823 }
1824 set_result(_gvn.transform(n));
1825 return true;
1826 }
1827
1828 //------------------------------runtime_math-----------------------------
1829 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1830 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1831 "must be (DD)D or (D)D type");
1832
1833 // Inputs
1834 Node* a = round_double_node(argument(0));
1835 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1836
1837 const TypePtr* no_memory_effects = NULL;
1838 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1839 no_memory_effects,
1840 a, top(), b, b ? top() : NULL);
1841 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1842 #ifdef ASSERT
1843 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1844 assert(value_top == top(), "second value must be top");
1845 #endif
1846
1847 set_result(value);
1848 return true;
1849 }
1850
1851 //------------------------------inline_math_native-----------------------------
1852 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1853 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1854 switch (id) {
1855 // These intrinsics are not properly supported on all hardware
1856 case vmIntrinsics::_dsin:
1857 return StubRoutines::dsin() != NULL ?
1858 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1859 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1860 case vmIntrinsics::_dcos:
1861 return StubRoutines::dcos() != NULL ?
1862 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1863 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1864 case vmIntrinsics::_dtan:
1865 return StubRoutines::dtan() != NULL ?
1866 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1867 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1868 case vmIntrinsics::_dlog:
1869 return StubRoutines::dlog() != NULL ?
1870 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1871 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1872 case vmIntrinsics::_dlog10:
1873 return StubRoutines::dlog10() != NULL ?
1874 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1875 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1876
1877 // These intrinsics are supported on all hardware
1878 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1879 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1880
1881 case vmIntrinsics::_dexp:
1882 return StubRoutines::dexp() != NULL ?
1883 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1884 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1885 case vmIntrinsics::_dpow: {
1886 Node* exp = round_double_node(argument(2));
1887 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1888 if (d != NULL && d->getd() == 2.0) {
1889 // Special case: pow(x, 2.0) => x * x
1890 Node* base = round_double_node(argument(0));
1891 set_result(_gvn.transform(new MulDNode(base, base)));
1892 return true;
1893 }
1894 return StubRoutines::dpow() != NULL ?
1895 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1896 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1897 }
1898 #undef FN_PTR
1899
1900 // These intrinsics are not yet correctly implemented
1901 case vmIntrinsics::_datan2:
1902 return false;
1903
1904 default:
1905 fatal_unexpected_iid(id);
1906 return false;
1907 }
1908 }
1909
1910 static bool is_simple_name(Node* n) {
1911 return (n->req() == 1 // constant
1912 || (n->is_Type() && n->as_Type()->type()->singleton())
1913 || n->is_Proj() // parameter or return value
1914 || n->is_Phi() // local of some sort
1915 );
1916 }
1917
1918 //----------------------------inline_notify-----------------------------------*
1919 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1920 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1921 address func;
1922 if (id == vmIntrinsics::_notify) {
1923 func = OptoRuntime::monitor_notify_Java();
1924 } else {
1925 func = OptoRuntime::monitor_notifyAll_Java();
1926 }
1927 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1928 make_slow_call_ex(call, env()->Throwable_klass(), false);
1929 return true;
1930 }
1931
1932
1933 //----------------------------inline_min_max-----------------------------------
1934 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1935 set_result(generate_min_max(id, argument(0), argument(1)));
1936 return true;
1937 }
1938
1939 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1940 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1941 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1942 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1943 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1944
1945 {
1946 PreserveJVMState pjvms(this);
1947 PreserveReexecuteState preexecs(this);
1948 jvms()->set_should_reexecute(true);
1949
1950 set_control(slow_path);
1951 set_i_o(i_o());
1952
1953 uncommon_trap(Deoptimization::Reason_intrinsic,
1954 Deoptimization::Action_none);
1955 }
1956
1957 set_control(fast_path);
1958 set_result(math);
1959 }
1960
1961 template <typename OverflowOp>
1962 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1963 typedef typename OverflowOp::MathOp MathOp;
1964
1965 MathOp* mathOp = new MathOp(arg1, arg2);
1966 Node* operation = _gvn.transform( mathOp );
1967 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1968 inline_math_mathExact(operation, ofcheck);
1969 return true;
1970 }
1971
1972 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1973 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1974 }
1975
1976 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1977 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1978 }
1979
1980 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1981 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1982 }
1983
1984 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1985 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1986 }
1987
1988 bool LibraryCallKit::inline_math_negateExactI() {
1989 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1990 }
1991
1992 bool LibraryCallKit::inline_math_negateExactL() {
1993 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1994 }
1995
1996 bool LibraryCallKit::inline_math_multiplyExactI() {
1997 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1998 }
1999
2000 bool LibraryCallKit::inline_math_multiplyExactL() {
2001 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2002 }
2003
2004 bool LibraryCallKit::inline_math_multiplyHigh() {
2005 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
2006 return true;
2007 }
2008
2009 Node*
2010 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2011 // These are the candidate return value:
2012 Node* xvalue = x0;
2013 Node* yvalue = y0;
2014
2015 if (xvalue == yvalue) {
2016 return xvalue;
2017 }
2018
2019 bool want_max = (id == vmIntrinsics::_max);
2020
2021 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2022 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2023 if (txvalue == NULL || tyvalue == NULL) return top();
2024 // This is not really necessary, but it is consistent with a
2025 // hypothetical MaxINode::Value method:
2026 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2027
2028 // %%% This folding logic should (ideally) be in a different place.
2029 // Some should be inside IfNode, and there to be a more reliable
2030 // transformation of ?: style patterns into cmoves. We also want
2031 // more powerful optimizations around cmove and min/max.
2032
2033 // Try to find a dominating comparison of these guys.
2034 // It can simplify the index computation for Arrays.copyOf
2035 // and similar uses of System.arraycopy.
2036 // First, compute the normalized version of CmpI(x, y).
2037 int cmp_op = Op_CmpI;
2038 Node* xkey = xvalue;
2039 Node* ykey = yvalue;
2040 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2041 if (ideal_cmpxy->is_Cmp()) {
2042 // E.g., if we have CmpI(length - offset, count),
2043 // it might idealize to CmpI(length, count + offset)
2044 cmp_op = ideal_cmpxy->Opcode();
2045 xkey = ideal_cmpxy->in(1);
2046 ykey = ideal_cmpxy->in(2);
2047 }
2048
2049 // Start by locating any relevant comparisons.
2050 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2051 Node* cmpxy = NULL;
2052 Node* cmpyx = NULL;
2053 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2054 Node* cmp = start_from->fast_out(k);
2055 if (cmp->outcnt() > 0 && // must have prior uses
2056 cmp->in(0) == NULL && // must be context-independent
2057 cmp->Opcode() == cmp_op) { // right kind of compare
2058 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2059 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2060 }
2061 }
2062
2063 const int NCMPS = 2;
2064 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2065 int cmpn;
2066 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2067 if (cmps[cmpn] != NULL) break; // find a result
2068 }
2069 if (cmpn < NCMPS) {
2070 // Look for a dominating test that tells us the min and max.
2071 int depth = 0; // Limit search depth for speed
2072 Node* dom = control();
2073 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2074 if (++depth >= 100) break;
2075 Node* ifproj = dom;
2076 if (!ifproj->is_Proj()) continue;
2077 Node* iff = ifproj->in(0);
2078 if (!iff->is_If()) continue;
2079 Node* bol = iff->in(1);
2080 if (!bol->is_Bool()) continue;
2081 Node* cmp = bol->in(1);
2082 if (cmp == NULL) continue;
2083 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2084 if (cmps[cmpn] == cmp) break;
2085 if (cmpn == NCMPS) continue;
2086 BoolTest::mask btest = bol->as_Bool()->_test._test;
2087 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2088 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2089 // At this point, we know that 'x btest y' is true.
2090 switch (btest) {
2091 case BoolTest::eq:
2092 // They are proven equal, so we can collapse the min/max.
2093 // Either value is the answer. Choose the simpler.
2094 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2095 return yvalue;
2096 return xvalue;
2097 case BoolTest::lt: // x < y
2098 case BoolTest::le: // x <= y
2099 return (want_max ? yvalue : xvalue);
2100 case BoolTest::gt: // x > y
2101 case BoolTest::ge: // x >= y
2102 return (want_max ? xvalue : yvalue);
2103 default:
2104 break;
2105 }
2106 }
2107 }
2108
2109 // We failed to find a dominating test.
2110 // Let's pick a test that might GVN with prior tests.
2111 Node* best_bol = NULL;
2112 BoolTest::mask best_btest = BoolTest::illegal;
2113 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2114 Node* cmp = cmps[cmpn];
2115 if (cmp == NULL) continue;
2116 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2117 Node* bol = cmp->fast_out(j);
2118 if (!bol->is_Bool()) continue;
2119 BoolTest::mask btest = bol->as_Bool()->_test._test;
2120 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2121 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2122 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2123 best_bol = bol->as_Bool();
2124 best_btest = btest;
2125 }
2126 }
2127 }
2128
2129 Node* answer_if_true = NULL;
2130 Node* answer_if_false = NULL;
2131 switch (best_btest) {
2132 default:
2133 if (cmpxy == NULL)
2134 cmpxy = ideal_cmpxy;
2135 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2136 // and fall through:
2137 case BoolTest::lt: // x < y
2138 case BoolTest::le: // x <= y
2139 answer_if_true = (want_max ? yvalue : xvalue);
2140 answer_if_false = (want_max ? xvalue : yvalue);
2141 break;
2142 case BoolTest::gt: // x > y
2143 case BoolTest::ge: // x >= y
2144 answer_if_true = (want_max ? xvalue : yvalue);
2145 answer_if_false = (want_max ? yvalue : xvalue);
2146 break;
2147 }
2148
2149 jint hi, lo;
2150 if (want_max) {
2151 // We can sharpen the minimum.
2152 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2153 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2154 } else {
2155 // We can sharpen the maximum.
2156 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2157 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2158 }
2159
2160 // Use a flow-free graph structure, to avoid creating excess control edges
2161 // which could hinder other optimizations.
2162 // Since Math.min/max is often used with arraycopy, we want
2163 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2164 Node* cmov = CMoveNode::make(NULL, best_bol,
2165 answer_if_false, answer_if_true,
2166 TypeInt::make(lo, hi, widen));
2167
2168 return _gvn.transform(cmov);
2169
2170 /*
2171 // This is not as desirable as it may seem, since Min and Max
2172 // nodes do not have a full set of optimizations.
2173 // And they would interfere, anyway, with 'if' optimizations
2174 // and with CMoveI canonical forms.
2175 switch (id) {
2176 case vmIntrinsics::_min:
2177 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2178 case vmIntrinsics::_max:
2179 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2180 default:
2181 ShouldNotReachHere();
2182 }
2183 */
2184 }
2185
2186 inline int
2187 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2188 const TypePtr* base_type = TypePtr::NULL_PTR;
2189 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2190 if (base_type == NULL) {
2191 // Unknown type.
2192 return Type::AnyPtr;
2193 } else if (base_type == TypePtr::NULL_PTR) {
2194 // Since this is a NULL+long form, we have to switch to a rawptr.
2195 base = _gvn.transform(new CastX2PNode(offset));
2196 offset = MakeConX(0);
2197 return Type::RawPtr;
2198 } else if (base_type->base() == Type::RawPtr) {
2199 return Type::RawPtr;
2200 } else if (base_type->isa_oopptr()) {
2201 // Base is never null => always a heap address.
2202 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2203 return Type::OopPtr;
2204 }
2205 // Offset is small => always a heap address.
2206 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2207 if (offset_type != NULL &&
2208 base_type->offset() == 0 && // (should always be?)
2209 offset_type->_lo >= 0 &&
2210 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2211 return Type::OopPtr;
2212 } else if (type == T_OBJECT) {
2213 // off heap access to an oop doesn't make any sense. Has to be on
2214 // heap.
2215 return Type::OopPtr;
2216 }
2217 // Otherwise, it might either be oop+off or NULL+addr.
2218 return Type::AnyPtr;
2219 } else {
2220 // No information:
2221 return Type::AnyPtr;
2222 }
2223 }
2224
2225 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2226 Node* uncasted_base = base;
2227 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2228 if (kind == Type::RawPtr) {
2229 return basic_plus_adr(top(), uncasted_base, offset);
2230 } else if (kind == Type::AnyPtr) {
2231 assert(base == uncasted_base, "unexpected base change");
2232 if (can_cast) {
2233 if (!_gvn.type(base)->speculative_maybe_null() &&
2234 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2235 // According to profiling, this access is always on
2236 // heap. Casting the base to not null and thus avoiding membars
2237 // around the access should allow better optimizations
2238 Node* null_ctl = top();
2239 base = null_check_oop(base, &null_ctl, true, true, true);
2240 assert(null_ctl->is_top(), "no null control here");
2241 return basic_plus_adr(base, offset);
2242 } else if (_gvn.type(base)->speculative_always_null() &&
2243 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2244 // According to profiling, this access is always off
2245 // heap.
2246 base = null_assert(base);
2247 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2248 offset = MakeConX(0);
2249 return basic_plus_adr(top(), raw_base, offset);
2250 }
2251 }
2252 // We don't know if it's an on heap or off heap access. Fall back
2253 // to raw memory access.
2254 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2255 return basic_plus_adr(top(), raw, offset);
2256 } else {
2257 assert(base == uncasted_base, "unexpected base change");
2258 // We know it's an on heap access so base can't be null
2259 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2260 base = must_be_not_null(base, true);
2261 }
2262 return basic_plus_adr(base, offset);
2263 }
2264 }
2265
2266 //--------------------------inline_number_methods-----------------------------
2267 // inline int Integer.numberOfLeadingZeros(int)
2268 // inline int Long.numberOfLeadingZeros(long)
2269 //
2270 // inline int Integer.numberOfTrailingZeros(int)
2271 // inline int Long.numberOfTrailingZeros(long)
2272 //
2273 // inline int Integer.bitCount(int)
2274 // inline int Long.bitCount(long)
2275 //
2276 // inline char Character.reverseBytes(char)
2277 // inline short Short.reverseBytes(short)
2278 // inline int Integer.reverseBytes(int)
2279 // inline long Long.reverseBytes(long)
2280 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2281 Node* arg = argument(0);
2282 Node* n = NULL;
2283 switch (id) {
2284 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2285 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2286 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2287 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2288 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2289 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2290 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2291 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2292 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2293 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2294 default: fatal_unexpected_iid(id); break;
2295 }
2296 set_result(_gvn.transform(n));
2297 return true;
2298 }
2299
2300 //----------------------------inline_unsafe_access----------------------------
2301
2302 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2303 // Attempt to infer a sharper value type from the offset and base type.
2304 ciKlass* sharpened_klass = NULL;
2305
2306 // See if it is an instance field, with an object type.
2307 if (alias_type->field() != NULL) {
2308 if (alias_type->field()->type()->is_klass()) {
2309 sharpened_klass = alias_type->field()->type()->as_klass();
2310 }
2311 }
2312
2313 // See if it is a narrow oop array.
2314 if (adr_type->isa_aryptr()) {
2315 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2316 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2317 if (elem_type != NULL) {
2318 sharpened_klass = elem_type->klass();
2319 }
2320 }
2321 }
2322
2323 // The sharpened class might be unloaded if there is no class loader
2324 // contraint in place.
2325 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2326 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2327
2328 #ifndef PRODUCT
2329 if (C->print_intrinsics() || C->print_inlining()) {
2330 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2331 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2332 }
2333 #endif
2334 // Sharpen the value type.
2335 return tjp;
2336 }
2337 return NULL;
2338 }
2339
2340 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) {
2341 switch (kind) {
2342 case Relaxed:
2343 return MO_UNORDERED;
2344 case Opaque:
2345 return MO_RELAXED;
2346 case Acquire:
2347 return MO_ACQUIRE;
2348 case Release:
2349 return MO_RELEASE;
2350 case Volatile:
2351 return MO_SEQ_CST;
2352 default:
2353 ShouldNotReachHere();
2354 return 0;
2355 }
2356 }
2357
2358 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2359 if (callee()->is_static()) return false; // caller must have the capability!
2360 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2361 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2362 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2363 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2364
2365 if (type == T_OBJECT || type == T_ARRAY) {
2366 decorators |= ON_UNKNOWN_OOP_REF;
2367 }
2368
2369 if (unaligned) {
2370 decorators |= C2_UNALIGNED;
2371 }
2372
2373 #ifndef PRODUCT
2374 {
2375 ResourceMark rm;
2376 // Check the signatures.
2377 ciSignature* sig = callee()->signature();
2378 #ifdef ASSERT
2379 if (!is_store) {
2380 // Object getObject(Object base, int/long offset), etc.
2381 BasicType rtype = sig->return_type()->basic_type();
2382 assert(rtype == type, "getter must return the expected value");
2383 assert(sig->count() == 2, "oop getter has 2 arguments");
2384 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2385 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2386 } else {
2387 // void putObject(Object base, int/long offset, Object x), etc.
2388 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2389 assert(sig->count() == 3, "oop putter has 3 arguments");
2390 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2391 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2392 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2393 assert(vtype == type, "putter must accept the expected value");
2394 }
2395 #endif // ASSERT
2396 }
2397 #endif //PRODUCT
2398
2399 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2400
2401 Node* receiver = argument(0); // type: oop
2402
2403 // Build address expression.
2404 Node* adr;
2405 Node* heap_base_oop = top();
2406 Node* offset = top();
2407 Node* val;
2408
2409 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2410 Node* base = argument(1); // type: oop
2411 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2412 offset = argument(2); // type: long
2413 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2414 // to be plain byte offsets, which are also the same as those accepted
2415 // by oopDesc::field_addr.
2416 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2417 "fieldOffset must be byte-scaled");
2418 // 32-bit machines ignore the high half!
2419 offset = ConvL2X(offset);
2420 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2421
2422 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2423 heap_base_oop = base;
2424 } else if (type == T_OBJECT) {
2425 return false; // off-heap oop accesses are not supported
2426 }
2427
2428 // Can base be NULL? Otherwise, always on-heap access.
2429 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2430
2431 if (!can_access_non_heap) {
2432 decorators |= IN_HEAP;
2433 }
2434
2435 val = is_store ? argument(4) : NULL;
2436
2437 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2438
2439 // Try to categorize the address.
2440 Compile::AliasType* alias_type = C->alias_type(adr_type);
2441 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2442
2443 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2444 alias_type->adr_type() == TypeAryPtr::RANGE) {
2445 return false; // not supported
2446 }
2447
2448 bool mismatched = false;
2449 BasicType bt = alias_type->basic_type();
2450 if (bt != T_ILLEGAL) {
2451 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2452 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2453 // Alias type doesn't differentiate between byte[] and boolean[]).
2454 // Use address type to get the element type.
2455 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2456 }
2457 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2458 // accessing an array field with getObject is not a mismatch
2459 bt = T_OBJECT;
2460 }
2461 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2462 // Don't intrinsify mismatched object accesses
2463 return false;
2464 }
2465 mismatched = (bt != type);
2466 } else if (alias_type->adr_type()->isa_oopptr()) {
2467 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2468 }
2469
2470 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2471
2472 if (mismatched) {
2473 decorators |= C2_MISMATCHED;
2474 }
2475
2476 // First guess at the value type.
2477 const Type *value_type = Type::get_const_basic_type(type);
2478
2479 // Figure out the memory ordering.
2480 decorators |= mo_decorator_for_access_kind(kind);
2481
2482 if (!is_store && type == T_OBJECT) {
2483 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2484 if (tjp != NULL) {
2485 value_type = tjp;
2486 }
2487 }
2488
2489 receiver = null_check(receiver);
2490 if (stopped()) {
2491 return true;
2492 }
2493 // Heap pointers get a null-check from the interpreter,
2494 // as a courtesy. However, this is not guaranteed by Unsafe,
2495 // and it is not possible to fully distinguish unintended nulls
2496 // from intended ones in this API.
2497
2498 if (!is_store) {
2499 Node* p = NULL;
2500 // Try to constant fold a load from a constant field
2501 ciField* field = alias_type->field();
2502 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2503 // final or stable field
2504 p = make_constant_from_field(field, heap_base_oop);
2505 }
2506
2507 if (p == NULL) { // Could not constant fold the load
2508 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators);
2509 // Normalize the value returned by getBoolean in the following cases
2510 if (type == T_BOOLEAN &&
2511 (mismatched ||
2512 heap_base_oop == top() || // - heap_base_oop is NULL or
2513 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL
2514 // and the unsafe access is made to large offset
2515 // (i.e., larger than the maximum offset necessary for any
2516 // field access)
2517 ) {
2518 IdealKit ideal = IdealKit(this);
2519 #define __ ideal.
2520 IdealVariable normalized_result(ideal);
2521 __ declarations_done();
2522 __ set(normalized_result, p);
2523 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2524 __ set(normalized_result, ideal.ConI(1));
2525 ideal.end_if();
2526 final_sync(ideal);
2527 p = __ value(normalized_result);
2528 #undef __
2529 }
2530 }
2531 if (type == T_ADDRESS) {
2532 p = gvn().transform(new CastP2XNode(NULL, p));
2533 p = ConvX2UL(p);
2534 }
2535 // The load node has the control of the preceding MemBarCPUOrder. All
2536 // following nodes will have the control of the MemBarCPUOrder inserted at
2537 // the end of this method. So, pushing the load onto the stack at a later
2538 // point is fine.
2539 set_result(p);
2540 } else {
2541 if (bt == T_ADDRESS) {
2542 // Repackage the long as a pointer.
2543 val = ConvL2X(val);
2544 val = gvn().transform(new CastX2PNode(val));
2545 }
2546 access_store_at(control(), heap_base_oop, adr, adr_type, val, value_type, type, decorators);
2547 }
2548
2549 return true;
2550 }
2551
2552 //----------------------------inline_unsafe_load_store----------------------------
2553 // This method serves a couple of different customers (depending on LoadStoreKind):
2554 //
2555 // LS_cmp_swap:
2556 //
2557 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2558 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2559 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2560 //
2561 // LS_cmp_swap_weak:
2562 //
2563 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2564 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2565 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2566 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2567 //
2568 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2569 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2570 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2571 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2572 //
2573 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2574 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2575 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2576 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2577 //
2578 // LS_cmp_exchange:
2579 //
2580 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2581 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2582 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2583 //
2584 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2585 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2586 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2587 //
2588 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2589 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2590 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2591 //
2592 // LS_get_add:
2593 //
2594 // int getAndAddInt( Object o, long offset, int delta)
2595 // long getAndAddLong(Object o, long offset, long delta)
2596 //
2597 // LS_get_set:
2598 //
2599 // int getAndSet(Object o, long offset, int newValue)
2600 // long getAndSet(Object o, long offset, long newValue)
2601 // Object getAndSet(Object o, long offset, Object newValue)
2602 //
2603 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2604 // This basic scheme here is the same as inline_unsafe_access, but
2605 // differs in enough details that combining them would make the code
2606 // overly confusing. (This is a true fact! I originally combined
2607 // them, but even I was confused by it!) As much code/comments as
2608 // possible are retained from inline_unsafe_access though to make
2609 // the correspondences clearer. - dl
2610
2611 if (callee()->is_static()) return false; // caller must have the capability!
2612
2613 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2614 decorators |= mo_decorator_for_access_kind(access_kind);
2615
2616 #ifndef PRODUCT
2617 BasicType rtype;
2618 {
2619 ResourceMark rm;
2620 // Check the signatures.
2621 ciSignature* sig = callee()->signature();
2622 rtype = sig->return_type()->basic_type();
2623 switch(kind) {
2624 case LS_get_add:
2625 case LS_get_set: {
2626 // Check the signatures.
2627 #ifdef ASSERT
2628 assert(rtype == type, "get and set must return the expected type");
2629 assert(sig->count() == 3, "get and set has 3 arguments");
2630 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2631 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2632 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2633 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2634 #endif // ASSERT
2635 break;
2636 }
2637 case LS_cmp_swap:
2638 case LS_cmp_swap_weak: {
2639 // Check the signatures.
2640 #ifdef ASSERT
2641 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2642 assert(sig->count() == 4, "CAS has 4 arguments");
2643 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2644 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2645 #endif // ASSERT
2646 break;
2647 }
2648 case LS_cmp_exchange: {
2649 // Check the signatures.
2650 #ifdef ASSERT
2651 assert(rtype == type, "CAS must return the expected type");
2652 assert(sig->count() == 4, "CAS has 4 arguments");
2653 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2654 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2655 #endif // ASSERT
2656 break;
2657 }
2658 default:
2659 ShouldNotReachHere();
2660 }
2661 }
2662 #endif //PRODUCT
2663
2664 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2665
2666 // Get arguments:
2667 Node* receiver = NULL;
2668 Node* base = NULL;
2669 Node* offset = NULL;
2670 Node* oldval = NULL;
2671 Node* newval = NULL;
2672 switch(kind) {
2673 case LS_cmp_swap:
2674 case LS_cmp_swap_weak:
2675 case LS_cmp_exchange: {
2676 const bool two_slot_type = type2size[type] == 2;
2677 receiver = argument(0); // type: oop
2678 base = argument(1); // type: oop
2679 offset = argument(2); // type: long
2680 oldval = argument(4); // type: oop, int, or long
2681 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2682 break;
2683 }
2684 case LS_get_add:
2685 case LS_get_set: {
2686 receiver = argument(0); // type: oop
2687 base = argument(1); // type: oop
2688 offset = argument(2); // type: long
2689 oldval = NULL;
2690 newval = argument(4); // type: oop, int, or long
2691 break;
2692 }
2693 default:
2694 ShouldNotReachHere();
2695 }
2696
2697 // Build field offset expression.
2698 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2699 // to be plain byte offsets, which are also the same as those accepted
2700 // by oopDesc::field_addr.
2701 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2702 // 32-bit machines ignore the high half of long offsets
2703 offset = ConvL2X(offset);
2704 Node* adr = make_unsafe_address(base, offset, type, false);
2705 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2706
2707 Compile::AliasType* alias_type = C->alias_type(adr_type);
2708 BasicType bt = alias_type->basic_type();
2709 if (bt != T_ILLEGAL &&
2710 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2711 // Don't intrinsify mismatched object accesses.
2712 return false;
2713 }
2714
2715 // For CAS, unlike inline_unsafe_access, there seems no point in
2716 // trying to refine types. Just use the coarse types here.
2717 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2718 const Type *value_type = Type::get_const_basic_type(type);
2719
2720 switch (kind) {
2721 case LS_get_set:
2722 case LS_cmp_exchange: {
2723 if (type == T_OBJECT) {
2724 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2725 if (tjp != NULL) {
2726 value_type = tjp;
2727 }
2728 }
2729 break;
2730 }
2731 case LS_cmp_swap:
2732 case LS_cmp_swap_weak:
2733 case LS_get_add:
2734 break;
2735 default:
2736 ShouldNotReachHere();
2737 }
2738
2739 // Null check receiver.
2740 receiver = null_check(receiver);
2741 if (stopped()) {
2742 return true;
2743 }
2744
2745 int alias_idx = C->get_alias_index(adr_type);
2746
2747 if (type == T_OBJECT || type == T_ARRAY) {
2748 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF;
2749
2750 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2751 // could be delayed during Parse (for example, in adjust_map_after_if()).
2752 // Execute transformation here to avoid barrier generation in such case.
2753 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2754 newval = _gvn.makecon(TypePtr::NULL_PTR);
2755
2756 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) {
2757 // Refine the value to a null constant, when it is known to be null
2758 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2759 }
2760 }
2761
2762 Node* result = NULL;
2763 switch (kind) {
2764 case LS_cmp_exchange: {
2765 result = access_atomic_cmpxchg_val_at(control(), base, adr, adr_type, alias_idx,
2766 oldval, newval, value_type, type, decorators);
2767 break;
2768 }
2769 case LS_cmp_swap_weak:
2770 decorators |= C2_WEAK_CMPXCHG;
2771 case LS_cmp_swap: {
2772 result = access_atomic_cmpxchg_bool_at(control(), base, adr, adr_type, alias_idx,
2773 oldval, newval, value_type, type, decorators);
2774 break;
2775 }
2776 case LS_get_set: {
2777 result = access_atomic_xchg_at(control(), base, adr, adr_type, alias_idx,
2778 newval, value_type, type, decorators);
2779 break;
2780 }
2781 case LS_get_add: {
2782 result = access_atomic_add_at(control(), base, adr, adr_type, alias_idx,
2783 newval, value_type, type, decorators);
2784 break;
2785 }
2786 default:
2787 ShouldNotReachHere();
2788 }
2789
2790 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2791 set_result(result);
2792 return true;
2793 }
2794
2795 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2796 // Regardless of form, don't allow previous ld/st to move down,
2797 // then issue acquire, release, or volatile mem_bar.
2798 insert_mem_bar(Op_MemBarCPUOrder);
2799 switch(id) {
2800 case vmIntrinsics::_loadFence:
2801 insert_mem_bar(Op_LoadFence);
2802 return true;
2803 case vmIntrinsics::_storeFence:
2804 insert_mem_bar(Op_StoreFence);
2805 return true;
2806 case vmIntrinsics::_fullFence:
2807 insert_mem_bar(Op_MemBarVolatile);
2808 return true;
2809 default:
2810 fatal_unexpected_iid(id);
2811 return false;
2812 }
2813 }
2814
2815 bool LibraryCallKit::inline_onspinwait() {
2816 insert_mem_bar(Op_OnSpinWait);
2817 return true;
2818 }
2819
2820 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2821 if (!kls->is_Con()) {
2822 return true;
2823 }
2824 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2825 if (klsptr == NULL) {
2826 return true;
2827 }
2828 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2829 // don't need a guard for a klass that is already initialized
2830 return !ik->is_initialized();
2831 }
2832
2833 //----------------------------inline_unsafe_allocate---------------------------
2834 // public native Object Unsafe.allocateInstance(Class<?> cls);
2835 bool LibraryCallKit::inline_unsafe_allocate() {
2836 if (callee()->is_static()) return false; // caller must have the capability!
2837
2838 null_check_receiver(); // null-check, then ignore
2839 Node* cls = null_check(argument(1));
2840 if (stopped()) return true;
2841
2842 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2843 kls = null_check(kls);
2844 if (stopped()) return true; // argument was like int.class
2845
2846 Node* test = NULL;
2847 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2848 // Note: The argument might still be an illegal value like
2849 // Serializable.class or Object[].class. The runtime will handle it.
2850 // But we must make an explicit check for initialization.
2851 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2852 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2853 // can generate code to load it as unsigned byte.
2854 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2855 Node* bits = intcon(InstanceKlass::fully_initialized);
2856 test = _gvn.transform(new SubINode(inst, bits));
2857 // The 'test' is non-zero if we need to take a slow path.
2858 }
2859
2860 Node* obj = new_instance(kls, test);
2861 set_result(obj);
2862 return true;
2863 }
2864
2865 //------------------------inline_native_time_funcs--------------
2866 // inline code for System.currentTimeMillis() and System.nanoTime()
2867 // these have the same type and signature
2868 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2869 const TypeFunc* tf = OptoRuntime::void_long_Type();
2870 const TypePtr* no_memory_effects = NULL;
2871 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2872 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2873 #ifdef ASSERT
2874 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2875 assert(value_top == top(), "second value must be top");
2876 #endif
2877 set_result(value);
2878 return true;
2879 }
2880
2881 #ifdef JFR_HAVE_INTRINSICS
2882
2883 /*
2884 * oop -> myklass
2885 * myklass->trace_id |= USED
2886 * return myklass->trace_id & ~0x3
2887 */
2888 bool LibraryCallKit::inline_native_classID() {
2889 Node* cls = null_check(argument(0), T_OBJECT);
2890 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2891 kls = null_check(kls, T_OBJECT);
2892
2893 ByteSize offset = KLASS_TRACE_ID_OFFSET;
2894 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2895 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2896
2897 Node* clsused = longcon(0x01l); // set the class bit
2898 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2899 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2900 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2901
2902 #ifdef TRACE_ID_META_BITS
2903 Node* mbits = longcon(~TRACE_ID_META_BITS);
2904 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
2905 #endif
2906 #ifdef TRACE_ID_SHIFT
2907 Node* cbits = intcon(TRACE_ID_SHIFT);
2908 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
2909 #endif
2910
2911 set_result(tvalue);
2912 return true;
2913
2914 }
2915
2916 bool LibraryCallKit::inline_native_getEventWriter() {
2917 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
2918
2919 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
2920 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR));
2921
2922 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
2923
2924 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
2925 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
2926
2927 IfNode* iff_jobj_null =
2928 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
2929
2930 enum { _normal_path = 1,
2931 _null_path = 2,
2932 PATH_LIMIT };
2933
2934 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
2935 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM);
2936
2937 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
2938 result_rgn->init_req(_null_path, jobj_is_null);
2939 result_val->init_req(_null_path, null());
2940
2941 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
2942 set_control(jobj_is_not_null);
2943 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT,
2944 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD);
2945 result_rgn->init_req(_normal_path, control());
2946 result_val->init_req(_normal_path, res);
2947
2948 set_result(result_rgn, result_val);
2949
2950 return true;
2951 }
2952
2953 #endif // JFR_HAVE_INTRINSICS
2954
2955 //------------------------inline_native_currentThread------------------
2956 bool LibraryCallKit::inline_native_currentThread() {
2957 Node* junk = NULL;
2958 set_result(generate_current_thread(junk));
2959 return true;
2960 }
2961
2962 //------------------------inline_native_isInterrupted------------------
2963 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2964 bool LibraryCallKit::inline_native_isInterrupted() {
2965 // Add a fast path to t.isInterrupted(clear_int):
2966 // (t == Thread.current() &&
2967 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
2968 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2969 // So, in the common case that the interrupt bit is false,
2970 // we avoid making a call into the VM. Even if the interrupt bit
2971 // is true, if the clear_int argument is false, we avoid the VM call.
2972 // However, if the receiver is not currentThread, we must call the VM,
2973 // because there must be some locking done around the operation.
2974
2975 // We only go to the fast case code if we pass two guards.
2976 // Paths which do not pass are accumulated in the slow_region.
2977
2978 enum {
2979 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2980 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2981 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2982 PATH_LIMIT
2983 };
2984
2985 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
2986 // out of the function.
2987 insert_mem_bar(Op_MemBarCPUOrder);
2988
2989 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
2990 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
2991
2992 RegionNode* slow_region = new RegionNode(1);
2993 record_for_igvn(slow_region);
2994
2995 // (a) Receiving thread must be the current thread.
2996 Node* rec_thr = argument(0);
2997 Node* tls_ptr = NULL;
2998 Node* cur_thr = generate_current_thread(tls_ptr);
2999 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3000 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3001
3002 generate_slow_guard(bol_thr, slow_region);
3003
3004 // (b) Interrupt bit on TLS must be false.
3005 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3006 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3007 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3008
3009 // Set the control input on the field _interrupted read to prevent it floating up.
3010 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3011 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3012 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3013
3014 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3015
3016 // First fast path: if (!TLS._interrupted) return false;
3017 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3018 result_rgn->init_req(no_int_result_path, false_bit);
3019 result_val->init_req(no_int_result_path, intcon(0));
3020
3021 // drop through to next case
3022 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3023
3024 #ifndef _WINDOWS
3025 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3026 Node* clr_arg = argument(1);
3027 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3028 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3029 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3030
3031 // Second fast path: ... else if (!clear_int) return true;
3032 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3033 result_rgn->init_req(no_clear_result_path, false_arg);
3034 result_val->init_req(no_clear_result_path, intcon(1));
3035
3036 // drop through to next case
3037 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3038 #else
3039 // To return true on Windows you must read the _interrupted field
3040 // and check the event state i.e. take the slow path.
3041 #endif // _WINDOWS
3042
3043 // (d) Otherwise, go to the slow path.
3044 slow_region->add_req(control());
3045 set_control( _gvn.transform(slow_region));
3046
3047 if (stopped()) {
3048 // There is no slow path.
3049 result_rgn->init_req(slow_result_path, top());
3050 result_val->init_req(slow_result_path, top());
3051 } else {
3052 // non-virtual because it is a private non-static
3053 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3054
3055 Node* slow_val = set_results_for_java_call(slow_call);
3056 // this->control() comes from set_results_for_java_call
3057
3058 Node* fast_io = slow_call->in(TypeFunc::I_O);
3059 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3060
3061 // These two phis are pre-filled with copies of of the fast IO and Memory
3062 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3063 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3064
3065 result_rgn->init_req(slow_result_path, control());
3066 result_io ->init_req(slow_result_path, i_o());
3067 result_mem->init_req(slow_result_path, reset_memory());
3068 result_val->init_req(slow_result_path, slow_val);
3069
3070 set_all_memory(_gvn.transform(result_mem));
3071 set_i_o( _gvn.transform(result_io));
3072 }
3073
3074 C->set_has_split_ifs(true); // Has chance for split-if optimization
3075 set_result(result_rgn, result_val);
3076 return true;
3077 }
3078
3079 //---------------------------load_mirror_from_klass----------------------------
3080 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3081 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3082 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3083 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3084 // mirror = ((OopHandle)mirror)->resolve();
3085 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
3086 }
3087
3088 //-----------------------load_klass_from_mirror_common-------------------------
3089 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3090 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3091 // and branch to the given path on the region.
3092 // If never_see_null, take an uncommon trap on null, so we can optimistically
3093 // compile for the non-null case.
3094 // If the region is NULL, force never_see_null = true.
3095 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3096 bool never_see_null,
3097 RegionNode* region,
3098 int null_path,
3099 int offset) {
3100 if (region == NULL) never_see_null = true;
3101 Node* p = basic_plus_adr(mirror, offset);
3102 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3103 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3104 Node* null_ctl = top();
3105 kls = null_check_oop(kls, &null_ctl, never_see_null);
3106 if (region != NULL) {
3107 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3108 region->init_req(null_path, null_ctl);
3109 } else {
3110 assert(null_ctl == top(), "no loose ends");
3111 }
3112 return kls;
3113 }
3114
3115 //--------------------(inline_native_Class_query helpers)---------------------
3116 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3117 // Fall through if (mods & mask) == bits, take the guard otherwise.
3118 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3119 // Branch around if the given klass has the given modifier bit set.
3120 // Like generate_guard, adds a new path onto the region.
3121 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3122 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3123 Node* mask = intcon(modifier_mask);
3124 Node* bits = intcon(modifier_bits);
3125 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3126 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3127 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3128 return generate_fair_guard(bol, region);
3129 }
3130 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3131 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3132 }
3133
3134 //-------------------------inline_native_Class_query-------------------
3135 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3136 const Type* return_type = TypeInt::BOOL;
3137 Node* prim_return_value = top(); // what happens if it's a primitive class?
3138 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3139 bool expect_prim = false; // most of these guys expect to work on refs
3140
3141 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3142
3143 Node* mirror = argument(0);
3144 Node* obj = top();
3145
3146 switch (id) {
3147 case vmIntrinsics::_isInstance:
3148 // nothing is an instance of a primitive type
3149 prim_return_value = intcon(0);
3150 obj = argument(1);
3151 break;
3152 case vmIntrinsics::_getModifiers:
3153 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3154 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3155 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3156 break;
3157 case vmIntrinsics::_isInterface:
3158 prim_return_value = intcon(0);
3159 break;
3160 case vmIntrinsics::_isArray:
3161 prim_return_value = intcon(0);
3162 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3163 break;
3164 case vmIntrinsics::_isPrimitive:
3165 prim_return_value = intcon(1);
3166 expect_prim = true; // obviously
3167 break;
3168 case vmIntrinsics::_getSuperclass:
3169 prim_return_value = null();
3170 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3171 break;
3172 case vmIntrinsics::_getClassAccessFlags:
3173 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3174 return_type = TypeInt::INT; // not bool! 6297094
3175 break;
3176 default:
3177 fatal_unexpected_iid(id);
3178 break;
3179 }
3180
3181 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3182 if (mirror_con == NULL) return false; // cannot happen?
3183
3184 #ifndef PRODUCT
3185 if (C->print_intrinsics() || C->print_inlining()) {
3186 ciType* k = mirror_con->java_mirror_type();
3187 if (k) {
3188 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3189 k->print_name();
3190 tty->cr();
3191 }
3192 }
3193 #endif
3194
3195 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3196 RegionNode* region = new RegionNode(PATH_LIMIT);
3197 record_for_igvn(region);
3198 PhiNode* phi = new PhiNode(region, return_type);
3199
3200 // The mirror will never be null of Reflection.getClassAccessFlags, however
3201 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3202 // if it is. See bug 4774291.
3203
3204 // For Reflection.getClassAccessFlags(), the null check occurs in
3205 // the wrong place; see inline_unsafe_access(), above, for a similar
3206 // situation.
3207 mirror = null_check(mirror);
3208 // If mirror or obj is dead, only null-path is taken.
3209 if (stopped()) return true;
3210
3211 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3212
3213 // Now load the mirror's klass metaobject, and null-check it.
3214 // Side-effects region with the control path if the klass is null.
3215 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3216 // If kls is null, we have a primitive mirror.
3217 phi->init_req(_prim_path, prim_return_value);
3218 if (stopped()) { set_result(region, phi); return true; }
3219 bool safe_for_replace = (region->in(_prim_path) == top());
3220
3221 Node* p; // handy temp
3222 Node* null_ctl;
3223
3224 // Now that we have the non-null klass, we can perform the real query.
3225 // For constant classes, the query will constant-fold in LoadNode::Value.
3226 Node* query_value = top();
3227 switch (id) {
3228 case vmIntrinsics::_isInstance:
3229 // nothing is an instance of a primitive type
3230 query_value = gen_instanceof(obj, kls, safe_for_replace);
3231 break;
3232
3233 case vmIntrinsics::_getModifiers:
3234 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3235 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3236 break;
3237
3238 case vmIntrinsics::_isInterface:
3239 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3240 if (generate_interface_guard(kls, region) != NULL)
3241 // A guard was added. If the guard is taken, it was an interface.
3242 phi->add_req(intcon(1));
3243 // If we fall through, it's a plain class.
3244 query_value = intcon(0);
3245 break;
3246
3247 case vmIntrinsics::_isArray:
3248 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3249 if (generate_array_guard(kls, region) != NULL)
3250 // A guard was added. If the guard is taken, it was an array.
3251 phi->add_req(intcon(1));
3252 // If we fall through, it's a plain class.
3253 query_value = intcon(0);
3254 break;
3255
3256 case vmIntrinsics::_isPrimitive:
3257 query_value = intcon(0); // "normal" path produces false
3258 break;
3259
3260 case vmIntrinsics::_getSuperclass:
3261 // The rules here are somewhat unfortunate, but we can still do better
3262 // with random logic than with a JNI call.
3263 // Interfaces store null or Object as _super, but must report null.
3264 // Arrays store an intermediate super as _super, but must report Object.
3265 // Other types can report the actual _super.
3266 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3267 if (generate_interface_guard(kls, region) != NULL)
3268 // A guard was added. If the guard is taken, it was an interface.
3269 phi->add_req(null());
3270 if (generate_array_guard(kls, region) != NULL)
3271 // A guard was added. If the guard is taken, it was an array.
3272 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3273 // If we fall through, it's a plain class. Get its _super.
3274 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3275 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3276 null_ctl = top();
3277 kls = null_check_oop(kls, &null_ctl);
3278 if (null_ctl != top()) {
3279 // If the guard is taken, Object.superClass is null (both klass and mirror).
3280 region->add_req(null_ctl);
3281 phi ->add_req(null());
3282 }
3283 if (!stopped()) {
3284 query_value = load_mirror_from_klass(kls);
3285 }
3286 break;
3287
3288 case vmIntrinsics::_getClassAccessFlags:
3289 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3290 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3291 break;
3292
3293 default:
3294 fatal_unexpected_iid(id);
3295 break;
3296 }
3297
3298 // Fall-through is the normal case of a query to a real class.
3299 phi->init_req(1, query_value);
3300 region->init_req(1, control());
3301
3302 C->set_has_split_ifs(true); // Has chance for split-if optimization
3303 set_result(region, phi);
3304 return true;
3305 }
3306
3307 //-------------------------inline_Class_cast-------------------
3308 bool LibraryCallKit::inline_Class_cast() {
3309 Node* mirror = argument(0); // Class
3310 Node* obj = argument(1);
3311 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3312 if (mirror_con == NULL) {
3313 return false; // dead path (mirror->is_top()).
3314 }
3315 if (obj == NULL || obj->is_top()) {
3316 return false; // dead path
3317 }
3318 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3319
3320 // First, see if Class.cast() can be folded statically.
3321 // java_mirror_type() returns non-null for compile-time Class constants.
3322 ciType* tm = mirror_con->java_mirror_type();
3323 if (tm != NULL && tm->is_klass() &&
3324 tp != NULL && tp->klass() != NULL) {
3325 if (!tp->klass()->is_loaded()) {
3326 // Don't use intrinsic when class is not loaded.
3327 return false;
3328 } else {
3329 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3330 if (static_res == Compile::SSC_always_true) {
3331 // isInstance() is true - fold the code.
3332 set_result(obj);
3333 return true;
3334 } else if (static_res == Compile::SSC_always_false) {
3335 // Don't use intrinsic, have to throw ClassCastException.
3336 // If the reference is null, the non-intrinsic bytecode will
3337 // be optimized appropriately.
3338 return false;
3339 }
3340 }
3341 }
3342
3343 // Bailout intrinsic and do normal inlining if exception path is frequent.
3344 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3345 return false;
3346 }
3347
3348 // Generate dynamic checks.
3349 // Class.cast() is java implementation of _checkcast bytecode.
3350 // Do checkcast (Parse::do_checkcast()) optimizations here.
3351
3352 mirror = null_check(mirror);
3353 // If mirror is dead, only null-path is taken.
3354 if (stopped()) {
3355 return true;
3356 }
3357
3358 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3359 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3360 RegionNode* region = new RegionNode(PATH_LIMIT);
3361 record_for_igvn(region);
3362
3363 // Now load the mirror's klass metaobject, and null-check it.
3364 // If kls is null, we have a primitive mirror and
3365 // nothing is an instance of a primitive type.
3366 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3367
3368 Node* res = top();
3369 if (!stopped()) {
3370 Node* bad_type_ctrl = top();
3371 // Do checkcast optimizations.
3372 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3373 region->init_req(_bad_type_path, bad_type_ctrl);
3374 }
3375 if (region->in(_prim_path) != top() ||
3376 region->in(_bad_type_path) != top()) {
3377 // Let Interpreter throw ClassCastException.
3378 PreserveJVMState pjvms(this);
3379 set_control(_gvn.transform(region));
3380 uncommon_trap(Deoptimization::Reason_intrinsic,
3381 Deoptimization::Action_maybe_recompile);
3382 }
3383 if (!stopped()) {
3384 set_result(res);
3385 }
3386 return true;
3387 }
3388
3389
3390 //--------------------------inline_native_subtype_check------------------------
3391 // This intrinsic takes the JNI calls out of the heart of
3392 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3393 bool LibraryCallKit::inline_native_subtype_check() {
3394 // Pull both arguments off the stack.
3395 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3396 args[0] = argument(0);
3397 args[1] = argument(1);
3398 Node* klasses[2]; // corresponding Klasses: superk, subk
3399 klasses[0] = klasses[1] = top();
3400
3401 enum {
3402 // A full decision tree on {superc is prim, subc is prim}:
3403 _prim_0_path = 1, // {P,N} => false
3404 // {P,P} & superc!=subc => false
3405 _prim_same_path, // {P,P} & superc==subc => true
3406 _prim_1_path, // {N,P} => false
3407 _ref_subtype_path, // {N,N} & subtype check wins => true
3408 _both_ref_path, // {N,N} & subtype check loses => false
3409 PATH_LIMIT
3410 };
3411
3412 RegionNode* region = new RegionNode(PATH_LIMIT);
3413 Node* phi = new PhiNode(region, TypeInt::BOOL);
3414 record_for_igvn(region);
3415
3416 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3417 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3418 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3419
3420 // First null-check both mirrors and load each mirror's klass metaobject.
3421 int which_arg;
3422 for (which_arg = 0; which_arg <= 1; which_arg++) {
3423 Node* arg = args[which_arg];
3424 arg = null_check(arg);
3425 if (stopped()) break;
3426 args[which_arg] = arg;
3427
3428 Node* p = basic_plus_adr(arg, class_klass_offset);
3429 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3430 klasses[which_arg] = _gvn.transform(kls);
3431 }
3432
3433 // Having loaded both klasses, test each for null.
3434 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3435 for (which_arg = 0; which_arg <= 1; which_arg++) {
3436 Node* kls = klasses[which_arg];
3437 Node* null_ctl = top();
3438 kls = null_check_oop(kls, &null_ctl, never_see_null);
3439 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3440 region->init_req(prim_path, null_ctl);
3441 if (stopped()) break;
3442 klasses[which_arg] = kls;
3443 }
3444
3445 if (!stopped()) {
3446 // now we have two reference types, in klasses[0..1]
3447 Node* subk = klasses[1]; // the argument to isAssignableFrom
3448 Node* superk = klasses[0]; // the receiver
3449 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3450 // now we have a successful reference subtype check
3451 region->set_req(_ref_subtype_path, control());
3452 }
3453
3454 // If both operands are primitive (both klasses null), then
3455 // we must return true when they are identical primitives.
3456 // It is convenient to test this after the first null klass check.
3457 set_control(region->in(_prim_0_path)); // go back to first null check
3458 if (!stopped()) {
3459 // Since superc is primitive, make a guard for the superc==subc case.
3460 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3461 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3462 generate_guard(bol_eq, region, PROB_FAIR);
3463 if (region->req() == PATH_LIMIT+1) {
3464 // A guard was added. If the added guard is taken, superc==subc.
3465 region->swap_edges(PATH_LIMIT, _prim_same_path);
3466 region->del_req(PATH_LIMIT);
3467 }
3468 region->set_req(_prim_0_path, control()); // Not equal after all.
3469 }
3470
3471 // these are the only paths that produce 'true':
3472 phi->set_req(_prim_same_path, intcon(1));
3473 phi->set_req(_ref_subtype_path, intcon(1));
3474
3475 // pull together the cases:
3476 assert(region->req() == PATH_LIMIT, "sane region");
3477 for (uint i = 1; i < region->req(); i++) {
3478 Node* ctl = region->in(i);
3479 if (ctl == NULL || ctl == top()) {
3480 region->set_req(i, top());
3481 phi ->set_req(i, top());
3482 } else if (phi->in(i) == NULL) {
3483 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3484 }
3485 }
3486
3487 set_control(_gvn.transform(region));
3488 set_result(_gvn.transform(phi));
3489 return true;
3490 }
3491
3492 //---------------------generate_array_guard_common------------------------
3493 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3494 bool obj_array, bool not_array) {
3495
3496 if (stopped()) {
3497 return NULL;
3498 }
3499
3500 // If obj_array/non_array==false/false:
3501 // Branch around if the given klass is in fact an array (either obj or prim).
3502 // If obj_array/non_array==false/true:
3503 // Branch around if the given klass is not an array klass of any kind.
3504 // If obj_array/non_array==true/true:
3505 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3506 // If obj_array/non_array==true/false:
3507 // Branch around if the kls is an oop array (Object[] or subtype)
3508 //
3509 // Like generate_guard, adds a new path onto the region.
3510 jint layout_con = 0;
3511 Node* layout_val = get_layout_helper(kls, layout_con);
3512 if (layout_val == NULL) {
3513 bool query = (obj_array
3514 ? Klass::layout_helper_is_objArray(layout_con)
3515 : Klass::layout_helper_is_array(layout_con));
3516 if (query == not_array) {
3517 return NULL; // never a branch
3518 } else { // always a branch
3519 Node* always_branch = control();
3520 if (region != NULL)
3521 region->add_req(always_branch);
3522 set_control(top());
3523 return always_branch;
3524 }
3525 }
3526 // Now test the correct condition.
3527 jint nval = (obj_array
3528 ? (jint)(Klass::_lh_array_tag_type_value
3529 << Klass::_lh_array_tag_shift)
3530 : Klass::_lh_neutral_value);
3531 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3532 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3533 // invert the test if we are looking for a non-array
3534 if (not_array) btest = BoolTest(btest).negate();
3535 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3536 return generate_fair_guard(bol, region);
3537 }
3538
3539
3540 //-----------------------inline_native_newArray--------------------------
3541 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3542 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3543 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3544 Node* mirror;
3545 Node* count_val;
3546 if (uninitialized) {
3547 mirror = argument(1);
3548 count_val = argument(2);
3549 } else {
3550 mirror = argument(0);
3551 count_val = argument(1);
3552 }
3553
3554 mirror = null_check(mirror);
3555 // If mirror or obj is dead, only null-path is taken.
3556 if (stopped()) return true;
3557
3558 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3559 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3560 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3561 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3562 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3563
3564 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3565 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3566 result_reg, _slow_path);
3567 Node* normal_ctl = control();
3568 Node* no_array_ctl = result_reg->in(_slow_path);
3569
3570 // Generate code for the slow case. We make a call to newArray().
3571 set_control(no_array_ctl);
3572 if (!stopped()) {
3573 // Either the input type is void.class, or else the
3574 // array klass has not yet been cached. Either the
3575 // ensuing call will throw an exception, or else it
3576 // will cache the array klass for next time.
3577 PreserveJVMState pjvms(this);
3578 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3579 Node* slow_result = set_results_for_java_call(slow_call);
3580 // this->control() comes from set_results_for_java_call
3581 result_reg->set_req(_slow_path, control());
3582 result_val->set_req(_slow_path, slow_result);
3583 result_io ->set_req(_slow_path, i_o());
3584 result_mem->set_req(_slow_path, reset_memory());
3585 }
3586
3587 set_control(normal_ctl);
3588 if (!stopped()) {
3589 // Normal case: The array type has been cached in the java.lang.Class.
3590 // The following call works fine even if the array type is polymorphic.
3591 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3592 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3593 result_reg->init_req(_normal_path, control());
3594 result_val->init_req(_normal_path, obj);
3595 result_io ->init_req(_normal_path, i_o());
3596 result_mem->init_req(_normal_path, reset_memory());
3597
3598 if (uninitialized) {
3599 // Mark the allocation so that zeroing is skipped
3600 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3601 alloc->maybe_set_complete(&_gvn);
3602 }
3603 }
3604
3605 // Return the combined state.
3606 set_i_o( _gvn.transform(result_io) );
3607 set_all_memory( _gvn.transform(result_mem));
3608
3609 C->set_has_split_ifs(true); // Has chance for split-if optimization
3610 set_result(result_reg, result_val);
3611 return true;
3612 }
3613
3614 //----------------------inline_native_getLength--------------------------
3615 // public static native int java.lang.reflect.Array.getLength(Object array);
3616 bool LibraryCallKit::inline_native_getLength() {
3617 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3618
3619 Node* array = null_check(argument(0));
3620 // If array is dead, only null-path is taken.
3621 if (stopped()) return true;
3622
3623 // Deoptimize if it is a non-array.
3624 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3625
3626 if (non_array != NULL) {
3627 PreserveJVMState pjvms(this);
3628 set_control(non_array);
3629 uncommon_trap(Deoptimization::Reason_intrinsic,
3630 Deoptimization::Action_maybe_recompile);
3631 }
3632
3633 // If control is dead, only non-array-path is taken.
3634 if (stopped()) return true;
3635
3636 // The works fine even if the array type is polymorphic.
3637 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3638 Node* result = load_array_length(array);
3639
3640 C->set_has_split_ifs(true); // Has chance for split-if optimization
3641 set_result(result);
3642 return true;
3643 }
3644
3645 //------------------------inline_array_copyOf----------------------------
3646 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3647 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3648 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3649 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3650
3651 // Get the arguments.
3652 Node* original = argument(0);
3653 Node* start = is_copyOfRange? argument(1): intcon(0);
3654 Node* end = is_copyOfRange? argument(2): argument(1);
3655 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3656
3657 Node* newcopy = NULL;
3658
3659 // Set the original stack and the reexecute bit for the interpreter to reexecute
3660 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3661 { PreserveReexecuteState preexecs(this);
3662 jvms()->set_should_reexecute(true);
3663
3664 array_type_mirror = null_check(array_type_mirror);
3665 original = null_check(original);
3666
3667 // Check if a null path was taken unconditionally.
3668 if (stopped()) return true;
3669
3670 Node* orig_length = load_array_length(original);
3671
3672 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3673 klass_node = null_check(klass_node);
3674
3675 RegionNode* bailout = new RegionNode(1);
3676 record_for_igvn(bailout);
3677
3678 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3679 // Bail out if that is so.
3680 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3681 if (not_objArray != NULL) {
3682 // Improve the klass node's type from the new optimistic assumption:
3683 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3684 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3685 Node* cast = new CastPPNode(klass_node, akls);
3686 cast->init_req(0, control());
3687 klass_node = _gvn.transform(cast);
3688 }
3689
3690 // Bail out if either start or end is negative.
3691 generate_negative_guard(start, bailout, &start);
3692 generate_negative_guard(end, bailout, &end);
3693
3694 Node* length = end;
3695 if (_gvn.type(start) != TypeInt::ZERO) {
3696 length = _gvn.transform(new SubINode(end, start));
3697 }
3698
3699 // Bail out if length is negative.
3700 // Without this the new_array would throw
3701 // NegativeArraySizeException but IllegalArgumentException is what
3702 // should be thrown
3703 generate_negative_guard(length, bailout, &length);
3704
3705 if (bailout->req() > 1) {
3706 PreserveJVMState pjvms(this);
3707 set_control(_gvn.transform(bailout));
3708 uncommon_trap(Deoptimization::Reason_intrinsic,
3709 Deoptimization::Action_maybe_recompile);
3710 }
3711
3712 if (!stopped()) {
3713 // How many elements will we copy from the original?
3714 // The answer is MinI(orig_length - start, length).
3715 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3716 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3717
3718 // Generate a direct call to the right arraycopy function(s).
3719 // We know the copy is disjoint but we might not know if the
3720 // oop stores need checking.
3721 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3722 // This will fail a store-check if x contains any non-nulls.
3723
3724 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3725 // loads/stores but it is legal only if we're sure the
3726 // Arrays.copyOf would succeed. So we need all input arguments
3727 // to the copyOf to be validated, including that the copy to the
3728 // new array won't trigger an ArrayStoreException. That subtype
3729 // check can be optimized if we know something on the type of
3730 // the input array from type speculation.
3731 if (_gvn.type(klass_node)->singleton()) {
3732 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3733 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3734
3735 int test = C->static_subtype_check(superk, subk);
3736 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3737 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3738 if (t_original->speculative_type() != NULL) {
3739 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3740 }
3741 }
3742 }
3743
3744 bool validated = false;
3745 // Reason_class_check rather than Reason_intrinsic because we
3746 // want to intrinsify even if this traps.
3747 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3748 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3749 klass_node);
3750
3751 if (not_subtype_ctrl != top()) {
3752 PreserveJVMState pjvms(this);
3753 set_control(not_subtype_ctrl);
3754 uncommon_trap(Deoptimization::Reason_class_check,
3755 Deoptimization::Action_make_not_entrant);
3756 assert(stopped(), "Should be stopped");
3757 }
3758 validated = true;
3759 }
3760
3761 if (!stopped()) {
3762 newcopy = new_array(klass_node, length, 0); // no arguments to push
3763
3764 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
3765 load_object_klass(original), klass_node);
3766 if (!is_copyOfRange) {
3767 ac->set_copyof(validated);
3768 } else {
3769 ac->set_copyofrange(validated);
3770 }
3771 Node* n = _gvn.transform(ac);
3772 if (n == ac) {
3773 ac->connect_outputs(this);
3774 } else {
3775 assert(validated, "shouldn't transform if all arguments not validated");
3776 set_all_memory(n);
3777 }
3778 }
3779 }
3780 } // original reexecute is set back here
3781
3782 C->set_has_split_ifs(true); // Has chance for split-if optimization
3783 if (!stopped()) {
3784 set_result(newcopy);
3785 }
3786 return true;
3787 }
3788
3789
3790 //----------------------generate_virtual_guard---------------------------
3791 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3792 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3793 RegionNode* slow_region) {
3794 ciMethod* method = callee();
3795 int vtable_index = method->vtable_index();
3796 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3797 "bad index %d", vtable_index);
3798 // Get the Method* out of the appropriate vtable entry.
3799 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
3800 vtable_index*vtableEntry::size_in_bytes() +
3801 vtableEntry::method_offset_in_bytes();
3802 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3803 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3804
3805 // Compare the target method with the expected method (e.g., Object.hashCode).
3806 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3807
3808 Node* native_call = makecon(native_call_addr);
3809 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3810 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3811
3812 return generate_slow_guard(test_native, slow_region);
3813 }
3814
3815 //-----------------------generate_method_call----------------------------
3816 // Use generate_method_call to make a slow-call to the real
3817 // method if the fast path fails. An alternative would be to
3818 // use a stub like OptoRuntime::slow_arraycopy_Java.
3819 // This only works for expanding the current library call,
3820 // not another intrinsic. (E.g., don't use this for making an
3821 // arraycopy call inside of the copyOf intrinsic.)
3822 CallJavaNode*
3823 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3824 // When compiling the intrinsic method itself, do not use this technique.
3825 guarantee(callee() != C->method(), "cannot make slow-call to self");
3826
3827 ciMethod* method = callee();
3828 // ensure the JVMS we have will be correct for this call
3829 guarantee(method_id == method->intrinsic_id(), "must match");
3830
3831 const TypeFunc* tf = TypeFunc::make(method);
3832 CallJavaNode* slow_call;
3833 if (is_static) {
3834 assert(!is_virtual, "");
3835 slow_call = new CallStaticJavaNode(C, tf,
3836 SharedRuntime::get_resolve_static_call_stub(),
3837 method, bci());
3838 } else if (is_virtual) {
3839 null_check_receiver();
3840 int vtable_index = Method::invalid_vtable_index;
3841 if (UseInlineCaches) {
3842 // Suppress the vtable call
3843 } else {
3844 // hashCode and clone are not a miranda methods,
3845 // so the vtable index is fixed.
3846 // No need to use the linkResolver to get it.
3847 vtable_index = method->vtable_index();
3848 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3849 "bad index %d", vtable_index);
3850 }
3851 slow_call = new CallDynamicJavaNode(tf,
3852 SharedRuntime::get_resolve_virtual_call_stub(),
3853 method, vtable_index, bci());
3854 } else { // neither virtual nor static: opt_virtual
3855 null_check_receiver();
3856 slow_call = new CallStaticJavaNode(C, tf,
3857 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3858 method, bci());
3859 slow_call->set_optimized_virtual(true);
3860 }
3861 set_arguments_for_java_call(slow_call);
3862 set_edges_for_java_call(slow_call);
3863 return slow_call;
3864 }
3865
3866
3867 /**
3868 * Build special case code for calls to hashCode on an object. This call may
3869 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3870 * slightly different code.
3871 */
3872 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3873 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3874 assert(!(is_virtual && is_static), "either virtual, special, or static");
3875
3876 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3877
3878 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3879 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3880 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3881 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3882 Node* obj = NULL;
3883 if (!is_static) {
3884 // Check for hashing null object
3885 obj = null_check_receiver();
3886 if (stopped()) return true; // unconditionally null
3887 result_reg->init_req(_null_path, top());
3888 result_val->init_req(_null_path, top());
3889 } else {
3890 // Do a null check, and return zero if null.
3891 // System.identityHashCode(null) == 0
3892 obj = argument(0);
3893 Node* null_ctl = top();
3894 obj = null_check_oop(obj, &null_ctl);
3895 result_reg->init_req(_null_path, null_ctl);
3896 result_val->init_req(_null_path, _gvn.intcon(0));
3897 }
3898
3899 // Unconditionally null? Then return right away.
3900 if (stopped()) {
3901 set_control( result_reg->in(_null_path));
3902 if (!stopped())
3903 set_result(result_val->in(_null_path));
3904 return true;
3905 }
3906
3907 // We only go to the fast case code if we pass a number of guards. The
3908 // paths which do not pass are accumulated in the slow_region.
3909 RegionNode* slow_region = new RegionNode(1);
3910 record_for_igvn(slow_region);
3911
3912 // If this is a virtual call, we generate a funny guard. We pull out
3913 // the vtable entry corresponding to hashCode() from the target object.
3914 // If the target method which we are calling happens to be the native
3915 // Object hashCode() method, we pass the guard. We do not need this
3916 // guard for non-virtual calls -- the caller is known to be the native
3917 // Object hashCode().
3918 if (is_virtual) {
3919 // After null check, get the object's klass.
3920 Node* obj_klass = load_object_klass(obj);
3921 generate_virtual_guard(obj_klass, slow_region);
3922 }
3923
3924 // Get the header out of the object, use LoadMarkNode when available
3925 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3926 // The control of the load must be NULL. Otherwise, the load can move before
3927 // the null check after castPP removal.
3928 Node* no_ctrl = NULL;
3929 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3930
3931 // Test the header to see if it is unlocked.
3932 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3933 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3934 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3935 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3936 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3937
3938 generate_slow_guard(test_unlocked, slow_region);
3939
3940 // Get the hash value and check to see that it has been properly assigned.
3941 // We depend on hash_mask being at most 32 bits and avoid the use of
3942 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3943 // vm: see markOop.hpp.
3944 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3945 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3946 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3947 // This hack lets the hash bits live anywhere in the mark object now, as long
3948 // as the shift drops the relevant bits into the low 32 bits. Note that
3949 // Java spec says that HashCode is an int so there's no point in capturing
3950 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3951 hshifted_header = ConvX2I(hshifted_header);
3952 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3953
3954 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3955 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3956 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3957
3958 generate_slow_guard(test_assigned, slow_region);
3959
3960 Node* init_mem = reset_memory();
3961 // fill in the rest of the null path:
3962 result_io ->init_req(_null_path, i_o());
3963 result_mem->init_req(_null_path, init_mem);
3964
3965 result_val->init_req(_fast_path, hash_val);
3966 result_reg->init_req(_fast_path, control());
3967 result_io ->init_req(_fast_path, i_o());
3968 result_mem->init_req(_fast_path, init_mem);
3969
3970 // Generate code for the slow case. We make a call to hashCode().
3971 set_control(_gvn.transform(slow_region));
3972 if (!stopped()) {
3973 // No need for PreserveJVMState, because we're using up the present state.
3974 set_all_memory(init_mem);
3975 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3976 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3977 Node* slow_result = set_results_for_java_call(slow_call);
3978 // this->control() comes from set_results_for_java_call
3979 result_reg->init_req(_slow_path, control());
3980 result_val->init_req(_slow_path, slow_result);
3981 result_io ->set_req(_slow_path, i_o());
3982 result_mem ->set_req(_slow_path, reset_memory());
3983 }
3984
3985 // Return the combined state.
3986 set_i_o( _gvn.transform(result_io) );
3987 set_all_memory( _gvn.transform(result_mem));
3988
3989 set_result(result_reg, result_val);
3990 return true;
3991 }
3992
3993 //---------------------------inline_native_getClass----------------------------
3994 // public final native Class<?> java.lang.Object.getClass();
3995 //
3996 // Build special case code for calls to getClass on an object.
3997 bool LibraryCallKit::inline_native_getClass() {
3998 Node* obj = null_check_receiver();
3999 if (stopped()) return true;
4000 set_result(load_mirror_from_klass(load_object_klass(obj)));
4001 return true;
4002 }
4003
4004 //-----------------inline_native_Reflection_getCallerClass---------------------
4005 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4006 //
4007 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4008 //
4009 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4010 // in that it must skip particular security frames and checks for
4011 // caller sensitive methods.
4012 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4013 #ifndef PRODUCT
4014 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4015 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4016 }
4017 #endif
4018
4019 if (!jvms()->has_method()) {
4020 #ifndef PRODUCT
4021 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4022 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4023 }
4024 #endif
4025 return false;
4026 }
4027
4028 // Walk back up the JVM state to find the caller at the required
4029 // depth.
4030 JVMState* caller_jvms = jvms();
4031
4032 // Cf. JVM_GetCallerClass
4033 // NOTE: Start the loop at depth 1 because the current JVM state does
4034 // not include the Reflection.getCallerClass() frame.
4035 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4036 ciMethod* m = caller_jvms->method();
4037 switch (n) {
4038 case 0:
4039 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4040 break;
4041 case 1:
4042 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4043 if (!m->caller_sensitive()) {
4044 #ifndef PRODUCT
4045 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4046 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4047 }
4048 #endif
4049 return false; // bail-out; let JVM_GetCallerClass do the work
4050 }
4051 break;
4052 default:
4053 if (!m->is_ignored_by_security_stack_walk()) {
4054 // We have reached the desired frame; return the holder class.
4055 // Acquire method holder as java.lang.Class and push as constant.
4056 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4057 ciInstance* caller_mirror = caller_klass->java_mirror();
4058 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4059
4060 #ifndef PRODUCT
4061 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4062 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());
4063 tty->print_cr(" JVM state at this point:");
4064 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4065 ciMethod* m = jvms()->of_depth(i)->method();
4066 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4067 }
4068 }
4069 #endif
4070 return true;
4071 }
4072 break;
4073 }
4074 }
4075
4076 #ifndef PRODUCT
4077 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4078 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4079 tty->print_cr(" JVM state at this point:");
4080 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4081 ciMethod* m = jvms()->of_depth(i)->method();
4082 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4083 }
4084 }
4085 #endif
4086
4087 return false; // bail-out; let JVM_GetCallerClass do the work
4088 }
4089
4090 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4091 Node* arg = argument(0);
4092 Node* result = NULL;
4093
4094 switch (id) {
4095 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4096 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4097 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4098 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4099
4100 case vmIntrinsics::_doubleToLongBits: {
4101 // two paths (plus control) merge in a wood
4102 RegionNode *r = new RegionNode(3);
4103 Node *phi = new PhiNode(r, TypeLong::LONG);
4104
4105 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4106 // Build the boolean node
4107 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4108
4109 // Branch either way.
4110 // NaN case is less traveled, which makes all the difference.
4111 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4112 Node *opt_isnan = _gvn.transform(ifisnan);
4113 assert( opt_isnan->is_If(), "Expect an IfNode");
4114 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4115 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4116
4117 set_control(iftrue);
4118
4119 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4120 Node *slow_result = longcon(nan_bits); // return NaN
4121 phi->init_req(1, _gvn.transform( slow_result ));
4122 r->init_req(1, iftrue);
4123
4124 // Else fall through
4125 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4126 set_control(iffalse);
4127
4128 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4129 r->init_req(2, iffalse);
4130
4131 // Post merge
4132 set_control(_gvn.transform(r));
4133 record_for_igvn(r);
4134
4135 C->set_has_split_ifs(true); // Has chance for split-if optimization
4136 result = phi;
4137 assert(result->bottom_type()->isa_long(), "must be");
4138 break;
4139 }
4140
4141 case vmIntrinsics::_floatToIntBits: {
4142 // two paths (plus control) merge in a wood
4143 RegionNode *r = new RegionNode(3);
4144 Node *phi = new PhiNode(r, TypeInt::INT);
4145
4146 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4147 // Build the boolean node
4148 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4149
4150 // Branch either way.
4151 // NaN case is less traveled, which makes all the difference.
4152 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4153 Node *opt_isnan = _gvn.transform(ifisnan);
4154 assert( opt_isnan->is_If(), "Expect an IfNode");
4155 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4156 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4157
4158 set_control(iftrue);
4159
4160 static const jint nan_bits = 0x7fc00000;
4161 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4162 phi->init_req(1, _gvn.transform( slow_result ));
4163 r->init_req(1, iftrue);
4164
4165 // Else fall through
4166 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4167 set_control(iffalse);
4168
4169 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4170 r->init_req(2, iffalse);
4171
4172 // Post merge
4173 set_control(_gvn.transform(r));
4174 record_for_igvn(r);
4175
4176 C->set_has_split_ifs(true); // Has chance for split-if optimization
4177 result = phi;
4178 assert(result->bottom_type()->isa_int(), "must be");
4179 break;
4180 }
4181
4182 default:
4183 fatal_unexpected_iid(id);
4184 break;
4185 }
4186 set_result(_gvn.transform(result));
4187 return true;
4188 }
4189
4190 //----------------------inline_unsafe_copyMemory-------------------------
4191 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4192 bool LibraryCallKit::inline_unsafe_copyMemory() {
4193 if (callee()->is_static()) return false; // caller must have the capability!
4194 null_check_receiver(); // null-check receiver
4195 if (stopped()) return true;
4196
4197 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4198
4199 Node* src_ptr = argument(1); // type: oop
4200 Node* src_off = ConvL2X(argument(2)); // type: long
4201 Node* dst_ptr = argument(4); // type: oop
4202 Node* dst_off = ConvL2X(argument(5)); // type: long
4203 Node* size = ConvL2X(argument(7)); // type: long
4204
4205 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4206 "fieldOffset must be byte-scaled");
4207
4208 Node* src = make_unsafe_address(src_ptr, src_off);
4209 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4210
4211 // Conservatively insert a memory barrier on all memory slices.
4212 // Do not let writes of the copy source or destination float below the copy.
4213 insert_mem_bar(Op_MemBarCPUOrder);
4214
4215 // Call it. Note that the length argument is not scaled.
4216 make_runtime_call(RC_LEAF|RC_NO_FP,
4217 OptoRuntime::fast_arraycopy_Type(),
4218 StubRoutines::unsafe_arraycopy(),
4219 "unsafe_arraycopy",
4220 TypeRawPtr::BOTTOM,
4221 src, dst, size XTOP);
4222
4223 // Do not let reads of the copy destination float above the copy.
4224 insert_mem_bar(Op_MemBarCPUOrder);
4225
4226 return true;
4227 }
4228
4229 //------------------------clone_coping-----------------------------------
4230 // Helper function for inline_native_clone.
4231 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) {
4232 assert(obj_size != NULL, "");
4233 Node* raw_obj = alloc_obj->in(1);
4234 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4235
4236 AllocateNode* alloc = NULL;
4237 if (ReduceBulkZeroing) {
4238 // We will be completely responsible for initializing this object -
4239 // mark Initialize node as complete.
4240 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4241 // The object was just allocated - there should be no any stores!
4242 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4243 // Mark as complete_with_arraycopy so that on AllocateNode
4244 // expansion, we know this AllocateNode is initialized by an array
4245 // copy and a StoreStore barrier exists after the array copy.
4246 alloc->initialization()->set_complete_with_arraycopy();
4247 }
4248
4249 // Copy the fastest available way.
4250 // TODO: generate fields copies for small objects instead.
4251 Node* size = _gvn.transform(obj_size);
4252
4253 access_clone(control(), obj, alloc_obj, size, is_array);
4254
4255 // Do not let reads from the cloned object float above the arraycopy.
4256 if (alloc != NULL) {
4257 // Do not let stores that initialize this object be reordered with
4258 // a subsequent store that would make this object accessible by
4259 // other threads.
4260 // Record what AllocateNode this StoreStore protects so that
4261 // escape analysis can go from the MemBarStoreStoreNode to the
4262 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4263 // based on the escape status of the AllocateNode.
4264 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4265 } else {
4266 insert_mem_bar(Op_MemBarCPUOrder);
4267 }
4268 }
4269
4270 //------------------------inline_native_clone----------------------------
4271 // protected native Object java.lang.Object.clone();
4272 //
4273 // Here are the simple edge cases:
4274 // null receiver => normal trap
4275 // virtual and clone was overridden => slow path to out-of-line clone
4276 // not cloneable or finalizer => slow path to out-of-line Object.clone
4277 //
4278 // The general case has two steps, allocation and copying.
4279 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4280 //
4281 // Copying also has two cases, oop arrays and everything else.
4282 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4283 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4284 //
4285 // These steps fold up nicely if and when the cloned object's klass
4286 // can be sharply typed as an object array, a type array, or an instance.
4287 //
4288 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4289 PhiNode* result_val;
4290
4291 // Set the reexecute bit for the interpreter to reexecute
4292 // the bytecode that invokes Object.clone if deoptimization happens.
4293 { PreserveReexecuteState preexecs(this);
4294 jvms()->set_should_reexecute(true);
4295
4296 Node* obj = null_check_receiver();
4297 if (stopped()) return true;
4298
4299 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4300
4301 // If we are going to clone an instance, we need its exact type to
4302 // know the number and types of fields to convert the clone to
4303 // loads/stores. Maybe a speculative type can help us.
4304 if (!obj_type->klass_is_exact() &&
4305 obj_type->speculative_type() != NULL &&
4306 obj_type->speculative_type()->is_instance_klass()) {
4307 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4308 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4309 !spec_ik->has_injected_fields()) {
4310 ciKlass* k = obj_type->klass();
4311 if (!k->is_instance_klass() ||
4312 k->as_instance_klass()->is_interface() ||
4313 k->as_instance_klass()->has_subklass()) {
4314 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4315 }
4316 }
4317 }
4318
4319 Node* obj_klass = load_object_klass(obj);
4320 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4321 const TypeOopPtr* toop = ((tklass != NULL)
4322 ? tklass->as_instance_type()
4323 : TypeInstPtr::NOTNULL);
4324
4325 // Conservatively insert a memory barrier on all memory slices.
4326 // Do not let writes into the original float below the clone.
4327 insert_mem_bar(Op_MemBarCPUOrder);
4328
4329 // paths into result_reg:
4330 enum {
4331 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4332 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4333 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4334 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4335 PATH_LIMIT
4336 };
4337 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4338 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4339 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4340 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4341 record_for_igvn(result_reg);
4342
4343 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4344 if (array_ctl != NULL) {
4345 // It's an array.
4346 PreserveJVMState pjvms(this);
4347 set_control(array_ctl);
4348 Node* obj_length = load_array_length(obj);
4349 Node* obj_size = NULL;
4350 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4351
4352 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4353 if (bs->array_copy_requires_gc_barriers(T_OBJECT)) {
4354 // If it is an oop array, it requires very special treatment,
4355 // because gc barriers are required when accessing the array.
4356 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4357 if (is_obja != NULL) {
4358 PreserveJVMState pjvms2(this);
4359 set_control(is_obja);
4360 // Generate a direct call to the right arraycopy function(s).
4361 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4362 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4363 ac->set_cloneoop();
4364 Node* n = _gvn.transform(ac);
4365 assert(n == ac, "cannot disappear");
4366 ac->connect_outputs(this);
4367
4368 result_reg->init_req(_objArray_path, control());
4369 result_val->init_req(_objArray_path, alloc_obj);
4370 result_i_o ->set_req(_objArray_path, i_o());
4371 result_mem ->set_req(_objArray_path, reset_memory());
4372 }
4373 }
4374 // Otherwise, there are no barriers to worry about.
4375 // (We can dispense with card marks if we know the allocation
4376 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4377 // causes the non-eden paths to take compensating steps to
4378 // simulate a fresh allocation, so that no further
4379 // card marks are required in compiled code to initialize
4380 // the object.)
4381
4382 if (!stopped()) {
4383 copy_to_clone(obj, alloc_obj, obj_size, true);
4384
4385 // Present the results of the copy.
4386 result_reg->init_req(_array_path, control());
4387 result_val->init_req(_array_path, alloc_obj);
4388 result_i_o ->set_req(_array_path, i_o());
4389 result_mem ->set_req(_array_path, reset_memory());
4390 }
4391 }
4392
4393 // We only go to the instance fast case code if we pass a number of guards.
4394 // The paths which do not pass are accumulated in the slow_region.
4395 RegionNode* slow_region = new RegionNode(1);
4396 record_for_igvn(slow_region);
4397 if (!stopped()) {
4398 // It's an instance (we did array above). Make the slow-path tests.
4399 // If this is a virtual call, we generate a funny guard. We grab
4400 // the vtable entry corresponding to clone() from the target object.
4401 // If the target method which we are calling happens to be the
4402 // Object clone() method, we pass the guard. We do not need this
4403 // guard for non-virtual calls; the caller is known to be the native
4404 // Object clone().
4405 if (is_virtual) {
4406 generate_virtual_guard(obj_klass, slow_region);
4407 }
4408
4409 // The object must be easily cloneable and must not have a finalizer.
4410 // Both of these conditions may be checked in a single test.
4411 // We could optimize the test further, but we don't care.
4412 generate_access_flags_guard(obj_klass,
4413 // Test both conditions:
4414 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4415 // Must be cloneable but not finalizer:
4416 JVM_ACC_IS_CLONEABLE_FAST,
4417 slow_region);
4418 }
4419
4420 if (!stopped()) {
4421 // It's an instance, and it passed the slow-path tests.
4422 PreserveJVMState pjvms(this);
4423 Node* obj_size = NULL;
4424 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4425 // is reexecuted if deoptimization occurs and there could be problems when merging
4426 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4427 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4428
4429 copy_to_clone(obj, alloc_obj, obj_size, false);
4430
4431 // Present the results of the slow call.
4432 result_reg->init_req(_instance_path, control());
4433 result_val->init_req(_instance_path, alloc_obj);
4434 result_i_o ->set_req(_instance_path, i_o());
4435 result_mem ->set_req(_instance_path, reset_memory());
4436 }
4437
4438 // Generate code for the slow case. We make a call to clone().
4439 set_control(_gvn.transform(slow_region));
4440 if (!stopped()) {
4441 PreserveJVMState pjvms(this);
4442 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4443 // We need to deoptimize on exception (see comment above)
4444 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true);
4445 // this->control() comes from set_results_for_java_call
4446 result_reg->init_req(_slow_path, control());
4447 result_val->init_req(_slow_path, slow_result);
4448 result_i_o ->set_req(_slow_path, i_o());
4449 result_mem ->set_req(_slow_path, reset_memory());
4450 }
4451
4452 // Return the combined state.
4453 set_control( _gvn.transform(result_reg));
4454 set_i_o( _gvn.transform(result_i_o));
4455 set_all_memory( _gvn.transform(result_mem));
4456 } // original reexecute is set back here
4457
4458 set_result(_gvn.transform(result_val));
4459 return true;
4460 }
4461
4462 // If we have a tighly coupled allocation, the arraycopy may take care
4463 // of the array initialization. If one of the guards we insert between
4464 // the allocation and the arraycopy causes a deoptimization, an
4465 // unitialized array will escape the compiled method. To prevent that
4466 // we set the JVM state for uncommon traps between the allocation and
4467 // the arraycopy to the state before the allocation so, in case of
4468 // deoptimization, we'll reexecute the allocation and the
4469 // initialization.
4470 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4471 if (alloc != NULL) {
4472 ciMethod* trap_method = alloc->jvms()->method();
4473 int trap_bci = alloc->jvms()->bci();
4474
4475 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4476 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4477 // Make sure there's no store between the allocation and the
4478 // arraycopy otherwise visible side effects could be rexecuted
4479 // in case of deoptimization and cause incorrect execution.
4480 bool no_interfering_store = true;
4481 Node* mem = alloc->in(TypeFunc::Memory);
4482 if (mem->is_MergeMem()) {
4483 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4484 Node* n = mms.memory();
4485 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4486 assert(n->is_Store(), "what else?");
4487 no_interfering_store = false;
4488 break;
4489 }
4490 }
4491 } else {
4492 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4493 Node* n = mms.memory();
4494 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4495 assert(n->is_Store(), "what else?");
4496 no_interfering_store = false;
4497 break;
4498 }
4499 }
4500 }
4501
4502 if (no_interfering_store) {
4503 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4504 uint size = alloc->req();
4505 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4506 old_jvms->set_map(sfpt);
4507 for (uint i = 0; i < size; i++) {
4508 sfpt->init_req(i, alloc->in(i));
4509 }
4510 // re-push array length for deoptimization
4511 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4512 old_jvms->set_sp(old_jvms->sp()+1);
4513 old_jvms->set_monoff(old_jvms->monoff()+1);
4514 old_jvms->set_scloff(old_jvms->scloff()+1);
4515 old_jvms->set_endoff(old_jvms->endoff()+1);
4516 old_jvms->set_should_reexecute(true);
4517
4518 sfpt->set_i_o(map()->i_o());
4519 sfpt->set_memory(map()->memory());
4520 sfpt->set_control(map()->control());
4521
4522 JVMState* saved_jvms = jvms();
4523 saved_reexecute_sp = _reexecute_sp;
4524
4525 set_jvms(sfpt->jvms());
4526 _reexecute_sp = jvms()->sp();
4527
4528 return saved_jvms;
4529 }
4530 }
4531 }
4532 return NULL;
4533 }
4534
4535 // In case of a deoptimization, we restart execution at the
4536 // allocation, allocating a new array. We would leave an uninitialized
4537 // array in the heap that GCs wouldn't expect. Move the allocation
4538 // after the traps so we don't allocate the array if we
4539 // deoptimize. This is possible because tightly_coupled_allocation()
4540 // guarantees there's no observer of the allocated array at this point
4541 // and the control flow is simple enough.
4542 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4543 int saved_reexecute_sp, uint new_idx) {
4544 if (saved_jvms != NULL && !stopped()) {
4545 assert(alloc != NULL, "only with a tightly coupled allocation");
4546 // restore JVM state to the state at the arraycopy
4547 saved_jvms->map()->set_control(map()->control());
4548 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4549 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4550 // If we've improved the types of some nodes (null check) while
4551 // emitting the guards, propagate them to the current state
4552 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4553 set_jvms(saved_jvms);
4554 _reexecute_sp = saved_reexecute_sp;
4555
4556 // Remove the allocation from above the guards
4557 CallProjections callprojs;
4558 alloc->extract_projections(&callprojs, true);
4559 InitializeNode* init = alloc->initialization();
4560 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4561 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4562 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4563 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4564
4565 // move the allocation here (after the guards)
4566 _gvn.hash_delete(alloc);
4567 alloc->set_req(TypeFunc::Control, control());
4568 alloc->set_req(TypeFunc::I_O, i_o());
4569 Node *mem = reset_memory();
4570 set_all_memory(mem);
4571 alloc->set_req(TypeFunc::Memory, mem);
4572 set_control(init->proj_out_or_null(TypeFunc::Control));
4573 set_i_o(callprojs.fallthrough_ioproj);
4574
4575 // Update memory as done in GraphKit::set_output_for_allocation()
4576 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4577 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4578 if (ary_type->isa_aryptr() && length_type != NULL) {
4579 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4580 }
4581 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4582 int elemidx = C->get_alias_index(telemref);
4583 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
4584 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
4585
4586 Node* allocx = _gvn.transform(alloc);
4587 assert(allocx == alloc, "where has the allocation gone?");
4588 assert(dest->is_CheckCastPP(), "not an allocation result?");
4589
4590 _gvn.hash_delete(dest);
4591 dest->set_req(0, control());
4592 Node* destx = _gvn.transform(dest);
4593 assert(destx == dest, "where has the allocation result gone?");
4594 }
4595 }
4596
4597
4598 //------------------------------inline_arraycopy-----------------------
4599 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4600 // Object dest, int destPos,
4601 // int length);
4602 bool LibraryCallKit::inline_arraycopy() {
4603 // Get the arguments.
4604 Node* src = argument(0); // type: oop
4605 Node* src_offset = argument(1); // type: int
4606 Node* dest = argument(2); // type: oop
4607 Node* dest_offset = argument(3); // type: int
4608 Node* length = argument(4); // type: int
4609
4610 uint new_idx = C->unique();
4611
4612 // Check for allocation before we add nodes that would confuse
4613 // tightly_coupled_allocation()
4614 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4615
4616 int saved_reexecute_sp = -1;
4617 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4618 // See arraycopy_restore_alloc_state() comment
4619 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4620 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4621 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
4622 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4623
4624 // The following tests must be performed
4625 // (1) src and dest are arrays.
4626 // (2) src and dest arrays must have elements of the same BasicType
4627 // (3) src and dest must not be null.
4628 // (4) src_offset must not be negative.
4629 // (5) dest_offset must not be negative.
4630 // (6) length must not be negative.
4631 // (7) src_offset + length must not exceed length of src.
4632 // (8) dest_offset + length must not exceed length of dest.
4633 // (9) each element of an oop array must be assignable
4634
4635 // (3) src and dest must not be null.
4636 // always do this here because we need the JVM state for uncommon traps
4637 Node* null_ctl = top();
4638 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4639 assert(null_ctl->is_top(), "no null control here");
4640 dest = null_check(dest, T_ARRAY);
4641
4642 if (!can_emit_guards) {
4643 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4644 // guards but the arraycopy node could still take advantage of a
4645 // tightly allocated allocation. tightly_coupled_allocation() is
4646 // called again to make sure it takes the null check above into
4647 // account: the null check is mandatory and if it caused an
4648 // uncommon trap to be emitted then the allocation can't be
4649 // considered tightly coupled in this context.
4650 alloc = tightly_coupled_allocation(dest, NULL);
4651 }
4652
4653 bool validated = false;
4654
4655 const Type* src_type = _gvn.type(src);
4656 const Type* dest_type = _gvn.type(dest);
4657 const TypeAryPtr* top_src = src_type->isa_aryptr();
4658 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4659
4660 // Do we have the type of src?
4661 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4662 // Do we have the type of dest?
4663 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4664 // Is the type for src from speculation?
4665 bool src_spec = false;
4666 // Is the type for dest from speculation?
4667 bool dest_spec = false;
4668
4669 if ((!has_src || !has_dest) && can_emit_guards) {
4670 // We don't have sufficient type information, let's see if
4671 // speculative types can help. We need to have types for both src
4672 // and dest so that it pays off.
4673
4674 // Do we already have or could we have type information for src
4675 bool could_have_src = has_src;
4676 // Do we already have or could we have type information for dest
4677 bool could_have_dest = has_dest;
4678
4679 ciKlass* src_k = NULL;
4680 if (!has_src) {
4681 src_k = src_type->speculative_type_not_null();
4682 if (src_k != NULL && src_k->is_array_klass()) {
4683 could_have_src = true;
4684 }
4685 }
4686
4687 ciKlass* dest_k = NULL;
4688 if (!has_dest) {
4689 dest_k = dest_type->speculative_type_not_null();
4690 if (dest_k != NULL && dest_k->is_array_klass()) {
4691 could_have_dest = true;
4692 }
4693 }
4694
4695 if (could_have_src && could_have_dest) {
4696 // This is going to pay off so emit the required guards
4697 if (!has_src) {
4698 src = maybe_cast_profiled_obj(src, src_k, true);
4699 src_type = _gvn.type(src);
4700 top_src = src_type->isa_aryptr();
4701 has_src = (top_src != NULL && top_src->klass() != NULL);
4702 src_spec = true;
4703 }
4704 if (!has_dest) {
4705 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4706 dest_type = _gvn.type(dest);
4707 top_dest = dest_type->isa_aryptr();
4708 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4709 dest_spec = true;
4710 }
4711 }
4712 }
4713
4714 if (has_src && has_dest && can_emit_guards) {
4715 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4716 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4717 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4718 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4719
4720 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4721 // If both arrays are object arrays then having the exact types
4722 // for both will remove the need for a subtype check at runtime
4723 // before the call and may make it possible to pick a faster copy
4724 // routine (without a subtype check on every element)
4725 // Do we have the exact type of src?
4726 bool could_have_src = src_spec;
4727 // Do we have the exact type of dest?
4728 bool could_have_dest = dest_spec;
4729 ciKlass* src_k = top_src->klass();
4730 ciKlass* dest_k = top_dest->klass();
4731 if (!src_spec) {
4732 src_k = src_type->speculative_type_not_null();
4733 if (src_k != NULL && src_k->is_array_klass()) {
4734 could_have_src = true;
4735 }
4736 }
4737 if (!dest_spec) {
4738 dest_k = dest_type->speculative_type_not_null();
4739 if (dest_k != NULL && dest_k->is_array_klass()) {
4740 could_have_dest = true;
4741 }
4742 }
4743 if (could_have_src && could_have_dest) {
4744 // If we can have both exact types, emit the missing guards
4745 if (could_have_src && !src_spec) {
4746 src = maybe_cast_profiled_obj(src, src_k, true);
4747 }
4748 if (could_have_dest && !dest_spec) {
4749 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4750 }
4751 }
4752 }
4753 }
4754
4755 ciMethod* trap_method = method();
4756 int trap_bci = bci();
4757 if (saved_jvms != NULL) {
4758 trap_method = alloc->jvms()->method();
4759 trap_bci = alloc->jvms()->bci();
4760 }
4761
4762 bool negative_length_guard_generated = false;
4763
4764 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4765 can_emit_guards &&
4766 !src->is_top() && !dest->is_top()) {
4767 // validate arguments: enables transformation the ArrayCopyNode
4768 validated = true;
4769
4770 RegionNode* slow_region = new RegionNode(1);
4771 record_for_igvn(slow_region);
4772
4773 // (1) src and dest are arrays.
4774 generate_non_array_guard(load_object_klass(src), slow_region);
4775 generate_non_array_guard(load_object_klass(dest), slow_region);
4776
4777 // (2) src and dest arrays must have elements of the same BasicType
4778 // done at macro expansion or at Ideal transformation time
4779
4780 // (4) src_offset must not be negative.
4781 generate_negative_guard(src_offset, slow_region);
4782
4783 // (5) dest_offset must not be negative.
4784 generate_negative_guard(dest_offset, slow_region);
4785
4786 // (7) src_offset + length must not exceed length of src.
4787 generate_limit_guard(src_offset, length,
4788 load_array_length(src),
4789 slow_region);
4790
4791 // (8) dest_offset + length must not exceed length of dest.
4792 generate_limit_guard(dest_offset, length,
4793 load_array_length(dest),
4794 slow_region);
4795
4796 // (6) length must not be negative.
4797 // This is also checked in generate_arraycopy() during macro expansion, but
4798 // we also have to check it here for the case where the ArrayCopyNode will
4799 // be eliminated by Escape Analysis.
4800 if (EliminateAllocations) {
4801 generate_negative_guard(length, slow_region);
4802 negative_length_guard_generated = true;
4803 }
4804
4805 // (9) each element of an oop array must be assignable
4806 Node* src_klass = load_object_klass(src);
4807 Node* dest_klass = load_object_klass(dest);
4808 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4809
4810 if (not_subtype_ctrl != top()) {
4811 PreserveJVMState pjvms(this);
4812 set_control(not_subtype_ctrl);
4813 uncommon_trap(Deoptimization::Reason_intrinsic,
4814 Deoptimization::Action_make_not_entrant);
4815 assert(stopped(), "Should be stopped");
4816 }
4817 {
4818 PreserveJVMState pjvms(this);
4819 set_control(_gvn.transform(slow_region));
4820 uncommon_trap(Deoptimization::Reason_intrinsic,
4821 Deoptimization::Action_make_not_entrant);
4822 assert(stopped(), "Should be stopped");
4823 }
4824
4825 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
4826 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
4827 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
4828 }
4829
4830 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
4831
4832 if (stopped()) {
4833 return true;
4834 }
4835
4836 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
4837 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4838 // so the compiler has a chance to eliminate them: during macro expansion,
4839 // we have to set their control (CastPP nodes are eliminated).
4840 load_object_klass(src), load_object_klass(dest),
4841 load_array_length(src), load_array_length(dest));
4842
4843 ac->set_arraycopy(validated);
4844
4845 Node* n = _gvn.transform(ac);
4846 if (n == ac) {
4847 ac->connect_outputs(this);
4848 } else {
4849 assert(validated, "shouldn't transform if all arguments not validated");
4850 set_all_memory(n);
4851 }
4852 clear_upper_avx();
4853
4854
4855 return true;
4856 }
4857
4858
4859 // Helper function which determines if an arraycopy immediately follows
4860 // an allocation, with no intervening tests or other escapes for the object.
4861 AllocateArrayNode*
4862 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4863 RegionNode* slow_region) {
4864 if (stopped()) return NULL; // no fast path
4865 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4866
4867 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4868 if (alloc == NULL) return NULL;
4869
4870 Node* rawmem = memory(Compile::AliasIdxRaw);
4871 // Is the allocation's memory state untouched?
4872 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4873 // Bail out if there have been raw-memory effects since the allocation.
4874 // (Example: There might have been a call or safepoint.)
4875 return NULL;
4876 }
4877 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4878 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4879 return NULL;
4880 }
4881
4882 // There must be no unexpected observers of this allocation.
4883 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4884 Node* obs = ptr->fast_out(i);
4885 if (obs != this->map()) {
4886 return NULL;
4887 }
4888 }
4889
4890 // This arraycopy must unconditionally follow the allocation of the ptr.
4891 Node* alloc_ctl = ptr->in(0);
4892 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4893
4894 Node* ctl = control();
4895 while (ctl != alloc_ctl) {
4896 // There may be guards which feed into the slow_region.
4897 // Any other control flow means that we might not get a chance
4898 // to finish initializing the allocated object.
4899 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4900 IfNode* iff = ctl->in(0)->as_If();
4901 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
4902 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4903 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4904 ctl = iff->in(0); // This test feeds the known slow_region.
4905 continue;
4906 }
4907 // One more try: Various low-level checks bottom out in
4908 // uncommon traps. If the debug-info of the trap omits
4909 // any reference to the allocation, as we've already
4910 // observed, then there can be no objection to the trap.
4911 bool found_trap = false;
4912 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4913 Node* obs = not_ctl->fast_out(j);
4914 if (obs->in(0) == not_ctl && obs->is_Call() &&
4915 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4916 found_trap = true; break;
4917 }
4918 }
4919 if (found_trap) {
4920 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4921 continue;
4922 }
4923 }
4924 return NULL;
4925 }
4926
4927 // If we get this far, we have an allocation which immediately
4928 // precedes the arraycopy, and we can take over zeroing the new object.
4929 // The arraycopy will finish the initialization, and provide
4930 // a new control state to which we will anchor the destination pointer.
4931
4932 return alloc;
4933 }
4934
4935 //-------------inline_encodeISOArray-----------------------------------
4936 // encode char[] to byte[] in ISO_8859_1
4937 bool LibraryCallKit::inline_encodeISOArray() {
4938 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4939 // no receiver since it is static method
4940 Node *src = argument(0);
4941 Node *src_offset = argument(1);
4942 Node *dst = argument(2);
4943 Node *dst_offset = argument(3);
4944 Node *length = argument(4);
4945
4946 src = must_be_not_null(src, true);
4947 dst = must_be_not_null(dst, true);
4948
4949 const Type* src_type = src->Value(&_gvn);
4950 const Type* dst_type = dst->Value(&_gvn);
4951 const TypeAryPtr* top_src = src_type->isa_aryptr();
4952 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4953 if (top_src == NULL || top_src->klass() == NULL ||
4954 top_dest == NULL || top_dest->klass() == NULL) {
4955 // failed array check
4956 return false;
4957 }
4958
4959 // Figure out the size and type of the elements we will be copying.
4960 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4961 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4962 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
4963 return false;
4964 }
4965
4966 Node* src_start = array_element_address(src, src_offset, T_CHAR);
4967 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
4968 // 'src_start' points to src array + scaled offset
4969 // 'dst_start' points to dst array + scaled offset
4970
4971 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
4972 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
4973 enc = _gvn.transform(enc);
4974 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
4975 set_memory(res_mem, mtype);
4976 set_result(enc);
4977 clear_upper_avx();
4978
4979 return true;
4980 }
4981
4982 //-------------inline_multiplyToLen-----------------------------------
4983 bool LibraryCallKit::inline_multiplyToLen() {
4984 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
4985
4986 address stubAddr = StubRoutines::multiplyToLen();
4987 if (stubAddr == NULL) {
4988 return false; // Intrinsic's stub is not implemented on this platform
4989 }
4990 const char* stubName = "multiplyToLen";
4991
4992 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
4993
4994 // no receiver because it is a static method
4995 Node* x = argument(0);
4996 Node* xlen = argument(1);
4997 Node* y = argument(2);
4998 Node* ylen = argument(3);
4999 Node* z = argument(4);
5000
5001 x = must_be_not_null(x, true);
5002 y = must_be_not_null(y, true);
5003
5004 const Type* x_type = x->Value(&_gvn);
5005 const Type* y_type = y->Value(&_gvn);
5006 const TypeAryPtr* top_x = x_type->isa_aryptr();
5007 const TypeAryPtr* top_y = y_type->isa_aryptr();
5008 if (top_x == NULL || top_x->klass() == NULL ||
5009 top_y == NULL || top_y->klass() == NULL) {
5010 // failed array check
5011 return false;
5012 }
5013
5014 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5015 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5016 if (x_elem != T_INT || y_elem != T_INT) {
5017 return false;
5018 }
5019
5020 // Set the original stack and the reexecute bit for the interpreter to reexecute
5021 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5022 // on the return from z array allocation in runtime.
5023 { PreserveReexecuteState preexecs(this);
5024 jvms()->set_should_reexecute(true);
5025
5026 Node* x_start = array_element_address(x, intcon(0), x_elem);
5027 Node* y_start = array_element_address(y, intcon(0), y_elem);
5028 // 'x_start' points to x array + scaled xlen
5029 // 'y_start' points to y array + scaled ylen
5030
5031 // Allocate the result array
5032 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5033 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5034 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5035
5036 IdealKit ideal(this);
5037
5038 #define __ ideal.
5039 Node* one = __ ConI(1);
5040 Node* zero = __ ConI(0);
5041 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5042 __ set(need_alloc, zero);
5043 __ set(z_alloc, z);
5044 __ if_then(z, BoolTest::eq, null()); {
5045 __ increment (need_alloc, one);
5046 } __ else_(); {
5047 // Update graphKit memory and control from IdealKit.
5048 sync_kit(ideal);
5049 Node *cast = new CastPPNode(z, TypePtr::NOTNULL);
5050 cast->init_req(0, control());
5051 _gvn.set_type(cast, cast->bottom_type());
5052 C->record_for_igvn(cast);
5053
5054 Node* zlen_arg = load_array_length(cast);
5055 // Update IdealKit memory and control from graphKit.
5056 __ sync_kit(this);
5057 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5058 __ increment (need_alloc, one);
5059 } __ end_if();
5060 } __ end_if();
5061
5062 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5063 // Update graphKit memory and control from IdealKit.
5064 sync_kit(ideal);
5065 Node * narr = new_array(klass_node, zlen, 1);
5066 // Update IdealKit memory and control from graphKit.
5067 __ sync_kit(this);
5068 __ set(z_alloc, narr);
5069 } __ end_if();
5070
5071 sync_kit(ideal);
5072 z = __ value(z_alloc);
5073 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5074 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5075 // Final sync IdealKit and GraphKit.
5076 final_sync(ideal);
5077 #undef __
5078
5079 Node* z_start = array_element_address(z, intcon(0), T_INT);
5080
5081 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5082 OptoRuntime::multiplyToLen_Type(),
5083 stubAddr, stubName, TypePtr::BOTTOM,
5084 x_start, xlen, y_start, ylen, z_start, zlen);
5085 } // original reexecute is set back here
5086
5087 C->set_has_split_ifs(true); // Has chance for split-if optimization
5088 set_result(z);
5089 return true;
5090 }
5091
5092 //-------------inline_squareToLen------------------------------------
5093 bool LibraryCallKit::inline_squareToLen() {
5094 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5095
5096 address stubAddr = StubRoutines::squareToLen();
5097 if (stubAddr == NULL) {
5098 return false; // Intrinsic's stub is not implemented on this platform
5099 }
5100 const char* stubName = "squareToLen";
5101
5102 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5103
5104 Node* x = argument(0);
5105 Node* len = argument(1);
5106 Node* z = argument(2);
5107 Node* zlen = argument(3);
5108
5109 x = must_be_not_null(x, true);
5110 z = must_be_not_null(z, true);
5111
5112 const Type* x_type = x->Value(&_gvn);
5113 const Type* z_type = z->Value(&_gvn);
5114 const TypeAryPtr* top_x = x_type->isa_aryptr();
5115 const TypeAryPtr* top_z = z_type->isa_aryptr();
5116 if (top_x == NULL || top_x->klass() == NULL ||
5117 top_z == NULL || top_z->klass() == NULL) {
5118 // failed array check
5119 return false;
5120 }
5121
5122 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5123 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5124 if (x_elem != T_INT || z_elem != T_INT) {
5125 return false;
5126 }
5127
5128
5129 Node* x_start = array_element_address(x, intcon(0), x_elem);
5130 Node* z_start = array_element_address(z, intcon(0), z_elem);
5131
5132 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5133 OptoRuntime::squareToLen_Type(),
5134 stubAddr, stubName, TypePtr::BOTTOM,
5135 x_start, len, z_start, zlen);
5136
5137 set_result(z);
5138 return true;
5139 }
5140
5141 //-------------inline_mulAdd------------------------------------------
5142 bool LibraryCallKit::inline_mulAdd() {
5143 assert(UseMulAddIntrinsic, "not implemented on this platform");
5144
5145 address stubAddr = StubRoutines::mulAdd();
5146 if (stubAddr == NULL) {
5147 return false; // Intrinsic's stub is not implemented on this platform
5148 }
5149 const char* stubName = "mulAdd";
5150
5151 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5152
5153 Node* out = argument(0);
5154 Node* in = argument(1);
5155 Node* offset = argument(2);
5156 Node* len = argument(3);
5157 Node* k = argument(4);
5158
5159 out = must_be_not_null(out, true);
5160
5161 const Type* out_type = out->Value(&_gvn);
5162 const Type* in_type = in->Value(&_gvn);
5163 const TypeAryPtr* top_out = out_type->isa_aryptr();
5164 const TypeAryPtr* top_in = in_type->isa_aryptr();
5165 if (top_out == NULL || top_out->klass() == NULL ||
5166 top_in == NULL || top_in->klass() == NULL) {
5167 // failed array check
5168 return false;
5169 }
5170
5171 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5172 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5173 if (out_elem != T_INT || in_elem != T_INT) {
5174 return false;
5175 }
5176
5177 Node* outlen = load_array_length(out);
5178 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5179 Node* out_start = array_element_address(out, intcon(0), out_elem);
5180 Node* in_start = array_element_address(in, intcon(0), in_elem);
5181
5182 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5183 OptoRuntime::mulAdd_Type(),
5184 stubAddr, stubName, TypePtr::BOTTOM,
5185 out_start,in_start, new_offset, len, k);
5186 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5187 set_result(result);
5188 return true;
5189 }
5190
5191 //-------------inline_montgomeryMultiply-----------------------------------
5192 bool LibraryCallKit::inline_montgomeryMultiply() {
5193 address stubAddr = StubRoutines::montgomeryMultiply();
5194 if (stubAddr == NULL) {
5195 return false; // Intrinsic's stub is not implemented on this platform
5196 }
5197
5198 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5199 const char* stubName = "montgomery_multiply";
5200
5201 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5202
5203 Node* a = argument(0);
5204 Node* b = argument(1);
5205 Node* n = argument(2);
5206 Node* len = argument(3);
5207 Node* inv = argument(4);
5208 Node* m = argument(6);
5209
5210 const Type* a_type = a->Value(&_gvn);
5211 const TypeAryPtr* top_a = a_type->isa_aryptr();
5212 const Type* b_type = b->Value(&_gvn);
5213 const TypeAryPtr* top_b = b_type->isa_aryptr();
5214 const Type* n_type = a->Value(&_gvn);
5215 const TypeAryPtr* top_n = n_type->isa_aryptr();
5216 const Type* m_type = a->Value(&_gvn);
5217 const TypeAryPtr* top_m = m_type->isa_aryptr();
5218 if (top_a == NULL || top_a->klass() == NULL ||
5219 top_b == NULL || top_b->klass() == NULL ||
5220 top_n == NULL || top_n->klass() == NULL ||
5221 top_m == NULL || top_m->klass() == NULL) {
5222 // failed array check
5223 return false;
5224 }
5225
5226 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5227 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5228 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5229 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5230 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5231 return false;
5232 }
5233
5234 // Make the call
5235 {
5236 Node* a_start = array_element_address(a, intcon(0), a_elem);
5237 Node* b_start = array_element_address(b, intcon(0), b_elem);
5238 Node* n_start = array_element_address(n, intcon(0), n_elem);
5239 Node* m_start = array_element_address(m, intcon(0), m_elem);
5240
5241 Node* call = make_runtime_call(RC_LEAF,
5242 OptoRuntime::montgomeryMultiply_Type(),
5243 stubAddr, stubName, TypePtr::BOTTOM,
5244 a_start, b_start, n_start, len, inv, top(),
5245 m_start);
5246 set_result(m);
5247 }
5248
5249 return true;
5250 }
5251
5252 bool LibraryCallKit::inline_montgomerySquare() {
5253 address stubAddr = StubRoutines::montgomerySquare();
5254 if (stubAddr == NULL) {
5255 return false; // Intrinsic's stub is not implemented on this platform
5256 }
5257
5258 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5259 const char* stubName = "montgomery_square";
5260
5261 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5262
5263 Node* a = argument(0);
5264 Node* n = argument(1);
5265 Node* len = argument(2);
5266 Node* inv = argument(3);
5267 Node* m = argument(5);
5268
5269 const Type* a_type = a->Value(&_gvn);
5270 const TypeAryPtr* top_a = a_type->isa_aryptr();
5271 const Type* n_type = a->Value(&_gvn);
5272 const TypeAryPtr* top_n = n_type->isa_aryptr();
5273 const Type* m_type = a->Value(&_gvn);
5274 const TypeAryPtr* top_m = m_type->isa_aryptr();
5275 if (top_a == NULL || top_a->klass() == NULL ||
5276 top_n == NULL || top_n->klass() == NULL ||
5277 top_m == NULL || top_m->klass() == NULL) {
5278 // failed array check
5279 return false;
5280 }
5281
5282 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5283 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5284 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5285 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5286 return false;
5287 }
5288
5289 // Make the call
5290 {
5291 Node* a_start = array_element_address(a, intcon(0), a_elem);
5292 Node* n_start = array_element_address(n, intcon(0), n_elem);
5293 Node* m_start = array_element_address(m, intcon(0), m_elem);
5294
5295 Node* call = make_runtime_call(RC_LEAF,
5296 OptoRuntime::montgomerySquare_Type(),
5297 stubAddr, stubName, TypePtr::BOTTOM,
5298 a_start, n_start, len, inv, top(),
5299 m_start);
5300 set_result(m);
5301 }
5302
5303 return true;
5304 }
5305
5306 //-------------inline_vectorizedMismatch------------------------------
5307 bool LibraryCallKit::inline_vectorizedMismatch() {
5308 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5309
5310 address stubAddr = StubRoutines::vectorizedMismatch();
5311 if (stubAddr == NULL) {
5312 return false; // Intrinsic's stub is not implemented on this platform
5313 }
5314 const char* stubName = "vectorizedMismatch";
5315 int size_l = callee()->signature()->size();
5316 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5317
5318 Node* obja = argument(0);
5319 Node* aoffset = argument(1);
5320 Node* objb = argument(3);
5321 Node* boffset = argument(4);
5322 Node* length = argument(6);
5323 Node* scale = argument(7);
5324
5325 const Type* a_type = obja->Value(&_gvn);
5326 const Type* b_type = objb->Value(&_gvn);
5327 const TypeAryPtr* top_a = a_type->isa_aryptr();
5328 const TypeAryPtr* top_b = b_type->isa_aryptr();
5329 if (top_a == NULL || top_a->klass() == NULL ||
5330 top_b == NULL || top_b->klass() == NULL) {
5331 // failed array check
5332 return false;
5333 }
5334
5335 Node* call;
5336 jvms()->set_should_reexecute(true);
5337
5338 Node* obja_adr = make_unsafe_address(obja, aoffset);
5339 Node* objb_adr = make_unsafe_address(objb, boffset);
5340
5341 call = make_runtime_call(RC_LEAF,
5342 OptoRuntime::vectorizedMismatch_Type(),
5343 stubAddr, stubName, TypePtr::BOTTOM,
5344 obja_adr, objb_adr, length, scale);
5345
5346 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5347 set_result(result);
5348 return true;
5349 }
5350
5351 /**
5352 * Calculate CRC32 for byte.
5353 * int java.util.zip.CRC32.update(int crc, int b)
5354 */
5355 bool LibraryCallKit::inline_updateCRC32() {
5356 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5357 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5358 // no receiver since it is static method
5359 Node* crc = argument(0); // type: int
5360 Node* b = argument(1); // type: int
5361
5362 /*
5363 * int c = ~ crc;
5364 * b = timesXtoThe32[(b ^ c) & 0xFF];
5365 * b = b ^ (c >>> 8);
5366 * crc = ~b;
5367 */
5368
5369 Node* M1 = intcon(-1);
5370 crc = _gvn.transform(new XorINode(crc, M1));
5371 Node* result = _gvn.transform(new XorINode(crc, b));
5372 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5373
5374 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5375 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5376 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5377 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5378
5379 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5380 result = _gvn.transform(new XorINode(crc, result));
5381 result = _gvn.transform(new XorINode(result, M1));
5382 set_result(result);
5383 return true;
5384 }
5385
5386 /**
5387 * Calculate CRC32 for byte[] array.
5388 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5389 */
5390 bool LibraryCallKit::inline_updateBytesCRC32() {
5391 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5392 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5393 // no receiver since it is static method
5394 Node* crc = argument(0); // type: int
5395 Node* src = argument(1); // type: oop
5396 Node* offset = argument(2); // type: int
5397 Node* length = argument(3); // type: int
5398
5399 const Type* src_type = src->Value(&_gvn);
5400 const TypeAryPtr* top_src = src_type->isa_aryptr();
5401 if (top_src == NULL || top_src->klass() == NULL) {
5402 // failed array check
5403 return false;
5404 }
5405
5406 // Figure out the size and type of the elements we will be copying.
5407 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5408 if (src_elem != T_BYTE) {
5409 return false;
5410 }
5411
5412 // 'src_start' points to src array + scaled offset
5413 src = must_be_not_null(src, true);
5414 Node* src_start = array_element_address(src, offset, src_elem);
5415
5416 // We assume that range check is done by caller.
5417 // TODO: generate range check (offset+length < src.length) in debug VM.
5418
5419 // Call the stub.
5420 address stubAddr = StubRoutines::updateBytesCRC32();
5421 const char *stubName = "updateBytesCRC32";
5422
5423 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5424 stubAddr, stubName, TypePtr::BOTTOM,
5425 crc, src_start, length);
5426 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5427 set_result(result);
5428 return true;
5429 }
5430
5431 /**
5432 * Calculate CRC32 for ByteBuffer.
5433 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5434 */
5435 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5436 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5437 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5438 // no receiver since it is static method
5439 Node* crc = argument(0); // type: int
5440 Node* src = argument(1); // type: long
5441 Node* offset = argument(3); // type: int
5442 Node* length = argument(4); // type: int
5443
5444 src = ConvL2X(src); // adjust Java long to machine word
5445 Node* base = _gvn.transform(new CastX2PNode(src));
5446 offset = ConvI2X(offset);
5447
5448 // 'src_start' points to src array + scaled offset
5449 Node* src_start = basic_plus_adr(top(), base, offset);
5450
5451 // Call the stub.
5452 address stubAddr = StubRoutines::updateBytesCRC32();
5453 const char *stubName = "updateBytesCRC32";
5454
5455 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5456 stubAddr, stubName, TypePtr::BOTTOM,
5457 crc, src_start, length);
5458 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5459 set_result(result);
5460 return true;
5461 }
5462
5463 //------------------------------get_table_from_crc32c_class-----------------------
5464 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5465 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5466 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5467
5468 return table;
5469 }
5470
5471 //------------------------------inline_updateBytesCRC32C-----------------------
5472 //
5473 // Calculate CRC32C for byte[] array.
5474 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5475 //
5476 bool LibraryCallKit::inline_updateBytesCRC32C() {
5477 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5478 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5479 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5480 // no receiver since it is a static method
5481 Node* crc = argument(0); // type: int
5482 Node* src = argument(1); // type: oop
5483 Node* offset = argument(2); // type: int
5484 Node* end = argument(3); // type: int
5485
5486 Node* length = _gvn.transform(new SubINode(end, offset));
5487
5488 const Type* src_type = src->Value(&_gvn);
5489 const TypeAryPtr* top_src = src_type->isa_aryptr();
5490 if (top_src == NULL || top_src->klass() == NULL) {
5491 // failed array check
5492 return false;
5493 }
5494
5495 // Figure out the size and type of the elements we will be copying.
5496 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5497 if (src_elem != T_BYTE) {
5498 return false;
5499 }
5500
5501 // 'src_start' points to src array + scaled offset
5502 src = must_be_not_null(src, true);
5503 Node* src_start = array_element_address(src, offset, src_elem);
5504
5505 // static final int[] byteTable in class CRC32C
5506 Node* table = get_table_from_crc32c_class(callee()->holder());
5507 table = must_be_not_null(table, true);
5508 Node* table_start = array_element_address(table, intcon(0), T_INT);
5509
5510 // We assume that range check is done by caller.
5511 // TODO: generate range check (offset+length < src.length) in debug VM.
5512
5513 // Call the stub.
5514 address stubAddr = StubRoutines::updateBytesCRC32C();
5515 const char *stubName = "updateBytesCRC32C";
5516
5517 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5518 stubAddr, stubName, TypePtr::BOTTOM,
5519 crc, src_start, length, table_start);
5520 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5521 set_result(result);
5522 return true;
5523 }
5524
5525 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5526 //
5527 // Calculate CRC32C for DirectByteBuffer.
5528 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5529 //
5530 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5531 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5532 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5533 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5534 // no receiver since it is a static method
5535 Node* crc = argument(0); // type: int
5536 Node* src = argument(1); // type: long
5537 Node* offset = argument(3); // type: int
5538 Node* end = argument(4); // type: int
5539
5540 Node* length = _gvn.transform(new SubINode(end, offset));
5541
5542 src = ConvL2X(src); // adjust Java long to machine word
5543 Node* base = _gvn.transform(new CastX2PNode(src));
5544 offset = ConvI2X(offset);
5545
5546 // 'src_start' points to src array + scaled offset
5547 Node* src_start = basic_plus_adr(top(), base, offset);
5548
5549 // static final int[] byteTable in class CRC32C
5550 Node* table = get_table_from_crc32c_class(callee()->holder());
5551 table = must_be_not_null(table, true);
5552 Node* table_start = array_element_address(table, intcon(0), T_INT);
5553
5554 // Call the stub.
5555 address stubAddr = StubRoutines::updateBytesCRC32C();
5556 const char *stubName = "updateBytesCRC32C";
5557
5558 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5559 stubAddr, stubName, TypePtr::BOTTOM,
5560 crc, src_start, length, table_start);
5561 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5562 set_result(result);
5563 return true;
5564 }
5565
5566 //------------------------------inline_updateBytesAdler32----------------------
5567 //
5568 // Calculate Adler32 checksum for byte[] array.
5569 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5570 //
5571 bool LibraryCallKit::inline_updateBytesAdler32() {
5572 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5573 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5574 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5575 // no receiver since it is static method
5576 Node* crc = argument(0); // type: int
5577 Node* src = argument(1); // type: oop
5578 Node* offset = argument(2); // type: int
5579 Node* length = argument(3); // type: int
5580
5581 const Type* src_type = src->Value(&_gvn);
5582 const TypeAryPtr* top_src = src_type->isa_aryptr();
5583 if (top_src == NULL || top_src->klass() == NULL) {
5584 // failed array check
5585 return false;
5586 }
5587
5588 // Figure out the size and type of the elements we will be copying.
5589 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5590 if (src_elem != T_BYTE) {
5591 return false;
5592 }
5593
5594 // 'src_start' points to src array + scaled offset
5595 Node* src_start = array_element_address(src, offset, src_elem);
5596
5597 // We assume that range check is done by caller.
5598 // TODO: generate range check (offset+length < src.length) in debug VM.
5599
5600 // Call the stub.
5601 address stubAddr = StubRoutines::updateBytesAdler32();
5602 const char *stubName = "updateBytesAdler32";
5603
5604 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5605 stubAddr, stubName, TypePtr::BOTTOM,
5606 crc, src_start, length);
5607 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5608 set_result(result);
5609 return true;
5610 }
5611
5612 //------------------------------inline_updateByteBufferAdler32---------------
5613 //
5614 // Calculate Adler32 checksum for DirectByteBuffer.
5615 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5616 //
5617 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5618 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5619 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5620 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5621 // no receiver since it is static method
5622 Node* crc = argument(0); // type: int
5623 Node* src = argument(1); // type: long
5624 Node* offset = argument(3); // type: int
5625 Node* length = argument(4); // type: int
5626
5627 src = ConvL2X(src); // adjust Java long to machine word
5628 Node* base = _gvn.transform(new CastX2PNode(src));
5629 offset = ConvI2X(offset);
5630
5631 // 'src_start' points to src array + scaled offset
5632 Node* src_start = basic_plus_adr(top(), base, offset);
5633
5634 // Call the stub.
5635 address stubAddr = StubRoutines::updateBytesAdler32();
5636 const char *stubName = "updateBytesAdler32";
5637
5638 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5639 stubAddr, stubName, TypePtr::BOTTOM,
5640 crc, src_start, length);
5641
5642 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5643 set_result(result);
5644 return true;
5645 }
5646
5647 //----------------------------inline_reference_get----------------------------
5648 // public T java.lang.ref.Reference.get();
5649 bool LibraryCallKit::inline_reference_get() {
5650 const int referent_offset = java_lang_ref_Reference::referent_offset;
5651 guarantee(referent_offset > 0, "should have already been set");
5652
5653 // Get the argument:
5654 Node* reference_obj = null_check_receiver();
5655 if (stopped()) return true;
5656
5657 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr();
5658 assert(tinst != NULL, "obj is null");
5659 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5660 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass();
5661 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"),
5662 ciSymbol::make("Ljava/lang/Object;"),
5663 false);
5664 assert (field != NULL, "undefined field");
5665
5666 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5667 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5668
5669 ciInstanceKlass* klass = env()->Object_klass();
5670 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5671
5672 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF;
5673 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators);
5674 // Add memory barrier to prevent commoning reads from this field
5675 // across safepoint since GC can change its value.
5676 insert_mem_bar(Op_MemBarCPUOrder);
5677
5678 set_result(result);
5679 return true;
5680 }
5681
5682
5683 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5684 bool is_exact=true, bool is_static=false,
5685 ciInstanceKlass * fromKls=NULL) {
5686 if (fromKls == NULL) {
5687 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5688 assert(tinst != NULL, "obj is null");
5689 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5690 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5691 fromKls = tinst->klass()->as_instance_klass();
5692 } else {
5693 assert(is_static, "only for static field access");
5694 }
5695 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5696 ciSymbol::make(fieldTypeString),
5697 is_static);
5698
5699 assert (field != NULL, "undefined field");
5700 if (field == NULL) return (Node *) NULL;
5701
5702 if (is_static) {
5703 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5704 fromObj = makecon(tip);
5705 }
5706
5707 // Next code copied from Parse::do_get_xxx():
5708
5709 // Compute address and memory type.
5710 int offset = field->offset_in_bytes();
5711 bool is_vol = field->is_volatile();
5712 ciType* field_klass = field->type();
5713 assert(field_klass->is_loaded(), "should be loaded");
5714 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5715 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5716 BasicType bt = field->layout_type();
5717
5718 // Build the resultant type of the load
5719 const Type *type;
5720 if (bt == T_OBJECT) {
5721 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5722 } else {
5723 type = Type::get_const_basic_type(bt);
5724 }
5725
5726 DecoratorSet decorators = IN_HEAP;
5727
5728 if (is_vol) {
5729 decorators |= MO_SEQ_CST;
5730 }
5731
5732 return access_load_at(fromObj, adr, adr_type, type, bt, decorators);
5733 }
5734
5735 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5736 bool is_exact = true, bool is_static = false,
5737 ciInstanceKlass * fromKls = NULL) {
5738 if (fromKls == NULL) {
5739 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5740 assert(tinst != NULL, "obj is null");
5741 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5742 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5743 fromKls = tinst->klass()->as_instance_klass();
5744 }
5745 else {
5746 assert(is_static, "only for static field access");
5747 }
5748 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5749 ciSymbol::make(fieldTypeString),
5750 is_static);
5751
5752 assert(field != NULL, "undefined field");
5753 assert(!field->is_volatile(), "not defined for volatile fields");
5754
5755 if (is_static) {
5756 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5757 fromObj = makecon(tip);
5758 }
5759
5760 // Next code copied from Parse::do_get_xxx():
5761
5762 // Compute address and memory type.
5763 int offset = field->offset_in_bytes();
5764 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5765
5766 return adr;
5767 }
5768
5769 //------------------------------inline_aescrypt_Block-----------------------
5770 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5771 address stubAddr = NULL;
5772 const char *stubName;
5773 assert(UseAES, "need AES instruction support");
5774
5775 switch(id) {
5776 case vmIntrinsics::_aescrypt_encryptBlock:
5777 stubAddr = StubRoutines::aescrypt_encryptBlock();
5778 stubName = "aescrypt_encryptBlock";
5779 break;
5780 case vmIntrinsics::_aescrypt_decryptBlock:
5781 stubAddr = StubRoutines::aescrypt_decryptBlock();
5782 stubName = "aescrypt_decryptBlock";
5783 break;
5784 default:
5785 break;
5786 }
5787 if (stubAddr == NULL) return false;
5788
5789 Node* aescrypt_object = argument(0);
5790 Node* src = argument(1);
5791 Node* src_offset = argument(2);
5792 Node* dest = argument(3);
5793 Node* dest_offset = argument(4);
5794
5795 src = must_be_not_null(src, true);
5796 dest = must_be_not_null(dest, true);
5797
5798 // (1) src and dest are arrays.
5799 const Type* src_type = src->Value(&_gvn);
5800 const Type* dest_type = dest->Value(&_gvn);
5801 const TypeAryPtr* top_src = src_type->isa_aryptr();
5802 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5803 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5804
5805 // for the quick and dirty code we will skip all the checks.
5806 // we are just trying to get the call to be generated.
5807 Node* src_start = src;
5808 Node* dest_start = dest;
5809 if (src_offset != NULL || dest_offset != NULL) {
5810 assert(src_offset != NULL && dest_offset != NULL, "");
5811 src_start = array_element_address(src, src_offset, T_BYTE);
5812 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5813 }
5814
5815 // now need to get the start of its expanded key array
5816 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5817 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5818 if (k_start == NULL) return false;
5819
5820 if (Matcher::pass_original_key_for_aes()) {
5821 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5822 // compatibility issues between Java key expansion and SPARC crypto instructions
5823 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5824 if (original_k_start == NULL) return false;
5825
5826 // Call the stub.
5827 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5828 stubAddr, stubName, TypePtr::BOTTOM,
5829 src_start, dest_start, k_start, original_k_start);
5830 } else {
5831 // Call the stub.
5832 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5833 stubAddr, stubName, TypePtr::BOTTOM,
5834 src_start, dest_start, k_start);
5835 }
5836
5837 return true;
5838 }
5839
5840 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5841 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5842 address stubAddr = NULL;
5843 const char *stubName = NULL;
5844
5845 assert(UseAES, "need AES instruction support");
5846
5847 switch(id) {
5848 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5849 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5850 stubName = "cipherBlockChaining_encryptAESCrypt";
5851 break;
5852 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5853 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5854 stubName = "cipherBlockChaining_decryptAESCrypt";
5855 break;
5856 default:
5857 break;
5858 }
5859 if (stubAddr == NULL) return false;
5860
5861 Node* cipherBlockChaining_object = argument(0);
5862 Node* src = argument(1);
5863 Node* src_offset = argument(2);
5864 Node* len = argument(3);
5865 Node* dest = argument(4);
5866 Node* dest_offset = argument(5);
5867
5868 src = must_be_not_null(src, false);
5869 dest = must_be_not_null(dest, false);
5870
5871 // (1) src and dest are arrays.
5872 const Type* src_type = src->Value(&_gvn);
5873 const Type* dest_type = dest->Value(&_gvn);
5874 const TypeAryPtr* top_src = src_type->isa_aryptr();
5875 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5876 assert (top_src != NULL && top_src->klass() != NULL
5877 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5878
5879 // checks are the responsibility of the caller
5880 Node* src_start = src;
5881 Node* dest_start = dest;
5882 if (src_offset != NULL || dest_offset != NULL) {
5883 assert(src_offset != NULL && dest_offset != NULL, "");
5884 src_start = array_element_address(src, src_offset, T_BYTE);
5885 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5886 }
5887
5888 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5889 // (because of the predicated logic executed earlier).
5890 // so we cast it here safely.
5891 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5892
5893 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5894 if (embeddedCipherObj == NULL) return false;
5895
5896 // cast it to what we know it will be at runtime
5897 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5898 assert(tinst != NULL, "CBC obj is null");
5899 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5900 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5901 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5902
5903 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5904 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5905 const TypeOopPtr* xtype = aklass->as_instance_type();
5906 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5907 aescrypt_object = _gvn.transform(aescrypt_object);
5908
5909 // we need to get the start of the aescrypt_object's expanded key array
5910 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5911 if (k_start == NULL) return false;
5912
5913 // similarly, get the start address of the r vector
5914 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5915 if (objRvec == NULL) return false;
5916 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5917
5918 Node* cbcCrypt;
5919 if (Matcher::pass_original_key_for_aes()) {
5920 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5921 // compatibility issues between Java key expansion and SPARC crypto instructions
5922 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5923 if (original_k_start == NULL) return false;
5924
5925 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5926 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5927 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5928 stubAddr, stubName, TypePtr::BOTTOM,
5929 src_start, dest_start, k_start, r_start, len, original_k_start);
5930 } else {
5931 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5932 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5933 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5934 stubAddr, stubName, TypePtr::BOTTOM,
5935 src_start, dest_start, k_start, r_start, len);
5936 }
5937
5938 // return cipher length (int)
5939 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5940 set_result(retvalue);
5941 return true;
5942 }
5943
5944 //------------------------------inline_counterMode_AESCrypt-----------------------
5945 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
5946 assert(UseAES, "need AES instruction support");
5947 if (!UseAESCTRIntrinsics) return false;
5948
5949 address stubAddr = NULL;
5950 const char *stubName = NULL;
5951 if (id == vmIntrinsics::_counterMode_AESCrypt) {
5952 stubAddr = StubRoutines::counterMode_AESCrypt();
5953 stubName = "counterMode_AESCrypt";
5954 }
5955 if (stubAddr == NULL) return false;
5956
5957 Node* counterMode_object = argument(0);
5958 Node* src = argument(1);
5959 Node* src_offset = argument(2);
5960 Node* len = argument(3);
5961 Node* dest = argument(4);
5962 Node* dest_offset = argument(5);
5963
5964 // (1) src and dest are arrays.
5965 const Type* src_type = src->Value(&_gvn);
5966 const Type* dest_type = dest->Value(&_gvn);
5967 const TypeAryPtr* top_src = src_type->isa_aryptr();
5968 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5969 assert(top_src != NULL && top_src->klass() != NULL &&
5970 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5971
5972 // checks are the responsibility of the caller
5973 Node* src_start = src;
5974 Node* dest_start = dest;
5975 if (src_offset != NULL || dest_offset != NULL) {
5976 assert(src_offset != NULL && dest_offset != NULL, "");
5977 src_start = array_element_address(src, src_offset, T_BYTE);
5978 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5979 }
5980
5981 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5982 // (because of the predicated logic executed earlier).
5983 // so we cast it here safely.
5984 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5985 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5986 if (embeddedCipherObj == NULL) return false;
5987 // cast it to what we know it will be at runtime
5988 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
5989 assert(tinst != NULL, "CTR obj is null");
5990 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
5991 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5992 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5993 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5994 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5995 const TypeOopPtr* xtype = aklass->as_instance_type();
5996 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5997 aescrypt_object = _gvn.transform(aescrypt_object);
5998 // we need to get the start of the aescrypt_object's expanded key array
5999 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6000 if (k_start == NULL) return false;
6001 // similarly, get the start address of the r vector
6002 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6003 if (obj_counter == NULL) return false;
6004 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6005
6006 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6007 if (saved_encCounter == NULL) return false;
6008 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6009 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6010
6011 Node* ctrCrypt;
6012 if (Matcher::pass_original_key_for_aes()) {
6013 // no SPARC version for AES/CTR intrinsics now.
6014 return false;
6015 }
6016 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6017 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6018 OptoRuntime::counterMode_aescrypt_Type(),
6019 stubAddr, stubName, TypePtr::BOTTOM,
6020 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6021
6022 // return cipher length (int)
6023 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6024 set_result(retvalue);
6025 return true;
6026 }
6027
6028 //------------------------------get_key_start_from_aescrypt_object-----------------------
6029 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6030 #if defined(PPC64) || defined(S390)
6031 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6032 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6033 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6034 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6035 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6036 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6037 if (objSessionK == NULL) {
6038 return (Node *) NULL;
6039 }
6040 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6041 #else
6042 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6043 #endif // PPC64
6044 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6045 if (objAESCryptKey == NULL) return (Node *) NULL;
6046
6047 // now have the array, need to get the start address of the K array
6048 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6049 return k_start;
6050 }
6051
6052 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6053 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6054 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6055 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6056 if (objAESCryptKey == NULL) return (Node *) NULL;
6057
6058 // now have the array, need to get the start address of the lastKey array
6059 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6060 return original_k_start;
6061 }
6062
6063 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6064 // Return node representing slow path of predicate check.
6065 // the pseudo code we want to emulate with this predicate is:
6066 // for encryption:
6067 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6068 // for decryption:
6069 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6070 // note cipher==plain is more conservative than the original java code but that's OK
6071 //
6072 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6073 // The receiver was checked for NULL already.
6074 Node* objCBC = argument(0);
6075
6076 Node* src = argument(1);
6077 Node* dest = argument(4);
6078
6079 // Load embeddedCipher field of CipherBlockChaining object.
6080 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6081
6082 // get AESCrypt klass for instanceOf check
6083 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6084 // will have same classloader as CipherBlockChaining object
6085 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6086 assert(tinst != NULL, "CBCobj is null");
6087 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6088
6089 // we want to do an instanceof comparison against the AESCrypt class
6090 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6091 if (!klass_AESCrypt->is_loaded()) {
6092 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6093 Node* ctrl = control();
6094 set_control(top()); // no regular fast path
6095 return ctrl;
6096 }
6097
6098 src = must_be_not_null(src, true);
6099 dest = must_be_not_null(dest, true);
6100
6101 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6102
6103 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6104 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6105 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6106
6107 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6108
6109 // for encryption, we are done
6110 if (!decrypting)
6111 return instof_false; // even if it is NULL
6112
6113 // for decryption, we need to add a further check to avoid
6114 // taking the intrinsic path when cipher and plain are the same
6115 // see the original java code for why.
6116 RegionNode* region = new RegionNode(3);
6117 region->init_req(1, instof_false);
6118
6119 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6120 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6121 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6122 region->init_req(2, src_dest_conjoint);
6123
6124 record_for_igvn(region);
6125 return _gvn.transform(region);
6126 }
6127
6128 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6129 // Return node representing slow path of predicate check.
6130 // the pseudo code we want to emulate with this predicate is:
6131 // for encryption:
6132 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6133 // for decryption:
6134 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6135 // note cipher==plain is more conservative than the original java code but that's OK
6136 //
6137
6138 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6139 // The receiver was checked for NULL already.
6140 Node* objCTR = argument(0);
6141
6142 // Load embeddedCipher field of CipherBlockChaining object.
6143 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6144
6145 // get AESCrypt klass for instanceOf check
6146 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6147 // will have same classloader as CipherBlockChaining object
6148 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6149 assert(tinst != NULL, "CTRobj is null");
6150 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6151
6152 // we want to do an instanceof comparison against the AESCrypt class
6153 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6154 if (!klass_AESCrypt->is_loaded()) {
6155 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6156 Node* ctrl = control();
6157 set_control(top()); // no regular fast path
6158 return ctrl;
6159 }
6160
6161 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6162 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6163 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6164 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6165 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6166
6167 return instof_false; // even if it is NULL
6168 }
6169
6170 //------------------------------inline_ghash_processBlocks
6171 bool LibraryCallKit::inline_ghash_processBlocks() {
6172 address stubAddr;
6173 const char *stubName;
6174 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6175
6176 stubAddr = StubRoutines::ghash_processBlocks();
6177 stubName = "ghash_processBlocks";
6178
6179 Node* data = argument(0);
6180 Node* offset = argument(1);
6181 Node* len = argument(2);
6182 Node* state = argument(3);
6183 Node* subkeyH = argument(4);
6184
6185 state = must_be_not_null(state, true);
6186 subkeyH = must_be_not_null(subkeyH, true);
6187 data = must_be_not_null(data, true);
6188
6189 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6190 assert(state_start, "state is NULL");
6191 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6192 assert(subkeyH_start, "subkeyH is NULL");
6193 Node* data_start = array_element_address(data, offset, T_BYTE);
6194 assert(data_start, "data is NULL");
6195
6196 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6197 OptoRuntime::ghash_processBlocks_Type(),
6198 stubAddr, stubName, TypePtr::BOTTOM,
6199 state_start, subkeyH_start, data_start, len);
6200 return true;
6201 }
6202
6203 bool LibraryCallKit::inline_base64_encodeBlock() {
6204 address stubAddr;
6205 const char *stubName;
6206 assert(UseBASE64Intrinsics, "need Base64 intrinsics support");
6207 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters");
6208 stubAddr = StubRoutines::base64_encodeBlock();
6209 stubName = "encodeBlock";
6210
6211 if (!stubAddr) return false;
6212 Node* base64obj = argument(0);
6213 Node* src = argument(1);
6214 Node* offset = argument(2);
6215 Node* len = argument(3);
6216 Node* dest = argument(4);
6217 Node* dp = argument(5);
6218 Node* isURL = argument(6);
6219
6220 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
6221 assert(src_start, "source array is NULL");
6222 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE);
6223 assert(dest_start, "destination array is NULL");
6224
6225 Node* base64 = make_runtime_call(RC_LEAF,
6226 OptoRuntime::base64_encodeBlock_Type(),
6227 stubAddr, stubName, TypePtr::BOTTOM,
6228 src_start, offset, len, dest_start, dp, isURL);
6229 return true;
6230 }
6231
6232 //------------------------------inline_sha_implCompress-----------------------
6233 //
6234 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6235 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6236 //
6237 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6238 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6239 //
6240 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6241 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6242 //
6243 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6244 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6245
6246 Node* sha_obj = argument(0);
6247 Node* src = argument(1); // type oop
6248 Node* ofs = argument(2); // type int
6249
6250 const Type* src_type = src->Value(&_gvn);
6251 const TypeAryPtr* top_src = src_type->isa_aryptr();
6252 if (top_src == NULL || top_src->klass() == NULL) {
6253 // failed array check
6254 return false;
6255 }
6256 // Figure out the size and type of the elements we will be copying.
6257 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6258 if (src_elem != T_BYTE) {
6259 return false;
6260 }
6261 // 'src_start' points to src array + offset
6262 src = must_be_not_null(src, true);
6263 Node* src_start = array_element_address(src, ofs, src_elem);
6264 Node* state = NULL;
6265 address stubAddr;
6266 const char *stubName;
6267
6268 switch(id) {
6269 case vmIntrinsics::_sha_implCompress:
6270 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6271 state = get_state_from_sha_object(sha_obj);
6272 stubAddr = StubRoutines::sha1_implCompress();
6273 stubName = "sha1_implCompress";
6274 break;
6275 case vmIntrinsics::_sha2_implCompress:
6276 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6277 state = get_state_from_sha_object(sha_obj);
6278 stubAddr = StubRoutines::sha256_implCompress();
6279 stubName = "sha256_implCompress";
6280 break;
6281 case vmIntrinsics::_sha5_implCompress:
6282 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6283 state = get_state_from_sha5_object(sha_obj);
6284 stubAddr = StubRoutines::sha512_implCompress();
6285 stubName = "sha512_implCompress";
6286 break;
6287 default:
6288 fatal_unexpected_iid(id);
6289 return false;
6290 }
6291 if (state == NULL) return false;
6292
6293 // Call the stub.
6294 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6295 stubAddr, stubName, TypePtr::BOTTOM,
6296 src_start, state);
6297
6298 return true;
6299 }
6300
6301 //------------------------------inline_digestBase_implCompressMB-----------------------
6302 //
6303 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6304 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6305 //
6306 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6307 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6308 "need SHA1/SHA256/SHA512 instruction support");
6309 assert((uint)predicate < 3, "sanity");
6310 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6311
6312 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6313 Node* src = argument(1); // byte[] array
6314 Node* ofs = argument(2); // type int
6315 Node* limit = argument(3); // type int
6316
6317 const Type* src_type = src->Value(&_gvn);
6318 const TypeAryPtr* top_src = src_type->isa_aryptr();
6319 if (top_src == NULL || top_src->klass() == NULL) {
6320 // failed array check
6321 return false;
6322 }
6323 // Figure out the size and type of the elements we will be copying.
6324 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6325 if (src_elem != T_BYTE) {
6326 return false;
6327 }
6328 // 'src_start' points to src array + offset
6329 src = must_be_not_null(src, false);
6330 Node* src_start = array_element_address(src, ofs, src_elem);
6331
6332 const char* klass_SHA_name = NULL;
6333 const char* stub_name = NULL;
6334 address stub_addr = NULL;
6335 bool long_state = false;
6336
6337 switch (predicate) {
6338 case 0:
6339 if (UseSHA1Intrinsics) {
6340 klass_SHA_name = "sun/security/provider/SHA";
6341 stub_name = "sha1_implCompressMB";
6342 stub_addr = StubRoutines::sha1_implCompressMB();
6343 }
6344 break;
6345 case 1:
6346 if (UseSHA256Intrinsics) {
6347 klass_SHA_name = "sun/security/provider/SHA2";
6348 stub_name = "sha256_implCompressMB";
6349 stub_addr = StubRoutines::sha256_implCompressMB();
6350 }
6351 break;
6352 case 2:
6353 if (UseSHA512Intrinsics) {
6354 klass_SHA_name = "sun/security/provider/SHA5";
6355 stub_name = "sha512_implCompressMB";
6356 stub_addr = StubRoutines::sha512_implCompressMB();
6357 long_state = true;
6358 }
6359 break;
6360 default:
6361 fatal("unknown SHA intrinsic predicate: %d", predicate);
6362 }
6363 if (klass_SHA_name != NULL) {
6364 // get DigestBase klass to lookup for SHA klass
6365 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6366 assert(tinst != NULL, "digestBase_obj is not instance???");
6367 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6368
6369 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6370 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6371 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6372 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6373 }
6374 return false;
6375 }
6376 //------------------------------inline_sha_implCompressMB-----------------------
6377 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6378 bool long_state, address stubAddr, const char *stubName,
6379 Node* src_start, Node* ofs, Node* limit) {
6380 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6381 const TypeOopPtr* xtype = aklass->as_instance_type();
6382 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6383 sha_obj = _gvn.transform(sha_obj);
6384
6385 Node* state;
6386 if (long_state) {
6387 state = get_state_from_sha5_object(sha_obj);
6388 } else {
6389 state = get_state_from_sha_object(sha_obj);
6390 }
6391 if (state == NULL) return false;
6392
6393 // Call the stub.
6394 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6395 OptoRuntime::digestBase_implCompressMB_Type(),
6396 stubAddr, stubName, TypePtr::BOTTOM,
6397 src_start, state, ofs, limit);
6398 // return ofs (int)
6399 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6400 set_result(result);
6401
6402 return true;
6403 }
6404
6405 enum VectorApiObjectType {
6406 VECAPI_VECTOR,
6407 VECAPI_MASK,
6408 VECAPI_SPECIES,
6409 VECAPI_SHUFFLE,
6410 };
6411
6412 static ciKlass* get_exact_klass_for_vector_box(ciKlass* ctx, BasicType bt, int num_elem, VectorApiObjectType obj_type) {
6413 // FIXME: use VM symbol table instead
6414 ResourceMark rm;
6415 stringStream ss;
6416 ss.print_raw("jdk/incubator/vector/");
6417 switch (bt) {
6418 case T_BYTE: ss.print_raw("Byte"); break;
6419 case T_SHORT: ss.print_raw("Short"); break;
6420 case T_INT: ss.print_raw("Int"); break;
6421 case T_LONG: ss.print_raw("Long"); break;
6422 case T_FLOAT: ss.print_raw("Float"); break;
6423 case T_DOUBLE: ss.print_raw("Double"); break;
6424 default: fatal("unknown element type: %d", bt);
6425 }
6426 int bits = num_elem * BitsPerByte * type2aelembytes(bt);
6427 switch (bits) {
6428 case 64: // fall through
6429 case 128: // fall through
6430 case 256: // fall through
6431 case 512: ss.print("%d", bits); break;
6432 default: fatal("unknown vector size: %d", bits);
6433 }
6434 ss.print_raw("Vector");
6435 if (obj_type != VECAPI_VECTOR) {
6436 ss.print_raw("$");
6437 switch (bt) {
6438 case T_BYTE: ss.print_raw("Byte"); break;
6439 case T_SHORT: ss.print_raw("Short"); break;
6440 case T_INT: ss.print_raw("Int"); break;
6441 case T_LONG: ss.print_raw("Long"); break;
6442 case T_FLOAT: ss.print_raw("Float"); break;
6443 case T_DOUBLE: ss.print_raw("Double"); break;
6444 default: fatal("unknown element type: %d", bt);
6445 }
6446 switch (bits) {
6447 case 64: // fall through
6448 case 128: // fall through
6449 case 256: // fall through
6450 case 512: ss.print("%d", bits); break;
6451 default: fatal("unknown vector size: %d", bits);
6452 }
6453 switch (obj_type) {
6454 case VECAPI_MASK: ss.print_raw("Mask"); break;
6455 case VECAPI_SPECIES: ss.print_raw("Species"); break;
6456 case VECAPI_SHUFFLE: ss.print_raw("Shuffle"); break;
6457 default: fatal("unknown vector object type");
6458 }
6459 }
6460
6461 ciSymbol* vector_klass_name = ciSymbol::make(ss.as_string());
6462 ciKlass* vector_klass = ctx->find_klass(vector_klass_name);
6463 return vector_klass;
6464 }
6465
6466 enum VectorMaskUseType {
6467 VecMaskUseLoad,
6468 VecMaskUseStore,
6469 VecMaskUseAll,
6470 VecMaskNotUsed
6471 };
6472
6473 static bool arch_supports_vector(int op, int num_elem, BasicType type, VectorMaskUseType mask_use_type, int op_arity = 0) {
6474 // Check that the operation is valid.
6475 if (op <= 0) {
6476 #ifndef PRODUCT
6477 if (DebugVectorApi) {
6478 tty->print_cr("Rejected intrinsification because no valid vector op could be extracted");
6479 }
6480 #endif
6481 return false;
6482 }
6483
6484 // Check that architecture supports this op-size-type combination.
6485 if (!Matcher::match_rule_supported_vector(op, num_elem, type, op_arity)) {
6486 #ifndef PRODUCT
6487 if (DebugVectorApi) {
6488 tty->print_cr("Rejected vector op (%s,%s,%d) because architecture does not support it",
6489 NodeClassNames[op], type2name(type), num_elem);
6490 }
6491 #endif
6492 return false;
6493 } else {
6494 assert(Matcher::match_rule_supported(op), "must be supported");
6495 }
6496
6497 // Check whether mask unboxing is supported.
6498 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseLoad) {
6499 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type)) {
6500 #ifndef PRODUCT
6501 if (DebugVectorApi) {
6502 tty->print_cr("Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
6503 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
6504 }
6505 #endif
6506 return false;
6507 }
6508 }
6509
6510 // Check whether mask boxing is supported.
6511 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseStore) {
6512 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type)) {
6513 #ifndef PRODUCT
6514 if (DebugVectorApi) {
6515 tty->print_cr("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
6516 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
6517 }
6518 #endif
6519 return false;
6520 }
6521 }
6522
6523 return true;
6524 }
6525
6526 // Should be aligned with constants in jdk.incubator.vector.VectorIntrinsics.
6527 enum {
6528 // Unary
6529 OP_ABS = 0,
6530 OP_NEG = 1,
6531 OP_SQRT = 2,
6532 OP_NOT = 3,
6533 // Binary
6534 OP_ADD = 4,
6535 OP_SUB = 5,
6536 OP_MUL = 6,
6537 OP_DIV = 7,
6538 OP_MIN = 8,
6539 OP_MAX = 9,
6540 OP_AND = 10,
6541 OP_OR = 11,
6542 OP_XOR = 12,
6543 // Ternary
6544 OP_FMA = 13,
6545 // Broadcast int
6546 OP_LSHIFT = 14,
6547 OP_RSHIFT = 15,
6548 OP_URSHIFT = 16,
6549 // Vector Math Library
6550 OP_TAN = 101,
6551 OP_SVML_START = OP_TAN,
6552 OP_TANH = 102,
6553 OP_SIN = 103,
6554 OP_SINH = 104,
6555 OP_COS = 105,
6556 OP_COSH = 106,
6557 OP_ASIN = 107,
6558 OP_ACOS = 108,
6559 OP_ATAN = 109,
6560 OP_ATAN2 = 110,
6561 OP_CBRT = 111,
6562 OP_LOG = 112,
6563 OP_LOG10 = 113,
6564 OP_LOG1P = 114,
6565 OP_POW = 115,
6566 OP_EXP = 116,
6567 OP_EXPM1 = 117,
6568 OP_HYPOT = 118,
6569 OP_SVML_END = OP_HYPOT,
6570 };
6571
6572 static int get_opc(jint op, BasicType bt) {
6573 switch (op) {
6574 case OP_ADD: {
6575 switch (bt) {
6576 case T_BYTE: // fall-through
6577 case T_SHORT: // fall-through
6578 case T_INT: return Op_AddI;
6579 case T_LONG: return Op_AddL;
6580 case T_FLOAT: return Op_AddF;
6581 case T_DOUBLE: return Op_AddD;
6582 default: fatal("ADD: %s", type2name(bt));
6583 }
6584 break;
6585 }
6586 case OP_SUB: {
6587 switch (bt) {
6588 case T_BYTE: // fall-through
6589 case T_SHORT: // fall-through
6590 case T_INT: return Op_SubI;
6591 case T_LONG: return Op_SubL;
6592 case T_FLOAT: return Op_SubF;
6593 case T_DOUBLE: return Op_SubD;
6594 default: fatal("SUB: %s", type2name(bt));
6595 }
6596 break;
6597 }
6598 case OP_MUL: {
6599 switch (bt) {
6600 case T_BYTE: // fall-through
6601 case T_SHORT: // fall-through
6602 case T_INT: return Op_MulI;
6603 case T_LONG: return Op_MulL;
6604 case T_FLOAT: return Op_MulF;
6605 case T_DOUBLE: return Op_MulD;
6606 default: fatal("MUL: %s", type2name(bt));
6607 }
6608 break;
6609 }
6610 case OP_DIV: {
6611 switch (bt) {
6612 case T_BYTE: // fall-through
6613 case T_SHORT: // fall-through
6614 case T_INT: return Op_DivI;
6615 case T_LONG: return Op_DivL;
6616 case T_FLOAT: return Op_DivF;
6617 case T_DOUBLE: return Op_DivD;
6618 default: fatal("DIV: %s", type2name(bt));
6619 }
6620 break;
6621 }
6622 case OP_MIN: {
6623 switch (bt) {
6624 case T_BYTE:
6625 case T_SHORT:
6626 case T_INT: return Op_MinI;
6627 case T_LONG: return Op_MinL;
6628 case T_FLOAT: return Op_MinF;
6629 case T_DOUBLE: return Op_MinD;
6630 default: fatal("MIN: %s", type2name(bt));
6631 }
6632 break;
6633 }
6634 case OP_MAX: {
6635 switch (bt) {
6636 case T_BYTE:
6637 case T_SHORT:
6638 case T_INT: return Op_MaxI;
6639 case T_LONG: return Op_MaxL;
6640 case T_FLOAT: return Op_MaxF;
6641 case T_DOUBLE: return Op_MaxD;
6642 default: fatal("MAX: %s", type2name(bt));
6643 }
6644 break;
6645 }
6646 case OP_ABS: {
6647 switch (bt) {
6648 case T_BYTE: // fall-through
6649 case T_SHORT: // fall-through
6650 case T_LONG: // fall-through
6651 case T_INT: return Op_AbsI;
6652 case T_FLOAT: return Op_AbsF;
6653 case T_DOUBLE: return Op_AbsD;
6654 default: fatal("ABS: %s", type2name(bt));
6655 }
6656 break;
6657 }
6658 case OP_NEG: {
6659 switch (bt) {
6660 case T_BYTE: // fall-through
6661 case T_SHORT: // fall-through
6662 case T_INT: return Op_NegI;
6663 case T_FLOAT: return Op_NegF;
6664 case T_DOUBLE: return Op_NegD;
6665 default: fatal("NEG: %s", type2name(bt));
6666 }
6667 break;
6668 }
6669 case OP_AND: {
6670 switch (bt) {
6671 case T_BYTE: // fall-through
6672 case T_SHORT: // fall-through
6673 case T_INT: return Op_AndI;
6674 case T_LONG: return Op_AndL;
6675 default: fatal("AND: %s", type2name(bt));
6676 }
6677 break;
6678 }
6679 case OP_OR: {
6680 switch (bt) {
6681 case T_BYTE: // fall-through
6682 case T_SHORT: // fall-through
6683 case T_INT: return Op_OrI;
6684 case T_LONG: return Op_OrL;
6685 default: fatal("OR: %s", type2name(bt));
6686 }
6687 break;
6688 }
6689 case OP_XOR: {
6690 switch (bt) {
6691 case T_BYTE: // fall-through
6692 case T_SHORT: // fall-through
6693 case T_INT: return Op_XorI;
6694 case T_LONG: return Op_XorL;
6695 default: fatal("XOR: %s", type2name(bt));
6696 }
6697 break;
6698 }
6699 case OP_SQRT: {
6700 switch (bt) {
6701 case T_FLOAT: return Op_SqrtF;
6702 case T_DOUBLE: return Op_SqrtD;
6703 default: fatal("SQRT: %s", type2name(bt));
6704 }
6705 break;
6706 }
6707 case OP_NOT: {
6708 switch (bt) {
6709 case T_BYTE: // fall-through
6710 case T_SHORT: // fall-through
6711 case T_INT: // fall-through
6712 case T_LONG: return Op_Not;
6713 default: fatal("NOT: %s", type2name(bt));
6714 }
6715 break;
6716 }
6717 case OP_FMA: {
6718 switch (bt) {
6719 case T_FLOAT: return Op_FmaF;
6720 case T_DOUBLE: return Op_FmaD;
6721 default: fatal("FMA: %s", type2name(bt));
6722 }
6723 break;
6724 }
6725 case OP_LSHIFT: {
6726 switch (bt) {
6727 case T_BYTE: // fall-through
6728 case T_SHORT: // fall-through
6729 case T_INT: return Op_LShiftI;
6730 case T_LONG: return Op_LShiftL;
6731 default: fatal("LSHIFT: %s", type2name(bt));
6732 }
6733 break;
6734 }
6735 case OP_RSHIFT: {
6736 switch (bt) {
6737 case T_BYTE: // fall-through
6738 case T_SHORT: // fall-through
6739 case T_INT: return Op_RShiftI;
6740 case T_LONG: return Op_RShiftL;
6741 default: fatal("RSHIFT: %s", type2name(bt));
6742 }
6743 break;
6744 }
6745 case OP_URSHIFT: {
6746 switch (bt) {
6747 case T_BYTE: // fall-through
6748 case T_SHORT: // fall-through
6749 case T_INT: return Op_URShiftI;
6750 case T_LONG: return Op_URShiftL;
6751 default: fatal("URSHIFT: %s", type2name(bt));
6752 }
6753 break;
6754 }
6755 case OP_TAN:
6756 case OP_TANH:
6757 case OP_SIN:
6758 case OP_SINH:
6759 case OP_COS:
6760 case OP_COSH:
6761 case OP_ASIN:
6762 case OP_ACOS:
6763 case OP_ATAN:
6764 case OP_ATAN2:
6765 case OP_CBRT:
6766 case OP_LOG:
6767 case OP_LOG10:
6768 case OP_LOG1P:
6769 case OP_POW:
6770 case OP_EXP:
6771 case OP_EXPM1:
6772 case OP_HYPOT:
6773 return Op_CallLeafVector;
6774 default: fatal("unknown op: %d", op);
6775 }
6776 return 0; // Unimplemented
6777 }
6778
6779 Node* LibraryCallKit::box_vector(Node* vector, const TypeInstPtr* vbox_type,
6780 BasicType elem_bt, int num_elem) {
6781
6782 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
6783
6784 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
6785 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
6786 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true);
6787 set_i_o(_gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
6788 set_all_memory(_gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
6789 Node* ret = _gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
6790
6791 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vec_type);
6792 return _gvn.transform(vbox);
6793 }
6794
6795 Node* LibraryCallKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) {
6796 const TypeInstPtr* vbox_type_v = gvn().type(v)->is_instptr();
6797 if (vbox_type->klass() != vbox_type_v->klass()) {
6798 return NULL; // arguments don't agree on vector shapes
6799 }
6800 if (vbox_type_v->maybe_null()) {
6801 return NULL; // no nulls are allowed
6802 }
6803 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
6804 Node* unbox = gvn().transform(new VectorUnboxNode(C, vec_type, v, merged_memory()));
6805 return unbox;
6806 }
6807
6808 void LibraryCallKit::set_vector_result(Node* result, bool set_res) {
6809 if (DebugVectorApi) {
6810 #ifndef PRODUCT
6811 tty->print("============ ");
6812 callee()->print();
6813 tty->print_cr(" ============");
6814 result->dump(5);
6815 tty->print_cr("----------------------------------------------------");
6816 #endif
6817 }
6818 if (set_res) {
6819 set_result(result);
6820 }
6821 }
6822
6823 bool LibraryCallKit::inline_vector_nary_operation(int n) {
6824 const TypeInt* opr = gvn().type(argument(0))->is_int();
6825 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
6826 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
6827 const TypeInt* vlen = gvn().type(argument(3))->is_int();
6828
6829 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
6830 return false; // not enough info for intrinsification
6831 }
6832 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
6833 if (!elem_type->is_primitive_type()) {
6834 return false; // should be primitive type
6835 }
6836 BasicType elem_bt = elem_type->basic_type();
6837 int num_elem = vlen->get_con();
6838 int opc = get_opc(opr->get_con(), elem_bt);
6839 int sopc = opc != Op_CallLeafVector ? VectorNode::opcode(opc, elem_bt) : Op_CallLeafVector;
6840 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
6841 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
6842
6843 if (opc == Op_CallLeafVector && !Matcher::supports_vector_calling_convention()) {
6844 return false;
6845 }
6846
6847 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
6848 if (!arch_supports_vector(sopc, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseAll : VecMaskNotUsed, n)) {
6849 return false; // not supported
6850 }
6851
6852 Node* opd1 = NULL; Node* opd2 = NULL; Node* opd3 = NULL;
6853 switch (n) {
6854 case 3: {
6855 opd3 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
6856 if (opd3 == NULL) {
6857 return false;
6858 }
6859 // fall-through
6860 }
6861 case 2: {
6862 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
6863 if (opd2 == NULL) {
6864 return false;
6865 }
6866 // fall-through
6867 }
6868 case 1: {
6869 opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
6870 if (opd1 == NULL) {
6871 return false;
6872 }
6873 break;
6874 }
6875 default: fatal("unsupported arity: %d", n);
6876 }
6877
6878 Node* operation = NULL;
6879 if (sopc == Op_CallLeafVector) {
6880 operation = gen_call_to_svml(opr->get_con(), elem_bt, num_elem, opd1, opd2);
6881 if (operation == NULL) {
6882 return false;
6883 }
6884 } else {
6885 switch (n) {
6886 case 1:
6887 case 2: {
6888 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt));
6889 break;
6890 }
6891 case 3: {
6892 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, opd3, num_elem, elem_bt));
6893 break;
6894 }
6895 default: fatal("unsupported arity: %d", n);
6896 }
6897 }
6898 // Wrap it up in VectorBox to keep object type information.
6899 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
6900 set_vector_result(operation);
6901
6902 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
6903 return true;
6904 }
6905
6906 // <V extends Vector<?,?>>
6907 // V broadcastCoerced(Class<?> vectorClass, Class<?> elementType, int vlen,
6908 // long bits,
6909 // LongFunction<V> defaultImpl)
6910 bool LibraryCallKit::inline_vector_broadcast_coerced() {
6911 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
6912 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
6913 const TypeInt* vlen = gvn().type(argument(2))->is_int();
6914
6915 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
6916 return false; // not enough info for intrinsification
6917 }
6918
6919 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
6920 if (!elem_type->is_primitive_type()) {
6921 return false; // should be primitive type
6922 }
6923 BasicType elem_bt = elem_type->basic_type();
6924 int num_elem = vlen->get_con();
6925 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
6926 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
6927
6928 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
6929 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt,
6930 vbox_klass->is_vectormask() ? VecMaskUseStore : VecMaskNotUsed)) {
6931 return false; // not supported
6932 }
6933
6934 Node* bits = argument(3); // long
6935
6936 Node* elem = NULL;
6937 switch (elem_bt) {
6938 case T_BOOLEAN: // fall-through
6939 case T_BYTE: // fall-through
6940 case T_SHORT: // fall-through
6941 case T_CHAR: // fall-through
6942 case T_INT: {
6943 elem = gvn().transform(new ConvL2INode(bits));
6944 break;
6945 }
6946 case T_DOUBLE: {
6947 elem = gvn().transform(new MoveL2DNode(bits));
6948 break;
6949 }
6950 case T_FLOAT: {
6951 bits = gvn().transform(new ConvL2INode(bits));
6952 elem = gvn().transform(new MoveI2FNode(bits));
6953 break;
6954 }
6955 case T_LONG: {
6956 elem = bits; // no conversion needed
6957 break;
6958 }
6959 default: fatal("%s", type2name(elem_bt));
6960 }
6961
6962 Node* broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt));
6963 broadcast = gvn().transform(broadcast);
6964 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
6965 set_vector_result(box);
6966
6967 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
6968 return true;
6969 }
6970
6971 // <C, V extends Vector<?,?>>
6972 // V load(Class<?> vectorClass, Class<?> elementType, int vlen,
6973 // Object base, long offset,
6974 // /* Vector.Mask<E,S> m*/
6975 // Object container, int index,
6976 // LoadVectorOperation<C, V> defaultImpl) {
6977 //
6978 // <C, V extends Vector<?,?>>
6979 // void store(Class<?> vectorClass, Class<?> elementType, int vlen,
6980 // Object base, long offset,
6981 // V v, /*Vector.Mask<E,S> m*/
6982 // Object container, int index,
6983 // StoreVectorOperation<C, V> defaultImpl) {
6984
6985 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
6986 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
6987 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
6988 const TypeInt* vlen = gvn().type(argument(2))->is_int();
6989
6990 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
6991 return false; // not enough info for intrinsification
6992 }
6993
6994 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
6995 if (!elem_type->is_primitive_type()) {
6996 return false; // should be primitive type
6997 }
6998 BasicType elem_bt = elem_type->basic_type();
6999 int num_elem = vlen->get_con();
7000
7001 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7002 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
7003 return false; // not supported
7004 }
7005
7006 Node* base = argument(3);
7007 Node* offset = ConvL2X(argument(4));
7008 Node* addr = make_unsafe_address(base, offset, elem_bt, true);
7009
7010 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
7011 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
7012
7013 // Now handle special case where load/store happens from/to byte array but element type is not byte.
7014 bool using_byte_array = arr_type != NULL && arr_type->elem()->array_element_basic_type() == T_BYTE && elem_bt != T_BYTE;
7015
7016 // It must be the case that if it is not special byte array case, there is consistency between
7017 // array and vector element types.
7018 if (!using_byte_array && arr_type != NULL && elem_bt != arr_type->elem()->array_element_basic_type()) {
7019 return false;
7020 }
7021
7022 // Since we are using byte array, we need to double check that the byte operations are supported by backend.
7023 if (using_byte_array) {
7024 int byte_num_elem = num_elem * type2aelembytes(elem_bt);
7025 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, byte_num_elem, T_BYTE, VecMaskNotUsed)) {
7026 return false; // not supported
7027 }
7028 }
7029
7030 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7031 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7032
7033 if (is_store) {
7034 Node* val = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7035 if (val == NULL) {
7036 return false; // operand unboxing failed
7037 }
7038 set_all_memory(reset_memory());
7039
7040 // In case the store needs to happen to byte array, reinterpret the incoming vector to byte vector.
7041 int store_num_elem = num_elem;
7042 if (using_byte_array) {
7043 store_num_elem = num_elem * type2aelembytes(elem_bt);
7044 const TypeVect* to_vect_type = TypeVect::make(T_BYTE, store_num_elem);
7045 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type));
7046 }
7047
7048 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(addr), addr, addr_type, val, store_num_elem));
7049 set_memory(vstore, addr_type);
7050 set_vector_result(vstore, false);
7051 } else {
7052 // When using byte array, we need to load as byte then reinterpret the value. Otherwise, do a simple vector load.
7053 Node* vload = NULL;
7054 if (using_byte_array) {
7055 int load_num_elem = num_elem * type2aelembytes(elem_bt);
7056 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, load_num_elem, T_BYTE));
7057 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
7058 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type));
7059 } else {
7060 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, elem_bt));
7061 }
7062
7063 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7064 set_vector_result(box);
7065 }
7066
7067 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7068 return true;
7069 }
7070
7071 // <V extends Vector<?,?>>
7072 // long reductionCoerced(int oprId, Class<?> vectorClass, Class<?> elementType, int vlen,
7073 // V v,
7074 // Function<V,Long> defaultImpl)
7075
7076 bool LibraryCallKit::inline_vector_reduction() {
7077 const TypeInt* opr = gvn().type(argument(0))->is_int();
7078 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7079 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7080 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7081
7082 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7083 return false; // not enough info for intrinsification
7084 }
7085 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7086 if (!elem_type->is_primitive_type()) {
7087 return false; // should be primitive type
7088 }
7089 BasicType elem_bt = elem_type->basic_type();
7090 int num_elem = vlen->get_con();
7091
7092 int opc = get_opc(opr->get_con(), elem_bt);
7093 int sopc = ReductionNode::opcode(opc, elem_bt);
7094
7095 // FIXME: When encountering a SubReduction, we want to check for support of
7096 // the corresponding AddReduction node.
7097 if (sopc == Op_SubReductionV) {
7098 if (gvn().type(argument(2))->isa_int()) {
7099 sopc = Op_AddReductionVI;
7100 } else if (gvn().type(argument(2))->isa_long()) {
7101 sopc = Op_AddReductionVL;
7102 }
7103 }
7104
7105 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7106 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7107 return false;
7108 }
7109
7110 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7111 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7112
7113 Node* opd = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7114 if (opd == NULL) {
7115 return false; // operand unboxing failed
7116 }
7117
7118 Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt);
7119 Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, elem_bt));
7120
7121 Node* bits = NULL;
7122 switch (elem_bt) {
7123 case T_BYTE:
7124 case T_SHORT:
7125 case T_INT: {
7126 bits = gvn().transform(new ConvI2LNode(rn));
7127 break;
7128 }
7129 case T_FLOAT: {
7130 rn = gvn().transform(new MoveF2INode(rn));
7131 bits = gvn().transform(new ConvI2LNode(rn));
7132 break;
7133 }
7134 case T_DOUBLE: {
7135 bits = gvn().transform(new MoveD2LNode(rn));
7136 break;
7137 }
7138 case T_LONG: {
7139 bits = rn; // no conversion needed
7140 break;
7141 }
7142 default: fatal("%s", type2name(elem_bt));
7143 }
7144 set_vector_result(bits);
7145 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7146 return true;
7147 }
7148
7149 // static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
7150 // V v1, V v2,
7151 // BiFunction<V, V, Boolean> defaultImpl) {
7152
7153 bool LibraryCallKit::inline_vector_test() {
7154 const TypeInt* cond = gvn().type(argument(0))->is_int();
7155 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7156 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7157 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7158
7159 if (!cond->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7160 return false; // not enough info for intrinsification
7161 }
7162 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7163 if (!elem_type->is_primitive_type()) {
7164 return false; // should be primitive type
7165 }
7166 BasicType elem_bt = elem_type->basic_type();
7167 int num_elem = vlen->get_con();
7168 Assembler::Condition booltest = (Assembler::Condition)cond->get_con();
7169 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7170 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7171
7172 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseLoad : VecMaskNotUsed)) {
7173 return false;
7174 }
7175
7176 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7177 Node* opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7178 if (opd1 == NULL || opd2 == NULL) {
7179 return false; // operand unboxing failed
7180 }
7181 Node* test = new VectorTestNode(opd1, opd2, booltest);
7182 test = _gvn.transform(test);
7183 set_vector_result(test);
7184 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7185 return true;
7186 }
7187
7188 // static
7189 // <V extends Vector, M extends Mask>
7190 // V blend(Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7191 // V v1, V v2, M m,
7192 // VectorBlendOp<V,M> defaultImpl) { ...
7193 //
7194 bool LibraryCallKit::inline_vector_blend() {
7195 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7196 const TypeInstPtr* mask_klass = gvn().type(argument(1))->is_instptr();
7197 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7198 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7199
7200 if (mask_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7201 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7202 return false; // not enough info for intrinsification
7203 }
7204 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7205 if (!elem_type->is_primitive_type()) {
7206 return false; // should be primitive type
7207 }
7208 BasicType elem_bt = elem_type->basic_type();
7209 BasicType mask_bt = elem_bt;
7210 int num_elem = vlen->get_con();
7211
7212 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
7213 return false; // not supported
7214 }
7215 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7216 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7217
7218 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7219 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7220
7221 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7222 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7223 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
7224
7225 if (v1 == NULL || v2 == NULL || mask == NULL) {
7226 return false; // operand unboxing failed
7227 }
7228
7229 Node* blend = _gvn.transform(new VectorBlendNode(v1, v2, mask));
7230 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
7231 set_vector_result(box);
7232
7233 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7234 return true;
7235 }
7236
7237 // static <V extends Vector<E,S>,
7238 // M extends Vector.Mask<E,S>,
7239 // S extends Vector.Shape, E>
7240 // M compare(int cond, Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7241 // V v1, V v2,
7242 // VectorCompareOp<V,M> defaultImpl) { ...
7243 //
7244 bool LibraryCallKit::inline_vector_compare() {
7245 const TypeInt* cond = gvn().type(argument(0))->is_int();
7246 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7247 const TypeInstPtr* mask_klass = gvn().type(argument(2))->is_instptr();
7248 const TypeInstPtr* elem_klass = gvn().type(argument(3))->is_instptr();
7249 const TypeInt* vlen = gvn().type(argument(4))->is_int();
7250
7251 if (!cond->is_con() || vector_klass->const_oop() == NULL || mask_klass->const_oop() == NULL ||
7252 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7253 return false; // not enough info for intrinsification
7254 }
7255 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7256 if (!elem_type->is_primitive_type()) {
7257 return false; // should be primitive type
7258 }
7259
7260 int num_elem = vlen->get_con();
7261 BasicType elem_bt = elem_type->basic_type();
7262 BasicType mask_bt = elem_bt;
7263
7264 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
7265 return false;
7266 }
7267
7268 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7269 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7270
7271 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7272 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7273
7274 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7275 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7276
7277 if (v1 == NULL || v2 == NULL) {
7278 return false; // operand unboxing failed
7279 }
7280 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
7281 const TypeVect* vt = TypeVect::make(mask_bt, num_elem);
7282 Node* operation = _gvn.transform(new VectorMaskCmpNode(pred, v1, v2, vt));
7283
7284 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
7285 set_vector_result(box);
7286
7287 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7288 return true;
7289 }
7290
7291 // static
7292 // <V extends Vector, Sh extends Shuffle>
7293 // V rearrangeOp(Class<V> vectorClass, Class<Sh> shuffleClass, Class< ? > elementType, int vlen,
7294 // V v1, Sh sh,
7295 // VectorSwizzleOp<V, Sh, S, E> defaultImpl) { ...
7296
7297 bool LibraryCallKit::inline_vector_rearrange() {
7298 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7299 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->is_instptr();
7300 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7301 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7302
7303 if (shuffle_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7304 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7305 return false; // not enough info for intrinsification
7306 }
7307 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7308 if (!elem_type->is_primitive_type()) {
7309 return false; // should be primitive type
7310 }
7311 BasicType elem_bt = elem_type->basic_type();
7312 BasicType shuffle_bt = elem_bt;
7313 int num_elem = vlen->get_con();
7314
7315 if (!arch_supports_vector(Op_VectorLoadShuffle, num_elem, elem_bt, VecMaskNotUsed)) {
7316 return false; // not supported
7317 }
7318 if (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed)) {
7319 return false; // not supported
7320 }
7321 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7322 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7323
7324 ciKlass* shbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
7325 const TypeInstPtr* shbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, shbox_klass);
7326
7327 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7328 Node* shuffle = unbox_vector(argument(5), shbox_type, shuffle_bt, num_elem);
7329
7330 if (v1 == NULL || shuffle == NULL) {
7331 return false; // operand unboxing failed
7332 }
7333
7334 Node* rearrange = _gvn.transform(new VectorRearrangeNode(v1, shuffle));
7335 Node* box = box_vector(rearrange, vbox_type, elem_bt, num_elem);
7336 set_result(box);
7337
7338 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7339 return true;
7340 }
7341
7342 Node* LibraryCallKit::shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
7343 assert(bt == T_INT || bt == T_LONG, "Only long and int are supported");
7344 juint mask = (bt == T_INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
7345 const TypeInt* t = cnt->find_int_type();
7346 if (t != NULL && t->is_con()) {
7347 juint shift = t->get_con();
7348 if (shift > mask) {
7349 return _gvn.transform(ConNode::make(TypeInt::make(shift & mask)));
7350 } else {
7351 return cnt;
7352 }
7353 } else {
7354 Node* maskedcnt = cnt;
7355 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
7356 Node* nmask = _gvn.transform(ConNode::make(TypeInt::make(mask)));
7357 maskedcnt = _gvn.transform(new AndINode(cnt, nmask));
7358 }
7359 return VectorNode::shift_count(shift_op, maskedcnt, num_elem, bt);
7360 }
7361 }
7362
7363 static void get_svml_address(int op, int bits, BasicType bt, const char** name_ptr, address* addr_ptr) {
7364 assert(name_ptr != NULL, "unexpected");
7365 assert(addr_ptr != NULL, "unexpected");
7366
7367 // Since the addresses are resolved at runtime, using switch instead of table - otherwise might get NULL addresses.
7368 if (bt == T_FLOAT) {
7369 switch(op) {
7370 case OP_EXP: {
7371 switch(bits) {
7372 case 64: *name_ptr = "vector_float64_exp"; *addr_ptr = StubRoutines::vector_float64_exp(); break;
7373 case 128: *name_ptr = "vector_float128_exp"; *addr_ptr = StubRoutines::vector_float128_exp(); break;
7374 case 256: *name_ptr = "vector_float256_exp"; *addr_ptr = StubRoutines::vector_float256_exp(); break;
7375 case 512: *name_ptr = "vector_float512_exp"; *addr_ptr = StubRoutines::vector_float512_exp(); break;
7376 default: Unimplemented(); break;
7377 }
7378 }
7379 break;
7380 case OP_LOG1P: {
7381 switch(bits) {
7382 case 64: *name_ptr = "vector_float64_log1p"; *addr_ptr = StubRoutines::vector_float64_log1p(); break;
7383 case 128: *name_ptr = "vector_float128_log1p"; *addr_ptr = StubRoutines::vector_float128_log1p(); break;
7384 case 256: *name_ptr = "vector_float256_log1p"; *addr_ptr = StubRoutines::vector_float256_log1p(); break;
7385 case 512: *name_ptr = "vector_float512_log1p"; *addr_ptr = StubRoutines::vector_float512_log1p(); break;
7386 default: Unimplemented(); break;
7387 }
7388 }
7389 break;
7390 case OP_LOG: {
7391 switch(bits) {
7392 case 64: *name_ptr = "vector_float64_log"; *addr_ptr = StubRoutines::vector_float64_log(); break;
7393 case 128: *name_ptr = "vector_float128_log"; *addr_ptr = StubRoutines::vector_float128_log(); break;
7394 case 256: *name_ptr = "vector_float256_log"; *addr_ptr = StubRoutines::vector_float256_log(); break;
7395 case 512: *name_ptr = "vector_float512_log"; *addr_ptr = StubRoutines::vector_float512_log(); break;
7396 default: Unimplemented(); break;
7397 }
7398 }
7399 break;
7400 case OP_LOG10: {
7401 switch(bits) {
7402 case 64: *name_ptr = "vector_float64_log10"; *addr_ptr = StubRoutines::vector_float64_log10(); break;
7403 case 128: *name_ptr = "vector_float128_log10"; *addr_ptr = StubRoutines::vector_float128_log10(); break;
7404 case 256: *name_ptr = "vector_float256_log10"; *addr_ptr = StubRoutines::vector_float256_log10(); break;
7405 case 512: *name_ptr = "vector_float512_log10"; *addr_ptr = StubRoutines::vector_float512_log10(); break;
7406 default: Unimplemented(); break;
7407 }
7408 }
7409 break;
7410 case OP_EXPM1: {
7411 switch(bits) {
7412 case 64: *name_ptr = "vector_float64_expm1"; *addr_ptr = StubRoutines::vector_float64_expm1(); break;
7413 case 128: *name_ptr = "vector_float128_expm1"; *addr_ptr = StubRoutines::vector_float128_expm1(); break;
7414 case 256: *name_ptr = "vector_float256_expm1"; *addr_ptr = StubRoutines::vector_float256_expm1(); break;
7415 case 512: *name_ptr = "vector_float512_expm1"; *addr_ptr = StubRoutines::vector_float512_expm1(); break;
7416 default: Unimplemented(); break;
7417 }
7418 }
7419 break;
7420 case OP_SIN: {
7421 switch(bits) {
7422 case 64: *name_ptr = "vector_float64_sin"; *addr_ptr = StubRoutines::vector_float64_sin(); break;
7423 case 128: *name_ptr = "vector_float128_sin"; *addr_ptr = StubRoutines::vector_float128_sin(); break;
7424 case 256: *name_ptr = "vector_float256_sin"; *addr_ptr = StubRoutines::vector_float256_sin(); break;
7425 case 512: *name_ptr = "vector_float512_sin"; *addr_ptr = StubRoutines::vector_float512_sin(); break;
7426 default: Unimplemented(); break;
7427 }
7428 }
7429 break;
7430 case OP_COS: {
7431 switch(bits) {
7432 case 64: *name_ptr = "vector_float64_cos"; *addr_ptr = StubRoutines::vector_float64_cos(); break;
7433 case 128: *name_ptr = "vector_float128_cos"; *addr_ptr = StubRoutines::vector_float128_cos(); break;
7434 case 256: *name_ptr = "vector_float256_cos"; *addr_ptr = StubRoutines::vector_float256_cos(); break;
7435 case 512: *name_ptr = "vector_float512_cos"; *addr_ptr = StubRoutines::vector_float512_cos(); break;
7436 default: Unimplemented(); break;
7437 }
7438 }
7439 break;
7440 case OP_TAN: {
7441 switch(bits) {
7442 case 64: *name_ptr = "vector_float64_tan"; *addr_ptr = StubRoutines::vector_float64_tan(); break;
7443 case 128: *name_ptr = "vector_float128_tan"; *addr_ptr = StubRoutines::vector_float128_tan(); break;
7444 case 256: *name_ptr = "vector_float256_tan"; *addr_ptr = StubRoutines::vector_float256_tan(); break;
7445 case 512: *name_ptr = "vector_float512_tan"; *addr_ptr = StubRoutines::vector_float512_tan(); break;
7446 default: Unimplemented(); break;
7447 }
7448 }
7449 break;
7450 case OP_SINH: {
7451 switch(bits) {
7452 case 64: *name_ptr = "vector_float64_sinh"; *addr_ptr = StubRoutines::vector_float64_sinh(); break;
7453 case 128: *name_ptr = "vector_float128_sinh"; *addr_ptr = StubRoutines::vector_float128_sinh(); break;
7454 case 256: *name_ptr = "vector_float256_sinh"; *addr_ptr = StubRoutines::vector_float256_sinh(); break;
7455 case 512: *name_ptr = "vector_float512_sinh"; *addr_ptr = StubRoutines::vector_float512_sinh(); break;
7456 default: Unimplemented(); break;
7457 }
7458 }
7459 break;
7460 case OP_COSH: {
7461 switch(bits) {
7462 case 64: *name_ptr = "vector_float64_cosh"; *addr_ptr = StubRoutines::vector_float64_cosh(); break;
7463 case 128: *name_ptr = "vector_float128_cosh"; *addr_ptr = StubRoutines::vector_float128_cosh(); break;
7464 case 256: *name_ptr = "vector_float256_cosh"; *addr_ptr = StubRoutines::vector_float256_cosh(); break;
7465 case 512: *name_ptr = "vector_float512_cosh"; *addr_ptr = StubRoutines::vector_float512_cosh(); break;
7466 default: Unimplemented(); break;
7467 }
7468 }
7469 break;
7470 case OP_TANH: {
7471 switch(bits) {
7472 case 64: *name_ptr = "vector_float64_tanh"; *addr_ptr = StubRoutines::vector_float64_tanh(); break;
7473 case 128: *name_ptr = "vector_float128_tanh"; *addr_ptr = StubRoutines::vector_float128_tanh(); break;
7474 case 256: *name_ptr = "vector_float256_tanh"; *addr_ptr = StubRoutines::vector_float256_tanh(); break;
7475 case 512: *name_ptr = "vector_float512_tanh"; *addr_ptr = StubRoutines::vector_float512_tanh(); break;
7476 default: Unimplemented(); break;
7477 }
7478 }
7479 break;
7480 case OP_ASIN: {
7481 switch(bits) {
7482 case 64: *name_ptr = "vector_float64_asin"; *addr_ptr = StubRoutines::vector_float64_asin(); break;
7483 case 128: *name_ptr = "vector_float128_asin"; *addr_ptr = StubRoutines::vector_float128_asin(); break;
7484 case 256: *name_ptr = "vector_float256_asin"; *addr_ptr = StubRoutines::vector_float256_asin(); break;
7485 case 512: *name_ptr = "vector_float512_asin"; *addr_ptr = StubRoutines::vector_float512_asin(); break;
7486 default: Unimplemented(); break;
7487 }
7488 }
7489 break;
7490 case OP_ACOS: {
7491 switch(bits) {
7492 case 64: *name_ptr = "vector_float64_acos"; *addr_ptr = StubRoutines::vector_float64_acos(); break;
7493 case 128: *name_ptr = "vector_float128_acos"; *addr_ptr = StubRoutines::vector_float128_acos(); break;
7494 case 256: *name_ptr = "vector_float256_acos"; *addr_ptr = StubRoutines::vector_float256_acos(); break;
7495 case 512: *name_ptr = "vector_float512_acos"; *addr_ptr = StubRoutines::vector_float512_acos(); break;
7496 default: Unimplemented(); break;
7497 }
7498 }
7499 break;
7500 case OP_ATAN: {
7501 switch(bits) {
7502 case 64: *name_ptr = "vector_float64_atan"; *addr_ptr = StubRoutines::vector_float64_atan(); break;
7503 case 128: *name_ptr = "vector_float128_atan"; *addr_ptr = StubRoutines::vector_float128_atan(); break;
7504 case 256: *name_ptr = "vector_float256_atan"; *addr_ptr = StubRoutines::vector_float256_atan(); break;
7505 case 512: *name_ptr = "vector_float512_atan"; *addr_ptr = StubRoutines::vector_float512_atan(); break;
7506 default: Unimplemented(); break;
7507 }
7508 }
7509 break;
7510 case OP_CBRT: {
7511 switch(bits) {
7512 case 64: *name_ptr = "vector_float64_cbrt"; *addr_ptr = StubRoutines::vector_float64_cbrt(); break;
7513 case 128: *name_ptr = "vector_float128_cbrt"; *addr_ptr = StubRoutines::vector_float128_cbrt(); break;
7514 case 256: *name_ptr = "vector_float256_cbrt"; *addr_ptr = StubRoutines::vector_float256_cbrt(); break;
7515 case 512: *name_ptr = "vector_float512_cbrt"; *addr_ptr = StubRoutines::vector_float512_cbrt(); break;
7516 default: Unimplemented(); break;
7517 }
7518 }
7519 break;
7520 case OP_HYPOT: {
7521 switch(bits) {
7522 case 64: *name_ptr = "vector_float64_hypot"; *addr_ptr = StubRoutines::vector_float64_hypot(); break;
7523 case 128: *name_ptr = "vector_float128_hypot"; *addr_ptr = StubRoutines::vector_float128_hypot(); break;
7524 case 256: *name_ptr = "vector_float256_hypot"; *addr_ptr = StubRoutines::vector_float256_hypot(); break;
7525 case 512: *name_ptr = "vector_float512_hypot"; *addr_ptr = StubRoutines::vector_float512_hypot(); break;
7526 default: Unimplemented(); break;
7527 }
7528 }
7529 break;
7530 case OP_POW: {
7531 switch(bits) {
7532 case 64: *name_ptr = "vector_float64_pow"; *addr_ptr = StubRoutines::vector_float64_pow(); break;
7533 case 128: *name_ptr = "vector_float128_pow"; *addr_ptr = StubRoutines::vector_float128_pow(); break;
7534 case 256: *name_ptr = "vector_float256_pow"; *addr_ptr = StubRoutines::vector_float256_pow(); break;
7535 case 512: *name_ptr = "vector_float512_pow"; *addr_ptr = StubRoutines::vector_float512_pow(); break;
7536 default: Unimplemented(); break;
7537 }
7538 }
7539 break;
7540 case OP_ATAN2: {
7541 switch(bits) {
7542 case 64: *name_ptr = "vector_float64_atan2"; *addr_ptr = StubRoutines::vector_float64_atan2(); break;
7543 case 128: *name_ptr = "vector_float128_atan2"; *addr_ptr = StubRoutines::vector_float128_atan2(); break;
7544 case 256: *name_ptr = "vector_float256_atan2"; *addr_ptr = StubRoutines::vector_float256_atan2(); break;
7545 case 512: *name_ptr = "vector_float512_atan2"; *addr_ptr = StubRoutines::vector_float512_atan2(); break;
7546 default: Unimplemented(); break;
7547 }
7548 }
7549 break;
7550 default:
7551 *name_ptr = "invalid";
7552 *addr_ptr = NULL;
7553 break;
7554 }
7555 } else {
7556 assert(bt == T_DOUBLE, "must be FP type only");
7557 switch(op) {
7558 case OP_EXP: {
7559 switch(bits) {
7560 case 64: *name_ptr = "vector_double64_exp"; *addr_ptr = StubRoutines::vector_double64_exp(); break;
7561 case 128: *name_ptr = "vector_double128_exp"; *addr_ptr = StubRoutines::vector_double128_exp(); break;
7562 case 256: *name_ptr = "vector_double256_exp"; *addr_ptr = StubRoutines::vector_double256_exp(); break;
7563 case 512: *name_ptr = "vector_double512_exp"; *addr_ptr = StubRoutines::vector_double512_exp(); break;
7564 default: Unimplemented(); break;
7565 }
7566 }
7567 break;
7568 case OP_LOG1P: {
7569 switch(bits) {
7570 case 64: *name_ptr = "vector_double64_log1p"; *addr_ptr = StubRoutines::vector_double64_log1p(); break;
7571 case 128: *name_ptr = "vector_double128_log1p"; *addr_ptr = StubRoutines::vector_double128_log1p(); break;
7572 case 256: *name_ptr = "vector_double256_log1p"; *addr_ptr = StubRoutines::vector_double256_log1p(); break;
7573 case 512: *name_ptr = "vector_double512_log1p"; *addr_ptr = StubRoutines::vector_double512_log1p(); break;
7574 default: Unimplemented(); break;
7575 }
7576 }
7577 break;
7578 case OP_LOG: {
7579 switch(bits) {
7580 case 64: *name_ptr = "vector_double64_log"; *addr_ptr = StubRoutines::vector_double64_log(); break;
7581 case 128: *name_ptr = "vector_double128_log"; *addr_ptr = StubRoutines::vector_double128_log(); break;
7582 case 256: *name_ptr = "vector_double256_log"; *addr_ptr = StubRoutines::vector_double256_log(); break;
7583 case 512: *name_ptr = "vector_double512_log"; *addr_ptr = StubRoutines::vector_double512_log(); break;
7584 default: Unimplemented(); break;
7585 }
7586 }
7587 break;
7588 case OP_LOG10: {
7589 switch(bits) {
7590 case 64: *name_ptr = "vector_double64_log10"; *addr_ptr = StubRoutines::vector_double64_log10(); break;
7591 case 128: *name_ptr = "vector_double128_log10"; *addr_ptr = StubRoutines::vector_double128_log10(); break;
7592 case 256: *name_ptr = "vector_double256_log10"; *addr_ptr = StubRoutines::vector_double256_log10(); break;
7593 case 512: *name_ptr = "vector_double512_log10"; *addr_ptr = StubRoutines::vector_double512_log10(); break;
7594 default: Unimplemented(); break;
7595 }
7596 }
7597 break;
7598 case OP_EXPM1: {
7599 switch(bits) {
7600 case 64: *name_ptr = "vector_double64_expm1"; *addr_ptr = StubRoutines::vector_double64_expm1(); break;
7601 case 128: *name_ptr = "vector_double128_expm1"; *addr_ptr = StubRoutines::vector_double128_expm1(); break;
7602 case 256: *name_ptr = "vector_double256_expm1"; *addr_ptr = StubRoutines::vector_double256_expm1(); break;
7603 case 512: *name_ptr = "vector_double512_expm1"; *addr_ptr = StubRoutines::vector_double512_expm1(); break;
7604 default: Unimplemented(); break;
7605 }
7606 }
7607 break;
7608 case OP_SIN: {
7609 switch(bits) {
7610 case 64: *name_ptr = "vector_double64_sin"; *addr_ptr = StubRoutines::vector_double64_sin(); break;
7611 case 128: *name_ptr = "vector_double128_sin"; *addr_ptr = StubRoutines::vector_double128_sin(); break;
7612 case 256: *name_ptr = "vector_double256_sin"; *addr_ptr = StubRoutines::vector_double256_sin(); break;
7613 case 512: *name_ptr = "vector_double512_sin"; *addr_ptr = StubRoutines::vector_double512_sin(); break;
7614 default: Unimplemented(); break;
7615 }
7616 }
7617 break;
7618 case OP_COS: {
7619 switch(bits) {
7620 case 64: *name_ptr = "vector_double64_cos"; *addr_ptr = StubRoutines::vector_double64_cos(); break;
7621 case 128: *name_ptr = "vector_double128_cos"; *addr_ptr = StubRoutines::vector_double128_cos(); break;
7622 case 256: *name_ptr = "vector_double256_cos"; *addr_ptr = StubRoutines::vector_double256_cos(); break;
7623 case 512: *name_ptr = "vector_double512_cos"; *addr_ptr = StubRoutines::vector_double512_cos(); break;
7624 default: Unimplemented(); break;
7625 }
7626 }
7627 break;
7628 case OP_TAN: {
7629 switch(bits) {
7630 case 64: *name_ptr = "vector_double64_tan"; *addr_ptr = StubRoutines::vector_double64_tan(); break;
7631 case 128: *name_ptr = "vector_double128_tan"; *addr_ptr = StubRoutines::vector_double128_tan(); break;
7632 case 256: *name_ptr = "vector_double256_tan"; *addr_ptr = StubRoutines::vector_double256_tan(); break;
7633 case 512: *name_ptr = "vector_double512_tan"; *addr_ptr = StubRoutines::vector_double512_tan(); break;
7634 default: Unimplemented(); break;
7635 }
7636 }
7637 break;
7638 case OP_SINH: {
7639 switch(bits) {
7640 case 64: *name_ptr = "vector_double64_sinh"; *addr_ptr = StubRoutines::vector_double64_sinh(); break;
7641 case 128: *name_ptr = "vector_double128_sinh"; *addr_ptr = StubRoutines::vector_double128_sinh(); break;
7642 case 256: *name_ptr = "vector_double256_sinh"; *addr_ptr = StubRoutines::vector_double256_sinh(); break;
7643 case 512: *name_ptr = "vector_double512_sinh"; *addr_ptr = StubRoutines::vector_double512_sinh(); break;
7644 default: Unimplemented(); break;
7645 }
7646 }
7647 break;
7648 case OP_COSH: {
7649 switch(bits) {
7650 case 64: *name_ptr = "vector_double64_cosh"; *addr_ptr = StubRoutines::vector_double64_cosh(); break;
7651 case 128: *name_ptr = "vector_double128_cosh"; *addr_ptr = StubRoutines::vector_double128_cosh(); break;
7652 case 256: *name_ptr = "vector_double256_cosh"; *addr_ptr = StubRoutines::vector_double256_cosh(); break;
7653 case 512: *name_ptr = "vector_double512_cosh"; *addr_ptr = StubRoutines::vector_double512_cosh(); break;
7654 default: Unimplemented(); break;
7655 }
7656 }
7657 break;
7658 case OP_TANH: {
7659 switch(bits) {
7660 case 64: *name_ptr = "vector_double64_tanh"; *addr_ptr = StubRoutines::vector_double64_tanh(); break;
7661 case 128: *name_ptr = "vector_double128_tanh"; *addr_ptr = StubRoutines::vector_double128_tanh(); break;
7662 case 256: *name_ptr = "vector_double256_tanh"; *addr_ptr = StubRoutines::vector_double256_tanh(); break;
7663 case 512: *name_ptr = "vector_double512_tanh"; *addr_ptr = StubRoutines::vector_double512_tanh(); break;
7664 default: Unimplemented(); break;
7665 }
7666 }
7667 break;
7668 case OP_ASIN: {
7669 switch(bits) {
7670 case 64: *name_ptr = "vector_double64_asin"; *addr_ptr = StubRoutines::vector_double64_asin(); break;
7671 case 128: *name_ptr = "vector_double128_asin"; *addr_ptr = StubRoutines::vector_double128_asin(); break;
7672 case 256: *name_ptr = "vector_double256_asin"; *addr_ptr = StubRoutines::vector_double256_asin(); break;
7673 case 512: *name_ptr = "vector_double512_asin"; *addr_ptr = StubRoutines::vector_double512_asin(); break;
7674 default: Unimplemented(); break;
7675 }
7676 }
7677 break;
7678 case OP_ACOS: {
7679 switch(bits) {
7680 case 64: *name_ptr = "vector_double64_acos"; *addr_ptr = StubRoutines::vector_double64_acos(); break;
7681 case 128: *name_ptr = "vector_double128_acos"; *addr_ptr = StubRoutines::vector_double128_acos(); break;
7682 case 256: *name_ptr = "vector_double256_acos"; *addr_ptr = StubRoutines::vector_double256_acos(); break;
7683 case 512: *name_ptr = "vector_double512_acos"; *addr_ptr = StubRoutines::vector_double512_acos(); break;
7684 default: Unimplemented(); break;
7685 }
7686 }
7687 break;
7688 case OP_ATAN: {
7689 switch(bits) {
7690 case 64: *name_ptr = "vector_double64_atan"; *addr_ptr = StubRoutines::vector_double64_atan(); break;
7691 case 128: *name_ptr = "vector_double128_atan"; *addr_ptr = StubRoutines::vector_double128_atan(); break;
7692 case 256: *name_ptr = "vector_double256_atan"; *addr_ptr = StubRoutines::vector_double256_atan(); break;
7693 case 512: *name_ptr = "vector_double512_atan"; *addr_ptr = StubRoutines::vector_double512_atan(); break;
7694 default: Unimplemented(); break;
7695 }
7696 }
7697 break;
7698 case OP_CBRT: {
7699 switch(bits) {
7700 case 64: *name_ptr = "vector_double64_cbrt"; *addr_ptr = StubRoutines::vector_double64_cbrt(); break;
7701 case 128: *name_ptr = "vector_double128_cbrt"; *addr_ptr = StubRoutines::vector_double128_cbrt(); break;
7702 case 256: *name_ptr = "vector_double256_cbrt"; *addr_ptr = StubRoutines::vector_double256_cbrt(); break;
7703 case 512: *name_ptr = "vector_double512_cbrt"; *addr_ptr = StubRoutines::vector_double512_cbrt(); break;
7704 default: Unimplemented(); break;
7705 }
7706 }
7707 break;
7708 case OP_HYPOT: {
7709 switch(bits) {
7710 case 64: *name_ptr = "vector_double64_hypot"; *addr_ptr = StubRoutines::vector_double64_hypot(); break;
7711 case 128: *name_ptr = "vector_double128_hypot"; *addr_ptr = StubRoutines::vector_double128_hypot(); break;
7712 case 256: *name_ptr = "vector_double256_hypot"; *addr_ptr = StubRoutines::vector_double256_hypot(); break;
7713 case 512: *name_ptr = "vector_double512_hypot"; *addr_ptr = StubRoutines::vector_double512_hypot(); break;
7714 default: Unimplemented(); break;
7715 }
7716 }
7717 break;
7718 case OP_POW: {
7719 switch(bits) {
7720 case 64: *name_ptr = "vector_double64_pow"; *addr_ptr = StubRoutines::vector_double64_pow(); break;
7721 case 128: *name_ptr = "vector_double128_pow"; *addr_ptr = StubRoutines::vector_double128_pow(); break;
7722 case 256: *name_ptr = "vector_double256_pow"; *addr_ptr = StubRoutines::vector_double256_pow(); break;
7723 case 512: *name_ptr = "vector_double512_pow"; *addr_ptr = StubRoutines::vector_double512_pow(); break;
7724 default: Unimplemented(); break;
7725 }
7726 }
7727 break;
7728 case OP_ATAN2: {
7729 switch(bits) {
7730 case 64: *name_ptr = "vector_double64_atan2"; *addr_ptr = StubRoutines::vector_double64_atan2(); break;
7731 case 128: *name_ptr = "vector_double128_atan2"; *addr_ptr = StubRoutines::vector_double128_atan2(); break;
7732 case 256: *name_ptr = "vector_double256_atan2"; *addr_ptr = StubRoutines::vector_double256_atan2(); break;
7733 case 512: *name_ptr = "vector_double512_atan2"; *addr_ptr = StubRoutines::vector_double512_atan2(); break;
7734 default: Unimplemented(); break;
7735 }
7736 }
7737 break;
7738
7739 default:
7740 *name_ptr = "invalid";
7741 *addr_ptr = NULL;
7742 break;
7743 }
7744 }
7745 }
7746
7747 Node* LibraryCallKit::gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2) {
7748 assert(vector_api_op_id >= OP_SVML_START && vector_api_op_id <= OP_SVML_END, "need valid op id");
7749 assert(opd1 != NULL, "must not be null");
7750 const TypeVect* vt = TypeVect::make(bt, num_elem);
7751 const TypeFunc* call_type = OptoRuntime::Math_Vector_Vector_Type(opd2 != NULL ? 2 : 1, vt, vt);
7752 const char* name = NULL;
7753 address addr = NULL;
7754
7755 // Get address for svml method.
7756 get_svml_address(vector_api_op_id, vt->length_in_bytes() * BitsPerByte, bt, &name, &addr);
7757
7758 if (addr == NULL) {
7759 return NULL;
7760 }
7761
7762 assert(name != NULL, "name must not be null");
7763 Node* operation = make_runtime_call(RC_VECTOR,
7764 call_type,
7765 addr,
7766 name,
7767 TypePtr::BOTTOM,
7768 opd1,
7769 opd2);
7770 return _gvn.transform(new ProjNode(_gvn.transform(operation), TypeFunc::Parms));
7771 }
7772
7773 // static
7774 // <V extends Vector<?,?>>
7775 // V broadcastInt(int opr, Class<V> vectorClass, Class<?> elementType, int vlen,
7776 // V v, int i,
7777 // VectorBroadcastIntOp<V> defaultImpl) {
7778 //
7779 bool LibraryCallKit::inline_vector_broadcast_int() {
7780 const TypeInt* opr = gvn().type(argument(0))->is_int();
7781 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7782 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7783 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7784
7785 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7786 return false; // not enough info for intrinsification
7787 }
7788 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7789 if (!elem_type->is_primitive_type()) {
7790 return false; // should be primitive type
7791 }
7792 BasicType elem_bt = elem_type->basic_type();
7793 int num_elem = vlen->get_con();
7794 int opc = get_opc(opr->get_con(), elem_bt);
7795 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
7796 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7797 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7798
7799 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7800 return false; // not supported
7801 }
7802 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7803 Node* opd2 = shift_count(argument(5), opc, elem_bt, num_elem);
7804 if (opd1 == NULL || opd2 == NULL) {
7805 return false;
7806 }
7807 Node* operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt));
7808
7809 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
7810 set_vector_result(vbox);
7811 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7812 return true;
7813 }
7814
7815 bool LibraryCallKit::inline_vector_cast_reinterpret(bool is_cast) {
7816 const TypeInstPtr* vector_klass_from = gvn().type(argument(0))->is_instptr();
7817 const TypeInstPtr* elem_klass_from = gvn().type(argument(1))->is_instptr();
7818 const TypeInt* vlen_from = gvn().type(argument(2))->is_int();
7819
7820 const TypeInstPtr* elem_klass_to = gvn().type(argument(3))->is_instptr();
7821 const TypeInt* vlen_to = gvn().type(argument(4))->is_int();
7822
7823 if (vector_klass_from->const_oop() == NULL || elem_klass_from->const_oop() == NULL || !vlen_from->is_con() ||
7824 elem_klass_to->const_oop() == NULL || !vlen_to->is_con()) {
7825 return false; // not enough info for intrinsification
7826 }
7827
7828 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
7829 if (!vbox_klass_from->is_vectorapi_vector()) {
7830 return false; // only vector & mask are supported
7831 }
7832 bool is_mask = vbox_klass_from->is_vectormask();
7833
7834 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
7835 if (!elem_type_from->is_primitive_type()) {
7836 return false; // should be primitive type
7837 }
7838 BasicType elem_bt_from = elem_type_from->basic_type();
7839 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
7840 if (!elem_type_to->is_primitive_type()) {
7841 return false; // should be primitive type
7842 }
7843 BasicType elem_bt_to = elem_type_to->basic_type();
7844 if (is_mask && elem_bt_from != elem_bt_to) {
7845 return false; // type mismatch
7846 }
7847 int num_elem_from = vlen_from->get_con();
7848 int num_elem_to = vlen_to->get_con();
7849
7850 // Check whether we can unbox to appropriate size. Even with casting, checking for reinterpret is needed
7851 // since we may need to change size.
7852 if (!arch_supports_vector(Op_VectorReinterpret,
7853 num_elem_from,
7854 elem_bt_from,
7855 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
7856 return false;
7857 }
7858
7859 // Check whether we can support resizing/reinterpreting to the new size.
7860 if (!arch_supports_vector(Op_VectorReinterpret,
7861 num_elem_to,
7862 elem_bt_to,
7863 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
7864 return false;
7865 }
7866
7867 // At this point, we know that both input and output vector registers are supported
7868 // by the architecture. Next check if the casted type is simply to same type - which means
7869 // that it is actually a resize and not a cast.
7870 if (is_cast && elem_bt_from == elem_bt_to) {
7871 is_cast = false;
7872 }
7873
7874 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
7875
7876 Node* opd1 = unbox_vector(argument(5), vbox_type_from, elem_bt_from, num_elem_from);
7877 if (opd1 == NULL) {
7878 return false;
7879 }
7880
7881 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from);
7882 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to);
7883
7884 Node* op = opd1;
7885 if (is_cast) {
7886 assert(!is_mask, "masks cannot be casted");
7887 int cast_vopc = VectorCastNode::opcode(elem_bt_from);
7888 // Make sure that cast is implemented to particular type/size combination.
7889 if (!arch_supports_vector(cast_vopc, num_elem_to, elem_bt_to, VecMaskNotUsed)) {
7890 return false;
7891 }
7892
7893 if (num_elem_from < num_elem_to) {
7894 // Since input and output number of elements are not consistent, we need to make sure we
7895 // properly size. Thus, first make a cast that retains the number of elements from source.
7896 // In case the size exceeds the arch size, we do the minimum.
7897 int num_elem_for_cast = MIN2(num_elem_from, Matcher::max_vector_size(elem_bt_to));
7898
7899 // It is possible that arch does not support this intermediate vector size
7900 // TODO More complex logic required here to handle this corner case for the sizes.
7901 if (!arch_supports_vector(cast_vopc, num_elem_for_cast, elem_bt_to, VecMaskNotUsed)) {
7902 return false;
7903 }
7904
7905 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_for_cast));
7906 // Now ensure that the destination gets properly resized to needed size.
7907 op = _gvn.transform(new VectorReinterpretNode(op, op->bottom_type()->is_vect(), dst_type));
7908 } else if (num_elem_from > num_elem_to) {
7909 // Since number elements from input is larger than output, simply reduce size of input (we are supposed to
7910 // drop top elements anyway).
7911 int num_elem_for_resize = MAX2(num_elem_to, Matcher::min_vector_size(elem_bt_to));
7912
7913 // It is possible that arch does not support this intermediate vector size
7914 // TODO More complex logic required here to handle this corner case for the sizes.
7915 if (!arch_supports_vector(Op_VectorReinterpret,
7916 num_elem_for_resize,
7917 elem_bt_from,
7918 VecMaskNotUsed)) {
7919 return false;
7920 }
7921
7922 op = _gvn.transform(new VectorReinterpretNode(op,
7923 src_type,
7924 TypeVect::make(elem_bt_from,
7925 num_elem_for_resize)));
7926 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
7927 } else {
7928 // Since input and output number of elements match, and since we know this vector size is
7929 // supported, simply do a cast with no resize needed.
7930 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
7931 }
7932 } else if (Type::cmp(src_type, dst_type) != 0) {
7933 assert(!is_cast, "must be reinterpret");
7934 op = _gvn.transform(new VectorReinterpretNode(op, src_type, dst_type));
7935 }
7936
7937 ciKlass* vbox_klass_to = get_exact_klass_for_vector_box(vbox_klass_from, elem_type_to->basic_type(),
7938 num_elem_to, is_mask ? VECAPI_MASK : VECAPI_VECTOR);
7939 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
7940 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
7941 set_vector_result(vbox);
7942 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem_to * type2aelembytes(elem_bt_to))));
7943 return true;
7944 }
7945
7946 bool LibraryCallKit::inline_vector_insert() {
7947 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7948 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7949 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7950 const TypeInt* idx = gvn().type(argument(4))->is_int();
7951
7952 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) {
7953 return false; // not enough info for intrinsification
7954 }
7955 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7956 if (!elem_type->is_primitive_type()) {
7957 return false; // should be primitive type
7958 }
7959 BasicType elem_bt = elem_type->basic_type();
7960 int num_elem = vlen->get_con();
7961 if (!arch_supports_vector(Op_VectorInsert, num_elem, elem_bt, VecMaskNotUsed)) {
7962 return false; // not supported
7963 }
7964
7965 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7966 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7967
7968 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
7969 if (opd == NULL) {
7970 return false;
7971 }
7972
7973 Node* insert_val = argument(5);
7974 assert(gvn().type(insert_val)->isa_long() != NULL, "expected to be long");
7975
7976 // Convert insert value back to its appropriate type.
7977 switch (elem_bt) {
7978 case T_BYTE:
7979 insert_val = gvn().transform(new ConvL2INode(insert_val));
7980 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::BYTE));
7981 break;
7982 case T_SHORT:
7983 insert_val = gvn().transform(new ConvL2INode(insert_val));
7984 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::SHORT));
7985 break;
7986 case T_INT:
7987 insert_val = gvn().transform(new ConvL2INode(insert_val));
7988 break;
7989 case T_FLOAT:
7990 insert_val = gvn().transform(new ConvL2INode(insert_val));
7991 insert_val = gvn().transform(new MoveI2FNode(insert_val));
7992 break;
7993 case T_DOUBLE:
7994 insert_val = gvn().transform(new MoveL2DNode(insert_val));
7995 break;
7996 case T_LONG:
7997 // no conversion needed
7998 break;
7999 default: fatal("%s", type2name(elem_bt)); break;
8000 }
8001
8002 Node* operation = VectorInsertNode::make(opd, insert_val, idx->get_con());
8003 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
8004 set_vector_result(operation);
8005
8006 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
8007 return true;
8008 }
8009
8010 bool LibraryCallKit::inline_vector_extract() {
8011 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
8012 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
8013 const TypeInt* vlen = gvn().type(argument(2))->is_int();
8014 const TypeInt* idx = gvn().type(argument(4))->is_int();
8015
8016 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) {
8017 return false; // not enough info for intrinsification
8018 }
8019 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
8020 if (!elem_type->is_primitive_type()) {
8021 return false; // should be primitive type
8022 }
8023 BasicType elem_bt = elem_type->basic_type();
8024 int num_elem = vlen->get_con();
8025 int vopc = ExtractNode::opcode(elem_bt);
8026 if (!arch_supports_vector(vopc, num_elem, elem_bt, VecMaskNotUsed)) {
8027 return false; // not supported
8028 }
8029
8030 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
8031 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
8032
8033 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
8034 if (opd == NULL) {
8035 return false;
8036 }
8037
8038 Node* operation = gvn().transform(ExtractNode::make(opd, idx->get_con(), elem_bt));
8039
8040 Node* bits = NULL;
8041 switch (elem_bt) {
8042 case T_BYTE:
8043 case T_SHORT:
8044 case T_INT: {
8045 bits = gvn().transform(new ConvI2LNode(operation));
8046 break;
8047 }
8048 case T_FLOAT: {
8049 bits = gvn().transform(new MoveF2INode(operation));
8050 bits = gvn().transform(new ConvI2LNode(bits));
8051 break;
8052 }
8053 case T_DOUBLE: {
8054 bits = gvn().transform(new MoveD2LNode(operation));
8055 break;
8056 }
8057 case T_LONG: {
8058 bits = operation; // no conversion needed
8059 break;
8060 }
8061 default: fatal("%s", type2name(elem_bt));
8062 }
8063
8064 set_vector_result(bits);
8065 return true;
8066 }
8067
8068 //------------------------------get_state_from_sha_object-----------------------
8069 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
8070 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
8071 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
8072 if (sha_state == NULL) return (Node *) NULL;
8073
8074 // now have the array, need to get the start address of the state array
8075 Node* state = array_element_address(sha_state, intcon(0), T_INT);
8076 return state;
8077 }
8078
8079 //------------------------------get_state_from_sha5_object-----------------------
8080 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
8081 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
8082 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
8083 if (sha_state == NULL) return (Node *) NULL;
8084
8085 // now have the array, need to get the start address of the state array
8086 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
8087 return state;
8088 }
8089
8090 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
8091 // Return node representing slow path of predicate check.
8092 // the pseudo code we want to emulate with this predicate is:
8093 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
8094 //
8095 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
8096 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
8097 "need SHA1/SHA256/SHA512 instruction support");
8098 assert((uint)predicate < 3, "sanity");
8099
8100 // The receiver was checked for NULL already.
8101 Node* digestBaseObj = argument(0);
8102
8103 // get DigestBase klass for instanceOf check
8104 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
8105 assert(tinst != NULL, "digestBaseObj is null");
8106 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
8107
8108 const char* klass_SHA_name = NULL;
8109 switch (predicate) {
8110 case 0:
8111 if (UseSHA1Intrinsics) {
8112 // we want to do an instanceof comparison against the SHA class
8113 klass_SHA_name = "sun/security/provider/SHA";
8114 }
8115 break;
8116 case 1:
8117 if (UseSHA256Intrinsics) {
8118 // we want to do an instanceof comparison against the SHA2 class
8119 klass_SHA_name = "sun/security/provider/SHA2";
8120 }
8121 break;
8122 case 2:
8123 if (UseSHA512Intrinsics) {
8124 // we want to do an instanceof comparison against the SHA5 class
8125 klass_SHA_name = "sun/security/provider/SHA5";
8126 }
8127 break;
8128 default:
8129 fatal("unknown SHA intrinsic predicate: %d", predicate);
8130 }
8131
8132 ciKlass* klass_SHA = NULL;
8133 if (klass_SHA_name != NULL) {
8134 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
8135 }
8136 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
8137 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
8138 Node* ctrl = control();
8139 set_control(top()); // no intrinsic path
8140 return ctrl;
8141 }
8142 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
8143
8144 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
8145 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
8146 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
8147 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
8148
8149 return instof_false; // even if it is NULL
8150 }
8151
8152 //-------------inline_fma-----------------------------------
8153 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
8154 Node *a = NULL;
8155 Node *b = NULL;
8156 Node *c = NULL;
8157 Node* result = NULL;
8158 switch (id) {
8159 case vmIntrinsics::_fmaD:
8160 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
8161 // no receiver since it is static method
8162 a = round_double_node(argument(0));
8163 b = round_double_node(argument(2));
8164 c = round_double_node(argument(4));
8165 result = _gvn.transform(new FmaDNode(control(), a, b, c));
8166 break;
8167 case vmIntrinsics::_fmaF:
8168 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
8169 a = argument(0);
8170 b = argument(1);
8171 c = argument(2);
8172 result = _gvn.transform(new FmaFNode(control(), a, b, c));
8173 break;
8174 default:
8175 fatal_unexpected_iid(id); break;
8176 }
8177 set_result(result);
8178 return true;
8179 }
8180
8181 bool LibraryCallKit::inline_profileBoolean() {
8182 Node* counts = argument(1);
8183 const TypeAryPtr* ary = NULL;
8184 ciArray* aobj = NULL;
8185 if (counts->is_Con()
8186 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
8187 && (aobj = ary->const_oop()->as_array()) != NULL
8188 && (aobj->length() == 2)) {
8189 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
8190 jint false_cnt = aobj->element_value(0).as_int();
8191 jint true_cnt = aobj->element_value(1).as_int();
8192
8193 if (C->log() != NULL) {
8194 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
8195 false_cnt, true_cnt);
8196 }
8197
8198 if (false_cnt + true_cnt == 0) {
8199 // According to profile, never executed.
8200 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
8201 Deoptimization::Action_reinterpret);
8202 return true;
8203 }
8204
8205 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
8206 // is a number of each value occurrences.
8207 Node* result = argument(0);
8208 if (false_cnt == 0 || true_cnt == 0) {
8209 // According to profile, one value has been never seen.
8210 int expected_val = (false_cnt == 0) ? 1 : 0;
8211
8212 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
8213 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
8214
8215 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
8216 Node* fast_path = _gvn.transform(new IfTrueNode(check));
8217 Node* slow_path = _gvn.transform(new IfFalseNode(check));
8218
8219 { // Slow path: uncommon trap for never seen value and then reexecute
8220 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
8221 // the value has been seen at least once.
8222 PreserveJVMState pjvms(this);
8223 PreserveReexecuteState preexecs(this);
8224 jvms()->set_should_reexecute(true);
8225
8226 set_control(slow_path);
8227 set_i_o(i_o());
8228
8229 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
8230 Deoptimization::Action_reinterpret);
8231 }
8232 // The guard for never seen value enables sharpening of the result and
8233 // returning a constant. It allows to eliminate branches on the same value
8234 // later on.
8235 set_control(fast_path);
8236 result = intcon(expected_val);
8237 }
8238 // Stop profiling.
8239 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
8240 // By replacing method body with profile data (represented as ProfileBooleanNode
8241 // on IR level) we effectively disable profiling.
8242 // It enables full speed execution once optimized code is generated.
8243 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
8244 C->record_for_igvn(profile);
8245 set_result(profile);
8246 return true;
8247 } else {
8248 // Continue profiling.
8249 // Profile data isn't available at the moment. So, execute method's bytecode version.
8250 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
8251 // is compiled and counters aren't available since corresponding MethodHandle
8252 // isn't a compile-time constant.
8253 return false;
8254 }
8255 }
8256
8257 bool LibraryCallKit::inline_isCompileConstant() {
8258 Node* n = argument(0);
8259 set_result(n->is_Con() ? intcon(1) : intcon(0));
8260 return true;
8261 }