1 /*
2 * Copyright (c) 1999, 2019, 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, DecoratorSet decorators, 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 bool inline_character_compare(vmIntrinsics::ID id);
329 bool inline_fp_min_max(vmIntrinsics::ID id);
330
331 bool inline_profileBoolean();
332 bool inline_isCompileConstant();
333
334 // Vector API support
335 bool inline_vector_nary_operation(int n);
336 bool inline_vector_broadcast_coerced();
337 bool inline_vector_mem_operation(bool is_store);
338 bool inline_vector_gather_scatter(bool is_scatter);
339 bool inline_vector_reduction();
340 bool inline_vector_test();
341 bool inline_vector_blend();
342 bool inline_vector_rearrange();
343 bool inline_vector_compare();
344 bool inline_vector_broadcast_int();
345 bool inline_vector_cast_reinterpret(bool is_cast);
346 bool inline_vector_extract();
347 bool inline_vector_insert();
348 Node* box_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
349 Node* unbox_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
350 Node* shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem);
351 Node* gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2);
352 void set_vector_result(Node* result, bool set_res = true);
353
354 void clear_upper_avx() {
355 #ifdef X86
356 if (UseAVX >= 2) {
357 C->set_clear_upper_avx(true);
358 }
359 #endif
360 }
361 };
362
363 //---------------------------make_vm_intrinsic----------------------------
364 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
365 vmIntrinsics::ID id = m->intrinsic_id();
366 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
367
368 if (!m->is_loaded()) {
369 // Do not attempt to inline unloaded methods.
370 return NULL;
371 }
372
373 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
374 bool is_available = false;
375
376 {
377 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
378 // the compiler must transition to '_thread_in_vm' state because both
379 // methods access VM-internal data.
380 VM_ENTRY_MARK;
381 methodHandle mh(THREAD, m->get_Method());
382 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
383 !C->directive()->is_intrinsic_disabled(mh) &&
384 !vmIntrinsics::is_disabled_by_flags(mh);
385
386 }
387
388 if (is_available) {
389 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
390 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
391 return new LibraryIntrinsic(m, is_virtual,
392 vmIntrinsics::predicates_needed(id),
393 vmIntrinsics::does_virtual_dispatch(id),
394 (vmIntrinsics::ID) id);
395 } else {
396 return NULL;
397 }
398 }
399
400 //----------------------register_library_intrinsics-----------------------
401 // Initialize this file's data structures, for each Compile instance.
402 void Compile::register_library_intrinsics() {
403 // Nothing to do here.
404 }
405
406 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
407 LibraryCallKit kit(jvms, this);
408 Compile* C = kit.C;
409 int nodes = C->unique();
410 #ifndef PRODUCT
411 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
412 char buf[1000];
413 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
414 tty->print_cr("Intrinsic %s", str);
415 }
416 #endif
417 ciMethod* callee = kit.callee();
418 const int bci = kit.bci();
419
420 // Try to inline the intrinsic.
421 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
422 kit.try_to_inline(_last_predicate)) {
423 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
424 : "(intrinsic)";
425 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
426 if (C->print_intrinsics() || C->print_inlining()) {
427 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
428 }
429 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
430 if (C->log()) {
431 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
432 vmIntrinsics::name_at(intrinsic_id()),
433 (is_virtual() ? " virtual='1'" : ""),
434 C->unique() - nodes);
435 }
436 // Push the result from the inlined method onto the stack.
437 kit.push_result();
438 C->print_inlining_update(this);
439 return kit.transfer_exceptions_into_jvms();
440 }
441
442 // The intrinsic bailed out
443 if (jvms->has_method()) {
444 // Not a root compile.
445 const char* msg;
446 if (callee->intrinsic_candidate()) {
447 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
448 } else {
449 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
450 : "failed to inline (intrinsic), method not annotated";
451 }
452 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
453 if (C->print_intrinsics() || C->print_inlining()) {
454 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
455 }
456 } else {
457 // Root compile
458 ResourceMark rm;
459 stringStream msg_stream;
460 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
461 vmIntrinsics::name_at(intrinsic_id()),
462 is_virtual() ? " (virtual)" : "", bci);
463 const char *msg = msg_stream.as_string();
464 log_debug(jit, inlining)("%s", msg);
465 if (C->print_intrinsics() || C->print_inlining()) {
466 tty->print("%s", msg);
467 }
468 }
469 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
470 C->print_inlining_update(this);
471 return NULL;
472 }
473
474 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
475 LibraryCallKit kit(jvms, this);
476 Compile* C = kit.C;
477 int nodes = C->unique();
478 _last_predicate = predicate;
479 #ifndef PRODUCT
480 assert(is_predicated() && predicate < predicates_count(), "sanity");
481 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
482 char buf[1000];
483 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
484 tty->print_cr("Predicate for intrinsic %s", str);
485 }
486 #endif
487 ciMethod* callee = kit.callee();
488 const int bci = kit.bci();
489
490 Node* slow_ctl = kit.try_to_predicate(predicate);
491 if (!kit.failing()) {
492 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
493 : "(intrinsic, predicate)";
494 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
495 if (C->print_intrinsics() || C->print_inlining()) {
496 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
497 }
498 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
499 if (C->log()) {
500 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
501 vmIntrinsics::name_at(intrinsic_id()),
502 (is_virtual() ? " virtual='1'" : ""),
503 C->unique() - nodes);
504 }
505 return slow_ctl; // Could be NULL if the check folds.
506 }
507
508 // The intrinsic bailed out
509 if (jvms->has_method()) {
510 // Not a root compile.
511 const char* msg = "failed to generate predicate for intrinsic";
512 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
513 if (C->print_intrinsics() || C->print_inlining()) {
514 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
515 }
516 } else {
517 // Root compile
518 ResourceMark rm;
519 stringStream msg_stream;
520 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
521 vmIntrinsics::name_at(intrinsic_id()),
522 is_virtual() ? " (virtual)" : "", bci);
523 const char *msg = msg_stream.as_string();
524 log_debug(jit, inlining)("%s", msg);
525 if (C->print_intrinsics() || C->print_inlining()) {
526 C->print_inlining_stream()->print("%s", msg);
527 }
528 }
529 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
530 return NULL;
531 }
532
533 bool LibraryCallKit::try_to_inline(int predicate) {
534 // Handle symbolic names for otherwise undistinguished boolean switches:
535 const bool is_store = true;
536 const bool is_compress = true;
537 const bool is_static = true;
538 const bool is_volatile = true;
539
540 if (!jvms()->has_method()) {
541 // Root JVMState has a null method.
542 assert(map()->memory()->Opcode() == Op_Parm, "");
543 // Insert the memory aliasing node
544 set_all_memory(reset_memory());
545 }
546 assert(merged_memory(), "");
547
548 switch (intrinsic_id()) {
549 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
550 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
551 case vmIntrinsics::_getClass: return inline_native_getClass();
552
553 case vmIntrinsics::_dsin:
554 case vmIntrinsics::_dcos:
555 case vmIntrinsics::_dtan:
556 case vmIntrinsics::_dabs:
557 case vmIntrinsics::_datan2:
558 case vmIntrinsics::_dsqrt:
559 case vmIntrinsics::_dexp:
560 case vmIntrinsics::_dlog:
561 case vmIntrinsics::_dlog10:
562 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
563
564 case vmIntrinsics::_min:
565 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
566
567 case vmIntrinsics::_notify:
568 case vmIntrinsics::_notifyAll:
569 return inline_notify(intrinsic_id());
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::_getReference: 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::_putReference: 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::_getReferenceVolatile: 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::_putReferenceVolatile: 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::_getReferenceAcquire: 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::_putReferenceRelease: 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::_getReferenceOpaque: 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::_putReferenceOpaque: 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::_compareAndSetReference: 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::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
710 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
711 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
712 case vmIntrinsics::_weakCompareAndSetReference: 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::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
731 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
732 case vmIntrinsics::_compareAndExchangeReferenceRelease: 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::_getAndSetReference: 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::_isDigit:
894 case vmIntrinsics::_isLowerCase:
895 case vmIntrinsics::_isUpperCase:
896 case vmIntrinsics::_isWhitespace:
897 return inline_character_compare(intrinsic_id());
898
899 case vmIntrinsics::_maxF:
900 case vmIntrinsics::_minF:
901 case vmIntrinsics::_maxD:
902 case vmIntrinsics::_minD:
903 return inline_fp_min_max(intrinsic_id());
904
905 case vmIntrinsics::_VectorUnaryOp:
906 return inline_vector_nary_operation(1);
907 case vmIntrinsics::_VectorBinaryOp:
908 return inline_vector_nary_operation(2);
909 case vmIntrinsics::_VectorTernaryOp:
910 return inline_vector_nary_operation(3);
911 case vmIntrinsics::_VectorBroadcastCoerced:
912 return inline_vector_broadcast_coerced();
913 case vmIntrinsics::_VectorLoadOp:
914 return inline_vector_mem_operation(/*is_store=*/false);
915 case vmIntrinsics::_VectorStoreOp:
916 return inline_vector_mem_operation(/*is_store=*/true);
917 case vmIntrinsics::_VectorGatherOp:
918 return inline_vector_gather_scatter(/*is_scatter*/ false);
919 case vmIntrinsics::_VectorScatterOp:
920 return inline_vector_gather_scatter(/*is_scatter*/ true);
921 case vmIntrinsics::_VectorReductionCoerced:
922 return inline_vector_reduction();
923 case vmIntrinsics::_VectorTest:
924 return inline_vector_test();
925 case vmIntrinsics::_VectorBlend:
926 return inline_vector_blend();
927 case vmIntrinsics::_VectorRearrange:
928 return inline_vector_rearrange();
929 case vmIntrinsics::_VectorCompare:
930 return inline_vector_compare();
931 case vmIntrinsics::_VectorBroadcastInt:
932 return inline_vector_broadcast_int();
933 case vmIntrinsics::_VectorReinterpret:
934 return inline_vector_cast_reinterpret(/*is_cast*/ false);
935 case vmIntrinsics::_VectorCast:
936 return inline_vector_cast_reinterpret(/*is_cast*/ true);
937 case vmIntrinsics::_VectorInsert:
938 return inline_vector_insert();
939 case vmIntrinsics::_VectorExtract:
940 return inline_vector_extract();
941
942 default:
943 // If you get here, it may be that someone has added a new intrinsic
944 // to the list in vmSymbols.hpp without implementing it here.
945 #ifndef PRODUCT
946 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
947 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
948 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
949 }
950 #endif
951 return false;
952 }
953 }
954
955 Node* LibraryCallKit::try_to_predicate(int predicate) {
956 if (!jvms()->has_method()) {
957 // Root JVMState has a null method.
958 assert(map()->memory()->Opcode() == Op_Parm, "");
959 // Insert the memory aliasing node
960 set_all_memory(reset_memory());
961 }
962 assert(merged_memory(), "");
963
964 switch (intrinsic_id()) {
965 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
966 return inline_cipherBlockChaining_AESCrypt_predicate(false);
967 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
968 return inline_cipherBlockChaining_AESCrypt_predicate(true);
969 case vmIntrinsics::_counterMode_AESCrypt:
970 return inline_counterMode_AESCrypt_predicate();
971 case vmIntrinsics::_digestBase_implCompressMB:
972 return inline_digestBase_implCompressMB_predicate(predicate);
973
974 default:
975 // If you get here, it may be that someone has added a new intrinsic
976 // to the list in vmSymbols.hpp without implementing it here.
977 #ifndef PRODUCT
978 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
979 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
980 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
981 }
982 #endif
983 Node* slow_ctl = control();
984 set_control(top()); // No fast path instrinsic
985 return slow_ctl;
986 }
987 }
988
989 //------------------------------set_result-------------------------------
990 // Helper function for finishing intrinsics.
991 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
992 record_for_igvn(region);
993 set_control(_gvn.transform(region));
994 set_result( _gvn.transform(value));
995 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
996 }
997
998 //------------------------------generate_guard---------------------------
999 // Helper function for generating guarded fast-slow graph structures.
1000 // The given 'test', if true, guards a slow path. If the test fails
1001 // then a fast path can be taken. (We generally hope it fails.)
1002 // In all cases, GraphKit::control() is updated to the fast path.
1003 // The returned value represents the control for the slow path.
1004 // The return value is never 'top'; it is either a valid control
1005 // or NULL if it is obvious that the slow path can never be taken.
1006 // Also, if region and the slow control are not NULL, the slow edge
1007 // is appended to the region.
1008 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1009 if (stopped()) {
1010 // Already short circuited.
1011 return NULL;
1012 }
1013
1014 // Build an if node and its projections.
1015 // If test is true we take the slow path, which we assume is uncommon.
1016 if (_gvn.type(test) == TypeInt::ZERO) {
1017 // The slow branch is never taken. No need to build this guard.
1018 return NULL;
1019 }
1020
1021 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1022
1023 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1024 if (if_slow == top()) {
1025 // The slow branch is never taken. No need to build this guard.
1026 return NULL;
1027 }
1028
1029 if (region != NULL)
1030 region->add_req(if_slow);
1031
1032 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1033 set_control(if_fast);
1034
1035 return if_slow;
1036 }
1037
1038 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1039 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1040 }
1041 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1042 return generate_guard(test, region, PROB_FAIR);
1043 }
1044
1045 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1046 Node* *pos_index) {
1047 if (stopped())
1048 return NULL; // already stopped
1049 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1050 return NULL; // index is already adequately typed
1051 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1052 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1053 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1054 if (is_neg != NULL && pos_index != NULL) {
1055 // Emulate effect of Parse::adjust_map_after_if.
1056 Node* ccast = new CastIINode(index, TypeInt::POS);
1057 ccast->set_req(0, control());
1058 (*pos_index) = _gvn.transform(ccast);
1059 }
1060 return is_neg;
1061 }
1062
1063 // Make sure that 'position' is a valid limit index, in [0..length].
1064 // There are two equivalent plans for checking this:
1065 // A. (offset + copyLength) unsigned<= arrayLength
1066 // B. offset <= (arrayLength - copyLength)
1067 // We require that all of the values above, except for the sum and
1068 // difference, are already known to be non-negative.
1069 // Plan A is robust in the face of overflow, if offset and copyLength
1070 // are both hugely positive.
1071 //
1072 // Plan B is less direct and intuitive, but it does not overflow at
1073 // all, since the difference of two non-negatives is always
1074 // representable. Whenever Java methods must perform the equivalent
1075 // check they generally use Plan B instead of Plan A.
1076 // For the moment we use Plan A.
1077 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1078 Node* subseq_length,
1079 Node* array_length,
1080 RegionNode* region) {
1081 if (stopped())
1082 return NULL; // already stopped
1083 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1084 if (zero_offset && subseq_length->eqv_uncast(array_length))
1085 return NULL; // common case of whole-array copy
1086 Node* last = subseq_length;
1087 if (!zero_offset) // last += offset
1088 last = _gvn.transform(new AddINode(last, offset));
1089 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1090 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1091 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1092 return is_over;
1093 }
1094
1095 // Emit range checks for the given String.value byte array
1096 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1097 if (stopped()) {
1098 return; // already stopped
1099 }
1100 RegionNode* bailout = new RegionNode(1);
1101 record_for_igvn(bailout);
1102 if (char_count) {
1103 // Convert char count to byte count
1104 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1105 }
1106
1107 // Offset and count must not be negative
1108 generate_negative_guard(offset, bailout);
1109 generate_negative_guard(count, bailout);
1110 // Offset + count must not exceed length of array
1111 generate_limit_guard(offset, count, load_array_length(array), bailout);
1112
1113 if (bailout->req() > 1) {
1114 PreserveJVMState pjvms(this);
1115 set_control(_gvn.transform(bailout));
1116 uncommon_trap(Deoptimization::Reason_intrinsic,
1117 Deoptimization::Action_maybe_recompile);
1118 }
1119 }
1120
1121 //--------------------------generate_current_thread--------------------
1122 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1123 ciKlass* thread_klass = env()->Thread_klass();
1124 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1125 Node* thread = _gvn.transform(new ThreadLocalNode());
1126 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1127 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1128 tls_output = thread;
1129 return threadObj;
1130 }
1131
1132
1133 //------------------------------make_string_method_node------------------------
1134 // Helper method for String intrinsic functions. This version is called with
1135 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1136 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1137 // containing the lengths of str1 and str2.
1138 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1139 Node* result = NULL;
1140 switch (opcode) {
1141 case Op_StrIndexOf:
1142 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1143 str1_start, cnt1, str2_start, cnt2, ae);
1144 break;
1145 case Op_StrComp:
1146 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1147 str1_start, cnt1, str2_start, cnt2, ae);
1148 break;
1149 case Op_StrEquals:
1150 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1151 // Use the constant length if there is one because optimized match rule may exist.
1152 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1153 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1154 break;
1155 default:
1156 ShouldNotReachHere();
1157 return NULL;
1158 }
1159
1160 // All these intrinsics have checks.
1161 C->set_has_split_ifs(true); // Has chance for split-if optimization
1162 clear_upper_avx();
1163
1164 return _gvn.transform(result);
1165 }
1166
1167 //------------------------------inline_string_compareTo------------------------
1168 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1169 Node* arg1 = argument(0);
1170 Node* arg2 = argument(1);
1171
1172 arg1 = must_be_not_null(arg1, true);
1173 arg2 = must_be_not_null(arg2, true);
1174
1175 arg1 = access_resolve(arg1, ACCESS_READ);
1176 arg2 = access_resolve(arg2, ACCESS_READ);
1177
1178 // Get start addr and length of first argument
1179 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1180 Node* arg1_cnt = load_array_length(arg1);
1181
1182 // Get start addr and length of second argument
1183 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1184 Node* arg2_cnt = load_array_length(arg2);
1185
1186 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1187 set_result(result);
1188 return true;
1189 }
1190
1191 //------------------------------inline_string_equals------------------------
1192 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1193 Node* arg1 = argument(0);
1194 Node* arg2 = argument(1);
1195
1196 // paths (plus control) merge
1197 RegionNode* region = new RegionNode(3);
1198 Node* phi = new PhiNode(region, TypeInt::BOOL);
1199
1200 if (!stopped()) {
1201
1202 arg1 = must_be_not_null(arg1, true);
1203 arg2 = must_be_not_null(arg2, true);
1204
1205 arg1 = access_resolve(arg1, ACCESS_READ);
1206 arg2 = access_resolve(arg2, ACCESS_READ);
1207
1208 // Get start addr and length of first argument
1209 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1210 Node* arg1_cnt = load_array_length(arg1);
1211
1212 // Get start addr and length of second argument
1213 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1214 Node* arg2_cnt = load_array_length(arg2);
1215
1216 // Check for arg1_cnt != arg2_cnt
1217 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1218 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1219 Node* if_ne = generate_slow_guard(bol, NULL);
1220 if (if_ne != NULL) {
1221 phi->init_req(2, intcon(0));
1222 region->init_req(2, if_ne);
1223 }
1224
1225 // Check for count == 0 is done by assembler code for StrEquals.
1226
1227 if (!stopped()) {
1228 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1229 phi->init_req(1, equals);
1230 region->init_req(1, control());
1231 }
1232 }
1233
1234 // post merge
1235 set_control(_gvn.transform(region));
1236 record_for_igvn(region);
1237
1238 set_result(_gvn.transform(phi));
1239 return true;
1240 }
1241
1242 //------------------------------inline_array_equals----------------------------
1243 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1244 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1245 Node* arg1 = argument(0);
1246 Node* arg2 = argument(1);
1247
1248 arg1 = access_resolve(arg1, ACCESS_READ);
1249 arg2 = access_resolve(arg2, ACCESS_READ);
1250
1251 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1252 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1253 clear_upper_avx();
1254
1255 return true;
1256 }
1257
1258 //------------------------------inline_hasNegatives------------------------------
1259 bool LibraryCallKit::inline_hasNegatives() {
1260 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1261 return false;
1262 }
1263
1264 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1265 // no receiver since it is static method
1266 Node* ba = argument(0);
1267 Node* offset = argument(1);
1268 Node* len = argument(2);
1269
1270 ba = must_be_not_null(ba, true);
1271
1272 // Range checks
1273 generate_string_range_check(ba, offset, len, false);
1274 if (stopped()) {
1275 return true;
1276 }
1277 ba = access_resolve(ba, ACCESS_READ);
1278 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1279 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1280 set_result(_gvn.transform(result));
1281 return true;
1282 }
1283
1284 bool LibraryCallKit::inline_preconditions_checkIndex() {
1285 Node* index = argument(0);
1286 Node* length = argument(1);
1287 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1288 return false;
1289 }
1290
1291 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1292 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1293
1294 {
1295 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1296 uncommon_trap(Deoptimization::Reason_intrinsic,
1297 Deoptimization::Action_make_not_entrant);
1298 }
1299
1300 if (stopped()) {
1301 return false;
1302 }
1303
1304 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1305 BoolTest::mask btest = BoolTest::lt;
1306 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1307 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1308 _gvn.set_type(rc, rc->Value(&_gvn));
1309 if (!rc_bool->is_Con()) {
1310 record_for_igvn(rc);
1311 }
1312 set_control(_gvn.transform(new IfTrueNode(rc)));
1313 {
1314 PreserveJVMState pjvms(this);
1315 set_control(_gvn.transform(new IfFalseNode(rc)));
1316 uncommon_trap(Deoptimization::Reason_range_check,
1317 Deoptimization::Action_make_not_entrant);
1318 }
1319
1320 if (stopped()) {
1321 return false;
1322 }
1323
1324 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1325 result->set_req(0, control());
1326 result = _gvn.transform(result);
1327 set_result(result);
1328 replace_in_map(index, result);
1329 clear_upper_avx();
1330 return true;
1331 }
1332
1333 //------------------------------inline_string_indexOf------------------------
1334 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1335 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1336 return false;
1337 }
1338 Node* src = argument(0);
1339 Node* tgt = argument(1);
1340
1341 // Make the merge point
1342 RegionNode* result_rgn = new RegionNode(4);
1343 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1344
1345 src = must_be_not_null(src, true);
1346 tgt = must_be_not_null(tgt, true);
1347
1348 src = access_resolve(src, ACCESS_READ);
1349 tgt = access_resolve(tgt, ACCESS_READ);
1350
1351 // Get start addr and length of source string
1352 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1353 Node* src_count = load_array_length(src);
1354
1355 // Get start addr and length of substring
1356 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1357 Node* tgt_count = load_array_length(tgt);
1358
1359 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1360 // Divide src size by 2 if String is UTF16 encoded
1361 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1362 }
1363 if (ae == StrIntrinsicNode::UU) {
1364 // Divide substring size by 2 if String is UTF16 encoded
1365 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1366 }
1367
1368 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1369 if (result != NULL) {
1370 result_phi->init_req(3, result);
1371 result_rgn->init_req(3, control());
1372 }
1373 set_control(_gvn.transform(result_rgn));
1374 record_for_igvn(result_rgn);
1375 set_result(_gvn.transform(result_phi));
1376
1377 return true;
1378 }
1379
1380 //-----------------------------inline_string_indexOf-----------------------
1381 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1382 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1383 return false;
1384 }
1385 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1386 return false;
1387 }
1388 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1389 Node* src = argument(0); // byte[]
1390 Node* src_count = argument(1); // char count
1391 Node* tgt = argument(2); // byte[]
1392 Node* tgt_count = argument(3); // char count
1393 Node* from_index = argument(4); // char index
1394
1395 src = must_be_not_null(src, true);
1396 tgt = must_be_not_null(tgt, true);
1397
1398 src = access_resolve(src, ACCESS_READ);
1399 tgt = access_resolve(tgt, ACCESS_READ);
1400
1401 // Multiply byte array index by 2 if String is UTF16 encoded
1402 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1403 src_count = _gvn.transform(new SubINode(src_count, from_index));
1404 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1405 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1406
1407 // Range checks
1408 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1409 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1410 if (stopped()) {
1411 return true;
1412 }
1413
1414 RegionNode* region = new RegionNode(5);
1415 Node* phi = new PhiNode(region, TypeInt::INT);
1416
1417 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1418 if (result != NULL) {
1419 // The result is index relative to from_index if substring was found, -1 otherwise.
1420 // Generate code which will fold into cmove.
1421 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1422 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1423
1424 Node* if_lt = generate_slow_guard(bol, NULL);
1425 if (if_lt != NULL) {
1426 // result == -1
1427 phi->init_req(3, result);
1428 region->init_req(3, if_lt);
1429 }
1430 if (!stopped()) {
1431 result = _gvn.transform(new AddINode(result, from_index));
1432 phi->init_req(4, result);
1433 region->init_req(4, control());
1434 }
1435 }
1436
1437 set_control(_gvn.transform(region));
1438 record_for_igvn(region);
1439 set_result(_gvn.transform(phi));
1440 clear_upper_avx();
1441
1442 return true;
1443 }
1444
1445 // Create StrIndexOfNode with fast path checks
1446 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1447 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1448 // Check for substr count > string count
1449 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1450 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1451 Node* if_gt = generate_slow_guard(bol, NULL);
1452 if (if_gt != NULL) {
1453 phi->init_req(1, intcon(-1));
1454 region->init_req(1, if_gt);
1455 }
1456 if (!stopped()) {
1457 // Check for substr count == 0
1458 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1459 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1460 Node* if_zero = generate_slow_guard(bol, NULL);
1461 if (if_zero != NULL) {
1462 phi->init_req(2, intcon(0));
1463 region->init_req(2, if_zero);
1464 }
1465 }
1466 if (!stopped()) {
1467 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1468 }
1469 return NULL;
1470 }
1471
1472 //-----------------------------inline_string_indexOfChar-----------------------
1473 bool LibraryCallKit::inline_string_indexOfChar() {
1474 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1475 return false;
1476 }
1477 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1478 return false;
1479 }
1480 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1481 Node* src = argument(0); // byte[]
1482 Node* tgt = argument(1); // tgt is int ch
1483 Node* from_index = argument(2);
1484 Node* max = argument(3);
1485
1486 src = must_be_not_null(src, true);
1487 src = access_resolve(src, ACCESS_READ);
1488
1489 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1490 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1491 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1492
1493 // Range checks
1494 generate_string_range_check(src, src_offset, src_count, true);
1495 if (stopped()) {
1496 return true;
1497 }
1498
1499 RegionNode* region = new RegionNode(3);
1500 Node* phi = new PhiNode(region, TypeInt::INT);
1501
1502 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1503 C->set_has_split_ifs(true); // Has chance for split-if optimization
1504 _gvn.transform(result);
1505
1506 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1507 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1508
1509 Node* if_lt = generate_slow_guard(bol, NULL);
1510 if (if_lt != NULL) {
1511 // result == -1
1512 phi->init_req(2, result);
1513 region->init_req(2, if_lt);
1514 }
1515 if (!stopped()) {
1516 result = _gvn.transform(new AddINode(result, from_index));
1517 phi->init_req(1, result);
1518 region->init_req(1, control());
1519 }
1520 set_control(_gvn.transform(region));
1521 record_for_igvn(region);
1522 set_result(_gvn.transform(phi));
1523
1524 return true;
1525 }
1526 //---------------------------inline_string_copy---------------------
1527 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1528 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1529 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1530 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1531 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1532 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1533 bool LibraryCallKit::inline_string_copy(bool compress) {
1534 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1535 return false;
1536 }
1537 int nargs = 5; // 2 oops, 3 ints
1538 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1539
1540 Node* src = argument(0);
1541 Node* src_offset = argument(1);
1542 Node* dst = argument(2);
1543 Node* dst_offset = argument(3);
1544 Node* length = argument(4);
1545
1546 // Check for allocation before we add nodes that would confuse
1547 // tightly_coupled_allocation()
1548 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1549
1550 // Figure out the size and type of the elements we will be copying.
1551 const Type* src_type = src->Value(&_gvn);
1552 const Type* dst_type = dst->Value(&_gvn);
1553 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1554 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1555 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1556 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1557 "Unsupported array types for inline_string_copy");
1558
1559 src = must_be_not_null(src, true);
1560 dst = must_be_not_null(dst, true);
1561
1562 // Convert char[] offsets to byte[] offsets
1563 bool convert_src = (compress && src_elem == T_BYTE);
1564 bool convert_dst = (!compress && dst_elem == T_BYTE);
1565 if (convert_src) {
1566 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1567 } else if (convert_dst) {
1568 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1569 }
1570
1571 // Range checks
1572 generate_string_range_check(src, src_offset, length, convert_src);
1573 generate_string_range_check(dst, dst_offset, length, convert_dst);
1574 if (stopped()) {
1575 return true;
1576 }
1577
1578 src = access_resolve(src, ACCESS_READ);
1579 dst = access_resolve(dst, ACCESS_WRITE);
1580
1581 Node* src_start = array_element_address(src, src_offset, src_elem);
1582 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1583 // 'src_start' points to src array + scaled offset
1584 // 'dst_start' points to dst array + scaled offset
1585 Node* count = NULL;
1586 if (compress) {
1587 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1588 } else {
1589 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1590 }
1591
1592 if (alloc != NULL) {
1593 if (alloc->maybe_set_complete(&_gvn)) {
1594 // "You break it, you buy it."
1595 InitializeNode* init = alloc->initialization();
1596 assert(init->is_complete(), "we just did this");
1597 init->set_complete_with_arraycopy();
1598 assert(dst->is_CheckCastPP(), "sanity");
1599 assert(dst->in(0)->in(0) == init, "dest pinned");
1600 }
1601 // Do not let stores that initialize this object be reordered with
1602 // a subsequent store that would make this object accessible by
1603 // other threads.
1604 // Record what AllocateNode this StoreStore protects so that
1605 // escape analysis can go from the MemBarStoreStoreNode to the
1606 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1607 // based on the escape status of the AllocateNode.
1608 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1609 }
1610 if (compress) {
1611 set_result(_gvn.transform(count));
1612 }
1613 clear_upper_avx();
1614
1615 return true;
1616 }
1617
1618 #ifdef _LP64
1619 #define XTOP ,top() /*additional argument*/
1620 #else //_LP64
1621 #define XTOP /*no additional argument*/
1622 #endif //_LP64
1623
1624 //------------------------inline_string_toBytesU--------------------------
1625 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1626 bool LibraryCallKit::inline_string_toBytesU() {
1627 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1628 return false;
1629 }
1630 // Get the arguments.
1631 Node* value = argument(0);
1632 Node* offset = argument(1);
1633 Node* length = argument(2);
1634
1635 Node* newcopy = NULL;
1636
1637 // Set the original stack and the reexecute bit for the interpreter to reexecute
1638 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1639 { PreserveReexecuteState preexecs(this);
1640 jvms()->set_should_reexecute(true);
1641
1642 // Check if a null path was taken unconditionally.
1643 value = null_check(value);
1644
1645 RegionNode* bailout = new RegionNode(1);
1646 record_for_igvn(bailout);
1647
1648 // Range checks
1649 generate_negative_guard(offset, bailout);
1650 generate_negative_guard(length, bailout);
1651 generate_limit_guard(offset, length, load_array_length(value), bailout);
1652 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1653 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1654
1655 if (bailout->req() > 1) {
1656 PreserveJVMState pjvms(this);
1657 set_control(_gvn.transform(bailout));
1658 uncommon_trap(Deoptimization::Reason_intrinsic,
1659 Deoptimization::Action_maybe_recompile);
1660 }
1661 if (stopped()) {
1662 return true;
1663 }
1664
1665 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1666 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1667 newcopy = new_array(klass_node, size, 0); // no arguments to push
1668 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1669
1670 // Calculate starting addresses.
1671 value = access_resolve(value, ACCESS_READ);
1672 Node* src_start = array_element_address(value, offset, T_CHAR);
1673 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1674
1675 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1676 const TypeInt* toffset = gvn().type(offset)->is_int();
1677 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1678
1679 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1680 const char* copyfunc_name = "arraycopy";
1681 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1682 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1683 OptoRuntime::fast_arraycopy_Type(),
1684 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1685 src_start, dst_start, ConvI2X(length) XTOP);
1686 // Do not let reads from the cloned object float above the arraycopy.
1687 if (alloc != NULL) {
1688 if (alloc->maybe_set_complete(&_gvn)) {
1689 // "You break it, you buy it."
1690 InitializeNode* init = alloc->initialization();
1691 assert(init->is_complete(), "we just did this");
1692 init->set_complete_with_arraycopy();
1693 assert(newcopy->is_CheckCastPP(), "sanity");
1694 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1695 }
1696 // Do not let stores that initialize this object be reordered with
1697 // a subsequent store that would make this object accessible by
1698 // other threads.
1699 // Record what AllocateNode this StoreStore protects so that
1700 // escape analysis can go from the MemBarStoreStoreNode to the
1701 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1702 // based on the escape status of the AllocateNode.
1703 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1704 } else {
1705 insert_mem_bar(Op_MemBarCPUOrder);
1706 }
1707 } // original reexecute is set back here
1708
1709 C->set_has_split_ifs(true); // Has chance for split-if optimization
1710 if (!stopped()) {
1711 set_result(newcopy);
1712 }
1713 clear_upper_avx();
1714
1715 return true;
1716 }
1717
1718 //------------------------inline_string_getCharsU--------------------------
1719 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1720 bool LibraryCallKit::inline_string_getCharsU() {
1721 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1722 return false;
1723 }
1724
1725 // Get the arguments.
1726 Node* src = argument(0);
1727 Node* src_begin = argument(1);
1728 Node* src_end = argument(2); // exclusive offset (i < src_end)
1729 Node* dst = argument(3);
1730 Node* dst_begin = argument(4);
1731
1732 // Check for allocation before we add nodes that would confuse
1733 // tightly_coupled_allocation()
1734 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1735
1736 // Check if a null path was taken unconditionally.
1737 src = null_check(src);
1738 dst = null_check(dst);
1739 if (stopped()) {
1740 return true;
1741 }
1742
1743 // Get length and convert char[] offset to byte[] offset
1744 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1745 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1746
1747 // Range checks
1748 generate_string_range_check(src, src_begin, length, true);
1749 generate_string_range_check(dst, dst_begin, length, false);
1750 if (stopped()) {
1751 return true;
1752 }
1753
1754 if (!stopped()) {
1755 src = access_resolve(src, ACCESS_READ);
1756 dst = access_resolve(dst, ACCESS_WRITE);
1757
1758 // Calculate starting addresses.
1759 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1760 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1761
1762 // Check if array addresses are aligned to HeapWordSize
1763 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1764 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1765 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1766 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1767
1768 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1769 const char* copyfunc_name = "arraycopy";
1770 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1771 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1772 OptoRuntime::fast_arraycopy_Type(),
1773 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1774 src_start, dst_start, ConvI2X(length) XTOP);
1775 // Do not let reads from the cloned object float above the arraycopy.
1776 if (alloc != NULL) {
1777 if (alloc->maybe_set_complete(&_gvn)) {
1778 // "You break it, you buy it."
1779 InitializeNode* init = alloc->initialization();
1780 assert(init->is_complete(), "we just did this");
1781 init->set_complete_with_arraycopy();
1782 assert(dst->is_CheckCastPP(), "sanity");
1783 assert(dst->in(0)->in(0) == init, "dest pinned");
1784 }
1785 // Do not let stores that initialize this object be reordered with
1786 // a subsequent store that would make this object accessible by
1787 // other threads.
1788 // Record what AllocateNode this StoreStore protects so that
1789 // escape analysis can go from the MemBarStoreStoreNode to the
1790 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1791 // based on the escape status of the AllocateNode.
1792 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1793 } else {
1794 insert_mem_bar(Op_MemBarCPUOrder);
1795 }
1796 }
1797
1798 C->set_has_split_ifs(true); // Has chance for split-if optimization
1799 return true;
1800 }
1801
1802 //----------------------inline_string_char_access----------------------------
1803 // Store/Load char to/from byte[] array.
1804 // static void StringUTF16.putChar(byte[] val, int index, int c)
1805 // static char StringUTF16.getChar(byte[] val, int index)
1806 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1807 Node* value = argument(0);
1808 Node* index = argument(1);
1809 Node* ch = is_store ? argument(2) : NULL;
1810
1811 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1812 // correctly requires matched array shapes.
1813 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1814 "sanity: byte[] and char[] bases agree");
1815 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1816 "sanity: byte[] and char[] scales agree");
1817
1818 // Bail when getChar over constants is requested: constant folding would
1819 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1820 // Java method would constant fold nicely instead.
1821 if (!is_store && value->is_Con() && index->is_Con()) {
1822 return false;
1823 }
1824
1825 value = must_be_not_null(value, true);
1826 value = access_resolve(value, is_store ? ACCESS_WRITE : ACCESS_READ);
1827
1828 Node* adr = array_element_address(value, index, T_CHAR);
1829 if (adr->is_top()) {
1830 return false;
1831 }
1832 if (is_store) {
1833 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED);
1834 } else {
1835 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD);
1836 set_result(ch);
1837 }
1838 return true;
1839 }
1840
1841 //--------------------------round_double_node--------------------------------
1842 // Round a double node if necessary.
1843 Node* LibraryCallKit::round_double_node(Node* n) {
1844 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1845 n = _gvn.transform(new RoundDoubleNode(0, n));
1846 return n;
1847 }
1848
1849 //------------------------------inline_math-----------------------------------
1850 // public static double Math.abs(double)
1851 // public static double Math.sqrt(double)
1852 // public static double Math.log(double)
1853 // public static double Math.log10(double)
1854 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1855 Node* arg = round_double_node(argument(0));
1856 Node* n = NULL;
1857 switch (id) {
1858 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1859 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1860 default: fatal_unexpected_iid(id); break;
1861 }
1862 set_result(_gvn.transform(n));
1863 return true;
1864 }
1865
1866 //------------------------------runtime_math-----------------------------
1867 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1868 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1869 "must be (DD)D or (D)D type");
1870
1871 // Inputs
1872 Node* a = round_double_node(argument(0));
1873 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1874
1875 const TypePtr* no_memory_effects = NULL;
1876 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1877 no_memory_effects,
1878 a, top(), b, b ? top() : NULL);
1879 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1880 #ifdef ASSERT
1881 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1882 assert(value_top == top(), "second value must be top");
1883 #endif
1884
1885 set_result(value);
1886 return true;
1887 }
1888
1889 //------------------------------inline_math_native-----------------------------
1890 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1891 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1892 switch (id) {
1893 // These intrinsics are not properly supported on all hardware
1894 case vmIntrinsics::_dsin:
1895 return StubRoutines::dsin() != NULL ?
1896 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1897 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1898 case vmIntrinsics::_dcos:
1899 return StubRoutines::dcos() != NULL ?
1900 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1901 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1902 case vmIntrinsics::_dtan:
1903 return StubRoutines::dtan() != NULL ?
1904 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1905 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1906 case vmIntrinsics::_dlog:
1907 return StubRoutines::dlog() != NULL ?
1908 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1909 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1910 case vmIntrinsics::_dlog10:
1911 return StubRoutines::dlog10() != NULL ?
1912 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1913 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1914
1915 // These intrinsics are supported on all hardware
1916 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1917 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1918
1919 case vmIntrinsics::_dexp:
1920 return StubRoutines::dexp() != NULL ?
1921 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1922 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1923 case vmIntrinsics::_dpow: {
1924 Node* exp = round_double_node(argument(2));
1925 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1926 if (d != NULL && d->getd() == 2.0) {
1927 // Special case: pow(x, 2.0) => x * x
1928 Node* base = round_double_node(argument(0));
1929 set_result(_gvn.transform(new MulDNode(base, base)));
1930 return true;
1931 }
1932 return StubRoutines::dpow() != NULL ?
1933 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1934 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1935 }
1936 #undef FN_PTR
1937
1938 // These intrinsics are not yet correctly implemented
1939 case vmIntrinsics::_datan2:
1940 return false;
1941
1942 default:
1943 fatal_unexpected_iid(id);
1944 return false;
1945 }
1946 }
1947
1948 static bool is_simple_name(Node* n) {
1949 return (n->req() == 1 // constant
1950 || (n->is_Type() && n->as_Type()->type()->singleton())
1951 || n->is_Proj() // parameter or return value
1952 || n->is_Phi() // local of some sort
1953 );
1954 }
1955
1956 //----------------------------inline_notify-----------------------------------*
1957 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1958 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1959 address func;
1960 if (id == vmIntrinsics::_notify) {
1961 func = OptoRuntime::monitor_notify_Java();
1962 } else {
1963 func = OptoRuntime::monitor_notifyAll_Java();
1964 }
1965 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1966 make_slow_call_ex(call, env()->Throwable_klass(), false);
1967 return true;
1968 }
1969
1970
1971 //----------------------------inline_min_max-----------------------------------
1972 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1973 set_result(generate_min_max(id, argument(0), argument(1)));
1974 return true;
1975 }
1976
1977 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1978 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1979 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1980 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1981 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1982
1983 {
1984 PreserveJVMState pjvms(this);
1985 PreserveReexecuteState preexecs(this);
1986 jvms()->set_should_reexecute(true);
1987
1988 set_control(slow_path);
1989 set_i_o(i_o());
1990
1991 uncommon_trap(Deoptimization::Reason_intrinsic,
1992 Deoptimization::Action_none);
1993 }
1994
1995 set_control(fast_path);
1996 set_result(math);
1997 }
1998
1999 template <typename OverflowOp>
2000 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2001 typedef typename OverflowOp::MathOp MathOp;
2002
2003 MathOp* mathOp = new MathOp(arg1, arg2);
2004 Node* operation = _gvn.transform( mathOp );
2005 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2006 inline_math_mathExact(operation, ofcheck);
2007 return true;
2008 }
2009
2010 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2011 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2012 }
2013
2014 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2015 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2016 }
2017
2018 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2019 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2020 }
2021
2022 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2023 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2024 }
2025
2026 bool LibraryCallKit::inline_math_negateExactI() {
2027 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2028 }
2029
2030 bool LibraryCallKit::inline_math_negateExactL() {
2031 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2032 }
2033
2034 bool LibraryCallKit::inline_math_multiplyExactI() {
2035 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2036 }
2037
2038 bool LibraryCallKit::inline_math_multiplyExactL() {
2039 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2040 }
2041
2042 bool LibraryCallKit::inline_math_multiplyHigh() {
2043 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
2044 return true;
2045 }
2046
2047 Node*
2048 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2049 // These are the candidate return value:
2050 Node* xvalue = x0;
2051 Node* yvalue = y0;
2052
2053 if (xvalue == yvalue) {
2054 return xvalue;
2055 }
2056
2057 bool want_max = (id == vmIntrinsics::_max);
2058
2059 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2060 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2061 if (txvalue == NULL || tyvalue == NULL) return top();
2062 // This is not really necessary, but it is consistent with a
2063 // hypothetical MaxINode::Value method:
2064 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2065
2066 // %%% This folding logic should (ideally) be in a different place.
2067 // Some should be inside IfNode, and there to be a more reliable
2068 // transformation of ?: style patterns into cmoves. We also want
2069 // more powerful optimizations around cmove and min/max.
2070
2071 // Try to find a dominating comparison of these guys.
2072 // It can simplify the index computation for Arrays.copyOf
2073 // and similar uses of System.arraycopy.
2074 // First, compute the normalized version of CmpI(x, y).
2075 int cmp_op = Op_CmpI;
2076 Node* xkey = xvalue;
2077 Node* ykey = yvalue;
2078 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2079 if (ideal_cmpxy->is_Cmp()) {
2080 // E.g., if we have CmpI(length - offset, count),
2081 // it might idealize to CmpI(length, count + offset)
2082 cmp_op = ideal_cmpxy->Opcode();
2083 xkey = ideal_cmpxy->in(1);
2084 ykey = ideal_cmpxy->in(2);
2085 }
2086
2087 // Start by locating any relevant comparisons.
2088 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2089 Node* cmpxy = NULL;
2090 Node* cmpyx = NULL;
2091 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2092 Node* cmp = start_from->fast_out(k);
2093 if (cmp->outcnt() > 0 && // must have prior uses
2094 cmp->in(0) == NULL && // must be context-independent
2095 cmp->Opcode() == cmp_op) { // right kind of compare
2096 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2097 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2098 }
2099 }
2100
2101 const int NCMPS = 2;
2102 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2103 int cmpn;
2104 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2105 if (cmps[cmpn] != NULL) break; // find a result
2106 }
2107 if (cmpn < NCMPS) {
2108 // Look for a dominating test that tells us the min and max.
2109 int depth = 0; // Limit search depth for speed
2110 Node* dom = control();
2111 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2112 if (++depth >= 100) break;
2113 Node* ifproj = dom;
2114 if (!ifproj->is_Proj()) continue;
2115 Node* iff = ifproj->in(0);
2116 if (!iff->is_If()) continue;
2117 Node* bol = iff->in(1);
2118 if (!bol->is_Bool()) continue;
2119 Node* cmp = bol->in(1);
2120 if (cmp == NULL) continue;
2121 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2122 if (cmps[cmpn] == cmp) break;
2123 if (cmpn == NCMPS) continue;
2124 BoolTest::mask btest = bol->as_Bool()->_test._test;
2125 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2126 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2127 // At this point, we know that 'x btest y' is true.
2128 switch (btest) {
2129 case BoolTest::eq:
2130 // They are proven equal, so we can collapse the min/max.
2131 // Either value is the answer. Choose the simpler.
2132 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2133 return yvalue;
2134 return xvalue;
2135 case BoolTest::lt: // x < y
2136 case BoolTest::le: // x <= y
2137 return (want_max ? yvalue : xvalue);
2138 case BoolTest::gt: // x > y
2139 case BoolTest::ge: // x >= y
2140 return (want_max ? xvalue : yvalue);
2141 default:
2142 break;
2143 }
2144 }
2145 }
2146
2147 // We failed to find a dominating test.
2148 // Let's pick a test that might GVN with prior tests.
2149 Node* best_bol = NULL;
2150 BoolTest::mask best_btest = BoolTest::illegal;
2151 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2152 Node* cmp = cmps[cmpn];
2153 if (cmp == NULL) continue;
2154 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2155 Node* bol = cmp->fast_out(j);
2156 if (!bol->is_Bool()) continue;
2157 BoolTest::mask btest = bol->as_Bool()->_test._test;
2158 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2159 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2160 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2161 best_bol = bol->as_Bool();
2162 best_btest = btest;
2163 }
2164 }
2165 }
2166
2167 Node* answer_if_true = NULL;
2168 Node* answer_if_false = NULL;
2169 switch (best_btest) {
2170 default:
2171 if (cmpxy == NULL)
2172 cmpxy = ideal_cmpxy;
2173 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2174 // and fall through:
2175 case BoolTest::lt: // x < y
2176 case BoolTest::le: // x <= y
2177 answer_if_true = (want_max ? yvalue : xvalue);
2178 answer_if_false = (want_max ? xvalue : yvalue);
2179 break;
2180 case BoolTest::gt: // x > y
2181 case BoolTest::ge: // x >= y
2182 answer_if_true = (want_max ? xvalue : yvalue);
2183 answer_if_false = (want_max ? yvalue : xvalue);
2184 break;
2185 }
2186
2187 jint hi, lo;
2188 if (want_max) {
2189 // We can sharpen the minimum.
2190 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2191 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2192 } else {
2193 // We can sharpen the maximum.
2194 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2195 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2196 }
2197
2198 // Use a flow-free graph structure, to avoid creating excess control edges
2199 // which could hinder other optimizations.
2200 // Since Math.min/max is often used with arraycopy, we want
2201 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2202 Node* cmov = CMoveNode::make(NULL, best_bol,
2203 answer_if_false, answer_if_true,
2204 TypeInt::make(lo, hi, widen));
2205
2206 return _gvn.transform(cmov);
2207
2208 /*
2209 // This is not as desirable as it may seem, since Min and Max
2210 // nodes do not have a full set of optimizations.
2211 // And they would interfere, anyway, with 'if' optimizations
2212 // and with CMoveI canonical forms.
2213 switch (id) {
2214 case vmIntrinsics::_min:
2215 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2216 case vmIntrinsics::_max:
2217 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2218 default:
2219 ShouldNotReachHere();
2220 }
2221 */
2222 }
2223
2224 inline int
2225 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2226 const TypePtr* base_type = TypePtr::NULL_PTR;
2227 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2228 if (base_type == NULL) {
2229 // Unknown type.
2230 return Type::AnyPtr;
2231 } else if (base_type == TypePtr::NULL_PTR) {
2232 // Since this is a NULL+long form, we have to switch to a rawptr.
2233 base = _gvn.transform(new CastX2PNode(offset));
2234 offset = MakeConX(0);
2235 return Type::RawPtr;
2236 } else if (base_type->base() == Type::RawPtr) {
2237 return Type::RawPtr;
2238 } else if (base_type->isa_oopptr()) {
2239 // Base is never null => always a heap address.
2240 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2241 return Type::OopPtr;
2242 }
2243 // Offset is small => always a heap address.
2244 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2245 if (offset_type != NULL &&
2246 base_type->offset() == 0 && // (should always be?)
2247 offset_type->_lo >= 0 &&
2248 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2249 return Type::OopPtr;
2250 } else if (type == T_OBJECT) {
2251 // off heap access to an oop doesn't make any sense. Has to be on
2252 // heap.
2253 return Type::OopPtr;
2254 }
2255 // Otherwise, it might either be oop+off or NULL+addr.
2256 return Type::AnyPtr;
2257 } else {
2258 // No information:
2259 return Type::AnyPtr;
2260 }
2261 }
2262
2263 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) {
2264 Node* uncasted_base = base;
2265 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2266 if (kind == Type::RawPtr) {
2267 return basic_plus_adr(top(), uncasted_base, offset);
2268 } else if (kind == Type::AnyPtr) {
2269 assert(base == uncasted_base, "unexpected base change");
2270 if (can_cast) {
2271 if (!_gvn.type(base)->speculative_maybe_null() &&
2272 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2273 // According to profiling, this access is always on
2274 // heap. Casting the base to not null and thus avoiding membars
2275 // around the access should allow better optimizations
2276 Node* null_ctl = top();
2277 base = null_check_oop(base, &null_ctl, true, true, true);
2278 assert(null_ctl->is_top(), "no null control here");
2279 return basic_plus_adr(base, offset);
2280 } else if (_gvn.type(base)->speculative_always_null() &&
2281 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2282 // According to profiling, this access is always off
2283 // heap.
2284 base = null_assert(base);
2285 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2286 offset = MakeConX(0);
2287 return basic_plus_adr(top(), raw_base, offset);
2288 }
2289 }
2290 // We don't know if it's an on heap or off heap access. Fall back
2291 // to raw memory access.
2292 base = access_resolve(base, decorators);
2293 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2294 return basic_plus_adr(top(), raw, offset);
2295 } else {
2296 assert(base == uncasted_base, "unexpected base change");
2297 // We know it's an on heap access so base can't be null
2298 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2299 base = must_be_not_null(base, true);
2300 }
2301 return basic_plus_adr(base, offset);
2302 }
2303 }
2304
2305 //--------------------------inline_number_methods-----------------------------
2306 // inline int Integer.numberOfLeadingZeros(int)
2307 // inline int Long.numberOfLeadingZeros(long)
2308 //
2309 // inline int Integer.numberOfTrailingZeros(int)
2310 // inline int Long.numberOfTrailingZeros(long)
2311 //
2312 // inline int Integer.bitCount(int)
2313 // inline int Long.bitCount(long)
2314 //
2315 // inline char Character.reverseBytes(char)
2316 // inline short Short.reverseBytes(short)
2317 // inline int Integer.reverseBytes(int)
2318 // inline long Long.reverseBytes(long)
2319 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2320 Node* arg = argument(0);
2321 Node* n = NULL;
2322 switch (id) {
2323 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2324 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2325 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2326 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2327 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2328 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2329 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2330 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2331 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2332 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2333 default: fatal_unexpected_iid(id); break;
2334 }
2335 set_result(_gvn.transform(n));
2336 return true;
2337 }
2338
2339 //----------------------------inline_unsafe_access----------------------------
2340
2341 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2342 // Attempt to infer a sharper value type from the offset and base type.
2343 ciKlass* sharpened_klass = NULL;
2344
2345 // See if it is an instance field, with an object type.
2346 if (alias_type->field() != NULL) {
2347 if (alias_type->field()->type()->is_klass()) {
2348 sharpened_klass = alias_type->field()->type()->as_klass();
2349 }
2350 }
2351
2352 // See if it is a narrow oop array.
2353 if (adr_type->isa_aryptr()) {
2354 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2355 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2356 if (elem_type != NULL) {
2357 sharpened_klass = elem_type->klass();
2358 }
2359 }
2360 }
2361
2362 // The sharpened class might be unloaded if there is no class loader
2363 // contraint in place.
2364 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2365 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2366
2367 #ifndef PRODUCT
2368 if (C->print_intrinsics() || C->print_inlining()) {
2369 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2370 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2371 }
2372 #endif
2373 // Sharpen the value type.
2374 return tjp;
2375 }
2376 return NULL;
2377 }
2378
2379 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) {
2380 switch (kind) {
2381 case Relaxed:
2382 return MO_UNORDERED;
2383 case Opaque:
2384 return MO_RELAXED;
2385 case Acquire:
2386 return MO_ACQUIRE;
2387 case Release:
2388 return MO_RELEASE;
2389 case Volatile:
2390 return MO_SEQ_CST;
2391 default:
2392 ShouldNotReachHere();
2393 return 0;
2394 }
2395 }
2396
2397 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2398 if (callee()->is_static()) return false; // caller must have the capability!
2399 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2400 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2401 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2402 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2403
2404 if (type == T_OBJECT || type == T_ARRAY) {
2405 decorators |= ON_UNKNOWN_OOP_REF;
2406 }
2407
2408 if (unaligned) {
2409 decorators |= C2_UNALIGNED;
2410 }
2411
2412 #ifndef PRODUCT
2413 {
2414 ResourceMark rm;
2415 // Check the signatures.
2416 ciSignature* sig = callee()->signature();
2417 #ifdef ASSERT
2418 if (!is_store) {
2419 // Object getReference(Object base, int/long offset), etc.
2420 BasicType rtype = sig->return_type()->basic_type();
2421 assert(rtype == type, "getter must return the expected value");
2422 assert(sig->count() == 2, "oop getter has 2 arguments");
2423 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2424 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2425 } else {
2426 // void putReference(Object base, int/long offset, Object x), etc.
2427 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2428 assert(sig->count() == 3, "oop putter has 3 arguments");
2429 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2430 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2431 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2432 assert(vtype == type, "putter must accept the expected value");
2433 }
2434 #endif // ASSERT
2435 }
2436 #endif //PRODUCT
2437
2438 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2439
2440 Node* receiver = argument(0); // type: oop
2441
2442 // Build address expression.
2443 Node* adr;
2444 Node* heap_base_oop = top();
2445 Node* offset = top();
2446 Node* val;
2447
2448 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2449 Node* base = argument(1); // type: oop
2450 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2451 offset = argument(2); // type: long
2452 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2453 // to be plain byte offsets, which are also the same as those accepted
2454 // by oopDesc::field_addr.
2455 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2456 "fieldOffset must be byte-scaled");
2457 // 32-bit machines ignore the high half!
2458 offset = ConvL2X(offset);
2459 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed);
2460
2461 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2462 heap_base_oop = base;
2463 } else if (type == T_OBJECT) {
2464 return false; // off-heap oop accesses are not supported
2465 }
2466
2467 // Can base be NULL? Otherwise, always on-heap access.
2468 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base));
2469
2470 if (!can_access_non_heap) {
2471 decorators |= IN_HEAP;
2472 }
2473
2474 val = is_store ? argument(4) : NULL;
2475
2476 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2477
2478 // Try to categorize the address.
2479 Compile::AliasType* alias_type = C->alias_type(adr_type);
2480 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2481
2482 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2483 alias_type->adr_type() == TypeAryPtr::RANGE) {
2484 return false; // not supported
2485 }
2486
2487 bool mismatched = false;
2488 BasicType bt = alias_type->basic_type();
2489 if (bt != T_ILLEGAL) {
2490 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2491 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2492 // Alias type doesn't differentiate between byte[] and boolean[]).
2493 // Use address type to get the element type.
2494 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2495 }
2496 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2497 // accessing an array field with getReference is not a mismatch
2498 bt = T_OBJECT;
2499 }
2500 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2501 // Don't intrinsify mismatched object accesses
2502 return false;
2503 }
2504 mismatched = (bt != type);
2505 } else if (alias_type->adr_type()->isa_oopptr()) {
2506 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2507 }
2508
2509 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2510
2511 if (mismatched) {
2512 decorators |= C2_MISMATCHED;
2513 }
2514
2515 // First guess at the value type.
2516 const Type *value_type = Type::get_const_basic_type(type);
2517
2518 // Figure out the memory ordering.
2519 decorators |= mo_decorator_for_access_kind(kind);
2520
2521 if (!is_store && type == T_OBJECT) {
2522 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2523 if (tjp != NULL) {
2524 value_type = tjp;
2525 }
2526 }
2527
2528 receiver = null_check(receiver);
2529 if (stopped()) {
2530 return true;
2531 }
2532 // Heap pointers get a null-check from the interpreter,
2533 // as a courtesy. However, this is not guaranteed by Unsafe,
2534 // and it is not possible to fully distinguish unintended nulls
2535 // from intended ones in this API.
2536
2537 if (!is_store) {
2538 Node* p = NULL;
2539 // Try to constant fold a load from a constant field
2540 ciField* field = alias_type->field();
2541 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2542 // final or stable field
2543 p = make_constant_from_field(field, heap_base_oop);
2544 }
2545
2546 if (p == NULL) { // Could not constant fold the load
2547 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators);
2548 // Normalize the value returned by getBoolean in the following cases
2549 if (type == T_BOOLEAN &&
2550 (mismatched ||
2551 heap_base_oop == top() || // - heap_base_oop is NULL or
2552 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL
2553 // and the unsafe access is made to large offset
2554 // (i.e., larger than the maximum offset necessary for any
2555 // field access)
2556 ) {
2557 IdealKit ideal = IdealKit(this);
2558 #define __ ideal.
2559 IdealVariable normalized_result(ideal);
2560 __ declarations_done();
2561 __ set(normalized_result, p);
2562 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2563 __ set(normalized_result, ideal.ConI(1));
2564 ideal.end_if();
2565 final_sync(ideal);
2566 p = __ value(normalized_result);
2567 #undef __
2568 }
2569 }
2570 if (type == T_ADDRESS) {
2571 p = gvn().transform(new CastP2XNode(NULL, p));
2572 p = ConvX2UL(p);
2573 }
2574 // The load node has the control of the preceding MemBarCPUOrder. All
2575 // following nodes will have the control of the MemBarCPUOrder inserted at
2576 // the end of this method. So, pushing the load onto the stack at a later
2577 // point is fine.
2578 set_result(p);
2579 } else {
2580 if (bt == T_ADDRESS) {
2581 // Repackage the long as a pointer.
2582 val = ConvL2X(val);
2583 val = gvn().transform(new CastX2PNode(val));
2584 }
2585 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators);
2586 }
2587
2588 return true;
2589 }
2590
2591 //----------------------------inline_unsafe_load_store----------------------------
2592 // This method serves a couple of different customers (depending on LoadStoreKind):
2593 //
2594 // LS_cmp_swap:
2595 //
2596 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x);
2597 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2598 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2599 //
2600 // LS_cmp_swap_weak:
2601 //
2602 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x);
2603 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x);
2604 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x);
2605 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x);
2606 //
2607 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2608 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2609 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2610 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2611 //
2612 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2613 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2614 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2615 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2616 //
2617 // LS_cmp_exchange:
2618 //
2619 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x);
2620 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x);
2621 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x);
2622 //
2623 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2624 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2625 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2626 //
2627 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2628 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2629 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2630 //
2631 // LS_get_add:
2632 //
2633 // int getAndAddInt( Object o, long offset, int delta)
2634 // long getAndAddLong(Object o, long offset, long delta)
2635 //
2636 // LS_get_set:
2637 //
2638 // int getAndSet(Object o, long offset, int newValue)
2639 // long getAndSet(Object o, long offset, long newValue)
2640 // Object getAndSet(Object o, long offset, Object newValue)
2641 //
2642 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2643 // This basic scheme here is the same as inline_unsafe_access, but
2644 // differs in enough details that combining them would make the code
2645 // overly confusing. (This is a true fact! I originally combined
2646 // them, but even I was confused by it!) As much code/comments as
2647 // possible are retained from inline_unsafe_access though to make
2648 // the correspondences clearer. - dl
2649
2650 if (callee()->is_static()) return false; // caller must have the capability!
2651
2652 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2653 decorators |= mo_decorator_for_access_kind(access_kind);
2654
2655 #ifndef PRODUCT
2656 BasicType rtype;
2657 {
2658 ResourceMark rm;
2659 // Check the signatures.
2660 ciSignature* sig = callee()->signature();
2661 rtype = sig->return_type()->basic_type();
2662 switch(kind) {
2663 case LS_get_add:
2664 case LS_get_set: {
2665 // Check the signatures.
2666 #ifdef ASSERT
2667 assert(rtype == type, "get and set must return the expected type");
2668 assert(sig->count() == 3, "get and set has 3 arguments");
2669 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2670 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2671 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2672 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2673 #endif // ASSERT
2674 break;
2675 }
2676 case LS_cmp_swap:
2677 case LS_cmp_swap_weak: {
2678 // Check the signatures.
2679 #ifdef ASSERT
2680 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2681 assert(sig->count() == 4, "CAS has 4 arguments");
2682 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2683 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2684 #endif // ASSERT
2685 break;
2686 }
2687 case LS_cmp_exchange: {
2688 // Check the signatures.
2689 #ifdef ASSERT
2690 assert(rtype == type, "CAS must return the expected type");
2691 assert(sig->count() == 4, "CAS has 4 arguments");
2692 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2693 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2694 #endif // ASSERT
2695 break;
2696 }
2697 default:
2698 ShouldNotReachHere();
2699 }
2700 }
2701 #endif //PRODUCT
2702
2703 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2704
2705 // Get arguments:
2706 Node* receiver = NULL;
2707 Node* base = NULL;
2708 Node* offset = NULL;
2709 Node* oldval = NULL;
2710 Node* newval = NULL;
2711 switch(kind) {
2712 case LS_cmp_swap:
2713 case LS_cmp_swap_weak:
2714 case LS_cmp_exchange: {
2715 const bool two_slot_type = type2size[type] == 2;
2716 receiver = argument(0); // type: oop
2717 base = argument(1); // type: oop
2718 offset = argument(2); // type: long
2719 oldval = argument(4); // type: oop, int, or long
2720 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2721 break;
2722 }
2723 case LS_get_add:
2724 case LS_get_set: {
2725 receiver = argument(0); // type: oop
2726 base = argument(1); // type: oop
2727 offset = argument(2); // type: long
2728 oldval = NULL;
2729 newval = argument(4); // type: oop, int, or long
2730 break;
2731 }
2732 default:
2733 ShouldNotReachHere();
2734 }
2735
2736 // Build field offset expression.
2737 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2738 // to be plain byte offsets, which are also the same as those accepted
2739 // by oopDesc::field_addr.
2740 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2741 // 32-bit machines ignore the high half of long offsets
2742 offset = ConvL2X(offset);
2743 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false);
2744 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2745
2746 Compile::AliasType* alias_type = C->alias_type(adr_type);
2747 BasicType bt = alias_type->basic_type();
2748 if (bt != T_ILLEGAL &&
2749 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2750 // Don't intrinsify mismatched object accesses.
2751 return false;
2752 }
2753
2754 // For CAS, unlike inline_unsafe_access, there seems no point in
2755 // trying to refine types. Just use the coarse types here.
2756 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2757 const Type *value_type = Type::get_const_basic_type(type);
2758
2759 switch (kind) {
2760 case LS_get_set:
2761 case LS_cmp_exchange: {
2762 if (type == T_OBJECT) {
2763 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2764 if (tjp != NULL) {
2765 value_type = tjp;
2766 }
2767 }
2768 break;
2769 }
2770 case LS_cmp_swap:
2771 case LS_cmp_swap_weak:
2772 case LS_get_add:
2773 break;
2774 default:
2775 ShouldNotReachHere();
2776 }
2777
2778 // Null check receiver.
2779 receiver = null_check(receiver);
2780 if (stopped()) {
2781 return true;
2782 }
2783
2784 int alias_idx = C->get_alias_index(adr_type);
2785
2786 if (type == T_OBJECT || type == T_ARRAY) {
2787 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF;
2788
2789 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2790 // could be delayed during Parse (for example, in adjust_map_after_if()).
2791 // Execute transformation here to avoid barrier generation in such case.
2792 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2793 newval = _gvn.makecon(TypePtr::NULL_PTR);
2794
2795 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) {
2796 // Refine the value to a null constant, when it is known to be null
2797 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2798 }
2799 }
2800
2801 Node* result = NULL;
2802 switch (kind) {
2803 case LS_cmp_exchange: {
2804 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx,
2805 oldval, newval, value_type, type, decorators);
2806 break;
2807 }
2808 case LS_cmp_swap_weak:
2809 decorators |= C2_WEAK_CMPXCHG;
2810 case LS_cmp_swap: {
2811 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx,
2812 oldval, newval, value_type, type, decorators);
2813 break;
2814 }
2815 case LS_get_set: {
2816 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx,
2817 newval, value_type, type, decorators);
2818 break;
2819 }
2820 case LS_get_add: {
2821 result = access_atomic_add_at(base, adr, adr_type, alias_idx,
2822 newval, value_type, type, decorators);
2823 break;
2824 }
2825 default:
2826 ShouldNotReachHere();
2827 }
2828
2829 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2830 set_result(result);
2831 return true;
2832 }
2833
2834 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2835 // Regardless of form, don't allow previous ld/st to move down,
2836 // then issue acquire, release, or volatile mem_bar.
2837 insert_mem_bar(Op_MemBarCPUOrder);
2838 switch(id) {
2839 case vmIntrinsics::_loadFence:
2840 insert_mem_bar(Op_LoadFence);
2841 return true;
2842 case vmIntrinsics::_storeFence:
2843 insert_mem_bar(Op_StoreFence);
2844 return true;
2845 case vmIntrinsics::_fullFence:
2846 insert_mem_bar(Op_MemBarVolatile);
2847 return true;
2848 default:
2849 fatal_unexpected_iid(id);
2850 return false;
2851 }
2852 }
2853
2854 bool LibraryCallKit::inline_onspinwait() {
2855 insert_mem_bar(Op_OnSpinWait);
2856 return true;
2857 }
2858
2859 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2860 if (!kls->is_Con()) {
2861 return true;
2862 }
2863 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2864 if (klsptr == NULL) {
2865 return true;
2866 }
2867 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2868 // don't need a guard for a klass that is already initialized
2869 return !ik->is_initialized();
2870 }
2871
2872 //----------------------------inline_unsafe_allocate---------------------------
2873 // public native Object Unsafe.allocateInstance(Class<?> cls);
2874 bool LibraryCallKit::inline_unsafe_allocate() {
2875 if (callee()->is_static()) return false; // caller must have the capability!
2876
2877 null_check_receiver(); // null-check, then ignore
2878 Node* cls = null_check(argument(1));
2879 if (stopped()) return true;
2880
2881 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2882 kls = null_check(kls);
2883 if (stopped()) return true; // argument was like int.class
2884
2885 Node* test = NULL;
2886 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2887 // Note: The argument might still be an illegal value like
2888 // Serializable.class or Object[].class. The runtime will handle it.
2889 // But we must make an explicit check for initialization.
2890 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2891 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2892 // can generate code to load it as unsigned byte.
2893 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2894 Node* bits = intcon(InstanceKlass::fully_initialized);
2895 test = _gvn.transform(new SubINode(inst, bits));
2896 // The 'test' is non-zero if we need to take a slow path.
2897 }
2898
2899 Node* obj = new_instance(kls, test);
2900 set_result(obj);
2901 return true;
2902 }
2903
2904 //------------------------inline_native_time_funcs--------------
2905 // inline code for System.currentTimeMillis() and System.nanoTime()
2906 // these have the same type and signature
2907 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2908 const TypeFunc* tf = OptoRuntime::void_long_Type();
2909 const TypePtr* no_memory_effects = NULL;
2910 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2911 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2912 #ifdef ASSERT
2913 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2914 assert(value_top == top(), "second value must be top");
2915 #endif
2916 set_result(value);
2917 return true;
2918 }
2919
2920 #ifdef JFR_HAVE_INTRINSICS
2921
2922 /*
2923 * oop -> myklass
2924 * myklass->trace_id |= USED
2925 * return myklass->trace_id & ~0x3
2926 */
2927 bool LibraryCallKit::inline_native_classID() {
2928 Node* cls = null_check(argument(0), T_OBJECT);
2929 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2930 kls = null_check(kls, T_OBJECT);
2931
2932 ByteSize offset = KLASS_TRACE_ID_OFFSET;
2933 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2934 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2935
2936 Node* clsused = longcon(0x01l); // set the class bit
2937 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2938 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2939 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2940
2941 #ifdef TRACE_ID_META_BITS
2942 Node* mbits = longcon(~TRACE_ID_META_BITS);
2943 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
2944 #endif
2945 #ifdef TRACE_ID_SHIFT
2946 Node* cbits = intcon(TRACE_ID_SHIFT);
2947 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
2948 #endif
2949
2950 set_result(tvalue);
2951 return true;
2952
2953 }
2954
2955 bool LibraryCallKit::inline_native_getEventWriter() {
2956 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
2957
2958 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
2959 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR));
2960
2961 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
2962
2963 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
2964 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
2965
2966 IfNode* iff_jobj_null =
2967 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
2968
2969 enum { _normal_path = 1,
2970 _null_path = 2,
2971 PATH_LIMIT };
2972
2973 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
2974 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM);
2975
2976 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
2977 result_rgn->init_req(_null_path, jobj_is_null);
2978 result_val->init_req(_null_path, null());
2979
2980 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
2981 set_control(jobj_is_not_null);
2982 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT,
2983 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD);
2984 result_rgn->init_req(_normal_path, control());
2985 result_val->init_req(_normal_path, res);
2986
2987 set_result(result_rgn, result_val);
2988
2989 return true;
2990 }
2991
2992 #endif // JFR_HAVE_INTRINSICS
2993
2994 //------------------------inline_native_currentThread------------------
2995 bool LibraryCallKit::inline_native_currentThread() {
2996 Node* junk = NULL;
2997 set_result(generate_current_thread(junk));
2998 return true;
2999 }
3000
3001 //------------------------inline_native_isInterrupted------------------
3002 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3003 bool LibraryCallKit::inline_native_isInterrupted() {
3004 // Add a fast path to t.isInterrupted(clear_int):
3005 // (t == Thread.current() &&
3006 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3007 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3008 // So, in the common case that the interrupt bit is false,
3009 // we avoid making a call into the VM. Even if the interrupt bit
3010 // is true, if the clear_int argument is false, we avoid the VM call.
3011 // However, if the receiver is not currentThread, we must call the VM,
3012 // because there must be some locking done around the operation.
3013
3014 // We only go to the fast case code if we pass two guards.
3015 // Paths which do not pass are accumulated in the slow_region.
3016
3017 enum {
3018 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3019 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3020 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3021 PATH_LIMIT
3022 };
3023
3024 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3025 // out of the function.
3026 insert_mem_bar(Op_MemBarCPUOrder);
3027
3028 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3029 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3030
3031 RegionNode* slow_region = new RegionNode(1);
3032 record_for_igvn(slow_region);
3033
3034 // (a) Receiving thread must be the current thread.
3035 Node* rec_thr = argument(0);
3036 Node* tls_ptr = NULL;
3037 Node* cur_thr = generate_current_thread(tls_ptr);
3038
3039 // Resolve oops to stable for CmpP below.
3040 cur_thr = access_resolve(cur_thr, 0);
3041 rec_thr = access_resolve(rec_thr, 0);
3042
3043 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3044 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3045
3046 generate_slow_guard(bol_thr, slow_region);
3047
3048 // (b) Interrupt bit on TLS must be false.
3049 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3050 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3051 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3052
3053 // Set the control input on the field _interrupted read to prevent it floating up.
3054 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3055 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3056 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3057
3058 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3059
3060 // First fast path: if (!TLS._interrupted) return false;
3061 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3062 result_rgn->init_req(no_int_result_path, false_bit);
3063 result_val->init_req(no_int_result_path, intcon(0));
3064
3065 // drop through to next case
3066 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3067
3068 #ifndef _WINDOWS
3069 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3070 Node* clr_arg = argument(1);
3071 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3072 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3073 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3074
3075 // Second fast path: ... else if (!clear_int) return true;
3076 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3077 result_rgn->init_req(no_clear_result_path, false_arg);
3078 result_val->init_req(no_clear_result_path, intcon(1));
3079
3080 // drop through to next case
3081 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3082 #else
3083 // To return true on Windows you must read the _interrupted field
3084 // and check the event state i.e. take the slow path.
3085 #endif // _WINDOWS
3086
3087 // (d) Otherwise, go to the slow path.
3088 slow_region->add_req(control());
3089 set_control( _gvn.transform(slow_region));
3090
3091 if (stopped()) {
3092 // There is no slow path.
3093 result_rgn->init_req(slow_result_path, top());
3094 result_val->init_req(slow_result_path, top());
3095 } else {
3096 // non-virtual because it is a private non-static
3097 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3098
3099 Node* slow_val = set_results_for_java_call(slow_call);
3100 // this->control() comes from set_results_for_java_call
3101
3102 Node* fast_io = slow_call->in(TypeFunc::I_O);
3103 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3104
3105 // These two phis are pre-filled with copies of of the fast IO and Memory
3106 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3107 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3108
3109 result_rgn->init_req(slow_result_path, control());
3110 result_io ->init_req(slow_result_path, i_o());
3111 result_mem->init_req(slow_result_path, reset_memory());
3112 result_val->init_req(slow_result_path, slow_val);
3113
3114 set_all_memory(_gvn.transform(result_mem));
3115 set_i_o( _gvn.transform(result_io));
3116 }
3117
3118 C->set_has_split_ifs(true); // Has chance for split-if optimization
3119 set_result(result_rgn, result_val);
3120 return true;
3121 }
3122
3123 //---------------------------load_mirror_from_klass----------------------------
3124 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3125 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3126 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3127 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3128 // mirror = ((OopHandle)mirror)->resolve();
3129 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
3130 }
3131
3132 //-----------------------load_klass_from_mirror_common-------------------------
3133 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3134 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3135 // and branch to the given path on the region.
3136 // If never_see_null, take an uncommon trap on null, so we can optimistically
3137 // compile for the non-null case.
3138 // If the region is NULL, force never_see_null = true.
3139 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3140 bool never_see_null,
3141 RegionNode* region,
3142 int null_path,
3143 int offset) {
3144 if (region == NULL) never_see_null = true;
3145 Node* p = basic_plus_adr(mirror, offset);
3146 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3147 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3148 Node* null_ctl = top();
3149 kls = null_check_oop(kls, &null_ctl, never_see_null);
3150 if (region != NULL) {
3151 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3152 region->init_req(null_path, null_ctl);
3153 } else {
3154 assert(null_ctl == top(), "no loose ends");
3155 }
3156 return kls;
3157 }
3158
3159 //--------------------(inline_native_Class_query helpers)---------------------
3160 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3161 // Fall through if (mods & mask) == bits, take the guard otherwise.
3162 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3163 // Branch around if the given klass has the given modifier bit set.
3164 // Like generate_guard, adds a new path onto the region.
3165 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3166 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3167 Node* mask = intcon(modifier_mask);
3168 Node* bits = intcon(modifier_bits);
3169 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3170 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3171 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3172 return generate_fair_guard(bol, region);
3173 }
3174 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3175 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3176 }
3177
3178 //-------------------------inline_native_Class_query-------------------
3179 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3180 const Type* return_type = TypeInt::BOOL;
3181 Node* prim_return_value = top(); // what happens if it's a primitive class?
3182 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3183 bool expect_prim = false; // most of these guys expect to work on refs
3184
3185 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3186
3187 Node* mirror = argument(0);
3188 Node* obj = top();
3189
3190 switch (id) {
3191 case vmIntrinsics::_isInstance:
3192 // nothing is an instance of a primitive type
3193 prim_return_value = intcon(0);
3194 obj = argument(1);
3195 break;
3196 case vmIntrinsics::_getModifiers:
3197 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3198 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3199 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3200 break;
3201 case vmIntrinsics::_isInterface:
3202 prim_return_value = intcon(0);
3203 break;
3204 case vmIntrinsics::_isArray:
3205 prim_return_value = intcon(0);
3206 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3207 break;
3208 case vmIntrinsics::_isPrimitive:
3209 prim_return_value = intcon(1);
3210 expect_prim = true; // obviously
3211 break;
3212 case vmIntrinsics::_getSuperclass:
3213 prim_return_value = null();
3214 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3215 break;
3216 case vmIntrinsics::_getClassAccessFlags:
3217 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3218 return_type = TypeInt::INT; // not bool! 6297094
3219 break;
3220 default:
3221 fatal_unexpected_iid(id);
3222 break;
3223 }
3224
3225 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3226 if (mirror_con == NULL) return false; // cannot happen?
3227
3228 #ifndef PRODUCT
3229 if (C->print_intrinsics() || C->print_inlining()) {
3230 ciType* k = mirror_con->java_mirror_type();
3231 if (k) {
3232 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3233 k->print_name();
3234 tty->cr();
3235 }
3236 }
3237 #endif
3238
3239 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3240 RegionNode* region = new RegionNode(PATH_LIMIT);
3241 record_for_igvn(region);
3242 PhiNode* phi = new PhiNode(region, return_type);
3243
3244 // The mirror will never be null of Reflection.getClassAccessFlags, however
3245 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3246 // if it is. See bug 4774291.
3247
3248 // For Reflection.getClassAccessFlags(), the null check occurs in
3249 // the wrong place; see inline_unsafe_access(), above, for a similar
3250 // situation.
3251 mirror = null_check(mirror);
3252 // If mirror or obj is dead, only null-path is taken.
3253 if (stopped()) return true;
3254
3255 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3256
3257 // Now load the mirror's klass metaobject, and null-check it.
3258 // Side-effects region with the control path if the klass is null.
3259 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3260 // If kls is null, we have a primitive mirror.
3261 phi->init_req(_prim_path, prim_return_value);
3262 if (stopped()) { set_result(region, phi); return true; }
3263 bool safe_for_replace = (region->in(_prim_path) == top());
3264
3265 Node* p; // handy temp
3266 Node* null_ctl;
3267
3268 // Now that we have the non-null klass, we can perform the real query.
3269 // For constant classes, the query will constant-fold in LoadNode::Value.
3270 Node* query_value = top();
3271 switch (id) {
3272 case vmIntrinsics::_isInstance:
3273 // nothing is an instance of a primitive type
3274 query_value = gen_instanceof(obj, kls, safe_for_replace);
3275 break;
3276
3277 case vmIntrinsics::_getModifiers:
3278 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3279 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3280 break;
3281
3282 case vmIntrinsics::_isInterface:
3283 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3284 if (generate_interface_guard(kls, region) != NULL)
3285 // A guard was added. If the guard is taken, it was an interface.
3286 phi->add_req(intcon(1));
3287 // If we fall through, it's a plain class.
3288 query_value = intcon(0);
3289 break;
3290
3291 case vmIntrinsics::_isArray:
3292 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3293 if (generate_array_guard(kls, region) != NULL)
3294 // A guard was added. If the guard is taken, it was an array.
3295 phi->add_req(intcon(1));
3296 // If we fall through, it's a plain class.
3297 query_value = intcon(0);
3298 break;
3299
3300 case vmIntrinsics::_isPrimitive:
3301 query_value = intcon(0); // "normal" path produces false
3302 break;
3303
3304 case vmIntrinsics::_getSuperclass:
3305 // The rules here are somewhat unfortunate, but we can still do better
3306 // with random logic than with a JNI call.
3307 // Interfaces store null or Object as _super, but must report null.
3308 // Arrays store an intermediate super as _super, but must report Object.
3309 // Other types can report the actual _super.
3310 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3311 if (generate_interface_guard(kls, region) != NULL)
3312 // A guard was added. If the guard is taken, it was an interface.
3313 phi->add_req(null());
3314 if (generate_array_guard(kls, region) != NULL)
3315 // A guard was added. If the guard is taken, it was an array.
3316 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3317 // If we fall through, it's a plain class. Get its _super.
3318 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3319 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3320 null_ctl = top();
3321 kls = null_check_oop(kls, &null_ctl);
3322 if (null_ctl != top()) {
3323 // If the guard is taken, Object.superClass is null (both klass and mirror).
3324 region->add_req(null_ctl);
3325 phi ->add_req(null());
3326 }
3327 if (!stopped()) {
3328 query_value = load_mirror_from_klass(kls);
3329 }
3330 break;
3331
3332 case vmIntrinsics::_getClassAccessFlags:
3333 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3334 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3335 break;
3336
3337 default:
3338 fatal_unexpected_iid(id);
3339 break;
3340 }
3341
3342 // Fall-through is the normal case of a query to a real class.
3343 phi->init_req(1, query_value);
3344 region->init_req(1, control());
3345
3346 C->set_has_split_ifs(true); // Has chance for split-if optimization
3347 set_result(region, phi);
3348 return true;
3349 }
3350
3351 //-------------------------inline_Class_cast-------------------
3352 bool LibraryCallKit::inline_Class_cast() {
3353 Node* mirror = argument(0); // Class
3354 Node* obj = argument(1);
3355 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3356 if (mirror_con == NULL) {
3357 return false; // dead path (mirror->is_top()).
3358 }
3359 if (obj == NULL || obj->is_top()) {
3360 return false; // dead path
3361 }
3362 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3363
3364 // First, see if Class.cast() can be folded statically.
3365 // java_mirror_type() returns non-null for compile-time Class constants.
3366 ciType* tm = mirror_con->java_mirror_type();
3367 if (tm != NULL && tm->is_klass() &&
3368 tp != NULL && tp->klass() != NULL) {
3369 if (!tp->klass()->is_loaded()) {
3370 // Don't use intrinsic when class is not loaded.
3371 return false;
3372 } else {
3373 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3374 if (static_res == Compile::SSC_always_true) {
3375 // isInstance() is true - fold the code.
3376 set_result(obj);
3377 return true;
3378 } else if (static_res == Compile::SSC_always_false) {
3379 // Don't use intrinsic, have to throw ClassCastException.
3380 // If the reference is null, the non-intrinsic bytecode will
3381 // be optimized appropriately.
3382 return false;
3383 }
3384 }
3385 }
3386
3387 // Bailout intrinsic and do normal inlining if exception path is frequent.
3388 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3389 return false;
3390 }
3391
3392 // Generate dynamic checks.
3393 // Class.cast() is java implementation of _checkcast bytecode.
3394 // Do checkcast (Parse::do_checkcast()) optimizations here.
3395
3396 mirror = null_check(mirror);
3397 // If mirror is dead, only null-path is taken.
3398 if (stopped()) {
3399 return true;
3400 }
3401
3402 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3403 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3404 RegionNode* region = new RegionNode(PATH_LIMIT);
3405 record_for_igvn(region);
3406
3407 // Now load the mirror's klass metaobject, and null-check it.
3408 // If kls is null, we have a primitive mirror and
3409 // nothing is an instance of a primitive type.
3410 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3411
3412 Node* res = top();
3413 if (!stopped()) {
3414 Node* bad_type_ctrl = top();
3415 // Do checkcast optimizations.
3416 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3417 region->init_req(_bad_type_path, bad_type_ctrl);
3418 }
3419 if (region->in(_prim_path) != top() ||
3420 region->in(_bad_type_path) != top()) {
3421 // Let Interpreter throw ClassCastException.
3422 PreserveJVMState pjvms(this);
3423 set_control(_gvn.transform(region));
3424 uncommon_trap(Deoptimization::Reason_intrinsic,
3425 Deoptimization::Action_maybe_recompile);
3426 }
3427 if (!stopped()) {
3428 set_result(res);
3429 }
3430 return true;
3431 }
3432
3433
3434 //--------------------------inline_native_subtype_check------------------------
3435 // This intrinsic takes the JNI calls out of the heart of
3436 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3437 bool LibraryCallKit::inline_native_subtype_check() {
3438 // Pull both arguments off the stack.
3439 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3440 args[0] = argument(0);
3441 args[1] = argument(1);
3442 Node* klasses[2]; // corresponding Klasses: superk, subk
3443 klasses[0] = klasses[1] = top();
3444
3445 enum {
3446 // A full decision tree on {superc is prim, subc is prim}:
3447 _prim_0_path = 1, // {P,N} => false
3448 // {P,P} & superc!=subc => false
3449 _prim_same_path, // {P,P} & superc==subc => true
3450 _prim_1_path, // {N,P} => false
3451 _ref_subtype_path, // {N,N} & subtype check wins => true
3452 _both_ref_path, // {N,N} & subtype check loses => false
3453 PATH_LIMIT
3454 };
3455
3456 RegionNode* region = new RegionNode(PATH_LIMIT);
3457 Node* phi = new PhiNode(region, TypeInt::BOOL);
3458 record_for_igvn(region);
3459
3460 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3461 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3462 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3463
3464 // First null-check both mirrors and load each mirror's klass metaobject.
3465 int which_arg;
3466 for (which_arg = 0; which_arg <= 1; which_arg++) {
3467 Node* arg = args[which_arg];
3468 arg = null_check(arg);
3469 if (stopped()) break;
3470 args[which_arg] = arg;
3471
3472 Node* p = basic_plus_adr(arg, class_klass_offset);
3473 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3474 klasses[which_arg] = _gvn.transform(kls);
3475 }
3476
3477 // Resolve oops to stable for CmpP below.
3478 args[0] = access_resolve(args[0], 0);
3479 args[1] = access_resolve(args[1], 0);
3480
3481 // Having loaded both klasses, test each for null.
3482 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3483 for (which_arg = 0; which_arg <= 1; which_arg++) {
3484 Node* kls = klasses[which_arg];
3485 Node* null_ctl = top();
3486 kls = null_check_oop(kls, &null_ctl, never_see_null);
3487 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3488 region->init_req(prim_path, null_ctl);
3489 if (stopped()) break;
3490 klasses[which_arg] = kls;
3491 }
3492
3493 if (!stopped()) {
3494 // now we have two reference types, in klasses[0..1]
3495 Node* subk = klasses[1]; // the argument to isAssignableFrom
3496 Node* superk = klasses[0]; // the receiver
3497 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3498 // now we have a successful reference subtype check
3499 region->set_req(_ref_subtype_path, control());
3500 }
3501
3502 // If both operands are primitive (both klasses null), then
3503 // we must return true when they are identical primitives.
3504 // It is convenient to test this after the first null klass check.
3505 set_control(region->in(_prim_0_path)); // go back to first null check
3506 if (!stopped()) {
3507 // Since superc is primitive, make a guard for the superc==subc case.
3508 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3509 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3510 generate_guard(bol_eq, region, PROB_FAIR);
3511 if (region->req() == PATH_LIMIT+1) {
3512 // A guard was added. If the added guard is taken, superc==subc.
3513 region->swap_edges(PATH_LIMIT, _prim_same_path);
3514 region->del_req(PATH_LIMIT);
3515 }
3516 region->set_req(_prim_0_path, control()); // Not equal after all.
3517 }
3518
3519 // these are the only paths that produce 'true':
3520 phi->set_req(_prim_same_path, intcon(1));
3521 phi->set_req(_ref_subtype_path, intcon(1));
3522
3523 // pull together the cases:
3524 assert(region->req() == PATH_LIMIT, "sane region");
3525 for (uint i = 1; i < region->req(); i++) {
3526 Node* ctl = region->in(i);
3527 if (ctl == NULL || ctl == top()) {
3528 region->set_req(i, top());
3529 phi ->set_req(i, top());
3530 } else if (phi->in(i) == NULL) {
3531 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3532 }
3533 }
3534
3535 set_control(_gvn.transform(region));
3536 set_result(_gvn.transform(phi));
3537 return true;
3538 }
3539
3540 //---------------------generate_array_guard_common------------------------
3541 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3542 bool obj_array, bool not_array) {
3543
3544 if (stopped()) {
3545 return NULL;
3546 }
3547
3548 // If obj_array/non_array==false/false:
3549 // Branch around if the given klass is in fact an array (either obj or prim).
3550 // If obj_array/non_array==false/true:
3551 // Branch around if the given klass is not an array klass of any kind.
3552 // If obj_array/non_array==true/true:
3553 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3554 // If obj_array/non_array==true/false:
3555 // Branch around if the kls is an oop array (Object[] or subtype)
3556 //
3557 // Like generate_guard, adds a new path onto the region.
3558 jint layout_con = 0;
3559 Node* layout_val = get_layout_helper(kls, layout_con);
3560 if (layout_val == NULL) {
3561 bool query = (obj_array
3562 ? Klass::layout_helper_is_objArray(layout_con)
3563 : Klass::layout_helper_is_array(layout_con));
3564 if (query == not_array) {
3565 return NULL; // never a branch
3566 } else { // always a branch
3567 Node* always_branch = control();
3568 if (region != NULL)
3569 region->add_req(always_branch);
3570 set_control(top());
3571 return always_branch;
3572 }
3573 }
3574 // Now test the correct condition.
3575 jint nval = (obj_array
3576 ? (jint)(Klass::_lh_array_tag_type_value
3577 << Klass::_lh_array_tag_shift)
3578 : Klass::_lh_neutral_value);
3579 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3580 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3581 // invert the test if we are looking for a non-array
3582 if (not_array) btest = BoolTest(btest).negate();
3583 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3584 return generate_fair_guard(bol, region);
3585 }
3586
3587
3588 //-----------------------inline_native_newArray--------------------------
3589 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3590 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3591 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3592 Node* mirror;
3593 Node* count_val;
3594 if (uninitialized) {
3595 mirror = argument(1);
3596 count_val = argument(2);
3597 } else {
3598 mirror = argument(0);
3599 count_val = argument(1);
3600 }
3601
3602 mirror = null_check(mirror);
3603 // If mirror or obj is dead, only null-path is taken.
3604 if (stopped()) return true;
3605
3606 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3607 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3608 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3609 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3610 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3611
3612 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3613 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3614 result_reg, _slow_path);
3615 Node* normal_ctl = control();
3616 Node* no_array_ctl = result_reg->in(_slow_path);
3617
3618 // Generate code for the slow case. We make a call to newArray().
3619 set_control(no_array_ctl);
3620 if (!stopped()) {
3621 // Either the input type is void.class, or else the
3622 // array klass has not yet been cached. Either the
3623 // ensuing call will throw an exception, or else it
3624 // will cache the array klass for next time.
3625 PreserveJVMState pjvms(this);
3626 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3627 Node* slow_result = set_results_for_java_call(slow_call);
3628 // this->control() comes from set_results_for_java_call
3629 result_reg->set_req(_slow_path, control());
3630 result_val->set_req(_slow_path, slow_result);
3631 result_io ->set_req(_slow_path, i_o());
3632 result_mem->set_req(_slow_path, reset_memory());
3633 }
3634
3635 set_control(normal_ctl);
3636 if (!stopped()) {
3637 // Normal case: The array type has been cached in the java.lang.Class.
3638 // The following call works fine even if the array type is polymorphic.
3639 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3640 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3641 result_reg->init_req(_normal_path, control());
3642 result_val->init_req(_normal_path, obj);
3643 result_io ->init_req(_normal_path, i_o());
3644 result_mem->init_req(_normal_path, reset_memory());
3645
3646 if (uninitialized) {
3647 // Mark the allocation so that zeroing is skipped
3648 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3649 alloc->maybe_set_complete(&_gvn);
3650 }
3651 }
3652
3653 // Return the combined state.
3654 set_i_o( _gvn.transform(result_io) );
3655 set_all_memory( _gvn.transform(result_mem));
3656
3657 C->set_has_split_ifs(true); // Has chance for split-if optimization
3658 set_result(result_reg, result_val);
3659 return true;
3660 }
3661
3662 //----------------------inline_native_getLength--------------------------
3663 // public static native int java.lang.reflect.Array.getLength(Object array);
3664 bool LibraryCallKit::inline_native_getLength() {
3665 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3666
3667 Node* array = null_check(argument(0));
3668 // If array is dead, only null-path is taken.
3669 if (stopped()) return true;
3670
3671 // Deoptimize if it is a non-array.
3672 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3673
3674 if (non_array != NULL) {
3675 PreserveJVMState pjvms(this);
3676 set_control(non_array);
3677 uncommon_trap(Deoptimization::Reason_intrinsic,
3678 Deoptimization::Action_maybe_recompile);
3679 }
3680
3681 // If control is dead, only non-array-path is taken.
3682 if (stopped()) return true;
3683
3684 // The works fine even if the array type is polymorphic.
3685 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3686 Node* result = load_array_length(array);
3687
3688 C->set_has_split_ifs(true); // Has chance for split-if optimization
3689 set_result(result);
3690 return true;
3691 }
3692
3693 //------------------------inline_array_copyOf----------------------------
3694 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3695 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3696 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3697 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3698
3699 // Get the arguments.
3700 Node* original = argument(0);
3701 Node* start = is_copyOfRange? argument(1): intcon(0);
3702 Node* end = is_copyOfRange? argument(2): argument(1);
3703 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3704
3705 Node* newcopy = NULL;
3706
3707 // Set the original stack and the reexecute bit for the interpreter to reexecute
3708 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3709 { PreserveReexecuteState preexecs(this);
3710 jvms()->set_should_reexecute(true);
3711
3712 array_type_mirror = null_check(array_type_mirror);
3713 original = null_check(original);
3714
3715 // Check if a null path was taken unconditionally.
3716 if (stopped()) return true;
3717
3718 Node* orig_length = load_array_length(original);
3719
3720 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3721 klass_node = null_check(klass_node);
3722
3723 RegionNode* bailout = new RegionNode(1);
3724 record_for_igvn(bailout);
3725
3726 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3727 // Bail out if that is so.
3728 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3729 if (not_objArray != NULL) {
3730 // Improve the klass node's type from the new optimistic assumption:
3731 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3732 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3733 Node* cast = new CastPPNode(klass_node, akls);
3734 cast->init_req(0, control());
3735 klass_node = _gvn.transform(cast);
3736 }
3737
3738 // Bail out if either start or end is negative.
3739 generate_negative_guard(start, bailout, &start);
3740 generate_negative_guard(end, bailout, &end);
3741
3742 Node* length = end;
3743 if (_gvn.type(start) != TypeInt::ZERO) {
3744 length = _gvn.transform(new SubINode(end, start));
3745 }
3746
3747 // Bail out if length is negative.
3748 // Without this the new_array would throw
3749 // NegativeArraySizeException but IllegalArgumentException is what
3750 // should be thrown
3751 generate_negative_guard(length, bailout, &length);
3752
3753 if (bailout->req() > 1) {
3754 PreserveJVMState pjvms(this);
3755 set_control(_gvn.transform(bailout));
3756 uncommon_trap(Deoptimization::Reason_intrinsic,
3757 Deoptimization::Action_maybe_recompile);
3758 }
3759
3760 if (!stopped()) {
3761 // How many elements will we copy from the original?
3762 // The answer is MinI(orig_length - start, length).
3763 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3764 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3765
3766 original = access_resolve(original, ACCESS_READ);
3767
3768 // Generate a direct call to the right arraycopy function(s).
3769 // We know the copy is disjoint but we might not know if the
3770 // oop stores need checking.
3771 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3772 // This will fail a store-check if x contains any non-nulls.
3773
3774 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3775 // loads/stores but it is legal only if we're sure the
3776 // Arrays.copyOf would succeed. So we need all input arguments
3777 // to the copyOf to be validated, including that the copy to the
3778 // new array won't trigger an ArrayStoreException. That subtype
3779 // check can be optimized if we know something on the type of
3780 // the input array from type speculation.
3781 if (_gvn.type(klass_node)->singleton()) {
3782 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3783 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3784
3785 int test = C->static_subtype_check(superk, subk);
3786 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3787 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3788 if (t_original->speculative_type() != NULL) {
3789 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3790 }
3791 }
3792 }
3793
3794 bool validated = false;
3795 // Reason_class_check rather than Reason_intrinsic because we
3796 // want to intrinsify even if this traps.
3797 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3798 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3799 klass_node);
3800
3801 if (not_subtype_ctrl != top()) {
3802 PreserveJVMState pjvms(this);
3803 set_control(not_subtype_ctrl);
3804 uncommon_trap(Deoptimization::Reason_class_check,
3805 Deoptimization::Action_make_not_entrant);
3806 assert(stopped(), "Should be stopped");
3807 }
3808 validated = true;
3809 }
3810
3811 if (!stopped()) {
3812 newcopy = new_array(klass_node, length, 0); // no arguments to push
3813
3814 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
3815 load_object_klass(original), klass_node);
3816 if (!is_copyOfRange) {
3817 ac->set_copyof(validated);
3818 } else {
3819 ac->set_copyofrange(validated);
3820 }
3821 Node* n = _gvn.transform(ac);
3822 if (n == ac) {
3823 ac->connect_outputs(this);
3824 } else {
3825 assert(validated, "shouldn't transform if all arguments not validated");
3826 set_all_memory(n);
3827 }
3828 }
3829 }
3830 } // original reexecute is set back here
3831
3832 C->set_has_split_ifs(true); // Has chance for split-if optimization
3833 if (!stopped()) {
3834 set_result(newcopy);
3835 }
3836 return true;
3837 }
3838
3839
3840 //----------------------generate_virtual_guard---------------------------
3841 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3842 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3843 RegionNode* slow_region) {
3844 ciMethod* method = callee();
3845 int vtable_index = method->vtable_index();
3846 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3847 "bad index %d", vtable_index);
3848 // Get the Method* out of the appropriate vtable entry.
3849 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
3850 vtable_index*vtableEntry::size_in_bytes() +
3851 vtableEntry::method_offset_in_bytes();
3852 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3853 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3854
3855 // Compare the target method with the expected method (e.g., Object.hashCode).
3856 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3857
3858 Node* native_call = makecon(native_call_addr);
3859 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3860 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3861
3862 return generate_slow_guard(test_native, slow_region);
3863 }
3864
3865 //-----------------------generate_method_call----------------------------
3866 // Use generate_method_call to make a slow-call to the real
3867 // method if the fast path fails. An alternative would be to
3868 // use a stub like OptoRuntime::slow_arraycopy_Java.
3869 // This only works for expanding the current library call,
3870 // not another intrinsic. (E.g., don't use this for making an
3871 // arraycopy call inside of the copyOf intrinsic.)
3872 CallJavaNode*
3873 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3874 // When compiling the intrinsic method itself, do not use this technique.
3875 guarantee(callee() != C->method(), "cannot make slow-call to self");
3876
3877 ciMethod* method = callee();
3878 // ensure the JVMS we have will be correct for this call
3879 guarantee(method_id == method->intrinsic_id(), "must match");
3880
3881 const TypeFunc* tf = TypeFunc::make(method);
3882 CallJavaNode* slow_call;
3883 if (is_static) {
3884 assert(!is_virtual, "");
3885 slow_call = new CallStaticJavaNode(C, tf,
3886 SharedRuntime::get_resolve_static_call_stub(),
3887 method, bci());
3888 } else if (is_virtual) {
3889 null_check_receiver();
3890 int vtable_index = Method::invalid_vtable_index;
3891 if (UseInlineCaches) {
3892 // Suppress the vtable call
3893 } else {
3894 // hashCode and clone are not a miranda methods,
3895 // so the vtable index is fixed.
3896 // No need to use the linkResolver to get it.
3897 vtable_index = method->vtable_index();
3898 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3899 "bad index %d", vtable_index);
3900 }
3901 slow_call = new CallDynamicJavaNode(tf,
3902 SharedRuntime::get_resolve_virtual_call_stub(),
3903 method, vtable_index, bci());
3904 } else { // neither virtual nor static: opt_virtual
3905 null_check_receiver();
3906 slow_call = new CallStaticJavaNode(C, tf,
3907 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3908 method, bci());
3909 slow_call->set_optimized_virtual(true);
3910 }
3911 if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) {
3912 // To be able to issue a direct call (optimized virtual or virtual)
3913 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information
3914 // about the method being invoked should be attached to the call site to
3915 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C).
3916 slow_call->set_override_symbolic_info(true);
3917 }
3918 set_arguments_for_java_call(slow_call);
3919 set_edges_for_java_call(slow_call);
3920 return slow_call;
3921 }
3922
3923
3924 /**
3925 * Build special case code for calls to hashCode on an object. This call may
3926 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3927 * slightly different code.
3928 */
3929 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3930 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3931 assert(!(is_virtual && is_static), "either virtual, special, or static");
3932
3933 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3934
3935 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3936 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3937 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3938 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3939 Node* obj = NULL;
3940 if (!is_static) {
3941 // Check for hashing null object
3942 obj = null_check_receiver();
3943 if (stopped()) return true; // unconditionally null
3944 result_reg->init_req(_null_path, top());
3945 result_val->init_req(_null_path, top());
3946 } else {
3947 // Do a null check, and return zero if null.
3948 // System.identityHashCode(null) == 0
3949 obj = argument(0);
3950 Node* null_ctl = top();
3951 obj = null_check_oop(obj, &null_ctl);
3952 result_reg->init_req(_null_path, null_ctl);
3953 result_val->init_req(_null_path, _gvn.intcon(0));
3954 }
3955
3956 // Unconditionally null? Then return right away.
3957 if (stopped()) {
3958 set_control( result_reg->in(_null_path));
3959 if (!stopped())
3960 set_result(result_val->in(_null_path));
3961 return true;
3962 }
3963
3964 // We only go to the fast case code if we pass a number of guards. The
3965 // paths which do not pass are accumulated in the slow_region.
3966 RegionNode* slow_region = new RegionNode(1);
3967 record_for_igvn(slow_region);
3968
3969 // If this is a virtual call, we generate a funny guard. We pull out
3970 // the vtable entry corresponding to hashCode() from the target object.
3971 // If the target method which we are calling happens to be the native
3972 // Object hashCode() method, we pass the guard. We do not need this
3973 // guard for non-virtual calls -- the caller is known to be the native
3974 // Object hashCode().
3975 if (is_virtual) {
3976 // After null check, get the object's klass.
3977 Node* obj_klass = load_object_klass(obj);
3978 generate_virtual_guard(obj_klass, slow_region);
3979 }
3980
3981 // Get the header out of the object, use LoadMarkNode when available
3982 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3983 // The control of the load must be NULL. Otherwise, the load can move before
3984 // the null check after castPP removal.
3985 Node* no_ctrl = NULL;
3986 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3987
3988 // Test the header to see if it is unlocked.
3989 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3990 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3991 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3992 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3993 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3994
3995 generate_slow_guard(test_unlocked, slow_region);
3996
3997 // Get the hash value and check to see that it has been properly assigned.
3998 // We depend on hash_mask being at most 32 bits and avoid the use of
3999 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4000 // vm: see markOop.hpp.
4001 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4002 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4003 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4004 // This hack lets the hash bits live anywhere in the mark object now, as long
4005 // as the shift drops the relevant bits into the low 32 bits. Note that
4006 // Java spec says that HashCode is an int so there's no point in capturing
4007 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4008 hshifted_header = ConvX2I(hshifted_header);
4009 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4010
4011 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4012 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4013 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4014
4015 generate_slow_guard(test_assigned, slow_region);
4016
4017 Node* init_mem = reset_memory();
4018 // fill in the rest of the null path:
4019 result_io ->init_req(_null_path, i_o());
4020 result_mem->init_req(_null_path, init_mem);
4021
4022 result_val->init_req(_fast_path, hash_val);
4023 result_reg->init_req(_fast_path, control());
4024 result_io ->init_req(_fast_path, i_o());
4025 result_mem->init_req(_fast_path, init_mem);
4026
4027 // Generate code for the slow case. We make a call to hashCode().
4028 set_control(_gvn.transform(slow_region));
4029 if (!stopped()) {
4030 // No need for PreserveJVMState, because we're using up the present state.
4031 set_all_memory(init_mem);
4032 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4033 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4034 Node* slow_result = set_results_for_java_call(slow_call);
4035 // this->control() comes from set_results_for_java_call
4036 result_reg->init_req(_slow_path, control());
4037 result_val->init_req(_slow_path, slow_result);
4038 result_io ->set_req(_slow_path, i_o());
4039 result_mem ->set_req(_slow_path, reset_memory());
4040 }
4041
4042 // Return the combined state.
4043 set_i_o( _gvn.transform(result_io) );
4044 set_all_memory( _gvn.transform(result_mem));
4045
4046 set_result(result_reg, result_val);
4047 return true;
4048 }
4049
4050 //---------------------------inline_native_getClass----------------------------
4051 // public final native Class<?> java.lang.Object.getClass();
4052 //
4053 // Build special case code for calls to getClass on an object.
4054 bool LibraryCallKit::inline_native_getClass() {
4055 Node* obj = null_check_receiver();
4056 if (stopped()) return true;
4057 set_result(load_mirror_from_klass(load_object_klass(obj)));
4058 return true;
4059 }
4060
4061 //-----------------inline_native_Reflection_getCallerClass---------------------
4062 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4063 //
4064 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4065 //
4066 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4067 // in that it must skip particular security frames and checks for
4068 // caller sensitive methods.
4069 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4070 #ifndef PRODUCT
4071 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4072 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4073 }
4074 #endif
4075
4076 if (!jvms()->has_method()) {
4077 #ifndef PRODUCT
4078 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4079 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4080 }
4081 #endif
4082 return false;
4083 }
4084
4085 // Walk back up the JVM state to find the caller at the required
4086 // depth.
4087 JVMState* caller_jvms = jvms();
4088
4089 // Cf. JVM_GetCallerClass
4090 // NOTE: Start the loop at depth 1 because the current JVM state does
4091 // not include the Reflection.getCallerClass() frame.
4092 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4093 ciMethod* m = caller_jvms->method();
4094 switch (n) {
4095 case 0:
4096 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4097 break;
4098 case 1:
4099 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4100 if (!m->caller_sensitive()) {
4101 #ifndef PRODUCT
4102 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4103 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4104 }
4105 #endif
4106 return false; // bail-out; let JVM_GetCallerClass do the work
4107 }
4108 break;
4109 default:
4110 if (!m->is_ignored_by_security_stack_walk()) {
4111 // We have reached the desired frame; return the holder class.
4112 // Acquire method holder as java.lang.Class and push as constant.
4113 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4114 ciInstance* caller_mirror = caller_klass->java_mirror();
4115 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4116
4117 #ifndef PRODUCT
4118 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4119 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());
4120 tty->print_cr(" JVM state at this point:");
4121 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4122 ciMethod* m = jvms()->of_depth(i)->method();
4123 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4124 }
4125 }
4126 #endif
4127 return true;
4128 }
4129 break;
4130 }
4131 }
4132
4133 #ifndef PRODUCT
4134 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4135 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4136 tty->print_cr(" JVM state at this point:");
4137 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4138 ciMethod* m = jvms()->of_depth(i)->method();
4139 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4140 }
4141 }
4142 #endif
4143
4144 return false; // bail-out; let JVM_GetCallerClass do the work
4145 }
4146
4147 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4148 Node* arg = argument(0);
4149 Node* result = NULL;
4150
4151 switch (id) {
4152 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4153 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4154 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4155 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4156
4157 case vmIntrinsics::_doubleToLongBits: {
4158 // two paths (plus control) merge in a wood
4159 RegionNode *r = new RegionNode(3);
4160 Node *phi = new PhiNode(r, TypeLong::LONG);
4161
4162 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4163 // Build the boolean node
4164 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4165
4166 // Branch either way.
4167 // NaN case is less traveled, which makes all the difference.
4168 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4169 Node *opt_isnan = _gvn.transform(ifisnan);
4170 assert( opt_isnan->is_If(), "Expect an IfNode");
4171 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4172 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4173
4174 set_control(iftrue);
4175
4176 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4177 Node *slow_result = longcon(nan_bits); // return NaN
4178 phi->init_req(1, _gvn.transform( slow_result ));
4179 r->init_req(1, iftrue);
4180
4181 // Else fall through
4182 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4183 set_control(iffalse);
4184
4185 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4186 r->init_req(2, iffalse);
4187
4188 // Post merge
4189 set_control(_gvn.transform(r));
4190 record_for_igvn(r);
4191
4192 C->set_has_split_ifs(true); // Has chance for split-if optimization
4193 result = phi;
4194 assert(result->bottom_type()->isa_long(), "must be");
4195 break;
4196 }
4197
4198 case vmIntrinsics::_floatToIntBits: {
4199 // two paths (plus control) merge in a wood
4200 RegionNode *r = new RegionNode(3);
4201 Node *phi = new PhiNode(r, TypeInt::INT);
4202
4203 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4204 // Build the boolean node
4205 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4206
4207 // Branch either way.
4208 // NaN case is less traveled, which makes all the difference.
4209 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4210 Node *opt_isnan = _gvn.transform(ifisnan);
4211 assert( opt_isnan->is_If(), "Expect an IfNode");
4212 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4213 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4214
4215 set_control(iftrue);
4216
4217 static const jint nan_bits = 0x7fc00000;
4218 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4219 phi->init_req(1, _gvn.transform( slow_result ));
4220 r->init_req(1, iftrue);
4221
4222 // Else fall through
4223 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4224 set_control(iffalse);
4225
4226 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4227 r->init_req(2, iffalse);
4228
4229 // Post merge
4230 set_control(_gvn.transform(r));
4231 record_for_igvn(r);
4232
4233 C->set_has_split_ifs(true); // Has chance for split-if optimization
4234 result = phi;
4235 assert(result->bottom_type()->isa_int(), "must be");
4236 break;
4237 }
4238
4239 default:
4240 fatal_unexpected_iid(id);
4241 break;
4242 }
4243 set_result(_gvn.transform(result));
4244 return true;
4245 }
4246
4247 //----------------------inline_unsafe_copyMemory-------------------------
4248 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4249 bool LibraryCallKit::inline_unsafe_copyMemory() {
4250 if (callee()->is_static()) return false; // caller must have the capability!
4251 null_check_receiver(); // null-check receiver
4252 if (stopped()) return true;
4253
4254 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4255
4256 Node* src_ptr = argument(1); // type: oop
4257 Node* src_off = ConvL2X(argument(2)); // type: long
4258 Node* dst_ptr = argument(4); // type: oop
4259 Node* dst_off = ConvL2X(argument(5)); // type: long
4260 Node* size = ConvL2X(argument(7)); // type: long
4261
4262 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4263 "fieldOffset must be byte-scaled");
4264
4265 src_ptr = access_resolve(src_ptr, ACCESS_READ);
4266 dst_ptr = access_resolve(dst_ptr, ACCESS_WRITE);
4267 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ);
4268 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE);
4269
4270 // Conservatively insert a memory barrier on all memory slices.
4271 // Do not let writes of the copy source or destination float below the copy.
4272 insert_mem_bar(Op_MemBarCPUOrder);
4273
4274 // Call it. Note that the length argument is not scaled.
4275 make_runtime_call(RC_LEAF|RC_NO_FP,
4276 OptoRuntime::fast_arraycopy_Type(),
4277 StubRoutines::unsafe_arraycopy(),
4278 "unsafe_arraycopy",
4279 TypeRawPtr::BOTTOM,
4280 src, dst, size XTOP);
4281
4282 // Do not let reads of the copy destination float above the copy.
4283 insert_mem_bar(Op_MemBarCPUOrder);
4284
4285 return true;
4286 }
4287
4288 //------------------------clone_coping-----------------------------------
4289 // Helper function for inline_native_clone.
4290 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) {
4291 assert(obj_size != NULL, "");
4292 Node* raw_obj = alloc_obj->in(1);
4293 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4294
4295 AllocateNode* alloc = NULL;
4296 if (ReduceBulkZeroing) {
4297 // We will be completely responsible for initializing this object -
4298 // mark Initialize node as complete.
4299 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4300 // The object was just allocated - there should be no any stores!
4301 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4302 // Mark as complete_with_arraycopy so that on AllocateNode
4303 // expansion, we know this AllocateNode is initialized by an array
4304 // copy and a StoreStore barrier exists after the array copy.
4305 alloc->initialization()->set_complete_with_arraycopy();
4306 }
4307
4308 // Copy the fastest available way.
4309 // TODO: generate fields copies for small objects instead.
4310 Node* size = _gvn.transform(obj_size);
4311
4312 access_clone(obj, alloc_obj, size, is_array);
4313
4314 // Do not let reads from the cloned object float above the arraycopy.
4315 if (alloc != NULL) {
4316 // Do not let stores that initialize this object be reordered with
4317 // a subsequent store that would make this object accessible by
4318 // other threads.
4319 // Record what AllocateNode this StoreStore protects so that
4320 // escape analysis can go from the MemBarStoreStoreNode to the
4321 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4322 // based on the escape status of the AllocateNode.
4323 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4324 } else {
4325 insert_mem_bar(Op_MemBarCPUOrder);
4326 }
4327 }
4328
4329 //------------------------inline_native_clone----------------------------
4330 // protected native Object java.lang.Object.clone();
4331 //
4332 // Here are the simple edge cases:
4333 // null receiver => normal trap
4334 // virtual and clone was overridden => slow path to out-of-line clone
4335 // not cloneable or finalizer => slow path to out-of-line Object.clone
4336 //
4337 // The general case has two steps, allocation and copying.
4338 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4339 //
4340 // Copying also has two cases, oop arrays and everything else.
4341 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4342 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4343 //
4344 // These steps fold up nicely if and when the cloned object's klass
4345 // can be sharply typed as an object array, a type array, or an instance.
4346 //
4347 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4348 PhiNode* result_val;
4349
4350 // Set the reexecute bit for the interpreter to reexecute
4351 // the bytecode that invokes Object.clone if deoptimization happens.
4352 { PreserveReexecuteState preexecs(this);
4353 jvms()->set_should_reexecute(true);
4354
4355 Node* obj = null_check_receiver();
4356 if (stopped()) return true;
4357
4358 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4359
4360 // If we are going to clone an instance, we need its exact type to
4361 // know the number and types of fields to convert the clone to
4362 // loads/stores. Maybe a speculative type can help us.
4363 if (!obj_type->klass_is_exact() &&
4364 obj_type->speculative_type() != NULL &&
4365 obj_type->speculative_type()->is_instance_klass()) {
4366 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4367 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4368 !spec_ik->has_injected_fields()) {
4369 ciKlass* k = obj_type->klass();
4370 if (!k->is_instance_klass() ||
4371 k->as_instance_klass()->is_interface() ||
4372 k->as_instance_klass()->has_subklass()) {
4373 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4374 }
4375 }
4376 }
4377
4378 Node* obj_klass = load_object_klass(obj);
4379 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4380 const TypeOopPtr* toop = ((tklass != NULL)
4381 ? tklass->as_instance_type()
4382 : TypeInstPtr::NOTNULL);
4383
4384 // Conservatively insert a memory barrier on all memory slices.
4385 // Do not let writes into the original float below the clone.
4386 insert_mem_bar(Op_MemBarCPUOrder);
4387
4388 // paths into result_reg:
4389 enum {
4390 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4391 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4392 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4393 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4394 PATH_LIMIT
4395 };
4396 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4397 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4398 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4399 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4400 record_for_igvn(result_reg);
4401
4402 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4403 if (array_ctl != NULL) {
4404 // It's an array.
4405 PreserveJVMState pjvms(this);
4406 set_control(array_ctl);
4407 Node* obj_length = load_array_length(obj);
4408 Node* obj_size = NULL;
4409 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4410
4411 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4412 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) {
4413 // If it is an oop array, it requires very special treatment,
4414 // because gc barriers are required when accessing the array.
4415 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4416 if (is_obja != NULL) {
4417 PreserveJVMState pjvms2(this);
4418 set_control(is_obja);
4419 obj = access_resolve(obj, ACCESS_READ);
4420 // Generate a direct call to the right arraycopy function(s).
4421 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4422 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4423 ac->set_cloneoop();
4424 Node* n = _gvn.transform(ac);
4425 assert(n == ac, "cannot disappear");
4426 ac->connect_outputs(this);
4427
4428 result_reg->init_req(_objArray_path, control());
4429 result_val->init_req(_objArray_path, alloc_obj);
4430 result_i_o ->set_req(_objArray_path, i_o());
4431 result_mem ->set_req(_objArray_path, reset_memory());
4432 }
4433 }
4434 // Otherwise, there are no barriers to worry about.
4435 // (We can dispense with card marks if we know the allocation
4436 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4437 // causes the non-eden paths to take compensating steps to
4438 // simulate a fresh allocation, so that no further
4439 // card marks are required in compiled code to initialize
4440 // the object.)
4441
4442 if (!stopped()) {
4443 copy_to_clone(obj, alloc_obj, obj_size, true);
4444
4445 // Present the results of the copy.
4446 result_reg->init_req(_array_path, control());
4447 result_val->init_req(_array_path, alloc_obj);
4448 result_i_o ->set_req(_array_path, i_o());
4449 result_mem ->set_req(_array_path, reset_memory());
4450 }
4451 }
4452
4453 // We only go to the instance fast case code if we pass a number of guards.
4454 // The paths which do not pass are accumulated in the slow_region.
4455 RegionNode* slow_region = new RegionNode(1);
4456 record_for_igvn(slow_region);
4457 if (!stopped()) {
4458 // It's an instance (we did array above). Make the slow-path tests.
4459 // If this is a virtual call, we generate a funny guard. We grab
4460 // the vtable entry corresponding to clone() from the target object.
4461 // If the target method which we are calling happens to be the
4462 // Object clone() method, we pass the guard. We do not need this
4463 // guard for non-virtual calls; the caller is known to be the native
4464 // Object clone().
4465 if (is_virtual) {
4466 generate_virtual_guard(obj_klass, slow_region);
4467 }
4468
4469 // The object must be easily cloneable and must not have a finalizer.
4470 // Both of these conditions may be checked in a single test.
4471 // We could optimize the test further, but we don't care.
4472 generate_access_flags_guard(obj_klass,
4473 // Test both conditions:
4474 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4475 // Must be cloneable but not finalizer:
4476 JVM_ACC_IS_CLONEABLE_FAST,
4477 slow_region);
4478 }
4479
4480 if (!stopped()) {
4481 // It's an instance, and it passed the slow-path tests.
4482 PreserveJVMState pjvms(this);
4483 Node* obj_size = NULL;
4484 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4485 // is reexecuted if deoptimization occurs and there could be problems when merging
4486 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4487 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4488
4489 copy_to_clone(obj, alloc_obj, obj_size, false);
4490
4491 // Present the results of the slow call.
4492 result_reg->init_req(_instance_path, control());
4493 result_val->init_req(_instance_path, alloc_obj);
4494 result_i_o ->set_req(_instance_path, i_o());
4495 result_mem ->set_req(_instance_path, reset_memory());
4496 }
4497
4498 // Generate code for the slow case. We make a call to clone().
4499 set_control(_gvn.transform(slow_region));
4500 if (!stopped()) {
4501 PreserveJVMState pjvms(this);
4502 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4503 // We need to deoptimize on exception (see comment above)
4504 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true);
4505 // this->control() comes from set_results_for_java_call
4506 result_reg->init_req(_slow_path, control());
4507 result_val->init_req(_slow_path, slow_result);
4508 result_i_o ->set_req(_slow_path, i_o());
4509 result_mem ->set_req(_slow_path, reset_memory());
4510 }
4511
4512 // Return the combined state.
4513 set_control( _gvn.transform(result_reg));
4514 set_i_o( _gvn.transform(result_i_o));
4515 set_all_memory( _gvn.transform(result_mem));
4516 } // original reexecute is set back here
4517
4518 set_result(_gvn.transform(result_val));
4519 return true;
4520 }
4521
4522 // If we have a tightly coupled allocation, the arraycopy may take care
4523 // of the array initialization. If one of the guards we insert between
4524 // the allocation and the arraycopy causes a deoptimization, an
4525 // unitialized array will escape the compiled method. To prevent that
4526 // we set the JVM state for uncommon traps between the allocation and
4527 // the arraycopy to the state before the allocation so, in case of
4528 // deoptimization, we'll reexecute the allocation and the
4529 // initialization.
4530 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4531 if (alloc != NULL) {
4532 ciMethod* trap_method = alloc->jvms()->method();
4533 int trap_bci = alloc->jvms()->bci();
4534
4535 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4536 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4537 // Make sure there's no store between the allocation and the
4538 // arraycopy otherwise visible side effects could be rexecuted
4539 // in case of deoptimization and cause incorrect execution.
4540 bool no_interfering_store = true;
4541 Node* mem = alloc->in(TypeFunc::Memory);
4542 if (mem->is_MergeMem()) {
4543 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4544 Node* n = mms.memory();
4545 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4546 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?");
4547 no_interfering_store = false;
4548 break;
4549 }
4550 }
4551 } else {
4552 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4553 Node* n = mms.memory();
4554 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4555 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?");
4556 no_interfering_store = false;
4557 break;
4558 }
4559 }
4560 }
4561
4562 if (no_interfering_store) {
4563 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4564 uint size = alloc->req();
4565 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4566 old_jvms->set_map(sfpt);
4567 for (uint i = 0; i < size; i++) {
4568 sfpt->init_req(i, alloc->in(i));
4569 }
4570 // re-push array length for deoptimization
4571 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4572 old_jvms->set_sp(old_jvms->sp()+1);
4573 old_jvms->set_monoff(old_jvms->monoff()+1);
4574 old_jvms->set_scloff(old_jvms->scloff()+1);
4575 old_jvms->set_endoff(old_jvms->endoff()+1);
4576 old_jvms->set_should_reexecute(true);
4577
4578 sfpt->set_i_o(map()->i_o());
4579 sfpt->set_memory(map()->memory());
4580 sfpt->set_control(map()->control());
4581
4582 JVMState* saved_jvms = jvms();
4583 saved_reexecute_sp = _reexecute_sp;
4584
4585 set_jvms(sfpt->jvms());
4586 _reexecute_sp = jvms()->sp();
4587
4588 return saved_jvms;
4589 }
4590 }
4591 }
4592 return NULL;
4593 }
4594
4595 // In case of a deoptimization, we restart execution at the
4596 // allocation, allocating a new array. We would leave an uninitialized
4597 // array in the heap that GCs wouldn't expect. Move the allocation
4598 // after the traps so we don't allocate the array if we
4599 // deoptimize. This is possible because tightly_coupled_allocation()
4600 // guarantees there's no observer of the allocated array at this point
4601 // and the control flow is simple enough.
4602 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4603 int saved_reexecute_sp, uint new_idx) {
4604 if (saved_jvms != NULL && !stopped()) {
4605 assert(alloc != NULL, "only with a tightly coupled allocation");
4606 // restore JVM state to the state at the arraycopy
4607 saved_jvms->map()->set_control(map()->control());
4608 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4609 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4610 // If we've improved the types of some nodes (null check) while
4611 // emitting the guards, propagate them to the current state
4612 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4613 set_jvms(saved_jvms);
4614 _reexecute_sp = saved_reexecute_sp;
4615
4616 // Remove the allocation from above the guards
4617 CallProjections callprojs;
4618 alloc->extract_projections(&callprojs, true);
4619 InitializeNode* init = alloc->initialization();
4620 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4621 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4622 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4623 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4624
4625 // move the allocation here (after the guards)
4626 _gvn.hash_delete(alloc);
4627 alloc->set_req(TypeFunc::Control, control());
4628 alloc->set_req(TypeFunc::I_O, i_o());
4629 Node *mem = reset_memory();
4630 set_all_memory(mem);
4631 alloc->set_req(TypeFunc::Memory, mem);
4632 set_control(init->proj_out_or_null(TypeFunc::Control));
4633 set_i_o(callprojs.fallthrough_ioproj);
4634
4635 // Update memory as done in GraphKit::set_output_for_allocation()
4636 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4637 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4638 if (ary_type->isa_aryptr() && length_type != NULL) {
4639 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4640 }
4641 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4642 int elemidx = C->get_alias_index(telemref);
4643 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
4644 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
4645
4646 Node* allocx = _gvn.transform(alloc);
4647 assert(allocx == alloc, "where has the allocation gone?");
4648 assert(dest->is_CheckCastPP(), "not an allocation result?");
4649
4650 _gvn.hash_delete(dest);
4651 dest->set_req(0, control());
4652 Node* destx = _gvn.transform(dest);
4653 assert(destx == dest, "where has the allocation result gone?");
4654 }
4655 }
4656
4657
4658 //------------------------------inline_arraycopy-----------------------
4659 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4660 // Object dest, int destPos,
4661 // int length);
4662 bool LibraryCallKit::inline_arraycopy() {
4663 // Get the arguments.
4664 Node* src = argument(0); // type: oop
4665 Node* src_offset = argument(1); // type: int
4666 Node* dest = argument(2); // type: oop
4667 Node* dest_offset = argument(3); // type: int
4668 Node* length = argument(4); // type: int
4669
4670 uint new_idx = C->unique();
4671
4672 // Check for allocation before we add nodes that would confuse
4673 // tightly_coupled_allocation()
4674 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4675
4676 int saved_reexecute_sp = -1;
4677 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4678 // See arraycopy_restore_alloc_state() comment
4679 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4680 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4681 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
4682 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4683
4684 // The following tests must be performed
4685 // (1) src and dest are arrays.
4686 // (2) src and dest arrays must have elements of the same BasicType
4687 // (3) src and dest must not be null.
4688 // (4) src_offset must not be negative.
4689 // (5) dest_offset must not be negative.
4690 // (6) length must not be negative.
4691 // (7) src_offset + length must not exceed length of src.
4692 // (8) dest_offset + length must not exceed length of dest.
4693 // (9) each element of an oop array must be assignable
4694
4695 // (3) src and dest must not be null.
4696 // always do this here because we need the JVM state for uncommon traps
4697 Node* null_ctl = top();
4698 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4699 assert(null_ctl->is_top(), "no null control here");
4700 dest = null_check(dest, T_ARRAY);
4701
4702 if (!can_emit_guards) {
4703 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4704 // guards but the arraycopy node could still take advantage of a
4705 // tightly allocated allocation. tightly_coupled_allocation() is
4706 // called again to make sure it takes the null check above into
4707 // account: the null check is mandatory and if it caused an
4708 // uncommon trap to be emitted then the allocation can't be
4709 // considered tightly coupled in this context.
4710 alloc = tightly_coupled_allocation(dest, NULL);
4711 }
4712
4713 bool validated = false;
4714
4715 const Type* src_type = _gvn.type(src);
4716 const Type* dest_type = _gvn.type(dest);
4717 const TypeAryPtr* top_src = src_type->isa_aryptr();
4718 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4719
4720 // Do we have the type of src?
4721 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4722 // Do we have the type of dest?
4723 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4724 // Is the type for src from speculation?
4725 bool src_spec = false;
4726 // Is the type for dest from speculation?
4727 bool dest_spec = false;
4728
4729 if ((!has_src || !has_dest) && can_emit_guards) {
4730 // We don't have sufficient type information, let's see if
4731 // speculative types can help. We need to have types for both src
4732 // and dest so that it pays off.
4733
4734 // Do we already have or could we have type information for src
4735 bool could_have_src = has_src;
4736 // Do we already have or could we have type information for dest
4737 bool could_have_dest = has_dest;
4738
4739 ciKlass* src_k = NULL;
4740 if (!has_src) {
4741 src_k = src_type->speculative_type_not_null();
4742 if (src_k != NULL && src_k->is_array_klass()) {
4743 could_have_src = true;
4744 }
4745 }
4746
4747 ciKlass* dest_k = NULL;
4748 if (!has_dest) {
4749 dest_k = dest_type->speculative_type_not_null();
4750 if (dest_k != NULL && dest_k->is_array_klass()) {
4751 could_have_dest = true;
4752 }
4753 }
4754
4755 if (could_have_src && could_have_dest) {
4756 // This is going to pay off so emit the required guards
4757 if (!has_src) {
4758 src = maybe_cast_profiled_obj(src, src_k, true);
4759 src_type = _gvn.type(src);
4760 top_src = src_type->isa_aryptr();
4761 has_src = (top_src != NULL && top_src->klass() != NULL);
4762 src_spec = true;
4763 }
4764 if (!has_dest) {
4765 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4766 dest_type = _gvn.type(dest);
4767 top_dest = dest_type->isa_aryptr();
4768 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4769 dest_spec = true;
4770 }
4771 }
4772 }
4773
4774 if (has_src && has_dest && can_emit_guards) {
4775 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4776 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4777 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4778 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4779
4780 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4781 // If both arrays are object arrays then having the exact types
4782 // for both will remove the need for a subtype check at runtime
4783 // before the call and may make it possible to pick a faster copy
4784 // routine (without a subtype check on every element)
4785 // Do we have the exact type of src?
4786 bool could_have_src = src_spec;
4787 // Do we have the exact type of dest?
4788 bool could_have_dest = dest_spec;
4789 ciKlass* src_k = top_src->klass();
4790 ciKlass* dest_k = top_dest->klass();
4791 if (!src_spec) {
4792 src_k = src_type->speculative_type_not_null();
4793 if (src_k != NULL && src_k->is_array_klass()) {
4794 could_have_src = true;
4795 }
4796 }
4797 if (!dest_spec) {
4798 dest_k = dest_type->speculative_type_not_null();
4799 if (dest_k != NULL && dest_k->is_array_klass()) {
4800 could_have_dest = true;
4801 }
4802 }
4803 if (could_have_src && could_have_dest) {
4804 // If we can have both exact types, emit the missing guards
4805 if (could_have_src && !src_spec) {
4806 src = maybe_cast_profiled_obj(src, src_k, true);
4807 }
4808 if (could_have_dest && !dest_spec) {
4809 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4810 }
4811 }
4812 }
4813 }
4814
4815 ciMethod* trap_method = method();
4816 int trap_bci = bci();
4817 if (saved_jvms != NULL) {
4818 trap_method = alloc->jvms()->method();
4819 trap_bci = alloc->jvms()->bci();
4820 }
4821
4822 bool negative_length_guard_generated = false;
4823
4824 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4825 can_emit_guards &&
4826 !src->is_top() && !dest->is_top()) {
4827 // validate arguments: enables transformation the ArrayCopyNode
4828 validated = true;
4829
4830 RegionNode* slow_region = new RegionNode(1);
4831 record_for_igvn(slow_region);
4832
4833 // (1) src and dest are arrays.
4834 generate_non_array_guard(load_object_klass(src), slow_region);
4835 generate_non_array_guard(load_object_klass(dest), slow_region);
4836
4837 // (2) src and dest arrays must have elements of the same BasicType
4838 // done at macro expansion or at Ideal transformation time
4839
4840 // (4) src_offset must not be negative.
4841 generate_negative_guard(src_offset, slow_region);
4842
4843 // (5) dest_offset must not be negative.
4844 generate_negative_guard(dest_offset, slow_region);
4845
4846 // (7) src_offset + length must not exceed length of src.
4847 generate_limit_guard(src_offset, length,
4848 load_array_length(src),
4849 slow_region);
4850
4851 // (8) dest_offset + length must not exceed length of dest.
4852 generate_limit_guard(dest_offset, length,
4853 load_array_length(dest),
4854 slow_region);
4855
4856 // (6) length must not be negative.
4857 // This is also checked in generate_arraycopy() during macro expansion, but
4858 // we also have to check it here for the case where the ArrayCopyNode will
4859 // be eliminated by Escape Analysis.
4860 if (EliminateAllocations) {
4861 generate_negative_guard(length, slow_region);
4862 negative_length_guard_generated = true;
4863 }
4864
4865 // (9) each element of an oop array must be assignable
4866 Node* src_klass = load_object_klass(src);
4867 Node* dest_klass = load_object_klass(dest);
4868 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4869
4870 if (not_subtype_ctrl != top()) {
4871 PreserveJVMState pjvms(this);
4872 set_control(not_subtype_ctrl);
4873 uncommon_trap(Deoptimization::Reason_intrinsic,
4874 Deoptimization::Action_make_not_entrant);
4875 assert(stopped(), "Should be stopped");
4876 }
4877 {
4878 PreserveJVMState pjvms(this);
4879 set_control(_gvn.transform(slow_region));
4880 uncommon_trap(Deoptimization::Reason_intrinsic,
4881 Deoptimization::Action_make_not_entrant);
4882 assert(stopped(), "Should be stopped");
4883 }
4884
4885 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
4886 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
4887 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
4888 }
4889
4890 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
4891
4892 if (stopped()) {
4893 return true;
4894 }
4895
4896 Node* new_src = access_resolve(src, ACCESS_READ);
4897 Node* new_dest = access_resolve(dest, ACCESS_WRITE);
4898
4899 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, new_src, src_offset, new_dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
4900 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4901 // so the compiler has a chance to eliminate them: during macro expansion,
4902 // we have to set their control (CastPP nodes are eliminated).
4903 load_object_klass(src), load_object_klass(dest),
4904 load_array_length(src), load_array_length(dest));
4905
4906 ac->set_arraycopy(validated);
4907
4908 Node* n = _gvn.transform(ac);
4909 if (n == ac) {
4910 ac->connect_outputs(this);
4911 } else {
4912 assert(validated, "shouldn't transform if all arguments not validated");
4913 set_all_memory(n);
4914 }
4915 clear_upper_avx();
4916
4917
4918 return true;
4919 }
4920
4921
4922 // Helper function which determines if an arraycopy immediately follows
4923 // an allocation, with no intervening tests or other escapes for the object.
4924 AllocateArrayNode*
4925 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4926 RegionNode* slow_region) {
4927 if (stopped()) return NULL; // no fast path
4928 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4929
4930 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4931 if (alloc == NULL) return NULL;
4932
4933 Node* rawmem = memory(Compile::AliasIdxRaw);
4934 // Is the allocation's memory state untouched?
4935 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4936 // Bail out if there have been raw-memory effects since the allocation.
4937 // (Example: There might have been a call or safepoint.)
4938 return NULL;
4939 }
4940 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4941 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4942 return NULL;
4943 }
4944
4945 // There must be no unexpected observers of this allocation.
4946 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4947 Node* obs = ptr->fast_out(i);
4948 if (obs != this->map()) {
4949 return NULL;
4950 }
4951 }
4952
4953 // This arraycopy must unconditionally follow the allocation of the ptr.
4954 Node* alloc_ctl = ptr->in(0);
4955 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4956
4957 Node* ctl = control();
4958 while (ctl != alloc_ctl) {
4959 // There may be guards which feed into the slow_region.
4960 // Any other control flow means that we might not get a chance
4961 // to finish initializing the allocated object.
4962 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4963 IfNode* iff = ctl->in(0)->as_If();
4964 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
4965 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4966 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4967 ctl = iff->in(0); // This test feeds the known slow_region.
4968 continue;
4969 }
4970 // One more try: Various low-level checks bottom out in
4971 // uncommon traps. If the debug-info of the trap omits
4972 // any reference to the allocation, as we've already
4973 // observed, then there can be no objection to the trap.
4974 bool found_trap = false;
4975 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4976 Node* obs = not_ctl->fast_out(j);
4977 if (obs->in(0) == not_ctl && obs->is_Call() &&
4978 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4979 found_trap = true; break;
4980 }
4981 }
4982 if (found_trap) {
4983 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4984 continue;
4985 }
4986 }
4987 return NULL;
4988 }
4989
4990 // If we get this far, we have an allocation which immediately
4991 // precedes the arraycopy, and we can take over zeroing the new object.
4992 // The arraycopy will finish the initialization, and provide
4993 // a new control state to which we will anchor the destination pointer.
4994
4995 return alloc;
4996 }
4997
4998 //-------------inline_encodeISOArray-----------------------------------
4999 // encode char[] to byte[] in ISO_8859_1
5000 bool LibraryCallKit::inline_encodeISOArray() {
5001 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5002 // no receiver since it is static method
5003 Node *src = argument(0);
5004 Node *src_offset = argument(1);
5005 Node *dst = argument(2);
5006 Node *dst_offset = argument(3);
5007 Node *length = argument(4);
5008
5009 src = must_be_not_null(src, true);
5010 dst = must_be_not_null(dst, true);
5011
5012 src = access_resolve(src, ACCESS_READ);
5013 dst = access_resolve(dst, ACCESS_WRITE);
5014
5015 const Type* src_type = src->Value(&_gvn);
5016 const Type* dst_type = dst->Value(&_gvn);
5017 const TypeAryPtr* top_src = src_type->isa_aryptr();
5018 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5019 if (top_src == NULL || top_src->klass() == NULL ||
5020 top_dest == NULL || top_dest->klass() == NULL) {
5021 // failed array check
5022 return false;
5023 }
5024
5025 // Figure out the size and type of the elements we will be copying.
5026 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5027 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5028 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5029 return false;
5030 }
5031
5032 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5033 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5034 // 'src_start' points to src array + scaled offset
5035 // 'dst_start' points to dst array + scaled offset
5036
5037 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5038 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5039 enc = _gvn.transform(enc);
5040 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5041 set_memory(res_mem, mtype);
5042 set_result(enc);
5043 clear_upper_avx();
5044
5045 return true;
5046 }
5047
5048 //-------------inline_multiplyToLen-----------------------------------
5049 bool LibraryCallKit::inline_multiplyToLen() {
5050 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5051
5052 address stubAddr = StubRoutines::multiplyToLen();
5053 if (stubAddr == NULL) {
5054 return false; // Intrinsic's stub is not implemented on this platform
5055 }
5056 const char* stubName = "multiplyToLen";
5057
5058 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5059
5060 // no receiver because it is a static method
5061 Node* x = argument(0);
5062 Node* xlen = argument(1);
5063 Node* y = argument(2);
5064 Node* ylen = argument(3);
5065 Node* z = argument(4);
5066
5067 x = must_be_not_null(x, true);
5068 y = must_be_not_null(y, true);
5069
5070 x = access_resolve(x, ACCESS_READ);
5071 y = access_resolve(y, ACCESS_READ);
5072 z = access_resolve(z, ACCESS_WRITE);
5073
5074 const Type* x_type = x->Value(&_gvn);
5075 const Type* y_type = y->Value(&_gvn);
5076 const TypeAryPtr* top_x = x_type->isa_aryptr();
5077 const TypeAryPtr* top_y = y_type->isa_aryptr();
5078 if (top_x == NULL || top_x->klass() == NULL ||
5079 top_y == NULL || top_y->klass() == NULL) {
5080 // failed array check
5081 return false;
5082 }
5083
5084 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5085 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5086 if (x_elem != T_INT || y_elem != T_INT) {
5087 return false;
5088 }
5089
5090 // Set the original stack and the reexecute bit for the interpreter to reexecute
5091 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5092 // on the return from z array allocation in runtime.
5093 { PreserveReexecuteState preexecs(this);
5094 jvms()->set_should_reexecute(true);
5095
5096 Node* x_start = array_element_address(x, intcon(0), x_elem);
5097 Node* y_start = array_element_address(y, intcon(0), y_elem);
5098 // 'x_start' points to x array + scaled xlen
5099 // 'y_start' points to y array + scaled ylen
5100
5101 // Allocate the result array
5102 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5103 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5104 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5105
5106 IdealKit ideal(this);
5107
5108 #define __ ideal.
5109 Node* one = __ ConI(1);
5110 Node* zero = __ ConI(0);
5111 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5112 __ set(need_alloc, zero);
5113 __ set(z_alloc, z);
5114 __ if_then(z, BoolTest::eq, null()); {
5115 __ increment (need_alloc, one);
5116 } __ else_(); {
5117 // Update graphKit memory and control from IdealKit.
5118 sync_kit(ideal);
5119 Node *cast = new CastPPNode(z, TypePtr::NOTNULL);
5120 cast->init_req(0, control());
5121 _gvn.set_type(cast, cast->bottom_type());
5122 C->record_for_igvn(cast);
5123
5124 Node* zlen_arg = load_array_length(cast);
5125 // Update IdealKit memory and control from graphKit.
5126 __ sync_kit(this);
5127 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5128 __ increment (need_alloc, one);
5129 } __ end_if();
5130 } __ end_if();
5131
5132 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5133 // Update graphKit memory and control from IdealKit.
5134 sync_kit(ideal);
5135 Node * narr = new_array(klass_node, zlen, 1);
5136 // Update IdealKit memory and control from graphKit.
5137 __ sync_kit(this);
5138 __ set(z_alloc, narr);
5139 } __ end_if();
5140
5141 sync_kit(ideal);
5142 z = __ value(z_alloc);
5143 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5144 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5145 // Final sync IdealKit and GraphKit.
5146 final_sync(ideal);
5147 #undef __
5148
5149 Node* z_start = array_element_address(z, intcon(0), T_INT);
5150
5151 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5152 OptoRuntime::multiplyToLen_Type(),
5153 stubAddr, stubName, TypePtr::BOTTOM,
5154 x_start, xlen, y_start, ylen, z_start, zlen);
5155 } // original reexecute is set back here
5156
5157 C->set_has_split_ifs(true); // Has chance for split-if optimization
5158 set_result(z);
5159 return true;
5160 }
5161
5162 //-------------inline_squareToLen------------------------------------
5163 bool LibraryCallKit::inline_squareToLen() {
5164 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5165
5166 address stubAddr = StubRoutines::squareToLen();
5167 if (stubAddr == NULL) {
5168 return false; // Intrinsic's stub is not implemented on this platform
5169 }
5170 const char* stubName = "squareToLen";
5171
5172 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5173
5174 Node* x = argument(0);
5175 Node* len = argument(1);
5176 Node* z = argument(2);
5177 Node* zlen = argument(3);
5178
5179 x = must_be_not_null(x, true);
5180 z = must_be_not_null(z, true);
5181
5182 x = access_resolve(x, ACCESS_READ);
5183 z = access_resolve(z, ACCESS_WRITE);
5184
5185 const Type* x_type = x->Value(&_gvn);
5186 const Type* z_type = z->Value(&_gvn);
5187 const TypeAryPtr* top_x = x_type->isa_aryptr();
5188 const TypeAryPtr* top_z = z_type->isa_aryptr();
5189 if (top_x == NULL || top_x->klass() == NULL ||
5190 top_z == NULL || top_z->klass() == NULL) {
5191 // failed array check
5192 return false;
5193 }
5194
5195 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5196 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5197 if (x_elem != T_INT || z_elem != T_INT) {
5198 return false;
5199 }
5200
5201
5202 Node* x_start = array_element_address(x, intcon(0), x_elem);
5203 Node* z_start = array_element_address(z, intcon(0), z_elem);
5204
5205 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5206 OptoRuntime::squareToLen_Type(),
5207 stubAddr, stubName, TypePtr::BOTTOM,
5208 x_start, len, z_start, zlen);
5209
5210 set_result(z);
5211 return true;
5212 }
5213
5214 //-------------inline_mulAdd------------------------------------------
5215 bool LibraryCallKit::inline_mulAdd() {
5216 assert(UseMulAddIntrinsic, "not implemented on this platform");
5217
5218 address stubAddr = StubRoutines::mulAdd();
5219 if (stubAddr == NULL) {
5220 return false; // Intrinsic's stub is not implemented on this platform
5221 }
5222 const char* stubName = "mulAdd";
5223
5224 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5225
5226 Node* out = argument(0);
5227 Node* in = argument(1);
5228 Node* offset = argument(2);
5229 Node* len = argument(3);
5230 Node* k = argument(4);
5231
5232 out = must_be_not_null(out, true);
5233
5234 in = access_resolve(in, ACCESS_READ);
5235 out = access_resolve(out, ACCESS_WRITE);
5236
5237 const Type* out_type = out->Value(&_gvn);
5238 const Type* in_type = in->Value(&_gvn);
5239 const TypeAryPtr* top_out = out_type->isa_aryptr();
5240 const TypeAryPtr* top_in = in_type->isa_aryptr();
5241 if (top_out == NULL || top_out->klass() == NULL ||
5242 top_in == NULL || top_in->klass() == NULL) {
5243 // failed array check
5244 return false;
5245 }
5246
5247 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5248 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5249 if (out_elem != T_INT || in_elem != T_INT) {
5250 return false;
5251 }
5252
5253 Node* outlen = load_array_length(out);
5254 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5255 Node* out_start = array_element_address(out, intcon(0), out_elem);
5256 Node* in_start = array_element_address(in, intcon(0), in_elem);
5257
5258 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5259 OptoRuntime::mulAdd_Type(),
5260 stubAddr, stubName, TypePtr::BOTTOM,
5261 out_start,in_start, new_offset, len, k);
5262 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5263 set_result(result);
5264 return true;
5265 }
5266
5267 //-------------inline_montgomeryMultiply-----------------------------------
5268 bool LibraryCallKit::inline_montgomeryMultiply() {
5269 address stubAddr = StubRoutines::montgomeryMultiply();
5270 if (stubAddr == NULL) {
5271 return false; // Intrinsic's stub is not implemented on this platform
5272 }
5273
5274 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5275 const char* stubName = "montgomery_multiply";
5276
5277 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5278
5279 Node* a = argument(0);
5280 Node* b = argument(1);
5281 Node* n = argument(2);
5282 Node* len = argument(3);
5283 Node* inv = argument(4);
5284 Node* m = argument(6);
5285
5286 a = access_resolve(a, ACCESS_READ);
5287 b = access_resolve(b, ACCESS_READ);
5288 n = access_resolve(n, ACCESS_READ);
5289 m = access_resolve(m, ACCESS_WRITE);
5290
5291 const Type* a_type = a->Value(&_gvn);
5292 const TypeAryPtr* top_a = a_type->isa_aryptr();
5293 const Type* b_type = b->Value(&_gvn);
5294 const TypeAryPtr* top_b = b_type->isa_aryptr();
5295 const Type* n_type = a->Value(&_gvn);
5296 const TypeAryPtr* top_n = n_type->isa_aryptr();
5297 const Type* m_type = a->Value(&_gvn);
5298 const TypeAryPtr* top_m = m_type->isa_aryptr();
5299 if (top_a == NULL || top_a->klass() == NULL ||
5300 top_b == NULL || top_b->klass() == NULL ||
5301 top_n == NULL || top_n->klass() == NULL ||
5302 top_m == NULL || top_m->klass() == NULL) {
5303 // failed array check
5304 return false;
5305 }
5306
5307 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5308 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5309 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5310 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5311 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5312 return false;
5313 }
5314
5315 // Make the call
5316 {
5317 Node* a_start = array_element_address(a, intcon(0), a_elem);
5318 Node* b_start = array_element_address(b, intcon(0), b_elem);
5319 Node* n_start = array_element_address(n, intcon(0), n_elem);
5320 Node* m_start = array_element_address(m, intcon(0), m_elem);
5321
5322 Node* call = make_runtime_call(RC_LEAF,
5323 OptoRuntime::montgomeryMultiply_Type(),
5324 stubAddr, stubName, TypePtr::BOTTOM,
5325 a_start, b_start, n_start, len, inv, top(),
5326 m_start);
5327 set_result(m);
5328 }
5329
5330 return true;
5331 }
5332
5333 bool LibraryCallKit::inline_montgomerySquare() {
5334 address stubAddr = StubRoutines::montgomerySquare();
5335 if (stubAddr == NULL) {
5336 return false; // Intrinsic's stub is not implemented on this platform
5337 }
5338
5339 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5340 const char* stubName = "montgomery_square";
5341
5342 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5343
5344 Node* a = argument(0);
5345 Node* n = argument(1);
5346 Node* len = argument(2);
5347 Node* inv = argument(3);
5348 Node* m = argument(5);
5349
5350 a = access_resolve(a, ACCESS_READ);
5351 n = access_resolve(n, ACCESS_READ);
5352 m = access_resolve(m, ACCESS_WRITE);
5353
5354 const Type* a_type = a->Value(&_gvn);
5355 const TypeAryPtr* top_a = a_type->isa_aryptr();
5356 const Type* n_type = a->Value(&_gvn);
5357 const TypeAryPtr* top_n = n_type->isa_aryptr();
5358 const Type* m_type = a->Value(&_gvn);
5359 const TypeAryPtr* top_m = m_type->isa_aryptr();
5360 if (top_a == NULL || top_a->klass() == NULL ||
5361 top_n == NULL || top_n->klass() == NULL ||
5362 top_m == NULL || top_m->klass() == NULL) {
5363 // failed array check
5364 return false;
5365 }
5366
5367 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5368 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5369 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5370 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5371 return false;
5372 }
5373
5374 // Make the call
5375 {
5376 Node* a_start = array_element_address(a, intcon(0), a_elem);
5377 Node* n_start = array_element_address(n, intcon(0), n_elem);
5378 Node* m_start = array_element_address(m, intcon(0), m_elem);
5379
5380 Node* call = make_runtime_call(RC_LEAF,
5381 OptoRuntime::montgomerySquare_Type(),
5382 stubAddr, stubName, TypePtr::BOTTOM,
5383 a_start, n_start, len, inv, top(),
5384 m_start);
5385 set_result(m);
5386 }
5387
5388 return true;
5389 }
5390
5391 //-------------inline_vectorizedMismatch------------------------------
5392 bool LibraryCallKit::inline_vectorizedMismatch() {
5393 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5394
5395 address stubAddr = StubRoutines::vectorizedMismatch();
5396 if (stubAddr == NULL) {
5397 return false; // Intrinsic's stub is not implemented on this platform
5398 }
5399 const char* stubName = "vectorizedMismatch";
5400 int size_l = callee()->signature()->size();
5401 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5402
5403 Node* obja = argument(0);
5404 Node* aoffset = argument(1);
5405 Node* objb = argument(3);
5406 Node* boffset = argument(4);
5407 Node* length = argument(6);
5408 Node* scale = argument(7);
5409
5410 const Type* a_type = obja->Value(&_gvn);
5411 const Type* b_type = objb->Value(&_gvn);
5412 const TypeAryPtr* top_a = a_type->isa_aryptr();
5413 const TypeAryPtr* top_b = b_type->isa_aryptr();
5414 if (top_a == NULL || top_a->klass() == NULL ||
5415 top_b == NULL || top_b->klass() == NULL) {
5416 // failed array check
5417 return false;
5418 }
5419
5420 Node* call;
5421 jvms()->set_should_reexecute(true);
5422
5423 obja = access_resolve(obja, ACCESS_READ);
5424 objb = access_resolve(objb, ACCESS_READ);
5425 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ);
5426 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ);
5427
5428 call = make_runtime_call(RC_LEAF,
5429 OptoRuntime::vectorizedMismatch_Type(),
5430 stubAddr, stubName, TypePtr::BOTTOM,
5431 obja_adr, objb_adr, length, scale);
5432
5433 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5434 set_result(result);
5435 return true;
5436 }
5437
5438 /**
5439 * Calculate CRC32 for byte.
5440 * int java.util.zip.CRC32.update(int crc, int b)
5441 */
5442 bool LibraryCallKit::inline_updateCRC32() {
5443 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5444 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5445 // no receiver since it is static method
5446 Node* crc = argument(0); // type: int
5447 Node* b = argument(1); // type: int
5448
5449 /*
5450 * int c = ~ crc;
5451 * b = timesXtoThe32[(b ^ c) & 0xFF];
5452 * b = b ^ (c >>> 8);
5453 * crc = ~b;
5454 */
5455
5456 Node* M1 = intcon(-1);
5457 crc = _gvn.transform(new XorINode(crc, M1));
5458 Node* result = _gvn.transform(new XorINode(crc, b));
5459 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5460
5461 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5462 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5463 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5464 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5465
5466 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5467 result = _gvn.transform(new XorINode(crc, result));
5468 result = _gvn.transform(new XorINode(result, M1));
5469 set_result(result);
5470 return true;
5471 }
5472
5473 /**
5474 * Calculate CRC32 for byte[] array.
5475 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5476 */
5477 bool LibraryCallKit::inline_updateBytesCRC32() {
5478 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5479 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5480 // no receiver since it is 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* length = argument(3); // type: int
5485
5486 const Type* src_type = src->Value(&_gvn);
5487 const TypeAryPtr* top_src = src_type->isa_aryptr();
5488 if (top_src == NULL || top_src->klass() == NULL) {
5489 // failed array check
5490 return false;
5491 }
5492
5493 // Figure out the size and type of the elements we will be copying.
5494 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5495 if (src_elem != T_BYTE) {
5496 return false;
5497 }
5498
5499 // 'src_start' points to src array + scaled offset
5500 src = must_be_not_null(src, true);
5501 src = access_resolve(src, ACCESS_READ);
5502 Node* src_start = array_element_address(src, offset, src_elem);
5503
5504 // We assume that range check is done by caller.
5505 // TODO: generate range check (offset+length < src.length) in debug VM.
5506
5507 // Call the stub.
5508 address stubAddr = StubRoutines::updateBytesCRC32();
5509 const char *stubName = "updateBytesCRC32";
5510
5511 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5512 stubAddr, stubName, TypePtr::BOTTOM,
5513 crc, src_start, length);
5514 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5515 set_result(result);
5516 return true;
5517 }
5518
5519 /**
5520 * Calculate CRC32 for ByteBuffer.
5521 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5522 */
5523 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5524 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5525 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5526 // no receiver since it is static method
5527 Node* crc = argument(0); // type: int
5528 Node* src = argument(1); // type: long
5529 Node* offset = argument(3); // type: int
5530 Node* length = argument(4); // type: int
5531
5532 src = ConvL2X(src); // adjust Java long to machine word
5533 Node* base = _gvn.transform(new CastX2PNode(src));
5534 offset = ConvI2X(offset);
5535
5536 // 'src_start' points to src array + scaled offset
5537 Node* src_start = basic_plus_adr(top(), base, offset);
5538
5539 // Call the stub.
5540 address stubAddr = StubRoutines::updateBytesCRC32();
5541 const char *stubName = "updateBytesCRC32";
5542
5543 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5544 stubAddr, stubName, TypePtr::BOTTOM,
5545 crc, src_start, length);
5546 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5547 set_result(result);
5548 return true;
5549 }
5550
5551 //------------------------------get_table_from_crc32c_class-----------------------
5552 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5553 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5554 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5555
5556 return table;
5557 }
5558
5559 //------------------------------inline_updateBytesCRC32C-----------------------
5560 //
5561 // Calculate CRC32C for byte[] array.
5562 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5563 //
5564 bool LibraryCallKit::inline_updateBytesCRC32C() {
5565 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5566 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5567 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5568 // no receiver since it is a static method
5569 Node* crc = argument(0); // type: int
5570 Node* src = argument(1); // type: oop
5571 Node* offset = argument(2); // type: int
5572 Node* end = argument(3); // type: int
5573
5574 Node* length = _gvn.transform(new SubINode(end, offset));
5575
5576 const Type* src_type = src->Value(&_gvn);
5577 const TypeAryPtr* top_src = src_type->isa_aryptr();
5578 if (top_src == NULL || top_src->klass() == NULL) {
5579 // failed array check
5580 return false;
5581 }
5582
5583 // Figure out the size and type of the elements we will be copying.
5584 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5585 if (src_elem != T_BYTE) {
5586 return false;
5587 }
5588
5589 // 'src_start' points to src array + scaled offset
5590 src = must_be_not_null(src, true);
5591 src = access_resolve(src, ACCESS_READ);
5592 Node* src_start = array_element_address(src, offset, src_elem);
5593
5594 // static final int[] byteTable in class CRC32C
5595 Node* table = get_table_from_crc32c_class(callee()->holder());
5596 table = must_be_not_null(table, true);
5597 table = access_resolve(table, ACCESS_READ);
5598 Node* table_start = array_element_address(table, intcon(0), T_INT);
5599
5600 // We assume that range check is done by caller.
5601 // TODO: generate range check (offset+length < src.length) in debug VM.
5602
5603 // Call the stub.
5604 address stubAddr = StubRoutines::updateBytesCRC32C();
5605 const char *stubName = "updateBytesCRC32C";
5606
5607 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5608 stubAddr, stubName, TypePtr::BOTTOM,
5609 crc, src_start, length, table_start);
5610 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5611 set_result(result);
5612 return true;
5613 }
5614
5615 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5616 //
5617 // Calculate CRC32C for DirectByteBuffer.
5618 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5619 //
5620 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5621 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5622 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5623 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5624 // no receiver since it is a static method
5625 Node* crc = argument(0); // type: int
5626 Node* src = argument(1); // type: long
5627 Node* offset = argument(3); // type: int
5628 Node* end = argument(4); // type: int
5629
5630 Node* length = _gvn.transform(new SubINode(end, offset));
5631
5632 src = ConvL2X(src); // adjust Java long to machine word
5633 Node* base = _gvn.transform(new CastX2PNode(src));
5634 offset = ConvI2X(offset);
5635
5636 // 'src_start' points to src array + scaled offset
5637 Node* src_start = basic_plus_adr(top(), base, offset);
5638
5639 // static final int[] byteTable in class CRC32C
5640 Node* table = get_table_from_crc32c_class(callee()->holder());
5641 table = must_be_not_null(table, true);
5642 table = access_resolve(table, ACCESS_READ);
5643 Node* table_start = array_element_address(table, intcon(0), T_INT);
5644
5645 // Call the stub.
5646 address stubAddr = StubRoutines::updateBytesCRC32C();
5647 const char *stubName = "updateBytesCRC32C";
5648
5649 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5650 stubAddr, stubName, TypePtr::BOTTOM,
5651 crc, src_start, length, table_start);
5652 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5653 set_result(result);
5654 return true;
5655 }
5656
5657 //------------------------------inline_updateBytesAdler32----------------------
5658 //
5659 // Calculate Adler32 checksum for byte[] array.
5660 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5661 //
5662 bool LibraryCallKit::inline_updateBytesAdler32() {
5663 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5664 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5665 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5666 // no receiver since it is static method
5667 Node* crc = argument(0); // type: int
5668 Node* src = argument(1); // type: oop
5669 Node* offset = argument(2); // type: int
5670 Node* length = argument(3); // type: int
5671
5672 const Type* src_type = src->Value(&_gvn);
5673 const TypeAryPtr* top_src = src_type->isa_aryptr();
5674 if (top_src == NULL || top_src->klass() == NULL) {
5675 // failed array check
5676 return false;
5677 }
5678
5679 // Figure out the size and type of the elements we will be copying.
5680 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5681 if (src_elem != T_BYTE) {
5682 return false;
5683 }
5684
5685 // 'src_start' points to src array + scaled offset
5686 src = access_resolve(src, ACCESS_READ);
5687 Node* src_start = array_element_address(src, offset, src_elem);
5688
5689 // We assume that range check is done by caller.
5690 // TODO: generate range check (offset+length < src.length) in debug VM.
5691
5692 // Call the stub.
5693 address stubAddr = StubRoutines::updateBytesAdler32();
5694 const char *stubName = "updateBytesAdler32";
5695
5696 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5697 stubAddr, stubName, TypePtr::BOTTOM,
5698 crc, src_start, length);
5699 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5700 set_result(result);
5701 return true;
5702 }
5703
5704 //------------------------------inline_updateByteBufferAdler32---------------
5705 //
5706 // Calculate Adler32 checksum for DirectByteBuffer.
5707 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5708 //
5709 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5710 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5711 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5712 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5713 // no receiver since it is static method
5714 Node* crc = argument(0); // type: int
5715 Node* src = argument(1); // type: long
5716 Node* offset = argument(3); // type: int
5717 Node* length = argument(4); // type: int
5718
5719 src = ConvL2X(src); // adjust Java long to machine word
5720 Node* base = _gvn.transform(new CastX2PNode(src));
5721 offset = ConvI2X(offset);
5722
5723 // 'src_start' points to src array + scaled offset
5724 Node* src_start = basic_plus_adr(top(), base, offset);
5725
5726 // Call the stub.
5727 address stubAddr = StubRoutines::updateBytesAdler32();
5728 const char *stubName = "updateBytesAdler32";
5729
5730 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5731 stubAddr, stubName, TypePtr::BOTTOM,
5732 crc, src_start, length);
5733
5734 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5735 set_result(result);
5736 return true;
5737 }
5738
5739 //----------------------------inline_reference_get----------------------------
5740 // public T java.lang.ref.Reference.get();
5741 bool LibraryCallKit::inline_reference_get() {
5742 const int referent_offset = java_lang_ref_Reference::referent_offset;
5743 guarantee(referent_offset > 0, "should have already been set");
5744
5745 // Get the argument:
5746 Node* reference_obj = null_check_receiver();
5747 if (stopped()) return true;
5748
5749 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr();
5750 assert(tinst != NULL, "obj is null");
5751 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5752 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass();
5753 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"),
5754 ciSymbol::make("Ljava/lang/Object;"),
5755 false);
5756 assert (field != NULL, "undefined field");
5757
5758 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5759 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5760
5761 ciInstanceKlass* klass = env()->Object_klass();
5762 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5763
5764 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF;
5765 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators);
5766 // Add memory barrier to prevent commoning reads from this field
5767 // across safepoint since GC can change its value.
5768 insert_mem_bar(Op_MemBarCPUOrder);
5769
5770 set_result(result);
5771 return true;
5772 }
5773
5774
5775 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5776 bool is_exact=true, bool is_static=false,
5777 ciInstanceKlass * fromKls=NULL) {
5778 if (fromKls == NULL) {
5779 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5780 assert(tinst != NULL, "obj is null");
5781 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5782 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5783 fromKls = tinst->klass()->as_instance_klass();
5784 } else {
5785 assert(is_static, "only for static field access");
5786 }
5787 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5788 ciSymbol::make(fieldTypeString),
5789 is_static);
5790
5791 assert (field != NULL, "undefined field");
5792 if (field == NULL) return (Node *) NULL;
5793
5794 if (is_static) {
5795 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5796 fromObj = makecon(tip);
5797 }
5798
5799 // Next code copied from Parse::do_get_xxx():
5800
5801 // Compute address and memory type.
5802 int offset = field->offset_in_bytes();
5803 bool is_vol = field->is_volatile();
5804 ciType* field_klass = field->type();
5805 assert(field_klass->is_loaded(), "should be loaded");
5806 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5807 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5808 BasicType bt = field->layout_type();
5809
5810 // Build the resultant type of the load
5811 const Type *type;
5812 if (bt == T_OBJECT) {
5813 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5814 } else {
5815 type = Type::get_const_basic_type(bt);
5816 }
5817
5818 DecoratorSet decorators = IN_HEAP;
5819
5820 if (is_vol) {
5821 decorators |= MO_SEQ_CST;
5822 }
5823
5824 return access_load_at(fromObj, adr, adr_type, type, bt, decorators);
5825 }
5826
5827 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5828 bool is_exact = true, bool is_static = false,
5829 ciInstanceKlass * fromKls = NULL) {
5830 if (fromKls == NULL) {
5831 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5832 assert(tinst != NULL, "obj is null");
5833 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5834 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5835 fromKls = tinst->klass()->as_instance_klass();
5836 }
5837 else {
5838 assert(is_static, "only for static field access");
5839 }
5840 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5841 ciSymbol::make(fieldTypeString),
5842 is_static);
5843
5844 assert(field != NULL, "undefined field");
5845 assert(!field->is_volatile(), "not defined for volatile fields");
5846
5847 if (is_static) {
5848 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5849 fromObj = makecon(tip);
5850 }
5851
5852 // Next code copied from Parse::do_get_xxx():
5853
5854 // Compute address and memory type.
5855 int offset = field->offset_in_bytes();
5856 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5857
5858 return adr;
5859 }
5860
5861 //------------------------------inline_aescrypt_Block-----------------------
5862 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5863 address stubAddr = NULL;
5864 const char *stubName;
5865 assert(UseAES, "need AES instruction support");
5866
5867 switch(id) {
5868 case vmIntrinsics::_aescrypt_encryptBlock:
5869 stubAddr = StubRoutines::aescrypt_encryptBlock();
5870 stubName = "aescrypt_encryptBlock";
5871 break;
5872 case vmIntrinsics::_aescrypt_decryptBlock:
5873 stubAddr = StubRoutines::aescrypt_decryptBlock();
5874 stubName = "aescrypt_decryptBlock";
5875 break;
5876 default:
5877 break;
5878 }
5879 if (stubAddr == NULL) return false;
5880
5881 Node* aescrypt_object = argument(0);
5882 Node* src = argument(1);
5883 Node* src_offset = argument(2);
5884 Node* dest = argument(3);
5885 Node* dest_offset = argument(4);
5886
5887 src = must_be_not_null(src, true);
5888 dest = must_be_not_null(dest, true);
5889
5890 src = access_resolve(src, ACCESS_READ);
5891 dest = access_resolve(dest, ACCESS_WRITE);
5892
5893 // (1) src and dest are arrays.
5894 const Type* src_type = src->Value(&_gvn);
5895 const Type* dest_type = dest->Value(&_gvn);
5896 const TypeAryPtr* top_src = src_type->isa_aryptr();
5897 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5898 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5899
5900 // for the quick and dirty code we will skip all the checks.
5901 // we are just trying to get the call to be generated.
5902 Node* src_start = src;
5903 Node* dest_start = dest;
5904 if (src_offset != NULL || dest_offset != NULL) {
5905 assert(src_offset != NULL && dest_offset != NULL, "");
5906 src_start = array_element_address(src, src_offset, T_BYTE);
5907 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5908 }
5909
5910 // now need to get the start of its expanded key array
5911 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5912 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5913 if (k_start == NULL) return false;
5914
5915 if (Matcher::pass_original_key_for_aes()) {
5916 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5917 // compatibility issues between Java key expansion and SPARC crypto instructions
5918 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5919 if (original_k_start == NULL) return false;
5920
5921 // Call the stub.
5922 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5923 stubAddr, stubName, TypePtr::BOTTOM,
5924 src_start, dest_start, k_start, original_k_start);
5925 } else {
5926 // Call the stub.
5927 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5928 stubAddr, stubName, TypePtr::BOTTOM,
5929 src_start, dest_start, k_start);
5930 }
5931
5932 return true;
5933 }
5934
5935 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5936 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5937 address stubAddr = NULL;
5938 const char *stubName = NULL;
5939
5940 assert(UseAES, "need AES instruction support");
5941
5942 switch(id) {
5943 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5944 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5945 stubName = "cipherBlockChaining_encryptAESCrypt";
5946 break;
5947 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5948 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5949 stubName = "cipherBlockChaining_decryptAESCrypt";
5950 break;
5951 default:
5952 break;
5953 }
5954 if (stubAddr == NULL) return false;
5955
5956 Node* cipherBlockChaining_object = argument(0);
5957 Node* src = argument(1);
5958 Node* src_offset = argument(2);
5959 Node* len = argument(3);
5960 Node* dest = argument(4);
5961 Node* dest_offset = argument(5);
5962
5963 src = must_be_not_null(src, false);
5964 dest = must_be_not_null(dest, false);
5965
5966 src = access_resolve(src, ACCESS_READ);
5967 dest = access_resolve(dest, ACCESS_WRITE);
5968
5969 // (1) src and dest are arrays.
5970 const Type* src_type = src->Value(&_gvn);
5971 const Type* dest_type = dest->Value(&_gvn);
5972 const TypeAryPtr* top_src = src_type->isa_aryptr();
5973 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5974 assert (top_src != NULL && top_src->klass() != NULL
5975 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5976
5977 // checks are the responsibility of the caller
5978 Node* src_start = src;
5979 Node* dest_start = dest;
5980 if (src_offset != NULL || dest_offset != NULL) {
5981 assert(src_offset != NULL && dest_offset != NULL, "");
5982 src_start = array_element_address(src, src_offset, T_BYTE);
5983 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5984 }
5985
5986 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5987 // (because of the predicated logic executed earlier).
5988 // so we cast it here safely.
5989 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5990
5991 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5992 if (embeddedCipherObj == NULL) return false;
5993
5994 // cast it to what we know it will be at runtime
5995 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5996 assert(tinst != NULL, "CBC obj is null");
5997 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5998 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5999 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6000
6001 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6002 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6003 const TypeOopPtr* xtype = aklass->as_instance_type();
6004 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6005 aescrypt_object = _gvn.transform(aescrypt_object);
6006
6007 // we need to get the start of the aescrypt_object's expanded key array
6008 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6009 if (k_start == NULL) return false;
6010
6011 // similarly, get the start address of the r vector
6012 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6013 if (objRvec == NULL) return false;
6014 objRvec = access_resolve(objRvec, ACCESS_WRITE);
6015 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6016
6017 Node* cbcCrypt;
6018 if (Matcher::pass_original_key_for_aes()) {
6019 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6020 // compatibility issues between Java key expansion and SPARC crypto instructions
6021 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6022 if (original_k_start == NULL) return false;
6023
6024 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6025 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6026 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6027 stubAddr, stubName, TypePtr::BOTTOM,
6028 src_start, dest_start, k_start, r_start, len, original_k_start);
6029 } else {
6030 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6031 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6032 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6033 stubAddr, stubName, TypePtr::BOTTOM,
6034 src_start, dest_start, k_start, r_start, len);
6035 }
6036
6037 // return cipher length (int)
6038 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6039 set_result(retvalue);
6040 return true;
6041 }
6042
6043 //------------------------------inline_counterMode_AESCrypt-----------------------
6044 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6045 assert(UseAES, "need AES instruction support");
6046 if (!UseAESCTRIntrinsics) return false;
6047
6048 address stubAddr = NULL;
6049 const char *stubName = NULL;
6050 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6051 stubAddr = StubRoutines::counterMode_AESCrypt();
6052 stubName = "counterMode_AESCrypt";
6053 }
6054 if (stubAddr == NULL) return false;
6055
6056 Node* counterMode_object = argument(0);
6057 Node* src = argument(1);
6058 Node* src_offset = argument(2);
6059 Node* len = argument(3);
6060 Node* dest = argument(4);
6061 Node* dest_offset = argument(5);
6062
6063 src = access_resolve(src, ACCESS_READ);
6064 dest = access_resolve(dest, ACCESS_WRITE);
6065 counterMode_object = access_resolve(counterMode_object, ACCESS_WRITE);
6066
6067 // (1) src and dest are arrays.
6068 const Type* src_type = src->Value(&_gvn);
6069 const Type* dest_type = dest->Value(&_gvn);
6070 const TypeAryPtr* top_src = src_type->isa_aryptr();
6071 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6072 assert(top_src != NULL && top_src->klass() != NULL &&
6073 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6074
6075 // checks are the responsibility of the caller
6076 Node* src_start = src;
6077 Node* dest_start = dest;
6078 if (src_offset != NULL || dest_offset != NULL) {
6079 assert(src_offset != NULL && dest_offset != NULL, "");
6080 src_start = array_element_address(src, src_offset, T_BYTE);
6081 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6082 }
6083
6084 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6085 // (because of the predicated logic executed earlier).
6086 // so we cast it here safely.
6087 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6088 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6089 if (embeddedCipherObj == NULL) return false;
6090 // cast it to what we know it will be at runtime
6091 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6092 assert(tinst != NULL, "CTR obj is null");
6093 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6094 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6095 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6096 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6097 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6098 const TypeOopPtr* xtype = aklass->as_instance_type();
6099 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6100 aescrypt_object = _gvn.transform(aescrypt_object);
6101 // we need to get the start of the aescrypt_object's expanded key array
6102 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6103 if (k_start == NULL) return false;
6104 // similarly, get the start address of the r vector
6105 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6106 if (obj_counter == NULL) return false;
6107 obj_counter = access_resolve(obj_counter, ACCESS_WRITE);
6108 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6109
6110 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6111 if (saved_encCounter == NULL) return false;
6112 saved_encCounter = access_resolve(saved_encCounter, ACCESS_WRITE);
6113 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6114 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6115
6116 Node* ctrCrypt;
6117 if (Matcher::pass_original_key_for_aes()) {
6118 // no SPARC version for AES/CTR intrinsics now.
6119 return false;
6120 }
6121 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6122 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6123 OptoRuntime::counterMode_aescrypt_Type(),
6124 stubAddr, stubName, TypePtr::BOTTOM,
6125 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6126
6127 // return cipher length (int)
6128 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6129 set_result(retvalue);
6130 return true;
6131 }
6132
6133 //------------------------------get_key_start_from_aescrypt_object-----------------------
6134 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6135 #if defined(PPC64) || defined(S390)
6136 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6137 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6138 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6139 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6140 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6141 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6142 if (objSessionK == NULL) {
6143 return (Node *) NULL;
6144 }
6145 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6146 #else
6147 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6148 #endif // PPC64
6149 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6150 if (objAESCryptKey == NULL) return (Node *) NULL;
6151
6152 // now have the array, need to get the start address of the K array
6153 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ);
6154 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6155 return k_start;
6156 }
6157
6158 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6159 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6160 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6161 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6162 if (objAESCryptKey == NULL) return (Node *) NULL;
6163
6164 // now have the array, need to get the start address of the lastKey array
6165 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ);
6166 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6167 return original_k_start;
6168 }
6169
6170 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6171 // Return node representing slow path of predicate check.
6172 // the pseudo code we want to emulate with this predicate is:
6173 // for encryption:
6174 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6175 // for decryption:
6176 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6177 // note cipher==plain is more conservative than the original java code but that's OK
6178 //
6179 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6180 // The receiver was checked for NULL already.
6181 Node* objCBC = argument(0);
6182
6183 Node* src = argument(1);
6184 Node* dest = argument(4);
6185
6186 // Load embeddedCipher field of CipherBlockChaining object.
6187 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6188
6189 // get AESCrypt klass for instanceOf check
6190 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6191 // will have same classloader as CipherBlockChaining object
6192 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6193 assert(tinst != NULL, "CBCobj is null");
6194 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6195
6196 // we want to do an instanceof comparison against the AESCrypt class
6197 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6198 if (!klass_AESCrypt->is_loaded()) {
6199 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6200 Node* ctrl = control();
6201 set_control(top()); // no regular fast path
6202 return ctrl;
6203 }
6204
6205 src = must_be_not_null(src, true);
6206 dest = must_be_not_null(dest, true);
6207
6208 // Resolve oops to stable for CmpP below.
6209 src = access_resolve(src, 0);
6210 dest = access_resolve(dest, 0);
6211
6212 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6213
6214 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6215 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6216 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6217
6218 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6219
6220 // for encryption, we are done
6221 if (!decrypting)
6222 return instof_false; // even if it is NULL
6223
6224 // for decryption, we need to add a further check to avoid
6225 // taking the intrinsic path when cipher and plain are the same
6226 // see the original java code for why.
6227 RegionNode* region = new RegionNode(3);
6228 region->init_req(1, instof_false);
6229
6230 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6231 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6232 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6233 region->init_req(2, src_dest_conjoint);
6234
6235 record_for_igvn(region);
6236 return _gvn.transform(region);
6237 }
6238
6239 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6240 // Return node representing slow path of predicate check.
6241 // the pseudo code we want to emulate with this predicate is:
6242 // for encryption:
6243 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6244 // for decryption:
6245 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6246 // note cipher==plain is more conservative than the original java code but that's OK
6247 //
6248
6249 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6250 // The receiver was checked for NULL already.
6251 Node* objCTR = argument(0);
6252
6253 // Load embeddedCipher field of CipherBlockChaining object.
6254 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6255
6256 // get AESCrypt klass for instanceOf check
6257 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6258 // will have same classloader as CipherBlockChaining object
6259 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6260 assert(tinst != NULL, "CTRobj is null");
6261 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6262
6263 // we want to do an instanceof comparison against the AESCrypt class
6264 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6265 if (!klass_AESCrypt->is_loaded()) {
6266 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6267 Node* ctrl = control();
6268 set_control(top()); // no regular fast path
6269 return ctrl;
6270 }
6271
6272 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6273 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6274 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6275 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6276 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6277
6278 return instof_false; // even if it is NULL
6279 }
6280
6281 //------------------------------inline_ghash_processBlocks
6282 bool LibraryCallKit::inline_ghash_processBlocks() {
6283 address stubAddr;
6284 const char *stubName;
6285 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6286
6287 stubAddr = StubRoutines::ghash_processBlocks();
6288 stubName = "ghash_processBlocks";
6289
6290 Node* data = argument(0);
6291 Node* offset = argument(1);
6292 Node* len = argument(2);
6293 Node* state = argument(3);
6294 Node* subkeyH = argument(4);
6295
6296 state = must_be_not_null(state, true);
6297 subkeyH = must_be_not_null(subkeyH, true);
6298 data = must_be_not_null(data, true);
6299
6300 state = access_resolve(state, ACCESS_WRITE);
6301 subkeyH = access_resolve(subkeyH, ACCESS_READ);
6302 data = access_resolve(data, ACCESS_READ);
6303
6304 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6305 assert(state_start, "state is NULL");
6306 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6307 assert(subkeyH_start, "subkeyH is NULL");
6308 Node* data_start = array_element_address(data, offset, T_BYTE);
6309 assert(data_start, "data is NULL");
6310
6311 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6312 OptoRuntime::ghash_processBlocks_Type(),
6313 stubAddr, stubName, TypePtr::BOTTOM,
6314 state_start, subkeyH_start, data_start, len);
6315 return true;
6316 }
6317
6318 bool LibraryCallKit::inline_base64_encodeBlock() {
6319 address stubAddr;
6320 const char *stubName;
6321 assert(UseBASE64Intrinsics, "need Base64 intrinsics support");
6322 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters");
6323 stubAddr = StubRoutines::base64_encodeBlock();
6324 stubName = "encodeBlock";
6325
6326 if (!stubAddr) return false;
6327 Node* base64obj = argument(0);
6328 Node* src = argument(1);
6329 Node* offset = argument(2);
6330 Node* len = argument(3);
6331 Node* dest = argument(4);
6332 Node* dp = argument(5);
6333 Node* isURL = argument(6);
6334
6335 src = must_be_not_null(src, true);
6336 src = access_resolve(src, ACCESS_READ);
6337 dest = must_be_not_null(dest, true);
6338 dest = access_resolve(dest, ACCESS_WRITE);
6339
6340 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
6341 assert(src_start, "source array is NULL");
6342 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE);
6343 assert(dest_start, "destination array is NULL");
6344
6345 Node* base64 = make_runtime_call(RC_LEAF,
6346 OptoRuntime::base64_encodeBlock_Type(),
6347 stubAddr, stubName, TypePtr::BOTTOM,
6348 src_start, offset, len, dest_start, dp, isURL);
6349 return true;
6350 }
6351
6352 //------------------------------inline_sha_implCompress-----------------------
6353 //
6354 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6355 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6356 //
6357 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6358 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6359 //
6360 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6361 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6362 //
6363 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6364 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6365
6366 Node* sha_obj = argument(0);
6367 Node* src = argument(1); // type oop
6368 Node* ofs = argument(2); // type int
6369
6370 const Type* src_type = src->Value(&_gvn);
6371 const TypeAryPtr* top_src = src_type->isa_aryptr();
6372 if (top_src == NULL || top_src->klass() == NULL) {
6373 // failed array check
6374 return false;
6375 }
6376 // Figure out the size and type of the elements we will be copying.
6377 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6378 if (src_elem != T_BYTE) {
6379 return false;
6380 }
6381 // 'src_start' points to src array + offset
6382 src = must_be_not_null(src, true);
6383 src = access_resolve(src, ACCESS_READ);
6384 Node* src_start = array_element_address(src, ofs, src_elem);
6385 Node* state = NULL;
6386 address stubAddr;
6387 const char *stubName;
6388
6389 switch(id) {
6390 case vmIntrinsics::_sha_implCompress:
6391 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6392 state = get_state_from_sha_object(sha_obj);
6393 stubAddr = StubRoutines::sha1_implCompress();
6394 stubName = "sha1_implCompress";
6395 break;
6396 case vmIntrinsics::_sha2_implCompress:
6397 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6398 state = get_state_from_sha_object(sha_obj);
6399 stubAddr = StubRoutines::sha256_implCompress();
6400 stubName = "sha256_implCompress";
6401 break;
6402 case vmIntrinsics::_sha5_implCompress:
6403 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6404 state = get_state_from_sha5_object(sha_obj);
6405 stubAddr = StubRoutines::sha512_implCompress();
6406 stubName = "sha512_implCompress";
6407 break;
6408 default:
6409 fatal_unexpected_iid(id);
6410 return false;
6411 }
6412 if (state == NULL) return false;
6413
6414 assert(stubAddr != NULL, "Stub is generated");
6415 if (stubAddr == NULL) return false;
6416
6417 // Call the stub.
6418 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6419 stubAddr, stubName, TypePtr::BOTTOM,
6420 src_start, state);
6421
6422 return true;
6423 }
6424
6425 //------------------------------inline_digestBase_implCompressMB-----------------------
6426 //
6427 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6428 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6429 //
6430 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6431 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6432 "need SHA1/SHA256/SHA512 instruction support");
6433 assert((uint)predicate < 3, "sanity");
6434 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6435
6436 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6437 Node* src = argument(1); // byte[] array
6438 Node* ofs = argument(2); // type int
6439 Node* limit = argument(3); // type int
6440
6441 const Type* src_type = src->Value(&_gvn);
6442 const TypeAryPtr* top_src = src_type->isa_aryptr();
6443 if (top_src == NULL || top_src->klass() == NULL) {
6444 // failed array check
6445 return false;
6446 }
6447 // Figure out the size and type of the elements we will be copying.
6448 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6449 if (src_elem != T_BYTE) {
6450 return false;
6451 }
6452 // 'src_start' points to src array + offset
6453 src = must_be_not_null(src, false);
6454 src = access_resolve(src, ACCESS_READ);
6455 Node* src_start = array_element_address(src, ofs, src_elem);
6456
6457 const char* klass_SHA_name = NULL;
6458 const char* stub_name = NULL;
6459 address stub_addr = NULL;
6460 bool long_state = false;
6461
6462 switch (predicate) {
6463 case 0:
6464 if (UseSHA1Intrinsics) {
6465 klass_SHA_name = "sun/security/provider/SHA";
6466 stub_name = "sha1_implCompressMB";
6467 stub_addr = StubRoutines::sha1_implCompressMB();
6468 }
6469 break;
6470 case 1:
6471 if (UseSHA256Intrinsics) {
6472 klass_SHA_name = "sun/security/provider/SHA2";
6473 stub_name = "sha256_implCompressMB";
6474 stub_addr = StubRoutines::sha256_implCompressMB();
6475 }
6476 break;
6477 case 2:
6478 if (UseSHA512Intrinsics) {
6479 klass_SHA_name = "sun/security/provider/SHA5";
6480 stub_name = "sha512_implCompressMB";
6481 stub_addr = StubRoutines::sha512_implCompressMB();
6482 long_state = true;
6483 }
6484 break;
6485 default:
6486 fatal("unknown SHA intrinsic predicate: %d", predicate);
6487 }
6488 if (klass_SHA_name != NULL) {
6489 assert(stub_addr != NULL, "Stub is generated");
6490 if (stub_addr == NULL) return false;
6491
6492 // get DigestBase klass to lookup for SHA klass
6493 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6494 assert(tinst != NULL, "digestBase_obj is not instance???");
6495 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6496
6497 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6498 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6499 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6500 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6501 }
6502 return false;
6503 }
6504 //------------------------------inline_sha_implCompressMB-----------------------
6505 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6506 bool long_state, address stubAddr, const char *stubName,
6507 Node* src_start, Node* ofs, Node* limit) {
6508 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6509 const TypeOopPtr* xtype = aklass->as_instance_type();
6510 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6511 sha_obj = _gvn.transform(sha_obj);
6512
6513 Node* state;
6514 if (long_state) {
6515 state = get_state_from_sha5_object(sha_obj);
6516 } else {
6517 state = get_state_from_sha_object(sha_obj);
6518 }
6519 if (state == NULL) return false;
6520
6521 // Call the stub.
6522 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6523 OptoRuntime::digestBase_implCompressMB_Type(),
6524 stubAddr, stubName, TypePtr::BOTTOM,
6525 src_start, state, ofs, limit);
6526 // return ofs (int)
6527 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6528 set_result(result);
6529
6530 return true;
6531 }
6532
6533 enum VectorApiObjectType {
6534 VECAPI_VECTOR,
6535 VECAPI_MASK,
6536 VECAPI_SPECIES,
6537 VECAPI_SHUFFLE,
6538 };
6539
6540 enum VectorMaskUseType {
6541 VecMaskUseLoad,
6542 VecMaskUseStore,
6543 VecMaskUseAll,
6544 VecMaskNotUsed
6545 };
6546
6547 static bool arch_supports_vector(int op, int num_elem, BasicType type, VectorMaskUseType mask_use_type, int op_arity = 0) {
6548 // Check that the operation is valid.
6549 if (op <= 0) {
6550 #ifndef PRODUCT
6551 if (DebugVectorApi) {
6552 tty->print_cr("Rejected intrinsification because no valid vector op could be extracted");
6553 }
6554 #endif
6555 return false;
6556 }
6557
6558 // Check that architecture supports this op-size-type combination.
6559 if (!Matcher::match_rule_supported_vector(op, num_elem, type, op_arity)) {
6560 #ifndef PRODUCT
6561 if (DebugVectorApi) {
6562 tty->print_cr("Rejected vector op (%s,%s,%d) because architecture does not support it",
6563 NodeClassNames[op], type2name(type), num_elem);
6564 }
6565 #endif
6566 return false;
6567 } else {
6568 assert(Matcher::match_rule_supported(op), "must be supported");
6569 }
6570
6571 // Check whether mask unboxing is supported.
6572 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseLoad) {
6573 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type)) {
6574 #ifndef PRODUCT
6575 if (DebugVectorApi) {
6576 tty->print_cr("Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
6577 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
6578 }
6579 #endif
6580 return false;
6581 }
6582 }
6583
6584 // Check whether mask boxing is supported.
6585 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseStore) {
6586 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type)) {
6587 #ifndef PRODUCT
6588 if (DebugVectorApi) {
6589 tty->print_cr("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
6590 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
6591 }
6592 #endif
6593 return false;
6594 }
6595 }
6596
6597 return true;
6598 }
6599
6600 // Should be aligned with constants in jdk.incubator.vector.VectorIntrinsics.
6601 enum {
6602 // Unary
6603 OP_ABS = 0,
6604 OP_NEG = 1,
6605 OP_SQRT = 2,
6606 OP_NOT = 3,
6607 // Binary
6608 OP_ADD = 4,
6609 OP_SUB = 5,
6610 OP_MUL = 6,
6611 OP_DIV = 7,
6612 OP_MIN = 8,
6613 OP_MAX = 9,
6614 OP_AND = 10,
6615 OP_OR = 11,
6616 OP_XOR = 12,
6617 // Ternary
6618 OP_FMA = 13,
6619 // Broadcast int
6620 OP_LSHIFT = 14,
6621 OP_RSHIFT = 15,
6622 OP_URSHIFT = 16,
6623 // Vector Math Library
6624 OP_TAN = 101,
6625 OP_SVML_START = OP_TAN,
6626 OP_TANH = 102,
6627 OP_SIN = 103,
6628 OP_SINH = 104,
6629 OP_COS = 105,
6630 OP_COSH = 106,
6631 OP_ASIN = 107,
6632 OP_ACOS = 108,
6633 OP_ATAN = 109,
6634 OP_ATAN2 = 110,
6635 OP_CBRT = 111,
6636 OP_LOG = 112,
6637 OP_LOG10 = 113,
6638 OP_LOG1P = 114,
6639 OP_POW = 115,
6640 OP_EXP = 116,
6641 OP_EXPM1 = 117,
6642 OP_HYPOT = 118,
6643 OP_SVML_END = OP_HYPOT,
6644 };
6645
6646 static int get_opc(jint op, BasicType bt) {
6647 switch (op) {
6648 case OP_ADD: {
6649 switch (bt) {
6650 case T_BYTE: // fall-through
6651 case T_SHORT: // fall-through
6652 case T_INT: return Op_AddI;
6653 case T_LONG: return Op_AddL;
6654 case T_FLOAT: return Op_AddF;
6655 case T_DOUBLE: return Op_AddD;
6656 default: fatal("ADD: %s", type2name(bt));
6657 }
6658 break;
6659 }
6660 case OP_SUB: {
6661 switch (bt) {
6662 case T_BYTE: // fall-through
6663 case T_SHORT: // fall-through
6664 case T_INT: return Op_SubI;
6665 case T_LONG: return Op_SubL;
6666 case T_FLOAT: return Op_SubF;
6667 case T_DOUBLE: return Op_SubD;
6668 default: fatal("SUB: %s", type2name(bt));
6669 }
6670 break;
6671 }
6672 case OP_MUL: {
6673 switch (bt) {
6674 case T_BYTE: // fall-through
6675 case T_SHORT: // fall-through
6676 case T_INT: return Op_MulI;
6677 case T_LONG: return Op_MulL;
6678 case T_FLOAT: return Op_MulF;
6679 case T_DOUBLE: return Op_MulD;
6680 default: fatal("MUL: %s", type2name(bt));
6681 }
6682 break;
6683 }
6684 case OP_DIV: {
6685 switch (bt) {
6686 case T_BYTE: // fall-through
6687 case T_SHORT: // fall-through
6688 case T_INT: return Op_DivI;
6689 case T_LONG: return Op_DivL;
6690 case T_FLOAT: return Op_DivF;
6691 case T_DOUBLE: return Op_DivD;
6692 default: fatal("DIV: %s", type2name(bt));
6693 }
6694 break;
6695 }
6696 case OP_MIN: {
6697 switch (bt) {
6698 case T_BYTE:
6699 case T_SHORT:
6700 case T_INT: return Op_MinI;
6701 case T_LONG: return Op_MinL;
6702 case T_FLOAT: return Op_MinF;
6703 case T_DOUBLE: return Op_MinD;
6704 default: fatal("MIN: %s", type2name(bt));
6705 }
6706 break;
6707 }
6708 case OP_MAX: {
6709 switch (bt) {
6710 case T_BYTE:
6711 case T_SHORT:
6712 case T_INT: return Op_MaxI;
6713 case T_LONG: return Op_MaxL;
6714 case T_FLOAT: return Op_MaxF;
6715 case T_DOUBLE: return Op_MaxD;
6716 default: fatal("MAX: %s", type2name(bt));
6717 }
6718 break;
6719 }
6720 case OP_ABS: {
6721 switch (bt) {
6722 case T_BYTE: // fall-through
6723 case T_SHORT: // fall-through
6724 case T_LONG: // fall-through
6725 case T_INT: return Op_AbsI;
6726 case T_FLOAT: return Op_AbsF;
6727 case T_DOUBLE: return Op_AbsD;
6728 default: fatal("ABS: %s", type2name(bt));
6729 }
6730 break;
6731 }
6732 case OP_NEG: {
6733 switch (bt) {
6734 case T_BYTE: // fall-through
6735 case T_SHORT: // fall-through
6736 case T_INT: return Op_NegI;
6737 case T_FLOAT: return Op_NegF;
6738 case T_DOUBLE: return Op_NegD;
6739 default: fatal("NEG: %s", type2name(bt));
6740 }
6741 break;
6742 }
6743 case OP_AND: {
6744 switch (bt) {
6745 case T_BYTE: // fall-through
6746 case T_SHORT: // fall-through
6747 case T_INT: return Op_AndI;
6748 case T_LONG: return Op_AndL;
6749 default: fatal("AND: %s", type2name(bt));
6750 }
6751 break;
6752 }
6753 case OP_OR: {
6754 switch (bt) {
6755 case T_BYTE: // fall-through
6756 case T_SHORT: // fall-through
6757 case T_INT: return Op_OrI;
6758 case T_LONG: return Op_OrL;
6759 default: fatal("OR: %s", type2name(bt));
6760 }
6761 break;
6762 }
6763 case OP_XOR: {
6764 switch (bt) {
6765 case T_BYTE: // fall-through
6766 case T_SHORT: // fall-through
6767 case T_INT: return Op_XorI;
6768 case T_LONG: return Op_XorL;
6769 default: fatal("XOR: %s", type2name(bt));
6770 }
6771 break;
6772 }
6773 case OP_SQRT: {
6774 switch (bt) {
6775 case T_FLOAT: return Op_SqrtF;
6776 case T_DOUBLE: return Op_SqrtD;
6777 default: fatal("SQRT: %s", type2name(bt));
6778 }
6779 break;
6780 }
6781 case OP_NOT: {
6782 switch (bt) {
6783 case T_BYTE: // fall-through
6784 case T_SHORT: // fall-through
6785 case T_INT: // fall-through
6786 case T_LONG: return Op_Not;
6787 default: fatal("NOT: %s", type2name(bt));
6788 }
6789 break;
6790 }
6791 case OP_FMA: {
6792 switch (bt) {
6793 case T_FLOAT: return Op_FmaF;
6794 case T_DOUBLE: return Op_FmaD;
6795 default: fatal("FMA: %s", type2name(bt));
6796 }
6797 break;
6798 }
6799 case OP_LSHIFT: {
6800 switch (bt) {
6801 case T_BYTE: // fall-through
6802 case T_SHORT: // fall-through
6803 case T_INT: return Op_LShiftI;
6804 case T_LONG: return Op_LShiftL;
6805 default: fatal("LSHIFT: %s", type2name(bt));
6806 }
6807 break;
6808 }
6809 case OP_RSHIFT: {
6810 switch (bt) {
6811 case T_BYTE: // fall-through
6812 case T_SHORT: // fall-through
6813 case T_INT: return Op_RShiftI;
6814 case T_LONG: return Op_RShiftL;
6815 default: fatal("RSHIFT: %s", type2name(bt));
6816 }
6817 break;
6818 }
6819 case OP_URSHIFT: {
6820 switch (bt) {
6821 case T_BYTE: return Op_URShiftB;
6822 case T_SHORT: return Op_URShiftS;
6823 case T_INT: return Op_URShiftI;
6824 case T_LONG: return Op_URShiftL;
6825 default: fatal("URSHIFT: %s", type2name(bt));
6826 }
6827 break;
6828 }
6829 case OP_TAN:
6830 case OP_TANH:
6831 case OP_SIN:
6832 case OP_SINH:
6833 case OP_COS:
6834 case OP_COSH:
6835 case OP_ASIN:
6836 case OP_ACOS:
6837 case OP_ATAN:
6838 case OP_ATAN2:
6839 case OP_CBRT:
6840 case OP_LOG:
6841 case OP_LOG10:
6842 case OP_LOG1P:
6843 case OP_POW:
6844 case OP_EXP:
6845 case OP_EXPM1:
6846 case OP_HYPOT:
6847 return Op_CallLeafVector;
6848 default: fatal("unknown op: %d", op);
6849 }
6850 return 0; // Unimplemented
6851 }
6852
6853 static int get_sopc(int opc, BasicType elem_bt) {
6854 int sopc = 0;
6855 switch (opc) {
6856 case Op_CallLeafVector:
6857 sopc = Op_CallLeafVector;
6858 break;
6859 default:
6860 sopc = VectorNode::opcode(opc, elem_bt);
6861 break;
6862 }
6863 return sopc;
6864 }
6865 Node* LibraryCallKit::box_vector(Node* vector, const TypeInstPtr* vbox_type,
6866 BasicType elem_bt, int num_elem) {
6867
6868 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
6869
6870 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
6871 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
6872 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true);
6873 set_i_o(_gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
6874 set_all_memory(_gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
6875 Node* ret = _gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
6876
6877 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vec_type);
6878 return _gvn.transform(vbox);
6879 }
6880
6881 Node* LibraryCallKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) {
6882 const TypeInstPtr* vbox_type_v = gvn().type(v)->is_instptr();
6883 if (vbox_type->klass() != vbox_type_v->klass()) {
6884 return NULL; // arguments don't agree on vector shapes
6885 }
6886 if (vbox_type_v->maybe_null()) {
6887 return NULL; // no nulls are allowed
6888 }
6889 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
6890 Node* unbox = gvn().transform(new VectorUnboxNode(C, vec_type, v, merged_memory()));
6891 return unbox;
6892 }
6893
6894 void LibraryCallKit::set_vector_result(Node* result, bool set_res) {
6895 if (DebugVectorApi) {
6896 #ifndef PRODUCT
6897 tty->print("============ ");
6898 callee()->print();
6899 tty->print_cr(" ============");
6900 result->dump(5);
6901 tty->print_cr("----------------------------------------------------");
6902 #endif
6903 }
6904 if (set_res) {
6905 set_result(result);
6906 }
6907 }
6908
6909 bool LibraryCallKit::inline_vector_nary_operation(int n) {
6910 const TypeInt* opr = gvn().type(argument(0))->is_int();
6911 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
6912 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
6913 const TypeInt* vlen = gvn().type(argument(3))->is_int();
6914
6915 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
6916 return false; // not enough info for intrinsification
6917 }
6918 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
6919 if (!elem_type->is_primitive_type()) {
6920 return false; // should be primitive type
6921 }
6922 BasicType elem_bt = elem_type->basic_type();
6923 int num_elem = vlen->get_con();
6924 int opc = get_opc(opr->get_con(), elem_bt);
6925 int sopc = get_sopc(opc, elem_bt);
6926 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
6927 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
6928
6929 if (opc == Op_CallLeafVector && !Matcher::supports_vector_calling_convention()) {
6930 return false;
6931 }
6932
6933 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
6934 if (!arch_supports_vector(sopc, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseAll : VecMaskNotUsed, n)) {
6935 return false; // not supported
6936 }
6937
6938 Node* opd1 = NULL; Node* opd2 = NULL; Node* opd3 = NULL;
6939 switch (n) {
6940 case 3: {
6941 opd3 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
6942 if (opd3 == NULL) {
6943 return false;
6944 }
6945 // fall-through
6946 }
6947 case 2: {
6948 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
6949 if (opd2 == NULL) {
6950 return false;
6951 }
6952 // fall-through
6953 }
6954 case 1: {
6955 opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
6956 if (opd1 == NULL) {
6957 return false;
6958 }
6959 break;
6960 }
6961 default: fatal("unsupported arity: %d", n);
6962 }
6963
6964 Node* operation = NULL;
6965 if (sopc == Op_CallLeafVector) {
6966 operation = gen_call_to_svml(opr->get_con(), elem_bt, num_elem, opd1, opd2);
6967 if (operation == NULL) {
6968 return false;
6969 }
6970 } else {
6971 switch (n) {
6972 case 1:
6973 case 2: {
6974 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt));
6975 break;
6976 }
6977 case 3: {
6978 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, opd3, num_elem, elem_bt));
6979 break;
6980 }
6981 default: fatal("unsupported arity: %d", n);
6982 }
6983 }
6984 // Wrap it up in VectorBox to keep object type information.
6985 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
6986 set_vector_result(operation);
6987
6988 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
6989 return true;
6990 }
6991
6992 // <V extends Vector<?,?>>
6993 // V broadcastCoerced(Class<?> vectorClass, Class<?> elementType, int vlen,
6994 // long bits,
6995 // LongFunction<V> defaultImpl)
6996 bool LibraryCallKit::inline_vector_broadcast_coerced() {
6997 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
6998 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
6999 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7000
7001 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7002 return false; // not enough info for intrinsification
7003 }
7004
7005 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7006 if (!elem_type->is_primitive_type()) {
7007 return false; // should be primitive type
7008 }
7009 BasicType elem_bt = elem_type->basic_type();
7010 int num_elem = vlen->get_con();
7011 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7012 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7013
7014 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7015 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt,
7016 vbox_klass->is_vectormask() ? VecMaskUseStore : VecMaskNotUsed)) {
7017 return false; // not supported
7018 }
7019
7020 Node* bits = argument(3); // long
7021
7022 Node* elem = NULL;
7023 switch (elem_bt) {
7024 case T_BOOLEAN: // fall-through
7025 case T_BYTE: // fall-through
7026 case T_SHORT: // fall-through
7027 case T_CHAR: // fall-through
7028 case T_INT: {
7029 elem = gvn().transform(new ConvL2INode(bits));
7030 break;
7031 }
7032 case T_DOUBLE: {
7033 elem = gvn().transform(new MoveL2DNode(bits));
7034 break;
7035 }
7036 case T_FLOAT: {
7037 bits = gvn().transform(new ConvL2INode(bits));
7038 elem = gvn().transform(new MoveI2FNode(bits));
7039 break;
7040 }
7041 case T_LONG: {
7042 elem = bits; // no conversion needed
7043 break;
7044 }
7045 default: fatal("%s", type2name(elem_bt));
7046 }
7047
7048 Node* broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt));
7049 broadcast = gvn().transform(broadcast);
7050 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
7051 set_vector_result(box);
7052
7053 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7054 return true;
7055 }
7056
7057 // <C, V extends Vector<?,?>>
7058 // V load(Class<?> vectorClass, Class<?> elementType, int vlen,
7059 // Object base, long offset,
7060 // /* Vector.Mask<E,S> m*/
7061 // Object container, int index,
7062 // LoadOperation<C, VM> defaultImpl) {
7063 //
7064 // <C, V extends Vector<?,?>>
7065 // void store(Class<?> vectorClass, Class<?> elementType, int vlen,
7066 // Object base, long offset,
7067 // V v, /*Vector.Mask<E,S> m*/
7068 // Object container, int index,
7069 // StoreVectorOperation<C, V> defaultImpl) {
7070
7071 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
7072 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7073 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7074 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7075
7076 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7077 return false; // not enough info for intrinsification
7078 }
7079
7080 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7081 if (!elem_type->is_primitive_type()) {
7082 return false; // should be primitive type
7083 }
7084 BasicType elem_bt = elem_type->basic_type();
7085 int num_elem = vlen->get_con();
7086
7087 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7088 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
7089 return false; // not supported
7090 }
7091
7092 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7093 bool is_mask = vbox_klass->is_vectormask();
7094
7095 Node* base = argument(3);
7096 Node* offset = ConvL2X(argument(4));
7097 DecoratorSet decorators = C2_UNSAFE_ACCESS;
7098 Node* addr = make_unsafe_address(base, offset, decorators, (is_mask ? T_BOOLEAN : elem_bt), true);
7099
7100 // Can base be NULL? Otherwise, always on-heap access.
7101 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base));
7102
7103 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
7104 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
7105
7106 // Now handle special case where load/store happens from/to byte array but element type is not byte.
7107 bool using_byte_array = arr_type != NULL && arr_type->elem()->array_element_basic_type() == T_BYTE && elem_bt != T_BYTE;
7108 // Handle loading masks.
7109 // If there is no consistency between array and vector element types, it must be special byte array case or loading masks
7110 if (arr_type != NULL && !using_byte_array && elem_bt != arr_type->elem()->array_element_basic_type() && !is_mask) {
7111 return false;
7112 }
7113 // Since we are using byte array, we need to double check that the byte operations are supported by backend.
7114 if (using_byte_array) {
7115 int byte_num_elem = num_elem * type2aelembytes(elem_bt);
7116 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, byte_num_elem, T_BYTE, VecMaskNotUsed)) {
7117 return false; // not supported
7118 }
7119 }
7120 if (is_mask) {
7121 if (!arch_supports_vector(Op_LoadVector, num_elem, T_BOOLEAN, VecMaskNotUsed)) {
7122 return false; // not supported
7123 }
7124 }
7125
7126 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7127
7128 if (can_access_non_heap) {
7129 insert_mem_bar(Op_MemBarCPUOrder);
7130 }
7131
7132 if (is_store) {
7133 Node* val = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7134 if (val == NULL) {
7135 return false; // operand unboxing failed
7136 }
7137 set_all_memory(reset_memory());
7138
7139 // In case the store needs to happen to byte array, reinterpret the incoming vector to byte vector.
7140 int store_num_elem = num_elem;
7141 if (using_byte_array) {
7142 store_num_elem = num_elem * type2aelembytes(elem_bt);
7143 const TypeVect* to_vect_type = TypeVect::make(T_BYTE, store_num_elem);
7144 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type));
7145 }
7146
7147 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(addr), addr, addr_type, val, store_num_elem));
7148 set_memory(vstore, addr_type);
7149 set_vector_result(vstore, false);
7150 } else {
7151 // When using byte array, we need to load as byte then reinterpret the value. Otherwise, do a simple vector load.
7152 Node* vload = NULL;
7153 if (using_byte_array) {
7154 int load_num_elem = num_elem * type2aelembytes(elem_bt);
7155 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, load_num_elem, T_BYTE));
7156 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
7157 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type));
7158 } else {
7159 // Special handle for masks
7160 if (is_mask) {
7161 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, T_BOOLEAN));
7162 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
7163 vload = gvn().transform(new VectorLoadMaskNode(vload, to_vect_type));
7164 } else {
7165 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, elem_bt));
7166 }
7167 }
7168 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7169 set_vector_result(box);
7170 }
7171
7172 if (can_access_non_heap) {
7173 insert_mem_bar(Op_MemBarCPUOrder);
7174 }
7175
7176 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7177 return true;
7178 }
7179
7180 // <C, V extends Vector<?>, W extends IntVector>
7181 // void storeWithMap(Class<?> vectorClass, Class<?> elementType, int length, Class<?> vectorIndexClass,
7182 // Object base, long offset, // Unsafe addressing
7183 // W index_vector, V v,
7184 // C container, int index, int[] indexMap, int indexM, // Arguments for default implementation
7185 // StoreVectorOperationWithMap<C, V> defaultImpl) {
7186 //
7187 bool LibraryCallKit::inline_vector_gather_scatter(bool is_scatter) {
7188 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7189 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7190 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7191 const TypeInstPtr* vector_idx_klass = gvn().type(argument(3))->is_instptr();
7192
7193 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || vector_idx_klass->const_oop() == NULL || !vlen->is_con()) {
7194 return false; // not enough info for intrinsification
7195 }
7196
7197 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7198 if (!elem_type->is_primitive_type()) {
7199 return false; // should be primitive type
7200 }
7201 BasicType elem_bt = elem_type->basic_type();
7202 int num_elem = vlen->get_con();
7203
7204 if (!arch_supports_vector(is_scatter ? Op_StoreVectorScatter : Op_LoadVectorGather, num_elem, elem_bt, VecMaskNotUsed)) {
7205 return false; // not supported
7206 }
7207
7208 // Check that the vector holding indices is supported by architecture
7209 if (!arch_supports_vector(Op_LoadVector, num_elem, T_INT, VecMaskNotUsed)) {
7210 return false; // not supported
7211 }
7212
7213 Node* base = argument(4);
7214 Node* offset = ConvL2X(argument(5));
7215 Node* addr = make_unsafe_address(base, offset, C2_UNSAFE_ACCESS, elem_bt, true);
7216
7217 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
7218 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
7219
7220 // The array must be consistent with vector type
7221 if (arr_type == NULL || (arr_type != NULL && elem_bt != arr_type->elem()->array_element_basic_type())) {
7222 return false;
7223 }
7224 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7225 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7226
7227 ciKlass* vbox_idx_klass = vector_idx_klass->const_oop()->as_instance()->java_lang_Class_klass();
7228
7229 if (vbox_idx_klass == NULL) {
7230 return false;
7231 }
7232
7233 const TypeInstPtr* vbox_idx_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_idx_klass);
7234
7235 Node* index_vect = unbox_vector(argument(7), vbox_idx_type, T_INT, num_elem);
7236 if (index_vect == NULL) {
7237 return false;
7238 }
7239
7240 const TypeVect* index_src_type = index_vect->as_Vector()->vect_type();
7241 const TypeVect* index_dst_type = TypeVect::make(T_INT, num_elem);
7242
7243 if (Type::cmp(index_src_type, index_dst_type) != 0) {
7244 // For DoubleMaxVector and LongMaxVector, indexVector's length is twice.
7245 assert((elem_bt == T_DOUBLE || elem_bt == T_LONG) &&
7246 index_src_type->element_basic_type() == T_INT &&
7247 index_src_type->length() == index_dst_type->length() * 2,
7248 "Expect double the length");
7249
7250 // Check whether resizing is support.
7251 if (!arch_supports_vector(Op_VectorReinterpret,
7252 index_src_type->length(),
7253 index_src_type->element_basic_type(),
7254 VecMaskNotUsed)) {
7255 return false;
7256 }
7257 if (!arch_supports_vector(Op_VectorReinterpret,
7258 num_elem,
7259 T_INT,
7260 VecMaskNotUsed)) {
7261 return false;
7262 }
7263 index_vect = gvn().transform(new VectorReinterpretNode(index_vect, index_src_type, index_dst_type));
7264 }
7265
7266 const TypeVect* vector_type = TypeVect::make(elem_bt, num_elem);
7267 if (is_scatter) {
7268 Node* val = unbox_vector(argument(8), vbox_type, elem_bt, num_elem);
7269 if (val == NULL) {
7270 return false; // operand unboxing failed
7271 }
7272 set_all_memory(reset_memory());
7273
7274 Node* vstore = gvn().transform(new StoreVectorScatterNode(control(), memory(addr), addr, addr_type, val, index_vect));
7275 set_memory(vstore, addr_type);
7276 set_vector_result(vstore, false);
7277 } else {
7278 Node* vload = gvn().transform(new LoadVectorGatherNode(control(), memory(addr), addr, addr_type, vector_type, index_vect));
7279
7280 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7281 set_vector_result(box);
7282 }
7283
7284 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7285 return true;
7286 }
7287
7288 // <V extends Vector<?,?>>
7289 // long reductionCoerced(int oprId, Class<?> vectorClass, Class<?> elementType, int vlen,
7290 // V v,
7291 // Function<V,Long> defaultImpl)
7292
7293 bool LibraryCallKit::inline_vector_reduction() {
7294 const TypeInt* opr = gvn().type(argument(0))->is_int();
7295 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7296 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7297 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7298
7299 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7300 return false; // not enough info for intrinsification
7301 }
7302 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7303 if (!elem_type->is_primitive_type()) {
7304 return false; // should be primitive type
7305 }
7306 BasicType elem_bt = elem_type->basic_type();
7307 int num_elem = vlen->get_con();
7308
7309 int opc = get_opc(opr->get_con(), elem_bt);
7310 int sopc = ReductionNode::opcode(opc, elem_bt);
7311
7312 // FIXME: When encountering a SubReduction, we want to check for support of
7313 // the corresponding AddReduction node.
7314 if (sopc == Op_SubReductionV) {
7315 if (gvn().type(argument(2))->isa_int()) {
7316 sopc = Op_AddReductionVI;
7317 } else if (gvn().type(argument(2))->isa_long()) {
7318 sopc = Op_AddReductionVL;
7319 }
7320 }
7321
7322 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7323 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7324 return false;
7325 }
7326
7327 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7328 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7329
7330 Node* opd = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7331 if (opd == NULL) {
7332 return false; // operand unboxing failed
7333 }
7334
7335 Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt);
7336 Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, elem_bt));
7337
7338 Node* bits = NULL;
7339 switch (elem_bt) {
7340 case T_BYTE:
7341 case T_SHORT:
7342 case T_INT: {
7343 bits = gvn().transform(new ConvI2LNode(rn));
7344 break;
7345 }
7346 case T_FLOAT: {
7347 rn = gvn().transform(new MoveF2INode(rn));
7348 bits = gvn().transform(new ConvI2LNode(rn));
7349 break;
7350 }
7351 case T_DOUBLE: {
7352 bits = gvn().transform(new MoveD2LNode(rn));
7353 break;
7354 }
7355 case T_LONG: {
7356 bits = rn; // no conversion needed
7357 break;
7358 }
7359 default: fatal("%s", type2name(elem_bt));
7360 }
7361 set_vector_result(bits);
7362 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7363 return true;
7364 }
7365
7366 // static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
7367 // V v1, V v2,
7368 // BiFunction<V, V, Boolean> defaultImpl) {
7369
7370 bool LibraryCallKit::inline_vector_test() {
7371 const TypeInt* cond = gvn().type(argument(0))->is_int();
7372 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7373 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7374 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7375
7376 if (!cond->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7377 return false; // not enough info for intrinsification
7378 }
7379 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7380 if (!elem_type->is_primitive_type()) {
7381 return false; // should be primitive type
7382 }
7383 BasicType elem_bt = elem_type->basic_type();
7384 int num_elem = vlen->get_con();
7385 BoolTest::mask booltest = (BoolTest::mask)cond->get_con();
7386 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7387 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7388
7389 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseLoad : VecMaskNotUsed)) {
7390 return false;
7391 }
7392
7393 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7394 Node* opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7395 if (opd1 == NULL || opd2 == NULL) {
7396 return false; // operand unboxing failed
7397 }
7398 Node* test = new VectorTestNode(opd1, opd2, booltest);
7399 test = _gvn.transform(test);
7400 set_vector_result(test);
7401 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7402 return true;
7403 }
7404
7405 // static
7406 // <V extends Vector, M extends Mask>
7407 // V blend(Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7408 // V v1, V v2, M m,
7409 // VectorBlendOp<V,M> defaultImpl) { ...
7410 //
7411 bool LibraryCallKit::inline_vector_blend() {
7412 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7413 const TypeInstPtr* mask_klass = gvn().type(argument(1))->is_instptr();
7414 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7415 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7416
7417 if (mask_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7418 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7419 return false; // not enough info for intrinsification
7420 }
7421 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7422 if (!elem_type->is_primitive_type()) {
7423 return false; // should be primitive type
7424 }
7425 BasicType elem_bt = elem_type->basic_type();
7426 BasicType mask_bt = elem_bt;
7427 int num_elem = vlen->get_con();
7428
7429 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
7430 return false; // not supported
7431 }
7432 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7433 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7434
7435 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7436 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7437
7438 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7439 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7440 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
7441
7442 if (v1 == NULL || v2 == NULL || mask == NULL) {
7443 return false; // operand unboxing failed
7444 }
7445
7446 Node* blend = _gvn.transform(new VectorBlendNode(v1, v2, mask));
7447 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
7448 set_vector_result(box);
7449
7450 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7451 return true;
7452 }
7453
7454 // static <V extends Vector<E,S>,
7455 // M extends Vector.Mask<E,S>,
7456 // S extends Vector.Shape, E>
7457 // M compare(int cond, Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7458 // V v1, V v2,
7459 // VectorCompareOp<V,M> defaultImpl) { ...
7460 //
7461 bool LibraryCallKit::inline_vector_compare() {
7462 const TypeInt* cond = gvn().type(argument(0))->is_int();
7463 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7464 const TypeInstPtr* mask_klass = gvn().type(argument(2))->is_instptr();
7465 const TypeInstPtr* elem_klass = gvn().type(argument(3))->is_instptr();
7466 const TypeInt* vlen = gvn().type(argument(4))->is_int();
7467
7468 if (!cond->is_con() || vector_klass->const_oop() == NULL || mask_klass->const_oop() == NULL ||
7469 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7470 return false; // not enough info for intrinsification
7471 }
7472 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7473 if (!elem_type->is_primitive_type()) {
7474 return false; // should be primitive type
7475 }
7476
7477 int num_elem = vlen->get_con();
7478 BasicType elem_bt = elem_type->basic_type();
7479 BasicType mask_bt = elem_bt;
7480
7481 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
7482 return false;
7483 }
7484
7485 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7486 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7487
7488 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7489 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7490
7491 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7492 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7493
7494 if (v1 == NULL || v2 == NULL) {
7495 return false; // operand unboxing failed
7496 }
7497 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
7498 const TypeVect* vt = TypeVect::make(mask_bt, num_elem);
7499 Node* operation = _gvn.transform(new VectorMaskCmpNode(pred, v1, v2, vt));
7500
7501 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
7502 set_vector_result(box);
7503
7504 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7505 return true;
7506 }
7507
7508 // static
7509 // <V extends Vector, Sh extends Shuffle>
7510 // V rearrangeOp(Class<V> vectorClass, Class<Sh> shuffleClass, Class< ? > elementType, int vlen,
7511 // V v1, Sh sh,
7512 // VectorSwizzleOp<V, Sh, S, E> defaultImpl) { ...
7513
7514 bool LibraryCallKit::inline_vector_rearrange() {
7515 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7516 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->is_instptr();
7517 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7518 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7519
7520 if (shuffle_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7521 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7522 return false; // not enough info for intrinsification
7523 }
7524 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7525 if (!elem_type->is_primitive_type()) {
7526 return false; // should be primitive type
7527 }
7528 BasicType elem_bt = elem_type->basic_type();
7529 BasicType shuffle_bt = elem_bt;
7530 int num_elem = vlen->get_con();
7531
7532 if (!arch_supports_vector(Op_VectorLoadShuffle, num_elem, elem_bt, VecMaskNotUsed)) {
7533 return false; // not supported
7534 }
7535 if (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed)) {
7536 return false; // not supported
7537 }
7538 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7539 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7540
7541 ciKlass* shbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
7542 const TypeInstPtr* shbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, shbox_klass);
7543
7544 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7545 Node* shuffle = unbox_vector(argument(5), shbox_type, shuffle_bt, num_elem);
7546
7547 if (v1 == NULL || shuffle == NULL) {
7548 return false; // operand unboxing failed
7549 }
7550
7551 Node* rearrange = _gvn.transform(new VectorRearrangeNode(v1, shuffle));
7552 Node* box = box_vector(rearrange, vbox_type, elem_bt, num_elem);
7553 set_result(box);
7554
7555 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7556 return true;
7557 }
7558
7559 Node* LibraryCallKit::shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
7560 assert(bt == T_INT || bt == T_LONG || bt == T_SHORT || bt == T_BYTE, "byte, short, long and int are supported");
7561 juint mask = (type2aelembytes(bt) * BitsPerByte - 1);
7562 Node* nmask = _gvn.transform(ConNode::make(TypeInt::make(mask)));
7563 Node* mcnt = _gvn.transform(new AndINode(cnt, nmask));
7564 return _gvn.transform(VectorNode::shift_count(shift_op, mcnt, num_elem, bt));
7565 }
7566
7567 static void get_svml_address(int op, int bits, BasicType bt, const char** name_ptr, address* addr_ptr) {
7568 assert(name_ptr != NULL, "unexpected");
7569 assert(addr_ptr != NULL, "unexpected");
7570
7571 // Since the addresses are resolved at runtime, using switch instead of table - otherwise might get NULL addresses.
7572 if (bt == T_FLOAT) {
7573 switch(op) {
7574 case OP_EXP: {
7575 switch(bits) {
7576 case 64: *name_ptr = "vector_float64_exp"; *addr_ptr = StubRoutines::vector_float64_exp(); break;
7577 case 128: *name_ptr = "vector_float128_exp"; *addr_ptr = StubRoutines::vector_float128_exp(); break;
7578 case 256: *name_ptr = "vector_float256_exp"; *addr_ptr = StubRoutines::vector_float256_exp(); break;
7579 case 512: *name_ptr = "vector_float512_exp"; *addr_ptr = StubRoutines::vector_float512_exp(); break;
7580 default: Unimplemented(); break;
7581 }
7582 }
7583 break;
7584 case OP_LOG1P: {
7585 switch(bits) {
7586 case 64: *name_ptr = "vector_float64_log1p"; *addr_ptr = StubRoutines::vector_float64_log1p(); break;
7587 case 128: *name_ptr = "vector_float128_log1p"; *addr_ptr = StubRoutines::vector_float128_log1p(); break;
7588 case 256: *name_ptr = "vector_float256_log1p"; *addr_ptr = StubRoutines::vector_float256_log1p(); break;
7589 case 512: *name_ptr = "vector_float512_log1p"; *addr_ptr = StubRoutines::vector_float512_log1p(); break;
7590 default: Unimplemented(); break;
7591 }
7592 }
7593 break;
7594 case OP_LOG: {
7595 switch(bits) {
7596 case 64: *name_ptr = "vector_float64_log"; *addr_ptr = StubRoutines::vector_float64_log(); break;
7597 case 128: *name_ptr = "vector_float128_log"; *addr_ptr = StubRoutines::vector_float128_log(); break;
7598 case 256: *name_ptr = "vector_float256_log"; *addr_ptr = StubRoutines::vector_float256_log(); break;
7599 case 512: *name_ptr = "vector_float512_log"; *addr_ptr = StubRoutines::vector_float512_log(); break;
7600 default: Unimplemented(); break;
7601 }
7602 }
7603 break;
7604 case OP_LOG10: {
7605 switch(bits) {
7606 case 64: *name_ptr = "vector_float64_log10"; *addr_ptr = StubRoutines::vector_float64_log10(); break;
7607 case 128: *name_ptr = "vector_float128_log10"; *addr_ptr = StubRoutines::vector_float128_log10(); break;
7608 case 256: *name_ptr = "vector_float256_log10"; *addr_ptr = StubRoutines::vector_float256_log10(); break;
7609 case 512: *name_ptr = "vector_float512_log10"; *addr_ptr = StubRoutines::vector_float512_log10(); break;
7610 default: Unimplemented(); break;
7611 }
7612 }
7613 break;
7614 case OP_EXPM1: {
7615 switch(bits) {
7616 case 64: *name_ptr = "vector_float64_expm1"; *addr_ptr = StubRoutines::vector_float64_expm1(); break;
7617 case 128: *name_ptr = "vector_float128_expm1"; *addr_ptr = StubRoutines::vector_float128_expm1(); break;
7618 case 256: *name_ptr = "vector_float256_expm1"; *addr_ptr = StubRoutines::vector_float256_expm1(); break;
7619 case 512: *name_ptr = "vector_float512_expm1"; *addr_ptr = StubRoutines::vector_float512_expm1(); break;
7620 default: Unimplemented(); break;
7621 }
7622 }
7623 break;
7624 case OP_SIN: {
7625 switch(bits) {
7626 case 64: *name_ptr = "vector_float64_sin"; *addr_ptr = StubRoutines::vector_float64_sin(); break;
7627 case 128: *name_ptr = "vector_float128_sin"; *addr_ptr = StubRoutines::vector_float128_sin(); break;
7628 case 256: *name_ptr = "vector_float256_sin"; *addr_ptr = StubRoutines::vector_float256_sin(); break;
7629 case 512: *name_ptr = "vector_float512_sin"; *addr_ptr = StubRoutines::vector_float512_sin(); break;
7630 default: Unimplemented(); break;
7631 }
7632 }
7633 break;
7634 case OP_COS: {
7635 switch(bits) {
7636 case 64: *name_ptr = "vector_float64_cos"; *addr_ptr = StubRoutines::vector_float64_cos(); break;
7637 case 128: *name_ptr = "vector_float128_cos"; *addr_ptr = StubRoutines::vector_float128_cos(); break;
7638 case 256: *name_ptr = "vector_float256_cos"; *addr_ptr = StubRoutines::vector_float256_cos(); break;
7639 case 512: *name_ptr = "vector_float512_cos"; *addr_ptr = StubRoutines::vector_float512_cos(); break;
7640 default: Unimplemented(); break;
7641 }
7642 }
7643 break;
7644 case OP_TAN: {
7645 switch(bits) {
7646 case 64: *name_ptr = "vector_float64_tan"; *addr_ptr = StubRoutines::vector_float64_tan(); break;
7647 case 128: *name_ptr = "vector_float128_tan"; *addr_ptr = StubRoutines::vector_float128_tan(); break;
7648 case 256: *name_ptr = "vector_float256_tan"; *addr_ptr = StubRoutines::vector_float256_tan(); break;
7649 case 512: *name_ptr = "vector_float512_tan"; *addr_ptr = StubRoutines::vector_float512_tan(); break;
7650 default: Unimplemented(); break;
7651 }
7652 }
7653 break;
7654 case OP_SINH: {
7655 switch(bits) {
7656 case 64: *name_ptr = "vector_float64_sinh"; *addr_ptr = StubRoutines::vector_float64_sinh(); break;
7657 case 128: *name_ptr = "vector_float128_sinh"; *addr_ptr = StubRoutines::vector_float128_sinh(); break;
7658 case 256: *name_ptr = "vector_float256_sinh"; *addr_ptr = StubRoutines::vector_float256_sinh(); break;
7659 case 512: *name_ptr = "vector_float512_sinh"; *addr_ptr = StubRoutines::vector_float512_sinh(); break;
7660 default: Unimplemented(); break;
7661 }
7662 }
7663 break;
7664 case OP_COSH: {
7665 switch(bits) {
7666 case 64: *name_ptr = "vector_float64_cosh"; *addr_ptr = StubRoutines::vector_float64_cosh(); break;
7667 case 128: *name_ptr = "vector_float128_cosh"; *addr_ptr = StubRoutines::vector_float128_cosh(); break;
7668 case 256: *name_ptr = "vector_float256_cosh"; *addr_ptr = StubRoutines::vector_float256_cosh(); break;
7669 case 512: *name_ptr = "vector_float512_cosh"; *addr_ptr = StubRoutines::vector_float512_cosh(); break;
7670 default: Unimplemented(); break;
7671 }
7672 }
7673 break;
7674 case OP_TANH: {
7675 switch(bits) {
7676 case 64: *name_ptr = "vector_float64_tanh"; *addr_ptr = StubRoutines::vector_float64_tanh(); break;
7677 case 128: *name_ptr = "vector_float128_tanh"; *addr_ptr = StubRoutines::vector_float128_tanh(); break;
7678 case 256: *name_ptr = "vector_float256_tanh"; *addr_ptr = StubRoutines::vector_float256_tanh(); break;
7679 case 512: *name_ptr = "vector_float512_tanh"; *addr_ptr = StubRoutines::vector_float512_tanh(); break;
7680 default: Unimplemented(); break;
7681 }
7682 }
7683 break;
7684 case OP_ASIN: {
7685 switch(bits) {
7686 case 64: *name_ptr = "vector_float64_asin"; *addr_ptr = StubRoutines::vector_float64_asin(); break;
7687 case 128: *name_ptr = "vector_float128_asin"; *addr_ptr = StubRoutines::vector_float128_asin(); break;
7688 case 256: *name_ptr = "vector_float256_asin"; *addr_ptr = StubRoutines::vector_float256_asin(); break;
7689 case 512: *name_ptr = "vector_float512_asin"; *addr_ptr = StubRoutines::vector_float512_asin(); break;
7690 default: Unimplemented(); break;
7691 }
7692 }
7693 break;
7694 case OP_ACOS: {
7695 switch(bits) {
7696 case 64: *name_ptr = "vector_float64_acos"; *addr_ptr = StubRoutines::vector_float64_acos(); break;
7697 case 128: *name_ptr = "vector_float128_acos"; *addr_ptr = StubRoutines::vector_float128_acos(); break;
7698 case 256: *name_ptr = "vector_float256_acos"; *addr_ptr = StubRoutines::vector_float256_acos(); break;
7699 case 512: *name_ptr = "vector_float512_acos"; *addr_ptr = StubRoutines::vector_float512_acos(); break;
7700 default: Unimplemented(); break;
7701 }
7702 }
7703 break;
7704 case OP_ATAN: {
7705 switch(bits) {
7706 case 64: *name_ptr = "vector_float64_atan"; *addr_ptr = StubRoutines::vector_float64_atan(); break;
7707 case 128: *name_ptr = "vector_float128_atan"; *addr_ptr = StubRoutines::vector_float128_atan(); break;
7708 case 256: *name_ptr = "vector_float256_atan"; *addr_ptr = StubRoutines::vector_float256_atan(); break;
7709 case 512: *name_ptr = "vector_float512_atan"; *addr_ptr = StubRoutines::vector_float512_atan(); break;
7710 default: Unimplemented(); break;
7711 }
7712 }
7713 break;
7714 case OP_CBRT: {
7715 switch(bits) {
7716 case 64: *name_ptr = "vector_float64_cbrt"; *addr_ptr = StubRoutines::vector_float64_cbrt(); break;
7717 case 128: *name_ptr = "vector_float128_cbrt"; *addr_ptr = StubRoutines::vector_float128_cbrt(); break;
7718 case 256: *name_ptr = "vector_float256_cbrt"; *addr_ptr = StubRoutines::vector_float256_cbrt(); break;
7719 case 512: *name_ptr = "vector_float512_cbrt"; *addr_ptr = StubRoutines::vector_float512_cbrt(); break;
7720 default: Unimplemented(); break;
7721 }
7722 }
7723 break;
7724 case OP_HYPOT: {
7725 switch(bits) {
7726 case 64: *name_ptr = "vector_float64_hypot"; *addr_ptr = StubRoutines::vector_float64_hypot(); break;
7727 case 128: *name_ptr = "vector_float128_hypot"; *addr_ptr = StubRoutines::vector_float128_hypot(); break;
7728 case 256: *name_ptr = "vector_float256_hypot"; *addr_ptr = StubRoutines::vector_float256_hypot(); break;
7729 case 512: *name_ptr = "vector_float512_hypot"; *addr_ptr = StubRoutines::vector_float512_hypot(); break;
7730 default: Unimplemented(); break;
7731 }
7732 }
7733 break;
7734 case OP_POW: {
7735 switch(bits) {
7736 case 64: *name_ptr = "vector_float64_pow"; *addr_ptr = StubRoutines::vector_float64_pow(); break;
7737 case 128: *name_ptr = "vector_float128_pow"; *addr_ptr = StubRoutines::vector_float128_pow(); break;
7738 case 256: *name_ptr = "vector_float256_pow"; *addr_ptr = StubRoutines::vector_float256_pow(); break;
7739 case 512: *name_ptr = "vector_float512_pow"; *addr_ptr = StubRoutines::vector_float512_pow(); break;
7740 default: Unimplemented(); break;
7741 }
7742 }
7743 break;
7744 case OP_ATAN2: {
7745 switch(bits) {
7746 case 64: *name_ptr = "vector_float64_atan2"; *addr_ptr = StubRoutines::vector_float64_atan2(); break;
7747 case 128: *name_ptr = "vector_float128_atan2"; *addr_ptr = StubRoutines::vector_float128_atan2(); break;
7748 case 256: *name_ptr = "vector_float256_atan2"; *addr_ptr = StubRoutines::vector_float256_atan2(); break;
7749 case 512: *name_ptr = "vector_float512_atan2"; *addr_ptr = StubRoutines::vector_float512_atan2(); break;
7750 default: Unimplemented(); break;
7751 }
7752 }
7753 break;
7754 default:
7755 *name_ptr = "invalid";
7756 *addr_ptr = NULL;
7757 break;
7758 }
7759 } else {
7760 assert(bt == T_DOUBLE, "must be FP type only");
7761 switch(op) {
7762 case OP_EXP: {
7763 switch(bits) {
7764 case 64: *name_ptr = "vector_double64_exp"; *addr_ptr = StubRoutines::vector_double64_exp(); break;
7765 case 128: *name_ptr = "vector_double128_exp"; *addr_ptr = StubRoutines::vector_double128_exp(); break;
7766 case 256: *name_ptr = "vector_double256_exp"; *addr_ptr = StubRoutines::vector_double256_exp(); break;
7767 case 512: *name_ptr = "vector_double512_exp"; *addr_ptr = StubRoutines::vector_double512_exp(); break;
7768 default: Unimplemented(); break;
7769 }
7770 }
7771 break;
7772 case OP_LOG1P: {
7773 switch(bits) {
7774 case 64: *name_ptr = "vector_double64_log1p"; *addr_ptr = StubRoutines::vector_double64_log1p(); break;
7775 case 128: *name_ptr = "vector_double128_log1p"; *addr_ptr = StubRoutines::vector_double128_log1p(); break;
7776 case 256: *name_ptr = "vector_double256_log1p"; *addr_ptr = StubRoutines::vector_double256_log1p(); break;
7777 case 512: *name_ptr = "vector_double512_log1p"; *addr_ptr = StubRoutines::vector_double512_log1p(); break;
7778 default: Unimplemented(); break;
7779 }
7780 }
7781 break;
7782 case OP_LOG: {
7783 switch(bits) {
7784 case 64: *name_ptr = "vector_double64_log"; *addr_ptr = StubRoutines::vector_double64_log(); break;
7785 case 128: *name_ptr = "vector_double128_log"; *addr_ptr = StubRoutines::vector_double128_log(); break;
7786 case 256: *name_ptr = "vector_double256_log"; *addr_ptr = StubRoutines::vector_double256_log(); break;
7787 case 512: *name_ptr = "vector_double512_log"; *addr_ptr = StubRoutines::vector_double512_log(); break;
7788 default: Unimplemented(); break;
7789 }
7790 }
7791 break;
7792 case OP_LOG10: {
7793 switch(bits) {
7794 case 64: *name_ptr = "vector_double64_log10"; *addr_ptr = StubRoutines::vector_double64_log10(); break;
7795 case 128: *name_ptr = "vector_double128_log10"; *addr_ptr = StubRoutines::vector_double128_log10(); break;
7796 case 256: *name_ptr = "vector_double256_log10"; *addr_ptr = StubRoutines::vector_double256_log10(); break;
7797 case 512: *name_ptr = "vector_double512_log10"; *addr_ptr = StubRoutines::vector_double512_log10(); break;
7798 default: Unimplemented(); break;
7799 }
7800 }
7801 break;
7802 case OP_EXPM1: {
7803 switch(bits) {
7804 case 64: *name_ptr = "vector_double64_expm1"; *addr_ptr = StubRoutines::vector_double64_expm1(); break;
7805 case 128: *name_ptr = "vector_double128_expm1"; *addr_ptr = StubRoutines::vector_double128_expm1(); break;
7806 case 256: *name_ptr = "vector_double256_expm1"; *addr_ptr = StubRoutines::vector_double256_expm1(); break;
7807 case 512: *name_ptr = "vector_double512_expm1"; *addr_ptr = StubRoutines::vector_double512_expm1(); break;
7808 default: Unimplemented(); break;
7809 }
7810 }
7811 break;
7812 case OP_SIN: {
7813 switch(bits) {
7814 case 64: *name_ptr = "vector_double64_sin"; *addr_ptr = StubRoutines::vector_double64_sin(); break;
7815 case 128: *name_ptr = "vector_double128_sin"; *addr_ptr = StubRoutines::vector_double128_sin(); break;
7816 case 256: *name_ptr = "vector_double256_sin"; *addr_ptr = StubRoutines::vector_double256_sin(); break;
7817 case 512: *name_ptr = "vector_double512_sin"; *addr_ptr = StubRoutines::vector_double512_sin(); break;
7818 default: Unimplemented(); break;
7819 }
7820 }
7821 break;
7822 case OP_COS: {
7823 switch(bits) {
7824 case 64: *name_ptr = "vector_double64_cos"; *addr_ptr = StubRoutines::vector_double64_cos(); break;
7825 case 128: *name_ptr = "vector_double128_cos"; *addr_ptr = StubRoutines::vector_double128_cos(); break;
7826 case 256: *name_ptr = "vector_double256_cos"; *addr_ptr = StubRoutines::vector_double256_cos(); break;
7827 case 512: *name_ptr = "vector_double512_cos"; *addr_ptr = StubRoutines::vector_double512_cos(); break;
7828 default: Unimplemented(); break;
7829 }
7830 }
7831 break;
7832 case OP_TAN: {
7833 switch(bits) {
7834 case 64: *name_ptr = "vector_double64_tan"; *addr_ptr = StubRoutines::vector_double64_tan(); break;
7835 case 128: *name_ptr = "vector_double128_tan"; *addr_ptr = StubRoutines::vector_double128_tan(); break;
7836 case 256: *name_ptr = "vector_double256_tan"; *addr_ptr = StubRoutines::vector_double256_tan(); break;
7837 case 512: *name_ptr = "vector_double512_tan"; *addr_ptr = StubRoutines::vector_double512_tan(); break;
7838 default: Unimplemented(); break;
7839 }
7840 }
7841 break;
7842 case OP_SINH: {
7843 switch(bits) {
7844 case 64: *name_ptr = "vector_double64_sinh"; *addr_ptr = StubRoutines::vector_double64_sinh(); break;
7845 case 128: *name_ptr = "vector_double128_sinh"; *addr_ptr = StubRoutines::vector_double128_sinh(); break;
7846 case 256: *name_ptr = "vector_double256_sinh"; *addr_ptr = StubRoutines::vector_double256_sinh(); break;
7847 case 512: *name_ptr = "vector_double512_sinh"; *addr_ptr = StubRoutines::vector_double512_sinh(); break;
7848 default: Unimplemented(); break;
7849 }
7850 }
7851 break;
7852 case OP_COSH: {
7853 switch(bits) {
7854 case 64: *name_ptr = "vector_double64_cosh"; *addr_ptr = StubRoutines::vector_double64_cosh(); break;
7855 case 128: *name_ptr = "vector_double128_cosh"; *addr_ptr = StubRoutines::vector_double128_cosh(); break;
7856 case 256: *name_ptr = "vector_double256_cosh"; *addr_ptr = StubRoutines::vector_double256_cosh(); break;
7857 case 512: *name_ptr = "vector_double512_cosh"; *addr_ptr = StubRoutines::vector_double512_cosh(); break;
7858 default: Unimplemented(); break;
7859 }
7860 }
7861 break;
7862 case OP_TANH: {
7863 switch(bits) {
7864 case 64: *name_ptr = "vector_double64_tanh"; *addr_ptr = StubRoutines::vector_double64_tanh(); break;
7865 case 128: *name_ptr = "vector_double128_tanh"; *addr_ptr = StubRoutines::vector_double128_tanh(); break;
7866 case 256: *name_ptr = "vector_double256_tanh"; *addr_ptr = StubRoutines::vector_double256_tanh(); break;
7867 case 512: *name_ptr = "vector_double512_tanh"; *addr_ptr = StubRoutines::vector_double512_tanh(); break;
7868 default: Unimplemented(); break;
7869 }
7870 }
7871 break;
7872 case OP_ASIN: {
7873 switch(bits) {
7874 case 64: *name_ptr = "vector_double64_asin"; *addr_ptr = StubRoutines::vector_double64_asin(); break;
7875 case 128: *name_ptr = "vector_double128_asin"; *addr_ptr = StubRoutines::vector_double128_asin(); break;
7876 case 256: *name_ptr = "vector_double256_asin"; *addr_ptr = StubRoutines::vector_double256_asin(); break;
7877 case 512: *name_ptr = "vector_double512_asin"; *addr_ptr = StubRoutines::vector_double512_asin(); break;
7878 default: Unimplemented(); break;
7879 }
7880 }
7881 break;
7882 case OP_ACOS: {
7883 switch(bits) {
7884 case 64: *name_ptr = "vector_double64_acos"; *addr_ptr = StubRoutines::vector_double64_acos(); break;
7885 case 128: *name_ptr = "vector_double128_acos"; *addr_ptr = StubRoutines::vector_double128_acos(); break;
7886 case 256: *name_ptr = "vector_double256_acos"; *addr_ptr = StubRoutines::vector_double256_acos(); break;
7887 case 512: *name_ptr = "vector_double512_acos"; *addr_ptr = StubRoutines::vector_double512_acos(); break;
7888 default: Unimplemented(); break;
7889 }
7890 }
7891 break;
7892 case OP_ATAN: {
7893 switch(bits) {
7894 case 64: *name_ptr = "vector_double64_atan"; *addr_ptr = StubRoutines::vector_double64_atan(); break;
7895 case 128: *name_ptr = "vector_double128_atan"; *addr_ptr = StubRoutines::vector_double128_atan(); break;
7896 case 256: *name_ptr = "vector_double256_atan"; *addr_ptr = StubRoutines::vector_double256_atan(); break;
7897 case 512: *name_ptr = "vector_double512_atan"; *addr_ptr = StubRoutines::vector_double512_atan(); break;
7898 default: Unimplemented(); break;
7899 }
7900 }
7901 break;
7902 case OP_CBRT: {
7903 switch(bits) {
7904 case 64: *name_ptr = "vector_double64_cbrt"; *addr_ptr = StubRoutines::vector_double64_cbrt(); break;
7905 case 128: *name_ptr = "vector_double128_cbrt"; *addr_ptr = StubRoutines::vector_double128_cbrt(); break;
7906 case 256: *name_ptr = "vector_double256_cbrt"; *addr_ptr = StubRoutines::vector_double256_cbrt(); break;
7907 case 512: *name_ptr = "vector_double512_cbrt"; *addr_ptr = StubRoutines::vector_double512_cbrt(); break;
7908 default: Unimplemented(); break;
7909 }
7910 }
7911 break;
7912 case OP_HYPOT: {
7913 switch(bits) {
7914 case 64: *name_ptr = "vector_double64_hypot"; *addr_ptr = StubRoutines::vector_double64_hypot(); break;
7915 case 128: *name_ptr = "vector_double128_hypot"; *addr_ptr = StubRoutines::vector_double128_hypot(); break;
7916 case 256: *name_ptr = "vector_double256_hypot"; *addr_ptr = StubRoutines::vector_double256_hypot(); break;
7917 case 512: *name_ptr = "vector_double512_hypot"; *addr_ptr = StubRoutines::vector_double512_hypot(); break;
7918 default: Unimplemented(); break;
7919 }
7920 }
7921 break;
7922 case OP_POW: {
7923 switch(bits) {
7924 case 64: *name_ptr = "vector_double64_pow"; *addr_ptr = StubRoutines::vector_double64_pow(); break;
7925 case 128: *name_ptr = "vector_double128_pow"; *addr_ptr = StubRoutines::vector_double128_pow(); break;
7926 case 256: *name_ptr = "vector_double256_pow"; *addr_ptr = StubRoutines::vector_double256_pow(); break;
7927 case 512: *name_ptr = "vector_double512_pow"; *addr_ptr = StubRoutines::vector_double512_pow(); break;
7928 default: Unimplemented(); break;
7929 }
7930 }
7931 break;
7932 case OP_ATAN2: {
7933 switch(bits) {
7934 case 64: *name_ptr = "vector_double64_atan2"; *addr_ptr = StubRoutines::vector_double64_atan2(); break;
7935 case 128: *name_ptr = "vector_double128_atan2"; *addr_ptr = StubRoutines::vector_double128_atan2(); break;
7936 case 256: *name_ptr = "vector_double256_atan2"; *addr_ptr = StubRoutines::vector_double256_atan2(); break;
7937 case 512: *name_ptr = "vector_double512_atan2"; *addr_ptr = StubRoutines::vector_double512_atan2(); break;
7938 default: Unimplemented(); break;
7939 }
7940 }
7941 break;
7942
7943 default:
7944 *name_ptr = "invalid";
7945 *addr_ptr = NULL;
7946 break;
7947 }
7948 }
7949 }
7950
7951 Node* LibraryCallKit::gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2) {
7952 assert(vector_api_op_id >= OP_SVML_START && vector_api_op_id <= OP_SVML_END, "need valid op id");
7953 assert(opd1 != NULL, "must not be null");
7954 const TypeVect* vt = TypeVect::make(bt, num_elem);
7955 const TypeFunc* call_type = OptoRuntime::Math_Vector_Vector_Type(opd2 != NULL ? 2 : 1, vt, vt);
7956 const char* name = NULL;
7957 address addr = NULL;
7958
7959 // Get address for svml method.
7960 get_svml_address(vector_api_op_id, vt->length_in_bytes() * BitsPerByte, bt, &name, &addr);
7961
7962 if (addr == NULL) {
7963 return NULL;
7964 }
7965
7966 assert(name != NULL, "name must not be null");
7967 Node* operation = make_runtime_call(RC_VECTOR,
7968 call_type,
7969 addr,
7970 name,
7971 TypePtr::BOTTOM,
7972 opd1,
7973 opd2);
7974 return _gvn.transform(new ProjNode(_gvn.transform(operation), TypeFunc::Parms));
7975 }
7976
7977 // static
7978 // <V extends Vector<?,?>>
7979 // V broadcastInt(int opr, Class<V> vectorClass, Class<?> elementType, int vlen,
7980 // V v, int i,
7981 // VectorBroadcastIntOp<V> defaultImpl) {
7982 //
7983 bool LibraryCallKit::inline_vector_broadcast_int() {
7984 const TypeInt* opr = gvn().type(argument(0))->is_int();
7985 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7986 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7987 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7988
7989 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7990 return false; // not enough info for intrinsification
7991 }
7992 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7993 if (!elem_type->is_primitive_type()) {
7994 return false; // should be primitive type
7995 }
7996 BasicType elem_bt = elem_type->basic_type();
7997 int num_elem = vlen->get_con();
7998 int opc = get_opc(opr->get_con(), elem_bt);
7999 int sopc = get_sopc(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
8000 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
8001 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
8002
8003 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
8004 return false; // not supported
8005 }
8006 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
8007 Node* opd2 = shift_count(argument(5), opc, elem_bt, num_elem);
8008 if (opd1 == NULL || opd2 == NULL) {
8009 return false;
8010 }
8011 Node* operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt));
8012
8013 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
8014 set_vector_result(vbox);
8015 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
8016 return true;
8017 }
8018
8019 bool LibraryCallKit::inline_vector_cast_reinterpret(bool is_cast) {
8020 const TypeInstPtr* vector_klass_from = gvn().type(argument(0))->is_instptr();
8021 const TypeInstPtr* elem_klass_from = gvn().type(argument(1))->is_instptr();
8022 const TypeInt* vlen_from = gvn().type(argument(2))->is_int();
8023
8024 const TypeInstPtr* vector_klass_to = gvn().type(argument(3))->is_instptr();
8025 const TypeInstPtr* elem_klass_to = gvn().type(argument(4))->is_instptr();
8026 const TypeInt* vlen_to = gvn().type(argument(5))->is_int();
8027
8028 if (vector_klass_from->const_oop() == NULL || elem_klass_from->const_oop() == NULL || !vlen_from->is_con() ||
8029 vector_klass_to->const_oop() == NULL || elem_klass_to->const_oop() == NULL || !vlen_to->is_con()) {
8030 return false; // not enough info for intrinsification
8031 }
8032
8033 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
8034 ciKlass* vbox_klass_to = vector_klass_to->const_oop()->as_instance()->java_lang_Class_klass();
8035 if (!vbox_klass_from->is_vectorapi_vector() || !vbox_klass_to->is_vectorapi_vector()) {
8036 return false; // only vector & mask are supported
8037 }
8038 bool is_mask = vbox_klass_from->is_vectormask();
8039
8040 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
8041 if (!elem_type_from->is_primitive_type()) {
8042 return false; // should be primitive type
8043 }
8044 BasicType elem_bt_from = elem_type_from->basic_type();
8045 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
8046 if (!elem_type_to->is_primitive_type()) {
8047 return false; // should be primitive type
8048 }
8049 BasicType elem_bt_to = elem_type_to->basic_type();
8050 if (is_mask && elem_bt_from != elem_bt_to) {
8051 return false; // type mismatch
8052 }
8053 int num_elem_from = vlen_from->get_con();
8054 int num_elem_to = vlen_to->get_con();
8055
8056 // Check whether we can unbox to appropriate size. Even with casting, checking for reinterpret is needed
8057 // since we may need to change size.
8058 if (!arch_supports_vector(Op_VectorReinterpret,
8059 num_elem_from,
8060 elem_bt_from,
8061 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
8062 return false;
8063 }
8064
8065 // Check whether we can support resizing/reinterpreting to the new size.
8066 if (!arch_supports_vector(Op_VectorReinterpret,
8067 num_elem_to,
8068 elem_bt_to,
8069 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
8070 return false;
8071 }
8072
8073 // At this point, we know that both input and output vector registers are supported
8074 // by the architecture. Next check if the casted type is simply to same type - which means
8075 // that it is actually a resize and not a cast.
8076 if (is_cast && elem_bt_from == elem_bt_to) {
8077 is_cast = false;
8078 }
8079
8080 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
8081
8082 Node* opd1 = unbox_vector(argument(6), vbox_type_from, elem_bt_from, num_elem_from);
8083 if (opd1 == NULL) {
8084 return false;
8085 }
8086
8087 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from);
8088 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to);
8089
8090 Node* op = opd1;
8091 if (is_cast) {
8092 assert(!is_mask, "masks cannot be casted");
8093 int cast_vopc = VectorCastNode::opcode(elem_bt_from);
8094 // Make sure that cast is implemented to particular type/size combination.
8095 if (!arch_supports_vector(cast_vopc, num_elem_to, elem_bt_to, VecMaskNotUsed)) {
8096 return false;
8097 }
8098
8099 if (num_elem_from < num_elem_to) {
8100 // Since input and output number of elements are not consistent, we need to make sure we
8101 // properly size. Thus, first make a cast that retains the number of elements from source.
8102 // In case the size exceeds the arch size, we do the minimum.
8103 int num_elem_for_cast = MIN2(num_elem_from, Matcher::max_vector_size(elem_bt_to));
8104
8105 // It is possible that arch does not support this intermediate vector size
8106 // TODO More complex logic required here to handle this corner case for the sizes.
8107 if (!arch_supports_vector(cast_vopc, num_elem_for_cast, elem_bt_to, VecMaskNotUsed)) {
8108 return false;
8109 }
8110
8111 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_for_cast));
8112 // Now ensure that the destination gets properly resized to needed size.
8113 op = _gvn.transform(new VectorReinterpretNode(op, op->bottom_type()->is_vect(), dst_type));
8114 } else if (num_elem_from > num_elem_to) {
8115 // Since number elements from input is larger than output, simply reduce size of input (we are supposed to
8116 // drop top elements anyway).
8117 int num_elem_for_resize = MAX2(num_elem_to, Matcher::min_vector_size(elem_bt_to));
8118
8119 // It is possible that arch does not support this intermediate vector size
8120 // TODO More complex logic required here to handle this corner case for the sizes.
8121 if (!arch_supports_vector(Op_VectorReinterpret,
8122 num_elem_for_resize,
8123 elem_bt_from,
8124 VecMaskNotUsed)) {
8125 return false;
8126 }
8127
8128 op = _gvn.transform(new VectorReinterpretNode(op,
8129 src_type,
8130 TypeVect::make(elem_bt_from,
8131 num_elem_for_resize)));
8132 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
8133 } else {
8134 // Since input and output number of elements match, and since we know this vector size is
8135 // supported, simply do a cast with no resize needed.
8136 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
8137 }
8138 } else if (Type::cmp(src_type, dst_type) != 0) {
8139 assert(!is_cast, "must be reinterpret");
8140 op = _gvn.transform(new VectorReinterpretNode(op, src_type, dst_type));
8141 }
8142
8143 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
8144 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
8145 set_vector_result(vbox);
8146 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem_to * type2aelembytes(elem_bt_to))));
8147 return true;
8148 }
8149
8150 bool LibraryCallKit::inline_vector_insert() {
8151 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
8152 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
8153 const TypeInt* vlen = gvn().type(argument(2))->is_int();
8154 const TypeInt* idx = gvn().type(argument(4))->is_int();
8155
8156 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) {
8157 return false; // not enough info for intrinsification
8158 }
8159 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
8160 if (!elem_type->is_primitive_type()) {
8161 return false; // should be primitive type
8162 }
8163 BasicType elem_bt = elem_type->basic_type();
8164 int num_elem = vlen->get_con();
8165 if (!arch_supports_vector(Op_VectorInsert, num_elem, elem_bt, VecMaskNotUsed)) {
8166 return false; // not supported
8167 }
8168
8169 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
8170 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
8171
8172 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
8173 if (opd == NULL) {
8174 return false;
8175 }
8176
8177 Node* insert_val = argument(5);
8178 assert(gvn().type(insert_val)->isa_long() != NULL, "expected to be long");
8179
8180 // Convert insert value back to its appropriate type.
8181 switch (elem_bt) {
8182 case T_BYTE:
8183 insert_val = gvn().transform(new ConvL2INode(insert_val));
8184 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::BYTE));
8185 break;
8186 case T_SHORT:
8187 insert_val = gvn().transform(new ConvL2INode(insert_val));
8188 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::SHORT));
8189 break;
8190 case T_INT:
8191 insert_val = gvn().transform(new ConvL2INode(insert_val));
8192 break;
8193 case T_FLOAT:
8194 insert_val = gvn().transform(new ConvL2INode(insert_val));
8195 insert_val = gvn().transform(new MoveI2FNode(insert_val));
8196 break;
8197 case T_DOUBLE:
8198 insert_val = gvn().transform(new MoveL2DNode(insert_val));
8199 break;
8200 case T_LONG:
8201 // no conversion needed
8202 break;
8203 default: fatal("%s", type2name(elem_bt)); break;
8204 }
8205
8206 Node* operation = _gvn.transform(VectorInsertNode::make(opd, insert_val, idx->get_con()));
8207 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
8208 set_vector_result(operation);
8209
8210 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
8211 return true;
8212 }
8213
8214 bool LibraryCallKit::inline_vector_extract() {
8215 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
8216 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
8217 const TypeInt* vlen = gvn().type(argument(2))->is_int();
8218 const TypeInt* idx = gvn().type(argument(4))->is_int();
8219
8220 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) {
8221 return false; // not enough info for intrinsification
8222 }
8223 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
8224 if (!elem_type->is_primitive_type()) {
8225 return false; // should be primitive type
8226 }
8227 BasicType elem_bt = elem_type->basic_type();
8228 int num_elem = vlen->get_con();
8229 int vopc = ExtractNode::opcode(elem_bt);
8230 if (!arch_supports_vector(vopc, num_elem, elem_bt, VecMaskNotUsed)) {
8231 return false; // not supported
8232 }
8233
8234 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
8235 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
8236
8237 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
8238 if (opd == NULL) {
8239 return false;
8240 }
8241
8242 Node* operation = gvn().transform(ExtractNode::make(opd, idx->get_con(), elem_bt));
8243
8244 Node* bits = NULL;
8245 switch (elem_bt) {
8246 case T_BYTE:
8247 case T_SHORT:
8248 case T_INT: {
8249 bits = gvn().transform(new ConvI2LNode(operation));
8250 break;
8251 }
8252 case T_FLOAT: {
8253 bits = gvn().transform(new MoveF2INode(operation));
8254 bits = gvn().transform(new ConvI2LNode(bits));
8255 break;
8256 }
8257 case T_DOUBLE: {
8258 bits = gvn().transform(new MoveD2LNode(operation));
8259 break;
8260 }
8261 case T_LONG: {
8262 bits = operation; // no conversion needed
8263 break;
8264 }
8265 default: fatal("%s", type2name(elem_bt));
8266 }
8267
8268 set_vector_result(bits);
8269 return true;
8270 }
8271
8272 //------------------------------get_state_from_sha_object-----------------------
8273 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
8274 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
8275 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
8276 if (sha_state == NULL) return (Node *) NULL;
8277
8278 // now have the array, need to get the start address of the state array
8279 sha_state = access_resolve(sha_state, ACCESS_WRITE);
8280 Node* state = array_element_address(sha_state, intcon(0), T_INT);
8281 return state;
8282 }
8283
8284 //------------------------------get_state_from_sha5_object-----------------------
8285 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
8286 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
8287 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
8288 if (sha_state == NULL) return (Node *) NULL;
8289
8290 // now have the array, need to get the start address of the state array
8291 sha_state = access_resolve(sha_state, ACCESS_WRITE);
8292 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
8293 return state;
8294 }
8295
8296 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
8297 // Return node representing slow path of predicate check.
8298 // the pseudo code we want to emulate with this predicate is:
8299 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
8300 //
8301 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
8302 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
8303 "need SHA1/SHA256/SHA512 instruction support");
8304 assert((uint)predicate < 3, "sanity");
8305
8306 // The receiver was checked for NULL already.
8307 Node* digestBaseObj = argument(0);
8308
8309 // get DigestBase klass for instanceOf check
8310 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
8311 assert(tinst != NULL, "digestBaseObj is null");
8312 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
8313
8314 const char* klass_SHA_name = NULL;
8315 switch (predicate) {
8316 case 0:
8317 if (UseSHA1Intrinsics) {
8318 // we want to do an instanceof comparison against the SHA class
8319 klass_SHA_name = "sun/security/provider/SHA";
8320 }
8321 break;
8322 case 1:
8323 if (UseSHA256Intrinsics) {
8324 // we want to do an instanceof comparison against the SHA2 class
8325 klass_SHA_name = "sun/security/provider/SHA2";
8326 }
8327 break;
8328 case 2:
8329 if (UseSHA512Intrinsics) {
8330 // we want to do an instanceof comparison against the SHA5 class
8331 klass_SHA_name = "sun/security/provider/SHA5";
8332 }
8333 break;
8334 default:
8335 fatal("unknown SHA intrinsic predicate: %d", predicate);
8336 }
8337
8338 ciKlass* klass_SHA = NULL;
8339 if (klass_SHA_name != NULL) {
8340 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
8341 }
8342 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
8343 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
8344 Node* ctrl = control();
8345 set_control(top()); // no intrinsic path
8346 return ctrl;
8347 }
8348 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
8349
8350 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
8351 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
8352 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
8353 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
8354
8355 return instof_false; // even if it is NULL
8356 }
8357
8358 //-------------inline_fma-----------------------------------
8359 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
8360 Node *a = NULL;
8361 Node *b = NULL;
8362 Node *c = NULL;
8363 Node* result = NULL;
8364 switch (id) {
8365 case vmIntrinsics::_fmaD:
8366 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
8367 // no receiver since it is static method
8368 a = round_double_node(argument(0));
8369 b = round_double_node(argument(2));
8370 c = round_double_node(argument(4));
8371 result = _gvn.transform(new FmaDNode(control(), a, b, c));
8372 break;
8373 case vmIntrinsics::_fmaF:
8374 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
8375 a = argument(0);
8376 b = argument(1);
8377 c = argument(2);
8378 result = _gvn.transform(new FmaFNode(control(), a, b, c));
8379 break;
8380 default:
8381 fatal_unexpected_iid(id); break;
8382 }
8383 set_result(result);
8384 return true;
8385 }
8386
8387 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) {
8388 // argument(0) is receiver
8389 Node* codePoint = argument(1);
8390 Node* n = NULL;
8391
8392 switch (id) {
8393 case vmIntrinsics::_isDigit :
8394 n = new DigitNode(control(), codePoint);
8395 break;
8396 case vmIntrinsics::_isLowerCase :
8397 n = new LowerCaseNode(control(), codePoint);
8398 break;
8399 case vmIntrinsics::_isUpperCase :
8400 n = new UpperCaseNode(control(), codePoint);
8401 break;
8402 case vmIntrinsics::_isWhitespace :
8403 n = new WhitespaceNode(control(), codePoint);
8404 break;
8405 default:
8406 fatal_unexpected_iid(id);
8407 }
8408
8409 set_result(_gvn.transform(n));
8410 return true;
8411 }
8412
8413 //------------------------------inline_fp_min_max------------------------------
8414 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) {
8415 /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416.
8416
8417 // The intrinsic should be used only when the API branches aren't predictable,
8418 // the last one performing the most important comparison. The following heuristic
8419 // uses the branch statistics to eventually bail out if necessary.
8420
8421 ciMethodData *md = callee()->method_data();
8422
8423 if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) {
8424 ciCallProfile cp = caller()->call_profile_at_bci(bci());
8425
8426 if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) {
8427 // Bail out if the call-site didn't contribute enough to the statistics.
8428 return false;
8429 }
8430
8431 uint taken = 0, not_taken = 0;
8432
8433 for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) {
8434 if (p->is_BranchData()) {
8435 taken = ((ciBranchData*)p)->taken();
8436 not_taken = ((ciBranchData*)p)->not_taken();
8437 }
8438 }
8439
8440 double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count());
8441 balance = balance < 0 ? -balance : balance;
8442 if ( balance > 0.2 ) {
8443 // Bail out if the most important branch is predictable enough.
8444 return false;
8445 }
8446 }
8447 */
8448
8449 Node *a = NULL;
8450 Node *b = NULL;
8451 Node *n = NULL;
8452 switch (id) {
8453 case vmIntrinsics::_maxF:
8454 case vmIntrinsics::_minF:
8455 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each.");
8456 a = argument(0);
8457 b = argument(1);
8458 break;
8459 case vmIntrinsics::_maxD:
8460 case vmIntrinsics::_minD:
8461 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each.");
8462 a = round_double_node(argument(0));
8463 b = round_double_node(argument(2));
8464 break;
8465 default:
8466 fatal_unexpected_iid(id);
8467 break;
8468 }
8469 if (a->is_Con() || b->is_Con()) {
8470 return false;
8471 }
8472 switch (id) {
8473 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break;
8474 case vmIntrinsics::_minF: n = new MinFNode(a, b); break;
8475 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break;
8476 case vmIntrinsics::_minD: n = new MinDNode(a, b); break;
8477 default: fatal_unexpected_iid(id); break;
8478 }
8479 set_result(_gvn.transform(n));
8480 return true;
8481 }
8482
8483 bool LibraryCallKit::inline_profileBoolean() {
8484 Node* counts = argument(1);
8485 const TypeAryPtr* ary = NULL;
8486 ciArray* aobj = NULL;
8487 if (counts->is_Con()
8488 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
8489 && (aobj = ary->const_oop()->as_array()) != NULL
8490 && (aobj->length() == 2)) {
8491 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
8492 jint false_cnt = aobj->element_value(0).as_int();
8493 jint true_cnt = aobj->element_value(1).as_int();
8494
8495 if (C->log() != NULL) {
8496 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
8497 false_cnt, true_cnt);
8498 }
8499
8500 if (false_cnt + true_cnt == 0) {
8501 // According to profile, never executed.
8502 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
8503 Deoptimization::Action_reinterpret);
8504 return true;
8505 }
8506
8507 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
8508 // is a number of each value occurrences.
8509 Node* result = argument(0);
8510 if (false_cnt == 0 || true_cnt == 0) {
8511 // According to profile, one value has been never seen.
8512 int expected_val = (false_cnt == 0) ? 1 : 0;
8513
8514 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
8515 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
8516
8517 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
8518 Node* fast_path = _gvn.transform(new IfTrueNode(check));
8519 Node* slow_path = _gvn.transform(new IfFalseNode(check));
8520
8521 { // Slow path: uncommon trap for never seen value and then reexecute
8522 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
8523 // the value has been seen at least once.
8524 PreserveJVMState pjvms(this);
8525 PreserveReexecuteState preexecs(this);
8526 jvms()->set_should_reexecute(true);
8527
8528 set_control(slow_path);
8529 set_i_o(i_o());
8530
8531 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
8532 Deoptimization::Action_reinterpret);
8533 }
8534 // The guard for never seen value enables sharpening of the result and
8535 // returning a constant. It allows to eliminate branches on the same value
8536 // later on.
8537 set_control(fast_path);
8538 result = intcon(expected_val);
8539 }
8540 // Stop profiling.
8541 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
8542 // By replacing method body with profile data (represented as ProfileBooleanNode
8543 // on IR level) we effectively disable profiling.
8544 // It enables full speed execution once optimized code is generated.
8545 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
8546 C->record_for_igvn(profile);
8547 set_result(profile);
8548 return true;
8549 } else {
8550 // Continue profiling.
8551 // Profile data isn't available at the moment. So, execute method's bytecode version.
8552 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
8553 // is compiled and counters aren't available since corresponding MethodHandle
8554 // isn't a compile-time constant.
8555 return false;
8556 }
8557 }
8558
8559 bool LibraryCallKit::inline_isCompileConstant() {
8560 Node* n = argument(0);
8561 set_result(n->is_Con() ? intcon(1) : intcon(0));
8562 return true;
8563 }