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_shuffle_to_vector();
338 bool inline_vector_shuffle_iota();
339 bool inline_vector_mem_operation(bool is_store);
340 bool inline_vector_gather_scatter(bool is_scatter);
341 bool inline_vector_reduction();
342 bool inline_vector_test();
343 bool inline_vector_blend();
344 bool inline_vector_rearrange();
345 bool inline_vector_compare();
346 bool inline_vector_broadcast_int();
347 bool inline_vector_cast_reinterpret(bool is_cast);
348 bool inline_vector_extract();
349 bool inline_vector_insert();
350 Node* box_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
351 Node* unbox_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem);
352 Node* shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem);
353 Node* gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2);
354 void set_vector_result(Node* result, bool set_res = true);
355
356 void clear_upper_avx() {
357 #ifdef X86
358 if (UseAVX >= 2) {
359 C->set_clear_upper_avx(true);
360 }
361 #endif
362 }
363 };
364
365 //---------------------------make_vm_intrinsic----------------------------
366 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
367 vmIntrinsics::ID id = m->intrinsic_id();
368 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
369
370 if (!m->is_loaded()) {
371 // Do not attempt to inline unloaded methods.
372 return NULL;
373 }
374
375 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
376 bool is_available = false;
377
378 {
379 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
380 // the compiler must transition to '_thread_in_vm' state because both
381 // methods access VM-internal data.
382 VM_ENTRY_MARK;
383 methodHandle mh(THREAD, m->get_Method());
384 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
385 !C->directive()->is_intrinsic_disabled(mh) &&
386 !vmIntrinsics::is_disabled_by_flags(mh);
387
388 }
389
390 if (is_available) {
391 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
392 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
393 return new LibraryIntrinsic(m, is_virtual,
394 vmIntrinsics::predicates_needed(id),
395 vmIntrinsics::does_virtual_dispatch(id),
396 (vmIntrinsics::ID) id);
397 } else {
398 return NULL;
399 }
400 }
401
402 //----------------------register_library_intrinsics-----------------------
403 // Initialize this file's data structures, for each Compile instance.
404 void Compile::register_library_intrinsics() {
405 // Nothing to do here.
406 }
407
408 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
409 LibraryCallKit kit(jvms, this);
410 Compile* C = kit.C;
411 int nodes = C->unique();
412 #ifndef PRODUCT
413 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
414 char buf[1000];
415 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
416 tty->print_cr("Intrinsic %s", str);
417 }
418 #endif
419 ciMethod* callee = kit.callee();
420 const int bci = kit.bci();
421
422 // Try to inline the intrinsic.
423 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
424 kit.try_to_inline(_last_predicate)) {
425 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
426 : "(intrinsic)";
427 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
428 if (C->print_intrinsics() || C->print_inlining()) {
429 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
430 }
431 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
432 if (C->log()) {
433 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
434 vmIntrinsics::name_at(intrinsic_id()),
435 (is_virtual() ? " virtual='1'" : ""),
436 C->unique() - nodes);
437 }
438 // Push the result from the inlined method onto the stack.
439 kit.push_result();
440 C->print_inlining_update(this);
441 return kit.transfer_exceptions_into_jvms();
442 }
443
444 // The intrinsic bailed out
445 if (jvms->has_method()) {
446 // Not a root compile.
447 const char* msg;
448 if (callee->intrinsic_candidate()) {
449 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
450 } else {
451 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
452 : "failed to inline (intrinsic), method not annotated";
453 }
454 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
455 if (C->print_intrinsics() || C->print_inlining()) {
456 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
457 }
458 } else {
459 // Root compile
460 ResourceMark rm;
461 stringStream msg_stream;
462 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
463 vmIntrinsics::name_at(intrinsic_id()),
464 is_virtual() ? " (virtual)" : "", bci);
465 const char *msg = msg_stream.as_string();
466 log_debug(jit, inlining)("%s", msg);
467 if (C->print_intrinsics() || C->print_inlining()) {
468 tty->print("%s", msg);
469 }
470 }
471 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
472 C->print_inlining_update(this);
473 return NULL;
474 }
475
476 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
477 LibraryCallKit kit(jvms, this);
478 Compile* C = kit.C;
479 int nodes = C->unique();
480 _last_predicate = predicate;
481 #ifndef PRODUCT
482 assert(is_predicated() && predicate < predicates_count(), "sanity");
483 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
484 char buf[1000];
485 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
486 tty->print_cr("Predicate for intrinsic %s", str);
487 }
488 #endif
489 ciMethod* callee = kit.callee();
490 const int bci = kit.bci();
491
492 Node* slow_ctl = kit.try_to_predicate(predicate);
493 if (!kit.failing()) {
494 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
495 : "(intrinsic, predicate)";
496 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
497 if (C->print_intrinsics() || C->print_inlining()) {
498 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
499 }
500 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
501 if (C->log()) {
502 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
503 vmIntrinsics::name_at(intrinsic_id()),
504 (is_virtual() ? " virtual='1'" : ""),
505 C->unique() - nodes);
506 }
507 return slow_ctl; // Could be NULL if the check folds.
508 }
509
510 // The intrinsic bailed out
511 if (jvms->has_method()) {
512 // Not a root compile.
513 const char* msg = "failed to generate predicate for intrinsic";
514 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
515 if (C->print_intrinsics() || C->print_inlining()) {
516 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
517 }
518 } else {
519 // Root compile
520 ResourceMark rm;
521 stringStream msg_stream;
522 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
523 vmIntrinsics::name_at(intrinsic_id()),
524 is_virtual() ? " (virtual)" : "", bci);
525 const char *msg = msg_stream.as_string();
526 log_debug(jit, inlining)("%s", msg);
527 if (C->print_intrinsics() || C->print_inlining()) {
528 C->print_inlining_stream()->print("%s", msg);
529 }
530 }
531 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
532 return NULL;
533 }
534
535 bool LibraryCallKit::try_to_inline(int predicate) {
536 // Handle symbolic names for otherwise undistinguished boolean switches:
537 const bool is_store = true;
538 const bool is_compress = true;
539 const bool is_static = true;
540 const bool is_volatile = true;
541
542 if (!jvms()->has_method()) {
543 // Root JVMState has a null method.
544 assert(map()->memory()->Opcode() == Op_Parm, "");
545 // Insert the memory aliasing node
546 set_all_memory(reset_memory());
547 }
548 assert(merged_memory(), "");
549
550 switch (intrinsic_id()) {
551 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
552 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
553 case vmIntrinsics::_getClass: return inline_native_getClass();
554
555 case vmIntrinsics::_dsin:
556 case vmIntrinsics::_dcos:
557 case vmIntrinsics::_dtan:
558 case vmIntrinsics::_dabs:
559 case vmIntrinsics::_datan2:
560 case vmIntrinsics::_dsqrt:
561 case vmIntrinsics::_dexp:
562 case vmIntrinsics::_dlog:
563 case vmIntrinsics::_dlog10:
564 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
565
566 case vmIntrinsics::_min:
567 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
568
569 case vmIntrinsics::_notify:
570 case vmIntrinsics::_notifyAll:
571 return inline_notify(intrinsic_id());
572
573 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
574 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
575 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
576 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
577 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
578 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
579 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
580 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
581 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
582 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
583 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
584 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
585 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
586
587 case vmIntrinsics::_arraycopy: return inline_arraycopy();
588
589 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
590 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
591 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
592 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
593
594 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
595 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
596 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
597 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
598 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
599 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
600 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
601
602 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
603 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
604
605 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
606 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
607 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
608 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
609
610 case vmIntrinsics::_compressStringC:
611 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
612 case vmIntrinsics::_inflateStringC:
613 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
614
615 case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
616 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
617 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
618 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
619 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
620 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
621 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
622 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
623 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
624
625 case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
626 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
627 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
628 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
629 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
630 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
631 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
632 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
633 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
634
635 case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
636 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
637 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
638 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
639 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
640 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
641 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
642 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
643 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
644
645 case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
646 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
647 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
648 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
649 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
650 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
651 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
652 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
653 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
654
655 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
656 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
657 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
658 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
659
660 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
661 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
662 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
663 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
664
665 case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
666 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
667 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
668 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
669 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
670 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
671 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
672 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
673 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
674
675 case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
676 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
677 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
678 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
679 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
680 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
681 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
682 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
683 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
684
685 case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
686 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
687 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
688 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
689 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
690 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
691 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
692 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
693 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
694
695 case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
696 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
697 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
698 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
699 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
700 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
701 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
702 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
703 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
704
705 case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
706 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
707 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
708 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
709 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
710
711 case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
712 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
713 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
714 case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
715 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
716 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
717 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
718 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
719 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
720 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
721 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
722 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
723 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
724 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
725 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
726 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
727 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
728 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
729 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
730 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
731
732 case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
733 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
734 case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
735 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
736 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
737 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
738 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
739 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
740 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
741 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
742 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
743 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
744 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
745 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
746 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
747
748 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
749 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
750 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
751 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
752
753 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
754 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
755 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
756 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
757 case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
758
759 case vmIntrinsics::_loadFence:
760 case vmIntrinsics::_storeFence:
761 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
762
763 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
764
765 case vmIntrinsics::_currentThread: return inline_native_currentThread();
766 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
767
768 #ifdef JFR_HAVE_INTRINSICS
769 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime");
770 case vmIntrinsics::_getClassId: return inline_native_classID();
771 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter();
772 #endif
773 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
774 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
775 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
776 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
777 case vmIntrinsics::_getLength: return inline_native_getLength();
778 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
779 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
780 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
781 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
782 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
783 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
784
785 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
786 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
787
788 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
789
790 case vmIntrinsics::_isInstance:
791 case vmIntrinsics::_getModifiers:
792 case vmIntrinsics::_isInterface:
793 case vmIntrinsics::_isArray:
794 case vmIntrinsics::_isPrimitive:
795 case vmIntrinsics::_getSuperclass:
796 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
797
798 case vmIntrinsics::_floatToRawIntBits:
799 case vmIntrinsics::_floatToIntBits:
800 case vmIntrinsics::_intBitsToFloat:
801 case vmIntrinsics::_doubleToRawLongBits:
802 case vmIntrinsics::_doubleToLongBits:
803 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
804
805 case vmIntrinsics::_numberOfLeadingZeros_i:
806 case vmIntrinsics::_numberOfLeadingZeros_l:
807 case vmIntrinsics::_numberOfTrailingZeros_i:
808 case vmIntrinsics::_numberOfTrailingZeros_l:
809 case vmIntrinsics::_bitCount_i:
810 case vmIntrinsics::_bitCount_l:
811 case vmIntrinsics::_reverseBytes_i:
812 case vmIntrinsics::_reverseBytes_l:
813 case vmIntrinsics::_reverseBytes_s:
814 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
815
816 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
817
818 case vmIntrinsics::_Reference_get: return inline_reference_get();
819
820 case vmIntrinsics::_Class_cast: return inline_Class_cast();
821
822 case vmIntrinsics::_aescrypt_encryptBlock:
823 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
824
825 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
826 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
827 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
828
829 case vmIntrinsics::_counterMode_AESCrypt:
830 return inline_counterMode_AESCrypt(intrinsic_id());
831
832 case vmIntrinsics::_sha_implCompress:
833 case vmIntrinsics::_sha2_implCompress:
834 case vmIntrinsics::_sha5_implCompress:
835 return inline_sha_implCompress(intrinsic_id());
836
837 case vmIntrinsics::_digestBase_implCompressMB:
838 return inline_digestBase_implCompressMB(predicate);
839
840 case vmIntrinsics::_multiplyToLen:
841 return inline_multiplyToLen();
842
843 case vmIntrinsics::_squareToLen:
844 return inline_squareToLen();
845
846 case vmIntrinsics::_mulAdd:
847 return inline_mulAdd();
848
849 case vmIntrinsics::_montgomeryMultiply:
850 return inline_montgomeryMultiply();
851 case vmIntrinsics::_montgomerySquare:
852 return inline_montgomerySquare();
853
854 case vmIntrinsics::_vectorizedMismatch:
855 return inline_vectorizedMismatch();
856
857 case vmIntrinsics::_ghash_processBlocks:
858 return inline_ghash_processBlocks();
859 case vmIntrinsics::_base64_encodeBlock:
860 return inline_base64_encodeBlock();
861
862 case vmIntrinsics::_encodeISOArray:
863 case vmIntrinsics::_encodeByteISOArray:
864 return inline_encodeISOArray();
865
866 case vmIntrinsics::_updateCRC32:
867 return inline_updateCRC32();
868 case vmIntrinsics::_updateBytesCRC32:
869 return inline_updateBytesCRC32();
870 case vmIntrinsics::_updateByteBufferCRC32:
871 return inline_updateByteBufferCRC32();
872
873 case vmIntrinsics::_updateBytesCRC32C:
874 return inline_updateBytesCRC32C();
875 case vmIntrinsics::_updateDirectByteBufferCRC32C:
876 return inline_updateDirectByteBufferCRC32C();
877
878 case vmIntrinsics::_updateBytesAdler32:
879 return inline_updateBytesAdler32();
880 case vmIntrinsics::_updateByteBufferAdler32:
881 return inline_updateByteBufferAdler32();
882
883 case vmIntrinsics::_profileBoolean:
884 return inline_profileBoolean();
885 case vmIntrinsics::_isCompileConstant:
886 return inline_isCompileConstant();
887
888 case vmIntrinsics::_hasNegatives:
889 return inline_hasNegatives();
890
891 case vmIntrinsics::_fmaD:
892 case vmIntrinsics::_fmaF:
893 return inline_fma(intrinsic_id());
894
895 case vmIntrinsics::_isDigit:
896 case vmIntrinsics::_isLowerCase:
897 case vmIntrinsics::_isUpperCase:
898 case vmIntrinsics::_isWhitespace:
899 return inline_character_compare(intrinsic_id());
900
901 case vmIntrinsics::_maxF:
902 case vmIntrinsics::_minF:
903 case vmIntrinsics::_maxD:
904 case vmIntrinsics::_minD:
905 return inline_fp_min_max(intrinsic_id());
906
907 case vmIntrinsics::_VectorUnaryOp:
908 return inline_vector_nary_operation(1);
909 case vmIntrinsics::_VectorBinaryOp:
910 return inline_vector_nary_operation(2);
911 case vmIntrinsics::_VectorTernaryOp:
912 return inline_vector_nary_operation(3);
913 case vmIntrinsics::_VectorBroadcastCoerced:
914 return inline_vector_broadcast_coerced();
915 case vmIntrinsics::_VectorShuffleIota:
916 return inline_vector_shuffle_iota();
917 case vmIntrinsics::_VectorShuffleToVector:
918 return inline_vector_shuffle_to_vector();
919 case vmIntrinsics::_VectorLoadOp:
920 return inline_vector_mem_operation(/*is_store=*/false);
921 case vmIntrinsics::_VectorStoreOp:
922 return inline_vector_mem_operation(/*is_store=*/true);
923 case vmIntrinsics::_VectorGatherOp:
924 return inline_vector_gather_scatter(/*is_scatter*/ false);
925 case vmIntrinsics::_VectorScatterOp:
926 return inline_vector_gather_scatter(/*is_scatter*/ true);
927 case vmIntrinsics::_VectorReductionCoerced:
928 return inline_vector_reduction();
929 case vmIntrinsics::_VectorTest:
930 return inline_vector_test();
931 case vmIntrinsics::_VectorBlend:
932 return inline_vector_blend();
933 case vmIntrinsics::_VectorRearrange:
934 return inline_vector_rearrange();
935 case vmIntrinsics::_VectorCompare:
936 return inline_vector_compare();
937 case vmIntrinsics::_VectorBroadcastInt:
938 return inline_vector_broadcast_int();
939 case vmIntrinsics::_VectorReinterpret:
940 return inline_vector_cast_reinterpret(/*is_cast*/ false);
941 case vmIntrinsics::_VectorCast:
942 return inline_vector_cast_reinterpret(/*is_cast*/ true);
943 case vmIntrinsics::_VectorInsert:
944 return inline_vector_insert();
945 case vmIntrinsics::_VectorExtract:
946 return inline_vector_extract();
947
948 default:
949 // If you get here, it may be that someone has added a new intrinsic
950 // to the list in vmSymbols.hpp without implementing it here.
951 #ifndef PRODUCT
952 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
953 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
954 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
955 }
956 #endif
957 return false;
958 }
959 }
960
961 Node* LibraryCallKit::try_to_predicate(int predicate) {
962 if (!jvms()->has_method()) {
963 // Root JVMState has a null method.
964 assert(map()->memory()->Opcode() == Op_Parm, "");
965 // Insert the memory aliasing node
966 set_all_memory(reset_memory());
967 }
968 assert(merged_memory(), "");
969
970 switch (intrinsic_id()) {
971 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
972 return inline_cipherBlockChaining_AESCrypt_predicate(false);
973 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
974 return inline_cipherBlockChaining_AESCrypt_predicate(true);
975 case vmIntrinsics::_counterMode_AESCrypt:
976 return inline_counterMode_AESCrypt_predicate();
977 case vmIntrinsics::_digestBase_implCompressMB:
978 return inline_digestBase_implCompressMB_predicate(predicate);
979
980 default:
981 // If you get here, it may be that someone has added a new intrinsic
982 // to the list in vmSymbols.hpp without implementing it here.
983 #ifndef PRODUCT
984 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
985 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
986 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
987 }
988 #endif
989 Node* slow_ctl = control();
990 set_control(top()); // No fast path instrinsic
991 return slow_ctl;
992 }
993 }
994
995 //------------------------------set_result-------------------------------
996 // Helper function for finishing intrinsics.
997 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
998 record_for_igvn(region);
999 set_control(_gvn.transform(region));
1000 set_result( _gvn.transform(value));
1001 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1002 }
1003
1004 //------------------------------generate_guard---------------------------
1005 // Helper function for generating guarded fast-slow graph structures.
1006 // The given 'test', if true, guards a slow path. If the test fails
1007 // then a fast path can be taken. (We generally hope it fails.)
1008 // In all cases, GraphKit::control() is updated to the fast path.
1009 // The returned value represents the control for the slow path.
1010 // The return value is never 'top'; it is either a valid control
1011 // or NULL if it is obvious that the slow path can never be taken.
1012 // Also, if region and the slow control are not NULL, the slow edge
1013 // is appended to the region.
1014 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1015 if (stopped()) {
1016 // Already short circuited.
1017 return NULL;
1018 }
1019
1020 // Build an if node and its projections.
1021 // If test is true we take the slow path, which we assume is uncommon.
1022 if (_gvn.type(test) == TypeInt::ZERO) {
1023 // The slow branch is never taken. No need to build this guard.
1024 return NULL;
1025 }
1026
1027 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1028
1029 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1030 if (if_slow == top()) {
1031 // The slow branch is never taken. No need to build this guard.
1032 return NULL;
1033 }
1034
1035 if (region != NULL)
1036 region->add_req(if_slow);
1037
1038 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1039 set_control(if_fast);
1040
1041 return if_slow;
1042 }
1043
1044 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1045 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1046 }
1047 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1048 return generate_guard(test, region, PROB_FAIR);
1049 }
1050
1051 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1052 Node* *pos_index) {
1053 if (stopped())
1054 return NULL; // already stopped
1055 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1056 return NULL; // index is already adequately typed
1057 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1058 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1059 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1060 if (is_neg != NULL && pos_index != NULL) {
1061 // Emulate effect of Parse::adjust_map_after_if.
1062 Node* ccast = new CastIINode(index, TypeInt::POS);
1063 ccast->set_req(0, control());
1064 (*pos_index) = _gvn.transform(ccast);
1065 }
1066 return is_neg;
1067 }
1068
1069 // Make sure that 'position' is a valid limit index, in [0..length].
1070 // There are two equivalent plans for checking this:
1071 // A. (offset + copyLength) unsigned<= arrayLength
1072 // B. offset <= (arrayLength - copyLength)
1073 // We require that all of the values above, except for the sum and
1074 // difference, are already known to be non-negative.
1075 // Plan A is robust in the face of overflow, if offset and copyLength
1076 // are both hugely positive.
1077 //
1078 // Plan B is less direct and intuitive, but it does not overflow at
1079 // all, since the difference of two non-negatives is always
1080 // representable. Whenever Java methods must perform the equivalent
1081 // check they generally use Plan B instead of Plan A.
1082 // For the moment we use Plan A.
1083 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1084 Node* subseq_length,
1085 Node* array_length,
1086 RegionNode* region) {
1087 if (stopped())
1088 return NULL; // already stopped
1089 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1090 if (zero_offset && subseq_length->eqv_uncast(array_length))
1091 return NULL; // common case of whole-array copy
1092 Node* last = subseq_length;
1093 if (!zero_offset) // last += offset
1094 last = _gvn.transform(new AddINode(last, offset));
1095 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1096 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1097 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1098 return is_over;
1099 }
1100
1101 // Emit range checks for the given String.value byte array
1102 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1103 if (stopped()) {
1104 return; // already stopped
1105 }
1106 RegionNode* bailout = new RegionNode(1);
1107 record_for_igvn(bailout);
1108 if (char_count) {
1109 // Convert char count to byte count
1110 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1111 }
1112
1113 // Offset and count must not be negative
1114 generate_negative_guard(offset, bailout);
1115 generate_negative_guard(count, bailout);
1116 // Offset + count must not exceed length of array
1117 generate_limit_guard(offset, count, load_array_length(array), bailout);
1118
1119 if (bailout->req() > 1) {
1120 PreserveJVMState pjvms(this);
1121 set_control(_gvn.transform(bailout));
1122 uncommon_trap(Deoptimization::Reason_intrinsic,
1123 Deoptimization::Action_maybe_recompile);
1124 }
1125 }
1126
1127 //--------------------------generate_current_thread--------------------
1128 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1129 ciKlass* thread_klass = env()->Thread_klass();
1130 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1131 Node* thread = _gvn.transform(new ThreadLocalNode());
1132 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1133 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1134 tls_output = thread;
1135 return threadObj;
1136 }
1137
1138
1139 //------------------------------make_string_method_node------------------------
1140 // Helper method for String intrinsic functions. This version is called with
1141 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1142 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1143 // containing the lengths of str1 and str2.
1144 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1145 Node* result = NULL;
1146 switch (opcode) {
1147 case Op_StrIndexOf:
1148 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1149 str1_start, cnt1, str2_start, cnt2, ae);
1150 break;
1151 case Op_StrComp:
1152 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1153 str1_start, cnt1, str2_start, cnt2, ae);
1154 break;
1155 case Op_StrEquals:
1156 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1157 // Use the constant length if there is one because optimized match rule may exist.
1158 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1159 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1160 break;
1161 default:
1162 ShouldNotReachHere();
1163 return NULL;
1164 }
1165
1166 // All these intrinsics have checks.
1167 C->set_has_split_ifs(true); // Has chance for split-if optimization
1168 clear_upper_avx();
1169
1170 return _gvn.transform(result);
1171 }
1172
1173 //------------------------------inline_string_compareTo------------------------
1174 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1175 Node* arg1 = argument(0);
1176 Node* arg2 = argument(1);
1177
1178 arg1 = must_be_not_null(arg1, true);
1179 arg2 = must_be_not_null(arg2, true);
1180
1181 arg1 = access_resolve(arg1, ACCESS_READ);
1182 arg2 = access_resolve(arg2, ACCESS_READ);
1183
1184 // Get start addr and length of first argument
1185 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1186 Node* arg1_cnt = load_array_length(arg1);
1187
1188 // Get start addr and length of second argument
1189 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1190 Node* arg2_cnt = load_array_length(arg2);
1191
1192 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1193 set_result(result);
1194 return true;
1195 }
1196
1197 //------------------------------inline_string_equals------------------------
1198 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1199 Node* arg1 = argument(0);
1200 Node* arg2 = argument(1);
1201
1202 // paths (plus control) merge
1203 RegionNode* region = new RegionNode(3);
1204 Node* phi = new PhiNode(region, TypeInt::BOOL);
1205
1206 if (!stopped()) {
1207
1208 arg1 = must_be_not_null(arg1, true);
1209 arg2 = must_be_not_null(arg2, true);
1210
1211 arg1 = access_resolve(arg1, ACCESS_READ);
1212 arg2 = access_resolve(arg2, ACCESS_READ);
1213
1214 // Get start addr and length of first argument
1215 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1216 Node* arg1_cnt = load_array_length(arg1);
1217
1218 // Get start addr and length of second argument
1219 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1220 Node* arg2_cnt = load_array_length(arg2);
1221
1222 // Check for arg1_cnt != arg2_cnt
1223 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1224 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1225 Node* if_ne = generate_slow_guard(bol, NULL);
1226 if (if_ne != NULL) {
1227 phi->init_req(2, intcon(0));
1228 region->init_req(2, if_ne);
1229 }
1230
1231 // Check for count == 0 is done by assembler code for StrEquals.
1232
1233 if (!stopped()) {
1234 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1235 phi->init_req(1, equals);
1236 region->init_req(1, control());
1237 }
1238 }
1239
1240 // post merge
1241 set_control(_gvn.transform(region));
1242 record_for_igvn(region);
1243
1244 set_result(_gvn.transform(phi));
1245 return true;
1246 }
1247
1248 //------------------------------inline_array_equals----------------------------
1249 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1250 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1251 Node* arg1 = argument(0);
1252 Node* arg2 = argument(1);
1253
1254 arg1 = access_resolve(arg1, ACCESS_READ);
1255 arg2 = access_resolve(arg2, ACCESS_READ);
1256
1257 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1258 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1259 clear_upper_avx();
1260
1261 return true;
1262 }
1263
1264 //------------------------------inline_hasNegatives------------------------------
1265 bool LibraryCallKit::inline_hasNegatives() {
1266 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1267 return false;
1268 }
1269
1270 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1271 // no receiver since it is static method
1272 Node* ba = argument(0);
1273 Node* offset = argument(1);
1274 Node* len = argument(2);
1275
1276 ba = must_be_not_null(ba, true);
1277
1278 // Range checks
1279 generate_string_range_check(ba, offset, len, false);
1280 if (stopped()) {
1281 return true;
1282 }
1283 ba = access_resolve(ba, ACCESS_READ);
1284 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1285 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1286 set_result(_gvn.transform(result));
1287 return true;
1288 }
1289
1290 bool LibraryCallKit::inline_preconditions_checkIndex() {
1291 Node* index = argument(0);
1292 Node* length = argument(1);
1293 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1294 return false;
1295 }
1296
1297 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1298 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1299
1300 {
1301 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1302 uncommon_trap(Deoptimization::Reason_intrinsic,
1303 Deoptimization::Action_make_not_entrant);
1304 }
1305
1306 if (stopped()) {
1307 return false;
1308 }
1309
1310 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1311 BoolTest::mask btest = BoolTest::lt;
1312 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1313 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1314 _gvn.set_type(rc, rc->Value(&_gvn));
1315 if (!rc_bool->is_Con()) {
1316 record_for_igvn(rc);
1317 }
1318 set_control(_gvn.transform(new IfTrueNode(rc)));
1319 {
1320 PreserveJVMState pjvms(this);
1321 set_control(_gvn.transform(new IfFalseNode(rc)));
1322 uncommon_trap(Deoptimization::Reason_range_check,
1323 Deoptimization::Action_make_not_entrant);
1324 }
1325
1326 if (stopped()) {
1327 return false;
1328 }
1329
1330 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1331 result->set_req(0, control());
1332 result = _gvn.transform(result);
1333 set_result(result);
1334 replace_in_map(index, result);
1335 clear_upper_avx();
1336 return true;
1337 }
1338
1339 //------------------------------inline_string_indexOf------------------------
1340 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1341 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1342 return false;
1343 }
1344 Node* src = argument(0);
1345 Node* tgt = argument(1);
1346
1347 // Make the merge point
1348 RegionNode* result_rgn = new RegionNode(4);
1349 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1350
1351 src = must_be_not_null(src, true);
1352 tgt = must_be_not_null(tgt, true);
1353
1354 src = access_resolve(src, ACCESS_READ);
1355 tgt = access_resolve(tgt, ACCESS_READ);
1356
1357 // Get start addr and length of source string
1358 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1359 Node* src_count = load_array_length(src);
1360
1361 // Get start addr and length of substring
1362 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1363 Node* tgt_count = load_array_length(tgt);
1364
1365 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1366 // Divide src size by 2 if String is UTF16 encoded
1367 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1368 }
1369 if (ae == StrIntrinsicNode::UU) {
1370 // Divide substring size by 2 if String is UTF16 encoded
1371 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1372 }
1373
1374 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1375 if (result != NULL) {
1376 result_phi->init_req(3, result);
1377 result_rgn->init_req(3, control());
1378 }
1379 set_control(_gvn.transform(result_rgn));
1380 record_for_igvn(result_rgn);
1381 set_result(_gvn.transform(result_phi));
1382
1383 return true;
1384 }
1385
1386 //-----------------------------inline_string_indexOf-----------------------
1387 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1388 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1389 return false;
1390 }
1391 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1392 return false;
1393 }
1394 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1395 Node* src = argument(0); // byte[]
1396 Node* src_count = argument(1); // char count
1397 Node* tgt = argument(2); // byte[]
1398 Node* tgt_count = argument(3); // char count
1399 Node* from_index = argument(4); // char index
1400
1401 src = must_be_not_null(src, true);
1402 tgt = must_be_not_null(tgt, true);
1403
1404 src = access_resolve(src, ACCESS_READ);
1405 tgt = access_resolve(tgt, ACCESS_READ);
1406
1407 // Multiply byte array index by 2 if String is UTF16 encoded
1408 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1409 src_count = _gvn.transform(new SubINode(src_count, from_index));
1410 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1411 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1412
1413 // Range checks
1414 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1415 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1416 if (stopped()) {
1417 return true;
1418 }
1419
1420 RegionNode* region = new RegionNode(5);
1421 Node* phi = new PhiNode(region, TypeInt::INT);
1422
1423 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1424 if (result != NULL) {
1425 // The result is index relative to from_index if substring was found, -1 otherwise.
1426 // Generate code which will fold into cmove.
1427 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1428 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1429
1430 Node* if_lt = generate_slow_guard(bol, NULL);
1431 if (if_lt != NULL) {
1432 // result == -1
1433 phi->init_req(3, result);
1434 region->init_req(3, if_lt);
1435 }
1436 if (!stopped()) {
1437 result = _gvn.transform(new AddINode(result, from_index));
1438 phi->init_req(4, result);
1439 region->init_req(4, control());
1440 }
1441 }
1442
1443 set_control(_gvn.transform(region));
1444 record_for_igvn(region);
1445 set_result(_gvn.transform(phi));
1446 clear_upper_avx();
1447
1448 return true;
1449 }
1450
1451 // Create StrIndexOfNode with fast path checks
1452 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1453 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1454 // Check for substr count > string count
1455 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1456 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1457 Node* if_gt = generate_slow_guard(bol, NULL);
1458 if (if_gt != NULL) {
1459 phi->init_req(1, intcon(-1));
1460 region->init_req(1, if_gt);
1461 }
1462 if (!stopped()) {
1463 // Check for substr count == 0
1464 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1465 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1466 Node* if_zero = generate_slow_guard(bol, NULL);
1467 if (if_zero != NULL) {
1468 phi->init_req(2, intcon(0));
1469 region->init_req(2, if_zero);
1470 }
1471 }
1472 if (!stopped()) {
1473 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1474 }
1475 return NULL;
1476 }
1477
1478 //-----------------------------inline_string_indexOfChar-----------------------
1479 bool LibraryCallKit::inline_string_indexOfChar() {
1480 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1481 return false;
1482 }
1483 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1484 return false;
1485 }
1486 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1487 Node* src = argument(0); // byte[]
1488 Node* tgt = argument(1); // tgt is int ch
1489 Node* from_index = argument(2);
1490 Node* max = argument(3);
1491
1492 src = must_be_not_null(src, true);
1493 src = access_resolve(src, ACCESS_READ);
1494
1495 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1496 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1497 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1498
1499 // Range checks
1500 generate_string_range_check(src, src_offset, src_count, true);
1501 if (stopped()) {
1502 return true;
1503 }
1504
1505 RegionNode* region = new RegionNode(3);
1506 Node* phi = new PhiNode(region, TypeInt::INT);
1507
1508 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1509 C->set_has_split_ifs(true); // Has chance for split-if optimization
1510 _gvn.transform(result);
1511
1512 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1513 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1514
1515 Node* if_lt = generate_slow_guard(bol, NULL);
1516 if (if_lt != NULL) {
1517 // result == -1
1518 phi->init_req(2, result);
1519 region->init_req(2, if_lt);
1520 }
1521 if (!stopped()) {
1522 result = _gvn.transform(new AddINode(result, from_index));
1523 phi->init_req(1, result);
1524 region->init_req(1, control());
1525 }
1526 set_control(_gvn.transform(region));
1527 record_for_igvn(region);
1528 set_result(_gvn.transform(phi));
1529
1530 return true;
1531 }
1532 //---------------------------inline_string_copy---------------------
1533 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1534 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1535 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1536 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1537 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1538 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1539 bool LibraryCallKit::inline_string_copy(bool compress) {
1540 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1541 return false;
1542 }
1543 int nargs = 5; // 2 oops, 3 ints
1544 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1545
1546 Node* src = argument(0);
1547 Node* src_offset = argument(1);
1548 Node* dst = argument(2);
1549 Node* dst_offset = argument(3);
1550 Node* length = argument(4);
1551
1552 // Check for allocation before we add nodes that would confuse
1553 // tightly_coupled_allocation()
1554 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1555
1556 // Figure out the size and type of the elements we will be copying.
1557 const Type* src_type = src->Value(&_gvn);
1558 const Type* dst_type = dst->Value(&_gvn);
1559 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1560 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1561 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1562 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1563 "Unsupported array types for inline_string_copy");
1564
1565 src = must_be_not_null(src, true);
1566 dst = must_be_not_null(dst, true);
1567
1568 // Convert char[] offsets to byte[] offsets
1569 bool convert_src = (compress && src_elem == T_BYTE);
1570 bool convert_dst = (!compress && dst_elem == T_BYTE);
1571 if (convert_src) {
1572 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1573 } else if (convert_dst) {
1574 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1575 }
1576
1577 // Range checks
1578 generate_string_range_check(src, src_offset, length, convert_src);
1579 generate_string_range_check(dst, dst_offset, length, convert_dst);
1580 if (stopped()) {
1581 return true;
1582 }
1583
1584 src = access_resolve(src, ACCESS_READ);
1585 dst = access_resolve(dst, ACCESS_WRITE);
1586
1587 Node* src_start = array_element_address(src, src_offset, src_elem);
1588 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1589 // 'src_start' points to src array + scaled offset
1590 // 'dst_start' points to dst array + scaled offset
1591 Node* count = NULL;
1592 if (compress) {
1593 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1594 } else {
1595 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1596 }
1597
1598 if (alloc != NULL) {
1599 if (alloc->maybe_set_complete(&_gvn)) {
1600 // "You break it, you buy it."
1601 InitializeNode* init = alloc->initialization();
1602 assert(init->is_complete(), "we just did this");
1603 init->set_complete_with_arraycopy();
1604 assert(dst->is_CheckCastPP(), "sanity");
1605 assert(dst->in(0)->in(0) == init, "dest pinned");
1606 }
1607 // Do not let stores that initialize this object be reordered with
1608 // a subsequent store that would make this object accessible by
1609 // other threads.
1610 // Record what AllocateNode this StoreStore protects so that
1611 // escape analysis can go from the MemBarStoreStoreNode to the
1612 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1613 // based on the escape status of the AllocateNode.
1614 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1615 }
1616 if (compress) {
1617 set_result(_gvn.transform(count));
1618 }
1619 clear_upper_avx();
1620
1621 return true;
1622 }
1623
1624 #ifdef _LP64
1625 #define XTOP ,top() /*additional argument*/
1626 #else //_LP64
1627 #define XTOP /*no additional argument*/
1628 #endif //_LP64
1629
1630 //------------------------inline_string_toBytesU--------------------------
1631 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1632 bool LibraryCallKit::inline_string_toBytesU() {
1633 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1634 return false;
1635 }
1636 // Get the arguments.
1637 Node* value = argument(0);
1638 Node* offset = argument(1);
1639 Node* length = argument(2);
1640
1641 Node* newcopy = NULL;
1642
1643 // Set the original stack and the reexecute bit for the interpreter to reexecute
1644 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1645 { PreserveReexecuteState preexecs(this);
1646 jvms()->set_should_reexecute(true);
1647
1648 // Check if a null path was taken unconditionally.
1649 value = null_check(value);
1650
1651 RegionNode* bailout = new RegionNode(1);
1652 record_for_igvn(bailout);
1653
1654 // Range checks
1655 generate_negative_guard(offset, bailout);
1656 generate_negative_guard(length, bailout);
1657 generate_limit_guard(offset, length, load_array_length(value), bailout);
1658 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1659 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1660
1661 if (bailout->req() > 1) {
1662 PreserveJVMState pjvms(this);
1663 set_control(_gvn.transform(bailout));
1664 uncommon_trap(Deoptimization::Reason_intrinsic,
1665 Deoptimization::Action_maybe_recompile);
1666 }
1667 if (stopped()) {
1668 return true;
1669 }
1670
1671 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1672 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1673 newcopy = new_array(klass_node, size, 0); // no arguments to push
1674 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1675
1676 // Calculate starting addresses.
1677 value = access_resolve(value, ACCESS_READ);
1678 Node* src_start = array_element_address(value, offset, T_CHAR);
1679 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1680
1681 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1682 const TypeInt* toffset = gvn().type(offset)->is_int();
1683 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1684
1685 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1686 const char* copyfunc_name = "arraycopy";
1687 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1688 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1689 OptoRuntime::fast_arraycopy_Type(),
1690 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1691 src_start, dst_start, ConvI2X(length) XTOP);
1692 // Do not let reads from the cloned object float above the arraycopy.
1693 if (alloc != NULL) {
1694 if (alloc->maybe_set_complete(&_gvn)) {
1695 // "You break it, you buy it."
1696 InitializeNode* init = alloc->initialization();
1697 assert(init->is_complete(), "we just did this");
1698 init->set_complete_with_arraycopy();
1699 assert(newcopy->is_CheckCastPP(), "sanity");
1700 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1701 }
1702 // Do not let stores that initialize this object be reordered with
1703 // a subsequent store that would make this object accessible by
1704 // other threads.
1705 // Record what AllocateNode this StoreStore protects so that
1706 // escape analysis can go from the MemBarStoreStoreNode to the
1707 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1708 // based on the escape status of the AllocateNode.
1709 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1710 } else {
1711 insert_mem_bar(Op_MemBarCPUOrder);
1712 }
1713 } // original reexecute is set back here
1714
1715 C->set_has_split_ifs(true); // Has chance for split-if optimization
1716 if (!stopped()) {
1717 set_result(newcopy);
1718 }
1719 clear_upper_avx();
1720
1721 return true;
1722 }
1723
1724 //------------------------inline_string_getCharsU--------------------------
1725 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1726 bool LibraryCallKit::inline_string_getCharsU() {
1727 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1728 return false;
1729 }
1730
1731 // Get the arguments.
1732 Node* src = argument(0);
1733 Node* src_begin = argument(1);
1734 Node* src_end = argument(2); // exclusive offset (i < src_end)
1735 Node* dst = argument(3);
1736 Node* dst_begin = argument(4);
1737
1738 // Check for allocation before we add nodes that would confuse
1739 // tightly_coupled_allocation()
1740 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1741
1742 // Check if a null path was taken unconditionally.
1743 src = null_check(src);
1744 dst = null_check(dst);
1745 if (stopped()) {
1746 return true;
1747 }
1748
1749 // Get length and convert char[] offset to byte[] offset
1750 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1751 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1752
1753 // Range checks
1754 generate_string_range_check(src, src_begin, length, true);
1755 generate_string_range_check(dst, dst_begin, length, false);
1756 if (stopped()) {
1757 return true;
1758 }
1759
1760 if (!stopped()) {
1761 src = access_resolve(src, ACCESS_READ);
1762 dst = access_resolve(dst, ACCESS_WRITE);
1763
1764 // Calculate starting addresses.
1765 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1766 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1767
1768 // Check if array addresses are aligned to HeapWordSize
1769 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1770 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1771 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1772 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1773
1774 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1775 const char* copyfunc_name = "arraycopy";
1776 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1777 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1778 OptoRuntime::fast_arraycopy_Type(),
1779 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1780 src_start, dst_start, ConvI2X(length) XTOP);
1781 // Do not let reads from the cloned object float above the arraycopy.
1782 if (alloc != NULL) {
1783 if (alloc->maybe_set_complete(&_gvn)) {
1784 // "You break it, you buy it."
1785 InitializeNode* init = alloc->initialization();
1786 assert(init->is_complete(), "we just did this");
1787 init->set_complete_with_arraycopy();
1788 assert(dst->is_CheckCastPP(), "sanity");
1789 assert(dst->in(0)->in(0) == init, "dest pinned");
1790 }
1791 // Do not let stores that initialize this object be reordered with
1792 // a subsequent store that would make this object accessible by
1793 // other threads.
1794 // Record what AllocateNode this StoreStore protects so that
1795 // escape analysis can go from the MemBarStoreStoreNode to the
1796 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1797 // based on the escape status of the AllocateNode.
1798 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1799 } else {
1800 insert_mem_bar(Op_MemBarCPUOrder);
1801 }
1802 }
1803
1804 C->set_has_split_ifs(true); // Has chance for split-if optimization
1805 return true;
1806 }
1807
1808 //----------------------inline_string_char_access----------------------------
1809 // Store/Load char to/from byte[] array.
1810 // static void StringUTF16.putChar(byte[] val, int index, int c)
1811 // static char StringUTF16.getChar(byte[] val, int index)
1812 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1813 Node* value = argument(0);
1814 Node* index = argument(1);
1815 Node* ch = is_store ? argument(2) : NULL;
1816
1817 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1818 // correctly requires matched array shapes.
1819 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1820 "sanity: byte[] and char[] bases agree");
1821 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1822 "sanity: byte[] and char[] scales agree");
1823
1824 // Bail when getChar over constants is requested: constant folding would
1825 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1826 // Java method would constant fold nicely instead.
1827 if (!is_store && value->is_Con() && index->is_Con()) {
1828 return false;
1829 }
1830
1831 value = must_be_not_null(value, true);
1832 value = access_resolve(value, is_store ? ACCESS_WRITE : ACCESS_READ);
1833
1834 Node* adr = array_element_address(value, index, T_CHAR);
1835 if (adr->is_top()) {
1836 return false;
1837 }
1838 if (is_store) {
1839 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED);
1840 } else {
1841 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD);
1842 set_result(ch);
1843 }
1844 return true;
1845 }
1846
1847 //--------------------------round_double_node--------------------------------
1848 // Round a double node if necessary.
1849 Node* LibraryCallKit::round_double_node(Node* n) {
1850 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1851 n = _gvn.transform(new RoundDoubleNode(0, n));
1852 return n;
1853 }
1854
1855 //------------------------------inline_math-----------------------------------
1856 // public static double Math.abs(double)
1857 // public static double Math.sqrt(double)
1858 // public static double Math.log(double)
1859 // public static double Math.log10(double)
1860 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1861 Node* arg = round_double_node(argument(0));
1862 Node* n = NULL;
1863 switch (id) {
1864 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1865 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1866 default: fatal_unexpected_iid(id); break;
1867 }
1868 set_result(_gvn.transform(n));
1869 return true;
1870 }
1871
1872 //------------------------------runtime_math-----------------------------
1873 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1874 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1875 "must be (DD)D or (D)D type");
1876
1877 // Inputs
1878 Node* a = round_double_node(argument(0));
1879 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1880
1881 const TypePtr* no_memory_effects = NULL;
1882 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1883 no_memory_effects,
1884 a, top(), b, b ? top() : NULL);
1885 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1886 #ifdef ASSERT
1887 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1888 assert(value_top == top(), "second value must be top");
1889 #endif
1890
1891 set_result(value);
1892 return true;
1893 }
1894
1895 //------------------------------inline_math_native-----------------------------
1896 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1897 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1898 switch (id) {
1899 // These intrinsics are not properly supported on all hardware
1900 case vmIntrinsics::_dsin:
1901 return StubRoutines::dsin() != NULL ?
1902 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1903 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1904 case vmIntrinsics::_dcos:
1905 return StubRoutines::dcos() != NULL ?
1906 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1907 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1908 case vmIntrinsics::_dtan:
1909 return StubRoutines::dtan() != NULL ?
1910 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1911 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1912 case vmIntrinsics::_dlog:
1913 return StubRoutines::dlog() != NULL ?
1914 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1915 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1916 case vmIntrinsics::_dlog10:
1917 return StubRoutines::dlog10() != NULL ?
1918 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1919 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1920
1921 // These intrinsics are supported on all hardware
1922 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1923 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1924
1925 case vmIntrinsics::_dexp:
1926 return StubRoutines::dexp() != NULL ?
1927 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1928 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1929 case vmIntrinsics::_dpow: {
1930 Node* exp = round_double_node(argument(2));
1931 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1932 if (d != NULL && d->getd() == 2.0) {
1933 // Special case: pow(x, 2.0) => x * x
1934 Node* base = round_double_node(argument(0));
1935 set_result(_gvn.transform(new MulDNode(base, base)));
1936 return true;
1937 }
1938 return StubRoutines::dpow() != NULL ?
1939 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1940 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1941 }
1942 #undef FN_PTR
1943
1944 // These intrinsics are not yet correctly implemented
1945 case vmIntrinsics::_datan2:
1946 return false;
1947
1948 default:
1949 fatal_unexpected_iid(id);
1950 return false;
1951 }
1952 }
1953
1954 static bool is_simple_name(Node* n) {
1955 return (n->req() == 1 // constant
1956 || (n->is_Type() && n->as_Type()->type()->singleton())
1957 || n->is_Proj() // parameter or return value
1958 || n->is_Phi() // local of some sort
1959 );
1960 }
1961
1962 //----------------------------inline_notify-----------------------------------*
1963 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1964 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1965 address func;
1966 if (id == vmIntrinsics::_notify) {
1967 func = OptoRuntime::monitor_notify_Java();
1968 } else {
1969 func = OptoRuntime::monitor_notifyAll_Java();
1970 }
1971 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1972 make_slow_call_ex(call, env()->Throwable_klass(), false);
1973 return true;
1974 }
1975
1976
1977 //----------------------------inline_min_max-----------------------------------
1978 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1979 set_result(generate_min_max(id, argument(0), argument(1)));
1980 return true;
1981 }
1982
1983 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1984 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1985 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1986 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1987 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1988
1989 {
1990 PreserveJVMState pjvms(this);
1991 PreserveReexecuteState preexecs(this);
1992 jvms()->set_should_reexecute(true);
1993
1994 set_control(slow_path);
1995 set_i_o(i_o());
1996
1997 uncommon_trap(Deoptimization::Reason_intrinsic,
1998 Deoptimization::Action_none);
1999 }
2000
2001 set_control(fast_path);
2002 set_result(math);
2003 }
2004
2005 template <typename OverflowOp>
2006 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2007 typedef typename OverflowOp::MathOp MathOp;
2008
2009 MathOp* mathOp = new MathOp(arg1, arg2);
2010 Node* operation = _gvn.transform( mathOp );
2011 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2012 inline_math_mathExact(operation, ofcheck);
2013 return true;
2014 }
2015
2016 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2017 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2018 }
2019
2020 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2021 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2022 }
2023
2024 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2025 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2026 }
2027
2028 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2029 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2030 }
2031
2032 bool LibraryCallKit::inline_math_negateExactI() {
2033 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2034 }
2035
2036 bool LibraryCallKit::inline_math_negateExactL() {
2037 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2038 }
2039
2040 bool LibraryCallKit::inline_math_multiplyExactI() {
2041 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2042 }
2043
2044 bool LibraryCallKit::inline_math_multiplyExactL() {
2045 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2046 }
2047
2048 bool LibraryCallKit::inline_math_multiplyHigh() {
2049 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
2050 return true;
2051 }
2052
2053 Node*
2054 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2055 // These are the candidate return value:
2056 Node* xvalue = x0;
2057 Node* yvalue = y0;
2058
2059 if (xvalue == yvalue) {
2060 return xvalue;
2061 }
2062
2063 bool want_max = (id == vmIntrinsics::_max);
2064
2065 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2066 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2067 if (txvalue == NULL || tyvalue == NULL) return top();
2068 // This is not really necessary, but it is consistent with a
2069 // hypothetical MaxINode::Value method:
2070 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2071
2072 // %%% This folding logic should (ideally) be in a different place.
2073 // Some should be inside IfNode, and there to be a more reliable
2074 // transformation of ?: style patterns into cmoves. We also want
2075 // more powerful optimizations around cmove and min/max.
2076
2077 // Try to find a dominating comparison of these guys.
2078 // It can simplify the index computation for Arrays.copyOf
2079 // and similar uses of System.arraycopy.
2080 // First, compute the normalized version of CmpI(x, y).
2081 int cmp_op = Op_CmpI;
2082 Node* xkey = xvalue;
2083 Node* ykey = yvalue;
2084 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2085 if (ideal_cmpxy->is_Cmp()) {
2086 // E.g., if we have CmpI(length - offset, count),
2087 // it might idealize to CmpI(length, count + offset)
2088 cmp_op = ideal_cmpxy->Opcode();
2089 xkey = ideal_cmpxy->in(1);
2090 ykey = ideal_cmpxy->in(2);
2091 }
2092
2093 // Start by locating any relevant comparisons.
2094 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2095 Node* cmpxy = NULL;
2096 Node* cmpyx = NULL;
2097 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2098 Node* cmp = start_from->fast_out(k);
2099 if (cmp->outcnt() > 0 && // must have prior uses
2100 cmp->in(0) == NULL && // must be context-independent
2101 cmp->Opcode() == cmp_op) { // right kind of compare
2102 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2103 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2104 }
2105 }
2106
2107 const int NCMPS = 2;
2108 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2109 int cmpn;
2110 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2111 if (cmps[cmpn] != NULL) break; // find a result
2112 }
2113 if (cmpn < NCMPS) {
2114 // Look for a dominating test that tells us the min and max.
2115 int depth = 0; // Limit search depth for speed
2116 Node* dom = control();
2117 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2118 if (++depth >= 100) break;
2119 Node* ifproj = dom;
2120 if (!ifproj->is_Proj()) continue;
2121 Node* iff = ifproj->in(0);
2122 if (!iff->is_If()) continue;
2123 Node* bol = iff->in(1);
2124 if (!bol->is_Bool()) continue;
2125 Node* cmp = bol->in(1);
2126 if (cmp == NULL) continue;
2127 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2128 if (cmps[cmpn] == cmp) break;
2129 if (cmpn == NCMPS) continue;
2130 BoolTest::mask btest = bol->as_Bool()->_test._test;
2131 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2132 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2133 // At this point, we know that 'x btest y' is true.
2134 switch (btest) {
2135 case BoolTest::eq:
2136 // They are proven equal, so we can collapse the min/max.
2137 // Either value is the answer. Choose the simpler.
2138 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2139 return yvalue;
2140 return xvalue;
2141 case BoolTest::lt: // x < y
2142 case BoolTest::le: // x <= y
2143 return (want_max ? yvalue : xvalue);
2144 case BoolTest::gt: // x > y
2145 case BoolTest::ge: // x >= y
2146 return (want_max ? xvalue : yvalue);
2147 default:
2148 break;
2149 }
2150 }
2151 }
2152
2153 // We failed to find a dominating test.
2154 // Let's pick a test that might GVN with prior tests.
2155 Node* best_bol = NULL;
2156 BoolTest::mask best_btest = BoolTest::illegal;
2157 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2158 Node* cmp = cmps[cmpn];
2159 if (cmp == NULL) continue;
2160 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2161 Node* bol = cmp->fast_out(j);
2162 if (!bol->is_Bool()) continue;
2163 BoolTest::mask btest = bol->as_Bool()->_test._test;
2164 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2165 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2166 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2167 best_bol = bol->as_Bool();
2168 best_btest = btest;
2169 }
2170 }
2171 }
2172
2173 Node* answer_if_true = NULL;
2174 Node* answer_if_false = NULL;
2175 switch (best_btest) {
2176 default:
2177 if (cmpxy == NULL)
2178 cmpxy = ideal_cmpxy;
2179 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2180 // and fall through:
2181 case BoolTest::lt: // x < y
2182 case BoolTest::le: // x <= y
2183 answer_if_true = (want_max ? yvalue : xvalue);
2184 answer_if_false = (want_max ? xvalue : yvalue);
2185 break;
2186 case BoolTest::gt: // x > y
2187 case BoolTest::ge: // x >= y
2188 answer_if_true = (want_max ? xvalue : yvalue);
2189 answer_if_false = (want_max ? yvalue : xvalue);
2190 break;
2191 }
2192
2193 jint hi, lo;
2194 if (want_max) {
2195 // We can sharpen the minimum.
2196 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2197 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2198 } else {
2199 // We can sharpen the maximum.
2200 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2201 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2202 }
2203
2204 // Use a flow-free graph structure, to avoid creating excess control edges
2205 // which could hinder other optimizations.
2206 // Since Math.min/max is often used with arraycopy, we want
2207 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2208 Node* cmov = CMoveNode::make(NULL, best_bol,
2209 answer_if_false, answer_if_true,
2210 TypeInt::make(lo, hi, widen));
2211
2212 return _gvn.transform(cmov);
2213
2214 /*
2215 // This is not as desirable as it may seem, since Min and Max
2216 // nodes do not have a full set of optimizations.
2217 // And they would interfere, anyway, with 'if' optimizations
2218 // and with CMoveI canonical forms.
2219 switch (id) {
2220 case vmIntrinsics::_min:
2221 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2222 case vmIntrinsics::_max:
2223 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2224 default:
2225 ShouldNotReachHere();
2226 }
2227 */
2228 }
2229
2230 inline int
2231 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2232 const TypePtr* base_type = TypePtr::NULL_PTR;
2233 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2234 if (base_type == NULL) {
2235 // Unknown type.
2236 return Type::AnyPtr;
2237 } else if (base_type == TypePtr::NULL_PTR) {
2238 // Since this is a NULL+long form, we have to switch to a rawptr.
2239 base = _gvn.transform(new CastX2PNode(offset));
2240 offset = MakeConX(0);
2241 return Type::RawPtr;
2242 } else if (base_type->base() == Type::RawPtr) {
2243 return Type::RawPtr;
2244 } else if (base_type->isa_oopptr()) {
2245 // Base is never null => always a heap address.
2246 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2247 return Type::OopPtr;
2248 }
2249 // Offset is small => always a heap address.
2250 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2251 if (offset_type != NULL &&
2252 base_type->offset() == 0 && // (should always be?)
2253 offset_type->_lo >= 0 &&
2254 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2255 return Type::OopPtr;
2256 } else if (type == T_OBJECT) {
2257 // off heap access to an oop doesn't make any sense. Has to be on
2258 // heap.
2259 return Type::OopPtr;
2260 }
2261 // Otherwise, it might either be oop+off or NULL+addr.
2262 return Type::AnyPtr;
2263 } else {
2264 // No information:
2265 return Type::AnyPtr;
2266 }
2267 }
2268
2269 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) {
2270 Node* uncasted_base = base;
2271 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2272 if (kind == Type::RawPtr) {
2273 return basic_plus_adr(top(), uncasted_base, offset);
2274 } else if (kind == Type::AnyPtr) {
2275 assert(base == uncasted_base, "unexpected base change");
2276 if (can_cast) {
2277 if (!_gvn.type(base)->speculative_maybe_null() &&
2278 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2279 // According to profiling, this access is always on
2280 // heap. Casting the base to not null and thus avoiding membars
2281 // around the access should allow better optimizations
2282 Node* null_ctl = top();
2283 base = null_check_oop(base, &null_ctl, true, true, true);
2284 assert(null_ctl->is_top(), "no null control here");
2285 return basic_plus_adr(base, offset);
2286 } else if (_gvn.type(base)->speculative_always_null() &&
2287 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2288 // According to profiling, this access is always off
2289 // heap.
2290 base = null_assert(base);
2291 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2292 offset = MakeConX(0);
2293 return basic_plus_adr(top(), raw_base, offset);
2294 }
2295 }
2296 // We don't know if it's an on heap or off heap access. Fall back
2297 // to raw memory access.
2298 base = access_resolve(base, decorators);
2299 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2300 return basic_plus_adr(top(), raw, offset);
2301 } else {
2302 assert(base == uncasted_base, "unexpected base change");
2303 // We know it's an on heap access so base can't be null
2304 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2305 base = must_be_not_null(base, true);
2306 }
2307 return basic_plus_adr(base, offset);
2308 }
2309 }
2310
2311 //--------------------------inline_number_methods-----------------------------
2312 // inline int Integer.numberOfLeadingZeros(int)
2313 // inline int Long.numberOfLeadingZeros(long)
2314 //
2315 // inline int Integer.numberOfTrailingZeros(int)
2316 // inline int Long.numberOfTrailingZeros(long)
2317 //
2318 // inline int Integer.bitCount(int)
2319 // inline int Long.bitCount(long)
2320 //
2321 // inline char Character.reverseBytes(char)
2322 // inline short Short.reverseBytes(short)
2323 // inline int Integer.reverseBytes(int)
2324 // inline long Long.reverseBytes(long)
2325 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2326 Node* arg = argument(0);
2327 Node* n = NULL;
2328 switch (id) {
2329 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2330 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2331 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2332 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2333 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2334 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2335 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2336 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2337 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2338 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2339 default: fatal_unexpected_iid(id); break;
2340 }
2341 set_result(_gvn.transform(n));
2342 return true;
2343 }
2344
2345 //----------------------------inline_unsafe_access----------------------------
2346
2347 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2348 // Attempt to infer a sharper value type from the offset and base type.
2349 ciKlass* sharpened_klass = NULL;
2350
2351 // See if it is an instance field, with an object type.
2352 if (alias_type->field() != NULL) {
2353 if (alias_type->field()->type()->is_klass()) {
2354 sharpened_klass = alias_type->field()->type()->as_klass();
2355 }
2356 }
2357
2358 // See if it is a narrow oop array.
2359 if (adr_type->isa_aryptr()) {
2360 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2361 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2362 if (elem_type != NULL) {
2363 sharpened_klass = elem_type->klass();
2364 }
2365 }
2366 }
2367
2368 // The sharpened class might be unloaded if there is no class loader
2369 // contraint in place.
2370 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2371 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2372
2373 #ifndef PRODUCT
2374 if (C->print_intrinsics() || C->print_inlining()) {
2375 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2376 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2377 }
2378 #endif
2379 // Sharpen the value type.
2380 return tjp;
2381 }
2382 return NULL;
2383 }
2384
2385 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) {
2386 switch (kind) {
2387 case Relaxed:
2388 return MO_UNORDERED;
2389 case Opaque:
2390 return MO_RELAXED;
2391 case Acquire:
2392 return MO_ACQUIRE;
2393 case Release:
2394 return MO_RELEASE;
2395 case Volatile:
2396 return MO_SEQ_CST;
2397 default:
2398 ShouldNotReachHere();
2399 return 0;
2400 }
2401 }
2402
2403 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2404 if (callee()->is_static()) return false; // caller must have the capability!
2405 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2406 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2407 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2408 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2409
2410 if (type == T_OBJECT || type == T_ARRAY) {
2411 decorators |= ON_UNKNOWN_OOP_REF;
2412 }
2413
2414 if (unaligned) {
2415 decorators |= C2_UNALIGNED;
2416 }
2417
2418 #ifndef PRODUCT
2419 {
2420 ResourceMark rm;
2421 // Check the signatures.
2422 ciSignature* sig = callee()->signature();
2423 #ifdef ASSERT
2424 if (!is_store) {
2425 // Object getReference(Object base, int/long offset), etc.
2426 BasicType rtype = sig->return_type()->basic_type();
2427 assert(rtype == type, "getter must return the expected value");
2428 assert(sig->count() == 2, "oop getter has 2 arguments");
2429 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2430 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2431 } else {
2432 // void putReference(Object base, int/long offset, Object x), etc.
2433 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2434 assert(sig->count() == 3, "oop putter has 3 arguments");
2435 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2436 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2437 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2438 assert(vtype == type, "putter must accept the expected value");
2439 }
2440 #endif // ASSERT
2441 }
2442 #endif //PRODUCT
2443
2444 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2445
2446 Node* receiver = argument(0); // type: oop
2447
2448 // Build address expression.
2449 Node* adr;
2450 Node* heap_base_oop = top();
2451 Node* offset = top();
2452 Node* val;
2453
2454 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2455 Node* base = argument(1); // type: oop
2456 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2457 offset = argument(2); // type: long
2458 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2459 // to be plain byte offsets, which are also the same as those accepted
2460 // by oopDesc::field_addr.
2461 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2462 "fieldOffset must be byte-scaled");
2463 // 32-bit machines ignore the high half!
2464 offset = ConvL2X(offset);
2465 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed);
2466
2467 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2468 heap_base_oop = base;
2469 } else if (type == T_OBJECT) {
2470 return false; // off-heap oop accesses are not supported
2471 }
2472
2473 // Can base be NULL? Otherwise, always on-heap access.
2474 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base));
2475
2476 if (!can_access_non_heap) {
2477 decorators |= IN_HEAP;
2478 }
2479
2480 val = is_store ? argument(4) : NULL;
2481
2482 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2483
2484 // Try to categorize the address.
2485 Compile::AliasType* alias_type = C->alias_type(adr_type);
2486 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2487
2488 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2489 alias_type->adr_type() == TypeAryPtr::RANGE) {
2490 return false; // not supported
2491 }
2492
2493 bool mismatched = false;
2494 BasicType bt = alias_type->basic_type();
2495 if (bt != T_ILLEGAL) {
2496 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2497 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2498 // Alias type doesn't differentiate between byte[] and boolean[]).
2499 // Use address type to get the element type.
2500 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2501 }
2502 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2503 // accessing an array field with getReference is not a mismatch
2504 bt = T_OBJECT;
2505 }
2506 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2507 // Don't intrinsify mismatched object accesses
2508 return false;
2509 }
2510 mismatched = (bt != type);
2511 } else if (alias_type->adr_type()->isa_oopptr()) {
2512 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2513 }
2514
2515 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2516
2517 if (mismatched) {
2518 decorators |= C2_MISMATCHED;
2519 }
2520
2521 // First guess at the value type.
2522 const Type *value_type = Type::get_const_basic_type(type);
2523
2524 // Figure out the memory ordering.
2525 decorators |= mo_decorator_for_access_kind(kind);
2526
2527 if (!is_store && type == T_OBJECT) {
2528 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2529 if (tjp != NULL) {
2530 value_type = tjp;
2531 }
2532 }
2533
2534 receiver = null_check(receiver);
2535 if (stopped()) {
2536 return true;
2537 }
2538 // Heap pointers get a null-check from the interpreter,
2539 // as a courtesy. However, this is not guaranteed by Unsafe,
2540 // and it is not possible to fully distinguish unintended nulls
2541 // from intended ones in this API.
2542
2543 if (!is_store) {
2544 Node* p = NULL;
2545 // Try to constant fold a load from a constant field
2546 ciField* field = alias_type->field();
2547 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2548 // final or stable field
2549 p = make_constant_from_field(field, heap_base_oop);
2550 }
2551
2552 if (p == NULL) { // Could not constant fold the load
2553 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators);
2554 // Normalize the value returned by getBoolean in the following cases
2555 if (type == T_BOOLEAN &&
2556 (mismatched ||
2557 heap_base_oop == top() || // - heap_base_oop is NULL or
2558 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL
2559 // and the unsafe access is made to large offset
2560 // (i.e., larger than the maximum offset necessary for any
2561 // field access)
2562 ) {
2563 IdealKit ideal = IdealKit(this);
2564 #define __ ideal.
2565 IdealVariable normalized_result(ideal);
2566 __ declarations_done();
2567 __ set(normalized_result, p);
2568 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2569 __ set(normalized_result, ideal.ConI(1));
2570 ideal.end_if();
2571 final_sync(ideal);
2572 p = __ value(normalized_result);
2573 #undef __
2574 }
2575 }
2576 if (type == T_ADDRESS) {
2577 p = gvn().transform(new CastP2XNode(NULL, p));
2578 p = ConvX2UL(p);
2579 }
2580 // The load node has the control of the preceding MemBarCPUOrder. All
2581 // following nodes will have the control of the MemBarCPUOrder inserted at
2582 // the end of this method. So, pushing the load onto the stack at a later
2583 // point is fine.
2584 set_result(p);
2585 } else {
2586 if (bt == T_ADDRESS) {
2587 // Repackage the long as a pointer.
2588 val = ConvL2X(val);
2589 val = gvn().transform(new CastX2PNode(val));
2590 }
2591 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators);
2592 }
2593
2594 return true;
2595 }
2596
2597 //----------------------------inline_unsafe_load_store----------------------------
2598 // This method serves a couple of different customers (depending on LoadStoreKind):
2599 //
2600 // LS_cmp_swap:
2601 //
2602 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x);
2603 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2604 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2605 //
2606 // LS_cmp_swap_weak:
2607 //
2608 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x);
2609 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x);
2610 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x);
2611 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x);
2612 //
2613 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2614 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2615 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2616 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2617 //
2618 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2619 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2620 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2621 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2622 //
2623 // LS_cmp_exchange:
2624 //
2625 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x);
2626 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x);
2627 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x);
2628 //
2629 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2630 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2631 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2632 //
2633 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2634 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2635 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2636 //
2637 // LS_get_add:
2638 //
2639 // int getAndAddInt( Object o, long offset, int delta)
2640 // long getAndAddLong(Object o, long offset, long delta)
2641 //
2642 // LS_get_set:
2643 //
2644 // int getAndSet(Object o, long offset, int newValue)
2645 // long getAndSet(Object o, long offset, long newValue)
2646 // Object getAndSet(Object o, long offset, Object newValue)
2647 //
2648 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2649 // This basic scheme here is the same as inline_unsafe_access, but
2650 // differs in enough details that combining them would make the code
2651 // overly confusing. (This is a true fact! I originally combined
2652 // them, but even I was confused by it!) As much code/comments as
2653 // possible are retained from inline_unsafe_access though to make
2654 // the correspondences clearer. - dl
2655
2656 if (callee()->is_static()) return false; // caller must have the capability!
2657
2658 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2659 decorators |= mo_decorator_for_access_kind(access_kind);
2660
2661 #ifndef PRODUCT
2662 BasicType rtype;
2663 {
2664 ResourceMark rm;
2665 // Check the signatures.
2666 ciSignature* sig = callee()->signature();
2667 rtype = sig->return_type()->basic_type();
2668 switch(kind) {
2669 case LS_get_add:
2670 case LS_get_set: {
2671 // Check the signatures.
2672 #ifdef ASSERT
2673 assert(rtype == type, "get and set must return the expected type");
2674 assert(sig->count() == 3, "get and set has 3 arguments");
2675 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2676 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2677 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2678 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2679 #endif // ASSERT
2680 break;
2681 }
2682 case LS_cmp_swap:
2683 case LS_cmp_swap_weak: {
2684 // Check the signatures.
2685 #ifdef ASSERT
2686 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2687 assert(sig->count() == 4, "CAS has 4 arguments");
2688 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2689 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2690 #endif // ASSERT
2691 break;
2692 }
2693 case LS_cmp_exchange: {
2694 // Check the signatures.
2695 #ifdef ASSERT
2696 assert(rtype == type, "CAS must return the expected type");
2697 assert(sig->count() == 4, "CAS has 4 arguments");
2698 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2699 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2700 #endif // ASSERT
2701 break;
2702 }
2703 default:
2704 ShouldNotReachHere();
2705 }
2706 }
2707 #endif //PRODUCT
2708
2709 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2710
2711 // Get arguments:
2712 Node* receiver = NULL;
2713 Node* base = NULL;
2714 Node* offset = NULL;
2715 Node* oldval = NULL;
2716 Node* newval = NULL;
2717 switch(kind) {
2718 case LS_cmp_swap:
2719 case LS_cmp_swap_weak:
2720 case LS_cmp_exchange: {
2721 const bool two_slot_type = type2size[type] == 2;
2722 receiver = argument(0); // type: oop
2723 base = argument(1); // type: oop
2724 offset = argument(2); // type: long
2725 oldval = argument(4); // type: oop, int, or long
2726 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2727 break;
2728 }
2729 case LS_get_add:
2730 case LS_get_set: {
2731 receiver = argument(0); // type: oop
2732 base = argument(1); // type: oop
2733 offset = argument(2); // type: long
2734 oldval = NULL;
2735 newval = argument(4); // type: oop, int, or long
2736 break;
2737 }
2738 default:
2739 ShouldNotReachHere();
2740 }
2741
2742 // Build field offset expression.
2743 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2744 // to be plain byte offsets, which are also the same as those accepted
2745 // by oopDesc::field_addr.
2746 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2747 // 32-bit machines ignore the high half of long offsets
2748 offset = ConvL2X(offset);
2749 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false);
2750 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2751
2752 Compile::AliasType* alias_type = C->alias_type(adr_type);
2753 BasicType bt = alias_type->basic_type();
2754 if (bt != T_ILLEGAL &&
2755 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2756 // Don't intrinsify mismatched object accesses.
2757 return false;
2758 }
2759
2760 // For CAS, unlike inline_unsafe_access, there seems no point in
2761 // trying to refine types. Just use the coarse types here.
2762 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2763 const Type *value_type = Type::get_const_basic_type(type);
2764
2765 switch (kind) {
2766 case LS_get_set:
2767 case LS_cmp_exchange: {
2768 if (type == T_OBJECT) {
2769 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2770 if (tjp != NULL) {
2771 value_type = tjp;
2772 }
2773 }
2774 break;
2775 }
2776 case LS_cmp_swap:
2777 case LS_cmp_swap_weak:
2778 case LS_get_add:
2779 break;
2780 default:
2781 ShouldNotReachHere();
2782 }
2783
2784 // Null check receiver.
2785 receiver = null_check(receiver);
2786 if (stopped()) {
2787 return true;
2788 }
2789
2790 int alias_idx = C->get_alias_index(adr_type);
2791
2792 if (type == T_OBJECT || type == T_ARRAY) {
2793 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF;
2794
2795 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2796 // could be delayed during Parse (for example, in adjust_map_after_if()).
2797 // Execute transformation here to avoid barrier generation in such case.
2798 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2799 newval = _gvn.makecon(TypePtr::NULL_PTR);
2800
2801 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) {
2802 // Refine the value to a null constant, when it is known to be null
2803 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2804 }
2805 }
2806
2807 Node* result = NULL;
2808 switch (kind) {
2809 case LS_cmp_exchange: {
2810 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx,
2811 oldval, newval, value_type, type, decorators);
2812 break;
2813 }
2814 case LS_cmp_swap_weak:
2815 decorators |= C2_WEAK_CMPXCHG;
2816 case LS_cmp_swap: {
2817 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx,
2818 oldval, newval, value_type, type, decorators);
2819 break;
2820 }
2821 case LS_get_set: {
2822 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx,
2823 newval, value_type, type, decorators);
2824 break;
2825 }
2826 case LS_get_add: {
2827 result = access_atomic_add_at(base, adr, adr_type, alias_idx,
2828 newval, value_type, type, decorators);
2829 break;
2830 }
2831 default:
2832 ShouldNotReachHere();
2833 }
2834
2835 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2836 set_result(result);
2837 return true;
2838 }
2839
2840 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2841 // Regardless of form, don't allow previous ld/st to move down,
2842 // then issue acquire, release, or volatile mem_bar.
2843 insert_mem_bar(Op_MemBarCPUOrder);
2844 switch(id) {
2845 case vmIntrinsics::_loadFence:
2846 insert_mem_bar(Op_LoadFence);
2847 return true;
2848 case vmIntrinsics::_storeFence:
2849 insert_mem_bar(Op_StoreFence);
2850 return true;
2851 case vmIntrinsics::_fullFence:
2852 insert_mem_bar(Op_MemBarVolatile);
2853 return true;
2854 default:
2855 fatal_unexpected_iid(id);
2856 return false;
2857 }
2858 }
2859
2860 bool LibraryCallKit::inline_onspinwait() {
2861 insert_mem_bar(Op_OnSpinWait);
2862 return true;
2863 }
2864
2865 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2866 if (!kls->is_Con()) {
2867 return true;
2868 }
2869 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2870 if (klsptr == NULL) {
2871 return true;
2872 }
2873 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2874 // don't need a guard for a klass that is already initialized
2875 return !ik->is_initialized();
2876 }
2877
2878 //----------------------------inline_unsafe_allocate---------------------------
2879 // public native Object Unsafe.allocateInstance(Class<?> cls);
2880 bool LibraryCallKit::inline_unsafe_allocate() {
2881 if (callee()->is_static()) return false; // caller must have the capability!
2882
2883 null_check_receiver(); // null-check, then ignore
2884 Node* cls = null_check(argument(1));
2885 if (stopped()) return true;
2886
2887 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2888 kls = null_check(kls);
2889 if (stopped()) return true; // argument was like int.class
2890
2891 Node* test = NULL;
2892 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2893 // Note: The argument might still be an illegal value like
2894 // Serializable.class or Object[].class. The runtime will handle it.
2895 // But we must make an explicit check for initialization.
2896 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2897 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2898 // can generate code to load it as unsigned byte.
2899 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2900 Node* bits = intcon(InstanceKlass::fully_initialized);
2901 test = _gvn.transform(new SubINode(inst, bits));
2902 // The 'test' is non-zero if we need to take a slow path.
2903 }
2904
2905 Node* obj = new_instance(kls, test);
2906 set_result(obj);
2907 return true;
2908 }
2909
2910 //------------------------inline_native_time_funcs--------------
2911 // inline code for System.currentTimeMillis() and System.nanoTime()
2912 // these have the same type and signature
2913 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2914 const TypeFunc* tf = OptoRuntime::void_long_Type();
2915 const TypePtr* no_memory_effects = NULL;
2916 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2917 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2918 #ifdef ASSERT
2919 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2920 assert(value_top == top(), "second value must be top");
2921 #endif
2922 set_result(value);
2923 return true;
2924 }
2925
2926 #ifdef JFR_HAVE_INTRINSICS
2927
2928 /*
2929 * oop -> myklass
2930 * myklass->trace_id |= USED
2931 * return myklass->trace_id & ~0x3
2932 */
2933 bool LibraryCallKit::inline_native_classID() {
2934 Node* cls = null_check(argument(0), T_OBJECT);
2935 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2936 kls = null_check(kls, T_OBJECT);
2937
2938 ByteSize offset = KLASS_TRACE_ID_OFFSET;
2939 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2940 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2941
2942 Node* clsused = longcon(0x01l); // set the class bit
2943 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2944 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2945 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2946
2947 #ifdef TRACE_ID_META_BITS
2948 Node* mbits = longcon(~TRACE_ID_META_BITS);
2949 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
2950 #endif
2951 #ifdef TRACE_ID_SHIFT
2952 Node* cbits = intcon(TRACE_ID_SHIFT);
2953 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
2954 #endif
2955
2956 set_result(tvalue);
2957 return true;
2958
2959 }
2960
2961 bool LibraryCallKit::inline_native_getEventWriter() {
2962 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
2963
2964 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
2965 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR));
2966
2967 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
2968
2969 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
2970 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
2971
2972 IfNode* iff_jobj_null =
2973 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
2974
2975 enum { _normal_path = 1,
2976 _null_path = 2,
2977 PATH_LIMIT };
2978
2979 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
2980 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM);
2981
2982 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
2983 result_rgn->init_req(_null_path, jobj_is_null);
2984 result_val->init_req(_null_path, null());
2985
2986 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
2987 set_control(jobj_is_not_null);
2988 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT,
2989 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD);
2990 result_rgn->init_req(_normal_path, control());
2991 result_val->init_req(_normal_path, res);
2992
2993 set_result(result_rgn, result_val);
2994
2995 return true;
2996 }
2997
2998 #endif // JFR_HAVE_INTRINSICS
2999
3000 //------------------------inline_native_currentThread------------------
3001 bool LibraryCallKit::inline_native_currentThread() {
3002 Node* junk = NULL;
3003 set_result(generate_current_thread(junk));
3004 return true;
3005 }
3006
3007 //------------------------inline_native_isInterrupted------------------
3008 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3009 bool LibraryCallKit::inline_native_isInterrupted() {
3010 // Add a fast path to t.isInterrupted(clear_int):
3011 // (t == Thread.current() &&
3012 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3013 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3014 // So, in the common case that the interrupt bit is false,
3015 // we avoid making a call into the VM. Even if the interrupt bit
3016 // is true, if the clear_int argument is false, we avoid the VM call.
3017 // However, if the receiver is not currentThread, we must call the VM,
3018 // because there must be some locking done around the operation.
3019
3020 // We only go to the fast case code if we pass two guards.
3021 // Paths which do not pass are accumulated in the slow_region.
3022
3023 enum {
3024 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3025 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3026 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3027 PATH_LIMIT
3028 };
3029
3030 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3031 // out of the function.
3032 insert_mem_bar(Op_MemBarCPUOrder);
3033
3034 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3035 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3036
3037 RegionNode* slow_region = new RegionNode(1);
3038 record_for_igvn(slow_region);
3039
3040 // (a) Receiving thread must be the current thread.
3041 Node* rec_thr = argument(0);
3042 Node* tls_ptr = NULL;
3043 Node* cur_thr = generate_current_thread(tls_ptr);
3044
3045 // Resolve oops to stable for CmpP below.
3046 cur_thr = access_resolve(cur_thr, 0);
3047 rec_thr = access_resolve(rec_thr, 0);
3048
3049 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3050 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3051
3052 generate_slow_guard(bol_thr, slow_region);
3053
3054 // (b) Interrupt bit on TLS must be false.
3055 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3056 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3057 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3058
3059 // Set the control input on the field _interrupted read to prevent it floating up.
3060 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3061 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3062 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3063
3064 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3065
3066 // First fast path: if (!TLS._interrupted) return false;
3067 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3068 result_rgn->init_req(no_int_result_path, false_bit);
3069 result_val->init_req(no_int_result_path, intcon(0));
3070
3071 // drop through to next case
3072 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3073
3074 #ifndef _WINDOWS
3075 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3076 Node* clr_arg = argument(1);
3077 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3078 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3079 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3080
3081 // Second fast path: ... else if (!clear_int) return true;
3082 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3083 result_rgn->init_req(no_clear_result_path, false_arg);
3084 result_val->init_req(no_clear_result_path, intcon(1));
3085
3086 // drop through to next case
3087 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3088 #else
3089 // To return true on Windows you must read the _interrupted field
3090 // and check the event state i.e. take the slow path.
3091 #endif // _WINDOWS
3092
3093 // (d) Otherwise, go to the slow path.
3094 slow_region->add_req(control());
3095 set_control( _gvn.transform(slow_region));
3096
3097 if (stopped()) {
3098 // There is no slow path.
3099 result_rgn->init_req(slow_result_path, top());
3100 result_val->init_req(slow_result_path, top());
3101 } else {
3102 // non-virtual because it is a private non-static
3103 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3104
3105 Node* slow_val = set_results_for_java_call(slow_call);
3106 // this->control() comes from set_results_for_java_call
3107
3108 Node* fast_io = slow_call->in(TypeFunc::I_O);
3109 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3110
3111 // These two phis are pre-filled with copies of of the fast IO and Memory
3112 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3113 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3114
3115 result_rgn->init_req(slow_result_path, control());
3116 result_io ->init_req(slow_result_path, i_o());
3117 result_mem->init_req(slow_result_path, reset_memory());
3118 result_val->init_req(slow_result_path, slow_val);
3119
3120 set_all_memory(_gvn.transform(result_mem));
3121 set_i_o( _gvn.transform(result_io));
3122 }
3123
3124 C->set_has_split_ifs(true); // Has chance for split-if optimization
3125 set_result(result_rgn, result_val);
3126 return true;
3127 }
3128
3129 //---------------------------load_mirror_from_klass----------------------------
3130 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3131 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3132 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3133 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3134 // mirror = ((OopHandle)mirror)->resolve();
3135 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
3136 }
3137
3138 //-----------------------load_klass_from_mirror_common-------------------------
3139 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3140 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3141 // and branch to the given path on the region.
3142 // If never_see_null, take an uncommon trap on null, so we can optimistically
3143 // compile for the non-null case.
3144 // If the region is NULL, force never_see_null = true.
3145 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3146 bool never_see_null,
3147 RegionNode* region,
3148 int null_path,
3149 int offset) {
3150 if (region == NULL) never_see_null = true;
3151 Node* p = basic_plus_adr(mirror, offset);
3152 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3153 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3154 Node* null_ctl = top();
3155 kls = null_check_oop(kls, &null_ctl, never_see_null);
3156 if (region != NULL) {
3157 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3158 region->init_req(null_path, null_ctl);
3159 } else {
3160 assert(null_ctl == top(), "no loose ends");
3161 }
3162 return kls;
3163 }
3164
3165 //--------------------(inline_native_Class_query helpers)---------------------
3166 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3167 // Fall through if (mods & mask) == bits, take the guard otherwise.
3168 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3169 // Branch around if the given klass has the given modifier bit set.
3170 // Like generate_guard, adds a new path onto the region.
3171 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3172 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3173 Node* mask = intcon(modifier_mask);
3174 Node* bits = intcon(modifier_bits);
3175 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3176 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3177 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3178 return generate_fair_guard(bol, region);
3179 }
3180 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3181 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3182 }
3183
3184 //-------------------------inline_native_Class_query-------------------
3185 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3186 const Type* return_type = TypeInt::BOOL;
3187 Node* prim_return_value = top(); // what happens if it's a primitive class?
3188 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3189 bool expect_prim = false; // most of these guys expect to work on refs
3190
3191 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3192
3193 Node* mirror = argument(0);
3194 Node* obj = top();
3195
3196 switch (id) {
3197 case vmIntrinsics::_isInstance:
3198 // nothing is an instance of a primitive type
3199 prim_return_value = intcon(0);
3200 obj = argument(1);
3201 break;
3202 case vmIntrinsics::_getModifiers:
3203 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3204 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3205 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3206 break;
3207 case vmIntrinsics::_isInterface:
3208 prim_return_value = intcon(0);
3209 break;
3210 case vmIntrinsics::_isArray:
3211 prim_return_value = intcon(0);
3212 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3213 break;
3214 case vmIntrinsics::_isPrimitive:
3215 prim_return_value = intcon(1);
3216 expect_prim = true; // obviously
3217 break;
3218 case vmIntrinsics::_getSuperclass:
3219 prim_return_value = null();
3220 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3221 break;
3222 case vmIntrinsics::_getClassAccessFlags:
3223 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3224 return_type = TypeInt::INT; // not bool! 6297094
3225 break;
3226 default:
3227 fatal_unexpected_iid(id);
3228 break;
3229 }
3230
3231 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3232 if (mirror_con == NULL) return false; // cannot happen?
3233
3234 #ifndef PRODUCT
3235 if (C->print_intrinsics() || C->print_inlining()) {
3236 ciType* k = mirror_con->java_mirror_type();
3237 if (k) {
3238 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3239 k->print_name();
3240 tty->cr();
3241 }
3242 }
3243 #endif
3244
3245 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3246 RegionNode* region = new RegionNode(PATH_LIMIT);
3247 record_for_igvn(region);
3248 PhiNode* phi = new PhiNode(region, return_type);
3249
3250 // The mirror will never be null of Reflection.getClassAccessFlags, however
3251 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3252 // if it is. See bug 4774291.
3253
3254 // For Reflection.getClassAccessFlags(), the null check occurs in
3255 // the wrong place; see inline_unsafe_access(), above, for a similar
3256 // situation.
3257 mirror = null_check(mirror);
3258 // If mirror or obj is dead, only null-path is taken.
3259 if (stopped()) return true;
3260
3261 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3262
3263 // Now load the mirror's klass metaobject, and null-check it.
3264 // Side-effects region with the control path if the klass is null.
3265 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3266 // If kls is null, we have a primitive mirror.
3267 phi->init_req(_prim_path, prim_return_value);
3268 if (stopped()) { set_result(region, phi); return true; }
3269 bool safe_for_replace = (region->in(_prim_path) == top());
3270
3271 Node* p; // handy temp
3272 Node* null_ctl;
3273
3274 // Now that we have the non-null klass, we can perform the real query.
3275 // For constant classes, the query will constant-fold in LoadNode::Value.
3276 Node* query_value = top();
3277 switch (id) {
3278 case vmIntrinsics::_isInstance:
3279 // nothing is an instance of a primitive type
3280 query_value = gen_instanceof(obj, kls, safe_for_replace);
3281 break;
3282
3283 case vmIntrinsics::_getModifiers:
3284 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3285 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3286 break;
3287
3288 case vmIntrinsics::_isInterface:
3289 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3290 if (generate_interface_guard(kls, region) != NULL)
3291 // A guard was added. If the guard is taken, it was an interface.
3292 phi->add_req(intcon(1));
3293 // If we fall through, it's a plain class.
3294 query_value = intcon(0);
3295 break;
3296
3297 case vmIntrinsics::_isArray:
3298 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3299 if (generate_array_guard(kls, region) != NULL)
3300 // A guard was added. If the guard is taken, it was an array.
3301 phi->add_req(intcon(1));
3302 // If we fall through, it's a plain class.
3303 query_value = intcon(0);
3304 break;
3305
3306 case vmIntrinsics::_isPrimitive:
3307 query_value = intcon(0); // "normal" path produces false
3308 break;
3309
3310 case vmIntrinsics::_getSuperclass:
3311 // The rules here are somewhat unfortunate, but we can still do better
3312 // with random logic than with a JNI call.
3313 // Interfaces store null or Object as _super, but must report null.
3314 // Arrays store an intermediate super as _super, but must report Object.
3315 // Other types can report the actual _super.
3316 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3317 if (generate_interface_guard(kls, region) != NULL)
3318 // A guard was added. If the guard is taken, it was an interface.
3319 phi->add_req(null());
3320 if (generate_array_guard(kls, region) != NULL)
3321 // A guard was added. If the guard is taken, it was an array.
3322 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3323 // If we fall through, it's a plain class. Get its _super.
3324 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3325 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3326 null_ctl = top();
3327 kls = null_check_oop(kls, &null_ctl);
3328 if (null_ctl != top()) {
3329 // If the guard is taken, Object.superClass is null (both klass and mirror).
3330 region->add_req(null_ctl);
3331 phi ->add_req(null());
3332 }
3333 if (!stopped()) {
3334 query_value = load_mirror_from_klass(kls);
3335 }
3336 break;
3337
3338 case vmIntrinsics::_getClassAccessFlags:
3339 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3340 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3341 break;
3342
3343 default:
3344 fatal_unexpected_iid(id);
3345 break;
3346 }
3347
3348 // Fall-through is the normal case of a query to a real class.
3349 phi->init_req(1, query_value);
3350 region->init_req(1, control());
3351
3352 C->set_has_split_ifs(true); // Has chance for split-if optimization
3353 set_result(region, phi);
3354 return true;
3355 }
3356
3357 //-------------------------inline_Class_cast-------------------
3358 bool LibraryCallKit::inline_Class_cast() {
3359 Node* mirror = argument(0); // Class
3360 Node* obj = argument(1);
3361 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3362 if (mirror_con == NULL) {
3363 return false; // dead path (mirror->is_top()).
3364 }
3365 if (obj == NULL || obj->is_top()) {
3366 return false; // dead path
3367 }
3368 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3369
3370 // First, see if Class.cast() can be folded statically.
3371 // java_mirror_type() returns non-null for compile-time Class constants.
3372 ciType* tm = mirror_con->java_mirror_type();
3373 if (tm != NULL && tm->is_klass() &&
3374 tp != NULL && tp->klass() != NULL) {
3375 if (!tp->klass()->is_loaded()) {
3376 // Don't use intrinsic when class is not loaded.
3377 return false;
3378 } else {
3379 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3380 if (static_res == Compile::SSC_always_true) {
3381 // isInstance() is true - fold the code.
3382 set_result(obj);
3383 return true;
3384 } else if (static_res == Compile::SSC_always_false) {
3385 // Don't use intrinsic, have to throw ClassCastException.
3386 // If the reference is null, the non-intrinsic bytecode will
3387 // be optimized appropriately.
3388 return false;
3389 }
3390 }
3391 }
3392
3393 // Bailout intrinsic and do normal inlining if exception path is frequent.
3394 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3395 return false;
3396 }
3397
3398 // Generate dynamic checks.
3399 // Class.cast() is java implementation of _checkcast bytecode.
3400 // Do checkcast (Parse::do_checkcast()) optimizations here.
3401
3402 mirror = null_check(mirror);
3403 // If mirror is dead, only null-path is taken.
3404 if (stopped()) {
3405 return true;
3406 }
3407
3408 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3409 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3410 RegionNode* region = new RegionNode(PATH_LIMIT);
3411 record_for_igvn(region);
3412
3413 // Now load the mirror's klass metaobject, and null-check it.
3414 // If kls is null, we have a primitive mirror and
3415 // nothing is an instance of a primitive type.
3416 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3417
3418 Node* res = top();
3419 if (!stopped()) {
3420 Node* bad_type_ctrl = top();
3421 // Do checkcast optimizations.
3422 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3423 region->init_req(_bad_type_path, bad_type_ctrl);
3424 }
3425 if (region->in(_prim_path) != top() ||
3426 region->in(_bad_type_path) != top()) {
3427 // Let Interpreter throw ClassCastException.
3428 PreserveJVMState pjvms(this);
3429 set_control(_gvn.transform(region));
3430 uncommon_trap(Deoptimization::Reason_intrinsic,
3431 Deoptimization::Action_maybe_recompile);
3432 }
3433 if (!stopped()) {
3434 set_result(res);
3435 }
3436 return true;
3437 }
3438
3439
3440 //--------------------------inline_native_subtype_check------------------------
3441 // This intrinsic takes the JNI calls out of the heart of
3442 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3443 bool LibraryCallKit::inline_native_subtype_check() {
3444 // Pull both arguments off the stack.
3445 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3446 args[0] = argument(0);
3447 args[1] = argument(1);
3448 Node* klasses[2]; // corresponding Klasses: superk, subk
3449 klasses[0] = klasses[1] = top();
3450
3451 enum {
3452 // A full decision tree on {superc is prim, subc is prim}:
3453 _prim_0_path = 1, // {P,N} => false
3454 // {P,P} & superc!=subc => false
3455 _prim_same_path, // {P,P} & superc==subc => true
3456 _prim_1_path, // {N,P} => false
3457 _ref_subtype_path, // {N,N} & subtype check wins => true
3458 _both_ref_path, // {N,N} & subtype check loses => false
3459 PATH_LIMIT
3460 };
3461
3462 RegionNode* region = new RegionNode(PATH_LIMIT);
3463 Node* phi = new PhiNode(region, TypeInt::BOOL);
3464 record_for_igvn(region);
3465
3466 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3467 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3468 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3469
3470 // First null-check both mirrors and load each mirror's klass metaobject.
3471 int which_arg;
3472 for (which_arg = 0; which_arg <= 1; which_arg++) {
3473 Node* arg = args[which_arg];
3474 arg = null_check(arg);
3475 if (stopped()) break;
3476 args[which_arg] = arg;
3477
3478 Node* p = basic_plus_adr(arg, class_klass_offset);
3479 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3480 klasses[which_arg] = _gvn.transform(kls);
3481 }
3482
3483 // Resolve oops to stable for CmpP below.
3484 args[0] = access_resolve(args[0], 0);
3485 args[1] = access_resolve(args[1], 0);
3486
3487 // Having loaded both klasses, test each for null.
3488 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3489 for (which_arg = 0; which_arg <= 1; which_arg++) {
3490 Node* kls = klasses[which_arg];
3491 Node* null_ctl = top();
3492 kls = null_check_oop(kls, &null_ctl, never_see_null);
3493 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3494 region->init_req(prim_path, null_ctl);
3495 if (stopped()) break;
3496 klasses[which_arg] = kls;
3497 }
3498
3499 if (!stopped()) {
3500 // now we have two reference types, in klasses[0..1]
3501 Node* subk = klasses[1]; // the argument to isAssignableFrom
3502 Node* superk = klasses[0]; // the receiver
3503 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3504 // now we have a successful reference subtype check
3505 region->set_req(_ref_subtype_path, control());
3506 }
3507
3508 // If both operands are primitive (both klasses null), then
3509 // we must return true when they are identical primitives.
3510 // It is convenient to test this after the first null klass check.
3511 set_control(region->in(_prim_0_path)); // go back to first null check
3512 if (!stopped()) {
3513 // Since superc is primitive, make a guard for the superc==subc case.
3514 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3515 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3516 generate_guard(bol_eq, region, PROB_FAIR);
3517 if (region->req() == PATH_LIMIT+1) {
3518 // A guard was added. If the added guard is taken, superc==subc.
3519 region->swap_edges(PATH_LIMIT, _prim_same_path);
3520 region->del_req(PATH_LIMIT);
3521 }
3522 region->set_req(_prim_0_path, control()); // Not equal after all.
3523 }
3524
3525 // these are the only paths that produce 'true':
3526 phi->set_req(_prim_same_path, intcon(1));
3527 phi->set_req(_ref_subtype_path, intcon(1));
3528
3529 // pull together the cases:
3530 assert(region->req() == PATH_LIMIT, "sane region");
3531 for (uint i = 1; i < region->req(); i++) {
3532 Node* ctl = region->in(i);
3533 if (ctl == NULL || ctl == top()) {
3534 region->set_req(i, top());
3535 phi ->set_req(i, top());
3536 } else if (phi->in(i) == NULL) {
3537 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3538 }
3539 }
3540
3541 set_control(_gvn.transform(region));
3542 set_result(_gvn.transform(phi));
3543 return true;
3544 }
3545
3546 //---------------------generate_array_guard_common------------------------
3547 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3548 bool obj_array, bool not_array) {
3549
3550 if (stopped()) {
3551 return NULL;
3552 }
3553
3554 // If obj_array/non_array==false/false:
3555 // Branch around if the given klass is in fact an array (either obj or prim).
3556 // If obj_array/non_array==false/true:
3557 // Branch around if the given klass is not an array klass of any kind.
3558 // If obj_array/non_array==true/true:
3559 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3560 // If obj_array/non_array==true/false:
3561 // Branch around if the kls is an oop array (Object[] or subtype)
3562 //
3563 // Like generate_guard, adds a new path onto the region.
3564 jint layout_con = 0;
3565 Node* layout_val = get_layout_helper(kls, layout_con);
3566 if (layout_val == NULL) {
3567 bool query = (obj_array
3568 ? Klass::layout_helper_is_objArray(layout_con)
3569 : Klass::layout_helper_is_array(layout_con));
3570 if (query == not_array) {
3571 return NULL; // never a branch
3572 } else { // always a branch
3573 Node* always_branch = control();
3574 if (region != NULL)
3575 region->add_req(always_branch);
3576 set_control(top());
3577 return always_branch;
3578 }
3579 }
3580 // Now test the correct condition.
3581 jint nval = (obj_array
3582 ? (jint)(Klass::_lh_array_tag_type_value
3583 << Klass::_lh_array_tag_shift)
3584 : Klass::_lh_neutral_value);
3585 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3586 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3587 // invert the test if we are looking for a non-array
3588 if (not_array) btest = BoolTest(btest).negate();
3589 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3590 return generate_fair_guard(bol, region);
3591 }
3592
3593
3594 //-----------------------inline_native_newArray--------------------------
3595 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3596 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3597 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3598 Node* mirror;
3599 Node* count_val;
3600 if (uninitialized) {
3601 mirror = argument(1);
3602 count_val = argument(2);
3603 } else {
3604 mirror = argument(0);
3605 count_val = argument(1);
3606 }
3607
3608 mirror = null_check(mirror);
3609 // If mirror or obj is dead, only null-path is taken.
3610 if (stopped()) return true;
3611
3612 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3613 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3614 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3615 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3616 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3617
3618 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3619 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3620 result_reg, _slow_path);
3621 Node* normal_ctl = control();
3622 Node* no_array_ctl = result_reg->in(_slow_path);
3623
3624 // Generate code for the slow case. We make a call to newArray().
3625 set_control(no_array_ctl);
3626 if (!stopped()) {
3627 // Either the input type is void.class, or else the
3628 // array klass has not yet been cached. Either the
3629 // ensuing call will throw an exception, or else it
3630 // will cache the array klass for next time.
3631 PreserveJVMState pjvms(this);
3632 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3633 Node* slow_result = set_results_for_java_call(slow_call);
3634 // this->control() comes from set_results_for_java_call
3635 result_reg->set_req(_slow_path, control());
3636 result_val->set_req(_slow_path, slow_result);
3637 result_io ->set_req(_slow_path, i_o());
3638 result_mem->set_req(_slow_path, reset_memory());
3639 }
3640
3641 set_control(normal_ctl);
3642 if (!stopped()) {
3643 // Normal case: The array type has been cached in the java.lang.Class.
3644 // The following call works fine even if the array type is polymorphic.
3645 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3646 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3647 result_reg->init_req(_normal_path, control());
3648 result_val->init_req(_normal_path, obj);
3649 result_io ->init_req(_normal_path, i_o());
3650 result_mem->init_req(_normal_path, reset_memory());
3651
3652 if (uninitialized) {
3653 // Mark the allocation so that zeroing is skipped
3654 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3655 alloc->maybe_set_complete(&_gvn);
3656 }
3657 }
3658
3659 // Return the combined state.
3660 set_i_o( _gvn.transform(result_io) );
3661 set_all_memory( _gvn.transform(result_mem));
3662
3663 C->set_has_split_ifs(true); // Has chance for split-if optimization
3664 set_result(result_reg, result_val);
3665 return true;
3666 }
3667
3668 //----------------------inline_native_getLength--------------------------
3669 // public static native int java.lang.reflect.Array.getLength(Object array);
3670 bool LibraryCallKit::inline_native_getLength() {
3671 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3672
3673 Node* array = null_check(argument(0));
3674 // If array is dead, only null-path is taken.
3675 if (stopped()) return true;
3676
3677 // Deoptimize if it is a non-array.
3678 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3679
3680 if (non_array != NULL) {
3681 PreserveJVMState pjvms(this);
3682 set_control(non_array);
3683 uncommon_trap(Deoptimization::Reason_intrinsic,
3684 Deoptimization::Action_maybe_recompile);
3685 }
3686
3687 // If control is dead, only non-array-path is taken.
3688 if (stopped()) return true;
3689
3690 // The works fine even if the array type is polymorphic.
3691 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3692 Node* result = load_array_length(array);
3693
3694 C->set_has_split_ifs(true); // Has chance for split-if optimization
3695 set_result(result);
3696 return true;
3697 }
3698
3699 //------------------------inline_array_copyOf----------------------------
3700 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3701 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3702 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3703 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3704
3705 // Get the arguments.
3706 Node* original = argument(0);
3707 Node* start = is_copyOfRange? argument(1): intcon(0);
3708 Node* end = is_copyOfRange? argument(2): argument(1);
3709 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3710
3711 Node* newcopy = NULL;
3712
3713 // Set the original stack and the reexecute bit for the interpreter to reexecute
3714 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3715 { PreserveReexecuteState preexecs(this);
3716 jvms()->set_should_reexecute(true);
3717
3718 array_type_mirror = null_check(array_type_mirror);
3719 original = null_check(original);
3720
3721 // Check if a null path was taken unconditionally.
3722 if (stopped()) return true;
3723
3724 Node* orig_length = load_array_length(original);
3725
3726 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3727 klass_node = null_check(klass_node);
3728
3729 RegionNode* bailout = new RegionNode(1);
3730 record_for_igvn(bailout);
3731
3732 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3733 // Bail out if that is so.
3734 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3735 if (not_objArray != NULL) {
3736 // Improve the klass node's type from the new optimistic assumption:
3737 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3738 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3739 Node* cast = new CastPPNode(klass_node, akls);
3740 cast->init_req(0, control());
3741 klass_node = _gvn.transform(cast);
3742 }
3743
3744 // Bail out if either start or end is negative.
3745 generate_negative_guard(start, bailout, &start);
3746 generate_negative_guard(end, bailout, &end);
3747
3748 Node* length = end;
3749 if (_gvn.type(start) != TypeInt::ZERO) {
3750 length = _gvn.transform(new SubINode(end, start));
3751 }
3752
3753 // Bail out if length is negative.
3754 // Without this the new_array would throw
3755 // NegativeArraySizeException but IllegalArgumentException is what
3756 // should be thrown
3757 generate_negative_guard(length, bailout, &length);
3758
3759 if (bailout->req() > 1) {
3760 PreserveJVMState pjvms(this);
3761 set_control(_gvn.transform(bailout));
3762 uncommon_trap(Deoptimization::Reason_intrinsic,
3763 Deoptimization::Action_maybe_recompile);
3764 }
3765
3766 if (!stopped()) {
3767 // How many elements will we copy from the original?
3768 // The answer is MinI(orig_length - start, length).
3769 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3770 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3771
3772 original = access_resolve(original, ACCESS_READ);
3773
3774 // Generate a direct call to the right arraycopy function(s).
3775 // We know the copy is disjoint but we might not know if the
3776 // oop stores need checking.
3777 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3778 // This will fail a store-check if x contains any non-nulls.
3779
3780 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3781 // loads/stores but it is legal only if we're sure the
3782 // Arrays.copyOf would succeed. So we need all input arguments
3783 // to the copyOf to be validated, including that the copy to the
3784 // new array won't trigger an ArrayStoreException. That subtype
3785 // check can be optimized if we know something on the type of
3786 // the input array from type speculation.
3787 if (_gvn.type(klass_node)->singleton()) {
3788 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3789 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3790
3791 int test = C->static_subtype_check(superk, subk);
3792 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3793 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3794 if (t_original->speculative_type() != NULL) {
3795 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3796 }
3797 }
3798 }
3799
3800 bool validated = false;
3801 // Reason_class_check rather than Reason_intrinsic because we
3802 // want to intrinsify even if this traps.
3803 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3804 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3805 klass_node);
3806
3807 if (not_subtype_ctrl != top()) {
3808 PreserveJVMState pjvms(this);
3809 set_control(not_subtype_ctrl);
3810 uncommon_trap(Deoptimization::Reason_class_check,
3811 Deoptimization::Action_make_not_entrant);
3812 assert(stopped(), "Should be stopped");
3813 }
3814 validated = true;
3815 }
3816
3817 if (!stopped()) {
3818 newcopy = new_array(klass_node, length, 0); // no arguments to push
3819
3820 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
3821 load_object_klass(original), klass_node);
3822 if (!is_copyOfRange) {
3823 ac->set_copyof(validated);
3824 } else {
3825 ac->set_copyofrange(validated);
3826 }
3827 Node* n = _gvn.transform(ac);
3828 if (n == ac) {
3829 ac->connect_outputs(this);
3830 } else {
3831 assert(validated, "shouldn't transform if all arguments not validated");
3832 set_all_memory(n);
3833 }
3834 }
3835 }
3836 } // original reexecute is set back here
3837
3838 C->set_has_split_ifs(true); // Has chance for split-if optimization
3839 if (!stopped()) {
3840 set_result(newcopy);
3841 }
3842 return true;
3843 }
3844
3845
3846 //----------------------generate_virtual_guard---------------------------
3847 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3848 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3849 RegionNode* slow_region) {
3850 ciMethod* method = callee();
3851 int vtable_index = method->vtable_index();
3852 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3853 "bad index %d", vtable_index);
3854 // Get the Method* out of the appropriate vtable entry.
3855 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
3856 vtable_index*vtableEntry::size_in_bytes() +
3857 vtableEntry::method_offset_in_bytes();
3858 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3859 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3860
3861 // Compare the target method with the expected method (e.g., Object.hashCode).
3862 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3863
3864 Node* native_call = makecon(native_call_addr);
3865 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3866 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3867
3868 return generate_slow_guard(test_native, slow_region);
3869 }
3870
3871 //-----------------------generate_method_call----------------------------
3872 // Use generate_method_call to make a slow-call to the real
3873 // method if the fast path fails. An alternative would be to
3874 // use a stub like OptoRuntime::slow_arraycopy_Java.
3875 // This only works for expanding the current library call,
3876 // not another intrinsic. (E.g., don't use this for making an
3877 // arraycopy call inside of the copyOf intrinsic.)
3878 CallJavaNode*
3879 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3880 // When compiling the intrinsic method itself, do not use this technique.
3881 guarantee(callee() != C->method(), "cannot make slow-call to self");
3882
3883 ciMethod* method = callee();
3884 // ensure the JVMS we have will be correct for this call
3885 guarantee(method_id == method->intrinsic_id(), "must match");
3886
3887 const TypeFunc* tf = TypeFunc::make(method);
3888 CallJavaNode* slow_call;
3889 if (is_static) {
3890 assert(!is_virtual, "");
3891 slow_call = new CallStaticJavaNode(C, tf,
3892 SharedRuntime::get_resolve_static_call_stub(),
3893 method, bci());
3894 } else if (is_virtual) {
3895 null_check_receiver();
3896 int vtable_index = Method::invalid_vtable_index;
3897 if (UseInlineCaches) {
3898 // Suppress the vtable call
3899 } else {
3900 // hashCode and clone are not a miranda methods,
3901 // so the vtable index is fixed.
3902 // No need to use the linkResolver to get it.
3903 vtable_index = method->vtable_index();
3904 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3905 "bad index %d", vtable_index);
3906 }
3907 slow_call = new CallDynamicJavaNode(tf,
3908 SharedRuntime::get_resolve_virtual_call_stub(),
3909 method, vtable_index, bci());
3910 } else { // neither virtual nor static: opt_virtual
3911 null_check_receiver();
3912 slow_call = new CallStaticJavaNode(C, tf,
3913 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3914 method, bci());
3915 slow_call->set_optimized_virtual(true);
3916 }
3917 if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) {
3918 // To be able to issue a direct call (optimized virtual or virtual)
3919 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information
3920 // about the method being invoked should be attached to the call site to
3921 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C).
3922 slow_call->set_override_symbolic_info(true);
3923 }
3924 set_arguments_for_java_call(slow_call);
3925 set_edges_for_java_call(slow_call);
3926 return slow_call;
3927 }
3928
3929
3930 /**
3931 * Build special case code for calls to hashCode on an object. This call may
3932 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3933 * slightly different code.
3934 */
3935 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3936 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3937 assert(!(is_virtual && is_static), "either virtual, special, or static");
3938
3939 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3940
3941 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3942 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3943 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3944 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3945 Node* obj = NULL;
3946 if (!is_static) {
3947 // Check for hashing null object
3948 obj = null_check_receiver();
3949 if (stopped()) return true; // unconditionally null
3950 result_reg->init_req(_null_path, top());
3951 result_val->init_req(_null_path, top());
3952 } else {
3953 // Do a null check, and return zero if null.
3954 // System.identityHashCode(null) == 0
3955 obj = argument(0);
3956 Node* null_ctl = top();
3957 obj = null_check_oop(obj, &null_ctl);
3958 result_reg->init_req(_null_path, null_ctl);
3959 result_val->init_req(_null_path, _gvn.intcon(0));
3960 }
3961
3962 // Unconditionally null? Then return right away.
3963 if (stopped()) {
3964 set_control( result_reg->in(_null_path));
3965 if (!stopped())
3966 set_result(result_val->in(_null_path));
3967 return true;
3968 }
3969
3970 // We only go to the fast case code if we pass a number of guards. The
3971 // paths which do not pass are accumulated in the slow_region.
3972 RegionNode* slow_region = new RegionNode(1);
3973 record_for_igvn(slow_region);
3974
3975 // If this is a virtual call, we generate a funny guard. We pull out
3976 // the vtable entry corresponding to hashCode() from the target object.
3977 // If the target method which we are calling happens to be the native
3978 // Object hashCode() method, we pass the guard. We do not need this
3979 // guard for non-virtual calls -- the caller is known to be the native
3980 // Object hashCode().
3981 if (is_virtual) {
3982 // After null check, get the object's klass.
3983 Node* obj_klass = load_object_klass(obj);
3984 generate_virtual_guard(obj_klass, slow_region);
3985 }
3986
3987 // Get the header out of the object, use LoadMarkNode when available
3988 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3989 // The control of the load must be NULL. Otherwise, the load can move before
3990 // the null check after castPP removal.
3991 Node* no_ctrl = NULL;
3992 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3993
3994 // Test the header to see if it is unlocked.
3995 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3996 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3997 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3998 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3999 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4000
4001 generate_slow_guard(test_unlocked, slow_region);
4002
4003 // Get the hash value and check to see that it has been properly assigned.
4004 // We depend on hash_mask being at most 32 bits and avoid the use of
4005 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4006 // vm: see markOop.hpp.
4007 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4008 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4009 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4010 // This hack lets the hash bits live anywhere in the mark object now, as long
4011 // as the shift drops the relevant bits into the low 32 bits. Note that
4012 // Java spec says that HashCode is an int so there's no point in capturing
4013 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4014 hshifted_header = ConvX2I(hshifted_header);
4015 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4016
4017 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4018 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4019 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4020
4021 generate_slow_guard(test_assigned, slow_region);
4022
4023 Node* init_mem = reset_memory();
4024 // fill in the rest of the null path:
4025 result_io ->init_req(_null_path, i_o());
4026 result_mem->init_req(_null_path, init_mem);
4027
4028 result_val->init_req(_fast_path, hash_val);
4029 result_reg->init_req(_fast_path, control());
4030 result_io ->init_req(_fast_path, i_o());
4031 result_mem->init_req(_fast_path, init_mem);
4032
4033 // Generate code for the slow case. We make a call to hashCode().
4034 set_control(_gvn.transform(slow_region));
4035 if (!stopped()) {
4036 // No need for PreserveJVMState, because we're using up the present state.
4037 set_all_memory(init_mem);
4038 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4039 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4040 Node* slow_result = set_results_for_java_call(slow_call);
4041 // this->control() comes from set_results_for_java_call
4042 result_reg->init_req(_slow_path, control());
4043 result_val->init_req(_slow_path, slow_result);
4044 result_io ->set_req(_slow_path, i_o());
4045 result_mem ->set_req(_slow_path, reset_memory());
4046 }
4047
4048 // Return the combined state.
4049 set_i_o( _gvn.transform(result_io) );
4050 set_all_memory( _gvn.transform(result_mem));
4051
4052 set_result(result_reg, result_val);
4053 return true;
4054 }
4055
4056 //---------------------------inline_native_getClass----------------------------
4057 // public final native Class<?> java.lang.Object.getClass();
4058 //
4059 // Build special case code for calls to getClass on an object.
4060 bool LibraryCallKit::inline_native_getClass() {
4061 Node* obj = null_check_receiver();
4062 if (stopped()) return true;
4063 set_result(load_mirror_from_klass(load_object_klass(obj)));
4064 return true;
4065 }
4066
4067 //-----------------inline_native_Reflection_getCallerClass---------------------
4068 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4069 //
4070 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4071 //
4072 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4073 // in that it must skip particular security frames and checks for
4074 // caller sensitive methods.
4075 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4076 #ifndef PRODUCT
4077 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4078 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4079 }
4080 #endif
4081
4082 if (!jvms()->has_method()) {
4083 #ifndef PRODUCT
4084 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4085 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4086 }
4087 #endif
4088 return false;
4089 }
4090
4091 // Walk back up the JVM state to find the caller at the required
4092 // depth.
4093 JVMState* caller_jvms = jvms();
4094
4095 // Cf. JVM_GetCallerClass
4096 // NOTE: Start the loop at depth 1 because the current JVM state does
4097 // not include the Reflection.getCallerClass() frame.
4098 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4099 ciMethod* m = caller_jvms->method();
4100 switch (n) {
4101 case 0:
4102 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4103 break;
4104 case 1:
4105 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4106 if (!m->caller_sensitive()) {
4107 #ifndef PRODUCT
4108 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4109 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4110 }
4111 #endif
4112 return false; // bail-out; let JVM_GetCallerClass do the work
4113 }
4114 break;
4115 default:
4116 if (!m->is_ignored_by_security_stack_walk()) {
4117 // We have reached the desired frame; return the holder class.
4118 // Acquire method holder as java.lang.Class and push as constant.
4119 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4120 ciInstance* caller_mirror = caller_klass->java_mirror();
4121 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4122
4123 #ifndef PRODUCT
4124 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4125 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());
4126 tty->print_cr(" JVM state at this point:");
4127 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4128 ciMethod* m = jvms()->of_depth(i)->method();
4129 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4130 }
4131 }
4132 #endif
4133 return true;
4134 }
4135 break;
4136 }
4137 }
4138
4139 #ifndef PRODUCT
4140 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4141 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4142 tty->print_cr(" JVM state at this point:");
4143 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4144 ciMethod* m = jvms()->of_depth(i)->method();
4145 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4146 }
4147 }
4148 #endif
4149
4150 return false; // bail-out; let JVM_GetCallerClass do the work
4151 }
4152
4153 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4154 Node* arg = argument(0);
4155 Node* result = NULL;
4156
4157 switch (id) {
4158 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4159 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4160 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4161 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4162
4163 case vmIntrinsics::_doubleToLongBits: {
4164 // two paths (plus control) merge in a wood
4165 RegionNode *r = new RegionNode(3);
4166 Node *phi = new PhiNode(r, TypeLong::LONG);
4167
4168 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4169 // Build the boolean node
4170 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4171
4172 // Branch either way.
4173 // NaN case is less traveled, which makes all the difference.
4174 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4175 Node *opt_isnan = _gvn.transform(ifisnan);
4176 assert( opt_isnan->is_If(), "Expect an IfNode");
4177 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4178 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4179
4180 set_control(iftrue);
4181
4182 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4183 Node *slow_result = longcon(nan_bits); // return NaN
4184 phi->init_req(1, _gvn.transform( slow_result ));
4185 r->init_req(1, iftrue);
4186
4187 // Else fall through
4188 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4189 set_control(iffalse);
4190
4191 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4192 r->init_req(2, iffalse);
4193
4194 // Post merge
4195 set_control(_gvn.transform(r));
4196 record_for_igvn(r);
4197
4198 C->set_has_split_ifs(true); // Has chance for split-if optimization
4199 result = phi;
4200 assert(result->bottom_type()->isa_long(), "must be");
4201 break;
4202 }
4203
4204 case vmIntrinsics::_floatToIntBits: {
4205 // two paths (plus control) merge in a wood
4206 RegionNode *r = new RegionNode(3);
4207 Node *phi = new PhiNode(r, TypeInt::INT);
4208
4209 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4210 // Build the boolean node
4211 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4212
4213 // Branch either way.
4214 // NaN case is less traveled, which makes all the difference.
4215 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4216 Node *opt_isnan = _gvn.transform(ifisnan);
4217 assert( opt_isnan->is_If(), "Expect an IfNode");
4218 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4219 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4220
4221 set_control(iftrue);
4222
4223 static const jint nan_bits = 0x7fc00000;
4224 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4225 phi->init_req(1, _gvn.transform( slow_result ));
4226 r->init_req(1, iftrue);
4227
4228 // Else fall through
4229 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4230 set_control(iffalse);
4231
4232 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4233 r->init_req(2, iffalse);
4234
4235 // Post merge
4236 set_control(_gvn.transform(r));
4237 record_for_igvn(r);
4238
4239 C->set_has_split_ifs(true); // Has chance for split-if optimization
4240 result = phi;
4241 assert(result->bottom_type()->isa_int(), "must be");
4242 break;
4243 }
4244
4245 default:
4246 fatal_unexpected_iid(id);
4247 break;
4248 }
4249 set_result(_gvn.transform(result));
4250 return true;
4251 }
4252
4253 //----------------------inline_unsafe_copyMemory-------------------------
4254 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4255 bool LibraryCallKit::inline_unsafe_copyMemory() {
4256 if (callee()->is_static()) return false; // caller must have the capability!
4257 null_check_receiver(); // null-check receiver
4258 if (stopped()) return true;
4259
4260 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4261
4262 Node* src_ptr = argument(1); // type: oop
4263 Node* src_off = ConvL2X(argument(2)); // type: long
4264 Node* dst_ptr = argument(4); // type: oop
4265 Node* dst_off = ConvL2X(argument(5)); // type: long
4266 Node* size = ConvL2X(argument(7)); // type: long
4267
4268 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4269 "fieldOffset must be byte-scaled");
4270
4271 src_ptr = access_resolve(src_ptr, ACCESS_READ);
4272 dst_ptr = access_resolve(dst_ptr, ACCESS_WRITE);
4273 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ);
4274 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE);
4275
4276 // Conservatively insert a memory barrier on all memory slices.
4277 // Do not let writes of the copy source or destination float below the copy.
4278 insert_mem_bar(Op_MemBarCPUOrder);
4279
4280 // Call it. Note that the length argument is not scaled.
4281 make_runtime_call(RC_LEAF|RC_NO_FP,
4282 OptoRuntime::fast_arraycopy_Type(),
4283 StubRoutines::unsafe_arraycopy(),
4284 "unsafe_arraycopy",
4285 TypeRawPtr::BOTTOM,
4286 src, dst, size XTOP);
4287
4288 // Do not let reads of the copy destination float above the copy.
4289 insert_mem_bar(Op_MemBarCPUOrder);
4290
4291 return true;
4292 }
4293
4294 //------------------------clone_coping-----------------------------------
4295 // Helper function for inline_native_clone.
4296 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) {
4297 assert(obj_size != NULL, "");
4298 Node* raw_obj = alloc_obj->in(1);
4299 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4300
4301 AllocateNode* alloc = NULL;
4302 if (ReduceBulkZeroing) {
4303 // We will be completely responsible for initializing this object -
4304 // mark Initialize node as complete.
4305 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4306 // The object was just allocated - there should be no any stores!
4307 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4308 // Mark as complete_with_arraycopy so that on AllocateNode
4309 // expansion, we know this AllocateNode is initialized by an array
4310 // copy and a StoreStore barrier exists after the array copy.
4311 alloc->initialization()->set_complete_with_arraycopy();
4312 }
4313
4314 // Copy the fastest available way.
4315 // TODO: generate fields copies for small objects instead.
4316 Node* size = _gvn.transform(obj_size);
4317
4318 access_clone(obj, alloc_obj, size, is_array);
4319
4320 // Do not let reads from the cloned object float above the arraycopy.
4321 if (alloc != NULL) {
4322 // Do not let stores that initialize this object be reordered with
4323 // a subsequent store that would make this object accessible by
4324 // other threads.
4325 // Record what AllocateNode this StoreStore protects so that
4326 // escape analysis can go from the MemBarStoreStoreNode to the
4327 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4328 // based on the escape status of the AllocateNode.
4329 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4330 } else {
4331 insert_mem_bar(Op_MemBarCPUOrder);
4332 }
4333 }
4334
4335 //------------------------inline_native_clone----------------------------
4336 // protected native Object java.lang.Object.clone();
4337 //
4338 // Here are the simple edge cases:
4339 // null receiver => normal trap
4340 // virtual and clone was overridden => slow path to out-of-line clone
4341 // not cloneable or finalizer => slow path to out-of-line Object.clone
4342 //
4343 // The general case has two steps, allocation and copying.
4344 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4345 //
4346 // Copying also has two cases, oop arrays and everything else.
4347 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4348 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4349 //
4350 // These steps fold up nicely if and when the cloned object's klass
4351 // can be sharply typed as an object array, a type array, or an instance.
4352 //
4353 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4354 PhiNode* result_val;
4355
4356 // Set the reexecute bit for the interpreter to reexecute
4357 // the bytecode that invokes Object.clone if deoptimization happens.
4358 { PreserveReexecuteState preexecs(this);
4359 jvms()->set_should_reexecute(true);
4360
4361 Node* obj = null_check_receiver();
4362 if (stopped()) return true;
4363
4364 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4365
4366 // If we are going to clone an instance, we need its exact type to
4367 // know the number and types of fields to convert the clone to
4368 // loads/stores. Maybe a speculative type can help us.
4369 if (!obj_type->klass_is_exact() &&
4370 obj_type->speculative_type() != NULL &&
4371 obj_type->speculative_type()->is_instance_klass()) {
4372 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4373 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4374 !spec_ik->has_injected_fields()) {
4375 ciKlass* k = obj_type->klass();
4376 if (!k->is_instance_klass() ||
4377 k->as_instance_klass()->is_interface() ||
4378 k->as_instance_klass()->has_subklass()) {
4379 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4380 }
4381 }
4382 }
4383
4384 Node* obj_klass = load_object_klass(obj);
4385 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4386 const TypeOopPtr* toop = ((tklass != NULL)
4387 ? tklass->as_instance_type()
4388 : TypeInstPtr::NOTNULL);
4389
4390 // Conservatively insert a memory barrier on all memory slices.
4391 // Do not let writes into the original float below the clone.
4392 insert_mem_bar(Op_MemBarCPUOrder);
4393
4394 // paths into result_reg:
4395 enum {
4396 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4397 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4398 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4399 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4400 PATH_LIMIT
4401 };
4402 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4403 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4404 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4405 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4406 record_for_igvn(result_reg);
4407
4408 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4409 if (array_ctl != NULL) {
4410 // It's an array.
4411 PreserveJVMState pjvms(this);
4412 set_control(array_ctl);
4413 Node* obj_length = load_array_length(obj);
4414 Node* obj_size = NULL;
4415 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4416
4417 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4418 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) {
4419 // If it is an oop array, it requires very special treatment,
4420 // because gc barriers are required when accessing the array.
4421 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4422 if (is_obja != NULL) {
4423 PreserveJVMState pjvms2(this);
4424 set_control(is_obja);
4425 obj = access_resolve(obj, ACCESS_READ);
4426 // Generate a direct call to the right arraycopy function(s).
4427 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4428 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4429 ac->set_cloneoop();
4430 Node* n = _gvn.transform(ac);
4431 assert(n == ac, "cannot disappear");
4432 ac->connect_outputs(this);
4433
4434 result_reg->init_req(_objArray_path, control());
4435 result_val->init_req(_objArray_path, alloc_obj);
4436 result_i_o ->set_req(_objArray_path, i_o());
4437 result_mem ->set_req(_objArray_path, reset_memory());
4438 }
4439 }
4440 // Otherwise, there are no barriers to worry about.
4441 // (We can dispense with card marks if we know the allocation
4442 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4443 // causes the non-eden paths to take compensating steps to
4444 // simulate a fresh allocation, so that no further
4445 // card marks are required in compiled code to initialize
4446 // the object.)
4447
4448 if (!stopped()) {
4449 copy_to_clone(obj, alloc_obj, obj_size, true);
4450
4451 // Present the results of the copy.
4452 result_reg->init_req(_array_path, control());
4453 result_val->init_req(_array_path, alloc_obj);
4454 result_i_o ->set_req(_array_path, i_o());
4455 result_mem ->set_req(_array_path, reset_memory());
4456 }
4457 }
4458
4459 // We only go to the instance fast case code if we pass a number of guards.
4460 // The paths which do not pass are accumulated in the slow_region.
4461 RegionNode* slow_region = new RegionNode(1);
4462 record_for_igvn(slow_region);
4463 if (!stopped()) {
4464 // It's an instance (we did array above). Make the slow-path tests.
4465 // If this is a virtual call, we generate a funny guard. We grab
4466 // the vtable entry corresponding to clone() from the target object.
4467 // If the target method which we are calling happens to be the
4468 // Object clone() method, we pass the guard. We do not need this
4469 // guard for non-virtual calls; the caller is known to be the native
4470 // Object clone().
4471 if (is_virtual) {
4472 generate_virtual_guard(obj_klass, slow_region);
4473 }
4474
4475 // The object must be easily cloneable and must not have a finalizer.
4476 // Both of these conditions may be checked in a single test.
4477 // We could optimize the test further, but we don't care.
4478 generate_access_flags_guard(obj_klass,
4479 // Test both conditions:
4480 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4481 // Must be cloneable but not finalizer:
4482 JVM_ACC_IS_CLONEABLE_FAST,
4483 slow_region);
4484 }
4485
4486 if (!stopped()) {
4487 // It's an instance, and it passed the slow-path tests.
4488 PreserveJVMState pjvms(this);
4489 Node* obj_size = NULL;
4490 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4491 // is reexecuted if deoptimization occurs and there could be problems when merging
4492 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4493 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4494
4495 copy_to_clone(obj, alloc_obj, obj_size, false);
4496
4497 // Present the results of the slow call.
4498 result_reg->init_req(_instance_path, control());
4499 result_val->init_req(_instance_path, alloc_obj);
4500 result_i_o ->set_req(_instance_path, i_o());
4501 result_mem ->set_req(_instance_path, reset_memory());
4502 }
4503
4504 // Generate code for the slow case. We make a call to clone().
4505 set_control(_gvn.transform(slow_region));
4506 if (!stopped()) {
4507 PreserveJVMState pjvms(this);
4508 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4509 // We need to deoptimize on exception (see comment above)
4510 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true);
4511 // this->control() comes from set_results_for_java_call
4512 result_reg->init_req(_slow_path, control());
4513 result_val->init_req(_slow_path, slow_result);
4514 result_i_o ->set_req(_slow_path, i_o());
4515 result_mem ->set_req(_slow_path, reset_memory());
4516 }
4517
4518 // Return the combined state.
4519 set_control( _gvn.transform(result_reg));
4520 set_i_o( _gvn.transform(result_i_o));
4521 set_all_memory( _gvn.transform(result_mem));
4522 } // original reexecute is set back here
4523
4524 set_result(_gvn.transform(result_val));
4525 return true;
4526 }
4527
4528 // If we have a tightly coupled allocation, the arraycopy may take care
4529 // of the array initialization. If one of the guards we insert between
4530 // the allocation and the arraycopy causes a deoptimization, an
4531 // unitialized array will escape the compiled method. To prevent that
4532 // we set the JVM state for uncommon traps between the allocation and
4533 // the arraycopy to the state before the allocation so, in case of
4534 // deoptimization, we'll reexecute the allocation and the
4535 // initialization.
4536 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4537 if (alloc != NULL) {
4538 ciMethod* trap_method = alloc->jvms()->method();
4539 int trap_bci = alloc->jvms()->bci();
4540
4541 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4542 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4543 // Make sure there's no store between the allocation and the
4544 // arraycopy otherwise visible side effects could be rexecuted
4545 // in case of deoptimization and cause incorrect execution.
4546 bool no_interfering_store = true;
4547 Node* mem = alloc->in(TypeFunc::Memory);
4548 if (mem->is_MergeMem()) {
4549 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4550 Node* n = mms.memory();
4551 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4552 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?");
4553 no_interfering_store = false;
4554 break;
4555 }
4556 }
4557 } else {
4558 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4559 Node* n = mms.memory();
4560 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4561 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?");
4562 no_interfering_store = false;
4563 break;
4564 }
4565 }
4566 }
4567
4568 if (no_interfering_store) {
4569 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4570 uint size = alloc->req();
4571 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4572 old_jvms->set_map(sfpt);
4573 for (uint i = 0; i < size; i++) {
4574 sfpt->init_req(i, alloc->in(i));
4575 }
4576 // re-push array length for deoptimization
4577 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4578 old_jvms->set_sp(old_jvms->sp()+1);
4579 old_jvms->set_monoff(old_jvms->monoff()+1);
4580 old_jvms->set_scloff(old_jvms->scloff()+1);
4581 old_jvms->set_endoff(old_jvms->endoff()+1);
4582 old_jvms->set_should_reexecute(true);
4583
4584 sfpt->set_i_o(map()->i_o());
4585 sfpt->set_memory(map()->memory());
4586 sfpt->set_control(map()->control());
4587
4588 JVMState* saved_jvms = jvms();
4589 saved_reexecute_sp = _reexecute_sp;
4590
4591 set_jvms(sfpt->jvms());
4592 _reexecute_sp = jvms()->sp();
4593
4594 return saved_jvms;
4595 }
4596 }
4597 }
4598 return NULL;
4599 }
4600
4601 // In case of a deoptimization, we restart execution at the
4602 // allocation, allocating a new array. We would leave an uninitialized
4603 // array in the heap that GCs wouldn't expect. Move the allocation
4604 // after the traps so we don't allocate the array if we
4605 // deoptimize. This is possible because tightly_coupled_allocation()
4606 // guarantees there's no observer of the allocated array at this point
4607 // and the control flow is simple enough.
4608 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4609 int saved_reexecute_sp, uint new_idx) {
4610 if (saved_jvms != NULL && !stopped()) {
4611 assert(alloc != NULL, "only with a tightly coupled allocation");
4612 // restore JVM state to the state at the arraycopy
4613 saved_jvms->map()->set_control(map()->control());
4614 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4615 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4616 // If we've improved the types of some nodes (null check) while
4617 // emitting the guards, propagate them to the current state
4618 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4619 set_jvms(saved_jvms);
4620 _reexecute_sp = saved_reexecute_sp;
4621
4622 // Remove the allocation from above the guards
4623 CallProjections callprojs;
4624 alloc->extract_projections(&callprojs, true);
4625 InitializeNode* init = alloc->initialization();
4626 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4627 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4628 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4629 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4630
4631 // move the allocation here (after the guards)
4632 _gvn.hash_delete(alloc);
4633 alloc->set_req(TypeFunc::Control, control());
4634 alloc->set_req(TypeFunc::I_O, i_o());
4635 Node *mem = reset_memory();
4636 set_all_memory(mem);
4637 alloc->set_req(TypeFunc::Memory, mem);
4638 set_control(init->proj_out_or_null(TypeFunc::Control));
4639 set_i_o(callprojs.fallthrough_ioproj);
4640
4641 // Update memory as done in GraphKit::set_output_for_allocation()
4642 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4643 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4644 if (ary_type->isa_aryptr() && length_type != NULL) {
4645 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4646 }
4647 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4648 int elemidx = C->get_alias_index(telemref);
4649 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
4650 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
4651
4652 Node* allocx = _gvn.transform(alloc);
4653 assert(allocx == alloc, "where has the allocation gone?");
4654 assert(dest->is_CheckCastPP(), "not an allocation result?");
4655
4656 _gvn.hash_delete(dest);
4657 dest->set_req(0, control());
4658 Node* destx = _gvn.transform(dest);
4659 assert(destx == dest, "where has the allocation result gone?");
4660 }
4661 }
4662
4663
4664 //------------------------------inline_arraycopy-----------------------
4665 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4666 // Object dest, int destPos,
4667 // int length);
4668 bool LibraryCallKit::inline_arraycopy() {
4669 // Get the arguments.
4670 Node* src = argument(0); // type: oop
4671 Node* src_offset = argument(1); // type: int
4672 Node* dest = argument(2); // type: oop
4673 Node* dest_offset = argument(3); // type: int
4674 Node* length = argument(4); // type: int
4675
4676 uint new_idx = C->unique();
4677
4678 // Check for allocation before we add nodes that would confuse
4679 // tightly_coupled_allocation()
4680 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4681
4682 int saved_reexecute_sp = -1;
4683 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4684 // See arraycopy_restore_alloc_state() comment
4685 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4686 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4687 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
4688 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4689
4690 // The following tests must be performed
4691 // (1) src and dest are arrays.
4692 // (2) src and dest arrays must have elements of the same BasicType
4693 // (3) src and dest must not be null.
4694 // (4) src_offset must not be negative.
4695 // (5) dest_offset must not be negative.
4696 // (6) length must not be negative.
4697 // (7) src_offset + length must not exceed length of src.
4698 // (8) dest_offset + length must not exceed length of dest.
4699 // (9) each element of an oop array must be assignable
4700
4701 // (3) src and dest must not be null.
4702 // always do this here because we need the JVM state for uncommon traps
4703 Node* null_ctl = top();
4704 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4705 assert(null_ctl->is_top(), "no null control here");
4706 dest = null_check(dest, T_ARRAY);
4707
4708 if (!can_emit_guards) {
4709 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4710 // guards but the arraycopy node could still take advantage of a
4711 // tightly allocated allocation. tightly_coupled_allocation() is
4712 // called again to make sure it takes the null check above into
4713 // account: the null check is mandatory and if it caused an
4714 // uncommon trap to be emitted then the allocation can't be
4715 // considered tightly coupled in this context.
4716 alloc = tightly_coupled_allocation(dest, NULL);
4717 }
4718
4719 bool validated = false;
4720
4721 const Type* src_type = _gvn.type(src);
4722 const Type* dest_type = _gvn.type(dest);
4723 const TypeAryPtr* top_src = src_type->isa_aryptr();
4724 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4725
4726 // Do we have the type of src?
4727 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4728 // Do we have the type of dest?
4729 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4730 // Is the type for src from speculation?
4731 bool src_spec = false;
4732 // Is the type for dest from speculation?
4733 bool dest_spec = false;
4734
4735 if ((!has_src || !has_dest) && can_emit_guards) {
4736 // We don't have sufficient type information, let's see if
4737 // speculative types can help. We need to have types for both src
4738 // and dest so that it pays off.
4739
4740 // Do we already have or could we have type information for src
4741 bool could_have_src = has_src;
4742 // Do we already have or could we have type information for dest
4743 bool could_have_dest = has_dest;
4744
4745 ciKlass* src_k = NULL;
4746 if (!has_src) {
4747 src_k = src_type->speculative_type_not_null();
4748 if (src_k != NULL && src_k->is_array_klass()) {
4749 could_have_src = true;
4750 }
4751 }
4752
4753 ciKlass* dest_k = NULL;
4754 if (!has_dest) {
4755 dest_k = dest_type->speculative_type_not_null();
4756 if (dest_k != NULL && dest_k->is_array_klass()) {
4757 could_have_dest = true;
4758 }
4759 }
4760
4761 if (could_have_src && could_have_dest) {
4762 // This is going to pay off so emit the required guards
4763 if (!has_src) {
4764 src = maybe_cast_profiled_obj(src, src_k, true);
4765 src_type = _gvn.type(src);
4766 top_src = src_type->isa_aryptr();
4767 has_src = (top_src != NULL && top_src->klass() != NULL);
4768 src_spec = true;
4769 }
4770 if (!has_dest) {
4771 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4772 dest_type = _gvn.type(dest);
4773 top_dest = dest_type->isa_aryptr();
4774 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4775 dest_spec = true;
4776 }
4777 }
4778 }
4779
4780 if (has_src && has_dest && can_emit_guards) {
4781 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4782 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4783 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4784 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4785
4786 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4787 // If both arrays are object arrays then having the exact types
4788 // for both will remove the need for a subtype check at runtime
4789 // before the call and may make it possible to pick a faster copy
4790 // routine (without a subtype check on every element)
4791 // Do we have the exact type of src?
4792 bool could_have_src = src_spec;
4793 // Do we have the exact type of dest?
4794 bool could_have_dest = dest_spec;
4795 ciKlass* src_k = top_src->klass();
4796 ciKlass* dest_k = top_dest->klass();
4797 if (!src_spec) {
4798 src_k = src_type->speculative_type_not_null();
4799 if (src_k != NULL && src_k->is_array_klass()) {
4800 could_have_src = true;
4801 }
4802 }
4803 if (!dest_spec) {
4804 dest_k = dest_type->speculative_type_not_null();
4805 if (dest_k != NULL && dest_k->is_array_klass()) {
4806 could_have_dest = true;
4807 }
4808 }
4809 if (could_have_src && could_have_dest) {
4810 // If we can have both exact types, emit the missing guards
4811 if (could_have_src && !src_spec) {
4812 src = maybe_cast_profiled_obj(src, src_k, true);
4813 }
4814 if (could_have_dest && !dest_spec) {
4815 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4816 }
4817 }
4818 }
4819 }
4820
4821 ciMethod* trap_method = method();
4822 int trap_bci = bci();
4823 if (saved_jvms != NULL) {
4824 trap_method = alloc->jvms()->method();
4825 trap_bci = alloc->jvms()->bci();
4826 }
4827
4828 bool negative_length_guard_generated = false;
4829
4830 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4831 can_emit_guards &&
4832 !src->is_top() && !dest->is_top()) {
4833 // validate arguments: enables transformation the ArrayCopyNode
4834 validated = true;
4835
4836 RegionNode* slow_region = new RegionNode(1);
4837 record_for_igvn(slow_region);
4838
4839 // (1) src and dest are arrays.
4840 generate_non_array_guard(load_object_klass(src), slow_region);
4841 generate_non_array_guard(load_object_klass(dest), slow_region);
4842
4843 // (2) src and dest arrays must have elements of the same BasicType
4844 // done at macro expansion or at Ideal transformation time
4845
4846 // (4) src_offset must not be negative.
4847 generate_negative_guard(src_offset, slow_region);
4848
4849 // (5) dest_offset must not be negative.
4850 generate_negative_guard(dest_offset, slow_region);
4851
4852 // (7) src_offset + length must not exceed length of src.
4853 generate_limit_guard(src_offset, length,
4854 load_array_length(src),
4855 slow_region);
4856
4857 // (8) dest_offset + length must not exceed length of dest.
4858 generate_limit_guard(dest_offset, length,
4859 load_array_length(dest),
4860 slow_region);
4861
4862 // (6) length must not be negative.
4863 // This is also checked in generate_arraycopy() during macro expansion, but
4864 // we also have to check it here for the case where the ArrayCopyNode will
4865 // be eliminated by Escape Analysis.
4866 if (EliminateAllocations) {
4867 generate_negative_guard(length, slow_region);
4868 negative_length_guard_generated = true;
4869 }
4870
4871 // (9) each element of an oop array must be assignable
4872 Node* src_klass = load_object_klass(src);
4873 Node* dest_klass = load_object_klass(dest);
4874 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4875
4876 if (not_subtype_ctrl != top()) {
4877 PreserveJVMState pjvms(this);
4878 set_control(not_subtype_ctrl);
4879 uncommon_trap(Deoptimization::Reason_intrinsic,
4880 Deoptimization::Action_make_not_entrant);
4881 assert(stopped(), "Should be stopped");
4882 }
4883 {
4884 PreserveJVMState pjvms(this);
4885 set_control(_gvn.transform(slow_region));
4886 uncommon_trap(Deoptimization::Reason_intrinsic,
4887 Deoptimization::Action_make_not_entrant);
4888 assert(stopped(), "Should be stopped");
4889 }
4890
4891 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
4892 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
4893 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
4894 }
4895
4896 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
4897
4898 if (stopped()) {
4899 return true;
4900 }
4901
4902 Node* new_src = access_resolve(src, ACCESS_READ);
4903 Node* new_dest = access_resolve(dest, ACCESS_WRITE);
4904
4905 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, new_src, src_offset, new_dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
4906 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4907 // so the compiler has a chance to eliminate them: during macro expansion,
4908 // we have to set their control (CastPP nodes are eliminated).
4909 load_object_klass(src), load_object_klass(dest),
4910 load_array_length(src), load_array_length(dest));
4911
4912 ac->set_arraycopy(validated);
4913
4914 Node* n = _gvn.transform(ac);
4915 if (n == ac) {
4916 ac->connect_outputs(this);
4917 } else {
4918 assert(validated, "shouldn't transform if all arguments not validated");
4919 set_all_memory(n);
4920 }
4921 clear_upper_avx();
4922
4923
4924 return true;
4925 }
4926
4927
4928 // Helper function which determines if an arraycopy immediately follows
4929 // an allocation, with no intervening tests or other escapes for the object.
4930 AllocateArrayNode*
4931 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4932 RegionNode* slow_region) {
4933 if (stopped()) return NULL; // no fast path
4934 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4935
4936 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4937 if (alloc == NULL) return NULL;
4938
4939 Node* rawmem = memory(Compile::AliasIdxRaw);
4940 // Is the allocation's memory state untouched?
4941 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4942 // Bail out if there have been raw-memory effects since the allocation.
4943 // (Example: There might have been a call or safepoint.)
4944 return NULL;
4945 }
4946 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4947 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4948 return NULL;
4949 }
4950
4951 // There must be no unexpected observers of this allocation.
4952 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4953 Node* obs = ptr->fast_out(i);
4954 if (obs != this->map()) {
4955 return NULL;
4956 }
4957 }
4958
4959 // This arraycopy must unconditionally follow the allocation of the ptr.
4960 Node* alloc_ctl = ptr->in(0);
4961 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4962
4963 Node* ctl = control();
4964 while (ctl != alloc_ctl) {
4965 // There may be guards which feed into the slow_region.
4966 // Any other control flow means that we might not get a chance
4967 // to finish initializing the allocated object.
4968 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4969 IfNode* iff = ctl->in(0)->as_If();
4970 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
4971 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4972 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4973 ctl = iff->in(0); // This test feeds the known slow_region.
4974 continue;
4975 }
4976 // One more try: Various low-level checks bottom out in
4977 // uncommon traps. If the debug-info of the trap omits
4978 // any reference to the allocation, as we've already
4979 // observed, then there can be no objection to the trap.
4980 bool found_trap = false;
4981 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4982 Node* obs = not_ctl->fast_out(j);
4983 if (obs->in(0) == not_ctl && obs->is_Call() &&
4984 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4985 found_trap = true; break;
4986 }
4987 }
4988 if (found_trap) {
4989 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4990 continue;
4991 }
4992 }
4993 return NULL;
4994 }
4995
4996 // If we get this far, we have an allocation which immediately
4997 // precedes the arraycopy, and we can take over zeroing the new object.
4998 // The arraycopy will finish the initialization, and provide
4999 // a new control state to which we will anchor the destination pointer.
5000
5001 return alloc;
5002 }
5003
5004 //-------------inline_encodeISOArray-----------------------------------
5005 // encode char[] to byte[] in ISO_8859_1
5006 bool LibraryCallKit::inline_encodeISOArray() {
5007 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5008 // no receiver since it is static method
5009 Node *src = argument(0);
5010 Node *src_offset = argument(1);
5011 Node *dst = argument(2);
5012 Node *dst_offset = argument(3);
5013 Node *length = argument(4);
5014
5015 src = must_be_not_null(src, true);
5016 dst = must_be_not_null(dst, true);
5017
5018 src = access_resolve(src, ACCESS_READ);
5019 dst = access_resolve(dst, ACCESS_WRITE);
5020
5021 const Type* src_type = src->Value(&_gvn);
5022 const Type* dst_type = dst->Value(&_gvn);
5023 const TypeAryPtr* top_src = src_type->isa_aryptr();
5024 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5025 if (top_src == NULL || top_src->klass() == NULL ||
5026 top_dest == NULL || top_dest->klass() == NULL) {
5027 // failed array check
5028 return false;
5029 }
5030
5031 // Figure out the size and type of the elements we will be copying.
5032 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5033 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5034 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5035 return false;
5036 }
5037
5038 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5039 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5040 // 'src_start' points to src array + scaled offset
5041 // 'dst_start' points to dst array + scaled offset
5042
5043 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5044 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5045 enc = _gvn.transform(enc);
5046 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5047 set_memory(res_mem, mtype);
5048 set_result(enc);
5049 clear_upper_avx();
5050
5051 return true;
5052 }
5053
5054 //-------------inline_multiplyToLen-----------------------------------
5055 bool LibraryCallKit::inline_multiplyToLen() {
5056 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5057
5058 address stubAddr = StubRoutines::multiplyToLen();
5059 if (stubAddr == NULL) {
5060 return false; // Intrinsic's stub is not implemented on this platform
5061 }
5062 const char* stubName = "multiplyToLen";
5063
5064 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5065
5066 // no receiver because it is a static method
5067 Node* x = argument(0);
5068 Node* xlen = argument(1);
5069 Node* y = argument(2);
5070 Node* ylen = argument(3);
5071 Node* z = argument(4);
5072
5073 x = must_be_not_null(x, true);
5074 y = must_be_not_null(y, true);
5075
5076 x = access_resolve(x, ACCESS_READ);
5077 y = access_resolve(y, ACCESS_READ);
5078 z = access_resolve(z, ACCESS_WRITE);
5079
5080 const Type* x_type = x->Value(&_gvn);
5081 const Type* y_type = y->Value(&_gvn);
5082 const TypeAryPtr* top_x = x_type->isa_aryptr();
5083 const TypeAryPtr* top_y = y_type->isa_aryptr();
5084 if (top_x == NULL || top_x->klass() == NULL ||
5085 top_y == NULL || top_y->klass() == NULL) {
5086 // failed array check
5087 return false;
5088 }
5089
5090 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5091 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5092 if (x_elem != T_INT || y_elem != T_INT) {
5093 return false;
5094 }
5095
5096 // Set the original stack and the reexecute bit for the interpreter to reexecute
5097 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5098 // on the return from z array allocation in runtime.
5099 { PreserveReexecuteState preexecs(this);
5100 jvms()->set_should_reexecute(true);
5101
5102 Node* x_start = array_element_address(x, intcon(0), x_elem);
5103 Node* y_start = array_element_address(y, intcon(0), y_elem);
5104 // 'x_start' points to x array + scaled xlen
5105 // 'y_start' points to y array + scaled ylen
5106
5107 // Allocate the result array
5108 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5109 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5110 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5111
5112 IdealKit ideal(this);
5113
5114 #define __ ideal.
5115 Node* one = __ ConI(1);
5116 Node* zero = __ ConI(0);
5117 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5118 __ set(need_alloc, zero);
5119 __ set(z_alloc, z);
5120 __ if_then(z, BoolTest::eq, null()); {
5121 __ increment (need_alloc, one);
5122 } __ else_(); {
5123 // Update graphKit memory and control from IdealKit.
5124 sync_kit(ideal);
5125 Node *cast = new CastPPNode(z, TypePtr::NOTNULL);
5126 cast->init_req(0, control());
5127 _gvn.set_type(cast, cast->bottom_type());
5128 C->record_for_igvn(cast);
5129
5130 Node* zlen_arg = load_array_length(cast);
5131 // Update IdealKit memory and control from graphKit.
5132 __ sync_kit(this);
5133 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5134 __ increment (need_alloc, one);
5135 } __ end_if();
5136 } __ end_if();
5137
5138 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5139 // Update graphKit memory and control from IdealKit.
5140 sync_kit(ideal);
5141 Node * narr = new_array(klass_node, zlen, 1);
5142 // Update IdealKit memory and control from graphKit.
5143 __ sync_kit(this);
5144 __ set(z_alloc, narr);
5145 } __ end_if();
5146
5147 sync_kit(ideal);
5148 z = __ value(z_alloc);
5149 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5150 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5151 // Final sync IdealKit and GraphKit.
5152 final_sync(ideal);
5153 #undef __
5154
5155 Node* z_start = array_element_address(z, intcon(0), T_INT);
5156
5157 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5158 OptoRuntime::multiplyToLen_Type(),
5159 stubAddr, stubName, TypePtr::BOTTOM,
5160 x_start, xlen, y_start, ylen, z_start, zlen);
5161 } // original reexecute is set back here
5162
5163 C->set_has_split_ifs(true); // Has chance for split-if optimization
5164 set_result(z);
5165 return true;
5166 }
5167
5168 //-------------inline_squareToLen------------------------------------
5169 bool LibraryCallKit::inline_squareToLen() {
5170 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5171
5172 address stubAddr = StubRoutines::squareToLen();
5173 if (stubAddr == NULL) {
5174 return false; // Intrinsic's stub is not implemented on this platform
5175 }
5176 const char* stubName = "squareToLen";
5177
5178 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5179
5180 Node* x = argument(0);
5181 Node* len = argument(1);
5182 Node* z = argument(2);
5183 Node* zlen = argument(3);
5184
5185 x = must_be_not_null(x, true);
5186 z = must_be_not_null(z, true);
5187
5188 x = access_resolve(x, ACCESS_READ);
5189 z = access_resolve(z, ACCESS_WRITE);
5190
5191 const Type* x_type = x->Value(&_gvn);
5192 const Type* z_type = z->Value(&_gvn);
5193 const TypeAryPtr* top_x = x_type->isa_aryptr();
5194 const TypeAryPtr* top_z = z_type->isa_aryptr();
5195 if (top_x == NULL || top_x->klass() == NULL ||
5196 top_z == NULL || top_z->klass() == NULL) {
5197 // failed array check
5198 return false;
5199 }
5200
5201 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5202 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5203 if (x_elem != T_INT || z_elem != T_INT) {
5204 return false;
5205 }
5206
5207
5208 Node* x_start = array_element_address(x, intcon(0), x_elem);
5209 Node* z_start = array_element_address(z, intcon(0), z_elem);
5210
5211 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5212 OptoRuntime::squareToLen_Type(),
5213 stubAddr, stubName, TypePtr::BOTTOM,
5214 x_start, len, z_start, zlen);
5215
5216 set_result(z);
5217 return true;
5218 }
5219
5220 //-------------inline_mulAdd------------------------------------------
5221 bool LibraryCallKit::inline_mulAdd() {
5222 assert(UseMulAddIntrinsic, "not implemented on this platform");
5223
5224 address stubAddr = StubRoutines::mulAdd();
5225 if (stubAddr == NULL) {
5226 return false; // Intrinsic's stub is not implemented on this platform
5227 }
5228 const char* stubName = "mulAdd";
5229
5230 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5231
5232 Node* out = argument(0);
5233 Node* in = argument(1);
5234 Node* offset = argument(2);
5235 Node* len = argument(3);
5236 Node* k = argument(4);
5237
5238 out = must_be_not_null(out, true);
5239
5240 in = access_resolve(in, ACCESS_READ);
5241 out = access_resolve(out, ACCESS_WRITE);
5242
5243 const Type* out_type = out->Value(&_gvn);
5244 const Type* in_type = in->Value(&_gvn);
5245 const TypeAryPtr* top_out = out_type->isa_aryptr();
5246 const TypeAryPtr* top_in = in_type->isa_aryptr();
5247 if (top_out == NULL || top_out->klass() == NULL ||
5248 top_in == NULL || top_in->klass() == NULL) {
5249 // failed array check
5250 return false;
5251 }
5252
5253 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5254 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5255 if (out_elem != T_INT || in_elem != T_INT) {
5256 return false;
5257 }
5258
5259 Node* outlen = load_array_length(out);
5260 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5261 Node* out_start = array_element_address(out, intcon(0), out_elem);
5262 Node* in_start = array_element_address(in, intcon(0), in_elem);
5263
5264 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5265 OptoRuntime::mulAdd_Type(),
5266 stubAddr, stubName, TypePtr::BOTTOM,
5267 out_start,in_start, new_offset, len, k);
5268 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5269 set_result(result);
5270 return true;
5271 }
5272
5273 //-------------inline_montgomeryMultiply-----------------------------------
5274 bool LibraryCallKit::inline_montgomeryMultiply() {
5275 address stubAddr = StubRoutines::montgomeryMultiply();
5276 if (stubAddr == NULL) {
5277 return false; // Intrinsic's stub is not implemented on this platform
5278 }
5279
5280 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5281 const char* stubName = "montgomery_multiply";
5282
5283 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5284
5285 Node* a = argument(0);
5286 Node* b = argument(1);
5287 Node* n = argument(2);
5288 Node* len = argument(3);
5289 Node* inv = argument(4);
5290 Node* m = argument(6);
5291
5292 a = access_resolve(a, ACCESS_READ);
5293 b = access_resolve(b, ACCESS_READ);
5294 n = access_resolve(n, ACCESS_READ);
5295 m = access_resolve(m, ACCESS_WRITE);
5296
5297 const Type* a_type = a->Value(&_gvn);
5298 const TypeAryPtr* top_a = a_type->isa_aryptr();
5299 const Type* b_type = b->Value(&_gvn);
5300 const TypeAryPtr* top_b = b_type->isa_aryptr();
5301 const Type* n_type = a->Value(&_gvn);
5302 const TypeAryPtr* top_n = n_type->isa_aryptr();
5303 const Type* m_type = a->Value(&_gvn);
5304 const TypeAryPtr* top_m = m_type->isa_aryptr();
5305 if (top_a == NULL || top_a->klass() == NULL ||
5306 top_b == NULL || top_b->klass() == NULL ||
5307 top_n == NULL || top_n->klass() == NULL ||
5308 top_m == NULL || top_m->klass() == NULL) {
5309 // failed array check
5310 return false;
5311 }
5312
5313 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5314 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5315 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5316 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5317 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5318 return false;
5319 }
5320
5321 // Make the call
5322 {
5323 Node* a_start = array_element_address(a, intcon(0), a_elem);
5324 Node* b_start = array_element_address(b, intcon(0), b_elem);
5325 Node* n_start = array_element_address(n, intcon(0), n_elem);
5326 Node* m_start = array_element_address(m, intcon(0), m_elem);
5327
5328 Node* call = make_runtime_call(RC_LEAF,
5329 OptoRuntime::montgomeryMultiply_Type(),
5330 stubAddr, stubName, TypePtr::BOTTOM,
5331 a_start, b_start, n_start, len, inv, top(),
5332 m_start);
5333 set_result(m);
5334 }
5335
5336 return true;
5337 }
5338
5339 bool LibraryCallKit::inline_montgomerySquare() {
5340 address stubAddr = StubRoutines::montgomerySquare();
5341 if (stubAddr == NULL) {
5342 return false; // Intrinsic's stub is not implemented on this platform
5343 }
5344
5345 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5346 const char* stubName = "montgomery_square";
5347
5348 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5349
5350 Node* a = argument(0);
5351 Node* n = argument(1);
5352 Node* len = argument(2);
5353 Node* inv = argument(3);
5354 Node* m = argument(5);
5355
5356 a = access_resolve(a, ACCESS_READ);
5357 n = access_resolve(n, ACCESS_READ);
5358 m = access_resolve(m, ACCESS_WRITE);
5359
5360 const Type* a_type = a->Value(&_gvn);
5361 const TypeAryPtr* top_a = a_type->isa_aryptr();
5362 const Type* n_type = a->Value(&_gvn);
5363 const TypeAryPtr* top_n = n_type->isa_aryptr();
5364 const Type* m_type = a->Value(&_gvn);
5365 const TypeAryPtr* top_m = m_type->isa_aryptr();
5366 if (top_a == NULL || top_a->klass() == NULL ||
5367 top_n == NULL || top_n->klass() == NULL ||
5368 top_m == NULL || top_m->klass() == NULL) {
5369 // failed array check
5370 return false;
5371 }
5372
5373 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5374 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5375 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5376 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5377 return false;
5378 }
5379
5380 // Make the call
5381 {
5382 Node* a_start = array_element_address(a, intcon(0), a_elem);
5383 Node* n_start = array_element_address(n, intcon(0), n_elem);
5384 Node* m_start = array_element_address(m, intcon(0), m_elem);
5385
5386 Node* call = make_runtime_call(RC_LEAF,
5387 OptoRuntime::montgomerySquare_Type(),
5388 stubAddr, stubName, TypePtr::BOTTOM,
5389 a_start, n_start, len, inv, top(),
5390 m_start);
5391 set_result(m);
5392 }
5393
5394 return true;
5395 }
5396
5397 //-------------inline_vectorizedMismatch------------------------------
5398 bool LibraryCallKit::inline_vectorizedMismatch() {
5399 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5400
5401 address stubAddr = StubRoutines::vectorizedMismatch();
5402 if (stubAddr == NULL) {
5403 return false; // Intrinsic's stub is not implemented on this platform
5404 }
5405 const char* stubName = "vectorizedMismatch";
5406 int size_l = callee()->signature()->size();
5407 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5408
5409 Node* obja = argument(0);
5410 Node* aoffset = argument(1);
5411 Node* objb = argument(3);
5412 Node* boffset = argument(4);
5413 Node* length = argument(6);
5414 Node* scale = argument(7);
5415
5416 const Type* a_type = obja->Value(&_gvn);
5417 const Type* b_type = objb->Value(&_gvn);
5418 const TypeAryPtr* top_a = a_type->isa_aryptr();
5419 const TypeAryPtr* top_b = b_type->isa_aryptr();
5420 if (top_a == NULL || top_a->klass() == NULL ||
5421 top_b == NULL || top_b->klass() == NULL) {
5422 // failed array check
5423 return false;
5424 }
5425
5426 Node* call;
5427 jvms()->set_should_reexecute(true);
5428
5429 obja = access_resolve(obja, ACCESS_READ);
5430 objb = access_resolve(objb, ACCESS_READ);
5431 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ);
5432 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ);
5433
5434 call = make_runtime_call(RC_LEAF,
5435 OptoRuntime::vectorizedMismatch_Type(),
5436 stubAddr, stubName, TypePtr::BOTTOM,
5437 obja_adr, objb_adr, length, scale);
5438
5439 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5440 set_result(result);
5441 return true;
5442 }
5443
5444 /**
5445 * Calculate CRC32 for byte.
5446 * int java.util.zip.CRC32.update(int crc, int b)
5447 */
5448 bool LibraryCallKit::inline_updateCRC32() {
5449 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5450 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5451 // no receiver since it is static method
5452 Node* crc = argument(0); // type: int
5453 Node* b = argument(1); // type: int
5454
5455 /*
5456 * int c = ~ crc;
5457 * b = timesXtoThe32[(b ^ c) & 0xFF];
5458 * b = b ^ (c >>> 8);
5459 * crc = ~b;
5460 */
5461
5462 Node* M1 = intcon(-1);
5463 crc = _gvn.transform(new XorINode(crc, M1));
5464 Node* result = _gvn.transform(new XorINode(crc, b));
5465 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5466
5467 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5468 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5469 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5470 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5471
5472 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5473 result = _gvn.transform(new XorINode(crc, result));
5474 result = _gvn.transform(new XorINode(result, M1));
5475 set_result(result);
5476 return true;
5477 }
5478
5479 /**
5480 * Calculate CRC32 for byte[] array.
5481 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5482 */
5483 bool LibraryCallKit::inline_updateBytesCRC32() {
5484 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5485 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5486 // no receiver since it is static method
5487 Node* crc = argument(0); // type: int
5488 Node* src = argument(1); // type: oop
5489 Node* offset = argument(2); // type: int
5490 Node* length = argument(3); // type: int
5491
5492 const Type* src_type = src->Value(&_gvn);
5493 const TypeAryPtr* top_src = src_type->isa_aryptr();
5494 if (top_src == NULL || top_src->klass() == NULL) {
5495 // failed array check
5496 return false;
5497 }
5498
5499 // Figure out the size and type of the elements we will be copying.
5500 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5501 if (src_elem != T_BYTE) {
5502 return false;
5503 }
5504
5505 // 'src_start' points to src array + scaled offset
5506 src = must_be_not_null(src, true);
5507 src = access_resolve(src, ACCESS_READ);
5508 Node* src_start = array_element_address(src, offset, src_elem);
5509
5510 // We assume that range check is done by caller.
5511 // TODO: generate range check (offset+length < src.length) in debug VM.
5512
5513 // Call the stub.
5514 address stubAddr = StubRoutines::updateBytesCRC32();
5515 const char *stubName = "updateBytesCRC32";
5516
5517 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5518 stubAddr, stubName, TypePtr::BOTTOM,
5519 crc, src_start, length);
5520 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5521 set_result(result);
5522 return true;
5523 }
5524
5525 /**
5526 * Calculate CRC32 for ByteBuffer.
5527 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5528 */
5529 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5530 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5531 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5532 // no receiver since it is static method
5533 Node* crc = argument(0); // type: int
5534 Node* src = argument(1); // type: long
5535 Node* offset = argument(3); // type: int
5536 Node* length = argument(4); // type: int
5537
5538 src = ConvL2X(src); // adjust Java long to machine word
5539 Node* base = _gvn.transform(new CastX2PNode(src));
5540 offset = ConvI2X(offset);
5541
5542 // 'src_start' points to src array + scaled offset
5543 Node* src_start = basic_plus_adr(top(), base, offset);
5544
5545 // Call the stub.
5546 address stubAddr = StubRoutines::updateBytesCRC32();
5547 const char *stubName = "updateBytesCRC32";
5548
5549 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5550 stubAddr, stubName, TypePtr::BOTTOM,
5551 crc, src_start, length);
5552 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5553 set_result(result);
5554 return true;
5555 }
5556
5557 //------------------------------get_table_from_crc32c_class-----------------------
5558 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5559 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5560 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5561
5562 return table;
5563 }
5564
5565 //------------------------------inline_updateBytesCRC32C-----------------------
5566 //
5567 // Calculate CRC32C for byte[] array.
5568 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5569 //
5570 bool LibraryCallKit::inline_updateBytesCRC32C() {
5571 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5572 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5573 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5574 // no receiver since it is a static method
5575 Node* crc = argument(0); // type: int
5576 Node* src = argument(1); // type: oop
5577 Node* offset = argument(2); // type: int
5578 Node* end = argument(3); // type: int
5579
5580 Node* length = _gvn.transform(new SubINode(end, offset));
5581
5582 const Type* src_type = src->Value(&_gvn);
5583 const TypeAryPtr* top_src = src_type->isa_aryptr();
5584 if (top_src == NULL || top_src->klass() == NULL) {
5585 // failed array check
5586 return false;
5587 }
5588
5589 // Figure out the size and type of the elements we will be copying.
5590 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5591 if (src_elem != T_BYTE) {
5592 return false;
5593 }
5594
5595 // 'src_start' points to src array + scaled offset
5596 src = must_be_not_null(src, true);
5597 src = access_resolve(src, ACCESS_READ);
5598 Node* src_start = array_element_address(src, offset, src_elem);
5599
5600 // static final int[] byteTable in class CRC32C
5601 Node* table = get_table_from_crc32c_class(callee()->holder());
5602 table = must_be_not_null(table, true);
5603 table = access_resolve(table, ACCESS_READ);
5604 Node* table_start = array_element_address(table, intcon(0), T_INT);
5605
5606 // We assume that range check is done by caller.
5607 // TODO: generate range check (offset+length < src.length) in debug VM.
5608
5609 // Call the stub.
5610 address stubAddr = StubRoutines::updateBytesCRC32C();
5611 const char *stubName = "updateBytesCRC32C";
5612
5613 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5614 stubAddr, stubName, TypePtr::BOTTOM,
5615 crc, src_start, length, table_start);
5616 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5617 set_result(result);
5618 return true;
5619 }
5620
5621 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5622 //
5623 // Calculate CRC32C for DirectByteBuffer.
5624 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5625 //
5626 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5627 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5628 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5629 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5630 // no receiver since it is a static method
5631 Node* crc = argument(0); // type: int
5632 Node* src = argument(1); // type: long
5633 Node* offset = argument(3); // type: int
5634 Node* end = argument(4); // type: int
5635
5636 Node* length = _gvn.transform(new SubINode(end, offset));
5637
5638 src = ConvL2X(src); // adjust Java long to machine word
5639 Node* base = _gvn.transform(new CastX2PNode(src));
5640 offset = ConvI2X(offset);
5641
5642 // 'src_start' points to src array + scaled offset
5643 Node* src_start = basic_plus_adr(top(), base, offset);
5644
5645 // static final int[] byteTable in class CRC32C
5646 Node* table = get_table_from_crc32c_class(callee()->holder());
5647 table = must_be_not_null(table, true);
5648 table = access_resolve(table, ACCESS_READ);
5649 Node* table_start = array_element_address(table, intcon(0), T_INT);
5650
5651 // Call the stub.
5652 address stubAddr = StubRoutines::updateBytesCRC32C();
5653 const char *stubName = "updateBytesCRC32C";
5654
5655 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5656 stubAddr, stubName, TypePtr::BOTTOM,
5657 crc, src_start, length, table_start);
5658 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5659 set_result(result);
5660 return true;
5661 }
5662
5663 //------------------------------inline_updateBytesAdler32----------------------
5664 //
5665 // Calculate Adler32 checksum for byte[] array.
5666 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5667 //
5668 bool LibraryCallKit::inline_updateBytesAdler32() {
5669 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5670 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5671 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5672 // no receiver since it is static method
5673 Node* crc = argument(0); // type: int
5674 Node* src = argument(1); // type: oop
5675 Node* offset = argument(2); // type: int
5676 Node* length = argument(3); // type: int
5677
5678 const Type* src_type = src->Value(&_gvn);
5679 const TypeAryPtr* top_src = src_type->isa_aryptr();
5680 if (top_src == NULL || top_src->klass() == NULL) {
5681 // failed array check
5682 return false;
5683 }
5684
5685 // Figure out the size and type of the elements we will be copying.
5686 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5687 if (src_elem != T_BYTE) {
5688 return false;
5689 }
5690
5691 // 'src_start' points to src array + scaled offset
5692 src = access_resolve(src, ACCESS_READ);
5693 Node* src_start = array_element_address(src, offset, src_elem);
5694
5695 // We assume that range check is done by caller.
5696 // TODO: generate range check (offset+length < src.length) in debug VM.
5697
5698 // Call the stub.
5699 address stubAddr = StubRoutines::updateBytesAdler32();
5700 const char *stubName = "updateBytesAdler32";
5701
5702 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5703 stubAddr, stubName, TypePtr::BOTTOM,
5704 crc, src_start, length);
5705 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5706 set_result(result);
5707 return true;
5708 }
5709
5710 //------------------------------inline_updateByteBufferAdler32---------------
5711 //
5712 // Calculate Adler32 checksum for DirectByteBuffer.
5713 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5714 //
5715 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5716 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5717 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5718 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5719 // no receiver since it is static method
5720 Node* crc = argument(0); // type: int
5721 Node* src = argument(1); // type: long
5722 Node* offset = argument(3); // type: int
5723 Node* length = argument(4); // type: int
5724
5725 src = ConvL2X(src); // adjust Java long to machine word
5726 Node* base = _gvn.transform(new CastX2PNode(src));
5727 offset = ConvI2X(offset);
5728
5729 // 'src_start' points to src array + scaled offset
5730 Node* src_start = basic_plus_adr(top(), base, offset);
5731
5732 // Call the stub.
5733 address stubAddr = StubRoutines::updateBytesAdler32();
5734 const char *stubName = "updateBytesAdler32";
5735
5736 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5737 stubAddr, stubName, TypePtr::BOTTOM,
5738 crc, src_start, length);
5739
5740 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5741 set_result(result);
5742 return true;
5743 }
5744
5745 //----------------------------inline_reference_get----------------------------
5746 // public T java.lang.ref.Reference.get();
5747 bool LibraryCallKit::inline_reference_get() {
5748 const int referent_offset = java_lang_ref_Reference::referent_offset;
5749 guarantee(referent_offset > 0, "should have already been set");
5750
5751 // Get the argument:
5752 Node* reference_obj = null_check_receiver();
5753 if (stopped()) return true;
5754
5755 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr();
5756 assert(tinst != NULL, "obj is null");
5757 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5758 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass();
5759 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"),
5760 ciSymbol::make("Ljava/lang/Object;"),
5761 false);
5762 assert (field != NULL, "undefined field");
5763
5764 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5765 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5766
5767 ciInstanceKlass* klass = env()->Object_klass();
5768 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5769
5770 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF;
5771 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators);
5772 // Add memory barrier to prevent commoning reads from this field
5773 // across safepoint since GC can change its value.
5774 insert_mem_bar(Op_MemBarCPUOrder);
5775
5776 set_result(result);
5777 return true;
5778 }
5779
5780
5781 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5782 bool is_exact=true, bool is_static=false,
5783 ciInstanceKlass * fromKls=NULL) {
5784 if (fromKls == NULL) {
5785 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5786 assert(tinst != NULL, "obj is null");
5787 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5788 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5789 fromKls = tinst->klass()->as_instance_klass();
5790 } else {
5791 assert(is_static, "only for static field access");
5792 }
5793 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5794 ciSymbol::make(fieldTypeString),
5795 is_static);
5796
5797 assert (field != NULL, "undefined field");
5798 if (field == NULL) return (Node *) NULL;
5799
5800 if (is_static) {
5801 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5802 fromObj = makecon(tip);
5803 }
5804
5805 // Next code copied from Parse::do_get_xxx():
5806
5807 // Compute address and memory type.
5808 int offset = field->offset_in_bytes();
5809 bool is_vol = field->is_volatile();
5810 ciType* field_klass = field->type();
5811 assert(field_klass->is_loaded(), "should be loaded");
5812 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5813 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5814 BasicType bt = field->layout_type();
5815
5816 // Build the resultant type of the load
5817 const Type *type;
5818 if (bt == T_OBJECT) {
5819 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5820 } else {
5821 type = Type::get_const_basic_type(bt);
5822 }
5823
5824 DecoratorSet decorators = IN_HEAP;
5825
5826 if (is_vol) {
5827 decorators |= MO_SEQ_CST;
5828 }
5829
5830 return access_load_at(fromObj, adr, adr_type, type, bt, decorators);
5831 }
5832
5833 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5834 bool is_exact = true, bool is_static = false,
5835 ciInstanceKlass * fromKls = NULL) {
5836 if (fromKls == NULL) {
5837 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5838 assert(tinst != NULL, "obj is null");
5839 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5840 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5841 fromKls = tinst->klass()->as_instance_klass();
5842 }
5843 else {
5844 assert(is_static, "only for static field access");
5845 }
5846 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5847 ciSymbol::make(fieldTypeString),
5848 is_static);
5849
5850 assert(field != NULL, "undefined field");
5851 assert(!field->is_volatile(), "not defined for volatile fields");
5852
5853 if (is_static) {
5854 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5855 fromObj = makecon(tip);
5856 }
5857
5858 // Next code copied from Parse::do_get_xxx():
5859
5860 // Compute address and memory type.
5861 int offset = field->offset_in_bytes();
5862 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5863
5864 return adr;
5865 }
5866
5867 //------------------------------inline_aescrypt_Block-----------------------
5868 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5869 address stubAddr = NULL;
5870 const char *stubName;
5871 assert(UseAES, "need AES instruction support");
5872
5873 switch(id) {
5874 case vmIntrinsics::_aescrypt_encryptBlock:
5875 stubAddr = StubRoutines::aescrypt_encryptBlock();
5876 stubName = "aescrypt_encryptBlock";
5877 break;
5878 case vmIntrinsics::_aescrypt_decryptBlock:
5879 stubAddr = StubRoutines::aescrypt_decryptBlock();
5880 stubName = "aescrypt_decryptBlock";
5881 break;
5882 default:
5883 break;
5884 }
5885 if (stubAddr == NULL) return false;
5886
5887 Node* aescrypt_object = argument(0);
5888 Node* src = argument(1);
5889 Node* src_offset = argument(2);
5890 Node* dest = argument(3);
5891 Node* dest_offset = argument(4);
5892
5893 src = must_be_not_null(src, true);
5894 dest = must_be_not_null(dest, true);
5895
5896 src = access_resolve(src, ACCESS_READ);
5897 dest = access_resolve(dest, ACCESS_WRITE);
5898
5899 // (1) src and dest are arrays.
5900 const Type* src_type = src->Value(&_gvn);
5901 const Type* dest_type = dest->Value(&_gvn);
5902 const TypeAryPtr* top_src = src_type->isa_aryptr();
5903 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5904 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5905
5906 // for the quick and dirty code we will skip all the checks.
5907 // we are just trying to get the call to be generated.
5908 Node* src_start = src;
5909 Node* dest_start = dest;
5910 if (src_offset != NULL || dest_offset != NULL) {
5911 assert(src_offset != NULL && dest_offset != NULL, "");
5912 src_start = array_element_address(src, src_offset, T_BYTE);
5913 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5914 }
5915
5916 // now need to get the start of its expanded key array
5917 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5918 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5919 if (k_start == NULL) return false;
5920
5921 if (Matcher::pass_original_key_for_aes()) {
5922 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5923 // compatibility issues between Java key expansion and SPARC crypto instructions
5924 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5925 if (original_k_start == NULL) return false;
5926
5927 // Call the stub.
5928 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5929 stubAddr, stubName, TypePtr::BOTTOM,
5930 src_start, dest_start, k_start, original_k_start);
5931 } else {
5932 // Call the stub.
5933 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5934 stubAddr, stubName, TypePtr::BOTTOM,
5935 src_start, dest_start, k_start);
5936 }
5937
5938 return true;
5939 }
5940
5941 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5942 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5943 address stubAddr = NULL;
5944 const char *stubName = NULL;
5945
5946 assert(UseAES, "need AES instruction support");
5947
5948 switch(id) {
5949 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5950 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5951 stubName = "cipherBlockChaining_encryptAESCrypt";
5952 break;
5953 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5954 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5955 stubName = "cipherBlockChaining_decryptAESCrypt";
5956 break;
5957 default:
5958 break;
5959 }
5960 if (stubAddr == NULL) return false;
5961
5962 Node* cipherBlockChaining_object = argument(0);
5963 Node* src = argument(1);
5964 Node* src_offset = argument(2);
5965 Node* len = argument(3);
5966 Node* dest = argument(4);
5967 Node* dest_offset = argument(5);
5968
5969 src = must_be_not_null(src, false);
5970 dest = must_be_not_null(dest, false);
5971
5972 src = access_resolve(src, ACCESS_READ);
5973 dest = access_resolve(dest, ACCESS_WRITE);
5974
5975 // (1) src and dest are arrays.
5976 const Type* src_type = src->Value(&_gvn);
5977 const Type* dest_type = dest->Value(&_gvn);
5978 const TypeAryPtr* top_src = src_type->isa_aryptr();
5979 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5980 assert (top_src != NULL && top_src->klass() != NULL
5981 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5982
5983 // checks are the responsibility of the caller
5984 Node* src_start = src;
5985 Node* dest_start = dest;
5986 if (src_offset != NULL || dest_offset != NULL) {
5987 assert(src_offset != NULL && dest_offset != NULL, "");
5988 src_start = array_element_address(src, src_offset, T_BYTE);
5989 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5990 }
5991
5992 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5993 // (because of the predicated logic executed earlier).
5994 // so we cast it here safely.
5995 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5996
5997 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5998 if (embeddedCipherObj == NULL) return false;
5999
6000 // cast it to what we know it will be at runtime
6001 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6002 assert(tinst != NULL, "CBC obj is null");
6003 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6004 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6005 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6006
6007 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6008 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6009 const TypeOopPtr* xtype = aklass->as_instance_type();
6010 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6011 aescrypt_object = _gvn.transform(aescrypt_object);
6012
6013 // we need to get the start of the aescrypt_object's expanded key array
6014 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6015 if (k_start == NULL) return false;
6016
6017 // similarly, get the start address of the r vector
6018 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6019 if (objRvec == NULL) return false;
6020 objRvec = access_resolve(objRvec, ACCESS_WRITE);
6021 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6022
6023 Node* cbcCrypt;
6024 if (Matcher::pass_original_key_for_aes()) {
6025 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6026 // compatibility issues between Java key expansion and SPARC crypto instructions
6027 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6028 if (original_k_start == NULL) return false;
6029
6030 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
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, original_k_start);
6035 } else {
6036 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6037 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6038 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6039 stubAddr, stubName, TypePtr::BOTTOM,
6040 src_start, dest_start, k_start, r_start, len);
6041 }
6042
6043 // return cipher length (int)
6044 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6045 set_result(retvalue);
6046 return true;
6047 }
6048
6049 //------------------------------inline_counterMode_AESCrypt-----------------------
6050 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6051 assert(UseAES, "need AES instruction support");
6052 if (!UseAESCTRIntrinsics) return false;
6053
6054 address stubAddr = NULL;
6055 const char *stubName = NULL;
6056 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6057 stubAddr = StubRoutines::counterMode_AESCrypt();
6058 stubName = "counterMode_AESCrypt";
6059 }
6060 if (stubAddr == NULL) return false;
6061
6062 Node* counterMode_object = argument(0);
6063 Node* src = argument(1);
6064 Node* src_offset = argument(2);
6065 Node* len = argument(3);
6066 Node* dest = argument(4);
6067 Node* dest_offset = argument(5);
6068
6069 src = access_resolve(src, ACCESS_READ);
6070 dest = access_resolve(dest, ACCESS_WRITE);
6071 counterMode_object = access_resolve(counterMode_object, ACCESS_WRITE);
6072
6073 // (1) src and dest are arrays.
6074 const Type* src_type = src->Value(&_gvn);
6075 const Type* dest_type = dest->Value(&_gvn);
6076 const TypeAryPtr* top_src = src_type->isa_aryptr();
6077 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6078 assert(top_src != NULL && top_src->klass() != NULL &&
6079 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6080
6081 // checks are the responsibility of the caller
6082 Node* src_start = src;
6083 Node* dest_start = dest;
6084 if (src_offset != NULL || dest_offset != NULL) {
6085 assert(src_offset != NULL && dest_offset != NULL, "");
6086 src_start = array_element_address(src, src_offset, T_BYTE);
6087 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6088 }
6089
6090 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6091 // (because of the predicated logic executed earlier).
6092 // so we cast it here safely.
6093 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6094 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6095 if (embeddedCipherObj == NULL) return false;
6096 // cast it to what we know it will be at runtime
6097 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6098 assert(tinst != NULL, "CTR obj is null");
6099 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6100 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6101 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6102 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6103 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6104 const TypeOopPtr* xtype = aklass->as_instance_type();
6105 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6106 aescrypt_object = _gvn.transform(aescrypt_object);
6107 // we need to get the start of the aescrypt_object's expanded key array
6108 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6109 if (k_start == NULL) return false;
6110 // similarly, get the start address of the r vector
6111 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6112 if (obj_counter == NULL) return false;
6113 obj_counter = access_resolve(obj_counter, ACCESS_WRITE);
6114 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6115
6116 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6117 if (saved_encCounter == NULL) return false;
6118 saved_encCounter = access_resolve(saved_encCounter, ACCESS_WRITE);
6119 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6120 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6121
6122 Node* ctrCrypt;
6123 if (Matcher::pass_original_key_for_aes()) {
6124 // no SPARC version for AES/CTR intrinsics now.
6125 return false;
6126 }
6127 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6128 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6129 OptoRuntime::counterMode_aescrypt_Type(),
6130 stubAddr, stubName, TypePtr::BOTTOM,
6131 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6132
6133 // return cipher length (int)
6134 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6135 set_result(retvalue);
6136 return true;
6137 }
6138
6139 //------------------------------get_key_start_from_aescrypt_object-----------------------
6140 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6141 #if defined(PPC64) || defined(S390)
6142 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6143 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6144 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6145 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6146 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6147 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6148 if (objSessionK == NULL) {
6149 return (Node *) NULL;
6150 }
6151 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6152 #else
6153 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6154 #endif // PPC64
6155 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6156 if (objAESCryptKey == NULL) return (Node *) NULL;
6157
6158 // now have the array, need to get the start address of the K array
6159 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ);
6160 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6161 return k_start;
6162 }
6163
6164 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6165 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6166 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6167 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6168 if (objAESCryptKey == NULL) return (Node *) NULL;
6169
6170 // now have the array, need to get the start address of the lastKey array
6171 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ);
6172 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6173 return original_k_start;
6174 }
6175
6176 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6177 // Return node representing slow path of predicate check.
6178 // the pseudo code we want to emulate with this predicate is:
6179 // for encryption:
6180 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6181 // for decryption:
6182 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6183 // note cipher==plain is more conservative than the original java code but that's OK
6184 //
6185 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6186 // The receiver was checked for NULL already.
6187 Node* objCBC = argument(0);
6188
6189 Node* src = argument(1);
6190 Node* dest = argument(4);
6191
6192 // Load embeddedCipher field of CipherBlockChaining object.
6193 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6194
6195 // get AESCrypt klass for instanceOf check
6196 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6197 // will have same classloader as CipherBlockChaining object
6198 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6199 assert(tinst != NULL, "CBCobj is null");
6200 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6201
6202 // we want to do an instanceof comparison against the AESCrypt class
6203 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6204 if (!klass_AESCrypt->is_loaded()) {
6205 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6206 Node* ctrl = control();
6207 set_control(top()); // no regular fast path
6208 return ctrl;
6209 }
6210
6211 src = must_be_not_null(src, true);
6212 dest = must_be_not_null(dest, true);
6213
6214 // Resolve oops to stable for CmpP below.
6215 src = access_resolve(src, 0);
6216 dest = access_resolve(dest, 0);
6217
6218 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6219
6220 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6221 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6222 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6223
6224 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6225
6226 // for encryption, we are done
6227 if (!decrypting)
6228 return instof_false; // even if it is NULL
6229
6230 // for decryption, we need to add a further check to avoid
6231 // taking the intrinsic path when cipher and plain are the same
6232 // see the original java code for why.
6233 RegionNode* region = new RegionNode(3);
6234 region->init_req(1, instof_false);
6235
6236 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6237 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6238 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6239 region->init_req(2, src_dest_conjoint);
6240
6241 record_for_igvn(region);
6242 return _gvn.transform(region);
6243 }
6244
6245 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6246 // Return node representing slow path of predicate check.
6247 // the pseudo code we want to emulate with this predicate is:
6248 // for encryption:
6249 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6250 // for decryption:
6251 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6252 // note cipher==plain is more conservative than the original java code but that's OK
6253 //
6254
6255 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6256 // The receiver was checked for NULL already.
6257 Node* objCTR = argument(0);
6258
6259 // Load embeddedCipher field of CipherBlockChaining object.
6260 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6261
6262 // get AESCrypt klass for instanceOf check
6263 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6264 // will have same classloader as CipherBlockChaining object
6265 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6266 assert(tinst != NULL, "CTRobj is null");
6267 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6268
6269 // we want to do an instanceof comparison against the AESCrypt class
6270 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6271 if (!klass_AESCrypt->is_loaded()) {
6272 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6273 Node* ctrl = control();
6274 set_control(top()); // no regular fast path
6275 return ctrl;
6276 }
6277
6278 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6279 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6280 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6281 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6282 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6283
6284 return instof_false; // even if it is NULL
6285 }
6286
6287 //------------------------------inline_ghash_processBlocks
6288 bool LibraryCallKit::inline_ghash_processBlocks() {
6289 address stubAddr;
6290 const char *stubName;
6291 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6292
6293 stubAddr = StubRoutines::ghash_processBlocks();
6294 stubName = "ghash_processBlocks";
6295
6296 Node* data = argument(0);
6297 Node* offset = argument(1);
6298 Node* len = argument(2);
6299 Node* state = argument(3);
6300 Node* subkeyH = argument(4);
6301
6302 state = must_be_not_null(state, true);
6303 subkeyH = must_be_not_null(subkeyH, true);
6304 data = must_be_not_null(data, true);
6305
6306 state = access_resolve(state, ACCESS_WRITE);
6307 subkeyH = access_resolve(subkeyH, ACCESS_READ);
6308 data = access_resolve(data, ACCESS_READ);
6309
6310 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6311 assert(state_start, "state is NULL");
6312 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6313 assert(subkeyH_start, "subkeyH is NULL");
6314 Node* data_start = array_element_address(data, offset, T_BYTE);
6315 assert(data_start, "data is NULL");
6316
6317 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6318 OptoRuntime::ghash_processBlocks_Type(),
6319 stubAddr, stubName, TypePtr::BOTTOM,
6320 state_start, subkeyH_start, data_start, len);
6321 return true;
6322 }
6323
6324 bool LibraryCallKit::inline_base64_encodeBlock() {
6325 address stubAddr;
6326 const char *stubName;
6327 assert(UseBASE64Intrinsics, "need Base64 intrinsics support");
6328 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters");
6329 stubAddr = StubRoutines::base64_encodeBlock();
6330 stubName = "encodeBlock";
6331
6332 if (!stubAddr) return false;
6333 Node* base64obj = argument(0);
6334 Node* src = argument(1);
6335 Node* offset = argument(2);
6336 Node* len = argument(3);
6337 Node* dest = argument(4);
6338 Node* dp = argument(5);
6339 Node* isURL = argument(6);
6340
6341 src = must_be_not_null(src, true);
6342 src = access_resolve(src, ACCESS_READ);
6343 dest = must_be_not_null(dest, true);
6344 dest = access_resolve(dest, ACCESS_WRITE);
6345
6346 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
6347 assert(src_start, "source array is NULL");
6348 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE);
6349 assert(dest_start, "destination array is NULL");
6350
6351 Node* base64 = make_runtime_call(RC_LEAF,
6352 OptoRuntime::base64_encodeBlock_Type(),
6353 stubAddr, stubName, TypePtr::BOTTOM,
6354 src_start, offset, len, dest_start, dp, isURL);
6355 return true;
6356 }
6357
6358 //------------------------------inline_sha_implCompress-----------------------
6359 //
6360 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6361 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6362 //
6363 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6364 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6365 //
6366 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6367 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6368 //
6369 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6370 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6371
6372 Node* sha_obj = argument(0);
6373 Node* src = argument(1); // type oop
6374 Node* ofs = argument(2); // type int
6375
6376 const Type* src_type = src->Value(&_gvn);
6377 const TypeAryPtr* top_src = src_type->isa_aryptr();
6378 if (top_src == NULL || top_src->klass() == NULL) {
6379 // failed array check
6380 return false;
6381 }
6382 // Figure out the size and type of the elements we will be copying.
6383 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6384 if (src_elem != T_BYTE) {
6385 return false;
6386 }
6387 // 'src_start' points to src array + offset
6388 src = must_be_not_null(src, true);
6389 src = access_resolve(src, ACCESS_READ);
6390 Node* src_start = array_element_address(src, ofs, src_elem);
6391 Node* state = NULL;
6392 address stubAddr;
6393 const char *stubName;
6394
6395 switch(id) {
6396 case vmIntrinsics::_sha_implCompress:
6397 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6398 state = get_state_from_sha_object(sha_obj);
6399 stubAddr = StubRoutines::sha1_implCompress();
6400 stubName = "sha1_implCompress";
6401 break;
6402 case vmIntrinsics::_sha2_implCompress:
6403 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6404 state = get_state_from_sha_object(sha_obj);
6405 stubAddr = StubRoutines::sha256_implCompress();
6406 stubName = "sha256_implCompress";
6407 break;
6408 case vmIntrinsics::_sha5_implCompress:
6409 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6410 state = get_state_from_sha5_object(sha_obj);
6411 stubAddr = StubRoutines::sha512_implCompress();
6412 stubName = "sha512_implCompress";
6413 break;
6414 default:
6415 fatal_unexpected_iid(id);
6416 return false;
6417 }
6418 if (state == NULL) return false;
6419
6420 assert(stubAddr != NULL, "Stub is generated");
6421 if (stubAddr == NULL) return false;
6422
6423 // Call the stub.
6424 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6425 stubAddr, stubName, TypePtr::BOTTOM,
6426 src_start, state);
6427
6428 return true;
6429 }
6430
6431 //------------------------------inline_digestBase_implCompressMB-----------------------
6432 //
6433 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6434 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6435 //
6436 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6437 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6438 "need SHA1/SHA256/SHA512 instruction support");
6439 assert((uint)predicate < 3, "sanity");
6440 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6441
6442 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6443 Node* src = argument(1); // byte[] array
6444 Node* ofs = argument(2); // type int
6445 Node* limit = argument(3); // type int
6446
6447 const Type* src_type = src->Value(&_gvn);
6448 const TypeAryPtr* top_src = src_type->isa_aryptr();
6449 if (top_src == NULL || top_src->klass() == NULL) {
6450 // failed array check
6451 return false;
6452 }
6453 // Figure out the size and type of the elements we will be copying.
6454 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6455 if (src_elem != T_BYTE) {
6456 return false;
6457 }
6458 // 'src_start' points to src array + offset
6459 src = must_be_not_null(src, false);
6460 src = access_resolve(src, ACCESS_READ);
6461 Node* src_start = array_element_address(src, ofs, src_elem);
6462
6463 const char* klass_SHA_name = NULL;
6464 const char* stub_name = NULL;
6465 address stub_addr = NULL;
6466 bool long_state = false;
6467
6468 switch (predicate) {
6469 case 0:
6470 if (UseSHA1Intrinsics) {
6471 klass_SHA_name = "sun/security/provider/SHA";
6472 stub_name = "sha1_implCompressMB";
6473 stub_addr = StubRoutines::sha1_implCompressMB();
6474 }
6475 break;
6476 case 1:
6477 if (UseSHA256Intrinsics) {
6478 klass_SHA_name = "sun/security/provider/SHA2";
6479 stub_name = "sha256_implCompressMB";
6480 stub_addr = StubRoutines::sha256_implCompressMB();
6481 }
6482 break;
6483 case 2:
6484 if (UseSHA512Intrinsics) {
6485 klass_SHA_name = "sun/security/provider/SHA5";
6486 stub_name = "sha512_implCompressMB";
6487 stub_addr = StubRoutines::sha512_implCompressMB();
6488 long_state = true;
6489 }
6490 break;
6491 default:
6492 fatal("unknown SHA intrinsic predicate: %d", predicate);
6493 }
6494 if (klass_SHA_name != NULL) {
6495 assert(stub_addr != NULL, "Stub is generated");
6496 if (stub_addr == NULL) return false;
6497
6498 // get DigestBase klass to lookup for SHA klass
6499 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6500 assert(tinst != NULL, "digestBase_obj is not instance???");
6501 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6502
6503 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6504 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6505 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6506 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6507 }
6508 return false;
6509 }
6510 //------------------------------inline_sha_implCompressMB-----------------------
6511 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6512 bool long_state, address stubAddr, const char *stubName,
6513 Node* src_start, Node* ofs, Node* limit) {
6514 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6515 const TypeOopPtr* xtype = aklass->as_instance_type();
6516 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6517 sha_obj = _gvn.transform(sha_obj);
6518
6519 Node* state;
6520 if (long_state) {
6521 state = get_state_from_sha5_object(sha_obj);
6522 } else {
6523 state = get_state_from_sha_object(sha_obj);
6524 }
6525 if (state == NULL) return false;
6526
6527 // Call the stub.
6528 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6529 OptoRuntime::digestBase_implCompressMB_Type(),
6530 stubAddr, stubName, TypePtr::BOTTOM,
6531 src_start, state, ofs, limit);
6532 // return ofs (int)
6533 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6534 set_result(result);
6535
6536 return true;
6537 }
6538
6539 enum VectorApiObjectType {
6540 VECAPI_VECTOR,
6541 VECAPI_MASK,
6542 VECAPI_SPECIES,
6543 VECAPI_SHUFFLE,
6544 };
6545
6546 enum VectorMaskUseType {
6547 VecMaskUseLoad,
6548 VecMaskUseStore,
6549 VecMaskUseAll,
6550 VecMaskNotUsed
6551 };
6552
6553 static bool arch_supports_vector(int op, int num_elem, BasicType type, VectorMaskUseType mask_use_type, int op_arity = 0) {
6554 // Check that the operation is valid.
6555 if (op <= 0) {
6556 #ifndef PRODUCT
6557 if (DebugVectorApi) {
6558 tty->print_cr("Rejected intrinsification because no valid vector op could be extracted");
6559 }
6560 #endif
6561 return false;
6562 }
6563
6564 // Check that architecture supports this op-size-type combination.
6565 if (!Matcher::match_rule_supported_vector(op, num_elem, type, op_arity)) {
6566 #ifndef PRODUCT
6567 if (DebugVectorApi) {
6568 tty->print_cr("Rejected vector op (%s,%s,%d) because architecture does not support it",
6569 NodeClassNames[op], type2name(type), num_elem);
6570 }
6571 #endif
6572 return false;
6573 } else {
6574 assert(Matcher::match_rule_supported(op), "must be supported");
6575 }
6576
6577 // Check whether mask unboxing is supported.
6578 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseLoad) {
6579 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type)) {
6580 #ifndef PRODUCT
6581 if (DebugVectorApi) {
6582 tty->print_cr("Rejected vector mask loading (%s,%s,%d) because architecture does not support it",
6583 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem);
6584 }
6585 #endif
6586 return false;
6587 }
6588 }
6589
6590 // Check whether mask boxing is supported.
6591 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseStore) {
6592 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type)) {
6593 #ifndef PRODUCT
6594 if (DebugVectorApi) {
6595 tty->print_cr("Rejected vector mask storing (%s,%s,%d) because architecture does not support it",
6596 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem);
6597 }
6598 #endif
6599 return false;
6600 }
6601 }
6602
6603 return true;
6604 }
6605
6606 // Should be aligned with constants in jdk.incubator.vector.VectorIntrinsics.
6607 enum {
6608 // Unary
6609 OP_ABS = 0,
6610 OP_NEG = 1,
6611 OP_SQRT = 2,
6612 OP_NOT = 3,
6613 // Binary
6614 OP_ADD = 4,
6615 OP_SUB = 5,
6616 OP_MUL = 6,
6617 OP_DIV = 7,
6618 OP_MIN = 8,
6619 OP_MAX = 9,
6620 OP_AND = 10,
6621 OP_OR = 11,
6622 OP_XOR = 12,
6623 // Ternary
6624 OP_FMA = 13,
6625 // Broadcast int
6626 OP_LSHIFT = 14,
6627 OP_RSHIFT = 15,
6628 OP_URSHIFT = 16,
6629 // Vector Math Library
6630 OP_TAN = 101,
6631 OP_SVML_START = OP_TAN,
6632 OP_TANH = 102,
6633 OP_SIN = 103,
6634 OP_SINH = 104,
6635 OP_COS = 105,
6636 OP_COSH = 106,
6637 OP_ASIN = 107,
6638 OP_ACOS = 108,
6639 OP_ATAN = 109,
6640 OP_ATAN2 = 110,
6641 OP_CBRT = 111,
6642 OP_LOG = 112,
6643 OP_LOG10 = 113,
6644 OP_LOG1P = 114,
6645 OP_POW = 115,
6646 OP_EXP = 116,
6647 OP_EXPM1 = 117,
6648 OP_HYPOT = 118,
6649 OP_SVML_END = OP_HYPOT,
6650 };
6651
6652 static int get_opc(jint op, BasicType bt) {
6653 switch (op) {
6654 case OP_ADD: {
6655 switch (bt) {
6656 case T_BYTE: // fall-through
6657 case T_SHORT: // fall-through
6658 case T_INT: return Op_AddI;
6659 case T_LONG: return Op_AddL;
6660 case T_FLOAT: return Op_AddF;
6661 case T_DOUBLE: return Op_AddD;
6662 default: fatal("ADD: %s", type2name(bt));
6663 }
6664 break;
6665 }
6666 case OP_SUB: {
6667 switch (bt) {
6668 case T_BYTE: // fall-through
6669 case T_SHORT: // fall-through
6670 case T_INT: return Op_SubI;
6671 case T_LONG: return Op_SubL;
6672 case T_FLOAT: return Op_SubF;
6673 case T_DOUBLE: return Op_SubD;
6674 default: fatal("SUB: %s", type2name(bt));
6675 }
6676 break;
6677 }
6678 case OP_MUL: {
6679 switch (bt) {
6680 case T_BYTE: // fall-through
6681 case T_SHORT: // fall-through
6682 case T_INT: return Op_MulI;
6683 case T_LONG: return Op_MulL;
6684 case T_FLOAT: return Op_MulF;
6685 case T_DOUBLE: return Op_MulD;
6686 default: fatal("MUL: %s", type2name(bt));
6687 }
6688 break;
6689 }
6690 case OP_DIV: {
6691 switch (bt) {
6692 case T_BYTE: // fall-through
6693 case T_SHORT: // fall-through
6694 case T_INT: return Op_DivI;
6695 case T_LONG: return Op_DivL;
6696 case T_FLOAT: return Op_DivF;
6697 case T_DOUBLE: return Op_DivD;
6698 default: fatal("DIV: %s", type2name(bt));
6699 }
6700 break;
6701 }
6702 case OP_MIN: {
6703 switch (bt) {
6704 case T_BYTE:
6705 case T_SHORT:
6706 case T_INT: return Op_MinI;
6707 case T_LONG: return Op_MinL;
6708 case T_FLOAT: return Op_MinF;
6709 case T_DOUBLE: return Op_MinD;
6710 default: fatal("MIN: %s", type2name(bt));
6711 }
6712 break;
6713 }
6714 case OP_MAX: {
6715 switch (bt) {
6716 case T_BYTE:
6717 case T_SHORT:
6718 case T_INT: return Op_MaxI;
6719 case T_LONG: return Op_MaxL;
6720 case T_FLOAT: return Op_MaxF;
6721 case T_DOUBLE: return Op_MaxD;
6722 default: fatal("MAX: %s", type2name(bt));
6723 }
6724 break;
6725 }
6726 case OP_ABS: {
6727 switch (bt) {
6728 case T_BYTE: // fall-through
6729 case T_SHORT: // fall-through
6730 case T_LONG: // fall-through
6731 case T_INT: return Op_AbsI;
6732 case T_FLOAT: return Op_AbsF;
6733 case T_DOUBLE: return Op_AbsD;
6734 default: fatal("ABS: %s", type2name(bt));
6735 }
6736 break;
6737 }
6738 case OP_NEG: {
6739 switch (bt) {
6740 case T_BYTE: // fall-through
6741 case T_SHORT: // fall-through
6742 case T_INT: return Op_NegI;
6743 case T_FLOAT: return Op_NegF;
6744 case T_DOUBLE: return Op_NegD;
6745 default: fatal("NEG: %s", type2name(bt));
6746 }
6747 break;
6748 }
6749 case OP_AND: {
6750 switch (bt) {
6751 case T_BYTE: // fall-through
6752 case T_SHORT: // fall-through
6753 case T_INT: return Op_AndI;
6754 case T_LONG: return Op_AndL;
6755 default: fatal("AND: %s", type2name(bt));
6756 }
6757 break;
6758 }
6759 case OP_OR: {
6760 switch (bt) {
6761 case T_BYTE: // fall-through
6762 case T_SHORT: // fall-through
6763 case T_INT: return Op_OrI;
6764 case T_LONG: return Op_OrL;
6765 default: fatal("OR: %s", type2name(bt));
6766 }
6767 break;
6768 }
6769 case OP_XOR: {
6770 switch (bt) {
6771 case T_BYTE: // fall-through
6772 case T_SHORT: // fall-through
6773 case T_INT: return Op_XorI;
6774 case T_LONG: return Op_XorL;
6775 default: fatal("XOR: %s", type2name(bt));
6776 }
6777 break;
6778 }
6779 case OP_SQRT: {
6780 switch (bt) {
6781 case T_FLOAT: return Op_SqrtF;
6782 case T_DOUBLE: return Op_SqrtD;
6783 default: fatal("SQRT: %s", type2name(bt));
6784 }
6785 break;
6786 }
6787 case OP_NOT: {
6788 switch (bt) {
6789 case T_BYTE: // fall-through
6790 case T_SHORT: // fall-through
6791 case T_INT: // fall-through
6792 case T_LONG: return Op_Not;
6793 default: fatal("NOT: %s", type2name(bt));
6794 }
6795 break;
6796 }
6797 case OP_FMA: {
6798 switch (bt) {
6799 case T_FLOAT: return Op_FmaF;
6800 case T_DOUBLE: return Op_FmaD;
6801 default: fatal("FMA: %s", type2name(bt));
6802 }
6803 break;
6804 }
6805 case OP_LSHIFT: {
6806 switch (bt) {
6807 case T_BYTE: // fall-through
6808 case T_SHORT: // fall-through
6809 case T_INT: return Op_LShiftI;
6810 case T_LONG: return Op_LShiftL;
6811 default: fatal("LSHIFT: %s", type2name(bt));
6812 }
6813 break;
6814 }
6815 case OP_RSHIFT: {
6816 switch (bt) {
6817 case T_BYTE: // fall-through
6818 case T_SHORT: // fall-through
6819 case T_INT: return Op_RShiftI;
6820 case T_LONG: return Op_RShiftL;
6821 default: fatal("RSHIFT: %s", type2name(bt));
6822 }
6823 break;
6824 }
6825 case OP_URSHIFT: {
6826 switch (bt) {
6827 case T_BYTE: return Op_URShiftB;
6828 case T_SHORT: return Op_URShiftS;
6829 case T_INT: return Op_URShiftI;
6830 case T_LONG: return Op_URShiftL;
6831 default: fatal("URSHIFT: %s", type2name(bt));
6832 }
6833 break;
6834 }
6835 case OP_TAN:
6836 case OP_TANH:
6837 case OP_SIN:
6838 case OP_SINH:
6839 case OP_COS:
6840 case OP_COSH:
6841 case OP_ASIN:
6842 case OP_ACOS:
6843 case OP_ATAN:
6844 case OP_ATAN2:
6845 case OP_CBRT:
6846 case OP_LOG:
6847 case OP_LOG10:
6848 case OP_LOG1P:
6849 case OP_POW:
6850 case OP_EXP:
6851 case OP_EXPM1:
6852 case OP_HYPOT:
6853 return Op_CallLeafVector;
6854 default: fatal("unknown op: %d", op);
6855 }
6856 return 0; // Unimplemented
6857 }
6858
6859 static int get_sopc(int opc, BasicType elem_bt) {
6860 int sopc = 0;
6861 switch (opc) {
6862 case Op_CallLeafVector:
6863 sopc = Op_CallLeafVector;
6864 break;
6865 default:
6866 sopc = VectorNode::opcode(opc, elem_bt);
6867 break;
6868 }
6869 return sopc;
6870 }
6871 Node* LibraryCallKit::box_vector(Node* vector, const TypeInstPtr* vbox_type,
6872 BasicType elem_bt, int num_elem) {
6873
6874 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
6875
6876 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
6877 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
6878 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true);
6879 set_i_o(_gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
6880 set_all_memory(_gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
6881 Node* ret = _gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
6882
6883 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vec_type);
6884 return _gvn.transform(vbox);
6885 }
6886
6887 Node* LibraryCallKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) {
6888 const TypeInstPtr* vbox_type_v = gvn().type(v)->is_instptr();
6889 if (vbox_type->klass() != vbox_type_v->klass()) {
6890 return NULL; // arguments don't agree on vector shapes
6891 }
6892 if (vbox_type_v->maybe_null()) {
6893 return NULL; // no nulls are allowed
6894 }
6895 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem);
6896 Node* unbox = gvn().transform(new VectorUnboxNode(C, vec_type, v, merged_memory()));
6897 return unbox;
6898 }
6899
6900 void LibraryCallKit::set_vector_result(Node* result, bool set_res) {
6901 if (DebugVectorApi) {
6902 #ifndef PRODUCT
6903 tty->print("============ ");
6904 callee()->print();
6905 tty->print_cr(" ============");
6906 result->dump(5);
6907 tty->print_cr("----------------------------------------------------");
6908 #endif
6909 }
6910 if (set_res) {
6911 set_result(result);
6912 }
6913 }
6914
6915 bool LibraryCallKit::inline_vector_nary_operation(int n) {
6916 const TypeInt* opr = gvn().type(argument(0))->is_int();
6917 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
6918 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
6919 const TypeInt* vlen = gvn().type(argument(3))->is_int();
6920
6921 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
6922 return false; // not enough info for intrinsification
6923 }
6924 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
6925 if (!elem_type->is_primitive_type()) {
6926 return false; // should be primitive type
6927 }
6928 BasicType elem_bt = elem_type->basic_type();
6929 int num_elem = vlen->get_con();
6930 int opc = get_opc(opr->get_con(), elem_bt);
6931 int sopc = get_sopc(opc, elem_bt);
6932 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
6933 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
6934
6935 if (opc == Op_CallLeafVector && !Matcher::supports_vector_calling_convention()) {
6936 return false;
6937 }
6938
6939 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
6940 if (!arch_supports_vector(sopc, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseAll : VecMaskNotUsed, n)) {
6941 return false; // not supported
6942 }
6943
6944 Node* opd1 = NULL; Node* opd2 = NULL; Node* opd3 = NULL;
6945 switch (n) {
6946 case 3: {
6947 opd3 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
6948 if (opd3 == NULL) {
6949 return false;
6950 }
6951 // fall-through
6952 }
6953 case 2: {
6954 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
6955 if (opd2 == NULL) {
6956 return false;
6957 }
6958 // fall-through
6959 }
6960 case 1: {
6961 opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
6962 if (opd1 == NULL) {
6963 return false;
6964 }
6965 break;
6966 }
6967 default: fatal("unsupported arity: %d", n);
6968 }
6969
6970 Node* operation = NULL;
6971 if (sopc == Op_CallLeafVector) {
6972 operation = gen_call_to_svml(opr->get_con(), elem_bt, num_elem, opd1, opd2);
6973 if (operation == NULL) {
6974 return false;
6975 }
6976 } else {
6977 switch (n) {
6978 case 1:
6979 case 2: {
6980 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt));
6981 break;
6982 }
6983 case 3: {
6984 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, opd3, num_elem, elem_bt));
6985 break;
6986 }
6987 default: fatal("unsupported arity: %d", n);
6988 }
6989 }
6990 // Wrap it up in VectorBox to keep object type information.
6991 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
6992 set_vector_result(operation);
6993
6994 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
6995 return true;
6996 }
6997
6998 // <Sh extends VectorShuffle<E>, E>
6999 // Sh ShuffleIota(Class<?> E, Class<?> ShuffleClass, Vector.Species<E> s, int length,
7000 // int step, ShuffleIotaOperation<Sh, E> defaultImpl)
7001 bool LibraryCallKit::inline_vector_shuffle_iota() {
7002 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->is_instptr();
7003 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7004 Node* step = argument(4);
7005
7006 if (!vlen->is_con() || shuffle_klass->const_oop() == NULL) {
7007 return false; // not enough info for intrinsification
7008 }
7009
7010 int num_elem = vlen->get_con();
7011 BasicType elem_bt = T_BYTE;
7012
7013 if (num_elem < 4)
7014 return false;
7015
7016 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt, VecMaskNotUsed)) {
7017 return false;
7018 }
7019 if (!arch_supports_vector(Op_AddVB, num_elem, elem_bt, VecMaskNotUsed)) {
7020 return false;
7021 }
7022 if (!arch_supports_vector(Op_AndV, num_elem, elem_bt, VecMaskNotUsed)) {
7023 return false;
7024 }
7025
7026 const TypeVect * vt = TypeVect::make(Type::get_const_basic_type(elem_bt), num_elem);
7027
7028 Node* iota = _gvn.transform(new VectorLoadConstNode(gvn().makecon(TypeInt::ZERO), vt));
7029 Node* bcast_step = _gvn.transform(VectorNode::scalar2vector(step, num_elem, Type::get_const_basic_type(elem_bt)));
7030
7031 Node* bcast_mod = _gvn.transform(VectorNode::scalar2vector(gvn().makecon(TypeInt::make(num_elem-1)),
7032 num_elem, Type::get_const_basic_type(elem_bt)));
7033 Node* add = _gvn.transform(VectorNode::make(Op_AddI, iota, bcast_step, num_elem, elem_bt));
7034 Node* res = _gvn.transform(VectorNode::make(Op_AndI, add, bcast_mod, num_elem, elem_bt));
7035
7036 ciKlass* sbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
7037 const TypeInstPtr* shuffle_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, sbox_klass);
7038
7039 // Wrap it up in VectorBox to keep object type information.
7040 res = box_vector(res, shuffle_box_type, elem_bt, num_elem);
7041
7042 set_vector_result(res);
7043 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7044 return true;
7045 }
7046
7047 // <VM ,Sh extends VectorShuffle<E>, E>
7048 // VM shuffleToVector(Class<VM> VecClass, Class<?>E , Class<?> ShuffleClass, Sh s, int length,
7049 // ShuffleToVectorOperation<VM,Sh,E> defaultImpl)
7050 bool LibraryCallKit::inline_vector_shuffle_to_vector() {
7051 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7052 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7053 const TypeInstPtr* shuffle_klass = gvn().type(argument(2))->is_instptr();
7054 Node* shuffle = argument(3);
7055 const TypeInt* vlen = gvn().type(argument(4))->is_int();
7056
7057 if (!vlen->is_con() || shuffle_klass->const_oop() == NULL) {
7058 return false; // not enough info for intrinsification
7059 }
7060
7061 int num_elem = vlen->get_con();
7062 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7063 BasicType elem_bt = elem_type->basic_type();
7064
7065 if (num_elem < 4) {
7066 return false;
7067 }
7068
7069 if (!arch_supports_vector(Op_VectorLoadShuffle, num_elem, elem_bt, VecMaskNotUsed)) {
7070 return false; // not supported
7071 }
7072
7073 ciKlass* sbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
7074 const TypeInstPtr* shuffle_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, sbox_klass);
7075
7076 // Unbox shuffle
7077 Node* shuffle_vec = unbox_vector(shuffle, shuffle_box_type, elem_bt, num_elem);
7078
7079 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7080 const TypeInstPtr* vec_box_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7081
7082 // Box vector
7083 Node* res = box_vector(shuffle_vec, vec_box_type, elem_bt, num_elem);
7084 set_vector_result(res);
7085 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7086 return true;
7087 }
7088
7089 // <V extends Vector<?,?>>
7090 // V broadcastCoerced(Class<?> vectorClass, Class<?> elementType, int vlen,
7091 // long bits,
7092 // LongFunction<V> defaultImpl)
7093 bool LibraryCallKit::inline_vector_broadcast_coerced() {
7094 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7095 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7096 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7097
7098 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7099 return false; // not enough info for intrinsification
7100 }
7101
7102 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7103 if (!elem_type->is_primitive_type()) {
7104 return false; // should be primitive type
7105 }
7106 BasicType elem_bt = elem_type->basic_type();
7107 int num_elem = vlen->get_con();
7108 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7109 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7110
7111 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7112 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt,
7113 vbox_klass->is_vectormask() ? VecMaskUseStore : VecMaskNotUsed)) {
7114 return false; // not supported
7115 }
7116
7117 Node* bits = argument(3); // long
7118
7119 Node* elem = NULL;
7120 switch (elem_bt) {
7121 case T_BOOLEAN: // fall-through
7122 case T_BYTE: // fall-through
7123 case T_SHORT: // fall-through
7124 case T_CHAR: // fall-through
7125 case T_INT: {
7126 elem = gvn().transform(new ConvL2INode(bits));
7127 break;
7128 }
7129 case T_DOUBLE: {
7130 elem = gvn().transform(new MoveL2DNode(bits));
7131 break;
7132 }
7133 case T_FLOAT: {
7134 bits = gvn().transform(new ConvL2INode(bits));
7135 elem = gvn().transform(new MoveI2FNode(bits));
7136 break;
7137 }
7138 case T_LONG: {
7139 elem = bits; // no conversion needed
7140 break;
7141 }
7142 default: fatal("%s", type2name(elem_bt));
7143 }
7144
7145 Node* broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt));
7146 broadcast = gvn().transform(broadcast);
7147 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem);
7148 set_vector_result(box);
7149
7150 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7151 return true;
7152 }
7153
7154 // <C, V extends Vector<?,?>>
7155 // V load(Class<?> vectorClass, Class<?> elementType, int vlen,
7156 // Object base, long offset,
7157 // /* Vector.Mask<E,S> m*/
7158 // Object container, int index,
7159 // LoadOperation<C, VM> defaultImpl) {
7160 //
7161 // <C, V extends Vector<?,?>>
7162 // void store(Class<?> vectorClass, Class<?> elementType, int vlen,
7163 // Object base, long offset,
7164 // V v, /*Vector.Mask<E,S> m*/
7165 // Object container, int index,
7166 // StoreVectorOperation<C, V> defaultImpl) {
7167
7168 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) {
7169 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7170 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7171 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7172
7173 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7174 return false; // not enough info for intrinsification
7175 }
7176
7177 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7178 if (!elem_type->is_primitive_type()) {
7179 return false; // should be primitive type
7180 }
7181 BasicType elem_bt = elem_type->basic_type();
7182 int num_elem = vlen->get_con();
7183
7184 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7185 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) {
7186 return false; // not supported
7187 }
7188
7189 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7190 bool is_mask = vbox_klass->is_vectormask();
7191
7192 Node* base = argument(3);
7193 Node* offset = ConvL2X(argument(4));
7194 DecoratorSet decorators = C2_UNSAFE_ACCESS;
7195 Node* addr = make_unsafe_address(base, offset, decorators, (is_mask ? T_BOOLEAN : elem_bt), true);
7196
7197 // Can base be NULL? Otherwise, always on-heap access.
7198 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base));
7199
7200 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
7201 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
7202
7203 // Now handle special case where load/store happens from/to byte array but element type is not byte.
7204 bool using_byte_array = arr_type != NULL && arr_type->elem()->array_element_basic_type() == T_BYTE && elem_bt != T_BYTE;
7205 // Handle loading masks.
7206 // If there is no consistency between array and vector element types, it must be special byte array case or loading masks
7207 if (arr_type != NULL && !using_byte_array && elem_bt != arr_type->elem()->array_element_basic_type() && !is_mask) {
7208 return false;
7209 }
7210 // Since we are using byte array, we need to double check that the byte operations are supported by backend.
7211 if (using_byte_array) {
7212 int byte_num_elem = num_elem * type2aelembytes(elem_bt);
7213 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, byte_num_elem, T_BYTE, VecMaskNotUsed)) {
7214 return false; // not supported
7215 }
7216 }
7217 if (is_mask) {
7218 if (!arch_supports_vector(Op_LoadVector, num_elem, T_BOOLEAN, VecMaskNotUsed)) {
7219 return false; // not supported
7220 }
7221 }
7222
7223 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7224
7225 if (can_access_non_heap) {
7226 insert_mem_bar(Op_MemBarCPUOrder);
7227 }
7228
7229 if (is_store) {
7230 Node* val = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7231 if (val == NULL) {
7232 return false; // operand unboxing failed
7233 }
7234 set_all_memory(reset_memory());
7235
7236 // In case the store needs to happen to byte array, reinterpret the incoming vector to byte vector.
7237 int store_num_elem = num_elem;
7238 if (using_byte_array) {
7239 store_num_elem = num_elem * type2aelembytes(elem_bt);
7240 const TypeVect* to_vect_type = TypeVect::make(T_BYTE, store_num_elem);
7241 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type));
7242 }
7243
7244 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(addr), addr, addr_type, val, store_num_elem));
7245 set_memory(vstore, addr_type);
7246 set_vector_result(vstore, false);
7247 } else {
7248 // When using byte array, we need to load as byte then reinterpret the value. Otherwise, do a simple vector load.
7249 Node* vload = NULL;
7250 if (using_byte_array) {
7251 int load_num_elem = num_elem * type2aelembytes(elem_bt);
7252 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, load_num_elem, T_BYTE));
7253 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
7254 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type));
7255 } else {
7256 // Special handle for masks
7257 if (is_mask) {
7258 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, T_BOOLEAN));
7259 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem);
7260 vload = gvn().transform(new VectorLoadMaskNode(vload, to_vect_type));
7261 } else {
7262 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, elem_bt));
7263 }
7264 }
7265 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7266 set_vector_result(box);
7267 }
7268
7269 if (can_access_non_heap) {
7270 insert_mem_bar(Op_MemBarCPUOrder);
7271 }
7272
7273 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7274 return true;
7275 }
7276
7277 // <C, V extends Vector<?>, W extends IntVector>
7278 // void storeWithMap(Class<?> vectorClass, Class<?> elementType, int length, Class<?> vectorIndexClass,
7279 // Object base, long offset, // Unsafe addressing
7280 // W index_vector, V v,
7281 // C container, int index, int[] indexMap, int indexM, // Arguments for default implementation
7282 // StoreVectorOperationWithMap<C, V> defaultImpl) {
7283 //
7284 bool LibraryCallKit::inline_vector_gather_scatter(bool is_scatter) {
7285 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7286 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
7287 const TypeInt* vlen = gvn().type(argument(2))->is_int();
7288 const TypeInstPtr* vector_idx_klass = gvn().type(argument(3))->is_instptr();
7289
7290 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || vector_idx_klass->const_oop() == NULL || !vlen->is_con()) {
7291 return false; // not enough info for intrinsification
7292 }
7293
7294 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7295 if (!elem_type->is_primitive_type()) {
7296 return false; // should be primitive type
7297 }
7298 BasicType elem_bt = elem_type->basic_type();
7299 int num_elem = vlen->get_con();
7300
7301 if (!arch_supports_vector(is_scatter ? Op_StoreVectorScatter : Op_LoadVectorGather, num_elem, elem_bt, VecMaskNotUsed)) {
7302 return false; // not supported
7303 }
7304
7305 // Check that the vector holding indices is supported by architecture
7306 if (!arch_supports_vector(Op_LoadVector, num_elem, T_INT, VecMaskNotUsed)) {
7307 return false; // not supported
7308 }
7309
7310 Node* base = argument(4);
7311 Node* offset = ConvL2X(argument(5));
7312 Node* addr = make_unsafe_address(base, offset, C2_UNSAFE_ACCESS, elem_bt, true);
7313
7314 const TypePtr *addr_type = gvn().type(addr)->isa_ptr();
7315 const TypeAryPtr* arr_type = addr_type->isa_aryptr();
7316
7317 // The array must be consistent with vector type
7318 if (arr_type == NULL || (arr_type != NULL && elem_bt != arr_type->elem()->array_element_basic_type())) {
7319 return false;
7320 }
7321 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7322 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7323
7324 ciKlass* vbox_idx_klass = vector_idx_klass->const_oop()->as_instance()->java_lang_Class_klass();
7325
7326 if (vbox_idx_klass == NULL) {
7327 return false;
7328 }
7329
7330 const TypeInstPtr* vbox_idx_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_idx_klass);
7331
7332 Node* index_vect = unbox_vector(argument(7), vbox_idx_type, T_INT, num_elem);
7333 if (index_vect == NULL) {
7334 return false;
7335 }
7336 const TypeVect* vector_type = TypeVect::make(elem_bt, num_elem);
7337 if (is_scatter) {
7338 Node* val = unbox_vector(argument(8), vbox_type, elem_bt, num_elem);
7339 if (val == NULL) {
7340 return false; // operand unboxing failed
7341 }
7342 set_all_memory(reset_memory());
7343
7344 Node* vstore = gvn().transform(new StoreVectorScatterNode(control(), memory(addr), addr, addr_type, val, index_vect));
7345 set_memory(vstore, addr_type);
7346 set_vector_result(vstore, false);
7347 } else {
7348 Node* vload = gvn().transform(new LoadVectorGatherNode(control(), memory(addr), addr, addr_type, vector_type, index_vect));
7349
7350 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem);
7351 set_vector_result(box);
7352 }
7353
7354 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7355 return true;
7356 }
7357
7358 // <V extends Vector<?,?>>
7359 // long reductionCoerced(int oprId, Class<?> vectorClass, Class<?> elementType, int vlen,
7360 // V v,
7361 // Function<V,Long> defaultImpl)
7362
7363 bool LibraryCallKit::inline_vector_reduction() {
7364 const TypeInt* opr = gvn().type(argument(0))->is_int();
7365 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7366 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7367 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7368
7369 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7370 return false; // not enough info for intrinsification
7371 }
7372 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7373 if (!elem_type->is_primitive_type()) {
7374 return false; // should be primitive type
7375 }
7376 BasicType elem_bt = elem_type->basic_type();
7377 int num_elem = vlen->get_con();
7378
7379 int opc = get_opc(opr->get_con(), elem_bt);
7380 int sopc = ReductionNode::opcode(opc, elem_bt);
7381
7382 // FIXME: When encountering a SubReduction, we want to check for support of
7383 // the corresponding AddReduction node.
7384 if (sopc == Op_SubReductionV) {
7385 if (gvn().type(argument(2))->isa_int()) {
7386 sopc = Op_AddReductionVI;
7387 } else if (gvn().type(argument(2))->isa_long()) {
7388 sopc = Op_AddReductionVL;
7389 }
7390 }
7391
7392 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad.
7393 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
7394 return false;
7395 }
7396
7397 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7398 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7399
7400 Node* opd = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7401 if (opd == NULL) {
7402 return false; // operand unboxing failed
7403 }
7404
7405 Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt);
7406 Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, elem_bt));
7407
7408 Node* bits = NULL;
7409 switch (elem_bt) {
7410 case T_BYTE:
7411 case T_SHORT:
7412 case T_INT: {
7413 bits = gvn().transform(new ConvI2LNode(rn));
7414 break;
7415 }
7416 case T_FLOAT: {
7417 rn = gvn().transform(new MoveF2INode(rn));
7418 bits = gvn().transform(new ConvI2LNode(rn));
7419 break;
7420 }
7421 case T_DOUBLE: {
7422 bits = gvn().transform(new MoveD2LNode(rn));
7423 break;
7424 }
7425 case T_LONG: {
7426 bits = rn; // no conversion needed
7427 break;
7428 }
7429 default: fatal("%s", type2name(elem_bt));
7430 }
7431 set_vector_result(bits);
7432 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7433 return true;
7434 }
7435
7436 // static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen,
7437 // V v1, V v2,
7438 // BiFunction<V, V, Boolean> defaultImpl) {
7439
7440 bool LibraryCallKit::inline_vector_test() {
7441 const TypeInt* cond = gvn().type(argument(0))->is_int();
7442 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7443 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7444 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7445
7446 if (!cond->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
7447 return false; // not enough info for intrinsification
7448 }
7449 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7450 if (!elem_type->is_primitive_type()) {
7451 return false; // should be primitive type
7452 }
7453 BasicType elem_bt = elem_type->basic_type();
7454 int num_elem = vlen->get_con();
7455 BoolTest::mask booltest = (BoolTest::mask)cond->get_con();
7456 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7457 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7458
7459 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseLoad : VecMaskNotUsed)) {
7460 return false;
7461 }
7462
7463 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7464 Node* opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7465 if (opd1 == NULL || opd2 == NULL) {
7466 return false; // operand unboxing failed
7467 }
7468 Node* test = new VectorTestNode(opd1, opd2, booltest);
7469 test = _gvn.transform(test);
7470 set_vector_result(test);
7471 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7472 return true;
7473 }
7474
7475 // static
7476 // <V extends Vector, M extends Mask>
7477 // V blend(Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7478 // V v1, V v2, M m,
7479 // VectorBlendOp<V,M> defaultImpl) { ...
7480 //
7481 bool LibraryCallKit::inline_vector_blend() {
7482 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7483 const TypeInstPtr* mask_klass = gvn().type(argument(1))->is_instptr();
7484 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7485 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7486
7487 if (mask_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7488 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7489 return false; // not enough info for intrinsification
7490 }
7491 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7492 if (!elem_type->is_primitive_type()) {
7493 return false; // should be primitive type
7494 }
7495 BasicType elem_bt = elem_type->basic_type();
7496 BasicType mask_bt = elem_bt;
7497 int num_elem = vlen->get_con();
7498
7499 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) {
7500 return false; // not supported
7501 }
7502 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7503 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7504
7505 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7506 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7507
7508 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7509 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7510 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem);
7511
7512 if (v1 == NULL || v2 == NULL || mask == NULL) {
7513 return false; // operand unboxing failed
7514 }
7515
7516 Node* blend = _gvn.transform(new VectorBlendNode(v1, v2, mask));
7517 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem);
7518 set_vector_result(box);
7519
7520 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7521 return true;
7522 }
7523
7524 // static <V extends Vector<E,S>,
7525 // M extends Vector.Mask<E,S>,
7526 // S extends Vector.Shape, E>
7527 // M compare(int cond, Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen,
7528 // V v1, V v2,
7529 // VectorCompareOp<V,M> defaultImpl) { ...
7530 //
7531 bool LibraryCallKit::inline_vector_compare() {
7532 const TypeInt* cond = gvn().type(argument(0))->is_int();
7533 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
7534 const TypeInstPtr* mask_klass = gvn().type(argument(2))->is_instptr();
7535 const TypeInstPtr* elem_klass = gvn().type(argument(3))->is_instptr();
7536 const TypeInt* vlen = gvn().type(argument(4))->is_int();
7537
7538 if (!cond->is_con() || vector_klass->const_oop() == NULL || mask_klass->const_oop() == NULL ||
7539 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7540 return false; // not enough info for intrinsification
7541 }
7542 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7543 if (!elem_type->is_primitive_type()) {
7544 return false; // should be primitive type
7545 }
7546
7547 int num_elem = vlen->get_con();
7548 BasicType elem_bt = elem_type->basic_type();
7549 BasicType mask_bt = elem_bt;
7550
7551 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) {
7552 return false;
7553 }
7554
7555 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7556 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7557
7558 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass();
7559 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass);
7560
7561 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem);
7562 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem);
7563
7564 if (v1 == NULL || v2 == NULL) {
7565 return false; // operand unboxing failed
7566 }
7567 BoolTest::mask pred = (BoolTest::mask)cond->get_con();
7568 const TypeVect* vt = TypeVect::make(mask_bt, num_elem);
7569 Node* operation = _gvn.transform(new VectorMaskCmpNode(pred, v1, v2, vt));
7570
7571 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem);
7572 set_vector_result(box);
7573
7574 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7575 return true;
7576 }
7577
7578 // static
7579 // <V extends Vector, Sh extends Shuffle>
7580 // V rearrangeOp(Class<V> vectorClass, Class<Sh> shuffleClass, Class< ? > elementType, int vlen,
7581 // V v1, Sh sh,
7582 // VectorSwizzleOp<V, Sh, S, E> defaultImpl) { ...
7583
7584 bool LibraryCallKit::inline_vector_rearrange() {
7585 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
7586 const TypeInstPtr* shuffle_klass = gvn().type(argument(1))->is_instptr();
7587 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
7588 const TypeInt* vlen = gvn().type(argument(3))->is_int();
7589
7590 if (shuffle_klass->const_oop() == NULL || vector_klass->const_oop() == NULL ||
7591 elem_klass->const_oop() == NULL || !vlen->is_con()) {
7592 return false; // not enough info for intrinsification
7593 }
7594 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
7595 if (!elem_type->is_primitive_type()) {
7596 return false; // should be primitive type
7597 }
7598 BasicType elem_bt = elem_type->basic_type();
7599 BasicType shuffle_bt = elem_bt;
7600 int num_elem = vlen->get_con();
7601
7602 if (!arch_supports_vector(Op_VectorLoadShuffle, num_elem, elem_bt, VecMaskNotUsed)) {
7603 return false; // not supported
7604 }
7605 if (!arch_supports_vector(Op_VectorRearrange, num_elem, elem_bt, VecMaskNotUsed)) {
7606 return false; // not supported
7607 }
7608 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
7609 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
7610
7611 ciKlass* shbox_klass = shuffle_klass->const_oop()->as_instance()->java_lang_Class_klass();
7612 const TypeInstPtr* shbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, shbox_klass);
7613
7614 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
7615 Node* shuffle = unbox_vector(argument(5), shbox_type, shuffle_bt, num_elem);
7616
7617 if (v1 == NULL || shuffle == NULL) {
7618 return false; // operand unboxing failed
7619 }
7620
7621 Node* rearrange = _gvn.transform(new VectorRearrangeNode(v1, shuffle));
7622 Node* box = box_vector(rearrange, vbox_type, elem_bt, num_elem);
7623 set_result(box);
7624
7625 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
7626 return true;
7627 }
7628
7629 Node* LibraryCallKit::shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) {
7630 assert(bt == T_INT || bt == T_LONG || bt == T_SHORT || bt == T_BYTE, "byte, short, long and int are supported");
7631 juint mask = (type2aelembytes(bt) * BitsPerByte - 1);
7632 Node* nmask = _gvn.transform(ConNode::make(TypeInt::make(mask)));
7633 Node* mcnt = _gvn.transform(new AndINode(cnt, nmask));
7634 return _gvn.transform(VectorNode::shift_count(shift_op, mcnt, num_elem, bt));
7635 }
7636
7637 static void get_svml_address(int op, int bits, BasicType bt, const char** name_ptr, address* addr_ptr) {
7638 assert(name_ptr != NULL, "unexpected");
7639 assert(addr_ptr != NULL, "unexpected");
7640
7641 // Since the addresses are resolved at runtime, using switch instead of table - otherwise might get NULL addresses.
7642 if (bt == T_FLOAT) {
7643 switch(op) {
7644 case OP_EXP: {
7645 switch(bits) {
7646 case 64: *name_ptr = "vector_float64_exp"; *addr_ptr = StubRoutines::vector_float64_exp(); break;
7647 case 128: *name_ptr = "vector_float128_exp"; *addr_ptr = StubRoutines::vector_float128_exp(); break;
7648 case 256: *name_ptr = "vector_float256_exp"; *addr_ptr = StubRoutines::vector_float256_exp(); break;
7649 case 512: *name_ptr = "vector_float512_exp"; *addr_ptr = StubRoutines::vector_float512_exp(); break;
7650 default: Unimplemented(); break;
7651 }
7652 }
7653 break;
7654 case OP_LOG1P: {
7655 switch(bits) {
7656 case 64: *name_ptr = "vector_float64_log1p"; *addr_ptr = StubRoutines::vector_float64_log1p(); break;
7657 case 128: *name_ptr = "vector_float128_log1p"; *addr_ptr = StubRoutines::vector_float128_log1p(); break;
7658 case 256: *name_ptr = "vector_float256_log1p"; *addr_ptr = StubRoutines::vector_float256_log1p(); break;
7659 case 512: *name_ptr = "vector_float512_log1p"; *addr_ptr = StubRoutines::vector_float512_log1p(); break;
7660 default: Unimplemented(); break;
7661 }
7662 }
7663 break;
7664 case OP_LOG: {
7665 switch(bits) {
7666 case 64: *name_ptr = "vector_float64_log"; *addr_ptr = StubRoutines::vector_float64_log(); break;
7667 case 128: *name_ptr = "vector_float128_log"; *addr_ptr = StubRoutines::vector_float128_log(); break;
7668 case 256: *name_ptr = "vector_float256_log"; *addr_ptr = StubRoutines::vector_float256_log(); break;
7669 case 512: *name_ptr = "vector_float512_log"; *addr_ptr = StubRoutines::vector_float512_log(); break;
7670 default: Unimplemented(); break;
7671 }
7672 }
7673 break;
7674 case OP_LOG10: {
7675 switch(bits) {
7676 case 64: *name_ptr = "vector_float64_log10"; *addr_ptr = StubRoutines::vector_float64_log10(); break;
7677 case 128: *name_ptr = "vector_float128_log10"; *addr_ptr = StubRoutines::vector_float128_log10(); break;
7678 case 256: *name_ptr = "vector_float256_log10"; *addr_ptr = StubRoutines::vector_float256_log10(); break;
7679 case 512: *name_ptr = "vector_float512_log10"; *addr_ptr = StubRoutines::vector_float512_log10(); break;
7680 default: Unimplemented(); break;
7681 }
7682 }
7683 break;
7684 case OP_EXPM1: {
7685 switch(bits) {
7686 case 64: *name_ptr = "vector_float64_expm1"; *addr_ptr = StubRoutines::vector_float64_expm1(); break;
7687 case 128: *name_ptr = "vector_float128_expm1"; *addr_ptr = StubRoutines::vector_float128_expm1(); break;
7688 case 256: *name_ptr = "vector_float256_expm1"; *addr_ptr = StubRoutines::vector_float256_expm1(); break;
7689 case 512: *name_ptr = "vector_float512_expm1"; *addr_ptr = StubRoutines::vector_float512_expm1(); break;
7690 default: Unimplemented(); break;
7691 }
7692 }
7693 break;
7694 case OP_SIN: {
7695 switch(bits) {
7696 case 64: *name_ptr = "vector_float64_sin"; *addr_ptr = StubRoutines::vector_float64_sin(); break;
7697 case 128: *name_ptr = "vector_float128_sin"; *addr_ptr = StubRoutines::vector_float128_sin(); break;
7698 case 256: *name_ptr = "vector_float256_sin"; *addr_ptr = StubRoutines::vector_float256_sin(); break;
7699 case 512: *name_ptr = "vector_float512_sin"; *addr_ptr = StubRoutines::vector_float512_sin(); break;
7700 default: Unimplemented(); break;
7701 }
7702 }
7703 break;
7704 case OP_COS: {
7705 switch(bits) {
7706 case 64: *name_ptr = "vector_float64_cos"; *addr_ptr = StubRoutines::vector_float64_cos(); break;
7707 case 128: *name_ptr = "vector_float128_cos"; *addr_ptr = StubRoutines::vector_float128_cos(); break;
7708 case 256: *name_ptr = "vector_float256_cos"; *addr_ptr = StubRoutines::vector_float256_cos(); break;
7709 case 512: *name_ptr = "vector_float512_cos"; *addr_ptr = StubRoutines::vector_float512_cos(); break;
7710 default: Unimplemented(); break;
7711 }
7712 }
7713 break;
7714 case OP_TAN: {
7715 switch(bits) {
7716 case 64: *name_ptr = "vector_float64_tan"; *addr_ptr = StubRoutines::vector_float64_tan(); break;
7717 case 128: *name_ptr = "vector_float128_tan"; *addr_ptr = StubRoutines::vector_float128_tan(); break;
7718 case 256: *name_ptr = "vector_float256_tan"; *addr_ptr = StubRoutines::vector_float256_tan(); break;
7719 case 512: *name_ptr = "vector_float512_tan"; *addr_ptr = StubRoutines::vector_float512_tan(); break;
7720 default: Unimplemented(); break;
7721 }
7722 }
7723 break;
7724 case OP_SINH: {
7725 switch(bits) {
7726 case 64: *name_ptr = "vector_float64_sinh"; *addr_ptr = StubRoutines::vector_float64_sinh(); break;
7727 case 128: *name_ptr = "vector_float128_sinh"; *addr_ptr = StubRoutines::vector_float128_sinh(); break;
7728 case 256: *name_ptr = "vector_float256_sinh"; *addr_ptr = StubRoutines::vector_float256_sinh(); break;
7729 case 512: *name_ptr = "vector_float512_sinh"; *addr_ptr = StubRoutines::vector_float512_sinh(); break;
7730 default: Unimplemented(); break;
7731 }
7732 }
7733 break;
7734 case OP_COSH: {
7735 switch(bits) {
7736 case 64: *name_ptr = "vector_float64_cosh"; *addr_ptr = StubRoutines::vector_float64_cosh(); break;
7737 case 128: *name_ptr = "vector_float128_cosh"; *addr_ptr = StubRoutines::vector_float128_cosh(); break;
7738 case 256: *name_ptr = "vector_float256_cosh"; *addr_ptr = StubRoutines::vector_float256_cosh(); break;
7739 case 512: *name_ptr = "vector_float512_cosh"; *addr_ptr = StubRoutines::vector_float512_cosh(); break;
7740 default: Unimplemented(); break;
7741 }
7742 }
7743 break;
7744 case OP_TANH: {
7745 switch(bits) {
7746 case 64: *name_ptr = "vector_float64_tanh"; *addr_ptr = StubRoutines::vector_float64_tanh(); break;
7747 case 128: *name_ptr = "vector_float128_tanh"; *addr_ptr = StubRoutines::vector_float128_tanh(); break;
7748 case 256: *name_ptr = "vector_float256_tanh"; *addr_ptr = StubRoutines::vector_float256_tanh(); break;
7749 case 512: *name_ptr = "vector_float512_tanh"; *addr_ptr = StubRoutines::vector_float512_tanh(); break;
7750 default: Unimplemented(); break;
7751 }
7752 }
7753 break;
7754 case OP_ASIN: {
7755 switch(bits) {
7756 case 64: *name_ptr = "vector_float64_asin"; *addr_ptr = StubRoutines::vector_float64_asin(); break;
7757 case 128: *name_ptr = "vector_float128_asin"; *addr_ptr = StubRoutines::vector_float128_asin(); break;
7758 case 256: *name_ptr = "vector_float256_asin"; *addr_ptr = StubRoutines::vector_float256_asin(); break;
7759 case 512: *name_ptr = "vector_float512_asin"; *addr_ptr = StubRoutines::vector_float512_asin(); break;
7760 default: Unimplemented(); break;
7761 }
7762 }
7763 break;
7764 case OP_ACOS: {
7765 switch(bits) {
7766 case 64: *name_ptr = "vector_float64_acos"; *addr_ptr = StubRoutines::vector_float64_acos(); break;
7767 case 128: *name_ptr = "vector_float128_acos"; *addr_ptr = StubRoutines::vector_float128_acos(); break;
7768 case 256: *name_ptr = "vector_float256_acos"; *addr_ptr = StubRoutines::vector_float256_acos(); break;
7769 case 512: *name_ptr = "vector_float512_acos"; *addr_ptr = StubRoutines::vector_float512_acos(); break;
7770 default: Unimplemented(); break;
7771 }
7772 }
7773 break;
7774 case OP_ATAN: {
7775 switch(bits) {
7776 case 64: *name_ptr = "vector_float64_atan"; *addr_ptr = StubRoutines::vector_float64_atan(); break;
7777 case 128: *name_ptr = "vector_float128_atan"; *addr_ptr = StubRoutines::vector_float128_atan(); break;
7778 case 256: *name_ptr = "vector_float256_atan"; *addr_ptr = StubRoutines::vector_float256_atan(); break;
7779 case 512: *name_ptr = "vector_float512_atan"; *addr_ptr = StubRoutines::vector_float512_atan(); break;
7780 default: Unimplemented(); break;
7781 }
7782 }
7783 break;
7784 case OP_CBRT: {
7785 switch(bits) {
7786 case 64: *name_ptr = "vector_float64_cbrt"; *addr_ptr = StubRoutines::vector_float64_cbrt(); break;
7787 case 128: *name_ptr = "vector_float128_cbrt"; *addr_ptr = StubRoutines::vector_float128_cbrt(); break;
7788 case 256: *name_ptr = "vector_float256_cbrt"; *addr_ptr = StubRoutines::vector_float256_cbrt(); break;
7789 case 512: *name_ptr = "vector_float512_cbrt"; *addr_ptr = StubRoutines::vector_float512_cbrt(); break;
7790 default: Unimplemented(); break;
7791 }
7792 }
7793 break;
7794 case OP_HYPOT: {
7795 switch(bits) {
7796 case 64: *name_ptr = "vector_float64_hypot"; *addr_ptr = StubRoutines::vector_float64_hypot(); break;
7797 case 128: *name_ptr = "vector_float128_hypot"; *addr_ptr = StubRoutines::vector_float128_hypot(); break;
7798 case 256: *name_ptr = "vector_float256_hypot"; *addr_ptr = StubRoutines::vector_float256_hypot(); break;
7799 case 512: *name_ptr = "vector_float512_hypot"; *addr_ptr = StubRoutines::vector_float512_hypot(); break;
7800 default: Unimplemented(); break;
7801 }
7802 }
7803 break;
7804 case OP_POW: {
7805 switch(bits) {
7806 case 64: *name_ptr = "vector_float64_pow"; *addr_ptr = StubRoutines::vector_float64_pow(); break;
7807 case 128: *name_ptr = "vector_float128_pow"; *addr_ptr = StubRoutines::vector_float128_pow(); break;
7808 case 256: *name_ptr = "vector_float256_pow"; *addr_ptr = StubRoutines::vector_float256_pow(); break;
7809 case 512: *name_ptr = "vector_float512_pow"; *addr_ptr = StubRoutines::vector_float512_pow(); break;
7810 default: Unimplemented(); break;
7811 }
7812 }
7813 break;
7814 case OP_ATAN2: {
7815 switch(bits) {
7816 case 64: *name_ptr = "vector_float64_atan2"; *addr_ptr = StubRoutines::vector_float64_atan2(); break;
7817 case 128: *name_ptr = "vector_float128_atan2"; *addr_ptr = StubRoutines::vector_float128_atan2(); break;
7818 case 256: *name_ptr = "vector_float256_atan2"; *addr_ptr = StubRoutines::vector_float256_atan2(); break;
7819 case 512: *name_ptr = "vector_float512_atan2"; *addr_ptr = StubRoutines::vector_float512_atan2(); break;
7820 default: Unimplemented(); break;
7821 }
7822 }
7823 break;
7824 default:
7825 *name_ptr = "invalid";
7826 *addr_ptr = NULL;
7827 break;
7828 }
7829 } else {
7830 assert(bt == T_DOUBLE, "must be FP type only");
7831 switch(op) {
7832 case OP_EXP: {
7833 switch(bits) {
7834 case 64: *name_ptr = "vector_double64_exp"; *addr_ptr = StubRoutines::vector_double64_exp(); break;
7835 case 128: *name_ptr = "vector_double128_exp"; *addr_ptr = StubRoutines::vector_double128_exp(); break;
7836 case 256: *name_ptr = "vector_double256_exp"; *addr_ptr = StubRoutines::vector_double256_exp(); break;
7837 case 512: *name_ptr = "vector_double512_exp"; *addr_ptr = StubRoutines::vector_double512_exp(); break;
7838 default: Unimplemented(); break;
7839 }
7840 }
7841 break;
7842 case OP_LOG1P: {
7843 switch(bits) {
7844 case 64: *name_ptr = "vector_double64_log1p"; *addr_ptr = StubRoutines::vector_double64_log1p(); break;
7845 case 128: *name_ptr = "vector_double128_log1p"; *addr_ptr = StubRoutines::vector_double128_log1p(); break;
7846 case 256: *name_ptr = "vector_double256_log1p"; *addr_ptr = StubRoutines::vector_double256_log1p(); break;
7847 case 512: *name_ptr = "vector_double512_log1p"; *addr_ptr = StubRoutines::vector_double512_log1p(); break;
7848 default: Unimplemented(); break;
7849 }
7850 }
7851 break;
7852 case OP_LOG: {
7853 switch(bits) {
7854 case 64: *name_ptr = "vector_double64_log"; *addr_ptr = StubRoutines::vector_double64_log(); break;
7855 case 128: *name_ptr = "vector_double128_log"; *addr_ptr = StubRoutines::vector_double128_log(); break;
7856 case 256: *name_ptr = "vector_double256_log"; *addr_ptr = StubRoutines::vector_double256_log(); break;
7857 case 512: *name_ptr = "vector_double512_log"; *addr_ptr = StubRoutines::vector_double512_log(); break;
7858 default: Unimplemented(); break;
7859 }
7860 }
7861 break;
7862 case OP_LOG10: {
7863 switch(bits) {
7864 case 64: *name_ptr = "vector_double64_log10"; *addr_ptr = StubRoutines::vector_double64_log10(); break;
7865 case 128: *name_ptr = "vector_double128_log10"; *addr_ptr = StubRoutines::vector_double128_log10(); break;
7866 case 256: *name_ptr = "vector_double256_log10"; *addr_ptr = StubRoutines::vector_double256_log10(); break;
7867 case 512: *name_ptr = "vector_double512_log10"; *addr_ptr = StubRoutines::vector_double512_log10(); break;
7868 default: Unimplemented(); break;
7869 }
7870 }
7871 break;
7872 case OP_EXPM1: {
7873 switch(bits) {
7874 case 64: *name_ptr = "vector_double64_expm1"; *addr_ptr = StubRoutines::vector_double64_expm1(); break;
7875 case 128: *name_ptr = "vector_double128_expm1"; *addr_ptr = StubRoutines::vector_double128_expm1(); break;
7876 case 256: *name_ptr = "vector_double256_expm1"; *addr_ptr = StubRoutines::vector_double256_expm1(); break;
7877 case 512: *name_ptr = "vector_double512_expm1"; *addr_ptr = StubRoutines::vector_double512_expm1(); break;
7878 default: Unimplemented(); break;
7879 }
7880 }
7881 break;
7882 case OP_SIN: {
7883 switch(bits) {
7884 case 64: *name_ptr = "vector_double64_sin"; *addr_ptr = StubRoutines::vector_double64_sin(); break;
7885 case 128: *name_ptr = "vector_double128_sin"; *addr_ptr = StubRoutines::vector_double128_sin(); break;
7886 case 256: *name_ptr = "vector_double256_sin"; *addr_ptr = StubRoutines::vector_double256_sin(); break;
7887 case 512: *name_ptr = "vector_double512_sin"; *addr_ptr = StubRoutines::vector_double512_sin(); break;
7888 default: Unimplemented(); break;
7889 }
7890 }
7891 break;
7892 case OP_COS: {
7893 switch(bits) {
7894 case 64: *name_ptr = "vector_double64_cos"; *addr_ptr = StubRoutines::vector_double64_cos(); break;
7895 case 128: *name_ptr = "vector_double128_cos"; *addr_ptr = StubRoutines::vector_double128_cos(); break;
7896 case 256: *name_ptr = "vector_double256_cos"; *addr_ptr = StubRoutines::vector_double256_cos(); break;
7897 case 512: *name_ptr = "vector_double512_cos"; *addr_ptr = StubRoutines::vector_double512_cos(); break;
7898 default: Unimplemented(); break;
7899 }
7900 }
7901 break;
7902 case OP_TAN: {
7903 switch(bits) {
7904 case 64: *name_ptr = "vector_double64_tan"; *addr_ptr = StubRoutines::vector_double64_tan(); break;
7905 case 128: *name_ptr = "vector_double128_tan"; *addr_ptr = StubRoutines::vector_double128_tan(); break;
7906 case 256: *name_ptr = "vector_double256_tan"; *addr_ptr = StubRoutines::vector_double256_tan(); break;
7907 case 512: *name_ptr = "vector_double512_tan"; *addr_ptr = StubRoutines::vector_double512_tan(); break;
7908 default: Unimplemented(); break;
7909 }
7910 }
7911 break;
7912 case OP_SINH: {
7913 switch(bits) {
7914 case 64: *name_ptr = "vector_double64_sinh"; *addr_ptr = StubRoutines::vector_double64_sinh(); break;
7915 case 128: *name_ptr = "vector_double128_sinh"; *addr_ptr = StubRoutines::vector_double128_sinh(); break;
7916 case 256: *name_ptr = "vector_double256_sinh"; *addr_ptr = StubRoutines::vector_double256_sinh(); break;
7917 case 512: *name_ptr = "vector_double512_sinh"; *addr_ptr = StubRoutines::vector_double512_sinh(); break;
7918 default: Unimplemented(); break;
7919 }
7920 }
7921 break;
7922 case OP_COSH: {
7923 switch(bits) {
7924 case 64: *name_ptr = "vector_double64_cosh"; *addr_ptr = StubRoutines::vector_double64_cosh(); break;
7925 case 128: *name_ptr = "vector_double128_cosh"; *addr_ptr = StubRoutines::vector_double128_cosh(); break;
7926 case 256: *name_ptr = "vector_double256_cosh"; *addr_ptr = StubRoutines::vector_double256_cosh(); break;
7927 case 512: *name_ptr = "vector_double512_cosh"; *addr_ptr = StubRoutines::vector_double512_cosh(); break;
7928 default: Unimplemented(); break;
7929 }
7930 }
7931 break;
7932 case OP_TANH: {
7933 switch(bits) {
7934 case 64: *name_ptr = "vector_double64_tanh"; *addr_ptr = StubRoutines::vector_double64_tanh(); break;
7935 case 128: *name_ptr = "vector_double128_tanh"; *addr_ptr = StubRoutines::vector_double128_tanh(); break;
7936 case 256: *name_ptr = "vector_double256_tanh"; *addr_ptr = StubRoutines::vector_double256_tanh(); break;
7937 case 512: *name_ptr = "vector_double512_tanh"; *addr_ptr = StubRoutines::vector_double512_tanh(); break;
7938 default: Unimplemented(); break;
7939 }
7940 }
7941 break;
7942 case OP_ASIN: {
7943 switch(bits) {
7944 case 64: *name_ptr = "vector_double64_asin"; *addr_ptr = StubRoutines::vector_double64_asin(); break;
7945 case 128: *name_ptr = "vector_double128_asin"; *addr_ptr = StubRoutines::vector_double128_asin(); break;
7946 case 256: *name_ptr = "vector_double256_asin"; *addr_ptr = StubRoutines::vector_double256_asin(); break;
7947 case 512: *name_ptr = "vector_double512_asin"; *addr_ptr = StubRoutines::vector_double512_asin(); break;
7948 default: Unimplemented(); break;
7949 }
7950 }
7951 break;
7952 case OP_ACOS: {
7953 switch(bits) {
7954 case 64: *name_ptr = "vector_double64_acos"; *addr_ptr = StubRoutines::vector_double64_acos(); break;
7955 case 128: *name_ptr = "vector_double128_acos"; *addr_ptr = StubRoutines::vector_double128_acos(); break;
7956 case 256: *name_ptr = "vector_double256_acos"; *addr_ptr = StubRoutines::vector_double256_acos(); break;
7957 case 512: *name_ptr = "vector_double512_acos"; *addr_ptr = StubRoutines::vector_double512_acos(); break;
7958 default: Unimplemented(); break;
7959 }
7960 }
7961 break;
7962 case OP_ATAN: {
7963 switch(bits) {
7964 case 64: *name_ptr = "vector_double64_atan"; *addr_ptr = StubRoutines::vector_double64_atan(); break;
7965 case 128: *name_ptr = "vector_double128_atan"; *addr_ptr = StubRoutines::vector_double128_atan(); break;
7966 case 256: *name_ptr = "vector_double256_atan"; *addr_ptr = StubRoutines::vector_double256_atan(); break;
7967 case 512: *name_ptr = "vector_double512_atan"; *addr_ptr = StubRoutines::vector_double512_atan(); break;
7968 default: Unimplemented(); break;
7969 }
7970 }
7971 break;
7972 case OP_CBRT: {
7973 switch(bits) {
7974 case 64: *name_ptr = "vector_double64_cbrt"; *addr_ptr = StubRoutines::vector_double64_cbrt(); break;
7975 case 128: *name_ptr = "vector_double128_cbrt"; *addr_ptr = StubRoutines::vector_double128_cbrt(); break;
7976 case 256: *name_ptr = "vector_double256_cbrt"; *addr_ptr = StubRoutines::vector_double256_cbrt(); break;
7977 case 512: *name_ptr = "vector_double512_cbrt"; *addr_ptr = StubRoutines::vector_double512_cbrt(); break;
7978 default: Unimplemented(); break;
7979 }
7980 }
7981 break;
7982 case OP_HYPOT: {
7983 switch(bits) {
7984 case 64: *name_ptr = "vector_double64_hypot"; *addr_ptr = StubRoutines::vector_double64_hypot(); break;
7985 case 128: *name_ptr = "vector_double128_hypot"; *addr_ptr = StubRoutines::vector_double128_hypot(); break;
7986 case 256: *name_ptr = "vector_double256_hypot"; *addr_ptr = StubRoutines::vector_double256_hypot(); break;
7987 case 512: *name_ptr = "vector_double512_hypot"; *addr_ptr = StubRoutines::vector_double512_hypot(); break;
7988 default: Unimplemented(); break;
7989 }
7990 }
7991 break;
7992 case OP_POW: {
7993 switch(bits) {
7994 case 64: *name_ptr = "vector_double64_pow"; *addr_ptr = StubRoutines::vector_double64_pow(); break;
7995 case 128: *name_ptr = "vector_double128_pow"; *addr_ptr = StubRoutines::vector_double128_pow(); break;
7996 case 256: *name_ptr = "vector_double256_pow"; *addr_ptr = StubRoutines::vector_double256_pow(); break;
7997 case 512: *name_ptr = "vector_double512_pow"; *addr_ptr = StubRoutines::vector_double512_pow(); break;
7998 default: Unimplemented(); break;
7999 }
8000 }
8001 break;
8002 case OP_ATAN2: {
8003 switch(bits) {
8004 case 64: *name_ptr = "vector_double64_atan2"; *addr_ptr = StubRoutines::vector_double64_atan2(); break;
8005 case 128: *name_ptr = "vector_double128_atan2"; *addr_ptr = StubRoutines::vector_double128_atan2(); break;
8006 case 256: *name_ptr = "vector_double256_atan2"; *addr_ptr = StubRoutines::vector_double256_atan2(); break;
8007 case 512: *name_ptr = "vector_double512_atan2"; *addr_ptr = StubRoutines::vector_double512_atan2(); break;
8008 default: Unimplemented(); break;
8009 }
8010 }
8011 break;
8012
8013 default:
8014 *name_ptr = "invalid";
8015 *addr_ptr = NULL;
8016 break;
8017 }
8018 }
8019 }
8020
8021 Node* LibraryCallKit::gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2) {
8022 assert(vector_api_op_id >= OP_SVML_START && vector_api_op_id <= OP_SVML_END, "need valid op id");
8023 assert(opd1 != NULL, "must not be null");
8024 const TypeVect* vt = TypeVect::make(bt, num_elem);
8025 const TypeFunc* call_type = OptoRuntime::Math_Vector_Vector_Type(opd2 != NULL ? 2 : 1, vt, vt);
8026 const char* name = NULL;
8027 address addr = NULL;
8028
8029 // Get address for svml method.
8030 get_svml_address(vector_api_op_id, vt->length_in_bytes() * BitsPerByte, bt, &name, &addr);
8031
8032 if (addr == NULL) {
8033 return NULL;
8034 }
8035
8036 assert(name != NULL, "name must not be null");
8037 Node* operation = make_runtime_call(RC_VECTOR,
8038 call_type,
8039 addr,
8040 name,
8041 TypePtr::BOTTOM,
8042 opd1,
8043 opd2);
8044 return _gvn.transform(new ProjNode(_gvn.transform(operation), TypeFunc::Parms));
8045 }
8046
8047 // static
8048 // <V extends Vector<?,?>>
8049 // V broadcastInt(int opr, Class<V> vectorClass, Class<?> elementType, int vlen,
8050 // V v, int i,
8051 // VectorBroadcastIntOp<V> defaultImpl) {
8052 //
8053 bool LibraryCallKit::inline_vector_broadcast_int() {
8054 const TypeInt* opr = gvn().type(argument(0))->is_int();
8055 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr();
8056 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr();
8057 const TypeInt* vlen = gvn().type(argument(3))->is_int();
8058
8059 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) {
8060 return false; // not enough info for intrinsification
8061 }
8062 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
8063 if (!elem_type->is_primitive_type()) {
8064 return false; // should be primitive type
8065 }
8066 BasicType elem_bt = elem_type->basic_type();
8067 int num_elem = vlen->get_con();
8068 int opc = get_opc(opr->get_con(), elem_bt);
8069 int sopc = get_sopc(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt);
8070 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
8071 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
8072
8073 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) {
8074 return false; // not supported
8075 }
8076 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem);
8077 Node* opd2 = shift_count(argument(5), opc, elem_bt, num_elem);
8078 if (opd1 == NULL || opd2 == NULL) {
8079 return false;
8080 }
8081 Node* operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt));
8082
8083 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem);
8084 set_vector_result(vbox);
8085 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
8086 return true;
8087 }
8088
8089 bool LibraryCallKit::inline_vector_cast_reinterpret(bool is_cast) {
8090 const TypeInstPtr* vector_klass_from = gvn().type(argument(0))->is_instptr();
8091 const TypeInstPtr* elem_klass_from = gvn().type(argument(1))->is_instptr();
8092 const TypeInt* vlen_from = gvn().type(argument(2))->is_int();
8093
8094 const TypeInstPtr* vector_klass_to = gvn().type(argument(3))->is_instptr();
8095 const TypeInstPtr* elem_klass_to = gvn().type(argument(4))->is_instptr();
8096 const TypeInt* vlen_to = gvn().type(argument(5))->is_int();
8097
8098 if (vector_klass_from->const_oop() == NULL || elem_klass_from->const_oop() == NULL || !vlen_from->is_con() ||
8099 vector_klass_to->const_oop() == NULL || elem_klass_to->const_oop() == NULL || !vlen_to->is_con()) {
8100 return false; // not enough info for intrinsification
8101 }
8102
8103 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass();
8104 ciKlass* vbox_klass_to = vector_klass_to->const_oop()->as_instance()->java_lang_Class_klass();
8105 if (!vbox_klass_from->is_vectorapi_vector() || !vbox_klass_to->is_vectorapi_vector()) {
8106 return false; // only vector & mask are supported
8107 }
8108 bool is_mask = vbox_klass_from->is_vectormask();
8109
8110 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type();
8111 if (!elem_type_from->is_primitive_type()) {
8112 return false; // should be primitive type
8113 }
8114 BasicType elem_bt_from = elem_type_from->basic_type();
8115 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type();
8116 if (!elem_type_to->is_primitive_type()) {
8117 return false; // should be primitive type
8118 }
8119 BasicType elem_bt_to = elem_type_to->basic_type();
8120 if (is_mask && elem_bt_from != elem_bt_to) {
8121 return false; // type mismatch
8122 }
8123 int num_elem_from = vlen_from->get_con();
8124 int num_elem_to = vlen_to->get_con();
8125
8126 // Check whether we can unbox to appropriate size. Even with casting, checking for reinterpret is needed
8127 // since we may need to change size.
8128 if (!arch_supports_vector(Op_VectorReinterpret,
8129 num_elem_from,
8130 elem_bt_from,
8131 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
8132 return false;
8133 }
8134
8135 // Check whether we can support resizing/reinterpreting to the new size.
8136 if (!arch_supports_vector(Op_VectorReinterpret,
8137 num_elem_to,
8138 elem_bt_to,
8139 is_mask ? VecMaskUseAll : VecMaskNotUsed)) {
8140 return false;
8141 }
8142
8143 // At this point, we know that both input and output vector registers are supported
8144 // by the architecture. Next check if the casted type is simply to same type - which means
8145 // that it is actually a resize and not a cast.
8146 if (is_cast && elem_bt_from == elem_bt_to) {
8147 is_cast = false;
8148 }
8149
8150 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from);
8151
8152 Node* opd1 = unbox_vector(argument(6), vbox_type_from, elem_bt_from, num_elem_from);
8153 if (opd1 == NULL) {
8154 return false;
8155 }
8156
8157 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from);
8158 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to);
8159
8160 Node* op = opd1;
8161 if (is_cast) {
8162 assert(!is_mask, "masks cannot be casted");
8163 int cast_vopc = VectorCastNode::opcode(elem_bt_from);
8164 // Make sure that cast is implemented to particular type/size combination.
8165 if (!arch_supports_vector(cast_vopc, num_elem_to, elem_bt_to, VecMaskNotUsed)) {
8166 return false;
8167 }
8168
8169 if (num_elem_from < num_elem_to) {
8170 // Since input and output number of elements are not consistent, we need to make sure we
8171 // properly size. Thus, first make a cast that retains the number of elements from source.
8172 // In case the size exceeds the arch size, we do the minimum.
8173 int num_elem_for_cast = MIN2(num_elem_from, Matcher::max_vector_size(elem_bt_to));
8174
8175 // It is possible that arch does not support this intermediate vector size
8176 // TODO More complex logic required here to handle this corner case for the sizes.
8177 if (!arch_supports_vector(cast_vopc, num_elem_for_cast, elem_bt_to, VecMaskNotUsed)) {
8178 return false;
8179 }
8180
8181 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_for_cast));
8182 // Now ensure that the destination gets properly resized to needed size.
8183 op = _gvn.transform(new VectorReinterpretNode(op, op->bottom_type()->is_vect(), dst_type));
8184 } else if (num_elem_from > num_elem_to) {
8185 // Since number elements from input is larger than output, simply reduce size of input (we are supposed to
8186 // drop top elements anyway).
8187 int num_elem_for_resize = MAX2(num_elem_to, Matcher::min_vector_size(elem_bt_to));
8188
8189 // It is possible that arch does not support this intermediate vector size
8190 // TODO More complex logic required here to handle this corner case for the sizes.
8191 if (!arch_supports_vector(Op_VectorReinterpret,
8192 num_elem_for_resize,
8193 elem_bt_from,
8194 VecMaskNotUsed)) {
8195 return false;
8196 }
8197
8198 op = _gvn.transform(new VectorReinterpretNode(op,
8199 src_type,
8200 TypeVect::make(elem_bt_from,
8201 num_elem_for_resize)));
8202 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
8203 } else {
8204 // Since input and output number of elements match, and since we know this vector size is
8205 // supported, simply do a cast with no resize needed.
8206 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to));
8207 }
8208 } else if (Type::cmp(src_type, dst_type) != 0) {
8209 assert(!is_cast, "must be reinterpret");
8210 op = _gvn.transform(new VectorReinterpretNode(op, src_type, dst_type));
8211 }
8212
8213 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to);
8214 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to);
8215 set_vector_result(vbox);
8216 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem_to * type2aelembytes(elem_bt_to))));
8217 return true;
8218 }
8219
8220 bool LibraryCallKit::inline_vector_insert() {
8221 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
8222 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
8223 const TypeInt* vlen = gvn().type(argument(2))->is_int();
8224 const TypeInt* idx = gvn().type(argument(4))->is_int();
8225
8226 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) {
8227 return false; // not enough info for intrinsification
8228 }
8229 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
8230 if (!elem_type->is_primitive_type()) {
8231 return false; // should be primitive type
8232 }
8233 BasicType elem_bt = elem_type->basic_type();
8234 int num_elem = vlen->get_con();
8235 if (!arch_supports_vector(Op_VectorInsert, num_elem, elem_bt, VecMaskNotUsed)) {
8236 return false; // not supported
8237 }
8238
8239 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
8240 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
8241
8242 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
8243 if (opd == NULL) {
8244 return false;
8245 }
8246
8247 Node* insert_val = argument(5);
8248 assert(gvn().type(insert_val)->isa_long() != NULL, "expected to be long");
8249
8250 // Convert insert value back to its appropriate type.
8251 switch (elem_bt) {
8252 case T_BYTE:
8253 insert_val = gvn().transform(new ConvL2INode(insert_val));
8254 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::BYTE));
8255 break;
8256 case T_SHORT:
8257 insert_val = gvn().transform(new ConvL2INode(insert_val));
8258 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::SHORT));
8259 break;
8260 case T_INT:
8261 insert_val = gvn().transform(new ConvL2INode(insert_val));
8262 break;
8263 case T_FLOAT:
8264 insert_val = gvn().transform(new ConvL2INode(insert_val));
8265 insert_val = gvn().transform(new MoveI2FNode(insert_val));
8266 break;
8267 case T_DOUBLE:
8268 insert_val = gvn().transform(new MoveL2DNode(insert_val));
8269 break;
8270 case T_LONG:
8271 // no conversion needed
8272 break;
8273 default: fatal("%s", type2name(elem_bt)); break;
8274 }
8275
8276 Node* operation = _gvn.transform(VectorInsertNode::make(opd, insert_val, idx->get_con()));
8277 operation = box_vector(operation, vbox_type, elem_bt, num_elem);
8278 set_vector_result(operation);
8279
8280 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt))));
8281 return true;
8282 }
8283
8284 bool LibraryCallKit::inline_vector_extract() {
8285 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr();
8286 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr();
8287 const TypeInt* vlen = gvn().type(argument(2))->is_int();
8288 const TypeInt* idx = gvn().type(argument(4))->is_int();
8289
8290 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) {
8291 return false; // not enough info for intrinsification
8292 }
8293 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type();
8294 if (!elem_type->is_primitive_type()) {
8295 return false; // should be primitive type
8296 }
8297 BasicType elem_bt = elem_type->basic_type();
8298 int num_elem = vlen->get_con();
8299 int vopc = ExtractNode::opcode(elem_bt);
8300 if (!arch_supports_vector(vopc, num_elem, elem_bt, VecMaskNotUsed)) {
8301 return false; // not supported
8302 }
8303
8304 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass();
8305 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass);
8306
8307 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem);
8308 if (opd == NULL) {
8309 return false;
8310 }
8311
8312 Node* operation = gvn().transform(ExtractNode::make(opd, idx->get_con(), elem_bt));
8313
8314 Node* bits = NULL;
8315 switch (elem_bt) {
8316 case T_BYTE:
8317 case T_SHORT:
8318 case T_INT: {
8319 bits = gvn().transform(new ConvI2LNode(operation));
8320 break;
8321 }
8322 case T_FLOAT: {
8323 bits = gvn().transform(new MoveF2INode(operation));
8324 bits = gvn().transform(new ConvI2LNode(bits));
8325 break;
8326 }
8327 case T_DOUBLE: {
8328 bits = gvn().transform(new MoveD2LNode(operation));
8329 break;
8330 }
8331 case T_LONG: {
8332 bits = operation; // no conversion needed
8333 break;
8334 }
8335 default: fatal("%s", type2name(elem_bt));
8336 }
8337
8338 set_vector_result(bits);
8339 return true;
8340 }
8341
8342 //------------------------------get_state_from_sha_object-----------------------
8343 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
8344 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
8345 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
8346 if (sha_state == NULL) return (Node *) NULL;
8347
8348 // now have the array, need to get the start address of the state array
8349 sha_state = access_resolve(sha_state, ACCESS_WRITE);
8350 Node* state = array_element_address(sha_state, intcon(0), T_INT);
8351 return state;
8352 }
8353
8354 //------------------------------get_state_from_sha5_object-----------------------
8355 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
8356 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
8357 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
8358 if (sha_state == NULL) return (Node *) NULL;
8359
8360 // now have the array, need to get the start address of the state array
8361 sha_state = access_resolve(sha_state, ACCESS_WRITE);
8362 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
8363 return state;
8364 }
8365
8366 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
8367 // Return node representing slow path of predicate check.
8368 // the pseudo code we want to emulate with this predicate is:
8369 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
8370 //
8371 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
8372 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
8373 "need SHA1/SHA256/SHA512 instruction support");
8374 assert((uint)predicate < 3, "sanity");
8375
8376 // The receiver was checked for NULL already.
8377 Node* digestBaseObj = argument(0);
8378
8379 // get DigestBase klass for instanceOf check
8380 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
8381 assert(tinst != NULL, "digestBaseObj is null");
8382 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
8383
8384 const char* klass_SHA_name = NULL;
8385 switch (predicate) {
8386 case 0:
8387 if (UseSHA1Intrinsics) {
8388 // we want to do an instanceof comparison against the SHA class
8389 klass_SHA_name = "sun/security/provider/SHA";
8390 }
8391 break;
8392 case 1:
8393 if (UseSHA256Intrinsics) {
8394 // we want to do an instanceof comparison against the SHA2 class
8395 klass_SHA_name = "sun/security/provider/SHA2";
8396 }
8397 break;
8398 case 2:
8399 if (UseSHA512Intrinsics) {
8400 // we want to do an instanceof comparison against the SHA5 class
8401 klass_SHA_name = "sun/security/provider/SHA5";
8402 }
8403 break;
8404 default:
8405 fatal("unknown SHA intrinsic predicate: %d", predicate);
8406 }
8407
8408 ciKlass* klass_SHA = NULL;
8409 if (klass_SHA_name != NULL) {
8410 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
8411 }
8412 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
8413 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
8414 Node* ctrl = control();
8415 set_control(top()); // no intrinsic path
8416 return ctrl;
8417 }
8418 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
8419
8420 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
8421 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
8422 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
8423 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
8424
8425 return instof_false; // even if it is NULL
8426 }
8427
8428 //-------------inline_fma-----------------------------------
8429 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
8430 Node *a = NULL;
8431 Node *b = NULL;
8432 Node *c = NULL;
8433 Node* result = NULL;
8434 switch (id) {
8435 case vmIntrinsics::_fmaD:
8436 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
8437 // no receiver since it is static method
8438 a = round_double_node(argument(0));
8439 b = round_double_node(argument(2));
8440 c = round_double_node(argument(4));
8441 result = _gvn.transform(new FmaDNode(control(), a, b, c));
8442 break;
8443 case vmIntrinsics::_fmaF:
8444 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
8445 a = argument(0);
8446 b = argument(1);
8447 c = argument(2);
8448 result = _gvn.transform(new FmaFNode(control(), a, b, c));
8449 break;
8450 default:
8451 fatal_unexpected_iid(id); break;
8452 }
8453 set_result(result);
8454 return true;
8455 }
8456
8457 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) {
8458 // argument(0) is receiver
8459 Node* codePoint = argument(1);
8460 Node* n = NULL;
8461
8462 switch (id) {
8463 case vmIntrinsics::_isDigit :
8464 n = new DigitNode(control(), codePoint);
8465 break;
8466 case vmIntrinsics::_isLowerCase :
8467 n = new LowerCaseNode(control(), codePoint);
8468 break;
8469 case vmIntrinsics::_isUpperCase :
8470 n = new UpperCaseNode(control(), codePoint);
8471 break;
8472 case vmIntrinsics::_isWhitespace :
8473 n = new WhitespaceNode(control(), codePoint);
8474 break;
8475 default:
8476 fatal_unexpected_iid(id);
8477 }
8478
8479 set_result(_gvn.transform(n));
8480 return true;
8481 }
8482
8483 //------------------------------inline_fp_min_max------------------------------
8484 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) {
8485 /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416.
8486
8487 // The intrinsic should be used only when the API branches aren't predictable,
8488 // the last one performing the most important comparison. The following heuristic
8489 // uses the branch statistics to eventually bail out if necessary.
8490
8491 ciMethodData *md = callee()->method_data();
8492
8493 if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) {
8494 ciCallProfile cp = caller()->call_profile_at_bci(bci());
8495
8496 if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) {
8497 // Bail out if the call-site didn't contribute enough to the statistics.
8498 return false;
8499 }
8500
8501 uint taken = 0, not_taken = 0;
8502
8503 for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) {
8504 if (p->is_BranchData()) {
8505 taken = ((ciBranchData*)p)->taken();
8506 not_taken = ((ciBranchData*)p)->not_taken();
8507 }
8508 }
8509
8510 double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count());
8511 balance = balance < 0 ? -balance : balance;
8512 if ( balance > 0.2 ) {
8513 // Bail out if the most important branch is predictable enough.
8514 return false;
8515 }
8516 }
8517 */
8518
8519 Node *a = NULL;
8520 Node *b = NULL;
8521 Node *n = NULL;
8522 switch (id) {
8523 case vmIntrinsics::_maxF:
8524 case vmIntrinsics::_minF:
8525 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each.");
8526 a = argument(0);
8527 b = argument(1);
8528 break;
8529 case vmIntrinsics::_maxD:
8530 case vmIntrinsics::_minD:
8531 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each.");
8532 a = round_double_node(argument(0));
8533 b = round_double_node(argument(2));
8534 break;
8535 default:
8536 fatal_unexpected_iid(id);
8537 break;
8538 }
8539 if (a->is_Con() || b->is_Con()) {
8540 return false;
8541 }
8542 switch (id) {
8543 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break;
8544 case vmIntrinsics::_minF: n = new MinFNode(a, b); break;
8545 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break;
8546 case vmIntrinsics::_minD: n = new MinDNode(a, b); break;
8547 default: fatal_unexpected_iid(id); break;
8548 }
8549 set_result(_gvn.transform(n));
8550 return true;
8551 }
8552
8553 bool LibraryCallKit::inline_profileBoolean() {
8554 Node* counts = argument(1);
8555 const TypeAryPtr* ary = NULL;
8556 ciArray* aobj = NULL;
8557 if (counts->is_Con()
8558 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
8559 && (aobj = ary->const_oop()->as_array()) != NULL
8560 && (aobj->length() == 2)) {
8561 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
8562 jint false_cnt = aobj->element_value(0).as_int();
8563 jint true_cnt = aobj->element_value(1).as_int();
8564
8565 if (C->log() != NULL) {
8566 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
8567 false_cnt, true_cnt);
8568 }
8569
8570 if (false_cnt + true_cnt == 0) {
8571 // According to profile, never executed.
8572 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
8573 Deoptimization::Action_reinterpret);
8574 return true;
8575 }
8576
8577 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
8578 // is a number of each value occurrences.
8579 Node* result = argument(0);
8580 if (false_cnt == 0 || true_cnt == 0) {
8581 // According to profile, one value has been never seen.
8582 int expected_val = (false_cnt == 0) ? 1 : 0;
8583
8584 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
8585 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
8586
8587 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
8588 Node* fast_path = _gvn.transform(new IfTrueNode(check));
8589 Node* slow_path = _gvn.transform(new IfFalseNode(check));
8590
8591 { // Slow path: uncommon trap for never seen value and then reexecute
8592 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
8593 // the value has been seen at least once.
8594 PreserveJVMState pjvms(this);
8595 PreserveReexecuteState preexecs(this);
8596 jvms()->set_should_reexecute(true);
8597
8598 set_control(slow_path);
8599 set_i_o(i_o());
8600
8601 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
8602 Deoptimization::Action_reinterpret);
8603 }
8604 // The guard for never seen value enables sharpening of the result and
8605 // returning a constant. It allows to eliminate branches on the same value
8606 // later on.
8607 set_control(fast_path);
8608 result = intcon(expected_val);
8609 }
8610 // Stop profiling.
8611 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
8612 // By replacing method body with profile data (represented as ProfileBooleanNode
8613 // on IR level) we effectively disable profiling.
8614 // It enables full speed execution once optimized code is generated.
8615 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
8616 C->record_for_igvn(profile);
8617 set_result(profile);
8618 return true;
8619 } else {
8620 // Continue profiling.
8621 // Profile data isn't available at the moment. So, execute method's bytecode version.
8622 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
8623 // is compiled and counters aren't available since corresponding MethodHandle
8624 // isn't a compile-time constant.
8625 return false;
8626 }
8627 }
8628
8629 bool LibraryCallKit::inline_isCompileConstant() {
8630 Node* n = argument(0);
8631 set_result(n->is_Con() ? intcon(1) : intcon(0));
8632 return true;
8633 }