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 "prims/nativeLookup.hpp" |
56 | #include "prims/unsafe.hpp" |
57 | #include "runtime/objectMonitor.hpp" |
58 | #include "runtime/sharedRuntime.hpp" |
59 | #include "utilities/macros.hpp" |
60 | |
61 | |
62 | class LibraryIntrinsic : public InlineCallGenerator { |
63 | // Extend the set of intrinsics known to the runtime: |
64 | public: |
65 | private: |
66 | bool _is_virtual; |
67 | bool _does_virtual_dispatch; |
68 | int8_t _predicates_count; // Intrinsic is predicated by several conditions |
69 | int8_t _last_predicate; // Last generated predicate |
70 | vmIntrinsics::ID _intrinsic_id; |
71 | |
72 | public: |
73 | LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) |
74 | : InlineCallGenerator(m), |
75 | _is_virtual(is_virtual), |
76 | _does_virtual_dispatch(does_virtual_dispatch), |
77 | _predicates_count((int8_t)predicates_count), |
78 | _last_predicate((int8_t)-1), |
79 | _intrinsic_id(id) |
80 | { |
81 | } |
82 | virtual bool is_intrinsic() const { return true; } |
83 | virtual bool is_virtual() const { return _is_virtual; } |
84 | virtual bool is_predicated() const { return _predicates_count > 0; } |
85 | virtual int predicates_count() const { return _predicates_count; } |
86 | virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } |
87 | virtual JVMState* generate(JVMState* jvms); |
88 | virtual Node* generate_predicate(JVMState* jvms, int predicate); |
89 | vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } |
90 | }; |
91 | |
92 | |
93 | // Local helper class for LibraryIntrinsic: |
94 | class LibraryCallKit : public GraphKit { |
95 | private: |
96 | LibraryIntrinsic* _intrinsic; // the library intrinsic being called |
97 | Node* _result; // the result node, if any |
98 | int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted |
99 | |
100 | const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type); |
101 | |
102 | public: |
103 | LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) |
104 | : GraphKit(jvms), |
105 | _intrinsic(intrinsic), |
106 | _result(NULL) |
107 | { |
108 | // Check if this is a root compile. In that case we don't have a caller. |
109 | if (!jvms->has_method()) { |
110 | _reexecute_sp = sp(); |
111 | } else { |
112 | // Find out how many arguments the interpreter needs when deoptimizing |
113 | // and save the stack pointer value so it can used by uncommon_trap. |
114 | // We find the argument count by looking at the declared signature. |
115 | bool ignored_will_link; |
116 | ciSignature* declared_signature = NULL; |
117 | ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); |
118 | const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); |
119 | _reexecute_sp = sp() + nargs; // "push" arguments back on stack |
120 | } |
121 | } |
122 | |
123 | virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } |
124 | |
125 | ciMethod* caller() const { return jvms()->method(); } |
126 | int bci() const { return jvms()->bci(); } |
127 | LibraryIntrinsic* intrinsic() const { return _intrinsic; } |
128 | vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } |
129 | ciMethod* callee() const { return _intrinsic->method(); } |
130 | |
131 | bool try_to_inline(int predicate); |
132 | Node* try_to_predicate(int predicate); |
133 | |
134 | void push_result() { |
135 | // Push the result onto the stack. |
136 | if (!stopped() && result() != NULL) { |
137 | BasicType bt = result()->bottom_type()->basic_type(); |
138 | push_node(bt, result()); |
139 | } |
140 | } |
141 | |
142 | private: |
143 | void fatal_unexpected_iid(vmIntrinsics::ID iid) { |
144 | fatal("unexpected intrinsic %d: %s" , iid, vmIntrinsics::name_at(iid)); |
145 | } |
146 | |
147 | void set_result(Node* n) { assert(_result == NULL, "only set once" ); _result = n; } |
148 | void set_result(RegionNode* region, PhiNode* value); |
149 | Node* result() { return _result; } |
150 | |
151 | virtual int reexecute_sp() { return _reexecute_sp; } |
152 | |
153 | // Helper functions to inline natives |
154 | Node* generate_guard(Node* test, RegionNode* region, float true_prob); |
155 | Node* generate_slow_guard(Node* test, RegionNode* region); |
156 | Node* generate_fair_guard(Node* test, RegionNode* region); |
157 | Node* generate_negative_guard(Node* index, RegionNode* region, |
158 | // resulting CastII of index: |
159 | Node* *pos_index = NULL); |
160 | Node* generate_limit_guard(Node* offset, Node* subseq_length, |
161 | Node* array_length, |
162 | RegionNode* region); |
163 | void generate_string_range_check(Node* array, Node* offset, |
164 | Node* length, bool char_count); |
165 | Node* generate_current_thread(Node* &tls_output); |
166 | Node* load_mirror_from_klass(Node* klass); |
167 | Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, |
168 | RegionNode* region, int null_path, |
169 | int offset); |
170 | Node* load_klass_from_mirror(Node* mirror, bool never_see_null, |
171 | RegionNode* region, int null_path) { |
172 | int offset = java_lang_Class::klass_offset_in_bytes(); |
173 | return load_klass_from_mirror_common(mirror, never_see_null, |
174 | region, null_path, |
175 | offset); |
176 | } |
177 | Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, |
178 | RegionNode* region, int null_path) { |
179 | int offset = java_lang_Class::array_klass_offset_in_bytes(); |
180 | return load_klass_from_mirror_common(mirror, never_see_null, |
181 | region, null_path, |
182 | offset); |
183 | } |
184 | Node* generate_access_flags_guard(Node* kls, |
185 | int modifier_mask, int modifier_bits, |
186 | RegionNode* region); |
187 | Node* generate_interface_guard(Node* kls, RegionNode* region); |
188 | Node* generate_array_guard(Node* kls, RegionNode* region) { |
189 | return generate_array_guard_common(kls, region, false, false); |
190 | } |
191 | Node* generate_non_array_guard(Node* kls, RegionNode* region) { |
192 | return generate_array_guard_common(kls, region, false, true); |
193 | } |
194 | Node* generate_objArray_guard(Node* kls, RegionNode* region) { |
195 | return generate_array_guard_common(kls, region, true, false); |
196 | } |
197 | Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { |
198 | return generate_array_guard_common(kls, region, true, true); |
199 | } |
200 | Node* generate_array_guard_common(Node* kls, RegionNode* region, |
201 | bool obj_array, bool not_array); |
202 | Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); |
203 | CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, |
204 | bool is_virtual = false, bool is_static = false); |
205 | CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { |
206 | return generate_method_call(method_id, false, true); |
207 | } |
208 | CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { |
209 | return generate_method_call(method_id, true, false); |
210 | } |
211 | Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); |
212 | Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); |
213 | |
214 | Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); |
215 | bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); |
216 | bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); |
217 | bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); |
218 | Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, |
219 | RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); |
220 | bool inline_string_indexOfChar(); |
221 | bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); |
222 | bool inline_string_toBytesU(); |
223 | bool inline_string_getCharsU(); |
224 | bool inline_string_copy(bool compress); |
225 | bool inline_string_char_access(bool is_store); |
226 | Node* round_double_node(Node* n); |
227 | bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); |
228 | bool inline_math_native(vmIntrinsics::ID id); |
229 | bool inline_math(vmIntrinsics::ID id); |
230 | bool inline_double_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 | void clear_upper_avx() { |
334 | #ifdef X86 |
335 | if (UseAVX >= 2) { |
336 | C->set_clear_upper_avx(true); |
337 | } |
338 | #endif |
339 | } |
340 | }; |
341 | |
342 | //---------------------------make_vm_intrinsic---------------------------- |
343 | CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { |
344 | vmIntrinsics::ID id = m->intrinsic_id(); |
345 | assert(id != vmIntrinsics::_none, "must be a VM intrinsic" ); |
346 | |
347 | if (!m->is_loaded()) { |
348 | // Do not attempt to inline unloaded methods. |
349 | return NULL; |
350 | } |
351 | |
352 | C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); |
353 | bool is_available = false; |
354 | |
355 | { |
356 | // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag |
357 | // the compiler must transition to '_thread_in_vm' state because both |
358 | // methods access VM-internal data. |
359 | VM_ENTRY_MARK; |
360 | methodHandle mh(THREAD, m->get_Method()); |
361 | is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && |
362 | !C->directive()->is_intrinsic_disabled(mh) && |
363 | !vmIntrinsics::is_disabled_by_flags(mh); |
364 | |
365 | } |
366 | |
367 | if (is_available) { |
368 | assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility" ); |
369 | assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?" ); |
370 | return new LibraryIntrinsic(m, is_virtual, |
371 | vmIntrinsics::predicates_needed(id), |
372 | vmIntrinsics::does_virtual_dispatch(id), |
373 | (vmIntrinsics::ID) id); |
374 | } else { |
375 | return NULL; |
376 | } |
377 | } |
378 | |
379 | //----------------------register_library_intrinsics----------------------- |
380 | // Initialize this file's data structures, for each Compile instance. |
381 | void Compile::register_library_intrinsics() { |
382 | // Nothing to do here. |
383 | } |
384 | |
385 | JVMState* LibraryIntrinsic::generate(JVMState* jvms) { |
386 | LibraryCallKit kit(jvms, this); |
387 | Compile* C = kit.C; |
388 | int nodes = C->unique(); |
389 | #ifndef PRODUCT |
390 | if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { |
391 | char buf[1000]; |
392 | const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); |
393 | tty->print_cr("Intrinsic %s" , str); |
394 | } |
395 | #endif |
396 | ciMethod* callee = kit.callee(); |
397 | const int bci = kit.bci(); |
398 | |
399 | // Try to inline the intrinsic. |
400 | if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && |
401 | kit.try_to_inline(_last_predicate)) { |
402 | const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" |
403 | : "(intrinsic)" ; |
404 | CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); |
405 | if (C->print_intrinsics() || C->print_inlining()) { |
406 | C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); |
407 | } |
408 | C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); |
409 | if (C->log()) { |
410 | C->log()->elem("intrinsic id='%s'%s nodes='%d'" , |
411 | vmIntrinsics::name_at(intrinsic_id()), |
412 | (is_virtual() ? " virtual='1'" : "" ), |
413 | C->unique() - nodes); |
414 | } |
415 | // Push the result from the inlined method onto the stack. |
416 | kit.push_result(); |
417 | C->print_inlining_update(this); |
418 | return kit.transfer_exceptions_into_jvms(); |
419 | } |
420 | |
421 | // The intrinsic bailed out |
422 | if (jvms->has_method()) { |
423 | // Not a root compile. |
424 | const char* msg; |
425 | if (callee->intrinsic_candidate()) { |
426 | msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)" ; |
427 | } else { |
428 | msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" |
429 | : "failed to inline (intrinsic), method not annotated" ; |
430 | } |
431 | CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); |
432 | if (C->print_intrinsics() || C->print_inlining()) { |
433 | C->print_inlining(callee, jvms->depth() - 1, bci, msg); |
434 | } |
435 | } else { |
436 | // Root compile |
437 | ResourceMark rm; |
438 | stringStream msg_stream; |
439 | msg_stream.print("Did not generate intrinsic %s%s at bci:%d in" , |
440 | vmIntrinsics::name_at(intrinsic_id()), |
441 | is_virtual() ? " (virtual)" : "" , bci); |
442 | const char *msg = msg_stream.as_string(); |
443 | log_debug(jit, inlining)("%s" , msg); |
444 | if (C->print_intrinsics() || C->print_inlining()) { |
445 | tty->print("%s" , msg); |
446 | } |
447 | } |
448 | C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); |
449 | C->print_inlining_update(this); |
450 | return NULL; |
451 | } |
452 | |
453 | Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { |
454 | LibraryCallKit kit(jvms, this); |
455 | Compile* C = kit.C; |
456 | int nodes = C->unique(); |
457 | _last_predicate = predicate; |
458 | #ifndef PRODUCT |
459 | assert(is_predicated() && predicate < predicates_count(), "sanity" ); |
460 | if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { |
461 | char buf[1000]; |
462 | const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); |
463 | tty->print_cr("Predicate for intrinsic %s" , str); |
464 | } |
465 | #endif |
466 | ciMethod* callee = kit.callee(); |
467 | const int bci = kit.bci(); |
468 | |
469 | Node* slow_ctl = kit.try_to_predicate(predicate); |
470 | if (!kit.failing()) { |
471 | const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" |
472 | : "(intrinsic, predicate)" ; |
473 | CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); |
474 | if (C->print_intrinsics() || C->print_inlining()) { |
475 | C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); |
476 | } |
477 | C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); |
478 | if (C->log()) { |
479 | C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'" , |
480 | vmIntrinsics::name_at(intrinsic_id()), |
481 | (is_virtual() ? " virtual='1'" : "" ), |
482 | C->unique() - nodes); |
483 | } |
484 | return slow_ctl; // Could be NULL if the check folds. |
485 | } |
486 | |
487 | // The intrinsic bailed out |
488 | if (jvms->has_method()) { |
489 | // Not a root compile. |
490 | const char* msg = "failed to generate predicate for intrinsic" ; |
491 | CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); |
492 | if (C->print_intrinsics() || C->print_inlining()) { |
493 | C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); |
494 | } |
495 | } else { |
496 | // Root compile |
497 | ResourceMark rm; |
498 | stringStream msg_stream; |
499 | msg_stream.print("Did not generate intrinsic %s%s at bci:%d in" , |
500 | vmIntrinsics::name_at(intrinsic_id()), |
501 | is_virtual() ? " (virtual)" : "" , bci); |
502 | const char *msg = msg_stream.as_string(); |
503 | log_debug(jit, inlining)("%s" , msg); |
504 | if (C->print_intrinsics() || C->print_inlining()) { |
505 | C->print_inlining_stream()->print("%s" , msg); |
506 | } |
507 | } |
508 | C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); |
509 | return NULL; |
510 | } |
511 | |
512 | bool LibraryCallKit::try_to_inline(int predicate) { |
513 | // Handle symbolic names for otherwise undistinguished boolean switches: |
514 | const bool is_store = true; |
515 | const bool is_compress = true; |
516 | const bool is_static = true; |
517 | const bool is_volatile = true; |
518 | |
519 | if (!jvms()->has_method()) { |
520 | // Root JVMState has a null method. |
521 | assert(map()->memory()->Opcode() == Op_Parm, "" ); |
522 | // Insert the memory aliasing node |
523 | set_all_memory(reset_memory()); |
524 | } |
525 | assert(merged_memory(), "" ); |
526 | |
527 | |
528 | switch (intrinsic_id()) { |
529 | case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); |
530 | case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); |
531 | case vmIntrinsics::_getClass: return inline_native_getClass(); |
532 | |
533 | case vmIntrinsics::_dsin: |
534 | case vmIntrinsics::_dcos: |
535 | case vmIntrinsics::_dtan: |
536 | case vmIntrinsics::_dabs: |
537 | case vmIntrinsics::_fabs: |
538 | case vmIntrinsics::_iabs: |
539 | case vmIntrinsics::_labs: |
540 | case vmIntrinsics::_datan2: |
541 | case vmIntrinsics::_dsqrt: |
542 | case vmIntrinsics::_dexp: |
543 | case vmIntrinsics::_dlog: |
544 | case vmIntrinsics::_dlog10: |
545 | case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); |
546 | |
547 | case vmIntrinsics::_min: |
548 | case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); |
549 | |
550 | case vmIntrinsics::_notify: |
551 | case vmIntrinsics::_notifyAll: |
552 | return inline_notify(intrinsic_id()); |
553 | |
554 | case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); |
555 | case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); |
556 | case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); |
557 | case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); |
558 | case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); |
559 | case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); |
560 | case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); |
561 | case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); |
562 | case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); |
563 | case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); |
564 | case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); |
565 | case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); |
566 | case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); |
567 | |
568 | case vmIntrinsics::_arraycopy: return inline_arraycopy(); |
569 | |
570 | case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); |
571 | case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); |
572 | case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); |
573 | case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); |
574 | |
575 | case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); |
576 | case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); |
577 | case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); |
578 | case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); |
579 | case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); |
580 | case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); |
581 | case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); |
582 | |
583 | case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); |
584 | case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); |
585 | |
586 | case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); |
587 | case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); |
588 | case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); |
589 | case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); |
590 | |
591 | case vmIntrinsics::_compressStringC: |
592 | case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); |
593 | case vmIntrinsics::_inflateStringC: |
594 | case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); |
595 | |
596 | case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); |
597 | case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); |
598 | case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); |
599 | case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); |
600 | case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); |
601 | case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); |
602 | case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); |
603 | case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); |
604 | case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); |
605 | |
606 | case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); |
607 | case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); |
608 | case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); |
609 | case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); |
610 | case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); |
611 | case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); |
612 | case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); |
613 | case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); |
614 | case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); |
615 | |
616 | case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); |
617 | case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); |
618 | case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); |
619 | case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); |
620 | case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); |
621 | case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); |
622 | case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); |
623 | case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); |
624 | case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); |
625 | |
626 | case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); |
627 | case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); |
628 | case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); |
629 | case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); |
630 | case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); |
631 | case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); |
632 | case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); |
633 | case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); |
634 | case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); |
635 | |
636 | case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); |
637 | case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); |
638 | case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); |
639 | case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); |
640 | |
641 | case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); |
642 | case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); |
643 | case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); |
644 | case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); |
645 | |
646 | case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); |
647 | case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); |
648 | case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); |
649 | case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); |
650 | case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); |
651 | case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); |
652 | case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); |
653 | case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); |
654 | case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); |
655 | |
656 | case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); |
657 | case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); |
658 | case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); |
659 | case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); |
660 | case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); |
661 | case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); |
662 | case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); |
663 | case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); |
664 | case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); |
665 | |
666 | case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); |
667 | case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); |
668 | case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); |
669 | case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); |
670 | case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); |
671 | case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); |
672 | case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); |
673 | case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); |
674 | case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); |
675 | |
676 | case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); |
677 | case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); |
678 | case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); |
679 | case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); |
680 | case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); |
681 | case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); |
682 | case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); |
683 | case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); |
684 | case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); |
685 | |
686 | case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); |
687 | case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); |
688 | case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); |
689 | case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); |
690 | case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); |
691 | |
692 | case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); |
693 | case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); |
694 | case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); |
695 | case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); |
696 | case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); |
697 | case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); |
698 | case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); |
699 | case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); |
700 | case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); |
701 | case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); |
702 | case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); |
703 | case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); |
704 | case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); |
705 | case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); |
706 | case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); |
707 | case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); |
708 | case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); |
709 | case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); |
710 | case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); |
711 | case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); |
712 | |
713 | case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); |
714 | case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); |
715 | case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); |
716 | case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); |
717 | case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); |
718 | case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); |
719 | case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); |
720 | case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); |
721 | case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); |
722 | case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); |
723 | case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); |
724 | case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); |
725 | case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); |
726 | case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); |
727 | case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); |
728 | |
729 | case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); |
730 | case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); |
731 | case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); |
732 | case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); |
733 | |
734 | case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); |
735 | case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); |
736 | case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); |
737 | case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); |
738 | case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); |
739 | |
740 | case vmIntrinsics::_loadFence: |
741 | case vmIntrinsics::_storeFence: |
742 | case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); |
743 | |
744 | case vmIntrinsics::_onSpinWait: return inline_onspinwait(); |
745 | |
746 | case vmIntrinsics::_currentThread: return inline_native_currentThread(); |
747 | case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); |
748 | |
749 | #ifdef JFR_HAVE_INTRINSICS |
750 | case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime" ); |
751 | case vmIntrinsics::_getClassId: return inline_native_classID(); |
752 | case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); |
753 | #endif |
754 | case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis" ); |
755 | case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime" ); |
756 | case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); |
757 | case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); |
758 | case vmIntrinsics::_getLength: return inline_native_getLength(); |
759 | case vmIntrinsics::_copyOf: return inline_array_copyOf(false); |
760 | case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); |
761 | case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); |
762 | case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); |
763 | case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); |
764 | case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); |
765 | |
766 | case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); |
767 | case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); |
768 | |
769 | case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); |
770 | |
771 | case vmIntrinsics::_isInstance: |
772 | case vmIntrinsics::_getModifiers: |
773 | case vmIntrinsics::_isInterface: |
774 | case vmIntrinsics::_isArray: |
775 | case vmIntrinsics::_isPrimitive: |
776 | case vmIntrinsics::_getSuperclass: |
777 | case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); |
778 | |
779 | case vmIntrinsics::_floatToRawIntBits: |
780 | case vmIntrinsics::_floatToIntBits: |
781 | case vmIntrinsics::_intBitsToFloat: |
782 | case vmIntrinsics::_doubleToRawLongBits: |
783 | case vmIntrinsics::_doubleToLongBits: |
784 | case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); |
785 | |
786 | case vmIntrinsics::_numberOfLeadingZeros_i: |
787 | case vmIntrinsics::_numberOfLeadingZeros_l: |
788 | case vmIntrinsics::_numberOfTrailingZeros_i: |
789 | case vmIntrinsics::_numberOfTrailingZeros_l: |
790 | case vmIntrinsics::_bitCount_i: |
791 | case vmIntrinsics::_bitCount_l: |
792 | case vmIntrinsics::_reverseBytes_i: |
793 | case vmIntrinsics::_reverseBytes_l: |
794 | case vmIntrinsics::_reverseBytes_s: |
795 | case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); |
796 | |
797 | case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); |
798 | |
799 | case vmIntrinsics::_Reference_get: return inline_reference_get(); |
800 | |
801 | case vmIntrinsics::_Class_cast: return inline_Class_cast(); |
802 | |
803 | case vmIntrinsics::_aescrypt_encryptBlock: |
804 | case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); |
805 | |
806 | case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: |
807 | case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: |
808 | return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); |
809 | |
810 | case vmIntrinsics::_counterMode_AESCrypt: |
811 | return inline_counterMode_AESCrypt(intrinsic_id()); |
812 | |
813 | case vmIntrinsics::_sha_implCompress: |
814 | case vmIntrinsics::_sha2_implCompress: |
815 | case vmIntrinsics::_sha5_implCompress: |
816 | return inline_sha_implCompress(intrinsic_id()); |
817 | |
818 | case vmIntrinsics::_digestBase_implCompressMB: |
819 | return inline_digestBase_implCompressMB(predicate); |
820 | |
821 | case vmIntrinsics::_multiplyToLen: |
822 | return inline_multiplyToLen(); |
823 | |
824 | case vmIntrinsics::_squareToLen: |
825 | return inline_squareToLen(); |
826 | |
827 | case vmIntrinsics::_mulAdd: |
828 | return inline_mulAdd(); |
829 | |
830 | case vmIntrinsics::_montgomeryMultiply: |
831 | return inline_montgomeryMultiply(); |
832 | case vmIntrinsics::_montgomerySquare: |
833 | return inline_montgomerySquare(); |
834 | |
835 | case vmIntrinsics::_vectorizedMismatch: |
836 | return inline_vectorizedMismatch(); |
837 | |
838 | case vmIntrinsics::_ghash_processBlocks: |
839 | return inline_ghash_processBlocks(); |
840 | case vmIntrinsics::_base64_encodeBlock: |
841 | return inline_base64_encodeBlock(); |
842 | |
843 | case vmIntrinsics::_encodeISOArray: |
844 | case vmIntrinsics::_encodeByteISOArray: |
845 | return inline_encodeISOArray(); |
846 | |
847 | case vmIntrinsics::_updateCRC32: |
848 | return inline_updateCRC32(); |
849 | case vmIntrinsics::_updateBytesCRC32: |
850 | return inline_updateBytesCRC32(); |
851 | case vmIntrinsics::_updateByteBufferCRC32: |
852 | return inline_updateByteBufferCRC32(); |
853 | |
854 | case vmIntrinsics::_updateBytesCRC32C: |
855 | return inline_updateBytesCRC32C(); |
856 | case vmIntrinsics::_updateDirectByteBufferCRC32C: |
857 | return inline_updateDirectByteBufferCRC32C(); |
858 | |
859 | case vmIntrinsics::_updateBytesAdler32: |
860 | return inline_updateBytesAdler32(); |
861 | case vmIntrinsics::_updateByteBufferAdler32: |
862 | return inline_updateByteBufferAdler32(); |
863 | |
864 | case vmIntrinsics::_profileBoolean: |
865 | return inline_profileBoolean(); |
866 | case vmIntrinsics::_isCompileConstant: |
867 | return inline_isCompileConstant(); |
868 | |
869 | case vmIntrinsics::_hasNegatives: |
870 | return inline_hasNegatives(); |
871 | |
872 | case vmIntrinsics::_fmaD: |
873 | case vmIntrinsics::_fmaF: |
874 | return inline_fma(intrinsic_id()); |
875 | |
876 | case vmIntrinsics::_isDigit: |
877 | case vmIntrinsics::_isLowerCase: |
878 | case vmIntrinsics::_isUpperCase: |
879 | case vmIntrinsics::_isWhitespace: |
880 | return inline_character_compare(intrinsic_id()); |
881 | |
882 | case vmIntrinsics::_maxF: |
883 | case vmIntrinsics::_minF: |
884 | case vmIntrinsics::_maxD: |
885 | case vmIntrinsics::_minD: |
886 | return inline_fp_min_max(intrinsic_id()); |
887 | |
888 | default: |
889 | // If you get here, it may be that someone has added a new intrinsic |
890 | // to the list in vmSymbols.hpp without implementing it here. |
891 | #ifndef PRODUCT |
892 | if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { |
893 | tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)" , |
894 | vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); |
895 | } |
896 | #endif |
897 | return false; |
898 | } |
899 | } |
900 | |
901 | Node* LibraryCallKit::try_to_predicate(int predicate) { |
902 | if (!jvms()->has_method()) { |
903 | // Root JVMState has a null method. |
904 | assert(map()->memory()->Opcode() == Op_Parm, "" ); |
905 | // Insert the memory aliasing node |
906 | set_all_memory(reset_memory()); |
907 | } |
908 | assert(merged_memory(), "" ); |
909 | |
910 | switch (intrinsic_id()) { |
911 | case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: |
912 | return inline_cipherBlockChaining_AESCrypt_predicate(false); |
913 | case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: |
914 | return inline_cipherBlockChaining_AESCrypt_predicate(true); |
915 | case vmIntrinsics::_counterMode_AESCrypt: |
916 | return inline_counterMode_AESCrypt_predicate(); |
917 | case vmIntrinsics::_digestBase_implCompressMB: |
918 | return inline_digestBase_implCompressMB_predicate(predicate); |
919 | |
920 | default: |
921 | // If you get here, it may be that someone has added a new intrinsic |
922 | // to the list in vmSymbols.hpp without implementing it here. |
923 | #ifndef PRODUCT |
924 | if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { |
925 | tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)" , |
926 | vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); |
927 | } |
928 | #endif |
929 | Node* slow_ctl = control(); |
930 | set_control(top()); // No fast path instrinsic |
931 | return slow_ctl; |
932 | } |
933 | } |
934 | |
935 | //------------------------------set_result------------------------------- |
936 | // Helper function for finishing intrinsics. |
937 | void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { |
938 | record_for_igvn(region); |
939 | set_control(_gvn.transform(region)); |
940 | set_result( _gvn.transform(value)); |
941 | assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity" ); |
942 | } |
943 | |
944 | //------------------------------generate_guard--------------------------- |
945 | // Helper function for generating guarded fast-slow graph structures. |
946 | // The given 'test', if true, guards a slow path. If the test fails |
947 | // then a fast path can be taken. (We generally hope it fails.) |
948 | // In all cases, GraphKit::control() is updated to the fast path. |
949 | // The returned value represents the control for the slow path. |
950 | // The return value is never 'top'; it is either a valid control |
951 | // or NULL if it is obvious that the slow path can never be taken. |
952 | // Also, if region and the slow control are not NULL, the slow edge |
953 | // is appended to the region. |
954 | Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { |
955 | if (stopped()) { |
956 | // Already short circuited. |
957 | return NULL; |
958 | } |
959 | |
960 | // Build an if node and its projections. |
961 | // If test is true we take the slow path, which we assume is uncommon. |
962 | if (_gvn.type(test) == TypeInt::ZERO) { |
963 | // The slow branch is never taken. No need to build this guard. |
964 | return NULL; |
965 | } |
966 | |
967 | IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); |
968 | |
969 | Node* if_slow = _gvn.transform(new IfTrueNode(iff)); |
970 | if (if_slow == top()) { |
971 | // The slow branch is never taken. No need to build this guard. |
972 | return NULL; |
973 | } |
974 | |
975 | if (region != NULL) |
976 | region->add_req(if_slow); |
977 | |
978 | Node* if_fast = _gvn.transform(new IfFalseNode(iff)); |
979 | set_control(if_fast); |
980 | |
981 | return if_slow; |
982 | } |
983 | |
984 | inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { |
985 | return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); |
986 | } |
987 | inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { |
988 | return generate_guard(test, region, PROB_FAIR); |
989 | } |
990 | |
991 | inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, |
992 | Node* *pos_index) { |
993 | if (stopped()) |
994 | return NULL; // already stopped |
995 | if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] |
996 | return NULL; // index is already adequately typed |
997 | Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); |
998 | Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); |
999 | Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); |
1000 | if (is_neg != NULL && pos_index != NULL) { |
1001 | // Emulate effect of Parse::adjust_map_after_if. |
1002 | Node* ccast = new CastIINode(index, TypeInt::POS); |
1003 | ccast->set_req(0, control()); |
1004 | (*pos_index) = _gvn.transform(ccast); |
1005 | } |
1006 | return is_neg; |
1007 | } |
1008 | |
1009 | // Make sure that 'position' is a valid limit index, in [0..length]. |
1010 | // There are two equivalent plans for checking this: |
1011 | // A. (offset + copyLength) unsigned<= arrayLength |
1012 | // B. offset <= (arrayLength - copyLength) |
1013 | // We require that all of the values above, except for the sum and |
1014 | // difference, are already known to be non-negative. |
1015 | // Plan A is robust in the face of overflow, if offset and copyLength |
1016 | // are both hugely positive. |
1017 | // |
1018 | // Plan B is less direct and intuitive, but it does not overflow at |
1019 | // all, since the difference of two non-negatives is always |
1020 | // representable. Whenever Java methods must perform the equivalent |
1021 | // check they generally use Plan B instead of Plan A. |
1022 | // For the moment we use Plan A. |
1023 | inline Node* LibraryCallKit::generate_limit_guard(Node* offset, |
1024 | Node* subseq_length, |
1025 | Node* array_length, |
1026 | RegionNode* region) { |
1027 | if (stopped()) |
1028 | return NULL; // already stopped |
1029 | bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; |
1030 | if (zero_offset && subseq_length->eqv_uncast(array_length)) |
1031 | return NULL; // common case of whole-array copy |
1032 | Node* last = subseq_length; |
1033 | if (!zero_offset) // last += offset |
1034 | last = _gvn.transform(new AddINode(last, offset)); |
1035 | Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); |
1036 | Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); |
1037 | Node* is_over = generate_guard(bol_lt, region, PROB_MIN); |
1038 | return is_over; |
1039 | } |
1040 | |
1041 | // Emit range checks for the given String.value byte array |
1042 | void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { |
1043 | if (stopped()) { |
1044 | return; // already stopped |
1045 | } |
1046 | RegionNode* bailout = new RegionNode(1); |
1047 | record_for_igvn(bailout); |
1048 | if (char_count) { |
1049 | // Convert char count to byte count |
1050 | count = _gvn.transform(new LShiftINode(count, intcon(1))); |
1051 | } |
1052 | |
1053 | // Offset and count must not be negative |
1054 | generate_negative_guard(offset, bailout); |
1055 | generate_negative_guard(count, bailout); |
1056 | // Offset + count must not exceed length of array |
1057 | generate_limit_guard(offset, count, load_array_length(array), bailout); |
1058 | |
1059 | if (bailout->req() > 1) { |
1060 | PreserveJVMState pjvms(this); |
1061 | set_control(_gvn.transform(bailout)); |
1062 | uncommon_trap(Deoptimization::Reason_intrinsic, |
1063 | Deoptimization::Action_maybe_recompile); |
1064 | } |
1065 | } |
1066 | |
1067 | //--------------------------generate_current_thread-------------------- |
1068 | Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { |
1069 | ciKlass* thread_klass = env()->Thread_klass(); |
1070 | const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); |
1071 | Node* thread = _gvn.transform(new ThreadLocalNode()); |
1072 | Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); |
1073 | Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); |
1074 | tls_output = thread; |
1075 | return threadObj; |
1076 | } |
1077 | |
1078 | |
1079 | //------------------------------make_string_method_node------------------------ |
1080 | // Helper method for String intrinsic functions. This version is called with |
1081 | // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded |
1082 | // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes |
1083 | // containing the lengths of str1 and str2. |
1084 | Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { |
1085 | Node* result = NULL; |
1086 | switch (opcode) { |
1087 | case Op_StrIndexOf: |
1088 | result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), |
1089 | str1_start, cnt1, str2_start, cnt2, ae); |
1090 | break; |
1091 | case Op_StrComp: |
1092 | result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), |
1093 | str1_start, cnt1, str2_start, cnt2, ae); |
1094 | break; |
1095 | case Op_StrEquals: |
1096 | // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). |
1097 | // Use the constant length if there is one because optimized match rule may exist. |
1098 | result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), |
1099 | str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); |
1100 | break; |
1101 | default: |
1102 | ShouldNotReachHere(); |
1103 | return NULL; |
1104 | } |
1105 | |
1106 | // All these intrinsics have checks. |
1107 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
1108 | clear_upper_avx(); |
1109 | |
1110 | return _gvn.transform(result); |
1111 | } |
1112 | |
1113 | //------------------------------inline_string_compareTo------------------------ |
1114 | bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { |
1115 | Node* arg1 = argument(0); |
1116 | Node* arg2 = argument(1); |
1117 | |
1118 | arg1 = must_be_not_null(arg1, true); |
1119 | arg2 = must_be_not_null(arg2, true); |
1120 | |
1121 | arg1 = access_resolve(arg1, ACCESS_READ); |
1122 | arg2 = access_resolve(arg2, ACCESS_READ); |
1123 | |
1124 | // Get start addr and length of first argument |
1125 | Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); |
1126 | Node* arg1_cnt = load_array_length(arg1); |
1127 | |
1128 | // Get start addr and length of second argument |
1129 | Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); |
1130 | Node* arg2_cnt = load_array_length(arg2); |
1131 | |
1132 | Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); |
1133 | set_result(result); |
1134 | return true; |
1135 | } |
1136 | |
1137 | //------------------------------inline_string_equals------------------------ |
1138 | bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { |
1139 | Node* arg1 = argument(0); |
1140 | Node* arg2 = argument(1); |
1141 | |
1142 | // paths (plus control) merge |
1143 | RegionNode* region = new RegionNode(3); |
1144 | Node* phi = new PhiNode(region, TypeInt::BOOL); |
1145 | |
1146 | if (!stopped()) { |
1147 | |
1148 | arg1 = must_be_not_null(arg1, true); |
1149 | arg2 = must_be_not_null(arg2, true); |
1150 | |
1151 | arg1 = access_resolve(arg1, ACCESS_READ); |
1152 | arg2 = access_resolve(arg2, ACCESS_READ); |
1153 | |
1154 | // Get start addr and length of first argument |
1155 | Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); |
1156 | Node* arg1_cnt = load_array_length(arg1); |
1157 | |
1158 | // Get start addr and length of second argument |
1159 | Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); |
1160 | Node* arg2_cnt = load_array_length(arg2); |
1161 | |
1162 | // Check for arg1_cnt != arg2_cnt |
1163 | Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); |
1164 | Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); |
1165 | Node* if_ne = generate_slow_guard(bol, NULL); |
1166 | if (if_ne != NULL) { |
1167 | phi->init_req(2, intcon(0)); |
1168 | region->init_req(2, if_ne); |
1169 | } |
1170 | |
1171 | // Check for count == 0 is done by assembler code for StrEquals. |
1172 | |
1173 | if (!stopped()) { |
1174 | Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); |
1175 | phi->init_req(1, equals); |
1176 | region->init_req(1, control()); |
1177 | } |
1178 | } |
1179 | |
1180 | // post merge |
1181 | set_control(_gvn.transform(region)); |
1182 | record_for_igvn(region); |
1183 | |
1184 | set_result(_gvn.transform(phi)); |
1185 | return true; |
1186 | } |
1187 | |
1188 | //------------------------------inline_array_equals---------------------------- |
1189 | bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { |
1190 | assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types" ); |
1191 | Node* arg1 = argument(0); |
1192 | Node* arg2 = argument(1); |
1193 | |
1194 | arg1 = access_resolve(arg1, ACCESS_READ); |
1195 | arg2 = access_resolve(arg2, ACCESS_READ); |
1196 | |
1197 | const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; |
1198 | set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); |
1199 | clear_upper_avx(); |
1200 | |
1201 | return true; |
1202 | } |
1203 | |
1204 | //------------------------------inline_hasNegatives------------------------------ |
1205 | bool LibraryCallKit::inline_hasNegatives() { |
1206 | if (too_many_traps(Deoptimization::Reason_intrinsic)) { |
1207 | return false; |
1208 | } |
1209 | |
1210 | assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters" ); |
1211 | // no receiver since it is static method |
1212 | Node* ba = argument(0); |
1213 | Node* offset = argument(1); |
1214 | Node* len = argument(2); |
1215 | |
1216 | ba = must_be_not_null(ba, true); |
1217 | |
1218 | // Range checks |
1219 | generate_string_range_check(ba, offset, len, false); |
1220 | if (stopped()) { |
1221 | return true; |
1222 | } |
1223 | ba = access_resolve(ba, ACCESS_READ); |
1224 | Node* ba_start = array_element_address(ba, offset, T_BYTE); |
1225 | Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); |
1226 | set_result(_gvn.transform(result)); |
1227 | return true; |
1228 | } |
1229 | |
1230 | bool LibraryCallKit::inline_preconditions_checkIndex() { |
1231 | Node* index = argument(0); |
1232 | Node* length = argument(1); |
1233 | if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { |
1234 | return false; |
1235 | } |
1236 | |
1237 | Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); |
1238 | Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); |
1239 | |
1240 | { |
1241 | BuildCutout unless(this, len_pos_bol, PROB_MAX); |
1242 | uncommon_trap(Deoptimization::Reason_intrinsic, |
1243 | Deoptimization::Action_make_not_entrant); |
1244 | } |
1245 | |
1246 | if (stopped()) { |
1247 | return false; |
1248 | } |
1249 | |
1250 | Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); |
1251 | BoolTest::mask btest = BoolTest::lt; |
1252 | Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); |
1253 | RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); |
1254 | _gvn.set_type(rc, rc->Value(&_gvn)); |
1255 | if (!rc_bool->is_Con()) { |
1256 | record_for_igvn(rc); |
1257 | } |
1258 | set_control(_gvn.transform(new IfTrueNode(rc))); |
1259 | { |
1260 | PreserveJVMState pjvms(this); |
1261 | set_control(_gvn.transform(new IfFalseNode(rc))); |
1262 | uncommon_trap(Deoptimization::Reason_range_check, |
1263 | Deoptimization::Action_make_not_entrant); |
1264 | } |
1265 | |
1266 | if (stopped()) { |
1267 | return false; |
1268 | } |
1269 | |
1270 | Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); |
1271 | result->set_req(0, control()); |
1272 | result = _gvn.transform(result); |
1273 | set_result(result); |
1274 | replace_in_map(index, result); |
1275 | clear_upper_avx(); |
1276 | return true; |
1277 | } |
1278 | |
1279 | //------------------------------inline_string_indexOf------------------------ |
1280 | bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { |
1281 | if (!Matcher::match_rule_supported(Op_StrIndexOf)) { |
1282 | return false; |
1283 | } |
1284 | Node* src = argument(0); |
1285 | Node* tgt = argument(1); |
1286 | |
1287 | // Make the merge point |
1288 | RegionNode* result_rgn = new RegionNode(4); |
1289 | Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); |
1290 | |
1291 | src = must_be_not_null(src, true); |
1292 | tgt = must_be_not_null(tgt, true); |
1293 | |
1294 | src = access_resolve(src, ACCESS_READ); |
1295 | tgt = access_resolve(tgt, ACCESS_READ); |
1296 | |
1297 | // Get start addr and length of source string |
1298 | Node* src_start = array_element_address(src, intcon(0), T_BYTE); |
1299 | Node* src_count = load_array_length(src); |
1300 | |
1301 | // Get start addr and length of substring |
1302 | Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); |
1303 | Node* tgt_count = load_array_length(tgt); |
1304 | |
1305 | if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { |
1306 | // Divide src size by 2 if String is UTF16 encoded |
1307 | src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); |
1308 | } |
1309 | if (ae == StrIntrinsicNode::UU) { |
1310 | // Divide substring size by 2 if String is UTF16 encoded |
1311 | tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); |
1312 | } |
1313 | |
1314 | Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); |
1315 | if (result != NULL) { |
1316 | result_phi->init_req(3, result); |
1317 | result_rgn->init_req(3, control()); |
1318 | } |
1319 | set_control(_gvn.transform(result_rgn)); |
1320 | record_for_igvn(result_rgn); |
1321 | set_result(_gvn.transform(result_phi)); |
1322 | |
1323 | return true; |
1324 | } |
1325 | |
1326 | //-----------------------------inline_string_indexOf----------------------- |
1327 | bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { |
1328 | if (too_many_traps(Deoptimization::Reason_intrinsic)) { |
1329 | return false; |
1330 | } |
1331 | if (!Matcher::match_rule_supported(Op_StrIndexOf)) { |
1332 | return false; |
1333 | } |
1334 | assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments" ); |
1335 | Node* src = argument(0); // byte[] |
1336 | Node* src_count = argument(1); // char count |
1337 | Node* tgt = argument(2); // byte[] |
1338 | Node* tgt_count = argument(3); // char count |
1339 | Node* from_index = argument(4); // char index |
1340 | |
1341 | src = must_be_not_null(src, true); |
1342 | tgt = must_be_not_null(tgt, true); |
1343 | |
1344 | src = access_resolve(src, ACCESS_READ); |
1345 | tgt = access_resolve(tgt, ACCESS_READ); |
1346 | |
1347 | // Multiply byte array index by 2 if String is UTF16 encoded |
1348 | Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); |
1349 | src_count = _gvn.transform(new SubINode(src_count, from_index)); |
1350 | Node* src_start = array_element_address(src, src_offset, T_BYTE); |
1351 | Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); |
1352 | |
1353 | // Range checks |
1354 | generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); |
1355 | generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); |
1356 | if (stopped()) { |
1357 | return true; |
1358 | } |
1359 | |
1360 | RegionNode* region = new RegionNode(5); |
1361 | Node* phi = new PhiNode(region, TypeInt::INT); |
1362 | |
1363 | Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); |
1364 | if (result != NULL) { |
1365 | // The result is index relative to from_index if substring was found, -1 otherwise. |
1366 | // Generate code which will fold into cmove. |
1367 | Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); |
1368 | Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); |
1369 | |
1370 | Node* if_lt = generate_slow_guard(bol, NULL); |
1371 | if (if_lt != NULL) { |
1372 | // result == -1 |
1373 | phi->init_req(3, result); |
1374 | region->init_req(3, if_lt); |
1375 | } |
1376 | if (!stopped()) { |
1377 | result = _gvn.transform(new AddINode(result, from_index)); |
1378 | phi->init_req(4, result); |
1379 | region->init_req(4, control()); |
1380 | } |
1381 | } |
1382 | |
1383 | set_control(_gvn.transform(region)); |
1384 | record_for_igvn(region); |
1385 | set_result(_gvn.transform(phi)); |
1386 | clear_upper_avx(); |
1387 | |
1388 | return true; |
1389 | } |
1390 | |
1391 | // Create StrIndexOfNode with fast path checks |
1392 | Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, |
1393 | RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { |
1394 | // Check for substr count > string count |
1395 | Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); |
1396 | Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); |
1397 | Node* if_gt = generate_slow_guard(bol, NULL); |
1398 | if (if_gt != NULL) { |
1399 | phi->init_req(1, intcon(-1)); |
1400 | region->init_req(1, if_gt); |
1401 | } |
1402 | if (!stopped()) { |
1403 | // Check for substr count == 0 |
1404 | cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); |
1405 | bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); |
1406 | Node* if_zero = generate_slow_guard(bol, NULL); |
1407 | if (if_zero != NULL) { |
1408 | phi->init_req(2, intcon(0)); |
1409 | region->init_req(2, if_zero); |
1410 | } |
1411 | } |
1412 | if (!stopped()) { |
1413 | return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); |
1414 | } |
1415 | return NULL; |
1416 | } |
1417 | |
1418 | //-----------------------------inline_string_indexOfChar----------------------- |
1419 | bool LibraryCallKit::inline_string_indexOfChar() { |
1420 | if (too_many_traps(Deoptimization::Reason_intrinsic)) { |
1421 | return false; |
1422 | } |
1423 | if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { |
1424 | return false; |
1425 | } |
1426 | assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments" ); |
1427 | Node* src = argument(0); // byte[] |
1428 | Node* tgt = argument(1); // tgt is int ch |
1429 | Node* from_index = argument(2); |
1430 | Node* max = argument(3); |
1431 | |
1432 | src = must_be_not_null(src, true); |
1433 | src = access_resolve(src, ACCESS_READ); |
1434 | |
1435 | Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); |
1436 | Node* src_start = array_element_address(src, src_offset, T_BYTE); |
1437 | Node* src_count = _gvn.transform(new SubINode(max, from_index)); |
1438 | |
1439 | // Range checks |
1440 | generate_string_range_check(src, src_offset, src_count, true); |
1441 | if (stopped()) { |
1442 | return true; |
1443 | } |
1444 | |
1445 | RegionNode* region = new RegionNode(3); |
1446 | Node* phi = new PhiNode(region, TypeInt::INT); |
1447 | |
1448 | Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); |
1449 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
1450 | _gvn.transform(result); |
1451 | |
1452 | Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); |
1453 | Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); |
1454 | |
1455 | Node* if_lt = generate_slow_guard(bol, NULL); |
1456 | if (if_lt != NULL) { |
1457 | // result == -1 |
1458 | phi->init_req(2, result); |
1459 | region->init_req(2, if_lt); |
1460 | } |
1461 | if (!stopped()) { |
1462 | result = _gvn.transform(new AddINode(result, from_index)); |
1463 | phi->init_req(1, result); |
1464 | region->init_req(1, control()); |
1465 | } |
1466 | set_control(_gvn.transform(region)); |
1467 | record_for_igvn(region); |
1468 | set_result(_gvn.transform(phi)); |
1469 | |
1470 | return true; |
1471 | } |
1472 | //---------------------------inline_string_copy--------------------- |
1473 | // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) |
1474 | // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) |
1475 | // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) |
1476 | // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) |
1477 | // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) |
1478 | // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) |
1479 | bool LibraryCallKit::inline_string_copy(bool compress) { |
1480 | if (too_many_traps(Deoptimization::Reason_intrinsic)) { |
1481 | return false; |
1482 | } |
1483 | int nargs = 5; // 2 oops, 3 ints |
1484 | assert(callee()->signature()->size() == nargs, "string copy has 5 arguments" ); |
1485 | |
1486 | Node* src = argument(0); |
1487 | Node* src_offset = argument(1); |
1488 | Node* dst = argument(2); |
1489 | Node* dst_offset = argument(3); |
1490 | Node* length = argument(4); |
1491 | |
1492 | // Check for allocation before we add nodes that would confuse |
1493 | // tightly_coupled_allocation() |
1494 | AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); |
1495 | |
1496 | // Figure out the size and type of the elements we will be copying. |
1497 | const Type* src_type = src->Value(&_gvn); |
1498 | const Type* dst_type = dst->Value(&_gvn); |
1499 | BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
1500 | BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
1501 | assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || |
1502 | (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), |
1503 | "Unsupported array types for inline_string_copy" ); |
1504 | |
1505 | src = must_be_not_null(src, true); |
1506 | dst = must_be_not_null(dst, true); |
1507 | |
1508 | // Convert char[] offsets to byte[] offsets |
1509 | bool convert_src = (compress && src_elem == T_BYTE); |
1510 | bool convert_dst = (!compress && dst_elem == T_BYTE); |
1511 | if (convert_src) { |
1512 | src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); |
1513 | } else if (convert_dst) { |
1514 | dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); |
1515 | } |
1516 | |
1517 | // Range checks |
1518 | generate_string_range_check(src, src_offset, length, convert_src); |
1519 | generate_string_range_check(dst, dst_offset, length, convert_dst); |
1520 | if (stopped()) { |
1521 | return true; |
1522 | } |
1523 | |
1524 | src = access_resolve(src, ACCESS_READ); |
1525 | dst = access_resolve(dst, ACCESS_WRITE); |
1526 | |
1527 | Node* src_start = array_element_address(src, src_offset, src_elem); |
1528 | Node* dst_start = array_element_address(dst, dst_offset, dst_elem); |
1529 | // 'src_start' points to src array + scaled offset |
1530 | // 'dst_start' points to dst array + scaled offset |
1531 | Node* count = NULL; |
1532 | if (compress) { |
1533 | count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); |
1534 | } else { |
1535 | inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); |
1536 | } |
1537 | |
1538 | if (alloc != NULL) { |
1539 | if (alloc->maybe_set_complete(&_gvn)) { |
1540 | // "You break it, you buy it." |
1541 | InitializeNode* init = alloc->initialization(); |
1542 | assert(init->is_complete(), "we just did this" ); |
1543 | init->set_complete_with_arraycopy(); |
1544 | assert(dst->is_CheckCastPP(), "sanity" ); |
1545 | assert(dst->in(0)->in(0) == init, "dest pinned" ); |
1546 | } |
1547 | // Do not let stores that initialize this object be reordered with |
1548 | // a subsequent store that would make this object accessible by |
1549 | // other threads. |
1550 | // Record what AllocateNode this StoreStore protects so that |
1551 | // escape analysis can go from the MemBarStoreStoreNode to the |
1552 | // AllocateNode and eliminate the MemBarStoreStoreNode if possible |
1553 | // based on the escape status of the AllocateNode. |
1554 | insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); |
1555 | } |
1556 | if (compress) { |
1557 | set_result(_gvn.transform(count)); |
1558 | } |
1559 | clear_upper_avx(); |
1560 | |
1561 | return true; |
1562 | } |
1563 | |
1564 | #ifdef _LP64 |
1565 | #define XTOP ,top() /*additional argument*/ |
1566 | #else //_LP64 |
1567 | #define XTOP /*no additional argument*/ |
1568 | #endif //_LP64 |
1569 | |
1570 | //------------------------inline_string_toBytesU-------------------------- |
1571 | // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) |
1572 | bool LibraryCallKit::inline_string_toBytesU() { |
1573 | if (too_many_traps(Deoptimization::Reason_intrinsic)) { |
1574 | return false; |
1575 | } |
1576 | // Get the arguments. |
1577 | Node* value = argument(0); |
1578 | Node* offset = argument(1); |
1579 | Node* length = argument(2); |
1580 | |
1581 | Node* newcopy = NULL; |
1582 | |
1583 | // Set the original stack and the reexecute bit for the interpreter to reexecute |
1584 | // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. |
1585 | { PreserveReexecuteState preexecs(this); |
1586 | jvms()->set_should_reexecute(true); |
1587 | |
1588 | // Check if a null path was taken unconditionally. |
1589 | value = null_check(value); |
1590 | |
1591 | RegionNode* bailout = new RegionNode(1); |
1592 | record_for_igvn(bailout); |
1593 | |
1594 | // Range checks |
1595 | generate_negative_guard(offset, bailout); |
1596 | generate_negative_guard(length, bailout); |
1597 | generate_limit_guard(offset, length, load_array_length(value), bailout); |
1598 | // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE |
1599 | generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); |
1600 | |
1601 | if (bailout->req() > 1) { |
1602 | PreserveJVMState pjvms(this); |
1603 | set_control(_gvn.transform(bailout)); |
1604 | uncommon_trap(Deoptimization::Reason_intrinsic, |
1605 | Deoptimization::Action_maybe_recompile); |
1606 | } |
1607 | if (stopped()) { |
1608 | return true; |
1609 | } |
1610 | |
1611 | Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); |
1612 | Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); |
1613 | newcopy = new_array(klass_node, size, 0); // no arguments to push |
1614 | AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); |
1615 | |
1616 | // Calculate starting addresses. |
1617 | value = access_resolve(value, ACCESS_READ); |
1618 | Node* src_start = array_element_address(value, offset, T_CHAR); |
1619 | Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); |
1620 | |
1621 | // Check if src array address is aligned to HeapWordSize (dst is always aligned) |
1622 | const TypeInt* toffset = gvn().type(offset)->is_int(); |
1623 | bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); |
1624 | |
1625 | // Figure out which arraycopy runtime method to call (disjoint, uninitialized). |
1626 | const char* copyfunc_name = "arraycopy" ; |
1627 | address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); |
1628 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
1629 | OptoRuntime::fast_arraycopy_Type(), |
1630 | copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, |
1631 | src_start, dst_start, ConvI2X(length) XTOP); |
1632 | // Do not let reads from the cloned object float above the arraycopy. |
1633 | if (alloc != NULL) { |
1634 | if (alloc->maybe_set_complete(&_gvn)) { |
1635 | // "You break it, you buy it." |
1636 | InitializeNode* init = alloc->initialization(); |
1637 | assert(init->is_complete(), "we just did this" ); |
1638 | init->set_complete_with_arraycopy(); |
1639 | assert(newcopy->is_CheckCastPP(), "sanity" ); |
1640 | assert(newcopy->in(0)->in(0) == init, "dest pinned" ); |
1641 | } |
1642 | // Do not let stores that initialize this object be reordered with |
1643 | // a subsequent store that would make this object accessible by |
1644 | // other threads. |
1645 | // Record what AllocateNode this StoreStore protects so that |
1646 | // escape analysis can go from the MemBarStoreStoreNode to the |
1647 | // AllocateNode and eliminate the MemBarStoreStoreNode if possible |
1648 | // based on the escape status of the AllocateNode. |
1649 | insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); |
1650 | } else { |
1651 | insert_mem_bar(Op_MemBarCPUOrder); |
1652 | } |
1653 | } // original reexecute is set back here |
1654 | |
1655 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
1656 | if (!stopped()) { |
1657 | set_result(newcopy); |
1658 | } |
1659 | clear_upper_avx(); |
1660 | |
1661 | return true; |
1662 | } |
1663 | |
1664 | //------------------------inline_string_getCharsU-------------------------- |
1665 | // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) |
1666 | bool LibraryCallKit::inline_string_getCharsU() { |
1667 | if (too_many_traps(Deoptimization::Reason_intrinsic)) { |
1668 | return false; |
1669 | } |
1670 | |
1671 | // Get the arguments. |
1672 | Node* src = argument(0); |
1673 | Node* src_begin = argument(1); |
1674 | Node* src_end = argument(2); // exclusive offset (i < src_end) |
1675 | Node* dst = argument(3); |
1676 | Node* dst_begin = argument(4); |
1677 | |
1678 | // Check for allocation before we add nodes that would confuse |
1679 | // tightly_coupled_allocation() |
1680 | AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); |
1681 | |
1682 | // Check if a null path was taken unconditionally. |
1683 | src = null_check(src); |
1684 | dst = null_check(dst); |
1685 | if (stopped()) { |
1686 | return true; |
1687 | } |
1688 | |
1689 | // Get length and convert char[] offset to byte[] offset |
1690 | Node* length = _gvn.transform(new SubINode(src_end, src_begin)); |
1691 | src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); |
1692 | |
1693 | // Range checks |
1694 | generate_string_range_check(src, src_begin, length, true); |
1695 | generate_string_range_check(dst, dst_begin, length, false); |
1696 | if (stopped()) { |
1697 | return true; |
1698 | } |
1699 | |
1700 | if (!stopped()) { |
1701 | src = access_resolve(src, ACCESS_READ); |
1702 | dst = access_resolve(dst, ACCESS_WRITE); |
1703 | |
1704 | // Calculate starting addresses. |
1705 | Node* src_start = array_element_address(src, src_begin, T_BYTE); |
1706 | Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); |
1707 | |
1708 | // Check if array addresses are aligned to HeapWordSize |
1709 | const TypeInt* tsrc = gvn().type(src_begin)->is_int(); |
1710 | const TypeInt* tdst = gvn().type(dst_begin)->is_int(); |
1711 | bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && |
1712 | tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); |
1713 | |
1714 | // Figure out which arraycopy runtime method to call (disjoint, uninitialized). |
1715 | const char* copyfunc_name = "arraycopy" ; |
1716 | address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); |
1717 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
1718 | OptoRuntime::fast_arraycopy_Type(), |
1719 | copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, |
1720 | src_start, dst_start, ConvI2X(length) XTOP); |
1721 | // Do not let reads from the cloned object float above the arraycopy. |
1722 | if (alloc != NULL) { |
1723 | if (alloc->maybe_set_complete(&_gvn)) { |
1724 | // "You break it, you buy it." |
1725 | InitializeNode* init = alloc->initialization(); |
1726 | assert(init->is_complete(), "we just did this" ); |
1727 | init->set_complete_with_arraycopy(); |
1728 | assert(dst->is_CheckCastPP(), "sanity" ); |
1729 | assert(dst->in(0)->in(0) == init, "dest pinned" ); |
1730 | } |
1731 | // Do not let stores that initialize this object be reordered with |
1732 | // a subsequent store that would make this object accessible by |
1733 | // other threads. |
1734 | // Record what AllocateNode this StoreStore protects so that |
1735 | // escape analysis can go from the MemBarStoreStoreNode to the |
1736 | // AllocateNode and eliminate the MemBarStoreStoreNode if possible |
1737 | // based on the escape status of the AllocateNode. |
1738 | insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); |
1739 | } else { |
1740 | insert_mem_bar(Op_MemBarCPUOrder); |
1741 | } |
1742 | } |
1743 | |
1744 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
1745 | return true; |
1746 | } |
1747 | |
1748 | //----------------------inline_string_char_access---------------------------- |
1749 | // Store/Load char to/from byte[] array. |
1750 | // static void StringUTF16.putChar(byte[] val, int index, int c) |
1751 | // static char StringUTF16.getChar(byte[] val, int index) |
1752 | bool LibraryCallKit::inline_string_char_access(bool is_store) { |
1753 | Node* value = argument(0); |
1754 | Node* index = argument(1); |
1755 | Node* ch = is_store ? argument(2) : NULL; |
1756 | |
1757 | // This intrinsic accesses byte[] array as char[] array. Computing the offsets |
1758 | // correctly requires matched array shapes. |
1759 | assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), |
1760 | "sanity: byte[] and char[] bases agree" ); |
1761 | assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, |
1762 | "sanity: byte[] and char[] scales agree" ); |
1763 | |
1764 | // Bail when getChar over constants is requested: constant folding would |
1765 | // reject folding mismatched char access over byte[]. A normal inlining for getChar |
1766 | // Java method would constant fold nicely instead. |
1767 | if (!is_store && value->is_Con() && index->is_Con()) { |
1768 | return false; |
1769 | } |
1770 | |
1771 | value = must_be_not_null(value, true); |
1772 | value = access_resolve(value, is_store ? ACCESS_WRITE : ACCESS_READ); |
1773 | |
1774 | Node* adr = array_element_address(value, index, T_CHAR); |
1775 | if (adr->is_top()) { |
1776 | return false; |
1777 | } |
1778 | if (is_store) { |
1779 | access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED); |
1780 | } else { |
1781 | ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD); |
1782 | set_result(ch); |
1783 | } |
1784 | return true; |
1785 | } |
1786 | |
1787 | //--------------------------round_double_node-------------------------------- |
1788 | // Round a double node if necessary. |
1789 | Node* LibraryCallKit::round_double_node(Node* n) { |
1790 | if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) |
1791 | n = _gvn.transform(new RoundDoubleNode(0, n)); |
1792 | return n; |
1793 | } |
1794 | |
1795 | //------------------------------inline_math----------------------------------- |
1796 | // public static double Math.abs(double) |
1797 | // public static double Math.sqrt(double) |
1798 | // public static double Math.log(double) |
1799 | // public static double Math.log10(double) |
1800 | bool LibraryCallKit::inline_double_math(vmIntrinsics::ID id) { |
1801 | Node* arg = round_double_node(argument(0)); |
1802 | Node* n = NULL; |
1803 | switch (id) { |
1804 | case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; |
1805 | case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; |
1806 | default: fatal_unexpected_iid(id); break; |
1807 | } |
1808 | set_result(_gvn.transform(n)); |
1809 | return true; |
1810 | } |
1811 | |
1812 | //------------------------------inline_math----------------------------------- |
1813 | // public static float Math.abs(float) |
1814 | // public static int Math.abs(int) |
1815 | // public static long Math.abs(long) |
1816 | bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { |
1817 | Node* arg = argument(0); |
1818 | Node* n = NULL; |
1819 | switch (id) { |
1820 | case vmIntrinsics::_fabs: n = new AbsFNode( arg); break; |
1821 | case vmIntrinsics::_iabs: n = new AbsINode( arg); break; |
1822 | case vmIntrinsics::_labs: n = new AbsLNode( arg); break; |
1823 | default: fatal_unexpected_iid(id); break; |
1824 | } |
1825 | set_result(_gvn.transform(n)); |
1826 | return true; |
1827 | } |
1828 | |
1829 | //------------------------------runtime_math----------------------------- |
1830 | bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { |
1831 | assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), |
1832 | "must be (DD)D or (D)D type" ); |
1833 | |
1834 | // Inputs |
1835 | Node* a = round_double_node(argument(0)); |
1836 | Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; |
1837 | |
1838 | const TypePtr* no_memory_effects = NULL; |
1839 | Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, |
1840 | no_memory_effects, |
1841 | a, top(), b, b ? top() : NULL); |
1842 | Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); |
1843 | #ifdef ASSERT |
1844 | Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); |
1845 | assert(value_top == top(), "second value must be top" ); |
1846 | #endif |
1847 | |
1848 | set_result(value); |
1849 | return true; |
1850 | } |
1851 | |
1852 | //------------------------------inline_math_native----------------------------- |
1853 | bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { |
1854 | #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) |
1855 | switch (id) { |
1856 | // These intrinsics are not properly supported on all hardware |
1857 | case vmIntrinsics::_dsin: |
1858 | return StubRoutines::dsin() != NULL ? |
1859 | runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin" ) : |
1860 | runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN" ); |
1861 | case vmIntrinsics::_dcos: |
1862 | return StubRoutines::dcos() != NULL ? |
1863 | runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos" ) : |
1864 | runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS" ); |
1865 | case vmIntrinsics::_dtan: |
1866 | return StubRoutines::dtan() != NULL ? |
1867 | runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan" ) : |
1868 | runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN" ); |
1869 | case vmIntrinsics::_dlog: |
1870 | return StubRoutines::dlog() != NULL ? |
1871 | runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog" ) : |
1872 | runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG" ); |
1873 | case vmIntrinsics::_dlog10: |
1874 | return StubRoutines::dlog10() != NULL ? |
1875 | runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10" ) : |
1876 | runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10" ); |
1877 | |
1878 | // These intrinsics are supported on all hardware |
1879 | case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_double_math(id) : false; |
1880 | case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_double_math(id) : false; |
1881 | case vmIntrinsics::_fabs: return Matcher::match_rule_supported(Op_AbsF) ? inline_math(id) : false; |
1882 | case vmIntrinsics::_iabs: return Matcher::match_rule_supported(Op_AbsI) ? inline_math(id) : false; |
1883 | case vmIntrinsics::_labs: return Matcher::match_rule_supported(Op_AbsL) ? inline_math(id) : false; |
1884 | |
1885 | case vmIntrinsics::_dexp: |
1886 | return StubRoutines::dexp() != NULL ? |
1887 | runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp" ) : |
1888 | runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP" ); |
1889 | case vmIntrinsics::_dpow: { |
1890 | Node* exp = round_double_node(argument(2)); |
1891 | const TypeD* d = _gvn.type(exp)->isa_double_constant(); |
1892 | if (d != NULL && d->getd() == 2.0) { |
1893 | // Special case: pow(x, 2.0) => x * x |
1894 | Node* base = round_double_node(argument(0)); |
1895 | set_result(_gvn.transform(new MulDNode(base, base))); |
1896 | return true; |
1897 | } |
1898 | return StubRoutines::dpow() != NULL ? |
1899 | runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow" ) : |
1900 | runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW" ); |
1901 | } |
1902 | #undef FN_PTR |
1903 | |
1904 | // These intrinsics are not yet correctly implemented |
1905 | case vmIntrinsics::_datan2: |
1906 | return false; |
1907 | |
1908 | default: |
1909 | fatal_unexpected_iid(id); |
1910 | return false; |
1911 | } |
1912 | } |
1913 | |
1914 | static bool is_simple_name(Node* n) { |
1915 | return (n->req() == 1 // constant |
1916 | || (n->is_Type() && n->as_Type()->type()->singleton()) |
1917 | || n->is_Proj() // parameter or return value |
1918 | || n->is_Phi() // local of some sort |
1919 | ); |
1920 | } |
1921 | |
1922 | //----------------------------inline_notify-----------------------------------* |
1923 | bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { |
1924 | const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); |
1925 | address func; |
1926 | if (id == vmIntrinsics::_notify) { |
1927 | func = OptoRuntime::monitor_notify_Java(); |
1928 | } else { |
1929 | func = OptoRuntime::monitor_notifyAll_Java(); |
1930 | } |
1931 | Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); |
1932 | make_slow_call_ex(call, env()->Throwable_klass(), false); |
1933 | return true; |
1934 | } |
1935 | |
1936 | |
1937 | //----------------------------inline_min_max----------------------------------- |
1938 | bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { |
1939 | set_result(generate_min_max(id, argument(0), argument(1))); |
1940 | return true; |
1941 | } |
1942 | |
1943 | void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { |
1944 | Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); |
1945 | IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); |
1946 | Node* fast_path = _gvn.transform( new IfFalseNode(check)); |
1947 | Node* slow_path = _gvn.transform( new IfTrueNode(check) ); |
1948 | |
1949 | { |
1950 | PreserveJVMState pjvms(this); |
1951 | PreserveReexecuteState preexecs(this); |
1952 | jvms()->set_should_reexecute(true); |
1953 | |
1954 | set_control(slow_path); |
1955 | set_i_o(i_o()); |
1956 | |
1957 | uncommon_trap(Deoptimization::Reason_intrinsic, |
1958 | Deoptimization::Action_none); |
1959 | } |
1960 | |
1961 | set_control(fast_path); |
1962 | set_result(math); |
1963 | } |
1964 | |
1965 | template <typename OverflowOp> |
1966 | bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { |
1967 | typedef typename OverflowOp::MathOp MathOp; |
1968 | |
1969 | MathOp* mathOp = new MathOp(arg1, arg2); |
1970 | Node* operation = _gvn.transform( mathOp ); |
1971 | Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); |
1972 | inline_math_mathExact(operation, ofcheck); |
1973 | return true; |
1974 | } |
1975 | |
1976 | bool LibraryCallKit::inline_math_addExactI(bool is_increment) { |
1977 | return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); |
1978 | } |
1979 | |
1980 | bool LibraryCallKit::inline_math_addExactL(bool is_increment) { |
1981 | return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); |
1982 | } |
1983 | |
1984 | bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { |
1985 | return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); |
1986 | } |
1987 | |
1988 | bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { |
1989 | return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); |
1990 | } |
1991 | |
1992 | bool LibraryCallKit::inline_math_negateExactI() { |
1993 | return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); |
1994 | } |
1995 | |
1996 | bool LibraryCallKit::inline_math_negateExactL() { |
1997 | return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); |
1998 | } |
1999 | |
2000 | bool LibraryCallKit::inline_math_multiplyExactI() { |
2001 | return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); |
2002 | } |
2003 | |
2004 | bool LibraryCallKit::inline_math_multiplyExactL() { |
2005 | return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); |
2006 | } |
2007 | |
2008 | bool LibraryCallKit::inline_math_multiplyHigh() { |
2009 | set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); |
2010 | return true; |
2011 | } |
2012 | |
2013 | Node* |
2014 | LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { |
2015 | // These are the candidate return value: |
2016 | Node* xvalue = x0; |
2017 | Node* yvalue = y0; |
2018 | |
2019 | if (xvalue == yvalue) { |
2020 | return xvalue; |
2021 | } |
2022 | |
2023 | bool want_max = (id == vmIntrinsics::_max); |
2024 | |
2025 | const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); |
2026 | const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); |
2027 | if (txvalue == NULL || tyvalue == NULL) return top(); |
2028 | // This is not really necessary, but it is consistent with a |
2029 | // hypothetical MaxINode::Value method: |
2030 | int widen = MAX2(txvalue->_widen, tyvalue->_widen); |
2031 | |
2032 | // %%% This folding logic should (ideally) be in a different place. |
2033 | // Some should be inside IfNode, and there to be a more reliable |
2034 | // transformation of ?: style patterns into cmoves. We also want |
2035 | // more powerful optimizations around cmove and min/max. |
2036 | |
2037 | // Try to find a dominating comparison of these guys. |
2038 | // It can simplify the index computation for Arrays.copyOf |
2039 | // and similar uses of System.arraycopy. |
2040 | // First, compute the normalized version of CmpI(x, y). |
2041 | int cmp_op = Op_CmpI; |
2042 | Node* xkey = xvalue; |
2043 | Node* ykey = yvalue; |
2044 | Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); |
2045 | if (ideal_cmpxy->is_Cmp()) { |
2046 | // E.g., if we have CmpI(length - offset, count), |
2047 | // it might idealize to CmpI(length, count + offset) |
2048 | cmp_op = ideal_cmpxy->Opcode(); |
2049 | xkey = ideal_cmpxy->in(1); |
2050 | ykey = ideal_cmpxy->in(2); |
2051 | } |
2052 | |
2053 | // Start by locating any relevant comparisons. |
2054 | Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; |
2055 | Node* cmpxy = NULL; |
2056 | Node* cmpyx = NULL; |
2057 | for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { |
2058 | Node* cmp = start_from->fast_out(k); |
2059 | if (cmp->outcnt() > 0 && // must have prior uses |
2060 | cmp->in(0) == NULL && // must be context-independent |
2061 | cmp->Opcode() == cmp_op) { // right kind of compare |
2062 | if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; |
2063 | if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; |
2064 | } |
2065 | } |
2066 | |
2067 | const int NCMPS = 2; |
2068 | Node* cmps[NCMPS] = { cmpxy, cmpyx }; |
2069 | int cmpn; |
2070 | for (cmpn = 0; cmpn < NCMPS; cmpn++) { |
2071 | if (cmps[cmpn] != NULL) break; // find a result |
2072 | } |
2073 | if (cmpn < NCMPS) { |
2074 | // Look for a dominating test that tells us the min and max. |
2075 | int depth = 0; // Limit search depth for speed |
2076 | Node* dom = control(); |
2077 | for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { |
2078 | if (++depth >= 100) break; |
2079 | Node* ifproj = dom; |
2080 | if (!ifproj->is_Proj()) continue; |
2081 | Node* iff = ifproj->in(0); |
2082 | if (!iff->is_If()) continue; |
2083 | Node* bol = iff->in(1); |
2084 | if (!bol->is_Bool()) continue; |
2085 | Node* cmp = bol->in(1); |
2086 | if (cmp == NULL) continue; |
2087 | for (cmpn = 0; cmpn < NCMPS; cmpn++) |
2088 | if (cmps[cmpn] == cmp) break; |
2089 | if (cmpn == NCMPS) continue; |
2090 | BoolTest::mask btest = bol->as_Bool()->_test._test; |
2091 | if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); |
2092 | if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); |
2093 | // At this point, we know that 'x btest y' is true. |
2094 | switch (btest) { |
2095 | case BoolTest::eq: |
2096 | // They are proven equal, so we can collapse the min/max. |
2097 | // Either value is the answer. Choose the simpler. |
2098 | if (is_simple_name(yvalue) && !is_simple_name(xvalue)) |
2099 | return yvalue; |
2100 | return xvalue; |
2101 | case BoolTest::lt: // x < y |
2102 | case BoolTest::le: // x <= y |
2103 | return (want_max ? yvalue : xvalue); |
2104 | case BoolTest::gt: // x > y |
2105 | case BoolTest::ge: // x >= y |
2106 | return (want_max ? xvalue : yvalue); |
2107 | default: |
2108 | break; |
2109 | } |
2110 | } |
2111 | } |
2112 | |
2113 | // We failed to find a dominating test. |
2114 | // Let's pick a test that might GVN with prior tests. |
2115 | Node* best_bol = NULL; |
2116 | BoolTest::mask best_btest = BoolTest::illegal; |
2117 | for (cmpn = 0; cmpn < NCMPS; cmpn++) { |
2118 | Node* cmp = cmps[cmpn]; |
2119 | if (cmp == NULL) continue; |
2120 | for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { |
2121 | Node* bol = cmp->fast_out(j); |
2122 | if (!bol->is_Bool()) continue; |
2123 | BoolTest::mask btest = bol->as_Bool()->_test._test; |
2124 | if (btest == BoolTest::eq || btest == BoolTest::ne) continue; |
2125 | if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); |
2126 | if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { |
2127 | best_bol = bol->as_Bool(); |
2128 | best_btest = btest; |
2129 | } |
2130 | } |
2131 | } |
2132 | |
2133 | Node* answer_if_true = NULL; |
2134 | Node* answer_if_false = NULL; |
2135 | switch (best_btest) { |
2136 | default: |
2137 | if (cmpxy == NULL) |
2138 | cmpxy = ideal_cmpxy; |
2139 | best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); |
2140 | // and fall through: |
2141 | case BoolTest::lt: // x < y |
2142 | case BoolTest::le: // x <= y |
2143 | answer_if_true = (want_max ? yvalue : xvalue); |
2144 | answer_if_false = (want_max ? xvalue : yvalue); |
2145 | break; |
2146 | case BoolTest::gt: // x > y |
2147 | case BoolTest::ge: // x >= y |
2148 | answer_if_true = (want_max ? xvalue : yvalue); |
2149 | answer_if_false = (want_max ? yvalue : xvalue); |
2150 | break; |
2151 | } |
2152 | |
2153 | jint hi, lo; |
2154 | if (want_max) { |
2155 | // We can sharpen the minimum. |
2156 | hi = MAX2(txvalue->_hi, tyvalue->_hi); |
2157 | lo = MAX2(txvalue->_lo, tyvalue->_lo); |
2158 | } else { |
2159 | // We can sharpen the maximum. |
2160 | hi = MIN2(txvalue->_hi, tyvalue->_hi); |
2161 | lo = MIN2(txvalue->_lo, tyvalue->_lo); |
2162 | } |
2163 | |
2164 | // Use a flow-free graph structure, to avoid creating excess control edges |
2165 | // which could hinder other optimizations. |
2166 | // Since Math.min/max is often used with arraycopy, we want |
2167 | // tightly_coupled_allocation to be able to see beyond min/max expressions. |
2168 | Node* cmov = CMoveNode::make(NULL, best_bol, |
2169 | answer_if_false, answer_if_true, |
2170 | TypeInt::make(lo, hi, widen)); |
2171 | |
2172 | return _gvn.transform(cmov); |
2173 | |
2174 | /* |
2175 | // This is not as desirable as it may seem, since Min and Max |
2176 | // nodes do not have a full set of optimizations. |
2177 | // And they would interfere, anyway, with 'if' optimizations |
2178 | // and with CMoveI canonical forms. |
2179 | switch (id) { |
2180 | case vmIntrinsics::_min: |
2181 | result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; |
2182 | case vmIntrinsics::_max: |
2183 | result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; |
2184 | default: |
2185 | ShouldNotReachHere(); |
2186 | } |
2187 | */ |
2188 | } |
2189 | |
2190 | inline int |
2191 | LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { |
2192 | const TypePtr* base_type = TypePtr::NULL_PTR; |
2193 | if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); |
2194 | if (base_type == NULL) { |
2195 | // Unknown type. |
2196 | return Type::AnyPtr; |
2197 | } else if (base_type == TypePtr::NULL_PTR) { |
2198 | // Since this is a NULL+long form, we have to switch to a rawptr. |
2199 | base = _gvn.transform(new CastX2PNode(offset)); |
2200 | offset = MakeConX(0); |
2201 | return Type::RawPtr; |
2202 | } else if (base_type->base() == Type::RawPtr) { |
2203 | return Type::RawPtr; |
2204 | } else if (base_type->isa_oopptr()) { |
2205 | // Base is never null => always a heap address. |
2206 | if (!TypePtr::NULL_PTR->higher_equal(base_type)) { |
2207 | return Type::OopPtr; |
2208 | } |
2209 | // Offset is small => always a heap address. |
2210 | const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); |
2211 | if (offset_type != NULL && |
2212 | base_type->offset() == 0 && // (should always be?) |
2213 | offset_type->_lo >= 0 && |
2214 | !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { |
2215 | return Type::OopPtr; |
2216 | } else if (type == T_OBJECT) { |
2217 | // off heap access to an oop doesn't make any sense. Has to be on |
2218 | // heap. |
2219 | return Type::OopPtr; |
2220 | } |
2221 | // Otherwise, it might either be oop+off or NULL+addr. |
2222 | return Type::AnyPtr; |
2223 | } else { |
2224 | // No information: |
2225 | return Type::AnyPtr; |
2226 | } |
2227 | } |
2228 | |
2229 | inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) { |
2230 | Node* uncasted_base = base; |
2231 | int kind = classify_unsafe_addr(uncasted_base, offset, type); |
2232 | if (kind == Type::RawPtr) { |
2233 | return basic_plus_adr(top(), uncasted_base, offset); |
2234 | } else if (kind == Type::AnyPtr) { |
2235 | assert(base == uncasted_base, "unexpected base change" ); |
2236 | if (can_cast) { |
2237 | if (!_gvn.type(base)->speculative_maybe_null() && |
2238 | !too_many_traps(Deoptimization::Reason_speculate_null_check)) { |
2239 | // According to profiling, this access is always on |
2240 | // heap. Casting the base to not null and thus avoiding membars |
2241 | // around the access should allow better optimizations |
2242 | Node* null_ctl = top(); |
2243 | base = null_check_oop(base, &null_ctl, true, true, true); |
2244 | assert(null_ctl->is_top(), "no null control here" ); |
2245 | return basic_plus_adr(base, offset); |
2246 | } else if (_gvn.type(base)->speculative_always_null() && |
2247 | !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { |
2248 | // According to profiling, this access is always off |
2249 | // heap. |
2250 | base = null_assert(base); |
2251 | Node* raw_base = _gvn.transform(new CastX2PNode(offset)); |
2252 | offset = MakeConX(0); |
2253 | return basic_plus_adr(top(), raw_base, offset); |
2254 | } |
2255 | } |
2256 | // We don't know if it's an on heap or off heap access. Fall back |
2257 | // to raw memory access. |
2258 | base = access_resolve(base, decorators); |
2259 | Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); |
2260 | return basic_plus_adr(top(), raw, offset); |
2261 | } else { |
2262 | assert(base == uncasted_base, "unexpected base change" ); |
2263 | // We know it's an on heap access so base can't be null |
2264 | if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { |
2265 | base = must_be_not_null(base, true); |
2266 | } |
2267 | return basic_plus_adr(base, offset); |
2268 | } |
2269 | } |
2270 | |
2271 | //--------------------------inline_number_methods----------------------------- |
2272 | // inline int Integer.numberOfLeadingZeros(int) |
2273 | // inline int Long.numberOfLeadingZeros(long) |
2274 | // |
2275 | // inline int Integer.numberOfTrailingZeros(int) |
2276 | // inline int Long.numberOfTrailingZeros(long) |
2277 | // |
2278 | // inline int Integer.bitCount(int) |
2279 | // inline int Long.bitCount(long) |
2280 | // |
2281 | // inline char Character.reverseBytes(char) |
2282 | // inline short Short.reverseBytes(short) |
2283 | // inline int Integer.reverseBytes(int) |
2284 | // inline long Long.reverseBytes(long) |
2285 | bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { |
2286 | Node* arg = argument(0); |
2287 | Node* n = NULL; |
2288 | switch (id) { |
2289 | case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; |
2290 | case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; |
2291 | case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; |
2292 | case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; |
2293 | case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; |
2294 | case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; |
2295 | case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; |
2296 | case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; |
2297 | case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; |
2298 | case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; |
2299 | default: fatal_unexpected_iid(id); break; |
2300 | } |
2301 | set_result(_gvn.transform(n)); |
2302 | return true; |
2303 | } |
2304 | |
2305 | //----------------------------inline_unsafe_access---------------------------- |
2306 | |
2307 | const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { |
2308 | // Attempt to infer a sharper value type from the offset and base type. |
2309 | ciKlass* sharpened_klass = NULL; |
2310 | |
2311 | // See if it is an instance field, with an object type. |
2312 | if (alias_type->field() != NULL) { |
2313 | if (alias_type->field()->type()->is_klass()) { |
2314 | sharpened_klass = alias_type->field()->type()->as_klass(); |
2315 | } |
2316 | } |
2317 | |
2318 | // See if it is a narrow oop array. |
2319 | if (adr_type->isa_aryptr()) { |
2320 | if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { |
2321 | const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); |
2322 | if (elem_type != NULL) { |
2323 | sharpened_klass = elem_type->klass(); |
2324 | } |
2325 | } |
2326 | } |
2327 | |
2328 | // The sharpened class might be unloaded if there is no class loader |
2329 | // contraint in place. |
2330 | if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { |
2331 | const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); |
2332 | |
2333 | #ifndef PRODUCT |
2334 | if (C->print_intrinsics() || C->print_inlining()) { |
2335 | tty->print(" from base type: " ); adr_type->dump(); tty->cr(); |
2336 | tty->print(" sharpened value: " ); tjp->dump(); tty->cr(); |
2337 | } |
2338 | #endif |
2339 | // Sharpen the value type. |
2340 | return tjp; |
2341 | } |
2342 | return NULL; |
2343 | } |
2344 | |
2345 | DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) { |
2346 | switch (kind) { |
2347 | case Relaxed: |
2348 | return MO_UNORDERED; |
2349 | case Opaque: |
2350 | return MO_RELAXED; |
2351 | case Acquire: |
2352 | return MO_ACQUIRE; |
2353 | case Release: |
2354 | return MO_RELEASE; |
2355 | case Volatile: |
2356 | return MO_SEQ_CST; |
2357 | default: |
2358 | ShouldNotReachHere(); |
2359 | return 0; |
2360 | } |
2361 | } |
2362 | |
2363 | bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { |
2364 | if (callee()->is_static()) return false; // caller must have the capability! |
2365 | DecoratorSet decorators = C2_UNSAFE_ACCESS; |
2366 | guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads" ); |
2367 | guarantee( is_store || kind != Release, "Release accesses can be produced only for stores" ); |
2368 | assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type" ); |
2369 | |
2370 | if (type == T_OBJECT || type == T_ARRAY) { |
2371 | decorators |= ON_UNKNOWN_OOP_REF; |
2372 | } |
2373 | |
2374 | if (unaligned) { |
2375 | decorators |= C2_UNALIGNED; |
2376 | } |
2377 | |
2378 | #ifndef PRODUCT |
2379 | { |
2380 | ResourceMark rm; |
2381 | // Check the signatures. |
2382 | ciSignature* sig = callee()->signature(); |
2383 | #ifdef ASSERT |
2384 | if (!is_store) { |
2385 | // Object getReference(Object base, int/long offset), etc. |
2386 | BasicType rtype = sig->return_type()->basic_type(); |
2387 | assert(rtype == type, "getter must return the expected value" ); |
2388 | assert(sig->count() == 2, "oop getter has 2 arguments" ); |
2389 | assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object" ); |
2390 | assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct" ); |
2391 | } else { |
2392 | // void putReference(Object base, int/long offset, Object x), etc. |
2393 | assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value" ); |
2394 | assert(sig->count() == 3, "oop putter has 3 arguments" ); |
2395 | assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object" ); |
2396 | assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct" ); |
2397 | BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); |
2398 | assert(vtype == type, "putter must accept the expected value" ); |
2399 | } |
2400 | #endif // ASSERT |
2401 | } |
2402 | #endif //PRODUCT |
2403 | |
2404 | C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
2405 | |
2406 | Node* receiver = argument(0); // type: oop |
2407 | |
2408 | // Build address expression. |
2409 | Node* adr; |
2410 | Node* heap_base_oop = top(); |
2411 | Node* offset = top(); |
2412 | Node* val; |
2413 | |
2414 | // The base is either a Java object or a value produced by Unsafe.staticFieldBase |
2415 | Node* base = argument(1); // type: oop |
2416 | // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset |
2417 | offset = argument(2); // type: long |
2418 | // We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
2419 | // to be plain byte offsets, which are also the same as those accepted |
2420 | // by oopDesc::field_addr. |
2421 | assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
2422 | "fieldOffset must be byte-scaled" ); |
2423 | // 32-bit machines ignore the high half! |
2424 | offset = ConvL2X(offset); |
2425 | adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed); |
2426 | |
2427 | if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { |
2428 | heap_base_oop = base; |
2429 | } else if (type == T_OBJECT) { |
2430 | return false; // off-heap oop accesses are not supported |
2431 | } |
2432 | |
2433 | // Can base be NULL? Otherwise, always on-heap access. |
2434 | bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base)); |
2435 | |
2436 | if (!can_access_non_heap) { |
2437 | decorators |= IN_HEAP; |
2438 | } |
2439 | |
2440 | val = is_store ? argument(4) : NULL; |
2441 | |
2442 | const TypePtr* adr_type = _gvn.type(adr)->isa_ptr(); |
2443 | if (adr_type == TypePtr::NULL_PTR) { |
2444 | return false; // off-heap access with zero address |
2445 | } |
2446 | |
2447 | // Try to categorize the address. |
2448 | Compile::AliasType* alias_type = C->alias_type(adr_type); |
2449 | assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here" ); |
2450 | |
2451 | if (alias_type->adr_type() == TypeInstPtr::KLASS || |
2452 | alias_type->adr_type() == TypeAryPtr::RANGE) { |
2453 | return false; // not supported |
2454 | } |
2455 | |
2456 | bool mismatched = false; |
2457 | BasicType bt = alias_type->basic_type(); |
2458 | if (bt != T_ILLEGAL) { |
2459 | assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access" ); |
2460 | if (bt == T_BYTE && adr_type->isa_aryptr()) { |
2461 | // Alias type doesn't differentiate between byte[] and boolean[]). |
2462 | // Use address type to get the element type. |
2463 | bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); |
2464 | } |
2465 | if (bt == T_ARRAY || bt == T_NARROWOOP) { |
2466 | // accessing an array field with getReference is not a mismatch |
2467 | bt = T_OBJECT; |
2468 | } |
2469 | if ((bt == T_OBJECT) != (type == T_OBJECT)) { |
2470 | // Don't intrinsify mismatched object accesses |
2471 | return false; |
2472 | } |
2473 | mismatched = (bt != type); |
2474 | } else if (alias_type->adr_type()->isa_oopptr()) { |
2475 | mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched |
2476 | } |
2477 | |
2478 | assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched" ); |
2479 | |
2480 | if (mismatched) { |
2481 | decorators |= C2_MISMATCHED; |
2482 | } |
2483 | |
2484 | // First guess at the value type. |
2485 | const Type *value_type = Type::get_const_basic_type(type); |
2486 | |
2487 | // Figure out the memory ordering. |
2488 | decorators |= mo_decorator_for_access_kind(kind); |
2489 | |
2490 | if (!is_store && type == T_OBJECT) { |
2491 | const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); |
2492 | if (tjp != NULL) { |
2493 | value_type = tjp; |
2494 | } |
2495 | } |
2496 | |
2497 | receiver = null_check(receiver); |
2498 | if (stopped()) { |
2499 | return true; |
2500 | } |
2501 | // Heap pointers get a null-check from the interpreter, |
2502 | // as a courtesy. However, this is not guaranteed by Unsafe, |
2503 | // and it is not possible to fully distinguish unintended nulls |
2504 | // from intended ones in this API. |
2505 | |
2506 | if (!is_store) { |
2507 | Node* p = NULL; |
2508 | // Try to constant fold a load from a constant field |
2509 | ciField* field = alias_type->field(); |
2510 | if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { |
2511 | // final or stable field |
2512 | p = make_constant_from_field(field, heap_base_oop); |
2513 | } |
2514 | |
2515 | if (p == NULL) { // Could not constant fold the load |
2516 | p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators); |
2517 | // Normalize the value returned by getBoolean in the following cases |
2518 | if (type == T_BOOLEAN && |
2519 | (mismatched || |
2520 | heap_base_oop == top() || // - heap_base_oop is NULL or |
2521 | (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL |
2522 | // and the unsafe access is made to large offset |
2523 | // (i.e., larger than the maximum offset necessary for any |
2524 | // field access) |
2525 | ) { |
2526 | IdealKit ideal = IdealKit(this); |
2527 | #define __ ideal. |
2528 | IdealVariable normalized_result(ideal); |
2529 | __ declarations_done(); |
2530 | __ set(normalized_result, p); |
2531 | __ if_then(p, BoolTest::ne, ideal.ConI(0)); |
2532 | __ set(normalized_result, ideal.ConI(1)); |
2533 | ideal.end_if(); |
2534 | final_sync(ideal); |
2535 | p = __ value(normalized_result); |
2536 | #undef __ |
2537 | } |
2538 | } |
2539 | if (type == T_ADDRESS) { |
2540 | p = gvn().transform(new CastP2XNode(NULL, p)); |
2541 | p = ConvX2UL(p); |
2542 | } |
2543 | // The load node has the control of the preceding MemBarCPUOrder. All |
2544 | // following nodes will have the control of the MemBarCPUOrder inserted at |
2545 | // the end of this method. So, pushing the load onto the stack at a later |
2546 | // point is fine. |
2547 | set_result(p); |
2548 | } else { |
2549 | if (bt == T_ADDRESS) { |
2550 | // Repackage the long as a pointer. |
2551 | val = ConvL2X(val); |
2552 | val = gvn().transform(new CastX2PNode(val)); |
2553 | } |
2554 | access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators); |
2555 | } |
2556 | |
2557 | return true; |
2558 | } |
2559 | |
2560 | //----------------------------inline_unsafe_load_store---------------------------- |
2561 | // This method serves a couple of different customers (depending on LoadStoreKind): |
2562 | // |
2563 | // LS_cmp_swap: |
2564 | // |
2565 | // boolean compareAndSetReference(Object o, long offset, Object expected, Object x); |
2566 | // boolean compareAndSetInt( Object o, long offset, int expected, int x); |
2567 | // boolean compareAndSetLong( Object o, long offset, long expected, long x); |
2568 | // |
2569 | // LS_cmp_swap_weak: |
2570 | // |
2571 | // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x); |
2572 | // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x); |
2573 | // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x); |
2574 | // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x); |
2575 | // |
2576 | // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); |
2577 | // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); |
2578 | // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); |
2579 | // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); |
2580 | // |
2581 | // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); |
2582 | // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); |
2583 | // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); |
2584 | // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); |
2585 | // |
2586 | // LS_cmp_exchange: |
2587 | // |
2588 | // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x); |
2589 | // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x); |
2590 | // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x); |
2591 | // |
2592 | // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); |
2593 | // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); |
2594 | // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); |
2595 | // |
2596 | // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); |
2597 | // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); |
2598 | // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); |
2599 | // |
2600 | // LS_get_add: |
2601 | // |
2602 | // int getAndAddInt( Object o, long offset, int delta) |
2603 | // long getAndAddLong(Object o, long offset, long delta) |
2604 | // |
2605 | // LS_get_set: |
2606 | // |
2607 | // int getAndSet(Object o, long offset, int newValue) |
2608 | // long getAndSet(Object o, long offset, long newValue) |
2609 | // Object getAndSet(Object o, long offset, Object newValue) |
2610 | // |
2611 | bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { |
2612 | // This basic scheme here is the same as inline_unsafe_access, but |
2613 | // differs in enough details that combining them would make the code |
2614 | // overly confusing. (This is a true fact! I originally combined |
2615 | // them, but even I was confused by it!) As much code/comments as |
2616 | // possible are retained from inline_unsafe_access though to make |
2617 | // the correspondences clearer. - dl |
2618 | |
2619 | if (callee()->is_static()) return false; // caller must have the capability! |
2620 | |
2621 | DecoratorSet decorators = C2_UNSAFE_ACCESS; |
2622 | decorators |= mo_decorator_for_access_kind(access_kind); |
2623 | |
2624 | #ifndef PRODUCT |
2625 | BasicType rtype; |
2626 | { |
2627 | ResourceMark rm; |
2628 | // Check the signatures. |
2629 | ciSignature* sig = callee()->signature(); |
2630 | rtype = sig->return_type()->basic_type(); |
2631 | switch(kind) { |
2632 | case LS_get_add: |
2633 | case LS_get_set: { |
2634 | // Check the signatures. |
2635 | #ifdef ASSERT |
2636 | assert(rtype == type, "get and set must return the expected type" ); |
2637 | assert(sig->count() == 3, "get and set has 3 arguments" ); |
2638 | assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object" ); |
2639 | assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long" ); |
2640 | assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta" ); |
2641 | assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation" ); |
2642 | #endif // ASSERT |
2643 | break; |
2644 | } |
2645 | case LS_cmp_swap: |
2646 | case LS_cmp_swap_weak: { |
2647 | // Check the signatures. |
2648 | #ifdef ASSERT |
2649 | assert(rtype == T_BOOLEAN, "CAS must return boolean" ); |
2650 | assert(sig->count() == 4, "CAS has 4 arguments" ); |
2651 | assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object" ); |
2652 | assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long" ); |
2653 | #endif // ASSERT |
2654 | break; |
2655 | } |
2656 | case LS_cmp_exchange: { |
2657 | // Check the signatures. |
2658 | #ifdef ASSERT |
2659 | assert(rtype == type, "CAS must return the expected type" ); |
2660 | assert(sig->count() == 4, "CAS has 4 arguments" ); |
2661 | assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object" ); |
2662 | assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long" ); |
2663 | #endif // ASSERT |
2664 | break; |
2665 | } |
2666 | default: |
2667 | ShouldNotReachHere(); |
2668 | } |
2669 | } |
2670 | #endif //PRODUCT |
2671 | |
2672 | C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
2673 | |
2674 | // Get arguments: |
2675 | Node* receiver = NULL; |
2676 | Node* base = NULL; |
2677 | Node* offset = NULL; |
2678 | Node* oldval = NULL; |
2679 | Node* newval = NULL; |
2680 | switch(kind) { |
2681 | case LS_cmp_swap: |
2682 | case LS_cmp_swap_weak: |
2683 | case LS_cmp_exchange: { |
2684 | const bool two_slot_type = type2size[type] == 2; |
2685 | receiver = argument(0); // type: oop |
2686 | base = argument(1); // type: oop |
2687 | offset = argument(2); // type: long |
2688 | oldval = argument(4); // type: oop, int, or long |
2689 | newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long |
2690 | break; |
2691 | } |
2692 | case LS_get_add: |
2693 | case LS_get_set: { |
2694 | receiver = argument(0); // type: oop |
2695 | base = argument(1); // type: oop |
2696 | offset = argument(2); // type: long |
2697 | oldval = NULL; |
2698 | newval = argument(4); // type: oop, int, or long |
2699 | break; |
2700 | } |
2701 | default: |
2702 | ShouldNotReachHere(); |
2703 | } |
2704 | |
2705 | // Build field offset expression. |
2706 | // We currently rely on the cookies produced by Unsafe.xxxFieldOffset |
2707 | // to be plain byte offsets, which are also the same as those accepted |
2708 | // by oopDesc::field_addr. |
2709 | assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled" ); |
2710 | // 32-bit machines ignore the high half of long offsets |
2711 | offset = ConvL2X(offset); |
2712 | Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false); |
2713 | const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); |
2714 | |
2715 | Compile::AliasType* alias_type = C->alias_type(adr_type); |
2716 | BasicType bt = alias_type->basic_type(); |
2717 | if (bt != T_ILLEGAL && |
2718 | ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { |
2719 | // Don't intrinsify mismatched object accesses. |
2720 | return false; |
2721 | } |
2722 | |
2723 | // For CAS, unlike inline_unsafe_access, there seems no point in |
2724 | // trying to refine types. Just use the coarse types here. |
2725 | assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here" ); |
2726 | const Type *value_type = Type::get_const_basic_type(type); |
2727 | |
2728 | switch (kind) { |
2729 | case LS_get_set: |
2730 | case LS_cmp_exchange: { |
2731 | if (type == T_OBJECT) { |
2732 | const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); |
2733 | if (tjp != NULL) { |
2734 | value_type = tjp; |
2735 | } |
2736 | } |
2737 | break; |
2738 | } |
2739 | case LS_cmp_swap: |
2740 | case LS_cmp_swap_weak: |
2741 | case LS_get_add: |
2742 | break; |
2743 | default: |
2744 | ShouldNotReachHere(); |
2745 | } |
2746 | |
2747 | // Null check receiver. |
2748 | receiver = null_check(receiver); |
2749 | if (stopped()) { |
2750 | return true; |
2751 | } |
2752 | |
2753 | int alias_idx = C->get_alias_index(adr_type); |
2754 | |
2755 | if (type == T_OBJECT || type == T_ARRAY) { |
2756 | decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF; |
2757 | |
2758 | // Transformation of a value which could be NULL pointer (CastPP #NULL) |
2759 | // could be delayed during Parse (for example, in adjust_map_after_if()). |
2760 | // Execute transformation here to avoid barrier generation in such case. |
2761 | if (_gvn.type(newval) == TypePtr::NULL_PTR) |
2762 | newval = _gvn.makecon(TypePtr::NULL_PTR); |
2763 | |
2764 | if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) { |
2765 | // Refine the value to a null constant, when it is known to be null |
2766 | oldval = _gvn.makecon(TypePtr::NULL_PTR); |
2767 | } |
2768 | } |
2769 | |
2770 | Node* result = NULL; |
2771 | switch (kind) { |
2772 | case LS_cmp_exchange: { |
2773 | result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx, |
2774 | oldval, newval, value_type, type, decorators); |
2775 | break; |
2776 | } |
2777 | case LS_cmp_swap_weak: |
2778 | decorators |= C2_WEAK_CMPXCHG; |
2779 | case LS_cmp_swap: { |
2780 | result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx, |
2781 | oldval, newval, value_type, type, decorators); |
2782 | break; |
2783 | } |
2784 | case LS_get_set: { |
2785 | result = access_atomic_xchg_at(base, adr, adr_type, alias_idx, |
2786 | newval, value_type, type, decorators); |
2787 | break; |
2788 | } |
2789 | case LS_get_add: { |
2790 | result = access_atomic_add_at(base, adr, adr_type, alias_idx, |
2791 | newval, value_type, type, decorators); |
2792 | break; |
2793 | } |
2794 | default: |
2795 | ShouldNotReachHere(); |
2796 | } |
2797 | |
2798 | assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match" ); |
2799 | set_result(result); |
2800 | return true; |
2801 | } |
2802 | |
2803 | bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { |
2804 | // Regardless of form, don't allow previous ld/st to move down, |
2805 | // then issue acquire, release, or volatile mem_bar. |
2806 | insert_mem_bar(Op_MemBarCPUOrder); |
2807 | switch(id) { |
2808 | case vmIntrinsics::_loadFence: |
2809 | insert_mem_bar(Op_LoadFence); |
2810 | return true; |
2811 | case vmIntrinsics::_storeFence: |
2812 | insert_mem_bar(Op_StoreFence); |
2813 | return true; |
2814 | case vmIntrinsics::_fullFence: |
2815 | insert_mem_bar(Op_MemBarVolatile); |
2816 | return true; |
2817 | default: |
2818 | fatal_unexpected_iid(id); |
2819 | return false; |
2820 | } |
2821 | } |
2822 | |
2823 | bool LibraryCallKit::inline_onspinwait() { |
2824 | insert_mem_bar(Op_OnSpinWait); |
2825 | return true; |
2826 | } |
2827 | |
2828 | bool LibraryCallKit::klass_needs_init_guard(Node* kls) { |
2829 | if (!kls->is_Con()) { |
2830 | return true; |
2831 | } |
2832 | const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); |
2833 | if (klsptr == NULL) { |
2834 | return true; |
2835 | } |
2836 | ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); |
2837 | // don't need a guard for a klass that is already initialized |
2838 | return !ik->is_initialized(); |
2839 | } |
2840 | |
2841 | //----------------------------inline_unsafe_allocate--------------------------- |
2842 | // public native Object Unsafe.allocateInstance(Class<?> cls); |
2843 | bool LibraryCallKit::inline_unsafe_allocate() { |
2844 | if (callee()->is_static()) return false; // caller must have the capability! |
2845 | |
2846 | null_check_receiver(); // null-check, then ignore |
2847 | Node* cls = null_check(argument(1)); |
2848 | if (stopped()) return true; |
2849 | |
2850 | Node* kls = load_klass_from_mirror(cls, false, NULL, 0); |
2851 | kls = null_check(kls); |
2852 | if (stopped()) return true; // argument was like int.class |
2853 | |
2854 | Node* test = NULL; |
2855 | if (LibraryCallKit::klass_needs_init_guard(kls)) { |
2856 | // Note: The argument might still be an illegal value like |
2857 | // Serializable.class or Object[].class. The runtime will handle it. |
2858 | // But we must make an explicit check for initialization. |
2859 | Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); |
2860 | // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler |
2861 | // can generate code to load it as unsigned byte. |
2862 | Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); |
2863 | Node* bits = intcon(InstanceKlass::fully_initialized); |
2864 | test = _gvn.transform(new SubINode(inst, bits)); |
2865 | // The 'test' is non-zero if we need to take a slow path. |
2866 | } |
2867 | |
2868 | Node* obj = new_instance(kls, test); |
2869 | set_result(obj); |
2870 | return true; |
2871 | } |
2872 | |
2873 | //------------------------inline_native_time_funcs-------------- |
2874 | // inline code for System.currentTimeMillis() and System.nanoTime() |
2875 | // these have the same type and signature |
2876 | bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { |
2877 | const TypeFunc* tf = OptoRuntime::void_long_Type(); |
2878 | const TypePtr* no_memory_effects = NULL; |
2879 | Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); |
2880 | Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); |
2881 | #ifdef ASSERT |
2882 | Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); |
2883 | assert(value_top == top(), "second value must be top" ); |
2884 | #endif |
2885 | set_result(value); |
2886 | return true; |
2887 | } |
2888 | |
2889 | #ifdef JFR_HAVE_INTRINSICS |
2890 | |
2891 | /* |
2892 | * oop -> myklass |
2893 | * myklass->trace_id |= USED |
2894 | * return myklass->trace_id & ~0x3 |
2895 | */ |
2896 | bool LibraryCallKit::inline_native_classID() { |
2897 | Node* cls = null_check(argument(0), T_OBJECT); |
2898 | Node* kls = load_klass_from_mirror(cls, false, NULL, 0); |
2899 | kls = null_check(kls, T_OBJECT); |
2900 | |
2901 | ByteSize offset = KLASS_TRACE_ID_OFFSET; |
2902 | Node* insp = basic_plus_adr(kls, in_bytes(offset)); |
2903 | Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); |
2904 | |
2905 | Node* clsused = longcon(0x01l); // set the class bit |
2906 | Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); |
2907 | const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); |
2908 | store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); |
2909 | |
2910 | #ifdef TRACE_ID_META_BITS |
2911 | Node* mbits = longcon(~TRACE_ID_META_BITS); |
2912 | tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); |
2913 | #endif |
2914 | #ifdef TRACE_ID_SHIFT |
2915 | Node* cbits = intcon(TRACE_ID_SHIFT); |
2916 | tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); |
2917 | #endif |
2918 | |
2919 | set_result(tvalue); |
2920 | return true; |
2921 | |
2922 | } |
2923 | |
2924 | bool LibraryCallKit::inline_native_getEventWriter() { |
2925 | Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); |
2926 | |
2927 | Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, |
2928 | in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)); |
2929 | |
2930 | Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); |
2931 | |
2932 | Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); |
2933 | Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); |
2934 | |
2935 | IfNode* iff_jobj_null = |
2936 | create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); |
2937 | |
2938 | enum { _normal_path = 1, |
2939 | _null_path = 2, |
2940 | PATH_LIMIT }; |
2941 | |
2942 | RegionNode* result_rgn = new RegionNode(PATH_LIMIT); |
2943 | PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM); |
2944 | |
2945 | Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); |
2946 | result_rgn->init_req(_null_path, jobj_is_null); |
2947 | result_val->init_req(_null_path, null()); |
2948 | |
2949 | Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); |
2950 | set_control(jobj_is_not_null); |
2951 | Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT, |
2952 | IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD); |
2953 | result_rgn->init_req(_normal_path, control()); |
2954 | result_val->init_req(_normal_path, res); |
2955 | |
2956 | set_result(result_rgn, result_val); |
2957 | |
2958 | return true; |
2959 | } |
2960 | |
2961 | #endif // JFR_HAVE_INTRINSICS |
2962 | |
2963 | //------------------------inline_native_currentThread------------------ |
2964 | bool LibraryCallKit::inline_native_currentThread() { |
2965 | Node* junk = NULL; |
2966 | set_result(generate_current_thread(junk)); |
2967 | return true; |
2968 | } |
2969 | |
2970 | //------------------------inline_native_isInterrupted------------------ |
2971 | // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); |
2972 | bool LibraryCallKit::inline_native_isInterrupted() { |
2973 | // Add a fast path to t.isInterrupted(clear_int): |
2974 | // (t == Thread.current() && |
2975 | // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) |
2976 | // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) |
2977 | // So, in the common case that the interrupt bit is false, |
2978 | // we avoid making a call into the VM. Even if the interrupt bit |
2979 | // is true, if the clear_int argument is false, we avoid the VM call. |
2980 | // However, if the receiver is not currentThread, we must call the VM, |
2981 | // because there must be some locking done around the operation. |
2982 | |
2983 | // We only go to the fast case code if we pass two guards. |
2984 | // Paths which do not pass are accumulated in the slow_region. |
2985 | |
2986 | enum { |
2987 | no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted |
2988 | no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int |
2989 | slow_result_path = 3, // slow path: t.isInterrupted(clear_int) |
2990 | PATH_LIMIT |
2991 | }; |
2992 | |
2993 | // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag |
2994 | // out of the function. |
2995 | insert_mem_bar(Op_MemBarCPUOrder); |
2996 | |
2997 | RegionNode* result_rgn = new RegionNode(PATH_LIMIT); |
2998 | PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); |
2999 | |
3000 | RegionNode* slow_region = new RegionNode(1); |
3001 | record_for_igvn(slow_region); |
3002 | |
3003 | // (a) Receiving thread must be the current thread. |
3004 | Node* rec_thr = argument(0); |
3005 | Node* tls_ptr = NULL; |
3006 | Node* cur_thr = generate_current_thread(tls_ptr); |
3007 | |
3008 | // Resolve oops to stable for CmpP below. |
3009 | cur_thr = access_resolve(cur_thr, 0); |
3010 | rec_thr = access_resolve(rec_thr, 0); |
3011 | |
3012 | Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); |
3013 | Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); |
3014 | |
3015 | generate_slow_guard(bol_thr, slow_region); |
3016 | |
3017 | // (b) Interrupt bit on TLS must be false. |
3018 | Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); |
3019 | Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); |
3020 | p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); |
3021 | |
3022 | // Set the control input on the field _interrupted read to prevent it floating up. |
3023 | Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); |
3024 | Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); |
3025 | Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); |
3026 | |
3027 | IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); |
3028 | |
3029 | // First fast path: if (!TLS._interrupted) return false; |
3030 | Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); |
3031 | result_rgn->init_req(no_int_result_path, false_bit); |
3032 | result_val->init_req(no_int_result_path, intcon(0)); |
3033 | |
3034 | // drop through to next case |
3035 | set_control( _gvn.transform(new IfTrueNode(iff_bit))); |
3036 | |
3037 | #ifndef _WINDOWS |
3038 | // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. |
3039 | Node* clr_arg = argument(1); |
3040 | Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); |
3041 | Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); |
3042 | IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); |
3043 | |
3044 | // Second fast path: ... else if (!clear_int) return true; |
3045 | Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); |
3046 | result_rgn->init_req(no_clear_result_path, false_arg); |
3047 | result_val->init_req(no_clear_result_path, intcon(1)); |
3048 | |
3049 | // drop through to next case |
3050 | set_control( _gvn.transform(new IfTrueNode(iff_arg))); |
3051 | #else |
3052 | // To return true on Windows you must read the _interrupted field |
3053 | // and check the event state i.e. take the slow path. |
3054 | #endif // _WINDOWS |
3055 | |
3056 | // (d) Otherwise, go to the slow path. |
3057 | slow_region->add_req(control()); |
3058 | set_control( _gvn.transform(slow_region)); |
3059 | |
3060 | if (stopped()) { |
3061 | // There is no slow path. |
3062 | result_rgn->init_req(slow_result_path, top()); |
3063 | result_val->init_req(slow_result_path, top()); |
3064 | } else { |
3065 | // non-virtual because it is a private non-static |
3066 | CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); |
3067 | |
3068 | Node* slow_val = set_results_for_java_call(slow_call); |
3069 | // this->control() comes from set_results_for_java_call |
3070 | |
3071 | Node* fast_io = slow_call->in(TypeFunc::I_O); |
3072 | Node* fast_mem = slow_call->in(TypeFunc::Memory); |
3073 | |
3074 | // These two phis are pre-filled with copies of of the fast IO and Memory |
3075 | PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); |
3076 | PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); |
3077 | |
3078 | result_rgn->init_req(slow_result_path, control()); |
3079 | result_io ->init_req(slow_result_path, i_o()); |
3080 | result_mem->init_req(slow_result_path, reset_memory()); |
3081 | result_val->init_req(slow_result_path, slow_val); |
3082 | |
3083 | set_all_memory(_gvn.transform(result_mem)); |
3084 | set_i_o( _gvn.transform(result_io)); |
3085 | } |
3086 | |
3087 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
3088 | set_result(result_rgn, result_val); |
3089 | return true; |
3090 | } |
3091 | |
3092 | //---------------------------load_mirror_from_klass---------------------------- |
3093 | // Given a klass oop, load its java mirror (a java.lang.Class oop). |
3094 | Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { |
3095 | Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); |
3096 | Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); |
3097 | // mirror = ((OopHandle)mirror)->resolve(); |
3098 | return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); |
3099 | } |
3100 | |
3101 | //-----------------------load_klass_from_mirror_common------------------------- |
3102 | // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. |
3103 | // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), |
3104 | // and branch to the given path on the region. |
3105 | // If never_see_null, take an uncommon trap on null, so we can optimistically |
3106 | // compile for the non-null case. |
3107 | // If the region is NULL, force never_see_null = true. |
3108 | Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, |
3109 | bool never_see_null, |
3110 | RegionNode* region, |
3111 | int null_path, |
3112 | int offset) { |
3113 | if (region == NULL) never_see_null = true; |
3114 | Node* p = basic_plus_adr(mirror, offset); |
3115 | const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; |
3116 | Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); |
3117 | Node* null_ctl = top(); |
3118 | kls = null_check_oop(kls, &null_ctl, never_see_null); |
3119 | if (region != NULL) { |
3120 | // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). |
3121 | region->init_req(null_path, null_ctl); |
3122 | } else { |
3123 | assert(null_ctl == top(), "no loose ends" ); |
3124 | } |
3125 | return kls; |
3126 | } |
3127 | |
3128 | //--------------------(inline_native_Class_query helpers)--------------------- |
3129 | // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. |
3130 | // Fall through if (mods & mask) == bits, take the guard otherwise. |
3131 | Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { |
3132 | // Branch around if the given klass has the given modifier bit set. |
3133 | // Like generate_guard, adds a new path onto the region. |
3134 | Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); |
3135 | Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); |
3136 | Node* mask = intcon(modifier_mask); |
3137 | Node* bits = intcon(modifier_bits); |
3138 | Node* mbit = _gvn.transform(new AndINode(mods, mask)); |
3139 | Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); |
3140 | Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); |
3141 | return generate_fair_guard(bol, region); |
3142 | } |
3143 | Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { |
3144 | return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); |
3145 | } |
3146 | |
3147 | //-------------------------inline_native_Class_query------------------- |
3148 | bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { |
3149 | const Type* return_type = TypeInt::BOOL; |
3150 | Node* prim_return_value = top(); // what happens if it's a primitive class? |
3151 | bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
3152 | bool expect_prim = false; // most of these guys expect to work on refs |
3153 | |
3154 | enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; |
3155 | |
3156 | Node* mirror = argument(0); |
3157 | Node* obj = top(); |
3158 | |
3159 | switch (id) { |
3160 | case vmIntrinsics::_isInstance: |
3161 | // nothing is an instance of a primitive type |
3162 | prim_return_value = intcon(0); |
3163 | obj = argument(1); |
3164 | break; |
3165 | case vmIntrinsics::_getModifiers: |
3166 | prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); |
3167 | assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line" ); |
3168 | return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); |
3169 | break; |
3170 | case vmIntrinsics::_isInterface: |
3171 | prim_return_value = intcon(0); |
3172 | break; |
3173 | case vmIntrinsics::_isArray: |
3174 | prim_return_value = intcon(0); |
3175 | expect_prim = true; // cf. ObjectStreamClass.getClassSignature |
3176 | break; |
3177 | case vmIntrinsics::_isPrimitive: |
3178 | prim_return_value = intcon(1); |
3179 | expect_prim = true; // obviously |
3180 | break; |
3181 | case vmIntrinsics::_getSuperclass: |
3182 | prim_return_value = null(); |
3183 | return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); |
3184 | break; |
3185 | case vmIntrinsics::_getClassAccessFlags: |
3186 | prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); |
3187 | return_type = TypeInt::INT; // not bool! 6297094 |
3188 | break; |
3189 | default: |
3190 | fatal_unexpected_iid(id); |
3191 | break; |
3192 | } |
3193 | |
3194 | const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); |
3195 | if (mirror_con == NULL) return false; // cannot happen? |
3196 | |
3197 | #ifndef PRODUCT |
3198 | if (C->print_intrinsics() || C->print_inlining()) { |
3199 | ciType* k = mirror_con->java_mirror_type(); |
3200 | if (k) { |
3201 | tty->print("Inlining %s on constant Class " , vmIntrinsics::name_at(intrinsic_id())); |
3202 | k->print_name(); |
3203 | tty->cr(); |
3204 | } |
3205 | } |
3206 | #endif |
3207 | |
3208 | // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). |
3209 | RegionNode* region = new RegionNode(PATH_LIMIT); |
3210 | record_for_igvn(region); |
3211 | PhiNode* phi = new PhiNode(region, return_type); |
3212 | |
3213 | // The mirror will never be null of Reflection.getClassAccessFlags, however |
3214 | // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE |
3215 | // if it is. See bug 4774291. |
3216 | |
3217 | // For Reflection.getClassAccessFlags(), the null check occurs in |
3218 | // the wrong place; see inline_unsafe_access(), above, for a similar |
3219 | // situation. |
3220 | mirror = null_check(mirror); |
3221 | // If mirror or obj is dead, only null-path is taken. |
3222 | if (stopped()) return true; |
3223 | |
3224 | if (expect_prim) never_see_null = false; // expect nulls (meaning prims) |
3225 | |
3226 | // Now load the mirror's klass metaobject, and null-check it. |
3227 | // Side-effects region with the control path if the klass is null. |
3228 | Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); |
3229 | // If kls is null, we have a primitive mirror. |
3230 | phi->init_req(_prim_path, prim_return_value); |
3231 | if (stopped()) { set_result(region, phi); return true; } |
3232 | bool safe_for_replace = (region->in(_prim_path) == top()); |
3233 | |
3234 | Node* p; // handy temp |
3235 | Node* null_ctl; |
3236 | |
3237 | // Now that we have the non-null klass, we can perform the real query. |
3238 | // For constant classes, the query will constant-fold in LoadNode::Value. |
3239 | Node* query_value = top(); |
3240 | switch (id) { |
3241 | case vmIntrinsics::_isInstance: |
3242 | // nothing is an instance of a primitive type |
3243 | query_value = gen_instanceof(obj, kls, safe_for_replace); |
3244 | break; |
3245 | |
3246 | case vmIntrinsics::_getModifiers: |
3247 | p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); |
3248 | query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); |
3249 | break; |
3250 | |
3251 | case vmIntrinsics::_isInterface: |
3252 | // (To verify this code sequence, check the asserts in JVM_IsInterface.) |
3253 | if (generate_interface_guard(kls, region) != NULL) |
3254 | // A guard was added. If the guard is taken, it was an interface. |
3255 | phi->add_req(intcon(1)); |
3256 | // If we fall through, it's a plain class. |
3257 | query_value = intcon(0); |
3258 | break; |
3259 | |
3260 | case vmIntrinsics::_isArray: |
3261 | // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) |
3262 | if (generate_array_guard(kls, region) != NULL) |
3263 | // A guard was added. If the guard is taken, it was an array. |
3264 | phi->add_req(intcon(1)); |
3265 | // If we fall through, it's a plain class. |
3266 | query_value = intcon(0); |
3267 | break; |
3268 | |
3269 | case vmIntrinsics::_isPrimitive: |
3270 | query_value = intcon(0); // "normal" path produces false |
3271 | break; |
3272 | |
3273 | case vmIntrinsics::_getSuperclass: |
3274 | // The rules here are somewhat unfortunate, but we can still do better |
3275 | // with random logic than with a JNI call. |
3276 | // Interfaces store null or Object as _super, but must report null. |
3277 | // Arrays store an intermediate super as _super, but must report Object. |
3278 | // Other types can report the actual _super. |
3279 | // (To verify this code sequence, check the asserts in JVM_IsInterface.) |
3280 | if (generate_interface_guard(kls, region) != NULL) |
3281 | // A guard was added. If the guard is taken, it was an interface. |
3282 | phi->add_req(null()); |
3283 | if (generate_array_guard(kls, region) != NULL) |
3284 | // A guard was added. If the guard is taken, it was an array. |
3285 | phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); |
3286 | // If we fall through, it's a plain class. Get its _super. |
3287 | p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); |
3288 | kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); |
3289 | null_ctl = top(); |
3290 | kls = null_check_oop(kls, &null_ctl); |
3291 | if (null_ctl != top()) { |
3292 | // If the guard is taken, Object.superClass is null (both klass and mirror). |
3293 | region->add_req(null_ctl); |
3294 | phi ->add_req(null()); |
3295 | } |
3296 | if (!stopped()) { |
3297 | query_value = load_mirror_from_klass(kls); |
3298 | } |
3299 | break; |
3300 | |
3301 | case vmIntrinsics::_getClassAccessFlags: |
3302 | p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); |
3303 | query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); |
3304 | break; |
3305 | |
3306 | default: |
3307 | fatal_unexpected_iid(id); |
3308 | break; |
3309 | } |
3310 | |
3311 | // Fall-through is the normal case of a query to a real class. |
3312 | phi->init_req(1, query_value); |
3313 | region->init_req(1, control()); |
3314 | |
3315 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
3316 | set_result(region, phi); |
3317 | return true; |
3318 | } |
3319 | |
3320 | //-------------------------inline_Class_cast------------------- |
3321 | bool LibraryCallKit::inline_Class_cast() { |
3322 | Node* mirror = argument(0); // Class |
3323 | Node* obj = argument(1); |
3324 | const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); |
3325 | if (mirror_con == NULL) { |
3326 | return false; // dead path (mirror->is_top()). |
3327 | } |
3328 | if (obj == NULL || obj->is_top()) { |
3329 | return false; // dead path |
3330 | } |
3331 | const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); |
3332 | |
3333 | // First, see if Class.cast() can be folded statically. |
3334 | // java_mirror_type() returns non-null for compile-time Class constants. |
3335 | ciType* tm = mirror_con->java_mirror_type(); |
3336 | if (tm != NULL && tm->is_klass() && |
3337 | tp != NULL && tp->klass() != NULL) { |
3338 | if (!tp->klass()->is_loaded()) { |
3339 | // Don't use intrinsic when class is not loaded. |
3340 | return false; |
3341 | } else { |
3342 | int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); |
3343 | if (static_res == Compile::SSC_always_true) { |
3344 | // isInstance() is true - fold the code. |
3345 | set_result(obj); |
3346 | return true; |
3347 | } else if (static_res == Compile::SSC_always_false) { |
3348 | // Don't use intrinsic, have to throw ClassCastException. |
3349 | // If the reference is null, the non-intrinsic bytecode will |
3350 | // be optimized appropriately. |
3351 | return false; |
3352 | } |
3353 | } |
3354 | } |
3355 | |
3356 | // Bailout intrinsic and do normal inlining if exception path is frequent. |
3357 | if (too_many_traps(Deoptimization::Reason_intrinsic)) { |
3358 | return false; |
3359 | } |
3360 | |
3361 | // Generate dynamic checks. |
3362 | // Class.cast() is java implementation of _checkcast bytecode. |
3363 | // Do checkcast (Parse::do_checkcast()) optimizations here. |
3364 | |
3365 | mirror = null_check(mirror); |
3366 | // If mirror is dead, only null-path is taken. |
3367 | if (stopped()) { |
3368 | return true; |
3369 | } |
3370 | |
3371 | // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). |
3372 | enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; |
3373 | RegionNode* region = new RegionNode(PATH_LIMIT); |
3374 | record_for_igvn(region); |
3375 | |
3376 | // Now load the mirror's klass metaobject, and null-check it. |
3377 | // If kls is null, we have a primitive mirror and |
3378 | // nothing is an instance of a primitive type. |
3379 | Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); |
3380 | |
3381 | Node* res = top(); |
3382 | if (!stopped()) { |
3383 | Node* bad_type_ctrl = top(); |
3384 | // Do checkcast optimizations. |
3385 | res = gen_checkcast(obj, kls, &bad_type_ctrl); |
3386 | region->init_req(_bad_type_path, bad_type_ctrl); |
3387 | } |
3388 | if (region->in(_prim_path) != top() || |
3389 | region->in(_bad_type_path) != top()) { |
3390 | // Let Interpreter throw ClassCastException. |
3391 | PreserveJVMState pjvms(this); |
3392 | set_control(_gvn.transform(region)); |
3393 | uncommon_trap(Deoptimization::Reason_intrinsic, |
3394 | Deoptimization::Action_maybe_recompile); |
3395 | } |
3396 | if (!stopped()) { |
3397 | set_result(res); |
3398 | } |
3399 | return true; |
3400 | } |
3401 | |
3402 | |
3403 | //--------------------------inline_native_subtype_check------------------------ |
3404 | // This intrinsic takes the JNI calls out of the heart of |
3405 | // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. |
3406 | bool LibraryCallKit::inline_native_subtype_check() { |
3407 | // Pull both arguments off the stack. |
3408 | Node* args[2]; // two java.lang.Class mirrors: superc, subc |
3409 | args[0] = argument(0); |
3410 | args[1] = argument(1); |
3411 | Node* klasses[2]; // corresponding Klasses: superk, subk |
3412 | klasses[0] = klasses[1] = top(); |
3413 | |
3414 | enum { |
3415 | // A full decision tree on {superc is prim, subc is prim}: |
3416 | _prim_0_path = 1, // {P,N} => false |
3417 | // {P,P} & superc!=subc => false |
3418 | _prim_same_path, // {P,P} & superc==subc => true |
3419 | _prim_1_path, // {N,P} => false |
3420 | _ref_subtype_path, // {N,N} & subtype check wins => true |
3421 | _both_ref_path, // {N,N} & subtype check loses => false |
3422 | PATH_LIMIT |
3423 | }; |
3424 | |
3425 | RegionNode* region = new RegionNode(PATH_LIMIT); |
3426 | Node* phi = new PhiNode(region, TypeInt::BOOL); |
3427 | record_for_igvn(region); |
3428 | |
3429 | const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads |
3430 | const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; |
3431 | int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); |
3432 | |
3433 | // First null-check both mirrors and load each mirror's klass metaobject. |
3434 | int which_arg; |
3435 | for (which_arg = 0; which_arg <= 1; which_arg++) { |
3436 | Node* arg = args[which_arg]; |
3437 | arg = null_check(arg); |
3438 | if (stopped()) break; |
3439 | args[which_arg] = arg; |
3440 | |
3441 | Node* p = basic_plus_adr(arg, class_klass_offset); |
3442 | Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); |
3443 | klasses[which_arg] = _gvn.transform(kls); |
3444 | } |
3445 | |
3446 | // Resolve oops to stable for CmpP below. |
3447 | args[0] = access_resolve(args[0], 0); |
3448 | args[1] = access_resolve(args[1], 0); |
3449 | |
3450 | // Having loaded both klasses, test each for null. |
3451 | bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
3452 | for (which_arg = 0; which_arg <= 1; which_arg++) { |
3453 | Node* kls = klasses[which_arg]; |
3454 | Node* null_ctl = top(); |
3455 | kls = null_check_oop(kls, &null_ctl, never_see_null); |
3456 | int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); |
3457 | region->init_req(prim_path, null_ctl); |
3458 | if (stopped()) break; |
3459 | klasses[which_arg] = kls; |
3460 | } |
3461 | |
3462 | if (!stopped()) { |
3463 | // now we have two reference types, in klasses[0..1] |
3464 | Node* subk = klasses[1]; // the argument to isAssignableFrom |
3465 | Node* superk = klasses[0]; // the receiver |
3466 | region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); |
3467 | // now we have a successful reference subtype check |
3468 | region->set_req(_ref_subtype_path, control()); |
3469 | } |
3470 | |
3471 | // If both operands are primitive (both klasses null), then |
3472 | // we must return true when they are identical primitives. |
3473 | // It is convenient to test this after the first null klass check. |
3474 | set_control(region->in(_prim_0_path)); // go back to first null check |
3475 | if (!stopped()) { |
3476 | // Since superc is primitive, make a guard for the superc==subc case. |
3477 | Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); |
3478 | Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); |
3479 | generate_guard(bol_eq, region, PROB_FAIR); |
3480 | if (region->req() == PATH_LIMIT+1) { |
3481 | // A guard was added. If the added guard is taken, superc==subc. |
3482 | region->swap_edges(PATH_LIMIT, _prim_same_path); |
3483 | region->del_req(PATH_LIMIT); |
3484 | } |
3485 | region->set_req(_prim_0_path, control()); // Not equal after all. |
3486 | } |
3487 | |
3488 | // these are the only paths that produce 'true': |
3489 | phi->set_req(_prim_same_path, intcon(1)); |
3490 | phi->set_req(_ref_subtype_path, intcon(1)); |
3491 | |
3492 | // pull together the cases: |
3493 | assert(region->req() == PATH_LIMIT, "sane region" ); |
3494 | for (uint i = 1; i < region->req(); i++) { |
3495 | Node* ctl = region->in(i); |
3496 | if (ctl == NULL || ctl == top()) { |
3497 | region->set_req(i, top()); |
3498 | phi ->set_req(i, top()); |
3499 | } else if (phi->in(i) == NULL) { |
3500 | phi->set_req(i, intcon(0)); // all other paths produce 'false' |
3501 | } |
3502 | } |
3503 | |
3504 | set_control(_gvn.transform(region)); |
3505 | set_result(_gvn.transform(phi)); |
3506 | return true; |
3507 | } |
3508 | |
3509 | //---------------------generate_array_guard_common------------------------ |
3510 | Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, |
3511 | bool obj_array, bool not_array) { |
3512 | |
3513 | if (stopped()) { |
3514 | return NULL; |
3515 | } |
3516 | |
3517 | // If obj_array/non_array==false/false: |
3518 | // Branch around if the given klass is in fact an array (either obj or prim). |
3519 | // If obj_array/non_array==false/true: |
3520 | // Branch around if the given klass is not an array klass of any kind. |
3521 | // If obj_array/non_array==true/true: |
3522 | // Branch around if the kls is not an oop array (kls is int[], String, etc.) |
3523 | // If obj_array/non_array==true/false: |
3524 | // Branch around if the kls is an oop array (Object[] or subtype) |
3525 | // |
3526 | // Like generate_guard, adds a new path onto the region. |
3527 | jint layout_con = 0; |
3528 | Node* layout_val = get_layout_helper(kls, layout_con); |
3529 | if (layout_val == NULL) { |
3530 | bool query = (obj_array |
3531 | ? Klass::layout_helper_is_objArray(layout_con) |
3532 | : Klass::layout_helper_is_array(layout_con)); |
3533 | if (query == not_array) { |
3534 | return NULL; // never a branch |
3535 | } else { // always a branch |
3536 | Node* always_branch = control(); |
3537 | if (region != NULL) |
3538 | region->add_req(always_branch); |
3539 | set_control(top()); |
3540 | return always_branch; |
3541 | } |
3542 | } |
3543 | // Now test the correct condition. |
3544 | jint nval = (obj_array |
3545 | ? (jint)(Klass::_lh_array_tag_type_value |
3546 | << Klass::_lh_array_tag_shift) |
3547 | : Klass::_lh_neutral_value); |
3548 | Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); |
3549 | BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array |
3550 | // invert the test if we are looking for a non-array |
3551 | if (not_array) btest = BoolTest(btest).negate(); |
3552 | Node* bol = _gvn.transform(new BoolNode(cmp, btest)); |
3553 | return generate_fair_guard(bol, region); |
3554 | } |
3555 | |
3556 | |
3557 | //-----------------------inline_native_newArray-------------------------- |
3558 | // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); |
3559 | // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); |
3560 | bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { |
3561 | Node* mirror; |
3562 | Node* count_val; |
3563 | if (uninitialized) { |
3564 | mirror = argument(1); |
3565 | count_val = argument(2); |
3566 | } else { |
3567 | mirror = argument(0); |
3568 | count_val = argument(1); |
3569 | } |
3570 | |
3571 | mirror = null_check(mirror); |
3572 | // If mirror or obj is dead, only null-path is taken. |
3573 | if (stopped()) return true; |
3574 | |
3575 | enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; |
3576 | RegionNode* result_reg = new RegionNode(PATH_LIMIT); |
3577 | PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); |
3578 | PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); |
3579 | PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); |
3580 | |
3581 | bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); |
3582 | Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, |
3583 | result_reg, _slow_path); |
3584 | Node* normal_ctl = control(); |
3585 | Node* no_array_ctl = result_reg->in(_slow_path); |
3586 | |
3587 | // Generate code for the slow case. We make a call to newArray(). |
3588 | set_control(no_array_ctl); |
3589 | if (!stopped()) { |
3590 | // Either the input type is void.class, or else the |
3591 | // array klass has not yet been cached. Either the |
3592 | // ensuing call will throw an exception, or else it |
3593 | // will cache the array klass for next time. |
3594 | PreserveJVMState pjvms(this); |
3595 | CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); |
3596 | Node* slow_result = set_results_for_java_call(slow_call); |
3597 | // this->control() comes from set_results_for_java_call |
3598 | result_reg->set_req(_slow_path, control()); |
3599 | result_val->set_req(_slow_path, slow_result); |
3600 | result_io ->set_req(_slow_path, i_o()); |
3601 | result_mem->set_req(_slow_path, reset_memory()); |
3602 | } |
3603 | |
3604 | set_control(normal_ctl); |
3605 | if (!stopped()) { |
3606 | // Normal case: The array type has been cached in the java.lang.Class. |
3607 | // The following call works fine even if the array type is polymorphic. |
3608 | // It could be a dynamic mix of int[], boolean[], Object[], etc. |
3609 | Node* obj = new_array(klass_node, count_val, 0); // no arguments to push |
3610 | result_reg->init_req(_normal_path, control()); |
3611 | result_val->init_req(_normal_path, obj); |
3612 | result_io ->init_req(_normal_path, i_o()); |
3613 | result_mem->init_req(_normal_path, reset_memory()); |
3614 | |
3615 | if (uninitialized) { |
3616 | // Mark the allocation so that zeroing is skipped |
3617 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); |
3618 | alloc->maybe_set_complete(&_gvn); |
3619 | } |
3620 | } |
3621 | |
3622 | // Return the combined state. |
3623 | set_i_o( _gvn.transform(result_io) ); |
3624 | set_all_memory( _gvn.transform(result_mem)); |
3625 | |
3626 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
3627 | set_result(result_reg, result_val); |
3628 | return true; |
3629 | } |
3630 | |
3631 | //----------------------inline_native_getLength-------------------------- |
3632 | // public static native int java.lang.reflect.Array.getLength(Object array); |
3633 | bool LibraryCallKit::inline_native_getLength() { |
3634 | if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
3635 | |
3636 | Node* array = null_check(argument(0)); |
3637 | // If array is dead, only null-path is taken. |
3638 | if (stopped()) return true; |
3639 | |
3640 | // Deoptimize if it is a non-array. |
3641 | Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); |
3642 | |
3643 | if (non_array != NULL) { |
3644 | PreserveJVMState pjvms(this); |
3645 | set_control(non_array); |
3646 | uncommon_trap(Deoptimization::Reason_intrinsic, |
3647 | Deoptimization::Action_maybe_recompile); |
3648 | } |
3649 | |
3650 | // If control is dead, only non-array-path is taken. |
3651 | if (stopped()) return true; |
3652 | |
3653 | // The works fine even if the array type is polymorphic. |
3654 | // It could be a dynamic mix of int[], boolean[], Object[], etc. |
3655 | Node* result = load_array_length(array); |
3656 | |
3657 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
3658 | set_result(result); |
3659 | return true; |
3660 | } |
3661 | |
3662 | //------------------------inline_array_copyOf---------------------------- |
3663 | // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); |
3664 | // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); |
3665 | bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { |
3666 | if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; |
3667 | |
3668 | // Get the arguments. |
3669 | Node* original = argument(0); |
3670 | Node* start = is_copyOfRange? argument(1): intcon(0); |
3671 | Node* end = is_copyOfRange? argument(2): argument(1); |
3672 | Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); |
3673 | |
3674 | Node* newcopy = NULL; |
3675 | |
3676 | // Set the original stack and the reexecute bit for the interpreter to reexecute |
3677 | // the bytecode that invokes Arrays.copyOf if deoptimization happens. |
3678 | { PreserveReexecuteState preexecs(this); |
3679 | jvms()->set_should_reexecute(true); |
3680 | |
3681 | array_type_mirror = null_check(array_type_mirror); |
3682 | original = null_check(original); |
3683 | |
3684 | // Check if a null path was taken unconditionally. |
3685 | if (stopped()) return true; |
3686 | |
3687 | Node* orig_length = load_array_length(original); |
3688 | |
3689 | Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); |
3690 | klass_node = null_check(klass_node); |
3691 | |
3692 | RegionNode* bailout = new RegionNode(1); |
3693 | record_for_igvn(bailout); |
3694 | |
3695 | // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. |
3696 | // Bail out if that is so. |
3697 | Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); |
3698 | if (not_objArray != NULL) { |
3699 | // Improve the klass node's type from the new optimistic assumption: |
3700 | ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); |
3701 | const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); |
3702 | Node* cast = new CastPPNode(klass_node, akls); |
3703 | cast->init_req(0, control()); |
3704 | klass_node = _gvn.transform(cast); |
3705 | } |
3706 | |
3707 | // Bail out if either start or end is negative. |
3708 | generate_negative_guard(start, bailout, &start); |
3709 | generate_negative_guard(end, bailout, &end); |
3710 | |
3711 | Node* length = end; |
3712 | if (_gvn.type(start) != TypeInt::ZERO) { |
3713 | length = _gvn.transform(new SubINode(end, start)); |
3714 | } |
3715 | |
3716 | // Bail out if length is negative. |
3717 | // Without this the new_array would throw |
3718 | // NegativeArraySizeException but IllegalArgumentException is what |
3719 | // should be thrown |
3720 | generate_negative_guard(length, bailout, &length); |
3721 | |
3722 | if (bailout->req() > 1) { |
3723 | PreserveJVMState pjvms(this); |
3724 | set_control(_gvn.transform(bailout)); |
3725 | uncommon_trap(Deoptimization::Reason_intrinsic, |
3726 | Deoptimization::Action_maybe_recompile); |
3727 | } |
3728 | |
3729 | if (!stopped()) { |
3730 | // How many elements will we copy from the original? |
3731 | // The answer is MinI(orig_length - start, length). |
3732 | Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); |
3733 | Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); |
3734 | |
3735 | original = access_resolve(original, ACCESS_READ); |
3736 | |
3737 | // Generate a direct call to the right arraycopy function(s). |
3738 | // We know the copy is disjoint but we might not know if the |
3739 | // oop stores need checking. |
3740 | // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). |
3741 | // This will fail a store-check if x contains any non-nulls. |
3742 | |
3743 | // ArrayCopyNode:Ideal may transform the ArrayCopyNode to |
3744 | // loads/stores but it is legal only if we're sure the |
3745 | // Arrays.copyOf would succeed. So we need all input arguments |
3746 | // to the copyOf to be validated, including that the copy to the |
3747 | // new array won't trigger an ArrayStoreException. That subtype |
3748 | // check can be optimized if we know something on the type of |
3749 | // the input array from type speculation. |
3750 | if (_gvn.type(klass_node)->singleton()) { |
3751 | ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); |
3752 | ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); |
3753 | |
3754 | int test = C->static_subtype_check(superk, subk); |
3755 | if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { |
3756 | const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); |
3757 | if (t_original->speculative_type() != NULL) { |
3758 | original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); |
3759 | } |
3760 | } |
3761 | } |
3762 | |
3763 | bool validated = false; |
3764 | // Reason_class_check rather than Reason_intrinsic because we |
3765 | // want to intrinsify even if this traps. |
3766 | if (!too_many_traps(Deoptimization::Reason_class_check)) { |
3767 | Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), |
3768 | klass_node); |
3769 | |
3770 | if (not_subtype_ctrl != top()) { |
3771 | PreserveJVMState pjvms(this); |
3772 | set_control(not_subtype_ctrl); |
3773 | uncommon_trap(Deoptimization::Reason_class_check, |
3774 | Deoptimization::Action_make_not_entrant); |
3775 | assert(stopped(), "Should be stopped" ); |
3776 | } |
3777 | validated = true; |
3778 | } |
3779 | |
3780 | if (!stopped()) { |
3781 | newcopy = new_array(klass_node, length, 0); // no arguments to push |
3782 | |
3783 | ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, |
3784 | load_object_klass(original), klass_node); |
3785 | if (!is_copyOfRange) { |
3786 | ac->set_copyof(validated); |
3787 | } else { |
3788 | ac->set_copyofrange(validated); |
3789 | } |
3790 | Node* n = _gvn.transform(ac); |
3791 | if (n == ac) { |
3792 | ac->connect_outputs(this); |
3793 | } else { |
3794 | assert(validated, "shouldn't transform if all arguments not validated" ); |
3795 | set_all_memory(n); |
3796 | } |
3797 | } |
3798 | } |
3799 | } // original reexecute is set back here |
3800 | |
3801 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
3802 | if (!stopped()) { |
3803 | set_result(newcopy); |
3804 | } |
3805 | return true; |
3806 | } |
3807 | |
3808 | |
3809 | //----------------------generate_virtual_guard--------------------------- |
3810 | // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. |
3811 | Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, |
3812 | RegionNode* slow_region) { |
3813 | ciMethod* method = callee(); |
3814 | int vtable_index = method->vtable_index(); |
3815 | assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, |
3816 | "bad index %d" , vtable_index); |
3817 | // Get the Method* out of the appropriate vtable entry. |
3818 | int entry_offset = in_bytes(Klass::vtable_start_offset()) + |
3819 | vtable_index*vtableEntry::size_in_bytes() + |
3820 | vtableEntry::method_offset_in_bytes(); |
3821 | Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); |
3822 | Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); |
3823 | |
3824 | // Compare the target method with the expected method (e.g., Object.hashCode). |
3825 | const TypePtr* native_call_addr = TypeMetadataPtr::make(method); |
3826 | |
3827 | Node* native_call = makecon(native_call_addr); |
3828 | Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); |
3829 | Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); |
3830 | |
3831 | return generate_slow_guard(test_native, slow_region); |
3832 | } |
3833 | |
3834 | //-----------------------generate_method_call---------------------------- |
3835 | // Use generate_method_call to make a slow-call to the real |
3836 | // method if the fast path fails. An alternative would be to |
3837 | // use a stub like OptoRuntime::slow_arraycopy_Java. |
3838 | // This only works for expanding the current library call, |
3839 | // not another intrinsic. (E.g., don't use this for making an |
3840 | // arraycopy call inside of the copyOf intrinsic.) |
3841 | CallJavaNode* |
3842 | LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { |
3843 | // When compiling the intrinsic method itself, do not use this technique. |
3844 | guarantee(callee() != C->method(), "cannot make slow-call to self" ); |
3845 | |
3846 | ciMethod* method = callee(); |
3847 | // ensure the JVMS we have will be correct for this call |
3848 | guarantee(method_id == method->intrinsic_id(), "must match" ); |
3849 | |
3850 | const TypeFunc* tf = TypeFunc::make(method); |
3851 | CallJavaNode* slow_call; |
3852 | if (is_static) { |
3853 | assert(!is_virtual, "" ); |
3854 | slow_call = new CallStaticJavaNode(C, tf, |
3855 | SharedRuntime::get_resolve_static_call_stub(), |
3856 | method, bci()); |
3857 | } else if (is_virtual) { |
3858 | null_check_receiver(); |
3859 | int vtable_index = Method::invalid_vtable_index; |
3860 | if (UseInlineCaches) { |
3861 | // Suppress the vtable call |
3862 | } else { |
3863 | // hashCode and clone are not a miranda methods, |
3864 | // so the vtable index is fixed. |
3865 | // No need to use the linkResolver to get it. |
3866 | vtable_index = method->vtable_index(); |
3867 | assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, |
3868 | "bad index %d" , vtable_index); |
3869 | } |
3870 | slow_call = new CallDynamicJavaNode(tf, |
3871 | SharedRuntime::get_resolve_virtual_call_stub(), |
3872 | method, vtable_index, bci()); |
3873 | } else { // neither virtual nor static: opt_virtual |
3874 | null_check_receiver(); |
3875 | slow_call = new CallStaticJavaNode(C, tf, |
3876 | SharedRuntime::get_resolve_opt_virtual_call_stub(), |
3877 | method, bci()); |
3878 | slow_call->set_optimized_virtual(true); |
3879 | } |
3880 | if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) { |
3881 | // To be able to issue a direct call (optimized virtual or virtual) |
3882 | // and skip a call to MH.linkTo*/invokeBasic adapter, additional information |
3883 | // about the method being invoked should be attached to the call site to |
3884 | // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C). |
3885 | slow_call->set_override_symbolic_info(true); |
3886 | } |
3887 | set_arguments_for_java_call(slow_call); |
3888 | set_edges_for_java_call(slow_call); |
3889 | return slow_call; |
3890 | } |
3891 | |
3892 | |
3893 | /** |
3894 | * Build special case code for calls to hashCode on an object. This call may |
3895 | * be virtual (invokevirtual) or bound (invokespecial). For each case we generate |
3896 | * slightly different code. |
3897 | */ |
3898 | bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { |
3899 | assert(is_static == callee()->is_static(), "correct intrinsic selection" ); |
3900 | assert(!(is_virtual && is_static), "either virtual, special, or static" ); |
3901 | |
3902 | enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; |
3903 | |
3904 | RegionNode* result_reg = new RegionNode(PATH_LIMIT); |
3905 | PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); |
3906 | PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); |
3907 | PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); |
3908 | Node* obj = NULL; |
3909 | if (!is_static) { |
3910 | // Check for hashing null object |
3911 | obj = null_check_receiver(); |
3912 | if (stopped()) return true; // unconditionally null |
3913 | result_reg->init_req(_null_path, top()); |
3914 | result_val->init_req(_null_path, top()); |
3915 | } else { |
3916 | // Do a null check, and return zero if null. |
3917 | // System.identityHashCode(null) == 0 |
3918 | obj = argument(0); |
3919 | Node* null_ctl = top(); |
3920 | obj = null_check_oop(obj, &null_ctl); |
3921 | result_reg->init_req(_null_path, null_ctl); |
3922 | result_val->init_req(_null_path, _gvn.intcon(0)); |
3923 | } |
3924 | |
3925 | // Unconditionally null? Then return right away. |
3926 | if (stopped()) { |
3927 | set_control( result_reg->in(_null_path)); |
3928 | if (!stopped()) |
3929 | set_result(result_val->in(_null_path)); |
3930 | return true; |
3931 | } |
3932 | |
3933 | // We only go to the fast case code if we pass a number of guards. The |
3934 | // paths which do not pass are accumulated in the slow_region. |
3935 | RegionNode* slow_region = new RegionNode(1); |
3936 | record_for_igvn(slow_region); |
3937 | |
3938 | // If this is a virtual call, we generate a funny guard. We pull out |
3939 | // the vtable entry corresponding to hashCode() from the target object. |
3940 | // If the target method which we are calling happens to be the native |
3941 | // Object hashCode() method, we pass the guard. We do not need this |
3942 | // guard for non-virtual calls -- the caller is known to be the native |
3943 | // Object hashCode(). |
3944 | if (is_virtual) { |
3945 | // After null check, get the object's klass. |
3946 | Node* obj_klass = load_object_klass(obj); |
3947 | generate_virtual_guard(obj_klass, slow_region); |
3948 | } |
3949 | |
3950 | // Get the header out of the object, use LoadMarkNode when available |
3951 | Node* = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); |
3952 | // The control of the load must be NULL. Otherwise, the load can move before |
3953 | // the null check after castPP removal. |
3954 | Node* no_ctrl = NULL; |
3955 | Node* = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); |
3956 | |
3957 | // Test the header to see if it is unlocked. |
3958 | Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); |
3959 | Node * = _gvn.transform(new AndXNode(header, lock_mask)); |
3960 | Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); |
3961 | Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); |
3962 | Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); |
3963 | |
3964 | generate_slow_guard(test_unlocked, slow_region); |
3965 | |
3966 | // Get the hash value and check to see that it has been properly assigned. |
3967 | // We depend on hash_mask being at most 32 bits and avoid the use of |
3968 | // hash_mask_in_place because it could be larger than 32 bits in a 64-bit |
3969 | // vm: see markOop.hpp. |
3970 | Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); |
3971 | Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); |
3972 | Node *= _gvn.transform(new URShiftXNode(header, hash_shift)); |
3973 | // This hack lets the hash bits live anywhere in the mark object now, as long |
3974 | // as the shift drops the relevant bits into the low 32 bits. Note that |
3975 | // Java spec says that HashCode is an int so there's no point in capturing |
3976 | // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). |
3977 | hshifted_header = ConvX2I(hshifted_header); |
3978 | Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); |
3979 | |
3980 | Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); |
3981 | Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); |
3982 | Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); |
3983 | |
3984 | generate_slow_guard(test_assigned, slow_region); |
3985 | |
3986 | Node* init_mem = reset_memory(); |
3987 | // fill in the rest of the null path: |
3988 | result_io ->init_req(_null_path, i_o()); |
3989 | result_mem->init_req(_null_path, init_mem); |
3990 | |
3991 | result_val->init_req(_fast_path, hash_val); |
3992 | result_reg->init_req(_fast_path, control()); |
3993 | result_io ->init_req(_fast_path, i_o()); |
3994 | result_mem->init_req(_fast_path, init_mem); |
3995 | |
3996 | // Generate code for the slow case. We make a call to hashCode(). |
3997 | set_control(_gvn.transform(slow_region)); |
3998 | if (!stopped()) { |
3999 | // No need for PreserveJVMState, because we're using up the present state. |
4000 | set_all_memory(init_mem); |
4001 | vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; |
4002 | CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); |
4003 | Node* slow_result = set_results_for_java_call(slow_call); |
4004 | // this->control() comes from set_results_for_java_call |
4005 | result_reg->init_req(_slow_path, control()); |
4006 | result_val->init_req(_slow_path, slow_result); |
4007 | result_io ->set_req(_slow_path, i_o()); |
4008 | result_mem ->set_req(_slow_path, reset_memory()); |
4009 | } |
4010 | |
4011 | // Return the combined state. |
4012 | set_i_o( _gvn.transform(result_io) ); |
4013 | set_all_memory( _gvn.transform(result_mem)); |
4014 | |
4015 | set_result(result_reg, result_val); |
4016 | return true; |
4017 | } |
4018 | |
4019 | //---------------------------inline_native_getClass---------------------------- |
4020 | // public final native Class<?> java.lang.Object.getClass(); |
4021 | // |
4022 | // Build special case code for calls to getClass on an object. |
4023 | bool LibraryCallKit::inline_native_getClass() { |
4024 | Node* obj = null_check_receiver(); |
4025 | if (stopped()) return true; |
4026 | set_result(load_mirror_from_klass(load_object_klass(obj))); |
4027 | return true; |
4028 | } |
4029 | |
4030 | //-----------------inline_native_Reflection_getCallerClass--------------------- |
4031 | // public static native Class<?> sun.reflect.Reflection.getCallerClass(); |
4032 | // |
4033 | // In the presence of deep enough inlining, getCallerClass() becomes a no-op. |
4034 | // |
4035 | // NOTE: This code must perform the same logic as JVM_GetCallerClass |
4036 | // in that it must skip particular security frames and checks for |
4037 | // caller sensitive methods. |
4038 | bool LibraryCallKit::inline_native_Reflection_getCallerClass() { |
4039 | #ifndef PRODUCT |
4040 | if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { |
4041 | tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass" ); |
4042 | } |
4043 | #endif |
4044 | |
4045 | if (!jvms()->has_method()) { |
4046 | #ifndef PRODUCT |
4047 | if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { |
4048 | tty->print_cr(" Bailing out because intrinsic was inlined at top level" ); |
4049 | } |
4050 | #endif |
4051 | return false; |
4052 | } |
4053 | |
4054 | // Walk back up the JVM state to find the caller at the required |
4055 | // depth. |
4056 | JVMState* caller_jvms = jvms(); |
4057 | |
4058 | // Cf. JVM_GetCallerClass |
4059 | // NOTE: Start the loop at depth 1 because the current JVM state does |
4060 | // not include the Reflection.getCallerClass() frame. |
4061 | for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { |
4062 | ciMethod* m = caller_jvms->method(); |
4063 | switch (n) { |
4064 | case 0: |
4065 | fatal("current JVM state does not include the Reflection.getCallerClass frame" ); |
4066 | break; |
4067 | case 1: |
4068 | // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). |
4069 | if (!m->caller_sensitive()) { |
4070 | #ifndef PRODUCT |
4071 | if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { |
4072 | tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d" , n); |
4073 | } |
4074 | #endif |
4075 | return false; // bail-out; let JVM_GetCallerClass do the work |
4076 | } |
4077 | break; |
4078 | default: |
4079 | if (!m->is_ignored_by_security_stack_walk()) { |
4080 | // We have reached the desired frame; return the holder class. |
4081 | // Acquire method holder as java.lang.Class and push as constant. |
4082 | ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); |
4083 | ciInstance* caller_mirror = caller_klass->java_mirror(); |
4084 | set_result(makecon(TypeInstPtr::make(caller_mirror))); |
4085 | |
4086 | #ifndef PRODUCT |
4087 | if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { |
4088 | 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()); |
4089 | tty->print_cr(" JVM state at this point:" ); |
4090 | for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { |
4091 | ciMethod* m = jvms()->of_depth(i)->method(); |
4092 | tty->print_cr(" %d) %s.%s" , n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); |
4093 | } |
4094 | } |
4095 | #endif |
4096 | return true; |
4097 | } |
4098 | break; |
4099 | } |
4100 | } |
4101 | |
4102 | #ifndef PRODUCT |
4103 | if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { |
4104 | tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d" , jvms()->depth()); |
4105 | tty->print_cr(" JVM state at this point:" ); |
4106 | for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { |
4107 | ciMethod* m = jvms()->of_depth(i)->method(); |
4108 | tty->print_cr(" %d) %s.%s" , n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); |
4109 | } |
4110 | } |
4111 | #endif |
4112 | |
4113 | return false; // bail-out; let JVM_GetCallerClass do the work |
4114 | } |
4115 | |
4116 | bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { |
4117 | Node* arg = argument(0); |
4118 | Node* result = NULL; |
4119 | |
4120 | switch (id) { |
4121 | case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; |
4122 | case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; |
4123 | case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; |
4124 | case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; |
4125 | |
4126 | case vmIntrinsics::_doubleToLongBits: { |
4127 | // two paths (plus control) merge in a wood |
4128 | RegionNode *r = new RegionNode(3); |
4129 | Node *phi = new PhiNode(r, TypeLong::LONG); |
4130 | |
4131 | Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); |
4132 | // Build the boolean node |
4133 | Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); |
4134 | |
4135 | // Branch either way. |
4136 | // NaN case is less traveled, which makes all the difference. |
4137 | IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
4138 | Node *opt_isnan = _gvn.transform(ifisnan); |
4139 | assert( opt_isnan->is_If(), "Expect an IfNode" ); |
4140 | IfNode *opt_ifisnan = (IfNode*)opt_isnan; |
4141 | Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); |
4142 | |
4143 | set_control(iftrue); |
4144 | |
4145 | static const jlong nan_bits = CONST64(0x7ff8000000000000); |
4146 | Node *slow_result = longcon(nan_bits); // return NaN |
4147 | phi->init_req(1, _gvn.transform( slow_result )); |
4148 | r->init_req(1, iftrue); |
4149 | |
4150 | // Else fall through |
4151 | Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); |
4152 | set_control(iffalse); |
4153 | |
4154 | phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); |
4155 | r->init_req(2, iffalse); |
4156 | |
4157 | // Post merge |
4158 | set_control(_gvn.transform(r)); |
4159 | record_for_igvn(r); |
4160 | |
4161 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
4162 | result = phi; |
4163 | assert(result->bottom_type()->isa_long(), "must be" ); |
4164 | break; |
4165 | } |
4166 | |
4167 | case vmIntrinsics::_floatToIntBits: { |
4168 | // two paths (plus control) merge in a wood |
4169 | RegionNode *r = new RegionNode(3); |
4170 | Node *phi = new PhiNode(r, TypeInt::INT); |
4171 | |
4172 | Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); |
4173 | // Build the boolean node |
4174 | Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); |
4175 | |
4176 | // Branch either way. |
4177 | // NaN case is less traveled, which makes all the difference. |
4178 | IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
4179 | Node *opt_isnan = _gvn.transform(ifisnan); |
4180 | assert( opt_isnan->is_If(), "Expect an IfNode" ); |
4181 | IfNode *opt_ifisnan = (IfNode*)opt_isnan; |
4182 | Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); |
4183 | |
4184 | set_control(iftrue); |
4185 | |
4186 | static const jint nan_bits = 0x7fc00000; |
4187 | Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN |
4188 | phi->init_req(1, _gvn.transform( slow_result )); |
4189 | r->init_req(1, iftrue); |
4190 | |
4191 | // Else fall through |
4192 | Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); |
4193 | set_control(iffalse); |
4194 | |
4195 | phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); |
4196 | r->init_req(2, iffalse); |
4197 | |
4198 | // Post merge |
4199 | set_control(_gvn.transform(r)); |
4200 | record_for_igvn(r); |
4201 | |
4202 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
4203 | result = phi; |
4204 | assert(result->bottom_type()->isa_int(), "must be" ); |
4205 | break; |
4206 | } |
4207 | |
4208 | default: |
4209 | fatal_unexpected_iid(id); |
4210 | break; |
4211 | } |
4212 | set_result(_gvn.transform(result)); |
4213 | return true; |
4214 | } |
4215 | |
4216 | //----------------------inline_unsafe_copyMemory------------------------- |
4217 | // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); |
4218 | bool LibraryCallKit::inline_unsafe_copyMemory() { |
4219 | if (callee()->is_static()) return false; // caller must have the capability! |
4220 | null_check_receiver(); // null-check receiver |
4221 | if (stopped()) return true; |
4222 | |
4223 | C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". |
4224 | |
4225 | Node* src_ptr = argument(1); // type: oop |
4226 | Node* src_off = ConvL2X(argument(2)); // type: long |
4227 | Node* dst_ptr = argument(4); // type: oop |
4228 | Node* dst_off = ConvL2X(argument(5)); // type: long |
4229 | Node* size = ConvL2X(argument(7)); // type: long |
4230 | |
4231 | assert(Unsafe_field_offset_to_byte_offset(11) == 11, |
4232 | "fieldOffset must be byte-scaled" ); |
4233 | |
4234 | src_ptr = access_resolve(src_ptr, ACCESS_READ); |
4235 | dst_ptr = access_resolve(dst_ptr, ACCESS_WRITE); |
4236 | Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ); |
4237 | Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE); |
4238 | |
4239 | // Conservatively insert a memory barrier on all memory slices. |
4240 | // Do not let writes of the copy source or destination float below the copy. |
4241 | insert_mem_bar(Op_MemBarCPUOrder); |
4242 | |
4243 | // Call it. Note that the length argument is not scaled. |
4244 | make_runtime_call(RC_LEAF|RC_NO_FP, |
4245 | OptoRuntime::fast_arraycopy_Type(), |
4246 | StubRoutines::unsafe_arraycopy(), |
4247 | "unsafe_arraycopy" , |
4248 | TypeRawPtr::BOTTOM, |
4249 | src, dst, size XTOP); |
4250 | |
4251 | // Do not let reads of the copy destination float above the copy. |
4252 | insert_mem_bar(Op_MemBarCPUOrder); |
4253 | |
4254 | return true; |
4255 | } |
4256 | |
4257 | //------------------------clone_coping----------------------------------- |
4258 | // Helper function for inline_native_clone. |
4259 | void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) { |
4260 | assert(obj_size != NULL, "" ); |
4261 | Node* raw_obj = alloc_obj->in(1); |
4262 | assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "" ); |
4263 | |
4264 | AllocateNode* alloc = NULL; |
4265 | if (ReduceBulkZeroing) { |
4266 | // We will be completely responsible for initializing this object - |
4267 | // mark Initialize node as complete. |
4268 | alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); |
4269 | // The object was just allocated - there should be no any stores! |
4270 | guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "" ); |
4271 | // Mark as complete_with_arraycopy so that on AllocateNode |
4272 | // expansion, we know this AllocateNode is initialized by an array |
4273 | // copy and a StoreStore barrier exists after the array copy. |
4274 | alloc->initialization()->set_complete_with_arraycopy(); |
4275 | } |
4276 | |
4277 | // Copy the fastest available way. |
4278 | // TODO: generate fields copies for small objects instead. |
4279 | Node* size = _gvn.transform(obj_size); |
4280 | |
4281 | access_clone(obj, alloc_obj, size, is_array); |
4282 | |
4283 | // Do not let reads from the cloned object float above the arraycopy. |
4284 | if (alloc != NULL) { |
4285 | // Do not let stores that initialize this object be reordered with |
4286 | // a subsequent store that would make this object accessible by |
4287 | // other threads. |
4288 | // Record what AllocateNode this StoreStore protects so that |
4289 | // escape analysis can go from the MemBarStoreStoreNode to the |
4290 | // AllocateNode and eliminate the MemBarStoreStoreNode if possible |
4291 | // based on the escape status of the AllocateNode. |
4292 | insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); |
4293 | } else { |
4294 | insert_mem_bar(Op_MemBarCPUOrder); |
4295 | } |
4296 | } |
4297 | |
4298 | //------------------------inline_native_clone---------------------------- |
4299 | // protected native Object java.lang.Object.clone(); |
4300 | // |
4301 | // Here are the simple edge cases: |
4302 | // null receiver => normal trap |
4303 | // virtual and clone was overridden => slow path to out-of-line clone |
4304 | // not cloneable or finalizer => slow path to out-of-line Object.clone |
4305 | // |
4306 | // The general case has two steps, allocation and copying. |
4307 | // Allocation has two cases, and uses GraphKit::new_instance or new_array. |
4308 | // |
4309 | // Copying also has two cases, oop arrays and everything else. |
4310 | // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). |
4311 | // Everything else uses the tight inline loop supplied by CopyArrayNode. |
4312 | // |
4313 | // These steps fold up nicely if and when the cloned object's klass |
4314 | // can be sharply typed as an object array, a type array, or an instance. |
4315 | // |
4316 | bool LibraryCallKit::inline_native_clone(bool is_virtual) { |
4317 | PhiNode* result_val; |
4318 | |
4319 | // Set the reexecute bit for the interpreter to reexecute |
4320 | // the bytecode that invokes Object.clone if deoptimization happens. |
4321 | { PreserveReexecuteState preexecs(this); |
4322 | jvms()->set_should_reexecute(true); |
4323 | |
4324 | Node* obj = null_check_receiver(); |
4325 | if (stopped()) return true; |
4326 | |
4327 | const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); |
4328 | |
4329 | // If we are going to clone an instance, we need its exact type to |
4330 | // know the number and types of fields to convert the clone to |
4331 | // loads/stores. Maybe a speculative type can help us. |
4332 | if (!obj_type->klass_is_exact() && |
4333 | obj_type->speculative_type() != NULL && |
4334 | obj_type->speculative_type()->is_instance_klass()) { |
4335 | ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); |
4336 | if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && |
4337 | !spec_ik->has_injected_fields()) { |
4338 | ciKlass* k = obj_type->klass(); |
4339 | if (!k->is_instance_klass() || |
4340 | k->as_instance_klass()->is_interface() || |
4341 | k->as_instance_klass()->has_subklass()) { |
4342 | obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); |
4343 | } |
4344 | } |
4345 | } |
4346 | |
4347 | Node* obj_klass = load_object_klass(obj); |
4348 | const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); |
4349 | const TypeOopPtr* toop = ((tklass != NULL) |
4350 | ? tklass->as_instance_type() |
4351 | : TypeInstPtr::NOTNULL); |
4352 | |
4353 | // Conservatively insert a memory barrier on all memory slices. |
4354 | // Do not let writes into the original float below the clone. |
4355 | insert_mem_bar(Op_MemBarCPUOrder); |
4356 | |
4357 | // paths into result_reg: |
4358 | enum { |
4359 | _slow_path = 1, // out-of-line call to clone method (virtual or not) |
4360 | _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy |
4361 | _array_path, // plain array allocation, plus arrayof_long_arraycopy |
4362 | _instance_path, // plain instance allocation, plus arrayof_long_arraycopy |
4363 | PATH_LIMIT |
4364 | }; |
4365 | RegionNode* result_reg = new RegionNode(PATH_LIMIT); |
4366 | result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); |
4367 | PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); |
4368 | PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); |
4369 | record_for_igvn(result_reg); |
4370 | |
4371 | Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); |
4372 | if (array_ctl != NULL) { |
4373 | // It's an array. |
4374 | PreserveJVMState pjvms(this); |
4375 | set_control(array_ctl); |
4376 | Node* obj_length = load_array_length(obj); |
4377 | Node* obj_size = NULL; |
4378 | Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push |
4379 | |
4380 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
4381 | if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { |
4382 | // If it is an oop array, it requires very special treatment, |
4383 | // because gc barriers are required when accessing the array. |
4384 | Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); |
4385 | if (is_obja != NULL) { |
4386 | PreserveJVMState pjvms2(this); |
4387 | set_control(is_obja); |
4388 | obj = access_resolve(obj, ACCESS_READ); |
4389 | // Generate a direct call to the right arraycopy function(s). |
4390 | Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); |
4391 | ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); |
4392 | ac->set_cloneoop(); |
4393 | Node* n = _gvn.transform(ac); |
4394 | assert(n == ac, "cannot disappear" ); |
4395 | ac->connect_outputs(this); |
4396 | |
4397 | result_reg->init_req(_objArray_path, control()); |
4398 | result_val->init_req(_objArray_path, alloc_obj); |
4399 | result_i_o ->set_req(_objArray_path, i_o()); |
4400 | result_mem ->set_req(_objArray_path, reset_memory()); |
4401 | } |
4402 | } |
4403 | // Otherwise, there are no barriers to worry about. |
4404 | // (We can dispense with card marks if we know the allocation |
4405 | // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks |
4406 | // causes the non-eden paths to take compensating steps to |
4407 | // simulate a fresh allocation, so that no further |
4408 | // card marks are required in compiled code to initialize |
4409 | // the object.) |
4410 | |
4411 | if (!stopped()) { |
4412 | copy_to_clone(obj, alloc_obj, obj_size, true); |
4413 | |
4414 | // Present the results of the copy. |
4415 | result_reg->init_req(_array_path, control()); |
4416 | result_val->init_req(_array_path, alloc_obj); |
4417 | result_i_o ->set_req(_array_path, i_o()); |
4418 | result_mem ->set_req(_array_path, reset_memory()); |
4419 | } |
4420 | } |
4421 | |
4422 | // We only go to the instance fast case code if we pass a number of guards. |
4423 | // The paths which do not pass are accumulated in the slow_region. |
4424 | RegionNode* slow_region = new RegionNode(1); |
4425 | record_for_igvn(slow_region); |
4426 | if (!stopped()) { |
4427 | // It's an instance (we did array above). Make the slow-path tests. |
4428 | // If this is a virtual call, we generate a funny guard. We grab |
4429 | // the vtable entry corresponding to clone() from the target object. |
4430 | // If the target method which we are calling happens to be the |
4431 | // Object clone() method, we pass the guard. We do not need this |
4432 | // guard for non-virtual calls; the caller is known to be the native |
4433 | // Object clone(). |
4434 | if (is_virtual) { |
4435 | generate_virtual_guard(obj_klass, slow_region); |
4436 | } |
4437 | |
4438 | // The object must be easily cloneable and must not have a finalizer. |
4439 | // Both of these conditions may be checked in a single test. |
4440 | // We could optimize the test further, but we don't care. |
4441 | generate_access_flags_guard(obj_klass, |
4442 | // Test both conditions: |
4443 | JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, |
4444 | // Must be cloneable but not finalizer: |
4445 | JVM_ACC_IS_CLONEABLE_FAST, |
4446 | slow_region); |
4447 | } |
4448 | |
4449 | if (!stopped()) { |
4450 | // It's an instance, and it passed the slow-path tests. |
4451 | PreserveJVMState pjvms(this); |
4452 | Node* obj_size = NULL; |
4453 | // Need to deoptimize on exception from allocation since Object.clone intrinsic |
4454 | // is reexecuted if deoptimization occurs and there could be problems when merging |
4455 | // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). |
4456 | Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); |
4457 | |
4458 | copy_to_clone(obj, alloc_obj, obj_size, false); |
4459 | |
4460 | // Present the results of the slow call. |
4461 | result_reg->init_req(_instance_path, control()); |
4462 | result_val->init_req(_instance_path, alloc_obj); |
4463 | result_i_o ->set_req(_instance_path, i_o()); |
4464 | result_mem ->set_req(_instance_path, reset_memory()); |
4465 | } |
4466 | |
4467 | // Generate code for the slow case. We make a call to clone(). |
4468 | set_control(_gvn.transform(slow_region)); |
4469 | if (!stopped()) { |
4470 | PreserveJVMState pjvms(this); |
4471 | CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); |
4472 | // We need to deoptimize on exception (see comment above) |
4473 | Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true); |
4474 | // this->control() comes from set_results_for_java_call |
4475 | result_reg->init_req(_slow_path, control()); |
4476 | result_val->init_req(_slow_path, slow_result); |
4477 | result_i_o ->set_req(_slow_path, i_o()); |
4478 | result_mem ->set_req(_slow_path, reset_memory()); |
4479 | } |
4480 | |
4481 | // Return the combined state. |
4482 | set_control( _gvn.transform(result_reg)); |
4483 | set_i_o( _gvn.transform(result_i_o)); |
4484 | set_all_memory( _gvn.transform(result_mem)); |
4485 | } // original reexecute is set back here |
4486 | |
4487 | set_result(_gvn.transform(result_val)); |
4488 | return true; |
4489 | } |
4490 | |
4491 | // If we have a tightly coupled allocation, the arraycopy may take care |
4492 | // of the array initialization. If one of the guards we insert between |
4493 | // the allocation and the arraycopy causes a deoptimization, an |
4494 | // unitialized array will escape the compiled method. To prevent that |
4495 | // we set the JVM state for uncommon traps between the allocation and |
4496 | // the arraycopy to the state before the allocation so, in case of |
4497 | // deoptimization, we'll reexecute the allocation and the |
4498 | // initialization. |
4499 | JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { |
4500 | if (alloc != NULL) { |
4501 | ciMethod* trap_method = alloc->jvms()->method(); |
4502 | int trap_bci = alloc->jvms()->bci(); |
4503 | |
4504 | if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && |
4505 | !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { |
4506 | // Make sure there's no store between the allocation and the |
4507 | // arraycopy otherwise visible side effects could be rexecuted |
4508 | // in case of deoptimization and cause incorrect execution. |
4509 | bool no_interfering_store = true; |
4510 | Node* mem = alloc->in(TypeFunc::Memory); |
4511 | if (mem->is_MergeMem()) { |
4512 | for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { |
4513 | Node* n = mms.memory(); |
4514 | if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { |
4515 | assert(n->is_Store(), "what else?" ); |
4516 | no_interfering_store = false; |
4517 | break; |
4518 | } |
4519 | } |
4520 | } else { |
4521 | for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { |
4522 | Node* n = mms.memory(); |
4523 | if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { |
4524 | assert(n->is_Store(), "what else?" ); |
4525 | no_interfering_store = false; |
4526 | break; |
4527 | } |
4528 | } |
4529 | } |
4530 | |
4531 | if (no_interfering_store) { |
4532 | JVMState* old_jvms = alloc->jvms()->clone_shallow(C); |
4533 | uint size = alloc->req(); |
4534 | SafePointNode* sfpt = new SafePointNode(size, old_jvms); |
4535 | old_jvms->set_map(sfpt); |
4536 | for (uint i = 0; i < size; i++) { |
4537 | sfpt->init_req(i, alloc->in(i)); |
4538 | } |
4539 | // re-push array length for deoptimization |
4540 | sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); |
4541 | old_jvms->set_sp(old_jvms->sp()+1); |
4542 | old_jvms->set_monoff(old_jvms->monoff()+1); |
4543 | old_jvms->set_scloff(old_jvms->scloff()+1); |
4544 | old_jvms->set_endoff(old_jvms->endoff()+1); |
4545 | old_jvms->set_should_reexecute(true); |
4546 | |
4547 | sfpt->set_i_o(map()->i_o()); |
4548 | sfpt->set_memory(map()->memory()); |
4549 | sfpt->set_control(map()->control()); |
4550 | |
4551 | JVMState* saved_jvms = jvms(); |
4552 | saved_reexecute_sp = _reexecute_sp; |
4553 | |
4554 | set_jvms(sfpt->jvms()); |
4555 | _reexecute_sp = jvms()->sp(); |
4556 | |
4557 | return saved_jvms; |
4558 | } |
4559 | } |
4560 | } |
4561 | return NULL; |
4562 | } |
4563 | |
4564 | // In case of a deoptimization, we restart execution at the |
4565 | // allocation, allocating a new array. We would leave an uninitialized |
4566 | // array in the heap that GCs wouldn't expect. Move the allocation |
4567 | // after the traps so we don't allocate the array if we |
4568 | // deoptimize. This is possible because tightly_coupled_allocation() |
4569 | // guarantees there's no observer of the allocated array at this point |
4570 | // and the control flow is simple enough. |
4571 | void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, |
4572 | int saved_reexecute_sp, uint new_idx) { |
4573 | if (saved_jvms != NULL && !stopped()) { |
4574 | assert(alloc != NULL, "only with a tightly coupled allocation" ); |
4575 | // restore JVM state to the state at the arraycopy |
4576 | saved_jvms->map()->set_control(map()->control()); |
4577 | assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?" ); |
4578 | assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?" ); |
4579 | // If we've improved the types of some nodes (null check) while |
4580 | // emitting the guards, propagate them to the current state |
4581 | map()->replaced_nodes().apply(saved_jvms->map(), new_idx); |
4582 | set_jvms(saved_jvms); |
4583 | _reexecute_sp = saved_reexecute_sp; |
4584 | |
4585 | // Remove the allocation from above the guards |
4586 | CallProjections callprojs; |
4587 | alloc->extract_projections(&callprojs, true); |
4588 | InitializeNode* init = alloc->initialization(); |
4589 | Node* alloc_mem = alloc->in(TypeFunc::Memory); |
4590 | C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); |
4591 | C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); |
4592 | C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); |
4593 | |
4594 | // move the allocation here (after the guards) |
4595 | _gvn.hash_delete(alloc); |
4596 | alloc->set_req(TypeFunc::Control, control()); |
4597 | alloc->set_req(TypeFunc::I_O, i_o()); |
4598 | Node *mem = reset_memory(); |
4599 | set_all_memory(mem); |
4600 | alloc->set_req(TypeFunc::Memory, mem); |
4601 | set_control(init->proj_out_or_null(TypeFunc::Control)); |
4602 | set_i_o(callprojs.fallthrough_ioproj); |
4603 | |
4604 | // Update memory as done in GraphKit::set_output_for_allocation() |
4605 | const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); |
4606 | const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); |
4607 | if (ary_type->isa_aryptr() && length_type != NULL) { |
4608 | ary_type = ary_type->is_aryptr()->cast_to_size(length_type); |
4609 | } |
4610 | const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); |
4611 | int elemidx = C->get_alias_index(telemref); |
4612 | set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); |
4613 | set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); |
4614 | |
4615 | Node* allocx = _gvn.transform(alloc); |
4616 | assert(allocx == alloc, "where has the allocation gone?" ); |
4617 | assert(dest->is_CheckCastPP(), "not an allocation result?" ); |
4618 | |
4619 | _gvn.hash_delete(dest); |
4620 | dest->set_req(0, control()); |
4621 | Node* destx = _gvn.transform(dest); |
4622 | assert(destx == dest, "where has the allocation result gone?" ); |
4623 | } |
4624 | } |
4625 | |
4626 | |
4627 | //------------------------------inline_arraycopy----------------------- |
4628 | // public static native void java.lang.System.arraycopy(Object src, int srcPos, |
4629 | // Object dest, int destPos, |
4630 | // int length); |
4631 | bool LibraryCallKit::inline_arraycopy() { |
4632 | // Get the arguments. |
4633 | Node* src = argument(0); // type: oop |
4634 | Node* src_offset = argument(1); // type: int |
4635 | Node* dest = argument(2); // type: oop |
4636 | Node* dest_offset = argument(3); // type: int |
4637 | Node* length = argument(4); // type: int |
4638 | |
4639 | uint new_idx = C->unique(); |
4640 | |
4641 | // Check for allocation before we add nodes that would confuse |
4642 | // tightly_coupled_allocation() |
4643 | AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); |
4644 | |
4645 | int saved_reexecute_sp = -1; |
4646 | JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); |
4647 | // See arraycopy_restore_alloc_state() comment |
4648 | // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards |
4649 | // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation |
4650 | // if saved_jvms == NULL and alloc != NULL, we can't emit any guards |
4651 | bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); |
4652 | |
4653 | // The following tests must be performed |
4654 | // (1) src and dest are arrays. |
4655 | // (2) src and dest arrays must have elements of the same BasicType |
4656 | // (3) src and dest must not be null. |
4657 | // (4) src_offset must not be negative. |
4658 | // (5) dest_offset must not be negative. |
4659 | // (6) length must not be negative. |
4660 | // (7) src_offset + length must not exceed length of src. |
4661 | // (8) dest_offset + length must not exceed length of dest. |
4662 | // (9) each element of an oop array must be assignable |
4663 | |
4664 | // (3) src and dest must not be null. |
4665 | // always do this here because we need the JVM state for uncommon traps |
4666 | Node* null_ctl = top(); |
4667 | src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); |
4668 | assert(null_ctl->is_top(), "no null control here" ); |
4669 | dest = null_check(dest, T_ARRAY); |
4670 | |
4671 | if (!can_emit_guards) { |
4672 | // if saved_jvms == NULL and alloc != NULL, we don't emit any |
4673 | // guards but the arraycopy node could still take advantage of a |
4674 | // tightly allocated allocation. tightly_coupled_allocation() is |
4675 | // called again to make sure it takes the null check above into |
4676 | // account: the null check is mandatory and if it caused an |
4677 | // uncommon trap to be emitted then the allocation can't be |
4678 | // considered tightly coupled in this context. |
4679 | alloc = tightly_coupled_allocation(dest, NULL); |
4680 | } |
4681 | |
4682 | bool validated = false; |
4683 | |
4684 | const Type* src_type = _gvn.type(src); |
4685 | const Type* dest_type = _gvn.type(dest); |
4686 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
4687 | const TypeAryPtr* top_dest = dest_type->isa_aryptr(); |
4688 | |
4689 | // Do we have the type of src? |
4690 | bool has_src = (top_src != NULL && top_src->klass() != NULL); |
4691 | // Do we have the type of dest? |
4692 | bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); |
4693 | // Is the type for src from speculation? |
4694 | bool src_spec = false; |
4695 | // Is the type for dest from speculation? |
4696 | bool dest_spec = false; |
4697 | |
4698 | if ((!has_src || !has_dest) && can_emit_guards) { |
4699 | // We don't have sufficient type information, let's see if |
4700 | // speculative types can help. We need to have types for both src |
4701 | // and dest so that it pays off. |
4702 | |
4703 | // Do we already have or could we have type information for src |
4704 | bool could_have_src = has_src; |
4705 | // Do we already have or could we have type information for dest |
4706 | bool could_have_dest = has_dest; |
4707 | |
4708 | ciKlass* src_k = NULL; |
4709 | if (!has_src) { |
4710 | src_k = src_type->speculative_type_not_null(); |
4711 | if (src_k != NULL && src_k->is_array_klass()) { |
4712 | could_have_src = true; |
4713 | } |
4714 | } |
4715 | |
4716 | ciKlass* dest_k = NULL; |
4717 | if (!has_dest) { |
4718 | dest_k = dest_type->speculative_type_not_null(); |
4719 | if (dest_k != NULL && dest_k->is_array_klass()) { |
4720 | could_have_dest = true; |
4721 | } |
4722 | } |
4723 | |
4724 | if (could_have_src && could_have_dest) { |
4725 | // This is going to pay off so emit the required guards |
4726 | if (!has_src) { |
4727 | src = maybe_cast_profiled_obj(src, src_k, true); |
4728 | src_type = _gvn.type(src); |
4729 | top_src = src_type->isa_aryptr(); |
4730 | has_src = (top_src != NULL && top_src->klass() != NULL); |
4731 | src_spec = true; |
4732 | } |
4733 | if (!has_dest) { |
4734 | dest = maybe_cast_profiled_obj(dest, dest_k, true); |
4735 | dest_type = _gvn.type(dest); |
4736 | top_dest = dest_type->isa_aryptr(); |
4737 | has_dest = (top_dest != NULL && top_dest->klass() != NULL); |
4738 | dest_spec = true; |
4739 | } |
4740 | } |
4741 | } |
4742 | |
4743 | if (has_src && has_dest && can_emit_guards) { |
4744 | BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); |
4745 | BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); |
4746 | if (src_elem == T_ARRAY) src_elem = T_OBJECT; |
4747 | if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; |
4748 | |
4749 | if (src_elem == dest_elem && src_elem == T_OBJECT) { |
4750 | // If both arrays are object arrays then having the exact types |
4751 | // for both will remove the need for a subtype check at runtime |
4752 | // before the call and may make it possible to pick a faster copy |
4753 | // routine (without a subtype check on every element) |
4754 | // Do we have the exact type of src? |
4755 | bool could_have_src = src_spec; |
4756 | // Do we have the exact type of dest? |
4757 | bool could_have_dest = dest_spec; |
4758 | ciKlass* src_k = top_src->klass(); |
4759 | ciKlass* dest_k = top_dest->klass(); |
4760 | if (!src_spec) { |
4761 | src_k = src_type->speculative_type_not_null(); |
4762 | if (src_k != NULL && src_k->is_array_klass()) { |
4763 | could_have_src = true; |
4764 | } |
4765 | } |
4766 | if (!dest_spec) { |
4767 | dest_k = dest_type->speculative_type_not_null(); |
4768 | if (dest_k != NULL && dest_k->is_array_klass()) { |
4769 | could_have_dest = true; |
4770 | } |
4771 | } |
4772 | if (could_have_src && could_have_dest) { |
4773 | // If we can have both exact types, emit the missing guards |
4774 | if (could_have_src && !src_spec) { |
4775 | src = maybe_cast_profiled_obj(src, src_k, true); |
4776 | } |
4777 | if (could_have_dest && !dest_spec) { |
4778 | dest = maybe_cast_profiled_obj(dest, dest_k, true); |
4779 | } |
4780 | } |
4781 | } |
4782 | } |
4783 | |
4784 | ciMethod* trap_method = method(); |
4785 | int trap_bci = bci(); |
4786 | if (saved_jvms != NULL) { |
4787 | trap_method = alloc->jvms()->method(); |
4788 | trap_bci = alloc->jvms()->bci(); |
4789 | } |
4790 | |
4791 | bool negative_length_guard_generated = false; |
4792 | |
4793 | if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && |
4794 | can_emit_guards && |
4795 | !src->is_top() && !dest->is_top()) { |
4796 | // validate arguments: enables transformation the ArrayCopyNode |
4797 | validated = true; |
4798 | |
4799 | RegionNode* slow_region = new RegionNode(1); |
4800 | record_for_igvn(slow_region); |
4801 | |
4802 | // (1) src and dest are arrays. |
4803 | generate_non_array_guard(load_object_klass(src), slow_region); |
4804 | generate_non_array_guard(load_object_klass(dest), slow_region); |
4805 | |
4806 | // (2) src and dest arrays must have elements of the same BasicType |
4807 | // done at macro expansion or at Ideal transformation time |
4808 | |
4809 | // (4) src_offset must not be negative. |
4810 | generate_negative_guard(src_offset, slow_region); |
4811 | |
4812 | // (5) dest_offset must not be negative. |
4813 | generate_negative_guard(dest_offset, slow_region); |
4814 | |
4815 | // (7) src_offset + length must not exceed length of src. |
4816 | generate_limit_guard(src_offset, length, |
4817 | load_array_length(src), |
4818 | slow_region); |
4819 | |
4820 | // (8) dest_offset + length must not exceed length of dest. |
4821 | generate_limit_guard(dest_offset, length, |
4822 | load_array_length(dest), |
4823 | slow_region); |
4824 | |
4825 | // (6) length must not be negative. |
4826 | // This is also checked in generate_arraycopy() during macro expansion, but |
4827 | // we also have to check it here for the case where the ArrayCopyNode will |
4828 | // be eliminated by Escape Analysis. |
4829 | if (EliminateAllocations) { |
4830 | generate_negative_guard(length, slow_region); |
4831 | negative_length_guard_generated = true; |
4832 | } |
4833 | |
4834 | // (9) each element of an oop array must be assignable |
4835 | Node* src_klass = load_object_klass(src); |
4836 | Node* dest_klass = load_object_klass(dest); |
4837 | Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); |
4838 | |
4839 | if (not_subtype_ctrl != top()) { |
4840 | PreserveJVMState pjvms(this); |
4841 | set_control(not_subtype_ctrl); |
4842 | uncommon_trap(Deoptimization::Reason_intrinsic, |
4843 | Deoptimization::Action_make_not_entrant); |
4844 | assert(stopped(), "Should be stopped" ); |
4845 | } |
4846 | { |
4847 | PreserveJVMState pjvms(this); |
4848 | set_control(_gvn.transform(slow_region)); |
4849 | uncommon_trap(Deoptimization::Reason_intrinsic, |
4850 | Deoptimization::Action_make_not_entrant); |
4851 | assert(stopped(), "Should be stopped" ); |
4852 | } |
4853 | |
4854 | const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); |
4855 | const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); |
4856 | src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); |
4857 | } |
4858 | |
4859 | arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); |
4860 | |
4861 | if (stopped()) { |
4862 | return true; |
4863 | } |
4864 | |
4865 | Node* new_src = access_resolve(src, ACCESS_READ); |
4866 | Node* new_dest = access_resolve(dest, ACCESS_WRITE); |
4867 | |
4868 | ArrayCopyNode* ac = ArrayCopyNode::make(this, true, new_src, src_offset, new_dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, |
4869 | // Create LoadRange and LoadKlass nodes for use during macro expansion here |
4870 | // so the compiler has a chance to eliminate them: during macro expansion, |
4871 | // we have to set their control (CastPP nodes are eliminated). |
4872 | load_object_klass(src), load_object_klass(dest), |
4873 | load_array_length(src), load_array_length(dest)); |
4874 | |
4875 | ac->set_arraycopy(validated); |
4876 | |
4877 | Node* n = _gvn.transform(ac); |
4878 | if (n == ac) { |
4879 | ac->connect_outputs(this); |
4880 | } else { |
4881 | assert(validated, "shouldn't transform if all arguments not validated" ); |
4882 | set_all_memory(n); |
4883 | } |
4884 | clear_upper_avx(); |
4885 | |
4886 | |
4887 | return true; |
4888 | } |
4889 | |
4890 | |
4891 | // Helper function which determines if an arraycopy immediately follows |
4892 | // an allocation, with no intervening tests or other escapes for the object. |
4893 | AllocateArrayNode* |
4894 | LibraryCallKit::tightly_coupled_allocation(Node* ptr, |
4895 | RegionNode* slow_region) { |
4896 | if (stopped()) return NULL; // no fast path |
4897 | if (C->AliasLevel() == 0) return NULL; // no MergeMems around |
4898 | |
4899 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); |
4900 | if (alloc == NULL) return NULL; |
4901 | |
4902 | Node* rawmem = memory(Compile::AliasIdxRaw); |
4903 | // Is the allocation's memory state untouched? |
4904 | if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { |
4905 | // Bail out if there have been raw-memory effects since the allocation. |
4906 | // (Example: There might have been a call or safepoint.) |
4907 | return NULL; |
4908 | } |
4909 | rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); |
4910 | if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { |
4911 | return NULL; |
4912 | } |
4913 | |
4914 | // There must be no unexpected observers of this allocation. |
4915 | for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { |
4916 | Node* obs = ptr->fast_out(i); |
4917 | if (obs != this->map()) { |
4918 | return NULL; |
4919 | } |
4920 | } |
4921 | |
4922 | // This arraycopy must unconditionally follow the allocation of the ptr. |
4923 | Node* alloc_ctl = ptr->in(0); |
4924 | assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo" ); |
4925 | |
4926 | Node* ctl = control(); |
4927 | while (ctl != alloc_ctl) { |
4928 | // There may be guards which feed into the slow_region. |
4929 | // Any other control flow means that we might not get a chance |
4930 | // to finish initializing the allocated object. |
4931 | if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { |
4932 | IfNode* iff = ctl->in(0)->as_If(); |
4933 | Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); |
4934 | assert(not_ctl != NULL && not_ctl != ctl, "found alternate" ); |
4935 | if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { |
4936 | ctl = iff->in(0); // This test feeds the known slow_region. |
4937 | continue; |
4938 | } |
4939 | // One more try: Various low-level checks bottom out in |
4940 | // uncommon traps. If the debug-info of the trap omits |
4941 | // any reference to the allocation, as we've already |
4942 | // observed, then there can be no objection to the trap. |
4943 | bool found_trap = false; |
4944 | for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { |
4945 | Node* obs = not_ctl->fast_out(j); |
4946 | if (obs->in(0) == not_ctl && obs->is_Call() && |
4947 | (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { |
4948 | found_trap = true; break; |
4949 | } |
4950 | } |
4951 | if (found_trap) { |
4952 | ctl = iff->in(0); // This test feeds a harmless uncommon trap. |
4953 | continue; |
4954 | } |
4955 | } |
4956 | return NULL; |
4957 | } |
4958 | |
4959 | // If we get this far, we have an allocation which immediately |
4960 | // precedes the arraycopy, and we can take over zeroing the new object. |
4961 | // The arraycopy will finish the initialization, and provide |
4962 | // a new control state to which we will anchor the destination pointer. |
4963 | |
4964 | return alloc; |
4965 | } |
4966 | |
4967 | //-------------inline_encodeISOArray----------------------------------- |
4968 | // encode char[] to byte[] in ISO_8859_1 |
4969 | bool LibraryCallKit::inline_encodeISOArray() { |
4970 | assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters" ); |
4971 | // no receiver since it is static method |
4972 | Node *src = argument(0); |
4973 | Node *src_offset = argument(1); |
4974 | Node *dst = argument(2); |
4975 | Node *dst_offset = argument(3); |
4976 | Node *length = argument(4); |
4977 | |
4978 | src = must_be_not_null(src, true); |
4979 | dst = must_be_not_null(dst, true); |
4980 | |
4981 | src = access_resolve(src, ACCESS_READ); |
4982 | dst = access_resolve(dst, ACCESS_WRITE); |
4983 | |
4984 | const Type* src_type = src->Value(&_gvn); |
4985 | const Type* dst_type = dst->Value(&_gvn); |
4986 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
4987 | const TypeAryPtr* top_dest = dst_type->isa_aryptr(); |
4988 | if (top_src == NULL || top_src->klass() == NULL || |
4989 | top_dest == NULL || top_dest->klass() == NULL) { |
4990 | // failed array check |
4991 | return false; |
4992 | } |
4993 | |
4994 | // Figure out the size and type of the elements we will be copying. |
4995 | BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
4996 | BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
4997 | if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { |
4998 | return false; |
4999 | } |
5000 | |
5001 | Node* src_start = array_element_address(src, src_offset, T_CHAR); |
5002 | Node* dst_start = array_element_address(dst, dst_offset, dst_elem); |
5003 | // 'src_start' points to src array + scaled offset |
5004 | // 'dst_start' points to dst array + scaled offset |
5005 | |
5006 | const TypeAryPtr* mtype = TypeAryPtr::BYTES; |
5007 | Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); |
5008 | enc = _gvn.transform(enc); |
5009 | Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); |
5010 | set_memory(res_mem, mtype); |
5011 | set_result(enc); |
5012 | clear_upper_avx(); |
5013 | |
5014 | return true; |
5015 | } |
5016 | |
5017 | //-------------inline_multiplyToLen----------------------------------- |
5018 | bool LibraryCallKit::inline_multiplyToLen() { |
5019 | assert(UseMultiplyToLenIntrinsic, "not implemented on this platform" ); |
5020 | |
5021 | address stubAddr = StubRoutines::multiplyToLen(); |
5022 | if (stubAddr == NULL) { |
5023 | return false; // Intrinsic's stub is not implemented on this platform |
5024 | } |
5025 | const char* stubName = "multiplyToLen" ; |
5026 | |
5027 | assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters" ); |
5028 | |
5029 | // no receiver because it is a static method |
5030 | Node* x = argument(0); |
5031 | Node* xlen = argument(1); |
5032 | Node* y = argument(2); |
5033 | Node* ylen = argument(3); |
5034 | Node* z = argument(4); |
5035 | |
5036 | x = must_be_not_null(x, true); |
5037 | y = must_be_not_null(y, true); |
5038 | |
5039 | x = access_resolve(x, ACCESS_READ); |
5040 | y = access_resolve(y, ACCESS_READ); |
5041 | z = access_resolve(z, ACCESS_WRITE); |
5042 | |
5043 | const Type* x_type = x->Value(&_gvn); |
5044 | const Type* y_type = y->Value(&_gvn); |
5045 | const TypeAryPtr* top_x = x_type->isa_aryptr(); |
5046 | const TypeAryPtr* top_y = y_type->isa_aryptr(); |
5047 | if (top_x == NULL || top_x->klass() == NULL || |
5048 | top_y == NULL || top_y->klass() == NULL) { |
5049 | // failed array check |
5050 | return false; |
5051 | } |
5052 | |
5053 | BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5054 | BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5055 | if (x_elem != T_INT || y_elem != T_INT) { |
5056 | return false; |
5057 | } |
5058 | |
5059 | // Set the original stack and the reexecute bit for the interpreter to reexecute |
5060 | // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens |
5061 | // on the return from z array allocation in runtime. |
5062 | { PreserveReexecuteState preexecs(this); |
5063 | jvms()->set_should_reexecute(true); |
5064 | |
5065 | Node* x_start = array_element_address(x, intcon(0), x_elem); |
5066 | Node* y_start = array_element_address(y, intcon(0), y_elem); |
5067 | // 'x_start' points to x array + scaled xlen |
5068 | // 'y_start' points to y array + scaled ylen |
5069 | |
5070 | // Allocate the result array |
5071 | Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); |
5072 | ciKlass* klass = ciTypeArrayKlass::make(T_INT); |
5073 | Node* klass_node = makecon(TypeKlassPtr::make(klass)); |
5074 | |
5075 | IdealKit ideal(this); |
5076 | |
5077 | #define __ ideal. |
5078 | Node* one = __ ConI(1); |
5079 | Node* zero = __ ConI(0); |
5080 | IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); |
5081 | __ set(need_alloc, zero); |
5082 | __ set(z_alloc, z); |
5083 | __ if_then(z, BoolTest::eq, null()); { |
5084 | __ increment (need_alloc, one); |
5085 | } __ else_(); { |
5086 | // Update graphKit memory and control from IdealKit. |
5087 | sync_kit(ideal); |
5088 | Node *cast = new CastPPNode(z, TypePtr::NOTNULL); |
5089 | cast->init_req(0, control()); |
5090 | _gvn.set_type(cast, cast->bottom_type()); |
5091 | C->record_for_igvn(cast); |
5092 | |
5093 | Node* zlen_arg = load_array_length(cast); |
5094 | // Update IdealKit memory and control from graphKit. |
5095 | __ sync_kit(this); |
5096 | __ if_then(zlen_arg, BoolTest::lt, zlen); { |
5097 | __ increment (need_alloc, one); |
5098 | } __ end_if(); |
5099 | } __ end_if(); |
5100 | |
5101 | __ if_then(__ value(need_alloc), BoolTest::ne, zero); { |
5102 | // Update graphKit memory and control from IdealKit. |
5103 | sync_kit(ideal); |
5104 | Node * narr = new_array(klass_node, zlen, 1); |
5105 | // Update IdealKit memory and control from graphKit. |
5106 | __ sync_kit(this); |
5107 | __ set(z_alloc, narr); |
5108 | } __ end_if(); |
5109 | |
5110 | sync_kit(ideal); |
5111 | z = __ value(z_alloc); |
5112 | // Can't use TypeAryPtr::INTS which uses Bottom offset. |
5113 | _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); |
5114 | // Final sync IdealKit and GraphKit. |
5115 | final_sync(ideal); |
5116 | #undef __ |
5117 | |
5118 | Node* z_start = array_element_address(z, intcon(0), T_INT); |
5119 | |
5120 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
5121 | OptoRuntime::multiplyToLen_Type(), |
5122 | stubAddr, stubName, TypePtr::BOTTOM, |
5123 | x_start, xlen, y_start, ylen, z_start, zlen); |
5124 | } // original reexecute is set back here |
5125 | |
5126 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
5127 | set_result(z); |
5128 | return true; |
5129 | } |
5130 | |
5131 | //-------------inline_squareToLen------------------------------------ |
5132 | bool LibraryCallKit::inline_squareToLen() { |
5133 | assert(UseSquareToLenIntrinsic, "not implemented on this platform" ); |
5134 | |
5135 | address stubAddr = StubRoutines::squareToLen(); |
5136 | if (stubAddr == NULL) { |
5137 | return false; // Intrinsic's stub is not implemented on this platform |
5138 | } |
5139 | const char* stubName = "squareToLen" ; |
5140 | |
5141 | assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters" ); |
5142 | |
5143 | Node* x = argument(0); |
5144 | Node* len = argument(1); |
5145 | Node* z = argument(2); |
5146 | Node* zlen = argument(3); |
5147 | |
5148 | x = must_be_not_null(x, true); |
5149 | z = must_be_not_null(z, true); |
5150 | |
5151 | x = access_resolve(x, ACCESS_READ); |
5152 | z = access_resolve(z, ACCESS_WRITE); |
5153 | |
5154 | const Type* x_type = x->Value(&_gvn); |
5155 | const Type* z_type = z->Value(&_gvn); |
5156 | const TypeAryPtr* top_x = x_type->isa_aryptr(); |
5157 | const TypeAryPtr* top_z = z_type->isa_aryptr(); |
5158 | if (top_x == NULL || top_x->klass() == NULL || |
5159 | top_z == NULL || top_z->klass() == NULL) { |
5160 | // failed array check |
5161 | return false; |
5162 | } |
5163 | |
5164 | BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5165 | BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5166 | if (x_elem != T_INT || z_elem != T_INT) { |
5167 | return false; |
5168 | } |
5169 | |
5170 | |
5171 | Node* x_start = array_element_address(x, intcon(0), x_elem); |
5172 | Node* z_start = array_element_address(z, intcon(0), z_elem); |
5173 | |
5174 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
5175 | OptoRuntime::squareToLen_Type(), |
5176 | stubAddr, stubName, TypePtr::BOTTOM, |
5177 | x_start, len, z_start, zlen); |
5178 | |
5179 | set_result(z); |
5180 | return true; |
5181 | } |
5182 | |
5183 | //-------------inline_mulAdd------------------------------------------ |
5184 | bool LibraryCallKit::inline_mulAdd() { |
5185 | assert(UseMulAddIntrinsic, "not implemented on this platform" ); |
5186 | |
5187 | address stubAddr = StubRoutines::mulAdd(); |
5188 | if (stubAddr == NULL) { |
5189 | return false; // Intrinsic's stub is not implemented on this platform |
5190 | } |
5191 | const char* stubName = "mulAdd" ; |
5192 | |
5193 | assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters" ); |
5194 | |
5195 | Node* out = argument(0); |
5196 | Node* in = argument(1); |
5197 | Node* offset = argument(2); |
5198 | Node* len = argument(3); |
5199 | Node* k = argument(4); |
5200 | |
5201 | out = must_be_not_null(out, true); |
5202 | |
5203 | in = access_resolve(in, ACCESS_READ); |
5204 | out = access_resolve(out, ACCESS_WRITE); |
5205 | |
5206 | const Type* out_type = out->Value(&_gvn); |
5207 | const Type* in_type = in->Value(&_gvn); |
5208 | const TypeAryPtr* top_out = out_type->isa_aryptr(); |
5209 | const TypeAryPtr* top_in = in_type->isa_aryptr(); |
5210 | if (top_out == NULL || top_out->klass() == NULL || |
5211 | top_in == NULL || top_in->klass() == NULL) { |
5212 | // failed array check |
5213 | return false; |
5214 | } |
5215 | |
5216 | BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5217 | BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5218 | if (out_elem != T_INT || in_elem != T_INT) { |
5219 | return false; |
5220 | } |
5221 | |
5222 | Node* outlen = load_array_length(out); |
5223 | Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); |
5224 | Node* out_start = array_element_address(out, intcon(0), out_elem); |
5225 | Node* in_start = array_element_address(in, intcon(0), in_elem); |
5226 | |
5227 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
5228 | OptoRuntime::mulAdd_Type(), |
5229 | stubAddr, stubName, TypePtr::BOTTOM, |
5230 | out_start,in_start, new_offset, len, k); |
5231 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5232 | set_result(result); |
5233 | return true; |
5234 | } |
5235 | |
5236 | //-------------inline_montgomeryMultiply----------------------------------- |
5237 | bool LibraryCallKit::inline_montgomeryMultiply() { |
5238 | address stubAddr = StubRoutines::montgomeryMultiply(); |
5239 | if (stubAddr == NULL) { |
5240 | return false; // Intrinsic's stub is not implemented on this platform |
5241 | } |
5242 | |
5243 | assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform" ); |
5244 | const char* stubName = "montgomery_multiply" ; |
5245 | |
5246 | assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters" ); |
5247 | |
5248 | Node* a = argument(0); |
5249 | Node* b = argument(1); |
5250 | Node* n = argument(2); |
5251 | Node* len = argument(3); |
5252 | Node* inv = argument(4); |
5253 | Node* m = argument(6); |
5254 | |
5255 | a = access_resolve(a, ACCESS_READ); |
5256 | b = access_resolve(b, ACCESS_READ); |
5257 | n = access_resolve(n, ACCESS_READ); |
5258 | m = access_resolve(m, ACCESS_WRITE); |
5259 | |
5260 | const Type* a_type = a->Value(&_gvn); |
5261 | const TypeAryPtr* top_a = a_type->isa_aryptr(); |
5262 | const Type* b_type = b->Value(&_gvn); |
5263 | const TypeAryPtr* top_b = b_type->isa_aryptr(); |
5264 | const Type* n_type = a->Value(&_gvn); |
5265 | const TypeAryPtr* top_n = n_type->isa_aryptr(); |
5266 | const Type* m_type = a->Value(&_gvn); |
5267 | const TypeAryPtr* top_m = m_type->isa_aryptr(); |
5268 | if (top_a == NULL || top_a->klass() == NULL || |
5269 | top_b == NULL || top_b->klass() == NULL || |
5270 | top_n == NULL || top_n->klass() == NULL || |
5271 | top_m == NULL || top_m->klass() == NULL) { |
5272 | // failed array check |
5273 | return false; |
5274 | } |
5275 | |
5276 | BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5277 | BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5278 | BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5279 | BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5280 | if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { |
5281 | return false; |
5282 | } |
5283 | |
5284 | // Make the call |
5285 | { |
5286 | Node* a_start = array_element_address(a, intcon(0), a_elem); |
5287 | Node* b_start = array_element_address(b, intcon(0), b_elem); |
5288 | Node* n_start = array_element_address(n, intcon(0), n_elem); |
5289 | Node* m_start = array_element_address(m, intcon(0), m_elem); |
5290 | |
5291 | Node* call = make_runtime_call(RC_LEAF, |
5292 | OptoRuntime::montgomeryMultiply_Type(), |
5293 | stubAddr, stubName, TypePtr::BOTTOM, |
5294 | a_start, b_start, n_start, len, inv, top(), |
5295 | m_start); |
5296 | set_result(m); |
5297 | } |
5298 | |
5299 | return true; |
5300 | } |
5301 | |
5302 | bool LibraryCallKit::inline_montgomerySquare() { |
5303 | address stubAddr = StubRoutines::montgomerySquare(); |
5304 | if (stubAddr == NULL) { |
5305 | return false; // Intrinsic's stub is not implemented on this platform |
5306 | } |
5307 | |
5308 | assert(UseMontgomerySquareIntrinsic, "not implemented on this platform" ); |
5309 | const char* stubName = "montgomery_square" ; |
5310 | |
5311 | assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters" ); |
5312 | |
5313 | Node* a = argument(0); |
5314 | Node* n = argument(1); |
5315 | Node* len = argument(2); |
5316 | Node* inv = argument(3); |
5317 | Node* m = argument(5); |
5318 | |
5319 | a = access_resolve(a, ACCESS_READ); |
5320 | n = access_resolve(n, ACCESS_READ); |
5321 | m = access_resolve(m, ACCESS_WRITE); |
5322 | |
5323 | const Type* a_type = a->Value(&_gvn); |
5324 | const TypeAryPtr* top_a = a_type->isa_aryptr(); |
5325 | const Type* n_type = a->Value(&_gvn); |
5326 | const TypeAryPtr* top_n = n_type->isa_aryptr(); |
5327 | const Type* m_type = a->Value(&_gvn); |
5328 | const TypeAryPtr* top_m = m_type->isa_aryptr(); |
5329 | if (top_a == NULL || top_a->klass() == NULL || |
5330 | top_n == NULL || top_n->klass() == NULL || |
5331 | top_m == NULL || top_m->klass() == NULL) { |
5332 | // failed array check |
5333 | return false; |
5334 | } |
5335 | |
5336 | BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5337 | BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5338 | BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5339 | if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { |
5340 | return false; |
5341 | } |
5342 | |
5343 | // Make the call |
5344 | { |
5345 | Node* a_start = array_element_address(a, intcon(0), a_elem); |
5346 | Node* n_start = array_element_address(n, intcon(0), n_elem); |
5347 | Node* m_start = array_element_address(m, intcon(0), m_elem); |
5348 | |
5349 | Node* call = make_runtime_call(RC_LEAF, |
5350 | OptoRuntime::montgomerySquare_Type(), |
5351 | stubAddr, stubName, TypePtr::BOTTOM, |
5352 | a_start, n_start, len, inv, top(), |
5353 | m_start); |
5354 | set_result(m); |
5355 | } |
5356 | |
5357 | return true; |
5358 | } |
5359 | |
5360 | //-------------inline_vectorizedMismatch------------------------------ |
5361 | bool LibraryCallKit::inline_vectorizedMismatch() { |
5362 | assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform" ); |
5363 | |
5364 | address stubAddr = StubRoutines::vectorizedMismatch(); |
5365 | if (stubAddr == NULL) { |
5366 | return false; // Intrinsic's stub is not implemented on this platform |
5367 | } |
5368 | const char* stubName = "vectorizedMismatch" ; |
5369 | int size_l = callee()->signature()->size(); |
5370 | assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters" ); |
5371 | |
5372 | Node* obja = argument(0); |
5373 | Node* aoffset = argument(1); |
5374 | Node* objb = argument(3); |
5375 | Node* boffset = argument(4); |
5376 | Node* length = argument(6); |
5377 | Node* scale = argument(7); |
5378 | |
5379 | const Type* a_type = obja->Value(&_gvn); |
5380 | const Type* b_type = objb->Value(&_gvn); |
5381 | const TypeAryPtr* top_a = a_type->isa_aryptr(); |
5382 | const TypeAryPtr* top_b = b_type->isa_aryptr(); |
5383 | if (top_a == NULL || top_a->klass() == NULL || |
5384 | top_b == NULL || top_b->klass() == NULL) { |
5385 | // failed array check |
5386 | return false; |
5387 | } |
5388 | |
5389 | Node* call; |
5390 | jvms()->set_should_reexecute(true); |
5391 | |
5392 | obja = access_resolve(obja, ACCESS_READ); |
5393 | objb = access_resolve(objb, ACCESS_READ); |
5394 | Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ); |
5395 | Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ); |
5396 | |
5397 | call = make_runtime_call(RC_LEAF, |
5398 | OptoRuntime::vectorizedMismatch_Type(), |
5399 | stubAddr, stubName, TypePtr::BOTTOM, |
5400 | obja_adr, objb_adr, length, scale); |
5401 | |
5402 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5403 | set_result(result); |
5404 | return true; |
5405 | } |
5406 | |
5407 | /** |
5408 | * Calculate CRC32 for byte. |
5409 | * int java.util.zip.CRC32.update(int crc, int b) |
5410 | */ |
5411 | bool LibraryCallKit::inline_updateCRC32() { |
5412 | assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support" ); |
5413 | assert(callee()->signature()->size() == 2, "update has 2 parameters" ); |
5414 | // no receiver since it is static method |
5415 | Node* crc = argument(0); // type: int |
5416 | Node* b = argument(1); // type: int |
5417 | |
5418 | /* |
5419 | * int c = ~ crc; |
5420 | * b = timesXtoThe32[(b ^ c) & 0xFF]; |
5421 | * b = b ^ (c >>> 8); |
5422 | * crc = ~b; |
5423 | */ |
5424 | |
5425 | Node* M1 = intcon(-1); |
5426 | crc = _gvn.transform(new XorINode(crc, M1)); |
5427 | Node* result = _gvn.transform(new XorINode(crc, b)); |
5428 | result = _gvn.transform(new AndINode(result, intcon(0xFF))); |
5429 | |
5430 | Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); |
5431 | Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); |
5432 | Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); |
5433 | result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); |
5434 | |
5435 | crc = _gvn.transform(new URShiftINode(crc, intcon(8))); |
5436 | result = _gvn.transform(new XorINode(crc, result)); |
5437 | result = _gvn.transform(new XorINode(result, M1)); |
5438 | set_result(result); |
5439 | return true; |
5440 | } |
5441 | |
5442 | /** |
5443 | * Calculate CRC32 for byte[] array. |
5444 | * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) |
5445 | */ |
5446 | bool LibraryCallKit::inline_updateBytesCRC32() { |
5447 | assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support" ); |
5448 | assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters" ); |
5449 | // no receiver since it is static method |
5450 | Node* crc = argument(0); // type: int |
5451 | Node* src = argument(1); // type: oop |
5452 | Node* offset = argument(2); // type: int |
5453 | Node* length = argument(3); // type: int |
5454 | |
5455 | const Type* src_type = src->Value(&_gvn); |
5456 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
5457 | if (top_src == NULL || top_src->klass() == NULL) { |
5458 | // failed array check |
5459 | return false; |
5460 | } |
5461 | |
5462 | // Figure out the size and type of the elements we will be copying. |
5463 | BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5464 | if (src_elem != T_BYTE) { |
5465 | return false; |
5466 | } |
5467 | |
5468 | // 'src_start' points to src array + scaled offset |
5469 | src = must_be_not_null(src, true); |
5470 | src = access_resolve(src, ACCESS_READ); |
5471 | Node* src_start = array_element_address(src, offset, src_elem); |
5472 | |
5473 | // We assume that range check is done by caller. |
5474 | // TODO: generate range check (offset+length < src.length) in debug VM. |
5475 | |
5476 | // Call the stub. |
5477 | address stubAddr = StubRoutines::updateBytesCRC32(); |
5478 | const char *stubName = "updateBytesCRC32" ; |
5479 | |
5480 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), |
5481 | stubAddr, stubName, TypePtr::BOTTOM, |
5482 | crc, src_start, length); |
5483 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5484 | set_result(result); |
5485 | return true; |
5486 | } |
5487 | |
5488 | /** |
5489 | * Calculate CRC32 for ByteBuffer. |
5490 | * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) |
5491 | */ |
5492 | bool LibraryCallKit::inline_updateByteBufferCRC32() { |
5493 | assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support" ); |
5494 | assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long" ); |
5495 | // no receiver since it is static method |
5496 | Node* crc = argument(0); // type: int |
5497 | Node* src = argument(1); // type: long |
5498 | Node* offset = argument(3); // type: int |
5499 | Node* length = argument(4); // type: int |
5500 | |
5501 | src = ConvL2X(src); // adjust Java long to machine word |
5502 | Node* base = _gvn.transform(new CastX2PNode(src)); |
5503 | offset = ConvI2X(offset); |
5504 | |
5505 | // 'src_start' points to src array + scaled offset |
5506 | Node* src_start = basic_plus_adr(top(), base, offset); |
5507 | |
5508 | // Call the stub. |
5509 | address stubAddr = StubRoutines::updateBytesCRC32(); |
5510 | const char *stubName = "updateBytesCRC32" ; |
5511 | |
5512 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), |
5513 | stubAddr, stubName, TypePtr::BOTTOM, |
5514 | crc, src_start, length); |
5515 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5516 | set_result(result); |
5517 | return true; |
5518 | } |
5519 | |
5520 | //------------------------------get_table_from_crc32c_class----------------------- |
5521 | Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { |
5522 | Node* table = load_field_from_object(NULL, "byteTable" , "[I" , /*is_exact*/ false, /*is_static*/ true, crc32c_class); |
5523 | assert (table != NULL, "wrong version of java.util.zip.CRC32C" ); |
5524 | |
5525 | return table; |
5526 | } |
5527 | |
5528 | //------------------------------inline_updateBytesCRC32C----------------------- |
5529 | // |
5530 | // Calculate CRC32C for byte[] array. |
5531 | // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) |
5532 | // |
5533 | bool LibraryCallKit::inline_updateBytesCRC32C() { |
5534 | assert(UseCRC32CIntrinsics, "need CRC32C instruction support" ); |
5535 | assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters" ); |
5536 | assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded" ); |
5537 | // no receiver since it is a static method |
5538 | Node* crc = argument(0); // type: int |
5539 | Node* src = argument(1); // type: oop |
5540 | Node* offset = argument(2); // type: int |
5541 | Node* end = argument(3); // type: int |
5542 | |
5543 | Node* length = _gvn.transform(new SubINode(end, offset)); |
5544 | |
5545 | const Type* src_type = src->Value(&_gvn); |
5546 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
5547 | if (top_src == NULL || top_src->klass() == NULL) { |
5548 | // failed array check |
5549 | return false; |
5550 | } |
5551 | |
5552 | // Figure out the size and type of the elements we will be copying. |
5553 | BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5554 | if (src_elem != T_BYTE) { |
5555 | return false; |
5556 | } |
5557 | |
5558 | // 'src_start' points to src array + scaled offset |
5559 | src = must_be_not_null(src, true); |
5560 | src = access_resolve(src, ACCESS_READ); |
5561 | Node* src_start = array_element_address(src, offset, src_elem); |
5562 | |
5563 | // static final int[] byteTable in class CRC32C |
5564 | Node* table = get_table_from_crc32c_class(callee()->holder()); |
5565 | table = must_be_not_null(table, true); |
5566 | table = access_resolve(table, ACCESS_READ); |
5567 | Node* table_start = array_element_address(table, intcon(0), T_INT); |
5568 | |
5569 | // We assume that range check is done by caller. |
5570 | // TODO: generate range check (offset+length < src.length) in debug VM. |
5571 | |
5572 | // Call the stub. |
5573 | address stubAddr = StubRoutines::updateBytesCRC32C(); |
5574 | const char *stubName = "updateBytesCRC32C" ; |
5575 | |
5576 | Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), |
5577 | stubAddr, stubName, TypePtr::BOTTOM, |
5578 | crc, src_start, length, table_start); |
5579 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5580 | set_result(result); |
5581 | return true; |
5582 | } |
5583 | |
5584 | //------------------------------inline_updateDirectByteBufferCRC32C----------------------- |
5585 | // |
5586 | // Calculate CRC32C for DirectByteBuffer. |
5587 | // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) |
5588 | // |
5589 | bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { |
5590 | assert(UseCRC32CIntrinsics, "need CRC32C instruction support" ); |
5591 | assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long" ); |
5592 | assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded" ); |
5593 | // no receiver since it is a static method |
5594 | Node* crc = argument(0); // type: int |
5595 | Node* src = argument(1); // type: long |
5596 | Node* offset = argument(3); // type: int |
5597 | Node* end = argument(4); // type: int |
5598 | |
5599 | Node* length = _gvn.transform(new SubINode(end, offset)); |
5600 | |
5601 | src = ConvL2X(src); // adjust Java long to machine word |
5602 | Node* base = _gvn.transform(new CastX2PNode(src)); |
5603 | offset = ConvI2X(offset); |
5604 | |
5605 | // 'src_start' points to src array + scaled offset |
5606 | Node* src_start = basic_plus_adr(top(), base, offset); |
5607 | |
5608 | // static final int[] byteTable in class CRC32C |
5609 | Node* table = get_table_from_crc32c_class(callee()->holder()); |
5610 | table = must_be_not_null(table, true); |
5611 | table = access_resolve(table, ACCESS_READ); |
5612 | Node* table_start = array_element_address(table, intcon(0), T_INT); |
5613 | |
5614 | // Call the stub. |
5615 | address stubAddr = StubRoutines::updateBytesCRC32C(); |
5616 | const char *stubName = "updateBytesCRC32C" ; |
5617 | |
5618 | Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), |
5619 | stubAddr, stubName, TypePtr::BOTTOM, |
5620 | crc, src_start, length, table_start); |
5621 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5622 | set_result(result); |
5623 | return true; |
5624 | } |
5625 | |
5626 | //------------------------------inline_updateBytesAdler32---------------------- |
5627 | // |
5628 | // Calculate Adler32 checksum for byte[] array. |
5629 | // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) |
5630 | // |
5631 | bool LibraryCallKit::inline_updateBytesAdler32() { |
5632 | assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need" ); // check if we actually need to check this flag or check a different one |
5633 | assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters" ); |
5634 | assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded" ); |
5635 | // no receiver since it is static method |
5636 | Node* crc = argument(0); // type: int |
5637 | Node* src = argument(1); // type: oop |
5638 | Node* offset = argument(2); // type: int |
5639 | Node* length = argument(3); // type: int |
5640 | |
5641 | const Type* src_type = src->Value(&_gvn); |
5642 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
5643 | if (top_src == NULL || top_src->klass() == NULL) { |
5644 | // failed array check |
5645 | return false; |
5646 | } |
5647 | |
5648 | // Figure out the size and type of the elements we will be copying. |
5649 | BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
5650 | if (src_elem != T_BYTE) { |
5651 | return false; |
5652 | } |
5653 | |
5654 | // 'src_start' points to src array + scaled offset |
5655 | src = access_resolve(src, ACCESS_READ); |
5656 | Node* src_start = array_element_address(src, offset, src_elem); |
5657 | |
5658 | // We assume that range check is done by caller. |
5659 | // TODO: generate range check (offset+length < src.length) in debug VM. |
5660 | |
5661 | // Call the stub. |
5662 | address stubAddr = StubRoutines::updateBytesAdler32(); |
5663 | const char *stubName = "updateBytesAdler32" ; |
5664 | |
5665 | Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), |
5666 | stubAddr, stubName, TypePtr::BOTTOM, |
5667 | crc, src_start, length); |
5668 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5669 | set_result(result); |
5670 | return true; |
5671 | } |
5672 | |
5673 | //------------------------------inline_updateByteBufferAdler32--------------- |
5674 | // |
5675 | // Calculate Adler32 checksum for DirectByteBuffer. |
5676 | // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) |
5677 | // |
5678 | bool LibraryCallKit::inline_updateByteBufferAdler32() { |
5679 | assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need" ); // check if we actually need to check this flag or check a different one |
5680 | assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long" ); |
5681 | assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded" ); |
5682 | // no receiver since it is static method |
5683 | Node* crc = argument(0); // type: int |
5684 | Node* src = argument(1); // type: long |
5685 | Node* offset = argument(3); // type: int |
5686 | Node* length = argument(4); // type: int |
5687 | |
5688 | src = ConvL2X(src); // adjust Java long to machine word |
5689 | Node* base = _gvn.transform(new CastX2PNode(src)); |
5690 | offset = ConvI2X(offset); |
5691 | |
5692 | // 'src_start' points to src array + scaled offset |
5693 | Node* src_start = basic_plus_adr(top(), base, offset); |
5694 | |
5695 | // Call the stub. |
5696 | address stubAddr = StubRoutines::updateBytesAdler32(); |
5697 | const char *stubName = "updateBytesAdler32" ; |
5698 | |
5699 | Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), |
5700 | stubAddr, stubName, TypePtr::BOTTOM, |
5701 | crc, src_start, length); |
5702 | |
5703 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
5704 | set_result(result); |
5705 | return true; |
5706 | } |
5707 | |
5708 | //----------------------------inline_reference_get---------------------------- |
5709 | // public T java.lang.ref.Reference.get(); |
5710 | bool LibraryCallKit::inline_reference_get() { |
5711 | const int referent_offset = java_lang_ref_Reference::referent_offset; |
5712 | guarantee(referent_offset > 0, "should have already been set" ); |
5713 | |
5714 | // Get the argument: |
5715 | Node* reference_obj = null_check_receiver(); |
5716 | if (stopped()) return true; |
5717 | |
5718 | const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr(); |
5719 | assert(tinst != NULL, "obj is null" ); |
5720 | assert(tinst->klass()->is_loaded(), "obj is not loaded" ); |
5721 | ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass(); |
5722 | ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent" ), |
5723 | ciSymbol::make("Ljava/lang/Object;" ), |
5724 | false); |
5725 | assert (field != NULL, "undefined field" ); |
5726 | |
5727 | Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); |
5728 | const TypePtr* adr_type = C->alias_type(field)->adr_type(); |
5729 | |
5730 | ciInstanceKlass* klass = env()->Object_klass(); |
5731 | const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); |
5732 | |
5733 | DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF; |
5734 | Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators); |
5735 | // Add memory barrier to prevent commoning reads from this field |
5736 | // across safepoint since GC can change its value. |
5737 | insert_mem_bar(Op_MemBarCPUOrder); |
5738 | |
5739 | set_result(result); |
5740 | return true; |
5741 | } |
5742 | |
5743 | |
5744 | Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, |
5745 | bool is_exact=true, bool is_static=false, |
5746 | ciInstanceKlass * fromKls=NULL) { |
5747 | if (fromKls == NULL) { |
5748 | const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); |
5749 | assert(tinst != NULL, "obj is null" ); |
5750 | assert(tinst->klass()->is_loaded(), "obj is not loaded" ); |
5751 | assert(!is_exact || tinst->klass_is_exact(), "klass not exact" ); |
5752 | fromKls = tinst->klass()->as_instance_klass(); |
5753 | } else { |
5754 | assert(is_static, "only for static field access" ); |
5755 | } |
5756 | ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), |
5757 | ciSymbol::make(fieldTypeString), |
5758 | is_static); |
5759 | |
5760 | assert (field != NULL, "undefined field" ); |
5761 | if (field == NULL) return (Node *) NULL; |
5762 | |
5763 | if (is_static) { |
5764 | const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); |
5765 | fromObj = makecon(tip); |
5766 | } |
5767 | |
5768 | // Next code copied from Parse::do_get_xxx(): |
5769 | |
5770 | // Compute address and memory type. |
5771 | int offset = field->offset_in_bytes(); |
5772 | bool is_vol = field->is_volatile(); |
5773 | ciType* field_klass = field->type(); |
5774 | assert(field_klass->is_loaded(), "should be loaded" ); |
5775 | const TypePtr* adr_type = C->alias_type(field)->adr_type(); |
5776 | Node *adr = basic_plus_adr(fromObj, fromObj, offset); |
5777 | BasicType bt = field->layout_type(); |
5778 | |
5779 | // Build the resultant type of the load |
5780 | const Type *type; |
5781 | if (bt == T_OBJECT) { |
5782 | type = TypeOopPtr::make_from_klass(field_klass->as_klass()); |
5783 | } else { |
5784 | type = Type::get_const_basic_type(bt); |
5785 | } |
5786 | |
5787 | DecoratorSet decorators = IN_HEAP; |
5788 | |
5789 | if (is_vol) { |
5790 | decorators |= MO_SEQ_CST; |
5791 | } |
5792 | |
5793 | return access_load_at(fromObj, adr, adr_type, type, bt, decorators); |
5794 | } |
5795 | |
5796 | Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, |
5797 | bool is_exact = true, bool is_static = false, |
5798 | ciInstanceKlass * fromKls = NULL) { |
5799 | if (fromKls == NULL) { |
5800 | const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); |
5801 | assert(tinst != NULL, "obj is null" ); |
5802 | assert(tinst->klass()->is_loaded(), "obj is not loaded" ); |
5803 | assert(!is_exact || tinst->klass_is_exact(), "klass not exact" ); |
5804 | fromKls = tinst->klass()->as_instance_klass(); |
5805 | } |
5806 | else { |
5807 | assert(is_static, "only for static field access" ); |
5808 | } |
5809 | ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), |
5810 | ciSymbol::make(fieldTypeString), |
5811 | is_static); |
5812 | |
5813 | assert(field != NULL, "undefined field" ); |
5814 | assert(!field->is_volatile(), "not defined for volatile fields" ); |
5815 | |
5816 | if (is_static) { |
5817 | const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); |
5818 | fromObj = makecon(tip); |
5819 | } |
5820 | |
5821 | // Next code copied from Parse::do_get_xxx(): |
5822 | |
5823 | // Compute address and memory type. |
5824 | int offset = field->offset_in_bytes(); |
5825 | Node *adr = basic_plus_adr(fromObj, fromObj, offset); |
5826 | |
5827 | return adr; |
5828 | } |
5829 | |
5830 | //------------------------------inline_aescrypt_Block----------------------- |
5831 | bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { |
5832 | address stubAddr = NULL; |
5833 | const char *stubName; |
5834 | assert(UseAES, "need AES instruction support" ); |
5835 | |
5836 | switch(id) { |
5837 | case vmIntrinsics::_aescrypt_encryptBlock: |
5838 | stubAddr = StubRoutines::aescrypt_encryptBlock(); |
5839 | stubName = "aescrypt_encryptBlock" ; |
5840 | break; |
5841 | case vmIntrinsics::_aescrypt_decryptBlock: |
5842 | stubAddr = StubRoutines::aescrypt_decryptBlock(); |
5843 | stubName = "aescrypt_decryptBlock" ; |
5844 | break; |
5845 | default: |
5846 | break; |
5847 | } |
5848 | if (stubAddr == NULL) return false; |
5849 | |
5850 | Node* aescrypt_object = argument(0); |
5851 | Node* src = argument(1); |
5852 | Node* src_offset = argument(2); |
5853 | Node* dest = argument(3); |
5854 | Node* dest_offset = argument(4); |
5855 | |
5856 | src = must_be_not_null(src, true); |
5857 | dest = must_be_not_null(dest, true); |
5858 | |
5859 | src = access_resolve(src, ACCESS_READ); |
5860 | dest = access_resolve(dest, ACCESS_WRITE); |
5861 | |
5862 | // (1) src and dest are arrays. |
5863 | const Type* src_type = src->Value(&_gvn); |
5864 | const Type* dest_type = dest->Value(&_gvn); |
5865 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
5866 | const TypeAryPtr* top_dest = dest_type->isa_aryptr(); |
5867 | assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange" ); |
5868 | |
5869 | // for the quick and dirty code we will skip all the checks. |
5870 | // we are just trying to get the call to be generated. |
5871 | Node* src_start = src; |
5872 | Node* dest_start = dest; |
5873 | if (src_offset != NULL || dest_offset != NULL) { |
5874 | assert(src_offset != NULL && dest_offset != NULL, "" ); |
5875 | src_start = array_element_address(src, src_offset, T_BYTE); |
5876 | dest_start = array_element_address(dest, dest_offset, T_BYTE); |
5877 | } |
5878 | |
5879 | // now need to get the start of its expanded key array |
5880 | // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java |
5881 | Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); |
5882 | if (k_start == NULL) return false; |
5883 | |
5884 | if (Matcher::pass_original_key_for_aes()) { |
5885 | // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to |
5886 | // compatibility issues between Java key expansion and SPARC crypto instructions |
5887 | Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); |
5888 | if (original_k_start == NULL) return false; |
5889 | |
5890 | // Call the stub. |
5891 | make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), |
5892 | stubAddr, stubName, TypePtr::BOTTOM, |
5893 | src_start, dest_start, k_start, original_k_start); |
5894 | } else { |
5895 | // Call the stub. |
5896 | make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), |
5897 | stubAddr, stubName, TypePtr::BOTTOM, |
5898 | src_start, dest_start, k_start); |
5899 | } |
5900 | |
5901 | return true; |
5902 | } |
5903 | |
5904 | //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- |
5905 | bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { |
5906 | address stubAddr = NULL; |
5907 | const char *stubName = NULL; |
5908 | |
5909 | assert(UseAES, "need AES instruction support" ); |
5910 | |
5911 | switch(id) { |
5912 | case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: |
5913 | stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); |
5914 | stubName = "cipherBlockChaining_encryptAESCrypt" ; |
5915 | break; |
5916 | case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: |
5917 | stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); |
5918 | stubName = "cipherBlockChaining_decryptAESCrypt" ; |
5919 | break; |
5920 | default: |
5921 | break; |
5922 | } |
5923 | if (stubAddr == NULL) return false; |
5924 | |
5925 | Node* cipherBlockChaining_object = argument(0); |
5926 | Node* src = argument(1); |
5927 | Node* src_offset = argument(2); |
5928 | Node* len = argument(3); |
5929 | Node* dest = argument(4); |
5930 | Node* dest_offset = argument(5); |
5931 | |
5932 | src = must_be_not_null(src, false); |
5933 | dest = must_be_not_null(dest, false); |
5934 | |
5935 | src = access_resolve(src, ACCESS_READ); |
5936 | dest = access_resolve(dest, ACCESS_WRITE); |
5937 | |
5938 | // (1) src and dest are arrays. |
5939 | const Type* src_type = src->Value(&_gvn); |
5940 | const Type* dest_type = dest->Value(&_gvn); |
5941 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
5942 | const TypeAryPtr* top_dest = dest_type->isa_aryptr(); |
5943 | assert (top_src != NULL && top_src->klass() != NULL |
5944 | && top_dest != NULL && top_dest->klass() != NULL, "args are strange" ); |
5945 | |
5946 | // checks are the responsibility of the caller |
5947 | Node* src_start = src; |
5948 | Node* dest_start = dest; |
5949 | if (src_offset != NULL || dest_offset != NULL) { |
5950 | assert(src_offset != NULL && dest_offset != NULL, "" ); |
5951 | src_start = array_element_address(src, src_offset, T_BYTE); |
5952 | dest_start = array_element_address(dest, dest_offset, T_BYTE); |
5953 | } |
5954 | |
5955 | // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object |
5956 | // (because of the predicated logic executed earlier). |
5957 | // so we cast it here safely. |
5958 | // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java |
5959 | |
5960 | Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher" , "Lcom/sun/crypto/provider/SymmetricCipher;" , /*is_exact*/ false); |
5961 | if (embeddedCipherObj == NULL) return false; |
5962 | |
5963 | // cast it to what we know it will be at runtime |
5964 | const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); |
5965 | assert(tinst != NULL, "CBC obj is null" ); |
5966 | assert(tinst->klass()->is_loaded(), "CBC obj is not loaded" ); |
5967 | ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt" )); |
5968 | assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded" ); |
5969 | |
5970 | ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); |
5971 | const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); |
5972 | const TypeOopPtr* xtype = aklass->as_instance_type(); |
5973 | Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); |
5974 | aescrypt_object = _gvn.transform(aescrypt_object); |
5975 | |
5976 | // we need to get the start of the aescrypt_object's expanded key array |
5977 | Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); |
5978 | if (k_start == NULL) return false; |
5979 | |
5980 | // similarly, get the start address of the r vector |
5981 | Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r" , "[B" , /*is_exact*/ false); |
5982 | if (objRvec == NULL) return false; |
5983 | objRvec = access_resolve(objRvec, ACCESS_WRITE); |
5984 | Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); |
5985 | |
5986 | Node* cbcCrypt; |
5987 | if (Matcher::pass_original_key_for_aes()) { |
5988 | // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to |
5989 | // compatibility issues between Java key expansion and SPARC crypto instructions |
5990 | Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); |
5991 | if (original_k_start == NULL) return false; |
5992 | |
5993 | // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start |
5994 | cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, |
5995 | OptoRuntime::cipherBlockChaining_aescrypt_Type(), |
5996 | stubAddr, stubName, TypePtr::BOTTOM, |
5997 | src_start, dest_start, k_start, r_start, len, original_k_start); |
5998 | } else { |
5999 | // Call the stub, passing src_start, dest_start, k_start, r_start and src_len |
6000 | cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, |
6001 | OptoRuntime::cipherBlockChaining_aescrypt_Type(), |
6002 | stubAddr, stubName, TypePtr::BOTTOM, |
6003 | src_start, dest_start, k_start, r_start, len); |
6004 | } |
6005 | |
6006 | // return cipher length (int) |
6007 | Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); |
6008 | set_result(retvalue); |
6009 | return true; |
6010 | } |
6011 | |
6012 | //------------------------------inline_counterMode_AESCrypt----------------------- |
6013 | bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { |
6014 | assert(UseAES, "need AES instruction support" ); |
6015 | if (!UseAESCTRIntrinsics) return false; |
6016 | |
6017 | address stubAddr = NULL; |
6018 | const char *stubName = NULL; |
6019 | if (id == vmIntrinsics::_counterMode_AESCrypt) { |
6020 | stubAddr = StubRoutines::counterMode_AESCrypt(); |
6021 | stubName = "counterMode_AESCrypt" ; |
6022 | } |
6023 | if (stubAddr == NULL) return false; |
6024 | |
6025 | Node* counterMode_object = argument(0); |
6026 | Node* src = argument(1); |
6027 | Node* src_offset = argument(2); |
6028 | Node* len = argument(3); |
6029 | Node* dest = argument(4); |
6030 | Node* dest_offset = argument(5); |
6031 | |
6032 | src = access_resolve(src, ACCESS_READ); |
6033 | dest = access_resolve(dest, ACCESS_WRITE); |
6034 | counterMode_object = access_resolve(counterMode_object, ACCESS_WRITE); |
6035 | |
6036 | // (1) src and dest are arrays. |
6037 | const Type* src_type = src->Value(&_gvn); |
6038 | const Type* dest_type = dest->Value(&_gvn); |
6039 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
6040 | const TypeAryPtr* top_dest = dest_type->isa_aryptr(); |
6041 | assert(top_src != NULL && top_src->klass() != NULL && |
6042 | top_dest != NULL && top_dest->klass() != NULL, "args are strange" ); |
6043 | |
6044 | // checks are the responsibility of the caller |
6045 | Node* src_start = src; |
6046 | Node* dest_start = dest; |
6047 | if (src_offset != NULL || dest_offset != NULL) { |
6048 | assert(src_offset != NULL && dest_offset != NULL, "" ); |
6049 | src_start = array_element_address(src, src_offset, T_BYTE); |
6050 | dest_start = array_element_address(dest, dest_offset, T_BYTE); |
6051 | } |
6052 | |
6053 | // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object |
6054 | // (because of the predicated logic executed earlier). |
6055 | // so we cast it here safely. |
6056 | // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java |
6057 | Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher" , "Lcom/sun/crypto/provider/SymmetricCipher;" , /*is_exact*/ false); |
6058 | if (embeddedCipherObj == NULL) return false; |
6059 | // cast it to what we know it will be at runtime |
6060 | const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); |
6061 | assert(tinst != NULL, "CTR obj is null" ); |
6062 | assert(tinst->klass()->is_loaded(), "CTR obj is not loaded" ); |
6063 | ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt" )); |
6064 | assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded" ); |
6065 | ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); |
6066 | const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); |
6067 | const TypeOopPtr* xtype = aklass->as_instance_type(); |
6068 | Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); |
6069 | aescrypt_object = _gvn.transform(aescrypt_object); |
6070 | // we need to get the start of the aescrypt_object's expanded key array |
6071 | Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); |
6072 | if (k_start == NULL) return false; |
6073 | // similarly, get the start address of the r vector |
6074 | Node* obj_counter = load_field_from_object(counterMode_object, "counter" , "[B" , /*is_exact*/ false); |
6075 | if (obj_counter == NULL) return false; |
6076 | obj_counter = access_resolve(obj_counter, ACCESS_WRITE); |
6077 | Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); |
6078 | |
6079 | Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter" , "[B" , /*is_exact*/ false); |
6080 | if (saved_encCounter == NULL) return false; |
6081 | saved_encCounter = access_resolve(saved_encCounter, ACCESS_WRITE); |
6082 | Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); |
6083 | Node* used = field_address_from_object(counterMode_object, "used" , "I" , /*is_exact*/ false); |
6084 | |
6085 | Node* ctrCrypt; |
6086 | if (Matcher::pass_original_key_for_aes()) { |
6087 | // no SPARC version for AES/CTR intrinsics now. |
6088 | return false; |
6089 | } |
6090 | // Call the stub, passing src_start, dest_start, k_start, r_start and src_len |
6091 | ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, |
6092 | OptoRuntime::counterMode_aescrypt_Type(), |
6093 | stubAddr, stubName, TypePtr::BOTTOM, |
6094 | src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); |
6095 | |
6096 | // return cipher length (int) |
6097 | Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); |
6098 | set_result(retvalue); |
6099 | return true; |
6100 | } |
6101 | |
6102 | //------------------------------get_key_start_from_aescrypt_object----------------------- |
6103 | Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { |
6104 | #if defined(PPC64) || defined(S390) |
6105 | // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. |
6106 | // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. |
6107 | // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. |
6108 | // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). |
6109 | Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK" , "[[I" , /*is_exact*/ false); |
6110 | assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt" ); |
6111 | if (objSessionK == NULL) { |
6112 | return (Node *) NULL; |
6113 | } |
6114 | Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); |
6115 | #else |
6116 | Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K" , "[I" , /*is_exact*/ false); |
6117 | #endif // PPC64 |
6118 | assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt" ); |
6119 | if (objAESCryptKey == NULL) return (Node *) NULL; |
6120 | |
6121 | // now have the array, need to get the start address of the K array |
6122 | objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); |
6123 | Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); |
6124 | return k_start; |
6125 | } |
6126 | |
6127 | //------------------------------get_original_key_start_from_aescrypt_object----------------------- |
6128 | Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { |
6129 | Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey" , "[B" , /*is_exact*/ false); |
6130 | assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt" ); |
6131 | if (objAESCryptKey == NULL) return (Node *) NULL; |
6132 | |
6133 | // now have the array, need to get the start address of the lastKey array |
6134 | objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); |
6135 | Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); |
6136 | return original_k_start; |
6137 | } |
6138 | |
6139 | //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- |
6140 | // Return node representing slow path of predicate check. |
6141 | // the pseudo code we want to emulate with this predicate is: |
6142 | // for encryption: |
6143 | // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath |
6144 | // for decryption: |
6145 | // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath |
6146 | // note cipher==plain is more conservative than the original java code but that's OK |
6147 | // |
6148 | Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { |
6149 | // The receiver was checked for NULL already. |
6150 | Node* objCBC = argument(0); |
6151 | |
6152 | Node* src = argument(1); |
6153 | Node* dest = argument(4); |
6154 | |
6155 | // Load embeddedCipher field of CipherBlockChaining object. |
6156 | Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher" , "Lcom/sun/crypto/provider/SymmetricCipher;" , /*is_exact*/ false); |
6157 | |
6158 | // get AESCrypt klass for instanceOf check |
6159 | // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point |
6160 | // will have same classloader as CipherBlockChaining object |
6161 | const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); |
6162 | assert(tinst != NULL, "CBCobj is null" ); |
6163 | assert(tinst->klass()->is_loaded(), "CBCobj is not loaded" ); |
6164 | |
6165 | // we want to do an instanceof comparison against the AESCrypt class |
6166 | ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt" )); |
6167 | if (!klass_AESCrypt->is_loaded()) { |
6168 | // if AESCrypt is not even loaded, we never take the intrinsic fast path |
6169 | Node* ctrl = control(); |
6170 | set_control(top()); // no regular fast path |
6171 | return ctrl; |
6172 | } |
6173 | |
6174 | src = must_be_not_null(src, true); |
6175 | dest = must_be_not_null(dest, true); |
6176 | |
6177 | // Resolve oops to stable for CmpP below. |
6178 | src = access_resolve(src, 0); |
6179 | dest = access_resolve(dest, 0); |
6180 | |
6181 | ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); |
6182 | |
6183 | Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); |
6184 | Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); |
6185 | Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); |
6186 | |
6187 | Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); |
6188 | |
6189 | // for encryption, we are done |
6190 | if (!decrypting) |
6191 | return instof_false; // even if it is NULL |
6192 | |
6193 | // for decryption, we need to add a further check to avoid |
6194 | // taking the intrinsic path when cipher and plain are the same |
6195 | // see the original java code for why. |
6196 | RegionNode* region = new RegionNode(3); |
6197 | region->init_req(1, instof_false); |
6198 | |
6199 | Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); |
6200 | Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); |
6201 | Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); |
6202 | region->init_req(2, src_dest_conjoint); |
6203 | |
6204 | record_for_igvn(region); |
6205 | return _gvn.transform(region); |
6206 | } |
6207 | |
6208 | //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- |
6209 | // Return node representing slow path of predicate check. |
6210 | // the pseudo code we want to emulate with this predicate is: |
6211 | // for encryption: |
6212 | // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath |
6213 | // for decryption: |
6214 | // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath |
6215 | // note cipher==plain is more conservative than the original java code but that's OK |
6216 | // |
6217 | |
6218 | Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { |
6219 | // The receiver was checked for NULL already. |
6220 | Node* objCTR = argument(0); |
6221 | |
6222 | // Load embeddedCipher field of CipherBlockChaining object. |
6223 | Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher" , "Lcom/sun/crypto/provider/SymmetricCipher;" , /*is_exact*/ false); |
6224 | |
6225 | // get AESCrypt klass for instanceOf check |
6226 | // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point |
6227 | // will have same classloader as CipherBlockChaining object |
6228 | const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); |
6229 | assert(tinst != NULL, "CTRobj is null" ); |
6230 | assert(tinst->klass()->is_loaded(), "CTRobj is not loaded" ); |
6231 | |
6232 | // we want to do an instanceof comparison against the AESCrypt class |
6233 | ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt" )); |
6234 | if (!klass_AESCrypt->is_loaded()) { |
6235 | // if AESCrypt is not even loaded, we never take the intrinsic fast path |
6236 | Node* ctrl = control(); |
6237 | set_control(top()); // no regular fast path |
6238 | return ctrl; |
6239 | } |
6240 | |
6241 | ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); |
6242 | Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); |
6243 | Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); |
6244 | Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); |
6245 | Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); |
6246 | |
6247 | return instof_false; // even if it is NULL |
6248 | } |
6249 | |
6250 | //------------------------------inline_ghash_processBlocks |
6251 | bool LibraryCallKit::inline_ghash_processBlocks() { |
6252 | address stubAddr; |
6253 | const char *stubName; |
6254 | assert(UseGHASHIntrinsics, "need GHASH intrinsics support" ); |
6255 | |
6256 | stubAddr = StubRoutines::ghash_processBlocks(); |
6257 | stubName = "ghash_processBlocks" ; |
6258 | |
6259 | Node* data = argument(0); |
6260 | Node* offset = argument(1); |
6261 | Node* len = argument(2); |
6262 | Node* state = argument(3); |
6263 | Node* subkeyH = argument(4); |
6264 | |
6265 | state = must_be_not_null(state, true); |
6266 | subkeyH = must_be_not_null(subkeyH, true); |
6267 | data = must_be_not_null(data, true); |
6268 | |
6269 | state = access_resolve(state, ACCESS_WRITE); |
6270 | subkeyH = access_resolve(subkeyH, ACCESS_READ); |
6271 | data = access_resolve(data, ACCESS_READ); |
6272 | |
6273 | Node* state_start = array_element_address(state, intcon(0), T_LONG); |
6274 | assert(state_start, "state is NULL" ); |
6275 | Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); |
6276 | assert(subkeyH_start, "subkeyH is NULL" ); |
6277 | Node* data_start = array_element_address(data, offset, T_BYTE); |
6278 | assert(data_start, "data is NULL" ); |
6279 | |
6280 | Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, |
6281 | OptoRuntime::ghash_processBlocks_Type(), |
6282 | stubAddr, stubName, TypePtr::BOTTOM, |
6283 | state_start, subkeyH_start, data_start, len); |
6284 | return true; |
6285 | } |
6286 | |
6287 | bool LibraryCallKit::inline_base64_encodeBlock() { |
6288 | address stubAddr; |
6289 | const char *stubName; |
6290 | assert(UseBASE64Intrinsics, "need Base64 intrinsics support" ); |
6291 | assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters" ); |
6292 | stubAddr = StubRoutines::base64_encodeBlock(); |
6293 | stubName = "encodeBlock" ; |
6294 | |
6295 | if (!stubAddr) return false; |
6296 | Node* base64obj = argument(0); |
6297 | Node* src = argument(1); |
6298 | Node* offset = argument(2); |
6299 | Node* len = argument(3); |
6300 | Node* dest = argument(4); |
6301 | Node* dp = argument(5); |
6302 | Node* isURL = argument(6); |
6303 | |
6304 | src = must_be_not_null(src, true); |
6305 | src = access_resolve(src, ACCESS_READ); |
6306 | dest = must_be_not_null(dest, true); |
6307 | dest = access_resolve(dest, ACCESS_WRITE); |
6308 | |
6309 | Node* src_start = array_element_address(src, intcon(0), T_BYTE); |
6310 | assert(src_start, "source array is NULL" ); |
6311 | Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); |
6312 | assert(dest_start, "destination array is NULL" ); |
6313 | |
6314 | Node* base64 = make_runtime_call(RC_LEAF, |
6315 | OptoRuntime::base64_encodeBlock_Type(), |
6316 | stubAddr, stubName, TypePtr::BOTTOM, |
6317 | src_start, offset, len, dest_start, dp, isURL); |
6318 | return true; |
6319 | } |
6320 | |
6321 | //------------------------------inline_sha_implCompress----------------------- |
6322 | // |
6323 | // Calculate SHA (i.e., SHA-1) for single-block byte[] array. |
6324 | // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) |
6325 | // |
6326 | // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. |
6327 | // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) |
6328 | // |
6329 | // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. |
6330 | // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) |
6331 | // |
6332 | bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { |
6333 | assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters" ); |
6334 | |
6335 | Node* sha_obj = argument(0); |
6336 | Node* src = argument(1); // type oop |
6337 | Node* ofs = argument(2); // type int |
6338 | |
6339 | const Type* src_type = src->Value(&_gvn); |
6340 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
6341 | if (top_src == NULL || top_src->klass() == NULL) { |
6342 | // failed array check |
6343 | return false; |
6344 | } |
6345 | // Figure out the size and type of the elements we will be copying. |
6346 | BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
6347 | if (src_elem != T_BYTE) { |
6348 | return false; |
6349 | } |
6350 | // 'src_start' points to src array + offset |
6351 | src = must_be_not_null(src, true); |
6352 | src = access_resolve(src, ACCESS_READ); |
6353 | Node* src_start = array_element_address(src, ofs, src_elem); |
6354 | Node* state = NULL; |
6355 | address stubAddr; |
6356 | const char *stubName; |
6357 | |
6358 | switch(id) { |
6359 | case vmIntrinsics::_sha_implCompress: |
6360 | assert(UseSHA1Intrinsics, "need SHA1 instruction support" ); |
6361 | state = get_state_from_sha_object(sha_obj); |
6362 | stubAddr = StubRoutines::sha1_implCompress(); |
6363 | stubName = "sha1_implCompress" ; |
6364 | break; |
6365 | case vmIntrinsics::_sha2_implCompress: |
6366 | assert(UseSHA256Intrinsics, "need SHA256 instruction support" ); |
6367 | state = get_state_from_sha_object(sha_obj); |
6368 | stubAddr = StubRoutines::sha256_implCompress(); |
6369 | stubName = "sha256_implCompress" ; |
6370 | break; |
6371 | case vmIntrinsics::_sha5_implCompress: |
6372 | assert(UseSHA512Intrinsics, "need SHA512 instruction support" ); |
6373 | state = get_state_from_sha5_object(sha_obj); |
6374 | stubAddr = StubRoutines::sha512_implCompress(); |
6375 | stubName = "sha512_implCompress" ; |
6376 | break; |
6377 | default: |
6378 | fatal_unexpected_iid(id); |
6379 | return false; |
6380 | } |
6381 | if (state == NULL) return false; |
6382 | |
6383 | assert(stubAddr != NULL, "Stub is generated" ); |
6384 | if (stubAddr == NULL) return false; |
6385 | |
6386 | // Call the stub. |
6387 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), |
6388 | stubAddr, stubName, TypePtr::BOTTOM, |
6389 | src_start, state); |
6390 | |
6391 | return true; |
6392 | } |
6393 | |
6394 | //------------------------------inline_digestBase_implCompressMB----------------------- |
6395 | // |
6396 | // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. |
6397 | // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) |
6398 | // |
6399 | bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { |
6400 | assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, |
6401 | "need SHA1/SHA256/SHA512 instruction support" ); |
6402 | assert((uint)predicate < 3, "sanity" ); |
6403 | assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters" ); |
6404 | |
6405 | Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. |
6406 | Node* src = argument(1); // byte[] array |
6407 | Node* ofs = argument(2); // type int |
6408 | Node* limit = argument(3); // type int |
6409 | |
6410 | const Type* src_type = src->Value(&_gvn); |
6411 | const TypeAryPtr* top_src = src_type->isa_aryptr(); |
6412 | if (top_src == NULL || top_src->klass() == NULL) { |
6413 | // failed array check |
6414 | return false; |
6415 | } |
6416 | // Figure out the size and type of the elements we will be copying. |
6417 | BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); |
6418 | if (src_elem != T_BYTE) { |
6419 | return false; |
6420 | } |
6421 | // 'src_start' points to src array + offset |
6422 | src = must_be_not_null(src, false); |
6423 | src = access_resolve(src, ACCESS_READ); |
6424 | Node* src_start = array_element_address(src, ofs, src_elem); |
6425 | |
6426 | const char* klass_SHA_name = NULL; |
6427 | const char* stub_name = NULL; |
6428 | address stub_addr = NULL; |
6429 | bool long_state = false; |
6430 | |
6431 | switch (predicate) { |
6432 | case 0: |
6433 | if (UseSHA1Intrinsics) { |
6434 | klass_SHA_name = "sun/security/provider/SHA" ; |
6435 | stub_name = "sha1_implCompressMB" ; |
6436 | stub_addr = StubRoutines::sha1_implCompressMB(); |
6437 | } |
6438 | break; |
6439 | case 1: |
6440 | if (UseSHA256Intrinsics) { |
6441 | klass_SHA_name = "sun/security/provider/SHA2" ; |
6442 | stub_name = "sha256_implCompressMB" ; |
6443 | stub_addr = StubRoutines::sha256_implCompressMB(); |
6444 | } |
6445 | break; |
6446 | case 2: |
6447 | if (UseSHA512Intrinsics) { |
6448 | klass_SHA_name = "sun/security/provider/SHA5" ; |
6449 | stub_name = "sha512_implCompressMB" ; |
6450 | stub_addr = StubRoutines::sha512_implCompressMB(); |
6451 | long_state = true; |
6452 | } |
6453 | break; |
6454 | default: |
6455 | fatal("unknown SHA intrinsic predicate: %d" , predicate); |
6456 | } |
6457 | if (klass_SHA_name != NULL) { |
6458 | assert(stub_addr != NULL, "Stub is generated" ); |
6459 | if (stub_addr == NULL) return false; |
6460 | |
6461 | // get DigestBase klass to lookup for SHA klass |
6462 | const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); |
6463 | assert(tinst != NULL, "digestBase_obj is not instance???" ); |
6464 | assert(tinst->klass()->is_loaded(), "DigestBase is not loaded" ); |
6465 | |
6466 | ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); |
6467 | assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded" ); |
6468 | ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); |
6469 | return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); |
6470 | } |
6471 | return false; |
6472 | } |
6473 | //------------------------------inline_sha_implCompressMB----------------------- |
6474 | bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, |
6475 | bool long_state, address stubAddr, const char *stubName, |
6476 | Node* src_start, Node* ofs, Node* limit) { |
6477 | const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); |
6478 | const TypeOopPtr* xtype = aklass->as_instance_type(); |
6479 | Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); |
6480 | sha_obj = _gvn.transform(sha_obj); |
6481 | |
6482 | Node* state; |
6483 | if (long_state) { |
6484 | state = get_state_from_sha5_object(sha_obj); |
6485 | } else { |
6486 | state = get_state_from_sha_object(sha_obj); |
6487 | } |
6488 | if (state == NULL) return false; |
6489 | |
6490 | // Call the stub. |
6491 | Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, |
6492 | OptoRuntime::digestBase_implCompressMB_Type(), |
6493 | stubAddr, stubName, TypePtr::BOTTOM, |
6494 | src_start, state, ofs, limit); |
6495 | // return ofs (int) |
6496 | Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); |
6497 | set_result(result); |
6498 | |
6499 | return true; |
6500 | } |
6501 | |
6502 | //------------------------------get_state_from_sha_object----------------------- |
6503 | Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { |
6504 | Node* sha_state = load_field_from_object(sha_object, "state" , "[I" , /*is_exact*/ false); |
6505 | assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2" ); |
6506 | if (sha_state == NULL) return (Node *) NULL; |
6507 | |
6508 | // now have the array, need to get the start address of the state array |
6509 | sha_state = access_resolve(sha_state, ACCESS_WRITE); |
6510 | Node* state = array_element_address(sha_state, intcon(0), T_INT); |
6511 | return state; |
6512 | } |
6513 | |
6514 | //------------------------------get_state_from_sha5_object----------------------- |
6515 | Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { |
6516 | Node* sha_state = load_field_from_object(sha_object, "state" , "[J" , /*is_exact*/ false); |
6517 | assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5" ); |
6518 | if (sha_state == NULL) return (Node *) NULL; |
6519 | |
6520 | // now have the array, need to get the start address of the state array |
6521 | sha_state = access_resolve(sha_state, ACCESS_WRITE); |
6522 | Node* state = array_element_address(sha_state, intcon(0), T_LONG); |
6523 | return state; |
6524 | } |
6525 | |
6526 | //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- |
6527 | // Return node representing slow path of predicate check. |
6528 | // the pseudo code we want to emulate with this predicate is: |
6529 | // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath |
6530 | // |
6531 | Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { |
6532 | assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, |
6533 | "need SHA1/SHA256/SHA512 instruction support" ); |
6534 | assert((uint)predicate < 3, "sanity" ); |
6535 | |
6536 | // The receiver was checked for NULL already. |
6537 | Node* digestBaseObj = argument(0); |
6538 | |
6539 | // get DigestBase klass for instanceOf check |
6540 | const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); |
6541 | assert(tinst != NULL, "digestBaseObj is null" ); |
6542 | assert(tinst->klass()->is_loaded(), "DigestBase is not loaded" ); |
6543 | |
6544 | const char* klass_SHA_name = NULL; |
6545 | switch (predicate) { |
6546 | case 0: |
6547 | if (UseSHA1Intrinsics) { |
6548 | // we want to do an instanceof comparison against the SHA class |
6549 | klass_SHA_name = "sun/security/provider/SHA" ; |
6550 | } |
6551 | break; |
6552 | case 1: |
6553 | if (UseSHA256Intrinsics) { |
6554 | // we want to do an instanceof comparison against the SHA2 class |
6555 | klass_SHA_name = "sun/security/provider/SHA2" ; |
6556 | } |
6557 | break; |
6558 | case 2: |
6559 | if (UseSHA512Intrinsics) { |
6560 | // we want to do an instanceof comparison against the SHA5 class |
6561 | klass_SHA_name = "sun/security/provider/SHA5" ; |
6562 | } |
6563 | break; |
6564 | default: |
6565 | fatal("unknown SHA intrinsic predicate: %d" , predicate); |
6566 | } |
6567 | |
6568 | ciKlass* klass_SHA = NULL; |
6569 | if (klass_SHA_name != NULL) { |
6570 | klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); |
6571 | } |
6572 | if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { |
6573 | // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path |
6574 | Node* ctrl = control(); |
6575 | set_control(top()); // no intrinsic path |
6576 | return ctrl; |
6577 | } |
6578 | ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); |
6579 | |
6580 | Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); |
6581 | Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); |
6582 | Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); |
6583 | Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); |
6584 | |
6585 | return instof_false; // even if it is NULL |
6586 | } |
6587 | |
6588 | //-------------inline_fma----------------------------------- |
6589 | bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { |
6590 | Node *a = NULL; |
6591 | Node *b = NULL; |
6592 | Node *c = NULL; |
6593 | Node* result = NULL; |
6594 | switch (id) { |
6595 | case vmIntrinsics::_fmaD: |
6596 | assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each." ); |
6597 | // no receiver since it is static method |
6598 | a = round_double_node(argument(0)); |
6599 | b = round_double_node(argument(2)); |
6600 | c = round_double_node(argument(4)); |
6601 | result = _gvn.transform(new FmaDNode(control(), a, b, c)); |
6602 | break; |
6603 | case vmIntrinsics::_fmaF: |
6604 | assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each." ); |
6605 | a = argument(0); |
6606 | b = argument(1); |
6607 | c = argument(2); |
6608 | result = _gvn.transform(new FmaFNode(control(), a, b, c)); |
6609 | break; |
6610 | default: |
6611 | fatal_unexpected_iid(id); break; |
6612 | } |
6613 | set_result(result); |
6614 | return true; |
6615 | } |
6616 | |
6617 | bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) { |
6618 | // argument(0) is receiver |
6619 | Node* codePoint = argument(1); |
6620 | Node* n = NULL; |
6621 | |
6622 | switch (id) { |
6623 | case vmIntrinsics::_isDigit : |
6624 | n = new DigitNode(control(), codePoint); |
6625 | break; |
6626 | case vmIntrinsics::_isLowerCase : |
6627 | n = new LowerCaseNode(control(), codePoint); |
6628 | break; |
6629 | case vmIntrinsics::_isUpperCase : |
6630 | n = new UpperCaseNode(control(), codePoint); |
6631 | break; |
6632 | case vmIntrinsics::_isWhitespace : |
6633 | n = new WhitespaceNode(control(), codePoint); |
6634 | break; |
6635 | default: |
6636 | fatal_unexpected_iid(id); |
6637 | } |
6638 | |
6639 | set_result(_gvn.transform(n)); |
6640 | return true; |
6641 | } |
6642 | |
6643 | //------------------------------inline_fp_min_max------------------------------ |
6644 | bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) { |
6645 | /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416. |
6646 | |
6647 | // The intrinsic should be used only when the API branches aren't predictable, |
6648 | // the last one performing the most important comparison. The following heuristic |
6649 | // uses the branch statistics to eventually bail out if necessary. |
6650 | |
6651 | ciMethodData *md = callee()->method_data(); |
6652 | |
6653 | if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) { |
6654 | ciCallProfile cp = caller()->call_profile_at_bci(bci()); |
6655 | |
6656 | if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) { |
6657 | // Bail out if the call-site didn't contribute enough to the statistics. |
6658 | return false; |
6659 | } |
6660 | |
6661 | uint taken = 0, not_taken = 0; |
6662 | |
6663 | for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) { |
6664 | if (p->is_BranchData()) { |
6665 | taken = ((ciBranchData*)p)->taken(); |
6666 | not_taken = ((ciBranchData*)p)->not_taken(); |
6667 | } |
6668 | } |
6669 | |
6670 | double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count()); |
6671 | balance = balance < 0 ? -balance : balance; |
6672 | if ( balance > 0.2 ) { |
6673 | // Bail out if the most important branch is predictable enough. |
6674 | return false; |
6675 | } |
6676 | } |
6677 | */ |
6678 | |
6679 | Node *a = NULL; |
6680 | Node *b = NULL; |
6681 | Node *n = NULL; |
6682 | switch (id) { |
6683 | case vmIntrinsics::_maxF: |
6684 | case vmIntrinsics::_minF: |
6685 | assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each." ); |
6686 | a = argument(0); |
6687 | b = argument(1); |
6688 | break; |
6689 | case vmIntrinsics::_maxD: |
6690 | case vmIntrinsics::_minD: |
6691 | assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each." ); |
6692 | a = round_double_node(argument(0)); |
6693 | b = round_double_node(argument(2)); |
6694 | break; |
6695 | default: |
6696 | fatal_unexpected_iid(id); |
6697 | break; |
6698 | } |
6699 | switch (id) { |
6700 | case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break; |
6701 | case vmIntrinsics::_minF: n = new MinFNode(a, b); break; |
6702 | case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break; |
6703 | case vmIntrinsics::_minD: n = new MinDNode(a, b); break; |
6704 | default: fatal_unexpected_iid(id); break; |
6705 | } |
6706 | set_result(_gvn.transform(n)); |
6707 | return true; |
6708 | } |
6709 | |
6710 | bool LibraryCallKit::inline_profileBoolean() { |
6711 | Node* counts = argument(1); |
6712 | const TypeAryPtr* ary = NULL; |
6713 | ciArray* aobj = NULL; |
6714 | if (counts->is_Con() |
6715 | && (ary = counts->bottom_type()->isa_aryptr()) != NULL |
6716 | && (aobj = ary->const_oop()->as_array()) != NULL |
6717 | && (aobj->length() == 2)) { |
6718 | // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. |
6719 | jint false_cnt = aobj->element_value(0).as_int(); |
6720 | jint true_cnt = aobj->element_value(1).as_int(); |
6721 | |
6722 | if (C->log() != NULL) { |
6723 | C->log()->elem("observe source='profileBoolean' false='%d' true='%d'" , |
6724 | false_cnt, true_cnt); |
6725 | } |
6726 | |
6727 | if (false_cnt + true_cnt == 0) { |
6728 | // According to profile, never executed. |
6729 | uncommon_trap_exact(Deoptimization::Reason_intrinsic, |
6730 | Deoptimization::Action_reinterpret); |
6731 | return true; |
6732 | } |
6733 | |
6734 | // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) |
6735 | // is a number of each value occurrences. |
6736 | Node* result = argument(0); |
6737 | if (false_cnt == 0 || true_cnt == 0) { |
6738 | // According to profile, one value has been never seen. |
6739 | int expected_val = (false_cnt == 0) ? 1 : 0; |
6740 | |
6741 | Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); |
6742 | Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); |
6743 | |
6744 | IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); |
6745 | Node* fast_path = _gvn.transform(new IfTrueNode(check)); |
6746 | Node* slow_path = _gvn.transform(new IfFalseNode(check)); |
6747 | |
6748 | { // Slow path: uncommon trap for never seen value and then reexecute |
6749 | // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows |
6750 | // the value has been seen at least once. |
6751 | PreserveJVMState pjvms(this); |
6752 | PreserveReexecuteState preexecs(this); |
6753 | jvms()->set_should_reexecute(true); |
6754 | |
6755 | set_control(slow_path); |
6756 | set_i_o(i_o()); |
6757 | |
6758 | uncommon_trap_exact(Deoptimization::Reason_intrinsic, |
6759 | Deoptimization::Action_reinterpret); |
6760 | } |
6761 | // The guard for never seen value enables sharpening of the result and |
6762 | // returning a constant. It allows to eliminate branches on the same value |
6763 | // later on. |
6764 | set_control(fast_path); |
6765 | result = intcon(expected_val); |
6766 | } |
6767 | // Stop profiling. |
6768 | // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. |
6769 | // By replacing method body with profile data (represented as ProfileBooleanNode |
6770 | // on IR level) we effectively disable profiling. |
6771 | // It enables full speed execution once optimized code is generated. |
6772 | Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); |
6773 | C->record_for_igvn(profile); |
6774 | set_result(profile); |
6775 | return true; |
6776 | } else { |
6777 | // Continue profiling. |
6778 | // Profile data isn't available at the moment. So, execute method's bytecode version. |
6779 | // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod |
6780 | // is compiled and counters aren't available since corresponding MethodHandle |
6781 | // isn't a compile-time constant. |
6782 | return false; |
6783 | } |
6784 | } |
6785 | |
6786 | bool LibraryCallKit::inline_isCompileConstant() { |
6787 | Node* n = argument(0); |
6788 | set_result(n->is_Con() ? intcon(1) : intcon(0)); |
6789 | return true; |
6790 | } |
6791 | |