1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/opto/library_call.cpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,6124 @@
1.4 +/*
1.5 + * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include "precompiled.hpp"
1.29 +#include "classfile/systemDictionary.hpp"
1.30 +#include "classfile/vmSymbols.hpp"
1.31 +#include "compiler/compileBroker.hpp"
1.32 +#include "compiler/compileLog.hpp"
1.33 +#include "oops/objArrayKlass.hpp"
1.34 +#include "opto/addnode.hpp"
1.35 +#include "opto/callGenerator.hpp"
1.36 +#include "opto/cfgnode.hpp"
1.37 +#include "opto/idealKit.hpp"
1.38 +#include "opto/mathexactnode.hpp"
1.39 +#include "opto/mulnode.hpp"
1.40 +#include "opto/parse.hpp"
1.41 +#include "opto/runtime.hpp"
1.42 +#include "opto/subnode.hpp"
1.43 +#include "prims/nativeLookup.hpp"
1.44 +#include "runtime/sharedRuntime.hpp"
1.45 +#include "trace/traceMacros.hpp"
1.46 +
1.47 +class LibraryIntrinsic : public InlineCallGenerator {
1.48 + // Extend the set of intrinsics known to the runtime:
1.49 + public:
1.50 + private:
1.51 + bool _is_virtual;
1.52 + bool _is_predicted;
1.53 + bool _does_virtual_dispatch;
1.54 + vmIntrinsics::ID _intrinsic_id;
1.55 +
1.56 + public:
1.57 + LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
1.58 + : InlineCallGenerator(m),
1.59 + _is_virtual(is_virtual),
1.60 + _is_predicted(is_predicted),
1.61 + _does_virtual_dispatch(does_virtual_dispatch),
1.62 + _intrinsic_id(id)
1.63 + {
1.64 + }
1.65 + virtual bool is_intrinsic() const { return true; }
1.66 + virtual bool is_virtual() const { return _is_virtual; }
1.67 + virtual bool is_predicted() const { return _is_predicted; }
1.68 + virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
1.69 + virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
1.70 + virtual Node* generate_predicate(JVMState* jvms);
1.71 + vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
1.72 +};
1.73 +
1.74 +
1.75 +// Local helper class for LibraryIntrinsic:
1.76 +class LibraryCallKit : public GraphKit {
1.77 + private:
1.78 + LibraryIntrinsic* _intrinsic; // the library intrinsic being called
1.79 + Node* _result; // the result node, if any
1.80 + int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
1.81 +
1.82 + const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
1.83 +
1.84 + public:
1.85 + LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
1.86 + : GraphKit(jvms),
1.87 + _intrinsic(intrinsic),
1.88 + _result(NULL)
1.89 + {
1.90 + // Check if this is a root compile. In that case we don't have a caller.
1.91 + if (!jvms->has_method()) {
1.92 + _reexecute_sp = sp();
1.93 + } else {
1.94 + // Find out how many arguments the interpreter needs when deoptimizing
1.95 + // and save the stack pointer value so it can used by uncommon_trap.
1.96 + // We find the argument count by looking at the declared signature.
1.97 + bool ignored_will_link;
1.98 + ciSignature* declared_signature = NULL;
1.99 + ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1.100 + const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
1.101 + _reexecute_sp = sp() + nargs; // "push" arguments back on stack
1.102 + }
1.103 + }
1.104 +
1.105 + virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
1.106 +
1.107 + ciMethod* caller() const { return jvms()->method(); }
1.108 + int bci() const { return jvms()->bci(); }
1.109 + LibraryIntrinsic* intrinsic() const { return _intrinsic; }
1.110 + vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
1.111 + ciMethod* callee() const { return _intrinsic->method(); }
1.112 +
1.113 + bool try_to_inline();
1.114 + Node* try_to_predicate();
1.115 +
1.116 + void push_result() {
1.117 + // Push the result onto the stack.
1.118 + if (!stopped() && result() != NULL) {
1.119 + BasicType bt = result()->bottom_type()->basic_type();
1.120 + push_node(bt, result());
1.121 + }
1.122 + }
1.123 +
1.124 + private:
1.125 + void fatal_unexpected_iid(vmIntrinsics::ID iid) {
1.126 + fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
1.127 + }
1.128 +
1.129 + void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
1.130 + void set_result(RegionNode* region, PhiNode* value);
1.131 + Node* result() { return _result; }
1.132 +
1.133 + virtual int reexecute_sp() { return _reexecute_sp; }
1.134 +
1.135 + // Helper functions to inline natives
1.136 + Node* generate_guard(Node* test, RegionNode* region, float true_prob);
1.137 + Node* generate_slow_guard(Node* test, RegionNode* region);
1.138 + Node* generate_fair_guard(Node* test, RegionNode* region);
1.139 + Node* generate_negative_guard(Node* index, RegionNode* region,
1.140 + // resulting CastII of index:
1.141 + Node* *pos_index = NULL);
1.142 + Node* generate_nonpositive_guard(Node* index, bool never_negative,
1.143 + // resulting CastII of index:
1.144 + Node* *pos_index = NULL);
1.145 + Node* generate_limit_guard(Node* offset, Node* subseq_length,
1.146 + Node* array_length,
1.147 + RegionNode* region);
1.148 + Node* generate_current_thread(Node* &tls_output);
1.149 + address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
1.150 + bool disjoint_bases, const char* &name, bool dest_uninitialized);
1.151 + Node* load_mirror_from_klass(Node* klass);
1.152 + Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
1.153 + RegionNode* region, int null_path,
1.154 + int offset);
1.155 + Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
1.156 + RegionNode* region, int null_path) {
1.157 + int offset = java_lang_Class::klass_offset_in_bytes();
1.158 + return load_klass_from_mirror_common(mirror, never_see_null,
1.159 + region, null_path,
1.160 + offset);
1.161 + }
1.162 + Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
1.163 + RegionNode* region, int null_path) {
1.164 + int offset = java_lang_Class::array_klass_offset_in_bytes();
1.165 + return load_klass_from_mirror_common(mirror, never_see_null,
1.166 + region, null_path,
1.167 + offset);
1.168 + }
1.169 + Node* generate_access_flags_guard(Node* kls,
1.170 + int modifier_mask, int modifier_bits,
1.171 + RegionNode* region);
1.172 + Node* generate_interface_guard(Node* kls, RegionNode* region);
1.173 + Node* generate_array_guard(Node* kls, RegionNode* region) {
1.174 + return generate_array_guard_common(kls, region, false, false);
1.175 + }
1.176 + Node* generate_non_array_guard(Node* kls, RegionNode* region) {
1.177 + return generate_array_guard_common(kls, region, false, true);
1.178 + }
1.179 + Node* generate_objArray_guard(Node* kls, RegionNode* region) {
1.180 + return generate_array_guard_common(kls, region, true, false);
1.181 + }
1.182 + Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
1.183 + return generate_array_guard_common(kls, region, true, true);
1.184 + }
1.185 + Node* generate_array_guard_common(Node* kls, RegionNode* region,
1.186 + bool obj_array, bool not_array);
1.187 + Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
1.188 + CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
1.189 + bool is_virtual = false, bool is_static = false);
1.190 + CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
1.191 + return generate_method_call(method_id, false, true);
1.192 + }
1.193 + CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
1.194 + return generate_method_call(method_id, true, false);
1.195 + }
1.196 + Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
1.197 +
1.198 + Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
1.199 + Node* make_string_method_node(int opcode, Node* str1, Node* str2);
1.200 + bool inline_string_compareTo();
1.201 + bool inline_string_indexOf();
1.202 + Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
1.203 + bool inline_string_equals();
1.204 + Node* round_double_node(Node* n);
1.205 + bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
1.206 + bool inline_math_native(vmIntrinsics::ID id);
1.207 + bool inline_trig(vmIntrinsics::ID id);
1.208 + bool inline_math(vmIntrinsics::ID id);
1.209 + template <typename OverflowOp>
1.210 + bool inline_math_overflow(Node* arg1, Node* arg2);
1.211 + void inline_math_mathExact(Node* math, Node* test);
1.212 + bool inline_math_addExactI(bool is_increment);
1.213 + bool inline_math_addExactL(bool is_increment);
1.214 + bool inline_math_multiplyExactI();
1.215 + bool inline_math_multiplyExactL();
1.216 + bool inline_math_negateExactI();
1.217 + bool inline_math_negateExactL();
1.218 + bool inline_math_subtractExactI(bool is_decrement);
1.219 + bool inline_math_subtractExactL(bool is_decrement);
1.220 + bool inline_exp();
1.221 + bool inline_pow();
1.222 + Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
1.223 + bool inline_min_max(vmIntrinsics::ID id);
1.224 + Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
1.225 + // This returns Type::AnyPtr, RawPtr, or OopPtr.
1.226 + int classify_unsafe_addr(Node* &base, Node* &offset);
1.227 + Node* make_unsafe_address(Node* base, Node* offset);
1.228 + // Helper for inline_unsafe_access.
1.229 + // Generates the guards that check whether the result of
1.230 + // Unsafe.getObject should be recorded in an SATB log buffer.
1.231 + void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
1.232 + bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
1.233 + bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
1.234 + static bool klass_needs_init_guard(Node* kls);
1.235 + bool inline_unsafe_allocate();
1.236 + bool inline_unsafe_copyMemory();
1.237 + bool inline_native_currentThread();
1.238 +#ifdef TRACE_HAVE_INTRINSICS
1.239 + bool inline_native_classID();
1.240 + bool inline_native_threadID();
1.241 +#endif
1.242 + bool inline_native_time_funcs(address method, const char* funcName);
1.243 + bool inline_native_isInterrupted();
1.244 + bool inline_native_Class_query(vmIntrinsics::ID id);
1.245 + bool inline_native_subtype_check();
1.246 +
1.247 + bool inline_native_newArray();
1.248 + bool inline_native_getLength();
1.249 + bool inline_array_copyOf(bool is_copyOfRange);
1.250 + bool inline_array_equals();
1.251 + void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
1.252 + bool inline_native_clone(bool is_virtual);
1.253 + bool inline_native_Reflection_getCallerClass();
1.254 + // Helper function for inlining native object hash method
1.255 + bool inline_native_hashcode(bool is_virtual, bool is_static);
1.256 + bool inline_native_getClass();
1.257 +
1.258 + // Helper functions for inlining arraycopy
1.259 + bool inline_arraycopy();
1.260 + void generate_arraycopy(const TypePtr* adr_type,
1.261 + BasicType basic_elem_type,
1.262 + Node* src, Node* src_offset,
1.263 + Node* dest, Node* dest_offset,
1.264 + Node* copy_length,
1.265 + bool disjoint_bases = false,
1.266 + bool length_never_negative = false,
1.267 + RegionNode* slow_region = NULL);
1.268 + AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
1.269 + RegionNode* slow_region);
1.270 + void generate_clear_array(const TypePtr* adr_type,
1.271 + Node* dest,
1.272 + BasicType basic_elem_type,
1.273 + Node* slice_off,
1.274 + Node* slice_len,
1.275 + Node* slice_end);
1.276 + bool generate_block_arraycopy(const TypePtr* adr_type,
1.277 + BasicType basic_elem_type,
1.278 + AllocateNode* alloc,
1.279 + Node* src, Node* src_offset,
1.280 + Node* dest, Node* dest_offset,
1.281 + Node* dest_size, bool dest_uninitialized);
1.282 + void generate_slow_arraycopy(const TypePtr* adr_type,
1.283 + Node* src, Node* src_offset,
1.284 + Node* dest, Node* dest_offset,
1.285 + Node* copy_length, bool dest_uninitialized);
1.286 + Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
1.287 + Node* dest_elem_klass,
1.288 + Node* src, Node* src_offset,
1.289 + Node* dest, Node* dest_offset,
1.290 + Node* copy_length, bool dest_uninitialized);
1.291 + Node* generate_generic_arraycopy(const TypePtr* adr_type,
1.292 + Node* src, Node* src_offset,
1.293 + Node* dest, Node* dest_offset,
1.294 + Node* copy_length, bool dest_uninitialized);
1.295 + void generate_unchecked_arraycopy(const TypePtr* adr_type,
1.296 + BasicType basic_elem_type,
1.297 + bool disjoint_bases,
1.298 + Node* src, Node* src_offset,
1.299 + Node* dest, Node* dest_offset,
1.300 + Node* copy_length, bool dest_uninitialized);
1.301 + typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
1.302 + bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
1.303 + bool inline_unsafe_ordered_store(BasicType type);
1.304 + bool inline_unsafe_fence(vmIntrinsics::ID id);
1.305 + bool inline_fp_conversions(vmIntrinsics::ID id);
1.306 + bool inline_number_methods(vmIntrinsics::ID id);
1.307 + bool inline_reference_get();
1.308 + bool inline_aescrypt_Block(vmIntrinsics::ID id);
1.309 + bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
1.310 + Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
1.311 + Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
1.312 + Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
1.313 + bool inline_encodeISOArray();
1.314 + bool inline_updateCRC32();
1.315 + bool inline_updateBytesCRC32();
1.316 + bool inline_updateByteBufferCRC32();
1.317 +};
1.318 +
1.319 +
1.320 +//---------------------------make_vm_intrinsic----------------------------
1.321 +CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
1.322 + vmIntrinsics::ID id = m->intrinsic_id();
1.323 + assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
1.324 +
1.325 + if (DisableIntrinsic[0] != '\0'
1.326 + && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
1.327 + // disabled by a user request on the command line:
1.328 + // example: -XX:DisableIntrinsic=_hashCode,_getClass
1.329 + return NULL;
1.330 + }
1.331 +
1.332 + if (!m->is_loaded()) {
1.333 + // do not attempt to inline unloaded methods
1.334 + return NULL;
1.335 + }
1.336 +
1.337 + // Only a few intrinsics implement a virtual dispatch.
1.338 + // They are expensive calls which are also frequently overridden.
1.339 + if (is_virtual) {
1.340 + switch (id) {
1.341 + case vmIntrinsics::_hashCode:
1.342 + case vmIntrinsics::_clone:
1.343 + // OK, Object.hashCode and Object.clone intrinsics come in both flavors
1.344 + break;
1.345 + default:
1.346 + return NULL;
1.347 + }
1.348 + }
1.349 +
1.350 + // -XX:-InlineNatives disables nearly all intrinsics:
1.351 + if (!InlineNatives) {
1.352 + switch (id) {
1.353 + case vmIntrinsics::_indexOf:
1.354 + case vmIntrinsics::_compareTo:
1.355 + case vmIntrinsics::_equals:
1.356 + case vmIntrinsics::_equalsC:
1.357 + case vmIntrinsics::_getAndAddInt:
1.358 + case vmIntrinsics::_getAndAddLong:
1.359 + case vmIntrinsics::_getAndSetInt:
1.360 + case vmIntrinsics::_getAndSetLong:
1.361 + case vmIntrinsics::_getAndSetObject:
1.362 + case vmIntrinsics::_loadFence:
1.363 + case vmIntrinsics::_storeFence:
1.364 + case vmIntrinsics::_fullFence:
1.365 + break; // InlineNatives does not control String.compareTo
1.366 + case vmIntrinsics::_Reference_get:
1.367 + break; // InlineNatives does not control Reference.get
1.368 + default:
1.369 + return NULL;
1.370 + }
1.371 + }
1.372 +
1.373 + bool is_predicted = false;
1.374 + bool does_virtual_dispatch = false;
1.375 +
1.376 + switch (id) {
1.377 + case vmIntrinsics::_compareTo:
1.378 + if (!SpecialStringCompareTo) return NULL;
1.379 + if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
1.380 + break;
1.381 + case vmIntrinsics::_indexOf:
1.382 + if (!SpecialStringIndexOf) return NULL;
1.383 + break;
1.384 + case vmIntrinsics::_equals:
1.385 + if (!SpecialStringEquals) return NULL;
1.386 + if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
1.387 + break;
1.388 + case vmIntrinsics::_equalsC:
1.389 + if (!SpecialArraysEquals) return NULL;
1.390 + if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
1.391 + break;
1.392 + case vmIntrinsics::_arraycopy:
1.393 + if (!InlineArrayCopy) return NULL;
1.394 + break;
1.395 + case vmIntrinsics::_copyMemory:
1.396 + if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
1.397 + if (!InlineArrayCopy) return NULL;
1.398 + break;
1.399 + case vmIntrinsics::_hashCode:
1.400 + if (!InlineObjectHash) return NULL;
1.401 + does_virtual_dispatch = true;
1.402 + break;
1.403 + case vmIntrinsics::_clone:
1.404 + does_virtual_dispatch = true;
1.405 + case vmIntrinsics::_copyOf:
1.406 + case vmIntrinsics::_copyOfRange:
1.407 + if (!InlineObjectCopy) return NULL;
1.408 + // These also use the arraycopy intrinsic mechanism:
1.409 + if (!InlineArrayCopy) return NULL;
1.410 + break;
1.411 + case vmIntrinsics::_encodeISOArray:
1.412 + if (!SpecialEncodeISOArray) return NULL;
1.413 + if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
1.414 + break;
1.415 + case vmIntrinsics::_checkIndex:
1.416 + // We do not intrinsify this. The optimizer does fine with it.
1.417 + return NULL;
1.418 +
1.419 + case vmIntrinsics::_getCallerClass:
1.420 + if (!UseNewReflection) return NULL;
1.421 + if (!InlineReflectionGetCallerClass) return NULL;
1.422 + if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
1.423 + break;
1.424 +
1.425 + case vmIntrinsics::_bitCount_i:
1.426 + if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
1.427 + break;
1.428 +
1.429 + case vmIntrinsics::_bitCount_l:
1.430 + if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
1.431 + break;
1.432 +
1.433 + case vmIntrinsics::_numberOfLeadingZeros_i:
1.434 + if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
1.435 + break;
1.436 +
1.437 + case vmIntrinsics::_numberOfLeadingZeros_l:
1.438 + if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
1.439 + break;
1.440 +
1.441 + case vmIntrinsics::_numberOfTrailingZeros_i:
1.442 + if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
1.443 + break;
1.444 +
1.445 + case vmIntrinsics::_numberOfTrailingZeros_l:
1.446 + if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
1.447 + break;
1.448 +
1.449 + case vmIntrinsics::_reverseBytes_c:
1.450 + if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
1.451 + break;
1.452 + case vmIntrinsics::_reverseBytes_s:
1.453 + if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
1.454 + break;
1.455 + case vmIntrinsics::_reverseBytes_i:
1.456 + if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
1.457 + break;
1.458 + case vmIntrinsics::_reverseBytes_l:
1.459 + if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
1.460 + break;
1.461 +
1.462 + case vmIntrinsics::_Reference_get:
1.463 + // Use the intrinsic version of Reference.get() so that the value in
1.464 + // the referent field can be registered by the G1 pre-barrier code.
1.465 + // Also add memory barrier to prevent commoning reads from this field
1.466 + // across safepoint since GC can change it value.
1.467 + break;
1.468 +
1.469 + case vmIntrinsics::_compareAndSwapObject:
1.470 +#ifdef _LP64
1.471 + if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
1.472 +#endif
1.473 + break;
1.474 +
1.475 + case vmIntrinsics::_compareAndSwapLong:
1.476 + if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
1.477 + break;
1.478 +
1.479 + case vmIntrinsics::_getAndAddInt:
1.480 + if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
1.481 + break;
1.482 +
1.483 + case vmIntrinsics::_getAndAddLong:
1.484 + if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
1.485 + break;
1.486 +
1.487 + case vmIntrinsics::_getAndSetInt:
1.488 + if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
1.489 + break;
1.490 +
1.491 + case vmIntrinsics::_getAndSetLong:
1.492 + if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
1.493 + break;
1.494 +
1.495 + case vmIntrinsics::_getAndSetObject:
1.496 +#ifdef _LP64
1.497 + if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
1.498 + if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
1.499 + break;
1.500 +#else
1.501 + if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
1.502 + break;
1.503 +#endif
1.504 +
1.505 + case vmIntrinsics::_aescrypt_encryptBlock:
1.506 + case vmIntrinsics::_aescrypt_decryptBlock:
1.507 + if (!UseAESIntrinsics) return NULL;
1.508 + break;
1.509 +
1.510 + case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1.511 + case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1.512 + if (!UseAESIntrinsics) return NULL;
1.513 + // these two require the predicated logic
1.514 + is_predicted = true;
1.515 + break;
1.516 +
1.517 + case vmIntrinsics::_updateCRC32:
1.518 + case vmIntrinsics::_updateBytesCRC32:
1.519 + case vmIntrinsics::_updateByteBufferCRC32:
1.520 + if (!UseCRC32Intrinsics) return NULL;
1.521 + break;
1.522 +
1.523 + case vmIntrinsics::_incrementExactI:
1.524 + case vmIntrinsics::_addExactI:
1.525 + if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
1.526 + break;
1.527 + case vmIntrinsics::_incrementExactL:
1.528 + case vmIntrinsics::_addExactL:
1.529 + if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
1.530 + break;
1.531 + case vmIntrinsics::_decrementExactI:
1.532 + case vmIntrinsics::_subtractExactI:
1.533 + if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
1.534 + break;
1.535 + case vmIntrinsics::_decrementExactL:
1.536 + case vmIntrinsics::_subtractExactL:
1.537 + if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
1.538 + break;
1.539 + case vmIntrinsics::_negateExactI:
1.540 + if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
1.541 + break;
1.542 + case vmIntrinsics::_negateExactL:
1.543 + if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
1.544 + break;
1.545 + case vmIntrinsics::_multiplyExactI:
1.546 + if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
1.547 + break;
1.548 + case vmIntrinsics::_multiplyExactL:
1.549 + if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
1.550 + break;
1.551 +
1.552 + default:
1.553 + assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
1.554 + assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
1.555 + break;
1.556 + }
1.557 +
1.558 + // -XX:-InlineClassNatives disables natives from the Class class.
1.559 + // The flag applies to all reflective calls, notably Array.newArray
1.560 + // (visible to Java programmers as Array.newInstance).
1.561 + if (m->holder()->name() == ciSymbol::java_lang_Class() ||
1.562 + m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
1.563 + if (!InlineClassNatives) return NULL;
1.564 + }
1.565 +
1.566 + // -XX:-InlineThreadNatives disables natives from the Thread class.
1.567 + if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
1.568 + if (!InlineThreadNatives) return NULL;
1.569 + }
1.570 +
1.571 + // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
1.572 + if (m->holder()->name() == ciSymbol::java_lang_Math() ||
1.573 + m->holder()->name() == ciSymbol::java_lang_Float() ||
1.574 + m->holder()->name() == ciSymbol::java_lang_Double()) {
1.575 + if (!InlineMathNatives) return NULL;
1.576 + }
1.577 +
1.578 + // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
1.579 + if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
1.580 + if (!InlineUnsafeOps) return NULL;
1.581 + }
1.582 +
1.583 + return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
1.584 +}
1.585 +
1.586 +//----------------------register_library_intrinsics-----------------------
1.587 +// Initialize this file's data structures, for each Compile instance.
1.588 +void Compile::register_library_intrinsics() {
1.589 + // Nothing to do here.
1.590 +}
1.591 +
1.592 +JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
1.593 + LibraryCallKit kit(jvms, this);
1.594 + Compile* C = kit.C;
1.595 + int nodes = C->unique();
1.596 +#ifndef PRODUCT
1.597 + if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
1.598 + char buf[1000];
1.599 + const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
1.600 + tty->print_cr("Intrinsic %s", str);
1.601 + }
1.602 +#endif
1.603 + ciMethod* callee = kit.callee();
1.604 + const int bci = kit.bci();
1.605 +
1.606 + // Try to inline the intrinsic.
1.607 + if (kit.try_to_inline()) {
1.608 + if (C->print_intrinsics() || C->print_inlining()) {
1.609 + C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
1.610 + }
1.611 + C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
1.612 + if (C->log()) {
1.613 + C->log()->elem("intrinsic id='%s'%s nodes='%d'",
1.614 + vmIntrinsics::name_at(intrinsic_id()),
1.615 + (is_virtual() ? " virtual='1'" : ""),
1.616 + C->unique() - nodes);
1.617 + }
1.618 + // Push the result from the inlined method onto the stack.
1.619 + kit.push_result();
1.620 + return kit.transfer_exceptions_into_jvms();
1.621 + }
1.622 +
1.623 + // The intrinsic bailed out
1.624 + if (C->print_intrinsics() || C->print_inlining()) {
1.625 + if (jvms->has_method()) {
1.626 + // Not a root compile.
1.627 + const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
1.628 + C->print_inlining(callee, jvms->depth() - 1, bci, msg);
1.629 + } else {
1.630 + // Root compile
1.631 + tty->print("Did not generate intrinsic %s%s at bci:%d in",
1.632 + vmIntrinsics::name_at(intrinsic_id()),
1.633 + (is_virtual() ? " (virtual)" : ""), bci);
1.634 + }
1.635 + }
1.636 + C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
1.637 + return NULL;
1.638 +}
1.639 +
1.640 +Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
1.641 + LibraryCallKit kit(jvms, this);
1.642 + Compile* C = kit.C;
1.643 + int nodes = C->unique();
1.644 +#ifndef PRODUCT
1.645 + assert(is_predicted(), "sanity");
1.646 + if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
1.647 + char buf[1000];
1.648 + const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
1.649 + tty->print_cr("Predicate for intrinsic %s", str);
1.650 + }
1.651 +#endif
1.652 + ciMethod* callee = kit.callee();
1.653 + const int bci = kit.bci();
1.654 +
1.655 + Node* slow_ctl = kit.try_to_predicate();
1.656 + if (!kit.failing()) {
1.657 + if (C->print_intrinsics() || C->print_inlining()) {
1.658 + C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
1.659 + }
1.660 + C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
1.661 + if (C->log()) {
1.662 + C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
1.663 + vmIntrinsics::name_at(intrinsic_id()),
1.664 + (is_virtual() ? " virtual='1'" : ""),
1.665 + C->unique() - nodes);
1.666 + }
1.667 + return slow_ctl; // Could be NULL if the check folds.
1.668 + }
1.669 +
1.670 + // The intrinsic bailed out
1.671 + if (C->print_intrinsics() || C->print_inlining()) {
1.672 + if (jvms->has_method()) {
1.673 + // Not a root compile.
1.674 + const char* msg = "failed to generate predicate for intrinsic";
1.675 + C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
1.676 + } else {
1.677 + // Root compile
1.678 + C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
1.679 + vmIntrinsics::name_at(intrinsic_id()),
1.680 + (is_virtual() ? " (virtual)" : ""), bci);
1.681 + }
1.682 + }
1.683 + C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
1.684 + return NULL;
1.685 +}
1.686 +
1.687 +bool LibraryCallKit::try_to_inline() {
1.688 + // Handle symbolic names for otherwise undistinguished boolean switches:
1.689 + const bool is_store = true;
1.690 + const bool is_native_ptr = true;
1.691 + const bool is_static = true;
1.692 + const bool is_volatile = true;
1.693 +
1.694 + if (!jvms()->has_method()) {
1.695 + // Root JVMState has a null method.
1.696 + assert(map()->memory()->Opcode() == Op_Parm, "");
1.697 + // Insert the memory aliasing node
1.698 + set_all_memory(reset_memory());
1.699 + }
1.700 + assert(merged_memory(), "");
1.701 +
1.702 +
1.703 + switch (intrinsic_id()) {
1.704 + case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
1.705 + case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
1.706 + case vmIntrinsics::_getClass: return inline_native_getClass();
1.707 +
1.708 + case vmIntrinsics::_dsin:
1.709 + case vmIntrinsics::_dcos:
1.710 + case vmIntrinsics::_dtan:
1.711 + case vmIntrinsics::_dabs:
1.712 + case vmIntrinsics::_datan2:
1.713 + case vmIntrinsics::_dsqrt:
1.714 + case vmIntrinsics::_dexp:
1.715 + case vmIntrinsics::_dlog:
1.716 + case vmIntrinsics::_dlog10:
1.717 + case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
1.718 +
1.719 + case vmIntrinsics::_min:
1.720 + case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
1.721 +
1.722 + case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
1.723 + case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
1.724 + case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
1.725 + case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
1.726 + case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
1.727 + case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
1.728 + case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
1.729 + case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
1.730 + case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
1.731 + case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
1.732 + case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
1.733 + case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
1.734 +
1.735 + case vmIntrinsics::_arraycopy: return inline_arraycopy();
1.736 +
1.737 + case vmIntrinsics::_compareTo: return inline_string_compareTo();
1.738 + case vmIntrinsics::_indexOf: return inline_string_indexOf();
1.739 + case vmIntrinsics::_equals: return inline_string_equals();
1.740 +
1.741 + case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
1.742 + case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
1.743 + case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
1.744 + case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
1.745 + case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
1.746 + case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
1.747 + case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
1.748 + case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
1.749 + case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
1.750 +
1.751 + case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
1.752 + case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
1.753 + case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
1.754 + case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
1.755 + case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
1.756 + case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
1.757 + case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
1.758 + case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
1.759 + case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
1.760 +
1.761 + case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
1.762 + case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
1.763 + case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
1.764 + case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
1.765 + case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
1.766 + case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
1.767 + case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
1.768 + case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
1.769 +
1.770 + case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
1.771 + case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
1.772 + case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
1.773 + case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
1.774 + case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
1.775 + case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
1.776 + case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
1.777 + case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
1.778 +
1.779 + case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
1.780 + case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
1.781 + case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
1.782 + case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
1.783 + case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
1.784 + case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
1.785 + case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
1.786 + case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
1.787 + case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
1.788 +
1.789 + case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
1.790 + case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
1.791 + case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
1.792 + case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
1.793 + case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
1.794 + case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
1.795 + case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
1.796 + case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
1.797 + case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
1.798 +
1.799 + case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
1.800 + case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
1.801 + case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
1.802 + case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
1.803 +
1.804 + case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
1.805 + case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
1.806 + case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
1.807 +
1.808 + case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
1.809 + case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
1.810 + case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
1.811 +
1.812 + case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
1.813 + case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
1.814 + case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
1.815 + case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
1.816 + case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
1.817 +
1.818 + case vmIntrinsics::_loadFence:
1.819 + case vmIntrinsics::_storeFence:
1.820 + case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
1.821 +
1.822 + case vmIntrinsics::_currentThread: return inline_native_currentThread();
1.823 + case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
1.824 +
1.825 +#ifdef TRACE_HAVE_INTRINSICS
1.826 + case vmIntrinsics::_classID: return inline_native_classID();
1.827 + case vmIntrinsics::_threadID: return inline_native_threadID();
1.828 + case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
1.829 +#endif
1.830 + case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
1.831 + case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
1.832 + case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
1.833 + case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
1.834 + case vmIntrinsics::_newArray: return inline_native_newArray();
1.835 + case vmIntrinsics::_getLength: return inline_native_getLength();
1.836 + case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
1.837 + case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
1.838 + case vmIntrinsics::_equalsC: return inline_array_equals();
1.839 + case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
1.840 +
1.841 + case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
1.842 +
1.843 + case vmIntrinsics::_isInstance:
1.844 + case vmIntrinsics::_getModifiers:
1.845 + case vmIntrinsics::_isInterface:
1.846 + case vmIntrinsics::_isArray:
1.847 + case vmIntrinsics::_isPrimitive:
1.848 + case vmIntrinsics::_getSuperclass:
1.849 + case vmIntrinsics::_getComponentType:
1.850 + case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
1.851 +
1.852 + case vmIntrinsics::_floatToRawIntBits:
1.853 + case vmIntrinsics::_floatToIntBits:
1.854 + case vmIntrinsics::_intBitsToFloat:
1.855 + case vmIntrinsics::_doubleToRawLongBits:
1.856 + case vmIntrinsics::_doubleToLongBits:
1.857 + case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
1.858 +
1.859 + case vmIntrinsics::_numberOfLeadingZeros_i:
1.860 + case vmIntrinsics::_numberOfLeadingZeros_l:
1.861 + case vmIntrinsics::_numberOfTrailingZeros_i:
1.862 + case vmIntrinsics::_numberOfTrailingZeros_l:
1.863 + case vmIntrinsics::_bitCount_i:
1.864 + case vmIntrinsics::_bitCount_l:
1.865 + case vmIntrinsics::_reverseBytes_i:
1.866 + case vmIntrinsics::_reverseBytes_l:
1.867 + case vmIntrinsics::_reverseBytes_s:
1.868 + case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
1.869 +
1.870 + case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
1.871 +
1.872 + case vmIntrinsics::_Reference_get: return inline_reference_get();
1.873 +
1.874 + case vmIntrinsics::_aescrypt_encryptBlock:
1.875 + case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
1.876 +
1.877 + case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1.878 + case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1.879 + return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
1.880 +
1.881 + case vmIntrinsics::_encodeISOArray:
1.882 + return inline_encodeISOArray();
1.883 +
1.884 + case vmIntrinsics::_updateCRC32:
1.885 + return inline_updateCRC32();
1.886 + case vmIntrinsics::_updateBytesCRC32:
1.887 + return inline_updateBytesCRC32();
1.888 + case vmIntrinsics::_updateByteBufferCRC32:
1.889 + return inline_updateByteBufferCRC32();
1.890 +
1.891 + default:
1.892 + // If you get here, it may be that someone has added a new intrinsic
1.893 + // to the list in vmSymbols.hpp without implementing it here.
1.894 +#ifndef PRODUCT
1.895 + if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1.896 + tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
1.897 + vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1.898 + }
1.899 +#endif
1.900 + return false;
1.901 + }
1.902 +}
1.903 +
1.904 +Node* LibraryCallKit::try_to_predicate() {
1.905 + if (!jvms()->has_method()) {
1.906 + // Root JVMState has a null method.
1.907 + assert(map()->memory()->Opcode() == Op_Parm, "");
1.908 + // Insert the memory aliasing node
1.909 + set_all_memory(reset_memory());
1.910 + }
1.911 + assert(merged_memory(), "");
1.912 +
1.913 + switch (intrinsic_id()) {
1.914 + case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1.915 + return inline_cipherBlockChaining_AESCrypt_predicate(false);
1.916 + case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1.917 + return inline_cipherBlockChaining_AESCrypt_predicate(true);
1.918 +
1.919 + default:
1.920 + // If you get here, it may be that someone has added a new intrinsic
1.921 + // to the list in vmSymbols.hpp without implementing it here.
1.922 +#ifndef PRODUCT
1.923 + if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1.924 + tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1.925 + vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1.926 + }
1.927 +#endif
1.928 + Node* slow_ctl = control();
1.929 + set_control(top()); // No fast path instrinsic
1.930 + return slow_ctl;
1.931 + }
1.932 +}
1.933 +
1.934 +//------------------------------set_result-------------------------------
1.935 +// Helper function for finishing intrinsics.
1.936 +void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1.937 + record_for_igvn(region);
1.938 + set_control(_gvn.transform(region));
1.939 + set_result( _gvn.transform(value));
1.940 + assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1.941 +}
1.942 +
1.943 +//------------------------------generate_guard---------------------------
1.944 +// Helper function for generating guarded fast-slow graph structures.
1.945 +// The given 'test', if true, guards a slow path. If the test fails
1.946 +// then a fast path can be taken. (We generally hope it fails.)
1.947 +// In all cases, GraphKit::control() is updated to the fast path.
1.948 +// The returned value represents the control for the slow path.
1.949 +// The return value is never 'top'; it is either a valid control
1.950 +// or NULL if it is obvious that the slow path can never be taken.
1.951 +// Also, if region and the slow control are not NULL, the slow edge
1.952 +// is appended to the region.
1.953 +Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1.954 + if (stopped()) {
1.955 + // Already short circuited.
1.956 + return NULL;
1.957 + }
1.958 +
1.959 + // Build an if node and its projections.
1.960 + // If test is true we take the slow path, which we assume is uncommon.
1.961 + if (_gvn.type(test) == TypeInt::ZERO) {
1.962 + // The slow branch is never taken. No need to build this guard.
1.963 + return NULL;
1.964 + }
1.965 +
1.966 + IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1.967 +
1.968 + Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
1.969 + if (if_slow == top()) {
1.970 + // The slow branch is never taken. No need to build this guard.
1.971 + return NULL;
1.972 + }
1.973 +
1.974 + if (region != NULL)
1.975 + region->add_req(if_slow);
1.976 +
1.977 + Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
1.978 + set_control(if_fast);
1.979 +
1.980 + return if_slow;
1.981 +}
1.982 +
1.983 +inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1.984 + return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1.985 +}
1.986 +inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1.987 + return generate_guard(test, region, PROB_FAIR);
1.988 +}
1.989 +
1.990 +inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1.991 + Node* *pos_index) {
1.992 + if (stopped())
1.993 + return NULL; // already stopped
1.994 + if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1.995 + return NULL; // index is already adequately typed
1.996 + Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1.997 + Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1.998 + Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1.999 + if (is_neg != NULL && pos_index != NULL) {
1.1000 + // Emulate effect of Parse::adjust_map_after_if.
1.1001 + Node* ccast = new (C) CastIINode(index, TypeInt::POS);
1.1002 + ccast->set_req(0, control());
1.1003 + (*pos_index) = _gvn.transform(ccast);
1.1004 + }
1.1005 + return is_neg;
1.1006 +}
1.1007 +
1.1008 +inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1.1009 + Node* *pos_index) {
1.1010 + if (stopped())
1.1011 + return NULL; // already stopped
1.1012 + if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1.1013 + return NULL; // index is already adequately typed
1.1014 + Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1.1015 + BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1.1016 + Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1.1017 + Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1.1018 + if (is_notp != NULL && pos_index != NULL) {
1.1019 + // Emulate effect of Parse::adjust_map_after_if.
1.1020 + Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1.1021 + ccast->set_req(0, control());
1.1022 + (*pos_index) = _gvn.transform(ccast);
1.1023 + }
1.1024 + return is_notp;
1.1025 +}
1.1026 +
1.1027 +// Make sure that 'position' is a valid limit index, in [0..length].
1.1028 +// There are two equivalent plans for checking this:
1.1029 +// A. (offset + copyLength) unsigned<= arrayLength
1.1030 +// B. offset <= (arrayLength - copyLength)
1.1031 +// We require that all of the values above, except for the sum and
1.1032 +// difference, are already known to be non-negative.
1.1033 +// Plan A is robust in the face of overflow, if offset and copyLength
1.1034 +// are both hugely positive.
1.1035 +//
1.1036 +// Plan B is less direct and intuitive, but it does not overflow at
1.1037 +// all, since the difference of two non-negatives is always
1.1038 +// representable. Whenever Java methods must perform the equivalent
1.1039 +// check they generally use Plan B instead of Plan A.
1.1040 +// For the moment we use Plan A.
1.1041 +inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1.1042 + Node* subseq_length,
1.1043 + Node* array_length,
1.1044 + RegionNode* region) {
1.1045 + if (stopped())
1.1046 + return NULL; // already stopped
1.1047 + bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1.1048 + if (zero_offset && subseq_length->eqv_uncast(array_length))
1.1049 + return NULL; // common case of whole-array copy
1.1050 + Node* last = subseq_length;
1.1051 + if (!zero_offset) // last += offset
1.1052 + last = _gvn.transform(new (C) AddINode(last, offset));
1.1053 + Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1.1054 + Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1.1055 + Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1.1056 + return is_over;
1.1057 +}
1.1058 +
1.1059 +
1.1060 +//--------------------------generate_current_thread--------------------
1.1061 +Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1.1062 + ciKlass* thread_klass = env()->Thread_klass();
1.1063 + const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1.1064 + Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1.1065 + Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1.1066 + Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1.1067 + tls_output = thread;
1.1068 + return threadObj;
1.1069 +}
1.1070 +
1.1071 +
1.1072 +//------------------------------make_string_method_node------------------------
1.1073 +// Helper method for String intrinsic functions. This version is called
1.1074 +// with str1 and str2 pointing to String object nodes.
1.1075 +//
1.1076 +Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1.1077 + Node* no_ctrl = NULL;
1.1078 +
1.1079 + // Get start addr of string
1.1080 + Node* str1_value = load_String_value(no_ctrl, str1);
1.1081 + Node* str1_offset = load_String_offset(no_ctrl, str1);
1.1082 + Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1.1083 +
1.1084 + // Get length of string 1
1.1085 + Node* str1_len = load_String_length(no_ctrl, str1);
1.1086 +
1.1087 + Node* str2_value = load_String_value(no_ctrl, str2);
1.1088 + Node* str2_offset = load_String_offset(no_ctrl, str2);
1.1089 + Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1.1090 +
1.1091 + Node* str2_len = NULL;
1.1092 + Node* result = NULL;
1.1093 +
1.1094 + switch (opcode) {
1.1095 + case Op_StrIndexOf:
1.1096 + // Get length of string 2
1.1097 + str2_len = load_String_length(no_ctrl, str2);
1.1098 +
1.1099 + result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1.1100 + str1_start, str1_len, str2_start, str2_len);
1.1101 + break;
1.1102 + case Op_StrComp:
1.1103 + // Get length of string 2
1.1104 + str2_len = load_String_length(no_ctrl, str2);
1.1105 +
1.1106 + result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1.1107 + str1_start, str1_len, str2_start, str2_len);
1.1108 + break;
1.1109 + case Op_StrEquals:
1.1110 + result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1.1111 + str1_start, str2_start, str1_len);
1.1112 + break;
1.1113 + default:
1.1114 + ShouldNotReachHere();
1.1115 + return NULL;
1.1116 + }
1.1117 +
1.1118 + // All these intrinsics have checks.
1.1119 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1120 +
1.1121 + return _gvn.transform(result);
1.1122 +}
1.1123 +
1.1124 +// Helper method for String intrinsic functions. This version is called
1.1125 +// with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1.1126 +// to Int nodes containing the lenghts of str1 and str2.
1.1127 +//
1.1128 +Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1.1129 + Node* result = NULL;
1.1130 + switch (opcode) {
1.1131 + case Op_StrIndexOf:
1.1132 + result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1.1133 + str1_start, cnt1, str2_start, cnt2);
1.1134 + break;
1.1135 + case Op_StrComp:
1.1136 + result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1.1137 + str1_start, cnt1, str2_start, cnt2);
1.1138 + break;
1.1139 + case Op_StrEquals:
1.1140 + result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1.1141 + str1_start, str2_start, cnt1);
1.1142 + break;
1.1143 + default:
1.1144 + ShouldNotReachHere();
1.1145 + return NULL;
1.1146 + }
1.1147 +
1.1148 + // All these intrinsics have checks.
1.1149 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1150 +
1.1151 + return _gvn.transform(result);
1.1152 +}
1.1153 +
1.1154 +//------------------------------inline_string_compareTo------------------------
1.1155 +// public int java.lang.String.compareTo(String anotherString);
1.1156 +bool LibraryCallKit::inline_string_compareTo() {
1.1157 + Node* receiver = null_check(argument(0));
1.1158 + Node* arg = null_check(argument(1));
1.1159 + if (stopped()) {
1.1160 + return true;
1.1161 + }
1.1162 + set_result(make_string_method_node(Op_StrComp, receiver, arg));
1.1163 + return true;
1.1164 +}
1.1165 +
1.1166 +//------------------------------inline_string_equals------------------------
1.1167 +bool LibraryCallKit::inline_string_equals() {
1.1168 + Node* receiver = null_check_receiver();
1.1169 + // NOTE: Do not null check argument for String.equals() because spec
1.1170 + // allows to specify NULL as argument.
1.1171 + Node* argument = this->argument(1);
1.1172 + if (stopped()) {
1.1173 + return true;
1.1174 + }
1.1175 +
1.1176 + // paths (plus control) merge
1.1177 + RegionNode* region = new (C) RegionNode(5);
1.1178 + Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1.1179 +
1.1180 + // does source == target string?
1.1181 + Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1.1182 + Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1.1183 +
1.1184 + Node* if_eq = generate_slow_guard(bol, NULL);
1.1185 + if (if_eq != NULL) {
1.1186 + // receiver == argument
1.1187 + phi->init_req(2, intcon(1));
1.1188 + region->init_req(2, if_eq);
1.1189 + }
1.1190 +
1.1191 + // get String klass for instanceOf
1.1192 + ciInstanceKlass* klass = env()->String_klass();
1.1193 +
1.1194 + if (!stopped()) {
1.1195 + Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1.1196 + Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1.1197 + Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1.1198 +
1.1199 + Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1.1200 + //instanceOf == true, fallthrough
1.1201 +
1.1202 + if (inst_false != NULL) {
1.1203 + phi->init_req(3, intcon(0));
1.1204 + region->init_req(3, inst_false);
1.1205 + }
1.1206 + }
1.1207 +
1.1208 + if (!stopped()) {
1.1209 + const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1.1210 +
1.1211 + // Properly cast the argument to String
1.1212 + argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1.1213 + // This path is taken only when argument's type is String:NotNull.
1.1214 + argument = cast_not_null(argument, false);
1.1215 +
1.1216 + Node* no_ctrl = NULL;
1.1217 +
1.1218 + // Get start addr of receiver
1.1219 + Node* receiver_val = load_String_value(no_ctrl, receiver);
1.1220 + Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1.1221 + Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1.1222 +
1.1223 + // Get length of receiver
1.1224 + Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1.1225 +
1.1226 + // Get start addr of argument
1.1227 + Node* argument_val = load_String_value(no_ctrl, argument);
1.1228 + Node* argument_offset = load_String_offset(no_ctrl, argument);
1.1229 + Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1.1230 +
1.1231 + // Get length of argument
1.1232 + Node* argument_cnt = load_String_length(no_ctrl, argument);
1.1233 +
1.1234 + // Check for receiver count != argument count
1.1235 + Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1.1236 + Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1.1237 + Node* if_ne = generate_slow_guard(bol, NULL);
1.1238 + if (if_ne != NULL) {
1.1239 + phi->init_req(4, intcon(0));
1.1240 + region->init_req(4, if_ne);
1.1241 + }
1.1242 +
1.1243 + // Check for count == 0 is done by assembler code for StrEquals.
1.1244 +
1.1245 + if (!stopped()) {
1.1246 + Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1.1247 + phi->init_req(1, equals);
1.1248 + region->init_req(1, control());
1.1249 + }
1.1250 + }
1.1251 +
1.1252 + // post merge
1.1253 + set_control(_gvn.transform(region));
1.1254 + record_for_igvn(region);
1.1255 +
1.1256 + set_result(_gvn.transform(phi));
1.1257 + return true;
1.1258 +}
1.1259 +
1.1260 +//------------------------------inline_array_equals----------------------------
1.1261 +bool LibraryCallKit::inline_array_equals() {
1.1262 + Node* arg1 = argument(0);
1.1263 + Node* arg2 = argument(1);
1.1264 + set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1.1265 + return true;
1.1266 +}
1.1267 +
1.1268 +// Java version of String.indexOf(constant string)
1.1269 +// class StringDecl {
1.1270 +// StringDecl(char[] ca) {
1.1271 +// offset = 0;
1.1272 +// count = ca.length;
1.1273 +// value = ca;
1.1274 +// }
1.1275 +// int offset;
1.1276 +// int count;
1.1277 +// char[] value;
1.1278 +// }
1.1279 +//
1.1280 +// static int string_indexOf_J(StringDecl string_object, char[] target_object,
1.1281 +// int targetOffset, int cache_i, int md2) {
1.1282 +// int cache = cache_i;
1.1283 +// int sourceOffset = string_object.offset;
1.1284 +// int sourceCount = string_object.count;
1.1285 +// int targetCount = target_object.length;
1.1286 +//
1.1287 +// int targetCountLess1 = targetCount - 1;
1.1288 +// int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1.1289 +//
1.1290 +// char[] source = string_object.value;
1.1291 +// char[] target = target_object;
1.1292 +// int lastChar = target[targetCountLess1];
1.1293 +//
1.1294 +// outer_loop:
1.1295 +// for (int i = sourceOffset; i < sourceEnd; ) {
1.1296 +// int src = source[i + targetCountLess1];
1.1297 +// if (src == lastChar) {
1.1298 +// // With random strings and a 4-character alphabet,
1.1299 +// // reverse matching at this point sets up 0.8% fewer
1.1300 +// // frames, but (paradoxically) makes 0.3% more probes.
1.1301 +// // Since those probes are nearer the lastChar probe,
1.1302 +// // there is may be a net D$ win with reverse matching.
1.1303 +// // But, reversing loop inhibits unroll of inner loop
1.1304 +// // for unknown reason. So, does running outer loop from
1.1305 +// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1.1306 +// for (int j = 0; j < targetCountLess1; j++) {
1.1307 +// if (target[targetOffset + j] != source[i+j]) {
1.1308 +// if ((cache & (1 << source[i+j])) == 0) {
1.1309 +// if (md2 < j+1) {
1.1310 +// i += j+1;
1.1311 +// continue outer_loop;
1.1312 +// }
1.1313 +// }
1.1314 +// i += md2;
1.1315 +// continue outer_loop;
1.1316 +// }
1.1317 +// }
1.1318 +// return i - sourceOffset;
1.1319 +// }
1.1320 +// if ((cache & (1 << src)) == 0) {
1.1321 +// i += targetCountLess1;
1.1322 +// } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1.1323 +// i++;
1.1324 +// }
1.1325 +// return -1;
1.1326 +// }
1.1327 +
1.1328 +//------------------------------string_indexOf------------------------
1.1329 +Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1.1330 + jint cache_i, jint md2_i) {
1.1331 +
1.1332 + Node* no_ctrl = NULL;
1.1333 + float likely = PROB_LIKELY(0.9);
1.1334 + float unlikely = PROB_UNLIKELY(0.9);
1.1335 +
1.1336 + const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1.1337 +
1.1338 + Node* source = load_String_value(no_ctrl, string_object);
1.1339 + Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1.1340 + Node* sourceCount = load_String_length(no_ctrl, string_object);
1.1341 +
1.1342 + Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1.1343 + jint target_length = target_array->length();
1.1344 + const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1.1345 + const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1.1346 +
1.1347 + // String.value field is known to be @Stable.
1.1348 + if (UseImplicitStableValues) {
1.1349 + target = cast_array_to_stable(target, target_type);
1.1350 + }
1.1351 +
1.1352 + IdealKit kit(this, false, true);
1.1353 +#define __ kit.
1.1354 + Node* zero = __ ConI(0);
1.1355 + Node* one = __ ConI(1);
1.1356 + Node* cache = __ ConI(cache_i);
1.1357 + Node* md2 = __ ConI(md2_i);
1.1358 + Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1.1359 + Node* targetCount = __ ConI(target_length);
1.1360 + Node* targetCountLess1 = __ ConI(target_length - 1);
1.1361 + Node* targetOffset = __ ConI(targetOffset_i);
1.1362 + Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1.1363 +
1.1364 + IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1.1365 + Node* outer_loop = __ make_label(2 /* goto */);
1.1366 + Node* return_ = __ make_label(1);
1.1367 +
1.1368 + __ set(rtn,__ ConI(-1));
1.1369 + __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1.1370 + Node* i2 = __ AddI(__ value(i), targetCountLess1);
1.1371 + // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1.1372 + Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1.1373 + __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1.1374 + __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1.1375 + Node* tpj = __ AddI(targetOffset, __ value(j));
1.1376 + Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1.1377 + Node* ipj = __ AddI(__ value(i), __ value(j));
1.1378 + Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1.1379 + __ if_then(targ, BoolTest::ne, src2); {
1.1380 + __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1.1381 + __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1.1382 + __ increment(i, __ AddI(__ value(j), one));
1.1383 + __ goto_(outer_loop);
1.1384 + } __ end_if(); __ dead(j);
1.1385 + }__ end_if(); __ dead(j);
1.1386 + __ increment(i, md2);
1.1387 + __ goto_(outer_loop);
1.1388 + }__ end_if();
1.1389 + __ increment(j, one);
1.1390 + }__ end_loop(); __ dead(j);
1.1391 + __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1.1392 + __ goto_(return_);
1.1393 + }__ end_if();
1.1394 + __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1.1395 + __ increment(i, targetCountLess1);
1.1396 + }__ end_if();
1.1397 + __ increment(i, one);
1.1398 + __ bind(outer_loop);
1.1399 + }__ end_loop(); __ dead(i);
1.1400 + __ bind(return_);
1.1401 +
1.1402 + // Final sync IdealKit and GraphKit.
1.1403 + final_sync(kit);
1.1404 + Node* result = __ value(rtn);
1.1405 +#undef __
1.1406 + C->set_has_loops(true);
1.1407 + return result;
1.1408 +}
1.1409 +
1.1410 +//------------------------------inline_string_indexOf------------------------
1.1411 +bool LibraryCallKit::inline_string_indexOf() {
1.1412 + Node* receiver = argument(0);
1.1413 + Node* arg = argument(1);
1.1414 +
1.1415 + Node* result;
1.1416 + // Disable the use of pcmpestri until it can be guaranteed that
1.1417 + // the load doesn't cross into the uncommited space.
1.1418 + if (Matcher::has_match_rule(Op_StrIndexOf) &&
1.1419 + UseSSE42Intrinsics) {
1.1420 + // Generate SSE4.2 version of indexOf
1.1421 + // We currently only have match rules that use SSE4.2
1.1422 +
1.1423 + receiver = null_check(receiver);
1.1424 + arg = null_check(arg);
1.1425 + if (stopped()) {
1.1426 + return true;
1.1427 + }
1.1428 +
1.1429 + ciInstanceKlass* str_klass = env()->String_klass();
1.1430 + const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1.1431 +
1.1432 + // Make the merge point
1.1433 + RegionNode* result_rgn = new (C) RegionNode(4);
1.1434 + Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1.1435 + Node* no_ctrl = NULL;
1.1436 +
1.1437 + // Get start addr of source string
1.1438 + Node* source = load_String_value(no_ctrl, receiver);
1.1439 + Node* source_offset = load_String_offset(no_ctrl, receiver);
1.1440 + Node* source_start = array_element_address(source, source_offset, T_CHAR);
1.1441 +
1.1442 + // Get length of source string
1.1443 + Node* source_cnt = load_String_length(no_ctrl, receiver);
1.1444 +
1.1445 + // Get start addr of substring
1.1446 + Node* substr = load_String_value(no_ctrl, arg);
1.1447 + Node* substr_offset = load_String_offset(no_ctrl, arg);
1.1448 + Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1.1449 +
1.1450 + // Get length of source string
1.1451 + Node* substr_cnt = load_String_length(no_ctrl, arg);
1.1452 +
1.1453 + // Check for substr count > string count
1.1454 + Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1.1455 + Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1.1456 + Node* if_gt = generate_slow_guard(bol, NULL);
1.1457 + if (if_gt != NULL) {
1.1458 + result_phi->init_req(2, intcon(-1));
1.1459 + result_rgn->init_req(2, if_gt);
1.1460 + }
1.1461 +
1.1462 + if (!stopped()) {
1.1463 + // Check for substr count == 0
1.1464 + cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1.1465 + bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1.1466 + Node* if_zero = generate_slow_guard(bol, NULL);
1.1467 + if (if_zero != NULL) {
1.1468 + result_phi->init_req(3, intcon(0));
1.1469 + result_rgn->init_req(3, if_zero);
1.1470 + }
1.1471 + }
1.1472 +
1.1473 + if (!stopped()) {
1.1474 + result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1.1475 + result_phi->init_req(1, result);
1.1476 + result_rgn->init_req(1, control());
1.1477 + }
1.1478 + set_control(_gvn.transform(result_rgn));
1.1479 + record_for_igvn(result_rgn);
1.1480 + result = _gvn.transform(result_phi);
1.1481 +
1.1482 + } else { // Use LibraryCallKit::string_indexOf
1.1483 + // don't intrinsify if argument isn't a constant string.
1.1484 + if (!arg->is_Con()) {
1.1485 + return false;
1.1486 + }
1.1487 + const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1.1488 + if (str_type == NULL) {
1.1489 + return false;
1.1490 + }
1.1491 + ciInstanceKlass* klass = env()->String_klass();
1.1492 + ciObject* str_const = str_type->const_oop();
1.1493 + if (str_const == NULL || str_const->klass() != klass) {
1.1494 + return false;
1.1495 + }
1.1496 + ciInstance* str = str_const->as_instance();
1.1497 + assert(str != NULL, "must be instance");
1.1498 +
1.1499 + ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1.1500 + ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1.1501 +
1.1502 + int o;
1.1503 + int c;
1.1504 + if (java_lang_String::has_offset_field()) {
1.1505 + o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1.1506 + c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1.1507 + } else {
1.1508 + o = 0;
1.1509 + c = pat->length();
1.1510 + }
1.1511 +
1.1512 + // constant strings have no offset and count == length which
1.1513 + // simplifies the resulting code somewhat so lets optimize for that.
1.1514 + if (o != 0 || c != pat->length()) {
1.1515 + return false;
1.1516 + }
1.1517 +
1.1518 + receiver = null_check(receiver, T_OBJECT);
1.1519 + // NOTE: No null check on the argument is needed since it's a constant String oop.
1.1520 + if (stopped()) {
1.1521 + return true;
1.1522 + }
1.1523 +
1.1524 + // The null string as a pattern always returns 0 (match at beginning of string)
1.1525 + if (c == 0) {
1.1526 + set_result(intcon(0));
1.1527 + return true;
1.1528 + }
1.1529 +
1.1530 + // Generate default indexOf
1.1531 + jchar lastChar = pat->char_at(o + (c - 1));
1.1532 + int cache = 0;
1.1533 + int i;
1.1534 + for (i = 0; i < c - 1; i++) {
1.1535 + assert(i < pat->length(), "out of range");
1.1536 + cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1.1537 + }
1.1538 +
1.1539 + int md2 = c;
1.1540 + for (i = 0; i < c - 1; i++) {
1.1541 + assert(i < pat->length(), "out of range");
1.1542 + if (pat->char_at(o + i) == lastChar) {
1.1543 + md2 = (c - 1) - i;
1.1544 + }
1.1545 + }
1.1546 +
1.1547 + result = string_indexOf(receiver, pat, o, cache, md2);
1.1548 + }
1.1549 + set_result(result);
1.1550 + return true;
1.1551 +}
1.1552 +
1.1553 +//--------------------------round_double_node--------------------------------
1.1554 +// Round a double node if necessary.
1.1555 +Node* LibraryCallKit::round_double_node(Node* n) {
1.1556 + if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1.1557 + n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1.1558 + return n;
1.1559 +}
1.1560 +
1.1561 +//------------------------------inline_math-----------------------------------
1.1562 +// public static double Math.abs(double)
1.1563 +// public static double Math.sqrt(double)
1.1564 +// public static double Math.log(double)
1.1565 +// public static double Math.log10(double)
1.1566 +bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1.1567 + Node* arg = round_double_node(argument(0));
1.1568 + Node* n;
1.1569 + switch (id) {
1.1570 + case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1.1571 + case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1.1572 + case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1.1573 + case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1.1574 + default: fatal_unexpected_iid(id); break;
1.1575 + }
1.1576 + set_result(_gvn.transform(n));
1.1577 + return true;
1.1578 +}
1.1579 +
1.1580 +//------------------------------inline_trig----------------------------------
1.1581 +// Inline sin/cos/tan instructions, if possible. If rounding is required, do
1.1582 +// argument reduction which will turn into a fast/slow diamond.
1.1583 +bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1.1584 + Node* arg = round_double_node(argument(0));
1.1585 + Node* n = NULL;
1.1586 +
1.1587 + switch (id) {
1.1588 + case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1.1589 + case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1.1590 + case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1.1591 + default: fatal_unexpected_iid(id); break;
1.1592 + }
1.1593 + n = _gvn.transform(n);
1.1594 +
1.1595 + // Rounding required? Check for argument reduction!
1.1596 + if (Matcher::strict_fp_requires_explicit_rounding) {
1.1597 + static const double pi_4 = 0.7853981633974483;
1.1598 + static const double neg_pi_4 = -0.7853981633974483;
1.1599 + // pi/2 in 80-bit extended precision
1.1600 + // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1.1601 + // -pi/2 in 80-bit extended precision
1.1602 + // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1.1603 + // Cutoff value for using this argument reduction technique
1.1604 + //static const double pi_2_minus_epsilon = 1.564660403643354;
1.1605 + //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1.1606 +
1.1607 + // Pseudocode for sin:
1.1608 + // if (x <= Math.PI / 4.0) {
1.1609 + // if (x >= -Math.PI / 4.0) return fsin(x);
1.1610 + // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1.1611 + // } else {
1.1612 + // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1.1613 + // }
1.1614 + // return StrictMath.sin(x);
1.1615 +
1.1616 + // Pseudocode for cos:
1.1617 + // if (x <= Math.PI / 4.0) {
1.1618 + // if (x >= -Math.PI / 4.0) return fcos(x);
1.1619 + // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1.1620 + // } else {
1.1621 + // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1.1622 + // }
1.1623 + // return StrictMath.cos(x);
1.1624 +
1.1625 + // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1.1626 + // requires a special machine instruction to load it. Instead we'll try
1.1627 + // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1.1628 + // probably do the math inside the SIN encoding.
1.1629 +
1.1630 + // Make the merge point
1.1631 + RegionNode* r = new (C) RegionNode(3);
1.1632 + Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1.1633 +
1.1634 + // Flatten arg so we need only 1 test
1.1635 + Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1.1636 + // Node for PI/4 constant
1.1637 + Node *pi4 = makecon(TypeD::make(pi_4));
1.1638 + // Check PI/4 : abs(arg)
1.1639 + Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1.1640 + // Check: If PI/4 < abs(arg) then go slow
1.1641 + Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1.1642 + // Branch either way
1.1643 + IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1.1644 + set_control(opt_iff(r,iff));
1.1645 +
1.1646 + // Set fast path result
1.1647 + phi->init_req(2, n);
1.1648 +
1.1649 + // Slow path - non-blocking leaf call
1.1650 + Node* call = NULL;
1.1651 + switch (id) {
1.1652 + case vmIntrinsics::_dsin:
1.1653 + call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1.1654 + CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1.1655 + "Sin", NULL, arg, top());
1.1656 + break;
1.1657 + case vmIntrinsics::_dcos:
1.1658 + call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1.1659 + CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1.1660 + "Cos", NULL, arg, top());
1.1661 + break;
1.1662 + case vmIntrinsics::_dtan:
1.1663 + call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1.1664 + CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1.1665 + "Tan", NULL, arg, top());
1.1666 + break;
1.1667 + }
1.1668 + assert(control()->in(0) == call, "");
1.1669 + Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1.1670 + r->init_req(1, control());
1.1671 + phi->init_req(1, slow_result);
1.1672 +
1.1673 + // Post-merge
1.1674 + set_control(_gvn.transform(r));
1.1675 + record_for_igvn(r);
1.1676 + n = _gvn.transform(phi);
1.1677 +
1.1678 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1679 + }
1.1680 + set_result(n);
1.1681 + return true;
1.1682 +}
1.1683 +
1.1684 +Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1.1685 + //-------------------
1.1686 + //result=(result.isNaN())? funcAddr():result;
1.1687 + // Check: If isNaN() by checking result!=result? then either trap
1.1688 + // or go to runtime
1.1689 + Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1.1690 + // Build the boolean node
1.1691 + Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1.1692 +
1.1693 + if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1.1694 + { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1.1695 + // The pow or exp intrinsic returned a NaN, which requires a call
1.1696 + // to the runtime. Recompile with the runtime call.
1.1697 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.1698 + Deoptimization::Action_make_not_entrant);
1.1699 + }
1.1700 + return result;
1.1701 + } else {
1.1702 + // If this inlining ever returned NaN in the past, we compile a call
1.1703 + // to the runtime to properly handle corner cases
1.1704 +
1.1705 + IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1.1706 + Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1.1707 + Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1.1708 +
1.1709 + if (!if_slow->is_top()) {
1.1710 + RegionNode* result_region = new (C) RegionNode(3);
1.1711 + PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1.1712 +
1.1713 + result_region->init_req(1, if_fast);
1.1714 + result_val->init_req(1, result);
1.1715 +
1.1716 + set_control(if_slow);
1.1717 +
1.1718 + const TypePtr* no_memory_effects = NULL;
1.1719 + Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1.1720 + no_memory_effects,
1.1721 + x, top(), y, y ? top() : NULL);
1.1722 + Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1.1723 +#ifdef ASSERT
1.1724 + Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1.1725 + assert(value_top == top(), "second value must be top");
1.1726 +#endif
1.1727 +
1.1728 + result_region->init_req(2, control());
1.1729 + result_val->init_req(2, value);
1.1730 + set_control(_gvn.transform(result_region));
1.1731 + return _gvn.transform(result_val);
1.1732 + } else {
1.1733 + return result;
1.1734 + }
1.1735 + }
1.1736 +}
1.1737 +
1.1738 +//------------------------------inline_exp-------------------------------------
1.1739 +// Inline exp instructions, if possible. The Intel hardware only misses
1.1740 +// really odd corner cases (+/- Infinity). Just uncommon-trap them.
1.1741 +bool LibraryCallKit::inline_exp() {
1.1742 + Node* arg = round_double_node(argument(0));
1.1743 + Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1.1744 +
1.1745 + n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1.1746 + set_result(n);
1.1747 +
1.1748 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1749 + return true;
1.1750 +}
1.1751 +
1.1752 +//------------------------------inline_pow-------------------------------------
1.1753 +// Inline power instructions, if possible.
1.1754 +bool LibraryCallKit::inline_pow() {
1.1755 + // Pseudocode for pow
1.1756 + // if (y == 2) {
1.1757 + // return x * x;
1.1758 + // } else {
1.1759 + // if (x <= 0.0) {
1.1760 + // long longy = (long)y;
1.1761 + // if ((double)longy == y) { // if y is long
1.1762 + // if (y + 1 == y) longy = 0; // huge number: even
1.1763 + // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1.1764 + // } else {
1.1765 + // result = NaN;
1.1766 + // }
1.1767 + // } else {
1.1768 + // result = DPow(x,y);
1.1769 + // }
1.1770 + // if (result != result)? {
1.1771 + // result = uncommon_trap() or runtime_call();
1.1772 + // }
1.1773 + // return result;
1.1774 + // }
1.1775 +
1.1776 + Node* x = round_double_node(argument(0));
1.1777 + Node* y = round_double_node(argument(2));
1.1778 +
1.1779 + Node* result = NULL;
1.1780 +
1.1781 + Node* const_two_node = makecon(TypeD::make(2.0));
1.1782 + Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1.1783 + Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1.1784 + IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1.1785 + Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1.1786 + Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1.1787 +
1.1788 + RegionNode* region_node = new (C) RegionNode(3);
1.1789 + region_node->init_req(1, if_true);
1.1790 +
1.1791 + Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1.1792 + // special case for x^y where y == 2, we can convert it to x * x
1.1793 + phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1.1794 +
1.1795 + // set control to if_false since we will now process the false branch
1.1796 + set_control(if_false);
1.1797 +
1.1798 + if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1.1799 + // Short form: skip the fancy tests and just check for NaN result.
1.1800 + result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1.1801 + } else {
1.1802 + // If this inlining ever returned NaN in the past, include all
1.1803 + // checks + call to the runtime.
1.1804 +
1.1805 + // Set the merge point for If node with condition of (x <= 0.0)
1.1806 + // There are four possible paths to region node and phi node
1.1807 + RegionNode *r = new (C) RegionNode(4);
1.1808 + Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1.1809 +
1.1810 + // Build the first if node: if (x <= 0.0)
1.1811 + // Node for 0 constant
1.1812 + Node *zeronode = makecon(TypeD::ZERO);
1.1813 + // Check x:0
1.1814 + Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1.1815 + // Check: If (x<=0) then go complex path
1.1816 + Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1.1817 + // Branch either way
1.1818 + IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1.1819 + // Fast path taken; set region slot 3
1.1820 + Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1.1821 + r->init_req(3,fast_taken); // Capture fast-control
1.1822 +
1.1823 + // Fast path not-taken, i.e. slow path
1.1824 + Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1.1825 +
1.1826 + // Set fast path result
1.1827 + Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1.1828 + phi->init_req(3, fast_result);
1.1829 +
1.1830 + // Complex path
1.1831 + // Build the second if node (if y is long)
1.1832 + // Node for (long)y
1.1833 + Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1.1834 + // Node for (double)((long) y)
1.1835 + Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1.1836 + // Check (double)((long) y) : y
1.1837 + Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1.1838 + // Check if (y isn't long) then go to slow path
1.1839 +
1.1840 + Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1.1841 + // Branch either way
1.1842 + IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1.1843 + Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1.1844 +
1.1845 + Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1.1846 +
1.1847 + // Calculate DPow(abs(x), y)*(1 & (long)y)
1.1848 + // Node for constant 1
1.1849 + Node *conone = longcon(1);
1.1850 + // 1& (long)y
1.1851 + Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1.1852 +
1.1853 + // A huge number is always even. Detect a huge number by checking
1.1854 + // if y + 1 == y and set integer to be tested for parity to 0.
1.1855 + // Required for corner case:
1.1856 + // (long)9.223372036854776E18 = max_jlong
1.1857 + // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1.1858 + // max_jlong is odd but 9.223372036854776E18 is even
1.1859 + Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1.1860 + Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1.1861 + Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1.1862 + Node* correctedsign = NULL;
1.1863 + if (ConditionalMoveLimit != 0) {
1.1864 + correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1.1865 + } else {
1.1866 + IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1.1867 + RegionNode *r = new (C) RegionNode(3);
1.1868 + Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1.1869 + r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1.1870 + r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1.1871 + phi->init_req(1, signnode);
1.1872 + phi->init_req(2, longcon(0));
1.1873 + correctedsign = _gvn.transform(phi);
1.1874 + ylong_path = _gvn.transform(r);
1.1875 + record_for_igvn(r);
1.1876 + }
1.1877 +
1.1878 + // zero node
1.1879 + Node *conzero = longcon(0);
1.1880 + // Check (1&(long)y)==0?
1.1881 + Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1.1882 + // Check if (1&(long)y)!=0?, if so the result is negative
1.1883 + Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1.1884 + // abs(x)
1.1885 + Node *absx=_gvn.transform(new (C) AbsDNode(x));
1.1886 + // abs(x)^y
1.1887 + Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1.1888 + // -abs(x)^y
1.1889 + Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1.1890 + // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1.1891 + Node *signresult = NULL;
1.1892 + if (ConditionalMoveLimit != 0) {
1.1893 + signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1.1894 + } else {
1.1895 + IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1.1896 + RegionNode *r = new (C) RegionNode(3);
1.1897 + Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1.1898 + r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1.1899 + r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1.1900 + phi->init_req(1, absxpowy);
1.1901 + phi->init_req(2, negabsxpowy);
1.1902 + signresult = _gvn.transform(phi);
1.1903 + ylong_path = _gvn.transform(r);
1.1904 + record_for_igvn(r);
1.1905 + }
1.1906 + // Set complex path fast result
1.1907 + r->init_req(2, ylong_path);
1.1908 + phi->init_req(2, signresult);
1.1909 +
1.1910 + static const jlong nan_bits = CONST64(0x7ff8000000000000);
1.1911 + Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1.1912 + r->init_req(1,slow_path);
1.1913 + phi->init_req(1,slow_result);
1.1914 +
1.1915 + // Post merge
1.1916 + set_control(_gvn.transform(r));
1.1917 + record_for_igvn(r);
1.1918 + result = _gvn.transform(phi);
1.1919 + }
1.1920 +
1.1921 + result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1.1922 +
1.1923 + // control from finish_pow_exp is now input to the region node
1.1924 + region_node->set_req(2, control());
1.1925 + // the result from finish_pow_exp is now input to the phi node
1.1926 + phi_node->init_req(2, result);
1.1927 + set_control(_gvn.transform(region_node));
1.1928 + record_for_igvn(region_node);
1.1929 + set_result(_gvn.transform(phi_node));
1.1930 +
1.1931 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1932 + return true;
1.1933 +}
1.1934 +
1.1935 +//------------------------------runtime_math-----------------------------
1.1936 +bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1.1937 + assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1.1938 + "must be (DD)D or (D)D type");
1.1939 +
1.1940 + // Inputs
1.1941 + Node* a = round_double_node(argument(0));
1.1942 + Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1.1943 +
1.1944 + const TypePtr* no_memory_effects = NULL;
1.1945 + Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1.1946 + no_memory_effects,
1.1947 + a, top(), b, b ? top() : NULL);
1.1948 + Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1.1949 +#ifdef ASSERT
1.1950 + Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1.1951 + assert(value_top == top(), "second value must be top");
1.1952 +#endif
1.1953 +
1.1954 + set_result(value);
1.1955 + return true;
1.1956 +}
1.1957 +
1.1958 +//------------------------------inline_math_native-----------------------------
1.1959 +bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1.1960 +#define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1.1961 + switch (id) {
1.1962 + // These intrinsics are not properly supported on all hardware
1.1963 + case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1.1964 + runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1.1965 + case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1.1966 + runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1.1967 + case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1.1968 + runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1.1969 +
1.1970 + case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1.1971 + runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1.1972 + case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1.1973 + runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1.1974 +
1.1975 + // These intrinsics are supported on all hardware
1.1976 + case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1.1977 + case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1.1978 +
1.1979 + case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1.1980 + runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1.1981 + case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1.1982 + runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1.1983 +#undef FN_PTR
1.1984 +
1.1985 + // These intrinsics are not yet correctly implemented
1.1986 + case vmIntrinsics::_datan2:
1.1987 + return false;
1.1988 +
1.1989 + default:
1.1990 + fatal_unexpected_iid(id);
1.1991 + return false;
1.1992 + }
1.1993 +}
1.1994 +
1.1995 +static bool is_simple_name(Node* n) {
1.1996 + return (n->req() == 1 // constant
1.1997 + || (n->is_Type() && n->as_Type()->type()->singleton())
1.1998 + || n->is_Proj() // parameter or return value
1.1999 + || n->is_Phi() // local of some sort
1.2000 + );
1.2001 +}
1.2002 +
1.2003 +//----------------------------inline_min_max-----------------------------------
1.2004 +bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1.2005 + set_result(generate_min_max(id, argument(0), argument(1)));
1.2006 + return true;
1.2007 +}
1.2008 +
1.2009 +void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1.2010 + Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
1.2011 + IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1.2012 + Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1.2013 + Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1.2014 +
1.2015 + {
1.2016 + PreserveJVMState pjvms(this);
1.2017 + PreserveReexecuteState preexecs(this);
1.2018 + jvms()->set_should_reexecute(true);
1.2019 +
1.2020 + set_control(slow_path);
1.2021 + set_i_o(i_o());
1.2022 +
1.2023 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.2024 + Deoptimization::Action_none);
1.2025 + }
1.2026 +
1.2027 + set_control(fast_path);
1.2028 + set_result(math);
1.2029 +}
1.2030 +
1.2031 +template <typename OverflowOp>
1.2032 +bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1.2033 + typedef typename OverflowOp::MathOp MathOp;
1.2034 +
1.2035 + MathOp* mathOp = new(C) MathOp(arg1, arg2);
1.2036 + Node* operation = _gvn.transform( mathOp );
1.2037 + Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
1.2038 + inline_math_mathExact(operation, ofcheck);
1.2039 + return true;
1.2040 +}
1.2041 +
1.2042 +bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1.2043 + return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1.2044 +}
1.2045 +
1.2046 +bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1.2047 + return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1.2048 +}
1.2049 +
1.2050 +bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1.2051 + return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1.2052 +}
1.2053 +
1.2054 +bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1.2055 + return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1.2056 +}
1.2057 +
1.2058 +bool LibraryCallKit::inline_math_negateExactI() {
1.2059 + return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1.2060 +}
1.2061 +
1.2062 +bool LibraryCallKit::inline_math_negateExactL() {
1.2063 + return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1.2064 +}
1.2065 +
1.2066 +bool LibraryCallKit::inline_math_multiplyExactI() {
1.2067 + return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1.2068 +}
1.2069 +
1.2070 +bool LibraryCallKit::inline_math_multiplyExactL() {
1.2071 + return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1.2072 +}
1.2073 +
1.2074 +Node*
1.2075 +LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1.2076 + // These are the candidate return value:
1.2077 + Node* xvalue = x0;
1.2078 + Node* yvalue = y0;
1.2079 +
1.2080 + if (xvalue == yvalue) {
1.2081 + return xvalue;
1.2082 + }
1.2083 +
1.2084 + bool want_max = (id == vmIntrinsics::_max);
1.2085 +
1.2086 + const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1.2087 + const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1.2088 + if (txvalue == NULL || tyvalue == NULL) return top();
1.2089 + // This is not really necessary, but it is consistent with a
1.2090 + // hypothetical MaxINode::Value method:
1.2091 + int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1.2092 +
1.2093 + // %%% This folding logic should (ideally) be in a different place.
1.2094 + // Some should be inside IfNode, and there to be a more reliable
1.2095 + // transformation of ?: style patterns into cmoves. We also want
1.2096 + // more powerful optimizations around cmove and min/max.
1.2097 +
1.2098 + // Try to find a dominating comparison of these guys.
1.2099 + // It can simplify the index computation for Arrays.copyOf
1.2100 + // and similar uses of System.arraycopy.
1.2101 + // First, compute the normalized version of CmpI(x, y).
1.2102 + int cmp_op = Op_CmpI;
1.2103 + Node* xkey = xvalue;
1.2104 + Node* ykey = yvalue;
1.2105 + Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
1.2106 + if (ideal_cmpxy->is_Cmp()) {
1.2107 + // E.g., if we have CmpI(length - offset, count),
1.2108 + // it might idealize to CmpI(length, count + offset)
1.2109 + cmp_op = ideal_cmpxy->Opcode();
1.2110 + xkey = ideal_cmpxy->in(1);
1.2111 + ykey = ideal_cmpxy->in(2);
1.2112 + }
1.2113 +
1.2114 + // Start by locating any relevant comparisons.
1.2115 + Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1.2116 + Node* cmpxy = NULL;
1.2117 + Node* cmpyx = NULL;
1.2118 + for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1.2119 + Node* cmp = start_from->fast_out(k);
1.2120 + if (cmp->outcnt() > 0 && // must have prior uses
1.2121 + cmp->in(0) == NULL && // must be context-independent
1.2122 + cmp->Opcode() == cmp_op) { // right kind of compare
1.2123 + if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1.2124 + if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1.2125 + }
1.2126 + }
1.2127 +
1.2128 + const int NCMPS = 2;
1.2129 + Node* cmps[NCMPS] = { cmpxy, cmpyx };
1.2130 + int cmpn;
1.2131 + for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1.2132 + if (cmps[cmpn] != NULL) break; // find a result
1.2133 + }
1.2134 + if (cmpn < NCMPS) {
1.2135 + // Look for a dominating test that tells us the min and max.
1.2136 + int depth = 0; // Limit search depth for speed
1.2137 + Node* dom = control();
1.2138 + for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1.2139 + if (++depth >= 100) break;
1.2140 + Node* ifproj = dom;
1.2141 + if (!ifproj->is_Proj()) continue;
1.2142 + Node* iff = ifproj->in(0);
1.2143 + if (!iff->is_If()) continue;
1.2144 + Node* bol = iff->in(1);
1.2145 + if (!bol->is_Bool()) continue;
1.2146 + Node* cmp = bol->in(1);
1.2147 + if (cmp == NULL) continue;
1.2148 + for (cmpn = 0; cmpn < NCMPS; cmpn++)
1.2149 + if (cmps[cmpn] == cmp) break;
1.2150 + if (cmpn == NCMPS) continue;
1.2151 + BoolTest::mask btest = bol->as_Bool()->_test._test;
1.2152 + if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1.2153 + if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1.2154 + // At this point, we know that 'x btest y' is true.
1.2155 + switch (btest) {
1.2156 + case BoolTest::eq:
1.2157 + // They are proven equal, so we can collapse the min/max.
1.2158 + // Either value is the answer. Choose the simpler.
1.2159 + if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1.2160 + return yvalue;
1.2161 + return xvalue;
1.2162 + case BoolTest::lt: // x < y
1.2163 + case BoolTest::le: // x <= y
1.2164 + return (want_max ? yvalue : xvalue);
1.2165 + case BoolTest::gt: // x > y
1.2166 + case BoolTest::ge: // x >= y
1.2167 + return (want_max ? xvalue : yvalue);
1.2168 + }
1.2169 + }
1.2170 + }
1.2171 +
1.2172 + // We failed to find a dominating test.
1.2173 + // Let's pick a test that might GVN with prior tests.
1.2174 + Node* best_bol = NULL;
1.2175 + BoolTest::mask best_btest = BoolTest::illegal;
1.2176 + for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1.2177 + Node* cmp = cmps[cmpn];
1.2178 + if (cmp == NULL) continue;
1.2179 + for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1.2180 + Node* bol = cmp->fast_out(j);
1.2181 + if (!bol->is_Bool()) continue;
1.2182 + BoolTest::mask btest = bol->as_Bool()->_test._test;
1.2183 + if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1.2184 + if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1.2185 + if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1.2186 + best_bol = bol->as_Bool();
1.2187 + best_btest = btest;
1.2188 + }
1.2189 + }
1.2190 + }
1.2191 +
1.2192 + Node* answer_if_true = NULL;
1.2193 + Node* answer_if_false = NULL;
1.2194 + switch (best_btest) {
1.2195 + default:
1.2196 + if (cmpxy == NULL)
1.2197 + cmpxy = ideal_cmpxy;
1.2198 + best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
1.2199 + // and fall through:
1.2200 + case BoolTest::lt: // x < y
1.2201 + case BoolTest::le: // x <= y
1.2202 + answer_if_true = (want_max ? yvalue : xvalue);
1.2203 + answer_if_false = (want_max ? xvalue : yvalue);
1.2204 + break;
1.2205 + case BoolTest::gt: // x > y
1.2206 + case BoolTest::ge: // x >= y
1.2207 + answer_if_true = (want_max ? xvalue : yvalue);
1.2208 + answer_if_false = (want_max ? yvalue : xvalue);
1.2209 + break;
1.2210 + }
1.2211 +
1.2212 + jint hi, lo;
1.2213 + if (want_max) {
1.2214 + // We can sharpen the minimum.
1.2215 + hi = MAX2(txvalue->_hi, tyvalue->_hi);
1.2216 + lo = MAX2(txvalue->_lo, tyvalue->_lo);
1.2217 + } else {
1.2218 + // We can sharpen the maximum.
1.2219 + hi = MIN2(txvalue->_hi, tyvalue->_hi);
1.2220 + lo = MIN2(txvalue->_lo, tyvalue->_lo);
1.2221 + }
1.2222 +
1.2223 + // Use a flow-free graph structure, to avoid creating excess control edges
1.2224 + // which could hinder other optimizations.
1.2225 + // Since Math.min/max is often used with arraycopy, we want
1.2226 + // tightly_coupled_allocation to be able to see beyond min/max expressions.
1.2227 + Node* cmov = CMoveNode::make(C, NULL, best_bol,
1.2228 + answer_if_false, answer_if_true,
1.2229 + TypeInt::make(lo, hi, widen));
1.2230 +
1.2231 + return _gvn.transform(cmov);
1.2232 +
1.2233 + /*
1.2234 + // This is not as desirable as it may seem, since Min and Max
1.2235 + // nodes do not have a full set of optimizations.
1.2236 + // And they would interfere, anyway, with 'if' optimizations
1.2237 + // and with CMoveI canonical forms.
1.2238 + switch (id) {
1.2239 + case vmIntrinsics::_min:
1.2240 + result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1.2241 + case vmIntrinsics::_max:
1.2242 + result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1.2243 + default:
1.2244 + ShouldNotReachHere();
1.2245 + }
1.2246 + */
1.2247 +}
1.2248 +
1.2249 +inline int
1.2250 +LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1.2251 + const TypePtr* base_type = TypePtr::NULL_PTR;
1.2252 + if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1.2253 + if (base_type == NULL) {
1.2254 + // Unknown type.
1.2255 + return Type::AnyPtr;
1.2256 + } else if (base_type == TypePtr::NULL_PTR) {
1.2257 + // Since this is a NULL+long form, we have to switch to a rawptr.
1.2258 + base = _gvn.transform(new (C) CastX2PNode(offset));
1.2259 + offset = MakeConX(0);
1.2260 + return Type::RawPtr;
1.2261 + } else if (base_type->base() == Type::RawPtr) {
1.2262 + return Type::RawPtr;
1.2263 + } else if (base_type->isa_oopptr()) {
1.2264 + // Base is never null => always a heap address.
1.2265 + if (base_type->ptr() == TypePtr::NotNull) {
1.2266 + return Type::OopPtr;
1.2267 + }
1.2268 + // Offset is small => always a heap address.
1.2269 + const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1.2270 + if (offset_type != NULL &&
1.2271 + base_type->offset() == 0 && // (should always be?)
1.2272 + offset_type->_lo >= 0 &&
1.2273 + !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1.2274 + return Type::OopPtr;
1.2275 + }
1.2276 + // Otherwise, it might either be oop+off or NULL+addr.
1.2277 + return Type::AnyPtr;
1.2278 + } else {
1.2279 + // No information:
1.2280 + return Type::AnyPtr;
1.2281 + }
1.2282 +}
1.2283 +
1.2284 +inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1.2285 + int kind = classify_unsafe_addr(base, offset);
1.2286 + if (kind == Type::RawPtr) {
1.2287 + return basic_plus_adr(top(), base, offset);
1.2288 + } else {
1.2289 + return basic_plus_adr(base, offset);
1.2290 + }
1.2291 +}
1.2292 +
1.2293 +//--------------------------inline_number_methods-----------------------------
1.2294 +// inline int Integer.numberOfLeadingZeros(int)
1.2295 +// inline int Long.numberOfLeadingZeros(long)
1.2296 +//
1.2297 +// inline int Integer.numberOfTrailingZeros(int)
1.2298 +// inline int Long.numberOfTrailingZeros(long)
1.2299 +//
1.2300 +// inline int Integer.bitCount(int)
1.2301 +// inline int Long.bitCount(long)
1.2302 +//
1.2303 +// inline char Character.reverseBytes(char)
1.2304 +// inline short Short.reverseBytes(short)
1.2305 +// inline int Integer.reverseBytes(int)
1.2306 +// inline long Long.reverseBytes(long)
1.2307 +bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
1.2308 + Node* arg = argument(0);
1.2309 + Node* n;
1.2310 + switch (id) {
1.2311 + case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
1.2312 + case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
1.2313 + case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
1.2314 + case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
1.2315 + case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
1.2316 + case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
1.2317 + case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
1.2318 + case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
1.2319 + case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
1.2320 + case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
1.2321 + default: fatal_unexpected_iid(id); break;
1.2322 + }
1.2323 + set_result(_gvn.transform(n));
1.2324 + return true;
1.2325 +}
1.2326 +
1.2327 +//----------------------------inline_unsafe_access----------------------------
1.2328 +
1.2329 +const static BasicType T_ADDRESS_HOLDER = T_LONG;
1.2330 +
1.2331 +// Helper that guards and inserts a pre-barrier.
1.2332 +void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
1.2333 + Node* pre_val, bool need_mem_bar) {
1.2334 + // We could be accessing the referent field of a reference object. If so, when G1
1.2335 + // is enabled, we need to log the value in the referent field in an SATB buffer.
1.2336 + // This routine performs some compile time filters and generates suitable
1.2337 + // runtime filters that guard the pre-barrier code.
1.2338 + // Also add memory barrier for non volatile load from the referent field
1.2339 + // to prevent commoning of loads across safepoint.
1.2340 + if (!UseG1GC && !need_mem_bar)
1.2341 + return;
1.2342 +
1.2343 + // Some compile time checks.
1.2344 +
1.2345 + // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
1.2346 + const TypeX* otype = offset->find_intptr_t_type();
1.2347 + if (otype != NULL && otype->is_con() &&
1.2348 + otype->get_con() != java_lang_ref_Reference::referent_offset) {
1.2349 + // Constant offset but not the reference_offset so just return
1.2350 + return;
1.2351 + }
1.2352 +
1.2353 + // We only need to generate the runtime guards for instances.
1.2354 + const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
1.2355 + if (btype != NULL) {
1.2356 + if (btype->isa_aryptr()) {
1.2357 + // Array type so nothing to do
1.2358 + return;
1.2359 + }
1.2360 +
1.2361 + const TypeInstPtr* itype = btype->isa_instptr();
1.2362 + if (itype != NULL) {
1.2363 + // Can the klass of base_oop be statically determined to be
1.2364 + // _not_ a sub-class of Reference and _not_ Object?
1.2365 + ciKlass* klass = itype->klass();
1.2366 + if ( klass->is_loaded() &&
1.2367 + !klass->is_subtype_of(env()->Reference_klass()) &&
1.2368 + !env()->Object_klass()->is_subtype_of(klass)) {
1.2369 + return;
1.2370 + }
1.2371 + }
1.2372 + }
1.2373 +
1.2374 + // The compile time filters did not reject base_oop/offset so
1.2375 + // we need to generate the following runtime filters
1.2376 + //
1.2377 + // if (offset == java_lang_ref_Reference::_reference_offset) {
1.2378 + // if (instance_of(base, java.lang.ref.Reference)) {
1.2379 + // pre_barrier(_, pre_val, ...);
1.2380 + // }
1.2381 + // }
1.2382 +
1.2383 + float likely = PROB_LIKELY( 0.999);
1.2384 + float unlikely = PROB_UNLIKELY(0.999);
1.2385 +
1.2386 + IdealKit ideal(this);
1.2387 +#define __ ideal.
1.2388 +
1.2389 + Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
1.2390 +
1.2391 + __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
1.2392 + // Update graphKit memory and control from IdealKit.
1.2393 + sync_kit(ideal);
1.2394 +
1.2395 + Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
1.2396 + Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
1.2397 +
1.2398 + // Update IdealKit memory and control from graphKit.
1.2399 + __ sync_kit(this);
1.2400 +
1.2401 + Node* one = __ ConI(1);
1.2402 + // is_instof == 0 if base_oop == NULL
1.2403 + __ if_then(is_instof, BoolTest::eq, one, unlikely); {
1.2404 +
1.2405 + // Update graphKit from IdeakKit.
1.2406 + sync_kit(ideal);
1.2407 +
1.2408 + // Use the pre-barrier to record the value in the referent field
1.2409 + pre_barrier(false /* do_load */,
1.2410 + __ ctrl(),
1.2411 + NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
1.2412 + pre_val /* pre_val */,
1.2413 + T_OBJECT);
1.2414 + if (need_mem_bar) {
1.2415 + // Add memory barrier to prevent commoning reads from this field
1.2416 + // across safepoint since GC can change its value.
1.2417 + insert_mem_bar(Op_MemBarCPUOrder);
1.2418 + }
1.2419 + // Update IdealKit from graphKit.
1.2420 + __ sync_kit(this);
1.2421 +
1.2422 + } __ end_if(); // _ref_type != ref_none
1.2423 + } __ end_if(); // offset == referent_offset
1.2424 +
1.2425 + // Final sync IdealKit and GraphKit.
1.2426 + final_sync(ideal);
1.2427 +#undef __
1.2428 +}
1.2429 +
1.2430 +
1.2431 +// Interpret Unsafe.fieldOffset cookies correctly:
1.2432 +extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
1.2433 +
1.2434 +const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
1.2435 + // Attempt to infer a sharper value type from the offset and base type.
1.2436 + ciKlass* sharpened_klass = NULL;
1.2437 +
1.2438 + // See if it is an instance field, with an object type.
1.2439 + if (alias_type->field() != NULL) {
1.2440 + assert(!is_native_ptr, "native pointer op cannot use a java address");
1.2441 + if (alias_type->field()->type()->is_klass()) {
1.2442 + sharpened_klass = alias_type->field()->type()->as_klass();
1.2443 + }
1.2444 + }
1.2445 +
1.2446 + // See if it is a narrow oop array.
1.2447 + if (adr_type->isa_aryptr()) {
1.2448 + if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
1.2449 + const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
1.2450 + if (elem_type != NULL) {
1.2451 + sharpened_klass = elem_type->klass();
1.2452 + }
1.2453 + }
1.2454 + }
1.2455 +
1.2456 + // The sharpened class might be unloaded if there is no class loader
1.2457 + // contraint in place.
1.2458 + if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
1.2459 + const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
1.2460 +
1.2461 +#ifndef PRODUCT
1.2462 + if (C->print_intrinsics() || C->print_inlining()) {
1.2463 + tty->print(" from base type: "); adr_type->dump();
1.2464 + tty->print(" sharpened value: "); tjp->dump();
1.2465 + }
1.2466 +#endif
1.2467 + // Sharpen the value type.
1.2468 + return tjp;
1.2469 + }
1.2470 + return NULL;
1.2471 +}
1.2472 +
1.2473 +bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
1.2474 + if (callee()->is_static()) return false; // caller must have the capability!
1.2475 +
1.2476 +#ifndef PRODUCT
1.2477 + {
1.2478 + ResourceMark rm;
1.2479 + // Check the signatures.
1.2480 + ciSignature* sig = callee()->signature();
1.2481 +#ifdef ASSERT
1.2482 + if (!is_store) {
1.2483 + // Object getObject(Object base, int/long offset), etc.
1.2484 + BasicType rtype = sig->return_type()->basic_type();
1.2485 + if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
1.2486 + rtype = T_ADDRESS; // it is really a C void*
1.2487 + assert(rtype == type, "getter must return the expected value");
1.2488 + if (!is_native_ptr) {
1.2489 + assert(sig->count() == 2, "oop getter has 2 arguments");
1.2490 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
1.2491 + assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
1.2492 + } else {
1.2493 + assert(sig->count() == 1, "native getter has 1 argument");
1.2494 + assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
1.2495 + }
1.2496 + } else {
1.2497 + // void putObject(Object base, int/long offset, Object x), etc.
1.2498 + assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
1.2499 + if (!is_native_ptr) {
1.2500 + assert(sig->count() == 3, "oop putter has 3 arguments");
1.2501 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
1.2502 + assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
1.2503 + } else {
1.2504 + assert(sig->count() == 2, "native putter has 2 arguments");
1.2505 + assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
1.2506 + }
1.2507 + BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
1.2508 + if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
1.2509 + vtype = T_ADDRESS; // it is really a C void*
1.2510 + assert(vtype == type, "putter must accept the expected value");
1.2511 + }
1.2512 +#endif // ASSERT
1.2513 + }
1.2514 +#endif //PRODUCT
1.2515 +
1.2516 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.2517 +
1.2518 + Node* receiver = argument(0); // type: oop
1.2519 +
1.2520 + // Build address expression. See the code in inline_unsafe_prefetch.
1.2521 + Node* adr;
1.2522 + Node* heap_base_oop = top();
1.2523 + Node* offset = top();
1.2524 + Node* val;
1.2525 +
1.2526 + if (!is_native_ptr) {
1.2527 + // The base is either a Java object or a value produced by Unsafe.staticFieldBase
1.2528 + Node* base = argument(1); // type: oop
1.2529 + // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
1.2530 + offset = argument(2); // type: long
1.2531 + // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1.2532 + // to be plain byte offsets, which are also the same as those accepted
1.2533 + // by oopDesc::field_base.
1.2534 + assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1.2535 + "fieldOffset must be byte-scaled");
1.2536 + // 32-bit machines ignore the high half!
1.2537 + offset = ConvL2X(offset);
1.2538 + adr = make_unsafe_address(base, offset);
1.2539 + heap_base_oop = base;
1.2540 + val = is_store ? argument(4) : NULL;
1.2541 + } else {
1.2542 + Node* ptr = argument(1); // type: long
1.2543 + ptr = ConvL2X(ptr); // adjust Java long to machine word
1.2544 + adr = make_unsafe_address(NULL, ptr);
1.2545 + val = is_store ? argument(3) : NULL;
1.2546 + }
1.2547 +
1.2548 + const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1.2549 +
1.2550 + // First guess at the value type.
1.2551 + const Type *value_type = Type::get_const_basic_type(type);
1.2552 +
1.2553 + // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
1.2554 + // there was not enough information to nail it down.
1.2555 + Compile::AliasType* alias_type = C->alias_type(adr_type);
1.2556 + assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
1.2557 +
1.2558 + // We will need memory barriers unless we can determine a unique
1.2559 + // alias category for this reference. (Note: If for some reason
1.2560 + // the barriers get omitted and the unsafe reference begins to "pollute"
1.2561 + // the alias analysis of the rest of the graph, either Compile::can_alias
1.2562 + // or Compile::must_alias will throw a diagnostic assert.)
1.2563 + bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
1.2564 +
1.2565 + // If we are reading the value of the referent field of a Reference
1.2566 + // object (either by using Unsafe directly or through reflection)
1.2567 + // then, if G1 is enabled, we need to record the referent in an
1.2568 + // SATB log buffer using the pre-barrier mechanism.
1.2569 + // Also we need to add memory barrier to prevent commoning reads
1.2570 + // from this field across safepoint since GC can change its value.
1.2571 + bool need_read_barrier = !is_native_ptr && !is_store &&
1.2572 + offset != top() && heap_base_oop != top();
1.2573 +
1.2574 + if (!is_store && type == T_OBJECT) {
1.2575 + const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
1.2576 + if (tjp != NULL) {
1.2577 + value_type = tjp;
1.2578 + }
1.2579 + }
1.2580 +
1.2581 + receiver = null_check(receiver);
1.2582 + if (stopped()) {
1.2583 + return true;
1.2584 + }
1.2585 + // Heap pointers get a null-check from the interpreter,
1.2586 + // as a courtesy. However, this is not guaranteed by Unsafe,
1.2587 + // and it is not possible to fully distinguish unintended nulls
1.2588 + // from intended ones in this API.
1.2589 +
1.2590 + if (is_volatile) {
1.2591 + // We need to emit leading and trailing CPU membars (see below) in
1.2592 + // addition to memory membars when is_volatile. This is a little
1.2593 + // too strong, but avoids the need to insert per-alias-type
1.2594 + // volatile membars (for stores; compare Parse::do_put_xxx), which
1.2595 + // we cannot do effectively here because we probably only have a
1.2596 + // rough approximation of type.
1.2597 + need_mem_bar = true;
1.2598 + // For Stores, place a memory ordering barrier now.
1.2599 + if (is_store) {
1.2600 + insert_mem_bar(Op_MemBarRelease);
1.2601 + } else {
1.2602 + if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
1.2603 + insert_mem_bar(Op_MemBarVolatile);
1.2604 + }
1.2605 + }
1.2606 + }
1.2607 +
1.2608 + // Memory barrier to prevent normal and 'unsafe' accesses from
1.2609 + // bypassing each other. Happens after null checks, so the
1.2610 + // exception paths do not take memory state from the memory barrier,
1.2611 + // so there's no problems making a strong assert about mixing users
1.2612 + // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
1.2613 + // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
1.2614 + if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
1.2615 +
1.2616 + if (!is_store) {
1.2617 + Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
1.2618 + // load value
1.2619 + switch (type) {
1.2620 + case T_BOOLEAN:
1.2621 + case T_CHAR:
1.2622 + case T_BYTE:
1.2623 + case T_SHORT:
1.2624 + case T_INT:
1.2625 + case T_LONG:
1.2626 + case T_FLOAT:
1.2627 + case T_DOUBLE:
1.2628 + break;
1.2629 + case T_OBJECT:
1.2630 + if (need_read_barrier) {
1.2631 + insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
1.2632 + }
1.2633 + break;
1.2634 + case T_ADDRESS:
1.2635 + // Cast to an int type.
1.2636 + p = _gvn.transform(new (C) CastP2XNode(NULL, p));
1.2637 + p = ConvX2UL(p);
1.2638 + break;
1.2639 + default:
1.2640 + fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
1.2641 + break;
1.2642 + }
1.2643 + // The load node has the control of the preceding MemBarCPUOrder. All
1.2644 + // following nodes will have the control of the MemBarCPUOrder inserted at
1.2645 + // the end of this method. So, pushing the load onto the stack at a later
1.2646 + // point is fine.
1.2647 + set_result(p);
1.2648 + } else {
1.2649 + // place effect of store into memory
1.2650 + switch (type) {
1.2651 + case T_DOUBLE:
1.2652 + val = dstore_rounding(val);
1.2653 + break;
1.2654 + case T_ADDRESS:
1.2655 + // Repackage the long as a pointer.
1.2656 + val = ConvL2X(val);
1.2657 + val = _gvn.transform(new (C) CastX2PNode(val));
1.2658 + break;
1.2659 + }
1.2660 +
1.2661 + MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
1.2662 + if (type != T_OBJECT ) {
1.2663 + (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
1.2664 + } else {
1.2665 + // Possibly an oop being stored to Java heap or native memory
1.2666 + if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
1.2667 + // oop to Java heap.
1.2668 + (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
1.2669 + } else {
1.2670 + // We can't tell at compile time if we are storing in the Java heap or outside
1.2671 + // of it. So we need to emit code to conditionally do the proper type of
1.2672 + // store.
1.2673 +
1.2674 + IdealKit ideal(this);
1.2675 +#define __ ideal.
1.2676 + // QQQ who knows what probability is here??
1.2677 + __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
1.2678 + // Sync IdealKit and graphKit.
1.2679 + sync_kit(ideal);
1.2680 + Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
1.2681 + // Update IdealKit memory.
1.2682 + __ sync_kit(this);
1.2683 + } __ else_(); {
1.2684 + __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
1.2685 + } __ end_if();
1.2686 + // Final sync IdealKit and GraphKit.
1.2687 + final_sync(ideal);
1.2688 +#undef __
1.2689 + }
1.2690 + }
1.2691 + }
1.2692 +
1.2693 + if (is_volatile) {
1.2694 + if (!is_store) {
1.2695 + insert_mem_bar(Op_MemBarAcquire);
1.2696 + } else {
1.2697 + if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
1.2698 + insert_mem_bar(Op_MemBarVolatile);
1.2699 + }
1.2700 + }
1.2701 + }
1.2702 +
1.2703 + if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
1.2704 +
1.2705 + return true;
1.2706 +}
1.2707 +
1.2708 +//----------------------------inline_unsafe_prefetch----------------------------
1.2709 +
1.2710 +bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
1.2711 +#ifndef PRODUCT
1.2712 + {
1.2713 + ResourceMark rm;
1.2714 + // Check the signatures.
1.2715 + ciSignature* sig = callee()->signature();
1.2716 +#ifdef ASSERT
1.2717 + // Object getObject(Object base, int/long offset), etc.
1.2718 + BasicType rtype = sig->return_type()->basic_type();
1.2719 + if (!is_native_ptr) {
1.2720 + assert(sig->count() == 2, "oop prefetch has 2 arguments");
1.2721 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
1.2722 + assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
1.2723 + } else {
1.2724 + assert(sig->count() == 1, "native prefetch has 1 argument");
1.2725 + assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
1.2726 + }
1.2727 +#endif // ASSERT
1.2728 + }
1.2729 +#endif // !PRODUCT
1.2730 +
1.2731 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.2732 +
1.2733 + const int idx = is_static ? 0 : 1;
1.2734 + if (!is_static) {
1.2735 + null_check_receiver();
1.2736 + if (stopped()) {
1.2737 + return true;
1.2738 + }
1.2739 + }
1.2740 +
1.2741 + // Build address expression. See the code in inline_unsafe_access.
1.2742 + Node *adr;
1.2743 + if (!is_native_ptr) {
1.2744 + // The base is either a Java object or a value produced by Unsafe.staticFieldBase
1.2745 + Node* base = argument(idx + 0); // type: oop
1.2746 + // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
1.2747 + Node* offset = argument(idx + 1); // type: long
1.2748 + // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1.2749 + // to be plain byte offsets, which are also the same as those accepted
1.2750 + // by oopDesc::field_base.
1.2751 + assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1.2752 + "fieldOffset must be byte-scaled");
1.2753 + // 32-bit machines ignore the high half!
1.2754 + offset = ConvL2X(offset);
1.2755 + adr = make_unsafe_address(base, offset);
1.2756 + } else {
1.2757 + Node* ptr = argument(idx + 0); // type: long
1.2758 + ptr = ConvL2X(ptr); // adjust Java long to machine word
1.2759 + adr = make_unsafe_address(NULL, ptr);
1.2760 + }
1.2761 +
1.2762 + // Generate the read or write prefetch
1.2763 + Node *prefetch;
1.2764 + if (is_store) {
1.2765 + prefetch = new (C) PrefetchWriteNode(i_o(), adr);
1.2766 + } else {
1.2767 + prefetch = new (C) PrefetchReadNode(i_o(), adr);
1.2768 + }
1.2769 + prefetch->init_req(0, control());
1.2770 + set_i_o(_gvn.transform(prefetch));
1.2771 +
1.2772 + return true;
1.2773 +}
1.2774 +
1.2775 +//----------------------------inline_unsafe_load_store----------------------------
1.2776 +// This method serves a couple of different customers (depending on LoadStoreKind):
1.2777 +//
1.2778 +// LS_cmpxchg:
1.2779 +// public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
1.2780 +// public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
1.2781 +// public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
1.2782 +//
1.2783 +// LS_xadd:
1.2784 +// public int getAndAddInt( Object o, long offset, int delta)
1.2785 +// public long getAndAddLong(Object o, long offset, long delta)
1.2786 +//
1.2787 +// LS_xchg:
1.2788 +// int getAndSet(Object o, long offset, int newValue)
1.2789 +// long getAndSet(Object o, long offset, long newValue)
1.2790 +// Object getAndSet(Object o, long offset, Object newValue)
1.2791 +//
1.2792 +bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
1.2793 + // This basic scheme here is the same as inline_unsafe_access, but
1.2794 + // differs in enough details that combining them would make the code
1.2795 + // overly confusing. (This is a true fact! I originally combined
1.2796 + // them, but even I was confused by it!) As much code/comments as
1.2797 + // possible are retained from inline_unsafe_access though to make
1.2798 + // the correspondences clearer. - dl
1.2799 +
1.2800 + if (callee()->is_static()) return false; // caller must have the capability!
1.2801 +
1.2802 +#ifndef PRODUCT
1.2803 + BasicType rtype;
1.2804 + {
1.2805 + ResourceMark rm;
1.2806 + // Check the signatures.
1.2807 + ciSignature* sig = callee()->signature();
1.2808 + rtype = sig->return_type()->basic_type();
1.2809 + if (kind == LS_xadd || kind == LS_xchg) {
1.2810 + // Check the signatures.
1.2811 +#ifdef ASSERT
1.2812 + assert(rtype == type, "get and set must return the expected type");
1.2813 + assert(sig->count() == 3, "get and set has 3 arguments");
1.2814 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
1.2815 + assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
1.2816 + assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
1.2817 +#endif // ASSERT
1.2818 + } else if (kind == LS_cmpxchg) {
1.2819 + // Check the signatures.
1.2820 +#ifdef ASSERT
1.2821 + assert(rtype == T_BOOLEAN, "CAS must return boolean");
1.2822 + assert(sig->count() == 4, "CAS has 4 arguments");
1.2823 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
1.2824 + assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
1.2825 +#endif // ASSERT
1.2826 + } else {
1.2827 + ShouldNotReachHere();
1.2828 + }
1.2829 + }
1.2830 +#endif //PRODUCT
1.2831 +
1.2832 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.2833 +
1.2834 + // Get arguments:
1.2835 + Node* receiver = NULL;
1.2836 + Node* base = NULL;
1.2837 + Node* offset = NULL;
1.2838 + Node* oldval = NULL;
1.2839 + Node* newval = NULL;
1.2840 + if (kind == LS_cmpxchg) {
1.2841 + const bool two_slot_type = type2size[type] == 2;
1.2842 + receiver = argument(0); // type: oop
1.2843 + base = argument(1); // type: oop
1.2844 + offset = argument(2); // type: long
1.2845 + oldval = argument(4); // type: oop, int, or long
1.2846 + newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
1.2847 + } else if (kind == LS_xadd || kind == LS_xchg){
1.2848 + receiver = argument(0); // type: oop
1.2849 + base = argument(1); // type: oop
1.2850 + offset = argument(2); // type: long
1.2851 + oldval = NULL;
1.2852 + newval = argument(4); // type: oop, int, or long
1.2853 + }
1.2854 +
1.2855 + // Null check receiver.
1.2856 + receiver = null_check(receiver);
1.2857 + if (stopped()) {
1.2858 + return true;
1.2859 + }
1.2860 +
1.2861 + // Build field offset expression.
1.2862 + // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1.2863 + // to be plain byte offsets, which are also the same as those accepted
1.2864 + // by oopDesc::field_base.
1.2865 + assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
1.2866 + // 32-bit machines ignore the high half of long offsets
1.2867 + offset = ConvL2X(offset);
1.2868 + Node* adr = make_unsafe_address(base, offset);
1.2869 + const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1.2870 +
1.2871 + // For CAS, unlike inline_unsafe_access, there seems no point in
1.2872 + // trying to refine types. Just use the coarse types here.
1.2873 + const Type *value_type = Type::get_const_basic_type(type);
1.2874 + Compile::AliasType* alias_type = C->alias_type(adr_type);
1.2875 + assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
1.2876 +
1.2877 + if (kind == LS_xchg && type == T_OBJECT) {
1.2878 + const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
1.2879 + if (tjp != NULL) {
1.2880 + value_type = tjp;
1.2881 + }
1.2882 + }
1.2883 +
1.2884 + int alias_idx = C->get_alias_index(adr_type);
1.2885 +
1.2886 + // Memory-model-wise, a LoadStore acts like a little synchronized
1.2887 + // block, so needs barriers on each side. These don't translate
1.2888 + // into actual barriers on most machines, but we still need rest of
1.2889 + // compiler to respect ordering.
1.2890 +
1.2891 + insert_mem_bar(Op_MemBarRelease);
1.2892 + insert_mem_bar(Op_MemBarCPUOrder);
1.2893 +
1.2894 + // 4984716: MemBars must be inserted before this
1.2895 + // memory node in order to avoid a false
1.2896 + // dependency which will confuse the scheduler.
1.2897 + Node *mem = memory(alias_idx);
1.2898 +
1.2899 + // For now, we handle only those cases that actually exist: ints,
1.2900 + // longs, and Object. Adding others should be straightforward.
1.2901 + Node* load_store;
1.2902 + switch(type) {
1.2903 + case T_INT:
1.2904 + if (kind == LS_xadd) {
1.2905 + load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
1.2906 + } else if (kind == LS_xchg) {
1.2907 + load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
1.2908 + } else if (kind == LS_cmpxchg) {
1.2909 + load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
1.2910 + } else {
1.2911 + ShouldNotReachHere();
1.2912 + }
1.2913 + break;
1.2914 + case T_LONG:
1.2915 + if (kind == LS_xadd) {
1.2916 + load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
1.2917 + } else if (kind == LS_xchg) {
1.2918 + load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
1.2919 + } else if (kind == LS_cmpxchg) {
1.2920 + load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
1.2921 + } else {
1.2922 + ShouldNotReachHere();
1.2923 + }
1.2924 + break;
1.2925 + case T_OBJECT:
1.2926 + // Transformation of a value which could be NULL pointer (CastPP #NULL)
1.2927 + // could be delayed during Parse (for example, in adjust_map_after_if()).
1.2928 + // Execute transformation here to avoid barrier generation in such case.
1.2929 + if (_gvn.type(newval) == TypePtr::NULL_PTR)
1.2930 + newval = _gvn.makecon(TypePtr::NULL_PTR);
1.2931 +
1.2932 + // Reference stores need a store barrier.
1.2933 + if (kind == LS_xchg) {
1.2934 + // If pre-barrier must execute before the oop store, old value will require do_load here.
1.2935 + if (!can_move_pre_barrier()) {
1.2936 + pre_barrier(true /* do_load*/,
1.2937 + control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
1.2938 + NULL /* pre_val*/,
1.2939 + T_OBJECT);
1.2940 + } // Else move pre_barrier to use load_store value, see below.
1.2941 + } else if (kind == LS_cmpxchg) {
1.2942 + // Same as for newval above:
1.2943 + if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
1.2944 + oldval = _gvn.makecon(TypePtr::NULL_PTR);
1.2945 + }
1.2946 + // The only known value which might get overwritten is oldval.
1.2947 + pre_barrier(false /* do_load */,
1.2948 + control(), NULL, NULL, max_juint, NULL, NULL,
1.2949 + oldval /* pre_val */,
1.2950 + T_OBJECT);
1.2951 + } else {
1.2952 + ShouldNotReachHere();
1.2953 + }
1.2954 +
1.2955 +#ifdef _LP64
1.2956 + if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1.2957 + Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
1.2958 + if (kind == LS_xchg) {
1.2959 + load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
1.2960 + newval_enc, adr_type, value_type->make_narrowoop()));
1.2961 + } else {
1.2962 + assert(kind == LS_cmpxchg, "wrong LoadStore operation");
1.2963 + Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
1.2964 + load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
1.2965 + newval_enc, oldval_enc));
1.2966 + }
1.2967 + } else
1.2968 +#endif
1.2969 + {
1.2970 + if (kind == LS_xchg) {
1.2971 + load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
1.2972 + } else {
1.2973 + assert(kind == LS_cmpxchg, "wrong LoadStore operation");
1.2974 + load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
1.2975 + }
1.2976 + }
1.2977 + post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
1.2978 + break;
1.2979 + default:
1.2980 + fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
1.2981 + break;
1.2982 + }
1.2983 +
1.2984 + // SCMemProjNodes represent the memory state of a LoadStore. Their
1.2985 + // main role is to prevent LoadStore nodes from being optimized away
1.2986 + // when their results aren't used.
1.2987 + Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
1.2988 + set_memory(proj, alias_idx);
1.2989 +
1.2990 + if (type == T_OBJECT && kind == LS_xchg) {
1.2991 +#ifdef _LP64
1.2992 + if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1.2993 + load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
1.2994 + }
1.2995 +#endif
1.2996 + if (can_move_pre_barrier()) {
1.2997 + // Don't need to load pre_val. The old value is returned by load_store.
1.2998 + // The pre_barrier can execute after the xchg as long as no safepoint
1.2999 + // gets inserted between them.
1.3000 + pre_barrier(false /* do_load */,
1.3001 + control(), NULL, NULL, max_juint, NULL, NULL,
1.3002 + load_store /* pre_val */,
1.3003 + T_OBJECT);
1.3004 + }
1.3005 + }
1.3006 +
1.3007 + // Add the trailing membar surrounding the access
1.3008 + insert_mem_bar(Op_MemBarCPUOrder);
1.3009 + insert_mem_bar(Op_MemBarAcquire);
1.3010 +
1.3011 + assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
1.3012 + set_result(load_store);
1.3013 + return true;
1.3014 +}
1.3015 +
1.3016 +//----------------------------inline_unsafe_ordered_store----------------------
1.3017 +// public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
1.3018 +// public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
1.3019 +// public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
1.3020 +bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
1.3021 + // This is another variant of inline_unsafe_access, differing in
1.3022 + // that it always issues store-store ("release") barrier and ensures
1.3023 + // store-atomicity (which only matters for "long").
1.3024 +
1.3025 + if (callee()->is_static()) return false; // caller must have the capability!
1.3026 +
1.3027 +#ifndef PRODUCT
1.3028 + {
1.3029 + ResourceMark rm;
1.3030 + // Check the signatures.
1.3031 + ciSignature* sig = callee()->signature();
1.3032 +#ifdef ASSERT
1.3033 + BasicType rtype = sig->return_type()->basic_type();
1.3034 + assert(rtype == T_VOID, "must return void");
1.3035 + assert(sig->count() == 3, "has 3 arguments");
1.3036 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
1.3037 + assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
1.3038 +#endif // ASSERT
1.3039 + }
1.3040 +#endif //PRODUCT
1.3041 +
1.3042 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.3043 +
1.3044 + // Get arguments:
1.3045 + Node* receiver = argument(0); // type: oop
1.3046 + Node* base = argument(1); // type: oop
1.3047 + Node* offset = argument(2); // type: long
1.3048 + Node* val = argument(4); // type: oop, int, or long
1.3049 +
1.3050 + // Null check receiver.
1.3051 + receiver = null_check(receiver);
1.3052 + if (stopped()) {
1.3053 + return true;
1.3054 + }
1.3055 +
1.3056 + // Build field offset expression.
1.3057 + assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
1.3058 + // 32-bit machines ignore the high half of long offsets
1.3059 + offset = ConvL2X(offset);
1.3060 + Node* adr = make_unsafe_address(base, offset);
1.3061 + const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1.3062 + const Type *value_type = Type::get_const_basic_type(type);
1.3063 + Compile::AliasType* alias_type = C->alias_type(adr_type);
1.3064 +
1.3065 + insert_mem_bar(Op_MemBarRelease);
1.3066 + insert_mem_bar(Op_MemBarCPUOrder);
1.3067 + // Ensure that the store is atomic for longs:
1.3068 + const bool require_atomic_access = true;
1.3069 + Node* store;
1.3070 + if (type == T_OBJECT) // reference stores need a store barrier.
1.3071 + store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
1.3072 + else {
1.3073 + store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
1.3074 + }
1.3075 + insert_mem_bar(Op_MemBarCPUOrder);
1.3076 + return true;
1.3077 +}
1.3078 +
1.3079 +bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
1.3080 + // Regardless of form, don't allow previous ld/st to move down,
1.3081 + // then issue acquire, release, or volatile mem_bar.
1.3082 + insert_mem_bar(Op_MemBarCPUOrder);
1.3083 + switch(id) {
1.3084 + case vmIntrinsics::_loadFence:
1.3085 + insert_mem_bar(Op_LoadFence);
1.3086 + return true;
1.3087 + case vmIntrinsics::_storeFence:
1.3088 + insert_mem_bar(Op_StoreFence);
1.3089 + return true;
1.3090 + case vmIntrinsics::_fullFence:
1.3091 + insert_mem_bar(Op_MemBarVolatile);
1.3092 + return true;
1.3093 + default:
1.3094 + fatal_unexpected_iid(id);
1.3095 + return false;
1.3096 + }
1.3097 +}
1.3098 +
1.3099 +bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
1.3100 + if (!kls->is_Con()) {
1.3101 + return true;
1.3102 + }
1.3103 + const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
1.3104 + if (klsptr == NULL) {
1.3105 + return true;
1.3106 + }
1.3107 + ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
1.3108 + // don't need a guard for a klass that is already initialized
1.3109 + return !ik->is_initialized();
1.3110 +}
1.3111 +
1.3112 +//----------------------------inline_unsafe_allocate---------------------------
1.3113 +// public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
1.3114 +bool LibraryCallKit::inline_unsafe_allocate() {
1.3115 + if (callee()->is_static()) return false; // caller must have the capability!
1.3116 +
1.3117 + null_check_receiver(); // null-check, then ignore
1.3118 + Node* cls = null_check(argument(1));
1.3119 + if (stopped()) return true;
1.3120 +
1.3121 + Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
1.3122 + kls = null_check(kls);
1.3123 + if (stopped()) return true; // argument was like int.class
1.3124 +
1.3125 + Node* test = NULL;
1.3126 + if (LibraryCallKit::klass_needs_init_guard(kls)) {
1.3127 + // Note: The argument might still be an illegal value like
1.3128 + // Serializable.class or Object[].class. The runtime will handle it.
1.3129 + // But we must make an explicit check for initialization.
1.3130 + Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
1.3131 + // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
1.3132 + // can generate code to load it as unsigned byte.
1.3133 + Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
1.3134 + Node* bits = intcon(InstanceKlass::fully_initialized);
1.3135 + test = _gvn.transform(new (C) SubINode(inst, bits));
1.3136 + // The 'test' is non-zero if we need to take a slow path.
1.3137 + }
1.3138 +
1.3139 + Node* obj = new_instance(kls, test);
1.3140 + set_result(obj);
1.3141 + return true;
1.3142 +}
1.3143 +
1.3144 +#ifdef TRACE_HAVE_INTRINSICS
1.3145 +/*
1.3146 + * oop -> myklass
1.3147 + * myklass->trace_id |= USED
1.3148 + * return myklass->trace_id & ~0x3
1.3149 + */
1.3150 +bool LibraryCallKit::inline_native_classID() {
1.3151 + null_check_receiver(); // null-check, then ignore
1.3152 + Node* cls = null_check(argument(1), T_OBJECT);
1.3153 + Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
1.3154 + kls = null_check(kls, T_OBJECT);
1.3155 + ByteSize offset = TRACE_ID_OFFSET;
1.3156 + Node* insp = basic_plus_adr(kls, in_bytes(offset));
1.3157 + Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
1.3158 + Node* bits = longcon(~0x03l); // ignore bit 0 & 1
1.3159 + Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
1.3160 + Node* clsused = longcon(0x01l); // set the class bit
1.3161 + Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
1.3162 +
1.3163 + const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
1.3164 + store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
1.3165 + set_result(andl);
1.3166 + return true;
1.3167 +}
1.3168 +
1.3169 +bool LibraryCallKit::inline_native_threadID() {
1.3170 + Node* tls_ptr = NULL;
1.3171 + Node* cur_thr = generate_current_thread(tls_ptr);
1.3172 + Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
1.3173 + Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
1.3174 + p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
1.3175 +
1.3176 + Node* threadid = NULL;
1.3177 + size_t thread_id_size = OSThread::thread_id_size();
1.3178 + if (thread_id_size == (size_t) BytesPerLong) {
1.3179 + threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
1.3180 + } else if (thread_id_size == (size_t) BytesPerInt) {
1.3181 + threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
1.3182 + } else {
1.3183 + ShouldNotReachHere();
1.3184 + }
1.3185 + set_result(threadid);
1.3186 + return true;
1.3187 +}
1.3188 +#endif
1.3189 +
1.3190 +//------------------------inline_native_time_funcs--------------
1.3191 +// inline code for System.currentTimeMillis() and System.nanoTime()
1.3192 +// these have the same type and signature
1.3193 +bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
1.3194 + const TypeFunc* tf = OptoRuntime::void_long_Type();
1.3195 + const TypePtr* no_memory_effects = NULL;
1.3196 + Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
1.3197 + Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
1.3198 +#ifdef ASSERT
1.3199 + Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
1.3200 + assert(value_top == top(), "second value must be top");
1.3201 +#endif
1.3202 + set_result(value);
1.3203 + return true;
1.3204 +}
1.3205 +
1.3206 +//------------------------inline_native_currentThread------------------
1.3207 +bool LibraryCallKit::inline_native_currentThread() {
1.3208 + Node* junk = NULL;
1.3209 + set_result(generate_current_thread(junk));
1.3210 + return true;
1.3211 +}
1.3212 +
1.3213 +//------------------------inline_native_isInterrupted------------------
1.3214 +// private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
1.3215 +bool LibraryCallKit::inline_native_isInterrupted() {
1.3216 + // Add a fast path to t.isInterrupted(clear_int):
1.3217 + // (t == Thread.current() &&
1.3218 + // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
1.3219 + // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
1.3220 + // So, in the common case that the interrupt bit is false,
1.3221 + // we avoid making a call into the VM. Even if the interrupt bit
1.3222 + // is true, if the clear_int argument is false, we avoid the VM call.
1.3223 + // However, if the receiver is not currentThread, we must call the VM,
1.3224 + // because there must be some locking done around the operation.
1.3225 +
1.3226 + // We only go to the fast case code if we pass two guards.
1.3227 + // Paths which do not pass are accumulated in the slow_region.
1.3228 +
1.3229 + enum {
1.3230 + no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
1.3231 + no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
1.3232 + slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
1.3233 + PATH_LIMIT
1.3234 + };
1.3235 +
1.3236 + // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
1.3237 + // out of the function.
1.3238 + insert_mem_bar(Op_MemBarCPUOrder);
1.3239 +
1.3240 + RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
1.3241 + PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
1.3242 +
1.3243 + RegionNode* slow_region = new (C) RegionNode(1);
1.3244 + record_for_igvn(slow_region);
1.3245 +
1.3246 + // (a) Receiving thread must be the current thread.
1.3247 + Node* rec_thr = argument(0);
1.3248 + Node* tls_ptr = NULL;
1.3249 + Node* cur_thr = generate_current_thread(tls_ptr);
1.3250 + Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
1.3251 + Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
1.3252 +
1.3253 + generate_slow_guard(bol_thr, slow_region);
1.3254 +
1.3255 + // (b) Interrupt bit on TLS must be false.
1.3256 + Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
1.3257 + Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
1.3258 + p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
1.3259 +
1.3260 + // Set the control input on the field _interrupted read to prevent it floating up.
1.3261 + Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
1.3262 + Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
1.3263 + Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
1.3264 +
1.3265 + IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1.3266 +
1.3267 + // First fast path: if (!TLS._interrupted) return false;
1.3268 + Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
1.3269 + result_rgn->init_req(no_int_result_path, false_bit);
1.3270 + result_val->init_req(no_int_result_path, intcon(0));
1.3271 +
1.3272 + // drop through to next case
1.3273 + set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
1.3274 +
1.3275 +#ifndef TARGET_OS_FAMILY_windows
1.3276 + // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
1.3277 + Node* clr_arg = argument(1);
1.3278 + Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
1.3279 + Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
1.3280 + IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
1.3281 +
1.3282 + // Second fast path: ... else if (!clear_int) return true;
1.3283 + Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
1.3284 + result_rgn->init_req(no_clear_result_path, false_arg);
1.3285 + result_val->init_req(no_clear_result_path, intcon(1));
1.3286 +
1.3287 + // drop through to next case
1.3288 + set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
1.3289 +#else
1.3290 + // To return true on Windows you must read the _interrupted field
1.3291 + // and check the the event state i.e. take the slow path.
1.3292 +#endif // TARGET_OS_FAMILY_windows
1.3293 +
1.3294 + // (d) Otherwise, go to the slow path.
1.3295 + slow_region->add_req(control());
1.3296 + set_control( _gvn.transform(slow_region));
1.3297 +
1.3298 + if (stopped()) {
1.3299 + // There is no slow path.
1.3300 + result_rgn->init_req(slow_result_path, top());
1.3301 + result_val->init_req(slow_result_path, top());
1.3302 + } else {
1.3303 + // non-virtual because it is a private non-static
1.3304 + CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
1.3305 +
1.3306 + Node* slow_val = set_results_for_java_call(slow_call);
1.3307 + // this->control() comes from set_results_for_java_call
1.3308 +
1.3309 + Node* fast_io = slow_call->in(TypeFunc::I_O);
1.3310 + Node* fast_mem = slow_call->in(TypeFunc::Memory);
1.3311 +
1.3312 + // These two phis are pre-filled with copies of of the fast IO and Memory
1.3313 + PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
1.3314 + PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
1.3315 +
1.3316 + result_rgn->init_req(slow_result_path, control());
1.3317 + result_io ->init_req(slow_result_path, i_o());
1.3318 + result_mem->init_req(slow_result_path, reset_memory());
1.3319 + result_val->init_req(slow_result_path, slow_val);
1.3320 +
1.3321 + set_all_memory(_gvn.transform(result_mem));
1.3322 + set_i_o( _gvn.transform(result_io));
1.3323 + }
1.3324 +
1.3325 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3326 + set_result(result_rgn, result_val);
1.3327 + return true;
1.3328 +}
1.3329 +
1.3330 +//---------------------------load_mirror_from_klass----------------------------
1.3331 +// Given a klass oop, load its java mirror (a java.lang.Class oop).
1.3332 +Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
1.3333 + Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
1.3334 + return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
1.3335 +}
1.3336 +
1.3337 +//-----------------------load_klass_from_mirror_common-------------------------
1.3338 +// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
1.3339 +// Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
1.3340 +// and branch to the given path on the region.
1.3341 +// If never_see_null, take an uncommon trap on null, so we can optimistically
1.3342 +// compile for the non-null case.
1.3343 +// If the region is NULL, force never_see_null = true.
1.3344 +Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
1.3345 + bool never_see_null,
1.3346 + RegionNode* region,
1.3347 + int null_path,
1.3348 + int offset) {
1.3349 + if (region == NULL) never_see_null = true;
1.3350 + Node* p = basic_plus_adr(mirror, offset);
1.3351 + const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
1.3352 + Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
1.3353 + Node* null_ctl = top();
1.3354 + kls = null_check_oop(kls, &null_ctl, never_see_null);
1.3355 + if (region != NULL) {
1.3356 + // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
1.3357 + region->init_req(null_path, null_ctl);
1.3358 + } else {
1.3359 + assert(null_ctl == top(), "no loose ends");
1.3360 + }
1.3361 + return kls;
1.3362 +}
1.3363 +
1.3364 +//--------------------(inline_native_Class_query helpers)---------------------
1.3365 +// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
1.3366 +// Fall through if (mods & mask) == bits, take the guard otherwise.
1.3367 +Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
1.3368 + // Branch around if the given klass has the given modifier bit set.
1.3369 + // Like generate_guard, adds a new path onto the region.
1.3370 + Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
1.3371 + Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
1.3372 + Node* mask = intcon(modifier_mask);
1.3373 + Node* bits = intcon(modifier_bits);
1.3374 + Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
1.3375 + Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
1.3376 + Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1.3377 + return generate_fair_guard(bol, region);
1.3378 +}
1.3379 +Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
1.3380 + return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
1.3381 +}
1.3382 +
1.3383 +//-------------------------inline_native_Class_query-------------------
1.3384 +bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
1.3385 + const Type* return_type = TypeInt::BOOL;
1.3386 + Node* prim_return_value = top(); // what happens if it's a primitive class?
1.3387 + bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
1.3388 + bool expect_prim = false; // most of these guys expect to work on refs
1.3389 +
1.3390 + enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
1.3391 +
1.3392 + Node* mirror = argument(0);
1.3393 + Node* obj = top();
1.3394 +
1.3395 + switch (id) {
1.3396 + case vmIntrinsics::_isInstance:
1.3397 + // nothing is an instance of a primitive type
1.3398 + prim_return_value = intcon(0);
1.3399 + obj = argument(1);
1.3400 + break;
1.3401 + case vmIntrinsics::_getModifiers:
1.3402 + prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
1.3403 + assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
1.3404 + return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
1.3405 + break;
1.3406 + case vmIntrinsics::_isInterface:
1.3407 + prim_return_value = intcon(0);
1.3408 + break;
1.3409 + case vmIntrinsics::_isArray:
1.3410 + prim_return_value = intcon(0);
1.3411 + expect_prim = true; // cf. ObjectStreamClass.getClassSignature
1.3412 + break;
1.3413 + case vmIntrinsics::_isPrimitive:
1.3414 + prim_return_value = intcon(1);
1.3415 + expect_prim = true; // obviously
1.3416 + break;
1.3417 + case vmIntrinsics::_getSuperclass:
1.3418 + prim_return_value = null();
1.3419 + return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
1.3420 + break;
1.3421 + case vmIntrinsics::_getComponentType:
1.3422 + prim_return_value = null();
1.3423 + return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
1.3424 + break;
1.3425 + case vmIntrinsics::_getClassAccessFlags:
1.3426 + prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
1.3427 + return_type = TypeInt::INT; // not bool! 6297094
1.3428 + break;
1.3429 + default:
1.3430 + fatal_unexpected_iid(id);
1.3431 + break;
1.3432 + }
1.3433 +
1.3434 + const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
1.3435 + if (mirror_con == NULL) return false; // cannot happen?
1.3436 +
1.3437 +#ifndef PRODUCT
1.3438 + if (C->print_intrinsics() || C->print_inlining()) {
1.3439 + ciType* k = mirror_con->java_mirror_type();
1.3440 + if (k) {
1.3441 + tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
1.3442 + k->print_name();
1.3443 + tty->cr();
1.3444 + }
1.3445 + }
1.3446 +#endif
1.3447 +
1.3448 + // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
1.3449 + RegionNode* region = new (C) RegionNode(PATH_LIMIT);
1.3450 + record_for_igvn(region);
1.3451 + PhiNode* phi = new (C) PhiNode(region, return_type);
1.3452 +
1.3453 + // The mirror will never be null of Reflection.getClassAccessFlags, however
1.3454 + // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
1.3455 + // if it is. See bug 4774291.
1.3456 +
1.3457 + // For Reflection.getClassAccessFlags(), the null check occurs in
1.3458 + // the wrong place; see inline_unsafe_access(), above, for a similar
1.3459 + // situation.
1.3460 + mirror = null_check(mirror);
1.3461 + // If mirror or obj is dead, only null-path is taken.
1.3462 + if (stopped()) return true;
1.3463 +
1.3464 + if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
1.3465 +
1.3466 + // Now load the mirror's klass metaobject, and null-check it.
1.3467 + // Side-effects region with the control path if the klass is null.
1.3468 + Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
1.3469 + // If kls is null, we have a primitive mirror.
1.3470 + phi->init_req(_prim_path, prim_return_value);
1.3471 + if (stopped()) { set_result(region, phi); return true; }
1.3472 + bool safe_for_replace = (region->in(_prim_path) == top());
1.3473 +
1.3474 + Node* p; // handy temp
1.3475 + Node* null_ctl;
1.3476 +
1.3477 + // Now that we have the non-null klass, we can perform the real query.
1.3478 + // For constant classes, the query will constant-fold in LoadNode::Value.
1.3479 + Node* query_value = top();
1.3480 + switch (id) {
1.3481 + case vmIntrinsics::_isInstance:
1.3482 + // nothing is an instance of a primitive type
1.3483 + query_value = gen_instanceof(obj, kls, safe_for_replace);
1.3484 + break;
1.3485 +
1.3486 + case vmIntrinsics::_getModifiers:
1.3487 + p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
1.3488 + query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
1.3489 + break;
1.3490 +
1.3491 + case vmIntrinsics::_isInterface:
1.3492 + // (To verify this code sequence, check the asserts in JVM_IsInterface.)
1.3493 + if (generate_interface_guard(kls, region) != NULL)
1.3494 + // A guard was added. If the guard is taken, it was an interface.
1.3495 + phi->add_req(intcon(1));
1.3496 + // If we fall through, it's a plain class.
1.3497 + query_value = intcon(0);
1.3498 + break;
1.3499 +
1.3500 + case vmIntrinsics::_isArray:
1.3501 + // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
1.3502 + if (generate_array_guard(kls, region) != NULL)
1.3503 + // A guard was added. If the guard is taken, it was an array.
1.3504 + phi->add_req(intcon(1));
1.3505 + // If we fall through, it's a plain class.
1.3506 + query_value = intcon(0);
1.3507 + break;
1.3508 +
1.3509 + case vmIntrinsics::_isPrimitive:
1.3510 + query_value = intcon(0); // "normal" path produces false
1.3511 + break;
1.3512 +
1.3513 + case vmIntrinsics::_getSuperclass:
1.3514 + // The rules here are somewhat unfortunate, but we can still do better
1.3515 + // with random logic than with a JNI call.
1.3516 + // Interfaces store null or Object as _super, but must report null.
1.3517 + // Arrays store an intermediate super as _super, but must report Object.
1.3518 + // Other types can report the actual _super.
1.3519 + // (To verify this code sequence, check the asserts in JVM_IsInterface.)
1.3520 + if (generate_interface_guard(kls, region) != NULL)
1.3521 + // A guard was added. If the guard is taken, it was an interface.
1.3522 + phi->add_req(null());
1.3523 + if (generate_array_guard(kls, region) != NULL)
1.3524 + // A guard was added. If the guard is taken, it was an array.
1.3525 + phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
1.3526 + // If we fall through, it's a plain class. Get its _super.
1.3527 + p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
1.3528 + kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
1.3529 + null_ctl = top();
1.3530 + kls = null_check_oop(kls, &null_ctl);
1.3531 + if (null_ctl != top()) {
1.3532 + // If the guard is taken, Object.superClass is null (both klass and mirror).
1.3533 + region->add_req(null_ctl);
1.3534 + phi ->add_req(null());
1.3535 + }
1.3536 + if (!stopped()) {
1.3537 + query_value = load_mirror_from_klass(kls);
1.3538 + }
1.3539 + break;
1.3540 +
1.3541 + case vmIntrinsics::_getComponentType:
1.3542 + if (generate_array_guard(kls, region) != NULL) {
1.3543 + // Be sure to pin the oop load to the guard edge just created:
1.3544 + Node* is_array_ctrl = region->in(region->req()-1);
1.3545 + Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
1.3546 + Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
1.3547 + phi->add_req(cmo);
1.3548 + }
1.3549 + query_value = null(); // non-array case is null
1.3550 + break;
1.3551 +
1.3552 + case vmIntrinsics::_getClassAccessFlags:
1.3553 + p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
1.3554 + query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
1.3555 + break;
1.3556 +
1.3557 + default:
1.3558 + fatal_unexpected_iid(id);
1.3559 + break;
1.3560 + }
1.3561 +
1.3562 + // Fall-through is the normal case of a query to a real class.
1.3563 + phi->init_req(1, query_value);
1.3564 + region->init_req(1, control());
1.3565 +
1.3566 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3567 + set_result(region, phi);
1.3568 + return true;
1.3569 +}
1.3570 +
1.3571 +//--------------------------inline_native_subtype_check------------------------
1.3572 +// This intrinsic takes the JNI calls out of the heart of
1.3573 +// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
1.3574 +bool LibraryCallKit::inline_native_subtype_check() {
1.3575 + // Pull both arguments off the stack.
1.3576 + Node* args[2]; // two java.lang.Class mirrors: superc, subc
1.3577 + args[0] = argument(0);
1.3578 + args[1] = argument(1);
1.3579 + Node* klasses[2]; // corresponding Klasses: superk, subk
1.3580 + klasses[0] = klasses[1] = top();
1.3581 +
1.3582 + enum {
1.3583 + // A full decision tree on {superc is prim, subc is prim}:
1.3584 + _prim_0_path = 1, // {P,N} => false
1.3585 + // {P,P} & superc!=subc => false
1.3586 + _prim_same_path, // {P,P} & superc==subc => true
1.3587 + _prim_1_path, // {N,P} => false
1.3588 + _ref_subtype_path, // {N,N} & subtype check wins => true
1.3589 + _both_ref_path, // {N,N} & subtype check loses => false
1.3590 + PATH_LIMIT
1.3591 + };
1.3592 +
1.3593 + RegionNode* region = new (C) RegionNode(PATH_LIMIT);
1.3594 + Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1.3595 + record_for_igvn(region);
1.3596 +
1.3597 + const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
1.3598 + const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
1.3599 + int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
1.3600 +
1.3601 + // First null-check both mirrors and load each mirror's klass metaobject.
1.3602 + int which_arg;
1.3603 + for (which_arg = 0; which_arg <= 1; which_arg++) {
1.3604 + Node* arg = args[which_arg];
1.3605 + arg = null_check(arg);
1.3606 + if (stopped()) break;
1.3607 + args[which_arg] = arg;
1.3608 +
1.3609 + Node* p = basic_plus_adr(arg, class_klass_offset);
1.3610 + Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
1.3611 + klasses[which_arg] = _gvn.transform(kls);
1.3612 + }
1.3613 +
1.3614 + // Having loaded both klasses, test each for null.
1.3615 + bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
1.3616 + for (which_arg = 0; which_arg <= 1; which_arg++) {
1.3617 + Node* kls = klasses[which_arg];
1.3618 + Node* null_ctl = top();
1.3619 + kls = null_check_oop(kls, &null_ctl, never_see_null);
1.3620 + int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
1.3621 + region->init_req(prim_path, null_ctl);
1.3622 + if (stopped()) break;
1.3623 + klasses[which_arg] = kls;
1.3624 + }
1.3625 +
1.3626 + if (!stopped()) {
1.3627 + // now we have two reference types, in klasses[0..1]
1.3628 + Node* subk = klasses[1]; // the argument to isAssignableFrom
1.3629 + Node* superk = klasses[0]; // the receiver
1.3630 + region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
1.3631 + // now we have a successful reference subtype check
1.3632 + region->set_req(_ref_subtype_path, control());
1.3633 + }
1.3634 +
1.3635 + // If both operands are primitive (both klasses null), then
1.3636 + // we must return true when they are identical primitives.
1.3637 + // It is convenient to test this after the first null klass check.
1.3638 + set_control(region->in(_prim_0_path)); // go back to first null check
1.3639 + if (!stopped()) {
1.3640 + // Since superc is primitive, make a guard for the superc==subc case.
1.3641 + Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
1.3642 + Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
1.3643 + generate_guard(bol_eq, region, PROB_FAIR);
1.3644 + if (region->req() == PATH_LIMIT+1) {
1.3645 + // A guard was added. If the added guard is taken, superc==subc.
1.3646 + region->swap_edges(PATH_LIMIT, _prim_same_path);
1.3647 + region->del_req(PATH_LIMIT);
1.3648 + }
1.3649 + region->set_req(_prim_0_path, control()); // Not equal after all.
1.3650 + }
1.3651 +
1.3652 + // these are the only paths that produce 'true':
1.3653 + phi->set_req(_prim_same_path, intcon(1));
1.3654 + phi->set_req(_ref_subtype_path, intcon(1));
1.3655 +
1.3656 + // pull together the cases:
1.3657 + assert(region->req() == PATH_LIMIT, "sane region");
1.3658 + for (uint i = 1; i < region->req(); i++) {
1.3659 + Node* ctl = region->in(i);
1.3660 + if (ctl == NULL || ctl == top()) {
1.3661 + region->set_req(i, top());
1.3662 + phi ->set_req(i, top());
1.3663 + } else if (phi->in(i) == NULL) {
1.3664 + phi->set_req(i, intcon(0)); // all other paths produce 'false'
1.3665 + }
1.3666 + }
1.3667 +
1.3668 + set_control(_gvn.transform(region));
1.3669 + set_result(_gvn.transform(phi));
1.3670 + return true;
1.3671 +}
1.3672 +
1.3673 +//---------------------generate_array_guard_common------------------------
1.3674 +Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
1.3675 + bool obj_array, bool not_array) {
1.3676 + // If obj_array/non_array==false/false:
1.3677 + // Branch around if the given klass is in fact an array (either obj or prim).
1.3678 + // If obj_array/non_array==false/true:
1.3679 + // Branch around if the given klass is not an array klass of any kind.
1.3680 + // If obj_array/non_array==true/true:
1.3681 + // Branch around if the kls is not an oop array (kls is int[], String, etc.)
1.3682 + // If obj_array/non_array==true/false:
1.3683 + // Branch around if the kls is an oop array (Object[] or subtype)
1.3684 + //
1.3685 + // Like generate_guard, adds a new path onto the region.
1.3686 + jint layout_con = 0;
1.3687 + Node* layout_val = get_layout_helper(kls, layout_con);
1.3688 + if (layout_val == NULL) {
1.3689 + bool query = (obj_array
1.3690 + ? Klass::layout_helper_is_objArray(layout_con)
1.3691 + : Klass::layout_helper_is_array(layout_con));
1.3692 + if (query == not_array) {
1.3693 + return NULL; // never a branch
1.3694 + } else { // always a branch
1.3695 + Node* always_branch = control();
1.3696 + if (region != NULL)
1.3697 + region->add_req(always_branch);
1.3698 + set_control(top());
1.3699 + return always_branch;
1.3700 + }
1.3701 + }
1.3702 + // Now test the correct condition.
1.3703 + jint nval = (obj_array
1.3704 + ? ((jint)Klass::_lh_array_tag_type_value
1.3705 + << Klass::_lh_array_tag_shift)
1.3706 + : Klass::_lh_neutral_value);
1.3707 + Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
1.3708 + BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
1.3709 + // invert the test if we are looking for a non-array
1.3710 + if (not_array) btest = BoolTest(btest).negate();
1.3711 + Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
1.3712 + return generate_fair_guard(bol, region);
1.3713 +}
1.3714 +
1.3715 +
1.3716 +//-----------------------inline_native_newArray--------------------------
1.3717 +// private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
1.3718 +bool LibraryCallKit::inline_native_newArray() {
1.3719 + Node* mirror = argument(0);
1.3720 + Node* count_val = argument(1);
1.3721 +
1.3722 + mirror = null_check(mirror);
1.3723 + // If mirror or obj is dead, only null-path is taken.
1.3724 + if (stopped()) return true;
1.3725 +
1.3726 + enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
1.3727 + RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
1.3728 + PhiNode* result_val = new(C) PhiNode(result_reg,
1.3729 + TypeInstPtr::NOTNULL);
1.3730 + PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
1.3731 + PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
1.3732 + TypePtr::BOTTOM);
1.3733 +
1.3734 + bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
1.3735 + Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
1.3736 + result_reg, _slow_path);
1.3737 + Node* normal_ctl = control();
1.3738 + Node* no_array_ctl = result_reg->in(_slow_path);
1.3739 +
1.3740 + // Generate code for the slow case. We make a call to newArray().
1.3741 + set_control(no_array_ctl);
1.3742 + if (!stopped()) {
1.3743 + // Either the input type is void.class, or else the
1.3744 + // array klass has not yet been cached. Either the
1.3745 + // ensuing call will throw an exception, or else it
1.3746 + // will cache the array klass for next time.
1.3747 + PreserveJVMState pjvms(this);
1.3748 + CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
1.3749 + Node* slow_result = set_results_for_java_call(slow_call);
1.3750 + // this->control() comes from set_results_for_java_call
1.3751 + result_reg->set_req(_slow_path, control());
1.3752 + result_val->set_req(_slow_path, slow_result);
1.3753 + result_io ->set_req(_slow_path, i_o());
1.3754 + result_mem->set_req(_slow_path, reset_memory());
1.3755 + }
1.3756 +
1.3757 + set_control(normal_ctl);
1.3758 + if (!stopped()) {
1.3759 + // Normal case: The array type has been cached in the java.lang.Class.
1.3760 + // The following call works fine even if the array type is polymorphic.
1.3761 + // It could be a dynamic mix of int[], boolean[], Object[], etc.
1.3762 + Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
1.3763 + result_reg->init_req(_normal_path, control());
1.3764 + result_val->init_req(_normal_path, obj);
1.3765 + result_io ->init_req(_normal_path, i_o());
1.3766 + result_mem->init_req(_normal_path, reset_memory());
1.3767 + }
1.3768 +
1.3769 + // Return the combined state.
1.3770 + set_i_o( _gvn.transform(result_io) );
1.3771 + set_all_memory( _gvn.transform(result_mem));
1.3772 +
1.3773 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3774 + set_result(result_reg, result_val);
1.3775 + return true;
1.3776 +}
1.3777 +
1.3778 +//----------------------inline_native_getLength--------------------------
1.3779 +// public static native int java.lang.reflect.Array.getLength(Object array);
1.3780 +bool LibraryCallKit::inline_native_getLength() {
1.3781 + if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1.3782 +
1.3783 + Node* array = null_check(argument(0));
1.3784 + // If array is dead, only null-path is taken.
1.3785 + if (stopped()) return true;
1.3786 +
1.3787 + // Deoptimize if it is a non-array.
1.3788 + Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
1.3789 +
1.3790 + if (non_array != NULL) {
1.3791 + PreserveJVMState pjvms(this);
1.3792 + set_control(non_array);
1.3793 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.3794 + Deoptimization::Action_maybe_recompile);
1.3795 + }
1.3796 +
1.3797 + // If control is dead, only non-array-path is taken.
1.3798 + if (stopped()) return true;
1.3799 +
1.3800 + // The works fine even if the array type is polymorphic.
1.3801 + // It could be a dynamic mix of int[], boolean[], Object[], etc.
1.3802 + Node* result = load_array_length(array);
1.3803 +
1.3804 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3805 + set_result(result);
1.3806 + return true;
1.3807 +}
1.3808 +
1.3809 +//------------------------inline_array_copyOf----------------------------
1.3810 +// public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
1.3811 +// public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
1.3812 +bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
1.3813 + if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1.3814 +
1.3815 + // Get the arguments.
1.3816 + Node* original = argument(0);
1.3817 + Node* start = is_copyOfRange? argument(1): intcon(0);
1.3818 + Node* end = is_copyOfRange? argument(2): argument(1);
1.3819 + Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
1.3820 +
1.3821 + Node* newcopy;
1.3822 +
1.3823 + // Set the original stack and the reexecute bit for the interpreter to reexecute
1.3824 + // the bytecode that invokes Arrays.copyOf if deoptimization happens.
1.3825 + { PreserveReexecuteState preexecs(this);
1.3826 + jvms()->set_should_reexecute(true);
1.3827 +
1.3828 + array_type_mirror = null_check(array_type_mirror);
1.3829 + original = null_check(original);
1.3830 +
1.3831 + // Check if a null path was taken unconditionally.
1.3832 + if (stopped()) return true;
1.3833 +
1.3834 + Node* orig_length = load_array_length(original);
1.3835 +
1.3836 + Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
1.3837 + klass_node = null_check(klass_node);
1.3838 +
1.3839 + RegionNode* bailout = new (C) RegionNode(1);
1.3840 + record_for_igvn(bailout);
1.3841 +
1.3842 + // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
1.3843 + // Bail out if that is so.
1.3844 + Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
1.3845 + if (not_objArray != NULL) {
1.3846 + // Improve the klass node's type from the new optimistic assumption:
1.3847 + ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
1.3848 + const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1.3849 + Node* cast = new (C) CastPPNode(klass_node, akls);
1.3850 + cast->init_req(0, control());
1.3851 + klass_node = _gvn.transform(cast);
1.3852 + }
1.3853 +
1.3854 + // Bail out if either start or end is negative.
1.3855 + generate_negative_guard(start, bailout, &start);
1.3856 + generate_negative_guard(end, bailout, &end);
1.3857 +
1.3858 + Node* length = end;
1.3859 + if (_gvn.type(start) != TypeInt::ZERO) {
1.3860 + length = _gvn.transform(new (C) SubINode(end, start));
1.3861 + }
1.3862 +
1.3863 + // Bail out if length is negative.
1.3864 + // Without this the new_array would throw
1.3865 + // NegativeArraySizeException but IllegalArgumentException is what
1.3866 + // should be thrown
1.3867 + generate_negative_guard(length, bailout, &length);
1.3868 +
1.3869 + if (bailout->req() > 1) {
1.3870 + PreserveJVMState pjvms(this);
1.3871 + set_control(_gvn.transform(bailout));
1.3872 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.3873 + Deoptimization::Action_maybe_recompile);
1.3874 + }
1.3875 +
1.3876 + if (!stopped()) {
1.3877 + // How many elements will we copy from the original?
1.3878 + // The answer is MinI(orig_length - start, length).
1.3879 + Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
1.3880 + Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
1.3881 +
1.3882 + newcopy = new_array(klass_node, length, 0); // no argments to push
1.3883 +
1.3884 + // Generate a direct call to the right arraycopy function(s).
1.3885 + // We know the copy is disjoint but we might not know if the
1.3886 + // oop stores need checking.
1.3887 + // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
1.3888 + // This will fail a store-check if x contains any non-nulls.
1.3889 + bool disjoint_bases = true;
1.3890 + // if start > orig_length then the length of the copy may be
1.3891 + // negative.
1.3892 + bool length_never_negative = !is_copyOfRange;
1.3893 + generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
1.3894 + original, start, newcopy, intcon(0), moved,
1.3895 + disjoint_bases, length_never_negative);
1.3896 + }
1.3897 + } // original reexecute is set back here
1.3898 +
1.3899 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3900 + if (!stopped()) {
1.3901 + set_result(newcopy);
1.3902 + }
1.3903 + return true;
1.3904 +}
1.3905 +
1.3906 +
1.3907 +//----------------------generate_virtual_guard---------------------------
1.3908 +// Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
1.3909 +Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
1.3910 + RegionNode* slow_region) {
1.3911 + ciMethod* method = callee();
1.3912 + int vtable_index = method->vtable_index();
1.3913 + assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
1.3914 + err_msg_res("bad index %d", vtable_index));
1.3915 + // Get the Method* out of the appropriate vtable entry.
1.3916 + int entry_offset = (InstanceKlass::vtable_start_offset() +
1.3917 + vtable_index*vtableEntry::size()) * wordSize +
1.3918 + vtableEntry::method_offset_in_bytes();
1.3919 + Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
1.3920 + Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
1.3921 +
1.3922 + // Compare the target method with the expected method (e.g., Object.hashCode).
1.3923 + const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
1.3924 +
1.3925 + Node* native_call = makecon(native_call_addr);
1.3926 + Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
1.3927 + Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
1.3928 +
1.3929 + return generate_slow_guard(test_native, slow_region);
1.3930 +}
1.3931 +
1.3932 +//-----------------------generate_method_call----------------------------
1.3933 +// Use generate_method_call to make a slow-call to the real
1.3934 +// method if the fast path fails. An alternative would be to
1.3935 +// use a stub like OptoRuntime::slow_arraycopy_Java.
1.3936 +// This only works for expanding the current library call,
1.3937 +// not another intrinsic. (E.g., don't use this for making an
1.3938 +// arraycopy call inside of the copyOf intrinsic.)
1.3939 +CallJavaNode*
1.3940 +LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
1.3941 + // When compiling the intrinsic method itself, do not use this technique.
1.3942 + guarantee(callee() != C->method(), "cannot make slow-call to self");
1.3943 +
1.3944 + ciMethod* method = callee();
1.3945 + // ensure the JVMS we have will be correct for this call
1.3946 + guarantee(method_id == method->intrinsic_id(), "must match");
1.3947 +
1.3948 + const TypeFunc* tf = TypeFunc::make(method);
1.3949 + CallJavaNode* slow_call;
1.3950 + if (is_static) {
1.3951 + assert(!is_virtual, "");
1.3952 + slow_call = new(C) CallStaticJavaNode(C, tf,
1.3953 + SharedRuntime::get_resolve_static_call_stub(),
1.3954 + method, bci());
1.3955 + } else if (is_virtual) {
1.3956 + null_check_receiver();
1.3957 + int vtable_index = Method::invalid_vtable_index;
1.3958 + if (UseInlineCaches) {
1.3959 + // Suppress the vtable call
1.3960 + } else {
1.3961 + // hashCode and clone are not a miranda methods,
1.3962 + // so the vtable index is fixed.
1.3963 + // No need to use the linkResolver to get it.
1.3964 + vtable_index = method->vtable_index();
1.3965 + assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
1.3966 + err_msg_res("bad index %d", vtable_index));
1.3967 + }
1.3968 + slow_call = new(C) CallDynamicJavaNode(tf,
1.3969 + SharedRuntime::get_resolve_virtual_call_stub(),
1.3970 + method, vtable_index, bci());
1.3971 + } else { // neither virtual nor static: opt_virtual
1.3972 + null_check_receiver();
1.3973 + slow_call = new(C) CallStaticJavaNode(C, tf,
1.3974 + SharedRuntime::get_resolve_opt_virtual_call_stub(),
1.3975 + method, bci());
1.3976 + slow_call->set_optimized_virtual(true);
1.3977 + }
1.3978 + set_arguments_for_java_call(slow_call);
1.3979 + set_edges_for_java_call(slow_call);
1.3980 + return slow_call;
1.3981 +}
1.3982 +
1.3983 +
1.3984 +/**
1.3985 + * Build special case code for calls to hashCode on an object. This call may
1.3986 + * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
1.3987 + * slightly different code.
1.3988 + */
1.3989 +bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
1.3990 + assert(is_static == callee()->is_static(), "correct intrinsic selection");
1.3991 + assert(!(is_virtual && is_static), "either virtual, special, or static");
1.3992 +
1.3993 + enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
1.3994 +
1.3995 + RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
1.3996 + PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
1.3997 + PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
1.3998 + PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
1.3999 + Node* obj = NULL;
1.4000 + if (!is_static) {
1.4001 + // Check for hashing null object
1.4002 + obj = null_check_receiver();
1.4003 + if (stopped()) return true; // unconditionally null
1.4004 + result_reg->init_req(_null_path, top());
1.4005 + result_val->init_req(_null_path, top());
1.4006 + } else {
1.4007 + // Do a null check, and return zero if null.
1.4008 + // System.identityHashCode(null) == 0
1.4009 + obj = argument(0);
1.4010 + Node* null_ctl = top();
1.4011 + obj = null_check_oop(obj, &null_ctl);
1.4012 + result_reg->init_req(_null_path, null_ctl);
1.4013 + result_val->init_req(_null_path, _gvn.intcon(0));
1.4014 + }
1.4015 +
1.4016 + // Unconditionally null? Then return right away.
1.4017 + if (stopped()) {
1.4018 + set_control( result_reg->in(_null_path));
1.4019 + if (!stopped())
1.4020 + set_result(result_val->in(_null_path));
1.4021 + return true;
1.4022 + }
1.4023 +
1.4024 + // We only go to the fast case code if we pass a number of guards. The
1.4025 + // paths which do not pass are accumulated in the slow_region.
1.4026 + RegionNode* slow_region = new (C) RegionNode(1);
1.4027 + record_for_igvn(slow_region);
1.4028 +
1.4029 + // If this is a virtual call, we generate a funny guard. We pull out
1.4030 + // the vtable entry corresponding to hashCode() from the target object.
1.4031 + // If the target method which we are calling happens to be the native
1.4032 + // Object hashCode() method, we pass the guard. We do not need this
1.4033 + // guard for non-virtual calls -- the caller is known to be the native
1.4034 + // Object hashCode().
1.4035 + if (is_virtual) {
1.4036 + // After null check, get the object's klass.
1.4037 + Node* obj_klass = load_object_klass(obj);
1.4038 + generate_virtual_guard(obj_klass, slow_region);
1.4039 + }
1.4040 +
1.4041 + // Get the header out of the object, use LoadMarkNode when available
1.4042 + Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
1.4043 + // The control of the load must be NULL. Otherwise, the load can move before
1.4044 + // the null check after castPP removal.
1.4045 + Node* no_ctrl = NULL;
1.4046 + Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
1.4047 +
1.4048 + // Test the header to see if it is unlocked.
1.4049 + Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
1.4050 + Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
1.4051 + Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
1.4052 + Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
1.4053 + Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
1.4054 +
1.4055 + generate_slow_guard(test_unlocked, slow_region);
1.4056 +
1.4057 + // Get the hash value and check to see that it has been properly assigned.
1.4058 + // We depend on hash_mask being at most 32 bits and avoid the use of
1.4059 + // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
1.4060 + // vm: see markOop.hpp.
1.4061 + Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
1.4062 + Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
1.4063 + Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
1.4064 + // This hack lets the hash bits live anywhere in the mark object now, as long
1.4065 + // as the shift drops the relevant bits into the low 32 bits. Note that
1.4066 + // Java spec says that HashCode is an int so there's no point in capturing
1.4067 + // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
1.4068 + hshifted_header = ConvX2I(hshifted_header);
1.4069 + Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
1.4070 +
1.4071 + Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
1.4072 + Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
1.4073 + Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
1.4074 +
1.4075 + generate_slow_guard(test_assigned, slow_region);
1.4076 +
1.4077 + Node* init_mem = reset_memory();
1.4078 + // fill in the rest of the null path:
1.4079 + result_io ->init_req(_null_path, i_o());
1.4080 + result_mem->init_req(_null_path, init_mem);
1.4081 +
1.4082 + result_val->init_req(_fast_path, hash_val);
1.4083 + result_reg->init_req(_fast_path, control());
1.4084 + result_io ->init_req(_fast_path, i_o());
1.4085 + result_mem->init_req(_fast_path, init_mem);
1.4086 +
1.4087 + // Generate code for the slow case. We make a call to hashCode().
1.4088 + set_control(_gvn.transform(slow_region));
1.4089 + if (!stopped()) {
1.4090 + // No need for PreserveJVMState, because we're using up the present state.
1.4091 + set_all_memory(init_mem);
1.4092 + vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
1.4093 + CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
1.4094 + Node* slow_result = set_results_for_java_call(slow_call);
1.4095 + // this->control() comes from set_results_for_java_call
1.4096 + result_reg->init_req(_slow_path, control());
1.4097 + result_val->init_req(_slow_path, slow_result);
1.4098 + result_io ->set_req(_slow_path, i_o());
1.4099 + result_mem ->set_req(_slow_path, reset_memory());
1.4100 + }
1.4101 +
1.4102 + // Return the combined state.
1.4103 + set_i_o( _gvn.transform(result_io) );
1.4104 + set_all_memory( _gvn.transform(result_mem));
1.4105 +
1.4106 + set_result(result_reg, result_val);
1.4107 + return true;
1.4108 +}
1.4109 +
1.4110 +//---------------------------inline_native_getClass----------------------------
1.4111 +// public final native Class<?> java.lang.Object.getClass();
1.4112 +//
1.4113 +// Build special case code for calls to getClass on an object.
1.4114 +bool LibraryCallKit::inline_native_getClass() {
1.4115 + Node* obj = null_check_receiver();
1.4116 + if (stopped()) return true;
1.4117 + set_result(load_mirror_from_klass(load_object_klass(obj)));
1.4118 + return true;
1.4119 +}
1.4120 +
1.4121 +//-----------------inline_native_Reflection_getCallerClass---------------------
1.4122 +// public static native Class<?> sun.reflect.Reflection.getCallerClass();
1.4123 +//
1.4124 +// In the presence of deep enough inlining, getCallerClass() becomes a no-op.
1.4125 +//
1.4126 +// NOTE: This code must perform the same logic as JVM_GetCallerClass
1.4127 +// in that it must skip particular security frames and checks for
1.4128 +// caller sensitive methods.
1.4129 +bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
1.4130 +#ifndef PRODUCT
1.4131 + if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
1.4132 + tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
1.4133 + }
1.4134 +#endif
1.4135 +
1.4136 + if (!jvms()->has_method()) {
1.4137 +#ifndef PRODUCT
1.4138 + if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
1.4139 + tty->print_cr(" Bailing out because intrinsic was inlined at top level");
1.4140 + }
1.4141 +#endif
1.4142 + return false;
1.4143 + }
1.4144 +
1.4145 + // Walk back up the JVM state to find the caller at the required
1.4146 + // depth.
1.4147 + JVMState* caller_jvms = jvms();
1.4148 +
1.4149 + // Cf. JVM_GetCallerClass
1.4150 + // NOTE: Start the loop at depth 1 because the current JVM state does
1.4151 + // not include the Reflection.getCallerClass() frame.
1.4152 + for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
1.4153 + ciMethod* m = caller_jvms->method();
1.4154 + switch (n) {
1.4155 + case 0:
1.4156 + fatal("current JVM state does not include the Reflection.getCallerClass frame");
1.4157 + break;
1.4158 + case 1:
1.4159 + // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
1.4160 + if (!m->caller_sensitive()) {
1.4161 +#ifndef PRODUCT
1.4162 + if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
1.4163 + tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
1.4164 + }
1.4165 +#endif
1.4166 + return false; // bail-out; let JVM_GetCallerClass do the work
1.4167 + }
1.4168 + break;
1.4169 + default:
1.4170 + if (!m->is_ignored_by_security_stack_walk()) {
1.4171 + // We have reached the desired frame; return the holder class.
1.4172 + // Acquire method holder as java.lang.Class and push as constant.
1.4173 + ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
1.4174 + ciInstance* caller_mirror = caller_klass->java_mirror();
1.4175 + set_result(makecon(TypeInstPtr::make(caller_mirror)));
1.4176 +
1.4177 +#ifndef PRODUCT
1.4178 + if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
1.4179 + 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());
1.4180 + tty->print_cr(" JVM state at this point:");
1.4181 + for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
1.4182 + ciMethod* m = jvms()->of_depth(i)->method();
1.4183 + tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
1.4184 + }
1.4185 + }
1.4186 +#endif
1.4187 + return true;
1.4188 + }
1.4189 + break;
1.4190 + }
1.4191 + }
1.4192 +
1.4193 +#ifndef PRODUCT
1.4194 + if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
1.4195 + tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
1.4196 + tty->print_cr(" JVM state at this point:");
1.4197 + for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
1.4198 + ciMethod* m = jvms()->of_depth(i)->method();
1.4199 + tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
1.4200 + }
1.4201 + }
1.4202 +#endif
1.4203 +
1.4204 + return false; // bail-out; let JVM_GetCallerClass do the work
1.4205 +}
1.4206 +
1.4207 +bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
1.4208 + Node* arg = argument(0);
1.4209 + Node* result;
1.4210 +
1.4211 + switch (id) {
1.4212 + case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
1.4213 + case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
1.4214 + case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
1.4215 + case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
1.4216 +
1.4217 + case vmIntrinsics::_doubleToLongBits: {
1.4218 + // two paths (plus control) merge in a wood
1.4219 + RegionNode *r = new (C) RegionNode(3);
1.4220 + Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1.4221 +
1.4222 + Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
1.4223 + // Build the boolean node
1.4224 + Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
1.4225 +
1.4226 + // Branch either way.
1.4227 + // NaN case is less traveled, which makes all the difference.
1.4228 + IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1.4229 + Node *opt_isnan = _gvn.transform(ifisnan);
1.4230 + assert( opt_isnan->is_If(), "Expect an IfNode");
1.4231 + IfNode *opt_ifisnan = (IfNode*)opt_isnan;
1.4232 + Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
1.4233 +
1.4234 + set_control(iftrue);
1.4235 +
1.4236 + static const jlong nan_bits = CONST64(0x7ff8000000000000);
1.4237 + Node *slow_result = longcon(nan_bits); // return NaN
1.4238 + phi->init_req(1, _gvn.transform( slow_result ));
1.4239 + r->init_req(1, iftrue);
1.4240 +
1.4241 + // Else fall through
1.4242 + Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
1.4243 + set_control(iffalse);
1.4244 +
1.4245 + phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
1.4246 + r->init_req(2, iffalse);
1.4247 +
1.4248 + // Post merge
1.4249 + set_control(_gvn.transform(r));
1.4250 + record_for_igvn(r);
1.4251 +
1.4252 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.4253 + result = phi;
1.4254 + assert(result->bottom_type()->isa_long(), "must be");
1.4255 + break;
1.4256 + }
1.4257 +
1.4258 + case vmIntrinsics::_floatToIntBits: {
1.4259 + // two paths (plus control) merge in a wood
1.4260 + RegionNode *r = new (C) RegionNode(3);
1.4261 + Node *phi = new (C) PhiNode(r, TypeInt::INT);
1.4262 +
1.4263 + Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
1.4264 + // Build the boolean node
1.4265 + Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
1.4266 +
1.4267 + // Branch either way.
1.4268 + // NaN case is less traveled, which makes all the difference.
1.4269 + IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1.4270 + Node *opt_isnan = _gvn.transform(ifisnan);
1.4271 + assert( opt_isnan->is_If(), "Expect an IfNode");
1.4272 + IfNode *opt_ifisnan = (IfNode*)opt_isnan;
1.4273 + Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
1.4274 +
1.4275 + set_control(iftrue);
1.4276 +
1.4277 + static const jint nan_bits = 0x7fc00000;
1.4278 + Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
1.4279 + phi->init_req(1, _gvn.transform( slow_result ));
1.4280 + r->init_req(1, iftrue);
1.4281 +
1.4282 + // Else fall through
1.4283 + Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
1.4284 + set_control(iffalse);
1.4285 +
1.4286 + phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
1.4287 + r->init_req(2, iffalse);
1.4288 +
1.4289 + // Post merge
1.4290 + set_control(_gvn.transform(r));
1.4291 + record_for_igvn(r);
1.4292 +
1.4293 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.4294 + result = phi;
1.4295 + assert(result->bottom_type()->isa_int(), "must be");
1.4296 + break;
1.4297 + }
1.4298 +
1.4299 + default:
1.4300 + fatal_unexpected_iid(id);
1.4301 + break;
1.4302 + }
1.4303 + set_result(_gvn.transform(result));
1.4304 + return true;
1.4305 +}
1.4306 +
1.4307 +#ifdef _LP64
1.4308 +#define XTOP ,top() /*additional argument*/
1.4309 +#else //_LP64
1.4310 +#define XTOP /*no additional argument*/
1.4311 +#endif //_LP64
1.4312 +
1.4313 +//----------------------inline_unsafe_copyMemory-------------------------
1.4314 +// public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
1.4315 +bool LibraryCallKit::inline_unsafe_copyMemory() {
1.4316 + if (callee()->is_static()) return false; // caller must have the capability!
1.4317 + null_check_receiver(); // null-check receiver
1.4318 + if (stopped()) return true;
1.4319 +
1.4320 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.4321 +
1.4322 + Node* src_ptr = argument(1); // type: oop
1.4323 + Node* src_off = ConvL2X(argument(2)); // type: long
1.4324 + Node* dst_ptr = argument(4); // type: oop
1.4325 + Node* dst_off = ConvL2X(argument(5)); // type: long
1.4326 + Node* size = ConvL2X(argument(7)); // type: long
1.4327 +
1.4328 + assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1.4329 + "fieldOffset must be byte-scaled");
1.4330 +
1.4331 + Node* src = make_unsafe_address(src_ptr, src_off);
1.4332 + Node* dst = make_unsafe_address(dst_ptr, dst_off);
1.4333 +
1.4334 + // Conservatively insert a memory barrier on all memory slices.
1.4335 + // Do not let writes of the copy source or destination float below the copy.
1.4336 + insert_mem_bar(Op_MemBarCPUOrder);
1.4337 +
1.4338 + // Call it. Note that the length argument is not scaled.
1.4339 + make_runtime_call(RC_LEAF|RC_NO_FP,
1.4340 + OptoRuntime::fast_arraycopy_Type(),
1.4341 + StubRoutines::unsafe_arraycopy(),
1.4342 + "unsafe_arraycopy",
1.4343 + TypeRawPtr::BOTTOM,
1.4344 + src, dst, size XTOP);
1.4345 +
1.4346 + // Do not let reads of the copy destination float above the copy.
1.4347 + insert_mem_bar(Op_MemBarCPUOrder);
1.4348 +
1.4349 + return true;
1.4350 +}
1.4351 +
1.4352 +//------------------------clone_coping-----------------------------------
1.4353 +// Helper function for inline_native_clone.
1.4354 +void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
1.4355 + assert(obj_size != NULL, "");
1.4356 + Node* raw_obj = alloc_obj->in(1);
1.4357 + assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
1.4358 +
1.4359 + AllocateNode* alloc = NULL;
1.4360 + if (ReduceBulkZeroing) {
1.4361 + // We will be completely responsible for initializing this object -
1.4362 + // mark Initialize node as complete.
1.4363 + alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
1.4364 + // The object was just allocated - there should be no any stores!
1.4365 + guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
1.4366 + // Mark as complete_with_arraycopy so that on AllocateNode
1.4367 + // expansion, we know this AllocateNode is initialized by an array
1.4368 + // copy and a StoreStore barrier exists after the array copy.
1.4369 + alloc->initialization()->set_complete_with_arraycopy();
1.4370 + }
1.4371 +
1.4372 + // Copy the fastest available way.
1.4373 + // TODO: generate fields copies for small objects instead.
1.4374 + Node* src = obj;
1.4375 + Node* dest = alloc_obj;
1.4376 + Node* size = _gvn.transform(obj_size);
1.4377 +
1.4378 + // Exclude the header but include array length to copy by 8 bytes words.
1.4379 + // Can't use base_offset_in_bytes(bt) since basic type is unknown.
1.4380 + int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
1.4381 + instanceOopDesc::base_offset_in_bytes();
1.4382 + // base_off:
1.4383 + // 8 - 32-bit VM
1.4384 + // 12 - 64-bit VM, compressed klass
1.4385 + // 16 - 64-bit VM, normal klass
1.4386 + if (base_off % BytesPerLong != 0) {
1.4387 + assert(UseCompressedClassPointers, "");
1.4388 + if (is_array) {
1.4389 + // Exclude length to copy by 8 bytes words.
1.4390 + base_off += sizeof(int);
1.4391 + } else {
1.4392 + // Include klass to copy by 8 bytes words.
1.4393 + base_off = instanceOopDesc::klass_offset_in_bytes();
1.4394 + }
1.4395 + assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
1.4396 + }
1.4397 + src = basic_plus_adr(src, base_off);
1.4398 + dest = basic_plus_adr(dest, base_off);
1.4399 +
1.4400 + // Compute the length also, if needed:
1.4401 + Node* countx = size;
1.4402 + countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
1.4403 + countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
1.4404 +
1.4405 + const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1.4406 + bool disjoint_bases = true;
1.4407 + generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
1.4408 + src, NULL, dest, NULL, countx,
1.4409 + /*dest_uninitialized*/true);
1.4410 +
1.4411 + // If necessary, emit some card marks afterwards. (Non-arrays only.)
1.4412 + if (card_mark) {
1.4413 + assert(!is_array, "");
1.4414 + // Put in store barrier for any and all oops we are sticking
1.4415 + // into this object. (We could avoid this if we could prove
1.4416 + // that the object type contains no oop fields at all.)
1.4417 + Node* no_particular_value = NULL;
1.4418 + Node* no_particular_field = NULL;
1.4419 + int raw_adr_idx = Compile::AliasIdxRaw;
1.4420 + post_barrier(control(),
1.4421 + memory(raw_adr_type),
1.4422 + alloc_obj,
1.4423 + no_particular_field,
1.4424 + raw_adr_idx,
1.4425 + no_particular_value,
1.4426 + T_OBJECT,
1.4427 + false);
1.4428 + }
1.4429 +
1.4430 + // Do not let reads from the cloned object float above the arraycopy.
1.4431 + if (alloc != NULL) {
1.4432 + // Do not let stores that initialize this object be reordered with
1.4433 + // a subsequent store that would make this object accessible by
1.4434 + // other threads.
1.4435 + // Record what AllocateNode this StoreStore protects so that
1.4436 + // escape analysis can go from the MemBarStoreStoreNode to the
1.4437 + // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1.4438 + // based on the escape status of the AllocateNode.
1.4439 + insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1.4440 + } else {
1.4441 + insert_mem_bar(Op_MemBarCPUOrder);
1.4442 + }
1.4443 +}
1.4444 +
1.4445 +//------------------------inline_native_clone----------------------------
1.4446 +// protected native Object java.lang.Object.clone();
1.4447 +//
1.4448 +// Here are the simple edge cases:
1.4449 +// null receiver => normal trap
1.4450 +// virtual and clone was overridden => slow path to out-of-line clone
1.4451 +// not cloneable or finalizer => slow path to out-of-line Object.clone
1.4452 +//
1.4453 +// The general case has two steps, allocation and copying.
1.4454 +// Allocation has two cases, and uses GraphKit::new_instance or new_array.
1.4455 +//
1.4456 +// Copying also has two cases, oop arrays and everything else.
1.4457 +// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
1.4458 +// Everything else uses the tight inline loop supplied by CopyArrayNode.
1.4459 +//
1.4460 +// These steps fold up nicely if and when the cloned object's klass
1.4461 +// can be sharply typed as an object array, a type array, or an instance.
1.4462 +//
1.4463 +bool LibraryCallKit::inline_native_clone(bool is_virtual) {
1.4464 + PhiNode* result_val;
1.4465 +
1.4466 + // Set the reexecute bit for the interpreter to reexecute
1.4467 + // the bytecode that invokes Object.clone if deoptimization happens.
1.4468 + { PreserveReexecuteState preexecs(this);
1.4469 + jvms()->set_should_reexecute(true);
1.4470 +
1.4471 + Node* obj = null_check_receiver();
1.4472 + if (stopped()) return true;
1.4473 +
1.4474 + Node* obj_klass = load_object_klass(obj);
1.4475 + const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
1.4476 + const TypeOopPtr* toop = ((tklass != NULL)
1.4477 + ? tklass->as_instance_type()
1.4478 + : TypeInstPtr::NOTNULL);
1.4479 +
1.4480 + // Conservatively insert a memory barrier on all memory slices.
1.4481 + // Do not let writes into the original float below the clone.
1.4482 + insert_mem_bar(Op_MemBarCPUOrder);
1.4483 +
1.4484 + // paths into result_reg:
1.4485 + enum {
1.4486 + _slow_path = 1, // out-of-line call to clone method (virtual or not)
1.4487 + _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
1.4488 + _array_path, // plain array allocation, plus arrayof_long_arraycopy
1.4489 + _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
1.4490 + PATH_LIMIT
1.4491 + };
1.4492 + RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
1.4493 + result_val = new(C) PhiNode(result_reg,
1.4494 + TypeInstPtr::NOTNULL);
1.4495 + PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
1.4496 + PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
1.4497 + TypePtr::BOTTOM);
1.4498 + record_for_igvn(result_reg);
1.4499 +
1.4500 + const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1.4501 + int raw_adr_idx = Compile::AliasIdxRaw;
1.4502 +
1.4503 + Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
1.4504 + if (array_ctl != NULL) {
1.4505 + // It's an array.
1.4506 + PreserveJVMState pjvms(this);
1.4507 + set_control(array_ctl);
1.4508 + Node* obj_length = load_array_length(obj);
1.4509 + Node* obj_size = NULL;
1.4510 + Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
1.4511 +
1.4512 + if (!use_ReduceInitialCardMarks()) {
1.4513 + // If it is an oop array, it requires very special treatment,
1.4514 + // because card marking is required on each card of the array.
1.4515 + Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
1.4516 + if (is_obja != NULL) {
1.4517 + PreserveJVMState pjvms2(this);
1.4518 + set_control(is_obja);
1.4519 + // Generate a direct call to the right arraycopy function(s).
1.4520 + bool disjoint_bases = true;
1.4521 + bool length_never_negative = true;
1.4522 + generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
1.4523 + obj, intcon(0), alloc_obj, intcon(0),
1.4524 + obj_length,
1.4525 + disjoint_bases, length_never_negative);
1.4526 + result_reg->init_req(_objArray_path, control());
1.4527 + result_val->init_req(_objArray_path, alloc_obj);
1.4528 + result_i_o ->set_req(_objArray_path, i_o());
1.4529 + result_mem ->set_req(_objArray_path, reset_memory());
1.4530 + }
1.4531 + }
1.4532 + // Otherwise, there are no card marks to worry about.
1.4533 + // (We can dispense with card marks if we know the allocation
1.4534 + // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
1.4535 + // causes the non-eden paths to take compensating steps to
1.4536 + // simulate a fresh allocation, so that no further
1.4537 + // card marks are required in compiled code to initialize
1.4538 + // the object.)
1.4539 +
1.4540 + if (!stopped()) {
1.4541 + copy_to_clone(obj, alloc_obj, obj_size, true, false);
1.4542 +
1.4543 + // Present the results of the copy.
1.4544 + result_reg->init_req(_array_path, control());
1.4545 + result_val->init_req(_array_path, alloc_obj);
1.4546 + result_i_o ->set_req(_array_path, i_o());
1.4547 + result_mem ->set_req(_array_path, reset_memory());
1.4548 + }
1.4549 + }
1.4550 +
1.4551 + // We only go to the instance fast case code if we pass a number of guards.
1.4552 + // The paths which do not pass are accumulated in the slow_region.
1.4553 + RegionNode* slow_region = new (C) RegionNode(1);
1.4554 + record_for_igvn(slow_region);
1.4555 + if (!stopped()) {
1.4556 + // It's an instance (we did array above). Make the slow-path tests.
1.4557 + // If this is a virtual call, we generate a funny guard. We grab
1.4558 + // the vtable entry corresponding to clone() from the target object.
1.4559 + // If the target method which we are calling happens to be the
1.4560 + // Object clone() method, we pass the guard. We do not need this
1.4561 + // guard for non-virtual calls; the caller is known to be the native
1.4562 + // Object clone().
1.4563 + if (is_virtual) {
1.4564 + generate_virtual_guard(obj_klass, slow_region);
1.4565 + }
1.4566 +
1.4567 + // The object must be cloneable and must not have a finalizer.
1.4568 + // Both of these conditions may be checked in a single test.
1.4569 + // We could optimize the cloneable test further, but we don't care.
1.4570 + generate_access_flags_guard(obj_klass,
1.4571 + // Test both conditions:
1.4572 + JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
1.4573 + // Must be cloneable but not finalizer:
1.4574 + JVM_ACC_IS_CLONEABLE,
1.4575 + slow_region);
1.4576 + }
1.4577 +
1.4578 + if (!stopped()) {
1.4579 + // It's an instance, and it passed the slow-path tests.
1.4580 + PreserveJVMState pjvms(this);
1.4581 + Node* obj_size = NULL;
1.4582 + // Need to deoptimize on exception from allocation since Object.clone intrinsic
1.4583 + // is reexecuted if deoptimization occurs and there could be problems when merging
1.4584 + // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
1.4585 + Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
1.4586 +
1.4587 + copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
1.4588 +
1.4589 + // Present the results of the slow call.
1.4590 + result_reg->init_req(_instance_path, control());
1.4591 + result_val->init_req(_instance_path, alloc_obj);
1.4592 + result_i_o ->set_req(_instance_path, i_o());
1.4593 + result_mem ->set_req(_instance_path, reset_memory());
1.4594 + }
1.4595 +
1.4596 + // Generate code for the slow case. We make a call to clone().
1.4597 + set_control(_gvn.transform(slow_region));
1.4598 + if (!stopped()) {
1.4599 + PreserveJVMState pjvms(this);
1.4600 + CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
1.4601 + Node* slow_result = set_results_for_java_call(slow_call);
1.4602 + // this->control() comes from set_results_for_java_call
1.4603 + result_reg->init_req(_slow_path, control());
1.4604 + result_val->init_req(_slow_path, slow_result);
1.4605 + result_i_o ->set_req(_slow_path, i_o());
1.4606 + result_mem ->set_req(_slow_path, reset_memory());
1.4607 + }
1.4608 +
1.4609 + // Return the combined state.
1.4610 + set_control( _gvn.transform(result_reg));
1.4611 + set_i_o( _gvn.transform(result_i_o));
1.4612 + set_all_memory( _gvn.transform(result_mem));
1.4613 + } // original reexecute is set back here
1.4614 +
1.4615 + set_result(_gvn.transform(result_val));
1.4616 + return true;
1.4617 +}
1.4618 +
1.4619 +//------------------------------basictype2arraycopy----------------------------
1.4620 +address LibraryCallKit::basictype2arraycopy(BasicType t,
1.4621 + Node* src_offset,
1.4622 + Node* dest_offset,
1.4623 + bool disjoint_bases,
1.4624 + const char* &name,
1.4625 + bool dest_uninitialized) {
1.4626 + const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
1.4627 + const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
1.4628 +
1.4629 + bool aligned = false;
1.4630 + bool disjoint = disjoint_bases;
1.4631 +
1.4632 + // if the offsets are the same, we can treat the memory regions as
1.4633 + // disjoint, because either the memory regions are in different arrays,
1.4634 + // or they are identical (which we can treat as disjoint.) We can also
1.4635 + // treat a copy with a destination index less that the source index
1.4636 + // as disjoint since a low->high copy will work correctly in this case.
1.4637 + if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
1.4638 + dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
1.4639 + // both indices are constants
1.4640 + int s_offs = src_offset_inttype->get_con();
1.4641 + int d_offs = dest_offset_inttype->get_con();
1.4642 + int element_size = type2aelembytes(t);
1.4643 + aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
1.4644 + ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
1.4645 + if (s_offs >= d_offs) disjoint = true;
1.4646 + } else if (src_offset == dest_offset && src_offset != NULL) {
1.4647 + // This can occur if the offsets are identical non-constants.
1.4648 + disjoint = true;
1.4649 + }
1.4650 +
1.4651 + return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
1.4652 +}
1.4653 +
1.4654 +
1.4655 +//------------------------------inline_arraycopy-----------------------
1.4656 +// public static native void java.lang.System.arraycopy(Object src, int srcPos,
1.4657 +// Object dest, int destPos,
1.4658 +// int length);
1.4659 +bool LibraryCallKit::inline_arraycopy() {
1.4660 + // Get the arguments.
1.4661 + Node* src = argument(0); // type: oop
1.4662 + Node* src_offset = argument(1); // type: int
1.4663 + Node* dest = argument(2); // type: oop
1.4664 + Node* dest_offset = argument(3); // type: int
1.4665 + Node* length = argument(4); // type: int
1.4666 +
1.4667 + // Compile time checks. If any of these checks cannot be verified at compile time,
1.4668 + // we do not make a fast path for this call. Instead, we let the call remain as it
1.4669 + // is. The checks we choose to mandate at compile time are:
1.4670 + //
1.4671 + // (1) src and dest are arrays.
1.4672 + const Type* src_type = src->Value(&_gvn);
1.4673 + const Type* dest_type = dest->Value(&_gvn);
1.4674 + const TypeAryPtr* top_src = src_type->isa_aryptr();
1.4675 + const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1.4676 +
1.4677 + // Do we have the type of src?
1.4678 + bool has_src = (top_src != NULL && top_src->klass() != NULL);
1.4679 + // Do we have the type of dest?
1.4680 + bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
1.4681 + // Is the type for src from speculation?
1.4682 + bool src_spec = false;
1.4683 + // Is the type for dest from speculation?
1.4684 + bool dest_spec = false;
1.4685 +
1.4686 + if (!has_src || !has_dest) {
1.4687 + // We don't have sufficient type information, let's see if
1.4688 + // speculative types can help. We need to have types for both src
1.4689 + // and dest so that it pays off.
1.4690 +
1.4691 + // Do we already have or could we have type information for src
1.4692 + bool could_have_src = has_src;
1.4693 + // Do we already have or could we have type information for dest
1.4694 + bool could_have_dest = has_dest;
1.4695 +
1.4696 + ciKlass* src_k = NULL;
1.4697 + if (!has_src) {
1.4698 + src_k = src_type->speculative_type();
1.4699 + if (src_k != NULL && src_k->is_array_klass()) {
1.4700 + could_have_src = true;
1.4701 + }
1.4702 + }
1.4703 +
1.4704 + ciKlass* dest_k = NULL;
1.4705 + if (!has_dest) {
1.4706 + dest_k = dest_type->speculative_type();
1.4707 + if (dest_k != NULL && dest_k->is_array_klass()) {
1.4708 + could_have_dest = true;
1.4709 + }
1.4710 + }
1.4711 +
1.4712 + if (could_have_src && could_have_dest) {
1.4713 + // This is going to pay off so emit the required guards
1.4714 + if (!has_src) {
1.4715 + src = maybe_cast_profiled_obj(src, src_k);
1.4716 + src_type = _gvn.type(src);
1.4717 + top_src = src_type->isa_aryptr();
1.4718 + has_src = (top_src != NULL && top_src->klass() != NULL);
1.4719 + src_spec = true;
1.4720 + }
1.4721 + if (!has_dest) {
1.4722 + dest = maybe_cast_profiled_obj(dest, dest_k);
1.4723 + dest_type = _gvn.type(dest);
1.4724 + top_dest = dest_type->isa_aryptr();
1.4725 + has_dest = (top_dest != NULL && top_dest->klass() != NULL);
1.4726 + dest_spec = true;
1.4727 + }
1.4728 + }
1.4729 + }
1.4730 +
1.4731 + if (!has_src || !has_dest) {
1.4732 + // Conservatively insert a memory barrier on all memory slices.
1.4733 + // Do not let writes into the source float below the arraycopy.
1.4734 + insert_mem_bar(Op_MemBarCPUOrder);
1.4735 +
1.4736 + // Call StubRoutines::generic_arraycopy stub.
1.4737 + generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
1.4738 + src, src_offset, dest, dest_offset, length);
1.4739 +
1.4740 + // Do not let reads from the destination float above the arraycopy.
1.4741 + // Since we cannot type the arrays, we don't know which slices
1.4742 + // might be affected. We could restrict this barrier only to those
1.4743 + // memory slices which pertain to array elements--but don't bother.
1.4744 + if (!InsertMemBarAfterArraycopy)
1.4745 + // (If InsertMemBarAfterArraycopy, there is already one in place.)
1.4746 + insert_mem_bar(Op_MemBarCPUOrder);
1.4747 + return true;
1.4748 + }
1.4749 +
1.4750 + // (2) src and dest arrays must have elements of the same BasicType
1.4751 + // Figure out the size and type of the elements we will be copying.
1.4752 + BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
1.4753 + BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1.4754 + if (src_elem == T_ARRAY) src_elem = T_OBJECT;
1.4755 + if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
1.4756 +
1.4757 + if (src_elem != dest_elem || dest_elem == T_VOID) {
1.4758 + // The component types are not the same or are not recognized. Punt.
1.4759 + // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1.4760 + generate_slow_arraycopy(TypePtr::BOTTOM,
1.4761 + src, src_offset, dest, dest_offset, length,
1.4762 + /*dest_uninitialized*/false);
1.4763 + return true;
1.4764 + }
1.4765 +
1.4766 + if (src_elem == T_OBJECT) {
1.4767 + // If both arrays are object arrays then having the exact types
1.4768 + // for both will remove the need for a subtype check at runtime
1.4769 + // before the call and may make it possible to pick a faster copy
1.4770 + // routine (without a subtype check on every element)
1.4771 + // Do we have the exact type of src?
1.4772 + bool could_have_src = src_spec;
1.4773 + // Do we have the exact type of dest?
1.4774 + bool could_have_dest = dest_spec;
1.4775 + ciKlass* src_k = top_src->klass();
1.4776 + ciKlass* dest_k = top_dest->klass();
1.4777 + if (!src_spec) {
1.4778 + src_k = src_type->speculative_type();
1.4779 + if (src_k != NULL && src_k->is_array_klass()) {
1.4780 + could_have_src = true;
1.4781 + }
1.4782 + }
1.4783 + if (!dest_spec) {
1.4784 + dest_k = dest_type->speculative_type();
1.4785 + if (dest_k != NULL && dest_k->is_array_klass()) {
1.4786 + could_have_dest = true;
1.4787 + }
1.4788 + }
1.4789 + if (could_have_src && could_have_dest) {
1.4790 + // If we can have both exact types, emit the missing guards
1.4791 + if (could_have_src && !src_spec) {
1.4792 + src = maybe_cast_profiled_obj(src, src_k);
1.4793 + }
1.4794 + if (could_have_dest && !dest_spec) {
1.4795 + dest = maybe_cast_profiled_obj(dest, dest_k);
1.4796 + }
1.4797 + }
1.4798 + }
1.4799 +
1.4800 + //---------------------------------------------------------------------------
1.4801 + // We will make a fast path for this call to arraycopy.
1.4802 +
1.4803 + // We have the following tests left to perform:
1.4804 + //
1.4805 + // (3) src and dest must not be null.
1.4806 + // (4) src_offset must not be negative.
1.4807 + // (5) dest_offset must not be negative.
1.4808 + // (6) length must not be negative.
1.4809 + // (7) src_offset + length must not exceed length of src.
1.4810 + // (8) dest_offset + length must not exceed length of dest.
1.4811 + // (9) each element of an oop array must be assignable
1.4812 +
1.4813 + RegionNode* slow_region = new (C) RegionNode(1);
1.4814 + record_for_igvn(slow_region);
1.4815 +
1.4816 + // (3) operands must not be null
1.4817 + // We currently perform our null checks with the null_check routine.
1.4818 + // This means that the null exceptions will be reported in the caller
1.4819 + // rather than (correctly) reported inside of the native arraycopy call.
1.4820 + // This should be corrected, given time. We do our null check with the
1.4821 + // stack pointer restored.
1.4822 + src = null_check(src, T_ARRAY);
1.4823 + dest = null_check(dest, T_ARRAY);
1.4824 +
1.4825 + // (4) src_offset must not be negative.
1.4826 + generate_negative_guard(src_offset, slow_region);
1.4827 +
1.4828 + // (5) dest_offset must not be negative.
1.4829 + generate_negative_guard(dest_offset, slow_region);
1.4830 +
1.4831 + // (6) length must not be negative (moved to generate_arraycopy()).
1.4832 + // generate_negative_guard(length, slow_region);
1.4833 +
1.4834 + // (7) src_offset + length must not exceed length of src.
1.4835 + generate_limit_guard(src_offset, length,
1.4836 + load_array_length(src),
1.4837 + slow_region);
1.4838 +
1.4839 + // (8) dest_offset + length must not exceed length of dest.
1.4840 + generate_limit_guard(dest_offset, length,
1.4841 + load_array_length(dest),
1.4842 + slow_region);
1.4843 +
1.4844 + // (9) each element of an oop array must be assignable
1.4845 + // The generate_arraycopy subroutine checks this.
1.4846 +
1.4847 + // This is where the memory effects are placed:
1.4848 + const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1.4849 + generate_arraycopy(adr_type, dest_elem,
1.4850 + src, src_offset, dest, dest_offset, length,
1.4851 + false, false, slow_region);
1.4852 +
1.4853 + return true;
1.4854 +}
1.4855 +
1.4856 +//-----------------------------generate_arraycopy----------------------
1.4857 +// Generate an optimized call to arraycopy.
1.4858 +// Caller must guard against non-arrays.
1.4859 +// Caller must determine a common array basic-type for both arrays.
1.4860 +// Caller must validate offsets against array bounds.
1.4861 +// The slow_region has already collected guard failure paths
1.4862 +// (such as out of bounds length or non-conformable array types).
1.4863 +// The generated code has this shape, in general:
1.4864 +//
1.4865 +// if (length == 0) return // via zero_path
1.4866 +// slowval = -1
1.4867 +// if (types unknown) {
1.4868 +// slowval = call generic copy loop
1.4869 +// if (slowval == 0) return // via checked_path
1.4870 +// } else if (indexes in bounds) {
1.4871 +// if ((is object array) && !(array type check)) {
1.4872 +// slowval = call checked copy loop
1.4873 +// if (slowval == 0) return // via checked_path
1.4874 +// } else {
1.4875 +// call bulk copy loop
1.4876 +// return // via fast_path
1.4877 +// }
1.4878 +// }
1.4879 +// // adjust params for remaining work:
1.4880 +// if (slowval != -1) {
1.4881 +// n = -1^slowval; src_offset += n; dest_offset += n; length -= n
1.4882 +// }
1.4883 +// slow_region:
1.4884 +// call slow arraycopy(src, src_offset, dest, dest_offset, length)
1.4885 +// return // via slow_call_path
1.4886 +//
1.4887 +// This routine is used from several intrinsics: System.arraycopy,
1.4888 +// Object.clone (the array subcase), and Arrays.copyOf[Range].
1.4889 +//
1.4890 +void
1.4891 +LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
1.4892 + BasicType basic_elem_type,
1.4893 + Node* src, Node* src_offset,
1.4894 + Node* dest, Node* dest_offset,
1.4895 + Node* copy_length,
1.4896 + bool disjoint_bases,
1.4897 + bool length_never_negative,
1.4898 + RegionNode* slow_region) {
1.4899 +
1.4900 + if (slow_region == NULL) {
1.4901 + slow_region = new(C) RegionNode(1);
1.4902 + record_for_igvn(slow_region);
1.4903 + }
1.4904 +
1.4905 + Node* original_dest = dest;
1.4906 + AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
1.4907 + bool dest_uninitialized = false;
1.4908 +
1.4909 + // See if this is the initialization of a newly-allocated array.
1.4910 + // If so, we will take responsibility here for initializing it to zero.
1.4911 + // (Note: Because tightly_coupled_allocation performs checks on the
1.4912 + // out-edges of the dest, we need to avoid making derived pointers
1.4913 + // from it until we have checked its uses.)
1.4914 + if (ReduceBulkZeroing
1.4915 + && !ZeroTLAB // pointless if already zeroed
1.4916 + && basic_elem_type != T_CONFLICT // avoid corner case
1.4917 + && !src->eqv_uncast(dest)
1.4918 + && ((alloc = tightly_coupled_allocation(dest, slow_region))
1.4919 + != NULL)
1.4920 + && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
1.4921 + && alloc->maybe_set_complete(&_gvn)) {
1.4922 + // "You break it, you buy it."
1.4923 + InitializeNode* init = alloc->initialization();
1.4924 + assert(init->is_complete(), "we just did this");
1.4925 + init->set_complete_with_arraycopy();
1.4926 + assert(dest->is_CheckCastPP(), "sanity");
1.4927 + assert(dest->in(0)->in(0) == init, "dest pinned");
1.4928 + adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
1.4929 + // From this point on, every exit path is responsible for
1.4930 + // initializing any non-copied parts of the object to zero.
1.4931 + // Also, if this flag is set we make sure that arraycopy interacts properly
1.4932 + // with G1, eliding pre-barriers. See CR 6627983.
1.4933 + dest_uninitialized = true;
1.4934 + } else {
1.4935 + // No zeroing elimination here.
1.4936 + alloc = NULL;
1.4937 + //original_dest = dest;
1.4938 + //dest_uninitialized = false;
1.4939 + }
1.4940 +
1.4941 + // Results are placed here:
1.4942 + enum { fast_path = 1, // normal void-returning assembly stub
1.4943 + checked_path = 2, // special assembly stub with cleanup
1.4944 + slow_call_path = 3, // something went wrong; call the VM
1.4945 + zero_path = 4, // bypass when length of copy is zero
1.4946 + bcopy_path = 5, // copy primitive array by 64-bit blocks
1.4947 + PATH_LIMIT = 6
1.4948 + };
1.4949 + RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
1.4950 + PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
1.4951 + PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
1.4952 + record_for_igvn(result_region);
1.4953 + _gvn.set_type_bottom(result_i_o);
1.4954 + _gvn.set_type_bottom(result_memory);
1.4955 + assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
1.4956 +
1.4957 + // The slow_control path:
1.4958 + Node* slow_control;
1.4959 + Node* slow_i_o = i_o();
1.4960 + Node* slow_mem = memory(adr_type);
1.4961 + debug_only(slow_control = (Node*) badAddress);
1.4962 +
1.4963 + // Checked control path:
1.4964 + Node* checked_control = top();
1.4965 + Node* checked_mem = NULL;
1.4966 + Node* checked_i_o = NULL;
1.4967 + Node* checked_value = NULL;
1.4968 +
1.4969 + if (basic_elem_type == T_CONFLICT) {
1.4970 + assert(!dest_uninitialized, "");
1.4971 + Node* cv = generate_generic_arraycopy(adr_type,
1.4972 + src, src_offset, dest, dest_offset,
1.4973 + copy_length, dest_uninitialized);
1.4974 + if (cv == NULL) cv = intcon(-1); // failure (no stub available)
1.4975 + checked_control = control();
1.4976 + checked_i_o = i_o();
1.4977 + checked_mem = memory(adr_type);
1.4978 + checked_value = cv;
1.4979 + set_control(top()); // no fast path
1.4980 + }
1.4981 +
1.4982 + Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
1.4983 + if (not_pos != NULL) {
1.4984 + PreserveJVMState pjvms(this);
1.4985 + set_control(not_pos);
1.4986 +
1.4987 + // (6) length must not be negative.
1.4988 + if (!length_never_negative) {
1.4989 + generate_negative_guard(copy_length, slow_region);
1.4990 + }
1.4991 +
1.4992 + // copy_length is 0.
1.4993 + if (!stopped() && dest_uninitialized) {
1.4994 + Node* dest_length = alloc->in(AllocateNode::ALength);
1.4995 + if (copy_length->eqv_uncast(dest_length)
1.4996 + || _gvn.find_int_con(dest_length, 1) <= 0) {
1.4997 + // There is no zeroing to do. No need for a secondary raw memory barrier.
1.4998 + } else {
1.4999 + // Clear the whole thing since there are no source elements to copy.
1.5000 + generate_clear_array(adr_type, dest, basic_elem_type,
1.5001 + intcon(0), NULL,
1.5002 + alloc->in(AllocateNode::AllocSize));
1.5003 + // Use a secondary InitializeNode as raw memory barrier.
1.5004 + // Currently it is needed only on this path since other
1.5005 + // paths have stub or runtime calls as raw memory barriers.
1.5006 + InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
1.5007 + Compile::AliasIdxRaw,
1.5008 + top())->as_Initialize();
1.5009 + init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
1.5010 + }
1.5011 + }
1.5012 +
1.5013 + // Present the results of the fast call.
1.5014 + result_region->init_req(zero_path, control());
1.5015 + result_i_o ->init_req(zero_path, i_o());
1.5016 + result_memory->init_req(zero_path, memory(adr_type));
1.5017 + }
1.5018 +
1.5019 + if (!stopped() && dest_uninitialized) {
1.5020 + // We have to initialize the *uncopied* part of the array to zero.
1.5021 + // The copy destination is the slice dest[off..off+len]. The other slices
1.5022 + // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
1.5023 + Node* dest_size = alloc->in(AllocateNode::AllocSize);
1.5024 + Node* dest_length = alloc->in(AllocateNode::ALength);
1.5025 + Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
1.5026 + copy_length));
1.5027 +
1.5028 + // If there is a head section that needs zeroing, do it now.
1.5029 + if (find_int_con(dest_offset, -1) != 0) {
1.5030 + generate_clear_array(adr_type, dest, basic_elem_type,
1.5031 + intcon(0), dest_offset,
1.5032 + NULL);
1.5033 + }
1.5034 +
1.5035 + // Next, perform a dynamic check on the tail length.
1.5036 + // It is often zero, and we can win big if we prove this.
1.5037 + // There are two wins: Avoid generating the ClearArray
1.5038 + // with its attendant messy index arithmetic, and upgrade
1.5039 + // the copy to a more hardware-friendly word size of 64 bits.
1.5040 + Node* tail_ctl = NULL;
1.5041 + if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
1.5042 + Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
1.5043 + Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
1.5044 + tail_ctl = generate_slow_guard(bol_lt, NULL);
1.5045 + assert(tail_ctl != NULL || !stopped(), "must be an outcome");
1.5046 + }
1.5047 +
1.5048 + // At this point, let's assume there is no tail.
1.5049 + if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
1.5050 + // There is no tail. Try an upgrade to a 64-bit copy.
1.5051 + bool didit = false;
1.5052 + { PreserveJVMState pjvms(this);
1.5053 + didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
1.5054 + src, src_offset, dest, dest_offset,
1.5055 + dest_size, dest_uninitialized);
1.5056 + if (didit) {
1.5057 + // Present the results of the block-copying fast call.
1.5058 + result_region->init_req(bcopy_path, control());
1.5059 + result_i_o ->init_req(bcopy_path, i_o());
1.5060 + result_memory->init_req(bcopy_path, memory(adr_type));
1.5061 + }
1.5062 + }
1.5063 + if (didit)
1.5064 + set_control(top()); // no regular fast path
1.5065 + }
1.5066 +
1.5067 + // Clear the tail, if any.
1.5068 + if (tail_ctl != NULL) {
1.5069 + Node* notail_ctl = stopped() ? NULL : control();
1.5070 + set_control(tail_ctl);
1.5071 + if (notail_ctl == NULL) {
1.5072 + generate_clear_array(adr_type, dest, basic_elem_type,
1.5073 + dest_tail, NULL,
1.5074 + dest_size);
1.5075 + } else {
1.5076 + // Make a local merge.
1.5077 + Node* done_ctl = new(C) RegionNode(3);
1.5078 + Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
1.5079 + done_ctl->init_req(1, notail_ctl);
1.5080 + done_mem->init_req(1, memory(adr_type));
1.5081 + generate_clear_array(adr_type, dest, basic_elem_type,
1.5082 + dest_tail, NULL,
1.5083 + dest_size);
1.5084 + done_ctl->init_req(2, control());
1.5085 + done_mem->init_req(2, memory(adr_type));
1.5086 + set_control( _gvn.transform(done_ctl));
1.5087 + set_memory( _gvn.transform(done_mem), adr_type );
1.5088 + }
1.5089 + }
1.5090 + }
1.5091 +
1.5092 + BasicType copy_type = basic_elem_type;
1.5093 + assert(basic_elem_type != T_ARRAY, "caller must fix this");
1.5094 + if (!stopped() && copy_type == T_OBJECT) {
1.5095 + // If src and dest have compatible element types, we can copy bits.
1.5096 + // Types S[] and D[] are compatible if D is a supertype of S.
1.5097 + //
1.5098 + // If they are not, we will use checked_oop_disjoint_arraycopy,
1.5099 + // which performs a fast optimistic per-oop check, and backs off
1.5100 + // further to JVM_ArrayCopy on the first per-oop check that fails.
1.5101 + // (Actually, we don't move raw bits only; the GC requires card marks.)
1.5102 +
1.5103 + // Get the Klass* for both src and dest
1.5104 + Node* src_klass = load_object_klass(src);
1.5105 + Node* dest_klass = load_object_klass(dest);
1.5106 +
1.5107 + // Generate the subtype check.
1.5108 + // This might fold up statically, or then again it might not.
1.5109 + //
1.5110 + // Non-static example: Copying List<String>.elements to a new String[].
1.5111 + // The backing store for a List<String> is always an Object[],
1.5112 + // but its elements are always type String, if the generic types
1.5113 + // are correct at the source level.
1.5114 + //
1.5115 + // Test S[] against D[], not S against D, because (probably)
1.5116 + // the secondary supertype cache is less busy for S[] than S.
1.5117 + // This usually only matters when D is an interface.
1.5118 + Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
1.5119 + // Plug failing path into checked_oop_disjoint_arraycopy
1.5120 + if (not_subtype_ctrl != top()) {
1.5121 + PreserveJVMState pjvms(this);
1.5122 + set_control(not_subtype_ctrl);
1.5123 + // (At this point we can assume disjoint_bases, since types differ.)
1.5124 + int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
1.5125 + Node* p1 = basic_plus_adr(dest_klass, ek_offset);
1.5126 + Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
1.5127 + Node* dest_elem_klass = _gvn.transform(n1);
1.5128 + Node* cv = generate_checkcast_arraycopy(adr_type,
1.5129 + dest_elem_klass,
1.5130 + src, src_offset, dest, dest_offset,
1.5131 + ConvI2X(copy_length), dest_uninitialized);
1.5132 + if (cv == NULL) cv = intcon(-1); // failure (no stub available)
1.5133 + checked_control = control();
1.5134 + checked_i_o = i_o();
1.5135 + checked_mem = memory(adr_type);
1.5136 + checked_value = cv;
1.5137 + }
1.5138 + // At this point we know we do not need type checks on oop stores.
1.5139 +
1.5140 + // Let's see if we need card marks:
1.5141 + if (alloc != NULL && use_ReduceInitialCardMarks()) {
1.5142 + // If we do not need card marks, copy using the jint or jlong stub.
1.5143 + copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
1.5144 + assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
1.5145 + "sizes agree");
1.5146 + }
1.5147 + }
1.5148 +
1.5149 + if (!stopped()) {
1.5150 + // Generate the fast path, if possible.
1.5151 + PreserveJVMState pjvms(this);
1.5152 + generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
1.5153 + src, src_offset, dest, dest_offset,
1.5154 + ConvI2X(copy_length), dest_uninitialized);
1.5155 +
1.5156 + // Present the results of the fast call.
1.5157 + result_region->init_req(fast_path, control());
1.5158 + result_i_o ->init_req(fast_path, i_o());
1.5159 + result_memory->init_req(fast_path, memory(adr_type));
1.5160 + }
1.5161 +
1.5162 + // Here are all the slow paths up to this point, in one bundle:
1.5163 + slow_control = top();
1.5164 + if (slow_region != NULL)
1.5165 + slow_control = _gvn.transform(slow_region);
1.5166 + DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
1.5167 +
1.5168 + set_control(checked_control);
1.5169 + if (!stopped()) {
1.5170 + // Clean up after the checked call.
1.5171 + // The returned value is either 0 or -1^K,
1.5172 + // where K = number of partially transferred array elements.
1.5173 + Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
1.5174 + Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1.5175 + IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
1.5176 +
1.5177 + // If it is 0, we are done, so transfer to the end.
1.5178 + Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
1.5179 + result_region->init_req(checked_path, checks_done);
1.5180 + result_i_o ->init_req(checked_path, checked_i_o);
1.5181 + result_memory->init_req(checked_path, checked_mem);
1.5182 +
1.5183 + // If it is not zero, merge into the slow call.
1.5184 + set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
1.5185 + RegionNode* slow_reg2 = new(C) RegionNode(3);
1.5186 + PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
1.5187 + PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
1.5188 + record_for_igvn(slow_reg2);
1.5189 + slow_reg2 ->init_req(1, slow_control);
1.5190 + slow_i_o2 ->init_req(1, slow_i_o);
1.5191 + slow_mem2 ->init_req(1, slow_mem);
1.5192 + slow_reg2 ->init_req(2, control());
1.5193 + slow_i_o2 ->init_req(2, checked_i_o);
1.5194 + slow_mem2 ->init_req(2, checked_mem);
1.5195 +
1.5196 + slow_control = _gvn.transform(slow_reg2);
1.5197 + slow_i_o = _gvn.transform(slow_i_o2);
1.5198 + slow_mem = _gvn.transform(slow_mem2);
1.5199 +
1.5200 + if (alloc != NULL) {
1.5201 + // We'll restart from the very beginning, after zeroing the whole thing.
1.5202 + // This can cause double writes, but that's OK since dest is brand new.
1.5203 + // So we ignore the low 31 bits of the value returned from the stub.
1.5204 + } else {
1.5205 + // We must continue the copy exactly where it failed, or else
1.5206 + // another thread might see the wrong number of writes to dest.
1.5207 + Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
1.5208 + Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
1.5209 + slow_offset->init_req(1, intcon(0));
1.5210 + slow_offset->init_req(2, checked_offset);
1.5211 + slow_offset = _gvn.transform(slow_offset);
1.5212 +
1.5213 + // Adjust the arguments by the conditionally incoming offset.
1.5214 + Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
1.5215 + Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
1.5216 + Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
1.5217 +
1.5218 + // Tweak the node variables to adjust the code produced below:
1.5219 + src_offset = src_off_plus;
1.5220 + dest_offset = dest_off_plus;
1.5221 + copy_length = length_minus;
1.5222 + }
1.5223 + }
1.5224 +
1.5225 + set_control(slow_control);
1.5226 + if (!stopped()) {
1.5227 + // Generate the slow path, if needed.
1.5228 + PreserveJVMState pjvms(this); // replace_in_map may trash the map
1.5229 +
1.5230 + set_memory(slow_mem, adr_type);
1.5231 + set_i_o(slow_i_o);
1.5232 +
1.5233 + if (dest_uninitialized) {
1.5234 + generate_clear_array(adr_type, dest, basic_elem_type,
1.5235 + intcon(0), NULL,
1.5236 + alloc->in(AllocateNode::AllocSize));
1.5237 + }
1.5238 +
1.5239 + generate_slow_arraycopy(adr_type,
1.5240 + src, src_offset, dest, dest_offset,
1.5241 + copy_length, /*dest_uninitialized*/false);
1.5242 +
1.5243 + result_region->init_req(slow_call_path, control());
1.5244 + result_i_o ->init_req(slow_call_path, i_o());
1.5245 + result_memory->init_req(slow_call_path, memory(adr_type));
1.5246 + }
1.5247 +
1.5248 + // Remove unused edges.
1.5249 + for (uint i = 1; i < result_region->req(); i++) {
1.5250 + if (result_region->in(i) == NULL)
1.5251 + result_region->init_req(i, top());
1.5252 + }
1.5253 +
1.5254 + // Finished; return the combined state.
1.5255 + set_control( _gvn.transform(result_region));
1.5256 + set_i_o( _gvn.transform(result_i_o) );
1.5257 + set_memory( _gvn.transform(result_memory), adr_type );
1.5258 +
1.5259 + // The memory edges above are precise in order to model effects around
1.5260 + // array copies accurately to allow value numbering of field loads around
1.5261 + // arraycopy. Such field loads, both before and after, are common in Java
1.5262 + // collections and similar classes involving header/array data structures.
1.5263 + //
1.5264 + // But with low number of register or when some registers are used or killed
1.5265 + // by arraycopy calls it causes registers spilling on stack. See 6544710.
1.5266 + // The next memory barrier is added to avoid it. If the arraycopy can be
1.5267 + // optimized away (which it can, sometimes) then we can manually remove
1.5268 + // the membar also.
1.5269 + //
1.5270 + // Do not let reads from the cloned object float above the arraycopy.
1.5271 + if (alloc != NULL) {
1.5272 + // Do not let stores that initialize this object be reordered with
1.5273 + // a subsequent store that would make this object accessible by
1.5274 + // other threads.
1.5275 + // Record what AllocateNode this StoreStore protects so that
1.5276 + // escape analysis can go from the MemBarStoreStoreNode to the
1.5277 + // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1.5278 + // based on the escape status of the AllocateNode.
1.5279 + insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1.5280 + } else if (InsertMemBarAfterArraycopy)
1.5281 + insert_mem_bar(Op_MemBarCPUOrder);
1.5282 +}
1.5283 +
1.5284 +
1.5285 +// Helper function which determines if an arraycopy immediately follows
1.5286 +// an allocation, with no intervening tests or other escapes for the object.
1.5287 +AllocateArrayNode*
1.5288 +LibraryCallKit::tightly_coupled_allocation(Node* ptr,
1.5289 + RegionNode* slow_region) {
1.5290 + if (stopped()) return NULL; // no fast path
1.5291 + if (C->AliasLevel() == 0) return NULL; // no MergeMems around
1.5292 +
1.5293 + AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
1.5294 + if (alloc == NULL) return NULL;
1.5295 +
1.5296 + Node* rawmem = memory(Compile::AliasIdxRaw);
1.5297 + // Is the allocation's memory state untouched?
1.5298 + if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
1.5299 + // Bail out if there have been raw-memory effects since the allocation.
1.5300 + // (Example: There might have been a call or safepoint.)
1.5301 + return NULL;
1.5302 + }
1.5303 + rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
1.5304 + if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
1.5305 + return NULL;
1.5306 + }
1.5307 +
1.5308 + // There must be no unexpected observers of this allocation.
1.5309 + for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
1.5310 + Node* obs = ptr->fast_out(i);
1.5311 + if (obs != this->map()) {
1.5312 + return NULL;
1.5313 + }
1.5314 + }
1.5315 +
1.5316 + // This arraycopy must unconditionally follow the allocation of the ptr.
1.5317 + Node* alloc_ctl = ptr->in(0);
1.5318 + assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
1.5319 +
1.5320 + Node* ctl = control();
1.5321 + while (ctl != alloc_ctl) {
1.5322 + // There may be guards which feed into the slow_region.
1.5323 + // Any other control flow means that we might not get a chance
1.5324 + // to finish initializing the allocated object.
1.5325 + if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
1.5326 + IfNode* iff = ctl->in(0)->as_If();
1.5327 + Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
1.5328 + assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
1.5329 + if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
1.5330 + ctl = iff->in(0); // This test feeds the known slow_region.
1.5331 + continue;
1.5332 + }
1.5333 + // One more try: Various low-level checks bottom out in
1.5334 + // uncommon traps. If the debug-info of the trap omits
1.5335 + // any reference to the allocation, as we've already
1.5336 + // observed, then there can be no objection to the trap.
1.5337 + bool found_trap = false;
1.5338 + for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
1.5339 + Node* obs = not_ctl->fast_out(j);
1.5340 + if (obs->in(0) == not_ctl && obs->is_Call() &&
1.5341 + (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
1.5342 + found_trap = true; break;
1.5343 + }
1.5344 + }
1.5345 + if (found_trap) {
1.5346 + ctl = iff->in(0); // This test feeds a harmless uncommon trap.
1.5347 + continue;
1.5348 + }
1.5349 + }
1.5350 + return NULL;
1.5351 + }
1.5352 +
1.5353 + // If we get this far, we have an allocation which immediately
1.5354 + // precedes the arraycopy, and we can take over zeroing the new object.
1.5355 + // The arraycopy will finish the initialization, and provide
1.5356 + // a new control state to which we will anchor the destination pointer.
1.5357 +
1.5358 + return alloc;
1.5359 +}
1.5360 +
1.5361 +// Helper for initialization of arrays, creating a ClearArray.
1.5362 +// It writes zero bits in [start..end), within the body of an array object.
1.5363 +// The memory effects are all chained onto the 'adr_type' alias category.
1.5364 +//
1.5365 +// Since the object is otherwise uninitialized, we are free
1.5366 +// to put a little "slop" around the edges of the cleared area,
1.5367 +// as long as it does not go back into the array's header,
1.5368 +// or beyond the array end within the heap.
1.5369 +//
1.5370 +// The lower edge can be rounded down to the nearest jint and the
1.5371 +// upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
1.5372 +//
1.5373 +// Arguments:
1.5374 +// adr_type memory slice where writes are generated
1.5375 +// dest oop of the destination array
1.5376 +// basic_elem_type element type of the destination
1.5377 +// slice_idx array index of first element to store
1.5378 +// slice_len number of elements to store (or NULL)
1.5379 +// dest_size total size in bytes of the array object
1.5380 +//
1.5381 +// Exactly one of slice_len or dest_size must be non-NULL.
1.5382 +// If dest_size is non-NULL, zeroing extends to the end of the object.
1.5383 +// If slice_len is non-NULL, the slice_idx value must be a constant.
1.5384 +void
1.5385 +LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
1.5386 + Node* dest,
1.5387 + BasicType basic_elem_type,
1.5388 + Node* slice_idx,
1.5389 + Node* slice_len,
1.5390 + Node* dest_size) {
1.5391 + // one or the other but not both of slice_len and dest_size:
1.5392 + assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
1.5393 + if (slice_len == NULL) slice_len = top();
1.5394 + if (dest_size == NULL) dest_size = top();
1.5395 +
1.5396 + // operate on this memory slice:
1.5397 + Node* mem = memory(adr_type); // memory slice to operate on
1.5398 +
1.5399 + // scaling and rounding of indexes:
1.5400 + int scale = exact_log2(type2aelembytes(basic_elem_type));
1.5401 + int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1.5402 + int clear_low = (-1 << scale) & (BytesPerInt - 1);
1.5403 + int bump_bit = (-1 << scale) & BytesPerInt;
1.5404 +
1.5405 + // determine constant starts and ends
1.5406 + const intptr_t BIG_NEG = -128;
1.5407 + assert(BIG_NEG + 2*abase < 0, "neg enough");
1.5408 + intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
1.5409 + intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
1.5410 + if (slice_len_con == 0) {
1.5411 + return; // nothing to do here
1.5412 + }
1.5413 + intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
1.5414 + intptr_t end_con = find_intptr_t_con(dest_size, -1);
1.5415 + if (slice_idx_con >= 0 && slice_len_con >= 0) {
1.5416 + assert(end_con < 0, "not two cons");
1.5417 + end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
1.5418 + BytesPerLong);
1.5419 + }
1.5420 +
1.5421 + if (start_con >= 0 && end_con >= 0) {
1.5422 + // Constant start and end. Simple.
1.5423 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.5424 + start_con, end_con, &_gvn);
1.5425 + } else if (start_con >= 0 && dest_size != top()) {
1.5426 + // Constant start, pre-rounded end after the tail of the array.
1.5427 + Node* end = dest_size;
1.5428 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.5429 + start_con, end, &_gvn);
1.5430 + } else if (start_con >= 0 && slice_len != top()) {
1.5431 + // Constant start, non-constant end. End needs rounding up.
1.5432 + // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
1.5433 + intptr_t end_base = abase + (slice_idx_con << scale);
1.5434 + int end_round = (-1 << scale) & (BytesPerLong - 1);
1.5435 + Node* end = ConvI2X(slice_len);
1.5436 + if (scale != 0)
1.5437 + end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
1.5438 + end_base += end_round;
1.5439 + end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
1.5440 + end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
1.5441 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.5442 + start_con, end, &_gvn);
1.5443 + } else if (start_con < 0 && dest_size != top()) {
1.5444 + // Non-constant start, pre-rounded end after the tail of the array.
1.5445 + // This is almost certainly a "round-to-end" operation.
1.5446 + Node* start = slice_idx;
1.5447 + start = ConvI2X(start);
1.5448 + if (scale != 0)
1.5449 + start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
1.5450 + start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
1.5451 + if ((bump_bit | clear_low) != 0) {
1.5452 + int to_clear = (bump_bit | clear_low);
1.5453 + // Align up mod 8, then store a jint zero unconditionally
1.5454 + // just before the mod-8 boundary.
1.5455 + if (((abase + bump_bit) & ~to_clear) - bump_bit
1.5456 + < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1.5457 + bump_bit = 0;
1.5458 + assert((abase & to_clear) == 0, "array base must be long-aligned");
1.5459 + } else {
1.5460 + // Bump 'start' up to (or past) the next jint boundary:
1.5461 + start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
1.5462 + assert((abase & clear_low) == 0, "array base must be int-aligned");
1.5463 + }
1.5464 + // Round bumped 'start' down to jlong boundary in body of array.
1.5465 + start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
1.5466 + if (bump_bit != 0) {
1.5467 + // Store a zero to the immediately preceding jint:
1.5468 + Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
1.5469 + Node* p1 = basic_plus_adr(dest, x1);
1.5470 + mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1.5471 + mem = _gvn.transform(mem);
1.5472 + }
1.5473 + }
1.5474 + Node* end = dest_size; // pre-rounded
1.5475 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.5476 + start, end, &_gvn);
1.5477 + } else {
1.5478 + // Non-constant start, unrounded non-constant end.
1.5479 + // (Nobody zeroes a random midsection of an array using this routine.)
1.5480 + ShouldNotReachHere(); // fix caller
1.5481 + }
1.5482 +
1.5483 + // Done.
1.5484 + set_memory(mem, adr_type);
1.5485 +}
1.5486 +
1.5487 +
1.5488 +bool
1.5489 +LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
1.5490 + BasicType basic_elem_type,
1.5491 + AllocateNode* alloc,
1.5492 + Node* src, Node* src_offset,
1.5493 + Node* dest, Node* dest_offset,
1.5494 + Node* dest_size, bool dest_uninitialized) {
1.5495 + // See if there is an advantage from block transfer.
1.5496 + int scale = exact_log2(type2aelembytes(basic_elem_type));
1.5497 + if (scale >= LogBytesPerLong)
1.5498 + return false; // it is already a block transfer
1.5499 +
1.5500 + // Look at the alignment of the starting offsets.
1.5501 + int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1.5502 +
1.5503 + intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
1.5504 + intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
1.5505 + if (src_off_con < 0 || dest_off_con < 0)
1.5506 + // At present, we can only understand constants.
1.5507 + return false;
1.5508 +
1.5509 + intptr_t src_off = abase + (src_off_con << scale);
1.5510 + intptr_t dest_off = abase + (dest_off_con << scale);
1.5511 +
1.5512 + if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1.5513 + // Non-aligned; too bad.
1.5514 + // One more chance: Pick off an initial 32-bit word.
1.5515 + // This is a common case, since abase can be odd mod 8.
1.5516 + if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1.5517 + ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1.5518 + Node* sptr = basic_plus_adr(src, src_off);
1.5519 + Node* dptr = basic_plus_adr(dest, dest_off);
1.5520 + Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1.5521 + store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
1.5522 + src_off += BytesPerInt;
1.5523 + dest_off += BytesPerInt;
1.5524 + } else {
1.5525 + return false;
1.5526 + }
1.5527 + }
1.5528 + assert(src_off % BytesPerLong == 0, "");
1.5529 + assert(dest_off % BytesPerLong == 0, "");
1.5530 +
1.5531 + // Do this copy by giant steps.
1.5532 + Node* sptr = basic_plus_adr(src, src_off);
1.5533 + Node* dptr = basic_plus_adr(dest, dest_off);
1.5534 + Node* countx = dest_size;
1.5535 + countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
1.5536 + countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
1.5537 +
1.5538 + bool disjoint_bases = true; // since alloc != NULL
1.5539 + generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
1.5540 + sptr, NULL, dptr, NULL, countx, dest_uninitialized);
1.5541 +
1.5542 + return true;
1.5543 +}
1.5544 +
1.5545 +
1.5546 +// Helper function; generates code for the slow case.
1.5547 +// We make a call to a runtime method which emulates the native method,
1.5548 +// but without the native wrapper overhead.
1.5549 +void
1.5550 +LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
1.5551 + Node* src, Node* src_offset,
1.5552 + Node* dest, Node* dest_offset,
1.5553 + Node* copy_length, bool dest_uninitialized) {
1.5554 + assert(!dest_uninitialized, "Invariant");
1.5555 + Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
1.5556 + OptoRuntime::slow_arraycopy_Type(),
1.5557 + OptoRuntime::slow_arraycopy_Java(),
1.5558 + "slow_arraycopy", adr_type,
1.5559 + src, src_offset, dest, dest_offset,
1.5560 + copy_length);
1.5561 +
1.5562 + // Handle exceptions thrown by this fellow:
1.5563 + make_slow_call_ex(call, env()->Throwable_klass(), false);
1.5564 +}
1.5565 +
1.5566 +// Helper function; generates code for cases requiring runtime checks.
1.5567 +Node*
1.5568 +LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
1.5569 + Node* dest_elem_klass,
1.5570 + Node* src, Node* src_offset,
1.5571 + Node* dest, Node* dest_offset,
1.5572 + Node* copy_length, bool dest_uninitialized) {
1.5573 + if (stopped()) return NULL;
1.5574 +
1.5575 + address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1.5576 + if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1.5577 + return NULL;
1.5578 + }
1.5579 +
1.5580 + // Pick out the parameters required to perform a store-check
1.5581 + // for the target array. This is an optimistic check. It will
1.5582 + // look in each non-null element's class, at the desired klass's
1.5583 + // super_check_offset, for the desired klass.
1.5584 + int sco_offset = in_bytes(Klass::super_check_offset_offset());
1.5585 + Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1.5586 + Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1.5587 + Node* check_offset = ConvI2X(_gvn.transform(n3));
1.5588 + Node* check_value = dest_elem_klass;
1.5589 +
1.5590 + Node* src_start = array_element_address(src, src_offset, T_OBJECT);
1.5591 + Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1.5592 +
1.5593 + // (We know the arrays are never conjoint, because their types differ.)
1.5594 + Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1.5595 + OptoRuntime::checkcast_arraycopy_Type(),
1.5596 + copyfunc_addr, "checkcast_arraycopy", adr_type,
1.5597 + // five arguments, of which two are
1.5598 + // intptr_t (jlong in LP64)
1.5599 + src_start, dest_start,
1.5600 + copy_length XTOP,
1.5601 + check_offset XTOP,
1.5602 + check_value);
1.5603 +
1.5604 + return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1.5605 +}
1.5606 +
1.5607 +
1.5608 +// Helper function; generates code for cases requiring runtime checks.
1.5609 +Node*
1.5610 +LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
1.5611 + Node* src, Node* src_offset,
1.5612 + Node* dest, Node* dest_offset,
1.5613 + Node* copy_length, bool dest_uninitialized) {
1.5614 + assert(!dest_uninitialized, "Invariant");
1.5615 + if (stopped()) return NULL;
1.5616 + address copyfunc_addr = StubRoutines::generic_arraycopy();
1.5617 + if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1.5618 + return NULL;
1.5619 + }
1.5620 +
1.5621 + Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1.5622 + OptoRuntime::generic_arraycopy_Type(),
1.5623 + copyfunc_addr, "generic_arraycopy", adr_type,
1.5624 + src, src_offset, dest, dest_offset, copy_length);
1.5625 +
1.5626 + return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1.5627 +}
1.5628 +
1.5629 +// Helper function; generates the fast out-of-line call to an arraycopy stub.
1.5630 +void
1.5631 +LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
1.5632 + BasicType basic_elem_type,
1.5633 + bool disjoint_bases,
1.5634 + Node* src, Node* src_offset,
1.5635 + Node* dest, Node* dest_offset,
1.5636 + Node* copy_length, bool dest_uninitialized) {
1.5637 + if (stopped()) return; // nothing to do
1.5638 +
1.5639 + Node* src_start = src;
1.5640 + Node* dest_start = dest;
1.5641 + if (src_offset != NULL || dest_offset != NULL) {
1.5642 + assert(src_offset != NULL && dest_offset != NULL, "");
1.5643 + src_start = array_element_address(src, src_offset, basic_elem_type);
1.5644 + dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1.5645 + }
1.5646 +
1.5647 + // Figure out which arraycopy runtime method to call.
1.5648 + const char* copyfunc_name = "arraycopy";
1.5649 + address copyfunc_addr =
1.5650 + basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1.5651 + disjoint_bases, copyfunc_name, dest_uninitialized);
1.5652 +
1.5653 + // Call it. Note that the count_ix value is not scaled to a byte-size.
1.5654 + make_runtime_call(RC_LEAF|RC_NO_FP,
1.5655 + OptoRuntime::fast_arraycopy_Type(),
1.5656 + copyfunc_addr, copyfunc_name, adr_type,
1.5657 + src_start, dest_start, copy_length XTOP);
1.5658 +}
1.5659 +
1.5660 +//-------------inline_encodeISOArray-----------------------------------
1.5661 +// encode char[] to byte[] in ISO_8859_1
1.5662 +bool LibraryCallKit::inline_encodeISOArray() {
1.5663 + assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
1.5664 + // no receiver since it is static method
1.5665 + Node *src = argument(0);
1.5666 + Node *src_offset = argument(1);
1.5667 + Node *dst = argument(2);
1.5668 + Node *dst_offset = argument(3);
1.5669 + Node *length = argument(4);
1.5670 +
1.5671 + const Type* src_type = src->Value(&_gvn);
1.5672 + const Type* dst_type = dst->Value(&_gvn);
1.5673 + const TypeAryPtr* top_src = src_type->isa_aryptr();
1.5674 + const TypeAryPtr* top_dest = dst_type->isa_aryptr();
1.5675 + if (top_src == NULL || top_src->klass() == NULL ||
1.5676 + top_dest == NULL || top_dest->klass() == NULL) {
1.5677 + // failed array check
1.5678 + return false;
1.5679 + }
1.5680 +
1.5681 + // Figure out the size and type of the elements we will be copying.
1.5682 + BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1.5683 + BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1.5684 + if (src_elem != T_CHAR || dst_elem != T_BYTE) {
1.5685 + return false;
1.5686 + }
1.5687 + Node* src_start = array_element_address(src, src_offset, src_elem);
1.5688 + Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1.5689 + // 'src_start' points to src array + scaled offset
1.5690 + // 'dst_start' points to dst array + scaled offset
1.5691 +
1.5692 + const TypeAryPtr* mtype = TypeAryPtr::BYTES;
1.5693 + Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
1.5694 + enc = _gvn.transform(enc);
1.5695 + Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
1.5696 + set_memory(res_mem, mtype);
1.5697 + set_result(enc);
1.5698 + return true;
1.5699 +}
1.5700 +
1.5701 +/**
1.5702 + * Calculate CRC32 for byte.
1.5703 + * int java.util.zip.CRC32.update(int crc, int b)
1.5704 + */
1.5705 +bool LibraryCallKit::inline_updateCRC32() {
1.5706 + assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
1.5707 + assert(callee()->signature()->size() == 2, "update has 2 parameters");
1.5708 + // no receiver since it is static method
1.5709 + Node* crc = argument(0); // type: int
1.5710 + Node* b = argument(1); // type: int
1.5711 +
1.5712 + /*
1.5713 + * int c = ~ crc;
1.5714 + * b = timesXtoThe32[(b ^ c) & 0xFF];
1.5715 + * b = b ^ (c >>> 8);
1.5716 + * crc = ~b;
1.5717 + */
1.5718 +
1.5719 + Node* M1 = intcon(-1);
1.5720 + crc = _gvn.transform(new (C) XorINode(crc, M1));
1.5721 + Node* result = _gvn.transform(new (C) XorINode(crc, b));
1.5722 + result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
1.5723 +
1.5724 + Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
1.5725 + Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
1.5726 + Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
1.5727 + result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
1.5728 +
1.5729 + crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
1.5730 + result = _gvn.transform(new (C) XorINode(crc, result));
1.5731 + result = _gvn.transform(new (C) XorINode(result, M1));
1.5732 + set_result(result);
1.5733 + return true;
1.5734 +}
1.5735 +
1.5736 +/**
1.5737 + * Calculate CRC32 for byte[] array.
1.5738 + * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
1.5739 + */
1.5740 +bool LibraryCallKit::inline_updateBytesCRC32() {
1.5741 + assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
1.5742 + assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
1.5743 + // no receiver since it is static method
1.5744 + Node* crc = argument(0); // type: int
1.5745 + Node* src = argument(1); // type: oop
1.5746 + Node* offset = argument(2); // type: int
1.5747 + Node* length = argument(3); // type: int
1.5748 +
1.5749 + const Type* src_type = src->Value(&_gvn);
1.5750 + const TypeAryPtr* top_src = src_type->isa_aryptr();
1.5751 + if (top_src == NULL || top_src->klass() == NULL) {
1.5752 + // failed array check
1.5753 + return false;
1.5754 + }
1.5755 +
1.5756 + // Figure out the size and type of the elements we will be copying.
1.5757 + BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1.5758 + if (src_elem != T_BYTE) {
1.5759 + return false;
1.5760 + }
1.5761 +
1.5762 + // 'src_start' points to src array + scaled offset
1.5763 + Node* src_start = array_element_address(src, offset, src_elem);
1.5764 +
1.5765 + // We assume that range check is done by caller.
1.5766 + // TODO: generate range check (offset+length < src.length) in debug VM.
1.5767 +
1.5768 + // Call the stub.
1.5769 + address stubAddr = StubRoutines::updateBytesCRC32();
1.5770 + const char *stubName = "updateBytesCRC32";
1.5771 +
1.5772 + Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
1.5773 + stubAddr, stubName, TypePtr::BOTTOM,
1.5774 + crc, src_start, length);
1.5775 + Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1.5776 + set_result(result);
1.5777 + return true;
1.5778 +}
1.5779 +
1.5780 +/**
1.5781 + * Calculate CRC32 for ByteBuffer.
1.5782 + * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
1.5783 + */
1.5784 +bool LibraryCallKit::inline_updateByteBufferCRC32() {
1.5785 + assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
1.5786 + assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
1.5787 + // no receiver since it is static method
1.5788 + Node* crc = argument(0); // type: int
1.5789 + Node* src = argument(1); // type: long
1.5790 + Node* offset = argument(3); // type: int
1.5791 + Node* length = argument(4); // type: int
1.5792 +
1.5793 + src = ConvL2X(src); // adjust Java long to machine word
1.5794 + Node* base = _gvn.transform(new (C) CastX2PNode(src));
1.5795 + offset = ConvI2X(offset);
1.5796 +
1.5797 + // 'src_start' points to src array + scaled offset
1.5798 + Node* src_start = basic_plus_adr(top(), base, offset);
1.5799 +
1.5800 + // Call the stub.
1.5801 + address stubAddr = StubRoutines::updateBytesCRC32();
1.5802 + const char *stubName = "updateBytesCRC32";
1.5803 +
1.5804 + Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
1.5805 + stubAddr, stubName, TypePtr::BOTTOM,
1.5806 + crc, src_start, length);
1.5807 + Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1.5808 + set_result(result);
1.5809 + return true;
1.5810 +}
1.5811 +
1.5812 +//----------------------------inline_reference_get----------------------------
1.5813 +// public T java.lang.ref.Reference.get();
1.5814 +bool LibraryCallKit::inline_reference_get() {
1.5815 + const int referent_offset = java_lang_ref_Reference::referent_offset;
1.5816 + guarantee(referent_offset > 0, "should have already been set");
1.5817 +
1.5818 + // Get the argument:
1.5819 + Node* reference_obj = null_check_receiver();
1.5820 + if (stopped()) return true;
1.5821 +
1.5822 + Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
1.5823 +
1.5824 + ciInstanceKlass* klass = env()->Object_klass();
1.5825 + const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
1.5826 +
1.5827 + Node* no_ctrl = NULL;
1.5828 + Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
1.5829 +
1.5830 + // Use the pre-barrier to record the value in the referent field
1.5831 + pre_barrier(false /* do_load */,
1.5832 + control(),
1.5833 + NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
1.5834 + result /* pre_val */,
1.5835 + T_OBJECT);
1.5836 +
1.5837 + // Add memory barrier to prevent commoning reads from this field
1.5838 + // across safepoint since GC can change its value.
1.5839 + insert_mem_bar(Op_MemBarCPUOrder);
1.5840 +
1.5841 + set_result(result);
1.5842 + return true;
1.5843 +}
1.5844 +
1.5845 +
1.5846 +Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
1.5847 + bool is_exact=true, bool is_static=false) {
1.5848 +
1.5849 + const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
1.5850 + assert(tinst != NULL, "obj is null");
1.5851 + assert(tinst->klass()->is_loaded(), "obj is not loaded");
1.5852 + assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
1.5853 +
1.5854 + ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
1.5855 + ciSymbol::make(fieldTypeString),
1.5856 + is_static);
1.5857 + if (field == NULL) return (Node *) NULL;
1.5858 + assert (field != NULL, "undefined field");
1.5859 +
1.5860 + // Next code copied from Parse::do_get_xxx():
1.5861 +
1.5862 + // Compute address and memory type.
1.5863 + int offset = field->offset_in_bytes();
1.5864 + bool is_vol = field->is_volatile();
1.5865 + ciType* field_klass = field->type();
1.5866 + assert(field_klass->is_loaded(), "should be loaded");
1.5867 + const TypePtr* adr_type = C->alias_type(field)->adr_type();
1.5868 + Node *adr = basic_plus_adr(fromObj, fromObj, offset);
1.5869 + BasicType bt = field->layout_type();
1.5870 +
1.5871 + // Build the resultant type of the load
1.5872 + const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
1.5873 +
1.5874 + // Build the load.
1.5875 + Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
1.5876 + return loadedField;
1.5877 +}
1.5878 +
1.5879 +
1.5880 +//------------------------------inline_aescrypt_Block-----------------------
1.5881 +bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
1.5882 + address stubAddr;
1.5883 + const char *stubName;
1.5884 + assert(UseAES, "need AES instruction support");
1.5885 +
1.5886 + switch(id) {
1.5887 + case vmIntrinsics::_aescrypt_encryptBlock:
1.5888 + stubAddr = StubRoutines::aescrypt_encryptBlock();
1.5889 + stubName = "aescrypt_encryptBlock";
1.5890 + break;
1.5891 + case vmIntrinsics::_aescrypt_decryptBlock:
1.5892 + stubAddr = StubRoutines::aescrypt_decryptBlock();
1.5893 + stubName = "aescrypt_decryptBlock";
1.5894 + break;
1.5895 + }
1.5896 + if (stubAddr == NULL) return false;
1.5897 +
1.5898 + Node* aescrypt_object = argument(0);
1.5899 + Node* src = argument(1);
1.5900 + Node* src_offset = argument(2);
1.5901 + Node* dest = argument(3);
1.5902 + Node* dest_offset = argument(4);
1.5903 +
1.5904 + // (1) src and dest are arrays.
1.5905 + const Type* src_type = src->Value(&_gvn);
1.5906 + const Type* dest_type = dest->Value(&_gvn);
1.5907 + const TypeAryPtr* top_src = src_type->isa_aryptr();
1.5908 + const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1.5909 + assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
1.5910 +
1.5911 + // for the quick and dirty code we will skip all the checks.
1.5912 + // we are just trying to get the call to be generated.
1.5913 + Node* src_start = src;
1.5914 + Node* dest_start = dest;
1.5915 + if (src_offset != NULL || dest_offset != NULL) {
1.5916 + assert(src_offset != NULL && dest_offset != NULL, "");
1.5917 + src_start = array_element_address(src, src_offset, T_BYTE);
1.5918 + dest_start = array_element_address(dest, dest_offset, T_BYTE);
1.5919 + }
1.5920 +
1.5921 + // now need to get the start of its expanded key array
1.5922 + // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
1.5923 + Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
1.5924 + if (k_start == NULL) return false;
1.5925 +
1.5926 + if (Matcher::pass_original_key_for_aes()) {
1.5927 + // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
1.5928 + // compatibility issues between Java key expansion and SPARC crypto instructions
1.5929 + Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
1.5930 + if (original_k_start == NULL) return false;
1.5931 +
1.5932 + // Call the stub.
1.5933 + make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
1.5934 + stubAddr, stubName, TypePtr::BOTTOM,
1.5935 + src_start, dest_start, k_start, original_k_start);
1.5936 + } else {
1.5937 + // Call the stub.
1.5938 + make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
1.5939 + stubAddr, stubName, TypePtr::BOTTOM,
1.5940 + src_start, dest_start, k_start);
1.5941 + }
1.5942 +
1.5943 + return true;
1.5944 +}
1.5945 +
1.5946 +//------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
1.5947 +bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
1.5948 + address stubAddr;
1.5949 + const char *stubName;
1.5950 +
1.5951 + assert(UseAES, "need AES instruction support");
1.5952 +
1.5953 + switch(id) {
1.5954 + case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1.5955 + stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
1.5956 + stubName = "cipherBlockChaining_encryptAESCrypt";
1.5957 + break;
1.5958 + case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1.5959 + stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
1.5960 + stubName = "cipherBlockChaining_decryptAESCrypt";
1.5961 + break;
1.5962 + }
1.5963 + if (stubAddr == NULL) return false;
1.5964 +
1.5965 + Node* cipherBlockChaining_object = argument(0);
1.5966 + Node* src = argument(1);
1.5967 + Node* src_offset = argument(2);
1.5968 + Node* len = argument(3);
1.5969 + Node* dest = argument(4);
1.5970 + Node* dest_offset = argument(5);
1.5971 +
1.5972 + // (1) src and dest are arrays.
1.5973 + const Type* src_type = src->Value(&_gvn);
1.5974 + const Type* dest_type = dest->Value(&_gvn);
1.5975 + const TypeAryPtr* top_src = src_type->isa_aryptr();
1.5976 + const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1.5977 + assert (top_src != NULL && top_src->klass() != NULL
1.5978 + && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
1.5979 +
1.5980 + // checks are the responsibility of the caller
1.5981 + Node* src_start = src;
1.5982 + Node* dest_start = dest;
1.5983 + if (src_offset != NULL || dest_offset != NULL) {
1.5984 + assert(src_offset != NULL && dest_offset != NULL, "");
1.5985 + src_start = array_element_address(src, src_offset, T_BYTE);
1.5986 + dest_start = array_element_address(dest, dest_offset, T_BYTE);
1.5987 + }
1.5988 +
1.5989 + // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
1.5990 + // (because of the predicated logic executed earlier).
1.5991 + // so we cast it here safely.
1.5992 + // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
1.5993 +
1.5994 + Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
1.5995 + if (embeddedCipherObj == NULL) return false;
1.5996 +
1.5997 + // cast it to what we know it will be at runtime
1.5998 + const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
1.5999 + assert(tinst != NULL, "CBC obj is null");
1.6000 + assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
1.6001 + ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
1.6002 + if (!klass_AESCrypt->is_loaded()) return false;
1.6003 +
1.6004 + ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
1.6005 + const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
1.6006 + const TypeOopPtr* xtype = aklass->as_instance_type();
1.6007 + Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
1.6008 + aescrypt_object = _gvn.transform(aescrypt_object);
1.6009 +
1.6010 + // we need to get the start of the aescrypt_object's expanded key array
1.6011 + Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
1.6012 + if (k_start == NULL) return false;
1.6013 +
1.6014 + // similarly, get the start address of the r vector
1.6015 + Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
1.6016 + if (objRvec == NULL) return false;
1.6017 + Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
1.6018 +
1.6019 + Node* cbcCrypt;
1.6020 + if (Matcher::pass_original_key_for_aes()) {
1.6021 + // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
1.6022 + // compatibility issues between Java key expansion and SPARC crypto instructions
1.6023 + Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
1.6024 + if (original_k_start == NULL) return false;
1.6025 +
1.6026 + // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
1.6027 + cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
1.6028 + OptoRuntime::cipherBlockChaining_aescrypt_Type(),
1.6029 + stubAddr, stubName, TypePtr::BOTTOM,
1.6030 + src_start, dest_start, k_start, r_start, len, original_k_start);
1.6031 + } else {
1.6032 + // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
1.6033 + cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
1.6034 + OptoRuntime::cipherBlockChaining_aescrypt_Type(),
1.6035 + stubAddr, stubName, TypePtr::BOTTOM,
1.6036 + src_start, dest_start, k_start, r_start, len);
1.6037 + }
1.6038 +
1.6039 + // return cipher length (int)
1.6040 + Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
1.6041 + set_result(retvalue);
1.6042 + return true;
1.6043 +}
1.6044 +
1.6045 +//------------------------------get_key_start_from_aescrypt_object-----------------------
1.6046 +Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
1.6047 + Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
1.6048 + assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
1.6049 + if (objAESCryptKey == NULL) return (Node *) NULL;
1.6050 +
1.6051 + // now have the array, need to get the start address of the K array
1.6052 + Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
1.6053 + return k_start;
1.6054 +}
1.6055 +
1.6056 +//------------------------------get_original_key_start_from_aescrypt_object-----------------------
1.6057 +Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
1.6058 + Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
1.6059 + assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
1.6060 + if (objAESCryptKey == NULL) return (Node *) NULL;
1.6061 +
1.6062 + // now have the array, need to get the start address of the lastKey array
1.6063 + Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
1.6064 + return original_k_start;
1.6065 +}
1.6066 +
1.6067 +//----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
1.6068 +// Return node representing slow path of predicate check.
1.6069 +// the pseudo code we want to emulate with this predicate is:
1.6070 +// for encryption:
1.6071 +// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
1.6072 +// for decryption:
1.6073 +// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
1.6074 +// note cipher==plain is more conservative than the original java code but that's OK
1.6075 +//
1.6076 +Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
1.6077 + // First, check receiver for NULL since it is virtual method.
1.6078 + Node* objCBC = argument(0);
1.6079 + objCBC = null_check(objCBC);
1.6080 +
1.6081 + if (stopped()) return NULL; // Always NULL
1.6082 +
1.6083 + // Load embeddedCipher field of CipherBlockChaining object.
1.6084 + Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
1.6085 +
1.6086 + // get AESCrypt klass for instanceOf check
1.6087 + // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
1.6088 + // will have same classloader as CipherBlockChaining object
1.6089 + const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
1.6090 + assert(tinst != NULL, "CBCobj is null");
1.6091 + assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
1.6092 +
1.6093 + // we want to do an instanceof comparison against the AESCrypt class
1.6094 + ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
1.6095 + if (!klass_AESCrypt->is_loaded()) {
1.6096 + // if AESCrypt is not even loaded, we never take the intrinsic fast path
1.6097 + Node* ctrl = control();
1.6098 + set_control(top()); // no regular fast path
1.6099 + return ctrl;
1.6100 + }
1.6101 + ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
1.6102 +
1.6103 + Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
1.6104 + Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
1.6105 + Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
1.6106 +
1.6107 + Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
1.6108 +
1.6109 + // for encryption, we are done
1.6110 + if (!decrypting)
1.6111 + return instof_false; // even if it is NULL
1.6112 +
1.6113 + // for decryption, we need to add a further check to avoid
1.6114 + // taking the intrinsic path when cipher and plain are the same
1.6115 + // see the original java code for why.
1.6116 + RegionNode* region = new(C) RegionNode(3);
1.6117 + region->init_req(1, instof_false);
1.6118 + Node* src = argument(1);
1.6119 + Node* dest = argument(4);
1.6120 + Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
1.6121 + Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
1.6122 + Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
1.6123 + region->init_req(2, src_dest_conjoint);
1.6124 +
1.6125 + record_for_igvn(region);
1.6126 + return _gvn.transform(region);
1.6127 +}
