library_call.cpp revision 12076:ca56c5db7157
1155324Simp/* 2155324Simp * Copyright (c) 1999, 2016, Oracle and/or its affiliates. All rights reserved. 3213496Scognet * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4155324Simp * 5155324Simp * This code is free software; you can redistribute it and/or modify it 6155324Simp * under the terms of the GNU General Public License version 2 only, as 7155324Simp * published by the Free Software Foundation. 8155324Simp * 9155324Simp * This code is distributed in the hope that it will be useful, but WITHOUT 10155324Simp * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11155324Simp * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12155324Simp * version 2 for more details (a copy is included in the LICENSE file that 13155324Simp * accompanied this code). 14185265Simp * 15185265Simp * You should have received a copy of the GNU General Public License version 16185265Simp * 2 along with this work; if not, write to the Free Software Foundation, 17185265Simp * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18185265Simp * 19185265Simp * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20185265Simp * or visit www.oracle.com if you need additional information or have any 21185265Simp * questions. 22185265Simp * 23185265Simp */ 24185265Simp 25155324Simp#include "precompiled.hpp" 26155324Simp#include "asm/macroAssembler.hpp" 27266277Sian#include "classfile/systemDictionary.hpp" 28266277Sian#include "classfile/vmSymbols.hpp" 29155324Simp#include "compiler/compileBroker.hpp" 30155324Simp#include "compiler/compileLog.hpp" 31155324Simp#include "memory/resourceArea.hpp" 32155324Simp#include "oops/objArrayKlass.hpp" 33155324Simp#include "opto/addnode.hpp" 34155324Simp#include "opto/arraycopynode.hpp" 35155324Simp#include "opto/c2compiler.hpp" 36155324Simp#include "opto/callGenerator.hpp" 37155324Simp#include "opto/castnode.hpp" 38155324Simp#include "opto/cfgnode.hpp" 39155324Simp#include "opto/convertnode.hpp" 40155324Simp#include "opto/countbitsnode.hpp" 41155324Simp#include "opto/intrinsicnode.hpp" 42155324Simp#include "opto/idealKit.hpp" 43155324Simp#include "opto/mathexactnode.hpp" 44155324Simp#include "opto/movenode.hpp" 45278613Sian#include "opto/mulnode.hpp" 46155324Simp#include "opto/narrowptrnode.hpp" 47155324Simp#include "opto/opaquenode.hpp" 48259364Sian#include "opto/parse.hpp" 49155324Simp#include "opto/runtime.hpp" 50155324Simp#include "opto/subnode.hpp" 51155324Simp#include "prims/nativeLookup.hpp" 52213496Scognet#include "prims/unsafe.hpp" 53213496Scognet#include "runtime/sharedRuntime.hpp" 54155324Simp#ifdef TRACE_HAVE_INTRINSICS 55213496Scognet#include "trace/traceMacros.hpp" 56187599Simp#endif 57155324Simp 58278727Sianclass LibraryIntrinsic : public InlineCallGenerator { 59155324Simp // Extend the set of intrinsics known to the runtime: 60155324Simp public: 61155324Simp private: 62155324Simp bool _is_virtual; 63155324Simp bool _does_virtual_dispatch; 64235718Simp int8_t _predicates_count; // Intrinsic is predicated by several conditions 65266097Sian int8_t _last_predicate; // Last generated predicate 66155324Simp vmIntrinsics::ID _intrinsic_id; 67161704Scognet 68266097Sian public: 69266097Sian LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 70161704Scognet : InlineCallGenerator(m), 71266097Sian _is_virtual(is_virtual), 72155324Simp _does_virtual_dispatch(does_virtual_dispatch), 73155324Simp _predicates_count((int8_t)predicates_count), 74266110Sian _last_predicate((int8_t)-1), 75236989Simp _intrinsic_id(id) 76155324Simp { 77155324Simp } 78155324Simp virtual bool is_intrinsic() const { return true; } 79155324Simp virtual bool is_virtual() const { return _is_virtual; } 80278727Sian virtual bool is_predicated() const { return _predicates_count > 0; } 81155324Simp virtual int predicates_count() const { return _predicates_count; } 82278727Sian virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 83155324Simp virtual JVMState* generate(JVMState* jvms); 84278727Sian virtual Node* generate_predicate(JVMState* jvms, int predicate); 85266277Sian vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 86266277Sian}; 87278727Sian 88155324Simp 89155324Simp// Local helper class for LibraryIntrinsic: 90155324Simpclass LibraryCallKit : public GraphKit { 91278727Sian private: 92155324Simp LibraryIntrinsic* _intrinsic; // the library intrinsic being called 93155324Simp Node* _result; // the result node, if any 94155324Simp int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 95155324Simp 96155324Simp const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type); 97155324Simp 98155324Simp public: 99156828Simp LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 100278727Sian : GraphKit(jvms), 101156828Simp _intrinsic(intrinsic), 102156828Simp _result(NULL) 103156828Simp { 104156828Simp // Check if this is a root compile. In that case we don't have a caller. 105213496Scognet if (!jvms->has_method()) { 106213496Scognet _reexecute_sp = sp(); 107213496Scognet } else { 108213496Scognet // Find out how many arguments the interpreter needs when deoptimizing 109213496Scognet // and save the stack pointer value so it can used by uncommon_trap. 110213496Scognet // We find the argument count by looking at the declared signature. 111213496Scognet bool ignored_will_link; 112213496Scognet ciSignature* declared_signature = NULL; 113213496Scognet ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 114213496Scognet const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 115213496Scognet _reexecute_sp = sp() + nargs; // "push" arguments back on stack 116213496Scognet } 117213496Scognet } 118155324Simp 119155324Simp virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 120155324Simp 121278727Sian ciMethod* caller() const { return jvms()->method(); } 122278727Sian int bci() const { return jvms()->bci(); } 123155324Simp LibraryIntrinsic* intrinsic() const { return _intrinsic; } 124155324Simp vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 125278727Sian ciMethod* callee() const { return _intrinsic->method(); } 126278727Sian 127278727Sian bool try_to_inline(int predicate); 128155324Simp Node* try_to_predicate(int predicate); 129155324Simp 130278727Sian void push_result() { 131278727Sian // Push the result onto the stack. 132155324Simp if (!stopped() && result() != NULL) { 133155324Simp BasicType bt = result()->bottom_type()->basic_type(); 134278727Sian push_node(bt, result()); 135155324Simp } 136155324Simp } 137278727Sian 138278727Sian private: 139278727Sian void fatal_unexpected_iid(vmIntrinsics::ID iid) { 140278727Sian fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 141155324Simp } 142155324Simp 143278727Sian void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 144278727Sian void set_result(RegionNode* region, PhiNode* value); 145278727Sian Node* result() { return _result; } 146278727Sian 147155324Simp virtual int reexecute_sp() { return _reexecute_sp; } 148155324Simp 149278727Sian // Helper functions to inline natives 150278727Sian Node* generate_guard(Node* test, RegionNode* region, float true_prob); 151278727Sian Node* generate_slow_guard(Node* test, RegionNode* region); 152278727Sian Node* generate_fair_guard(Node* test, RegionNode* region); 153155324Simp Node* generate_negative_guard(Node* index, RegionNode* region, 154155324Simp // resulting CastII of index: 155278727Sian Node* *pos_index = NULL); 156278727Sian Node* generate_limit_guard(Node* offset, Node* subseq_length, 157278727Sian Node* array_length, 158278727Sian RegionNode* region); 159155324Simp void generate_string_range_check(Node* array, Node* offset, 160155324Simp Node* length, bool char_count); 161278727Sian Node* generate_current_thread(Node* &tls_output); 162278727Sian Node* load_mirror_from_klass(Node* klass); 163278727Sian Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 164278727Sian RegionNode* region, int null_path, 165155324Simp int offset); 166155324Simp Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 167278727Sian RegionNode* region, int null_path) { 168278727Sian int offset = java_lang_Class::klass_offset_in_bytes(); 169278727Sian return load_klass_from_mirror_common(mirror, never_see_null, 170278727Sian region, null_path, 171155324Simp offset); 172155324Simp } 173278727Sian Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 174278727Sian RegionNode* region, int null_path) { 175278727Sian int offset = java_lang_Class::array_klass_offset_in_bytes(); 176278727Sian return load_klass_from_mirror_common(mirror, never_see_null, 177155324Simp region, null_path, 178155324Simp offset); 179278727Sian } 180278727Sian Node* generate_access_flags_guard(Node* kls, 181278727Sian int modifier_mask, int modifier_bits, 182278727Sian RegionNode* region); 183155324Simp Node* generate_interface_guard(Node* kls, RegionNode* region); 184155324Simp Node* generate_array_guard(Node* kls, RegionNode* region) { 185278727Sian return generate_array_guard_common(kls, region, false, false); 186278727Sian } 187278727Sian Node* generate_non_array_guard(Node* kls, RegionNode* region) { 188278727Sian return generate_array_guard_common(kls, region, false, true); 189184310Sstas } 190278727Sian Node* generate_objArray_guard(Node* kls, RegionNode* region) { 191278727Sian return generate_array_guard_common(kls, region, true, false); 192278727Sian } 193278727Sian Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 194278727Sian return generate_array_guard_common(kls, region, true, true); 195184310Sstas } 196184310Sstas Node* generate_array_guard_common(Node* kls, RegionNode* region, 197278727Sian bool obj_array, bool not_array); 198278727Sian Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 199278727Sian CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 200278727Sian bool is_virtual = false, bool is_static = false); 201184310Sstas CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 202184310Sstas return generate_method_call(method_id, false, true); 203278727Sian } 204278727Sian CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 205278727Sian return generate_method_call(method_id, true, false); 206278727Sian } 207184310Sstas Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 208278727Sian Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 209278727Sian 210278727Sian Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 211278727Sian bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 212278727Sian bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 213184310Sstas bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 214184310Sstas Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 215278727Sian RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 216278727Sian bool inline_string_indexOfChar(); 217278727Sian bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 218278727Sian bool inline_string_toBytesU(); 219184310Sstas bool inline_string_getCharsU(); 220184310Sstas bool inline_string_copy(bool compress); 221278727Sian bool inline_string_char_access(bool is_store); 222278727Sian Node* round_double_node(Node* n); 223278727Sian bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 224278727Sian bool inline_math_native(vmIntrinsics::ID id); 225155324Simp bool inline_math(vmIntrinsics::ID id); 226155324Simp template <typename OverflowOp> 227266277Sian bool inline_math_overflow(Node* arg1, Node* arg2); 228266277Sian void inline_math_mathExact(Node* math, Node* test); 229266277Sian bool inline_math_addExactI(bool is_increment); 230266277Sian bool inline_math_addExactL(bool is_increment); 231266277Sian bool inline_math_multiplyExactI(); 232266277Sian bool inline_math_multiplyExactL(); 233155324Simp bool inline_math_negateExactI(); 234155324Simp bool inline_math_negateExactL(); 235155324Simp bool inline_math_subtractExactI(bool is_decrement); 236213496Scognet bool inline_math_subtractExactL(bool is_decrement); 237266065Sian bool inline_min_max(vmIntrinsics::ID id); 238265999Sian bool inline_notify(vmIntrinsics::ID id); 239155324Simp Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 240155324Simp // This returns Type::AnyPtr, RawPtr, or OopPtr. 241155324Simp int classify_unsafe_addr(Node* &base, Node* &offset); 242155324Simp Node* make_unsafe_address(Node* base, Node* offset); 243155324Simp // Helper for inline_unsafe_access. 244155324Simp // Generates the guards that check whether the result of 245155324Simp // Unsafe.getObject should be recorded in an SATB log buffer. 246155324Simp void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 247155324Simp 248238390Simp typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 249238390Simp bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 250238390Simp static bool klass_needs_init_guard(Node* kls); 251238390Simp bool inline_unsafe_allocate(); 252238390Simp bool inline_unsafe_newArray(bool uninitialized); 253238390Simp bool inline_unsafe_copyMemory(); 254238390Simp bool inline_native_currentThread(); 255238390Simp 256238390Simp bool inline_native_time_funcs(address method, const char* funcName); 257238390Simp bool inline_native_isInterrupted(); 258238390Simp bool inline_native_Class_query(vmIntrinsics::ID id); 259238390Simp bool inline_native_subtype_check(); 260155324Simp bool inline_native_getLength(); 261155324Simp bool inline_array_copyOf(bool is_copyOfRange); 262155324Simp bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 263155324Simp bool inline_preconditions_checkIndex(); 264155324Simp void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 265265999Sian bool inline_native_clone(bool is_virtual); 266265999Sian bool inline_native_Reflection_getCallerClass(); 267155324Simp // Helper function for inlining native object hash method 268155324Simp bool inline_native_hashcode(bool is_virtual, bool is_static); 269213496Scognet bool inline_native_getClass(); 270238389Simp 271155324Simp // Helper functions for inlining arraycopy 272213496Scognet bool inline_arraycopy(); 273155324Simp AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 274164432Simp RegionNode* slow_region); 275155324Simp JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 276155324Simp void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 277155324Simp 278213496Scognet typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 279213496Scognet MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store); 280213496Scognet MemNode::MemOrd access_kind_to_memord(AccessKind access_kind); 281213496Scognet bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 282155324Simp bool inline_unsafe_fence(vmIntrinsics::ID id); 283266110Sian bool inline_onspinwait(); 284266110Sian bool inline_fp_conversions(vmIntrinsics::ID id); 285266110Sian bool inline_number_methods(vmIntrinsics::ID id); 286266110Sian bool inline_reference_get(); 287266110Sian bool inline_Class_cast(); 288266110Sian bool inline_aescrypt_Block(vmIntrinsics::ID id); 289266110Sian bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 290266110Sian bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 291266110Sian Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 292155324Simp Node* inline_counterMode_AESCrypt_predicate(); 293238390Simp Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 294238390Simp Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 295238390Simp bool inline_ghash_processBlocks(); 296238390Simp bool inline_sha_implCompress(vmIntrinsics::ID id); 297238398Simp bool inline_digestBase_implCompressMB(int predicate); 298156828Simp bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 299155324Simp bool long_state, address stubAddr, const char *stubName, 300155324Simp Node* src_start, Node* ofs, Node* limit); 301278613Sian Node* get_state_from_sha_object(Node *sha_object); 302155324Simp Node* get_state_from_sha5_object(Node *sha_object); 303155324Simp Node* inline_digestBase_implCompressMB_predicate(int predicate); 304155324Simp bool inline_encodeISOArray(); 305155324Simp bool inline_updateCRC32(); 306155324Simp bool inline_updateBytesCRC32(); 307155324Simp bool inline_updateByteBufferCRC32(); 308155324Simp Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 309155324Simp bool inline_updateBytesCRC32C(); 310155324Simp bool inline_updateDirectByteBufferCRC32C(); 311155324Simp bool inline_updateBytesAdler32(); 312155324Simp bool inline_updateByteBufferAdler32(); 313266110Sian bool inline_multiplyToLen(); 314155324Simp bool inline_hasNegatives(); 315155324Simp bool inline_squareToLen(); 316155324Simp bool inline_mulAdd(); 317155324Simp bool inline_montgomeryMultiply(); 318155324Simp bool inline_montgomerySquare(); 319155324Simp bool inline_vectorizedMismatch(); 320155324Simp bool inline_fma(vmIntrinsics::ID id); 321155324Simp 322155324Simp bool inline_profileBoolean(); 323155324Simp bool inline_isCompileConstant(); 324155324Simp}; 325155324Simp 326155324Simp//---------------------------make_vm_intrinsic---------------------------- 327155324SimpCallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 328155324Simp vmIntrinsics::ID id = m->intrinsic_id(); 329155324Simp assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 330155324Simp 331155324Simp if (!m->is_loaded()) { 332155324Simp // Do not attempt to inline unloaded methods. 333155324Simp return NULL; 334155324Simp } 335155324Simp 336182805Simp C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 337182805Simp bool is_available = false; 338184309Sstas 339266110Sian { 340266110Sian // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 341184309Sstas // the compiler must transition to '_thread_in_vm' state because both 342266110Sian // methods access VM-internal data. 343184309Sstas VM_ENTRY_MARK; 344155324Simp methodHandle mh(THREAD, m->get_Method()); 345155324Simp is_available = compiler->is_intrinsic_supported(mh, is_virtual) && 346155324Simp !C->directive()->is_intrinsic_disabled(mh) && 347155324Simp !vmIntrinsics::is_disabled_by_flags(mh); 348155324Simp 349155324Simp } 350157891Simp 351155324Simp if (is_available) { 352155324Simp assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 353155324Simp assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 354155324Simp return new LibraryIntrinsic(m, is_virtual, 355155324Simp vmIntrinsics::predicates_needed(id), 356155324Simp vmIntrinsics::does_virtual_dispatch(id), 357155324Simp (vmIntrinsics::ID) id); 358155324Simp } else { 359155324Simp return NULL; 360155324Simp } 361155324Simp} 362155324Simp 363155324Simp//----------------------register_library_intrinsics----------------------- 364155324Simp// Initialize this file's data structures, for each Compile instance. 365155324Simpvoid Compile::register_library_intrinsics() { 366155324Simp // Nothing to do here. 367155324Simp} 368155324Simp 369155324SimpJVMState* LibraryIntrinsic::generate(JVMState* jvms) { 370155324Simp LibraryCallKit kit(jvms, this); 371155324Simp Compile* C = kit.C; 372155324Simp int nodes = C->unique(); 373155324Simp#ifndef PRODUCT 374155324Simp if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 375155324Simp char buf[1000]; 376155324Simp const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 377155324Simp tty->print_cr("Intrinsic %s", str); 378155324Simp } 379155324Simp#endif 380155324Simp ciMethod* callee = kit.callee(); 381155324Simp const int bci = kit.bci(); 382155324Simp 383155324Simp // Try to inline the intrinsic. 384236989Simp if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 385236989Simp kit.try_to_inline(_last_predicate)) { 386155324Simp if (C->print_intrinsics() || C->print_inlining()) { 387226832Skevlo C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 388161704Scognet } 389238348Simp C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 390167261Spiso if (C->log()) { 391226832Skevlo C->log()->elem("intrinsic id='%s'%s nodes='%d'", 392226832Skevlo vmIntrinsics::name_at(intrinsic_id()), 393226832Skevlo (is_virtual() ? " virtual='1'" : ""), 394226832Skevlo C->unique() - nodes); 395226832Skevlo } 396155324Simp // Push the result from the inlined method onto the stack. 397155324Simp kit.push_result(); 398155324Simp C->print_inlining_update(this); 399155324Simp return kit.transfer_exceptions_into_jvms(); 400155324Simp } 401155324Simp 402155324Simp // The intrinsic bailed out 403155324Simp if (C->print_intrinsics() || C->print_inlining()) { 404155324Simp if (jvms->has_method()) { 405236989Simp // Not a root compile. 406155324Simp const char* msg; 407155324Simp if (callee->intrinsic_candidate()) { 408155324Simp msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 409155324Simp } else { 410155324Simp msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 411155324Simp : "failed to inline (intrinsic), method not annotated"; 412155324Simp } 413155324Simp C->print_inlining(callee, jvms->depth() - 1, bci, msg); 414155324Simp } else { 415155324Simp // Root compile 416155324Simp tty->print("Did not generate intrinsic %s%s at bci:%d in", 417155324Simp vmIntrinsics::name_at(intrinsic_id()), 418155324Simp (is_virtual() ? " (virtual)" : ""), bci); 419155324Simp } 420155324Simp } 421236989Simp C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 422155324Simp C->print_inlining_update(this); 423155324Simp return NULL; 424155324Simp} 425155324Simp 426155324SimpNode* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 427155324Simp LibraryCallKit kit(jvms, this); 428155324Simp Compile* C = kit.C; 429155324Simp int nodes = C->unique(); 430155324Simp _last_predicate = predicate; 431155324Simp#ifndef PRODUCT 432155324Simp assert(is_predicated() && predicate < predicates_count(), "sanity"); 433155324Simp if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 434155324Simp char buf[1000]; 435155324Simp const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 436155324Simp tty->print_cr("Predicate for intrinsic %s", str); 437155324Simp } 438155324Simp#endif 439155324Simp ciMethod* callee = kit.callee(); 440155324Simp const int bci = kit.bci(); 441155324Simp 442155324Simp Node* slow_ctl = kit.try_to_predicate(predicate); 443155324Simp if (!kit.failing()) { 444155324Simp if (C->print_intrinsics() || C->print_inlining()) { 445155324Simp C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 446155324Simp } 447155324Simp C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 448155324Simp if (C->log()) { 449155324Simp C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 450155324Simp vmIntrinsics::name_at(intrinsic_id()), 451155324Simp (is_virtual() ? " virtual='1'" : ""), 452178366Scognet C->unique() - nodes); 453178366Scognet } 454178366Scognet return slow_ctl; // Could be NULL if the check folds. 455213496Scognet } 456178366Scognet 457178366Scognet // The intrinsic bailed out 458178366Scognet if (C->print_intrinsics() || C->print_inlining()) { 459238334Simp if (jvms->has_method()) { 460238334Simp // Not a root compile. 461238334Simp const char* msg = "failed to generate predicate for intrinsic"; 462238334Simp C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 463238334Simp } else { 464238334Simp // Root compile 465238334Simp C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 466238334Simp vmIntrinsics::name_at(intrinsic_id()), 467238334Simp (is_virtual() ? " (virtual)" : ""), bci); 468238334Simp } 469238334Simp } 470238334Simp C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 471238334Simp return NULL; 472238334Simp} 473238334Simp 474238334Simpbool LibraryCallKit::try_to_inline(int predicate) { 475238334Simp // Handle symbolic names for otherwise undistinguished boolean switches: 476238334Simp const bool is_store = true; 477238334Simp const bool is_compress = true; 478238334Simp const bool is_static = true; 479238334Simp const bool is_volatile = true; 480238334Simp 481238334Simp if (!jvms()->has_method()) { 482238334Simp // Root JVMState has a null method. 483238334Simp assert(map()->memory()->Opcode() == Op_Parm, ""); 484238334Simp // Insert the memory aliasing node 485238334Simp set_all_memory(reset_memory()); 486238334Simp } 487238334Simp assert(merged_memory(), ""); 488266110Sian 489266110Sian 490266110Sian switch (intrinsic_id()) { 491266110Sian case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 492266110Sian case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 493266110Sian case vmIntrinsics::_getClass: return inline_native_getClass(); 494266110Sian 495266110Sian case vmIntrinsics::_dsin: 496238334Simp case vmIntrinsics::_dcos: 497238334Simp case vmIntrinsics::_dtan: 498238334Simp case vmIntrinsics::_dabs: 499238334Simp case vmIntrinsics::_datan2: 500155324Simp case vmIntrinsics::_dsqrt: 501155324Simp case vmIntrinsics::_dexp: 502155324Simp case vmIntrinsics::_dlog: 503155324Simp case vmIntrinsics::_dlog10: 504155324Simp case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 505155324Simp 506155324Simp case vmIntrinsics::_min: 507155324Simp case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 508155324Simp 509155324Simp case vmIntrinsics::_notify: 510155324Simp case vmIntrinsics::_notifyAll: 511155324Simp if (InlineNotify) { 512155324Simp return inline_notify(intrinsic_id()); 513155324Simp } 514155324Simp return false; 515155324Simp 516155324Simp case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 517155324Simp case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 518155324Simp case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 519155324Simp case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 520155324Simp case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 521155324Simp case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 522155324Simp case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 523155324Simp case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 524213496Scognet case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 525155324Simp case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 526155324Simp case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 527155324Simp case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 528266277Sian 529 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 530 531 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 532 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 533 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 534 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 535 536 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 537 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 538 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 539 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 540 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 541 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 542 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 543 544 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 545 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 546 547 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 548 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 549 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 550 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 551 552 case vmIntrinsics::_compressStringC: 553 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 554 case vmIntrinsics::_inflateStringC: 555 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 556 557 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 558 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 559 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 560 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 561 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 562 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 563 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 564 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 565 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 566 567 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 568 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 569 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 570 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 571 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 572 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 573 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 574 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 575 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 576 577 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 578 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 579 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 580 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 581 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 582 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 583 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 584 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 585 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 586 587 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 588 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 589 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 590 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 591 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 592 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 593 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 594 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 595 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 596 597 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 598 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 599 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 600 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 601 602 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 603 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 604 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 605 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 606 607 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 608 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 609 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 610 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 611 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 612 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 613 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 614 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 615 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 616 617 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 618 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 619 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 620 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 621 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 622 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 623 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 624 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 625 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 626 627 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 628 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 629 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 630 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 631 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 632 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 633 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 634 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 635 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 636 637 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 638 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 639 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 640 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 641 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 642 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 643 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 644 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 645 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 646 647 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 648 case vmIntrinsics::_compareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 649 case vmIntrinsics::_compareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 650 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 651 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 652 653 case vmIntrinsics::_weakCompareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 654 case vmIntrinsics::_weakCompareAndSwapObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 655 case vmIntrinsics::_weakCompareAndSwapObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 656 case vmIntrinsics::_weakCompareAndSwapObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 657 case vmIntrinsics::_weakCompareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 658 case vmIntrinsics::_weakCompareAndSwapByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 659 case vmIntrinsics::_weakCompareAndSwapByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 660 case vmIntrinsics::_weakCompareAndSwapByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 661 case vmIntrinsics::_weakCompareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 662 case vmIntrinsics::_weakCompareAndSwapShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 663 case vmIntrinsics::_weakCompareAndSwapShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 664 case vmIntrinsics::_weakCompareAndSwapShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 665 case vmIntrinsics::_weakCompareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 666 case vmIntrinsics::_weakCompareAndSwapIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 667 case vmIntrinsics::_weakCompareAndSwapIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 668 case vmIntrinsics::_weakCompareAndSwapIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 669 case vmIntrinsics::_weakCompareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 670 case vmIntrinsics::_weakCompareAndSwapLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 671 case vmIntrinsics::_weakCompareAndSwapLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 672 case vmIntrinsics::_weakCompareAndSwapLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 673 674 case vmIntrinsics::_compareAndExchangeObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 675 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 676 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 677 case vmIntrinsics::_compareAndExchangeByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 678 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 679 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 680 case vmIntrinsics::_compareAndExchangeShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 681 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 682 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 683 case vmIntrinsics::_compareAndExchangeIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 684 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 685 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 686 case vmIntrinsics::_compareAndExchangeLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 687 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 688 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 689 690 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 691 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 692 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 693 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 694 695 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 696 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 697 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 698 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 699 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 700 701 case vmIntrinsics::_loadFence: 702 case vmIntrinsics::_storeFence: 703 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 704 705 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 706 707 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 708 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 709 710#ifdef TRACE_HAVE_INTRINSICS 711 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 712#endif 713 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 714 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 715 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 716 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 717 case vmIntrinsics::_getLength: return inline_native_getLength(); 718 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 719 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 720 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 721 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 722 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 723 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 724 725 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 726 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 727 728 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 729 730 case vmIntrinsics::_isInstance: 731 case vmIntrinsics::_getModifiers: 732 case vmIntrinsics::_isInterface: 733 case vmIntrinsics::_isArray: 734 case vmIntrinsics::_isPrimitive: 735 case vmIntrinsics::_getSuperclass: 736 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 737 738 case vmIntrinsics::_floatToRawIntBits: 739 case vmIntrinsics::_floatToIntBits: 740 case vmIntrinsics::_intBitsToFloat: 741 case vmIntrinsics::_doubleToRawLongBits: 742 case vmIntrinsics::_doubleToLongBits: 743 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 744 745 case vmIntrinsics::_numberOfLeadingZeros_i: 746 case vmIntrinsics::_numberOfLeadingZeros_l: 747 case vmIntrinsics::_numberOfTrailingZeros_i: 748 case vmIntrinsics::_numberOfTrailingZeros_l: 749 case vmIntrinsics::_bitCount_i: 750 case vmIntrinsics::_bitCount_l: 751 case vmIntrinsics::_reverseBytes_i: 752 case vmIntrinsics::_reverseBytes_l: 753 case vmIntrinsics::_reverseBytes_s: 754 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 755 756 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 757 758 case vmIntrinsics::_Reference_get: return inline_reference_get(); 759 760 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 761 762 case vmIntrinsics::_aescrypt_encryptBlock: 763 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 764 765 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 766 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 767 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 768 769 case vmIntrinsics::_counterMode_AESCrypt: 770 return inline_counterMode_AESCrypt(intrinsic_id()); 771 772 case vmIntrinsics::_sha_implCompress: 773 case vmIntrinsics::_sha2_implCompress: 774 case vmIntrinsics::_sha5_implCompress: 775 return inline_sha_implCompress(intrinsic_id()); 776 777 case vmIntrinsics::_digestBase_implCompressMB: 778 return inline_digestBase_implCompressMB(predicate); 779 780 case vmIntrinsics::_multiplyToLen: 781 return inline_multiplyToLen(); 782 783 case vmIntrinsics::_squareToLen: 784 return inline_squareToLen(); 785 786 case vmIntrinsics::_mulAdd: 787 return inline_mulAdd(); 788 789 case vmIntrinsics::_montgomeryMultiply: 790 return inline_montgomeryMultiply(); 791 case vmIntrinsics::_montgomerySquare: 792 return inline_montgomerySquare(); 793 794 case vmIntrinsics::_vectorizedMismatch: 795 return inline_vectorizedMismatch(); 796 797 case vmIntrinsics::_ghash_processBlocks: 798 return inline_ghash_processBlocks(); 799 800 case vmIntrinsics::_encodeISOArray: 801 case vmIntrinsics::_encodeByteISOArray: 802 return inline_encodeISOArray(); 803 804 case vmIntrinsics::_updateCRC32: 805 return inline_updateCRC32(); 806 case vmIntrinsics::_updateBytesCRC32: 807 return inline_updateBytesCRC32(); 808 case vmIntrinsics::_updateByteBufferCRC32: 809 return inline_updateByteBufferCRC32(); 810 811 case vmIntrinsics::_updateBytesCRC32C: 812 return inline_updateBytesCRC32C(); 813 case vmIntrinsics::_updateDirectByteBufferCRC32C: 814 return inline_updateDirectByteBufferCRC32C(); 815 816 case vmIntrinsics::_updateBytesAdler32: 817 return inline_updateBytesAdler32(); 818 case vmIntrinsics::_updateByteBufferAdler32: 819 return inline_updateByteBufferAdler32(); 820 821 case vmIntrinsics::_profileBoolean: 822 return inline_profileBoolean(); 823 case vmIntrinsics::_isCompileConstant: 824 return inline_isCompileConstant(); 825 826 case vmIntrinsics::_hasNegatives: 827 return inline_hasNegatives(); 828 829 case vmIntrinsics::_fmaD: 830 case vmIntrinsics::_fmaF: 831 return inline_fma(intrinsic_id()); 832 833 default: 834 // If you get here, it may be that someone has added a new intrinsic 835 // to the list in vmSymbols.hpp without implementing it here. 836#ifndef PRODUCT 837 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 838 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 839 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 840 } 841#endif 842 return false; 843 } 844} 845 846Node* LibraryCallKit::try_to_predicate(int predicate) { 847 if (!jvms()->has_method()) { 848 // Root JVMState has a null method. 849 assert(map()->memory()->Opcode() == Op_Parm, ""); 850 // Insert the memory aliasing node 851 set_all_memory(reset_memory()); 852 } 853 assert(merged_memory(), ""); 854 855 switch (intrinsic_id()) { 856 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 857 return inline_cipherBlockChaining_AESCrypt_predicate(false); 858 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 859 return inline_cipherBlockChaining_AESCrypt_predicate(true); 860 case vmIntrinsics::_counterMode_AESCrypt: 861 return inline_counterMode_AESCrypt_predicate(); 862 case vmIntrinsics::_digestBase_implCompressMB: 863 return inline_digestBase_implCompressMB_predicate(predicate); 864 865 default: 866 // If you get here, it may be that someone has added a new intrinsic 867 // to the list in vmSymbols.hpp without implementing it here. 868#ifndef PRODUCT 869 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 870 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 871 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 872 } 873#endif 874 Node* slow_ctl = control(); 875 set_control(top()); // No fast path instrinsic 876 return slow_ctl; 877 } 878} 879 880//------------------------------set_result------------------------------- 881// Helper function for finishing intrinsics. 882void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 883 record_for_igvn(region); 884 set_control(_gvn.transform(region)); 885 set_result( _gvn.transform(value)); 886 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 887} 888 889//------------------------------generate_guard--------------------------- 890// Helper function for generating guarded fast-slow graph structures. 891// The given 'test', if true, guards a slow path. If the test fails 892// then a fast path can be taken. (We generally hope it fails.) 893// In all cases, GraphKit::control() is updated to the fast path. 894// The returned value represents the control for the slow path. 895// The return value is never 'top'; it is either a valid control 896// or NULL if it is obvious that the slow path can never be taken. 897// Also, if region and the slow control are not NULL, the slow edge 898// is appended to the region. 899Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 900 if (stopped()) { 901 // Already short circuited. 902 return NULL; 903 } 904 905 // Build an if node and its projections. 906 // If test is true we take the slow path, which we assume is uncommon. 907 if (_gvn.type(test) == TypeInt::ZERO) { 908 // The slow branch is never taken. No need to build this guard. 909 return NULL; 910 } 911 912 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 913 914 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 915 if (if_slow == top()) { 916 // The slow branch is never taken. No need to build this guard. 917 return NULL; 918 } 919 920 if (region != NULL) 921 region->add_req(if_slow); 922 923 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 924 set_control(if_fast); 925 926 return if_slow; 927} 928 929inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 930 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 931} 932inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 933 return generate_guard(test, region, PROB_FAIR); 934} 935 936inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 937 Node* *pos_index) { 938 if (stopped()) 939 return NULL; // already stopped 940 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 941 return NULL; // index is already adequately typed 942 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 943 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 944 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 945 if (is_neg != NULL && pos_index != NULL) { 946 // Emulate effect of Parse::adjust_map_after_if. 947 Node* ccast = new CastIINode(index, TypeInt::POS); 948 ccast->set_req(0, control()); 949 (*pos_index) = _gvn.transform(ccast); 950 } 951 return is_neg; 952} 953 954// Make sure that 'position' is a valid limit index, in [0..length]. 955// There are two equivalent plans for checking this: 956// A. (offset + copyLength) unsigned<= arrayLength 957// B. offset <= (arrayLength - copyLength) 958// We require that all of the values above, except for the sum and 959// difference, are already known to be non-negative. 960// Plan A is robust in the face of overflow, if offset and copyLength 961// are both hugely positive. 962// 963// Plan B is less direct and intuitive, but it does not overflow at 964// all, since the difference of two non-negatives is always 965// representable. Whenever Java methods must perform the equivalent 966// check they generally use Plan B instead of Plan A. 967// For the moment we use Plan A. 968inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 969 Node* subseq_length, 970 Node* array_length, 971 RegionNode* region) { 972 if (stopped()) 973 return NULL; // already stopped 974 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 975 if (zero_offset && subseq_length->eqv_uncast(array_length)) 976 return NULL; // common case of whole-array copy 977 Node* last = subseq_length; 978 if (!zero_offset) // last += offset 979 last = _gvn.transform(new AddINode(last, offset)); 980 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 981 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 982 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 983 return is_over; 984} 985 986// Emit range checks for the given String.value byte array 987void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 988 if (stopped()) { 989 return; // already stopped 990 } 991 RegionNode* bailout = new RegionNode(1); 992 record_for_igvn(bailout); 993 if (char_count) { 994 // Convert char count to byte count 995 count = _gvn.transform(new LShiftINode(count, intcon(1))); 996 } 997 998 // Offset and count must not be negative 999 generate_negative_guard(offset, bailout); 1000 generate_negative_guard(count, bailout); 1001 // Offset + count must not exceed length of array 1002 generate_limit_guard(offset, count, load_array_length(array), bailout); 1003 1004 if (bailout->req() > 1) { 1005 PreserveJVMState pjvms(this); 1006 set_control(_gvn.transform(bailout)); 1007 uncommon_trap(Deoptimization::Reason_intrinsic, 1008 Deoptimization::Action_maybe_recompile); 1009 } 1010} 1011 1012//--------------------------generate_current_thread-------------------- 1013Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1014 ciKlass* thread_klass = env()->Thread_klass(); 1015 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1016 Node* thread = _gvn.transform(new ThreadLocalNode()); 1017 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1018 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1019 tls_output = thread; 1020 return threadObj; 1021} 1022 1023 1024//------------------------------make_string_method_node------------------------ 1025// Helper method for String intrinsic functions. This version is called with 1026// str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1027// characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1028// containing the lengths of str1 and str2. 1029Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1030 Node* result = NULL; 1031 switch (opcode) { 1032 case Op_StrIndexOf: 1033 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1034 str1_start, cnt1, str2_start, cnt2, ae); 1035 break; 1036 case Op_StrComp: 1037 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1038 str1_start, cnt1, str2_start, cnt2, ae); 1039 break; 1040 case Op_StrEquals: 1041 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1042 // Use the constant length if there is one because optimized match rule may exist. 1043 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1044 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1045 break; 1046 default: 1047 ShouldNotReachHere(); 1048 return NULL; 1049 } 1050 1051 // All these intrinsics have checks. 1052 C->set_has_split_ifs(true); // Has chance for split-if optimization 1053 1054 return _gvn.transform(result); 1055} 1056 1057//------------------------------inline_string_compareTo------------------------ 1058bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1059 Node* arg1 = argument(0); 1060 Node* arg2 = argument(1); 1061 1062 // Get start addr and length of first argument 1063 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1064 Node* arg1_cnt = load_array_length(arg1); 1065 1066 // Get start addr and length of second argument 1067 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1068 Node* arg2_cnt = load_array_length(arg2); 1069 1070 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1071 set_result(result); 1072 return true; 1073} 1074 1075//------------------------------inline_string_equals------------------------ 1076bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1077 Node* arg1 = argument(0); 1078 Node* arg2 = argument(1); 1079 1080 // paths (plus control) merge 1081 RegionNode* region = new RegionNode(3); 1082 Node* phi = new PhiNode(region, TypeInt::BOOL); 1083 1084 if (!stopped()) { 1085 // Get start addr and length of first argument 1086 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1087 Node* arg1_cnt = load_array_length(arg1); 1088 1089 // Get start addr and length of second argument 1090 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1091 Node* arg2_cnt = load_array_length(arg2); 1092 1093 // Check for arg1_cnt != arg2_cnt 1094 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1095 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1096 Node* if_ne = generate_slow_guard(bol, NULL); 1097 if (if_ne != NULL) { 1098 phi->init_req(2, intcon(0)); 1099 region->init_req(2, if_ne); 1100 } 1101 1102 // Check for count == 0 is done by assembler code for StrEquals. 1103 1104 if (!stopped()) { 1105 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1106 phi->init_req(1, equals); 1107 region->init_req(1, control()); 1108 } 1109 } 1110 1111 // post merge 1112 set_control(_gvn.transform(region)); 1113 record_for_igvn(region); 1114 1115 set_result(_gvn.transform(phi)); 1116 return true; 1117} 1118 1119//------------------------------inline_array_equals---------------------------- 1120bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1121 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1122 Node* arg1 = argument(0); 1123 Node* arg2 = argument(1); 1124 1125 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1126 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1127 return true; 1128} 1129 1130//------------------------------inline_hasNegatives------------------------------ 1131bool LibraryCallKit::inline_hasNegatives() { 1132 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1133 return false; 1134 } 1135 1136 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1137 // no receiver since it is static method 1138 Node* ba = argument(0); 1139 Node* offset = argument(1); 1140 Node* len = argument(2); 1141 1142 // Range checks 1143 generate_string_range_check(ba, offset, len, false); 1144 if (stopped()) { 1145 return true; 1146 } 1147 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1148 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1149 set_result(_gvn.transform(result)); 1150 return true; 1151} 1152 1153bool LibraryCallKit::inline_preconditions_checkIndex() { 1154 Node* index = argument(0); 1155 Node* length = argument(1); 1156 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1157 return false; 1158 } 1159 1160 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1161 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1162 1163 { 1164 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1165 uncommon_trap(Deoptimization::Reason_intrinsic, 1166 Deoptimization::Action_make_not_entrant); 1167 } 1168 1169 if (stopped()) { 1170 return false; 1171 } 1172 1173 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1174 BoolTest::mask btest = BoolTest::lt; 1175 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1176 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1177 _gvn.set_type(rc, rc->Value(&_gvn)); 1178 if (!rc_bool->is_Con()) { 1179 record_for_igvn(rc); 1180 } 1181 set_control(_gvn.transform(new IfTrueNode(rc))); 1182 { 1183 PreserveJVMState pjvms(this); 1184 set_control(_gvn.transform(new IfFalseNode(rc))); 1185 uncommon_trap(Deoptimization::Reason_range_check, 1186 Deoptimization::Action_make_not_entrant); 1187 } 1188 1189 if (stopped()) { 1190 return false; 1191 } 1192 1193 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1194 result->set_req(0, control()); 1195 result = _gvn.transform(result); 1196 set_result(result); 1197 replace_in_map(index, result); 1198 return true; 1199} 1200 1201//------------------------------inline_string_indexOf------------------------ 1202bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1203 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1204 return false; 1205 } 1206 Node* src = argument(0); 1207 Node* tgt = argument(1); 1208 1209 // Make the merge point 1210 RegionNode* result_rgn = new RegionNode(4); 1211 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1212 1213 // Get start addr and length of source string 1214 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1215 Node* src_count = load_array_length(src); 1216 1217 // Get start addr and length of substring 1218 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1219 Node* tgt_count = load_array_length(tgt); 1220 1221 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1222 // Divide src size by 2 if String is UTF16 encoded 1223 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1224 } 1225 if (ae == StrIntrinsicNode::UU) { 1226 // Divide substring size by 2 if String is UTF16 encoded 1227 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1228 } 1229 1230 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1231 if (result != NULL) { 1232 result_phi->init_req(3, result); 1233 result_rgn->init_req(3, control()); 1234 } 1235 set_control(_gvn.transform(result_rgn)); 1236 record_for_igvn(result_rgn); 1237 set_result(_gvn.transform(result_phi)); 1238 1239 return true; 1240} 1241 1242//-----------------------------inline_string_indexOf----------------------- 1243bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1244 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1245 return false; 1246 } 1247 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1248 return false; 1249 } 1250 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1251 Node* src = argument(0); // byte[] 1252 Node* src_count = argument(1); // char count 1253 Node* tgt = argument(2); // byte[] 1254 Node* tgt_count = argument(3); // char count 1255 Node* from_index = argument(4); // char index 1256 1257 // Multiply byte array index by 2 if String is UTF16 encoded 1258 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1259 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1260 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1261 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1262 1263 // Range checks 1264 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1265 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1266 if (stopped()) { 1267 return true; 1268 } 1269 1270 RegionNode* region = new RegionNode(5); 1271 Node* phi = new PhiNode(region, TypeInt::INT); 1272 1273 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1274 if (result != NULL) { 1275 // The result is index relative to from_index if substring was found, -1 otherwise. 1276 // Generate code which will fold into cmove. 1277 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1278 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1279 1280 Node* if_lt = generate_slow_guard(bol, NULL); 1281 if (if_lt != NULL) { 1282 // result == -1 1283 phi->init_req(3, result); 1284 region->init_req(3, if_lt); 1285 } 1286 if (!stopped()) { 1287 result = _gvn.transform(new AddINode(result, from_index)); 1288 phi->init_req(4, result); 1289 region->init_req(4, control()); 1290 } 1291 } 1292 1293 set_control(_gvn.transform(region)); 1294 record_for_igvn(region); 1295 set_result(_gvn.transform(phi)); 1296 1297 return true; 1298} 1299 1300// Create StrIndexOfNode with fast path checks 1301Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1302 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1303 // Check for substr count > string count 1304 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1305 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1306 Node* if_gt = generate_slow_guard(bol, NULL); 1307 if (if_gt != NULL) { 1308 phi->init_req(1, intcon(-1)); 1309 region->init_req(1, if_gt); 1310 } 1311 if (!stopped()) { 1312 // Check for substr count == 0 1313 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1314 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1315 Node* if_zero = generate_slow_guard(bol, NULL); 1316 if (if_zero != NULL) { 1317 phi->init_req(2, intcon(0)); 1318 region->init_req(2, if_zero); 1319 } 1320 } 1321 if (!stopped()) { 1322 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1323 } 1324 return NULL; 1325} 1326 1327//-----------------------------inline_string_indexOfChar----------------------- 1328bool LibraryCallKit::inline_string_indexOfChar() { 1329 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1330 return false; 1331 } 1332 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1333 return false; 1334 } 1335 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1336 Node* src = argument(0); // byte[] 1337 Node* tgt = argument(1); // tgt is int ch 1338 Node* from_index = argument(2); 1339 Node* max = argument(3); 1340 1341 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1342 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1343 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1344 1345 // Range checks 1346 generate_string_range_check(src, src_offset, src_count, true); 1347 if (stopped()) { 1348 return true; 1349 } 1350 1351 RegionNode* region = new RegionNode(3); 1352 Node* phi = new PhiNode(region, TypeInt::INT); 1353 1354 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1355 C->set_has_split_ifs(true); // Has chance for split-if optimization 1356 _gvn.transform(result); 1357 1358 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1359 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1360 1361 Node* if_lt = generate_slow_guard(bol, NULL); 1362 if (if_lt != NULL) { 1363 // result == -1 1364 phi->init_req(2, result); 1365 region->init_req(2, if_lt); 1366 } 1367 if (!stopped()) { 1368 result = _gvn.transform(new AddINode(result, from_index)); 1369 phi->init_req(1, result); 1370 region->init_req(1, control()); 1371 } 1372 set_control(_gvn.transform(region)); 1373 record_for_igvn(region); 1374 set_result(_gvn.transform(phi)); 1375 1376 return true; 1377} 1378//---------------------------inline_string_copy--------------------- 1379// compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1380// int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1381// int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1382// compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1383// void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1384// void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1385bool LibraryCallKit::inline_string_copy(bool compress) { 1386 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1387 return false; 1388 } 1389 int nargs = 5; // 2 oops, 3 ints 1390 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1391 1392 Node* src = argument(0); 1393 Node* src_offset = argument(1); 1394 Node* dst = argument(2); 1395 Node* dst_offset = argument(3); 1396 Node* length = argument(4); 1397 1398 // Check for allocation before we add nodes that would confuse 1399 // tightly_coupled_allocation() 1400 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1401 1402 // Figure out the size and type of the elements we will be copying. 1403 const Type* src_type = src->Value(&_gvn); 1404 const Type* dst_type = dst->Value(&_gvn); 1405 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1406 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1407 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1408 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1409 "Unsupported array types for inline_string_copy"); 1410 1411 // Convert char[] offsets to byte[] offsets 1412 bool convert_src = (compress && src_elem == T_BYTE); 1413 bool convert_dst = (!compress && dst_elem == T_BYTE); 1414 if (convert_src) { 1415 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1416 } else if (convert_dst) { 1417 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1418 } 1419 1420 // Range checks 1421 generate_string_range_check(src, src_offset, length, convert_src); 1422 generate_string_range_check(dst, dst_offset, length, convert_dst); 1423 if (stopped()) { 1424 return true; 1425 } 1426 1427 Node* src_start = array_element_address(src, src_offset, src_elem); 1428 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1429 // 'src_start' points to src array + scaled offset 1430 // 'dst_start' points to dst array + scaled offset 1431 Node* count = NULL; 1432 if (compress) { 1433 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1434 } else { 1435 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1436 } 1437 1438 if (alloc != NULL) { 1439 if (alloc->maybe_set_complete(&_gvn)) { 1440 // "You break it, you buy it." 1441 InitializeNode* init = alloc->initialization(); 1442 assert(init->is_complete(), "we just did this"); 1443 init->set_complete_with_arraycopy(); 1444 assert(dst->is_CheckCastPP(), "sanity"); 1445 assert(dst->in(0)->in(0) == init, "dest pinned"); 1446 } 1447 // Do not let stores that initialize this object be reordered with 1448 // a subsequent store that would make this object accessible by 1449 // other threads. 1450 // Record what AllocateNode this StoreStore protects so that 1451 // escape analysis can go from the MemBarStoreStoreNode to the 1452 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1453 // based on the escape status of the AllocateNode. 1454 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1455 } 1456 if (compress) { 1457 set_result(_gvn.transform(count)); 1458 } 1459 return true; 1460} 1461 1462#ifdef _LP64 1463#define XTOP ,top() /*additional argument*/ 1464#else //_LP64 1465#define XTOP /*no additional argument*/ 1466#endif //_LP64 1467 1468//------------------------inline_string_toBytesU-------------------------- 1469// public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1470bool LibraryCallKit::inline_string_toBytesU() { 1471 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1472 return false; 1473 } 1474 // Get the arguments. 1475 Node* value = argument(0); 1476 Node* offset = argument(1); 1477 Node* length = argument(2); 1478 1479 Node* newcopy = NULL; 1480 1481 // Set the original stack and the reexecute bit for the interpreter to reexecute 1482 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1483 { PreserveReexecuteState preexecs(this); 1484 jvms()->set_should_reexecute(true); 1485 1486 // Check if a null path was taken unconditionally. 1487 value = null_check(value); 1488 1489 RegionNode* bailout = new RegionNode(1); 1490 record_for_igvn(bailout); 1491 1492 // Range checks 1493 generate_negative_guard(offset, bailout); 1494 generate_negative_guard(length, bailout); 1495 generate_limit_guard(offset, length, load_array_length(value), bailout); 1496 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1497 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1498 1499 if (bailout->req() > 1) { 1500 PreserveJVMState pjvms(this); 1501 set_control(_gvn.transform(bailout)); 1502 uncommon_trap(Deoptimization::Reason_intrinsic, 1503 Deoptimization::Action_maybe_recompile); 1504 } 1505 if (stopped()) { 1506 return true; 1507 } 1508 1509 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1510 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1511 newcopy = new_array(klass_node, size, 0); // no arguments to push 1512 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1513 1514 // Calculate starting addresses. 1515 Node* src_start = array_element_address(value, offset, T_CHAR); 1516 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1517 1518 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1519 const TypeInt* toffset = gvn().type(offset)->is_int(); 1520 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1521 1522 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1523 const char* copyfunc_name = "arraycopy"; 1524 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1525 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1526 OptoRuntime::fast_arraycopy_Type(), 1527 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1528 src_start, dst_start, ConvI2X(length) XTOP); 1529 // Do not let reads from the cloned object float above the arraycopy. 1530 if (alloc != NULL) { 1531 if (alloc->maybe_set_complete(&_gvn)) { 1532 // "You break it, you buy it." 1533 InitializeNode* init = alloc->initialization(); 1534 assert(init->is_complete(), "we just did this"); 1535 init->set_complete_with_arraycopy(); 1536 assert(newcopy->is_CheckCastPP(), "sanity"); 1537 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1538 } 1539 // Do not let stores that initialize this object be reordered with 1540 // a subsequent store that would make this object accessible by 1541 // other threads. 1542 // Record what AllocateNode this StoreStore protects so that 1543 // escape analysis can go from the MemBarStoreStoreNode to the 1544 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1545 // based on the escape status of the AllocateNode. 1546 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1547 } else { 1548 insert_mem_bar(Op_MemBarCPUOrder); 1549 } 1550 } // original reexecute is set back here 1551 1552 C->set_has_split_ifs(true); // Has chance for split-if optimization 1553 if (!stopped()) { 1554 set_result(newcopy); 1555 } 1556 return true; 1557} 1558 1559//------------------------inline_string_getCharsU-------------------------- 1560// public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1561bool LibraryCallKit::inline_string_getCharsU() { 1562 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1563 return false; 1564 } 1565 1566 // Get the arguments. 1567 Node* src = argument(0); 1568 Node* src_begin = argument(1); 1569 Node* src_end = argument(2); // exclusive offset (i < src_end) 1570 Node* dst = argument(3); 1571 Node* dst_begin = argument(4); 1572 1573 // Check for allocation before we add nodes that would confuse 1574 // tightly_coupled_allocation() 1575 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1576 1577 // Check if a null path was taken unconditionally. 1578 src = null_check(src); 1579 dst = null_check(dst); 1580 if (stopped()) { 1581 return true; 1582 } 1583 1584 // Get length and convert char[] offset to byte[] offset 1585 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1586 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1587 1588 // Range checks 1589 generate_string_range_check(src, src_begin, length, true); 1590 generate_string_range_check(dst, dst_begin, length, false); 1591 if (stopped()) { 1592 return true; 1593 } 1594 1595 if (!stopped()) { 1596 // Calculate starting addresses. 1597 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1598 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1599 1600 // Check if array addresses are aligned to HeapWordSize 1601 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1602 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1603 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1604 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1605 1606 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1607 const char* copyfunc_name = "arraycopy"; 1608 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1609 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1610 OptoRuntime::fast_arraycopy_Type(), 1611 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1612 src_start, dst_start, ConvI2X(length) XTOP); 1613 // Do not let reads from the cloned object float above the arraycopy. 1614 if (alloc != NULL) { 1615 if (alloc->maybe_set_complete(&_gvn)) { 1616 // "You break it, you buy it." 1617 InitializeNode* init = alloc->initialization(); 1618 assert(init->is_complete(), "we just did this"); 1619 init->set_complete_with_arraycopy(); 1620 assert(dst->is_CheckCastPP(), "sanity"); 1621 assert(dst->in(0)->in(0) == init, "dest pinned"); 1622 } 1623 // Do not let stores that initialize this object be reordered with 1624 // a subsequent store that would make this object accessible by 1625 // other threads. 1626 // Record what AllocateNode this StoreStore protects so that 1627 // escape analysis can go from the MemBarStoreStoreNode to the 1628 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1629 // based on the escape status of the AllocateNode. 1630 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1631 } else { 1632 insert_mem_bar(Op_MemBarCPUOrder); 1633 } 1634 } 1635 1636 C->set_has_split_ifs(true); // Has chance for split-if optimization 1637 return true; 1638} 1639 1640//----------------------inline_string_char_access---------------------------- 1641// Store/Load char to/from byte[] array. 1642// static void StringUTF16.putChar(byte[] val, int index, int c) 1643// static char StringUTF16.getChar(byte[] val, int index) 1644bool LibraryCallKit::inline_string_char_access(bool is_store) { 1645 Node* value = argument(0); 1646 Node* index = argument(1); 1647 Node* ch = is_store ? argument(2) : NULL; 1648 1649 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1650 // correctly requires matched array shapes. 1651 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1652 "sanity: byte[] and char[] bases agree"); 1653 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1654 "sanity: byte[] and char[] scales agree"); 1655 1656 // Bail when getChar over constants is requested: constant folding would 1657 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1658 // Java method would constant fold nicely instead. 1659 if (!is_store && value->is_Con() && index->is_Con()) { 1660 return false; 1661 } 1662 1663 Node* adr = array_element_address(value, index, T_CHAR); 1664 if (is_store) { 1665 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1666 false, false, true /* mismatched */); 1667 } else { 1668 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1669 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */); 1670 set_result(ch); 1671 } 1672 return true; 1673} 1674 1675//--------------------------round_double_node-------------------------------- 1676// Round a double node if necessary. 1677Node* LibraryCallKit::round_double_node(Node* n) { 1678 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1679 n = _gvn.transform(new RoundDoubleNode(0, n)); 1680 return n; 1681} 1682 1683//------------------------------inline_math----------------------------------- 1684// public static double Math.abs(double) 1685// public static double Math.sqrt(double) 1686// public static double Math.log(double) 1687// public static double Math.log10(double) 1688bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1689 Node* arg = round_double_node(argument(0)); 1690 Node* n = NULL; 1691 switch (id) { 1692 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1693 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1694 default: fatal_unexpected_iid(id); break; 1695 } 1696 set_result(_gvn.transform(n)); 1697 return true; 1698} 1699 1700//------------------------------runtime_math----------------------------- 1701bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1702 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1703 "must be (DD)D or (D)D type"); 1704 1705 // Inputs 1706 Node* a = round_double_node(argument(0)); 1707 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1708 1709 const TypePtr* no_memory_effects = NULL; 1710 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1711 no_memory_effects, 1712 a, top(), b, b ? top() : NULL); 1713 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1714#ifdef ASSERT 1715 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1716 assert(value_top == top(), "second value must be top"); 1717#endif 1718 1719 set_result(value); 1720 return true; 1721} 1722 1723//------------------------------inline_math_native----------------------------- 1724bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1725#define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1726 switch (id) { 1727 // These intrinsics are not properly supported on all hardware 1728 case vmIntrinsics::_dsin: 1729 return StubRoutines::dsin() != NULL ? 1730 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1731 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1732 case vmIntrinsics::_dcos: 1733 return StubRoutines::dcos() != NULL ? 1734 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1735 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1736 case vmIntrinsics::_dtan: 1737 return StubRoutines::dtan() != NULL ? 1738 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1739 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1740 case vmIntrinsics::_dlog: 1741 return StubRoutines::dlog() != NULL ? 1742 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1743 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1744 case vmIntrinsics::_dlog10: 1745 return StubRoutines::dlog10() != NULL ? 1746 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1747 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1748 1749 // These intrinsics are supported on all hardware 1750 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1751 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1752 1753 case vmIntrinsics::_dexp: 1754 return StubRoutines::dexp() != NULL ? 1755 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1756 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1757 case vmIntrinsics::_dpow: 1758 return StubRoutines::dpow() != NULL ? 1759 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1760 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1761#undef FN_PTR 1762 1763 // These intrinsics are not yet correctly implemented 1764 case vmIntrinsics::_datan2: 1765 return false; 1766 1767 default: 1768 fatal_unexpected_iid(id); 1769 return false; 1770 } 1771} 1772 1773static bool is_simple_name(Node* n) { 1774 return (n->req() == 1 // constant 1775 || (n->is_Type() && n->as_Type()->type()->singleton()) 1776 || n->is_Proj() // parameter or return value 1777 || n->is_Phi() // local of some sort 1778 ); 1779} 1780 1781//----------------------------inline_notify-----------------------------------* 1782bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1783 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1784 address func; 1785 if (id == vmIntrinsics::_notify) { 1786 func = OptoRuntime::monitor_notify_Java(); 1787 } else { 1788 func = OptoRuntime::monitor_notifyAll_Java(); 1789 } 1790 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1791 make_slow_call_ex(call, env()->Throwable_klass(), false); 1792 return true; 1793} 1794 1795 1796//----------------------------inline_min_max----------------------------------- 1797bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1798 set_result(generate_min_max(id, argument(0), argument(1))); 1799 return true; 1800} 1801 1802void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1803 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1804 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1805 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1806 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1807 1808 { 1809 PreserveJVMState pjvms(this); 1810 PreserveReexecuteState preexecs(this); 1811 jvms()->set_should_reexecute(true); 1812 1813 set_control(slow_path); 1814 set_i_o(i_o()); 1815 1816 uncommon_trap(Deoptimization::Reason_intrinsic, 1817 Deoptimization::Action_none); 1818 } 1819 1820 set_control(fast_path); 1821 set_result(math); 1822} 1823 1824template <typename OverflowOp> 1825bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1826 typedef typename OverflowOp::MathOp MathOp; 1827 1828 MathOp* mathOp = new MathOp(arg1, arg2); 1829 Node* operation = _gvn.transform( mathOp ); 1830 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1831 inline_math_mathExact(operation, ofcheck); 1832 return true; 1833} 1834 1835bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1836 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1837} 1838 1839bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1840 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1841} 1842 1843bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1844 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1845} 1846 1847bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1848 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1849} 1850 1851bool LibraryCallKit::inline_math_negateExactI() { 1852 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1853} 1854 1855bool LibraryCallKit::inline_math_negateExactL() { 1856 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1857} 1858 1859bool LibraryCallKit::inline_math_multiplyExactI() { 1860 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1861} 1862 1863bool LibraryCallKit::inline_math_multiplyExactL() { 1864 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1865} 1866 1867Node* 1868LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1869 // These are the candidate return value: 1870 Node* xvalue = x0; 1871 Node* yvalue = y0; 1872 1873 if (xvalue == yvalue) { 1874 return xvalue; 1875 } 1876 1877 bool want_max = (id == vmIntrinsics::_max); 1878 1879 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1880 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1881 if (txvalue == NULL || tyvalue == NULL) return top(); 1882 // This is not really necessary, but it is consistent with a 1883 // hypothetical MaxINode::Value method: 1884 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1885 1886 // %%% This folding logic should (ideally) be in a different place. 1887 // Some should be inside IfNode, and there to be a more reliable 1888 // transformation of ?: style patterns into cmoves. We also want 1889 // more powerful optimizations around cmove and min/max. 1890 1891 // Try to find a dominating comparison of these guys. 1892 // It can simplify the index computation for Arrays.copyOf 1893 // and similar uses of System.arraycopy. 1894 // First, compute the normalized version of CmpI(x, y). 1895 int cmp_op = Op_CmpI; 1896 Node* xkey = xvalue; 1897 Node* ykey = yvalue; 1898 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 1899 if (ideal_cmpxy->is_Cmp()) { 1900 // E.g., if we have CmpI(length - offset, count), 1901 // it might idealize to CmpI(length, count + offset) 1902 cmp_op = ideal_cmpxy->Opcode(); 1903 xkey = ideal_cmpxy->in(1); 1904 ykey = ideal_cmpxy->in(2); 1905 } 1906 1907 // Start by locating any relevant comparisons. 1908 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1909 Node* cmpxy = NULL; 1910 Node* cmpyx = NULL; 1911 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1912 Node* cmp = start_from->fast_out(k); 1913 if (cmp->outcnt() > 0 && // must have prior uses 1914 cmp->in(0) == NULL && // must be context-independent 1915 cmp->Opcode() == cmp_op) { // right kind of compare 1916 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1917 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1918 } 1919 } 1920 1921 const int NCMPS = 2; 1922 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1923 int cmpn; 1924 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1925 if (cmps[cmpn] != NULL) break; // find a result 1926 } 1927 if (cmpn < NCMPS) { 1928 // Look for a dominating test that tells us the min and max. 1929 int depth = 0; // Limit search depth for speed 1930 Node* dom = control(); 1931 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1932 if (++depth >= 100) break; 1933 Node* ifproj = dom; 1934 if (!ifproj->is_Proj()) continue; 1935 Node* iff = ifproj->in(0); 1936 if (!iff->is_If()) continue; 1937 Node* bol = iff->in(1); 1938 if (!bol->is_Bool()) continue; 1939 Node* cmp = bol->in(1); 1940 if (cmp == NULL) continue; 1941 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1942 if (cmps[cmpn] == cmp) break; 1943 if (cmpn == NCMPS) continue; 1944 BoolTest::mask btest = bol->as_Bool()->_test._test; 1945 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1946 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1947 // At this point, we know that 'x btest y' is true. 1948 switch (btest) { 1949 case BoolTest::eq: 1950 // They are proven equal, so we can collapse the min/max. 1951 // Either value is the answer. Choose the simpler. 1952 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1953 return yvalue; 1954 return xvalue; 1955 case BoolTest::lt: // x < y 1956 case BoolTest::le: // x <= y 1957 return (want_max ? yvalue : xvalue); 1958 case BoolTest::gt: // x > y 1959 case BoolTest::ge: // x >= y 1960 return (want_max ? xvalue : yvalue); 1961 } 1962 } 1963 } 1964 1965 // We failed to find a dominating test. 1966 // Let's pick a test that might GVN with prior tests. 1967 Node* best_bol = NULL; 1968 BoolTest::mask best_btest = BoolTest::illegal; 1969 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1970 Node* cmp = cmps[cmpn]; 1971 if (cmp == NULL) continue; 1972 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1973 Node* bol = cmp->fast_out(j); 1974 if (!bol->is_Bool()) continue; 1975 BoolTest::mask btest = bol->as_Bool()->_test._test; 1976 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1977 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1978 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1979 best_bol = bol->as_Bool(); 1980 best_btest = btest; 1981 } 1982 } 1983 } 1984 1985 Node* answer_if_true = NULL; 1986 Node* answer_if_false = NULL; 1987 switch (best_btest) { 1988 default: 1989 if (cmpxy == NULL) 1990 cmpxy = ideal_cmpxy; 1991 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 1992 // and fall through: 1993 case BoolTest::lt: // x < y 1994 case BoolTest::le: // x <= y 1995 answer_if_true = (want_max ? yvalue : xvalue); 1996 answer_if_false = (want_max ? xvalue : yvalue); 1997 break; 1998 case BoolTest::gt: // x > y 1999 case BoolTest::ge: // x >= y 2000 answer_if_true = (want_max ? xvalue : yvalue); 2001 answer_if_false = (want_max ? yvalue : xvalue); 2002 break; 2003 } 2004 2005 jint hi, lo; 2006 if (want_max) { 2007 // We can sharpen the minimum. 2008 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2009 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2010 } else { 2011 // We can sharpen the maximum. 2012 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2013 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2014 } 2015 2016 // Use a flow-free graph structure, to avoid creating excess control edges 2017 // which could hinder other optimizations. 2018 // Since Math.min/max is often used with arraycopy, we want 2019 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2020 Node* cmov = CMoveNode::make(NULL, best_bol, 2021 answer_if_false, answer_if_true, 2022 TypeInt::make(lo, hi, widen)); 2023 2024 return _gvn.transform(cmov); 2025 2026 /* 2027 // This is not as desirable as it may seem, since Min and Max 2028 // nodes do not have a full set of optimizations. 2029 // And they would interfere, anyway, with 'if' optimizations 2030 // and with CMoveI canonical forms. 2031 switch (id) { 2032 case vmIntrinsics::_min: 2033 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2034 case vmIntrinsics::_max: 2035 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2036 default: 2037 ShouldNotReachHere(); 2038 } 2039 */ 2040} 2041 2042inline int 2043LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2044 const TypePtr* base_type = TypePtr::NULL_PTR; 2045 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2046 if (base_type == NULL) { 2047 // Unknown type. 2048 return Type::AnyPtr; 2049 } else if (base_type == TypePtr::NULL_PTR) { 2050 // Since this is a NULL+long form, we have to switch to a rawptr. 2051 base = _gvn.transform(new CastX2PNode(offset)); 2052 offset = MakeConX(0); 2053 return Type::RawPtr; 2054 } else if (base_type->base() == Type::RawPtr) { 2055 return Type::RawPtr; 2056 } else if (base_type->isa_oopptr()) { 2057 // Base is never null => always a heap address. 2058 if (base_type->ptr() == TypePtr::NotNull) { 2059 return Type::OopPtr; 2060 } 2061 // Offset is small => always a heap address. 2062 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2063 if (offset_type != NULL && 2064 base_type->offset() == 0 && // (should always be?) 2065 offset_type->_lo >= 0 && 2066 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2067 return Type::OopPtr; 2068 } 2069 // Otherwise, it might either be oop+off or NULL+addr. 2070 return Type::AnyPtr; 2071 } else { 2072 // No information: 2073 return Type::AnyPtr; 2074 } 2075} 2076 2077inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2078 int kind = classify_unsafe_addr(base, offset); 2079 if (kind == Type::RawPtr) { 2080 return basic_plus_adr(top(), base, offset); 2081 } else { 2082 return basic_plus_adr(base, offset); 2083 } 2084} 2085 2086//--------------------------inline_number_methods----------------------------- 2087// inline int Integer.numberOfLeadingZeros(int) 2088// inline int Long.numberOfLeadingZeros(long) 2089// 2090// inline int Integer.numberOfTrailingZeros(int) 2091// inline int Long.numberOfTrailingZeros(long) 2092// 2093// inline int Integer.bitCount(int) 2094// inline int Long.bitCount(long) 2095// 2096// inline char Character.reverseBytes(char) 2097// inline short Short.reverseBytes(short) 2098// inline int Integer.reverseBytes(int) 2099// inline long Long.reverseBytes(long) 2100bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2101 Node* arg = argument(0); 2102 Node* n = NULL; 2103 switch (id) { 2104 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2105 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2106 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2107 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2108 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2109 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2110 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2111 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2112 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2113 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2114 default: fatal_unexpected_iid(id); break; 2115 } 2116 set_result(_gvn.transform(n)); 2117 return true; 2118} 2119 2120//----------------------------inline_unsafe_access---------------------------- 2121 2122// Helper that guards and inserts a pre-barrier. 2123void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2124 Node* pre_val, bool need_mem_bar) { 2125 // We could be accessing the referent field of a reference object. If so, when G1 2126 // is enabled, we need to log the value in the referent field in an SATB buffer. 2127 // This routine performs some compile time filters and generates suitable 2128 // runtime filters that guard the pre-barrier code. 2129 // Also add memory barrier for non volatile load from the referent field 2130 // to prevent commoning of loads across safepoint. 2131 if (!UseG1GC && !need_mem_bar) 2132 return; 2133 2134 // Some compile time checks. 2135 2136 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2137 const TypeX* otype = offset->find_intptr_t_type(); 2138 if (otype != NULL && otype->is_con() && 2139 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2140 // Constant offset but not the reference_offset so just return 2141 return; 2142 } 2143 2144 // We only need to generate the runtime guards for instances. 2145 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2146 if (btype != NULL) { 2147 if (btype->isa_aryptr()) { 2148 // Array type so nothing to do 2149 return; 2150 } 2151 2152 const TypeInstPtr* itype = btype->isa_instptr(); 2153 if (itype != NULL) { 2154 // Can the klass of base_oop be statically determined to be 2155 // _not_ a sub-class of Reference and _not_ Object? 2156 ciKlass* klass = itype->klass(); 2157 if ( klass->is_loaded() && 2158 !klass->is_subtype_of(env()->Reference_klass()) && 2159 !env()->Object_klass()->is_subtype_of(klass)) { 2160 return; 2161 } 2162 } 2163 } 2164 2165 // The compile time filters did not reject base_oop/offset so 2166 // we need to generate the following runtime filters 2167 // 2168 // if (offset == java_lang_ref_Reference::_reference_offset) { 2169 // if (instance_of(base, java.lang.ref.Reference)) { 2170 // pre_barrier(_, pre_val, ...); 2171 // } 2172 // } 2173 2174 float likely = PROB_LIKELY( 0.999); 2175 float unlikely = PROB_UNLIKELY(0.999); 2176 2177 IdealKit ideal(this); 2178#define __ ideal. 2179 2180 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2181 2182 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2183 // Update graphKit memory and control from IdealKit. 2184 sync_kit(ideal); 2185 2186 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2187 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2188 2189 // Update IdealKit memory and control from graphKit. 2190 __ sync_kit(this); 2191 2192 Node* one = __ ConI(1); 2193 // is_instof == 0 if base_oop == NULL 2194 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2195 2196 // Update graphKit from IdeakKit. 2197 sync_kit(ideal); 2198 2199 // Use the pre-barrier to record the value in the referent field 2200 pre_barrier(false /* do_load */, 2201 __ ctrl(), 2202 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2203 pre_val /* pre_val */, 2204 T_OBJECT); 2205 if (need_mem_bar) { 2206 // Add memory barrier to prevent commoning reads from this field 2207 // across safepoint since GC can change its value. 2208 insert_mem_bar(Op_MemBarCPUOrder); 2209 } 2210 // Update IdealKit from graphKit. 2211 __ sync_kit(this); 2212 2213 } __ end_if(); // _ref_type != ref_none 2214 } __ end_if(); // offset == referent_offset 2215 2216 // Final sync IdealKit and GraphKit. 2217 final_sync(ideal); 2218#undef __ 2219} 2220 2221 2222const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2223 // Attempt to infer a sharper value type from the offset and base type. 2224 ciKlass* sharpened_klass = NULL; 2225 2226 // See if it is an instance field, with an object type. 2227 if (alias_type->field() != NULL) { 2228 if (alias_type->field()->type()->is_klass()) { 2229 sharpened_klass = alias_type->field()->type()->as_klass(); 2230 } 2231 } 2232 2233 // See if it is a narrow oop array. 2234 if (adr_type->isa_aryptr()) { 2235 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2236 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2237 if (elem_type != NULL) { 2238 sharpened_klass = elem_type->klass(); 2239 } 2240 } 2241 } 2242 2243 // The sharpened class might be unloaded if there is no class loader 2244 // contraint in place. 2245 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2246 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2247 2248#ifndef PRODUCT 2249 if (C->print_intrinsics() || C->print_inlining()) { 2250 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2251 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2252 } 2253#endif 2254 // Sharpen the value type. 2255 return tjp; 2256 } 2257 return NULL; 2258} 2259 2260bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2261 if (callee()->is_static()) return false; // caller must have the capability! 2262 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2263 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2264 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2265 2266#ifndef PRODUCT 2267 { 2268 ResourceMark rm; 2269 // Check the signatures. 2270 ciSignature* sig = callee()->signature(); 2271#ifdef ASSERT 2272 if (!is_store) { 2273 // Object getObject(Object base, int/long offset), etc. 2274 BasicType rtype = sig->return_type()->basic_type(); 2275 assert(rtype == type, "getter must return the expected value"); 2276 assert(sig->count() == 2, "oop getter has 2 arguments"); 2277 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2278 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2279 } else { 2280 // void putObject(Object base, int/long offset, Object x), etc. 2281 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2282 assert(sig->count() == 3, "oop putter has 3 arguments"); 2283 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2284 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2285 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2286 assert(vtype == type, "putter must accept the expected value"); 2287 } 2288#endif // ASSERT 2289 } 2290#endif //PRODUCT 2291 2292 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2293 2294 Node* receiver = argument(0); // type: oop 2295 2296 // Build address expression. 2297 Node* adr; 2298 Node* heap_base_oop = top(); 2299 Node* offset = top(); 2300 Node* val; 2301 2302 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2303 Node* base = argument(1); // type: oop 2304 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2305 offset = argument(2); // type: long 2306 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2307 // to be plain byte offsets, which are also the same as those accepted 2308 // by oopDesc::field_base. 2309 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2310 "fieldOffset must be byte-scaled"); 2311 // 32-bit machines ignore the high half! 2312 offset = ConvL2X(offset); 2313 adr = make_unsafe_address(base, offset); 2314 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2315 heap_base_oop = base; 2316 } else if (type == T_OBJECT) { 2317 return false; // off-heap oop accesses are not supported 2318 } 2319 2320 // Can base be NULL? Otherwise, always on-heap access. 2321 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop)); 2322 2323 val = is_store ? argument(4) : NULL; 2324 2325 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2326 2327 // Try to categorize the address. 2328 Compile::AliasType* alias_type = C->alias_type(adr_type); 2329 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2330 2331 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2332 alias_type->adr_type() == TypeAryPtr::RANGE) { 2333 return false; // not supported 2334 } 2335 2336 bool mismatched = false; 2337 BasicType bt = alias_type->basic_type(); 2338 if (bt != T_ILLEGAL) { 2339 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2340 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2341 // Alias type doesn't differentiate between byte[] and boolean[]). 2342 // Use address type to get the element type. 2343 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2344 } 2345 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2346 // accessing an array field with getObject is not a mismatch 2347 bt = T_OBJECT; 2348 } 2349 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2350 // Don't intrinsify mismatched object accesses 2351 return false; 2352 } 2353 mismatched = (bt != type); 2354 } else if (alias_type->adr_type()->isa_oopptr()) { 2355 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2356 } 2357 2358 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2359 2360 // First guess at the value type. 2361 const Type *value_type = Type::get_const_basic_type(type); 2362 2363 // We will need memory barriers unless we can determine a unique 2364 // alias category for this reference. (Note: If for some reason 2365 // the barriers get omitted and the unsafe reference begins to "pollute" 2366 // the alias analysis of the rest of the graph, either Compile::can_alias 2367 // or Compile::must_alias will throw a diagnostic assert.) 2368 bool need_mem_bar; 2369 switch (kind) { 2370 case Relaxed: 2371 need_mem_bar = mismatched || can_access_non_heap; 2372 break; 2373 case Opaque: 2374 // Opaque uses CPUOrder membars for protection against code movement. 2375 case Acquire: 2376 case Release: 2377 case Volatile: 2378 need_mem_bar = true; 2379 break; 2380 default: 2381 ShouldNotReachHere(); 2382 } 2383 2384 // Some accesses require access atomicity for all types, notably longs and doubles. 2385 // When AlwaysAtomicAccesses is enabled, all accesses are atomic. 2386 bool requires_atomic_access = false; 2387 switch (kind) { 2388 case Relaxed: 2389 requires_atomic_access = AlwaysAtomicAccesses; 2390 break; 2391 case Opaque: 2392 // Opaque accesses are atomic. 2393 case Acquire: 2394 case Release: 2395 case Volatile: 2396 requires_atomic_access = true; 2397 break; 2398 default: 2399 ShouldNotReachHere(); 2400 } 2401 2402 // Figure out the memory ordering. 2403 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd 2404 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store); 2405 2406 // If we are reading the value of the referent field of a Reference 2407 // object (either by using Unsafe directly or through reflection) 2408 // then, if G1 is enabled, we need to record the referent in an 2409 // SATB log buffer using the pre-barrier mechanism. 2410 // Also we need to add memory barrier to prevent commoning reads 2411 // from this field across safepoint since GC can change its value. 2412 bool need_read_barrier = !is_store && 2413 offset != top() && heap_base_oop != top(); 2414 2415 if (!is_store && type == T_OBJECT) { 2416 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2417 if (tjp != NULL) { 2418 value_type = tjp; 2419 } 2420 } 2421 2422 receiver = null_check(receiver); 2423 if (stopped()) { 2424 return true; 2425 } 2426 // Heap pointers get a null-check from the interpreter, 2427 // as a courtesy. However, this is not guaranteed by Unsafe, 2428 // and it is not possible to fully distinguish unintended nulls 2429 // from intended ones in this API. 2430 2431 // We need to emit leading and trailing CPU membars (see below) in 2432 // addition to memory membars for special access modes. This is a little 2433 // too strong, but avoids the need to insert per-alias-type 2434 // volatile membars (for stores; compare Parse::do_put_xxx), which 2435 // we cannot do effectively here because we probably only have a 2436 // rough approximation of type. 2437 2438 switch(kind) { 2439 case Relaxed: 2440 case Opaque: 2441 case Acquire: 2442 break; 2443 case Release: 2444 case Volatile: 2445 if (is_store) { 2446 insert_mem_bar(Op_MemBarRelease); 2447 } else { 2448 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2449 insert_mem_bar(Op_MemBarVolatile); 2450 } 2451 } 2452 break; 2453 default: 2454 ShouldNotReachHere(); 2455 } 2456 2457 // Memory barrier to prevent normal and 'unsafe' accesses from 2458 // bypassing each other. Happens after null checks, so the 2459 // exception paths do not take memory state from the memory barrier, 2460 // so there's no problems making a strong assert about mixing users 2461 // of safe & unsafe memory. 2462 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2463 2464 if (!is_store) { 2465 Node* p = NULL; 2466 // Try to constant fold a load from a constant field 2467 ciField* field = alias_type->field(); 2468 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2469 // final or stable field 2470 p = make_constant_from_field(field, heap_base_oop); 2471 } 2472 if (p == NULL) { 2473 // To be valid, unsafe loads may depend on other conditions than 2474 // the one that guards them: pin the Load node 2475 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched); 2476 // load value 2477 switch (type) { 2478 case T_BOOLEAN: 2479 { 2480 // Normalize the value returned by getBoolean in the following cases 2481 if (mismatched || 2482 heap_base_oop == top() || // - heap_base_oop is NULL or 2483 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL 2484 // and the unsafe access is made to large offset 2485 // (i.e., larger than the maximum offset necessary for any 2486 // field access) 2487 ) { 2488 IdealKit ideal = IdealKit(this); 2489#define __ ideal. 2490 IdealVariable normalized_result(ideal); 2491 __ declarations_done(); 2492 __ set(normalized_result, p); 2493 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2494 __ set(normalized_result, ideal.ConI(1)); 2495 ideal.end_if(); 2496 final_sync(ideal); 2497 p = __ value(normalized_result); 2498#undef __ 2499 } 2500 } 2501 case T_CHAR: 2502 case T_BYTE: 2503 case T_SHORT: 2504 case T_INT: 2505 case T_LONG: 2506 case T_FLOAT: 2507 case T_DOUBLE: 2508 break; 2509 case T_OBJECT: 2510 if (need_read_barrier) { 2511 // We do not require a mem bar inside pre_barrier if need_mem_bar 2512 // is set: the barriers would be emitted by us. 2513 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar); 2514 } 2515 break; 2516 case T_ADDRESS: 2517 // Cast to an int type. 2518 p = _gvn.transform(new CastP2XNode(NULL, p)); 2519 p = ConvX2UL(p); 2520 break; 2521 default: 2522 fatal("unexpected type %d: %s", type, type2name(type)); 2523 break; 2524 } 2525 } 2526 // The load node has the control of the preceding MemBarCPUOrder. All 2527 // following nodes will have the control of the MemBarCPUOrder inserted at 2528 // the end of this method. So, pushing the load onto the stack at a later 2529 // point is fine. 2530 set_result(p); 2531 } else { 2532 // place effect of store into memory 2533 switch (type) { 2534 case T_DOUBLE: 2535 val = dstore_rounding(val); 2536 break; 2537 case T_ADDRESS: 2538 // Repackage the long as a pointer. 2539 val = ConvL2X(val); 2540 val = _gvn.transform(new CastX2PNode(val)); 2541 break; 2542 } 2543 2544 if (type == T_OBJECT) { 2545 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2546 } else { 2547 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched); 2548 } 2549 } 2550 2551 switch(kind) { 2552 case Relaxed: 2553 case Opaque: 2554 case Release: 2555 break; 2556 case Acquire: 2557 case Volatile: 2558 if (!is_store) { 2559 insert_mem_bar(Op_MemBarAcquire); 2560 } else { 2561 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2562 insert_mem_bar(Op_MemBarVolatile); 2563 } 2564 } 2565 break; 2566 default: 2567 ShouldNotReachHere(); 2568 } 2569 2570 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2571 2572 return true; 2573} 2574 2575//----------------------------inline_unsafe_load_store---------------------------- 2576// This method serves a couple of different customers (depending on LoadStoreKind): 2577// 2578// LS_cmp_swap: 2579// 2580// boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2581// boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2582// boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2583// 2584// LS_cmp_swap_weak: 2585// 2586// boolean weakCompareAndSwapObject( Object o, long offset, Object expected, Object x); 2587// boolean weakCompareAndSwapObjectAcquire(Object o, long offset, Object expected, Object x); 2588// boolean weakCompareAndSwapObjectRelease(Object o, long offset, Object expected, Object x); 2589// 2590// boolean weakCompareAndSwapInt( Object o, long offset, int expected, int x); 2591// boolean weakCompareAndSwapIntAcquire( Object o, long offset, int expected, int x); 2592// boolean weakCompareAndSwapIntRelease( Object o, long offset, int expected, int x); 2593// 2594// boolean weakCompareAndSwapLong( Object o, long offset, long expected, long x); 2595// boolean weakCompareAndSwapLongAcquire( Object o, long offset, long expected, long x); 2596// boolean weakCompareAndSwapLongRelease( Object o, long offset, long expected, long x); 2597// 2598// LS_cmp_exchange: 2599// 2600// Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x); 2601// Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x); 2602// Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x); 2603// 2604// Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2605// Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2606// Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2607// 2608// Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2609// Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2610// Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2611// 2612// LS_get_add: 2613// 2614// int getAndAddInt( Object o, long offset, int delta) 2615// long getAndAddLong(Object o, long offset, long delta) 2616// 2617// LS_get_set: 2618// 2619// int getAndSet(Object o, long offset, int newValue) 2620// long getAndSet(Object o, long offset, long newValue) 2621// Object getAndSet(Object o, long offset, Object newValue) 2622// 2623bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2624 // This basic scheme here is the same as inline_unsafe_access, but 2625 // differs in enough details that combining them would make the code 2626 // overly confusing. (This is a true fact! I originally combined 2627 // them, but even I was confused by it!) As much code/comments as 2628 // possible are retained from inline_unsafe_access though to make 2629 // the correspondences clearer. - dl 2630 2631 if (callee()->is_static()) return false; // caller must have the capability! 2632 2633#ifndef PRODUCT 2634 BasicType rtype; 2635 { 2636 ResourceMark rm; 2637 // Check the signatures. 2638 ciSignature* sig = callee()->signature(); 2639 rtype = sig->return_type()->basic_type(); 2640 switch(kind) { 2641 case LS_get_add: 2642 case LS_get_set: { 2643 // Check the signatures. 2644#ifdef ASSERT 2645 assert(rtype == type, "get and set must return the expected type"); 2646 assert(sig->count() == 3, "get and set has 3 arguments"); 2647 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2648 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2649 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2650 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2651#endif // ASSERT 2652 break; 2653 } 2654 case LS_cmp_swap: 2655 case LS_cmp_swap_weak: { 2656 // Check the signatures. 2657#ifdef ASSERT 2658 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2659 assert(sig->count() == 4, "CAS has 4 arguments"); 2660 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2661 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2662#endif // ASSERT 2663 break; 2664 } 2665 case LS_cmp_exchange: { 2666 // Check the signatures. 2667#ifdef ASSERT 2668 assert(rtype == type, "CAS must return the expected type"); 2669 assert(sig->count() == 4, "CAS has 4 arguments"); 2670 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2671 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2672#endif // ASSERT 2673 break; 2674 } 2675 default: 2676 ShouldNotReachHere(); 2677 } 2678 } 2679#endif //PRODUCT 2680 2681 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2682 2683 // Get arguments: 2684 Node* receiver = NULL; 2685 Node* base = NULL; 2686 Node* offset = NULL; 2687 Node* oldval = NULL; 2688 Node* newval = NULL; 2689 switch(kind) { 2690 case LS_cmp_swap: 2691 case LS_cmp_swap_weak: 2692 case LS_cmp_exchange: { 2693 const bool two_slot_type = type2size[type] == 2; 2694 receiver = argument(0); // type: oop 2695 base = argument(1); // type: oop 2696 offset = argument(2); // type: long 2697 oldval = argument(4); // type: oop, int, or long 2698 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2699 break; 2700 } 2701 case LS_get_add: 2702 case LS_get_set: { 2703 receiver = argument(0); // type: oop 2704 base = argument(1); // type: oop 2705 offset = argument(2); // type: long 2706 oldval = NULL; 2707 newval = argument(4); // type: oop, int, or long 2708 break; 2709 } 2710 default: 2711 ShouldNotReachHere(); 2712 } 2713 2714 // Build field offset expression. 2715 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2716 // to be plain byte offsets, which are also the same as those accepted 2717 // by oopDesc::field_base. 2718 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2719 // 32-bit machines ignore the high half of long offsets 2720 offset = ConvL2X(offset); 2721 Node* adr = make_unsafe_address(base, offset); 2722 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2723 2724 Compile::AliasType* alias_type = C->alias_type(adr_type); 2725 BasicType bt = alias_type->basic_type(); 2726 if (bt != T_ILLEGAL && 2727 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2728 // Don't intrinsify mismatched object accesses. 2729 return false; 2730 } 2731 2732 // For CAS, unlike inline_unsafe_access, there seems no point in 2733 // trying to refine types. Just use the coarse types here. 2734 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2735 const Type *value_type = Type::get_const_basic_type(type); 2736 2737 switch (kind) { 2738 case LS_get_set: 2739 case LS_cmp_exchange: { 2740 if (type == T_OBJECT) { 2741 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2742 if (tjp != NULL) { 2743 value_type = tjp; 2744 } 2745 } 2746 break; 2747 } 2748 case LS_cmp_swap: 2749 case LS_cmp_swap_weak: 2750 case LS_get_add: 2751 break; 2752 default: 2753 ShouldNotReachHere(); 2754 } 2755 2756 // Null check receiver. 2757 receiver = null_check(receiver); 2758 if (stopped()) { 2759 return true; 2760 } 2761 2762 int alias_idx = C->get_alias_index(adr_type); 2763 2764 // Memory-model-wise, a LoadStore acts like a little synchronized 2765 // block, so needs barriers on each side. These don't translate 2766 // into actual barriers on most machines, but we still need rest of 2767 // compiler to respect ordering. 2768 2769 switch (access_kind) { 2770 case Relaxed: 2771 case Acquire: 2772 break; 2773 case Release: 2774 insert_mem_bar(Op_MemBarRelease); 2775 break; 2776 case Volatile: 2777 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2778 insert_mem_bar(Op_MemBarVolatile); 2779 } else { 2780 insert_mem_bar(Op_MemBarRelease); 2781 } 2782 break; 2783 default: 2784 ShouldNotReachHere(); 2785 } 2786 insert_mem_bar(Op_MemBarCPUOrder); 2787 2788 // Figure out the memory ordering. 2789 MemNode::MemOrd mo = access_kind_to_memord(access_kind); 2790 2791 // 4984716: MemBars must be inserted before this 2792 // memory node in order to avoid a false 2793 // dependency which will confuse the scheduler. 2794 Node *mem = memory(alias_idx); 2795 2796 // For now, we handle only those cases that actually exist: ints, 2797 // longs, and Object. Adding others should be straightforward. 2798 Node* load_store = NULL; 2799 switch(type) { 2800 case T_BYTE: 2801 switch(kind) { 2802 case LS_get_add: 2803 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type)); 2804 break; 2805 case LS_get_set: 2806 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type)); 2807 break; 2808 case LS_cmp_swap_weak: 2809 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo)); 2810 break; 2811 case LS_cmp_swap: 2812 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo)); 2813 break; 2814 case LS_cmp_exchange: 2815 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2816 break; 2817 default: 2818 ShouldNotReachHere(); 2819 } 2820 break; 2821 case T_SHORT: 2822 switch(kind) { 2823 case LS_get_add: 2824 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type)); 2825 break; 2826 case LS_get_set: 2827 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type)); 2828 break; 2829 case LS_cmp_swap_weak: 2830 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo)); 2831 break; 2832 case LS_cmp_swap: 2833 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo)); 2834 break; 2835 case LS_cmp_exchange: 2836 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2837 break; 2838 default: 2839 ShouldNotReachHere(); 2840 } 2841 break; 2842 case T_INT: 2843 switch(kind) { 2844 case LS_get_add: 2845 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2846 break; 2847 case LS_get_set: 2848 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2849 break; 2850 case LS_cmp_swap_weak: 2851 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 2852 break; 2853 case LS_cmp_swap: 2854 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 2855 break; 2856 case LS_cmp_exchange: 2857 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo)); 2858 break; 2859 default: 2860 ShouldNotReachHere(); 2861 } 2862 break; 2863 case T_LONG: 2864 switch(kind) { 2865 case LS_get_add: 2866 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2867 break; 2868 case LS_get_set: 2869 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2870 break; 2871 case LS_cmp_swap_weak: 2872 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 2873 break; 2874 case LS_cmp_swap: 2875 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 2876 break; 2877 case LS_cmp_exchange: 2878 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2879 break; 2880 default: 2881 ShouldNotReachHere(); 2882 } 2883 break; 2884 case T_OBJECT: 2885 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2886 // could be delayed during Parse (for example, in adjust_map_after_if()). 2887 // Execute transformation here to avoid barrier generation in such case. 2888 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2889 newval = _gvn.makecon(TypePtr::NULL_PTR); 2890 2891 // Reference stores need a store barrier. 2892 switch(kind) { 2893 case LS_get_set: { 2894 // If pre-barrier must execute before the oop store, old value will require do_load here. 2895 if (!can_move_pre_barrier()) { 2896 pre_barrier(true /* do_load*/, 2897 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2898 NULL /* pre_val*/, 2899 T_OBJECT); 2900 } // Else move pre_barrier to use load_store value, see below. 2901 break; 2902 } 2903 case LS_cmp_swap_weak: 2904 case LS_cmp_swap: 2905 case LS_cmp_exchange: { 2906 // Same as for newval above: 2907 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2908 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2909 } 2910 // The only known value which might get overwritten is oldval. 2911 pre_barrier(false /* do_load */, 2912 control(), NULL, NULL, max_juint, NULL, NULL, 2913 oldval /* pre_val */, 2914 T_OBJECT); 2915 break; 2916 } 2917 default: 2918 ShouldNotReachHere(); 2919 } 2920 2921#ifdef _LP64 2922 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2923 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2924 2925 switch(kind) { 2926 case LS_get_set: 2927 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop())); 2928 break; 2929 case LS_cmp_swap_weak: { 2930 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2931 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 2932 break; 2933 } 2934 case LS_cmp_swap: { 2935 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2936 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 2937 break; 2938 } 2939 case LS_cmp_exchange: { 2940 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2941 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 2942 break; 2943 } 2944 default: 2945 ShouldNotReachHere(); 2946 } 2947 } else 2948#endif 2949 switch (kind) { 2950 case LS_get_set: 2951 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2952 break; 2953 case LS_cmp_swap_weak: 2954 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 2955 break; 2956 case LS_cmp_swap: 2957 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 2958 break; 2959 case LS_cmp_exchange: 2960 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo)); 2961 break; 2962 default: 2963 ShouldNotReachHere(); 2964 } 2965 2966 // Emit the post barrier only when the actual store happened. This makes sense 2967 // to check only for LS_cmp_* that can fail to set the value. 2968 // LS_cmp_exchange does not produce any branches by default, so there is no 2969 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with 2970 // CompareAndExchange and move branches here, it would make sense to conditionalize 2971 // post_barriers for LS_cmp_exchange as well. 2972 // 2973 // CAS success path is marked more likely since we anticipate this is a performance 2974 // critical path, while CAS failure path can use the penalty for going through unlikely 2975 // path as backoff. Which is still better than doing a store barrier there. 2976 switch (kind) { 2977 case LS_get_set: 2978 case LS_cmp_exchange: { 2979 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2980 break; 2981 } 2982 case LS_cmp_swap_weak: 2983 case LS_cmp_swap: { 2984 IdealKit ideal(this); 2985 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 2986 sync_kit(ideal); 2987 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2988 ideal.sync_kit(this); 2989 } ideal.end_if(); 2990 final_sync(ideal); 2991 break; 2992 } 2993 default: 2994 ShouldNotReachHere(); 2995 } 2996 break; 2997 default: 2998 fatal("unexpected type %d: %s", type, type2name(type)); 2999 break; 3000 } 3001 3002 // SCMemProjNodes represent the memory state of a LoadStore. Their 3003 // main role is to prevent LoadStore nodes from being optimized away 3004 // when their results aren't used. 3005 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 3006 set_memory(proj, alias_idx); 3007 3008 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) { 3009#ifdef _LP64 3010 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3011 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 3012 } 3013#endif 3014 if (can_move_pre_barrier()) { 3015 // Don't need to load pre_val. The old value is returned by load_store. 3016 // The pre_barrier can execute after the xchg as long as no safepoint 3017 // gets inserted between them. 3018 pre_barrier(false /* do_load */, 3019 control(), NULL, NULL, max_juint, NULL, NULL, 3020 load_store /* pre_val */, 3021 T_OBJECT); 3022 } 3023 } 3024 3025 // Add the trailing membar surrounding the access 3026 insert_mem_bar(Op_MemBarCPUOrder); 3027 3028 switch (access_kind) { 3029 case Relaxed: 3030 case Release: 3031 break; // do nothing 3032 case Acquire: 3033 case Volatile: 3034 insert_mem_bar(Op_MemBarAcquire); 3035 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code 3036 break; 3037 default: 3038 ShouldNotReachHere(); 3039 } 3040 3041 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3042 set_result(load_store); 3043 return true; 3044} 3045 3046MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) { 3047 MemNode::MemOrd mo = MemNode::unset; 3048 switch(kind) { 3049 case Opaque: 3050 case Relaxed: mo = MemNode::unordered; break; 3051 case Acquire: mo = MemNode::acquire; break; 3052 case Release: mo = MemNode::release; break; 3053 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break; 3054 default: 3055 ShouldNotReachHere(); 3056 } 3057 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3058 return mo; 3059} 3060 3061MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) { 3062 MemNode::MemOrd mo = MemNode::unset; 3063 switch(kind) { 3064 case Opaque: 3065 case Relaxed: mo = MemNode::unordered; break; 3066 case Acquire: mo = MemNode::acquire; break; 3067 case Release: mo = MemNode::release; break; 3068 case Volatile: mo = MemNode::seqcst; break; 3069 default: 3070 ShouldNotReachHere(); 3071 } 3072 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3073 return mo; 3074} 3075 3076bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3077 // Regardless of form, don't allow previous ld/st to move down, 3078 // then issue acquire, release, or volatile mem_bar. 3079 insert_mem_bar(Op_MemBarCPUOrder); 3080 switch(id) { 3081 case vmIntrinsics::_loadFence: 3082 insert_mem_bar(Op_LoadFence); 3083 return true; 3084 case vmIntrinsics::_storeFence: 3085 insert_mem_bar(Op_StoreFence); 3086 return true; 3087 case vmIntrinsics::_fullFence: 3088 insert_mem_bar(Op_MemBarVolatile); 3089 return true; 3090 default: 3091 fatal_unexpected_iid(id); 3092 return false; 3093 } 3094} 3095 3096bool LibraryCallKit::inline_onspinwait() { 3097 insert_mem_bar(Op_OnSpinWait); 3098 return true; 3099} 3100 3101bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3102 if (!kls->is_Con()) { 3103 return true; 3104 } 3105 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3106 if (klsptr == NULL) { 3107 return true; 3108 } 3109 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3110 // don't need a guard for a klass that is already initialized 3111 return !ik->is_initialized(); 3112} 3113 3114//----------------------------inline_unsafe_allocate--------------------------- 3115// public native Object Unsafe.allocateInstance(Class<?> cls); 3116bool LibraryCallKit::inline_unsafe_allocate() { 3117 if (callee()->is_static()) return false; // caller must have the capability! 3118 3119 null_check_receiver(); // null-check, then ignore 3120 Node* cls = null_check(argument(1)); 3121 if (stopped()) return true; 3122 3123 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3124 kls = null_check(kls); 3125 if (stopped()) return true; // argument was like int.class 3126 3127 Node* test = NULL; 3128 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3129 // Note: The argument might still be an illegal value like 3130 // Serializable.class or Object[].class. The runtime will handle it. 3131 // But we must make an explicit check for initialization. 3132 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3133 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3134 // can generate code to load it as unsigned byte. 3135 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3136 Node* bits = intcon(InstanceKlass::fully_initialized); 3137 test = _gvn.transform(new SubINode(inst, bits)); 3138 // The 'test' is non-zero if we need to take a slow path. 3139 } 3140 3141 Node* obj = new_instance(kls, test); 3142 set_result(obj); 3143 return true; 3144} 3145 3146//------------------------inline_native_time_funcs-------------- 3147// inline code for System.currentTimeMillis() and System.nanoTime() 3148// these have the same type and signature 3149bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3150 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3151 const TypePtr* no_memory_effects = NULL; 3152 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3153 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3154#ifdef ASSERT 3155 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3156 assert(value_top == top(), "second value must be top"); 3157#endif 3158 set_result(value); 3159 return true; 3160} 3161 3162//------------------------inline_native_currentThread------------------ 3163bool LibraryCallKit::inline_native_currentThread() { 3164 Node* junk = NULL; 3165 set_result(generate_current_thread(junk)); 3166 return true; 3167} 3168 3169//------------------------inline_native_isInterrupted------------------ 3170// private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3171bool LibraryCallKit::inline_native_isInterrupted() { 3172 // Add a fast path to t.isInterrupted(clear_int): 3173 // (t == Thread.current() && 3174 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3175 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3176 // So, in the common case that the interrupt bit is false, 3177 // we avoid making a call into the VM. Even if the interrupt bit 3178 // is true, if the clear_int argument is false, we avoid the VM call. 3179 // However, if the receiver is not currentThread, we must call the VM, 3180 // because there must be some locking done around the operation. 3181 3182 // We only go to the fast case code if we pass two guards. 3183 // Paths which do not pass are accumulated in the slow_region. 3184 3185 enum { 3186 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3187 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3188 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3189 PATH_LIMIT 3190 }; 3191 3192 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3193 // out of the function. 3194 insert_mem_bar(Op_MemBarCPUOrder); 3195 3196 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3197 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3198 3199 RegionNode* slow_region = new RegionNode(1); 3200 record_for_igvn(slow_region); 3201 3202 // (a) Receiving thread must be the current thread. 3203 Node* rec_thr = argument(0); 3204 Node* tls_ptr = NULL; 3205 Node* cur_thr = generate_current_thread(tls_ptr); 3206 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3207 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3208 3209 generate_slow_guard(bol_thr, slow_region); 3210 3211 // (b) Interrupt bit on TLS must be false. 3212 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3213 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3214 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3215 3216 // Set the control input on the field _interrupted read to prevent it floating up. 3217 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3218 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3219 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3220 3221 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3222 3223 // First fast path: if (!TLS._interrupted) return false; 3224 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3225 result_rgn->init_req(no_int_result_path, false_bit); 3226 result_val->init_req(no_int_result_path, intcon(0)); 3227 3228 // drop through to next case 3229 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3230 3231#ifndef _WINDOWS 3232 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3233 Node* clr_arg = argument(1); 3234 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3235 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3236 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3237 3238 // Second fast path: ... else if (!clear_int) return true; 3239 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3240 result_rgn->init_req(no_clear_result_path, false_arg); 3241 result_val->init_req(no_clear_result_path, intcon(1)); 3242 3243 // drop through to next case 3244 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3245#else 3246 // To return true on Windows you must read the _interrupted field 3247 // and check the event state i.e. take the slow path. 3248#endif // _WINDOWS 3249 3250 // (d) Otherwise, go to the slow path. 3251 slow_region->add_req(control()); 3252 set_control( _gvn.transform(slow_region)); 3253 3254 if (stopped()) { 3255 // There is no slow path. 3256 result_rgn->init_req(slow_result_path, top()); 3257 result_val->init_req(slow_result_path, top()); 3258 } else { 3259 // non-virtual because it is a private non-static 3260 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3261 3262 Node* slow_val = set_results_for_java_call(slow_call); 3263 // this->control() comes from set_results_for_java_call 3264 3265 Node* fast_io = slow_call->in(TypeFunc::I_O); 3266 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3267 3268 // These two phis are pre-filled with copies of of the fast IO and Memory 3269 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3270 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3271 3272 result_rgn->init_req(slow_result_path, control()); 3273 result_io ->init_req(slow_result_path, i_o()); 3274 result_mem->init_req(slow_result_path, reset_memory()); 3275 result_val->init_req(slow_result_path, slow_val); 3276 3277 set_all_memory(_gvn.transform(result_mem)); 3278 set_i_o( _gvn.transform(result_io)); 3279 } 3280 3281 C->set_has_split_ifs(true); // Has chance for split-if optimization 3282 set_result(result_rgn, result_val); 3283 return true; 3284} 3285 3286//---------------------------load_mirror_from_klass---------------------------- 3287// Given a klass oop, load its java mirror (a java.lang.Class oop). 3288Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3289 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3290 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3291} 3292 3293//-----------------------load_klass_from_mirror_common------------------------- 3294// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3295// Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3296// and branch to the given path on the region. 3297// If never_see_null, take an uncommon trap on null, so we can optimistically 3298// compile for the non-null case. 3299// If the region is NULL, force never_see_null = true. 3300Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3301 bool never_see_null, 3302 RegionNode* region, 3303 int null_path, 3304 int offset) { 3305 if (region == NULL) never_see_null = true; 3306 Node* p = basic_plus_adr(mirror, offset); 3307 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3308 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3309 Node* null_ctl = top(); 3310 kls = null_check_oop(kls, &null_ctl, never_see_null); 3311 if (region != NULL) { 3312 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3313 region->init_req(null_path, null_ctl); 3314 } else { 3315 assert(null_ctl == top(), "no loose ends"); 3316 } 3317 return kls; 3318} 3319 3320//--------------------(inline_native_Class_query helpers)--------------------- 3321// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3322// Fall through if (mods & mask) == bits, take the guard otherwise. 3323Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3324 // Branch around if the given klass has the given modifier bit set. 3325 // Like generate_guard, adds a new path onto the region. 3326 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3327 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3328 Node* mask = intcon(modifier_mask); 3329 Node* bits = intcon(modifier_bits); 3330 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3331 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3332 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3333 return generate_fair_guard(bol, region); 3334} 3335Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3336 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3337} 3338 3339//-------------------------inline_native_Class_query------------------- 3340bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3341 const Type* return_type = TypeInt::BOOL; 3342 Node* prim_return_value = top(); // what happens if it's a primitive class? 3343 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3344 bool expect_prim = false; // most of these guys expect to work on refs 3345 3346 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3347 3348 Node* mirror = argument(0); 3349 Node* obj = top(); 3350 3351 switch (id) { 3352 case vmIntrinsics::_isInstance: 3353 // nothing is an instance of a primitive type 3354 prim_return_value = intcon(0); 3355 obj = argument(1); 3356 break; 3357 case vmIntrinsics::_getModifiers: 3358 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3359 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3360 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3361 break; 3362 case vmIntrinsics::_isInterface: 3363 prim_return_value = intcon(0); 3364 break; 3365 case vmIntrinsics::_isArray: 3366 prim_return_value = intcon(0); 3367 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3368 break; 3369 case vmIntrinsics::_isPrimitive: 3370 prim_return_value = intcon(1); 3371 expect_prim = true; // obviously 3372 break; 3373 case vmIntrinsics::_getSuperclass: 3374 prim_return_value = null(); 3375 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3376 break; 3377 case vmIntrinsics::_getClassAccessFlags: 3378 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3379 return_type = TypeInt::INT; // not bool! 6297094 3380 break; 3381 default: 3382 fatal_unexpected_iid(id); 3383 break; 3384 } 3385 3386 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3387 if (mirror_con == NULL) return false; // cannot happen? 3388 3389#ifndef PRODUCT 3390 if (C->print_intrinsics() || C->print_inlining()) { 3391 ciType* k = mirror_con->java_mirror_type(); 3392 if (k) { 3393 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3394 k->print_name(); 3395 tty->cr(); 3396 } 3397 } 3398#endif 3399 3400 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3401 RegionNode* region = new RegionNode(PATH_LIMIT); 3402 record_for_igvn(region); 3403 PhiNode* phi = new PhiNode(region, return_type); 3404 3405 // The mirror will never be null of Reflection.getClassAccessFlags, however 3406 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3407 // if it is. See bug 4774291. 3408 3409 // For Reflection.getClassAccessFlags(), the null check occurs in 3410 // the wrong place; see inline_unsafe_access(), above, for a similar 3411 // situation. 3412 mirror = null_check(mirror); 3413 // If mirror or obj is dead, only null-path is taken. 3414 if (stopped()) return true; 3415 3416 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3417 3418 // Now load the mirror's klass metaobject, and null-check it. 3419 // Side-effects region with the control path if the klass is null. 3420 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3421 // If kls is null, we have a primitive mirror. 3422 phi->init_req(_prim_path, prim_return_value); 3423 if (stopped()) { set_result(region, phi); return true; } 3424 bool safe_for_replace = (region->in(_prim_path) == top()); 3425 3426 Node* p; // handy temp 3427 Node* null_ctl; 3428 3429 // Now that we have the non-null klass, we can perform the real query. 3430 // For constant classes, the query will constant-fold in LoadNode::Value. 3431 Node* query_value = top(); 3432 switch (id) { 3433 case vmIntrinsics::_isInstance: 3434 // nothing is an instance of a primitive type 3435 query_value = gen_instanceof(obj, kls, safe_for_replace); 3436 break; 3437 3438 case vmIntrinsics::_getModifiers: 3439 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3440 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3441 break; 3442 3443 case vmIntrinsics::_isInterface: 3444 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3445 if (generate_interface_guard(kls, region) != NULL) 3446 // A guard was added. If the guard is taken, it was an interface. 3447 phi->add_req(intcon(1)); 3448 // If we fall through, it's a plain class. 3449 query_value = intcon(0); 3450 break; 3451 3452 case vmIntrinsics::_isArray: 3453 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3454 if (generate_array_guard(kls, region) != NULL) 3455 // A guard was added. If the guard is taken, it was an array. 3456 phi->add_req(intcon(1)); 3457 // If we fall through, it's a plain class. 3458 query_value = intcon(0); 3459 break; 3460 3461 case vmIntrinsics::_isPrimitive: 3462 query_value = intcon(0); // "normal" path produces false 3463 break; 3464 3465 case vmIntrinsics::_getSuperclass: 3466 // The rules here are somewhat unfortunate, but we can still do better 3467 // with random logic than with a JNI call. 3468 // Interfaces store null or Object as _super, but must report null. 3469 // Arrays store an intermediate super as _super, but must report Object. 3470 // Other types can report the actual _super. 3471 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3472 if (generate_interface_guard(kls, region) != NULL) 3473 // A guard was added. If the guard is taken, it was an interface. 3474 phi->add_req(null()); 3475 if (generate_array_guard(kls, region) != NULL) 3476 // A guard was added. If the guard is taken, it was an array. 3477 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3478 // If we fall through, it's a plain class. Get its _super. 3479 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3480 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3481 null_ctl = top(); 3482 kls = null_check_oop(kls, &null_ctl); 3483 if (null_ctl != top()) { 3484 // If the guard is taken, Object.superClass is null (both klass and mirror). 3485 region->add_req(null_ctl); 3486 phi ->add_req(null()); 3487 } 3488 if (!stopped()) { 3489 query_value = load_mirror_from_klass(kls); 3490 } 3491 break; 3492 3493 case vmIntrinsics::_getClassAccessFlags: 3494 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3495 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3496 break; 3497 3498 default: 3499 fatal_unexpected_iid(id); 3500 break; 3501 } 3502 3503 // Fall-through is the normal case of a query to a real class. 3504 phi->init_req(1, query_value); 3505 region->init_req(1, control()); 3506 3507 C->set_has_split_ifs(true); // Has chance for split-if optimization 3508 set_result(region, phi); 3509 return true; 3510} 3511 3512//-------------------------inline_Class_cast------------------- 3513bool LibraryCallKit::inline_Class_cast() { 3514 Node* mirror = argument(0); // Class 3515 Node* obj = argument(1); 3516 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3517 if (mirror_con == NULL) { 3518 return false; // dead path (mirror->is_top()). 3519 } 3520 if (obj == NULL || obj->is_top()) { 3521 return false; // dead path 3522 } 3523 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3524 3525 // First, see if Class.cast() can be folded statically. 3526 // java_mirror_type() returns non-null for compile-time Class constants. 3527 ciType* tm = mirror_con->java_mirror_type(); 3528 if (tm != NULL && tm->is_klass() && 3529 tp != NULL && tp->klass() != NULL) { 3530 if (!tp->klass()->is_loaded()) { 3531 // Don't use intrinsic when class is not loaded. 3532 return false; 3533 } else { 3534 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3535 if (static_res == Compile::SSC_always_true) { 3536 // isInstance() is true - fold the code. 3537 set_result(obj); 3538 return true; 3539 } else if (static_res == Compile::SSC_always_false) { 3540 // Don't use intrinsic, have to throw ClassCastException. 3541 // If the reference is null, the non-intrinsic bytecode will 3542 // be optimized appropriately. 3543 return false; 3544 } 3545 } 3546 } 3547 3548 // Bailout intrinsic and do normal inlining if exception path is frequent. 3549 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3550 return false; 3551 } 3552 3553 // Generate dynamic checks. 3554 // Class.cast() is java implementation of _checkcast bytecode. 3555 // Do checkcast (Parse::do_checkcast()) optimizations here. 3556 3557 mirror = null_check(mirror); 3558 // If mirror is dead, only null-path is taken. 3559 if (stopped()) { 3560 return true; 3561 } 3562 3563 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3564 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3565 RegionNode* region = new RegionNode(PATH_LIMIT); 3566 record_for_igvn(region); 3567 3568 // Now load the mirror's klass metaobject, and null-check it. 3569 // If kls is null, we have a primitive mirror and 3570 // nothing is an instance of a primitive type. 3571 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3572 3573 Node* res = top(); 3574 if (!stopped()) { 3575 Node* bad_type_ctrl = top(); 3576 // Do checkcast optimizations. 3577 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3578 region->init_req(_bad_type_path, bad_type_ctrl); 3579 } 3580 if (region->in(_prim_path) != top() || 3581 region->in(_bad_type_path) != top()) { 3582 // Let Interpreter throw ClassCastException. 3583 PreserveJVMState pjvms(this); 3584 set_control(_gvn.transform(region)); 3585 uncommon_trap(Deoptimization::Reason_intrinsic, 3586 Deoptimization::Action_maybe_recompile); 3587 } 3588 if (!stopped()) { 3589 set_result(res); 3590 } 3591 return true; 3592} 3593 3594 3595//--------------------------inline_native_subtype_check------------------------ 3596// This intrinsic takes the JNI calls out of the heart of 3597// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3598bool LibraryCallKit::inline_native_subtype_check() { 3599 // Pull both arguments off the stack. 3600 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3601 args[0] = argument(0); 3602 args[1] = argument(1); 3603 Node* klasses[2]; // corresponding Klasses: superk, subk 3604 klasses[0] = klasses[1] = top(); 3605 3606 enum { 3607 // A full decision tree on {superc is prim, subc is prim}: 3608 _prim_0_path = 1, // {P,N} => false 3609 // {P,P} & superc!=subc => false 3610 _prim_same_path, // {P,P} & superc==subc => true 3611 _prim_1_path, // {N,P} => false 3612 _ref_subtype_path, // {N,N} & subtype check wins => true 3613 _both_ref_path, // {N,N} & subtype check loses => false 3614 PATH_LIMIT 3615 }; 3616 3617 RegionNode* region = new RegionNode(PATH_LIMIT); 3618 Node* phi = new PhiNode(region, TypeInt::BOOL); 3619 record_for_igvn(region); 3620 3621 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3622 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3623 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3624 3625 // First null-check both mirrors and load each mirror's klass metaobject. 3626 int which_arg; 3627 for (which_arg = 0; which_arg <= 1; which_arg++) { 3628 Node* arg = args[which_arg]; 3629 arg = null_check(arg); 3630 if (stopped()) break; 3631 args[which_arg] = arg; 3632 3633 Node* p = basic_plus_adr(arg, class_klass_offset); 3634 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3635 klasses[which_arg] = _gvn.transform(kls); 3636 } 3637 3638 // Having loaded both klasses, test each for null. 3639 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3640 for (which_arg = 0; which_arg <= 1; which_arg++) { 3641 Node* kls = klasses[which_arg]; 3642 Node* null_ctl = top(); 3643 kls = null_check_oop(kls, &null_ctl, never_see_null); 3644 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3645 region->init_req(prim_path, null_ctl); 3646 if (stopped()) break; 3647 klasses[which_arg] = kls; 3648 } 3649 3650 if (!stopped()) { 3651 // now we have two reference types, in klasses[0..1] 3652 Node* subk = klasses[1]; // the argument to isAssignableFrom 3653 Node* superk = klasses[0]; // the receiver 3654 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3655 // now we have a successful reference subtype check 3656 region->set_req(_ref_subtype_path, control()); 3657 } 3658 3659 // If both operands are primitive (both klasses null), then 3660 // we must return true when they are identical primitives. 3661 // It is convenient to test this after the first null klass check. 3662 set_control(region->in(_prim_0_path)); // go back to first null check 3663 if (!stopped()) { 3664 // Since superc is primitive, make a guard for the superc==subc case. 3665 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3666 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3667 generate_guard(bol_eq, region, PROB_FAIR); 3668 if (region->req() == PATH_LIMIT+1) { 3669 // A guard was added. If the added guard is taken, superc==subc. 3670 region->swap_edges(PATH_LIMIT, _prim_same_path); 3671 region->del_req(PATH_LIMIT); 3672 } 3673 region->set_req(_prim_0_path, control()); // Not equal after all. 3674 } 3675 3676 // these are the only paths that produce 'true': 3677 phi->set_req(_prim_same_path, intcon(1)); 3678 phi->set_req(_ref_subtype_path, intcon(1)); 3679 3680 // pull together the cases: 3681 assert(region->req() == PATH_LIMIT, "sane region"); 3682 for (uint i = 1; i < region->req(); i++) { 3683 Node* ctl = region->in(i); 3684 if (ctl == NULL || ctl == top()) { 3685 region->set_req(i, top()); 3686 phi ->set_req(i, top()); 3687 } else if (phi->in(i) == NULL) { 3688 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3689 } 3690 } 3691 3692 set_control(_gvn.transform(region)); 3693 set_result(_gvn.transform(phi)); 3694 return true; 3695} 3696 3697//---------------------generate_array_guard_common------------------------ 3698Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3699 bool obj_array, bool not_array) { 3700 3701 if (stopped()) { 3702 return NULL; 3703 } 3704 3705 // If obj_array/non_array==false/false: 3706 // Branch around if the given klass is in fact an array (either obj or prim). 3707 // If obj_array/non_array==false/true: 3708 // Branch around if the given klass is not an array klass of any kind. 3709 // If obj_array/non_array==true/true: 3710 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3711 // If obj_array/non_array==true/false: 3712 // Branch around if the kls is an oop array (Object[] or subtype) 3713 // 3714 // Like generate_guard, adds a new path onto the region. 3715 jint layout_con = 0; 3716 Node* layout_val = get_layout_helper(kls, layout_con); 3717 if (layout_val == NULL) { 3718 bool query = (obj_array 3719 ? Klass::layout_helper_is_objArray(layout_con) 3720 : Klass::layout_helper_is_array(layout_con)); 3721 if (query == not_array) { 3722 return NULL; // never a branch 3723 } else { // always a branch 3724 Node* always_branch = control(); 3725 if (region != NULL) 3726 region->add_req(always_branch); 3727 set_control(top()); 3728 return always_branch; 3729 } 3730 } 3731 // Now test the correct condition. 3732 jint nval = (obj_array 3733 ? (jint)(Klass::_lh_array_tag_type_value 3734 << Klass::_lh_array_tag_shift) 3735 : Klass::_lh_neutral_value); 3736 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3737 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3738 // invert the test if we are looking for a non-array 3739 if (not_array) btest = BoolTest(btest).negate(); 3740 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3741 return generate_fair_guard(bol, region); 3742} 3743 3744 3745//-----------------------inline_native_newArray-------------------------- 3746// private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3747// private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3748bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3749 Node* mirror; 3750 Node* count_val; 3751 if (uninitialized) { 3752 mirror = argument(1); 3753 count_val = argument(2); 3754 } else { 3755 mirror = argument(0); 3756 count_val = argument(1); 3757 } 3758 3759 mirror = null_check(mirror); 3760 // If mirror or obj is dead, only null-path is taken. 3761 if (stopped()) return true; 3762 3763 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3764 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3765 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3766 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3767 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3768 3769 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3770 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3771 result_reg, _slow_path); 3772 Node* normal_ctl = control(); 3773 Node* no_array_ctl = result_reg->in(_slow_path); 3774 3775 // Generate code for the slow case. We make a call to newArray(). 3776 set_control(no_array_ctl); 3777 if (!stopped()) { 3778 // Either the input type is void.class, or else the 3779 // array klass has not yet been cached. Either the 3780 // ensuing call will throw an exception, or else it 3781 // will cache the array klass for next time. 3782 PreserveJVMState pjvms(this); 3783 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3784 Node* slow_result = set_results_for_java_call(slow_call); 3785 // this->control() comes from set_results_for_java_call 3786 result_reg->set_req(_slow_path, control()); 3787 result_val->set_req(_slow_path, slow_result); 3788 result_io ->set_req(_slow_path, i_o()); 3789 result_mem->set_req(_slow_path, reset_memory()); 3790 } 3791 3792 set_control(normal_ctl); 3793 if (!stopped()) { 3794 // Normal case: The array type has been cached in the java.lang.Class. 3795 // The following call works fine even if the array type is polymorphic. 3796 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3797 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3798 result_reg->init_req(_normal_path, control()); 3799 result_val->init_req(_normal_path, obj); 3800 result_io ->init_req(_normal_path, i_o()); 3801 result_mem->init_req(_normal_path, reset_memory()); 3802 3803 if (uninitialized) { 3804 // Mark the allocation so that zeroing is skipped 3805 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3806 alloc->maybe_set_complete(&_gvn); 3807 } 3808 } 3809 3810 // Return the combined state. 3811 set_i_o( _gvn.transform(result_io) ); 3812 set_all_memory( _gvn.transform(result_mem)); 3813 3814 C->set_has_split_ifs(true); // Has chance for split-if optimization 3815 set_result(result_reg, result_val); 3816 return true; 3817} 3818 3819//----------------------inline_native_getLength-------------------------- 3820// public static native int java.lang.reflect.Array.getLength(Object array); 3821bool LibraryCallKit::inline_native_getLength() { 3822 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3823 3824 Node* array = null_check(argument(0)); 3825 // If array is dead, only null-path is taken. 3826 if (stopped()) return true; 3827 3828 // Deoptimize if it is a non-array. 3829 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3830 3831 if (non_array != NULL) { 3832 PreserveJVMState pjvms(this); 3833 set_control(non_array); 3834 uncommon_trap(Deoptimization::Reason_intrinsic, 3835 Deoptimization::Action_maybe_recompile); 3836 } 3837 3838 // If control is dead, only non-array-path is taken. 3839 if (stopped()) return true; 3840 3841 // The works fine even if the array type is polymorphic. 3842 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3843 Node* result = load_array_length(array); 3844 3845 C->set_has_split_ifs(true); // Has chance for split-if optimization 3846 set_result(result); 3847 return true; 3848} 3849 3850//------------------------inline_array_copyOf---------------------------- 3851// public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3852// public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3853bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3854 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3855 3856 // Get the arguments. 3857 Node* original = argument(0); 3858 Node* start = is_copyOfRange? argument(1): intcon(0); 3859 Node* end = is_copyOfRange? argument(2): argument(1); 3860 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3861 3862 Node* newcopy = NULL; 3863 3864 // Set the original stack and the reexecute bit for the interpreter to reexecute 3865 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3866 { PreserveReexecuteState preexecs(this); 3867 jvms()->set_should_reexecute(true); 3868 3869 array_type_mirror = null_check(array_type_mirror); 3870 original = null_check(original); 3871 3872 // Check if a null path was taken unconditionally. 3873 if (stopped()) return true; 3874 3875 Node* orig_length = load_array_length(original); 3876 3877 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3878 klass_node = null_check(klass_node); 3879 3880 RegionNode* bailout = new RegionNode(1); 3881 record_for_igvn(bailout); 3882 3883 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3884 // Bail out if that is so. 3885 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3886 if (not_objArray != NULL) { 3887 // Improve the klass node's type from the new optimistic assumption: 3888 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3889 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3890 Node* cast = new CastPPNode(klass_node, akls); 3891 cast->init_req(0, control()); 3892 klass_node = _gvn.transform(cast); 3893 } 3894 3895 // Bail out if either start or end is negative. 3896 generate_negative_guard(start, bailout, &start); 3897 generate_negative_guard(end, bailout, &end); 3898 3899 Node* length = end; 3900 if (_gvn.type(start) != TypeInt::ZERO) { 3901 length = _gvn.transform(new SubINode(end, start)); 3902 } 3903 3904 // Bail out if length is negative. 3905 // Without this the new_array would throw 3906 // NegativeArraySizeException but IllegalArgumentException is what 3907 // should be thrown 3908 generate_negative_guard(length, bailout, &length); 3909 3910 if (bailout->req() > 1) { 3911 PreserveJVMState pjvms(this); 3912 set_control(_gvn.transform(bailout)); 3913 uncommon_trap(Deoptimization::Reason_intrinsic, 3914 Deoptimization::Action_maybe_recompile); 3915 } 3916 3917 if (!stopped()) { 3918 // How many elements will we copy from the original? 3919 // The answer is MinI(orig_length - start, length). 3920 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3921 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3922 3923 // Generate a direct call to the right arraycopy function(s). 3924 // We know the copy is disjoint but we might not know if the 3925 // oop stores need checking. 3926 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3927 // This will fail a store-check if x contains any non-nulls. 3928 3929 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3930 // loads/stores but it is legal only if we're sure the 3931 // Arrays.copyOf would succeed. So we need all input arguments 3932 // to the copyOf to be validated, including that the copy to the 3933 // new array won't trigger an ArrayStoreException. That subtype 3934 // check can be optimized if we know something on the type of 3935 // the input array from type speculation. 3936 if (_gvn.type(klass_node)->singleton()) { 3937 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3938 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3939 3940 int test = C->static_subtype_check(superk, subk); 3941 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3942 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3943 if (t_original->speculative_type() != NULL) { 3944 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3945 } 3946 } 3947 } 3948 3949 bool validated = false; 3950 // Reason_class_check rather than Reason_intrinsic because we 3951 // want to intrinsify even if this traps. 3952 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3953 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3954 klass_node); 3955 3956 if (not_subtype_ctrl != top()) { 3957 PreserveJVMState pjvms(this); 3958 set_control(not_subtype_ctrl); 3959 uncommon_trap(Deoptimization::Reason_class_check, 3960 Deoptimization::Action_make_not_entrant); 3961 assert(stopped(), "Should be stopped"); 3962 } 3963 validated = true; 3964 } 3965 3966 if (!stopped()) { 3967 newcopy = new_array(klass_node, length, 0); // no arguments to push 3968 3969 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3970 load_object_klass(original), klass_node); 3971 if (!is_copyOfRange) { 3972 ac->set_copyof(validated); 3973 } else { 3974 ac->set_copyofrange(validated); 3975 } 3976 Node* n = _gvn.transform(ac); 3977 if (n == ac) { 3978 ac->connect_outputs(this); 3979 } else { 3980 assert(validated, "shouldn't transform if all arguments not validated"); 3981 set_all_memory(n); 3982 } 3983 } 3984 } 3985 } // original reexecute is set back here 3986 3987 C->set_has_split_ifs(true); // Has chance for split-if optimization 3988 if (!stopped()) { 3989 set_result(newcopy); 3990 } 3991 return true; 3992} 3993 3994 3995//----------------------generate_virtual_guard--------------------------- 3996// Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3997Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3998 RegionNode* slow_region) { 3999 ciMethod* method = callee(); 4000 int vtable_index = method->vtable_index(); 4001 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4002 "bad index %d", vtable_index); 4003 // Get the Method* out of the appropriate vtable entry. 4004 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 4005 vtable_index*vtableEntry::size_in_bytes() + 4006 vtableEntry::method_offset_in_bytes(); 4007 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4008 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4009 4010 // Compare the target method with the expected method (e.g., Object.hashCode). 4011 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4012 4013 Node* native_call = makecon(native_call_addr); 4014 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4015 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4016 4017 return generate_slow_guard(test_native, slow_region); 4018} 4019 4020//-----------------------generate_method_call---------------------------- 4021// Use generate_method_call to make a slow-call to the real 4022// method if the fast path fails. An alternative would be to 4023// use a stub like OptoRuntime::slow_arraycopy_Java. 4024// This only works for expanding the current library call, 4025// not another intrinsic. (E.g., don't use this for making an 4026// arraycopy call inside of the copyOf intrinsic.) 4027CallJavaNode* 4028LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4029 // When compiling the intrinsic method itself, do not use this technique. 4030 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4031 4032 ciMethod* method = callee(); 4033 // ensure the JVMS we have will be correct for this call 4034 guarantee(method_id == method->intrinsic_id(), "must match"); 4035 4036 const TypeFunc* tf = TypeFunc::make(method); 4037 CallJavaNode* slow_call; 4038 if (is_static) { 4039 assert(!is_virtual, ""); 4040 slow_call = new CallStaticJavaNode(C, tf, 4041 SharedRuntime::get_resolve_static_call_stub(), 4042 method, bci()); 4043 } else if (is_virtual) { 4044 null_check_receiver(); 4045 int vtable_index = Method::invalid_vtable_index; 4046 if (UseInlineCaches) { 4047 // Suppress the vtable call 4048 } else { 4049 // hashCode and clone are not a miranda methods, 4050 // so the vtable index is fixed. 4051 // No need to use the linkResolver to get it. 4052 vtable_index = method->vtable_index(); 4053 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4054 "bad index %d", vtable_index); 4055 } 4056 slow_call = new CallDynamicJavaNode(tf, 4057 SharedRuntime::get_resolve_virtual_call_stub(), 4058 method, vtable_index, bci()); 4059 } else { // neither virtual nor static: opt_virtual 4060 null_check_receiver(); 4061 slow_call = new CallStaticJavaNode(C, tf, 4062 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4063 method, bci()); 4064 slow_call->set_optimized_virtual(true); 4065 } 4066 set_arguments_for_java_call(slow_call); 4067 set_edges_for_java_call(slow_call); 4068 return slow_call; 4069} 4070 4071 4072/** 4073 * Build special case code for calls to hashCode on an object. This call may 4074 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4075 * slightly different code. 4076 */ 4077bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4078 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4079 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4080 4081 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4082 4083 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4084 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4085 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4086 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4087 Node* obj = NULL; 4088 if (!is_static) { 4089 // Check for hashing null object 4090 obj = null_check_receiver(); 4091 if (stopped()) return true; // unconditionally null 4092 result_reg->init_req(_null_path, top()); 4093 result_val->init_req(_null_path, top()); 4094 } else { 4095 // Do a null check, and return zero if null. 4096 // System.identityHashCode(null) == 0 4097 obj = argument(0); 4098 Node* null_ctl = top(); 4099 obj = null_check_oop(obj, &null_ctl); 4100 result_reg->init_req(_null_path, null_ctl); 4101 result_val->init_req(_null_path, _gvn.intcon(0)); 4102 } 4103 4104 // Unconditionally null? Then return right away. 4105 if (stopped()) { 4106 set_control( result_reg->in(_null_path)); 4107 if (!stopped()) 4108 set_result(result_val->in(_null_path)); 4109 return true; 4110 } 4111 4112 // We only go to the fast case code if we pass a number of guards. The 4113 // paths which do not pass are accumulated in the slow_region. 4114 RegionNode* slow_region = new RegionNode(1); 4115 record_for_igvn(slow_region); 4116 4117 // If this is a virtual call, we generate a funny guard. We pull out 4118 // the vtable entry corresponding to hashCode() from the target object. 4119 // If the target method which we are calling happens to be the native 4120 // Object hashCode() method, we pass the guard. We do not need this 4121 // guard for non-virtual calls -- the caller is known to be the native 4122 // Object hashCode(). 4123 if (is_virtual) { 4124 // After null check, get the object's klass. 4125 Node* obj_klass = load_object_klass(obj); 4126 generate_virtual_guard(obj_klass, slow_region); 4127 } 4128 4129 // Get the header out of the object, use LoadMarkNode when available 4130 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4131 // The control of the load must be NULL. Otherwise, the load can move before 4132 // the null check after castPP removal. 4133 Node* no_ctrl = NULL; 4134 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4135 4136 // Test the header to see if it is unlocked. 4137 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4138 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4139 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4140 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4141 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4142 4143 generate_slow_guard(test_unlocked, slow_region); 4144 4145 // Get the hash value and check to see that it has been properly assigned. 4146 // We depend on hash_mask being at most 32 bits and avoid the use of 4147 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4148 // vm: see markOop.hpp. 4149 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4150 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4151 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4152 // This hack lets the hash bits live anywhere in the mark object now, as long 4153 // as the shift drops the relevant bits into the low 32 bits. Note that 4154 // Java spec says that HashCode is an int so there's no point in capturing 4155 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4156 hshifted_header = ConvX2I(hshifted_header); 4157 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4158 4159 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4160 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4161 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4162 4163 generate_slow_guard(test_assigned, slow_region); 4164 4165 Node* init_mem = reset_memory(); 4166 // fill in the rest of the null path: 4167 result_io ->init_req(_null_path, i_o()); 4168 result_mem->init_req(_null_path, init_mem); 4169 4170 result_val->init_req(_fast_path, hash_val); 4171 result_reg->init_req(_fast_path, control()); 4172 result_io ->init_req(_fast_path, i_o()); 4173 result_mem->init_req(_fast_path, init_mem); 4174 4175 // Generate code for the slow case. We make a call to hashCode(). 4176 set_control(_gvn.transform(slow_region)); 4177 if (!stopped()) { 4178 // No need for PreserveJVMState, because we're using up the present state. 4179 set_all_memory(init_mem); 4180 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4181 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4182 Node* slow_result = set_results_for_java_call(slow_call); 4183 // this->control() comes from set_results_for_java_call 4184 result_reg->init_req(_slow_path, control()); 4185 result_val->init_req(_slow_path, slow_result); 4186 result_io ->set_req(_slow_path, i_o()); 4187 result_mem ->set_req(_slow_path, reset_memory()); 4188 } 4189 4190 // Return the combined state. 4191 set_i_o( _gvn.transform(result_io) ); 4192 set_all_memory( _gvn.transform(result_mem)); 4193 4194 set_result(result_reg, result_val); 4195 return true; 4196} 4197 4198//---------------------------inline_native_getClass---------------------------- 4199// public final native Class<?> java.lang.Object.getClass(); 4200// 4201// Build special case code for calls to getClass on an object. 4202bool LibraryCallKit::inline_native_getClass() { 4203 Node* obj = null_check_receiver(); 4204 if (stopped()) return true; 4205 set_result(load_mirror_from_klass(load_object_klass(obj))); 4206 return true; 4207} 4208 4209//-----------------inline_native_Reflection_getCallerClass--------------------- 4210// public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4211// 4212// In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4213// 4214// NOTE: This code must perform the same logic as JVM_GetCallerClass 4215// in that it must skip particular security frames and checks for 4216// caller sensitive methods. 4217bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4218#ifndef PRODUCT 4219 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4220 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4221 } 4222#endif 4223 4224 if (!jvms()->has_method()) { 4225#ifndef PRODUCT 4226 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4227 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4228 } 4229#endif 4230 return false; 4231 } 4232 4233 // Walk back up the JVM state to find the caller at the required 4234 // depth. 4235 JVMState* caller_jvms = jvms(); 4236 4237 // Cf. JVM_GetCallerClass 4238 // NOTE: Start the loop at depth 1 because the current JVM state does 4239 // not include the Reflection.getCallerClass() frame. 4240 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4241 ciMethod* m = caller_jvms->method(); 4242 switch (n) { 4243 case 0: 4244 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4245 break; 4246 case 1: 4247 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4248 if (!m->caller_sensitive()) { 4249#ifndef PRODUCT 4250 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4251 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4252 } 4253#endif 4254 return false; // bail-out; let JVM_GetCallerClass do the work 4255 } 4256 break; 4257 default: 4258 if (!m->is_ignored_by_security_stack_walk()) { 4259 // We have reached the desired frame; return the holder class. 4260 // Acquire method holder as java.lang.Class and push as constant. 4261 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4262 ciInstance* caller_mirror = caller_klass->java_mirror(); 4263 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4264 4265#ifndef PRODUCT 4266 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4267 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()); 4268 tty->print_cr(" JVM state at this point:"); 4269 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4270 ciMethod* m = jvms()->of_depth(i)->method(); 4271 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4272 } 4273 } 4274#endif 4275 return true; 4276 } 4277 break; 4278 } 4279 } 4280 4281#ifndef PRODUCT 4282 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4283 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4284 tty->print_cr(" JVM state at this point:"); 4285 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4286 ciMethod* m = jvms()->of_depth(i)->method(); 4287 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4288 } 4289 } 4290#endif 4291 4292 return false; // bail-out; let JVM_GetCallerClass do the work 4293} 4294 4295bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4296 Node* arg = argument(0); 4297 Node* result = NULL; 4298 4299 switch (id) { 4300 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4301 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4302 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4303 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4304 4305 case vmIntrinsics::_doubleToLongBits: { 4306 // two paths (plus control) merge in a wood 4307 RegionNode *r = new RegionNode(3); 4308 Node *phi = new PhiNode(r, TypeLong::LONG); 4309 4310 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4311 // Build the boolean node 4312 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4313 4314 // Branch either way. 4315 // NaN case is less traveled, which makes all the difference. 4316 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4317 Node *opt_isnan = _gvn.transform(ifisnan); 4318 assert( opt_isnan->is_If(), "Expect an IfNode"); 4319 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4320 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4321 4322 set_control(iftrue); 4323 4324 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4325 Node *slow_result = longcon(nan_bits); // return NaN 4326 phi->init_req(1, _gvn.transform( slow_result )); 4327 r->init_req(1, iftrue); 4328 4329 // Else fall through 4330 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4331 set_control(iffalse); 4332 4333 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4334 r->init_req(2, iffalse); 4335 4336 // Post merge 4337 set_control(_gvn.transform(r)); 4338 record_for_igvn(r); 4339 4340 C->set_has_split_ifs(true); // Has chance for split-if optimization 4341 result = phi; 4342 assert(result->bottom_type()->isa_long(), "must be"); 4343 break; 4344 } 4345 4346 case vmIntrinsics::_floatToIntBits: { 4347 // two paths (plus control) merge in a wood 4348 RegionNode *r = new RegionNode(3); 4349 Node *phi = new PhiNode(r, TypeInt::INT); 4350 4351 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4352 // Build the boolean node 4353 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4354 4355 // Branch either way. 4356 // NaN case is less traveled, which makes all the difference. 4357 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4358 Node *opt_isnan = _gvn.transform(ifisnan); 4359 assert( opt_isnan->is_If(), "Expect an IfNode"); 4360 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4361 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4362 4363 set_control(iftrue); 4364 4365 static const jint nan_bits = 0x7fc00000; 4366 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4367 phi->init_req(1, _gvn.transform( slow_result )); 4368 r->init_req(1, iftrue); 4369 4370 // Else fall through 4371 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4372 set_control(iffalse); 4373 4374 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4375 r->init_req(2, iffalse); 4376 4377 // Post merge 4378 set_control(_gvn.transform(r)); 4379 record_for_igvn(r); 4380 4381 C->set_has_split_ifs(true); // Has chance for split-if optimization 4382 result = phi; 4383 assert(result->bottom_type()->isa_int(), "must be"); 4384 break; 4385 } 4386 4387 default: 4388 fatal_unexpected_iid(id); 4389 break; 4390 } 4391 set_result(_gvn.transform(result)); 4392 return true; 4393} 4394 4395//----------------------inline_unsafe_copyMemory------------------------- 4396// public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4397bool LibraryCallKit::inline_unsafe_copyMemory() { 4398 if (callee()->is_static()) return false; // caller must have the capability! 4399 null_check_receiver(); // null-check receiver 4400 if (stopped()) return true; 4401 4402 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4403 4404 Node* src_ptr = argument(1); // type: oop 4405 Node* src_off = ConvL2X(argument(2)); // type: long 4406 Node* dst_ptr = argument(4); // type: oop 4407 Node* dst_off = ConvL2X(argument(5)); // type: long 4408 Node* size = ConvL2X(argument(7)); // type: long 4409 4410 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4411 "fieldOffset must be byte-scaled"); 4412 4413 Node* src = make_unsafe_address(src_ptr, src_off); 4414 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4415 4416 // Conservatively insert a memory barrier on all memory slices. 4417 // Do not let writes of the copy source or destination float below the copy. 4418 insert_mem_bar(Op_MemBarCPUOrder); 4419 4420 // Call it. Note that the length argument is not scaled. 4421 make_runtime_call(RC_LEAF|RC_NO_FP, 4422 OptoRuntime::fast_arraycopy_Type(), 4423 StubRoutines::unsafe_arraycopy(), 4424 "unsafe_arraycopy", 4425 TypeRawPtr::BOTTOM, 4426 src, dst, size XTOP); 4427 4428 // Do not let reads of the copy destination float above the copy. 4429 insert_mem_bar(Op_MemBarCPUOrder); 4430 4431 return true; 4432} 4433 4434//------------------------clone_coping----------------------------------- 4435// Helper function for inline_native_clone. 4436void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4437 assert(obj_size != NULL, ""); 4438 Node* raw_obj = alloc_obj->in(1); 4439 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4440 4441 AllocateNode* alloc = NULL; 4442 if (ReduceBulkZeroing) { 4443 // We will be completely responsible for initializing this object - 4444 // mark Initialize node as complete. 4445 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4446 // The object was just allocated - there should be no any stores! 4447 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4448 // Mark as complete_with_arraycopy so that on AllocateNode 4449 // expansion, we know this AllocateNode is initialized by an array 4450 // copy and a StoreStore barrier exists after the array copy. 4451 alloc->initialization()->set_complete_with_arraycopy(); 4452 } 4453 4454 // Copy the fastest available way. 4455 // TODO: generate fields copies for small objects instead. 4456 Node* src = obj; 4457 Node* dest = alloc_obj; 4458 Node* size = _gvn.transform(obj_size); 4459 4460 // Exclude the header but include array length to copy by 8 bytes words. 4461 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4462 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4463 instanceOopDesc::base_offset_in_bytes(); 4464 // base_off: 4465 // 8 - 32-bit VM 4466 // 12 - 64-bit VM, compressed klass 4467 // 16 - 64-bit VM, normal klass 4468 if (base_off % BytesPerLong != 0) { 4469 assert(UseCompressedClassPointers, ""); 4470 if (is_array) { 4471 // Exclude length to copy by 8 bytes words. 4472 base_off += sizeof(int); 4473 } else { 4474 // Include klass to copy by 8 bytes words. 4475 base_off = instanceOopDesc::klass_offset_in_bytes(); 4476 } 4477 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4478 } 4479 src = basic_plus_adr(src, base_off); 4480 dest = basic_plus_adr(dest, base_off); 4481 4482 // Compute the length also, if needed: 4483 Node* countx = size; 4484 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4485 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4486 4487 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4488 4489 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4490 ac->set_clonebasic(); 4491 Node* n = _gvn.transform(ac); 4492 if (n == ac) { 4493 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4494 } else { 4495 set_all_memory(n); 4496 } 4497 4498 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4499 if (card_mark) { 4500 assert(!is_array, ""); 4501 // Put in store barrier for any and all oops we are sticking 4502 // into this object. (We could avoid this if we could prove 4503 // that the object type contains no oop fields at all.) 4504 Node* no_particular_value = NULL; 4505 Node* no_particular_field = NULL; 4506 int raw_adr_idx = Compile::AliasIdxRaw; 4507 post_barrier(control(), 4508 memory(raw_adr_type), 4509 alloc_obj, 4510 no_particular_field, 4511 raw_adr_idx, 4512 no_particular_value, 4513 T_OBJECT, 4514 false); 4515 } 4516 4517 // Do not let reads from the cloned object float above the arraycopy. 4518 if (alloc != NULL) { 4519 // Do not let stores that initialize this object be reordered with 4520 // a subsequent store that would make this object accessible by 4521 // other threads. 4522 // Record what AllocateNode this StoreStore protects so that 4523 // escape analysis can go from the MemBarStoreStoreNode to the 4524 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4525 // based on the escape status of the AllocateNode. 4526 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4527 } else { 4528 insert_mem_bar(Op_MemBarCPUOrder); 4529 } 4530} 4531 4532//------------------------inline_native_clone---------------------------- 4533// protected native Object java.lang.Object.clone(); 4534// 4535// Here are the simple edge cases: 4536// null receiver => normal trap 4537// virtual and clone was overridden => slow path to out-of-line clone 4538// not cloneable or finalizer => slow path to out-of-line Object.clone 4539// 4540// The general case has two steps, allocation and copying. 4541// Allocation has two cases, and uses GraphKit::new_instance or new_array. 4542// 4543// Copying also has two cases, oop arrays and everything else. 4544// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4545// Everything else uses the tight inline loop supplied by CopyArrayNode. 4546// 4547// These steps fold up nicely if and when the cloned object's klass 4548// can be sharply typed as an object array, a type array, or an instance. 4549// 4550bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4551 PhiNode* result_val; 4552 4553 // Set the reexecute bit for the interpreter to reexecute 4554 // the bytecode that invokes Object.clone if deoptimization happens. 4555 { PreserveReexecuteState preexecs(this); 4556 jvms()->set_should_reexecute(true); 4557 4558 Node* obj = null_check_receiver(); 4559 if (stopped()) return true; 4560 4561 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4562 4563 // If we are going to clone an instance, we need its exact type to 4564 // know the number and types of fields to convert the clone to 4565 // loads/stores. Maybe a speculative type can help us. 4566 if (!obj_type->klass_is_exact() && 4567 obj_type->speculative_type() != NULL && 4568 obj_type->speculative_type()->is_instance_klass()) { 4569 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4570 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4571 !spec_ik->has_injected_fields()) { 4572 ciKlass* k = obj_type->klass(); 4573 if (!k->is_instance_klass() || 4574 k->as_instance_klass()->is_interface() || 4575 k->as_instance_klass()->has_subklass()) { 4576 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4577 } 4578 } 4579 } 4580 4581 Node* obj_klass = load_object_klass(obj); 4582 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4583 const TypeOopPtr* toop = ((tklass != NULL) 4584 ? tklass->as_instance_type() 4585 : TypeInstPtr::NOTNULL); 4586 4587 // Conservatively insert a memory barrier on all memory slices. 4588 // Do not let writes into the original float below the clone. 4589 insert_mem_bar(Op_MemBarCPUOrder); 4590 4591 // paths into result_reg: 4592 enum { 4593 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4594 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4595 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4596 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4597 PATH_LIMIT 4598 }; 4599 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4600 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4601 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4602 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4603 record_for_igvn(result_reg); 4604 4605 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4606 int raw_adr_idx = Compile::AliasIdxRaw; 4607 4608 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4609 if (array_ctl != NULL) { 4610 // It's an array. 4611 PreserveJVMState pjvms(this); 4612 set_control(array_ctl); 4613 Node* obj_length = load_array_length(obj); 4614 Node* obj_size = NULL; 4615 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4616 4617 if (!use_ReduceInitialCardMarks()) { 4618 // If it is an oop array, it requires very special treatment, 4619 // because card marking is required on each card of the array. 4620 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4621 if (is_obja != NULL) { 4622 PreserveJVMState pjvms2(this); 4623 set_control(is_obja); 4624 // Generate a direct call to the right arraycopy function(s). 4625 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4626 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4627 ac->set_cloneoop(); 4628 Node* n = _gvn.transform(ac); 4629 assert(n == ac, "cannot disappear"); 4630 ac->connect_outputs(this); 4631 4632 result_reg->init_req(_objArray_path, control()); 4633 result_val->init_req(_objArray_path, alloc_obj); 4634 result_i_o ->set_req(_objArray_path, i_o()); 4635 result_mem ->set_req(_objArray_path, reset_memory()); 4636 } 4637 } 4638 // Otherwise, there are no card marks to worry about. 4639 // (We can dispense with card marks if we know the allocation 4640 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4641 // causes the non-eden paths to take compensating steps to 4642 // simulate a fresh allocation, so that no further 4643 // card marks are required in compiled code to initialize 4644 // the object.) 4645 4646 if (!stopped()) { 4647 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4648 4649 // Present the results of the copy. 4650 result_reg->init_req(_array_path, control()); 4651 result_val->init_req(_array_path, alloc_obj); 4652 result_i_o ->set_req(_array_path, i_o()); 4653 result_mem ->set_req(_array_path, reset_memory()); 4654 } 4655 } 4656 4657 // We only go to the instance fast case code if we pass a number of guards. 4658 // The paths which do not pass are accumulated in the slow_region. 4659 RegionNode* slow_region = new RegionNode(1); 4660 record_for_igvn(slow_region); 4661 if (!stopped()) { 4662 // It's an instance (we did array above). Make the slow-path tests. 4663 // If this is a virtual call, we generate a funny guard. We grab 4664 // the vtable entry corresponding to clone() from the target object. 4665 // If the target method which we are calling happens to be the 4666 // Object clone() method, we pass the guard. We do not need this 4667 // guard for non-virtual calls; the caller is known to be the native 4668 // Object clone(). 4669 if (is_virtual) { 4670 generate_virtual_guard(obj_klass, slow_region); 4671 } 4672 4673 // The object must be easily cloneable and must not have a finalizer. 4674 // Both of these conditions may be checked in a single test. 4675 // We could optimize the test further, but we don't care. 4676 generate_access_flags_guard(obj_klass, 4677 // Test both conditions: 4678 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4679 // Must be cloneable but not finalizer: 4680 JVM_ACC_IS_CLONEABLE_FAST, 4681 slow_region); 4682 } 4683 4684 if (!stopped()) { 4685 // It's an instance, and it passed the slow-path tests. 4686 PreserveJVMState pjvms(this); 4687 Node* obj_size = NULL; 4688 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4689 // is reexecuted if deoptimization occurs and there could be problems when merging 4690 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4691 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4692 4693 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4694 4695 // Present the results of the slow call. 4696 result_reg->init_req(_instance_path, control()); 4697 result_val->init_req(_instance_path, alloc_obj); 4698 result_i_o ->set_req(_instance_path, i_o()); 4699 result_mem ->set_req(_instance_path, reset_memory()); 4700 } 4701 4702 // Generate code for the slow case. We make a call to clone(). 4703 set_control(_gvn.transform(slow_region)); 4704 if (!stopped()) { 4705 PreserveJVMState pjvms(this); 4706 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4707 Node* slow_result = set_results_for_java_call(slow_call); 4708 // this->control() comes from set_results_for_java_call 4709 result_reg->init_req(_slow_path, control()); 4710 result_val->init_req(_slow_path, slow_result); 4711 result_i_o ->set_req(_slow_path, i_o()); 4712 result_mem ->set_req(_slow_path, reset_memory()); 4713 } 4714 4715 // Return the combined state. 4716 set_control( _gvn.transform(result_reg)); 4717 set_i_o( _gvn.transform(result_i_o)); 4718 set_all_memory( _gvn.transform(result_mem)); 4719 } // original reexecute is set back here 4720 4721 set_result(_gvn.transform(result_val)); 4722 return true; 4723} 4724 4725// If we have a tighly coupled allocation, the arraycopy may take care 4726// of the array initialization. If one of the guards we insert between 4727// the allocation and the arraycopy causes a deoptimization, an 4728// unitialized array will escape the compiled method. To prevent that 4729// we set the JVM state for uncommon traps between the allocation and 4730// the arraycopy to the state before the allocation so, in case of 4731// deoptimization, we'll reexecute the allocation and the 4732// initialization. 4733JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4734 if (alloc != NULL) { 4735 ciMethod* trap_method = alloc->jvms()->method(); 4736 int trap_bci = alloc->jvms()->bci(); 4737 4738 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4739 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4740 // Make sure there's no store between the allocation and the 4741 // arraycopy otherwise visible side effects could be rexecuted 4742 // in case of deoptimization and cause incorrect execution. 4743 bool no_interfering_store = true; 4744 Node* mem = alloc->in(TypeFunc::Memory); 4745 if (mem->is_MergeMem()) { 4746 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4747 Node* n = mms.memory(); 4748 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4749 assert(n->is_Store(), "what else?"); 4750 no_interfering_store = false; 4751 break; 4752 } 4753 } 4754 } else { 4755 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4756 Node* n = mms.memory(); 4757 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4758 assert(n->is_Store(), "what else?"); 4759 no_interfering_store = false; 4760 break; 4761 } 4762 } 4763 } 4764 4765 if (no_interfering_store) { 4766 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4767 uint size = alloc->req(); 4768 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4769 old_jvms->set_map(sfpt); 4770 for (uint i = 0; i < size; i++) { 4771 sfpt->init_req(i, alloc->in(i)); 4772 } 4773 // re-push array length for deoptimization 4774 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4775 old_jvms->set_sp(old_jvms->sp()+1); 4776 old_jvms->set_monoff(old_jvms->monoff()+1); 4777 old_jvms->set_scloff(old_jvms->scloff()+1); 4778 old_jvms->set_endoff(old_jvms->endoff()+1); 4779 old_jvms->set_should_reexecute(true); 4780 4781 sfpt->set_i_o(map()->i_o()); 4782 sfpt->set_memory(map()->memory()); 4783 sfpt->set_control(map()->control()); 4784 4785 JVMState* saved_jvms = jvms(); 4786 saved_reexecute_sp = _reexecute_sp; 4787 4788 set_jvms(sfpt->jvms()); 4789 _reexecute_sp = jvms()->sp(); 4790 4791 return saved_jvms; 4792 } 4793 } 4794 } 4795 return NULL; 4796} 4797 4798// In case of a deoptimization, we restart execution at the 4799// allocation, allocating a new array. We would leave an uninitialized 4800// array in the heap that GCs wouldn't expect. Move the allocation 4801// after the traps so we don't allocate the array if we 4802// deoptimize. This is possible because tightly_coupled_allocation() 4803// guarantees there's no observer of the allocated array at this point 4804// and the control flow is simple enough. 4805void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4806 if (saved_jvms != NULL && !stopped()) { 4807 assert(alloc != NULL, "only with a tightly coupled allocation"); 4808 // restore JVM state to the state at the arraycopy 4809 saved_jvms->map()->set_control(map()->control()); 4810 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4811 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4812 // If we've improved the types of some nodes (null check) while 4813 // emitting the guards, propagate them to the current state 4814 map()->replaced_nodes().apply(saved_jvms->map()); 4815 set_jvms(saved_jvms); 4816 _reexecute_sp = saved_reexecute_sp; 4817 4818 // Remove the allocation from above the guards 4819 CallProjections callprojs; 4820 alloc->extract_projections(&callprojs, true); 4821 InitializeNode* init = alloc->initialization(); 4822 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4823 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4824 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4825 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4826 4827 // move the allocation here (after the guards) 4828 _gvn.hash_delete(alloc); 4829 alloc->set_req(TypeFunc::Control, control()); 4830 alloc->set_req(TypeFunc::I_O, i_o()); 4831 Node *mem = reset_memory(); 4832 set_all_memory(mem); 4833 alloc->set_req(TypeFunc::Memory, mem); 4834 set_control(init->proj_out(TypeFunc::Control)); 4835 set_i_o(callprojs.fallthrough_ioproj); 4836 4837 // Update memory as done in GraphKit::set_output_for_allocation() 4838 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4839 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4840 if (ary_type->isa_aryptr() && length_type != NULL) { 4841 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4842 } 4843 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4844 int elemidx = C->get_alias_index(telemref); 4845 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4846 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4847 4848 Node* allocx = _gvn.transform(alloc); 4849 assert(allocx == alloc, "where has the allocation gone?"); 4850 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4851 4852 _gvn.hash_delete(dest); 4853 dest->set_req(0, control()); 4854 Node* destx = _gvn.transform(dest); 4855 assert(destx == dest, "where has the allocation result gone?"); 4856 } 4857} 4858 4859 4860//------------------------------inline_arraycopy----------------------- 4861// public static native void java.lang.System.arraycopy(Object src, int srcPos, 4862// Object dest, int destPos, 4863// int length); 4864bool LibraryCallKit::inline_arraycopy() { 4865 // Get the arguments. 4866 Node* src = argument(0); // type: oop 4867 Node* src_offset = argument(1); // type: int 4868 Node* dest = argument(2); // type: oop 4869 Node* dest_offset = argument(3); // type: int 4870 Node* length = argument(4); // type: int 4871 4872 4873 // Check for allocation before we add nodes that would confuse 4874 // tightly_coupled_allocation() 4875 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4876 4877 int saved_reexecute_sp = -1; 4878 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4879 // See arraycopy_restore_alloc_state() comment 4880 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4881 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4882 // if saved_jvms == NULL and alloc != NULL, we can���t emit any guards 4883 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4884 4885 // The following tests must be performed 4886 // (1) src and dest are arrays. 4887 // (2) src and dest arrays must have elements of the same BasicType 4888 // (3) src and dest must not be null. 4889 // (4) src_offset must not be negative. 4890 // (5) dest_offset must not be negative. 4891 // (6) length must not be negative. 4892 // (7) src_offset + length must not exceed length of src. 4893 // (8) dest_offset + length must not exceed length of dest. 4894 // (9) each element of an oop array must be assignable 4895 4896 // (3) src and dest must not be null. 4897 // always do this here because we need the JVM state for uncommon traps 4898 Node* null_ctl = top(); 4899 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4900 assert(null_ctl->is_top(), "no null control here"); 4901 dest = null_check(dest, T_ARRAY); 4902 4903 if (!can_emit_guards) { 4904 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4905 // guards but the arraycopy node could still take advantage of a 4906 // tightly allocated allocation. tightly_coupled_allocation() is 4907 // called again to make sure it takes the null check above into 4908 // account: the null check is mandatory and if it caused an 4909 // uncommon trap to be emitted then the allocation can't be 4910 // considered tightly coupled in this context. 4911 alloc = tightly_coupled_allocation(dest, NULL); 4912 } 4913 4914 bool validated = false; 4915 4916 const Type* src_type = _gvn.type(src); 4917 const Type* dest_type = _gvn.type(dest); 4918 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4919 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4920 4921 // Do we have the type of src? 4922 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4923 // Do we have the type of dest? 4924 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4925 // Is the type for src from speculation? 4926 bool src_spec = false; 4927 // Is the type for dest from speculation? 4928 bool dest_spec = false; 4929 4930 if ((!has_src || !has_dest) && can_emit_guards) { 4931 // We don't have sufficient type information, let's see if 4932 // speculative types can help. We need to have types for both src 4933 // and dest so that it pays off. 4934 4935 // Do we already have or could we have type information for src 4936 bool could_have_src = has_src; 4937 // Do we already have or could we have type information for dest 4938 bool could_have_dest = has_dest; 4939 4940 ciKlass* src_k = NULL; 4941 if (!has_src) { 4942 src_k = src_type->speculative_type_not_null(); 4943 if (src_k != NULL && src_k->is_array_klass()) { 4944 could_have_src = true; 4945 } 4946 } 4947 4948 ciKlass* dest_k = NULL; 4949 if (!has_dest) { 4950 dest_k = dest_type->speculative_type_not_null(); 4951 if (dest_k != NULL && dest_k->is_array_klass()) { 4952 could_have_dest = true; 4953 } 4954 } 4955 4956 if (could_have_src && could_have_dest) { 4957 // This is going to pay off so emit the required guards 4958 if (!has_src) { 4959 src = maybe_cast_profiled_obj(src, src_k, true); 4960 src_type = _gvn.type(src); 4961 top_src = src_type->isa_aryptr(); 4962 has_src = (top_src != NULL && top_src->klass() != NULL); 4963 src_spec = true; 4964 } 4965 if (!has_dest) { 4966 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4967 dest_type = _gvn.type(dest); 4968 top_dest = dest_type->isa_aryptr(); 4969 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4970 dest_spec = true; 4971 } 4972 } 4973 } 4974 4975 if (has_src && has_dest && can_emit_guards) { 4976 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4977 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4978 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4979 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4980 4981 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4982 // If both arrays are object arrays then having the exact types 4983 // for both will remove the need for a subtype check at runtime 4984 // before the call and may make it possible to pick a faster copy 4985 // routine (without a subtype check on every element) 4986 // Do we have the exact type of src? 4987 bool could_have_src = src_spec; 4988 // Do we have the exact type of dest? 4989 bool could_have_dest = dest_spec; 4990 ciKlass* src_k = top_src->klass(); 4991 ciKlass* dest_k = top_dest->klass(); 4992 if (!src_spec) { 4993 src_k = src_type->speculative_type_not_null(); 4994 if (src_k != NULL && src_k->is_array_klass()) { 4995 could_have_src = true; 4996 } 4997 } 4998 if (!dest_spec) { 4999 dest_k = dest_type->speculative_type_not_null(); 5000 if (dest_k != NULL && dest_k->is_array_klass()) { 5001 could_have_dest = true; 5002 } 5003 } 5004 if (could_have_src && could_have_dest) { 5005 // If we can have both exact types, emit the missing guards 5006 if (could_have_src && !src_spec) { 5007 src = maybe_cast_profiled_obj(src, src_k, true); 5008 } 5009 if (could_have_dest && !dest_spec) { 5010 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5011 } 5012 } 5013 } 5014 } 5015 5016 ciMethod* trap_method = method(); 5017 int trap_bci = bci(); 5018 if (saved_jvms != NULL) { 5019 trap_method = alloc->jvms()->method(); 5020 trap_bci = alloc->jvms()->bci(); 5021 } 5022 5023 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5024 can_emit_guards && 5025 !src->is_top() && !dest->is_top()) { 5026 // validate arguments: enables transformation the ArrayCopyNode 5027 validated = true; 5028 5029 RegionNode* slow_region = new RegionNode(1); 5030 record_for_igvn(slow_region); 5031 5032 // (1) src and dest are arrays. 5033 generate_non_array_guard(load_object_klass(src), slow_region); 5034 generate_non_array_guard(load_object_klass(dest), slow_region); 5035 5036 // (2) src and dest arrays must have elements of the same BasicType 5037 // done at macro expansion or at Ideal transformation time 5038 5039 // (4) src_offset must not be negative. 5040 generate_negative_guard(src_offset, slow_region); 5041 5042 // (5) dest_offset must not be negative. 5043 generate_negative_guard(dest_offset, slow_region); 5044 5045 // (7) src_offset + length must not exceed length of src. 5046 generate_limit_guard(src_offset, length, 5047 load_array_length(src), 5048 slow_region); 5049 5050 // (8) dest_offset + length must not exceed length of dest. 5051 generate_limit_guard(dest_offset, length, 5052 load_array_length(dest), 5053 slow_region); 5054 5055 // (9) each element of an oop array must be assignable 5056 Node* src_klass = load_object_klass(src); 5057 Node* dest_klass = load_object_klass(dest); 5058 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5059 5060 if (not_subtype_ctrl != top()) { 5061 PreserveJVMState pjvms(this); 5062 set_control(not_subtype_ctrl); 5063 uncommon_trap(Deoptimization::Reason_intrinsic, 5064 Deoptimization::Action_make_not_entrant); 5065 assert(stopped(), "Should be stopped"); 5066 } 5067 { 5068 PreserveJVMState pjvms(this); 5069 set_control(_gvn.transform(slow_region)); 5070 uncommon_trap(Deoptimization::Reason_intrinsic, 5071 Deoptimization::Action_make_not_entrant); 5072 assert(stopped(), "Should be stopped"); 5073 } 5074 } 5075 5076 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5077 5078 if (stopped()) { 5079 return true; 5080 } 5081 5082 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5083 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5084 // so the compiler has a chance to eliminate them: during macro expansion, 5085 // we have to set their control (CastPP nodes are eliminated). 5086 load_object_klass(src), load_object_klass(dest), 5087 load_array_length(src), load_array_length(dest)); 5088 5089 ac->set_arraycopy(validated); 5090 5091 Node* n = _gvn.transform(ac); 5092 if (n == ac) { 5093 ac->connect_outputs(this); 5094 } else { 5095 assert(validated, "shouldn't transform if all arguments not validated"); 5096 set_all_memory(n); 5097 } 5098 5099 return true; 5100} 5101 5102 5103// Helper function which determines if an arraycopy immediately follows 5104// an allocation, with no intervening tests or other escapes for the object. 5105AllocateArrayNode* 5106LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5107 RegionNode* slow_region) { 5108 if (stopped()) return NULL; // no fast path 5109 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5110 5111 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5112 if (alloc == NULL) return NULL; 5113 5114 Node* rawmem = memory(Compile::AliasIdxRaw); 5115 // Is the allocation's memory state untouched? 5116 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5117 // Bail out if there have been raw-memory effects since the allocation. 5118 // (Example: There might have been a call or safepoint.) 5119 return NULL; 5120 } 5121 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5122 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5123 return NULL; 5124 } 5125 5126 // There must be no unexpected observers of this allocation. 5127 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5128 Node* obs = ptr->fast_out(i); 5129 if (obs != this->map()) { 5130 return NULL; 5131 } 5132 } 5133 5134 // This arraycopy must unconditionally follow the allocation of the ptr. 5135 Node* alloc_ctl = ptr->in(0); 5136 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5137 5138 Node* ctl = control(); 5139 while (ctl != alloc_ctl) { 5140 // There may be guards which feed into the slow_region. 5141 // Any other control flow means that we might not get a chance 5142 // to finish initializing the allocated object. 5143 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5144 IfNode* iff = ctl->in(0)->as_If(); 5145 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5146 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5147 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5148 ctl = iff->in(0); // This test feeds the known slow_region. 5149 continue; 5150 } 5151 // One more try: Various low-level checks bottom out in 5152 // uncommon traps. If the debug-info of the trap omits 5153 // any reference to the allocation, as we've already 5154 // observed, then there can be no objection to the trap. 5155 bool found_trap = false; 5156 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5157 Node* obs = not_ctl->fast_out(j); 5158 if (obs->in(0) == not_ctl && obs->is_Call() && 5159 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5160 found_trap = true; break; 5161 } 5162 } 5163 if (found_trap) { 5164 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5165 continue; 5166 } 5167 } 5168 return NULL; 5169 } 5170 5171 // If we get this far, we have an allocation which immediately 5172 // precedes the arraycopy, and we can take over zeroing the new object. 5173 // The arraycopy will finish the initialization, and provide 5174 // a new control state to which we will anchor the destination pointer. 5175 5176 return alloc; 5177} 5178 5179//-------------inline_encodeISOArray----------------------------------- 5180// encode char[] to byte[] in ISO_8859_1 5181bool LibraryCallKit::inline_encodeISOArray() { 5182 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5183 // no receiver since it is static method 5184 Node *src = argument(0); 5185 Node *src_offset = argument(1); 5186 Node *dst = argument(2); 5187 Node *dst_offset = argument(3); 5188 Node *length = argument(4); 5189 5190 const Type* src_type = src->Value(&_gvn); 5191 const Type* dst_type = dst->Value(&_gvn); 5192 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5193 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5194 if (top_src == NULL || top_src->klass() == NULL || 5195 top_dest == NULL || top_dest->klass() == NULL) { 5196 // failed array check 5197 return false; 5198 } 5199 5200 // Figure out the size and type of the elements we will be copying. 5201 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5202 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5203 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5204 return false; 5205 } 5206 5207 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5208 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5209 // 'src_start' points to src array + scaled offset 5210 // 'dst_start' points to dst array + scaled offset 5211 5212 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5213 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5214 enc = _gvn.transform(enc); 5215 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5216 set_memory(res_mem, mtype); 5217 set_result(enc); 5218 return true; 5219} 5220 5221//-------------inline_multiplyToLen----------------------------------- 5222bool LibraryCallKit::inline_multiplyToLen() { 5223 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5224 5225 address stubAddr = StubRoutines::multiplyToLen(); 5226 if (stubAddr == NULL) { 5227 return false; // Intrinsic's stub is not implemented on this platform 5228 } 5229 const char* stubName = "multiplyToLen"; 5230 5231 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5232 5233 // no receiver because it is a static method 5234 Node* x = argument(0); 5235 Node* xlen = argument(1); 5236 Node* y = argument(2); 5237 Node* ylen = argument(3); 5238 Node* z = argument(4); 5239 5240 const Type* x_type = x->Value(&_gvn); 5241 const Type* y_type = y->Value(&_gvn); 5242 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5243 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5244 if (top_x == NULL || top_x->klass() == NULL || 5245 top_y == NULL || top_y->klass() == NULL) { 5246 // failed array check 5247 return false; 5248 } 5249 5250 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5251 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5252 if (x_elem != T_INT || y_elem != T_INT) { 5253 return false; 5254 } 5255 5256 // Set the original stack and the reexecute bit for the interpreter to reexecute 5257 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5258 // on the return from z array allocation in runtime. 5259 { PreserveReexecuteState preexecs(this); 5260 jvms()->set_should_reexecute(true); 5261 5262 Node* x_start = array_element_address(x, intcon(0), x_elem); 5263 Node* y_start = array_element_address(y, intcon(0), y_elem); 5264 // 'x_start' points to x array + scaled xlen 5265 // 'y_start' points to y array + scaled ylen 5266 5267 // Allocate the result array 5268 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5269 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5270 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5271 5272 IdealKit ideal(this); 5273 5274#define __ ideal. 5275 Node* one = __ ConI(1); 5276 Node* zero = __ ConI(0); 5277 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5278 __ set(need_alloc, zero); 5279 __ set(z_alloc, z); 5280 __ if_then(z, BoolTest::eq, null()); { 5281 __ increment (need_alloc, one); 5282 } __ else_(); { 5283 // Update graphKit memory and control from IdealKit. 5284 sync_kit(ideal); 5285 Node* zlen_arg = load_array_length(z); 5286 // Update IdealKit memory and control from graphKit. 5287 __ sync_kit(this); 5288 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5289 __ increment (need_alloc, one); 5290 } __ end_if(); 5291 } __ end_if(); 5292 5293 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5294 // Update graphKit memory and control from IdealKit. 5295 sync_kit(ideal); 5296 Node * narr = new_array(klass_node, zlen, 1); 5297 // Update IdealKit memory and control from graphKit. 5298 __ sync_kit(this); 5299 __ set(z_alloc, narr); 5300 } __ end_if(); 5301 5302 sync_kit(ideal); 5303 z = __ value(z_alloc); 5304 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5305 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5306 // Final sync IdealKit and GraphKit. 5307 final_sync(ideal); 5308#undef __ 5309 5310 Node* z_start = array_element_address(z, intcon(0), T_INT); 5311 5312 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5313 OptoRuntime::multiplyToLen_Type(), 5314 stubAddr, stubName, TypePtr::BOTTOM, 5315 x_start, xlen, y_start, ylen, z_start, zlen); 5316 } // original reexecute is set back here 5317 5318 C->set_has_split_ifs(true); // Has chance for split-if optimization 5319 set_result(z); 5320 return true; 5321} 5322 5323//-------------inline_squareToLen------------------------------------ 5324bool LibraryCallKit::inline_squareToLen() { 5325 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5326 5327 address stubAddr = StubRoutines::squareToLen(); 5328 if (stubAddr == NULL) { 5329 return false; // Intrinsic's stub is not implemented on this platform 5330 } 5331 const char* stubName = "squareToLen"; 5332 5333 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5334 5335 Node* x = argument(0); 5336 Node* len = argument(1); 5337 Node* z = argument(2); 5338 Node* zlen = argument(3); 5339 5340 const Type* x_type = x->Value(&_gvn); 5341 const Type* z_type = z->Value(&_gvn); 5342 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5343 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5344 if (top_x == NULL || top_x->klass() == NULL || 5345 top_z == NULL || top_z->klass() == NULL) { 5346 // failed array check 5347 return false; 5348 } 5349 5350 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5351 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5352 if (x_elem != T_INT || z_elem != T_INT) { 5353 return false; 5354 } 5355 5356 5357 Node* x_start = array_element_address(x, intcon(0), x_elem); 5358 Node* z_start = array_element_address(z, intcon(0), z_elem); 5359 5360 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5361 OptoRuntime::squareToLen_Type(), 5362 stubAddr, stubName, TypePtr::BOTTOM, 5363 x_start, len, z_start, zlen); 5364 5365 set_result(z); 5366 return true; 5367} 5368 5369//-------------inline_mulAdd------------------------------------------ 5370bool LibraryCallKit::inline_mulAdd() { 5371 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5372 5373 address stubAddr = StubRoutines::mulAdd(); 5374 if (stubAddr == NULL) { 5375 return false; // Intrinsic's stub is not implemented on this platform 5376 } 5377 const char* stubName = "mulAdd"; 5378 5379 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5380 5381 Node* out = argument(0); 5382 Node* in = argument(1); 5383 Node* offset = argument(2); 5384 Node* len = argument(3); 5385 Node* k = argument(4); 5386 5387 const Type* out_type = out->Value(&_gvn); 5388 const Type* in_type = in->Value(&_gvn); 5389 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5390 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5391 if (top_out == NULL || top_out->klass() == NULL || 5392 top_in == NULL || top_in->klass() == NULL) { 5393 // failed array check 5394 return false; 5395 } 5396 5397 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5398 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5399 if (out_elem != T_INT || in_elem != T_INT) { 5400 return false; 5401 } 5402 5403 Node* outlen = load_array_length(out); 5404 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5405 Node* out_start = array_element_address(out, intcon(0), out_elem); 5406 Node* in_start = array_element_address(in, intcon(0), in_elem); 5407 5408 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5409 OptoRuntime::mulAdd_Type(), 5410 stubAddr, stubName, TypePtr::BOTTOM, 5411 out_start,in_start, new_offset, len, k); 5412 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5413 set_result(result); 5414 return true; 5415} 5416 5417//-------------inline_montgomeryMultiply----------------------------------- 5418bool LibraryCallKit::inline_montgomeryMultiply() { 5419 address stubAddr = StubRoutines::montgomeryMultiply(); 5420 if (stubAddr == NULL) { 5421 return false; // Intrinsic's stub is not implemented on this platform 5422 } 5423 5424 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5425 const char* stubName = "montgomery_square"; 5426 5427 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5428 5429 Node* a = argument(0); 5430 Node* b = argument(1); 5431 Node* n = argument(2); 5432 Node* len = argument(3); 5433 Node* inv = argument(4); 5434 Node* m = argument(6); 5435 5436 const Type* a_type = a->Value(&_gvn); 5437 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5438 const Type* b_type = b->Value(&_gvn); 5439 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5440 const Type* n_type = a->Value(&_gvn); 5441 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5442 const Type* m_type = a->Value(&_gvn); 5443 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5444 if (top_a == NULL || top_a->klass() == NULL || 5445 top_b == NULL || top_b->klass() == NULL || 5446 top_n == NULL || top_n->klass() == NULL || 5447 top_m == NULL || top_m->klass() == NULL) { 5448 // failed array check 5449 return false; 5450 } 5451 5452 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5453 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5454 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5455 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5456 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5457 return false; 5458 } 5459 5460 // Make the call 5461 { 5462 Node* a_start = array_element_address(a, intcon(0), a_elem); 5463 Node* b_start = array_element_address(b, intcon(0), b_elem); 5464 Node* n_start = array_element_address(n, intcon(0), n_elem); 5465 Node* m_start = array_element_address(m, intcon(0), m_elem); 5466 5467 Node* call = make_runtime_call(RC_LEAF, 5468 OptoRuntime::montgomeryMultiply_Type(), 5469 stubAddr, stubName, TypePtr::BOTTOM, 5470 a_start, b_start, n_start, len, inv, top(), 5471 m_start); 5472 set_result(m); 5473 } 5474 5475 return true; 5476} 5477 5478bool LibraryCallKit::inline_montgomerySquare() { 5479 address stubAddr = StubRoutines::montgomerySquare(); 5480 if (stubAddr == NULL) { 5481 return false; // Intrinsic's stub is not implemented on this platform 5482 } 5483 5484 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5485 const char* stubName = "montgomery_square"; 5486 5487 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5488 5489 Node* a = argument(0); 5490 Node* n = argument(1); 5491 Node* len = argument(2); 5492 Node* inv = argument(3); 5493 Node* m = argument(5); 5494 5495 const Type* a_type = a->Value(&_gvn); 5496 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5497 const Type* n_type = a->Value(&_gvn); 5498 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5499 const Type* m_type = a->Value(&_gvn); 5500 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5501 if (top_a == NULL || top_a->klass() == NULL || 5502 top_n == NULL || top_n->klass() == NULL || 5503 top_m == NULL || top_m->klass() == NULL) { 5504 // failed array check 5505 return false; 5506 } 5507 5508 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5509 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5510 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5511 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5512 return false; 5513 } 5514 5515 // Make the call 5516 { 5517 Node* a_start = array_element_address(a, intcon(0), a_elem); 5518 Node* n_start = array_element_address(n, intcon(0), n_elem); 5519 Node* m_start = array_element_address(m, intcon(0), m_elem); 5520 5521 Node* call = make_runtime_call(RC_LEAF, 5522 OptoRuntime::montgomerySquare_Type(), 5523 stubAddr, stubName, TypePtr::BOTTOM, 5524 a_start, n_start, len, inv, top(), 5525 m_start); 5526 set_result(m); 5527 } 5528 5529 return true; 5530} 5531 5532//-------------inline_vectorizedMismatch------------------------------ 5533bool LibraryCallKit::inline_vectorizedMismatch() { 5534 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5535 5536 address stubAddr = StubRoutines::vectorizedMismatch(); 5537 if (stubAddr == NULL) { 5538 return false; // Intrinsic's stub is not implemented on this platform 5539 } 5540 const char* stubName = "vectorizedMismatch"; 5541 int size_l = callee()->signature()->size(); 5542 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5543 5544 Node* obja = argument(0); 5545 Node* aoffset = argument(1); 5546 Node* objb = argument(3); 5547 Node* boffset = argument(4); 5548 Node* length = argument(6); 5549 Node* scale = argument(7); 5550 5551 const Type* a_type = obja->Value(&_gvn); 5552 const Type* b_type = objb->Value(&_gvn); 5553 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5554 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5555 if (top_a == NULL || top_a->klass() == NULL || 5556 top_b == NULL || top_b->klass() == NULL) { 5557 // failed array check 5558 return false; 5559 } 5560 5561 Node* call; 5562 jvms()->set_should_reexecute(true); 5563 5564 Node* obja_adr = make_unsafe_address(obja, aoffset); 5565 Node* objb_adr = make_unsafe_address(objb, boffset); 5566 5567 call = make_runtime_call(RC_LEAF, 5568 OptoRuntime::vectorizedMismatch_Type(), 5569 stubAddr, stubName, TypePtr::BOTTOM, 5570 obja_adr, objb_adr, length, scale); 5571 5572 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5573 set_result(result); 5574 return true; 5575} 5576 5577/** 5578 * Calculate CRC32 for byte. 5579 * int java.util.zip.CRC32.update(int crc, int b) 5580 */ 5581bool LibraryCallKit::inline_updateCRC32() { 5582 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5583 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5584 // no receiver since it is static method 5585 Node* crc = argument(0); // type: int 5586 Node* b = argument(1); // type: int 5587 5588 /* 5589 * int c = ~ crc; 5590 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5591 * b = b ^ (c >>> 8); 5592 * crc = ~b; 5593 */ 5594 5595 Node* M1 = intcon(-1); 5596 crc = _gvn.transform(new XorINode(crc, M1)); 5597 Node* result = _gvn.transform(new XorINode(crc, b)); 5598 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5599 5600 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5601 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5602 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5603 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5604 5605 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5606 result = _gvn.transform(new XorINode(crc, result)); 5607 result = _gvn.transform(new XorINode(result, M1)); 5608 set_result(result); 5609 return true; 5610} 5611 5612/** 5613 * Calculate CRC32 for byte[] array. 5614 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5615 */ 5616bool LibraryCallKit::inline_updateBytesCRC32() { 5617 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5618 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5619 // no receiver since it is static method 5620 Node* crc = argument(0); // type: int 5621 Node* src = argument(1); // type: oop 5622 Node* offset = argument(2); // type: int 5623 Node* length = argument(3); // type: int 5624 5625 const Type* src_type = src->Value(&_gvn); 5626 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5627 if (top_src == NULL || top_src->klass() == NULL) { 5628 // failed array check 5629 return false; 5630 } 5631 5632 // Figure out the size and type of the elements we will be copying. 5633 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5634 if (src_elem != T_BYTE) { 5635 return false; 5636 } 5637 5638 // 'src_start' points to src array + scaled offset 5639 Node* src_start = array_element_address(src, offset, src_elem); 5640 5641 // We assume that range check is done by caller. 5642 // TODO: generate range check (offset+length < src.length) in debug VM. 5643 5644 // Call the stub. 5645 address stubAddr = StubRoutines::updateBytesCRC32(); 5646 const char *stubName = "updateBytesCRC32"; 5647 5648 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5649 stubAddr, stubName, TypePtr::BOTTOM, 5650 crc, src_start, length); 5651 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5652 set_result(result); 5653 return true; 5654} 5655 5656/** 5657 * Calculate CRC32 for ByteBuffer. 5658 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5659 */ 5660bool LibraryCallKit::inline_updateByteBufferCRC32() { 5661 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5662 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5663 // no receiver since it is static method 5664 Node* crc = argument(0); // type: int 5665 Node* src = argument(1); // type: long 5666 Node* offset = argument(3); // type: int 5667 Node* length = argument(4); // type: int 5668 5669 src = ConvL2X(src); // adjust Java long to machine word 5670 Node* base = _gvn.transform(new CastX2PNode(src)); 5671 offset = ConvI2X(offset); 5672 5673 // 'src_start' points to src array + scaled offset 5674 Node* src_start = basic_plus_adr(top(), base, offset); 5675 5676 // Call the stub. 5677 address stubAddr = StubRoutines::updateBytesCRC32(); 5678 const char *stubName = "updateBytesCRC32"; 5679 5680 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5681 stubAddr, stubName, TypePtr::BOTTOM, 5682 crc, src_start, length); 5683 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5684 set_result(result); 5685 return true; 5686} 5687 5688//------------------------------get_table_from_crc32c_class----------------------- 5689Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5690 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5691 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5692 5693 return table; 5694} 5695 5696//------------------------------inline_updateBytesCRC32C----------------------- 5697// 5698// Calculate CRC32C for byte[] array. 5699// int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5700// 5701bool LibraryCallKit::inline_updateBytesCRC32C() { 5702 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5703 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5704 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5705 // no receiver since it is a static method 5706 Node* crc = argument(0); // type: int 5707 Node* src = argument(1); // type: oop 5708 Node* offset = argument(2); // type: int 5709 Node* end = argument(3); // type: int 5710 5711 Node* length = _gvn.transform(new SubINode(end, offset)); 5712 5713 const Type* src_type = src->Value(&_gvn); 5714 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5715 if (top_src == NULL || top_src->klass() == NULL) { 5716 // failed array check 5717 return false; 5718 } 5719 5720 // Figure out the size and type of the elements we will be copying. 5721 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5722 if (src_elem != T_BYTE) { 5723 return false; 5724 } 5725 5726 // 'src_start' points to src array + scaled offset 5727 Node* src_start = array_element_address(src, offset, src_elem); 5728 5729 // static final int[] byteTable in class CRC32C 5730 Node* table = get_table_from_crc32c_class(callee()->holder()); 5731 Node* table_start = array_element_address(table, intcon(0), T_INT); 5732 5733 // We assume that range check is done by caller. 5734 // TODO: generate range check (offset+length < src.length) in debug VM. 5735 5736 // Call the stub. 5737 address stubAddr = StubRoutines::updateBytesCRC32C(); 5738 const char *stubName = "updateBytesCRC32C"; 5739 5740 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5741 stubAddr, stubName, TypePtr::BOTTOM, 5742 crc, src_start, length, table_start); 5743 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5744 set_result(result); 5745 return true; 5746} 5747 5748//------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5749// 5750// Calculate CRC32C for DirectByteBuffer. 5751// int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5752// 5753bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5754 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5755 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5756 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5757 // no receiver since it is a static method 5758 Node* crc = argument(0); // type: int 5759 Node* src = argument(1); // type: long 5760 Node* offset = argument(3); // type: int 5761 Node* end = argument(4); // type: int 5762 5763 Node* length = _gvn.transform(new SubINode(end, offset)); 5764 5765 src = ConvL2X(src); // adjust Java long to machine word 5766 Node* base = _gvn.transform(new CastX2PNode(src)); 5767 offset = ConvI2X(offset); 5768 5769 // 'src_start' points to src array + scaled offset 5770 Node* src_start = basic_plus_adr(top(), base, offset); 5771 5772 // static final int[] byteTable in class CRC32C 5773 Node* table = get_table_from_crc32c_class(callee()->holder()); 5774 Node* table_start = array_element_address(table, intcon(0), T_INT); 5775 5776 // Call the stub. 5777 address stubAddr = StubRoutines::updateBytesCRC32C(); 5778 const char *stubName = "updateBytesCRC32C"; 5779 5780 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5781 stubAddr, stubName, TypePtr::BOTTOM, 5782 crc, src_start, length, table_start); 5783 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5784 set_result(result); 5785 return true; 5786} 5787 5788//------------------------------inline_updateBytesAdler32---------------------- 5789// 5790// Calculate Adler32 checksum for byte[] array. 5791// int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5792// 5793bool LibraryCallKit::inline_updateBytesAdler32() { 5794 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5795 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5796 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5797 // no receiver since it is static method 5798 Node* crc = argument(0); // type: int 5799 Node* src = argument(1); // type: oop 5800 Node* offset = argument(2); // type: int 5801 Node* length = argument(3); // type: int 5802 5803 const Type* src_type = src->Value(&_gvn); 5804 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5805 if (top_src == NULL || top_src->klass() == NULL) { 5806 // failed array check 5807 return false; 5808 } 5809 5810 // Figure out the size and type of the elements we will be copying. 5811 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5812 if (src_elem != T_BYTE) { 5813 return false; 5814 } 5815 5816 // 'src_start' points to src array + scaled offset 5817 Node* src_start = array_element_address(src, offset, src_elem); 5818 5819 // We assume that range check is done by caller. 5820 // TODO: generate range check (offset+length < src.length) in debug VM. 5821 5822 // Call the stub. 5823 address stubAddr = StubRoutines::updateBytesAdler32(); 5824 const char *stubName = "updateBytesAdler32"; 5825 5826 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5827 stubAddr, stubName, TypePtr::BOTTOM, 5828 crc, src_start, length); 5829 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5830 set_result(result); 5831 return true; 5832} 5833 5834//------------------------------inline_updateByteBufferAdler32--------------- 5835// 5836// Calculate Adler32 checksum for DirectByteBuffer. 5837// int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5838// 5839bool LibraryCallKit::inline_updateByteBufferAdler32() { 5840 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5841 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5842 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5843 // no receiver since it is static method 5844 Node* crc = argument(0); // type: int 5845 Node* src = argument(1); // type: long 5846 Node* offset = argument(3); // type: int 5847 Node* length = argument(4); // type: int 5848 5849 src = ConvL2X(src); // adjust Java long to machine word 5850 Node* base = _gvn.transform(new CastX2PNode(src)); 5851 offset = ConvI2X(offset); 5852 5853 // 'src_start' points to src array + scaled offset 5854 Node* src_start = basic_plus_adr(top(), base, offset); 5855 5856 // Call the stub. 5857 address stubAddr = StubRoutines::updateBytesAdler32(); 5858 const char *stubName = "updateBytesAdler32"; 5859 5860 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5861 stubAddr, stubName, TypePtr::BOTTOM, 5862 crc, src_start, length); 5863 5864 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5865 set_result(result); 5866 return true; 5867} 5868 5869//----------------------------inline_reference_get---------------------------- 5870// public T java.lang.ref.Reference.get(); 5871bool LibraryCallKit::inline_reference_get() { 5872 const int referent_offset = java_lang_ref_Reference::referent_offset; 5873 guarantee(referent_offset > 0, "should have already been set"); 5874 5875 // Get the argument: 5876 Node* reference_obj = null_check_receiver(); 5877 if (stopped()) return true; 5878 5879 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5880 5881 ciInstanceKlass* klass = env()->Object_klass(); 5882 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5883 5884 Node* no_ctrl = NULL; 5885 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5886 5887 // Use the pre-barrier to record the value in the referent field 5888 pre_barrier(false /* do_load */, 5889 control(), 5890 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5891 result /* pre_val */, 5892 T_OBJECT); 5893 5894 // Add memory barrier to prevent commoning reads from this field 5895 // across safepoint since GC can change its value. 5896 insert_mem_bar(Op_MemBarCPUOrder); 5897 5898 set_result(result); 5899 return true; 5900} 5901 5902 5903Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5904 bool is_exact=true, bool is_static=false, 5905 ciInstanceKlass * fromKls=NULL) { 5906 if (fromKls == NULL) { 5907 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5908 assert(tinst != NULL, "obj is null"); 5909 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5910 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5911 fromKls = tinst->klass()->as_instance_klass(); 5912 } else { 5913 assert(is_static, "only for static field access"); 5914 } 5915 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5916 ciSymbol::make(fieldTypeString), 5917 is_static); 5918 5919 assert (field != NULL, "undefined field"); 5920 if (field == NULL) return (Node *) NULL; 5921 5922 if (is_static) { 5923 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5924 fromObj = makecon(tip); 5925 } 5926 5927 // Next code copied from Parse::do_get_xxx(): 5928 5929 // Compute address and memory type. 5930 int offset = field->offset_in_bytes(); 5931 bool is_vol = field->is_volatile(); 5932 ciType* field_klass = field->type(); 5933 assert(field_klass->is_loaded(), "should be loaded"); 5934 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5935 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5936 BasicType bt = field->layout_type(); 5937 5938 // Build the resultant type of the load 5939 const Type *type; 5940 if (bt == T_OBJECT) { 5941 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5942 } else { 5943 type = Type::get_const_basic_type(bt); 5944 } 5945 5946 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5947 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5948 } 5949 // Build the load. 5950 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5951 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5952 // If reference is volatile, prevent following memory ops from 5953 // floating up past the volatile read. Also prevents commoning 5954 // another volatile read. 5955 if (is_vol) { 5956 // Memory barrier includes bogus read of value to force load BEFORE membar 5957 insert_mem_bar(Op_MemBarAcquire, loadedField); 5958 } 5959 return loadedField; 5960} 5961 5962Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5963 bool is_exact = true, bool is_static = false, 5964 ciInstanceKlass * fromKls = NULL) { 5965 if (fromKls == NULL) { 5966 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5967 assert(tinst != NULL, "obj is null"); 5968 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5969 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5970 fromKls = tinst->klass()->as_instance_klass(); 5971 } 5972 else { 5973 assert(is_static, "only for static field access"); 5974 } 5975 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5976 ciSymbol::make(fieldTypeString), 5977 is_static); 5978 5979 assert(field != NULL, "undefined field"); 5980 assert(!field->is_volatile(), "not defined for volatile fields"); 5981 5982 if (is_static) { 5983 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5984 fromObj = makecon(tip); 5985 } 5986 5987 // Next code copied from Parse::do_get_xxx(): 5988 5989 // Compute address and memory type. 5990 int offset = field->offset_in_bytes(); 5991 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5992 5993 return adr; 5994} 5995 5996//------------------------------inline_aescrypt_Block----------------------- 5997bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5998 address stubAddr = NULL; 5999 const char *stubName; 6000 assert(UseAES, "need AES instruction support"); 6001 6002 switch(id) { 6003 case vmIntrinsics::_aescrypt_encryptBlock: 6004 stubAddr = StubRoutines::aescrypt_encryptBlock(); 6005 stubName = "aescrypt_encryptBlock"; 6006 break; 6007 case vmIntrinsics::_aescrypt_decryptBlock: 6008 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6009 stubName = "aescrypt_decryptBlock"; 6010 break; 6011 } 6012 if (stubAddr == NULL) return false; 6013 6014 Node* aescrypt_object = argument(0); 6015 Node* src = argument(1); 6016 Node* src_offset = argument(2); 6017 Node* dest = argument(3); 6018 Node* dest_offset = argument(4); 6019 6020 // (1) src and dest are arrays. 6021 const Type* src_type = src->Value(&_gvn); 6022 const Type* dest_type = dest->Value(&_gvn); 6023 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6024 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6025 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6026 6027 // for the quick and dirty code we will skip all the checks. 6028 // we are just trying to get the call to be generated. 6029 Node* src_start = src; 6030 Node* dest_start = dest; 6031 if (src_offset != NULL || dest_offset != NULL) { 6032 assert(src_offset != NULL && dest_offset != NULL, ""); 6033 src_start = array_element_address(src, src_offset, T_BYTE); 6034 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6035 } 6036 6037 // now need to get the start of its expanded key array 6038 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6039 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6040 if (k_start == NULL) return false; 6041 6042 if (Matcher::pass_original_key_for_aes()) { 6043 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6044 // compatibility issues between Java key expansion and SPARC crypto instructions 6045 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6046 if (original_k_start == NULL) return false; 6047 6048 // Call the stub. 6049 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6050 stubAddr, stubName, TypePtr::BOTTOM, 6051 src_start, dest_start, k_start, original_k_start); 6052 } else { 6053 // Call the stub. 6054 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6055 stubAddr, stubName, TypePtr::BOTTOM, 6056 src_start, dest_start, k_start); 6057 } 6058 6059 return true; 6060} 6061 6062//------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6063bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6064 address stubAddr = NULL; 6065 const char *stubName = NULL; 6066 6067 assert(UseAES, "need AES instruction support"); 6068 6069 switch(id) { 6070 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6071 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6072 stubName = "cipherBlockChaining_encryptAESCrypt"; 6073 break; 6074 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6075 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6076 stubName = "cipherBlockChaining_decryptAESCrypt"; 6077 break; 6078 } 6079 if (stubAddr == NULL) return false; 6080 6081 Node* cipherBlockChaining_object = argument(0); 6082 Node* src = argument(1); 6083 Node* src_offset = argument(2); 6084 Node* len = argument(3); 6085 Node* dest = argument(4); 6086 Node* dest_offset = argument(5); 6087 6088 // (1) src and dest are arrays. 6089 const Type* src_type = src->Value(&_gvn); 6090 const Type* dest_type = dest->Value(&_gvn); 6091 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6092 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6093 assert (top_src != NULL && top_src->klass() != NULL 6094 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6095 6096 // checks are the responsibility of the caller 6097 Node* src_start = src; 6098 Node* dest_start = dest; 6099 if (src_offset != NULL || dest_offset != NULL) { 6100 assert(src_offset != NULL && dest_offset != NULL, ""); 6101 src_start = array_element_address(src, src_offset, T_BYTE); 6102 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6103 } 6104 6105 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6106 // (because of the predicated logic executed earlier). 6107 // so we cast it here safely. 6108 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6109 6110 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6111 if (embeddedCipherObj == NULL) return false; 6112 6113 // cast it to what we know it will be at runtime 6114 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6115 assert(tinst != NULL, "CBC obj is null"); 6116 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6117 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6118 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6119 6120 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6121 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6122 const TypeOopPtr* xtype = aklass->as_instance_type(); 6123 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6124 aescrypt_object = _gvn.transform(aescrypt_object); 6125 6126 // we need to get the start of the aescrypt_object's expanded key array 6127 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6128 if (k_start == NULL) return false; 6129 6130 // similarly, get the start address of the r vector 6131 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6132 if (objRvec == NULL) return false; 6133 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6134 6135 Node* cbcCrypt; 6136 if (Matcher::pass_original_key_for_aes()) { 6137 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6138 // compatibility issues between Java key expansion and SPARC crypto instructions 6139 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6140 if (original_k_start == NULL) return false; 6141 6142 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6143 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6144 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6145 stubAddr, stubName, TypePtr::BOTTOM, 6146 src_start, dest_start, k_start, r_start, len, original_k_start); 6147 } else { 6148 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6149 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6150 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6151 stubAddr, stubName, TypePtr::BOTTOM, 6152 src_start, dest_start, k_start, r_start, len); 6153 } 6154 6155 // return cipher length (int) 6156 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6157 set_result(retvalue); 6158 return true; 6159} 6160 6161//------------------------------inline_counterMode_AESCrypt----------------------- 6162bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6163 assert(UseAES, "need AES instruction support"); 6164 if (!UseAESCTRIntrinsics) return false; 6165 6166 address stubAddr = NULL; 6167 const char *stubName = NULL; 6168 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6169 stubAddr = StubRoutines::counterMode_AESCrypt(); 6170 stubName = "counterMode_AESCrypt"; 6171 } 6172 if (stubAddr == NULL) return false; 6173 6174 Node* counterMode_object = argument(0); 6175 Node* src = argument(1); 6176 Node* src_offset = argument(2); 6177 Node* len = argument(3); 6178 Node* dest = argument(4); 6179 Node* dest_offset = argument(5); 6180 6181 // (1) src and dest are arrays. 6182 const Type* src_type = src->Value(&_gvn); 6183 const Type* dest_type = dest->Value(&_gvn); 6184 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6185 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6186 assert(top_src != NULL && top_src->klass() != NULL && 6187 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6188 6189 // checks are the responsibility of the caller 6190 Node* src_start = src; 6191 Node* dest_start = dest; 6192 if (src_offset != NULL || dest_offset != NULL) { 6193 assert(src_offset != NULL && dest_offset != NULL, ""); 6194 src_start = array_element_address(src, src_offset, T_BYTE); 6195 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6196 } 6197 6198 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6199 // (because of the predicated logic executed earlier). 6200 // so we cast it here safely. 6201 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6202 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6203 if (embeddedCipherObj == NULL) return false; 6204 // cast it to what we know it will be at runtime 6205 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6206 assert(tinst != NULL, "CTR obj is null"); 6207 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6208 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6209 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6210 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6211 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6212 const TypeOopPtr* xtype = aklass->as_instance_type(); 6213 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6214 aescrypt_object = _gvn.transform(aescrypt_object); 6215 // we need to get the start of the aescrypt_object's expanded key array 6216 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6217 if (k_start == NULL) return false; 6218 // similarly, get the start address of the r vector 6219 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6220 if (obj_counter == NULL) return false; 6221 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6222 6223 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6224 if (saved_encCounter == NULL) return false; 6225 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6226 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6227 6228 Node* ctrCrypt; 6229 if (Matcher::pass_original_key_for_aes()) { 6230 // no SPARC version for AES/CTR intrinsics now. 6231 return false; 6232 } 6233 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6234 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6235 OptoRuntime::counterMode_aescrypt_Type(), 6236 stubAddr, stubName, TypePtr::BOTTOM, 6237 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6238 6239 // return cipher length (int) 6240 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6241 set_result(retvalue); 6242 return true; 6243} 6244 6245//------------------------------get_key_start_from_aescrypt_object----------------------- 6246Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6247#ifdef PPC64 6248 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6249 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6250 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6251 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6252 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6253 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6254 if (objSessionK == NULL) { 6255 return (Node *) NULL; 6256 } 6257 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6258#else 6259 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6260#endif // PPC64 6261 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6262 if (objAESCryptKey == NULL) return (Node *) NULL; 6263 6264 // now have the array, need to get the start address of the K array 6265 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6266 return k_start; 6267} 6268 6269//------------------------------get_original_key_start_from_aescrypt_object----------------------- 6270Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6271 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6272 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6273 if (objAESCryptKey == NULL) return (Node *) NULL; 6274 6275 // now have the array, need to get the start address of the lastKey array 6276 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6277 return original_k_start; 6278} 6279 6280//----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6281// Return node representing slow path of predicate check. 6282// the pseudo code we want to emulate with this predicate is: 6283// for encryption: 6284// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6285// for decryption: 6286// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6287// note cipher==plain is more conservative than the original java code but that's OK 6288// 6289Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6290 // The receiver was checked for NULL already. 6291 Node* objCBC = argument(0); 6292 6293 // Load embeddedCipher field of CipherBlockChaining object. 6294 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6295 6296 // get AESCrypt klass for instanceOf check 6297 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6298 // will have same classloader as CipherBlockChaining object 6299 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6300 assert(tinst != NULL, "CBCobj is null"); 6301 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6302 6303 // we want to do an instanceof comparison against the AESCrypt class 6304 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6305 if (!klass_AESCrypt->is_loaded()) { 6306 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6307 Node* ctrl = control(); 6308 set_control(top()); // no regular fast path 6309 return ctrl; 6310 } 6311 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6312 6313 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6314 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6315 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6316 6317 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6318 6319 // for encryption, we are done 6320 if (!decrypting) 6321 return instof_false; // even if it is NULL 6322 6323 // for decryption, we need to add a further check to avoid 6324 // taking the intrinsic path when cipher and plain are the same 6325 // see the original java code for why. 6326 RegionNode* region = new RegionNode(3); 6327 region->init_req(1, instof_false); 6328 Node* src = argument(1); 6329 Node* dest = argument(4); 6330 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6331 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6332 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6333 region->init_req(2, src_dest_conjoint); 6334 6335 record_for_igvn(region); 6336 return _gvn.transform(region); 6337} 6338 6339//----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6340// Return node representing slow path of predicate check. 6341// the pseudo code we want to emulate with this predicate is: 6342// for encryption: 6343// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6344// for decryption: 6345// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6346// note cipher==plain is more conservative than the original java code but that's OK 6347// 6348 6349Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6350 // The receiver was checked for NULL already. 6351 Node* objCTR = argument(0); 6352 6353 // Load embeddedCipher field of CipherBlockChaining object. 6354 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6355 6356 // get AESCrypt klass for instanceOf check 6357 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6358 // will have same classloader as CipherBlockChaining object 6359 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6360 assert(tinst != NULL, "CTRobj is null"); 6361 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6362 6363 // we want to do an instanceof comparison against the AESCrypt class 6364 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6365 if (!klass_AESCrypt->is_loaded()) { 6366 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6367 Node* ctrl = control(); 6368 set_control(top()); // no regular fast path 6369 return ctrl; 6370 } 6371 6372 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6373 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6374 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6375 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6376 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6377 6378 return instof_false; // even if it is NULL 6379} 6380 6381//------------------------------inline_ghash_processBlocks 6382bool LibraryCallKit::inline_ghash_processBlocks() { 6383 address stubAddr; 6384 const char *stubName; 6385 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6386 6387 stubAddr = StubRoutines::ghash_processBlocks(); 6388 stubName = "ghash_processBlocks"; 6389 6390 Node* data = argument(0); 6391 Node* offset = argument(1); 6392 Node* len = argument(2); 6393 Node* state = argument(3); 6394 Node* subkeyH = argument(4); 6395 6396 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6397 assert(state_start, "state is NULL"); 6398 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6399 assert(subkeyH_start, "subkeyH is NULL"); 6400 Node* data_start = array_element_address(data, offset, T_BYTE); 6401 assert(data_start, "data is NULL"); 6402 6403 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6404 OptoRuntime::ghash_processBlocks_Type(), 6405 stubAddr, stubName, TypePtr::BOTTOM, 6406 state_start, subkeyH_start, data_start, len); 6407 return true; 6408} 6409 6410//------------------------------inline_sha_implCompress----------------------- 6411// 6412// Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6413// void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6414// 6415// Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6416// void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6417// 6418// Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6419// void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6420// 6421bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6422 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6423 6424 Node* sha_obj = argument(0); 6425 Node* src = argument(1); // type oop 6426 Node* ofs = argument(2); // type int 6427 6428 const Type* src_type = src->Value(&_gvn); 6429 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6430 if (top_src == NULL || top_src->klass() == NULL) { 6431 // failed array check 6432 return false; 6433 } 6434 // Figure out the size and type of the elements we will be copying. 6435 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6436 if (src_elem != T_BYTE) { 6437 return false; 6438 } 6439 // 'src_start' points to src array + offset 6440 Node* src_start = array_element_address(src, ofs, src_elem); 6441 Node* state = NULL; 6442 address stubAddr; 6443 const char *stubName; 6444 6445 switch(id) { 6446 case vmIntrinsics::_sha_implCompress: 6447 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6448 state = get_state_from_sha_object(sha_obj); 6449 stubAddr = StubRoutines::sha1_implCompress(); 6450 stubName = "sha1_implCompress"; 6451 break; 6452 case vmIntrinsics::_sha2_implCompress: 6453 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6454 state = get_state_from_sha_object(sha_obj); 6455 stubAddr = StubRoutines::sha256_implCompress(); 6456 stubName = "sha256_implCompress"; 6457 break; 6458 case vmIntrinsics::_sha5_implCompress: 6459 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6460 state = get_state_from_sha5_object(sha_obj); 6461 stubAddr = StubRoutines::sha512_implCompress(); 6462 stubName = "sha512_implCompress"; 6463 break; 6464 default: 6465 fatal_unexpected_iid(id); 6466 return false; 6467 } 6468 if (state == NULL) return false; 6469 6470 // Call the stub. 6471 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6472 stubAddr, stubName, TypePtr::BOTTOM, 6473 src_start, state); 6474 6475 return true; 6476} 6477 6478//------------------------------inline_digestBase_implCompressMB----------------------- 6479// 6480// Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6481// int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6482// 6483bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6484 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6485 "need SHA1/SHA256/SHA512 instruction support"); 6486 assert((uint)predicate < 3, "sanity"); 6487 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6488 6489 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6490 Node* src = argument(1); // byte[] array 6491 Node* ofs = argument(2); // type int 6492 Node* limit = argument(3); // type int 6493 6494 const Type* src_type = src->Value(&_gvn); 6495 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6496 if (top_src == NULL || top_src->klass() == NULL) { 6497 // failed array check 6498 return false; 6499 } 6500 // Figure out the size and type of the elements we will be copying. 6501 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6502 if (src_elem != T_BYTE) { 6503 return false; 6504 } 6505 // 'src_start' points to src array + offset 6506 Node* src_start = array_element_address(src, ofs, src_elem); 6507 6508 const char* klass_SHA_name = NULL; 6509 const char* stub_name = NULL; 6510 address stub_addr = NULL; 6511 bool long_state = false; 6512 6513 switch (predicate) { 6514 case 0: 6515 if (UseSHA1Intrinsics) { 6516 klass_SHA_name = "sun/security/provider/SHA"; 6517 stub_name = "sha1_implCompressMB"; 6518 stub_addr = StubRoutines::sha1_implCompressMB(); 6519 } 6520 break; 6521 case 1: 6522 if (UseSHA256Intrinsics) { 6523 klass_SHA_name = "sun/security/provider/SHA2"; 6524 stub_name = "sha256_implCompressMB"; 6525 stub_addr = StubRoutines::sha256_implCompressMB(); 6526 } 6527 break; 6528 case 2: 6529 if (UseSHA512Intrinsics) { 6530 klass_SHA_name = "sun/security/provider/SHA5"; 6531 stub_name = "sha512_implCompressMB"; 6532 stub_addr = StubRoutines::sha512_implCompressMB(); 6533 long_state = true; 6534 } 6535 break; 6536 default: 6537 fatal("unknown SHA intrinsic predicate: %d", predicate); 6538 } 6539 if (klass_SHA_name != NULL) { 6540 // get DigestBase klass to lookup for SHA klass 6541 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6542 assert(tinst != NULL, "digestBase_obj is not instance???"); 6543 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6544 6545 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6546 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6547 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6548 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6549 } 6550 return false; 6551} 6552//------------------------------inline_sha_implCompressMB----------------------- 6553bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6554 bool long_state, address stubAddr, const char *stubName, 6555 Node* src_start, Node* ofs, Node* limit) { 6556 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6557 const TypeOopPtr* xtype = aklass->as_instance_type(); 6558 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6559 sha_obj = _gvn.transform(sha_obj); 6560 6561 Node* state; 6562 if (long_state) { 6563 state = get_state_from_sha5_object(sha_obj); 6564 } else { 6565 state = get_state_from_sha_object(sha_obj); 6566 } 6567 if (state == NULL) return false; 6568 6569 // Call the stub. 6570 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6571 OptoRuntime::digestBase_implCompressMB_Type(), 6572 stubAddr, stubName, TypePtr::BOTTOM, 6573 src_start, state, ofs, limit); 6574 // return ofs (int) 6575 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6576 set_result(result); 6577 6578 return true; 6579} 6580 6581//------------------------------get_state_from_sha_object----------------------- 6582Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6583 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6584 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6585 if (sha_state == NULL) return (Node *) NULL; 6586 6587 // now have the array, need to get the start address of the state array 6588 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6589 return state; 6590} 6591 6592//------------------------------get_state_from_sha5_object----------------------- 6593Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6594 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6595 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6596 if (sha_state == NULL) return (Node *) NULL; 6597 6598 // now have the array, need to get the start address of the state array 6599 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6600 return state; 6601} 6602 6603//----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6604// Return node representing slow path of predicate check. 6605// the pseudo code we want to emulate with this predicate is: 6606// if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6607// 6608Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6609 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6610 "need SHA1/SHA256/SHA512 instruction support"); 6611 assert((uint)predicate < 3, "sanity"); 6612 6613 // The receiver was checked for NULL already. 6614 Node* digestBaseObj = argument(0); 6615 6616 // get DigestBase klass for instanceOf check 6617 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6618 assert(tinst != NULL, "digestBaseObj is null"); 6619 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6620 6621 const char* klass_SHA_name = NULL; 6622 switch (predicate) { 6623 case 0: 6624 if (UseSHA1Intrinsics) { 6625 // we want to do an instanceof comparison against the SHA class 6626 klass_SHA_name = "sun/security/provider/SHA"; 6627 } 6628 break; 6629 case 1: 6630 if (UseSHA256Intrinsics) { 6631 // we want to do an instanceof comparison against the SHA2 class 6632 klass_SHA_name = "sun/security/provider/SHA2"; 6633 } 6634 break; 6635 case 2: 6636 if (UseSHA512Intrinsics) { 6637 // we want to do an instanceof comparison against the SHA5 class 6638 klass_SHA_name = "sun/security/provider/SHA5"; 6639 } 6640 break; 6641 default: 6642 fatal("unknown SHA intrinsic predicate: %d", predicate); 6643 } 6644 6645 ciKlass* klass_SHA = NULL; 6646 if (klass_SHA_name != NULL) { 6647 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6648 } 6649 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6650 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6651 Node* ctrl = control(); 6652 set_control(top()); // no intrinsic path 6653 return ctrl; 6654 } 6655 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6656 6657 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6658 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6659 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6660 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6661 6662 return instof_false; // even if it is NULL 6663} 6664 6665//-------------inline_fma----------------------------------- 6666bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 6667 Node *a = NULL; 6668 Node *b = NULL; 6669 Node *c = NULL; 6670 Node* result = NULL; 6671 switch (id) { 6672 case vmIntrinsics::_fmaD: 6673 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 6674 // no receiver since it is static method 6675 a = round_double_node(argument(0)); 6676 b = round_double_node(argument(2)); 6677 c = round_double_node(argument(4)); 6678 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 6679 break; 6680 case vmIntrinsics::_fmaF: 6681 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 6682 a = argument(0); 6683 b = argument(1); 6684 c = argument(2); 6685 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 6686 break; 6687 default: 6688 fatal_unexpected_iid(id); break; 6689 } 6690 set_result(result); 6691 return true; 6692} 6693 6694bool LibraryCallKit::inline_profileBoolean() { 6695 Node* counts = argument(1); 6696 const TypeAryPtr* ary = NULL; 6697 ciArray* aobj = NULL; 6698 if (counts->is_Con() 6699 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6700 && (aobj = ary->const_oop()->as_array()) != NULL 6701 && (aobj->length() == 2)) { 6702 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6703 jint false_cnt = aobj->element_value(0).as_int(); 6704 jint true_cnt = aobj->element_value(1).as_int(); 6705 6706 if (C->log() != NULL) { 6707 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6708 false_cnt, true_cnt); 6709 } 6710 6711 if (false_cnt + true_cnt == 0) { 6712 // According to profile, never executed. 6713 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6714 Deoptimization::Action_reinterpret); 6715 return true; 6716 } 6717 6718 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6719 // is a number of each value occurrences. 6720 Node* result = argument(0); 6721 if (false_cnt == 0 || true_cnt == 0) { 6722 // According to profile, one value has been never seen. 6723 int expected_val = (false_cnt == 0) ? 1 : 0; 6724 6725 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6726 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6727 6728 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6729 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6730 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6731 6732 { // Slow path: uncommon trap for never seen value and then reexecute 6733 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6734 // the value has been seen at least once. 6735 PreserveJVMState pjvms(this); 6736 PreserveReexecuteState preexecs(this); 6737 jvms()->set_should_reexecute(true); 6738 6739 set_control(slow_path); 6740 set_i_o(i_o()); 6741 6742 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6743 Deoptimization::Action_reinterpret); 6744 } 6745 // The guard for never seen value enables sharpening of the result and 6746 // returning a constant. It allows to eliminate branches on the same value 6747 // later on. 6748 set_control(fast_path); 6749 result = intcon(expected_val); 6750 } 6751 // Stop profiling. 6752 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6753 // By replacing method body with profile data (represented as ProfileBooleanNode 6754 // on IR level) we effectively disable profiling. 6755 // It enables full speed execution once optimized code is generated. 6756 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6757 C->record_for_igvn(profile); 6758 set_result(profile); 6759 return true; 6760 } else { 6761 // Continue profiling. 6762 // Profile data isn't available at the moment. So, execute method's bytecode version. 6763 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6764 // is compiled and counters aren't available since corresponding MethodHandle 6765 // isn't a compile-time constant. 6766 return false; 6767 } 6768} 6769 6770bool LibraryCallKit::inline_isCompileConstant() { 6771 Node* n = argument(0); 6772 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6773 return true; 6774} 6775