1 /* 2 * Copyright (c) 1999, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "ci/ciUtilities.inline.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "classfile/vmSymbols.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "gc/shared/barrierSet.hpp" 33 #include "jfr/support/jfrIntrinsics.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "oops/klass.inline.hpp" 36 #include "oops/objArrayKlass.hpp" 37 #include "opto/addnode.hpp" 38 #include "opto/arraycopynode.hpp" 39 #include "opto/c2compiler.hpp" 40 #include "opto/castnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/convertnode.hpp" 43 #include "opto/countbitsnode.hpp" 44 #include "opto/idealKit.hpp" 45 #include "opto/library_call.hpp" 46 #include "opto/mathexactnode.hpp" 47 #include "opto/mulnode.hpp" 48 #include "opto/narrowptrnode.hpp" 49 #include "opto/opaquenode.hpp" 50 #include "opto/parse.hpp" 51 #include "opto/runtime.hpp" 52 #include "opto/rootnode.hpp" 53 #include "opto/subnode.hpp" 54 #include "prims/nativeLookup.hpp" 55 #include "prims/unsafe.hpp" 56 #include "runtime/objectMonitor.hpp" 57 #include "runtime/sharedRuntime.hpp" 58 #include "utilities/macros.hpp" 59 #include "utilities/powerOfTwo.hpp" 60 61 //---------------------------make_vm_intrinsic---------------------------- 62 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 63 vmIntrinsics::ID id = m->intrinsic_id(); 64 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 65 66 if (!m->is_loaded()) { 67 // Do not attempt to inline unloaded methods. 68 return NULL; 69 } 70 71 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 72 bool is_available = false; 73 74 { 75 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 76 // the compiler must transition to '_thread_in_vm' state because both 77 // methods access VM-internal data. 78 VM_ENTRY_MARK; 79 methodHandle mh(THREAD, m->get_Method()); 80 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && 81 !C->directive()->is_intrinsic_disabled(mh) && 82 !vmIntrinsics::is_disabled_by_flags(mh); 83 84 } 85 86 if (is_available) { 87 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 88 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 89 return new LibraryIntrinsic(m, is_virtual, 90 vmIntrinsics::predicates_needed(id), 91 vmIntrinsics::does_virtual_dispatch(id), 92 (vmIntrinsics::ID) id); 93 } else { 94 return NULL; 95 } 96 } 97 98 //----------------------register_library_intrinsics----------------------- 99 // Initialize this file's data structures, for each Compile instance. 100 void Compile::register_library_intrinsics() { 101 // Nothing to do here. 102 } 103 104 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 105 LibraryCallKit kit(jvms, this); 106 Compile* C = kit.C; 107 int nodes = C->unique(); 108 #ifndef PRODUCT 109 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 110 char buf[1000]; 111 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 112 tty->print_cr("Intrinsic %s", str); 113 } 114 #endif 115 ciMethod* callee = kit.callee(); 116 const int bci = kit.bci(); 117 118 // Try to inline the intrinsic. 119 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 120 kit.try_to_inline(_last_predicate)) { 121 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" 122 : "(intrinsic)"; 123 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 124 if (C->print_intrinsics() || C->print_inlining()) { 125 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 126 } 127 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 128 if (C->log()) { 129 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 130 vmIntrinsics::name_at(intrinsic_id()), 131 (is_virtual() ? " virtual='1'" : ""), 132 C->unique() - nodes); 133 } 134 // Push the result from the inlined method onto the stack. 135 kit.push_result(); 136 C->print_inlining_update(this); 137 return kit.transfer_exceptions_into_jvms(); 138 } 139 140 // The intrinsic bailed out 141 if (jvms->has_method()) { 142 // Not a root compile. 143 const char* msg; 144 if (callee->intrinsic_candidate()) { 145 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 146 } else { 147 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 148 : "failed to inline (intrinsic), method not annotated"; 149 } 150 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); 151 if (C->print_intrinsics() || C->print_inlining()) { 152 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 153 } 154 } else { 155 // Root compile 156 ResourceMark rm; 157 stringStream msg_stream; 158 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 159 vmIntrinsics::name_at(intrinsic_id()), 160 is_virtual() ? " (virtual)" : "", bci); 161 const char *msg = msg_stream.as_string(); 162 log_debug(jit, inlining)("%s", msg); 163 if (C->print_intrinsics() || C->print_inlining()) { 164 tty->print("%s", msg); 165 } 166 } 167 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 168 C->print_inlining_update(this); 169 170 return NULL; 171 } 172 173 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 174 LibraryCallKit kit(jvms, this); 175 Compile* C = kit.C; 176 int nodes = C->unique(); 177 _last_predicate = predicate; 178 #ifndef PRODUCT 179 assert(is_predicated() && predicate < predicates_count(), "sanity"); 180 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 181 char buf[1000]; 182 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 183 tty->print_cr("Predicate for intrinsic %s", str); 184 } 185 #endif 186 ciMethod* callee = kit.callee(); 187 const int bci = kit.bci(); 188 189 Node* slow_ctl = kit.try_to_predicate(predicate); 190 if (!kit.failing()) { 191 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" 192 : "(intrinsic, predicate)"; 193 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 194 if (C->print_intrinsics() || C->print_inlining()) { 195 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 196 } 197 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 198 if (C->log()) { 199 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 200 vmIntrinsics::name_at(intrinsic_id()), 201 (is_virtual() ? " virtual='1'" : ""), 202 C->unique() - nodes); 203 } 204 return slow_ctl; // Could be NULL if the check folds. 205 } 206 207 // The intrinsic bailed out 208 if (jvms->has_method()) { 209 // Not a root compile. 210 const char* msg = "failed to generate predicate for intrinsic"; 211 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); 212 if (C->print_intrinsics() || C->print_inlining()) { 213 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 214 } 215 } else { 216 // Root compile 217 ResourceMark rm; 218 stringStream msg_stream; 219 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 220 vmIntrinsics::name_at(intrinsic_id()), 221 is_virtual() ? " (virtual)" : "", bci); 222 const char *msg = msg_stream.as_string(); 223 log_debug(jit, inlining)("%s", msg); 224 if (C->print_intrinsics() || C->print_inlining()) { 225 C->print_inlining_stream()->print("%s", msg); 226 } 227 } 228 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 229 return NULL; 230 } 231 232 bool LibraryCallKit::try_to_inline(int predicate) { 233 // Handle symbolic names for otherwise undistinguished boolean switches: 234 const bool is_store = true; 235 const bool is_compress = true; 236 const bool is_static = true; 237 const bool is_volatile = true; 238 239 if (!jvms()->has_method()) { 240 // Root JVMState has a null method. 241 assert(map()->memory()->Opcode() == Op_Parm, ""); 242 // Insert the memory aliasing node 243 set_all_memory(reset_memory()); 244 } 245 assert(merged_memory(), ""); 246 247 switch (intrinsic_id()) { 248 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 249 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 250 case vmIntrinsics::_getClass: return inline_native_getClass(); 251 252 case vmIntrinsics::_ceil: 253 case vmIntrinsics::_floor: 254 case vmIntrinsics::_rint: 255 case vmIntrinsics::_dsin: 256 case vmIntrinsics::_dcos: 257 case vmIntrinsics::_dtan: 258 case vmIntrinsics::_dabs: 259 case vmIntrinsics::_fabs: 260 case vmIntrinsics::_iabs: 261 case vmIntrinsics::_labs: 262 case vmIntrinsics::_datan2: 263 case vmIntrinsics::_dsqrt: 264 case vmIntrinsics::_dexp: 265 case vmIntrinsics::_dlog: 266 case vmIntrinsics::_dlog10: 267 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 268 269 case vmIntrinsics::_min: 270 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 271 272 case vmIntrinsics::_notify: 273 case vmIntrinsics::_notifyAll: 274 return inline_notify(intrinsic_id()); 275 276 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 277 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 278 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 279 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 280 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 281 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 282 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 283 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 284 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); 285 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 286 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 287 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 288 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 289 290 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 291 292 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 293 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 294 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 295 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 296 297 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 298 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 299 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 300 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 301 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 302 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 303 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 304 305 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 306 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 307 308 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 309 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 310 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 311 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 312 313 case vmIntrinsics::_compressStringC: 314 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 315 case vmIntrinsics::_inflateStringC: 316 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 317 318 case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 319 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 320 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 321 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 322 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 323 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 324 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 325 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 326 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 327 328 case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 329 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 330 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 331 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 332 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 333 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 334 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 335 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 336 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 337 338 case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 339 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 340 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 341 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 342 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 343 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 344 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 345 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 346 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 347 348 case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 349 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 350 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 351 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 352 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 353 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 354 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 355 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 356 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 357 358 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 359 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 360 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 361 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 362 363 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 364 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 365 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 366 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 367 368 case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 369 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 370 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 371 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 372 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 373 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 374 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 375 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 376 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 377 378 case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 379 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 380 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 381 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 382 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 383 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 384 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 385 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 386 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 387 388 case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 389 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 390 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 391 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 392 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 393 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 394 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 395 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 396 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 397 398 case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 399 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 400 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 401 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 402 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 403 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 404 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 405 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 406 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 407 408 case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 409 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 410 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 411 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 412 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 413 414 case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 415 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 416 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 417 case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 418 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 419 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 420 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 421 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 422 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 423 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 424 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 425 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 426 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 427 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 428 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 429 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 430 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 431 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 432 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 433 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 434 435 case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 436 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 437 case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 438 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 439 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 440 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 441 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 442 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 443 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 444 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 445 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 446 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 447 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 448 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 449 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 450 451 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 452 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 453 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 454 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 455 456 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 457 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 458 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 459 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 460 case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 461 462 case vmIntrinsics::_loadFence: 463 case vmIntrinsics::_storeFence: 464 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 465 466 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 467 468 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 469 470 #ifdef JFR_HAVE_INTRINSICS 471 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime"); 472 case vmIntrinsics::_getClassId: return inline_native_classID(); 473 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); 474 #endif 475 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 476 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 477 case vmIntrinsics::_writeback0: return inline_unsafe_writeback0(); 478 case vmIntrinsics::_writebackPreSync0: return inline_unsafe_writebackSync0(true); 479 case vmIntrinsics::_writebackPostSync0: return inline_unsafe_writebackSync0(false); 480 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 481 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 482 case vmIntrinsics::_getLength: return inline_native_getLength(); 483 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 484 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 485 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 486 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 487 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 488 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 489 490 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 491 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 492 493 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 494 495 case vmIntrinsics::_isInstance: 496 case vmIntrinsics::_getModifiers: 497 case vmIntrinsics::_isInterface: 498 case vmIntrinsics::_isArray: 499 case vmIntrinsics::_isPrimitive: 500 case vmIntrinsics::_isHidden: 501 case vmIntrinsics::_getSuperclass: 502 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 503 504 case vmIntrinsics::_floatToRawIntBits: 505 case vmIntrinsics::_floatToIntBits: 506 case vmIntrinsics::_intBitsToFloat: 507 case vmIntrinsics::_doubleToRawLongBits: 508 case vmIntrinsics::_doubleToLongBits: 509 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 510 511 case vmIntrinsics::_numberOfLeadingZeros_i: 512 case vmIntrinsics::_numberOfLeadingZeros_l: 513 case vmIntrinsics::_numberOfTrailingZeros_i: 514 case vmIntrinsics::_numberOfTrailingZeros_l: 515 case vmIntrinsics::_bitCount_i: 516 case vmIntrinsics::_bitCount_l: 517 case vmIntrinsics::_reverseBytes_i: 518 case vmIntrinsics::_reverseBytes_l: 519 case vmIntrinsics::_reverseBytes_s: 520 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 521 522 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 523 524 case vmIntrinsics::_Reference_get: return inline_reference_get(); 525 526 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 527 528 case vmIntrinsics::_aescrypt_encryptBlock: 529 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 530 531 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 532 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 533 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 534 535 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt: 536 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt: 537 return inline_electronicCodeBook_AESCrypt(intrinsic_id()); 538 539 case vmIntrinsics::_counterMode_AESCrypt: 540 return inline_counterMode_AESCrypt(intrinsic_id()); 541 542 case vmIntrinsics::_md5_implCompress: 543 case vmIntrinsics::_sha_implCompress: 544 case vmIntrinsics::_sha2_implCompress: 545 case vmIntrinsics::_sha5_implCompress: 546 return inline_digestBase_implCompress(intrinsic_id()); 547 548 case vmIntrinsics::_digestBase_implCompressMB: 549 return inline_digestBase_implCompressMB(predicate); 550 551 case vmIntrinsics::_multiplyToLen: 552 return inline_multiplyToLen(); 553 554 case vmIntrinsics::_squareToLen: 555 return inline_squareToLen(); 556 557 case vmIntrinsics::_mulAdd: 558 return inline_mulAdd(); 559 560 case vmIntrinsics::_montgomeryMultiply: 561 return inline_montgomeryMultiply(); 562 case vmIntrinsics::_montgomerySquare: 563 return inline_montgomerySquare(); 564 565 case vmIntrinsics::_bigIntegerRightShiftWorker: 566 return inline_bigIntegerShift(true); 567 case vmIntrinsics::_bigIntegerLeftShiftWorker: 568 return inline_bigIntegerShift(false); 569 570 case vmIntrinsics::_vectorizedMismatch: 571 return inline_vectorizedMismatch(); 572 573 case vmIntrinsics::_ghash_processBlocks: 574 return inline_ghash_processBlocks(); 575 case vmIntrinsics::_base64_encodeBlock: 576 return inline_base64_encodeBlock(); 577 578 case vmIntrinsics::_encodeISOArray: 579 case vmIntrinsics::_encodeByteISOArray: 580 return inline_encodeISOArray(); 581 582 case vmIntrinsics::_updateCRC32: 583 return inline_updateCRC32(); 584 case vmIntrinsics::_updateBytesCRC32: 585 return inline_updateBytesCRC32(); 586 case vmIntrinsics::_updateByteBufferCRC32: 587 return inline_updateByteBufferCRC32(); 588 589 case vmIntrinsics::_updateBytesCRC32C: 590 return inline_updateBytesCRC32C(); 591 case vmIntrinsics::_updateDirectByteBufferCRC32C: 592 return inline_updateDirectByteBufferCRC32C(); 593 594 case vmIntrinsics::_updateBytesAdler32: 595 return inline_updateBytesAdler32(); 596 case vmIntrinsics::_updateByteBufferAdler32: 597 return inline_updateByteBufferAdler32(); 598 599 case vmIntrinsics::_profileBoolean: 600 return inline_profileBoolean(); 601 case vmIntrinsics::_isCompileConstant: 602 return inline_isCompileConstant(); 603 604 case vmIntrinsics::_hasNegatives: 605 return inline_hasNegatives(); 606 607 case vmIntrinsics::_fmaD: 608 case vmIntrinsics::_fmaF: 609 return inline_fma(intrinsic_id()); 610 611 case vmIntrinsics::_isDigit: 612 case vmIntrinsics::_isLowerCase: 613 case vmIntrinsics::_isUpperCase: 614 case vmIntrinsics::_isWhitespace: 615 return inline_character_compare(intrinsic_id()); 616 617 case vmIntrinsics::_maxF: 618 case vmIntrinsics::_minF: 619 case vmIntrinsics::_maxD: 620 case vmIntrinsics::_minD: 621 return inline_fp_min_max(intrinsic_id()); 622 623 case vmIntrinsics::_VectorUnaryOp: 624 return inline_vector_nary_operation(1); 625 case vmIntrinsics::_VectorBinaryOp: 626 return inline_vector_nary_operation(2); 627 case vmIntrinsics::_VectorTernaryOp: 628 return inline_vector_nary_operation(3); 629 case vmIntrinsics::_VectorBroadcastCoerced: 630 return inline_vector_broadcast_coerced(); 631 case vmIntrinsics::_VectorShuffleIota: 632 return inline_vector_shuffle_iota(); 633 case vmIntrinsics::_VectorShuffleToVector: 634 return inline_vector_shuffle_to_vector(); 635 case vmIntrinsics::_VectorLoadOp: 636 return inline_vector_mem_operation(/*is_store=*/false); 637 case vmIntrinsics::_VectorStoreOp: 638 return inline_vector_mem_operation(/*is_store=*/true); 639 case vmIntrinsics::_VectorGatherOp: 640 return inline_vector_gather_scatter(/*is_scatter*/ false); 641 case vmIntrinsics::_VectorScatterOp: 642 return inline_vector_gather_scatter(/*is_scatter*/ true); 643 case vmIntrinsics::_VectorReductionCoerced: 644 return inline_vector_reduction(); 645 case vmIntrinsics::_VectorTest: 646 return inline_vector_test(); 647 case vmIntrinsics::_VectorBlend: 648 return inline_vector_blend(); 649 case vmIntrinsics::_VectorRearrange: 650 return inline_vector_rearrange(); 651 case vmIntrinsics::_VectorCompare: 652 return inline_vector_compare(); 653 case vmIntrinsics::_VectorBroadcastInt: 654 return inline_vector_broadcast_int(); 655 case vmIntrinsics::_VectorConvert: 656 return inline_vector_convert(); 657 case vmIntrinsics::_VectorInsert: 658 return inline_vector_insert(); 659 case vmIntrinsics::_VectorExtract: 660 return inline_vector_extract(); 661 662 default: 663 // If you get here, it may be that someone has added a new intrinsic 664 // to the list in vmSymbols.hpp without implementing it here. 665 #ifndef PRODUCT 666 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 667 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 668 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 669 } 670 #endif 671 return false; 672 } 673 } 674 675 Node* LibraryCallKit::try_to_predicate(int predicate) { 676 if (!jvms()->has_method()) { 677 // Root JVMState has a null method. 678 assert(map()->memory()->Opcode() == Op_Parm, ""); 679 // Insert the memory aliasing node 680 set_all_memory(reset_memory()); 681 } 682 assert(merged_memory(), ""); 683 684 switch (intrinsic_id()) { 685 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 686 return inline_cipherBlockChaining_AESCrypt_predicate(false); 687 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 688 return inline_cipherBlockChaining_AESCrypt_predicate(true); 689 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt: 690 return inline_electronicCodeBook_AESCrypt_predicate(false); 691 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt: 692 return inline_electronicCodeBook_AESCrypt_predicate(true); 693 case vmIntrinsics::_counterMode_AESCrypt: 694 return inline_counterMode_AESCrypt_predicate(); 695 case vmIntrinsics::_digestBase_implCompressMB: 696 return inline_digestBase_implCompressMB_predicate(predicate); 697 698 default: 699 // If you get here, it may be that someone has added a new intrinsic 700 // to the list in vmSymbols.hpp without implementing it here. 701 #ifndef PRODUCT 702 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 703 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 704 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 705 } 706 #endif 707 Node* slow_ctl = control(); 708 set_control(top()); // No fast path instrinsic 709 return slow_ctl; 710 } 711 } 712 713 //------------------------------set_result------------------------------- 714 // Helper function for finishing intrinsics. 715 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 716 record_for_igvn(region); 717 set_control(_gvn.transform(region)); 718 set_result( _gvn.transform(value)); 719 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 720 } 721 722 //------------------------------generate_guard--------------------------- 723 // Helper function for generating guarded fast-slow graph structures. 724 // The given 'test', if true, guards a slow path. If the test fails 725 // then a fast path can be taken. (We generally hope it fails.) 726 // In all cases, GraphKit::control() is updated to the fast path. 727 // The returned value represents the control for the slow path. 728 // The return value is never 'top'; it is either a valid control 729 // or NULL if it is obvious that the slow path can never be taken. 730 // Also, if region and the slow control are not NULL, the slow edge 731 // is appended to the region. 732 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 733 if (stopped()) { 734 // Already short circuited. 735 return NULL; 736 } 737 738 // Build an if node and its projections. 739 // If test is true we take the slow path, which we assume is uncommon. 740 if (_gvn.type(test) == TypeInt::ZERO) { 741 // The slow branch is never taken. No need to build this guard. 742 return NULL; 743 } 744 745 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 746 747 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 748 if (if_slow == top()) { 749 // The slow branch is never taken. No need to build this guard. 750 return NULL; 751 } 752 753 if (region != NULL) 754 region->add_req(if_slow); 755 756 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 757 set_control(if_fast); 758 759 return if_slow; 760 } 761 762 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 763 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 764 } 765 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 766 return generate_guard(test, region, PROB_FAIR); 767 } 768 769 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 770 Node* *pos_index) { 771 if (stopped()) 772 return NULL; // already stopped 773 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 774 return NULL; // index is already adequately typed 775 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 776 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 777 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 778 if (is_neg != NULL && pos_index != NULL) { 779 // Emulate effect of Parse::adjust_map_after_if. 780 Node* ccast = new CastIINode(index, TypeInt::POS); 781 ccast->set_req(0, control()); 782 (*pos_index) = _gvn.transform(ccast); 783 } 784 return is_neg; 785 } 786 787 // Make sure that 'position' is a valid limit index, in [0..length]. 788 // There are two equivalent plans for checking this: 789 // A. (offset + copyLength) unsigned<= arrayLength 790 // B. offset <= (arrayLength - copyLength) 791 // We require that all of the values above, except for the sum and 792 // difference, are already known to be non-negative. 793 // Plan A is robust in the face of overflow, if offset and copyLength 794 // are both hugely positive. 795 // 796 // Plan B is less direct and intuitive, but it does not overflow at 797 // all, since the difference of two non-negatives is always 798 // representable. Whenever Java methods must perform the equivalent 799 // check they generally use Plan B instead of Plan A. 800 // For the moment we use Plan A. 801 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 802 Node* subseq_length, 803 Node* array_length, 804 RegionNode* region) { 805 if (stopped()) 806 return NULL; // already stopped 807 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 808 if (zero_offset && subseq_length->eqv_uncast(array_length)) 809 return NULL; // common case of whole-array copy 810 Node* last = subseq_length; 811 if (!zero_offset) // last += offset 812 last = _gvn.transform(new AddINode(last, offset)); 813 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 814 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 815 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 816 return is_over; 817 } 818 819 // Emit range checks for the given String.value byte array 820 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 821 if (stopped()) { 822 return; // already stopped 823 } 824 RegionNode* bailout = new RegionNode(1); 825 record_for_igvn(bailout); 826 if (char_count) { 827 // Convert char count to byte count 828 count = _gvn.transform(new LShiftINode(count, intcon(1))); 829 } 830 831 // Offset and count must not be negative 832 generate_negative_guard(offset, bailout); 833 generate_negative_guard(count, bailout); 834 // Offset + count must not exceed length of array 835 generate_limit_guard(offset, count, load_array_length(array), bailout); 836 837 if (bailout->req() > 1) { 838 PreserveJVMState pjvms(this); 839 set_control(_gvn.transform(bailout)); 840 uncommon_trap(Deoptimization::Reason_intrinsic, 841 Deoptimization::Action_maybe_recompile); 842 } 843 } 844 845 //--------------------------generate_current_thread-------------------- 846 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 847 ciKlass* thread_klass = env()->Thread_klass(); 848 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 849 Node* thread = _gvn.transform(new ThreadLocalNode()); 850 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 851 Node* threadObj = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), p, p->bottom_type()->is_ptr(), thread_type, T_OBJECT, MemNode::unordered)); 852 tls_output = thread; 853 return threadObj; 854 } 855 856 857 //------------------------------make_string_method_node------------------------ 858 // Helper method for String intrinsic functions. This version is called with 859 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 860 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 861 // containing the lengths of str1 and str2. 862 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 863 Node* result = NULL; 864 switch (opcode) { 865 case Op_StrIndexOf: 866 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 867 str1_start, cnt1, str2_start, cnt2, ae); 868 break; 869 case Op_StrComp: 870 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 871 str1_start, cnt1, str2_start, cnt2, ae); 872 break; 873 case Op_StrEquals: 874 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 875 // Use the constant length if there is one because optimized match rule may exist. 876 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 877 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 878 break; 879 default: 880 ShouldNotReachHere(); 881 return NULL; 882 } 883 884 // All these intrinsics have checks. 885 C->set_has_split_ifs(true); // Has chance for split-if optimization 886 clear_upper_avx(); 887 888 return _gvn.transform(result); 889 } 890 891 //------------------------------inline_string_compareTo------------------------ 892 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 893 Node* arg1 = argument(0); 894 Node* arg2 = argument(1); 895 896 arg1 = must_be_not_null(arg1, true); 897 arg2 = must_be_not_null(arg2, true); 898 899 // Get start addr and length of first argument 900 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 901 Node* arg1_cnt = load_array_length(arg1); 902 903 // Get start addr and length of second argument 904 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 905 Node* arg2_cnt = load_array_length(arg2); 906 907 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 908 set_result(result); 909 return true; 910 } 911 912 //------------------------------inline_string_equals------------------------ 913 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 914 Node* arg1 = argument(0); 915 Node* arg2 = argument(1); 916 917 // paths (plus control) merge 918 RegionNode* region = new RegionNode(3); 919 Node* phi = new PhiNode(region, TypeInt::BOOL); 920 921 if (!stopped()) { 922 923 arg1 = must_be_not_null(arg1, true); 924 arg2 = must_be_not_null(arg2, true); 925 926 // Get start addr and length of first argument 927 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 928 Node* arg1_cnt = load_array_length(arg1); 929 930 // Get start addr and length of second argument 931 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 932 Node* arg2_cnt = load_array_length(arg2); 933 934 // Check for arg1_cnt != arg2_cnt 935 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 936 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 937 Node* if_ne = generate_slow_guard(bol, NULL); 938 if (if_ne != NULL) { 939 phi->init_req(2, intcon(0)); 940 region->init_req(2, if_ne); 941 } 942 943 // Check for count == 0 is done by assembler code for StrEquals. 944 945 if (!stopped()) { 946 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 947 phi->init_req(1, equals); 948 region->init_req(1, control()); 949 } 950 } 951 952 // post merge 953 set_control(_gvn.transform(region)); 954 record_for_igvn(region); 955 956 set_result(_gvn.transform(phi)); 957 return true; 958 } 959 960 //------------------------------inline_array_equals---------------------------- 961 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 962 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 963 Node* arg1 = argument(0); 964 Node* arg2 = argument(1); 965 966 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 967 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 968 clear_upper_avx(); 969 970 return true; 971 } 972 973 //------------------------------inline_hasNegatives------------------------------ 974 bool LibraryCallKit::inline_hasNegatives() { 975 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 976 return false; 977 } 978 979 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 980 // no receiver since it is static method 981 Node* ba = argument(0); 982 Node* offset = argument(1); 983 Node* len = argument(2); 984 985 ba = must_be_not_null(ba, true); 986 987 // Range checks 988 generate_string_range_check(ba, offset, len, false); 989 if (stopped()) { 990 return true; 991 } 992 Node* ba_start = array_element_address(ba, offset, T_BYTE); 993 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 994 set_result(_gvn.transform(result)); 995 return true; 996 } 997 998 bool LibraryCallKit::inline_preconditions_checkIndex() { 999 Node* index = argument(0); 1000 Node* length = argument(1); 1001 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1002 return false; 1003 } 1004 1005 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1006 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1007 1008 { 1009 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1010 uncommon_trap(Deoptimization::Reason_intrinsic, 1011 Deoptimization::Action_make_not_entrant); 1012 } 1013 1014 if (stopped()) { 1015 return false; 1016 } 1017 1018 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1019 BoolTest::mask btest = BoolTest::lt; 1020 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1021 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1022 _gvn.set_type(rc, rc->Value(&_gvn)); 1023 if (!rc_bool->is_Con()) { 1024 record_for_igvn(rc); 1025 } 1026 set_control(_gvn.transform(new IfTrueNode(rc))); 1027 { 1028 PreserveJVMState pjvms(this); 1029 set_control(_gvn.transform(new IfFalseNode(rc))); 1030 uncommon_trap(Deoptimization::Reason_range_check, 1031 Deoptimization::Action_make_not_entrant); 1032 } 1033 1034 if (stopped()) { 1035 return false; 1036 } 1037 1038 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1039 result->set_req(0, control()); 1040 result = _gvn.transform(result); 1041 set_result(result); 1042 replace_in_map(index, result); 1043 clear_upper_avx(); 1044 return true; 1045 } 1046 1047 //------------------------------inline_string_indexOf------------------------ 1048 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1049 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1050 return false; 1051 } 1052 Node* src = argument(0); 1053 Node* tgt = argument(1); 1054 1055 // Make the merge point 1056 RegionNode* result_rgn = new RegionNode(4); 1057 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1058 1059 src = must_be_not_null(src, true); 1060 tgt = must_be_not_null(tgt, true); 1061 1062 // Get start addr and length of source string 1063 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1064 Node* src_count = load_array_length(src); 1065 1066 // Get start addr and length of substring 1067 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1068 Node* tgt_count = load_array_length(tgt); 1069 1070 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1071 // Divide src size by 2 if String is UTF16 encoded 1072 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1073 } 1074 if (ae == StrIntrinsicNode::UU) { 1075 // Divide substring size by 2 if String is UTF16 encoded 1076 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1077 } 1078 1079 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1080 if (result != NULL) { 1081 result_phi->init_req(3, result); 1082 result_rgn->init_req(3, control()); 1083 } 1084 set_control(_gvn.transform(result_rgn)); 1085 record_for_igvn(result_rgn); 1086 set_result(_gvn.transform(result_phi)); 1087 1088 return true; 1089 } 1090 1091 //-----------------------------inline_string_indexOf----------------------- 1092 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1093 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1094 return false; 1095 } 1096 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1097 return false; 1098 } 1099 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1100 Node* src = argument(0); // byte[] 1101 Node* src_count = argument(1); // char count 1102 Node* tgt = argument(2); // byte[] 1103 Node* tgt_count = argument(3); // char count 1104 Node* from_index = argument(4); // char index 1105 1106 src = must_be_not_null(src, true); 1107 tgt = must_be_not_null(tgt, true); 1108 1109 // Multiply byte array index by 2 if String is UTF16 encoded 1110 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1111 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1112 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1113 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1114 1115 // Range checks 1116 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1117 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1118 if (stopped()) { 1119 return true; 1120 } 1121 1122 RegionNode* region = new RegionNode(5); 1123 Node* phi = new PhiNode(region, TypeInt::INT); 1124 1125 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1126 if (result != NULL) { 1127 // The result is index relative to from_index if substring was found, -1 otherwise. 1128 // Generate code which will fold into cmove. 1129 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1130 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1131 1132 Node* if_lt = generate_slow_guard(bol, NULL); 1133 if (if_lt != NULL) { 1134 // result == -1 1135 phi->init_req(3, result); 1136 region->init_req(3, if_lt); 1137 } 1138 if (!stopped()) { 1139 result = _gvn.transform(new AddINode(result, from_index)); 1140 phi->init_req(4, result); 1141 region->init_req(4, control()); 1142 } 1143 } 1144 1145 set_control(_gvn.transform(region)); 1146 record_for_igvn(region); 1147 set_result(_gvn.transform(phi)); 1148 clear_upper_avx(); 1149 1150 return true; 1151 } 1152 1153 // Create StrIndexOfNode with fast path checks 1154 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1155 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1156 // Check for substr count > string count 1157 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1158 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1159 Node* if_gt = generate_slow_guard(bol, NULL); 1160 if (if_gt != NULL) { 1161 phi->init_req(1, intcon(-1)); 1162 region->init_req(1, if_gt); 1163 } 1164 if (!stopped()) { 1165 // Check for substr count == 0 1166 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1167 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1168 Node* if_zero = generate_slow_guard(bol, NULL); 1169 if (if_zero != NULL) { 1170 phi->init_req(2, intcon(0)); 1171 region->init_req(2, if_zero); 1172 } 1173 } 1174 if (!stopped()) { 1175 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1176 } 1177 return NULL; 1178 } 1179 1180 //-----------------------------inline_string_indexOfChar----------------------- 1181 bool LibraryCallKit::inline_string_indexOfChar() { 1182 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1183 return false; 1184 } 1185 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1186 return false; 1187 } 1188 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1189 Node* src = argument(0); // byte[] 1190 Node* tgt = argument(1); // tgt is int ch 1191 Node* from_index = argument(2); 1192 Node* max = argument(3); 1193 1194 src = must_be_not_null(src, true); 1195 1196 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1197 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1198 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1199 1200 // Range checks 1201 generate_string_range_check(src, src_offset, src_count, true); 1202 if (stopped()) { 1203 return true; 1204 } 1205 1206 RegionNode* region = new RegionNode(3); 1207 Node* phi = new PhiNode(region, TypeInt::INT); 1208 1209 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1210 C->set_has_split_ifs(true); // Has chance for split-if optimization 1211 _gvn.transform(result); 1212 1213 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1214 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1215 1216 Node* if_lt = generate_slow_guard(bol, NULL); 1217 if (if_lt != NULL) { 1218 // result == -1 1219 phi->init_req(2, result); 1220 region->init_req(2, if_lt); 1221 } 1222 if (!stopped()) { 1223 result = _gvn.transform(new AddINode(result, from_index)); 1224 phi->init_req(1, result); 1225 region->init_req(1, control()); 1226 } 1227 set_control(_gvn.transform(region)); 1228 record_for_igvn(region); 1229 set_result(_gvn.transform(phi)); 1230 1231 return true; 1232 } 1233 //---------------------------inline_string_copy--------------------- 1234 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1235 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1236 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1237 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1238 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1239 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1240 bool LibraryCallKit::inline_string_copy(bool compress) { 1241 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1242 return false; 1243 } 1244 int nargs = 5; // 2 oops, 3 ints 1245 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1246 1247 Node* src = argument(0); 1248 Node* src_offset = argument(1); 1249 Node* dst = argument(2); 1250 Node* dst_offset = argument(3); 1251 Node* length = argument(4); 1252 1253 // Check for allocation before we add nodes that would confuse 1254 // tightly_coupled_allocation() 1255 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1256 1257 // Figure out the size and type of the elements we will be copying. 1258 const Type* src_type = src->Value(&_gvn); 1259 const Type* dst_type = dst->Value(&_gvn); 1260 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1261 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1262 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1263 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1264 "Unsupported array types for inline_string_copy"); 1265 1266 src = must_be_not_null(src, true); 1267 dst = must_be_not_null(dst, true); 1268 1269 // Convert char[] offsets to byte[] offsets 1270 bool convert_src = (compress && src_elem == T_BYTE); 1271 bool convert_dst = (!compress && dst_elem == T_BYTE); 1272 if (convert_src) { 1273 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1274 } else if (convert_dst) { 1275 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1276 } 1277 1278 // Range checks 1279 generate_string_range_check(src, src_offset, length, convert_src); 1280 generate_string_range_check(dst, dst_offset, length, convert_dst); 1281 if (stopped()) { 1282 return true; 1283 } 1284 1285 Node* src_start = array_element_address(src, src_offset, src_elem); 1286 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1287 // 'src_start' points to src array + scaled offset 1288 // 'dst_start' points to dst array + scaled offset 1289 Node* count = NULL; 1290 if (compress) { 1291 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1292 } else { 1293 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1294 } 1295 1296 if (alloc != NULL) { 1297 if (alloc->maybe_set_complete(&_gvn)) { 1298 // "You break it, you buy it." 1299 InitializeNode* init = alloc->initialization(); 1300 assert(init->is_complete(), "we just did this"); 1301 init->set_complete_with_arraycopy(); 1302 assert(dst->is_CheckCastPP(), "sanity"); 1303 assert(dst->in(0)->in(0) == init, "dest pinned"); 1304 } 1305 // Do not let stores that initialize this object be reordered with 1306 // a subsequent store that would make this object accessible by 1307 // other threads. 1308 // Record what AllocateNode this StoreStore protects so that 1309 // escape analysis can go from the MemBarStoreStoreNode to the 1310 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1311 // based on the escape status of the AllocateNode. 1312 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1313 } 1314 if (compress) { 1315 set_result(_gvn.transform(count)); 1316 } 1317 clear_upper_avx(); 1318 1319 return true; 1320 } 1321 1322 #ifdef _LP64 1323 #define XTOP ,top() /*additional argument*/ 1324 #else //_LP64 1325 #define XTOP /*no additional argument*/ 1326 #endif //_LP64 1327 1328 //------------------------inline_string_toBytesU-------------------------- 1329 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1330 bool LibraryCallKit::inline_string_toBytesU() { 1331 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1332 return false; 1333 } 1334 // Get the arguments. 1335 Node* value = argument(0); 1336 Node* offset = argument(1); 1337 Node* length = argument(2); 1338 1339 Node* newcopy = NULL; 1340 1341 // Set the original stack and the reexecute bit for the interpreter to reexecute 1342 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1343 { PreserveReexecuteState preexecs(this); 1344 jvms()->set_should_reexecute(true); 1345 1346 // Check if a null path was taken unconditionally. 1347 value = null_check(value); 1348 1349 RegionNode* bailout = new RegionNode(1); 1350 record_for_igvn(bailout); 1351 1352 // Range checks 1353 generate_negative_guard(offset, bailout); 1354 generate_negative_guard(length, bailout); 1355 generate_limit_guard(offset, length, load_array_length(value), bailout); 1356 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1357 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1358 1359 if (bailout->req() > 1) { 1360 PreserveJVMState pjvms(this); 1361 set_control(_gvn.transform(bailout)); 1362 uncommon_trap(Deoptimization::Reason_intrinsic, 1363 Deoptimization::Action_maybe_recompile); 1364 } 1365 if (stopped()) { 1366 return true; 1367 } 1368 1369 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1370 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1371 newcopy = new_array(klass_node, size, 0); // no arguments to push 1372 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1373 1374 // Calculate starting addresses. 1375 Node* src_start = array_element_address(value, offset, T_CHAR); 1376 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1377 1378 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1379 const TypeInt* toffset = gvn().type(offset)->is_int(); 1380 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1381 1382 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1383 const char* copyfunc_name = "arraycopy"; 1384 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1385 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1386 OptoRuntime::fast_arraycopy_Type(), 1387 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1388 src_start, dst_start, ConvI2X(length) XTOP); 1389 // Do not let reads from the cloned object float above the arraycopy. 1390 if (alloc != NULL) { 1391 if (alloc->maybe_set_complete(&_gvn)) { 1392 // "You break it, you buy it." 1393 InitializeNode* init = alloc->initialization(); 1394 assert(init->is_complete(), "we just did this"); 1395 init->set_complete_with_arraycopy(); 1396 assert(newcopy->is_CheckCastPP(), "sanity"); 1397 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1398 } 1399 // Do not let stores that initialize this object be reordered with 1400 // a subsequent store that would make this object accessible by 1401 // other threads. 1402 // Record what AllocateNode this StoreStore protects so that 1403 // escape analysis can go from the MemBarStoreStoreNode to the 1404 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1405 // based on the escape status of the AllocateNode. 1406 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1407 } else { 1408 insert_mem_bar(Op_MemBarCPUOrder); 1409 } 1410 } // original reexecute is set back here 1411 1412 C->set_has_split_ifs(true); // Has chance for split-if optimization 1413 if (!stopped()) { 1414 set_result(newcopy); 1415 } 1416 clear_upper_avx(); 1417 1418 return true; 1419 } 1420 1421 //------------------------inline_string_getCharsU-------------------------- 1422 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1423 bool LibraryCallKit::inline_string_getCharsU() { 1424 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1425 return false; 1426 } 1427 1428 // Get the arguments. 1429 Node* src = argument(0); 1430 Node* src_begin = argument(1); 1431 Node* src_end = argument(2); // exclusive offset (i < src_end) 1432 Node* dst = argument(3); 1433 Node* dst_begin = argument(4); 1434 1435 // Check for allocation before we add nodes that would confuse 1436 // tightly_coupled_allocation() 1437 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1438 1439 // Check if a null path was taken unconditionally. 1440 src = null_check(src); 1441 dst = null_check(dst); 1442 if (stopped()) { 1443 return true; 1444 } 1445 1446 // Get length and convert char[] offset to byte[] offset 1447 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1448 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1449 1450 // Range checks 1451 generate_string_range_check(src, src_begin, length, true); 1452 generate_string_range_check(dst, dst_begin, length, false); 1453 if (stopped()) { 1454 return true; 1455 } 1456 1457 if (!stopped()) { 1458 // Calculate starting addresses. 1459 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1460 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1461 1462 // Check if array addresses are aligned to HeapWordSize 1463 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1464 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1465 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1466 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1467 1468 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1469 const char* copyfunc_name = "arraycopy"; 1470 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1471 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1472 OptoRuntime::fast_arraycopy_Type(), 1473 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1474 src_start, dst_start, ConvI2X(length) XTOP); 1475 // Do not let reads from the cloned object float above the arraycopy. 1476 if (alloc != NULL) { 1477 if (alloc->maybe_set_complete(&_gvn)) { 1478 // "You break it, you buy it." 1479 InitializeNode* init = alloc->initialization(); 1480 assert(init->is_complete(), "we just did this"); 1481 init->set_complete_with_arraycopy(); 1482 assert(dst->is_CheckCastPP(), "sanity"); 1483 assert(dst->in(0)->in(0) == init, "dest pinned"); 1484 } 1485 // Do not let stores that initialize this object be reordered with 1486 // a subsequent store that would make this object accessible by 1487 // other threads. 1488 // Record what AllocateNode this StoreStore protects so that 1489 // escape analysis can go from the MemBarStoreStoreNode to the 1490 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1491 // based on the escape status of the AllocateNode. 1492 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1493 } else { 1494 insert_mem_bar(Op_MemBarCPUOrder); 1495 } 1496 } 1497 1498 C->set_has_split_ifs(true); // Has chance for split-if optimization 1499 return true; 1500 } 1501 1502 //----------------------inline_string_char_access---------------------------- 1503 // Store/Load char to/from byte[] array. 1504 // static void StringUTF16.putChar(byte[] val, int index, int c) 1505 // static char StringUTF16.getChar(byte[] val, int index) 1506 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1507 Node* value = argument(0); 1508 Node* index = argument(1); 1509 Node* ch = is_store ? argument(2) : NULL; 1510 1511 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1512 // correctly requires matched array shapes. 1513 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1514 "sanity: byte[] and char[] bases agree"); 1515 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1516 "sanity: byte[] and char[] scales agree"); 1517 1518 // Bail when getChar over constants is requested: constant folding would 1519 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1520 // Java method would constant fold nicely instead. 1521 if (!is_store && value->is_Con() && index->is_Con()) { 1522 return false; 1523 } 1524 1525 value = must_be_not_null(value, true); 1526 1527 Node* adr = array_element_address(value, index, T_CHAR); 1528 if (adr->is_top()) { 1529 return false; 1530 } 1531 if (is_store) { 1532 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED); 1533 } else { 1534 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD); 1535 set_result(ch); 1536 } 1537 return true; 1538 } 1539 1540 //--------------------------round_double_node-------------------------------- 1541 // Round a double node if necessary. 1542 Node* LibraryCallKit::round_double_node(Node* n) { 1543 if (Matcher::strict_fp_requires_explicit_rounding) { 1544 #ifdef IA32 1545 if (UseSSE < 2) { 1546 n = _gvn.transform(new RoundDoubleNode(NULL, n)); 1547 } 1548 #else 1549 Unimplemented(); 1550 #endif // IA32 1551 } 1552 return n; 1553 } 1554 1555 //------------------------------inline_math----------------------------------- 1556 // public static double Math.abs(double) 1557 // public static double Math.sqrt(double) 1558 // public static double Math.log(double) 1559 // public static double Math.log10(double) 1560 bool LibraryCallKit::inline_double_math(vmIntrinsics::ID id) { 1561 Node* arg = round_double_node(argument(0)); 1562 Node* n = NULL; 1563 switch (id) { 1564 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1565 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1566 case vmIntrinsics::_ceil: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_ceil); break; 1567 case vmIntrinsics::_floor: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_floor); break; 1568 case vmIntrinsics::_rint: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_rint); break; 1569 default: fatal_unexpected_iid(id); break; 1570 } 1571 set_result(_gvn.transform(n)); 1572 return true; 1573 } 1574 1575 //------------------------------inline_math----------------------------------- 1576 // public static float Math.abs(float) 1577 // public static int Math.abs(int) 1578 // public static long Math.abs(long) 1579 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1580 Node* arg = argument(0); 1581 Node* n = NULL; 1582 switch (id) { 1583 case vmIntrinsics::_fabs: n = new AbsFNode( arg); break; 1584 case vmIntrinsics::_iabs: n = new AbsINode( arg); break; 1585 case vmIntrinsics::_labs: n = new AbsLNode( arg); break; 1586 default: fatal_unexpected_iid(id); break; 1587 } 1588 set_result(_gvn.transform(n)); 1589 return true; 1590 } 1591 1592 //------------------------------runtime_math----------------------------- 1593 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1594 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1595 "must be (DD)D or (D)D type"); 1596 1597 // Inputs 1598 Node* a = round_double_node(argument(0)); 1599 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1600 1601 const TypePtr* no_memory_effects = NULL; 1602 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1603 no_memory_effects, 1604 a, top(), b, b ? top() : NULL); 1605 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1606 #ifdef ASSERT 1607 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1608 assert(value_top == top(), "second value must be top"); 1609 #endif 1610 1611 set_result(value); 1612 return true; 1613 } 1614 1615 //------------------------------inline_math_native----------------------------- 1616 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1617 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1618 switch (id) { 1619 // These intrinsics are not properly supported on all hardware 1620 case vmIntrinsics::_dsin: 1621 return StubRoutines::dsin() != NULL ? 1622 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1623 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1624 case vmIntrinsics::_dcos: 1625 return StubRoutines::dcos() != NULL ? 1626 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1627 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1628 case vmIntrinsics::_dtan: 1629 return StubRoutines::dtan() != NULL ? 1630 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1631 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1632 case vmIntrinsics::_dlog: 1633 return StubRoutines::dlog() != NULL ? 1634 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1635 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1636 case vmIntrinsics::_dlog10: 1637 return StubRoutines::dlog10() != NULL ? 1638 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1639 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1640 1641 // These intrinsics are supported on all hardware 1642 case vmIntrinsics::_ceil: 1643 case vmIntrinsics::_floor: 1644 case vmIntrinsics::_rint: return Matcher::match_rule_supported(Op_RoundDoubleMode) ? inline_double_math(id) : false; 1645 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_double_math(id) : false; 1646 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_double_math(id) : false; 1647 case vmIntrinsics::_fabs: return Matcher::match_rule_supported(Op_AbsF) ? inline_math(id) : false; 1648 case vmIntrinsics::_iabs: return Matcher::match_rule_supported(Op_AbsI) ? inline_math(id) : false; 1649 case vmIntrinsics::_labs: return Matcher::match_rule_supported(Op_AbsL) ? inline_math(id) : false; 1650 1651 case vmIntrinsics::_dexp: 1652 return StubRoutines::dexp() != NULL ? 1653 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1654 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1655 case vmIntrinsics::_dpow: { 1656 Node* exp = round_double_node(argument(2)); 1657 const TypeD* d = _gvn.type(exp)->isa_double_constant(); 1658 if (d != NULL && d->getd() == 2.0) { 1659 // Special case: pow(x, 2.0) => x * x 1660 Node* base = round_double_node(argument(0)); 1661 set_result(_gvn.transform(new MulDNode(base, base))); 1662 return true; 1663 } 1664 return StubRoutines::dpow() != NULL ? 1665 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1666 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1667 } 1668 #undef FN_PTR 1669 1670 // These intrinsics are not yet correctly implemented 1671 case vmIntrinsics::_datan2: 1672 return false; 1673 1674 default: 1675 fatal_unexpected_iid(id); 1676 return false; 1677 } 1678 } 1679 1680 static bool is_simple_name(Node* n) { 1681 return (n->req() == 1 // constant 1682 || (n->is_Type() && n->as_Type()->type()->singleton()) 1683 || n->is_Proj() // parameter or return value 1684 || n->is_Phi() // local of some sort 1685 ); 1686 } 1687 1688 //----------------------------inline_notify-----------------------------------* 1689 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1690 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1691 address func; 1692 if (id == vmIntrinsics::_notify) { 1693 func = OptoRuntime::monitor_notify_Java(); 1694 } else { 1695 func = OptoRuntime::monitor_notifyAll_Java(); 1696 } 1697 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1698 make_slow_call_ex(call, env()->Throwable_klass(), false); 1699 return true; 1700 } 1701 1702 1703 //----------------------------inline_min_max----------------------------------- 1704 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1705 set_result(generate_min_max(id, argument(0), argument(1))); 1706 return true; 1707 } 1708 1709 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1710 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1711 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1712 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1713 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1714 1715 { 1716 PreserveJVMState pjvms(this); 1717 PreserveReexecuteState preexecs(this); 1718 jvms()->set_should_reexecute(true); 1719 1720 set_control(slow_path); 1721 set_i_o(i_o()); 1722 1723 uncommon_trap(Deoptimization::Reason_intrinsic, 1724 Deoptimization::Action_none); 1725 } 1726 1727 set_control(fast_path); 1728 set_result(math); 1729 } 1730 1731 template <typename OverflowOp> 1732 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1733 typedef typename OverflowOp::MathOp MathOp; 1734 1735 MathOp* mathOp = new MathOp(arg1, arg2); 1736 Node* operation = _gvn.transform( mathOp ); 1737 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1738 inline_math_mathExact(operation, ofcheck); 1739 return true; 1740 } 1741 1742 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1743 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1744 } 1745 1746 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1747 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1748 } 1749 1750 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1751 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1752 } 1753 1754 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1755 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1756 } 1757 1758 bool LibraryCallKit::inline_math_negateExactI() { 1759 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1760 } 1761 1762 bool LibraryCallKit::inline_math_negateExactL() { 1763 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1764 } 1765 1766 bool LibraryCallKit::inline_math_multiplyExactI() { 1767 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1768 } 1769 1770 bool LibraryCallKit::inline_math_multiplyExactL() { 1771 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1772 } 1773 1774 bool LibraryCallKit::inline_math_multiplyHigh() { 1775 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); 1776 return true; 1777 } 1778 1779 Node* 1780 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1781 // These are the candidate return value: 1782 Node* xvalue = x0; 1783 Node* yvalue = y0; 1784 1785 if (xvalue == yvalue) { 1786 return xvalue; 1787 } 1788 1789 bool want_max = (id == vmIntrinsics::_max); 1790 1791 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1792 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1793 if (txvalue == NULL || tyvalue == NULL) return top(); 1794 // This is not really necessary, but it is consistent with a 1795 // hypothetical MaxINode::Value method: 1796 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1797 1798 // %%% This folding logic should (ideally) be in a different place. 1799 // Some should be inside IfNode, and there to be a more reliable 1800 // transformation of ?: style patterns into cmoves. We also want 1801 // more powerful optimizations around cmove and min/max. 1802 1803 // Try to find a dominating comparison of these guys. 1804 // It can simplify the index computation for Arrays.copyOf 1805 // and similar uses of System.arraycopy. 1806 // First, compute the normalized version of CmpI(x, y). 1807 int cmp_op = Op_CmpI; 1808 Node* xkey = xvalue; 1809 Node* ykey = yvalue; 1810 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 1811 if (ideal_cmpxy->is_Cmp()) { 1812 // E.g., if we have CmpI(length - offset, count), 1813 // it might idealize to CmpI(length, count + offset) 1814 cmp_op = ideal_cmpxy->Opcode(); 1815 xkey = ideal_cmpxy->in(1); 1816 ykey = ideal_cmpxy->in(2); 1817 } 1818 1819 // Start by locating any relevant comparisons. 1820 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1821 Node* cmpxy = NULL; 1822 Node* cmpyx = NULL; 1823 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1824 Node* cmp = start_from->fast_out(k); 1825 if (cmp->outcnt() > 0 && // must have prior uses 1826 cmp->in(0) == NULL && // must be context-independent 1827 cmp->Opcode() == cmp_op) { // right kind of compare 1828 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1829 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1830 } 1831 } 1832 1833 const int NCMPS = 2; 1834 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1835 int cmpn; 1836 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1837 if (cmps[cmpn] != NULL) break; // find a result 1838 } 1839 if (cmpn < NCMPS) { 1840 // Look for a dominating test that tells us the min and max. 1841 int depth = 0; // Limit search depth for speed 1842 Node* dom = control(); 1843 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1844 if (++depth >= 100) break; 1845 Node* ifproj = dom; 1846 if (!ifproj->is_Proj()) continue; 1847 Node* iff = ifproj->in(0); 1848 if (!iff->is_If()) continue; 1849 Node* bol = iff->in(1); 1850 if (!bol->is_Bool()) continue; 1851 Node* cmp = bol->in(1); 1852 if (cmp == NULL) continue; 1853 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1854 if (cmps[cmpn] == cmp) break; 1855 if (cmpn == NCMPS) continue; 1856 BoolTest::mask btest = bol->as_Bool()->_test._test; 1857 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1858 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1859 // At this point, we know that 'x btest y' is true. 1860 switch (btest) { 1861 case BoolTest::eq: 1862 // They are proven equal, so we can collapse the min/max. 1863 // Either value is the answer. Choose the simpler. 1864 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1865 return yvalue; 1866 return xvalue; 1867 case BoolTest::lt: // x < y 1868 case BoolTest::le: // x <= y 1869 return (want_max ? yvalue : xvalue); 1870 case BoolTest::gt: // x > y 1871 case BoolTest::ge: // x >= y 1872 return (want_max ? xvalue : yvalue); 1873 default: 1874 break; 1875 } 1876 } 1877 } 1878 1879 // We failed to find a dominating test. 1880 // Let's pick a test that might GVN with prior tests. 1881 Node* best_bol = NULL; 1882 BoolTest::mask best_btest = BoolTest::illegal; 1883 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1884 Node* cmp = cmps[cmpn]; 1885 if (cmp == NULL) continue; 1886 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1887 Node* bol = cmp->fast_out(j); 1888 if (!bol->is_Bool()) continue; 1889 BoolTest::mask btest = bol->as_Bool()->_test._test; 1890 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1891 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1892 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1893 best_bol = bol->as_Bool(); 1894 best_btest = btest; 1895 } 1896 } 1897 } 1898 1899 Node* answer_if_true = NULL; 1900 Node* answer_if_false = NULL; 1901 switch (best_btest) { 1902 default: 1903 if (cmpxy == NULL) 1904 cmpxy = ideal_cmpxy; 1905 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 1906 // and fall through: 1907 case BoolTest::lt: // x < y 1908 case BoolTest::le: // x <= y 1909 answer_if_true = (want_max ? yvalue : xvalue); 1910 answer_if_false = (want_max ? xvalue : yvalue); 1911 break; 1912 case BoolTest::gt: // x > y 1913 case BoolTest::ge: // x >= y 1914 answer_if_true = (want_max ? xvalue : yvalue); 1915 answer_if_false = (want_max ? yvalue : xvalue); 1916 break; 1917 } 1918 1919 jint hi, lo; 1920 if (want_max) { 1921 // We can sharpen the minimum. 1922 hi = MAX2(txvalue->_hi, tyvalue->_hi); 1923 lo = MAX2(txvalue->_lo, tyvalue->_lo); 1924 } else { 1925 // We can sharpen the maximum. 1926 hi = MIN2(txvalue->_hi, tyvalue->_hi); 1927 lo = MIN2(txvalue->_lo, tyvalue->_lo); 1928 } 1929 1930 // Use a flow-free graph structure, to avoid creating excess control edges 1931 // which could hinder other optimizations. 1932 // Since Math.min/max is often used with arraycopy, we want 1933 // tightly_coupled_allocation to be able to see beyond min/max expressions. 1934 Node* cmov = CMoveNode::make(NULL, best_bol, 1935 answer_if_false, answer_if_true, 1936 TypeInt::make(lo, hi, widen)); 1937 1938 return _gvn.transform(cmov); 1939 1940 /* 1941 // This is not as desirable as it may seem, since Min and Max 1942 // nodes do not have a full set of optimizations. 1943 // And they would interfere, anyway, with 'if' optimizations 1944 // and with CMoveI canonical forms. 1945 switch (id) { 1946 case vmIntrinsics::_min: 1947 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 1948 case vmIntrinsics::_max: 1949 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 1950 default: 1951 ShouldNotReachHere(); 1952 } 1953 */ 1954 } 1955 1956 inline int 1957 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { 1958 const TypePtr* base_type = TypePtr::NULL_PTR; 1959 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 1960 if (base_type == NULL) { 1961 // Unknown type. 1962 return Type::AnyPtr; 1963 } else if (base_type == TypePtr::NULL_PTR) { 1964 // Since this is a NULL+long form, we have to switch to a rawptr. 1965 base = _gvn.transform(new CastX2PNode(offset)); 1966 offset = MakeConX(0); 1967 return Type::RawPtr; 1968 } else if (base_type->base() == Type::RawPtr) { 1969 return Type::RawPtr; 1970 } else if (base_type->isa_oopptr()) { 1971 // Base is never null => always a heap address. 1972 if (!TypePtr::NULL_PTR->higher_equal(base_type)) { 1973 return Type::OopPtr; 1974 } 1975 // Offset is small => always a heap address. 1976 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 1977 if (offset_type != NULL && 1978 base_type->offset() == 0 && // (should always be?) 1979 offset_type->_lo >= 0 && 1980 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 1981 return Type::OopPtr; 1982 } else if (type == T_OBJECT) { 1983 // off heap access to an oop doesn't make any sense. Has to be on 1984 // heap. 1985 return Type::OopPtr; 1986 } 1987 // Otherwise, it might either be oop+off or NULL+addr. 1988 return Type::AnyPtr; 1989 } else { 1990 // No information: 1991 return Type::AnyPtr; 1992 } 1993 } 1994 1995 Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) { 1996 Node* uncasted_base = base; 1997 int kind = classify_unsafe_addr(uncasted_base, offset, type); 1998 if (kind == Type::RawPtr) { 1999 return basic_plus_adr(top(), uncasted_base, offset); 2000 } else if (kind == Type::AnyPtr) { 2001 assert(base == uncasted_base, "unexpected base change"); 2002 if (can_cast) { 2003 if (!_gvn.type(base)->speculative_maybe_null() && 2004 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2005 // According to profiling, this access is always on 2006 // heap. Casting the base to not null and thus avoiding membars 2007 // around the access should allow better optimizations 2008 Node* null_ctl = top(); 2009 base = null_check_oop(base, &null_ctl, true, true, true); 2010 assert(null_ctl->is_top(), "no null control here"); 2011 return basic_plus_adr(base, offset); 2012 } else if (_gvn.type(base)->speculative_always_null() && 2013 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2014 // According to profiling, this access is always off 2015 // heap. 2016 base = null_assert(base); 2017 Node* raw_base = _gvn.transform(new CastX2PNode(offset)); 2018 offset = MakeConX(0); 2019 return basic_plus_adr(top(), raw_base, offset); 2020 } 2021 } 2022 // We don't know if it's an on heap or off heap access. Fall back 2023 // to raw memory access. 2024 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); 2025 return basic_plus_adr(top(), raw, offset); 2026 } else { 2027 assert(base == uncasted_base, "unexpected base change"); 2028 // We know it's an on heap access so base can't be null 2029 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { 2030 base = must_be_not_null(base, true); 2031 } 2032 return basic_plus_adr(base, offset); 2033 } 2034 } 2035 2036 //--------------------------inline_number_methods----------------------------- 2037 // inline int Integer.numberOfLeadingZeros(int) 2038 // inline int Long.numberOfLeadingZeros(long) 2039 // 2040 // inline int Integer.numberOfTrailingZeros(int) 2041 // inline int Long.numberOfTrailingZeros(long) 2042 // 2043 // inline int Integer.bitCount(int) 2044 // inline int Long.bitCount(long) 2045 // 2046 // inline char Character.reverseBytes(char) 2047 // inline short Short.reverseBytes(short) 2048 // inline int Integer.reverseBytes(int) 2049 // inline long Long.reverseBytes(long) 2050 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2051 Node* arg = argument(0); 2052 Node* n = NULL; 2053 switch (id) { 2054 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2055 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2056 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2057 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2058 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2059 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2060 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2061 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2062 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2063 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2064 default: fatal_unexpected_iid(id); break; 2065 } 2066 set_result(_gvn.transform(n)); 2067 return true; 2068 } 2069 2070 //----------------------------inline_unsafe_access---------------------------- 2071 2072 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2073 // Attempt to infer a sharper value type from the offset and base type. 2074 ciKlass* sharpened_klass = NULL; 2075 2076 // See if it is an instance field, with an object type. 2077 if (alias_type->field() != NULL) { 2078 if (alias_type->field()->type()->is_klass()) { 2079 sharpened_klass = alias_type->field()->type()->as_klass(); 2080 } 2081 } 2082 2083 // See if it is a narrow oop array. 2084 if (adr_type->isa_aryptr()) { 2085 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2086 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2087 if (elem_type != NULL) { 2088 sharpened_klass = elem_type->klass(); 2089 } 2090 } 2091 } 2092 2093 // The sharpened class might be unloaded if there is no class loader 2094 // contraint in place. 2095 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2096 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2097 2098 #ifndef PRODUCT 2099 if (C->print_intrinsics() || C->print_inlining()) { 2100 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2101 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2102 } 2103 #endif 2104 // Sharpen the value type. 2105 return tjp; 2106 } 2107 return NULL; 2108 } 2109 2110 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) { 2111 switch (kind) { 2112 case Relaxed: 2113 return MO_UNORDERED; 2114 case Opaque: 2115 return MO_RELAXED; 2116 case Acquire: 2117 return MO_ACQUIRE; 2118 case Release: 2119 return MO_RELEASE; 2120 case Volatile: 2121 return MO_SEQ_CST; 2122 default: 2123 ShouldNotReachHere(); 2124 return 0; 2125 } 2126 } 2127 2128 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2129 if (callee()->is_static()) return false; // caller must have the capability! 2130 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2131 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2132 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2133 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2134 2135 if (is_reference_type(type)) { 2136 decorators |= ON_UNKNOWN_OOP_REF; 2137 } 2138 2139 if (unaligned) { 2140 decorators |= C2_UNALIGNED; 2141 } 2142 2143 #ifndef PRODUCT 2144 { 2145 ResourceMark rm; 2146 // Check the signatures. 2147 ciSignature* sig = callee()->signature(); 2148 #ifdef ASSERT 2149 if (!is_store) { 2150 // Object getReference(Object base, int/long offset), etc. 2151 BasicType rtype = sig->return_type()->basic_type(); 2152 assert(rtype == type, "getter must return the expected value"); 2153 assert(sig->count() == 2, "oop getter has 2 arguments"); 2154 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2155 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2156 } else { 2157 // void putReference(Object base, int/long offset, Object x), etc. 2158 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2159 assert(sig->count() == 3, "oop putter has 3 arguments"); 2160 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2161 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2162 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2163 assert(vtype == type, "putter must accept the expected value"); 2164 } 2165 #endif // ASSERT 2166 } 2167 #endif //PRODUCT 2168 2169 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2170 2171 Node* receiver = argument(0); // type: oop 2172 2173 // Build address expression. 2174 Node* adr; 2175 Node* heap_base_oop = top(); 2176 Node* offset = top(); 2177 Node* val; 2178 2179 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2180 Node* base = argument(1); // type: oop 2181 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2182 offset = argument(2); // type: long 2183 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2184 // to be plain byte offsets, which are also the same as those accepted 2185 // by oopDesc::field_addr. 2186 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2187 "fieldOffset must be byte-scaled"); 2188 // 32-bit machines ignore the high half! 2189 offset = ConvL2X(offset); 2190 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed); 2191 2192 if (_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) { 2193 if (type != T_OBJECT) { 2194 decorators |= IN_NATIVE; // off-heap primitive access 2195 } else { 2196 return false; // off-heap oop accesses are not supported 2197 } 2198 } else { 2199 heap_base_oop = base; // on-heap or mixed access 2200 } 2201 2202 // Can base be NULL? Otherwise, always on-heap access. 2203 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base)); 2204 2205 if (!can_access_non_heap) { 2206 decorators |= IN_HEAP; 2207 } 2208 2209 val = is_store ? argument(4) : NULL; 2210 2211 const TypePtr* adr_type = _gvn.type(adr)->isa_ptr(); 2212 if (adr_type == TypePtr::NULL_PTR) { 2213 return false; // off-heap access with zero address 2214 } 2215 2216 // Try to categorize the address. 2217 Compile::AliasType* alias_type = C->alias_type(adr_type); 2218 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2219 2220 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2221 alias_type->adr_type() == TypeAryPtr::RANGE) { 2222 return false; // not supported 2223 } 2224 2225 bool mismatched = false; 2226 BasicType bt = alias_type->basic_type(); 2227 if (bt != T_ILLEGAL) { 2228 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2229 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2230 // Alias type doesn't differentiate between byte[] and boolean[]). 2231 // Use address type to get the element type. 2232 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2233 } 2234 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2235 // accessing an array field with getReference is not a mismatch 2236 bt = T_OBJECT; 2237 } 2238 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2239 // Don't intrinsify mismatched object accesses 2240 return false; 2241 } 2242 mismatched = (bt != type); 2243 } else if (alias_type->adr_type()->isa_oopptr()) { 2244 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2245 } 2246 2247 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2248 2249 if (mismatched) { 2250 decorators |= C2_MISMATCHED; 2251 } 2252 2253 // First guess at the value type. 2254 const Type *value_type = Type::get_const_basic_type(type); 2255 2256 // Figure out the memory ordering. 2257 decorators |= mo_decorator_for_access_kind(kind); 2258 2259 if (!is_store && type == T_OBJECT) { 2260 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2261 if (tjp != NULL) { 2262 value_type = tjp; 2263 } 2264 } 2265 2266 receiver = null_check(receiver); 2267 if (stopped()) { 2268 return true; 2269 } 2270 // Heap pointers get a null-check from the interpreter, 2271 // as a courtesy. However, this is not guaranteed by Unsafe, 2272 // and it is not possible to fully distinguish unintended nulls 2273 // from intended ones in this API. 2274 2275 if (!is_store) { 2276 Node* p = NULL; 2277 // Try to constant fold a load from a constant field 2278 ciField* field = alias_type->field(); 2279 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2280 // final or stable field 2281 p = make_constant_from_field(field, heap_base_oop); 2282 } 2283 2284 if (p == NULL) { // Could not constant fold the load 2285 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators); 2286 // Normalize the value returned by getBoolean in the following cases 2287 if (type == T_BOOLEAN && 2288 (mismatched || 2289 heap_base_oop == top() || // - heap_base_oop is NULL or 2290 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL 2291 // and the unsafe access is made to large offset 2292 // (i.e., larger than the maximum offset necessary for any 2293 // field access) 2294 ) { 2295 IdealKit ideal = IdealKit(this); 2296 #define __ ideal. 2297 IdealVariable normalized_result(ideal); 2298 __ declarations_done(); 2299 __ set(normalized_result, p); 2300 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2301 __ set(normalized_result, ideal.ConI(1)); 2302 ideal.end_if(); 2303 final_sync(ideal); 2304 p = __ value(normalized_result); 2305 #undef __ 2306 } 2307 } 2308 if (type == T_ADDRESS) { 2309 p = gvn().transform(new CastP2XNode(NULL, p)); 2310 p = ConvX2UL(p); 2311 } 2312 // The load node has the control of the preceding MemBarCPUOrder. All 2313 // following nodes will have the control of the MemBarCPUOrder inserted at 2314 // the end of this method. So, pushing the load onto the stack at a later 2315 // point is fine. 2316 set_result(p); 2317 } else { 2318 if (bt == T_ADDRESS) { 2319 // Repackage the long as a pointer. 2320 val = ConvL2X(val); 2321 val = gvn().transform(new CastX2PNode(val)); 2322 } 2323 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators); 2324 } 2325 2326 return true; 2327 } 2328 2329 //----------------------------inline_unsafe_load_store---------------------------- 2330 // This method serves a couple of different customers (depending on LoadStoreKind): 2331 // 2332 // LS_cmp_swap: 2333 // 2334 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x); 2335 // boolean compareAndSetInt( Object o, long offset, int expected, int x); 2336 // boolean compareAndSetLong( Object o, long offset, long expected, long x); 2337 // 2338 // LS_cmp_swap_weak: 2339 // 2340 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x); 2341 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x); 2342 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x); 2343 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x); 2344 // 2345 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); 2346 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); 2347 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); 2348 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); 2349 // 2350 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); 2351 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); 2352 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); 2353 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); 2354 // 2355 // LS_cmp_exchange: 2356 // 2357 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x); 2358 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x); 2359 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x); 2360 // 2361 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2362 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2363 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2364 // 2365 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2366 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2367 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2368 // 2369 // LS_get_add: 2370 // 2371 // int getAndAddInt( Object o, long offset, int delta) 2372 // long getAndAddLong(Object o, long offset, long delta) 2373 // 2374 // LS_get_set: 2375 // 2376 // int getAndSet(Object o, long offset, int newValue) 2377 // long getAndSet(Object o, long offset, long newValue) 2378 // Object getAndSet(Object o, long offset, Object newValue) 2379 // 2380 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2381 // This basic scheme here is the same as inline_unsafe_access, but 2382 // differs in enough details that combining them would make the code 2383 // overly confusing. (This is a true fact! I originally combined 2384 // them, but even I was confused by it!) As much code/comments as 2385 // possible are retained from inline_unsafe_access though to make 2386 // the correspondences clearer. - dl 2387 2388 if (callee()->is_static()) return false; // caller must have the capability! 2389 2390 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2391 decorators |= mo_decorator_for_access_kind(access_kind); 2392 2393 #ifndef PRODUCT 2394 BasicType rtype; 2395 { 2396 ResourceMark rm; 2397 // Check the signatures. 2398 ciSignature* sig = callee()->signature(); 2399 rtype = sig->return_type()->basic_type(); 2400 switch(kind) { 2401 case LS_get_add: 2402 case LS_get_set: { 2403 // Check the signatures. 2404 #ifdef ASSERT 2405 assert(rtype == type, "get and set must return the expected type"); 2406 assert(sig->count() == 3, "get and set has 3 arguments"); 2407 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2408 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2409 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2410 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2411 #endif // ASSERT 2412 break; 2413 } 2414 case LS_cmp_swap: 2415 case LS_cmp_swap_weak: { 2416 // Check the signatures. 2417 #ifdef ASSERT 2418 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2419 assert(sig->count() == 4, "CAS has 4 arguments"); 2420 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2421 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2422 #endif // ASSERT 2423 break; 2424 } 2425 case LS_cmp_exchange: { 2426 // Check the signatures. 2427 #ifdef ASSERT 2428 assert(rtype == type, "CAS must return the expected type"); 2429 assert(sig->count() == 4, "CAS has 4 arguments"); 2430 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2431 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2432 #endif // ASSERT 2433 break; 2434 } 2435 default: 2436 ShouldNotReachHere(); 2437 } 2438 } 2439 #endif //PRODUCT 2440 2441 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2442 2443 // Get arguments: 2444 Node* receiver = NULL; 2445 Node* base = NULL; 2446 Node* offset = NULL; 2447 Node* oldval = NULL; 2448 Node* newval = NULL; 2449 switch(kind) { 2450 case LS_cmp_swap: 2451 case LS_cmp_swap_weak: 2452 case LS_cmp_exchange: { 2453 const bool two_slot_type = type2size[type] == 2; 2454 receiver = argument(0); // type: oop 2455 base = argument(1); // type: oop 2456 offset = argument(2); // type: long 2457 oldval = argument(4); // type: oop, int, or long 2458 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2459 break; 2460 } 2461 case LS_get_add: 2462 case LS_get_set: { 2463 receiver = argument(0); // type: oop 2464 base = argument(1); // type: oop 2465 offset = argument(2); // type: long 2466 oldval = NULL; 2467 newval = argument(4); // type: oop, int, or long 2468 break; 2469 } 2470 default: 2471 ShouldNotReachHere(); 2472 } 2473 2474 // Build field offset expression. 2475 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2476 // to be plain byte offsets, which are also the same as those accepted 2477 // by oopDesc::field_addr. 2478 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2479 // 32-bit machines ignore the high half of long offsets 2480 offset = ConvL2X(offset); 2481 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false); 2482 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2483 2484 Compile::AliasType* alias_type = C->alias_type(adr_type); 2485 BasicType bt = alias_type->basic_type(); 2486 if (bt != T_ILLEGAL && 2487 (is_reference_type(bt) != (type == T_OBJECT))) { 2488 // Don't intrinsify mismatched object accesses. 2489 return false; 2490 } 2491 2492 // For CAS, unlike inline_unsafe_access, there seems no point in 2493 // trying to refine types. Just use the coarse types here. 2494 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2495 const Type *value_type = Type::get_const_basic_type(type); 2496 2497 switch (kind) { 2498 case LS_get_set: 2499 case LS_cmp_exchange: { 2500 if (type == T_OBJECT) { 2501 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2502 if (tjp != NULL) { 2503 value_type = tjp; 2504 } 2505 } 2506 break; 2507 } 2508 case LS_cmp_swap: 2509 case LS_cmp_swap_weak: 2510 case LS_get_add: 2511 break; 2512 default: 2513 ShouldNotReachHere(); 2514 } 2515 2516 // Null check receiver. 2517 receiver = null_check(receiver); 2518 if (stopped()) { 2519 return true; 2520 } 2521 2522 int alias_idx = C->get_alias_index(adr_type); 2523 2524 if (is_reference_type(type)) { 2525 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF; 2526 2527 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2528 // could be delayed during Parse (for example, in adjust_map_after_if()). 2529 // Execute transformation here to avoid barrier generation in such case. 2530 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2531 newval = _gvn.makecon(TypePtr::NULL_PTR); 2532 2533 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) { 2534 // Refine the value to a null constant, when it is known to be null 2535 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2536 } 2537 } 2538 2539 Node* result = NULL; 2540 switch (kind) { 2541 case LS_cmp_exchange: { 2542 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx, 2543 oldval, newval, value_type, type, decorators); 2544 break; 2545 } 2546 case LS_cmp_swap_weak: 2547 decorators |= C2_WEAK_CMPXCHG; 2548 case LS_cmp_swap: { 2549 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx, 2550 oldval, newval, value_type, type, decorators); 2551 break; 2552 } 2553 case LS_get_set: { 2554 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx, 2555 newval, value_type, type, decorators); 2556 break; 2557 } 2558 case LS_get_add: { 2559 result = access_atomic_add_at(base, adr, adr_type, alias_idx, 2560 newval, value_type, type, decorators); 2561 break; 2562 } 2563 default: 2564 ShouldNotReachHere(); 2565 } 2566 2567 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2568 set_result(result); 2569 return true; 2570 } 2571 2572 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2573 // Regardless of form, don't allow previous ld/st to move down, 2574 // then issue acquire, release, or volatile mem_bar. 2575 insert_mem_bar(Op_MemBarCPUOrder); 2576 switch(id) { 2577 case vmIntrinsics::_loadFence: 2578 insert_mem_bar(Op_LoadFence); 2579 return true; 2580 case vmIntrinsics::_storeFence: 2581 insert_mem_bar(Op_StoreFence); 2582 return true; 2583 case vmIntrinsics::_fullFence: 2584 insert_mem_bar(Op_MemBarVolatile); 2585 return true; 2586 default: 2587 fatal_unexpected_iid(id); 2588 return false; 2589 } 2590 } 2591 2592 bool LibraryCallKit::inline_onspinwait() { 2593 insert_mem_bar(Op_OnSpinWait); 2594 return true; 2595 } 2596 2597 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2598 if (!kls->is_Con()) { 2599 return true; 2600 } 2601 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2602 if (klsptr == NULL) { 2603 return true; 2604 } 2605 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2606 // don't need a guard for a klass that is already initialized 2607 return !ik->is_initialized(); 2608 } 2609 2610 //----------------------------inline_unsafe_writeback0------------------------- 2611 // public native void Unsafe.writeback0(long address) 2612 bool LibraryCallKit::inline_unsafe_writeback0() { 2613 if (!Matcher::has_match_rule(Op_CacheWB)) { 2614 return false; 2615 } 2616 #ifndef PRODUCT 2617 assert(Matcher::has_match_rule(Op_CacheWBPreSync), "found match rule for CacheWB but not CacheWBPreSync"); 2618 assert(Matcher::has_match_rule(Op_CacheWBPostSync), "found match rule for CacheWB but not CacheWBPostSync"); 2619 ciSignature* sig = callee()->signature(); 2620 assert(sig->type_at(0)->basic_type() == T_LONG, "Unsafe_writeback0 address is long!"); 2621 #endif 2622 null_check_receiver(); // null-check, then ignore 2623 Node *addr = argument(1); 2624 addr = new CastX2PNode(addr); 2625 addr = _gvn.transform(addr); 2626 Node *flush = new CacheWBNode(control(), memory(TypeRawPtr::BOTTOM), addr); 2627 flush = _gvn.transform(flush); 2628 set_memory(flush, TypeRawPtr::BOTTOM); 2629 return true; 2630 } 2631 2632 //----------------------------inline_unsafe_writeback0------------------------- 2633 // public native void Unsafe.writeback0(long address) 2634 bool LibraryCallKit::inline_unsafe_writebackSync0(bool is_pre) { 2635 if (is_pre && !Matcher::has_match_rule(Op_CacheWBPreSync)) { 2636 return false; 2637 } 2638 if (!is_pre && !Matcher::has_match_rule(Op_CacheWBPostSync)) { 2639 return false; 2640 } 2641 #ifndef PRODUCT 2642 assert(Matcher::has_match_rule(Op_CacheWB), 2643 (is_pre ? "found match rule for CacheWBPreSync but not CacheWB" 2644 : "found match rule for CacheWBPostSync but not CacheWB")); 2645 2646 #endif 2647 null_check_receiver(); // null-check, then ignore 2648 Node *sync; 2649 if (is_pre) { 2650 sync = new CacheWBPreSyncNode(control(), memory(TypeRawPtr::BOTTOM)); 2651 } else { 2652 sync = new CacheWBPostSyncNode(control(), memory(TypeRawPtr::BOTTOM)); 2653 } 2654 sync = _gvn.transform(sync); 2655 set_memory(sync, TypeRawPtr::BOTTOM); 2656 return true; 2657 } 2658 2659 //----------------------------inline_unsafe_allocate--------------------------- 2660 // public native Object Unsafe.allocateInstance(Class<?> cls); 2661 bool LibraryCallKit::inline_unsafe_allocate() { 2662 if (callee()->is_static()) return false; // caller must have the capability! 2663 2664 null_check_receiver(); // null-check, then ignore 2665 Node* cls = null_check(argument(1)); 2666 if (stopped()) return true; 2667 2668 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2669 kls = null_check(kls); 2670 if (stopped()) return true; // argument was like int.class 2671 2672 Node* test = NULL; 2673 if (LibraryCallKit::klass_needs_init_guard(kls)) { 2674 // Note: The argument might still be an illegal value like 2675 // Serializable.class or Object[].class. The runtime will handle it. 2676 // But we must make an explicit check for initialization. 2677 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2678 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2679 // can generate code to load it as unsigned byte. 2680 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 2681 Node* bits = intcon(InstanceKlass::fully_initialized); 2682 test = _gvn.transform(new SubINode(inst, bits)); 2683 // The 'test' is non-zero if we need to take a slow path. 2684 } 2685 2686 Node* obj = new_instance(kls, test); 2687 set_result(obj); 2688 return true; 2689 } 2690 2691 //------------------------inline_native_time_funcs-------------- 2692 // inline code for System.currentTimeMillis() and System.nanoTime() 2693 // these have the same type and signature 2694 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2695 const TypeFunc* tf = OptoRuntime::void_long_Type(); 2696 const TypePtr* no_memory_effects = NULL; 2697 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2698 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 2699 #ifdef ASSERT 2700 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 2701 assert(value_top == top(), "second value must be top"); 2702 #endif 2703 set_result(value); 2704 return true; 2705 } 2706 2707 #ifdef JFR_HAVE_INTRINSICS 2708 2709 /* 2710 * oop -> myklass 2711 * myklass->trace_id |= USED 2712 * return myklass->trace_id & ~0x3 2713 */ 2714 bool LibraryCallKit::inline_native_classID() { 2715 Node* cls = null_check(argument(0), T_OBJECT); 2716 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2717 kls = null_check(kls, T_OBJECT); 2718 2719 ByteSize offset = KLASS_TRACE_ID_OFFSET; 2720 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2721 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 2722 2723 Node* clsused = longcon(0x01l); // set the class bit 2724 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 2725 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2726 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 2727 2728 #ifdef TRACE_ID_META_BITS 2729 Node* mbits = longcon(~TRACE_ID_META_BITS); 2730 tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); 2731 #endif 2732 #ifdef TRACE_ID_SHIFT 2733 Node* cbits = intcon(TRACE_ID_SHIFT); 2734 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); 2735 #endif 2736 2737 set_result(tvalue); 2738 return true; 2739 2740 } 2741 2742 bool LibraryCallKit::inline_native_getEventWriter() { 2743 Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); 2744 2745 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 2746 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)); 2747 2748 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 2749 2750 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); 2751 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); 2752 2753 IfNode* iff_jobj_null = 2754 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 2755 2756 enum { _normal_path = 1, 2757 _null_path = 2, 2758 PATH_LIMIT }; 2759 2760 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 2761 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM); 2762 2763 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); 2764 result_rgn->init_req(_null_path, jobj_is_null); 2765 result_val->init_req(_null_path, null()); 2766 2767 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); 2768 set_control(jobj_is_not_null); 2769 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT, 2770 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD); 2771 result_rgn->init_req(_normal_path, control()); 2772 result_val->init_req(_normal_path, res); 2773 2774 set_result(result_rgn, result_val); 2775 2776 return true; 2777 } 2778 2779 #endif // JFR_HAVE_INTRINSICS 2780 2781 //------------------------inline_native_currentThread------------------ 2782 bool LibraryCallKit::inline_native_currentThread() { 2783 Node* junk = NULL; 2784 set_result(generate_current_thread(junk)); 2785 return true; 2786 } 2787 2788 //---------------------------load_mirror_from_klass---------------------------- 2789 // Given a klass oop, load its java mirror (a java.lang.Class oop). 2790 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 2791 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 2792 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 2793 // mirror = ((OopHandle)mirror)->resolve(); 2794 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 2795 } 2796 2797 //-----------------------load_klass_from_mirror_common------------------------- 2798 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 2799 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 2800 // and branch to the given path on the region. 2801 // If never_see_null, take an uncommon trap on null, so we can optimistically 2802 // compile for the non-null case. 2803 // If the region is NULL, force never_see_null = true. 2804 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 2805 bool never_see_null, 2806 RegionNode* region, 2807 int null_path, 2808 int offset) { 2809 if (region == NULL) never_see_null = true; 2810 Node* p = basic_plus_adr(mirror, offset); 2811 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2812 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 2813 Node* null_ctl = top(); 2814 kls = null_check_oop(kls, &null_ctl, never_see_null); 2815 if (region != NULL) { 2816 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 2817 region->init_req(null_path, null_ctl); 2818 } else { 2819 assert(null_ctl == top(), "no loose ends"); 2820 } 2821 return kls; 2822 } 2823 2824 //--------------------(inline_native_Class_query helpers)--------------------- 2825 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 2826 // Fall through if (mods & mask) == bits, take the guard otherwise. 2827 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 2828 // Branch around if the given klass has the given modifier bit set. 2829 // Like generate_guard, adds a new path onto the region. 2830 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 2831 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 2832 Node* mask = intcon(modifier_mask); 2833 Node* bits = intcon(modifier_bits); 2834 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 2835 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 2836 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 2837 return generate_fair_guard(bol, region); 2838 } 2839 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 2840 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 2841 } 2842 Node* LibraryCallKit::generate_hidden_class_guard(Node* kls, RegionNode* region) { 2843 return generate_access_flags_guard(kls, JVM_ACC_IS_HIDDEN_CLASS, 0, region); 2844 } 2845 2846 //-------------------------inline_native_Class_query------------------- 2847 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 2848 const Type* return_type = TypeInt::BOOL; 2849 Node* prim_return_value = top(); // what happens if it's a primitive class? 2850 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 2851 bool expect_prim = false; // most of these guys expect to work on refs 2852 2853 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 2854 2855 Node* mirror = argument(0); 2856 Node* obj = top(); 2857 2858 switch (id) { 2859 case vmIntrinsics::_isInstance: 2860 // nothing is an instance of a primitive type 2861 prim_return_value = intcon(0); 2862 obj = argument(1); 2863 break; 2864 case vmIntrinsics::_getModifiers: 2865 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2866 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 2867 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 2868 break; 2869 case vmIntrinsics::_isInterface: 2870 prim_return_value = intcon(0); 2871 break; 2872 case vmIntrinsics::_isArray: 2873 prim_return_value = intcon(0); 2874 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 2875 break; 2876 case vmIntrinsics::_isPrimitive: 2877 prim_return_value = intcon(1); 2878 expect_prim = true; // obviously 2879 break; 2880 case vmIntrinsics::_isHidden: 2881 prim_return_value = intcon(0); 2882 break; 2883 case vmIntrinsics::_getSuperclass: 2884 prim_return_value = null(); 2885 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2886 break; 2887 case vmIntrinsics::_getClassAccessFlags: 2888 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2889 return_type = TypeInt::INT; // not bool! 6297094 2890 break; 2891 default: 2892 fatal_unexpected_iid(id); 2893 break; 2894 } 2895 2896 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 2897 if (mirror_con == NULL) return false; // cannot happen? 2898 2899 #ifndef PRODUCT 2900 if (C->print_intrinsics() || C->print_inlining()) { 2901 ciType* k = mirror_con->java_mirror_type(); 2902 if (k) { 2903 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 2904 k->print_name(); 2905 tty->cr(); 2906 } 2907 } 2908 #endif 2909 2910 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 2911 RegionNode* region = new RegionNode(PATH_LIMIT); 2912 record_for_igvn(region); 2913 PhiNode* phi = new PhiNode(region, return_type); 2914 2915 // The mirror will never be null of Reflection.getClassAccessFlags, however 2916 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 2917 // if it is. See bug 4774291. 2918 2919 // For Reflection.getClassAccessFlags(), the null check occurs in 2920 // the wrong place; see inline_unsafe_access(), above, for a similar 2921 // situation. 2922 mirror = null_check(mirror); 2923 // If mirror or obj is dead, only null-path is taken. 2924 if (stopped()) return true; 2925 2926 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 2927 2928 // Now load the mirror's klass metaobject, and null-check it. 2929 // Side-effects region with the control path if the klass is null. 2930 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 2931 // If kls is null, we have a primitive mirror. 2932 phi->init_req(_prim_path, prim_return_value); 2933 if (stopped()) { set_result(region, phi); return true; } 2934 bool safe_for_replace = (region->in(_prim_path) == top()); 2935 2936 Node* p; // handy temp 2937 Node* null_ctl; 2938 2939 // Now that we have the non-null klass, we can perform the real query. 2940 // For constant classes, the query will constant-fold in LoadNode::Value. 2941 Node* query_value = top(); 2942 switch (id) { 2943 case vmIntrinsics::_isInstance: 2944 // nothing is an instance of a primitive type 2945 query_value = gen_instanceof(obj, kls, safe_for_replace); 2946 break; 2947 2948 case vmIntrinsics::_getModifiers: 2949 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 2950 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 2951 break; 2952 2953 case vmIntrinsics::_isInterface: 2954 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2955 if (generate_interface_guard(kls, region) != NULL) 2956 // A guard was added. If the guard is taken, it was an interface. 2957 phi->add_req(intcon(1)); 2958 // If we fall through, it's a plain class. 2959 query_value = intcon(0); 2960 break; 2961 2962 case vmIntrinsics::_isArray: 2963 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 2964 if (generate_array_guard(kls, region) != NULL) 2965 // A guard was added. If the guard is taken, it was an array. 2966 phi->add_req(intcon(1)); 2967 // If we fall through, it's a plain class. 2968 query_value = intcon(0); 2969 break; 2970 2971 case vmIntrinsics::_isPrimitive: 2972 query_value = intcon(0); // "normal" path produces false 2973 break; 2974 2975 case vmIntrinsics::_isHidden: 2976 // (To verify this code sequence, check the asserts in JVM_IsHiddenClass.) 2977 if (generate_hidden_class_guard(kls, region) != NULL) 2978 // A guard was added. If the guard is taken, it was an hidden class. 2979 phi->add_req(intcon(1)); 2980 // If we fall through, it's a plain class. 2981 query_value = intcon(0); 2982 break; 2983 2984 2985 case vmIntrinsics::_getSuperclass: 2986 // The rules here are somewhat unfortunate, but we can still do better 2987 // with random logic than with a JNI call. 2988 // Interfaces store null or Object as _super, but must report null. 2989 // Arrays store an intermediate super as _super, but must report Object. 2990 // Other types can report the actual _super. 2991 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2992 if (generate_interface_guard(kls, region) != NULL) 2993 // A guard was added. If the guard is taken, it was an interface. 2994 phi->add_req(null()); 2995 if (generate_array_guard(kls, region) != NULL) 2996 // A guard was added. If the guard is taken, it was an array. 2997 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 2998 // If we fall through, it's a plain class. Get its _super. 2999 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3000 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3001 null_ctl = top(); 3002 kls = null_check_oop(kls, &null_ctl); 3003 if (null_ctl != top()) { 3004 // If the guard is taken, Object.superClass is null (both klass and mirror). 3005 region->add_req(null_ctl); 3006 phi ->add_req(null()); 3007 } 3008 if (!stopped()) { 3009 query_value = load_mirror_from_klass(kls); 3010 } 3011 break; 3012 3013 case vmIntrinsics::_getClassAccessFlags: 3014 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3015 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3016 break; 3017 3018 default: 3019 fatal_unexpected_iid(id); 3020 break; 3021 } 3022 3023 // Fall-through is the normal case of a query to a real class. 3024 phi->init_req(1, query_value); 3025 region->init_req(1, control()); 3026 3027 C->set_has_split_ifs(true); // Has chance for split-if optimization 3028 set_result(region, phi); 3029 return true; 3030 } 3031 3032 //-------------------------inline_Class_cast------------------- 3033 bool LibraryCallKit::inline_Class_cast() { 3034 Node* mirror = argument(0); // Class 3035 Node* obj = argument(1); 3036 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3037 if (mirror_con == NULL) { 3038 return false; // dead path (mirror->is_top()). 3039 } 3040 if (obj == NULL || obj->is_top()) { 3041 return false; // dead path 3042 } 3043 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3044 3045 // First, see if Class.cast() can be folded statically. 3046 // java_mirror_type() returns non-null for compile-time Class constants. 3047 ciType* tm = mirror_con->java_mirror_type(); 3048 if (tm != NULL && tm->is_klass() && 3049 tp != NULL && tp->klass() != NULL) { 3050 if (!tp->klass()->is_loaded()) { 3051 // Don't use intrinsic when class is not loaded. 3052 return false; 3053 } else { 3054 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3055 if (static_res == Compile::SSC_always_true) { 3056 // isInstance() is true - fold the code. 3057 set_result(obj); 3058 return true; 3059 } else if (static_res == Compile::SSC_always_false) { 3060 // Don't use intrinsic, have to throw ClassCastException. 3061 // If the reference is null, the non-intrinsic bytecode will 3062 // be optimized appropriately. 3063 return false; 3064 } 3065 } 3066 } 3067 3068 // Bailout intrinsic and do normal inlining if exception path is frequent. 3069 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3070 return false; 3071 } 3072 3073 // Generate dynamic checks. 3074 // Class.cast() is java implementation of _checkcast bytecode. 3075 // Do checkcast (Parse::do_checkcast()) optimizations here. 3076 3077 mirror = null_check(mirror); 3078 // If mirror is dead, only null-path is taken. 3079 if (stopped()) { 3080 return true; 3081 } 3082 3083 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3084 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3085 RegionNode* region = new RegionNode(PATH_LIMIT); 3086 record_for_igvn(region); 3087 3088 // Now load the mirror's klass metaobject, and null-check it. 3089 // If kls is null, we have a primitive mirror and 3090 // nothing is an instance of a primitive type. 3091 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3092 3093 Node* res = top(); 3094 if (!stopped()) { 3095 Node* bad_type_ctrl = top(); 3096 // Do checkcast optimizations. 3097 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3098 region->init_req(_bad_type_path, bad_type_ctrl); 3099 } 3100 if (region->in(_prim_path) != top() || 3101 region->in(_bad_type_path) != top()) { 3102 // Let Interpreter throw ClassCastException. 3103 PreserveJVMState pjvms(this); 3104 set_control(_gvn.transform(region)); 3105 uncommon_trap(Deoptimization::Reason_intrinsic, 3106 Deoptimization::Action_maybe_recompile); 3107 } 3108 if (!stopped()) { 3109 set_result(res); 3110 } 3111 return true; 3112 } 3113 3114 3115 //--------------------------inline_native_subtype_check------------------------ 3116 // This intrinsic takes the JNI calls out of the heart of 3117 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3118 bool LibraryCallKit::inline_native_subtype_check() { 3119 // Pull both arguments off the stack. 3120 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3121 args[0] = argument(0); 3122 args[1] = argument(1); 3123 Node* klasses[2]; // corresponding Klasses: superk, subk 3124 klasses[0] = klasses[1] = top(); 3125 3126 enum { 3127 // A full decision tree on {superc is prim, subc is prim}: 3128 _prim_0_path = 1, // {P,N} => false 3129 // {P,P} & superc!=subc => false 3130 _prim_same_path, // {P,P} & superc==subc => true 3131 _prim_1_path, // {N,P} => false 3132 _ref_subtype_path, // {N,N} & subtype check wins => true 3133 _both_ref_path, // {N,N} & subtype check loses => false 3134 PATH_LIMIT 3135 }; 3136 3137 RegionNode* region = new RegionNode(PATH_LIMIT); 3138 Node* phi = new PhiNode(region, TypeInt::BOOL); 3139 record_for_igvn(region); 3140 3141 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3142 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3143 int class_klass_offset = java_lang_Class::klass_offset(); 3144 3145 // First null-check both mirrors and load each mirror's klass metaobject. 3146 int which_arg; 3147 for (which_arg = 0; which_arg <= 1; which_arg++) { 3148 Node* arg = args[which_arg]; 3149 arg = null_check(arg); 3150 if (stopped()) break; 3151 args[which_arg] = arg; 3152 3153 Node* p = basic_plus_adr(arg, class_klass_offset); 3154 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3155 klasses[which_arg] = _gvn.transform(kls); 3156 } 3157 3158 // Having loaded both klasses, test each for null. 3159 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3160 for (which_arg = 0; which_arg <= 1; which_arg++) { 3161 Node* kls = klasses[which_arg]; 3162 Node* null_ctl = top(); 3163 kls = null_check_oop(kls, &null_ctl, never_see_null); 3164 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3165 region->init_req(prim_path, null_ctl); 3166 if (stopped()) break; 3167 klasses[which_arg] = kls; 3168 } 3169 3170 if (!stopped()) { 3171 // now we have two reference types, in klasses[0..1] 3172 Node* subk = klasses[1]; // the argument to isAssignableFrom 3173 Node* superk = klasses[0]; // the receiver 3174 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3175 // now we have a successful reference subtype check 3176 region->set_req(_ref_subtype_path, control()); 3177 } 3178 3179 // If both operands are primitive (both klasses null), then 3180 // we must return true when they are identical primitives. 3181 // It is convenient to test this after the first null klass check. 3182 set_control(region->in(_prim_0_path)); // go back to first null check 3183 if (!stopped()) { 3184 // Since superc is primitive, make a guard for the superc==subc case. 3185 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3186 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3187 generate_guard(bol_eq, region, PROB_FAIR); 3188 if (region->req() == PATH_LIMIT+1) { 3189 // A guard was added. If the added guard is taken, superc==subc. 3190 region->swap_edges(PATH_LIMIT, _prim_same_path); 3191 region->del_req(PATH_LIMIT); 3192 } 3193 region->set_req(_prim_0_path, control()); // Not equal after all. 3194 } 3195 3196 // these are the only paths that produce 'true': 3197 phi->set_req(_prim_same_path, intcon(1)); 3198 phi->set_req(_ref_subtype_path, intcon(1)); 3199 3200 // pull together the cases: 3201 assert(region->req() == PATH_LIMIT, "sane region"); 3202 for (uint i = 1; i < region->req(); i++) { 3203 Node* ctl = region->in(i); 3204 if (ctl == NULL || ctl == top()) { 3205 region->set_req(i, top()); 3206 phi ->set_req(i, top()); 3207 } else if (phi->in(i) == NULL) { 3208 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3209 } 3210 } 3211 3212 set_control(_gvn.transform(region)); 3213 set_result(_gvn.transform(phi)); 3214 return true; 3215 } 3216 3217 //---------------------generate_array_guard_common------------------------ 3218 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3219 bool obj_array, bool not_array) { 3220 3221 if (stopped()) { 3222 return NULL; 3223 } 3224 3225 // If obj_array/non_array==false/false: 3226 // Branch around if the given klass is in fact an array (either obj or prim). 3227 // If obj_array/non_array==false/true: 3228 // Branch around if the given klass is not an array klass of any kind. 3229 // If obj_array/non_array==true/true: 3230 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3231 // If obj_array/non_array==true/false: 3232 // Branch around if the kls is an oop array (Object[] or subtype) 3233 // 3234 // Like generate_guard, adds a new path onto the region. 3235 jint layout_con = 0; 3236 Node* layout_val = get_layout_helper(kls, layout_con); 3237 if (layout_val == NULL) { 3238 bool query = (obj_array 3239 ? Klass::layout_helper_is_objArray(layout_con) 3240 : Klass::layout_helper_is_array(layout_con)); 3241 if (query == not_array) { 3242 return NULL; // never a branch 3243 } else { // always a branch 3244 Node* always_branch = control(); 3245 if (region != NULL) 3246 region->add_req(always_branch); 3247 set_control(top()); 3248 return always_branch; 3249 } 3250 } 3251 // Now test the correct condition. 3252 jint nval = (obj_array 3253 ? (jint)(Klass::_lh_array_tag_type_value 3254 << Klass::_lh_array_tag_shift) 3255 : Klass::_lh_neutral_value); 3256 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3257 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3258 // invert the test if we are looking for a non-array 3259 if (not_array) btest = BoolTest(btest).negate(); 3260 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3261 return generate_fair_guard(bol, region); 3262 } 3263 3264 3265 //-----------------------inline_native_newArray-------------------------- 3266 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3267 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3268 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3269 Node* mirror; 3270 Node* count_val; 3271 if (uninitialized) { 3272 mirror = argument(1); 3273 count_val = argument(2); 3274 } else { 3275 mirror = argument(0); 3276 count_val = argument(1); 3277 } 3278 3279 mirror = null_check(mirror); 3280 // If mirror or obj is dead, only null-path is taken. 3281 if (stopped()) return true; 3282 3283 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3284 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3285 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3286 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3287 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3288 3289 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3290 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3291 result_reg, _slow_path); 3292 Node* normal_ctl = control(); 3293 Node* no_array_ctl = result_reg->in(_slow_path); 3294 3295 // Generate code for the slow case. We make a call to newArray(). 3296 set_control(no_array_ctl); 3297 if (!stopped()) { 3298 // Either the input type is void.class, or else the 3299 // array klass has not yet been cached. Either the 3300 // ensuing call will throw an exception, or else it 3301 // will cache the array klass for next time. 3302 PreserveJVMState pjvms(this); 3303 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3304 Node* slow_result = set_results_for_java_call(slow_call); 3305 // this->control() comes from set_results_for_java_call 3306 result_reg->set_req(_slow_path, control()); 3307 result_val->set_req(_slow_path, slow_result); 3308 result_io ->set_req(_slow_path, i_o()); 3309 result_mem->set_req(_slow_path, reset_memory()); 3310 } 3311 3312 set_control(normal_ctl); 3313 if (!stopped()) { 3314 // Normal case: The array type has been cached in the java.lang.Class. 3315 // The following call works fine even if the array type is polymorphic. 3316 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3317 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3318 result_reg->init_req(_normal_path, control()); 3319 result_val->init_req(_normal_path, obj); 3320 result_io ->init_req(_normal_path, i_o()); 3321 result_mem->init_req(_normal_path, reset_memory()); 3322 3323 if (uninitialized) { 3324 // Mark the allocation so that zeroing is skipped 3325 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3326 alloc->maybe_set_complete(&_gvn); 3327 } 3328 } 3329 3330 // Return the combined state. 3331 set_i_o( _gvn.transform(result_io) ); 3332 set_all_memory( _gvn.transform(result_mem)); 3333 3334 C->set_has_split_ifs(true); // Has chance for split-if optimization 3335 set_result(result_reg, result_val); 3336 return true; 3337 } 3338 3339 //----------------------inline_native_getLength-------------------------- 3340 // public static native int java.lang.reflect.Array.getLength(Object array); 3341 bool LibraryCallKit::inline_native_getLength() { 3342 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3343 3344 Node* array = null_check(argument(0)); 3345 // If array is dead, only null-path is taken. 3346 if (stopped()) return true; 3347 3348 // Deoptimize if it is a non-array. 3349 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3350 3351 if (non_array != NULL) { 3352 PreserveJVMState pjvms(this); 3353 set_control(non_array); 3354 uncommon_trap(Deoptimization::Reason_intrinsic, 3355 Deoptimization::Action_maybe_recompile); 3356 } 3357 3358 // If control is dead, only non-array-path is taken. 3359 if (stopped()) return true; 3360 3361 // The works fine even if the array type is polymorphic. 3362 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3363 Node* result = load_array_length(array); 3364 3365 C->set_has_split_ifs(true); // Has chance for split-if optimization 3366 set_result(result); 3367 return true; 3368 } 3369 3370 //------------------------inline_array_copyOf---------------------------- 3371 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3372 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3373 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3374 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3375 3376 // Get the arguments. 3377 Node* original = argument(0); 3378 Node* start = is_copyOfRange? argument(1): intcon(0); 3379 Node* end = is_copyOfRange? argument(2): argument(1); 3380 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3381 3382 Node* newcopy = NULL; 3383 3384 // Set the original stack and the reexecute bit for the interpreter to reexecute 3385 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3386 { PreserveReexecuteState preexecs(this); 3387 jvms()->set_should_reexecute(true); 3388 3389 array_type_mirror = null_check(array_type_mirror); 3390 original = null_check(original); 3391 3392 // Check if a null path was taken unconditionally. 3393 if (stopped()) return true; 3394 3395 Node* orig_length = load_array_length(original); 3396 3397 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3398 klass_node = null_check(klass_node); 3399 3400 RegionNode* bailout = new RegionNode(1); 3401 record_for_igvn(bailout); 3402 3403 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3404 // Bail out if that is so. 3405 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3406 if (not_objArray != NULL) { 3407 // Improve the klass node's type from the new optimistic assumption: 3408 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3409 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3410 Node* cast = new CastPPNode(klass_node, akls); 3411 cast->init_req(0, control()); 3412 klass_node = _gvn.transform(cast); 3413 } 3414 3415 // Bail out if either start or end is negative. 3416 generate_negative_guard(start, bailout, &start); 3417 generate_negative_guard(end, bailout, &end); 3418 3419 Node* length = end; 3420 if (_gvn.type(start) != TypeInt::ZERO) { 3421 length = _gvn.transform(new SubINode(end, start)); 3422 } 3423 3424 // Bail out if length is negative. 3425 // Without this the new_array would throw 3426 // NegativeArraySizeException but IllegalArgumentException is what 3427 // should be thrown 3428 generate_negative_guard(length, bailout, &length); 3429 3430 if (bailout->req() > 1) { 3431 PreserveJVMState pjvms(this); 3432 set_control(_gvn.transform(bailout)); 3433 uncommon_trap(Deoptimization::Reason_intrinsic, 3434 Deoptimization::Action_maybe_recompile); 3435 } 3436 3437 if (!stopped()) { 3438 // How many elements will we copy from the original? 3439 // The answer is MinI(orig_length - start, length). 3440 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3441 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3442 3443 // Generate a direct call to the right arraycopy function(s). 3444 // We know the copy is disjoint but we might not know if the 3445 // oop stores need checking. 3446 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3447 // This will fail a store-check if x contains any non-nulls. 3448 3449 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3450 // loads/stores but it is legal only if we're sure the 3451 // Arrays.copyOf would succeed. So we need all input arguments 3452 // to the copyOf to be validated, including that the copy to the 3453 // new array won't trigger an ArrayStoreException. That subtype 3454 // check can be optimized if we know something on the type of 3455 // the input array from type speculation. 3456 if (_gvn.type(klass_node)->singleton()) { 3457 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3458 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3459 3460 int test = C->static_subtype_check(superk, subk); 3461 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3462 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3463 if (t_original->speculative_type() != NULL) { 3464 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3465 } 3466 } 3467 } 3468 3469 bool validated = false; 3470 // Reason_class_check rather than Reason_intrinsic because we 3471 // want to intrinsify even if this traps. 3472 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3473 Node* not_subtype_ctrl = gen_subtype_check(original, klass_node); 3474 3475 if (not_subtype_ctrl != top()) { 3476 PreserveJVMState pjvms(this); 3477 set_control(not_subtype_ctrl); 3478 uncommon_trap(Deoptimization::Reason_class_check, 3479 Deoptimization::Action_make_not_entrant); 3480 assert(stopped(), "Should be stopped"); 3481 } 3482 validated = true; 3483 } 3484 3485 if (!stopped()) { 3486 newcopy = new_array(klass_node, length, 0); // no arguments to push 3487 3488 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, 3489 load_object_klass(original), klass_node); 3490 if (!is_copyOfRange) { 3491 ac->set_copyof(validated); 3492 } else { 3493 ac->set_copyofrange(validated); 3494 } 3495 Node* n = _gvn.transform(ac); 3496 if (n == ac) { 3497 ac->connect_outputs(this); 3498 } else { 3499 assert(validated, "shouldn't transform if all arguments not validated"); 3500 set_all_memory(n); 3501 } 3502 } 3503 } 3504 } // original reexecute is set back here 3505 3506 C->set_has_split_ifs(true); // Has chance for split-if optimization 3507 if (!stopped()) { 3508 set_result(newcopy); 3509 } 3510 return true; 3511 } 3512 3513 3514 //----------------------generate_virtual_guard--------------------------- 3515 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3516 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3517 RegionNode* slow_region) { 3518 ciMethod* method = callee(); 3519 int vtable_index = method->vtable_index(); 3520 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3521 "bad index %d", vtable_index); 3522 // Get the Method* out of the appropriate vtable entry. 3523 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 3524 vtable_index*vtableEntry::size_in_bytes() + 3525 vtableEntry::method_offset_in_bytes(); 3526 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3527 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3528 3529 // Compare the target method with the expected method (e.g., Object.hashCode). 3530 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3531 3532 Node* native_call = makecon(native_call_addr); 3533 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3534 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3535 3536 return generate_slow_guard(test_native, slow_region); 3537 } 3538 3539 //-----------------------generate_method_call---------------------------- 3540 // Use generate_method_call to make a slow-call to the real 3541 // method if the fast path fails. An alternative would be to 3542 // use a stub like OptoRuntime::slow_arraycopy_Java. 3543 // This only works for expanding the current library call, 3544 // not another intrinsic. (E.g., don't use this for making an 3545 // arraycopy call inside of the copyOf intrinsic.) 3546 CallJavaNode* 3547 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3548 // When compiling the intrinsic method itself, do not use this technique. 3549 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3550 3551 ciMethod* method = callee(); 3552 // ensure the JVMS we have will be correct for this call 3553 guarantee(method_id == method->intrinsic_id(), "must match"); 3554 3555 const TypeFunc* tf = TypeFunc::make(method); 3556 CallJavaNode* slow_call; 3557 if (is_static) { 3558 assert(!is_virtual, ""); 3559 slow_call = new CallStaticJavaNode(C, tf, 3560 SharedRuntime::get_resolve_static_call_stub(), 3561 method, bci()); 3562 } else if (is_virtual) { 3563 null_check_receiver(); 3564 int vtable_index = Method::invalid_vtable_index; 3565 if (UseInlineCaches) { 3566 // Suppress the vtable call 3567 } else { 3568 // hashCode and clone are not a miranda methods, 3569 // so the vtable index is fixed. 3570 // No need to use the linkResolver to get it. 3571 vtable_index = method->vtable_index(); 3572 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3573 "bad index %d", vtable_index); 3574 } 3575 slow_call = new CallDynamicJavaNode(tf, 3576 SharedRuntime::get_resolve_virtual_call_stub(), 3577 method, vtable_index, bci()); 3578 } else { // neither virtual nor static: opt_virtual 3579 null_check_receiver(); 3580 slow_call = new CallStaticJavaNode(C, tf, 3581 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3582 method, bci()); 3583 slow_call->set_optimized_virtual(true); 3584 } 3585 if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) { 3586 // To be able to issue a direct call (optimized virtual or virtual) 3587 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information 3588 // about the method being invoked should be attached to the call site to 3589 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C). 3590 slow_call->set_override_symbolic_info(true); 3591 } 3592 set_arguments_for_java_call(slow_call); 3593 set_edges_for_java_call(slow_call); 3594 return slow_call; 3595 } 3596 3597 3598 /** 3599 * Build special case code for calls to hashCode on an object. This call may 3600 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 3601 * slightly different code. 3602 */ 3603 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3604 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3605 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3606 3607 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3608 3609 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3610 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 3611 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3612 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3613 Node* obj = NULL; 3614 if (!is_static) { 3615 // Check for hashing null object 3616 obj = null_check_receiver(); 3617 if (stopped()) return true; // unconditionally null 3618 result_reg->init_req(_null_path, top()); 3619 result_val->init_req(_null_path, top()); 3620 } else { 3621 // Do a null check, and return zero if null. 3622 // System.identityHashCode(null) == 0 3623 obj = argument(0); 3624 Node* null_ctl = top(); 3625 obj = null_check_oop(obj, &null_ctl); 3626 result_reg->init_req(_null_path, null_ctl); 3627 result_val->init_req(_null_path, _gvn.intcon(0)); 3628 } 3629 3630 // Unconditionally null? Then return right away. 3631 if (stopped()) { 3632 set_control( result_reg->in(_null_path)); 3633 if (!stopped()) 3634 set_result(result_val->in(_null_path)); 3635 return true; 3636 } 3637 3638 // We only go to the fast case code if we pass a number of guards. The 3639 // paths which do not pass are accumulated in the slow_region. 3640 RegionNode* slow_region = new RegionNode(1); 3641 record_for_igvn(slow_region); 3642 3643 // If this is a virtual call, we generate a funny guard. We pull out 3644 // the vtable entry corresponding to hashCode() from the target object. 3645 // If the target method which we are calling happens to be the native 3646 // Object hashCode() method, we pass the guard. We do not need this 3647 // guard for non-virtual calls -- the caller is known to be the native 3648 // Object hashCode(). 3649 if (is_virtual) { 3650 // After null check, get the object's klass. 3651 Node* obj_klass = load_object_klass(obj); 3652 generate_virtual_guard(obj_klass, slow_region); 3653 } 3654 3655 // Get the header out of the object, use LoadMarkNode when available 3656 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3657 // The control of the load must be NULL. Otherwise, the load can move before 3658 // the null check after castPP removal. 3659 Node* no_ctrl = NULL; 3660 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3661 3662 // Test the header to see if it is unlocked. 3663 Node *lock_mask = _gvn.MakeConX(markWord::biased_lock_mask_in_place); 3664 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 3665 Node *unlocked_val = _gvn.MakeConX(markWord::unlocked_value); 3666 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 3667 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 3668 3669 generate_slow_guard(test_unlocked, slow_region); 3670 3671 // Get the hash value and check to see that it has been properly assigned. 3672 // We depend on hash_mask being at most 32 bits and avoid the use of 3673 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3674 // vm: see markWord.hpp. 3675 Node *hash_mask = _gvn.intcon(markWord::hash_mask); 3676 Node *hash_shift = _gvn.intcon(markWord::hash_shift); 3677 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 3678 // This hack lets the hash bits live anywhere in the mark object now, as long 3679 // as the shift drops the relevant bits into the low 32 bits. Note that 3680 // Java spec says that HashCode is an int so there's no point in capturing 3681 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3682 hshifted_header = ConvX2I(hshifted_header); 3683 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 3684 3685 Node *no_hash_val = _gvn.intcon(markWord::no_hash); 3686 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 3687 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 3688 3689 generate_slow_guard(test_assigned, slow_region); 3690 3691 Node* init_mem = reset_memory(); 3692 // fill in the rest of the null path: 3693 result_io ->init_req(_null_path, i_o()); 3694 result_mem->init_req(_null_path, init_mem); 3695 3696 result_val->init_req(_fast_path, hash_val); 3697 result_reg->init_req(_fast_path, control()); 3698 result_io ->init_req(_fast_path, i_o()); 3699 result_mem->init_req(_fast_path, init_mem); 3700 3701 // Generate code for the slow case. We make a call to hashCode(). 3702 set_control(_gvn.transform(slow_region)); 3703 if (!stopped()) { 3704 // No need for PreserveJVMState, because we're using up the present state. 3705 set_all_memory(init_mem); 3706 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 3707 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3708 Node* slow_result = set_results_for_java_call(slow_call); 3709 // this->control() comes from set_results_for_java_call 3710 result_reg->init_req(_slow_path, control()); 3711 result_val->init_req(_slow_path, slow_result); 3712 result_io ->set_req(_slow_path, i_o()); 3713 result_mem ->set_req(_slow_path, reset_memory()); 3714 } 3715 3716 // Return the combined state. 3717 set_i_o( _gvn.transform(result_io) ); 3718 set_all_memory( _gvn.transform(result_mem)); 3719 3720 set_result(result_reg, result_val); 3721 return true; 3722 } 3723 3724 //---------------------------inline_native_getClass---------------------------- 3725 // public final native Class<?> java.lang.Object.getClass(); 3726 // 3727 // Build special case code for calls to getClass on an object. 3728 bool LibraryCallKit::inline_native_getClass() { 3729 Node* obj = null_check_receiver(); 3730 if (stopped()) return true; 3731 set_result(load_mirror_from_klass(load_object_klass(obj))); 3732 return true; 3733 } 3734 3735 //-----------------inline_native_Reflection_getCallerClass--------------------- 3736 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 3737 // 3738 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3739 // 3740 // NOTE: This code must perform the same logic as JVM_GetCallerClass 3741 // in that it must skip particular security frames and checks for 3742 // caller sensitive methods. 3743 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3744 #ifndef PRODUCT 3745 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 3746 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3747 } 3748 #endif 3749 3750 if (!jvms()->has_method()) { 3751 #ifndef PRODUCT 3752 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 3753 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3754 } 3755 #endif 3756 return false; 3757 } 3758 3759 // Walk back up the JVM state to find the caller at the required 3760 // depth. 3761 JVMState* caller_jvms = jvms(); 3762 3763 // Cf. JVM_GetCallerClass 3764 // NOTE: Start the loop at depth 1 because the current JVM state does 3765 // not include the Reflection.getCallerClass() frame. 3766 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 3767 ciMethod* m = caller_jvms->method(); 3768 switch (n) { 3769 case 0: 3770 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 3771 break; 3772 case 1: 3773 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 3774 if (!m->caller_sensitive()) { 3775 #ifndef PRODUCT 3776 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 3777 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 3778 } 3779 #endif 3780 return false; // bail-out; let JVM_GetCallerClass do the work 3781 } 3782 break; 3783 default: 3784 if (!m->is_ignored_by_security_stack_walk()) { 3785 // We have reached the desired frame; return the holder class. 3786 // Acquire method holder as java.lang.Class and push as constant. 3787 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3788 ciInstance* caller_mirror = caller_klass->java_mirror(); 3789 set_result(makecon(TypeInstPtr::make(caller_mirror))); 3790 3791 #ifndef PRODUCT 3792 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 3793 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()); 3794 tty->print_cr(" JVM state at this point:"); 3795 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 3796 ciMethod* m = jvms()->of_depth(i)->method(); 3797 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 3798 } 3799 } 3800 #endif 3801 return true; 3802 } 3803 break; 3804 } 3805 } 3806 3807 #ifndef PRODUCT 3808 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 3809 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 3810 tty->print_cr(" JVM state at this point:"); 3811 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 3812 ciMethod* m = jvms()->of_depth(i)->method(); 3813 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 3814 } 3815 } 3816 #endif 3817 3818 return false; // bail-out; let JVM_GetCallerClass do the work 3819 } 3820 3821 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 3822 Node* arg = argument(0); 3823 Node* result = NULL; 3824 3825 switch (id) { 3826 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 3827 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 3828 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 3829 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 3830 3831 case vmIntrinsics::_doubleToLongBits: { 3832 // two paths (plus control) merge in a wood 3833 RegionNode *r = new RegionNode(3); 3834 Node *phi = new PhiNode(r, TypeLong::LONG); 3835 3836 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 3837 // Build the boolean node 3838 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 3839 3840 // Branch either way. 3841 // NaN case is less traveled, which makes all the difference. 3842 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3843 Node *opt_isnan = _gvn.transform(ifisnan); 3844 assert( opt_isnan->is_If(), "Expect an IfNode"); 3845 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3846 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 3847 3848 set_control(iftrue); 3849 3850 static const jlong nan_bits = CONST64(0x7ff8000000000000); 3851 Node *slow_result = longcon(nan_bits); // return NaN 3852 phi->init_req(1, _gvn.transform( slow_result )); 3853 r->init_req(1, iftrue); 3854 3855 // Else fall through 3856 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 3857 set_control(iffalse); 3858 3859 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 3860 r->init_req(2, iffalse); 3861 3862 // Post merge 3863 set_control(_gvn.transform(r)); 3864 record_for_igvn(r); 3865 3866 C->set_has_split_ifs(true); // Has chance for split-if optimization 3867 result = phi; 3868 assert(result->bottom_type()->isa_long(), "must be"); 3869 break; 3870 } 3871 3872 case vmIntrinsics::_floatToIntBits: { 3873 // two paths (plus control) merge in a wood 3874 RegionNode *r = new RegionNode(3); 3875 Node *phi = new PhiNode(r, TypeInt::INT); 3876 3877 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 3878 // Build the boolean node 3879 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 3880 3881 // Branch either way. 3882 // NaN case is less traveled, which makes all the difference. 3883 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3884 Node *opt_isnan = _gvn.transform(ifisnan); 3885 assert( opt_isnan->is_If(), "Expect an IfNode"); 3886 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3887 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 3888 3889 set_control(iftrue); 3890 3891 static const jint nan_bits = 0x7fc00000; 3892 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 3893 phi->init_req(1, _gvn.transform( slow_result )); 3894 r->init_req(1, iftrue); 3895 3896 // Else fall through 3897 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 3898 set_control(iffalse); 3899 3900 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 3901 r->init_req(2, iffalse); 3902 3903 // Post merge 3904 set_control(_gvn.transform(r)); 3905 record_for_igvn(r); 3906 3907 C->set_has_split_ifs(true); // Has chance for split-if optimization 3908 result = phi; 3909 assert(result->bottom_type()->isa_int(), "must be"); 3910 break; 3911 } 3912 3913 default: 3914 fatal_unexpected_iid(id); 3915 break; 3916 } 3917 set_result(_gvn.transform(result)); 3918 return true; 3919 } 3920 3921 //----------------------inline_unsafe_copyMemory------------------------- 3922 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 3923 bool LibraryCallKit::inline_unsafe_copyMemory() { 3924 if (callee()->is_static()) return false; // caller must have the capability! 3925 null_check_receiver(); // null-check receiver 3926 if (stopped()) return true; 3927 3928 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3929 3930 Node* src_ptr = argument(1); // type: oop 3931 Node* src_off = ConvL2X(argument(2)); // type: long 3932 Node* dst_ptr = argument(4); // type: oop 3933 Node* dst_off = ConvL2X(argument(5)); // type: long 3934 Node* size = ConvL2X(argument(7)); // type: long 3935 3936 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 3937 "fieldOffset must be byte-scaled"); 3938 3939 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ); 3940 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE); 3941 3942 // Conservatively insert a memory barrier on all memory slices. 3943 // Do not let writes of the copy source or destination float below the copy. 3944 insert_mem_bar(Op_MemBarCPUOrder); 3945 3946 Node* thread = _gvn.transform(new ThreadLocalNode()); 3947 Node* doing_unsafe_access_addr = basic_plus_adr(top(), thread, in_bytes(JavaThread::doing_unsafe_access_offset())); 3948 BasicType doing_unsafe_access_bt = T_BYTE; 3949 assert((sizeof(bool) * CHAR_BIT) == 8, "not implemented"); 3950 3951 // update volatile field 3952 store_to_memory(control(), doing_unsafe_access_addr, intcon(1), doing_unsafe_access_bt, Compile::AliasIdxRaw, MemNode::unordered); 3953 3954 // Call it. Note that the length argument is not scaled. 3955 make_runtime_call(RC_LEAF|RC_NO_FP, 3956 OptoRuntime::fast_arraycopy_Type(), 3957 StubRoutines::unsafe_arraycopy(), 3958 "unsafe_arraycopy", 3959 TypeRawPtr::BOTTOM, 3960 src, dst, size XTOP); 3961 3962 store_to_memory(control(), doing_unsafe_access_addr, intcon(0), doing_unsafe_access_bt, Compile::AliasIdxRaw, MemNode::unordered); 3963 3964 // Do not let reads of the copy destination float above the copy. 3965 insert_mem_bar(Op_MemBarCPUOrder); 3966 3967 return true; 3968 } 3969 3970 //------------------------clone_coping----------------------------------- 3971 // Helper function for inline_native_clone. 3972 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) { 3973 assert(obj_size != NULL, ""); 3974 Node* raw_obj = alloc_obj->in(1); 3975 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 3976 3977 AllocateNode* alloc = NULL; 3978 if (ReduceBulkZeroing) { 3979 // We will be completely responsible for initializing this object - 3980 // mark Initialize node as complete. 3981 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 3982 // The object was just allocated - there should be no any stores! 3983 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 3984 // Mark as complete_with_arraycopy so that on AllocateNode 3985 // expansion, we know this AllocateNode is initialized by an array 3986 // copy and a StoreStore barrier exists after the array copy. 3987 alloc->initialization()->set_complete_with_arraycopy(); 3988 } 3989 3990 Node* size = _gvn.transform(obj_size); 3991 access_clone(obj, alloc_obj, size, is_array); 3992 3993 // Do not let reads from the cloned object float above the arraycopy. 3994 if (alloc != NULL) { 3995 // Do not let stores that initialize this object be reordered with 3996 // a subsequent store that would make this object accessible by 3997 // other threads. 3998 // Record what AllocateNode this StoreStore protects so that 3999 // escape analysis can go from the MemBarStoreStoreNode to the 4000 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4001 // based on the escape status of the AllocateNode. 4002 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 4003 } else { 4004 insert_mem_bar(Op_MemBarCPUOrder); 4005 } 4006 } 4007 4008 //------------------------inline_native_clone---------------------------- 4009 // protected native Object java.lang.Object.clone(); 4010 // 4011 // Here are the simple edge cases: 4012 // null receiver => normal trap 4013 // virtual and clone was overridden => slow path to out-of-line clone 4014 // not cloneable or finalizer => slow path to out-of-line Object.clone 4015 // 4016 // The general case has two steps, allocation and copying. 4017 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4018 // 4019 // Copying also has two cases, oop arrays and everything else. 4020 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4021 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4022 // 4023 // These steps fold up nicely if and when the cloned object's klass 4024 // can be sharply typed as an object array, a type array, or an instance. 4025 // 4026 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4027 PhiNode* result_val; 4028 4029 // Set the reexecute bit for the interpreter to reexecute 4030 // the bytecode that invokes Object.clone if deoptimization happens. 4031 { PreserveReexecuteState preexecs(this); 4032 jvms()->set_should_reexecute(true); 4033 4034 Node* obj = null_check_receiver(); 4035 if (stopped()) return true; 4036 4037 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4038 4039 // If we are going to clone an instance, we need its exact type to 4040 // know the number and types of fields to convert the clone to 4041 // loads/stores. Maybe a speculative type can help us. 4042 if (!obj_type->klass_is_exact() && 4043 obj_type->speculative_type() != NULL && 4044 obj_type->speculative_type()->is_instance_klass()) { 4045 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4046 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4047 !spec_ik->has_injected_fields()) { 4048 ciKlass* k = obj_type->klass(); 4049 if (!k->is_instance_klass() || 4050 k->as_instance_klass()->is_interface() || 4051 k->as_instance_klass()->has_subklass()) { 4052 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4053 } 4054 } 4055 } 4056 4057 // Conservatively insert a memory barrier on all memory slices. 4058 // Do not let writes into the original float below the clone. 4059 insert_mem_bar(Op_MemBarCPUOrder); 4060 4061 // paths into result_reg: 4062 enum { 4063 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4064 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4065 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4066 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4067 PATH_LIMIT 4068 }; 4069 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4070 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4071 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4072 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4073 record_for_igvn(result_reg); 4074 4075 Node* obj_klass = load_object_klass(obj); 4076 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4077 if (array_ctl != NULL) { 4078 // It's an array. 4079 PreserveJVMState pjvms(this); 4080 set_control(array_ctl); 4081 Node* obj_length = load_array_length(obj); 4082 Node* obj_size = NULL; 4083 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size, /*deoptimize_on_exception=*/true); 4084 4085 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4086 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { 4087 // If it is an oop array, it requires very special treatment, 4088 // because gc barriers are required when accessing the array. 4089 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4090 if (is_obja != NULL) { 4091 PreserveJVMState pjvms2(this); 4092 set_control(is_obja); 4093 // Generate a direct call to the right arraycopy function(s). 4094 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4095 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); 4096 ac->set_clone_oop_array(); 4097 Node* n = _gvn.transform(ac); 4098 assert(n == ac, "cannot disappear"); 4099 ac->connect_outputs(this, /*deoptimize_on_exception=*/true); 4100 4101 result_reg->init_req(_objArray_path, control()); 4102 result_val->init_req(_objArray_path, alloc_obj); 4103 result_i_o ->set_req(_objArray_path, i_o()); 4104 result_mem ->set_req(_objArray_path, reset_memory()); 4105 } 4106 } 4107 // Otherwise, there are no barriers to worry about. 4108 // (We can dispense with card marks if we know the allocation 4109 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4110 // causes the non-eden paths to take compensating steps to 4111 // simulate a fresh allocation, so that no further 4112 // card marks are required in compiled code to initialize 4113 // the object.) 4114 4115 if (!stopped()) { 4116 copy_to_clone(obj, alloc_obj, obj_size, true); 4117 4118 // Present the results of the copy. 4119 result_reg->init_req(_array_path, control()); 4120 result_val->init_req(_array_path, alloc_obj); 4121 result_i_o ->set_req(_array_path, i_o()); 4122 result_mem ->set_req(_array_path, reset_memory()); 4123 } 4124 } 4125 4126 // We only go to the instance fast case code if we pass a number of guards. 4127 // The paths which do not pass are accumulated in the slow_region. 4128 RegionNode* slow_region = new RegionNode(1); 4129 record_for_igvn(slow_region); 4130 if (!stopped()) { 4131 // It's an instance (we did array above). Make the slow-path tests. 4132 // If this is a virtual call, we generate a funny guard. We grab 4133 // the vtable entry corresponding to clone() from the target object. 4134 // If the target method which we are calling happens to be the 4135 // Object clone() method, we pass the guard. We do not need this 4136 // guard for non-virtual calls; the caller is known to be the native 4137 // Object clone(). 4138 if (is_virtual) { 4139 generate_virtual_guard(obj_klass, slow_region); 4140 } 4141 4142 // The object must be easily cloneable and must not have a finalizer. 4143 // Both of these conditions may be checked in a single test. 4144 // We could optimize the test further, but we don't care. 4145 generate_access_flags_guard(obj_klass, 4146 // Test both conditions: 4147 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4148 // Must be cloneable but not finalizer: 4149 JVM_ACC_IS_CLONEABLE_FAST, 4150 slow_region); 4151 } 4152 4153 if (!stopped()) { 4154 // It's an instance, and it passed the slow-path tests. 4155 PreserveJVMState pjvms(this); 4156 Node* obj_size = NULL; 4157 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4158 // is reexecuted if deoptimization occurs and there could be problems when merging 4159 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4160 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4161 4162 copy_to_clone(obj, alloc_obj, obj_size, false); 4163 4164 // Present the results of the slow call. 4165 result_reg->init_req(_instance_path, control()); 4166 result_val->init_req(_instance_path, alloc_obj); 4167 result_i_o ->set_req(_instance_path, i_o()); 4168 result_mem ->set_req(_instance_path, reset_memory()); 4169 } 4170 4171 // Generate code for the slow case. We make a call to clone(). 4172 set_control(_gvn.transform(slow_region)); 4173 if (!stopped()) { 4174 PreserveJVMState pjvms(this); 4175 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4176 // We need to deoptimize on exception (see comment above) 4177 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true); 4178 // this->control() comes from set_results_for_java_call 4179 result_reg->init_req(_slow_path, control()); 4180 result_val->init_req(_slow_path, slow_result); 4181 result_i_o ->set_req(_slow_path, i_o()); 4182 result_mem ->set_req(_slow_path, reset_memory()); 4183 } 4184 4185 // Return the combined state. 4186 set_control( _gvn.transform(result_reg)); 4187 set_i_o( _gvn.transform(result_i_o)); 4188 set_all_memory( _gvn.transform(result_mem)); 4189 } // original reexecute is set back here 4190 4191 set_result(_gvn.transform(result_val)); 4192 return true; 4193 } 4194 4195 // If we have a tightly coupled allocation, the arraycopy may take care 4196 // of the array initialization. If one of the guards we insert between 4197 // the allocation and the arraycopy causes a deoptimization, an 4198 // unitialized array will escape the compiled method. To prevent that 4199 // we set the JVM state for uncommon traps between the allocation and 4200 // the arraycopy to the state before the allocation so, in case of 4201 // deoptimization, we'll reexecute the allocation and the 4202 // initialization. 4203 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4204 if (alloc != NULL) { 4205 ciMethod* trap_method = alloc->jvms()->method(); 4206 int trap_bci = alloc->jvms()->bci(); 4207 4208 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4209 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4210 // Make sure there's no store between the allocation and the 4211 // arraycopy otherwise visible side effects could be rexecuted 4212 // in case of deoptimization and cause incorrect execution. 4213 bool no_interfering_store = true; 4214 Node* mem = alloc->in(TypeFunc::Memory); 4215 if (mem->is_MergeMem()) { 4216 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4217 Node* n = mms.memory(); 4218 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4219 assert(n->is_Store(), "what else?"); 4220 no_interfering_store = false; 4221 break; 4222 } 4223 } 4224 } else { 4225 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4226 Node* n = mms.memory(); 4227 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4228 assert(n->is_Store(), "what else?"); 4229 no_interfering_store = false; 4230 break; 4231 } 4232 } 4233 } 4234 4235 if (no_interfering_store) { 4236 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4237 uint size = alloc->req(); 4238 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4239 old_jvms->set_map(sfpt); 4240 for (uint i = 0; i < size; i++) { 4241 sfpt->init_req(i, alloc->in(i)); 4242 } 4243 // re-push array length for deoptimization 4244 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4245 old_jvms->set_sp(old_jvms->sp()+1); 4246 old_jvms->set_monoff(old_jvms->monoff()+1); 4247 old_jvms->set_scloff(old_jvms->scloff()+1); 4248 old_jvms->set_endoff(old_jvms->endoff()+1); 4249 old_jvms->set_should_reexecute(true); 4250 4251 sfpt->set_i_o(map()->i_o()); 4252 sfpt->set_memory(map()->memory()); 4253 sfpt->set_control(map()->control()); 4254 4255 JVMState* saved_jvms = jvms(); 4256 saved_reexecute_sp = _reexecute_sp; 4257 4258 set_jvms(sfpt->jvms()); 4259 _reexecute_sp = jvms()->sp(); 4260 4261 return saved_jvms; 4262 } 4263 } 4264 } 4265 return NULL; 4266 } 4267 4268 // In case of a deoptimization, we restart execution at the 4269 // allocation, allocating a new array. We would leave an uninitialized 4270 // array in the heap that GCs wouldn't expect. Move the allocation 4271 // after the traps so we don't allocate the array if we 4272 // deoptimize. This is possible because tightly_coupled_allocation() 4273 // guarantees there's no observer of the allocated array at this point 4274 // and the control flow is simple enough. 4275 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, 4276 int saved_reexecute_sp, uint new_idx) { 4277 if (saved_jvms != NULL && !stopped()) { 4278 assert(alloc != NULL, "only with a tightly coupled allocation"); 4279 // restore JVM state to the state at the arraycopy 4280 saved_jvms->map()->set_control(map()->control()); 4281 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4282 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4283 // If we've improved the types of some nodes (null check) while 4284 // emitting the guards, propagate them to the current state 4285 map()->replaced_nodes().apply(saved_jvms->map(), new_idx); 4286 set_jvms(saved_jvms); 4287 _reexecute_sp = saved_reexecute_sp; 4288 4289 // Remove the allocation from above the guards 4290 CallProjections callprojs; 4291 alloc->extract_projections(&callprojs, true); 4292 InitializeNode* init = alloc->initialization(); 4293 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4294 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4295 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4296 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4297 4298 // move the allocation here (after the guards) 4299 _gvn.hash_delete(alloc); 4300 alloc->set_req(TypeFunc::Control, control()); 4301 alloc->set_req(TypeFunc::I_O, i_o()); 4302 Node *mem = reset_memory(); 4303 set_all_memory(mem); 4304 alloc->set_req(TypeFunc::Memory, mem); 4305 set_control(init->proj_out_or_null(TypeFunc::Control)); 4306 set_i_o(callprojs.fallthrough_ioproj); 4307 4308 // Update memory as done in GraphKit::set_output_for_allocation() 4309 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4310 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4311 if (ary_type->isa_aryptr() && length_type != NULL) { 4312 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4313 } 4314 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4315 int elemidx = C->get_alias_index(telemref); 4316 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); 4317 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); 4318 4319 Node* allocx = _gvn.transform(alloc); 4320 assert(allocx == alloc, "where has the allocation gone?"); 4321 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4322 4323 _gvn.hash_delete(dest); 4324 dest->set_req(0, control()); 4325 Node* destx = _gvn.transform(dest); 4326 assert(destx == dest, "where has the allocation result gone?"); 4327 } 4328 } 4329 4330 4331 //------------------------------inline_arraycopy----------------------- 4332 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4333 // Object dest, int destPos, 4334 // int length); 4335 bool LibraryCallKit::inline_arraycopy() { 4336 // Get the arguments. 4337 Node* src = argument(0); // type: oop 4338 Node* src_offset = argument(1); // type: int 4339 Node* dest = argument(2); // type: oop 4340 Node* dest_offset = argument(3); // type: int 4341 Node* length = argument(4); // type: int 4342 4343 uint new_idx = C->unique(); 4344 4345 // Check for allocation before we add nodes that would confuse 4346 // tightly_coupled_allocation() 4347 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4348 4349 int saved_reexecute_sp = -1; 4350 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4351 // See arraycopy_restore_alloc_state() comment 4352 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4353 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4354 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 4355 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4356 4357 // The following tests must be performed 4358 // (1) src and dest are arrays. 4359 // (2) src and dest arrays must have elements of the same BasicType 4360 // (3) src and dest must not be null. 4361 // (4) src_offset must not be negative. 4362 // (5) dest_offset must not be negative. 4363 // (6) length must not be negative. 4364 // (7) src_offset + length must not exceed length of src. 4365 // (8) dest_offset + length must not exceed length of dest. 4366 // (9) each element of an oop array must be assignable 4367 4368 // (3) src and dest must not be null. 4369 // always do this here because we need the JVM state for uncommon traps 4370 Node* null_ctl = top(); 4371 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4372 assert(null_ctl->is_top(), "no null control here"); 4373 dest = null_check(dest, T_ARRAY); 4374 4375 if (!can_emit_guards) { 4376 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4377 // guards but the arraycopy node could still take advantage of a 4378 // tightly allocated allocation. tightly_coupled_allocation() is 4379 // called again to make sure it takes the null check above into 4380 // account: the null check is mandatory and if it caused an 4381 // uncommon trap to be emitted then the allocation can't be 4382 // considered tightly coupled in this context. 4383 alloc = tightly_coupled_allocation(dest, NULL); 4384 } 4385 4386 bool validated = false; 4387 4388 const Type* src_type = _gvn.type(src); 4389 const Type* dest_type = _gvn.type(dest); 4390 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4391 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4392 4393 // Do we have the type of src? 4394 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4395 // Do we have the type of dest? 4396 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4397 // Is the type for src from speculation? 4398 bool src_spec = false; 4399 // Is the type for dest from speculation? 4400 bool dest_spec = false; 4401 4402 if ((!has_src || !has_dest) && can_emit_guards) { 4403 // We don't have sufficient type information, let's see if 4404 // speculative types can help. We need to have types for both src 4405 // and dest so that it pays off. 4406 4407 // Do we already have or could we have type information for src 4408 bool could_have_src = has_src; 4409 // Do we already have or could we have type information for dest 4410 bool could_have_dest = has_dest; 4411 4412 ciKlass* src_k = NULL; 4413 if (!has_src) { 4414 src_k = src_type->speculative_type_not_null(); 4415 if (src_k != NULL && src_k->is_array_klass()) { 4416 could_have_src = true; 4417 } 4418 } 4419 4420 ciKlass* dest_k = NULL; 4421 if (!has_dest) { 4422 dest_k = dest_type->speculative_type_not_null(); 4423 if (dest_k != NULL && dest_k->is_array_klass()) { 4424 could_have_dest = true; 4425 } 4426 } 4427 4428 if (could_have_src && could_have_dest) { 4429 // This is going to pay off so emit the required guards 4430 if (!has_src) { 4431 src = maybe_cast_profiled_obj(src, src_k, true); 4432 src_type = _gvn.type(src); 4433 top_src = src_type->isa_aryptr(); 4434 has_src = (top_src != NULL && top_src->klass() != NULL); 4435 src_spec = true; 4436 } 4437 if (!has_dest) { 4438 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4439 dest_type = _gvn.type(dest); 4440 top_dest = dest_type->isa_aryptr(); 4441 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4442 dest_spec = true; 4443 } 4444 } 4445 } 4446 4447 if (has_src && has_dest && can_emit_guards) { 4448 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4449 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4450 if (is_reference_type(src_elem)) src_elem = T_OBJECT; 4451 if (is_reference_type(dest_elem)) dest_elem = T_OBJECT; 4452 4453 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4454 // If both arrays are object arrays then having the exact types 4455 // for both will remove the need for a subtype check at runtime 4456 // before the call and may make it possible to pick a faster copy 4457 // routine (without a subtype check on every element) 4458 // Do we have the exact type of src? 4459 bool could_have_src = src_spec; 4460 // Do we have the exact type of dest? 4461 bool could_have_dest = dest_spec; 4462 ciKlass* src_k = top_src->klass(); 4463 ciKlass* dest_k = top_dest->klass(); 4464 if (!src_spec) { 4465 src_k = src_type->speculative_type_not_null(); 4466 if (src_k != NULL && src_k->is_array_klass()) { 4467 could_have_src = true; 4468 } 4469 } 4470 if (!dest_spec) { 4471 dest_k = dest_type->speculative_type_not_null(); 4472 if (dest_k != NULL && dest_k->is_array_klass()) { 4473 could_have_dest = true; 4474 } 4475 } 4476 if (could_have_src && could_have_dest) { 4477 // If we can have both exact types, emit the missing guards 4478 if (could_have_src && !src_spec) { 4479 src = maybe_cast_profiled_obj(src, src_k, true); 4480 } 4481 if (could_have_dest && !dest_spec) { 4482 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4483 } 4484 } 4485 } 4486 } 4487 4488 ciMethod* trap_method = method(); 4489 int trap_bci = bci(); 4490 if (saved_jvms != NULL) { 4491 trap_method = alloc->jvms()->method(); 4492 trap_bci = alloc->jvms()->bci(); 4493 } 4494 4495 bool negative_length_guard_generated = false; 4496 4497 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4498 can_emit_guards && 4499 !src->is_top() && !dest->is_top()) { 4500 // validate arguments: enables transformation the ArrayCopyNode 4501 validated = true; 4502 4503 RegionNode* slow_region = new RegionNode(1); 4504 record_for_igvn(slow_region); 4505 4506 // (1) src and dest are arrays. 4507 generate_non_array_guard(load_object_klass(src), slow_region); 4508 generate_non_array_guard(load_object_klass(dest), slow_region); 4509 4510 // (2) src and dest arrays must have elements of the same BasicType 4511 // done at macro expansion or at Ideal transformation time 4512 4513 // (4) src_offset must not be negative. 4514 generate_negative_guard(src_offset, slow_region); 4515 4516 // (5) dest_offset must not be negative. 4517 generate_negative_guard(dest_offset, slow_region); 4518 4519 // (7) src_offset + length must not exceed length of src. 4520 generate_limit_guard(src_offset, length, 4521 load_array_length(src), 4522 slow_region); 4523 4524 // (8) dest_offset + length must not exceed length of dest. 4525 generate_limit_guard(dest_offset, length, 4526 load_array_length(dest), 4527 slow_region); 4528 4529 // (6) length must not be negative. 4530 // This is also checked in generate_arraycopy() during macro expansion, but 4531 // we also have to check it here for the case where the ArrayCopyNode will 4532 // be eliminated by Escape Analysis. 4533 if (EliminateAllocations) { 4534 generate_negative_guard(length, slow_region); 4535 negative_length_guard_generated = true; 4536 } 4537 4538 // (9) each element of an oop array must be assignable 4539 Node* dest_klass = load_object_klass(dest); 4540 if (src != dest) { 4541 Node* not_subtype_ctrl = gen_subtype_check(src, dest_klass); 4542 4543 if (not_subtype_ctrl != top()) { 4544 PreserveJVMState pjvms(this); 4545 set_control(not_subtype_ctrl); 4546 uncommon_trap(Deoptimization::Reason_intrinsic, 4547 Deoptimization::Action_make_not_entrant); 4548 assert(stopped(), "Should be stopped"); 4549 } 4550 } 4551 { 4552 PreserveJVMState pjvms(this); 4553 set_control(_gvn.transform(slow_region)); 4554 uncommon_trap(Deoptimization::Reason_intrinsic, 4555 Deoptimization::Action_make_not_entrant); 4556 assert(stopped(), "Should be stopped"); 4557 } 4558 4559 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); 4560 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); 4561 src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); 4562 } 4563 4564 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); 4565 4566 if (stopped()) { 4567 return true; 4568 } 4569 4570 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, 4571 // Create LoadRange and LoadKlass nodes for use during macro expansion here 4572 // so the compiler has a chance to eliminate them: during macro expansion, 4573 // we have to set their control (CastPP nodes are eliminated). 4574 load_object_klass(src), load_object_klass(dest), 4575 load_array_length(src), load_array_length(dest)); 4576 4577 ac->set_arraycopy(validated); 4578 4579 Node* n = _gvn.transform(ac); 4580 if (n == ac) { 4581 ac->connect_outputs(this); 4582 } else { 4583 assert(validated, "shouldn't transform if all arguments not validated"); 4584 set_all_memory(n); 4585 } 4586 clear_upper_avx(); 4587 4588 4589 return true; 4590 } 4591 4592 4593 // Helper function which determines if an arraycopy immediately follows 4594 // an allocation, with no intervening tests or other escapes for the object. 4595 AllocateArrayNode* 4596 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4597 RegionNode* slow_region) { 4598 if (stopped()) return NULL; // no fast path 4599 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4600 4601 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4602 if (alloc == NULL) return NULL; 4603 4604 Node* rawmem = memory(Compile::AliasIdxRaw); 4605 // Is the allocation's memory state untouched? 4606 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4607 // Bail out if there have been raw-memory effects since the allocation. 4608 // (Example: There might have been a call or safepoint.) 4609 return NULL; 4610 } 4611 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4612 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4613 return NULL; 4614 } 4615 4616 // There must be no unexpected observers of this allocation. 4617 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4618 Node* obs = ptr->fast_out(i); 4619 if (obs != this->map()) { 4620 return NULL; 4621 } 4622 } 4623 4624 // This arraycopy must unconditionally follow the allocation of the ptr. 4625 Node* alloc_ctl = ptr->in(0); 4626 Node* ctl = control(); 4627 while (ctl != alloc_ctl) { 4628 // There may be guards which feed into the slow_region. 4629 // Any other control flow means that we might not get a chance 4630 // to finish initializing the allocated object. 4631 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4632 IfNode* iff = ctl->in(0)->as_If(); 4633 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); 4634 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4635 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4636 ctl = iff->in(0); // This test feeds the known slow_region. 4637 continue; 4638 } 4639 // One more try: Various low-level checks bottom out in 4640 // uncommon traps. If the debug-info of the trap omits 4641 // any reference to the allocation, as we've already 4642 // observed, then there can be no objection to the trap. 4643 bool found_trap = false; 4644 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4645 Node* obs = not_ctl->fast_out(j); 4646 if (obs->in(0) == not_ctl && obs->is_Call() && 4647 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 4648 found_trap = true; break; 4649 } 4650 } 4651 if (found_trap) { 4652 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4653 continue; 4654 } 4655 } 4656 return NULL; 4657 } 4658 4659 // If we get this far, we have an allocation which immediately 4660 // precedes the arraycopy, and we can take over zeroing the new object. 4661 // The arraycopy will finish the initialization, and provide 4662 // a new control state to which we will anchor the destination pointer. 4663 4664 return alloc; 4665 } 4666 4667 //-------------inline_encodeISOArray----------------------------------- 4668 // encode char[] to byte[] in ISO_8859_1 4669 bool LibraryCallKit::inline_encodeISOArray() { 4670 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 4671 // no receiver since it is static method 4672 Node *src = argument(0); 4673 Node *src_offset = argument(1); 4674 Node *dst = argument(2); 4675 Node *dst_offset = argument(3); 4676 Node *length = argument(4); 4677 4678 src = must_be_not_null(src, true); 4679 dst = must_be_not_null(dst, true); 4680 4681 const Type* src_type = src->Value(&_gvn); 4682 const Type* dst_type = dst->Value(&_gvn); 4683 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4684 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 4685 if (top_src == NULL || top_src->klass() == NULL || 4686 top_dest == NULL || top_dest->klass() == NULL) { 4687 // failed array check 4688 return false; 4689 } 4690 4691 // Figure out the size and type of the elements we will be copying. 4692 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4693 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4694 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 4695 return false; 4696 } 4697 4698 Node* src_start = array_element_address(src, src_offset, T_CHAR); 4699 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 4700 // 'src_start' points to src array + scaled offset 4701 // 'dst_start' points to dst array + scaled offset 4702 4703 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 4704 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 4705 enc = _gvn.transform(enc); 4706 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 4707 set_memory(res_mem, mtype); 4708 set_result(enc); 4709 clear_upper_avx(); 4710 4711 return true; 4712 } 4713 4714 //-------------inline_multiplyToLen----------------------------------- 4715 bool LibraryCallKit::inline_multiplyToLen() { 4716 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 4717 4718 address stubAddr = StubRoutines::multiplyToLen(); 4719 if (stubAddr == NULL) { 4720 return false; // Intrinsic's stub is not implemented on this platform 4721 } 4722 const char* stubName = "multiplyToLen"; 4723 4724 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 4725 4726 // no receiver because it is a static method 4727 Node* x = argument(0); 4728 Node* xlen = argument(1); 4729 Node* y = argument(2); 4730 Node* ylen = argument(3); 4731 Node* z = argument(4); 4732 4733 x = must_be_not_null(x, true); 4734 y = must_be_not_null(y, true); 4735 4736 const Type* x_type = x->Value(&_gvn); 4737 const Type* y_type = y->Value(&_gvn); 4738 const TypeAryPtr* top_x = x_type->isa_aryptr(); 4739 const TypeAryPtr* top_y = y_type->isa_aryptr(); 4740 if (top_x == NULL || top_x->klass() == NULL || 4741 top_y == NULL || top_y->klass() == NULL) { 4742 // failed array check 4743 return false; 4744 } 4745 4746 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4747 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4748 if (x_elem != T_INT || y_elem != T_INT) { 4749 return false; 4750 } 4751 4752 // Set the original stack and the reexecute bit for the interpreter to reexecute 4753 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 4754 // on the return from z array allocation in runtime. 4755 { PreserveReexecuteState preexecs(this); 4756 jvms()->set_should_reexecute(true); 4757 4758 Node* x_start = array_element_address(x, intcon(0), x_elem); 4759 Node* y_start = array_element_address(y, intcon(0), y_elem); 4760 // 'x_start' points to x array + scaled xlen 4761 // 'y_start' points to y array + scaled ylen 4762 4763 // Allocate the result array 4764 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 4765 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 4766 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 4767 4768 IdealKit ideal(this); 4769 4770 #define __ ideal. 4771 Node* one = __ ConI(1); 4772 Node* zero = __ ConI(0); 4773 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 4774 __ set(need_alloc, zero); 4775 __ set(z_alloc, z); 4776 __ if_then(z, BoolTest::eq, null()); { 4777 __ increment (need_alloc, one); 4778 } __ else_(); { 4779 // Update graphKit memory and control from IdealKit. 4780 sync_kit(ideal); 4781 Node *cast = new CastPPNode(z, TypePtr::NOTNULL); 4782 cast->init_req(0, control()); 4783 _gvn.set_type(cast, cast->bottom_type()); 4784 C->record_for_igvn(cast); 4785 4786 Node* zlen_arg = load_array_length(cast); 4787 // Update IdealKit memory and control from graphKit. 4788 __ sync_kit(this); 4789 __ if_then(zlen_arg, BoolTest::lt, zlen); { 4790 __ increment (need_alloc, one); 4791 } __ end_if(); 4792 } __ end_if(); 4793 4794 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 4795 // Update graphKit memory and control from IdealKit. 4796 sync_kit(ideal); 4797 Node * narr = new_array(klass_node, zlen, 1); 4798 // Update IdealKit memory and control from graphKit. 4799 __ sync_kit(this); 4800 __ set(z_alloc, narr); 4801 } __ end_if(); 4802 4803 sync_kit(ideal); 4804 z = __ value(z_alloc); 4805 // Can't use TypeAryPtr::INTS which uses Bottom offset. 4806 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 4807 // Final sync IdealKit and GraphKit. 4808 final_sync(ideal); 4809 #undef __ 4810 4811 Node* z_start = array_element_address(z, intcon(0), T_INT); 4812 4813 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 4814 OptoRuntime::multiplyToLen_Type(), 4815 stubAddr, stubName, TypePtr::BOTTOM, 4816 x_start, xlen, y_start, ylen, z_start, zlen); 4817 } // original reexecute is set back here 4818 4819 C->set_has_split_ifs(true); // Has chance for split-if optimization 4820 set_result(z); 4821 return true; 4822 } 4823 4824 //-------------inline_squareToLen------------------------------------ 4825 bool LibraryCallKit::inline_squareToLen() { 4826 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 4827 4828 address stubAddr = StubRoutines::squareToLen(); 4829 if (stubAddr == NULL) { 4830 return false; // Intrinsic's stub is not implemented on this platform 4831 } 4832 const char* stubName = "squareToLen"; 4833 4834 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 4835 4836 Node* x = argument(0); 4837 Node* len = argument(1); 4838 Node* z = argument(2); 4839 Node* zlen = argument(3); 4840 4841 x = must_be_not_null(x, true); 4842 z = must_be_not_null(z, true); 4843 4844 const Type* x_type = x->Value(&_gvn); 4845 const Type* z_type = z->Value(&_gvn); 4846 const TypeAryPtr* top_x = x_type->isa_aryptr(); 4847 const TypeAryPtr* top_z = z_type->isa_aryptr(); 4848 if (top_x == NULL || top_x->klass() == NULL || 4849 top_z == NULL || top_z->klass() == NULL) { 4850 // failed array check 4851 return false; 4852 } 4853 4854 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4855 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4856 if (x_elem != T_INT || z_elem != T_INT) { 4857 return false; 4858 } 4859 4860 4861 Node* x_start = array_element_address(x, intcon(0), x_elem); 4862 Node* z_start = array_element_address(z, intcon(0), z_elem); 4863 4864 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 4865 OptoRuntime::squareToLen_Type(), 4866 stubAddr, stubName, TypePtr::BOTTOM, 4867 x_start, len, z_start, zlen); 4868 4869 set_result(z); 4870 return true; 4871 } 4872 4873 //-------------inline_mulAdd------------------------------------------ 4874 bool LibraryCallKit::inline_mulAdd() { 4875 assert(UseMulAddIntrinsic, "not implemented on this platform"); 4876 4877 address stubAddr = StubRoutines::mulAdd(); 4878 if (stubAddr == NULL) { 4879 return false; // Intrinsic's stub is not implemented on this platform 4880 } 4881 const char* stubName = "mulAdd"; 4882 4883 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 4884 4885 Node* out = argument(0); 4886 Node* in = argument(1); 4887 Node* offset = argument(2); 4888 Node* len = argument(3); 4889 Node* k = argument(4); 4890 4891 out = must_be_not_null(out, true); 4892 4893 const Type* out_type = out->Value(&_gvn); 4894 const Type* in_type = in->Value(&_gvn); 4895 const TypeAryPtr* top_out = out_type->isa_aryptr(); 4896 const TypeAryPtr* top_in = in_type->isa_aryptr(); 4897 if (top_out == NULL || top_out->klass() == NULL || 4898 top_in == NULL || top_in->klass() == NULL) { 4899 // failed array check 4900 return false; 4901 } 4902 4903 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4904 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4905 if (out_elem != T_INT || in_elem != T_INT) { 4906 return false; 4907 } 4908 4909 Node* outlen = load_array_length(out); 4910 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 4911 Node* out_start = array_element_address(out, intcon(0), out_elem); 4912 Node* in_start = array_element_address(in, intcon(0), in_elem); 4913 4914 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 4915 OptoRuntime::mulAdd_Type(), 4916 stubAddr, stubName, TypePtr::BOTTOM, 4917 out_start,in_start, new_offset, len, k); 4918 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 4919 set_result(result); 4920 return true; 4921 } 4922 4923 //-------------inline_montgomeryMultiply----------------------------------- 4924 bool LibraryCallKit::inline_montgomeryMultiply() { 4925 address stubAddr = StubRoutines::montgomeryMultiply(); 4926 if (stubAddr == NULL) { 4927 return false; // Intrinsic's stub is not implemented on this platform 4928 } 4929 4930 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 4931 const char* stubName = "montgomery_multiply"; 4932 4933 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 4934 4935 Node* a = argument(0); 4936 Node* b = argument(1); 4937 Node* n = argument(2); 4938 Node* len = argument(3); 4939 Node* inv = argument(4); 4940 Node* m = argument(6); 4941 4942 const Type* a_type = a->Value(&_gvn); 4943 const TypeAryPtr* top_a = a_type->isa_aryptr(); 4944 const Type* b_type = b->Value(&_gvn); 4945 const TypeAryPtr* top_b = b_type->isa_aryptr(); 4946 const Type* n_type = a->Value(&_gvn); 4947 const TypeAryPtr* top_n = n_type->isa_aryptr(); 4948 const Type* m_type = a->Value(&_gvn); 4949 const TypeAryPtr* top_m = m_type->isa_aryptr(); 4950 if (top_a == NULL || top_a->klass() == NULL || 4951 top_b == NULL || top_b->klass() == NULL || 4952 top_n == NULL || top_n->klass() == NULL || 4953 top_m == NULL || top_m->klass() == NULL) { 4954 // failed array check 4955 return false; 4956 } 4957 4958 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4959 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4960 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4961 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4962 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 4963 return false; 4964 } 4965 4966 // Make the call 4967 { 4968 Node* a_start = array_element_address(a, intcon(0), a_elem); 4969 Node* b_start = array_element_address(b, intcon(0), b_elem); 4970 Node* n_start = array_element_address(n, intcon(0), n_elem); 4971 Node* m_start = array_element_address(m, intcon(0), m_elem); 4972 4973 Node* call = make_runtime_call(RC_LEAF, 4974 OptoRuntime::montgomeryMultiply_Type(), 4975 stubAddr, stubName, TypePtr::BOTTOM, 4976 a_start, b_start, n_start, len, inv, top(), 4977 m_start); 4978 set_result(m); 4979 } 4980 4981 return true; 4982 } 4983 4984 bool LibraryCallKit::inline_montgomerySquare() { 4985 address stubAddr = StubRoutines::montgomerySquare(); 4986 if (stubAddr == NULL) { 4987 return false; // Intrinsic's stub is not implemented on this platform 4988 } 4989 4990 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 4991 const char* stubName = "montgomery_square"; 4992 4993 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 4994 4995 Node* a = argument(0); 4996 Node* n = argument(1); 4997 Node* len = argument(2); 4998 Node* inv = argument(3); 4999 Node* m = argument(5); 5000 5001 const Type* a_type = a->Value(&_gvn); 5002 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5003 const Type* n_type = a->Value(&_gvn); 5004 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5005 const Type* m_type = a->Value(&_gvn); 5006 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5007 if (top_a == NULL || top_a->klass() == NULL || 5008 top_n == NULL || top_n->klass() == NULL || 5009 top_m == NULL || top_m->klass() == NULL) { 5010 // failed array check 5011 return false; 5012 } 5013 5014 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5015 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5016 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5017 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5018 return false; 5019 } 5020 5021 // Make the call 5022 { 5023 Node* a_start = array_element_address(a, intcon(0), a_elem); 5024 Node* n_start = array_element_address(n, intcon(0), n_elem); 5025 Node* m_start = array_element_address(m, intcon(0), m_elem); 5026 5027 Node* call = make_runtime_call(RC_LEAF, 5028 OptoRuntime::montgomerySquare_Type(), 5029 stubAddr, stubName, TypePtr::BOTTOM, 5030 a_start, n_start, len, inv, top(), 5031 m_start); 5032 set_result(m); 5033 } 5034 5035 return true; 5036 } 5037 5038 bool LibraryCallKit::inline_bigIntegerShift(bool isRightShift) { 5039 address stubAddr = NULL; 5040 const char* stubName = NULL; 5041 5042 stubAddr = isRightShift? StubRoutines::bigIntegerRightShift(): StubRoutines::bigIntegerLeftShift(); 5043 if (stubAddr == NULL) { 5044 return false; // Intrinsic's stub is not implemented on this platform 5045 } 5046 5047 stubName = isRightShift? "bigIntegerRightShiftWorker" : "bigIntegerLeftShiftWorker"; 5048 5049 assert(callee()->signature()->size() == 5, "expected 5 arguments"); 5050 5051 Node* newArr = argument(0); 5052 Node* oldArr = argument(1); 5053 Node* newIdx = argument(2); 5054 Node* shiftCount = argument(3); 5055 Node* numIter = argument(4); 5056 5057 const Type* newArr_type = newArr->Value(&_gvn); 5058 const TypeAryPtr* top_newArr = newArr_type->isa_aryptr(); 5059 const Type* oldArr_type = oldArr->Value(&_gvn); 5060 const TypeAryPtr* top_oldArr = oldArr_type->isa_aryptr(); 5061 if (top_newArr == NULL || top_newArr->klass() == NULL || top_oldArr == NULL 5062 || top_oldArr->klass() == NULL) { 5063 return false; 5064 } 5065 5066 BasicType newArr_elem = newArr_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5067 BasicType oldArr_elem = oldArr_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5068 if (newArr_elem != T_INT || oldArr_elem != T_INT) { 5069 return false; 5070 } 5071 5072 // Make the call 5073 { 5074 Node* newArr_start = array_element_address(newArr, intcon(0), newArr_elem); 5075 Node* oldArr_start = array_element_address(oldArr, intcon(0), oldArr_elem); 5076 5077 Node* call = make_runtime_call(RC_LEAF, 5078 OptoRuntime::bigIntegerShift_Type(), 5079 stubAddr, 5080 stubName, 5081 TypePtr::BOTTOM, 5082 newArr_start, 5083 oldArr_start, 5084 newIdx, 5085 shiftCount, 5086 numIter); 5087 } 5088 5089 return true; 5090 } 5091 5092 //-------------inline_vectorizedMismatch------------------------------ 5093 bool LibraryCallKit::inline_vectorizedMismatch() { 5094 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5095 5096 address stubAddr = StubRoutines::vectorizedMismatch(); 5097 if (stubAddr == NULL) { 5098 return false; // Intrinsic's stub is not implemented on this platform 5099 } 5100 const char* stubName = "vectorizedMismatch"; 5101 int size_l = callee()->signature()->size(); 5102 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5103 5104 Node* obja = argument(0); 5105 Node* aoffset = argument(1); 5106 Node* objb = argument(3); 5107 Node* boffset = argument(4); 5108 Node* length = argument(6); 5109 Node* scale = argument(7); 5110 5111 const Type* a_type = obja->Value(&_gvn); 5112 const Type* b_type = objb->Value(&_gvn); 5113 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5114 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5115 if (top_a == NULL || top_a->klass() == NULL || 5116 top_b == NULL || top_b->klass() == NULL) { 5117 // failed array check 5118 return false; 5119 } 5120 5121 Node* call; 5122 jvms()->set_should_reexecute(true); 5123 5124 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ); 5125 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ); 5126 5127 call = make_runtime_call(RC_LEAF, 5128 OptoRuntime::vectorizedMismatch_Type(), 5129 stubAddr, stubName, TypePtr::BOTTOM, 5130 obja_adr, objb_adr, length, scale); 5131 5132 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5133 set_result(result); 5134 return true; 5135 } 5136 5137 /** 5138 * Calculate CRC32 for byte. 5139 * int java.util.zip.CRC32.update(int crc, int b) 5140 */ 5141 bool LibraryCallKit::inline_updateCRC32() { 5142 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5143 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5144 // no receiver since it is static method 5145 Node* crc = argument(0); // type: int 5146 Node* b = argument(1); // type: int 5147 5148 /* 5149 * int c = ~ crc; 5150 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5151 * b = b ^ (c >>> 8); 5152 * crc = ~b; 5153 */ 5154 5155 Node* M1 = intcon(-1); 5156 crc = _gvn.transform(new XorINode(crc, M1)); 5157 Node* result = _gvn.transform(new XorINode(crc, b)); 5158 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5159 5160 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5161 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5162 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5163 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5164 5165 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5166 result = _gvn.transform(new XorINode(crc, result)); 5167 result = _gvn.transform(new XorINode(result, M1)); 5168 set_result(result); 5169 return true; 5170 } 5171 5172 /** 5173 * Calculate CRC32 for byte[] array. 5174 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5175 */ 5176 bool LibraryCallKit::inline_updateBytesCRC32() { 5177 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5178 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5179 // no receiver since it is static method 5180 Node* crc = argument(0); // type: int 5181 Node* src = argument(1); // type: oop 5182 Node* offset = argument(2); // type: int 5183 Node* length = argument(3); // type: int 5184 5185 const Type* src_type = src->Value(&_gvn); 5186 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5187 if (top_src == NULL || top_src->klass() == NULL) { 5188 // failed array check 5189 return false; 5190 } 5191 5192 // Figure out the size and type of the elements we will be copying. 5193 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5194 if (src_elem != T_BYTE) { 5195 return false; 5196 } 5197 5198 // 'src_start' points to src array + scaled offset 5199 src = must_be_not_null(src, true); 5200 Node* src_start = array_element_address(src, offset, src_elem); 5201 5202 // We assume that range check is done by caller. 5203 // TODO: generate range check (offset+length < src.length) in debug VM. 5204 5205 // Call the stub. 5206 address stubAddr = StubRoutines::updateBytesCRC32(); 5207 const char *stubName = "updateBytesCRC32"; 5208 5209 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5210 stubAddr, stubName, TypePtr::BOTTOM, 5211 crc, src_start, length); 5212 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5213 set_result(result); 5214 return true; 5215 } 5216 5217 /** 5218 * Calculate CRC32 for ByteBuffer. 5219 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5220 */ 5221 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5222 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5223 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5224 // no receiver since it is static method 5225 Node* crc = argument(0); // type: int 5226 Node* src = argument(1); // type: long 5227 Node* offset = argument(3); // type: int 5228 Node* length = argument(4); // type: int 5229 5230 src = ConvL2X(src); // adjust Java long to machine word 5231 Node* base = _gvn.transform(new CastX2PNode(src)); 5232 offset = ConvI2X(offset); 5233 5234 // 'src_start' points to src array + scaled offset 5235 Node* src_start = basic_plus_adr(top(), base, offset); 5236 5237 // Call the stub. 5238 address stubAddr = StubRoutines::updateBytesCRC32(); 5239 const char *stubName = "updateBytesCRC32"; 5240 5241 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5242 stubAddr, stubName, TypePtr::BOTTOM, 5243 crc, src_start, length); 5244 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5245 set_result(result); 5246 return true; 5247 } 5248 5249 //------------------------------get_table_from_crc32c_class----------------------- 5250 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5251 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5252 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5253 5254 return table; 5255 } 5256 5257 //------------------------------inline_updateBytesCRC32C----------------------- 5258 // 5259 // Calculate CRC32C for byte[] array. 5260 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5261 // 5262 bool LibraryCallKit::inline_updateBytesCRC32C() { 5263 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5264 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5265 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5266 // no receiver since it is a static method 5267 Node* crc = argument(0); // type: int 5268 Node* src = argument(1); // type: oop 5269 Node* offset = argument(2); // type: int 5270 Node* end = argument(3); // type: int 5271 5272 Node* length = _gvn.transform(new SubINode(end, offset)); 5273 5274 const Type* src_type = src->Value(&_gvn); 5275 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5276 if (top_src == NULL || top_src->klass() == NULL) { 5277 // failed array check 5278 return false; 5279 } 5280 5281 // Figure out the size and type of the elements we will be copying. 5282 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5283 if (src_elem != T_BYTE) { 5284 return false; 5285 } 5286 5287 // 'src_start' points to src array + scaled offset 5288 src = must_be_not_null(src, true); 5289 Node* src_start = array_element_address(src, offset, src_elem); 5290 5291 // static final int[] byteTable in class CRC32C 5292 Node* table = get_table_from_crc32c_class(callee()->holder()); 5293 table = must_be_not_null(table, true); 5294 Node* table_start = array_element_address(table, intcon(0), T_INT); 5295 5296 // We assume that range check is done by caller. 5297 // TODO: generate range check (offset+length < src.length) in debug VM. 5298 5299 // Call the stub. 5300 address stubAddr = StubRoutines::updateBytesCRC32C(); 5301 const char *stubName = "updateBytesCRC32C"; 5302 5303 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5304 stubAddr, stubName, TypePtr::BOTTOM, 5305 crc, src_start, length, table_start); 5306 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5307 set_result(result); 5308 return true; 5309 } 5310 5311 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5312 // 5313 // Calculate CRC32C for DirectByteBuffer. 5314 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5315 // 5316 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5317 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5318 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5319 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5320 // no receiver since it is a static method 5321 Node* crc = argument(0); // type: int 5322 Node* src = argument(1); // type: long 5323 Node* offset = argument(3); // type: int 5324 Node* end = argument(4); // type: int 5325 5326 Node* length = _gvn.transform(new SubINode(end, offset)); 5327 5328 src = ConvL2X(src); // adjust Java long to machine word 5329 Node* base = _gvn.transform(new CastX2PNode(src)); 5330 offset = ConvI2X(offset); 5331 5332 // 'src_start' points to src array + scaled offset 5333 Node* src_start = basic_plus_adr(top(), base, offset); 5334 5335 // static final int[] byteTable in class CRC32C 5336 Node* table = get_table_from_crc32c_class(callee()->holder()); 5337 table = must_be_not_null(table, true); 5338 Node* table_start = array_element_address(table, intcon(0), T_INT); 5339 5340 // Call the stub. 5341 address stubAddr = StubRoutines::updateBytesCRC32C(); 5342 const char *stubName = "updateBytesCRC32C"; 5343 5344 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5345 stubAddr, stubName, TypePtr::BOTTOM, 5346 crc, src_start, length, table_start); 5347 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5348 set_result(result); 5349 return true; 5350 } 5351 5352 //------------------------------inline_updateBytesAdler32---------------------- 5353 // 5354 // Calculate Adler32 checksum for byte[] array. 5355 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5356 // 5357 bool LibraryCallKit::inline_updateBytesAdler32() { 5358 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5359 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5360 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5361 // no receiver since it is static method 5362 Node* crc = argument(0); // type: int 5363 Node* src = argument(1); // type: oop 5364 Node* offset = argument(2); // type: int 5365 Node* length = argument(3); // type: int 5366 5367 const Type* src_type = src->Value(&_gvn); 5368 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5369 if (top_src == NULL || top_src->klass() == NULL) { 5370 // failed array check 5371 return false; 5372 } 5373 5374 // Figure out the size and type of the elements we will be copying. 5375 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5376 if (src_elem != T_BYTE) { 5377 return false; 5378 } 5379 5380 // 'src_start' points to src array + scaled offset 5381 Node* src_start = array_element_address(src, offset, src_elem); 5382 5383 // We assume that range check is done by caller. 5384 // TODO: generate range check (offset+length < src.length) in debug VM. 5385 5386 // Call the stub. 5387 address stubAddr = StubRoutines::updateBytesAdler32(); 5388 const char *stubName = "updateBytesAdler32"; 5389 5390 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5391 stubAddr, stubName, TypePtr::BOTTOM, 5392 crc, src_start, length); 5393 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5394 set_result(result); 5395 return true; 5396 } 5397 5398 //------------------------------inline_updateByteBufferAdler32--------------- 5399 // 5400 // Calculate Adler32 checksum for DirectByteBuffer. 5401 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5402 // 5403 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5404 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5405 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5406 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5407 // no receiver since it is static method 5408 Node* crc = argument(0); // type: int 5409 Node* src = argument(1); // type: long 5410 Node* offset = argument(3); // type: int 5411 Node* length = argument(4); // type: int 5412 5413 src = ConvL2X(src); // adjust Java long to machine word 5414 Node* base = _gvn.transform(new CastX2PNode(src)); 5415 offset = ConvI2X(offset); 5416 5417 // 'src_start' points to src array + scaled offset 5418 Node* src_start = basic_plus_adr(top(), base, offset); 5419 5420 // Call the stub. 5421 address stubAddr = StubRoutines::updateBytesAdler32(); 5422 const char *stubName = "updateBytesAdler32"; 5423 5424 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5425 stubAddr, stubName, TypePtr::BOTTOM, 5426 crc, src_start, length); 5427 5428 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5429 set_result(result); 5430 return true; 5431 } 5432 5433 //----------------------------inline_reference_get---------------------------- 5434 // public T java.lang.ref.Reference.get(); 5435 bool LibraryCallKit::inline_reference_get() { 5436 const int referent_offset = java_lang_ref_Reference::referent_offset(); 5437 5438 // Get the argument: 5439 Node* reference_obj = null_check_receiver(); 5440 if (stopped()) return true; 5441 5442 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr(); 5443 assert(tinst != NULL, "obj is null"); 5444 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5445 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass(); 5446 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"), 5447 ciSymbol::make("Ljava/lang/Object;"), 5448 false); 5449 assert (field != NULL, "undefined field"); 5450 5451 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5452 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5453 5454 ciInstanceKlass* klass = env()->Object_klass(); 5455 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5456 5457 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF; 5458 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators); 5459 // Add memory barrier to prevent commoning reads from this field 5460 // across safepoint since GC can change its value. 5461 insert_mem_bar(Op_MemBarCPUOrder); 5462 5463 set_result(result); 5464 return true; 5465 } 5466 5467 5468 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5469 bool is_exact=true, bool is_static=false, 5470 ciInstanceKlass * fromKls=NULL) { 5471 if (fromKls == NULL) { 5472 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5473 assert(tinst != NULL, "obj is null"); 5474 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5475 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5476 fromKls = tinst->klass()->as_instance_klass(); 5477 } else { 5478 assert(is_static, "only for static field access"); 5479 } 5480 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5481 ciSymbol::make(fieldTypeString), 5482 is_static); 5483 5484 assert (field != NULL, "undefined field"); 5485 if (field == NULL) return (Node *) NULL; 5486 5487 if (is_static) { 5488 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5489 fromObj = makecon(tip); 5490 } 5491 5492 // Next code copied from Parse::do_get_xxx(): 5493 5494 // Compute address and memory type. 5495 int offset = field->offset_in_bytes(); 5496 bool is_vol = field->is_volatile(); 5497 ciType* field_klass = field->type(); 5498 assert(field_klass->is_loaded(), "should be loaded"); 5499 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5500 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5501 BasicType bt = field->layout_type(); 5502 5503 // Build the resultant type of the load 5504 const Type *type; 5505 if (bt == T_OBJECT) { 5506 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5507 } else { 5508 type = Type::get_const_basic_type(bt); 5509 } 5510 5511 DecoratorSet decorators = IN_HEAP; 5512 5513 if (is_vol) { 5514 decorators |= MO_SEQ_CST; 5515 } 5516 5517 return access_load_at(fromObj, adr, adr_type, type, bt, decorators); 5518 } 5519 5520 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5521 bool is_exact = true, bool is_static = false, 5522 ciInstanceKlass * fromKls = NULL) { 5523 if (fromKls == NULL) { 5524 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5525 assert(tinst != NULL, "obj is null"); 5526 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5527 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5528 fromKls = tinst->klass()->as_instance_klass(); 5529 } 5530 else { 5531 assert(is_static, "only for static field access"); 5532 } 5533 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5534 ciSymbol::make(fieldTypeString), 5535 is_static); 5536 5537 assert(field != NULL, "undefined field"); 5538 assert(!field->is_volatile(), "not defined for volatile fields"); 5539 5540 if (is_static) { 5541 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5542 fromObj = makecon(tip); 5543 } 5544 5545 // Next code copied from Parse::do_get_xxx(): 5546 5547 // Compute address and memory type. 5548 int offset = field->offset_in_bytes(); 5549 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5550 5551 return adr; 5552 } 5553 5554 //------------------------------inline_aescrypt_Block----------------------- 5555 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5556 address stubAddr = NULL; 5557 const char *stubName; 5558 assert(UseAES, "need AES instruction support"); 5559 5560 switch(id) { 5561 case vmIntrinsics::_aescrypt_encryptBlock: 5562 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5563 stubName = "aescrypt_encryptBlock"; 5564 break; 5565 case vmIntrinsics::_aescrypt_decryptBlock: 5566 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5567 stubName = "aescrypt_decryptBlock"; 5568 break; 5569 default: 5570 break; 5571 } 5572 if (stubAddr == NULL) return false; 5573 5574 Node* aescrypt_object = argument(0); 5575 Node* src = argument(1); 5576 Node* src_offset = argument(2); 5577 Node* dest = argument(3); 5578 Node* dest_offset = argument(4); 5579 5580 src = must_be_not_null(src, true); 5581 dest = must_be_not_null(dest, true); 5582 5583 // (1) src and dest are arrays. 5584 const Type* src_type = src->Value(&_gvn); 5585 const Type* dest_type = dest->Value(&_gvn); 5586 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5587 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5588 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5589 5590 // for the quick and dirty code we will skip all the checks. 5591 // we are just trying to get the call to be generated. 5592 Node* src_start = src; 5593 Node* dest_start = dest; 5594 if (src_offset != NULL || dest_offset != NULL) { 5595 assert(src_offset != NULL && dest_offset != NULL, ""); 5596 src_start = array_element_address(src, src_offset, T_BYTE); 5597 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5598 } 5599 5600 // now need to get the start of its expanded key array 5601 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5602 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5603 if (k_start == NULL) return false; 5604 5605 if (Matcher::pass_original_key_for_aes()) { 5606 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5607 // compatibility issues between Java key expansion and SPARC crypto instructions 5608 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5609 if (original_k_start == NULL) return false; 5610 5611 // Call the stub. 5612 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5613 stubAddr, stubName, TypePtr::BOTTOM, 5614 src_start, dest_start, k_start, original_k_start); 5615 } else { 5616 // Call the stub. 5617 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5618 stubAddr, stubName, TypePtr::BOTTOM, 5619 src_start, dest_start, k_start); 5620 } 5621 5622 return true; 5623 } 5624 5625 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5626 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5627 address stubAddr = NULL; 5628 const char *stubName = NULL; 5629 5630 assert(UseAES, "need AES instruction support"); 5631 5632 switch(id) { 5633 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5634 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5635 stubName = "cipherBlockChaining_encryptAESCrypt"; 5636 break; 5637 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5638 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5639 stubName = "cipherBlockChaining_decryptAESCrypt"; 5640 break; 5641 default: 5642 break; 5643 } 5644 if (stubAddr == NULL) return false; 5645 5646 Node* cipherBlockChaining_object = argument(0); 5647 Node* src = argument(1); 5648 Node* src_offset = argument(2); 5649 Node* len = argument(3); 5650 Node* dest = argument(4); 5651 Node* dest_offset = argument(5); 5652 5653 src = must_be_not_null(src, false); 5654 dest = must_be_not_null(dest, false); 5655 5656 // (1) src and dest are arrays. 5657 const Type* src_type = src->Value(&_gvn); 5658 const Type* dest_type = dest->Value(&_gvn); 5659 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5660 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5661 assert (top_src != NULL && top_src->klass() != NULL 5662 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5663 5664 // checks are the responsibility of the caller 5665 Node* src_start = src; 5666 Node* dest_start = dest; 5667 if (src_offset != NULL || dest_offset != NULL) { 5668 assert(src_offset != NULL && dest_offset != NULL, ""); 5669 src_start = array_element_address(src, src_offset, T_BYTE); 5670 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5671 } 5672 5673 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5674 // (because of the predicated logic executed earlier). 5675 // so we cast it here safely. 5676 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5677 5678 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5679 if (embeddedCipherObj == NULL) return false; 5680 5681 // cast it to what we know it will be at runtime 5682 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5683 assert(tinst != NULL, "CBC obj is null"); 5684 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5685 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5686 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5687 5688 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5689 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5690 const TypeOopPtr* xtype = aklass->as_instance_type(); 5691 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5692 aescrypt_object = _gvn.transform(aescrypt_object); 5693 5694 // we need to get the start of the aescrypt_object's expanded key array 5695 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5696 if (k_start == NULL) return false; 5697 5698 // similarly, get the start address of the r vector 5699 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5700 if (objRvec == NULL) return false; 5701 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5702 5703 Node* cbcCrypt; 5704 if (Matcher::pass_original_key_for_aes()) { 5705 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5706 // compatibility issues between Java key expansion and SPARC crypto instructions 5707 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5708 if (original_k_start == NULL) return false; 5709 5710 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5711 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5712 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5713 stubAddr, stubName, TypePtr::BOTTOM, 5714 src_start, dest_start, k_start, r_start, len, original_k_start); 5715 } else { 5716 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5717 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5718 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5719 stubAddr, stubName, TypePtr::BOTTOM, 5720 src_start, dest_start, k_start, r_start, len); 5721 } 5722 5723 // return cipher length (int) 5724 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5725 set_result(retvalue); 5726 return true; 5727 } 5728 5729 //------------------------------inline_electronicCodeBook_AESCrypt----------------------- 5730 bool LibraryCallKit::inline_electronicCodeBook_AESCrypt(vmIntrinsics::ID id) { 5731 address stubAddr = NULL; 5732 const char *stubName = NULL; 5733 5734 assert(UseAES, "need AES instruction support"); 5735 5736 switch (id) { 5737 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt: 5738 stubAddr = StubRoutines::electronicCodeBook_encryptAESCrypt(); 5739 stubName = "electronicCodeBook_encryptAESCrypt"; 5740 break; 5741 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt: 5742 stubAddr = StubRoutines::electronicCodeBook_decryptAESCrypt(); 5743 stubName = "electronicCodeBook_decryptAESCrypt"; 5744 break; 5745 default: 5746 break; 5747 } 5748 5749 if (stubAddr == NULL) return false; 5750 5751 Node* electronicCodeBook_object = argument(0); 5752 Node* src = argument(1); 5753 Node* src_offset = argument(2); 5754 Node* len = argument(3); 5755 Node* dest = argument(4); 5756 Node* dest_offset = argument(5); 5757 5758 // (1) src and dest are arrays. 5759 const Type* src_type = src->Value(&_gvn); 5760 const Type* dest_type = dest->Value(&_gvn); 5761 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5762 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5763 assert(top_src != NULL && top_src->klass() != NULL 5764 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5765 5766 // checks are the responsibility of the caller 5767 Node* src_start = src; 5768 Node* dest_start = dest; 5769 if (src_offset != NULL || dest_offset != NULL) { 5770 assert(src_offset != NULL && dest_offset != NULL, ""); 5771 src_start = array_element_address(src, src_offset, T_BYTE); 5772 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5773 } 5774 5775 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5776 // (because of the predicated logic executed earlier). 5777 // so we cast it here safely. 5778 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5779 5780 Node* embeddedCipherObj = load_field_from_object(electronicCodeBook_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5781 if (embeddedCipherObj == NULL) return false; 5782 5783 // cast it to what we know it will be at runtime 5784 const TypeInstPtr* tinst = _gvn.type(electronicCodeBook_object)->isa_instptr(); 5785 assert(tinst != NULL, "ECB obj is null"); 5786 assert(tinst->klass()->is_loaded(), "ECB obj is not loaded"); 5787 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5788 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5789 5790 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5791 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5792 const TypeOopPtr* xtype = aklass->as_instance_type(); 5793 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5794 aescrypt_object = _gvn.transform(aescrypt_object); 5795 5796 // we need to get the start of the aescrypt_object's expanded key array 5797 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5798 if (k_start == NULL) return false; 5799 5800 Node* ecbCrypt; 5801 if (Matcher::pass_original_key_for_aes()) { 5802 // no SPARC version for AES/ECB intrinsics now. 5803 return false; 5804 } 5805 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5806 ecbCrypt = make_runtime_call(RC_LEAF | RC_NO_FP, 5807 OptoRuntime::electronicCodeBook_aescrypt_Type(), 5808 stubAddr, stubName, TypePtr::BOTTOM, 5809 src_start, dest_start, k_start, len); 5810 5811 // return cipher length (int) 5812 Node* retvalue = _gvn.transform(new ProjNode(ecbCrypt, TypeFunc::Parms)); 5813 set_result(retvalue); 5814 return true; 5815 } 5816 5817 //------------------------------inline_counterMode_AESCrypt----------------------- 5818 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 5819 assert(UseAES, "need AES instruction support"); 5820 if (!UseAESCTRIntrinsics) return false; 5821 5822 address stubAddr = NULL; 5823 const char *stubName = NULL; 5824 if (id == vmIntrinsics::_counterMode_AESCrypt) { 5825 stubAddr = StubRoutines::counterMode_AESCrypt(); 5826 stubName = "counterMode_AESCrypt"; 5827 } 5828 if (stubAddr == NULL) return false; 5829 5830 Node* counterMode_object = argument(0); 5831 Node* src = argument(1); 5832 Node* src_offset = argument(2); 5833 Node* len = argument(3); 5834 Node* dest = argument(4); 5835 Node* dest_offset = argument(5); 5836 5837 // (1) src and dest are arrays. 5838 const Type* src_type = src->Value(&_gvn); 5839 const Type* dest_type = dest->Value(&_gvn); 5840 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5841 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5842 assert(top_src != NULL && top_src->klass() != NULL && 5843 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5844 5845 // checks are the responsibility of the caller 5846 Node* src_start = src; 5847 Node* dest_start = dest; 5848 if (src_offset != NULL || dest_offset != NULL) { 5849 assert(src_offset != NULL && dest_offset != NULL, ""); 5850 src_start = array_element_address(src, src_offset, T_BYTE); 5851 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5852 } 5853 5854 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5855 // (because of the predicated logic executed earlier). 5856 // so we cast it here safely. 5857 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5858 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5859 if (embeddedCipherObj == NULL) return false; 5860 // cast it to what we know it will be at runtime 5861 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 5862 assert(tinst != NULL, "CTR obj is null"); 5863 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 5864 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5865 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5866 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5867 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5868 const TypeOopPtr* xtype = aklass->as_instance_type(); 5869 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5870 aescrypt_object = _gvn.transform(aescrypt_object); 5871 // we need to get the start of the aescrypt_object's expanded key array 5872 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5873 if (k_start == NULL) return false; 5874 // similarly, get the start address of the r vector 5875 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 5876 if (obj_counter == NULL) return false; 5877 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 5878 5879 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 5880 if (saved_encCounter == NULL) return false; 5881 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 5882 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 5883 5884 Node* ctrCrypt; 5885 if (Matcher::pass_original_key_for_aes()) { 5886 // no SPARC version for AES/CTR intrinsics now. 5887 return false; 5888 } 5889 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5890 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5891 OptoRuntime::counterMode_aescrypt_Type(), 5892 stubAddr, stubName, TypePtr::BOTTOM, 5893 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 5894 5895 // return cipher length (int) 5896 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 5897 set_result(retvalue); 5898 return true; 5899 } 5900 5901 //------------------------------get_key_start_from_aescrypt_object----------------------- 5902 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5903 #if defined(PPC64) || defined(S390) 5904 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 5905 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 5906 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 5907 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 5908 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 5909 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5910 if (objSessionK == NULL) { 5911 return (Node *) NULL; 5912 } 5913 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 5914 #else 5915 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5916 #endif // PPC64 5917 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5918 if (objAESCryptKey == NULL) return (Node *) NULL; 5919 5920 // now have the array, need to get the start address of the K array 5921 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5922 return k_start; 5923 } 5924 5925 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5926 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5927 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5928 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5929 if (objAESCryptKey == NULL) return (Node *) NULL; 5930 5931 // now have the array, need to get the start address of the lastKey array 5932 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5933 return original_k_start; 5934 } 5935 5936 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5937 // Return node representing slow path of predicate check. 5938 // the pseudo code we want to emulate with this predicate is: 5939 // for encryption: 5940 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5941 // for decryption: 5942 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5943 // note cipher==plain is more conservative than the original java code but that's OK 5944 // 5945 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5946 // The receiver was checked for NULL already. 5947 Node* objCBC = argument(0); 5948 5949 Node* src = argument(1); 5950 Node* dest = argument(4); 5951 5952 // Load embeddedCipher field of CipherBlockChaining object. 5953 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5954 5955 // get AESCrypt klass for instanceOf check 5956 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5957 // will have same classloader as CipherBlockChaining object 5958 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5959 assert(tinst != NULL, "CBCobj is null"); 5960 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5961 5962 // we want to do an instanceof comparison against the AESCrypt class 5963 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5964 if (!klass_AESCrypt->is_loaded()) { 5965 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5966 Node* ctrl = control(); 5967 set_control(top()); // no regular fast path 5968 return ctrl; 5969 } 5970 5971 src = must_be_not_null(src, true); 5972 dest = must_be_not_null(dest, true); 5973 5974 // Resolve oops to stable for CmpP below. 5975 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5976 5977 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5978 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5979 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5980 5981 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5982 5983 // for encryption, we are done 5984 if (!decrypting) 5985 return instof_false; // even if it is NULL 5986 5987 // for decryption, we need to add a further check to avoid 5988 // taking the intrinsic path when cipher and plain are the same 5989 // see the original java code for why. 5990 RegionNode* region = new RegionNode(3); 5991 region->init_req(1, instof_false); 5992 5993 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5994 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5995 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5996 region->init_req(2, src_dest_conjoint); 5997 5998 record_for_igvn(region); 5999 return _gvn.transform(region); 6000 } 6001 6002 //----------------------------inline_electronicCodeBook_AESCrypt_predicate---------------------------- 6003 // Return node representing slow path of predicate check. 6004 // the pseudo code we want to emulate with this predicate is: 6005 // for encryption: 6006 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6007 // for decryption: 6008 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6009 // note cipher==plain is more conservative than the original java code but that's OK 6010 // 6011 Node* LibraryCallKit::inline_electronicCodeBook_AESCrypt_predicate(bool decrypting) { 6012 // The receiver was checked for NULL already. 6013 Node* objECB = argument(0); 6014 6015 // Load embeddedCipher field of ElectronicCodeBook object. 6016 Node* embeddedCipherObj = load_field_from_object(objECB, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6017 6018 // get AESCrypt klass for instanceOf check 6019 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6020 // will have same classloader as ElectronicCodeBook object 6021 const TypeInstPtr* tinst = _gvn.type(objECB)->isa_instptr(); 6022 assert(tinst != NULL, "ECBobj is null"); 6023 assert(tinst->klass()->is_loaded(), "ECBobj is not loaded"); 6024 6025 // we want to do an instanceof comparison against the AESCrypt class 6026 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6027 if (!klass_AESCrypt->is_loaded()) { 6028 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6029 Node* ctrl = control(); 6030 set_control(top()); // no regular fast path 6031 return ctrl; 6032 } 6033 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6034 6035 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6036 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6037 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6038 6039 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6040 6041 // for encryption, we are done 6042 if (!decrypting) 6043 return instof_false; // even if it is NULL 6044 6045 // for decryption, we need to add a further check to avoid 6046 // taking the intrinsic path when cipher and plain are the same 6047 // see the original java code for why. 6048 RegionNode* region = new RegionNode(3); 6049 region->init_req(1, instof_false); 6050 Node* src = argument(1); 6051 Node* dest = argument(4); 6052 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6053 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6054 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6055 region->init_req(2, src_dest_conjoint); 6056 6057 record_for_igvn(region); 6058 return _gvn.transform(region); 6059 } 6060 6061 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6062 // Return node representing slow path of predicate check. 6063 // the pseudo code we want to emulate with this predicate is: 6064 // for encryption: 6065 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6066 // for decryption: 6067 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6068 // note cipher==plain is more conservative than the original java code but that's OK 6069 // 6070 6071 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6072 // The receiver was checked for NULL already. 6073 Node* objCTR = argument(0); 6074 6075 // Load embeddedCipher field of CipherBlockChaining object. 6076 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6077 6078 // get AESCrypt klass for instanceOf check 6079 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6080 // will have same classloader as CipherBlockChaining object 6081 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6082 assert(tinst != NULL, "CTRobj is null"); 6083 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6084 6085 // we want to do an instanceof comparison against the AESCrypt class 6086 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6087 if (!klass_AESCrypt->is_loaded()) { 6088 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6089 Node* ctrl = control(); 6090 set_control(top()); // no regular fast path 6091 return ctrl; 6092 } 6093 6094 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6095 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6096 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6097 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6098 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6099 6100 return instof_false; // even if it is NULL 6101 } 6102 6103 //------------------------------inline_ghash_processBlocks 6104 bool LibraryCallKit::inline_ghash_processBlocks() { 6105 address stubAddr; 6106 const char *stubName; 6107 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6108 6109 stubAddr = StubRoutines::ghash_processBlocks(); 6110 stubName = "ghash_processBlocks"; 6111 6112 Node* data = argument(0); 6113 Node* offset = argument(1); 6114 Node* len = argument(2); 6115 Node* state = argument(3); 6116 Node* subkeyH = argument(4); 6117 6118 state = must_be_not_null(state, true); 6119 subkeyH = must_be_not_null(subkeyH, true); 6120 data = must_be_not_null(data, true); 6121 6122 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6123 assert(state_start, "state is NULL"); 6124 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6125 assert(subkeyH_start, "subkeyH is NULL"); 6126 Node* data_start = array_element_address(data, offset, T_BYTE); 6127 assert(data_start, "data is NULL"); 6128 6129 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6130 OptoRuntime::ghash_processBlocks_Type(), 6131 stubAddr, stubName, TypePtr::BOTTOM, 6132 state_start, subkeyH_start, data_start, len); 6133 return true; 6134 } 6135 6136 bool LibraryCallKit::inline_base64_encodeBlock() { 6137 address stubAddr; 6138 const char *stubName; 6139 assert(UseBASE64Intrinsics, "need Base64 intrinsics support"); 6140 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters"); 6141 stubAddr = StubRoutines::base64_encodeBlock(); 6142 stubName = "encodeBlock"; 6143 6144 if (!stubAddr) return false; 6145 Node* base64obj = argument(0); 6146 Node* src = argument(1); 6147 Node* offset = argument(2); 6148 Node* len = argument(3); 6149 Node* dest = argument(4); 6150 Node* dp = argument(5); 6151 Node* isURL = argument(6); 6152 6153 src = must_be_not_null(src, true); 6154 dest = must_be_not_null(dest, true); 6155 6156 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 6157 assert(src_start, "source array is NULL"); 6158 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); 6159 assert(dest_start, "destination array is NULL"); 6160 6161 Node* base64 = make_runtime_call(RC_LEAF, 6162 OptoRuntime::base64_encodeBlock_Type(), 6163 stubAddr, stubName, TypePtr::BOTTOM, 6164 src_start, offset, len, dest_start, dp, isURL); 6165 return true; 6166 } 6167 6168 //------------------------------inline_digestBase_implCompress----------------------- 6169 // 6170 // Calculate MD5 for single-block byte[] array. 6171 // void com.sun.security.provider.MD5.implCompress(byte[] buf, int ofs) 6172 // 6173 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6174 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6175 // 6176 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6177 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6178 // 6179 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6180 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6181 // 6182 bool LibraryCallKit::inline_digestBase_implCompress(vmIntrinsics::ID id) { 6183 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6184 6185 Node* digestBase_obj = argument(0); 6186 Node* src = argument(1); // type oop 6187 Node* ofs = argument(2); // type int 6188 6189 const Type* src_type = src->Value(&_gvn); 6190 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6191 if (top_src == NULL || top_src->klass() == NULL) { 6192 // failed array check 6193 return false; 6194 } 6195 // Figure out the size and type of the elements we will be copying. 6196 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6197 if (src_elem != T_BYTE) { 6198 return false; 6199 } 6200 // 'src_start' points to src array + offset 6201 src = must_be_not_null(src, true); 6202 Node* src_start = array_element_address(src, ofs, src_elem); 6203 Node* state = NULL; 6204 address stubAddr; 6205 const char *stubName; 6206 6207 switch(id) { 6208 case vmIntrinsics::_md5_implCompress: 6209 assert(UseMD5Intrinsics, "need MD5 instruction support"); 6210 state = get_state_from_digest_object(digestBase_obj); 6211 stubAddr = StubRoutines::md5_implCompress(); 6212 stubName = "md5_implCompress"; 6213 break; 6214 case vmIntrinsics::_sha_implCompress: 6215 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6216 state = get_state_from_digest_object(digestBase_obj); 6217 stubAddr = StubRoutines::sha1_implCompress(); 6218 stubName = "sha1_implCompress"; 6219 break; 6220 case vmIntrinsics::_sha2_implCompress: 6221 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6222 state = get_state_from_digest_object(digestBase_obj); 6223 stubAddr = StubRoutines::sha256_implCompress(); 6224 stubName = "sha256_implCompress"; 6225 break; 6226 case vmIntrinsics::_sha5_implCompress: 6227 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6228 state = get_long_state_from_digest_object(digestBase_obj); 6229 stubAddr = StubRoutines::sha512_implCompress(); 6230 stubName = "sha512_implCompress"; 6231 break; 6232 default: 6233 fatal_unexpected_iid(id); 6234 return false; 6235 } 6236 if (state == NULL) return false; 6237 6238 assert(stubAddr != NULL, "Stub is generated"); 6239 if (stubAddr == NULL) return false; 6240 6241 // Call the stub. 6242 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::digestBase_implCompress_Type(), 6243 stubAddr, stubName, TypePtr::BOTTOM, 6244 src_start, state); 6245 6246 return true; 6247 } 6248 6249 //------------------------------inline_digestBase_implCompressMB----------------------- 6250 // 6251 // Calculate MD5/SHA/SHA2/SHA5 for multi-block byte[] array. 6252 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6253 // 6254 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6255 assert(UseMD5Intrinsics || UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6256 "need MD5/SHA1/SHA256/SHA512 instruction support"); 6257 assert((uint)predicate < 4, "sanity"); 6258 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6259 6260 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6261 Node* src = argument(1); // byte[] array 6262 Node* ofs = argument(2); // type int 6263 Node* limit = argument(3); // type int 6264 6265 const Type* src_type = src->Value(&_gvn); 6266 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6267 if (top_src == NULL || top_src->klass() == NULL) { 6268 // failed array check 6269 return false; 6270 } 6271 // Figure out the size and type of the elements we will be copying. 6272 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6273 if (src_elem != T_BYTE) { 6274 return false; 6275 } 6276 // 'src_start' points to src array + offset 6277 src = must_be_not_null(src, false); 6278 Node* src_start = array_element_address(src, ofs, src_elem); 6279 6280 const char* klass_digestBase_name = NULL; 6281 const char* stub_name = NULL; 6282 address stub_addr = NULL; 6283 bool long_state = false; 6284 6285 switch (predicate) { 6286 case 0: 6287 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_md5_implCompress)) { 6288 klass_digestBase_name = "sun/security/provider/MD5"; 6289 stub_name = "md5_implCompressMB"; 6290 stub_addr = StubRoutines::md5_implCompressMB(); 6291 } 6292 break; 6293 case 1: 6294 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha_implCompress)) { 6295 klass_digestBase_name = "sun/security/provider/SHA"; 6296 stub_name = "sha1_implCompressMB"; 6297 stub_addr = StubRoutines::sha1_implCompressMB(); 6298 } 6299 break; 6300 case 2: 6301 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha2_implCompress)) { 6302 klass_digestBase_name = "sun/security/provider/SHA2"; 6303 stub_name = "sha256_implCompressMB"; 6304 stub_addr = StubRoutines::sha256_implCompressMB(); 6305 } 6306 break; 6307 case 3: 6308 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha5_implCompress)) { 6309 klass_digestBase_name = "sun/security/provider/SHA5"; 6310 stub_name = "sha512_implCompressMB"; 6311 stub_addr = StubRoutines::sha512_implCompressMB(); 6312 long_state = true; 6313 } 6314 break; 6315 default: 6316 fatal("unknown DigestBase intrinsic predicate: %d", predicate); 6317 } 6318 if (klass_digestBase_name != NULL) { 6319 assert(stub_addr != NULL, "Stub is generated"); 6320 if (stub_addr == NULL) return false; 6321 6322 // get DigestBase klass to lookup for SHA klass 6323 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6324 assert(tinst != NULL, "digestBase_obj is not instance???"); 6325 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6326 6327 ciKlass* klass_digestBase = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_digestBase_name)); 6328 assert(klass_digestBase->is_loaded(), "predicate checks that this class is loaded"); 6329 ciInstanceKlass* instklass_digestBase = klass_digestBase->as_instance_klass(); 6330 return inline_digestBase_implCompressMB(digestBase_obj, instklass_digestBase, long_state, stub_addr, stub_name, src_start, ofs, limit); 6331 } 6332 return false; 6333 } 6334 6335 //------------------------------inline_digestBase_implCompressMB----------------------- 6336 bool LibraryCallKit::inline_digestBase_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_digestBase, 6337 bool long_state, address stubAddr, const char *stubName, 6338 Node* src_start, Node* ofs, Node* limit) { 6339 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_digestBase); 6340 const TypeOopPtr* xtype = aklass->as_instance_type(); 6341 Node* digest_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6342 digest_obj = _gvn.transform(digest_obj); 6343 6344 Node* state; 6345 if (long_state) { 6346 state = get_long_state_from_digest_object(digest_obj); 6347 } else { 6348 state = get_state_from_digest_object(digest_obj); 6349 } 6350 if (state == NULL) return false; 6351 6352 // Call the stub. 6353 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6354 OptoRuntime::digestBase_implCompressMB_Type(), 6355 stubAddr, stubName, TypePtr::BOTTOM, 6356 src_start, state, ofs, limit); 6357 // return ofs (int) 6358 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6359 set_result(result); 6360 6361 return true; 6362 } 6363 6364 //------------------------------get_state_from_digest_object----------------------- 6365 Node * LibraryCallKit::get_state_from_digest_object(Node *digest_object) { 6366 Node* digest_state = load_field_from_object(digest_object, "state", "[I", /*is_exact*/ false); 6367 assert (digest_state != NULL, "wrong version of sun.security.provider.MD5/SHA/SHA2"); 6368 if (digest_state == NULL) return (Node *) NULL; 6369 6370 // now have the array, need to get the start address of the state array 6371 Node* state = array_element_address(digest_state, intcon(0), T_INT); 6372 return state; 6373 } 6374 6375 //------------------------------get_long_state_from_digest_object----------------------- 6376 Node * LibraryCallKit::get_long_state_from_digest_object(Node *digest_object) { 6377 Node* digest_state = load_field_from_object(digest_object, "state", "[J", /*is_exact*/ false); 6378 assert (digest_state != NULL, "wrong version of sun.security.provider.SHA5"); 6379 if (digest_state == NULL) return (Node *) NULL; 6380 6381 // now have the array, need to get the start address of the state array 6382 Node* state = array_element_address(digest_state, intcon(0), T_LONG); 6383 return state; 6384 } 6385 6386 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6387 // Return node representing slow path of predicate check. 6388 // the pseudo code we want to emulate with this predicate is: 6389 // if (digestBaseObj instanceof MD5/SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6390 // 6391 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6392 assert(UseMD5Intrinsics || UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6393 "need MD5/SHA1/SHA256/SHA512 instruction support"); 6394 assert((uint)predicate < 4, "sanity"); 6395 6396 // The receiver was checked for NULL already. 6397 Node* digestBaseObj = argument(0); 6398 6399 // get DigestBase klass for instanceOf check 6400 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6401 assert(tinst != NULL, "digestBaseObj is null"); 6402 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6403 6404 const char* klass_name = NULL; 6405 switch (predicate) { 6406 case 0: 6407 if (UseMD5Intrinsics) { 6408 // we want to do an instanceof comparison against the MD5 class 6409 klass_name = "sun/security/provider/MD5"; 6410 } 6411 break; 6412 case 1: 6413 if (UseSHA1Intrinsics) { 6414 // we want to do an instanceof comparison against the SHA class 6415 klass_name = "sun/security/provider/SHA"; 6416 } 6417 break; 6418 case 2: 6419 if (UseSHA256Intrinsics) { 6420 // we want to do an instanceof comparison against the SHA2 class 6421 klass_name = "sun/security/provider/SHA2"; 6422 } 6423 break; 6424 case 3: 6425 if (UseSHA512Intrinsics) { 6426 // we want to do an instanceof comparison against the SHA5 class 6427 klass_name = "sun/security/provider/SHA5"; 6428 } 6429 break; 6430 default: 6431 fatal("unknown SHA intrinsic predicate: %d", predicate); 6432 } 6433 6434 ciKlass* klass = NULL; 6435 if (klass_name != NULL) { 6436 klass = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_name)); 6437 } 6438 if ((klass == NULL) || !klass->is_loaded()) { 6439 // if none of MD5/SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6440 Node* ctrl = control(); 6441 set_control(top()); // no intrinsic path 6442 return ctrl; 6443 } 6444 ciInstanceKlass* instklass = klass->as_instance_klass(); 6445 6446 Node* instof = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass))); 6447 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6448 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6449 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6450 6451 return instof_false; // even if it is NULL 6452 } 6453 6454 //-------------inline_fma----------------------------------- 6455 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 6456 Node *a = NULL; 6457 Node *b = NULL; 6458 Node *c = NULL; 6459 Node* result = NULL; 6460 switch (id) { 6461 case vmIntrinsics::_fmaD: 6462 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 6463 // no receiver since it is static method 6464 a = round_double_node(argument(0)); 6465 b = round_double_node(argument(2)); 6466 c = round_double_node(argument(4)); 6467 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 6468 break; 6469 case vmIntrinsics::_fmaF: 6470 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 6471 a = argument(0); 6472 b = argument(1); 6473 c = argument(2); 6474 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 6475 break; 6476 default: 6477 fatal_unexpected_iid(id); break; 6478 } 6479 set_result(result); 6480 return true; 6481 } 6482 6483 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) { 6484 // argument(0) is receiver 6485 Node* codePoint = argument(1); 6486 Node* n = NULL; 6487 6488 switch (id) { 6489 case vmIntrinsics::_isDigit : 6490 n = new DigitNode(control(), codePoint); 6491 break; 6492 case vmIntrinsics::_isLowerCase : 6493 n = new LowerCaseNode(control(), codePoint); 6494 break; 6495 case vmIntrinsics::_isUpperCase : 6496 n = new UpperCaseNode(control(), codePoint); 6497 break; 6498 case vmIntrinsics::_isWhitespace : 6499 n = new WhitespaceNode(control(), codePoint); 6500 break; 6501 default: 6502 fatal_unexpected_iid(id); 6503 } 6504 6505 set_result(_gvn.transform(n)); 6506 return true; 6507 } 6508 6509 //------------------------------inline_fp_min_max------------------------------ 6510 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) { 6511 /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416. 6512 6513 // The intrinsic should be used only when the API branches aren't predictable, 6514 // the last one performing the most important comparison. The following heuristic 6515 // uses the branch statistics to eventually bail out if necessary. 6516 6517 ciMethodData *md = callee()->method_data(); 6518 6519 if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) { 6520 ciCallProfile cp = caller()->call_profile_at_bci(bci()); 6521 6522 if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) { 6523 // Bail out if the call-site didn't contribute enough to the statistics. 6524 return false; 6525 } 6526 6527 uint taken = 0, not_taken = 0; 6528 6529 for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) { 6530 if (p->is_BranchData()) { 6531 taken = ((ciBranchData*)p)->taken(); 6532 not_taken = ((ciBranchData*)p)->not_taken(); 6533 } 6534 } 6535 6536 double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count()); 6537 balance = balance < 0 ? -balance : balance; 6538 if ( balance > 0.2 ) { 6539 // Bail out if the most important branch is predictable enough. 6540 return false; 6541 } 6542 } 6543 */ 6544 6545 Node *a = NULL; 6546 Node *b = NULL; 6547 Node *n = NULL; 6548 switch (id) { 6549 case vmIntrinsics::_maxF: 6550 case vmIntrinsics::_minF: 6551 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each."); 6552 a = argument(0); 6553 b = argument(1); 6554 break; 6555 case vmIntrinsics::_maxD: 6556 case vmIntrinsics::_minD: 6557 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each."); 6558 a = round_double_node(argument(0)); 6559 b = round_double_node(argument(2)); 6560 break; 6561 default: 6562 fatal_unexpected_iid(id); 6563 break; 6564 } 6565 switch (id) { 6566 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break; 6567 case vmIntrinsics::_minF: n = new MinFNode(a, b); break; 6568 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break; 6569 case vmIntrinsics::_minD: n = new MinDNode(a, b); break; 6570 default: fatal_unexpected_iid(id); break; 6571 } 6572 set_result(_gvn.transform(n)); 6573 return true; 6574 } 6575 6576 bool LibraryCallKit::inline_profileBoolean() { 6577 Node* counts = argument(1); 6578 const TypeAryPtr* ary = NULL; 6579 ciArray* aobj = NULL; 6580 if (counts->is_Con() 6581 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6582 && (aobj = ary->const_oop()->as_array()) != NULL 6583 && (aobj->length() == 2)) { 6584 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6585 jint false_cnt = aobj->element_value(0).as_int(); 6586 jint true_cnt = aobj->element_value(1).as_int(); 6587 6588 if (C->log() != NULL) { 6589 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6590 false_cnt, true_cnt); 6591 } 6592 6593 if (false_cnt + true_cnt == 0) { 6594 // According to profile, never executed. 6595 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6596 Deoptimization::Action_reinterpret); 6597 return true; 6598 } 6599 6600 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6601 // is a number of each value occurrences. 6602 Node* result = argument(0); 6603 if (false_cnt == 0 || true_cnt == 0) { 6604 // According to profile, one value has been never seen. 6605 int expected_val = (false_cnt == 0) ? 1 : 0; 6606 6607 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6608 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6609 6610 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6611 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6612 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6613 6614 { // Slow path: uncommon trap for never seen value and then reexecute 6615 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6616 // the value has been seen at least once. 6617 PreserveJVMState pjvms(this); 6618 PreserveReexecuteState preexecs(this); 6619 jvms()->set_should_reexecute(true); 6620 6621 set_control(slow_path); 6622 set_i_o(i_o()); 6623 6624 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6625 Deoptimization::Action_reinterpret); 6626 } 6627 // The guard for never seen value enables sharpening of the result and 6628 // returning a constant. It allows to eliminate branches on the same value 6629 // later on. 6630 set_control(fast_path); 6631 result = intcon(expected_val); 6632 } 6633 // Stop profiling. 6634 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6635 // By replacing method body with profile data (represented as ProfileBooleanNode 6636 // on IR level) we effectively disable profiling. 6637 // It enables full speed execution once optimized code is generated. 6638 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6639 C->record_for_igvn(profile); 6640 set_result(profile); 6641 return true; 6642 } else { 6643 // Continue profiling. 6644 // Profile data isn't available at the moment. So, execute method's bytecode version. 6645 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6646 // is compiled and counters aren't available since corresponding MethodHandle 6647 // isn't a compile-time constant. 6648 return false; 6649 } 6650 } 6651 6652 bool LibraryCallKit::inline_isCompileConstant() { 6653 Node* n = argument(0); 6654 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6655 return true; 6656 }