annotate src/share/vm/opto/library_call.cpp @ 423:a1980da045cc

6462850: generate biased locking code in C2 ideal graph Summary: Inline biased locking code in C2 ideal graph during macro nodes expansion Reviewed-by: never
author kvn
date Fri, 07 Nov 2008 09:29:38 -0800
parents 9c2ecc2ffb12
children 98cb887364d3
rev   line source
duke@0 1 /*
xdono@196 2 * Copyright 1999-2008 Sun Microsystems, Inc. All Rights Reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
duke@0 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@0 20 * CA 95054 USA or visit www.sun.com if you need additional information or
duke@0 21 * have any questions.
duke@0 22 *
duke@0 23 */
duke@0 24
duke@0 25 #include "incls/_precompiled.incl"
duke@0 26 #include "incls/_library_call.cpp.incl"
duke@0 27
duke@0 28 class LibraryIntrinsic : public InlineCallGenerator {
duke@0 29 // Extend the set of intrinsics known to the runtime:
duke@0 30 public:
duke@0 31 private:
duke@0 32 bool _is_virtual;
duke@0 33 vmIntrinsics::ID _intrinsic_id;
duke@0 34
duke@0 35 public:
duke@0 36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
duke@0 37 : InlineCallGenerator(m),
duke@0 38 _is_virtual(is_virtual),
duke@0 39 _intrinsic_id(id)
duke@0 40 {
duke@0 41 }
duke@0 42 virtual bool is_intrinsic() const { return true; }
duke@0 43 virtual bool is_virtual() const { return _is_virtual; }
duke@0 44 virtual JVMState* generate(JVMState* jvms);
duke@0 45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
duke@0 46 };
duke@0 47
duke@0 48
duke@0 49 // Local helper class for LibraryIntrinsic:
duke@0 50 class LibraryCallKit : public GraphKit {
duke@0 51 private:
duke@0 52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
duke@0 53
duke@0 54 public:
duke@0 55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
duke@0 56 : GraphKit(caller),
duke@0 57 _intrinsic(intrinsic)
duke@0 58 {
duke@0 59 }
duke@0 60
duke@0 61 ciMethod* caller() const { return jvms()->method(); }
duke@0 62 int bci() const { return jvms()->bci(); }
duke@0 63 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
duke@0 64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
duke@0 65 ciMethod* callee() const { return _intrinsic->method(); }
duke@0 66 ciSignature* signature() const { return callee()->signature(); }
duke@0 67 int arg_size() const { return callee()->arg_size(); }
duke@0 68
duke@0 69 bool try_to_inline();
duke@0 70
duke@0 71 // Helper functions to inline natives
duke@0 72 void push_result(RegionNode* region, PhiNode* value);
duke@0 73 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
duke@0 74 Node* generate_slow_guard(Node* test, RegionNode* region);
duke@0 75 Node* generate_fair_guard(Node* test, RegionNode* region);
duke@0 76 Node* generate_negative_guard(Node* index, RegionNode* region,
duke@0 77 // resulting CastII of index:
duke@0 78 Node* *pos_index = NULL);
duke@0 79 Node* generate_nonpositive_guard(Node* index, bool never_negative,
duke@0 80 // resulting CastII of index:
duke@0 81 Node* *pos_index = NULL);
duke@0 82 Node* generate_limit_guard(Node* offset, Node* subseq_length,
duke@0 83 Node* array_length,
duke@0 84 RegionNode* region);
duke@0 85 Node* generate_current_thread(Node* &tls_output);
duke@0 86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
duke@0 87 bool disjoint_bases, const char* &name);
duke@0 88 Node* load_mirror_from_klass(Node* klass);
duke@0 89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
duke@0 90 int nargs,
duke@0 91 RegionNode* region, int null_path,
duke@0 92 int offset);
duke@0 93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
duke@0 94 RegionNode* region, int null_path) {
duke@0 95 int offset = java_lang_Class::klass_offset_in_bytes();
duke@0 96 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
duke@0 97 region, null_path,
duke@0 98 offset);
duke@0 99 }
duke@0 100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
duke@0 101 int nargs,
duke@0 102 RegionNode* region, int null_path) {
duke@0 103 int offset = java_lang_Class::array_klass_offset_in_bytes();
duke@0 104 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
duke@0 105 region, null_path,
duke@0 106 offset);
duke@0 107 }
duke@0 108 Node* generate_access_flags_guard(Node* kls,
duke@0 109 int modifier_mask, int modifier_bits,
duke@0 110 RegionNode* region);
duke@0 111 Node* generate_interface_guard(Node* kls, RegionNode* region);
duke@0 112 Node* generate_array_guard(Node* kls, RegionNode* region) {
duke@0 113 return generate_array_guard_common(kls, region, false, false);
duke@0 114 }
duke@0 115 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
duke@0 116 return generate_array_guard_common(kls, region, false, true);
duke@0 117 }
duke@0 118 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
duke@0 119 return generate_array_guard_common(kls, region, true, false);
duke@0 120 }
duke@0 121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
duke@0 122 return generate_array_guard_common(kls, region, true, true);
duke@0 123 }
duke@0 124 Node* generate_array_guard_common(Node* kls, RegionNode* region,
duke@0 125 bool obj_array, bool not_array);
duke@0 126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
duke@0 127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
duke@0 128 bool is_virtual = false, bool is_static = false);
duke@0 129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
duke@0 130 return generate_method_call(method_id, false, true);
duke@0 131 }
duke@0 132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
duke@0 133 return generate_method_call(method_id, true, false);
duke@0 134 }
duke@0 135
duke@0 136 bool inline_string_compareTo();
duke@0 137 bool inline_string_indexOf();
duke@0 138 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
duke@0 139 Node* pop_math_arg();
duke@0 140 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
duke@0 141 bool inline_math_native(vmIntrinsics::ID id);
duke@0 142 bool inline_trig(vmIntrinsics::ID id);
duke@0 143 bool inline_trans(vmIntrinsics::ID id);
duke@0 144 bool inline_abs(vmIntrinsics::ID id);
duke@0 145 bool inline_sqrt(vmIntrinsics::ID id);
duke@0 146 bool inline_pow(vmIntrinsics::ID id);
duke@0 147 bool inline_exp(vmIntrinsics::ID id);
duke@0 148 bool inline_min_max(vmIntrinsics::ID id);
duke@0 149 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
duke@0 150 // This returns Type::AnyPtr, RawPtr, or OopPtr.
duke@0 151 int classify_unsafe_addr(Node* &base, Node* &offset);
duke@0 152 Node* make_unsafe_address(Node* base, Node* offset);
duke@0 153 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
duke@0 154 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
duke@0 155 bool inline_unsafe_allocate();
duke@0 156 bool inline_unsafe_copyMemory();
duke@0 157 bool inline_native_currentThread();
duke@0 158 bool inline_native_time_funcs(bool isNano);
duke@0 159 bool inline_native_isInterrupted();
duke@0 160 bool inline_native_Class_query(vmIntrinsics::ID id);
duke@0 161 bool inline_native_subtype_check();
duke@0 162
duke@0 163 bool inline_native_newArray();
duke@0 164 bool inline_native_getLength();
duke@0 165 bool inline_array_copyOf(bool is_copyOfRange);
rasbold@169 166 bool inline_array_equals();
duke@0 167 bool inline_native_clone(bool is_virtual);
duke@0 168 bool inline_native_Reflection_getCallerClass();
duke@0 169 bool inline_native_AtomicLong_get();
duke@0 170 bool inline_native_AtomicLong_attemptUpdate();
duke@0 171 bool is_method_invoke_or_aux_frame(JVMState* jvms);
duke@0 172 // Helper function for inlining native object hash method
duke@0 173 bool inline_native_hashcode(bool is_virtual, bool is_static);
duke@0 174 bool inline_native_getClass();
duke@0 175
duke@0 176 // Helper functions for inlining arraycopy
duke@0 177 bool inline_arraycopy();
duke@0 178 void generate_arraycopy(const TypePtr* adr_type,
duke@0 179 BasicType basic_elem_type,
duke@0 180 Node* src, Node* src_offset,
duke@0 181 Node* dest, Node* dest_offset,
duke@0 182 Node* copy_length,
duke@0 183 int nargs, // arguments on stack for debug info
duke@0 184 bool disjoint_bases = false,
duke@0 185 bool length_never_negative = false,
duke@0 186 RegionNode* slow_region = NULL);
duke@0 187 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
duke@0 188 RegionNode* slow_region);
duke@0 189 void generate_clear_array(const TypePtr* adr_type,
duke@0 190 Node* dest,
duke@0 191 BasicType basic_elem_type,
duke@0 192 Node* slice_off,
duke@0 193 Node* slice_len,
duke@0 194 Node* slice_end);
duke@0 195 bool generate_block_arraycopy(const TypePtr* adr_type,
duke@0 196 BasicType basic_elem_type,
duke@0 197 AllocateNode* alloc,
duke@0 198 Node* src, Node* src_offset,
duke@0 199 Node* dest, Node* dest_offset,
duke@0 200 Node* dest_size);
duke@0 201 void generate_slow_arraycopy(const TypePtr* adr_type,
duke@0 202 Node* src, Node* src_offset,
duke@0 203 Node* dest, Node* dest_offset,
duke@0 204 Node* copy_length,
duke@0 205 int nargs);
duke@0 206 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
duke@0 207 Node* dest_elem_klass,
duke@0 208 Node* src, Node* src_offset,
duke@0 209 Node* dest, Node* dest_offset,
duke@0 210 Node* copy_length, int nargs);
duke@0 211 Node* generate_generic_arraycopy(const TypePtr* adr_type,
duke@0 212 Node* src, Node* src_offset,
duke@0 213 Node* dest, Node* dest_offset,
duke@0 214 Node* copy_length, int nargs);
duke@0 215 void generate_unchecked_arraycopy(const TypePtr* adr_type,
duke@0 216 BasicType basic_elem_type,
duke@0 217 bool disjoint_bases,
duke@0 218 Node* src, Node* src_offset,
duke@0 219 Node* dest, Node* dest_offset,
duke@0 220 Node* copy_length);
duke@0 221 bool inline_unsafe_CAS(BasicType type);
duke@0 222 bool inline_unsafe_ordered_store(BasicType type);
duke@0 223 bool inline_fp_conversions(vmIntrinsics::ID id);
duke@0 224 bool inline_reverseBytes(vmIntrinsics::ID id);
duke@0 225 };
duke@0 226
duke@0 227
duke@0 228 //---------------------------make_vm_intrinsic----------------------------
duke@0 229 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
duke@0 230 vmIntrinsics::ID id = m->intrinsic_id();
duke@0 231 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
duke@0 232
duke@0 233 if (DisableIntrinsic[0] != '\0'
duke@0 234 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
duke@0 235 // disabled by a user request on the command line:
duke@0 236 // example: -XX:DisableIntrinsic=_hashCode,_getClass
duke@0 237 return NULL;
duke@0 238 }
duke@0 239
duke@0 240 if (!m->is_loaded()) {
duke@0 241 // do not attempt to inline unloaded methods
duke@0 242 return NULL;
duke@0 243 }
duke@0 244
duke@0 245 // Only a few intrinsics implement a virtual dispatch.
duke@0 246 // They are expensive calls which are also frequently overridden.
duke@0 247 if (is_virtual) {
duke@0 248 switch (id) {
duke@0 249 case vmIntrinsics::_hashCode:
duke@0 250 case vmIntrinsics::_clone:
duke@0 251 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
duke@0 252 break;
duke@0 253 default:
duke@0 254 return NULL;
duke@0 255 }
duke@0 256 }
duke@0 257
duke@0 258 // -XX:-InlineNatives disables nearly all intrinsics:
duke@0 259 if (!InlineNatives) {
duke@0 260 switch (id) {
duke@0 261 case vmIntrinsics::_indexOf:
duke@0 262 case vmIntrinsics::_compareTo:
rasbold@169 263 case vmIntrinsics::_equalsC:
duke@0 264 break; // InlineNatives does not control String.compareTo
duke@0 265 default:
duke@0 266 return NULL;
duke@0 267 }
duke@0 268 }
duke@0 269
duke@0 270 switch (id) {
duke@0 271 case vmIntrinsics::_compareTo:
duke@0 272 if (!SpecialStringCompareTo) return NULL;
duke@0 273 break;
duke@0 274 case vmIntrinsics::_indexOf:
duke@0 275 if (!SpecialStringIndexOf) return NULL;
duke@0 276 break;
rasbold@169 277 case vmIntrinsics::_equalsC:
rasbold@169 278 if (!SpecialArraysEquals) return NULL;
rasbold@169 279 break;
duke@0 280 case vmIntrinsics::_arraycopy:
duke@0 281 if (!InlineArrayCopy) return NULL;
duke@0 282 break;
duke@0 283 case vmIntrinsics::_copyMemory:
duke@0 284 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
duke@0 285 if (!InlineArrayCopy) return NULL;
duke@0 286 break;
duke@0 287 case vmIntrinsics::_hashCode:
duke@0 288 if (!InlineObjectHash) return NULL;
duke@0 289 break;
duke@0 290 case vmIntrinsics::_clone:
duke@0 291 case vmIntrinsics::_copyOf:
duke@0 292 case vmIntrinsics::_copyOfRange:
duke@0 293 if (!InlineObjectCopy) return NULL;
duke@0 294 // These also use the arraycopy intrinsic mechanism:
duke@0 295 if (!InlineArrayCopy) return NULL;
duke@0 296 break;
duke@0 297 case vmIntrinsics::_checkIndex:
duke@0 298 // We do not intrinsify this. The optimizer does fine with it.
duke@0 299 return NULL;
duke@0 300
duke@0 301 case vmIntrinsics::_get_AtomicLong:
duke@0 302 case vmIntrinsics::_attemptUpdate:
duke@0 303 if (!InlineAtomicLong) return NULL;
duke@0 304 break;
duke@0 305
duke@0 306 case vmIntrinsics::_Object_init:
duke@0 307 case vmIntrinsics::_invoke:
duke@0 308 // We do not intrinsify these; they are marked for other purposes.
duke@0 309 return NULL;
duke@0 310
duke@0 311 case vmIntrinsics::_getCallerClass:
duke@0 312 if (!UseNewReflection) return NULL;
duke@0 313 if (!InlineReflectionGetCallerClass) return NULL;
duke@0 314 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
duke@0 315 break;
duke@0 316
duke@0 317 default:
duke@0 318 break;
duke@0 319 }
duke@0 320
duke@0 321 // -XX:-InlineClassNatives disables natives from the Class class.
duke@0 322 // The flag applies to all reflective calls, notably Array.newArray
duke@0 323 // (visible to Java programmers as Array.newInstance).
duke@0 324 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
duke@0 325 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
duke@0 326 if (!InlineClassNatives) return NULL;
duke@0 327 }
duke@0 328
duke@0 329 // -XX:-InlineThreadNatives disables natives from the Thread class.
duke@0 330 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
duke@0 331 if (!InlineThreadNatives) return NULL;
duke@0 332 }
duke@0 333
duke@0 334 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
duke@0 335 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
duke@0 336 m->holder()->name() == ciSymbol::java_lang_Float() ||
duke@0 337 m->holder()->name() == ciSymbol::java_lang_Double()) {
duke@0 338 if (!InlineMathNatives) return NULL;
duke@0 339 }
duke@0 340
duke@0 341 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
duke@0 342 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
duke@0 343 if (!InlineUnsafeOps) return NULL;
duke@0 344 }
duke@0 345
duke@0 346 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
duke@0 347 }
duke@0 348
duke@0 349 //----------------------register_library_intrinsics-----------------------
duke@0 350 // Initialize this file's data structures, for each Compile instance.
duke@0 351 void Compile::register_library_intrinsics() {
duke@0 352 // Nothing to do here.
duke@0 353 }
duke@0 354
duke@0 355 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
duke@0 356 LibraryCallKit kit(jvms, this);
duke@0 357 Compile* C = kit.C;
duke@0 358 int nodes = C->unique();
duke@0 359 #ifndef PRODUCT
duke@0 360 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
duke@0 361 char buf[1000];
duke@0 362 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
duke@0 363 tty->print_cr("Intrinsic %s", str);
duke@0 364 }
duke@0 365 #endif
duke@0 366 if (kit.try_to_inline()) {
duke@0 367 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
duke@0 368 tty->print("Inlining intrinsic %s%s at bci:%d in",
duke@0 369 vmIntrinsics::name_at(intrinsic_id()),
duke@0 370 (is_virtual() ? " (virtual)" : ""), kit.bci());
duke@0 371 kit.caller()->print_short_name(tty);
duke@0 372 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
duke@0 373 }
duke@0 374 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
duke@0 375 if (C->log()) {
duke@0 376 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
duke@0 377 vmIntrinsics::name_at(intrinsic_id()),
duke@0 378 (is_virtual() ? " virtual='1'" : ""),
duke@0 379 C->unique() - nodes);
duke@0 380 }
duke@0 381 return kit.transfer_exceptions_into_jvms();
duke@0 382 }
duke@0 383
duke@0 384 if (PrintIntrinsics) {
duke@0 385 switch (intrinsic_id()) {
duke@0 386 case vmIntrinsics::_invoke:
duke@0 387 case vmIntrinsics::_Object_init:
duke@0 388 // We do not expect to inline these, so do not produce any noise about them.
duke@0 389 break;
duke@0 390 default:
duke@0 391 tty->print("Did not inline intrinsic %s%s at bci:%d in",
duke@0 392 vmIntrinsics::name_at(intrinsic_id()),
duke@0 393 (is_virtual() ? " (virtual)" : ""), kit.bci());
duke@0 394 kit.caller()->print_short_name(tty);
duke@0 395 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
duke@0 396 }
duke@0 397 }
duke@0 398 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
duke@0 399 return NULL;
duke@0 400 }
duke@0 401
duke@0 402 bool LibraryCallKit::try_to_inline() {
duke@0 403 // Handle symbolic names for otherwise undistinguished boolean switches:
duke@0 404 const bool is_store = true;
duke@0 405 const bool is_native_ptr = true;
duke@0 406 const bool is_static = true;
duke@0 407
duke@0 408 switch (intrinsic_id()) {
duke@0 409 case vmIntrinsics::_hashCode:
duke@0 410 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
duke@0 411 case vmIntrinsics::_identityHashCode:
duke@0 412 return inline_native_hashcode(/*!virtual*/ false, is_static);
duke@0 413 case vmIntrinsics::_getClass:
duke@0 414 return inline_native_getClass();
duke@0 415
duke@0 416 case vmIntrinsics::_dsin:
duke@0 417 case vmIntrinsics::_dcos:
duke@0 418 case vmIntrinsics::_dtan:
duke@0 419 case vmIntrinsics::_dabs:
duke@0 420 case vmIntrinsics::_datan2:
duke@0 421 case vmIntrinsics::_dsqrt:
duke@0 422 case vmIntrinsics::_dexp:
duke@0 423 case vmIntrinsics::_dlog:
duke@0 424 case vmIntrinsics::_dlog10:
duke@0 425 case vmIntrinsics::_dpow:
duke@0 426 return inline_math_native(intrinsic_id());
duke@0 427
duke@0 428 case vmIntrinsics::_min:
duke@0 429 case vmIntrinsics::_max:
duke@0 430 return inline_min_max(intrinsic_id());
duke@0 431
duke@0 432 case vmIntrinsics::_arraycopy:
duke@0 433 return inline_arraycopy();
duke@0 434
duke@0 435 case vmIntrinsics::_compareTo:
duke@0 436 return inline_string_compareTo();
duke@0 437 case vmIntrinsics::_indexOf:
duke@0 438 return inline_string_indexOf();
duke@0 439
duke@0 440 case vmIntrinsics::_getObject:
duke@0 441 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
duke@0 442 case vmIntrinsics::_getBoolean:
duke@0 443 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
duke@0 444 case vmIntrinsics::_getByte:
duke@0 445 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
duke@0 446 case vmIntrinsics::_getShort:
duke@0 447 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
duke@0 448 case vmIntrinsics::_getChar:
duke@0 449 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
duke@0 450 case vmIntrinsics::_getInt:
duke@0 451 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
duke@0 452 case vmIntrinsics::_getLong:
duke@0 453 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
duke@0 454 case vmIntrinsics::_getFloat:
duke@0 455 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
duke@0 456 case vmIntrinsics::_getDouble:
duke@0 457 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
duke@0 458
duke@0 459 case vmIntrinsics::_putObject:
duke@0 460 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
duke@0 461 case vmIntrinsics::_putBoolean:
duke@0 462 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
duke@0 463 case vmIntrinsics::_putByte:
duke@0 464 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
duke@0 465 case vmIntrinsics::_putShort:
duke@0 466 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
duke@0 467 case vmIntrinsics::_putChar:
duke@0 468 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
duke@0 469 case vmIntrinsics::_putInt:
duke@0 470 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
duke@0 471 case vmIntrinsics::_putLong:
duke@0 472 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
duke@0 473 case vmIntrinsics::_putFloat:
duke@0 474 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
duke@0 475 case vmIntrinsics::_putDouble:
duke@0 476 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
duke@0 477
duke@0 478 case vmIntrinsics::_getByte_raw:
duke@0 479 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
duke@0 480 case vmIntrinsics::_getShort_raw:
duke@0 481 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
duke@0 482 case vmIntrinsics::_getChar_raw:
duke@0 483 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
duke@0 484 case vmIntrinsics::_getInt_raw:
duke@0 485 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
duke@0 486 case vmIntrinsics::_getLong_raw:
duke@0 487 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
duke@0 488 case vmIntrinsics::_getFloat_raw:
duke@0 489 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
duke@0 490 case vmIntrinsics::_getDouble_raw:
duke@0 491 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
duke@0 492 case vmIntrinsics::_getAddress_raw:
duke@0 493 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
duke@0 494
duke@0 495 case vmIntrinsics::_putByte_raw:
duke@0 496 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
duke@0 497 case vmIntrinsics::_putShort_raw:
duke@0 498 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
duke@0 499 case vmIntrinsics::_putChar_raw:
duke@0 500 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
duke@0 501 case vmIntrinsics::_putInt_raw:
duke@0 502 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
duke@0 503 case vmIntrinsics::_putLong_raw:
duke@0 504 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
duke@0 505 case vmIntrinsics::_putFloat_raw:
duke@0 506 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
duke@0 507 case vmIntrinsics::_putDouble_raw:
duke@0 508 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
duke@0 509 case vmIntrinsics::_putAddress_raw:
duke@0 510 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
duke@0 511
duke@0 512 case vmIntrinsics::_getObjectVolatile:
duke@0 513 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
duke@0 514 case vmIntrinsics::_getBooleanVolatile:
duke@0 515 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
duke@0 516 case vmIntrinsics::_getByteVolatile:
duke@0 517 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
duke@0 518 case vmIntrinsics::_getShortVolatile:
duke@0 519 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
duke@0 520 case vmIntrinsics::_getCharVolatile:
duke@0 521 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
duke@0 522 case vmIntrinsics::_getIntVolatile:
duke@0 523 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
duke@0 524 case vmIntrinsics::_getLongVolatile:
duke@0 525 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
duke@0 526 case vmIntrinsics::_getFloatVolatile:
duke@0 527 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
duke@0 528 case vmIntrinsics::_getDoubleVolatile:
duke@0 529 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
duke@0 530
duke@0 531 case vmIntrinsics::_putObjectVolatile:
duke@0 532 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
duke@0 533 case vmIntrinsics::_putBooleanVolatile:
duke@0 534 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
duke@0 535 case vmIntrinsics::_putByteVolatile:
duke@0 536 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
duke@0 537 case vmIntrinsics::_putShortVolatile:
duke@0 538 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
duke@0 539 case vmIntrinsics::_putCharVolatile:
duke@0 540 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
duke@0 541 case vmIntrinsics::_putIntVolatile:
duke@0 542 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
duke@0 543 case vmIntrinsics::_putLongVolatile:
duke@0 544 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
duke@0 545 case vmIntrinsics::_putFloatVolatile:
duke@0 546 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
duke@0 547 case vmIntrinsics::_putDoubleVolatile:
duke@0 548 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
duke@0 549
duke@0 550 case vmIntrinsics::_prefetchRead:
duke@0 551 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
duke@0 552 case vmIntrinsics::_prefetchWrite:
duke@0 553 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
duke@0 554 case vmIntrinsics::_prefetchReadStatic:
duke@0 555 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
duke@0 556 case vmIntrinsics::_prefetchWriteStatic:
duke@0 557 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
duke@0 558
duke@0 559 case vmIntrinsics::_compareAndSwapObject:
duke@0 560 return inline_unsafe_CAS(T_OBJECT);
duke@0 561 case vmIntrinsics::_compareAndSwapInt:
duke@0 562 return inline_unsafe_CAS(T_INT);
duke@0 563 case vmIntrinsics::_compareAndSwapLong:
duke@0 564 return inline_unsafe_CAS(T_LONG);
duke@0 565
duke@0 566 case vmIntrinsics::_putOrderedObject:
duke@0 567 return inline_unsafe_ordered_store(T_OBJECT);
duke@0 568 case vmIntrinsics::_putOrderedInt:
duke@0 569 return inline_unsafe_ordered_store(T_INT);
duke@0 570 case vmIntrinsics::_putOrderedLong:
duke@0 571 return inline_unsafe_ordered_store(T_LONG);
duke@0 572
duke@0 573 case vmIntrinsics::_currentThread:
duke@0 574 return inline_native_currentThread();
duke@0 575 case vmIntrinsics::_isInterrupted:
duke@0 576 return inline_native_isInterrupted();
duke@0 577
duke@0 578 case vmIntrinsics::_currentTimeMillis:
duke@0 579 return inline_native_time_funcs(false);
duke@0 580 case vmIntrinsics::_nanoTime:
duke@0 581 return inline_native_time_funcs(true);
duke@0 582 case vmIntrinsics::_allocateInstance:
duke@0 583 return inline_unsafe_allocate();
duke@0 584 case vmIntrinsics::_copyMemory:
duke@0 585 return inline_unsafe_copyMemory();
duke@0 586 case vmIntrinsics::_newArray:
duke@0 587 return inline_native_newArray();
duke@0 588 case vmIntrinsics::_getLength:
duke@0 589 return inline_native_getLength();
duke@0 590 case vmIntrinsics::_copyOf:
duke@0 591 return inline_array_copyOf(false);
duke@0 592 case vmIntrinsics::_copyOfRange:
duke@0 593 return inline_array_copyOf(true);
rasbold@169 594 case vmIntrinsics::_equalsC:
rasbold@169 595 return inline_array_equals();
duke@0 596 case vmIntrinsics::_clone:
duke@0 597 return inline_native_clone(intrinsic()->is_virtual());
duke@0 598
duke@0 599 case vmIntrinsics::_isAssignableFrom:
duke@0 600 return inline_native_subtype_check();
duke@0 601
duke@0 602 case vmIntrinsics::_isInstance:
duke@0 603 case vmIntrinsics::_getModifiers:
duke@0 604 case vmIntrinsics::_isInterface:
duke@0 605 case vmIntrinsics::_isArray:
duke@0 606 case vmIntrinsics::_isPrimitive:
duke@0 607 case vmIntrinsics::_getSuperclass:
duke@0 608 case vmIntrinsics::_getComponentType:
duke@0 609 case vmIntrinsics::_getClassAccessFlags:
duke@0 610 return inline_native_Class_query(intrinsic_id());
duke@0 611
duke@0 612 case vmIntrinsics::_floatToRawIntBits:
duke@0 613 case vmIntrinsics::_floatToIntBits:
duke@0 614 case vmIntrinsics::_intBitsToFloat:
duke@0 615 case vmIntrinsics::_doubleToRawLongBits:
duke@0 616 case vmIntrinsics::_doubleToLongBits:
duke@0 617 case vmIntrinsics::_longBitsToDouble:
duke@0 618 return inline_fp_conversions(intrinsic_id());
duke@0 619
duke@0 620 case vmIntrinsics::_reverseBytes_i:
duke@0 621 case vmIntrinsics::_reverseBytes_l:
duke@0 622 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
duke@0 623
duke@0 624 case vmIntrinsics::_get_AtomicLong:
duke@0 625 return inline_native_AtomicLong_get();
duke@0 626 case vmIntrinsics::_attemptUpdate:
duke@0 627 return inline_native_AtomicLong_attemptUpdate();
duke@0 628
duke@0 629 case vmIntrinsics::_getCallerClass:
duke@0 630 return inline_native_Reflection_getCallerClass();
duke@0 631
duke@0 632 default:
duke@0 633 // If you get here, it may be that someone has added a new intrinsic
duke@0 634 // to the list in vmSymbols.hpp without implementing it here.
duke@0 635 #ifndef PRODUCT
duke@0 636 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
duke@0 637 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
duke@0 638 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
duke@0 639 }
duke@0 640 #endif
duke@0 641 return false;
duke@0 642 }
duke@0 643 }
duke@0 644
duke@0 645 //------------------------------push_result------------------------------
duke@0 646 // Helper function for finishing intrinsics.
duke@0 647 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
duke@0 648 record_for_igvn(region);
duke@0 649 set_control(_gvn.transform(region));
duke@0 650 BasicType value_type = value->type()->basic_type();
duke@0 651 push_node(value_type, _gvn.transform(value));
duke@0 652 }
duke@0 653
duke@0 654 //------------------------------generate_guard---------------------------
duke@0 655 // Helper function for generating guarded fast-slow graph structures.
duke@0 656 // The given 'test', if true, guards a slow path. If the test fails
duke@0 657 // then a fast path can be taken. (We generally hope it fails.)
duke@0 658 // In all cases, GraphKit::control() is updated to the fast path.
duke@0 659 // The returned value represents the control for the slow path.
duke@0 660 // The return value is never 'top'; it is either a valid control
duke@0 661 // or NULL if it is obvious that the slow path can never be taken.
duke@0 662 // Also, if region and the slow control are not NULL, the slow edge
duke@0 663 // is appended to the region.
duke@0 664 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
duke@0 665 if (stopped()) {
duke@0 666 // Already short circuited.
duke@0 667 return NULL;
duke@0 668 }
duke@0 669
duke@0 670 // Build an if node and its projections.
duke@0 671 // If test is true we take the slow path, which we assume is uncommon.
duke@0 672 if (_gvn.type(test) == TypeInt::ZERO) {
duke@0 673 // The slow branch is never taken. No need to build this guard.
duke@0 674 return NULL;
duke@0 675 }
duke@0 676
duke@0 677 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
duke@0 678
duke@0 679 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
duke@0 680 if (if_slow == top()) {
duke@0 681 // The slow branch is never taken. No need to build this guard.
duke@0 682 return NULL;
duke@0 683 }
duke@0 684
duke@0 685 if (region != NULL)
duke@0 686 region->add_req(if_slow);
duke@0 687
duke@0 688 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
duke@0 689 set_control(if_fast);
duke@0 690
duke@0 691 return if_slow;
duke@0 692 }
duke@0 693
duke@0 694 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
duke@0 695 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
duke@0 696 }
duke@0 697 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
duke@0 698 return generate_guard(test, region, PROB_FAIR);
duke@0 699 }
duke@0 700
duke@0 701 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
duke@0 702 Node* *pos_index) {
duke@0 703 if (stopped())
duke@0 704 return NULL; // already stopped
duke@0 705 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
duke@0 706 return NULL; // index is already adequately typed
duke@0 707 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
duke@0 708 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
duke@0 709 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
duke@0 710 if (is_neg != NULL && pos_index != NULL) {
duke@0 711 // Emulate effect of Parse::adjust_map_after_if.
duke@0 712 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
duke@0 713 ccast->set_req(0, control());
duke@0 714 (*pos_index) = _gvn.transform(ccast);
duke@0 715 }
duke@0 716 return is_neg;
duke@0 717 }
duke@0 718
duke@0 719 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
duke@0 720 Node* *pos_index) {
duke@0 721 if (stopped())
duke@0 722 return NULL; // already stopped
duke@0 723 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
duke@0 724 return NULL; // index is already adequately typed
duke@0 725 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
duke@0 726 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
duke@0 727 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
duke@0 728 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
duke@0 729 if (is_notp != NULL && pos_index != NULL) {
duke@0 730 // Emulate effect of Parse::adjust_map_after_if.
duke@0 731 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
duke@0 732 ccast->set_req(0, control());
duke@0 733 (*pos_index) = _gvn.transform(ccast);
duke@0 734 }
duke@0 735 return is_notp;
duke@0 736 }
duke@0 737
duke@0 738 // Make sure that 'position' is a valid limit index, in [0..length].
duke@0 739 // There are two equivalent plans for checking this:
duke@0 740 // A. (offset + copyLength) unsigned<= arrayLength
duke@0 741 // B. offset <= (arrayLength - copyLength)
duke@0 742 // We require that all of the values above, except for the sum and
duke@0 743 // difference, are already known to be non-negative.
duke@0 744 // Plan A is robust in the face of overflow, if offset and copyLength
duke@0 745 // are both hugely positive.
duke@0 746 //
duke@0 747 // Plan B is less direct and intuitive, but it does not overflow at
duke@0 748 // all, since the difference of two non-negatives is always
duke@0 749 // representable. Whenever Java methods must perform the equivalent
duke@0 750 // check they generally use Plan B instead of Plan A.
duke@0 751 // For the moment we use Plan A.
duke@0 752 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
duke@0 753 Node* subseq_length,
duke@0 754 Node* array_length,
duke@0 755 RegionNode* region) {
duke@0 756 if (stopped())
duke@0 757 return NULL; // already stopped
duke@0 758 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
duke@0 759 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
duke@0 760 return NULL; // common case of whole-array copy
duke@0 761 Node* last = subseq_length;
duke@0 762 if (!zero_offset) // last += offset
duke@0 763 last = _gvn.transform( new (C, 3) AddINode(last, offset));
duke@0 764 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
duke@0 765 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
duke@0 766 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
duke@0 767 return is_over;
duke@0 768 }
duke@0 769
duke@0 770
duke@0 771 //--------------------------generate_current_thread--------------------
duke@0 772 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
duke@0 773 ciKlass* thread_klass = env()->Thread_klass();
duke@0 774 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
duke@0 775 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
duke@0 776 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
duke@0 777 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
duke@0 778 tls_output = thread;
duke@0 779 return threadObj;
duke@0 780 }
duke@0 781
duke@0 782
duke@0 783 //------------------------------inline_string_compareTo------------------------
duke@0 784 bool LibraryCallKit::inline_string_compareTo() {
duke@0 785
duke@0 786 const int value_offset = java_lang_String::value_offset_in_bytes();
duke@0 787 const int count_offset = java_lang_String::count_offset_in_bytes();
duke@0 788 const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@0 789
duke@0 790 _sp += 2;
duke@0 791 Node *argument = pop(); // pop non-receiver first: it was pushed second
duke@0 792 Node *receiver = pop();
duke@0 793
duke@0 794 // Null check on self without removing any arguments. The argument
duke@0 795 // null check technically happens in the wrong place, which can lead to
duke@0 796 // invalid stack traces when string compare is inlined into a method
duke@0 797 // which handles NullPointerExceptions.
duke@0 798 _sp += 2;
duke@0 799 receiver = do_null_check(receiver, T_OBJECT);
duke@0 800 argument = do_null_check(argument, T_OBJECT);
duke@0 801 _sp -= 2;
duke@0 802 if (stopped()) {
duke@0 803 return true;
duke@0 804 }
duke@0 805
duke@0 806 ciInstanceKlass* klass = env()->String_klass();
duke@0 807 const TypeInstPtr* string_type =
duke@0 808 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
duke@0 809
duke@0 810 Node* compare =
duke@0 811 _gvn.transform(new (C, 7) StrCompNode(
duke@0 812 control(),
duke@0 813 memory(TypeAryPtr::CHARS),
duke@0 814 memory(string_type->add_offset(value_offset)),
duke@0 815 memory(string_type->add_offset(count_offset)),
duke@0 816 memory(string_type->add_offset(offset_offset)),
duke@0 817 receiver,
duke@0 818 argument));
duke@0 819 push(compare);
duke@0 820 return true;
duke@0 821 }
duke@0 822
rasbold@169 823 //------------------------------inline_array_equals----------------------------
rasbold@169 824 bool LibraryCallKit::inline_array_equals() {
rasbold@169 825
rasbold@174 826 if (!Matcher::has_match_rule(Op_AryEq)) return false;
rasbold@174 827
rasbold@169 828 _sp += 2;
rasbold@169 829 Node *argument2 = pop();
rasbold@169 830 Node *argument1 = pop();
rasbold@169 831
rasbold@169 832 Node* equals =
rasbold@169 833 _gvn.transform(new (C, 3) AryEqNode(control(),
rasbold@169 834 argument1,
rasbold@169 835 argument2)
rasbold@169 836 );
rasbold@169 837 push(equals);
rasbold@169 838 return true;
rasbold@169 839 }
rasbold@169 840
duke@0 841 // Java version of String.indexOf(constant string)
duke@0 842 // class StringDecl {
duke@0 843 // StringDecl(char[] ca) {
duke@0 844 // offset = 0;
duke@0 845 // count = ca.length;
duke@0 846 // value = ca;
duke@0 847 // }
duke@0 848 // int offset;
duke@0 849 // int count;
duke@0 850 // char[] value;
duke@0 851 // }
duke@0 852 //
duke@0 853 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
duke@0 854 // int targetOffset, int cache_i, int md2) {
duke@0 855 // int cache = cache_i;
duke@0 856 // int sourceOffset = string_object.offset;
duke@0 857 // int sourceCount = string_object.count;
duke@0 858 // int targetCount = target_object.length;
duke@0 859 //
duke@0 860 // int targetCountLess1 = targetCount - 1;
duke@0 861 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
duke@0 862 //
duke@0 863 // char[] source = string_object.value;
duke@0 864 // char[] target = target_object;
duke@0 865 // int lastChar = target[targetCountLess1];
duke@0 866 //
duke@0 867 // outer_loop:
duke@0 868 // for (int i = sourceOffset; i < sourceEnd; ) {
duke@0 869 // int src = source[i + targetCountLess1];
duke@0 870 // if (src == lastChar) {
duke@0 871 // // With random strings and a 4-character alphabet,
duke@0 872 // // reverse matching at this point sets up 0.8% fewer
duke@0 873 // // frames, but (paradoxically) makes 0.3% more probes.
duke@0 874 // // Since those probes are nearer the lastChar probe,
duke@0 875 // // there is may be a net D$ win with reverse matching.
duke@0 876 // // But, reversing loop inhibits unroll of inner loop
duke@0 877 // // for unknown reason. So, does running outer loop from
duke@0 878 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
duke@0 879 // for (int j = 0; j < targetCountLess1; j++) {
duke@0 880 // if (target[targetOffset + j] != source[i+j]) {
duke@0 881 // if ((cache & (1 << source[i+j])) == 0) {
duke@0 882 // if (md2 < j+1) {
duke@0 883 // i += j+1;
duke@0 884 // continue outer_loop;
duke@0 885 // }
duke@0 886 // }
duke@0 887 // i += md2;
duke@0 888 // continue outer_loop;
duke@0 889 // }
duke@0 890 // }
duke@0 891 // return i - sourceOffset;
duke@0 892 // }
duke@0 893 // if ((cache & (1 << src)) == 0) {
duke@0 894 // i += targetCountLess1;
duke@0 895 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
duke@0 896 // i++;
duke@0 897 // }
duke@0 898 // return -1;
duke@0 899 // }
duke@0 900
duke@0 901 //------------------------------string_indexOf------------------------
duke@0 902 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
duke@0 903 jint cache_i, jint md2_i) {
duke@0 904
duke@0 905 Node* no_ctrl = NULL;
duke@0 906 float likely = PROB_LIKELY(0.9);
duke@0 907 float unlikely = PROB_UNLIKELY(0.9);
duke@0 908
duke@0 909 const int value_offset = java_lang_String::value_offset_in_bytes();
duke@0 910 const int count_offset = java_lang_String::count_offset_in_bytes();
duke@0 911 const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@0 912
duke@0 913 ciInstanceKlass* klass = env()->String_klass();
duke@0 914 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
duke@0 915 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
duke@0 916
duke@0 917 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
duke@0 918 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
duke@0 919 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
duke@0 920 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
duke@0 921 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
duke@0 922 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
duke@0 923
kvn@164 924 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
duke@0 925 jint target_length = target_array->length();
duke@0 926 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
duke@0 927 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
duke@0 928
duke@0 929 IdealKit kit(gvn(), control(), merged_memory());
duke@0 930 #define __ kit.
duke@0 931 Node* zero = __ ConI(0);
duke@0 932 Node* one = __ ConI(1);
duke@0 933 Node* cache = __ ConI(cache_i);
duke@0 934 Node* md2 = __ ConI(md2_i);
duke@0 935 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
duke@0 936 Node* targetCount = __ ConI(target_length);
duke@0 937 Node* targetCountLess1 = __ ConI(target_length - 1);
duke@0 938 Node* targetOffset = __ ConI(targetOffset_i);
duke@0 939 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
duke@0 940
duke@0 941 IdealVariable rtn(kit), i(kit), j(kit); __ declares_done();
duke@0 942 Node* outer_loop = __ make_label(2 /* goto */);
duke@0 943 Node* return_ = __ make_label(1);
duke@0 944
duke@0 945 __ set(rtn,__ ConI(-1));
duke@0 946 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
duke@0 947 Node* i2 = __ AddI(__ value(i), targetCountLess1);
duke@0 948 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
duke@0 949 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
duke@0 950 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
duke@0 951 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
duke@0 952 Node* tpj = __ AddI(targetOffset, __ value(j));
duke@0 953 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
duke@0 954 Node* ipj = __ AddI(__ value(i), __ value(j));
duke@0 955 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
duke@0 956 __ if_then(targ, BoolTest::ne, src2); {
duke@0 957 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
duke@0 958 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
duke@0 959 __ increment(i, __ AddI(__ value(j), one));
duke@0 960 __ goto_(outer_loop);
duke@0 961 } __ end_if(); __ dead(j);
duke@0 962 }__ end_if(); __ dead(j);
duke@0 963 __ increment(i, md2);
duke@0 964 __ goto_(outer_loop);
duke@0 965 }__ end_if();
duke@0 966 __ increment(j, one);
duke@0 967 }__ end_loop(); __ dead(j);
duke@0 968 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
duke@0 969 __ goto_(return_);
duke@0 970 }__ end_if();
duke@0 971 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
duke@0 972 __ increment(i, targetCountLess1);
duke@0 973 }__ end_if();
duke@0 974 __ increment(i, one);
duke@0 975 __ bind(outer_loop);
duke@0 976 }__ end_loop(); __ dead(i);
duke@0 977 __ bind(return_);
duke@0 978 __ drain_delay_transform();
duke@0 979
duke@0 980 set_control(__ ctrl());
duke@0 981 Node* result = __ value(rtn);
duke@0 982 #undef __
duke@0 983 C->set_has_loops(true);
duke@0 984 return result;
duke@0 985 }
duke@0 986
duke@0 987
duke@0 988 //------------------------------inline_string_indexOf------------------------
duke@0 989 bool LibraryCallKit::inline_string_indexOf() {
duke@0 990
duke@0 991 _sp += 2;
duke@0 992 Node *argument = pop(); // pop non-receiver first: it was pushed second
duke@0 993 Node *receiver = pop();
duke@0 994
duke@0 995 // don't intrinsify is argument isn't a constant string.
duke@0 996 if (!argument->is_Con()) {
duke@0 997 return false;
duke@0 998 }
duke@0 999 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
duke@0 1000 if (str_type == NULL) {
duke@0 1001 return false;
duke@0 1002 }
duke@0 1003 ciInstanceKlass* klass = env()->String_klass();
duke@0 1004 ciObject* str_const = str_type->const_oop();
duke@0 1005 if (str_const == NULL || str_const->klass() != klass) {
duke@0 1006 return false;
duke@0 1007 }
duke@0 1008 ciInstance* str = str_const->as_instance();
duke@0 1009 assert(str != NULL, "must be instance");
duke@0 1010
duke@0 1011 const int value_offset = java_lang_String::value_offset_in_bytes();
duke@0 1012 const int count_offset = java_lang_String::count_offset_in_bytes();
duke@0 1013 const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@0 1014
duke@0 1015 ciObject* v = str->field_value_by_offset(value_offset).as_object();
duke@0 1016 int o = str->field_value_by_offset(offset_offset).as_int();
duke@0 1017 int c = str->field_value_by_offset(count_offset).as_int();
duke@0 1018 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
duke@0 1019
duke@0 1020 // constant strings have no offset and count == length which
duke@0 1021 // simplifies the resulting code somewhat so lets optimize for that.
duke@0 1022 if (o != 0 || c != pat->length()) {
duke@0 1023 return false;
duke@0 1024 }
duke@0 1025
duke@0 1026 // Null check on self without removing any arguments. The argument
duke@0 1027 // null check technically happens in the wrong place, which can lead to
duke@0 1028 // invalid stack traces when string compare is inlined into a method
duke@0 1029 // which handles NullPointerExceptions.
duke@0 1030 _sp += 2;
duke@0 1031 receiver = do_null_check(receiver, T_OBJECT);
duke@0 1032 // No null check on the argument is needed since it's a constant String oop.
duke@0 1033 _sp -= 2;
duke@0 1034 if (stopped()) {
duke@0 1035 return true;
duke@0 1036 }
duke@0 1037
duke@0 1038 // The null string as a pattern always returns 0 (match at beginning of string)
duke@0 1039 if (c == 0) {
duke@0 1040 push(intcon(0));
duke@0 1041 return true;
duke@0 1042 }
duke@0 1043
duke@0 1044 jchar lastChar = pat->char_at(o + (c - 1));
duke@0 1045 int cache = 0;
duke@0 1046 int i;
duke@0 1047 for (i = 0; i < c - 1; i++) {
duke@0 1048 assert(i < pat->length(), "out of range");
duke@0 1049 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
duke@0 1050 }
duke@0 1051
duke@0 1052 int md2 = c;
duke@0 1053 for (i = 0; i < c - 1; i++) {
duke@0 1054 assert(i < pat->length(), "out of range");
duke@0 1055 if (pat->char_at(o + i) == lastChar) {
duke@0 1056 md2 = (c - 1) - i;
duke@0 1057 }
duke@0 1058 }
duke@0 1059
duke@0 1060 Node* result = string_indexOf(receiver, pat, o, cache, md2);
duke@0 1061 push(result);
duke@0 1062 return true;
duke@0 1063 }
duke@0 1064
duke@0 1065 //--------------------------pop_math_arg--------------------------------
duke@0 1066 // Pop a double argument to a math function from the stack
duke@0 1067 // rounding it if necessary.
duke@0 1068 Node * LibraryCallKit::pop_math_arg() {
duke@0 1069 Node *arg = pop_pair();
duke@0 1070 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
duke@0 1071 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
duke@0 1072 return arg;
duke@0 1073 }
duke@0 1074
duke@0 1075 //------------------------------inline_trig----------------------------------
duke@0 1076 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
duke@0 1077 // argument reduction which will turn into a fast/slow diamond.
duke@0 1078 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
duke@0 1079 _sp += arg_size(); // restore stack pointer
duke@0 1080 Node* arg = pop_math_arg();
duke@0 1081 Node* trig = NULL;
duke@0 1082
duke@0 1083 switch (id) {
duke@0 1084 case vmIntrinsics::_dsin:
duke@0 1085 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
duke@0 1086 break;
duke@0 1087 case vmIntrinsics::_dcos:
duke@0 1088 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
duke@0 1089 break;
duke@0 1090 case vmIntrinsics::_dtan:
duke@0 1091 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
duke@0 1092 break;
duke@0 1093 default:
duke@0 1094 assert(false, "bad intrinsic was passed in");
duke@0 1095 return false;
duke@0 1096 }
duke@0 1097
duke@0 1098 // Rounding required? Check for argument reduction!
duke@0 1099 if( Matcher::strict_fp_requires_explicit_rounding ) {
duke@0 1100
duke@0 1101 static const double pi_4 = 0.7853981633974483;
duke@0 1102 static const double neg_pi_4 = -0.7853981633974483;
duke@0 1103 // pi/2 in 80-bit extended precision
duke@0 1104 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
duke@0 1105 // -pi/2 in 80-bit extended precision
duke@0 1106 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
duke@0 1107 // Cutoff value for using this argument reduction technique
duke@0 1108 //static const double pi_2_minus_epsilon = 1.564660403643354;
duke@0 1109 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
duke@0 1110
duke@0 1111 // Pseudocode for sin:
duke@0 1112 // if (x <= Math.PI / 4.0) {
duke@0 1113 // if (x >= -Math.PI / 4.0) return fsin(x);
duke@0 1114 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
duke@0 1115 // } else {
duke@0 1116 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
duke@0 1117 // }
duke@0 1118 // return StrictMath.sin(x);
duke@0 1119
duke@0 1120 // Pseudocode for cos:
duke@0 1121 // if (x <= Math.PI / 4.0) {
duke@0 1122 // if (x >= -Math.PI / 4.0) return fcos(x);
duke@0 1123 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
duke@0 1124 // } else {
duke@0 1125 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
duke@0 1126 // }
duke@0 1127 // return StrictMath.cos(x);
duke@0 1128
duke@0 1129 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
duke@0 1130 // requires a special machine instruction to load it. Instead we'll try
duke@0 1131 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
duke@0 1132 // probably do the math inside the SIN encoding.
duke@0 1133
duke@0 1134 // Make the merge point
duke@0 1135 RegionNode *r = new (C, 3) RegionNode(3);
duke@0 1136 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
duke@0 1137
duke@0 1138 // Flatten arg so we need only 1 test
duke@0 1139 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
duke@0 1140 // Node for PI/4 constant
duke@0 1141 Node *pi4 = makecon(TypeD::make(pi_4));
duke@0 1142 // Check PI/4 : abs(arg)
duke@0 1143 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
duke@0 1144 // Check: If PI/4 < abs(arg) then go slow
duke@0 1145 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
duke@0 1146 // Branch either way
duke@0 1147 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@0 1148 set_control(opt_iff(r,iff));
duke@0 1149
duke@0 1150 // Set fast path result
duke@0 1151 phi->init_req(2,trig);
duke@0 1152
duke@0 1153 // Slow path - non-blocking leaf call
duke@0 1154 Node* call = NULL;
duke@0 1155 switch (id) {
duke@0 1156 case vmIntrinsics::_dsin:
duke@0 1157 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@0 1158 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
duke@0 1159 "Sin", NULL, arg, top());
duke@0 1160 break;
duke@0 1161 case vmIntrinsics::_dcos:
duke@0 1162 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@0 1163 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
duke@0 1164 "Cos", NULL, arg, top());
duke@0 1165 break;
duke@0 1166 case vmIntrinsics::_dtan:
duke@0 1167 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@0 1168 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
duke@0 1169 "Tan", NULL, arg, top());
duke@0 1170 break;
duke@0 1171 }
duke@0 1172 assert(control()->in(0) == call, "");
duke@0 1173 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
duke@0 1174 r->init_req(1,control());
duke@0 1175 phi->init_req(1,slow_result);
duke@0 1176
duke@0 1177 // Post-merge
duke@0 1178 set_control(_gvn.transform(r));
duke@0 1179 record_for_igvn(r);
duke@0 1180 trig = _gvn.transform(phi);
duke@0 1181
duke@0 1182 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 1183 }
duke@0 1184 // Push result back on JVM stack
duke@0 1185 push_pair(trig);
duke@0 1186 return true;
duke@0 1187 }
duke@0 1188
duke@0 1189 //------------------------------inline_sqrt-------------------------------------
duke@0 1190 // Inline square root instruction, if possible.
duke@0 1191 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
duke@0 1192 assert(id == vmIntrinsics::_dsqrt, "Not square root");
duke@0 1193 _sp += arg_size(); // restore stack pointer
duke@0 1194 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
duke@0 1195 return true;
duke@0 1196 }
duke@0 1197
duke@0 1198 //------------------------------inline_abs-------------------------------------
duke@0 1199 // Inline absolute value instruction, if possible.
duke@0 1200 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
duke@0 1201 assert(id == vmIntrinsics::_dabs, "Not absolute value");
duke@0 1202 _sp += arg_size(); // restore stack pointer
duke@0 1203 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
duke@0 1204 return true;
duke@0 1205 }
duke@0 1206
duke@0 1207 //------------------------------inline_exp-------------------------------------
duke@0 1208 // Inline exp instructions, if possible. The Intel hardware only misses
duke@0 1209 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
duke@0 1210 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
duke@0 1211 assert(id == vmIntrinsics::_dexp, "Not exp");
duke@0 1212
duke@0 1213 // If this inlining ever returned NaN in the past, we do not intrinsify it
duke@0 1214 // every again. NaN results requires StrictMath.exp handling.
duke@0 1215 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@0 1216
duke@0 1217 // Do not intrinsify on older platforms which lack cmove.
duke@0 1218 if (ConditionalMoveLimit == 0) return false;
duke@0 1219
duke@0 1220 _sp += arg_size(); // restore stack pointer
duke@0 1221 Node *x = pop_math_arg();
duke@0 1222 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
duke@0 1223
duke@0 1224 //-------------------
duke@0 1225 //result=(result.isNaN())? StrictMath::exp():result;
duke@0 1226 // Check: If isNaN() by checking result!=result? then go to Strict Math
duke@0 1227 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
duke@0 1228 // Build the boolean node
duke@0 1229 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
duke@0 1230
duke@0 1231 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
duke@0 1232 // End the current control-flow path
duke@0 1233 push_pair(x);
duke@0 1234 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
duke@0 1235 // to handle. Recompile without intrinsifying Math.exp
duke@0 1236 uncommon_trap(Deoptimization::Reason_intrinsic,
duke@0 1237 Deoptimization::Action_make_not_entrant);
duke@0 1238 }
duke@0 1239
duke@0 1240 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 1241
duke@0 1242 push_pair(result);
duke@0 1243
duke@0 1244 return true;
duke@0 1245 }
duke@0 1246
duke@0 1247 //------------------------------inline_pow-------------------------------------
duke@0 1248 // Inline power instructions, if possible.
duke@0 1249 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
duke@0 1250 assert(id == vmIntrinsics::_dpow, "Not pow");
duke@0 1251
duke@0 1252 // If this inlining ever returned NaN in the past, we do not intrinsify it
duke@0 1253 // every again. NaN results requires StrictMath.pow handling.
duke@0 1254 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@0 1255
duke@0 1256 // Do not intrinsify on older platforms which lack cmove.
duke@0 1257 if (ConditionalMoveLimit == 0) return false;
duke@0 1258
duke@0 1259 // Pseudocode for pow
duke@0 1260 // if (x <= 0.0) {
duke@0 1261 // if ((double)((int)y)==y) { // if y is int
duke@0 1262 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
duke@0 1263 // } else {
duke@0 1264 // result = NaN;
duke@0 1265 // }
duke@0 1266 // } else {
duke@0 1267 // result = DPow(x,y);
duke@0 1268 // }
duke@0 1269 // if (result != result)? {
duke@0 1270 // ucommon_trap();
duke@0 1271 // }
duke@0 1272 // return result;
duke@0 1273
duke@0 1274 _sp += arg_size(); // restore stack pointer
duke@0 1275 Node* y = pop_math_arg();
duke@0 1276 Node* x = pop_math_arg();
duke@0 1277
duke@0 1278 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
duke@0 1279
duke@0 1280 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
duke@0 1281 // inside of something) then skip the fancy tests and just check for
duke@0 1282 // NaN result.
duke@0 1283 Node *result = NULL;
duke@0 1284 if( jvms()->depth() >= 1 ) {
duke@0 1285 result = fast_result;
duke@0 1286 } else {
duke@0 1287
duke@0 1288 // Set the merge point for If node with condition of (x <= 0.0)
duke@0 1289 // There are four possible paths to region node and phi node
duke@0 1290 RegionNode *r = new (C, 4) RegionNode(4);
duke@0 1291 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
duke@0 1292
duke@0 1293 // Build the first if node: if (x <= 0.0)
duke@0 1294 // Node for 0 constant
duke@0 1295 Node *zeronode = makecon(TypeD::ZERO);
duke@0 1296 // Check x:0
duke@0 1297 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
duke@0 1298 // Check: If (x<=0) then go complex path
duke@0 1299 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
duke@0 1300 // Branch either way
duke@0 1301 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
duke@0 1302 Node *opt_test = _gvn.transform(if1);
duke@0 1303 //assert( opt_test->is_If(), "Expect an IfNode");
duke@0 1304 IfNode *opt_if1 = (IfNode*)opt_test;
duke@0 1305 // Fast path taken; set region slot 3
duke@0 1306 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
duke@0 1307 r->init_req(3,fast_taken); // Capture fast-control
duke@0 1308
duke@0 1309 // Fast path not-taken, i.e. slow path
duke@0 1310 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
duke@0 1311
duke@0 1312 // Set fast path result
duke@0 1313 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
duke@0 1314 phi->init_req(3, fast_result);
duke@0 1315
duke@0 1316 // Complex path
duke@0 1317 // Build the second if node (if y is int)
duke@0 1318 // Node for (int)y
duke@0 1319 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
duke@0 1320 // Node for (double)((int) y)
duke@0 1321 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
duke@0 1322 // Check (double)((int) y) : y
duke@0 1323 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
duke@0 1324 // Check if (y isn't int) then go to slow path
duke@0 1325
duke@0 1326 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
duke@0 1327 // Branch eith way
duke@0 1328 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
duke@0 1329 Node *slow_path = opt_iff(r,if2); // Set region path 2
duke@0 1330
duke@0 1331 // Calculate DPow(abs(x), y)*(1 & (int)y)
duke@0 1332 // Node for constant 1
duke@0 1333 Node *conone = intcon(1);
duke@0 1334 // 1& (int)y
duke@0 1335 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
duke@0 1336 // zero node
duke@0 1337 Node *conzero = intcon(0);
duke@0 1338 // Check (1&(int)y)==0?
duke@0 1339 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
duke@0 1340 // Check if (1&(int)y)!=0?, if so the result is negative
duke@0 1341 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
duke@0 1342 // abs(x)
duke@0 1343 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
duke@0 1344 // abs(x)^y
duke@0 1345 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
duke@0 1346 // -abs(x)^y
duke@0 1347 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
duke@0 1348 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
duke@0 1349 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
duke@0 1350 // Set complex path fast result
duke@0 1351 phi->init_req(2, signresult);
duke@0 1352
duke@0 1353 static const jlong nan_bits = CONST64(0x7ff8000000000000);
duke@0 1354 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
duke@0 1355 r->init_req(1,slow_path);
duke@0 1356 phi->init_req(1,slow_result);
duke@0 1357
duke@0 1358 // Post merge
duke@0 1359 set_control(_gvn.transform(r));
duke@0 1360 record_for_igvn(r);
duke@0 1361 result=_gvn.transform(phi);
duke@0 1362 }
duke@0 1363
duke@0 1364 //-------------------
duke@0 1365 //result=(result.isNaN())? uncommon_trap():result;
duke@0 1366 // Check: If isNaN() by checking result!=result? then go to Strict Math
duke@0 1367 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
duke@0 1368 // Build the boolean node
duke@0 1369 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
duke@0 1370
duke@0 1371 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
duke@0 1372 // End the current control-flow path
duke@0 1373 push_pair(x);
duke@0 1374 push_pair(y);
duke@0 1375 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
duke@0 1376 // to handle. Recompile without intrinsifying Math.pow.
duke@0 1377 uncommon_trap(Deoptimization::Reason_intrinsic,
duke@0 1378 Deoptimization::Action_make_not_entrant);
duke@0 1379 }
duke@0 1380
duke@0 1381 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 1382
duke@0 1383 push_pair(result);
duke@0 1384
duke@0 1385 return true;
duke@0 1386 }
duke@0 1387
duke@0 1388 //------------------------------inline_trans-------------------------------------
duke@0 1389 // Inline transcendental instructions, if possible. The Intel hardware gets
duke@0 1390 // these right, no funny corner cases missed.
duke@0 1391 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
duke@0 1392 _sp += arg_size(); // restore stack pointer
duke@0 1393 Node* arg = pop_math_arg();
duke@0 1394 Node* trans = NULL;
duke@0 1395
duke@0 1396 switch (id) {
duke@0 1397 case vmIntrinsics::_dlog:
duke@0 1398 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
duke@0 1399 break;
duke@0 1400 case vmIntrinsics::_dlog10:
duke@0 1401 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
duke@0 1402 break;
duke@0 1403 default:
duke@0 1404 assert(false, "bad intrinsic was passed in");
duke@0 1405 return false;
duke@0 1406 }
duke@0 1407
duke@0 1408 // Push result back on JVM stack
duke@0 1409 push_pair(trans);
duke@0 1410 return true;
duke@0 1411 }
duke@0 1412
duke@0 1413 //------------------------------runtime_math-----------------------------
duke@0 1414 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
duke@0 1415 Node* a = NULL;
duke@0 1416 Node* b = NULL;
duke@0 1417
duke@0 1418 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
duke@0 1419 "must be (DD)D or (D)D type");
duke@0 1420
duke@0 1421 // Inputs
duke@0 1422 _sp += arg_size(); // restore stack pointer
duke@0 1423 if (call_type == OptoRuntime::Math_DD_D_Type()) {
duke@0 1424 b = pop_math_arg();
duke@0 1425 }
duke@0 1426 a = pop_math_arg();
duke@0 1427
duke@0 1428 const TypePtr* no_memory_effects = NULL;
duke@0 1429 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
duke@0 1430 no_memory_effects,
duke@0 1431 a, top(), b, b ? top() : NULL);
duke@0 1432 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
duke@0 1433 #ifdef ASSERT
duke@0 1434 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
duke@0 1435 assert(value_top == top(), "second value must be top");
duke@0 1436 #endif
duke@0 1437
duke@0 1438 push_pair(value);
duke@0 1439 return true;
duke@0 1440 }
duke@0 1441
duke@0 1442 //------------------------------inline_math_native-----------------------------
duke@0 1443 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
duke@0 1444 switch (id) {
duke@0 1445 // These intrinsics are not properly supported on all hardware
duke@0 1446 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
duke@0 1447 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
duke@0 1448 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
duke@0 1449 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
duke@0 1450 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
duke@0 1451 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
duke@0 1452
duke@0 1453 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
duke@0 1454 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
duke@0 1455 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
duke@0 1456 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
duke@0 1457
duke@0 1458 // These intrinsics are supported on all hardware
duke@0 1459 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
duke@0 1460 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
duke@0 1461
duke@0 1462 // These intrinsics don't work on X86. The ad implementation doesn't
duke@0 1463 // handle NaN's properly. Instead of returning infinity, the ad
duke@0 1464 // implementation returns a NaN on overflow. See bug: 6304089
duke@0 1465 // Once the ad implementations are fixed, change the code below
duke@0 1466 // to match the intrinsics above
duke@0 1467
duke@0 1468 case vmIntrinsics::_dexp: return
duke@0 1469 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
duke@0 1470 case vmIntrinsics::_dpow: return
duke@0 1471 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
duke@0 1472
duke@0 1473 // These intrinsics are not yet correctly implemented
duke@0 1474 case vmIntrinsics::_datan2:
duke@0 1475 return false;
duke@0 1476
duke@0 1477 default:
duke@0 1478 ShouldNotReachHere();
duke@0 1479 return false;
duke@0 1480 }
duke@0 1481 }
duke@0 1482
duke@0 1483 static bool is_simple_name(Node* n) {
duke@0 1484 return (n->req() == 1 // constant
duke@0 1485 || (n->is_Type() && n->as_Type()->type()->singleton())
duke@0 1486 || n->is_Proj() // parameter or return value
duke@0 1487 || n->is_Phi() // local of some sort
duke@0 1488 );
duke@0 1489 }
duke@0 1490
duke@0 1491 //----------------------------inline_min_max-----------------------------------
duke@0 1492 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
duke@0 1493 push(generate_min_max(id, argument(0), argument(1)));
duke@0 1494
duke@0 1495 return true;
duke@0 1496 }
duke@0 1497
duke@0 1498 Node*
duke@0 1499 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
duke@0 1500 // These are the candidate return value:
duke@0 1501 Node* xvalue = x0;
duke@0 1502 Node* yvalue = y0;
duke@0 1503
duke@0 1504 if (xvalue == yvalue) {
duke@0 1505 return xvalue;
duke@0 1506 }
duke@0 1507
duke@0 1508 bool want_max = (id == vmIntrinsics::_max);
duke@0 1509
duke@0 1510 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
duke@0 1511 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
duke@0 1512 if (txvalue == NULL || tyvalue == NULL) return top();
duke@0 1513 // This is not really necessary, but it is consistent with a
duke@0 1514 // hypothetical MaxINode::Value method:
duke@0 1515 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
duke@0 1516
duke@0 1517 // %%% This folding logic should (ideally) be in a different place.
duke@0 1518 // Some should be inside IfNode, and there to be a more reliable
duke@0 1519 // transformation of ?: style patterns into cmoves. We also want
duke@0 1520 // more powerful optimizations around cmove and min/max.
duke@0 1521
duke@0 1522 // Try to find a dominating comparison of these guys.
duke@0 1523 // It can simplify the index computation for Arrays.copyOf
duke@0 1524 // and similar uses of System.arraycopy.
duke@0 1525 // First, compute the normalized version of CmpI(x, y).
duke@0 1526 int cmp_op = Op_CmpI;
duke@0 1527 Node* xkey = xvalue;
duke@0 1528 Node* ykey = yvalue;
duke@0 1529 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
duke@0 1530 if (ideal_cmpxy->is_Cmp()) {
duke@0 1531 // E.g., if we have CmpI(length - offset, count),
duke@0 1532 // it might idealize to CmpI(length, count + offset)
duke@0 1533 cmp_op = ideal_cmpxy->Opcode();
duke@0 1534 xkey = ideal_cmpxy->in(1);
duke@0 1535 ykey = ideal_cmpxy->in(2);
duke@0 1536 }
duke@0 1537
duke@0 1538 // Start by locating any relevant comparisons.
duke@0 1539 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
duke@0 1540 Node* cmpxy = NULL;
duke@0 1541 Node* cmpyx = NULL;
duke@0 1542 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
duke@0 1543 Node* cmp = start_from->fast_out(k);
duke@0 1544 if (cmp->outcnt() > 0 && // must have prior uses
duke@0 1545 cmp->in(0) == NULL && // must be context-independent
duke@0 1546 cmp->Opcode() == cmp_op) { // right kind of compare
duke@0 1547 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
duke@0 1548 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
duke@0 1549 }
duke@0 1550 }
duke@0 1551
duke@0 1552 const int NCMPS = 2;
duke@0 1553 Node* cmps[NCMPS] = { cmpxy, cmpyx };
duke@0 1554 int cmpn;
duke@0 1555 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
duke@0 1556 if (cmps[cmpn] != NULL) break; // find a result
duke@0 1557 }
duke@0 1558 if (cmpn < NCMPS) {
duke@0 1559 // Look for a dominating test that tells us the min and max.
duke@0 1560 int depth = 0; // Limit search depth for speed
duke@0 1561 Node* dom = control();
duke@0 1562 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
duke@0 1563 if (++depth >= 100) break;
duke@0 1564 Node* ifproj = dom;
duke@0 1565 if (!ifproj->is_Proj()) continue;
duke@0 1566 Node* iff = ifproj->in(0);
duke@0 1567 if (!iff->is_If()) continue;
duke@0 1568 Node* bol = iff->in(1);
duke@0 1569 if (!bol->is_Bool()) continue;
duke@0 1570 Node* cmp = bol->in(1);
duke@0 1571 if (cmp == NULL) continue;
duke@0 1572 for (cmpn = 0; cmpn < NCMPS; cmpn++)
duke@0 1573 if (cmps[cmpn] == cmp) break;
duke@0 1574 if (cmpn == NCMPS) continue;
duke@0 1575 BoolTest::mask btest = bol->as_Bool()->_test._test;
duke@0 1576 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
duke@0 1577 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
duke@0 1578 // At this point, we know that 'x btest y' is true.
duke@0 1579 switch (btest) {
duke@0 1580 case BoolTest::eq:
duke@0 1581 // They are proven equal, so we can collapse the min/max.
duke@0 1582 // Either value is the answer. Choose the simpler.
duke@0 1583 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
duke@0 1584 return yvalue;
duke@0 1585 return xvalue;
duke@0 1586 case BoolTest::lt: // x < y
duke@0 1587 case BoolTest::le: // x <= y
duke@0 1588 return (want_max ? yvalue : xvalue);
duke@0 1589 case BoolTest::gt: // x > y
duke@0 1590 case BoolTest::ge: // x >= y
duke@0 1591 return (want_max ? xvalue : yvalue);
duke@0 1592 }
duke@0 1593 }
duke@0 1594 }
duke@0 1595
duke@0 1596 // We failed to find a dominating test.
duke@0 1597 // Let's pick a test that might GVN with prior tests.
duke@0 1598 Node* best_bol = NULL;
duke@0 1599 BoolTest::mask best_btest = BoolTest::illegal;
duke@0 1600 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
duke@0 1601 Node* cmp = cmps[cmpn];
duke@0 1602 if (cmp == NULL) continue;
duke@0 1603 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
duke@0 1604 Node* bol = cmp->fast_out(j);
duke@0 1605 if (!bol->is_Bool()) continue;
duke@0 1606 BoolTest::mask btest = bol->as_Bool()->_test._test;
duke@0 1607 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
duke@0 1608 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
duke@0 1609 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
duke@0 1610 best_bol = bol->as_Bool();
duke@0 1611 best_btest = btest;
duke@0 1612 }
duke@0 1613 }
duke@0 1614 }
duke@0 1615
duke@0 1616 Node* answer_if_true = NULL;
duke@0 1617 Node* answer_if_false = NULL;
duke@0 1618 switch (best_btest) {
duke@0 1619 default:
duke@0 1620 if (cmpxy == NULL)
duke@0 1621 cmpxy = ideal_cmpxy;
duke@0 1622 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
duke@0 1623 // and fall through:
duke@0 1624 case BoolTest::lt: // x < y
duke@0 1625 case BoolTest::le: // x <= y
duke@0 1626 answer_if_true = (want_max ? yvalue : xvalue);
duke@0 1627 answer_if_false = (want_max ? xvalue : yvalue);
duke@0 1628 break;
duke@0 1629 case BoolTest::gt: // x > y
duke@0 1630 case BoolTest::ge: // x >= y
duke@0 1631 answer_if_true = (want_max ? xvalue : yvalue);
duke@0 1632 answer_if_false = (want_max ? yvalue : xvalue);
duke@0 1633 break;
duke@0 1634 }
duke@0 1635
duke@0 1636 jint hi, lo;
duke@0 1637 if (want_max) {
duke@0 1638 // We can sharpen the minimum.
duke@0 1639 hi = MAX2(txvalue->_hi, tyvalue->_hi);
duke@0 1640 lo = MAX2(txvalue->_lo, tyvalue->_lo);
duke@0 1641 } else {
duke@0 1642 // We can sharpen the maximum.
duke@0 1643 hi = MIN2(txvalue->_hi, tyvalue->_hi);
duke@0 1644 lo = MIN2(txvalue->_lo, tyvalue->_lo);
duke@0 1645 }
duke@0 1646
duke@0 1647 // Use a flow-free graph structure, to avoid creating excess control edges
duke@0 1648 // which could hinder other optimizations.
duke@0 1649 // Since Math.min/max is often used with arraycopy, we want
duke@0 1650 // tightly_coupled_allocation to be able to see beyond min/max expressions.
duke@0 1651 Node* cmov = CMoveNode::make(C, NULL, best_bol,
duke@0 1652 answer_if_false, answer_if_true,
duke@0 1653 TypeInt::make(lo, hi, widen));
duke@0 1654
duke@0 1655 return _gvn.transform(cmov);
duke@0 1656
duke@0 1657 /*
duke@0 1658 // This is not as desirable as it may seem, since Min and Max
duke@0 1659 // nodes do not have a full set of optimizations.
duke@0 1660 // And they would interfere, anyway, with 'if' optimizations
duke@0 1661 // and with CMoveI canonical forms.
duke@0 1662 switch (id) {
duke@0 1663 case vmIntrinsics::_min:
duke@0 1664 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
duke@0 1665 case vmIntrinsics::_max:
duke@0 1666 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
duke@0 1667 default:
duke@0 1668 ShouldNotReachHere();
duke@0 1669 }
duke@0 1670 */
duke@0 1671 }
duke@0 1672
duke@0 1673 inline int
duke@0 1674 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
duke@0 1675 const TypePtr* base_type = TypePtr::NULL_PTR;
duke@0 1676 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
duke@0 1677 if (base_type == NULL) {
duke@0 1678 // Unknown type.
duke@0 1679 return Type::AnyPtr;
duke@0 1680 } else if (base_type == TypePtr::NULL_PTR) {
duke@0 1681 // Since this is a NULL+long form, we have to switch to a rawptr.
duke@0 1682 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
duke@0 1683 offset = MakeConX(0);
duke@0 1684 return Type::RawPtr;
duke@0 1685 } else if (base_type->base() == Type::RawPtr) {
duke@0 1686 return Type::RawPtr;
duke@0 1687 } else if (base_type->isa_oopptr()) {
duke@0 1688 // Base is never null => always a heap address.
duke@0 1689 if (base_type->ptr() == TypePtr::NotNull) {
duke@0 1690 return Type::OopPtr;
duke@0 1691 }
duke@0 1692 // Offset is small => always a heap address.
duke@0 1693 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
duke@0 1694 if (offset_type != NULL &&
duke@0 1695 base_type->offset() == 0 && // (should always be?)
duke@0 1696 offset_type->_lo >= 0 &&
duke@0 1697 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
duke@0 1698 return Type::OopPtr;
duke@0 1699 }
duke@0 1700 // Otherwise, it might either be oop+off or NULL+addr.
duke@0 1701 return Type::AnyPtr;
duke@0 1702 } else {
duke@0 1703 // No information:
duke@0 1704 return Type::AnyPtr;
duke@0 1705 }
duke@0 1706 }
duke@0 1707
duke@0 1708 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
duke@0 1709 int kind = classify_unsafe_addr(base, offset);
duke@0 1710 if (kind == Type::RawPtr) {
duke@0 1711 return basic_plus_adr(top(), base, offset);
duke@0 1712 } else {
duke@0 1713 return basic_plus_adr(base, offset);
duke@0 1714 }
duke@0 1715 }
duke@0 1716
duke@0 1717 //----------------------------inline_reverseBytes_int/long-------------------
duke@0 1718 // inline Int.reverseBytes(int)
duke@0 1719 // inline Long.reverseByes(long)
duke@0 1720 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
duke@0 1721 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
duke@0 1722 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
duke@0 1723 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
duke@0 1724 _sp += arg_size(); // restore stack pointer
duke@0 1725 switch (id) {
duke@0 1726 case vmIntrinsics::_reverseBytes_i:
duke@0 1727 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
duke@0 1728 break;
duke@0 1729 case vmIntrinsics::_reverseBytes_l:
duke@0 1730 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
duke@0 1731 break;
duke@0 1732 default:
duke@0 1733 ;
duke@0 1734 }
duke@0 1735 return true;
duke@0 1736 }
duke@0 1737
duke@0 1738 //----------------------------inline_unsafe_access----------------------------
duke@0 1739
duke@0 1740 const static BasicType T_ADDRESS_HOLDER = T_LONG;
duke@0 1741
duke@0 1742 // Interpret Unsafe.fieldOffset cookies correctly:
duke@0 1743 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
duke@0 1744
duke@0 1745 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
duke@0 1746 if (callee()->is_static()) return false; // caller must have the capability!
duke@0 1747
duke@0 1748 #ifndef PRODUCT
duke@0 1749 {
duke@0 1750 ResourceMark rm;
duke@0 1751 // Check the signatures.
duke@0 1752 ciSignature* sig = signature();
duke@0 1753 #ifdef ASSERT
duke@0 1754 if (!is_store) {
duke@0 1755 // Object getObject(Object base, int/long offset), etc.
duke@0 1756 BasicType rtype = sig->return_type()->basic_type();
duke@0 1757 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
duke@0 1758 rtype = T_ADDRESS; // it is really a C void*
duke@0 1759 assert(rtype == type, "getter must return the expected value");
duke@0 1760 if (!is_native_ptr) {
duke@0 1761 assert(sig->count() == 2, "oop getter has 2 arguments");
duke@0 1762 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
duke@0 1763 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
duke@0 1764 } else {
duke@0 1765 assert(sig->count() == 1, "native getter has 1 argument");
duke@0 1766 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
duke@0 1767 }
duke@0 1768 } else {
duke@0 1769 // void putObject(Object base, int/long offset, Object x), etc.
duke@0 1770 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
duke@0 1771 if (!is_native_ptr) {
duke@0 1772 assert(sig->count() == 3, "oop putter has 3 arguments");
duke@0 1773 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
duke@0 1774 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
duke@0 1775 } else {
duke@0 1776 assert(sig->count() == 2, "native putter has 2 arguments");
duke@0 1777 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
duke@0 1778 }
duke@0 1779 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
duke@0 1780 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
duke@0 1781 vtype = T_ADDRESS; // it is really a C void*
duke@0 1782 assert(vtype == type, "putter must accept the expected value");
duke@0 1783 }
duke@0 1784 #endif // ASSERT
duke@0 1785 }
duke@0 1786 #endif //PRODUCT
duke@0 1787
duke@0 1788 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@0 1789
duke@0 1790 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
duke@0 1791
duke@0 1792 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
duke@0 1793 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
duke@0 1794
duke@0 1795 debug_only(int saved_sp = _sp);
duke@0 1796 _sp += nargs;
duke@0 1797
duke@0 1798 Node* val;
duke@0 1799 debug_only(val = (Node*)(uintptr_t)-1);
duke@0 1800
duke@0 1801
duke@0 1802 if (is_store) {
duke@0 1803 // Get the value being stored. (Pop it first; it was pushed last.)
duke@0 1804 switch (type) {
duke@0 1805 case T_DOUBLE:
duke@0 1806 case T_LONG:
duke@0 1807 case T_ADDRESS:
duke@0 1808 val = pop_pair();
duke@0 1809 break;
duke@0 1810 default:
duke@0 1811 val = pop();
duke@0 1812 }
duke@0 1813 }
duke@0 1814
duke@0 1815 // Build address expression. See the code in inline_unsafe_prefetch.
duke@0 1816 Node *adr;
duke@0 1817 Node *heap_base_oop = top();
duke@0 1818 if (!is_native_ptr) {
duke@0 1819 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
duke@0 1820 Node* offset = pop_pair();
duke@0 1821 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
duke@0 1822 Node* base = pop();
duke@0 1823 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@0 1824 // to be plain byte offsets, which are also the same as those accepted
duke@0 1825 // by oopDesc::field_base.
duke@0 1826 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@0 1827 "fieldOffset must be byte-scaled");
duke@0 1828 // 32-bit machines ignore the high half!
duke@0 1829 offset = ConvL2X(offset);
duke@0 1830 adr = make_unsafe_address(base, offset);
duke@0 1831 heap_base_oop = base;
duke@0 1832 } else {
duke@0 1833 Node* ptr = pop_pair();
duke@0 1834 // Adjust Java long to machine word:
duke@0 1835 ptr = ConvL2X(ptr);
duke@0 1836 adr = make_unsafe_address(NULL, ptr);
duke@0 1837 }
duke@0 1838
duke@0 1839 // Pop receiver last: it was pushed first.
duke@0 1840 Node *receiver = pop();
duke@0 1841
duke@0 1842 assert(saved_sp == _sp, "must have correct argument count");
duke@0 1843
duke@0 1844 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@0 1845
duke@0 1846 // First guess at the value type.
duke@0 1847 const Type *value_type = Type::get_const_basic_type(type);
duke@0 1848
duke@0 1849 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
duke@0 1850 // there was not enough information to nail it down.
duke@0 1851 Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@0 1852 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
duke@0 1853
duke@0 1854 // We will need memory barriers unless we can determine a unique
duke@0 1855 // alias category for this reference. (Note: If for some reason
duke@0 1856 // the barriers get omitted and the unsafe reference begins to "pollute"
duke@0 1857 // the alias analysis of the rest of the graph, either Compile::can_alias
duke@0 1858 // or Compile::must_alias will throw a diagnostic assert.)
duke@0 1859 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
duke@0 1860
duke@0 1861 if (!is_store && type == T_OBJECT) {
duke@0 1862 // Attempt to infer a sharper value type from the offset and base type.
duke@0 1863 ciKlass* sharpened_klass = NULL;
duke@0 1864
duke@0 1865 // See if it is an instance field, with an object type.
duke@0 1866 if (alias_type->field() != NULL) {
duke@0 1867 assert(!is_native_ptr, "native pointer op cannot use a java address");
duke@0 1868 if (alias_type->field()->type()->is_klass()) {
duke@0 1869 sharpened_klass = alias_type->field()->type()->as_klass();
duke@0 1870 }
duke@0 1871 }
duke@0 1872
duke@0 1873 // See if it is a narrow oop array.
duke@0 1874 if (adr_type->isa_aryptr()) {
coleenp@113 1875 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) {
duke@0 1876 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
duke@0 1877 if (elem_type != NULL) {
duke@0 1878 sharpened_klass = elem_type->klass();
duke@0 1879 }
duke@0 1880 }
duke@0 1881 }
duke@0 1882
duke@0 1883 if (sharpened_klass != NULL) {
duke@0 1884 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
duke@0 1885
duke@0 1886 // Sharpen the value type.
duke@0 1887 value_type = tjp;
duke@0 1888
duke@0 1889 #ifndef PRODUCT
duke@0 1890 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
duke@0 1891 tty->print(" from base type: "); adr_type->dump();
duke@0 1892 tty->print(" sharpened value: "); value_type->dump();
duke@0 1893 }
duke@0 1894 #endif
duke@0 1895 }
duke@0 1896 }
duke@0 1897
duke@0 1898 // Null check on self without removing any arguments. The argument
duke@0 1899 // null check technically happens in the wrong place, which can lead to
duke@0 1900 // invalid stack traces when the primitive is inlined into a method
duke@0 1901 // which handles NullPointerExceptions.
duke@0 1902 _sp += nargs;
duke@0 1903 do_null_check(receiver, T_OBJECT);
duke@0 1904 _sp -= nargs;
duke@0 1905 if (stopped()) {
duke@0 1906 return true;
duke@0 1907 }
duke@0 1908 // Heap pointers get a null-check from the interpreter,
duke@0 1909 // as a courtesy. However, this is not guaranteed by Unsafe,
duke@0 1910 // and it is not possible to fully distinguish unintended nulls
duke@0 1911 // from intended ones in this API.
duke@0 1912
duke@0 1913 if (is_volatile) {
duke@0 1914 // We need to emit leading and trailing CPU membars (see below) in
duke@0 1915 // addition to memory membars when is_volatile. This is a little
duke@0 1916 // too strong, but avoids the need to insert per-alias-type
duke@0 1917 // volatile membars (for stores; compare Parse::do_put_xxx), which
duke@0 1918 // we cannot do effctively here because we probably only have a
duke@0 1919 // rough approximation of type.
duke@0 1920 need_mem_bar = true;
duke@0 1921 // For Stores, place a memory ordering barrier now.
duke@0 1922 if (is_store)
duke@0 1923 insert_mem_bar(Op_MemBarRelease);
duke@0 1924 }
duke@0 1925
duke@0 1926 // Memory barrier to prevent normal and 'unsafe' accesses from
duke@0 1927 // bypassing each other. Happens after null checks, so the
duke@0 1928 // exception paths do not take memory state from the memory barrier,
duke@0 1929 // so there's no problems making a strong assert about mixing users
duke@0 1930 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
duke@0 1931 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
duke@0 1932 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
duke@0 1933
duke@0 1934 if (!is_store) {
duke@0 1935 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
duke@0 1936 // load value and push onto stack
duke@0 1937 switch (type) {
duke@0 1938 case T_BOOLEAN:
duke@0 1939 case T_CHAR:
duke@0 1940 case T_BYTE:
duke@0 1941 case T_SHORT:
duke@0 1942 case T_INT:
duke@0 1943 case T_FLOAT:
duke@0 1944 case T_OBJECT:
duke@0 1945 push( p );
duke@0 1946 break;
duke@0 1947 case T_ADDRESS:
duke@0 1948 // Cast to an int type.
duke@0 1949 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
duke@0 1950 p = ConvX2L(p);
duke@0 1951 push_pair(p);
duke@0 1952 break;
duke@0 1953 case T_DOUBLE:
duke@0 1954 case T_LONG:
duke@0 1955 push_pair( p );
duke@0 1956 break;
duke@0 1957 default: ShouldNotReachHere();
duke@0 1958 }
duke@0 1959 } else {
duke@0 1960 // place effect of store into memory
duke@0 1961 switch (type) {
duke@0 1962 case T_DOUBLE:
duke@0 1963 val = dstore_rounding(val);
duke@0 1964 break;
duke@0 1965 case T_ADDRESS:
duke@0 1966 // Repackage the long as a pointer.
duke@0 1967 val = ConvL2X(val);
duke@0 1968 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
duke@0 1969 break;
duke@0 1970 }
duke@0 1971
duke@0 1972 if (type != T_OBJECT ) {
duke@0 1973 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
duke@0 1974 } else {
duke@0 1975 // Possibly an oop being stored to Java heap or native memory
duke@0 1976 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
duke@0 1977 // oop to Java heap.
duke@0 1978 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
duke@0 1979 } else {
duke@0 1980
duke@0 1981 // We can't tell at compile time if we are storing in the Java heap or outside
duke@0 1982 // of it. So we need to emit code to conditionally do the proper type of
duke@0 1983 // store.
duke@0 1984
duke@0 1985 IdealKit kit(gvn(), control(), merged_memory());
duke@0 1986 kit.declares_done();
duke@0 1987 // QQQ who knows what probability is here??
duke@0 1988 kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
duke@0 1989 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
duke@0 1990 } kit.else_(); {
duke@0 1991 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
duke@0 1992 } kit.end_if();
duke@0 1993 }
duke@0 1994 }
duke@0 1995 }
duke@0 1996
duke@0 1997 if (is_volatile) {
duke@0 1998 if (!is_store)
duke@0 1999 insert_mem_bar(Op_MemBarAcquire);
duke@0 2000 else
duke@0 2001 insert_mem_bar(Op_MemBarVolatile);
duke@0 2002 }
duke@0 2003
duke@0 2004 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
duke@0 2005
duke@0 2006 return true;
duke@0 2007 }
duke@0 2008
duke@0 2009 //----------------------------inline_unsafe_prefetch----------------------------
duke@0 2010
duke@0 2011 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
duke@0 2012 #ifndef PRODUCT
duke@0 2013 {
duke@0 2014 ResourceMark rm;
duke@0 2015 // Check the signatures.
duke@0 2016 ciSignature* sig = signature();
duke@0 2017 #ifdef ASSERT
duke@0 2018 // Object getObject(Object base, int/long offset), etc.
duke@0 2019 BasicType rtype = sig->return_type()->basic_type();
duke@0 2020 if (!is_native_ptr) {
duke@0 2021 assert(sig->count() == 2, "oop prefetch has 2 arguments");
duke@0 2022 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
duke@0 2023 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
duke@0 2024 } else {
duke@0 2025 assert(sig->count() == 1, "native prefetch has 1 argument");
duke@0 2026 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
duke@0 2027 }
duke@0 2028 #endif // ASSERT
duke@0 2029 }
duke@0 2030 #endif // !PRODUCT
duke@0 2031
duke@0 2032 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@0 2033
duke@0 2034 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
duke@0 2035 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
duke@0 2036
duke@0 2037 debug_only(int saved_sp = _sp);
duke@0 2038 _sp += nargs;
duke@0 2039
duke@0 2040 // Build address expression. See the code in inline_unsafe_access.
duke@0 2041 Node *adr;
duke@0 2042 if (!is_native_ptr) {
duke@0 2043 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
duke@0 2044 Node* offset = pop_pair();
duke@0 2045 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
duke@0 2046 Node* base = pop();
duke@0 2047 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@0 2048 // to be plain byte offsets, which are also the same as those accepted
duke@0 2049 // by oopDesc::field_base.
duke@0 2050 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@0 2051 "fieldOffset must be byte-scaled");
duke@0 2052 // 32-bit machines ignore the high half!
duke@0 2053 offset = ConvL2X(offset);
duke@0 2054 adr = make_unsafe_address(base, offset);
duke@0 2055 } else {
duke@0 2056 Node* ptr = pop_pair();
duke@0 2057 // Adjust Java long to machine word:
duke@0 2058 ptr = ConvL2X(ptr);
duke@0 2059 adr = make_unsafe_address(NULL, ptr);
duke@0 2060 }
duke@0 2061
duke@0 2062 if (is_static) {
duke@0 2063 assert(saved_sp == _sp, "must have correct argument count");
duke@0 2064 } else {
duke@0 2065 // Pop receiver last: it was pushed first.
duke@0 2066 Node *receiver = pop();
duke@0 2067 assert(saved_sp == _sp, "must have correct argument count");
duke@0 2068
duke@0 2069 // Null check on self without removing any arguments. The argument
duke@0 2070 // null check technically happens in the wrong place, which can lead to
duke@0 2071 // invalid stack traces when the primitive is inlined into a method
duke@0 2072 // which handles NullPointerExceptions.
duke@0 2073 _sp += nargs;
duke@0 2074 do_null_check(receiver, T_OBJECT);
duke@0 2075 _sp -= nargs;
duke@0 2076 if (stopped()) {
duke@0 2077 return true;
duke@0 2078 }
duke@0 2079 }
duke@0 2080
duke@0 2081 // Generate the read or write prefetch
duke@0 2082 Node *prefetch;
duke@0 2083 if (is_store) {
duke@0 2084 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
duke@0 2085 } else {
duke@0 2086 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
duke@0 2087 }
duke@0 2088 prefetch->init_req(0, control());
duke@0 2089 set_i_o(_gvn.transform(prefetch));
duke@0 2090
duke@0 2091 return true;
duke@0 2092 }
duke@0 2093
duke@0 2094 //----------------------------inline_unsafe_CAS----------------------------
duke@0 2095
duke@0 2096 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
duke@0 2097 // This basic scheme here is the same as inline_unsafe_access, but
duke@0 2098 // differs in enough details that combining them would make the code
duke@0 2099 // overly confusing. (This is a true fact! I originally combined
duke@0 2100 // them, but even I was confused by it!) As much code/comments as
duke@0 2101 // possible are retained from inline_unsafe_access though to make
duke@0 2102 // the correspondances clearer. - dl
duke@0 2103
duke@0 2104 if (callee()->is_static()) return false; // caller must have the capability!
duke@0 2105
duke@0 2106 #ifndef PRODUCT
duke@0 2107 {
duke@0 2108 ResourceMark rm;
duke@0 2109 // Check the signatures.
duke@0 2110 ciSignature* sig = signature();
duke@0 2111 #ifdef ASSERT
duke@0 2112 BasicType rtype = sig->return_type()->basic_type();
duke@0 2113 assert(rtype == T_BOOLEAN, "CAS must return boolean");
duke@0 2114 assert(sig->count() == 4, "CAS has 4 arguments");
duke@0 2115 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
duke@0 2116 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
duke@0 2117 #endif // ASSERT
duke@0 2118 }
duke@0 2119 #endif //PRODUCT
duke@0 2120
duke@0 2121 // number of stack slots per value argument (1 or 2)
duke@0 2122 int type_words = type2size[type];
duke@0 2123
duke@0 2124 // Cannot inline wide CAS on machines that don't support it natively
kvn@29 2125 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
duke@0 2126 return false;
duke@0 2127
duke@0 2128 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@0 2129
duke@0 2130 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
duke@0 2131 int nargs = 1 + 1 + 2 + type_words + type_words;
duke@0 2132
duke@0 2133 // pop arguments: newval, oldval, offset, base, and receiver
duke@0 2134 debug_only(int saved_sp = _sp);
duke@0 2135 _sp += nargs;
duke@0 2136 Node* newval = (type_words == 1) ? pop() : pop_pair();
duke@0 2137 Node* oldval = (type_words == 1) ? pop() : pop_pair();
duke@0 2138 Node *offset = pop_pair();
duke@0 2139 Node *base = pop();
duke@0 2140 Node *receiver = pop();
duke@0 2141 assert(saved_sp == _sp, "must have correct argument count");
duke@0 2142
duke@0 2143 // Null check receiver.
duke@0 2144 _sp += nargs;
duke@0 2145 do_null_check(receiver, T_OBJECT);
duke@0 2146 _sp -= nargs;
duke@0 2147 if (stopped()) {
duke@0 2148 return true;
duke@0 2149 }
duke@0 2150
duke@0 2151 // Build field offset expression.
duke@0 2152 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@0 2153 // to be plain byte offsets, which are also the same as those accepted
duke@0 2154 // by oopDesc::field_base.
duke@0 2155 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
duke@0 2156 // 32-bit machines ignore the high half of long offsets
duke@0 2157 offset = ConvL2X(offset);
duke@0 2158 Node* adr = make_unsafe_address(base, offset);
duke@0 2159 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@0 2160
duke@0 2161 // (Unlike inline_unsafe_access, there seems no point in trying
duke@0 2162 // to refine types. Just use the coarse types here.
duke@0 2163 const Type *value_type = Type::get_const_basic_type(type);
duke@0 2164 Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@0 2165 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
duke@0 2166 int alias_idx = C->get_alias_index(adr_type);
duke@0 2167
duke@0 2168 // Memory-model-wise, a CAS acts like a little synchronized block,
duke@0 2169 // so needs barriers on each side. These don't't translate into
duke@0 2170 // actual barriers on most machines, but we still need rest of
duke@0 2171 // compiler to respect ordering.
duke@0 2172
duke@0 2173 insert_mem_bar(Op_MemBarRelease);
duke@0 2174 insert_mem_bar(Op_MemBarCPUOrder);
duke@0 2175
duke@0 2176 // 4984716: MemBars must be inserted before this
duke@0 2177 // memory node in order to avoid a false
duke@0 2178 // dependency which will confuse the scheduler.
duke@0 2179 Node *mem = memory(alias_idx);
duke@0 2180
duke@0 2181 // For now, we handle only those cases that actually exist: ints,
duke@0 2182 // longs, and Object. Adding others should be straightforward.
duke@0 2183 Node* cas;
duke@0 2184 switch(type) {
duke@0 2185 case T_INT:
duke@0 2186 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
duke@0 2187 break;
duke@0 2188 case T_LONG:
duke@0 2189 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
duke@0 2190 break;
duke@0 2191 case T_OBJECT:
coleenp@113 2192 // reference stores need a store barrier.
duke@0 2193 // (They don't if CAS fails, but it isn't worth checking.)
duke@0 2194 pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT);
coleenp@113 2195 #ifdef _LP64
kvn@163 2196 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@221 2197 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
kvn@221 2198 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
coleenp@113 2199 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
kvn@221 2200 newval_enc, oldval_enc));
coleenp@113 2201 } else
coleenp@113 2202 #endif
kvn@221 2203 {
kvn@221 2204 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
kvn@221 2205 }
duke@0 2206 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
duke@0 2207 break;
duke@0 2208 default:
duke@0 2209 ShouldNotReachHere();
duke@0 2210 break;
duke@0 2211 }
duke@0 2212
duke@0 2213 // SCMemProjNodes represent the memory state of CAS. Their main
duke@0 2214 // role is to prevent CAS nodes from being optimized away when their
duke@0 2215 // results aren't used.
duke@0 2216 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
duke@0 2217 set_memory(proj, alias_idx);
duke@0 2218
duke@0 2219 // Add the trailing membar surrounding the access
duke@0 2220 insert_mem_bar(Op_MemBarCPUOrder);
duke@0 2221 insert_mem_bar(Op_MemBarAcquire);
duke@0 2222
duke@0 2223 push(cas);
duke@0 2224 return true;
duke@0 2225 }
duke@0 2226
duke@0 2227 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
duke@0 2228 // This is another variant of inline_unsafe_access, differing in
duke@0 2229 // that it always issues store-store ("release") barrier and ensures
duke@0 2230 // store-atomicity (which only matters for "long").
duke@0 2231
duke@0 2232 if (callee()->is_static()) return false; // caller must have the capability!
duke@0 2233
duke@0 2234 #ifndef PRODUCT
duke@0 2235 {
duke@0 2236 ResourceMark rm;
duke@0 2237 // Check the signatures.
duke@0 2238 ciSignature* sig = signature();
duke@0 2239 #ifdef ASSERT
duke@0 2240 BasicType rtype = sig->return_type()->basic_type();
duke@0 2241 assert(rtype == T_VOID, "must return void");
duke@0 2242 assert(sig->count() == 3, "has 3 arguments");
duke@0 2243 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
duke@0 2244 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
duke@0 2245 #endif // ASSERT
duke@0 2246 }
duke@0 2247 #endif //PRODUCT
duke@0 2248
duke@0 2249 // number of stack slots per value argument (1 or 2)
duke@0 2250 int type_words = type2size[type];
duke@0 2251
duke@0 2252 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@0 2253
duke@0 2254 // Argument words: "this" plus oop plus offset plus value;
duke@0 2255 int nargs = 1 + 1 + 2 + type_words;
duke@0 2256
duke@0 2257 // pop arguments: val, offset, base, and receiver
duke@0 2258 debug_only(int saved_sp = _sp);
duke@0 2259 _sp += nargs;
duke@0 2260 Node* val = (type_words == 1) ? pop() : pop_pair();
duke@0 2261 Node *offset = pop_pair();
duke@0 2262 Node *base = pop();
duke@0 2263 Node *receiver = pop();
duke@0 2264 assert(saved_sp == _sp, "must have correct argument count");
duke@0 2265
duke@0 2266 // Null check receiver.
duke@0 2267 _sp += nargs;
duke@0 2268 do_null_check(receiver, T_OBJECT);
duke@0 2269 _sp -= nargs;
duke@0 2270 if (stopped()) {
duke@0 2271 return true;
duke@0 2272 }
duke@0 2273
duke@0 2274 // Build field offset expression.
duke@0 2275 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
duke@0 2276 // 32-bit machines ignore the high half of long offsets
duke@0 2277 offset = ConvL2X(offset);
duke@0 2278 Node* adr = make_unsafe_address(base, offset);
duke@0 2279 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@0 2280 const Type *value_type = Type::get_const_basic_type(type);
duke@0 2281 Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@0 2282
duke@0 2283 insert_mem_bar(Op_MemBarRelease);
duke@0 2284 insert_mem_bar(Op_MemBarCPUOrder);
duke@0 2285 // Ensure that the store is atomic for longs:
duke@0 2286 bool require_atomic_access = true;
duke@0 2287 Node* store;
duke@0 2288 if (type == T_OBJECT) // reference stores need a store barrier.
duke@0 2289 store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type);
duke@0 2290 else {
duke@0 2291 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
duke@0 2292 }
duke@0 2293 insert_mem_bar(Op_MemBarCPUOrder);
duke@0 2294 return true;
duke@0 2295 }
duke@0 2296
duke@0 2297 bool LibraryCallKit::inline_unsafe_allocate() {
duke@0 2298 if (callee()->is_static()) return false; // caller must have the capability!
duke@0 2299 int nargs = 1 + 1;
duke@0 2300 assert(signature()->size() == nargs-1, "alloc has 1 argument");
duke@0 2301 null_check_receiver(callee()); // check then ignore argument(0)
duke@0 2302 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2303 Node* cls = do_null_check(argument(1), T_OBJECT);
duke@0 2304 _sp -= nargs;
duke@0 2305 if (stopped()) return true;
duke@0 2306
duke@0 2307 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
duke@0 2308 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2309 kls = do_null_check(kls, T_OBJECT);
duke@0 2310 _sp -= nargs;
duke@0 2311 if (stopped()) return true; // argument was like int.class
duke@0 2312
duke@0 2313 // Note: The argument might still be an illegal value like
duke@0 2314 // Serializable.class or Object[].class. The runtime will handle it.
duke@0 2315 // But we must make an explicit check for initialization.
duke@0 2316 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
duke@0 2317 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
duke@0 2318 Node* bits = intcon(instanceKlass::fully_initialized);
duke@0 2319 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
duke@0 2320 // The 'test' is non-zero if we need to take a slow path.
duke@0 2321
duke@0 2322 Node* obj = new_instance(kls, test);
duke@0 2323 push(obj);
duke@0 2324
duke@0 2325 return true;
duke@0 2326 }
duke@0 2327
duke@0 2328 //------------------------inline_native_time_funcs--------------
duke@0 2329 // inline code for System.currentTimeMillis() and System.nanoTime()
duke@0 2330 // these have the same type and signature
duke@0 2331 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
duke@0 2332 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
duke@0 2333 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
duke@0 2334 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
duke@0 2335 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
duke@0 2336 const TypePtr* no_memory_effects = NULL;
duke@0 2337 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
duke@0 2338 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
duke@0 2339 #ifdef ASSERT
duke@0 2340 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
duke@0 2341 assert(value_top == top(), "second value must be top");
duke@0 2342 #endif
duke@0 2343 push_pair(value);
duke@0 2344 return true;
duke@0 2345 }
duke@0 2346
duke@0 2347 //------------------------inline_native_currentThread------------------
duke@0 2348 bool LibraryCallKit::inline_native_currentThread() {
duke@0 2349 Node* junk = NULL;
duke@0 2350 push(generate_current_thread(junk));
duke@0 2351 return true;
duke@0 2352 }
duke@0 2353
duke@0 2354 //------------------------inline_native_isInterrupted------------------
duke@0 2355 bool LibraryCallKit::inline_native_isInterrupted() {
duke@0 2356 const int nargs = 1+1; // receiver + boolean
duke@0 2357 assert(nargs == arg_size(), "sanity");
duke@0 2358 // Add a fast path to t.isInterrupted(clear_int):
duke@0 2359 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
duke@0 2360 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
duke@0 2361 // So, in the common case that the interrupt bit is false,
duke@0 2362 // we avoid making a call into the VM. Even if the interrupt bit
duke@0 2363 // is true, if the clear_int argument is false, we avoid the VM call.
duke@0 2364 // However, if the receiver is not currentThread, we must call the VM,
duke@0 2365 // because there must be some locking done around the operation.
duke@0 2366
duke@0 2367 // We only go to the fast case code if we pass two guards.
duke@0 2368 // Paths which do not pass are accumulated in the slow_region.
duke@0 2369 RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@0 2370 record_for_igvn(slow_region);
duke@0 2371 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
duke@0 2372 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
duke@0 2373 enum { no_int_result_path = 1,
duke@0 2374 no_clear_result_path = 2,
duke@0 2375 slow_result_path = 3
duke@0 2376 };
duke@0 2377
duke@0 2378 // (a) Receiving thread must be the current thread.
duke@0 2379 Node* rec_thr = argument(0);
duke@0 2380 Node* tls_ptr = NULL;
duke@0 2381 Node* cur_thr = generate_current_thread(tls_ptr);
duke@0 2382 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
duke@0 2383 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
duke@0 2384
duke@0 2385 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
duke@0 2386 if (!known_current_thread)
duke@0 2387 generate_slow_guard(bol_thr, slow_region);
duke@0 2388
duke@0 2389 // (b) Interrupt bit on TLS must be false.
duke@0 2390 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
duke@0 2391 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
duke@0 2392 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
duke@0 2393 Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT);
duke@0 2394 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
duke@0 2395 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
duke@0 2396
duke@0 2397 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
duke@0 2398
duke@0 2399 // First fast path: if (!TLS._interrupted) return false;
duke@0 2400 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
duke@0 2401 result_rgn->init_req(no_int_result_path, false_bit);
duke@0 2402 result_val->init_req(no_int_result_path, intcon(0));
duke@0 2403
duke@0 2404 // drop through to next case
duke@0 2405 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
duke@0 2406
duke@0 2407 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
duke@0 2408 Node* clr_arg = argument(1);
duke@0 2409 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
duke@0 2410 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
duke@0 2411 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
duke@0 2412
duke@0 2413 // Second fast path: ... else if (!clear_int) return true;
duke@0 2414 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
duke@0 2415 result_rgn->init_req(no_clear_result_path, false_arg);
duke@0 2416 result_val->init_req(no_clear_result_path, intcon(1));
duke@0 2417
duke@0 2418 // drop through to next case
duke@0 2419 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
duke@0 2420
duke@0 2421 // (d) Otherwise, go to the slow path.
duke@0 2422 slow_region->add_req(control());
duke@0 2423 set_control( _gvn.transform(slow_region) );
duke@0 2424
duke@0 2425 if (stopped()) {
duke@0 2426 // There is no slow path.
duke@0 2427 result_rgn->init_req(slow_result_path, top());
duke@0 2428 result_val->init_req(slow_result_path, top());
duke@0 2429 } else {
duke@0 2430 // non-virtual because it is a private non-static
duke@0 2431 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
duke@0 2432
duke@0 2433 Node* slow_val = set_results_for_java_call(slow_call);
duke@0 2434 // this->control() comes from set_results_for_java_call
duke@0 2435
duke@0 2436 // If we know that the result of the slow call will be true, tell the optimizer!
duke@0 2437 if (known_current_thread) slow_val = intcon(1);
duke@0 2438
duke@0 2439 Node* fast_io = slow_call->in(TypeFunc::I_O);
duke@0 2440 Node* fast_mem = slow_call->in(TypeFunc::Memory);
duke@0 2441 // These two phis are pre-filled with copies of of the fast IO and Memory
duke@0 2442 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
duke@0 2443 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
duke@0 2444
duke@0 2445 result_rgn->init_req(slow_result_path, control());
duke@0 2446 io_phi ->init_req(slow_result_path, i_o());
duke@0 2447 mem_phi ->init_req(slow_result_path, reset_memory());
duke@0 2448 result_val->init_req(slow_result_path, slow_val);
duke@0 2449
duke@0 2450 set_all_memory( _gvn.transform(mem_phi) );
duke@0 2451 set_i_o( _gvn.transform(io_phi) );
duke@0 2452 }
duke@0 2453
duke@0 2454 push_result(result_rgn, result_val);
duke@0 2455 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 2456
duke@0 2457 return true;
duke@0 2458 }
duke@0 2459
duke@0 2460 //---------------------------load_mirror_from_klass----------------------------
duke@0 2461 // Given a klass oop, load its java mirror (a java.lang.Class oop).
duke@0 2462 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
duke@0 2463 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
duke@0 2464 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
duke@0 2465 }
duke@0 2466
duke@0 2467 //-----------------------load_klass_from_mirror_common-------------------------
duke@0 2468 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
duke@0 2469 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
duke@0 2470 // and branch to the given path on the region.
duke@0 2471 // If never_see_null, take an uncommon trap on null, so we can optimistically
duke@0 2472 // compile for the non-null case.
duke@0 2473 // If the region is NULL, force never_see_null = true.
duke@0 2474 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
duke@0 2475 bool never_see_null,
duke@0 2476 int nargs,
duke@0 2477 RegionNode* region,
duke@0 2478 int null_path,
duke@0 2479 int offset) {
duke@0 2480 if (region == NULL) never_see_null = true;
duke@0 2481 Node* p = basic_plus_adr(mirror, offset);
duke@0 2482 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
kvn@164 2483 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
duke@0 2484 _sp += nargs; // any deopt will start just before call to enclosing method
duke@0 2485 Node* null_ctl = top();
duke@0 2486 kls = null_check_oop(kls, &null_ctl, never_see_null);
duke@0 2487 if (region != NULL) {
duke@0 2488 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
duke@0 2489 region->init_req(null_path, null_ctl);
duke@0 2490 } else {
duke@0 2491 assert(null_ctl == top(), "no loose ends");
duke@0 2492 }
duke@0 2493 _sp -= nargs;
duke@0 2494 return kls;
duke@0 2495 }
duke@0 2496
duke@0 2497 //--------------------(inline_native_Class_query helpers)---------------------
duke@0 2498 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
duke@0 2499 // Fall through if (mods & mask) == bits, take the guard otherwise.
duke@0 2500 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
duke@0 2501 // Branch around if the given klass has the given modifier bit set.
duke@0 2502 // Like generate_guard, adds a new path onto the region.
duke@0 2503 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
duke@0 2504 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
duke@0 2505 Node* mask = intcon(modifier_mask);
duke@0 2506 Node* bits = intcon(modifier_bits);
duke@0 2507 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
duke@0 2508 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
duke@0 2509 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
duke@0 2510 return generate_fair_guard(bol, region);
duke@0 2511 }
duke@0 2512 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
duke@0 2513 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
duke@0 2514 }
duke@0 2515
duke@0 2516 //-------------------------inline_native_Class_query-------------------
duke@0 2517 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
duke@0 2518 int nargs = 1+0; // just the Class mirror, in most cases
duke@0 2519 const Type* return_type = TypeInt::BOOL;
duke@0 2520 Node* prim_return_value = top(); // what happens if it's a primitive class?
duke@0 2521 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@0 2522 bool expect_prim = false; // most of these guys expect to work on refs
duke@0 2523
duke@0 2524 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
duke@0 2525
duke@0 2526 switch (id) {
duke@0 2527 case vmIntrinsics::_isInstance:
duke@0 2528 nargs = 1+1; // the Class mirror, plus the object getting queried about
duke@0 2529 // nothing is an instance of a primitive type
duke@0 2530 prim_return_value = intcon(0);
duke@0 2531 break;
duke@0 2532 case vmIntrinsics::_getModifiers:
duke@0 2533 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
duke@0 2534 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
duke@0 2535 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
duke@0 2536 break;
duke@0 2537 case vmIntrinsics::_isInterface:
duke@0 2538 prim_return_value = intcon(0);
duke@0 2539 break;
duke@0 2540 case vmIntrinsics::_isArray:
duke@0 2541 prim_return_value = intcon(0);
duke@0 2542 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
duke@0 2543 break;
duke@0 2544 case vmIntrinsics::_isPrimitive:
duke@0 2545 prim_return_value = intcon(1);
duke@0 2546 expect_prim = true; // obviously
duke@0 2547 break;
duke@0 2548 case vmIntrinsics::_getSuperclass:
duke@0 2549 prim_return_value = null();
duke@0 2550 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
duke@0 2551 break;
duke@0 2552 case vmIntrinsics::_getComponentType:
duke@0 2553 prim_return_value = null();
duke@0 2554 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
duke@0 2555 break;
duke@0 2556 case vmIntrinsics::_getClassAccessFlags:
duke@0 2557 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
duke@0 2558 return_type = TypeInt::INT; // not bool! 6297094
duke@0 2559 break;
duke@0 2560 default:
duke@0 2561 ShouldNotReachHere();
duke@0 2562 }
duke@0 2563
duke@0 2564 Node* mirror = argument(0);
duke@0 2565 Node* obj = (nargs <= 1)? top(): argument(1);
duke@0 2566
duke@0 2567 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
duke@0 2568 if (mirror_con == NULL) return false; // cannot happen?
duke@0 2569
duke@0 2570 #ifndef PRODUCT
duke@0 2571 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
duke@0 2572 ciType* k = mirror_con->java_mirror_type();
duke@0 2573 if (k) {
duke@0 2574 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
duke@0 2575 k->print_name();
duke@0 2576 tty->cr();
duke@0 2577 }
duke@0 2578 }
duke@0 2579 #endif
duke@0 2580
duke@0 2581 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
duke@0 2582 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@0 2583 record_for_igvn(region);
duke@0 2584 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
duke@0 2585
duke@0 2586 // The mirror will never be null of Reflection.getClassAccessFlags, however
duke@0 2587 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
duke@0 2588 // if it is. See bug 4774291.
duke@0 2589
duke@0 2590 // For Reflection.getClassAccessFlags(), the null check occurs in
duke@0 2591 // the wrong place; see inline_unsafe_access(), above, for a similar
duke@0 2592 // situation.
duke@0 2593 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2594 mirror = do_null_check(mirror, T_OBJECT);
duke@0 2595 _sp -= nargs;
duke@0 2596 // If mirror or obj is dead, only null-path is taken.
duke@0 2597 if (stopped()) return true;
duke@0 2598
duke@0 2599 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
duke@0 2600
duke@0 2601 // Now load the mirror's klass metaobject, and null-check it.
duke@0 2602 // Side-effects region with the control path if the klass is null.
duke@0 2603 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
duke@0 2604 region, _prim_path);
duke@0 2605 // If kls is null, we have a primitive mirror.
duke@0 2606 phi->init_req(_prim_path, prim_return_value);
duke@0 2607 if (stopped()) { push_result(region, phi); return true; }
duke@0 2608
duke@0 2609 Node* p; // handy temp
duke@0 2610 Node* null_ctl;
duke@0 2611
duke@0 2612 // Now that we have the non-null klass, we can perform the real query.
duke@0 2613 // For constant classes, the query will constant-fold in LoadNode::Value.
duke@0 2614 Node* query_value = top();
duke@0 2615 switch (id) {
duke@0 2616 case vmIntrinsics::_isInstance:
duke@0 2617 // nothing is an instance of a primitive type
duke@0 2618 query_value = gen_instanceof(obj, kls);
duke@0 2619 break;
duke@0 2620
duke@0 2621 case vmIntrinsics::_getModifiers:
duke@0 2622 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
duke@0 2623 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
duke@0 2624 break;
duke@0 2625
duke@0 2626 case vmIntrinsics::_isInterface:
duke@0 2627 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
duke@0 2628 if (generate_interface_guard(kls, region) != NULL)
duke@0 2629 // A guard was added. If the guard is taken, it was an interface.
duke@0 2630 phi->add_req(intcon(1));
duke@0 2631 // If we fall through, it's a plain class.
duke@0 2632 query_value = intcon(0);
duke@0 2633 break;
duke@0 2634
duke@0 2635 case vmIntrinsics::_isArray:
duke@0 2636 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
duke@0 2637 if (generate_array_guard(kls, region) != NULL)
duke@0 2638 // A guard was added. If the guard is taken, it was an array.
duke@0 2639 phi->add_req(intcon(1));
duke@0 2640 // If we fall through, it's a plain class.
duke@0 2641 query_value = intcon(0);
duke@0 2642 break;
duke@0 2643
duke@0 2644 case vmIntrinsics::_isPrimitive:
duke@0 2645 query_value = intcon(0); // "normal" path produces false
duke@0 2646 break;
duke@0 2647
duke@0 2648 case vmIntrinsics::_getSuperclass:
duke@0 2649 // The rules here are somewhat unfortunate, but we can still do better
duke@0 2650 // with random logic than with a JNI call.
duke@0 2651 // Interfaces store null or Object as _super, but must report null.
duke@0 2652 // Arrays store an intermediate super as _super, but must report Object.
duke@0 2653 // Other types can report the actual _super.
duke@0 2654 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
duke@0 2655 if (generate_interface_guard(kls, region) != NULL)
duke@0 2656 // A guard was added. If the guard is taken, it was an interface.
duke@0 2657 phi->add_req(null());
duke@0 2658 if (generate_array_guard(kls, region) != NULL)
duke@0 2659 // A guard was added. If the guard is taken, it was an array.
duke@0 2660 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
duke@0 2661 // If we fall through, it's a plain class. Get its _super.
duke@0 2662 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
kvn@164 2663 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
duke@0 2664 null_ctl = top();
duke@0 2665 kls = null_check_oop(kls, &null_ctl);
duke@0 2666 if (null_ctl != top()) {
duke@0 2667 // If the guard is taken, Object.superClass is null (both klass and mirror).
duke@0 2668 region->add_req(null_ctl);
duke@0 2669 phi ->add_req(null());
duke@0 2670 }
duke@0 2671 if (!stopped()) {
duke@0 2672 query_value = load_mirror_from_klass(kls);
duke@0 2673 }
duke@0 2674 break;
duke@0 2675
duke@0 2676 case vmIntrinsics::_getComponentType:
duke@0 2677 if (generate_array_guard(kls, region) != NULL) {
duke@0 2678 // Be sure to pin the oop load to the guard edge just created:
duke@0 2679 Node* is_array_ctrl = region->in(region->req()-1);
duke@0 2680 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
duke@0 2681 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
duke@0 2682 phi->add_req(cmo);
duke@0 2683 }
duke@0 2684 query_value = null(); // non-array case is null
duke@0 2685 break;
duke@0 2686
duke@0 2687 case vmIntrinsics::_getClassAccessFlags:
duke@0 2688 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
duke@0 2689 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
duke@0 2690 break;
duke@0 2691
duke@0 2692 default:
duke@0 2693 ShouldNotReachHere();
duke@0 2694 }
duke@0 2695
duke@0 2696 // Fall-through is the normal case of a query to a real class.
duke@0 2697 phi->init_req(1, query_value);
duke@0 2698 region->init_req(1, control());
duke@0 2699
duke@0 2700 push_result(region, phi);
duke@0 2701 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 2702
duke@0 2703 return true;
duke@0 2704 }
duke@0 2705
duke@0 2706 //--------------------------inline_native_subtype_check------------------------
duke@0 2707 // This intrinsic takes the JNI calls out of the heart of
duke@0 2708 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
duke@0 2709 bool LibraryCallKit::inline_native_subtype_check() {
duke@0 2710 int nargs = 1+1; // the Class mirror, plus the other class getting examined
duke@0 2711
duke@0 2712 // Pull both arguments off the stack.
duke@0 2713 Node* args[2]; // two java.lang.Class mirrors: superc, subc
duke@0 2714 args[0] = argument(0);
duke@0 2715 args[1] = argument(1);
duke@0 2716 Node* klasses[2]; // corresponding Klasses: superk, subk
duke@0 2717 klasses[0] = klasses[1] = top();
duke@0 2718
duke@0 2719 enum {
duke@0 2720 // A full decision tree on {superc is prim, subc is prim}:
duke@0 2721 _prim_0_path = 1, // {P,N} => false
duke@0 2722 // {P,P} & superc!=subc => false
duke@0 2723 _prim_same_path, // {P,P} & superc==subc => true
duke@0 2724 _prim_1_path, // {N,P} => false
duke@0 2725 _ref_subtype_path, // {N,N} & subtype check wins => true
duke@0 2726 _both_ref_path, // {N,N} & subtype check loses => false
duke@0 2727 PATH_LIMIT
duke@0 2728 };
duke@0 2729
duke@0 2730 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@0 2731 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
duke@0 2732 record_for_igvn(region);
duke@0 2733
duke@0 2734 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
duke@0 2735 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
duke@0 2736 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
duke@0 2737
duke@0 2738 // First null-check both mirrors and load each mirror's klass metaobject.
duke@0 2739 int which_arg;
duke@0 2740 for (which_arg = 0; which_arg <= 1; which_arg++) {
duke@0 2741 Node* arg = args[which_arg];
duke@0 2742 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2743 arg = do_null_check(arg, T_OBJECT);
duke@0 2744 _sp -= nargs;
duke@0 2745 if (stopped()) break;
duke@0 2746 args[which_arg] = _gvn.transform(arg);
duke@0 2747
duke@0 2748 Node* p = basic_plus_adr(arg, class_klass_offset);
kvn@164 2749 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
duke@0 2750 klasses[which_arg] = _gvn.transform(kls);
duke@0 2751 }
duke@0 2752
duke@0 2753 // Having loaded both klasses, test each for null.
duke@0 2754 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@0 2755 for (which_arg = 0; which_arg <= 1; which_arg++) {
duke@0 2756 Node* kls = klasses[which_arg];
duke@0 2757 Node* null_ctl = top();
duke@0 2758 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2759 kls = null_check_oop(kls, &null_ctl, never_see_null);
duke@0 2760 _sp -= nargs;
duke@0 2761 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
duke@0 2762 region->init_req(prim_path, null_ctl);
duke@0 2763 if (stopped()) break;
duke@0 2764 klasses[which_arg] = kls;
duke@0 2765 }
duke@0 2766
duke@0 2767 if (!stopped()) {
duke@0 2768 // now we have two reference types, in klasses[0..1]
duke@0 2769 Node* subk = klasses[1]; // the argument to isAssignableFrom
duke@0 2770 Node* superk = klasses[0]; // the receiver
duke@0 2771 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
duke@0 2772 // now we have a successful reference subtype check
duke@0 2773 region->set_req(_ref_subtype_path, control());
duke@0 2774 }
duke@0 2775
duke@0 2776 // If both operands are primitive (both klasses null), then
duke@0 2777 // we must return true when they are identical primitives.
duke@0 2778 // It is convenient to test this after the first null klass check.
duke@0 2779 set_control(region->in(_prim_0_path)); // go back to first null check
duke@0 2780 if (!stopped()) {
duke@0 2781 // Since superc is primitive, make a guard for the superc==subc case.
duke@0 2782 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
duke@0 2783 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
duke@0 2784 generate_guard(bol_eq, region, PROB_FAIR);
duke@0 2785 if (region->req() == PATH_LIMIT+1) {
duke@0 2786 // A guard was added. If the added guard is taken, superc==subc.
duke@0 2787 region->swap_edges(PATH_LIMIT, _prim_same_path);
duke@0 2788 region->del_req(PATH_LIMIT);
duke@0 2789 }
duke@0 2790 region->set_req(_prim_0_path, control()); // Not equal after all.
duke@0 2791 }
duke@0 2792
duke@0 2793 // these are the only paths that produce 'true':
duke@0 2794 phi->set_req(_prim_same_path, intcon(1));
duke@0 2795 phi->set_req(_ref_subtype_path, intcon(1));
duke@0 2796
duke@0 2797 // pull together the cases:
duke@0 2798 assert(region->req() == PATH_LIMIT, "sane region");
duke@0 2799 for (uint i = 1; i < region->req(); i++) {
duke@0 2800 Node* ctl = region->in(i);
duke@0 2801 if (ctl == NULL || ctl == top()) {
duke@0 2802 region->set_req(i, top());
duke@0 2803 phi ->set_req(i, top());
duke@0 2804 } else if (phi->in(i) == NULL) {
duke@0 2805 phi->set_req(i, intcon(0)); // all other paths produce 'false'
duke@0 2806 }
duke@0 2807 }
duke@0 2808
duke@0 2809 set_control(_gvn.transform(region));
duke@0 2810 push(_gvn.transform(phi));
duke@0 2811
duke@0 2812 return true;
duke@0 2813 }
duke@0 2814
duke@0 2815 //---------------------generate_array_guard_common------------------------
duke@0 2816 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
duke@0 2817 bool obj_array, bool not_array) {
duke@0 2818 // If obj_array/non_array==false/false:
duke@0 2819 // Branch around if the given klass is in fact an array (either obj or prim).
duke@0 2820 // If obj_array/non_array==false/true:
duke@0 2821 // Branch around if the given klass is not an array klass of any kind.
duke@0 2822 // If obj_array/non_array==true/true:
duke@0 2823 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
duke@0 2824 // If obj_array/non_array==true/false:
duke@0 2825 // Branch around if the kls is an oop array (Object[] or subtype)
duke@0 2826 //
duke@0 2827 // Like generate_guard, adds a new path onto the region.
duke@0 2828 jint layout_con = 0;
duke@0 2829 Node* layout_val = get_layout_helper(kls, layout_con);
duke@0 2830 if (layout_val == NULL) {
duke@0 2831 bool query = (obj_array
duke@0 2832 ? Klass::layout_helper_is_objArray(layout_con)
duke@0 2833 : Klass::layout_helper_is_javaArray(layout_con));
duke@0 2834 if (query == not_array) {
duke@0 2835 return NULL; // never a branch
duke@0 2836 } else { // always a branch
duke@0 2837 Node* always_branch = control();
duke@0 2838 if (region != NULL)
duke@0 2839 region->add_req(always_branch);
duke@0 2840 set_control(top());
duke@0 2841 return always_branch;
duke@0 2842 }
duke@0 2843 }
duke@0 2844 // Now test the correct condition.
duke@0 2845 jint nval = (obj_array
duke@0 2846 ? ((jint)Klass::_lh_array_tag_type_value
duke@0 2847 << Klass::_lh_array_tag_shift)
duke@0 2848 : Klass::_lh_neutral_value);
duke@0 2849 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
duke@0 2850 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
duke@0 2851 // invert the test if we are looking for a non-array
duke@0 2852 if (not_array) btest = BoolTest(btest).negate();
duke@0 2853 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
duke@0 2854 return generate_fair_guard(bol, region);
duke@0 2855 }
duke@0 2856
duke@0 2857
duke@0 2858 //-----------------------inline_native_newArray--------------------------
duke@0 2859 bool LibraryCallKit::inline_native_newArray() {
duke@0 2860 int nargs = 2;
duke@0 2861 Node* mirror = argument(0);
duke@0 2862 Node* count_val = argument(1);
duke@0 2863
duke@0 2864 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2865 mirror = do_null_check(mirror, T_OBJECT);
duke@0 2866 _sp -= nargs;
kvn@163 2867 // If mirror or obj is dead, only null-path is taken.
kvn@163 2868 if (stopped()) return true;
duke@0 2869
duke@0 2870 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
duke@0 2871 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@0 2872 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@0 2873 TypeInstPtr::NOTNULL);
duke@0 2874 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@0 2875 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@0 2876 TypePtr::BOTTOM);
duke@0 2877
duke@0 2878 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@0 2879 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
duke@0 2880 nargs,
duke@0 2881 result_reg, _slow_path);
duke@0 2882 Node* normal_ctl = control();
duke@0 2883 Node* no_array_ctl = result_reg->in(_slow_path);
duke@0 2884
duke@0 2885 // Generate code for the slow case. We make a call to newArray().
duke@0 2886 set_control(no_array_ctl);
duke@0 2887 if (!stopped()) {
duke@0 2888 // Either the input type is void.class, or else the
duke@0 2889 // array klass has not yet been cached. Either the
duke@0 2890 // ensuing call will throw an exception, or else it
duke@0 2891 // will cache the array klass for next time.
duke@0 2892 PreserveJVMState pjvms(this);
duke@0 2893 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
duke@0 2894 Node* slow_result = set_results_for_java_call(slow_call);
duke@0 2895 // this->control() comes from set_results_for_java_call
duke@0 2896 result_reg->set_req(_slow_path, control());
duke@0 2897 result_val->set_req(_slow_path, slow_result);
duke@0 2898 result_io ->set_req(_slow_path, i_o());
duke@0 2899 result_mem->set_req(_slow_path, reset_memory());
duke@0 2900 }
duke@0 2901
duke@0 2902 set_control(normal_ctl);
duke@0 2903 if (!stopped()) {
duke@0 2904 // Normal case: The array type has been cached in the java.lang.Class.
duke@0 2905 // The following call works fine even if the array type is polymorphic.
duke@0 2906 // It could be a dynamic mix of int[], boolean[], Object[], etc.
duke@0 2907 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2908 Node* obj = new_array(klass_node, count_val);
duke@0 2909 _sp -= nargs;
duke@0 2910 result_reg->init_req(_normal_path, control());
duke@0 2911 result_val->init_req(_normal_path, obj);
duke@0 2912 result_io ->init_req(_normal_path, i_o());
duke@0 2913 result_mem->init_req(_normal_path, reset_memory());
duke@0 2914 }
duke@0 2915
duke@0 2916 // Return the combined state.
duke@0 2917 set_i_o( _gvn.transform(result_io) );
duke@0 2918 set_all_memory( _gvn.transform(result_mem) );
duke@0 2919 push_result(result_reg, result_val);
duke@0 2920 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 2921
duke@0 2922 return true;
duke@0 2923 }
duke@0 2924
duke@0 2925 //----------------------inline_native_getLength--------------------------
duke@0 2926 bool LibraryCallKit::inline_native_getLength() {
duke@0 2927 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@0 2928
duke@0 2929 int nargs = 1;
duke@0 2930 Node* array = argument(0);
duke@0 2931
duke@0 2932 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2933 array = do_null_check(array, T_OBJECT);
duke@0 2934 _sp -= nargs;
duke@0 2935
duke@0 2936 // If array is dead, only null-path is taken.
duke@0 2937 if (stopped()) return true;
duke@0 2938
duke@0 2939 // Deoptimize if it is a non-array.
duke@0 2940 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
duke@0 2941
duke@0 2942 if (non_array != NULL) {
duke@0 2943 PreserveJVMState pjvms(this);
duke@0 2944 set_control(non_array);
duke@0 2945 _sp += nargs; // push the arguments back on the stack
duke@0 2946 uncommon_trap(Deoptimization::Reason_intrinsic,
duke@0 2947 Deoptimization::Action_maybe_recompile);
duke@0 2948 }
duke@0 2949
duke@0 2950 // If control is dead, only non-array-path is taken.
duke@0 2951 if (stopped()) return true;
duke@0 2952
duke@0 2953 // The works fine even if the array type is polymorphic.
duke@0 2954 // It could be a dynamic mix of int[], boolean[], Object[], etc.
duke@0 2955 push( load_array_length(array) );
duke@0 2956
duke@0 2957 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 2958
duke@0 2959 return true;
duke@0 2960 }
duke@0 2961
duke@0 2962 //------------------------inline_array_copyOf----------------------------
duke@0 2963 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
duke@0 2964 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@0 2965
duke@0 2966 // Restore the stack and pop off the arguments.
duke@0 2967 int nargs = 3 + (is_copyOfRange? 1: 0);
duke@0 2968 Node* original = argument(0);
duke@0 2969 Node* start = is_copyOfRange? argument(1): intcon(0);
duke@0 2970 Node* end = is_copyOfRange? argument(2): argument(1);
duke@0 2971 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
duke@0 2972
duke@0 2973 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2974 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
duke@0 2975 original = do_null_check(original, T_OBJECT);
duke@0 2976 _sp -= nargs;
duke@0 2977
duke@0 2978 // Check if a null path was taken unconditionally.
duke@0 2979 if (stopped()) return true;
duke@0 2980
duke@0 2981 Node* orig_length = load_array_length(original);
duke@0 2982
duke@0 2983 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs,
duke@0 2984 NULL, 0);
duke@0 2985 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 2986 klass_node = do_null_check(klass_node, T_OBJECT);
duke@0 2987 _sp -= nargs;
duke@0 2988
duke@0 2989 RegionNode* bailout = new (C, 1) RegionNode(1);
duke@0 2990 record_for_igvn(bailout);
duke@0 2991
duke@0 2992 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
duke@0 2993 // Bail out if that is so.
duke@0 2994 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
duke@0 2995 if (not_objArray != NULL) {
duke@0 2996 // Improve the klass node's type from the new optimistic assumption:
duke@0 2997 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
duke@0 2998 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 2999 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
duke@0 3000 cast->init_req(0, control());
duke@0 3001 klass_node = _gvn.transform(cast);
duke@0 3002 }
duke@0 3003
duke@0 3004 // Bail out if either start or end is negative.
duke@0 3005 generate_negative_guard(start, bailout, &start);
duke@0 3006 generate_negative_guard(end, bailout, &end);
duke@0 3007
duke@0 3008 Node* length = end;
duke@0 3009 if (_gvn.type(start) != TypeInt::ZERO) {
duke@0 3010 length = _gvn.transform( new (C, 3) SubINode(end, start) );
duke@0 3011 }
duke@0 3012
duke@0 3013 // Bail out if length is negative.
duke@0 3014 // ...Not needed, since the new_array will throw the right exception.
duke@0 3015 //generate_negative_guard(length, bailout, &length);
duke@0 3016
duke@0 3017 if (bailout->req() > 1) {
duke@0 3018 PreserveJVMState pjvms(this);
duke@0 3019 set_control( _gvn.transform(bailout) );
duke@0 3020 _sp += nargs; // push the arguments back on the stack
duke@0 3021 uncommon_trap(Deoptimization::Reason_intrinsic,
duke@0 3022 Deoptimization::Action_maybe_recompile);
duke@0 3023 }
duke@0 3024
duke@0 3025 if (!stopped()) {
duke@0 3026 // How many elements will we copy from the original?
duke@0 3027 // The answer is MinI(orig_length - start, length).
duke@0 3028 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
duke@0 3029 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
duke@0 3030
duke@0 3031 _sp += nargs; // set original stack for use by uncommon_trap
duke@0 3032 Node* newcopy = new_array(klass_node, length);
duke@0 3033 _sp -= nargs;
duke@0 3034
duke@0 3035 // Generate a direct call to the right arraycopy function(s).
duke@0 3036 // We know the copy is disjoint but we might not know if the
duke@0 3037 // oop stores need checking.
duke@0 3038 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
duke@0 3039 // This will fail a store-check if x contains any non-nulls.
duke@0 3040 bool disjoint_bases = true;
duke@0 3041 bool length_never_negative = true;
duke@0 3042 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
duke@0 3043 original, start, newcopy, intcon(0), moved,
duke@0 3044 nargs, disjoint_bases, length_never_negative);
duke@0 3045
duke@0 3046 push(newcopy);
duke@0 3047 }
duke@0 3048
duke@0 3049 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 3050
duke@0 3051 return true;
duke@0 3052 }
duke@0 3053
duke@0 3054
duke@0 3055 //----------------------generate_virtual_guard---------------------------
duke@0 3056 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
duke@0 3057 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
duke@0 3058 RegionNode* slow_region) {
duke@0 3059 ciMethod* method = callee();
duke@0 3060 int vtable_index = method->vtable_index();
duke@0 3061 // Get the methodOop out of the appropriate vtable entry.
duke@0 3062 int entry_offset = (instanceKlass::vtable_start_offset() +
duke@0 3063 vtable_index*vtableEntry::size()) * wordSize +
duke@0 3064 vtableEntry::method_offset_in_bytes();
duke@0 3065 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
duke@0 3066 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
duke@0 3067
duke@0 3068 // Compare the target method with the expected method (e.g., Object.hashCode).
duke@0 3069 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
duke@0 3070
duke@0 3071 Node* native_call = makecon(native_call_addr);
duke@0 3072 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
duke@0 3073 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
duke@0 3074
duke@0 3075 return generate_slow_guard(test_native, slow_region);
duke@0 3076 }
duke@0 3077
duke@0 3078 //-----------------------generate_method_call----------------------------
duke@0 3079 // Use generate_method_call to make a slow-call to the real
duke@0 3080 // method if the fast path fails. An alternative would be to
duke@0 3081 // use a stub like OptoRuntime::slow_arraycopy_Java.
duke@0 3082 // This only works for expanding the current library call,
duke@0 3083 // not another intrinsic. (E.g., don't use this for making an
duke@0 3084 // arraycopy call inside of the copyOf intrinsic.)
duke@0 3085 CallJavaNode*
duke@0 3086 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
duke@0 3087 // When compiling the intrinsic method itself, do not use this technique.
duke@0 3088 guarantee(callee() != C->method(), "cannot make slow-call to self");
duke@0 3089
duke@0 3090 ciMethod* method = callee();
duke@0 3091 // ensure the JVMS we have will be correct for this call
duke@0 3092 guarantee(method_id == method->intrinsic_id(), "must match");
duke@0 3093
duke@0 3094 const TypeFunc* tf = TypeFunc::make(method);
duke@0 3095 int tfdc = tf->domain()->cnt();
duke@0 3096 CallJavaNode* slow_call;
duke@0 3097 if (is_static) {
duke@0 3098 assert(!is_virtual, "");
duke@0 3099 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
duke@0 3100 SharedRuntime::get_resolve_static_call_stub(),
duke@0 3101 method, bci());
duke@0 3102 } else if (is_virtual) {
duke@0 3103 null_check_receiver(method);
duke@0 3104 int vtable_index = methodOopDesc::invalid_vtable_index;
duke@0 3105 if (UseInlineCaches) {
duke@0 3106 // Suppress the vtable call
duke@0 3107 } else {
duke@0 3108 // hashCode and clone are not a miranda methods,
duke@0 3109 // so the vtable index is fixed.
duke@0 3110 // No need to use the linkResolver to get it.
duke@0 3111 vtable_index = method->vtable_index();
duke@0 3112 }
duke@0 3113 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
duke@0 3114 SharedRuntime::get_resolve_virtual_call_stub(),
duke@0 3115 method, vtable_index, bci());
duke@0 3116 } else { // neither virtual nor static: opt_virtual
duke@0 3117 null_check_receiver(method);
duke@0 3118 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
duke@0 3119 SharedRuntime::get_resolve_opt_virtual_call_stub(),
duke@0 3120 method, bci());
duke@0 3121 slow_call->set_optimized_virtual(true);
duke@0 3122 }
duke@0 3123 set_arguments_for_java_call(slow_call);
duke@0 3124 set_edges_for_java_call(slow_call);
duke@0 3125 return slow_call;
duke@0 3126 }
duke@0 3127
duke@0 3128
duke@0 3129 //------------------------------inline_native_hashcode--------------------
duke@0 3130 // Build special case code for calls to hashCode on an object.
duke@0 3131 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
duke@0 3132 assert(is_static == callee()->is_static(), "correct intrinsic selection");
duke@0 3133 assert(!(is_virtual && is_static), "either virtual, special, or static");
duke@0 3134
duke@0 3135 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
duke@0 3136
duke@0 3137 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@0 3138 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@0 3139 TypeInt::INT);
duke@0 3140 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@0 3141 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@0 3142 TypePtr::BOTTOM);
duke@0 3143 Node* obj = NULL;
duke@0 3144 if (!is_static) {
duke@0 3145 // Check for hashing null object
duke@0 3146 obj = null_check_receiver(callee());
duke@0 3147 if (stopped()) return true; // unconditionally null
duke@0 3148 result_reg->init_req(_null_path, top());
duke@0 3149 result_val->init_req(_null_path, top());
duke@0 3150 } else {
duke@0 3151 // Do a null check, and return zero if null.
duke@0 3152 // System.identityHashCode(null) == 0
duke@0 3153 obj = argument(0);
duke@0 3154 Node* null_ctl = top();
duke@0 3155 obj = null_check_oop(obj, &null_ctl);
duke@0 3156 result_reg->init_req(_null_path, null_ctl);
duke@0 3157 result_val->init_req(_null_path, _gvn.intcon(0));
duke@0 3158 }
duke@0 3159
duke@0 3160 // Unconditionally null? Then return right away.
duke@0 3161 if (stopped()) {
duke@0 3162 set_control( result_reg->in(_null_path) );
duke@0 3163 if (!stopped())
duke@0 3164 push( result_val ->in(_null_path) );
duke@0 3165 return true;
duke@0 3166 }
duke@0 3167
duke@0 3168 // After null check, get the object's klass.
duke@0 3169 Node* obj_klass = load_object_klass(obj);
duke@0 3170
duke@0 3171 // This call may be virtual (invokevirtual) or bound (invokespecial).
duke@0 3172 // For each case we generate slightly different code.
duke@0 3173
duke@0 3174 // We only go to the fast case code if we pass a number of guards. The
duke@0 3175 // paths which do not pass are accumulated in the slow_region.
duke@0 3176 RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@0 3177 record_for_igvn(slow_region);
duke@0 3178
duke@0 3179 // If this is a virtual call, we generate a funny guard. We pull out
duke@0 3180 // the vtable entry corresponding to hashCode() from the target object.
duke@0 3181 // If the target method which we are calling happens to be the native
duke@0 3182 // Object hashCode() method, we pass the guard. We do not need this
duke@0 3183 // guard for non-virtual calls -- the caller is known to be the native
duke@0 3184 // Object hashCode().
duke@0 3185 if (is_virtual) {
duke@0 3186 generate_virtual_guard(obj_klass, slow_region);
duke@0 3187 }
duke@0 3188
duke@0 3189 // Get the header out of the object, use LoadMarkNode when available
duke@0 3190 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
duke@0 3191 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
duke@0 3192 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
duke@0 3193
duke@0 3194 // Test the header to see if it is unlocked.
duke@0 3195 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
duke@0 3196 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
duke@0 3197 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
duke@0 3198 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
duke@0 3199 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
duke@0 3200
duke@0 3201 generate_slow_guard(test_unlocked, slow_region);
duke@0 3202
duke@0 3203 // Get the hash value and check to see that it has been properly assigned.
duke@0 3204 // We depend on hash_mask being at most 32 bits and avoid the use of
duke@0 3205 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
duke@0 3206 // vm: see markOop.hpp.
duke@0 3207 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
duke@0 3208 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
duke@0 3209 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
duke@0 3210 // This hack lets the hash bits live anywhere in the mark object now, as long
duke@0 3211 // as the shift drops the relevent bits into the low 32 bits. Note that
duke@0 3212 // Java spec says that HashCode is an int so there's no point in capturing
duke@0 3213 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
duke@0 3214 hshifted_header = ConvX2I(hshifted_header);
duke@0 3215 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
duke@0 3216
duke@0 3217 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
duke@0 3218 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
duke@0 3219 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
duke@0 3220
duke@0 3221 generate_slow_guard(test_assigned, slow_region);
duke@0 3222
duke@0 3223 Node* init_mem = reset_memory();
duke@0 3224 // fill in the rest of the null path:
duke@0 3225 result_io ->init_req(_null_path, i_o());
duke@0 3226 result_mem->init_req(_null_path, init_mem);
duke@0 3227
duke@0 3228 result_val->init_req(_fast_path, hash_val);
duke@0 3229 result_reg->init_req(_fast_path, control());
duke@0 3230 result_io ->init_req(_fast_path, i_o());
duke@0 3231 result_mem->init_req(_fast_path, init_mem);
duke@0 3232
duke@0 3233 // Generate code for the slow case. We make a call to hashCode().
duke@0 3234 set_control(_gvn.transform(slow_region));
duke@0 3235 if (!stopped()) {
duke@0 3236 // No need for PreserveJVMState, because we're using up the present state.
duke@0 3237 set_all_memory(init_mem);
duke@0 3238 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
duke@0 3239 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
duke@0 3240 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
duke@0 3241 Node* slow_result = set_results_for_java_call(slow_call);
duke@0 3242 // this->control() comes from set_results_for_java_call
duke@0 3243 result_reg->init_req(_slow_path, control());
duke@0 3244 result_val->init_req(_slow_path, slow_result);
duke@0 3245 result_io ->set_req(_slow_path, i_o());
duke@0 3246 result_mem ->set_req(_slow_path, reset_memory());
duke@0 3247 }
duke@0 3248
duke@0 3249 // Return the combined state.
duke@0 3250 set_i_o( _gvn.transform(result_io) );
duke@0 3251 set_all_memory( _gvn.transform(result_mem) );
duke@0 3252 push_result(result_reg, result_val);
duke@0 3253
duke@0 3254 return true;
duke@0 3255 }
duke@0 3256
duke@0 3257 //---------------------------inline_native_getClass----------------------------
duke@0 3258 // Build special case code for calls to hashCode on an object.
duke@0 3259 bool LibraryCallKit::inline_native_getClass() {
duke@0 3260 Node* obj = null_check_receiver(callee());
duke@0 3261 if (stopped()) return true;
duke@0 3262 push( load_mirror_from_klass(load_object_klass(obj)) );
duke@0 3263 return true;
duke@0 3264 }
duke@0 3265
duke@0 3266 //-----------------inline_native_Reflection_getCallerClass---------------------
duke@0 3267 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
duke@0 3268 //
duke@0 3269 // NOTE that this code must perform the same logic as
duke@0 3270 // vframeStream::security_get_caller_frame in that it must skip
duke@0 3271 // Method.invoke() and auxiliary frames.
duke@0 3272
duke@0 3273
duke@0 3274
duke@0 3275
duke@0 3276 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
duke@0 3277 ciMethod* method = callee();
duke@0 3278
duke@0 3279 #ifndef PRODUCT
duke@0 3280 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@0 3281 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
duke@0 3282 }
duke@0 3283 #endif
duke@0 3284
duke@0 3285 debug_only(int saved_sp = _sp);
duke@0 3286
duke@0 3287 // Argument words: (int depth)
duke@0 3288 int nargs = 1;
duke@0 3289
duke@0 3290 _sp += nargs;
duke@0 3291 Node* caller_depth_node = pop();
duke@0 3292
duke@0 3293 assert(saved_sp == _sp, "must have correct argument count");
duke@0 3294
duke@0 3295 // The depth value must be a constant in order for the runtime call
duke@0 3296 // to be eliminated.
duke@0 3297 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
duke@0 3298 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
duke@0 3299 #ifndef PRODUCT
duke@0 3300 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@0 3301 tty->print_cr(" Bailing out because caller depth was not a constant");
duke@0 3302 }
duke@0 3303 #endif
duke@0 3304 return false;
duke@0 3305 }
duke@0 3306 // Note that the JVM state at this point does not include the
duke@0 3307 // getCallerClass() frame which we are trying to inline. The
duke@0 3308 // semantics of getCallerClass(), however, are that the "first"
duke@0 3309 // frame is the getCallerClass() frame, so we subtract one from the
duke@0 3310 // requested depth before continuing. We don't inline requests of
duke@0 3311 // getCallerClass(0).
duke@0 3312 int caller_depth = caller_depth_type->get_con() - 1;
duke@0 3313 if (caller_depth < 0) {
duke@0 3314 #ifndef PRODUCT
duke@0 3315 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@0 3316 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
duke@0 3317 }
duke@0 3318 #endif
duke@0 3319 return false;
duke@0 3320 }
duke@0 3321
duke@0 3322 if (!jvms()->has_method()) {
duke@0 3323 #ifndef PRODUCT
duke@0 3324 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@0 3325 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
duke@0 3326 }
duke@0 3327 #endif
duke@0 3328 return false;
duke@0 3329 }
duke@0 3330 int _depth = jvms()->depth(); // cache call chain depth
duke@0 3331
duke@0 3332 // Walk back up the JVM state to find the caller at the required
duke@0 3333 // depth. NOTE that this code must perform the same logic as
duke@0 3334 // vframeStream::security_get_caller_frame in that it must skip
duke@0 3335 // Method.invoke() and auxiliary frames. Note also that depth is
duke@0 3336 // 1-based (1 is the bottom of the inlining).
duke@0 3337 int inlining_depth = _depth;
duke@0 3338 JVMState* caller_jvms = NULL;
duke@0 3339
duke@0 3340 if (inlining_depth > 0) {
duke@0 3341 caller_jvms = jvms();
duke@0 3342 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
duke@0 3343 do {
duke@0 3344 // The following if-tests should be performed in this order
duke@0 3345 if (is_method_invoke_or_aux_frame(caller_jvms)) {
duke@0 3346 // Skip a Method.invoke() or auxiliary frame
duke@0 3347 } else if (caller_depth > 0) {
duke@0 3348 // Skip real frame
duke@0 3349 --caller_depth;
duke@0 3350 } else {
duke@0 3351 // We're done: reached desired caller after skipping.
duke@0 3352 break;
duke@0 3353 }
duke@0 3354 caller_jvms = caller_jvms->caller();
duke@0 3355 --inlining_depth;
duke@0 3356 } while (inlining_depth > 0);
duke@0 3357 }
duke@0 3358
duke@0 3359 if (inlining_depth == 0) {
duke@0 3360 #ifndef PRODUCT
duke@0 3361 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@0 3362 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
duke@0 3363 tty->print_cr(" JVM state at this point:");
duke@0 3364 for (int i = _depth; i >= 1; i--) {
duke@0 3365 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
duke@0 3366 }
duke@0 3367 }
duke@0 3368 #endif
duke@0 3369 return false; // Reached end of inlining
duke@0 3370 }
duke@0 3371
duke@0 3372 // Acquire method holder as java.lang.Class
duke@0 3373 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
duke@0 3374 ciInstance* caller_mirror = caller_klass->java_mirror();
duke@0 3375 // Push this as a constant
duke@0 3376 push(makecon(TypeInstPtr::make(caller_mirror)));
duke@0 3377 #ifndef PRODUCT
duke@0 3378 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@0 3379 tty->print_cr(" Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth);
duke@0 3380 tty->print_cr(" JVM state at this point:");
duke@0 3381 for (int i = _depth; i >= 1; i--) {
duke@0 3382 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
duke@0 3383 }
duke@0 3384 }
duke@0 3385 #endif
duke@0 3386 return true;
duke@0 3387 }
duke@0 3388
duke@0 3389 // Helper routine for above
duke@0 3390 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
duke@0 3391 // Is this the Method.invoke method itself?
duke@0 3392 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
duke@0 3393 return true;
duke@0 3394
duke@0 3395 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
duke@0 3396 ciKlass* k = jvms->method()->holder();
duke@0 3397 if (k->is_instance_klass()) {
duke@0 3398 ciInstanceKlass* ik = k->as_instance_klass();
duke@0 3399 for (; ik != NULL; ik = ik->super()) {
duke@0 3400 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
duke@0 3401 ik == env()->find_system_klass(ik->name())) {
duke@0 3402 return true;
duke@0 3403 }
duke@0 3404 }
duke@0 3405 }
duke@0 3406
duke@0 3407 return false;
duke@0 3408 }
duke@0 3409
duke@0 3410 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
duke@0 3411 // inline_native_AtomicLong_attemptUpdate() but it has no way of
duke@0 3412 // computing it since there is no lookup field by name function in the
duke@0 3413 // CI interface. This is computed and set by inline_native_AtomicLong_get().
duke@0 3414 // Using a static variable here is safe even if we have multiple compilation
duke@0 3415 // threads because the offset is constant. At worst the same offset will be
duke@0 3416 // computed and stored multiple
duke@0 3417
duke@0 3418 bool LibraryCallKit::inline_native_AtomicLong_get() {
duke@0 3419 // Restore the stack and pop off the argument
duke@0 3420 _sp+=1;
duke@0 3421 Node *obj = pop();
duke@0 3422
duke@0 3423 // get the offset of the "value" field. Since the CI interfaces
duke@0 3424 // does not provide a way to look up a field by name, we scan the bytecodes
duke@0 3425 // to get the field index. We expect the first 2 instructions of the method
duke@0 3426 // to be:
duke@0 3427 // 0 aload_0
duke@0 3428 // 1 getfield "value"
duke@0 3429 ciMethod* method = callee();
duke@0 3430 if (value_field_offset == -1)
duke@0 3431 {
duke@0 3432 ciField* value_field;
duke@0 3433 ciBytecodeStream iter(method);
duke@0 3434 Bytecodes::Code bc = iter.next();
duke@0 3435
duke@0 3436 if ((bc != Bytecodes::_aload_0) &&
duke@0 3437 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
duke@0 3438 return false;
duke@0 3439 bc = iter.next();
duke@0 3440 if (bc != Bytecodes::_getfield)
duke@0 3441 return false;
duke@0 3442 bool ignore;
duke@0 3443 value_field = iter.get_field(ignore);
duke@0 3444 value_field_offset = value_field->offset_in_bytes();
duke@0 3445 }
duke@0 3446
duke@0 3447 // Null check without removing any arguments.
duke@0 3448 _sp++;
duke@0 3449 obj = do_null_check(obj, T_OBJECT);
duke@0 3450 _sp--;
duke@0 3451 // Check for locking null object
duke@0 3452 if (stopped()) return true;
duke@0 3453
duke@0 3454 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
duke@0 3455 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
duke@0 3456 int alias_idx = C->get_alias_index(adr_type);
duke@0 3457
duke@0 3458 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
duke@0 3459
duke@0 3460 push_pair(result);
duke@0 3461
duke@0 3462 return true;
duke@0 3463 }
duke@0 3464
duke@0 3465 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
duke@0 3466 // Restore the stack and pop off the arguments
duke@0 3467 _sp+=5;
duke@0 3468 Node *newVal = pop_pair();
duke@0 3469 Node *oldVal = pop_pair();
duke@0 3470 Node *obj = pop();
duke@0 3471
duke@0 3472 // we need the offset of the "value" field which was computed when
duke@0 3473 // inlining the get() method. Give up if we don't have it.
duke@0 3474 if (value_field_offset == -1)
duke@0 3475 return false;
duke@0 3476
duke@0 3477 // Null check without removing any arguments.
duke@0 3478 _sp+=5;
duke@0 3479 obj = do_null_check(obj, T_OBJECT);
duke@0 3480 _sp-=5;
duke@0 3481 // Check for locking null object
duke@0 3482 if (stopped()) return true;
duke@0 3483
duke@0 3484 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
duke@0 3485 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
duke@0 3486 int alias_idx = C->get_alias_index(adr_type);
duke@0 3487
kvn@423 3488 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
kvn@423 3489 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
duke@0 3490 set_memory(store_proj, alias_idx);
kvn@423 3491 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
kvn@423 3492
kvn@423 3493 Node *result;
kvn@423 3494 // CMove node is not used to be able fold a possible check code
kvn@423 3495 // after attemptUpdate() call. This code could be transformed
kvn@423 3496 // into CMove node by loop optimizations.
kvn@423 3497 {
kvn@423 3498 RegionNode *r = new (C, 3) RegionNode(3);
kvn@423 3499 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
kvn@423 3500
kvn@423 3501 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
kvn@423 3502 Node *iftrue = opt_iff(r, iff);
kvn@423 3503 r->init_req(1, iftrue);
kvn@423 3504 result->init_req(1, intcon(1));
kvn@423 3505 result->init_req(2, intcon(0));
kvn@423 3506
kvn@423 3507 set_control(_gvn.transform(r));
kvn@423 3508 record_for_igvn(r);
kvn@423 3509
kvn@423 3510 C->set_has_split_ifs(true); // Has chance for split-if optimization
kvn@423 3511 }
kvn@423 3512
kvn@423 3513 push(_gvn.transform(result));
duke@0 3514 return true;
duke@0 3515 }
duke@0 3516
duke@0 3517 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
duke@0 3518 // restore the arguments
duke@0 3519 _sp += arg_size();
duke@0 3520
duke@0 3521 switch (id) {
duke@0 3522 case vmIntrinsics::_floatToRawIntBits:
duke@0 3523 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
duke@0 3524 break;
duke@0 3525
duke@0 3526 case vmIntrinsics::_intBitsToFloat:
duke@0 3527 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
duke@0 3528 break;
duke@0 3529
duke@0 3530 case vmIntrinsics::_doubleToRawLongBits:
duke@0 3531 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
duke@0 3532 break;
duke@0 3533
duke@0 3534 case vmIntrinsics::_longBitsToDouble:
duke@0 3535 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
duke@0 3536 break;
duke@0 3537
duke@0 3538 case vmIntrinsics::_doubleToLongBits: {
duke@0 3539 Node* value = pop_pair();
duke@0 3540
duke@0 3541 // two paths (plus control) merge in a wood
duke@0 3542 RegionNode *r = new (C, 3) RegionNode(3);
duke@0 3543 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
duke@0 3544
duke@0 3545 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
duke@0 3546 // Build the boolean node
duke@0 3547 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
duke@0 3548
duke@0 3549 // Branch either way.
duke@0 3550 // NaN case is less traveled, which makes all the difference.
duke@0 3551 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@0 3552 Node *opt_isnan = _gvn.transform(ifisnan);
duke@0 3553 assert( opt_isnan->is_If(), "Expect an IfNode");
duke@0 3554 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
duke@0 3555 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
duke@0 3556
duke@0 3557 set_control(iftrue);
duke@0 3558
duke@0 3559 static const jlong nan_bits = CONST64(0x7ff8000000000000);
duke@0 3560 Node *slow_result = longcon(nan_bits); // return NaN
duke@0 3561 phi->init_req(1, _gvn.transform( slow_result ));
duke@0 3562 r->init_req(1, iftrue);
duke@0 3563
duke@0 3564 // Else fall through
duke@0 3565 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
duke@0 3566 set_control(iffalse);
duke@0 3567
duke@0 3568 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
duke@0 3569 r->init_req(2, iffalse);
duke@0 3570
duke@0 3571 // Post merge
duke@0 3572 set_control(_gvn.transform(r));
duke@0 3573 record_for_igvn(r);
duke@0 3574
duke@0 3575 Node* result = _gvn.transform(phi);
duke@0 3576 assert(result->bottom_type()->isa_long(), "must be");
duke@0 3577 push_pair(result);
duke@0 3578
duke@0 3579 C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@0 3580
duke@0 3581 break;
duke@0 3582 }
duke@0 3583
duke@0 3584 case vmIntrinsics::_floatToIntBits: {
duke@0 3585 Node* value = pop();
duke@0 3586
duke@0 3587 // two paths (plus control) merge in a wood
duke@0 3588 RegionNode *r = new (C, 3) RegionNode(3);
duke@0 3589 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
duke@0 3590
duke@0 3591 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
duke@0 3592 // Build the boolean node
duke@0 3593 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
duke@0 3594
duke@0 3595 // Branch either way.
duke@0 3596 // NaN case is less traveled, which makes all the difference.
duke@0 3597 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@0 3598 Node *opt_isnan = _gvn.transform(ifisnan);
duke@0 3599 assert( opt_isnan->is_If(), "Expect an IfNode");
duke@0 3600 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
duke@0 3601 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
duke@0 3602
duke@0 3603 set_control(iftrue);
duke@0 3604
duke@0 3605 static const jint nan_bits = 0x7fc00000;