1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/opto/library_call.cpp Sat Dec 01 00:00:00 2007 +0000
1.3 @@ -0,0 +1,4921 @@
1.4 +/*
1.5 + * Copyright 1999-2007 Sun Microsystems, Inc. All Rights Reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
1.23 + * CA 95054 USA or visit www.sun.com if you need additional information or
1.24 + * have any questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include "incls/_precompiled.incl"
1.29 +#include "incls/_library_call.cpp.incl"
1.30 +
1.31 +class LibraryIntrinsic : public InlineCallGenerator {
1.32 + // Extend the set of intrinsics known to the runtime:
1.33 + public:
1.34 + private:
1.35 + bool _is_virtual;
1.36 + vmIntrinsics::ID _intrinsic_id;
1.37 +
1.38 + public:
1.39 + LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
1.40 + : InlineCallGenerator(m),
1.41 + _is_virtual(is_virtual),
1.42 + _intrinsic_id(id)
1.43 + {
1.44 + }
1.45 + virtual bool is_intrinsic() const { return true; }
1.46 + virtual bool is_virtual() const { return _is_virtual; }
1.47 + virtual JVMState* generate(JVMState* jvms);
1.48 + vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
1.49 +};
1.50 +
1.51 +
1.52 +// Local helper class for LibraryIntrinsic:
1.53 +class LibraryCallKit : public GraphKit {
1.54 + private:
1.55 + LibraryIntrinsic* _intrinsic; // the library intrinsic being called
1.56 +
1.57 + public:
1.58 + LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
1.59 + : GraphKit(caller),
1.60 + _intrinsic(intrinsic)
1.61 + {
1.62 + }
1.63 +
1.64 + ciMethod* caller() const { return jvms()->method(); }
1.65 + int bci() const { return jvms()->bci(); }
1.66 + LibraryIntrinsic* intrinsic() const { return _intrinsic; }
1.67 + vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
1.68 + ciMethod* callee() const { return _intrinsic->method(); }
1.69 + ciSignature* signature() const { return callee()->signature(); }
1.70 + int arg_size() const { return callee()->arg_size(); }
1.71 +
1.72 + bool try_to_inline();
1.73 +
1.74 + // Helper functions to inline natives
1.75 + void push_result(RegionNode* region, PhiNode* value);
1.76 + Node* generate_guard(Node* test, RegionNode* region, float true_prob);
1.77 + Node* generate_slow_guard(Node* test, RegionNode* region);
1.78 + Node* generate_fair_guard(Node* test, RegionNode* region);
1.79 + Node* generate_negative_guard(Node* index, RegionNode* region,
1.80 + // resulting CastII of index:
1.81 + Node* *pos_index = NULL);
1.82 + Node* generate_nonpositive_guard(Node* index, bool never_negative,
1.83 + // resulting CastII of index:
1.84 + Node* *pos_index = NULL);
1.85 + Node* generate_limit_guard(Node* offset, Node* subseq_length,
1.86 + Node* array_length,
1.87 + RegionNode* region);
1.88 + Node* generate_current_thread(Node* &tls_output);
1.89 + address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
1.90 + bool disjoint_bases, const char* &name);
1.91 + Node* load_mirror_from_klass(Node* klass);
1.92 + Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
1.93 + int nargs,
1.94 + RegionNode* region, int null_path,
1.95 + int offset);
1.96 + Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
1.97 + RegionNode* region, int null_path) {
1.98 + int offset = java_lang_Class::klass_offset_in_bytes();
1.99 + return load_klass_from_mirror_common(mirror, never_see_null, nargs,
1.100 + region, null_path,
1.101 + offset);
1.102 + }
1.103 + Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
1.104 + int nargs,
1.105 + RegionNode* region, int null_path) {
1.106 + int offset = java_lang_Class::array_klass_offset_in_bytes();
1.107 + return load_klass_from_mirror_common(mirror, never_see_null, nargs,
1.108 + region, null_path,
1.109 + offset);
1.110 + }
1.111 + Node* generate_access_flags_guard(Node* kls,
1.112 + int modifier_mask, int modifier_bits,
1.113 + RegionNode* region);
1.114 + Node* generate_interface_guard(Node* kls, RegionNode* region);
1.115 + Node* generate_array_guard(Node* kls, RegionNode* region) {
1.116 + return generate_array_guard_common(kls, region, false, false);
1.117 + }
1.118 + Node* generate_non_array_guard(Node* kls, RegionNode* region) {
1.119 + return generate_array_guard_common(kls, region, false, true);
1.120 + }
1.121 + Node* generate_objArray_guard(Node* kls, RegionNode* region) {
1.122 + return generate_array_guard_common(kls, region, true, false);
1.123 + }
1.124 + Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
1.125 + return generate_array_guard_common(kls, region, true, true);
1.126 + }
1.127 + Node* generate_array_guard_common(Node* kls, RegionNode* region,
1.128 + bool obj_array, bool not_array);
1.129 + Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
1.130 + CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
1.131 + bool is_virtual = false, bool is_static = false);
1.132 + CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
1.133 + return generate_method_call(method_id, false, true);
1.134 + }
1.135 + CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
1.136 + return generate_method_call(method_id, true, false);
1.137 + }
1.138 +
1.139 + bool inline_string_compareTo();
1.140 + bool inline_string_indexOf();
1.141 + Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
1.142 + Node* pop_math_arg();
1.143 + bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
1.144 + bool inline_math_native(vmIntrinsics::ID id);
1.145 + bool inline_trig(vmIntrinsics::ID id);
1.146 + bool inline_trans(vmIntrinsics::ID id);
1.147 + bool inline_abs(vmIntrinsics::ID id);
1.148 + bool inline_sqrt(vmIntrinsics::ID id);
1.149 + bool inline_pow(vmIntrinsics::ID id);
1.150 + bool inline_exp(vmIntrinsics::ID id);
1.151 + bool inline_min_max(vmIntrinsics::ID id);
1.152 + Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
1.153 + // This returns Type::AnyPtr, RawPtr, or OopPtr.
1.154 + int classify_unsafe_addr(Node* &base, Node* &offset);
1.155 + Node* make_unsafe_address(Node* base, Node* offset);
1.156 + bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
1.157 + bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
1.158 + bool inline_unsafe_allocate();
1.159 + bool inline_unsafe_copyMemory();
1.160 + bool inline_native_currentThread();
1.161 + bool inline_native_time_funcs(bool isNano);
1.162 + bool inline_native_isInterrupted();
1.163 + bool inline_native_Class_query(vmIntrinsics::ID id);
1.164 + bool inline_native_subtype_check();
1.165 +
1.166 + bool inline_native_newArray();
1.167 + bool inline_native_getLength();
1.168 + bool inline_array_copyOf(bool is_copyOfRange);
1.169 + bool inline_native_clone(bool is_virtual);
1.170 + bool inline_native_Reflection_getCallerClass();
1.171 + bool inline_native_AtomicLong_get();
1.172 + bool inline_native_AtomicLong_attemptUpdate();
1.173 + bool is_method_invoke_or_aux_frame(JVMState* jvms);
1.174 + // Helper function for inlining native object hash method
1.175 + bool inline_native_hashcode(bool is_virtual, bool is_static);
1.176 + bool inline_native_getClass();
1.177 +
1.178 + // Helper functions for inlining arraycopy
1.179 + bool inline_arraycopy();
1.180 + void generate_arraycopy(const TypePtr* adr_type,
1.181 + BasicType basic_elem_type,
1.182 + Node* src, Node* src_offset,
1.183 + Node* dest, Node* dest_offset,
1.184 + Node* copy_length,
1.185 + int nargs, // arguments on stack for debug info
1.186 + bool disjoint_bases = false,
1.187 + bool length_never_negative = false,
1.188 + RegionNode* slow_region = NULL);
1.189 + AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
1.190 + RegionNode* slow_region);
1.191 + void generate_clear_array(const TypePtr* adr_type,
1.192 + Node* dest,
1.193 + BasicType basic_elem_type,
1.194 + Node* slice_off,
1.195 + Node* slice_len,
1.196 + Node* slice_end);
1.197 + bool generate_block_arraycopy(const TypePtr* adr_type,
1.198 + BasicType basic_elem_type,
1.199 + AllocateNode* alloc,
1.200 + Node* src, Node* src_offset,
1.201 + Node* dest, Node* dest_offset,
1.202 + Node* dest_size);
1.203 + void generate_slow_arraycopy(const TypePtr* adr_type,
1.204 + Node* src, Node* src_offset,
1.205 + Node* dest, Node* dest_offset,
1.206 + Node* copy_length,
1.207 + int nargs);
1.208 + Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
1.209 + Node* dest_elem_klass,
1.210 + Node* src, Node* src_offset,
1.211 + Node* dest, Node* dest_offset,
1.212 + Node* copy_length, int nargs);
1.213 + Node* generate_generic_arraycopy(const TypePtr* adr_type,
1.214 + Node* src, Node* src_offset,
1.215 + Node* dest, Node* dest_offset,
1.216 + Node* copy_length, int nargs);
1.217 + void generate_unchecked_arraycopy(const TypePtr* adr_type,
1.218 + BasicType basic_elem_type,
1.219 + bool disjoint_bases,
1.220 + Node* src, Node* src_offset,
1.221 + Node* dest, Node* dest_offset,
1.222 + Node* copy_length);
1.223 + bool inline_unsafe_CAS(BasicType type);
1.224 + bool inline_unsafe_ordered_store(BasicType type);
1.225 + bool inline_fp_conversions(vmIntrinsics::ID id);
1.226 + bool inline_reverseBytes(vmIntrinsics::ID id);
1.227 +};
1.228 +
1.229 +
1.230 +//---------------------------make_vm_intrinsic----------------------------
1.231 +CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
1.232 + vmIntrinsics::ID id = m->intrinsic_id();
1.233 + assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
1.234 +
1.235 + if (DisableIntrinsic[0] != '\0'
1.236 + && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
1.237 + // disabled by a user request on the command line:
1.238 + // example: -XX:DisableIntrinsic=_hashCode,_getClass
1.239 + return NULL;
1.240 + }
1.241 +
1.242 + if (!m->is_loaded()) {
1.243 + // do not attempt to inline unloaded methods
1.244 + return NULL;
1.245 + }
1.246 +
1.247 + // Only a few intrinsics implement a virtual dispatch.
1.248 + // They are expensive calls which are also frequently overridden.
1.249 + if (is_virtual) {
1.250 + switch (id) {
1.251 + case vmIntrinsics::_hashCode:
1.252 + case vmIntrinsics::_clone:
1.253 + // OK, Object.hashCode and Object.clone intrinsics come in both flavors
1.254 + break;
1.255 + default:
1.256 + return NULL;
1.257 + }
1.258 + }
1.259 +
1.260 + // -XX:-InlineNatives disables nearly all intrinsics:
1.261 + if (!InlineNatives) {
1.262 + switch (id) {
1.263 + case vmIntrinsics::_indexOf:
1.264 + case vmIntrinsics::_compareTo:
1.265 + break; // InlineNatives does not control String.compareTo
1.266 + default:
1.267 + return NULL;
1.268 + }
1.269 + }
1.270 +
1.271 + switch (id) {
1.272 + case vmIntrinsics::_compareTo:
1.273 + if (!SpecialStringCompareTo) return NULL;
1.274 + break;
1.275 + case vmIntrinsics::_indexOf:
1.276 + if (!SpecialStringIndexOf) return NULL;
1.277 + break;
1.278 + case vmIntrinsics::_arraycopy:
1.279 + if (!InlineArrayCopy) return NULL;
1.280 + break;
1.281 + case vmIntrinsics::_copyMemory:
1.282 + if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
1.283 + if (!InlineArrayCopy) return NULL;
1.284 + break;
1.285 + case vmIntrinsics::_hashCode:
1.286 + if (!InlineObjectHash) return NULL;
1.287 + break;
1.288 + case vmIntrinsics::_clone:
1.289 + case vmIntrinsics::_copyOf:
1.290 + case vmIntrinsics::_copyOfRange:
1.291 + if (!InlineObjectCopy) return NULL;
1.292 + // These also use the arraycopy intrinsic mechanism:
1.293 + if (!InlineArrayCopy) return NULL;
1.294 + break;
1.295 + case vmIntrinsics::_checkIndex:
1.296 + // We do not intrinsify this. The optimizer does fine with it.
1.297 + return NULL;
1.298 +
1.299 + case vmIntrinsics::_get_AtomicLong:
1.300 + case vmIntrinsics::_attemptUpdate:
1.301 + if (!InlineAtomicLong) return NULL;
1.302 + break;
1.303 +
1.304 + case vmIntrinsics::_Object_init:
1.305 + case vmIntrinsics::_invoke:
1.306 + // We do not intrinsify these; they are marked for other purposes.
1.307 + return NULL;
1.308 +
1.309 + case vmIntrinsics::_getCallerClass:
1.310 + if (!UseNewReflection) return NULL;
1.311 + if (!InlineReflectionGetCallerClass) return NULL;
1.312 + if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
1.313 + break;
1.314 +
1.315 + default:
1.316 + break;
1.317 + }
1.318 +
1.319 + // -XX:-InlineClassNatives disables natives from the Class class.
1.320 + // The flag applies to all reflective calls, notably Array.newArray
1.321 + // (visible to Java programmers as Array.newInstance).
1.322 + if (m->holder()->name() == ciSymbol::java_lang_Class() ||
1.323 + m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
1.324 + if (!InlineClassNatives) return NULL;
1.325 + }
1.326 +
1.327 + // -XX:-InlineThreadNatives disables natives from the Thread class.
1.328 + if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
1.329 + if (!InlineThreadNatives) return NULL;
1.330 + }
1.331 +
1.332 + // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
1.333 + if (m->holder()->name() == ciSymbol::java_lang_Math() ||
1.334 + m->holder()->name() == ciSymbol::java_lang_Float() ||
1.335 + m->holder()->name() == ciSymbol::java_lang_Double()) {
1.336 + if (!InlineMathNatives) return NULL;
1.337 + }
1.338 +
1.339 + // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
1.340 + if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
1.341 + if (!InlineUnsafeOps) return NULL;
1.342 + }
1.343 +
1.344 + return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
1.345 +}
1.346 +
1.347 +//----------------------register_library_intrinsics-----------------------
1.348 +// Initialize this file's data structures, for each Compile instance.
1.349 +void Compile::register_library_intrinsics() {
1.350 + // Nothing to do here.
1.351 +}
1.352 +
1.353 +JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
1.354 + LibraryCallKit kit(jvms, this);
1.355 + Compile* C = kit.C;
1.356 + int nodes = C->unique();
1.357 +#ifndef PRODUCT
1.358 + if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
1.359 + char buf[1000];
1.360 + const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
1.361 + tty->print_cr("Intrinsic %s", str);
1.362 + }
1.363 +#endif
1.364 + if (kit.try_to_inline()) {
1.365 + if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
1.366 + tty->print("Inlining intrinsic %s%s at bci:%d in",
1.367 + vmIntrinsics::name_at(intrinsic_id()),
1.368 + (is_virtual() ? " (virtual)" : ""), kit.bci());
1.369 + kit.caller()->print_short_name(tty);
1.370 + tty->print_cr(" (%d bytes)", kit.caller()->code_size());
1.371 + }
1.372 + C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
1.373 + if (C->log()) {
1.374 + C->log()->elem("intrinsic id='%s'%s nodes='%d'",
1.375 + vmIntrinsics::name_at(intrinsic_id()),
1.376 + (is_virtual() ? " virtual='1'" : ""),
1.377 + C->unique() - nodes);
1.378 + }
1.379 + return kit.transfer_exceptions_into_jvms();
1.380 + }
1.381 +
1.382 + if (PrintIntrinsics) {
1.383 + switch (intrinsic_id()) {
1.384 + case vmIntrinsics::_invoke:
1.385 + case vmIntrinsics::_Object_init:
1.386 + // We do not expect to inline these, so do not produce any noise about them.
1.387 + break;
1.388 + default:
1.389 + tty->print("Did not inline intrinsic %s%s at bci:%d in",
1.390 + vmIntrinsics::name_at(intrinsic_id()),
1.391 + (is_virtual() ? " (virtual)" : ""), kit.bci());
1.392 + kit.caller()->print_short_name(tty);
1.393 + tty->print_cr(" (%d bytes)", kit.caller()->code_size());
1.394 + }
1.395 + }
1.396 + C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
1.397 + return NULL;
1.398 +}
1.399 +
1.400 +bool LibraryCallKit::try_to_inline() {
1.401 + // Handle symbolic names for otherwise undistinguished boolean switches:
1.402 + const bool is_store = true;
1.403 + const bool is_native_ptr = true;
1.404 + const bool is_static = true;
1.405 +
1.406 + switch (intrinsic_id()) {
1.407 + case vmIntrinsics::_hashCode:
1.408 + return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
1.409 + case vmIntrinsics::_identityHashCode:
1.410 + return inline_native_hashcode(/*!virtual*/ false, is_static);
1.411 + case vmIntrinsics::_getClass:
1.412 + return inline_native_getClass();
1.413 +
1.414 + case vmIntrinsics::_dsin:
1.415 + case vmIntrinsics::_dcos:
1.416 + case vmIntrinsics::_dtan:
1.417 + case vmIntrinsics::_dabs:
1.418 + case vmIntrinsics::_datan2:
1.419 + case vmIntrinsics::_dsqrt:
1.420 + case vmIntrinsics::_dexp:
1.421 + case vmIntrinsics::_dlog:
1.422 + case vmIntrinsics::_dlog10:
1.423 + case vmIntrinsics::_dpow:
1.424 + return inline_math_native(intrinsic_id());
1.425 +
1.426 + case vmIntrinsics::_min:
1.427 + case vmIntrinsics::_max:
1.428 + return inline_min_max(intrinsic_id());
1.429 +
1.430 + case vmIntrinsics::_arraycopy:
1.431 + return inline_arraycopy();
1.432 +
1.433 + case vmIntrinsics::_compareTo:
1.434 + return inline_string_compareTo();
1.435 + case vmIntrinsics::_indexOf:
1.436 + return inline_string_indexOf();
1.437 +
1.438 + case vmIntrinsics::_getObject:
1.439 + return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
1.440 + case vmIntrinsics::_getBoolean:
1.441 + return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
1.442 + case vmIntrinsics::_getByte:
1.443 + return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
1.444 + case vmIntrinsics::_getShort:
1.445 + return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
1.446 + case vmIntrinsics::_getChar:
1.447 + return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
1.448 + case vmIntrinsics::_getInt:
1.449 + return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
1.450 + case vmIntrinsics::_getLong:
1.451 + return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
1.452 + case vmIntrinsics::_getFloat:
1.453 + return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
1.454 + case vmIntrinsics::_getDouble:
1.455 + return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
1.456 +
1.457 + case vmIntrinsics::_putObject:
1.458 + return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
1.459 + case vmIntrinsics::_putBoolean:
1.460 + return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
1.461 + case vmIntrinsics::_putByte:
1.462 + return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
1.463 + case vmIntrinsics::_putShort:
1.464 + return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
1.465 + case vmIntrinsics::_putChar:
1.466 + return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
1.467 + case vmIntrinsics::_putInt:
1.468 + return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
1.469 + case vmIntrinsics::_putLong:
1.470 + return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
1.471 + case vmIntrinsics::_putFloat:
1.472 + return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
1.473 + case vmIntrinsics::_putDouble:
1.474 + return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
1.475 +
1.476 + case vmIntrinsics::_getByte_raw:
1.477 + return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
1.478 + case vmIntrinsics::_getShort_raw:
1.479 + return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
1.480 + case vmIntrinsics::_getChar_raw:
1.481 + return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
1.482 + case vmIntrinsics::_getInt_raw:
1.483 + return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
1.484 + case vmIntrinsics::_getLong_raw:
1.485 + return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
1.486 + case vmIntrinsics::_getFloat_raw:
1.487 + return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
1.488 + case vmIntrinsics::_getDouble_raw:
1.489 + return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
1.490 + case vmIntrinsics::_getAddress_raw:
1.491 + return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
1.492 +
1.493 + case vmIntrinsics::_putByte_raw:
1.494 + return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
1.495 + case vmIntrinsics::_putShort_raw:
1.496 + return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
1.497 + case vmIntrinsics::_putChar_raw:
1.498 + return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
1.499 + case vmIntrinsics::_putInt_raw:
1.500 + return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
1.501 + case vmIntrinsics::_putLong_raw:
1.502 + return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
1.503 + case vmIntrinsics::_putFloat_raw:
1.504 + return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
1.505 + case vmIntrinsics::_putDouble_raw:
1.506 + return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
1.507 + case vmIntrinsics::_putAddress_raw:
1.508 + return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
1.509 +
1.510 + case vmIntrinsics::_getObjectVolatile:
1.511 + return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
1.512 + case vmIntrinsics::_getBooleanVolatile:
1.513 + return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
1.514 + case vmIntrinsics::_getByteVolatile:
1.515 + return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
1.516 + case vmIntrinsics::_getShortVolatile:
1.517 + return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
1.518 + case vmIntrinsics::_getCharVolatile:
1.519 + return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
1.520 + case vmIntrinsics::_getIntVolatile:
1.521 + return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
1.522 + case vmIntrinsics::_getLongVolatile:
1.523 + return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
1.524 + case vmIntrinsics::_getFloatVolatile:
1.525 + return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
1.526 + case vmIntrinsics::_getDoubleVolatile:
1.527 + return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
1.528 +
1.529 + case vmIntrinsics::_putObjectVolatile:
1.530 + return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
1.531 + case vmIntrinsics::_putBooleanVolatile:
1.532 + return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
1.533 + case vmIntrinsics::_putByteVolatile:
1.534 + return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
1.535 + case vmIntrinsics::_putShortVolatile:
1.536 + return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
1.537 + case vmIntrinsics::_putCharVolatile:
1.538 + return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
1.539 + case vmIntrinsics::_putIntVolatile:
1.540 + return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
1.541 + case vmIntrinsics::_putLongVolatile:
1.542 + return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
1.543 + case vmIntrinsics::_putFloatVolatile:
1.544 + return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
1.545 + case vmIntrinsics::_putDoubleVolatile:
1.546 + return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
1.547 +
1.548 + case vmIntrinsics::_prefetchRead:
1.549 + return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
1.550 + case vmIntrinsics::_prefetchWrite:
1.551 + return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
1.552 + case vmIntrinsics::_prefetchReadStatic:
1.553 + return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
1.554 + case vmIntrinsics::_prefetchWriteStatic:
1.555 + return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
1.556 +
1.557 + case vmIntrinsics::_compareAndSwapObject:
1.558 + return inline_unsafe_CAS(T_OBJECT);
1.559 + case vmIntrinsics::_compareAndSwapInt:
1.560 + return inline_unsafe_CAS(T_INT);
1.561 + case vmIntrinsics::_compareAndSwapLong:
1.562 + return inline_unsafe_CAS(T_LONG);
1.563 +
1.564 + case vmIntrinsics::_putOrderedObject:
1.565 + return inline_unsafe_ordered_store(T_OBJECT);
1.566 + case vmIntrinsics::_putOrderedInt:
1.567 + return inline_unsafe_ordered_store(T_INT);
1.568 + case vmIntrinsics::_putOrderedLong:
1.569 + return inline_unsafe_ordered_store(T_LONG);
1.570 +
1.571 + case vmIntrinsics::_currentThread:
1.572 + return inline_native_currentThread();
1.573 + case vmIntrinsics::_isInterrupted:
1.574 + return inline_native_isInterrupted();
1.575 +
1.576 + case vmIntrinsics::_currentTimeMillis:
1.577 + return inline_native_time_funcs(false);
1.578 + case vmIntrinsics::_nanoTime:
1.579 + return inline_native_time_funcs(true);
1.580 + case vmIntrinsics::_allocateInstance:
1.581 + return inline_unsafe_allocate();
1.582 + case vmIntrinsics::_copyMemory:
1.583 + return inline_unsafe_copyMemory();
1.584 + case vmIntrinsics::_newArray:
1.585 + return inline_native_newArray();
1.586 + case vmIntrinsics::_getLength:
1.587 + return inline_native_getLength();
1.588 + case vmIntrinsics::_copyOf:
1.589 + return inline_array_copyOf(false);
1.590 + case vmIntrinsics::_copyOfRange:
1.591 + return inline_array_copyOf(true);
1.592 + case vmIntrinsics::_clone:
1.593 + return inline_native_clone(intrinsic()->is_virtual());
1.594 +
1.595 + case vmIntrinsics::_isAssignableFrom:
1.596 + return inline_native_subtype_check();
1.597 +
1.598 + case vmIntrinsics::_isInstance:
1.599 + case vmIntrinsics::_getModifiers:
1.600 + case vmIntrinsics::_isInterface:
1.601 + case vmIntrinsics::_isArray:
1.602 + case vmIntrinsics::_isPrimitive:
1.603 + case vmIntrinsics::_getSuperclass:
1.604 + case vmIntrinsics::_getComponentType:
1.605 + case vmIntrinsics::_getClassAccessFlags:
1.606 + return inline_native_Class_query(intrinsic_id());
1.607 +
1.608 + case vmIntrinsics::_floatToRawIntBits:
1.609 + case vmIntrinsics::_floatToIntBits:
1.610 + case vmIntrinsics::_intBitsToFloat:
1.611 + case vmIntrinsics::_doubleToRawLongBits:
1.612 + case vmIntrinsics::_doubleToLongBits:
1.613 + case vmIntrinsics::_longBitsToDouble:
1.614 + return inline_fp_conversions(intrinsic_id());
1.615 +
1.616 + case vmIntrinsics::_reverseBytes_i:
1.617 + case vmIntrinsics::_reverseBytes_l:
1.618 + return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
1.619 +
1.620 + case vmIntrinsics::_get_AtomicLong:
1.621 + return inline_native_AtomicLong_get();
1.622 + case vmIntrinsics::_attemptUpdate:
1.623 + return inline_native_AtomicLong_attemptUpdate();
1.624 +
1.625 + case vmIntrinsics::_getCallerClass:
1.626 + return inline_native_Reflection_getCallerClass();
1.627 +
1.628 + default:
1.629 + // If you get here, it may be that someone has added a new intrinsic
1.630 + // to the list in vmSymbols.hpp without implementing it here.
1.631 +#ifndef PRODUCT
1.632 + if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1.633 + tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
1.634 + vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1.635 + }
1.636 +#endif
1.637 + return false;
1.638 + }
1.639 +}
1.640 +
1.641 +//------------------------------push_result------------------------------
1.642 +// Helper function for finishing intrinsics.
1.643 +void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
1.644 + record_for_igvn(region);
1.645 + set_control(_gvn.transform(region));
1.646 + BasicType value_type = value->type()->basic_type();
1.647 + push_node(value_type, _gvn.transform(value));
1.648 +}
1.649 +
1.650 +//------------------------------generate_guard---------------------------
1.651 +// Helper function for generating guarded fast-slow graph structures.
1.652 +// The given 'test', if true, guards a slow path. If the test fails
1.653 +// then a fast path can be taken. (We generally hope it fails.)
1.654 +// In all cases, GraphKit::control() is updated to the fast path.
1.655 +// The returned value represents the control for the slow path.
1.656 +// The return value is never 'top'; it is either a valid control
1.657 +// or NULL if it is obvious that the slow path can never be taken.
1.658 +// Also, if region and the slow control are not NULL, the slow edge
1.659 +// is appended to the region.
1.660 +Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1.661 + if (stopped()) {
1.662 + // Already short circuited.
1.663 + return NULL;
1.664 + }
1.665 +
1.666 + // Build an if node and its projections.
1.667 + // If test is true we take the slow path, which we assume is uncommon.
1.668 + if (_gvn.type(test) == TypeInt::ZERO) {
1.669 + // The slow branch is never taken. No need to build this guard.
1.670 + return NULL;
1.671 + }
1.672 +
1.673 + IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1.674 +
1.675 + Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
1.676 + if (if_slow == top()) {
1.677 + // The slow branch is never taken. No need to build this guard.
1.678 + return NULL;
1.679 + }
1.680 +
1.681 + if (region != NULL)
1.682 + region->add_req(if_slow);
1.683 +
1.684 + Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
1.685 + set_control(if_fast);
1.686 +
1.687 + return if_slow;
1.688 +}
1.689 +
1.690 +inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1.691 + return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1.692 +}
1.693 +inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1.694 + return generate_guard(test, region, PROB_FAIR);
1.695 +}
1.696 +
1.697 +inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1.698 + Node* *pos_index) {
1.699 + if (stopped())
1.700 + return NULL; // already stopped
1.701 + if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1.702 + return NULL; // index is already adequately typed
1.703 + Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
1.704 + Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
1.705 + Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1.706 + if (is_neg != NULL && pos_index != NULL) {
1.707 + // Emulate effect of Parse::adjust_map_after_if.
1.708 + Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
1.709 + ccast->set_req(0, control());
1.710 + (*pos_index) = _gvn.transform(ccast);
1.711 + }
1.712 + return is_neg;
1.713 +}
1.714 +
1.715 +inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1.716 + Node* *pos_index) {
1.717 + if (stopped())
1.718 + return NULL; // already stopped
1.719 + if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1.720 + return NULL; // index is already adequately typed
1.721 + Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
1.722 + BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1.723 + Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
1.724 + Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1.725 + if (is_notp != NULL && pos_index != NULL) {
1.726 + // Emulate effect of Parse::adjust_map_after_if.
1.727 + Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
1.728 + ccast->set_req(0, control());
1.729 + (*pos_index) = _gvn.transform(ccast);
1.730 + }
1.731 + return is_notp;
1.732 +}
1.733 +
1.734 +// Make sure that 'position' is a valid limit index, in [0..length].
1.735 +// There are two equivalent plans for checking this:
1.736 +// A. (offset + copyLength) unsigned<= arrayLength
1.737 +// B. offset <= (arrayLength - copyLength)
1.738 +// We require that all of the values above, except for the sum and
1.739 +// difference, are already known to be non-negative.
1.740 +// Plan A is robust in the face of overflow, if offset and copyLength
1.741 +// are both hugely positive.
1.742 +//
1.743 +// Plan B is less direct and intuitive, but it does not overflow at
1.744 +// all, since the difference of two non-negatives is always
1.745 +// representable. Whenever Java methods must perform the equivalent
1.746 +// check they generally use Plan B instead of Plan A.
1.747 +// For the moment we use Plan A.
1.748 +inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1.749 + Node* subseq_length,
1.750 + Node* array_length,
1.751 + RegionNode* region) {
1.752 + if (stopped())
1.753 + return NULL; // already stopped
1.754 + bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1.755 + if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
1.756 + return NULL; // common case of whole-array copy
1.757 + Node* last = subseq_length;
1.758 + if (!zero_offset) // last += offset
1.759 + last = _gvn.transform( new (C, 3) AddINode(last, offset));
1.760 + Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
1.761 + Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
1.762 + Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1.763 + return is_over;
1.764 +}
1.765 +
1.766 +
1.767 +//--------------------------generate_current_thread--------------------
1.768 +Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1.769 + ciKlass* thread_klass = env()->Thread_klass();
1.770 + const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1.771 + Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
1.772 + Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1.773 + Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
1.774 + tls_output = thread;
1.775 + return threadObj;
1.776 +}
1.777 +
1.778 +
1.779 +//------------------------------inline_string_compareTo------------------------
1.780 +bool LibraryCallKit::inline_string_compareTo() {
1.781 +
1.782 + const int value_offset = java_lang_String::value_offset_in_bytes();
1.783 + const int count_offset = java_lang_String::count_offset_in_bytes();
1.784 + const int offset_offset = java_lang_String::offset_offset_in_bytes();
1.785 +
1.786 + _sp += 2;
1.787 + Node *argument = pop(); // pop non-receiver first: it was pushed second
1.788 + Node *receiver = pop();
1.789 +
1.790 + // Null check on self without removing any arguments. The argument
1.791 + // null check technically happens in the wrong place, which can lead to
1.792 + // invalid stack traces when string compare is inlined into a method
1.793 + // which handles NullPointerExceptions.
1.794 + _sp += 2;
1.795 + receiver = do_null_check(receiver, T_OBJECT);
1.796 + argument = do_null_check(argument, T_OBJECT);
1.797 + _sp -= 2;
1.798 + if (stopped()) {
1.799 + return true;
1.800 + }
1.801 +
1.802 + ciInstanceKlass* klass = env()->String_klass();
1.803 + const TypeInstPtr* string_type =
1.804 + TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1.805 +
1.806 + Node* compare =
1.807 + _gvn.transform(new (C, 7) StrCompNode(
1.808 + control(),
1.809 + memory(TypeAryPtr::CHARS),
1.810 + memory(string_type->add_offset(value_offset)),
1.811 + memory(string_type->add_offset(count_offset)),
1.812 + memory(string_type->add_offset(offset_offset)),
1.813 + receiver,
1.814 + argument));
1.815 + push(compare);
1.816 + return true;
1.817 +}
1.818 +
1.819 +// Java version of String.indexOf(constant string)
1.820 +// class StringDecl {
1.821 +// StringDecl(char[] ca) {
1.822 +// offset = 0;
1.823 +// count = ca.length;
1.824 +// value = ca;
1.825 +// }
1.826 +// int offset;
1.827 +// int count;
1.828 +// char[] value;
1.829 +// }
1.830 +//
1.831 +// static int string_indexOf_J(StringDecl string_object, char[] target_object,
1.832 +// int targetOffset, int cache_i, int md2) {
1.833 +// int cache = cache_i;
1.834 +// int sourceOffset = string_object.offset;
1.835 +// int sourceCount = string_object.count;
1.836 +// int targetCount = target_object.length;
1.837 +//
1.838 +// int targetCountLess1 = targetCount - 1;
1.839 +// int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1.840 +//
1.841 +// char[] source = string_object.value;
1.842 +// char[] target = target_object;
1.843 +// int lastChar = target[targetCountLess1];
1.844 +//
1.845 +// outer_loop:
1.846 +// for (int i = sourceOffset; i < sourceEnd; ) {
1.847 +// int src = source[i + targetCountLess1];
1.848 +// if (src == lastChar) {
1.849 +// // With random strings and a 4-character alphabet,
1.850 +// // reverse matching at this point sets up 0.8% fewer
1.851 +// // frames, but (paradoxically) makes 0.3% more probes.
1.852 +// // Since those probes are nearer the lastChar probe,
1.853 +// // there is may be a net D$ win with reverse matching.
1.854 +// // But, reversing loop inhibits unroll of inner loop
1.855 +// // for unknown reason. So, does running outer loop from
1.856 +// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1.857 +// for (int j = 0; j < targetCountLess1; j++) {
1.858 +// if (target[targetOffset + j] != source[i+j]) {
1.859 +// if ((cache & (1 << source[i+j])) == 0) {
1.860 +// if (md2 < j+1) {
1.861 +// i += j+1;
1.862 +// continue outer_loop;
1.863 +// }
1.864 +// }
1.865 +// i += md2;
1.866 +// continue outer_loop;
1.867 +// }
1.868 +// }
1.869 +// return i - sourceOffset;
1.870 +// }
1.871 +// if ((cache & (1 << src)) == 0) {
1.872 +// i += targetCountLess1;
1.873 +// } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1.874 +// i++;
1.875 +// }
1.876 +// return -1;
1.877 +// }
1.878 +
1.879 +//------------------------------string_indexOf------------------------
1.880 +Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1.881 + jint cache_i, jint md2_i) {
1.882 +
1.883 + Node* no_ctrl = NULL;
1.884 + float likely = PROB_LIKELY(0.9);
1.885 + float unlikely = PROB_UNLIKELY(0.9);
1.886 +
1.887 + const int value_offset = java_lang_String::value_offset_in_bytes();
1.888 + const int count_offset = java_lang_String::count_offset_in_bytes();
1.889 + const int offset_offset = java_lang_String::offset_offset_in_bytes();
1.890 +
1.891 + ciInstanceKlass* klass = env()->String_klass();
1.892 + const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1.893 + const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1.894 +
1.895 + Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1.896 + Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1.897 + Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1.898 + Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1.899 + Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1.900 + Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1.901 +
1.902 + Node* target = _gvn.transform(ConPNode::make(C, target_array));
1.903 + jint target_length = target_array->length();
1.904 + const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1.905 + const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1.906 +
1.907 + IdealKit kit(gvn(), control(), merged_memory());
1.908 +#define __ kit.
1.909 + Node* zero = __ ConI(0);
1.910 + Node* one = __ ConI(1);
1.911 + Node* cache = __ ConI(cache_i);
1.912 + Node* md2 = __ ConI(md2_i);
1.913 + Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1.914 + Node* targetCount = __ ConI(target_length);
1.915 + Node* targetCountLess1 = __ ConI(target_length - 1);
1.916 + Node* targetOffset = __ ConI(targetOffset_i);
1.917 + Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1.918 +
1.919 + IdealVariable rtn(kit), i(kit), j(kit); __ declares_done();
1.920 + Node* outer_loop = __ make_label(2 /* goto */);
1.921 + Node* return_ = __ make_label(1);
1.922 +
1.923 + __ set(rtn,__ ConI(-1));
1.924 + __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
1.925 + Node* i2 = __ AddI(__ value(i), targetCountLess1);
1.926 + // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1.927 + Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1.928 + __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1.929 + __ loop(j, zero, BoolTest::lt, targetCountLess1); {
1.930 + Node* tpj = __ AddI(targetOffset, __ value(j));
1.931 + Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1.932 + Node* ipj = __ AddI(__ value(i), __ value(j));
1.933 + Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1.934 + __ if_then(targ, BoolTest::ne, src2); {
1.935 + __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1.936 + __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1.937 + __ increment(i, __ AddI(__ value(j), one));
1.938 + __ goto_(outer_loop);
1.939 + } __ end_if(); __ dead(j);
1.940 + }__ end_if(); __ dead(j);
1.941 + __ increment(i, md2);
1.942 + __ goto_(outer_loop);
1.943 + }__ end_if();
1.944 + __ increment(j, one);
1.945 + }__ end_loop(); __ dead(j);
1.946 + __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1.947 + __ goto_(return_);
1.948 + }__ end_if();
1.949 + __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1.950 + __ increment(i, targetCountLess1);
1.951 + }__ end_if();
1.952 + __ increment(i, one);
1.953 + __ bind(outer_loop);
1.954 + }__ end_loop(); __ dead(i);
1.955 + __ bind(return_);
1.956 + __ drain_delay_transform();
1.957 +
1.958 + set_control(__ ctrl());
1.959 + Node* result = __ value(rtn);
1.960 +#undef __
1.961 + C->set_has_loops(true);
1.962 + return result;
1.963 +}
1.964 +
1.965 +
1.966 +//------------------------------inline_string_indexOf------------------------
1.967 +bool LibraryCallKit::inline_string_indexOf() {
1.968 +
1.969 + _sp += 2;
1.970 + Node *argument = pop(); // pop non-receiver first: it was pushed second
1.971 + Node *receiver = pop();
1.972 +
1.973 + // don't intrinsify is argument isn't a constant string.
1.974 + if (!argument->is_Con()) {
1.975 + return false;
1.976 + }
1.977 + const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1.978 + if (str_type == NULL) {
1.979 + return false;
1.980 + }
1.981 + ciInstanceKlass* klass = env()->String_klass();
1.982 + ciObject* str_const = str_type->const_oop();
1.983 + if (str_const == NULL || str_const->klass() != klass) {
1.984 + return false;
1.985 + }
1.986 + ciInstance* str = str_const->as_instance();
1.987 + assert(str != NULL, "must be instance");
1.988 +
1.989 + const int value_offset = java_lang_String::value_offset_in_bytes();
1.990 + const int count_offset = java_lang_String::count_offset_in_bytes();
1.991 + const int offset_offset = java_lang_String::offset_offset_in_bytes();
1.992 +
1.993 + ciObject* v = str->field_value_by_offset(value_offset).as_object();
1.994 + int o = str->field_value_by_offset(offset_offset).as_int();
1.995 + int c = str->field_value_by_offset(count_offset).as_int();
1.996 + ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1.997 +
1.998 + // constant strings have no offset and count == length which
1.999 + // simplifies the resulting code somewhat so lets optimize for that.
1.1000 + if (o != 0 || c != pat->length()) {
1.1001 + return false;
1.1002 + }
1.1003 +
1.1004 + // Null check on self without removing any arguments. The argument
1.1005 + // null check technically happens in the wrong place, which can lead to
1.1006 + // invalid stack traces when string compare is inlined into a method
1.1007 + // which handles NullPointerExceptions.
1.1008 + _sp += 2;
1.1009 + receiver = do_null_check(receiver, T_OBJECT);
1.1010 + // No null check on the argument is needed since it's a constant String oop.
1.1011 + _sp -= 2;
1.1012 + if (stopped()) {
1.1013 + return true;
1.1014 + }
1.1015 +
1.1016 + // The null string as a pattern always returns 0 (match at beginning of string)
1.1017 + if (c == 0) {
1.1018 + push(intcon(0));
1.1019 + return true;
1.1020 + }
1.1021 +
1.1022 + jchar lastChar = pat->char_at(o + (c - 1));
1.1023 + int cache = 0;
1.1024 + int i;
1.1025 + for (i = 0; i < c - 1; i++) {
1.1026 + assert(i < pat->length(), "out of range");
1.1027 + cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1.1028 + }
1.1029 +
1.1030 + int md2 = c;
1.1031 + for (i = 0; i < c - 1; i++) {
1.1032 + assert(i < pat->length(), "out of range");
1.1033 + if (pat->char_at(o + i) == lastChar) {
1.1034 + md2 = (c - 1) - i;
1.1035 + }
1.1036 + }
1.1037 +
1.1038 + Node* result = string_indexOf(receiver, pat, o, cache, md2);
1.1039 + push(result);
1.1040 + return true;
1.1041 +}
1.1042 +
1.1043 +//--------------------------pop_math_arg--------------------------------
1.1044 +// Pop a double argument to a math function from the stack
1.1045 +// rounding it if necessary.
1.1046 +Node * LibraryCallKit::pop_math_arg() {
1.1047 + Node *arg = pop_pair();
1.1048 + if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1.1049 + arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1.1050 + return arg;
1.1051 +}
1.1052 +
1.1053 +//------------------------------inline_trig----------------------------------
1.1054 +// Inline sin/cos/tan instructions, if possible. If rounding is required, do
1.1055 +// argument reduction which will turn into a fast/slow diamond.
1.1056 +bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1.1057 + _sp += arg_size(); // restore stack pointer
1.1058 + Node* arg = pop_math_arg();
1.1059 + Node* trig = NULL;
1.1060 +
1.1061 + switch (id) {
1.1062 + case vmIntrinsics::_dsin:
1.1063 + trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1.1064 + break;
1.1065 + case vmIntrinsics::_dcos:
1.1066 + trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1.1067 + break;
1.1068 + case vmIntrinsics::_dtan:
1.1069 + trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1.1070 + break;
1.1071 + default:
1.1072 + assert(false, "bad intrinsic was passed in");
1.1073 + return false;
1.1074 + }
1.1075 +
1.1076 + // Rounding required? Check for argument reduction!
1.1077 + if( Matcher::strict_fp_requires_explicit_rounding ) {
1.1078 +
1.1079 + static const double pi_4 = 0.7853981633974483;
1.1080 + static const double neg_pi_4 = -0.7853981633974483;
1.1081 + // pi/2 in 80-bit extended precision
1.1082 + // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1.1083 + // -pi/2 in 80-bit extended precision
1.1084 + // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1.1085 + // Cutoff value for using this argument reduction technique
1.1086 + //static const double pi_2_minus_epsilon = 1.564660403643354;
1.1087 + //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1.1088 +
1.1089 + // Pseudocode for sin:
1.1090 + // if (x <= Math.PI / 4.0) {
1.1091 + // if (x >= -Math.PI / 4.0) return fsin(x);
1.1092 + // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1.1093 + // } else {
1.1094 + // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1.1095 + // }
1.1096 + // return StrictMath.sin(x);
1.1097 +
1.1098 + // Pseudocode for cos:
1.1099 + // if (x <= Math.PI / 4.0) {
1.1100 + // if (x >= -Math.PI / 4.0) return fcos(x);
1.1101 + // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1.1102 + // } else {
1.1103 + // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1.1104 + // }
1.1105 + // return StrictMath.cos(x);
1.1106 +
1.1107 + // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1.1108 + // requires a special machine instruction to load it. Instead we'll try
1.1109 + // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1.1110 + // probably do the math inside the SIN encoding.
1.1111 +
1.1112 + // Make the merge point
1.1113 + RegionNode *r = new (C, 3) RegionNode(3);
1.1114 + Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1.1115 +
1.1116 + // Flatten arg so we need only 1 test
1.1117 + Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1.1118 + // Node for PI/4 constant
1.1119 + Node *pi4 = makecon(TypeD::make(pi_4));
1.1120 + // Check PI/4 : abs(arg)
1.1121 + Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1.1122 + // Check: If PI/4 < abs(arg) then go slow
1.1123 + Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1.1124 + // Branch either way
1.1125 + IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1.1126 + set_control(opt_iff(r,iff));
1.1127 +
1.1128 + // Set fast path result
1.1129 + phi->init_req(2,trig);
1.1130 +
1.1131 + // Slow path - non-blocking leaf call
1.1132 + Node* call = NULL;
1.1133 + switch (id) {
1.1134 + case vmIntrinsics::_dsin:
1.1135 + call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1.1136 + CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1.1137 + "Sin", NULL, arg, top());
1.1138 + break;
1.1139 + case vmIntrinsics::_dcos:
1.1140 + call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1.1141 + CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1.1142 + "Cos", NULL, arg, top());
1.1143 + break;
1.1144 + case vmIntrinsics::_dtan:
1.1145 + call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1.1146 + CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1.1147 + "Tan", NULL, arg, top());
1.1148 + break;
1.1149 + }
1.1150 + assert(control()->in(0) == call, "");
1.1151 + Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1.1152 + r->init_req(1,control());
1.1153 + phi->init_req(1,slow_result);
1.1154 +
1.1155 + // Post-merge
1.1156 + set_control(_gvn.transform(r));
1.1157 + record_for_igvn(r);
1.1158 + trig = _gvn.transform(phi);
1.1159 +
1.1160 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1161 + }
1.1162 + // Push result back on JVM stack
1.1163 + push_pair(trig);
1.1164 + return true;
1.1165 +}
1.1166 +
1.1167 +//------------------------------inline_sqrt-------------------------------------
1.1168 +// Inline square root instruction, if possible.
1.1169 +bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1.1170 + assert(id == vmIntrinsics::_dsqrt, "Not square root");
1.1171 + _sp += arg_size(); // restore stack pointer
1.1172 + push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1.1173 + return true;
1.1174 +}
1.1175 +
1.1176 +//------------------------------inline_abs-------------------------------------
1.1177 +// Inline absolute value instruction, if possible.
1.1178 +bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1.1179 + assert(id == vmIntrinsics::_dabs, "Not absolute value");
1.1180 + _sp += arg_size(); // restore stack pointer
1.1181 + push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1.1182 + return true;
1.1183 +}
1.1184 +
1.1185 +//------------------------------inline_exp-------------------------------------
1.1186 +// Inline exp instructions, if possible. The Intel hardware only misses
1.1187 +// really odd corner cases (+/- Infinity). Just uncommon-trap them.
1.1188 +bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1.1189 + assert(id == vmIntrinsics::_dexp, "Not exp");
1.1190 +
1.1191 + // If this inlining ever returned NaN in the past, we do not intrinsify it
1.1192 + // every again. NaN results requires StrictMath.exp handling.
1.1193 + if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1.1194 +
1.1195 + // Do not intrinsify on older platforms which lack cmove.
1.1196 + if (ConditionalMoveLimit == 0) return false;
1.1197 +
1.1198 + _sp += arg_size(); // restore stack pointer
1.1199 + Node *x = pop_math_arg();
1.1200 + Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1.1201 +
1.1202 + //-------------------
1.1203 + //result=(result.isNaN())? StrictMath::exp():result;
1.1204 + // Check: If isNaN() by checking result!=result? then go to Strict Math
1.1205 + Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1.1206 + // Build the boolean node
1.1207 + Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1.1208 +
1.1209 + { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1.1210 + // End the current control-flow path
1.1211 + push_pair(x);
1.1212 + // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1.1213 + // to handle. Recompile without intrinsifying Math.exp
1.1214 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.1215 + Deoptimization::Action_make_not_entrant);
1.1216 + }
1.1217 +
1.1218 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1219 +
1.1220 + push_pair(result);
1.1221 +
1.1222 + return true;
1.1223 +}
1.1224 +
1.1225 +//------------------------------inline_pow-------------------------------------
1.1226 +// Inline power instructions, if possible.
1.1227 +bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1.1228 + assert(id == vmIntrinsics::_dpow, "Not pow");
1.1229 +
1.1230 + // If this inlining ever returned NaN in the past, we do not intrinsify it
1.1231 + // every again. NaN results requires StrictMath.pow handling.
1.1232 + if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1.1233 +
1.1234 + // Do not intrinsify on older platforms which lack cmove.
1.1235 + if (ConditionalMoveLimit == 0) return false;
1.1236 +
1.1237 + // Pseudocode for pow
1.1238 + // if (x <= 0.0) {
1.1239 + // if ((double)((int)y)==y) { // if y is int
1.1240 + // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1.1241 + // } else {
1.1242 + // result = NaN;
1.1243 + // }
1.1244 + // } else {
1.1245 + // result = DPow(x,y);
1.1246 + // }
1.1247 + // if (result != result)? {
1.1248 + // ucommon_trap();
1.1249 + // }
1.1250 + // return result;
1.1251 +
1.1252 + _sp += arg_size(); // restore stack pointer
1.1253 + Node* y = pop_math_arg();
1.1254 + Node* x = pop_math_arg();
1.1255 +
1.1256 + Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1.1257 +
1.1258 + // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1.1259 + // inside of something) then skip the fancy tests and just check for
1.1260 + // NaN result.
1.1261 + Node *result = NULL;
1.1262 + if( jvms()->depth() >= 1 ) {
1.1263 + result = fast_result;
1.1264 + } else {
1.1265 +
1.1266 + // Set the merge point for If node with condition of (x <= 0.0)
1.1267 + // There are four possible paths to region node and phi node
1.1268 + RegionNode *r = new (C, 4) RegionNode(4);
1.1269 + Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1.1270 +
1.1271 + // Build the first if node: if (x <= 0.0)
1.1272 + // Node for 0 constant
1.1273 + Node *zeronode = makecon(TypeD::ZERO);
1.1274 + // Check x:0
1.1275 + Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1.1276 + // Check: If (x<=0) then go complex path
1.1277 + Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1.1278 + // Branch either way
1.1279 + IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1.1280 + Node *opt_test = _gvn.transform(if1);
1.1281 + //assert( opt_test->is_If(), "Expect an IfNode");
1.1282 + IfNode *opt_if1 = (IfNode*)opt_test;
1.1283 + // Fast path taken; set region slot 3
1.1284 + Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1.1285 + r->init_req(3,fast_taken); // Capture fast-control
1.1286 +
1.1287 + // Fast path not-taken, i.e. slow path
1.1288 + Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1.1289 +
1.1290 + // Set fast path result
1.1291 + Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1.1292 + phi->init_req(3, fast_result);
1.1293 +
1.1294 + // Complex path
1.1295 + // Build the second if node (if y is int)
1.1296 + // Node for (int)y
1.1297 + Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1.1298 + // Node for (double)((int) y)
1.1299 + Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1.1300 + // Check (double)((int) y) : y
1.1301 + Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1.1302 + // Check if (y isn't int) then go to slow path
1.1303 +
1.1304 + Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1.1305 + // Branch eith way
1.1306 + IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1.1307 + Node *slow_path = opt_iff(r,if2); // Set region path 2
1.1308 +
1.1309 + // Calculate DPow(abs(x), y)*(1 & (int)y)
1.1310 + // Node for constant 1
1.1311 + Node *conone = intcon(1);
1.1312 + // 1& (int)y
1.1313 + Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1.1314 + // zero node
1.1315 + Node *conzero = intcon(0);
1.1316 + // Check (1&(int)y)==0?
1.1317 + Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1.1318 + // Check if (1&(int)y)!=0?, if so the result is negative
1.1319 + Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1.1320 + // abs(x)
1.1321 + Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1.1322 + // abs(x)^y
1.1323 + Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1.1324 + // -abs(x)^y
1.1325 + Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1.1326 + // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1.1327 + Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1.1328 + // Set complex path fast result
1.1329 + phi->init_req(2, signresult);
1.1330 +
1.1331 + static const jlong nan_bits = CONST64(0x7ff8000000000000);
1.1332 + Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1.1333 + r->init_req(1,slow_path);
1.1334 + phi->init_req(1,slow_result);
1.1335 +
1.1336 + // Post merge
1.1337 + set_control(_gvn.transform(r));
1.1338 + record_for_igvn(r);
1.1339 + result=_gvn.transform(phi);
1.1340 + }
1.1341 +
1.1342 + //-------------------
1.1343 + //result=(result.isNaN())? uncommon_trap():result;
1.1344 + // Check: If isNaN() by checking result!=result? then go to Strict Math
1.1345 + Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1.1346 + // Build the boolean node
1.1347 + Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1.1348 +
1.1349 + { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1.1350 + // End the current control-flow path
1.1351 + push_pair(x);
1.1352 + push_pair(y);
1.1353 + // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1.1354 + // to handle. Recompile without intrinsifying Math.pow.
1.1355 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.1356 + Deoptimization::Action_make_not_entrant);
1.1357 + }
1.1358 +
1.1359 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.1360 +
1.1361 + push_pair(result);
1.1362 +
1.1363 + return true;
1.1364 +}
1.1365 +
1.1366 +//------------------------------inline_trans-------------------------------------
1.1367 +// Inline transcendental instructions, if possible. The Intel hardware gets
1.1368 +// these right, no funny corner cases missed.
1.1369 +bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1.1370 + _sp += arg_size(); // restore stack pointer
1.1371 + Node* arg = pop_math_arg();
1.1372 + Node* trans = NULL;
1.1373 +
1.1374 + switch (id) {
1.1375 + case vmIntrinsics::_dlog:
1.1376 + trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1.1377 + break;
1.1378 + case vmIntrinsics::_dlog10:
1.1379 + trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1.1380 + break;
1.1381 + default:
1.1382 + assert(false, "bad intrinsic was passed in");
1.1383 + return false;
1.1384 + }
1.1385 +
1.1386 + // Push result back on JVM stack
1.1387 + push_pair(trans);
1.1388 + return true;
1.1389 +}
1.1390 +
1.1391 +//------------------------------runtime_math-----------------------------
1.1392 +bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1.1393 + Node* a = NULL;
1.1394 + Node* b = NULL;
1.1395 +
1.1396 + assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1.1397 + "must be (DD)D or (D)D type");
1.1398 +
1.1399 + // Inputs
1.1400 + _sp += arg_size(); // restore stack pointer
1.1401 + if (call_type == OptoRuntime::Math_DD_D_Type()) {
1.1402 + b = pop_math_arg();
1.1403 + }
1.1404 + a = pop_math_arg();
1.1405 +
1.1406 + const TypePtr* no_memory_effects = NULL;
1.1407 + Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1.1408 + no_memory_effects,
1.1409 + a, top(), b, b ? top() : NULL);
1.1410 + Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1.1411 +#ifdef ASSERT
1.1412 + Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1.1413 + assert(value_top == top(), "second value must be top");
1.1414 +#endif
1.1415 +
1.1416 + push_pair(value);
1.1417 + return true;
1.1418 +}
1.1419 +
1.1420 +//------------------------------inline_math_native-----------------------------
1.1421 +bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1.1422 + switch (id) {
1.1423 + // These intrinsics are not properly supported on all hardware
1.1424 + case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1.1425 + runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1.1426 + case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1.1427 + runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1.1428 + case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1.1429 + runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1.1430 +
1.1431 + case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1.1432 + runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1.1433 + case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1.1434 + runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1.1435 +
1.1436 + // These intrinsics are supported on all hardware
1.1437 + case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1.1438 + case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1.1439 +
1.1440 + // These intrinsics don't work on X86. The ad implementation doesn't
1.1441 + // handle NaN's properly. Instead of returning infinity, the ad
1.1442 + // implementation returns a NaN on overflow. See bug: 6304089
1.1443 + // Once the ad implementations are fixed, change the code below
1.1444 + // to match the intrinsics above
1.1445 +
1.1446 + case vmIntrinsics::_dexp: return
1.1447 + runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1.1448 + case vmIntrinsics::_dpow: return
1.1449 + runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1.1450 +
1.1451 + // These intrinsics are not yet correctly implemented
1.1452 + case vmIntrinsics::_datan2:
1.1453 + return false;
1.1454 +
1.1455 + default:
1.1456 + ShouldNotReachHere();
1.1457 + return false;
1.1458 + }
1.1459 +}
1.1460 +
1.1461 +static bool is_simple_name(Node* n) {
1.1462 + return (n->req() == 1 // constant
1.1463 + || (n->is_Type() && n->as_Type()->type()->singleton())
1.1464 + || n->is_Proj() // parameter or return value
1.1465 + || n->is_Phi() // local of some sort
1.1466 + );
1.1467 +}
1.1468 +
1.1469 +//----------------------------inline_min_max-----------------------------------
1.1470 +bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1.1471 + push(generate_min_max(id, argument(0), argument(1)));
1.1472 +
1.1473 + return true;
1.1474 +}
1.1475 +
1.1476 +Node*
1.1477 +LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1.1478 + // These are the candidate return value:
1.1479 + Node* xvalue = x0;
1.1480 + Node* yvalue = y0;
1.1481 +
1.1482 + if (xvalue == yvalue) {
1.1483 + return xvalue;
1.1484 + }
1.1485 +
1.1486 + bool want_max = (id == vmIntrinsics::_max);
1.1487 +
1.1488 + const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1.1489 + const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1.1490 + if (txvalue == NULL || tyvalue == NULL) return top();
1.1491 + // This is not really necessary, but it is consistent with a
1.1492 + // hypothetical MaxINode::Value method:
1.1493 + int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1.1494 +
1.1495 + // %%% This folding logic should (ideally) be in a different place.
1.1496 + // Some should be inside IfNode, and there to be a more reliable
1.1497 + // transformation of ?: style patterns into cmoves. We also want
1.1498 + // more powerful optimizations around cmove and min/max.
1.1499 +
1.1500 + // Try to find a dominating comparison of these guys.
1.1501 + // It can simplify the index computation for Arrays.copyOf
1.1502 + // and similar uses of System.arraycopy.
1.1503 + // First, compute the normalized version of CmpI(x, y).
1.1504 + int cmp_op = Op_CmpI;
1.1505 + Node* xkey = xvalue;
1.1506 + Node* ykey = yvalue;
1.1507 + Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1.1508 + if (ideal_cmpxy->is_Cmp()) {
1.1509 + // E.g., if we have CmpI(length - offset, count),
1.1510 + // it might idealize to CmpI(length, count + offset)
1.1511 + cmp_op = ideal_cmpxy->Opcode();
1.1512 + xkey = ideal_cmpxy->in(1);
1.1513 + ykey = ideal_cmpxy->in(2);
1.1514 + }
1.1515 +
1.1516 + // Start by locating any relevant comparisons.
1.1517 + Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1.1518 + Node* cmpxy = NULL;
1.1519 + Node* cmpyx = NULL;
1.1520 + for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1.1521 + Node* cmp = start_from->fast_out(k);
1.1522 + if (cmp->outcnt() > 0 && // must have prior uses
1.1523 + cmp->in(0) == NULL && // must be context-independent
1.1524 + cmp->Opcode() == cmp_op) { // right kind of compare
1.1525 + if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1.1526 + if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1.1527 + }
1.1528 + }
1.1529 +
1.1530 + const int NCMPS = 2;
1.1531 + Node* cmps[NCMPS] = { cmpxy, cmpyx };
1.1532 + int cmpn;
1.1533 + for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1.1534 + if (cmps[cmpn] != NULL) break; // find a result
1.1535 + }
1.1536 + if (cmpn < NCMPS) {
1.1537 + // Look for a dominating test that tells us the min and max.
1.1538 + int depth = 0; // Limit search depth for speed
1.1539 + Node* dom = control();
1.1540 + for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1.1541 + if (++depth >= 100) break;
1.1542 + Node* ifproj = dom;
1.1543 + if (!ifproj->is_Proj()) continue;
1.1544 + Node* iff = ifproj->in(0);
1.1545 + if (!iff->is_If()) continue;
1.1546 + Node* bol = iff->in(1);
1.1547 + if (!bol->is_Bool()) continue;
1.1548 + Node* cmp = bol->in(1);
1.1549 + if (cmp == NULL) continue;
1.1550 + for (cmpn = 0; cmpn < NCMPS; cmpn++)
1.1551 + if (cmps[cmpn] == cmp) break;
1.1552 + if (cmpn == NCMPS) continue;
1.1553 + BoolTest::mask btest = bol->as_Bool()->_test._test;
1.1554 + if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1.1555 + if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1.1556 + // At this point, we know that 'x btest y' is true.
1.1557 + switch (btest) {
1.1558 + case BoolTest::eq:
1.1559 + // They are proven equal, so we can collapse the min/max.
1.1560 + // Either value is the answer. Choose the simpler.
1.1561 + if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1.1562 + return yvalue;
1.1563 + return xvalue;
1.1564 + case BoolTest::lt: // x < y
1.1565 + case BoolTest::le: // x <= y
1.1566 + return (want_max ? yvalue : xvalue);
1.1567 + case BoolTest::gt: // x > y
1.1568 + case BoolTest::ge: // x >= y
1.1569 + return (want_max ? xvalue : yvalue);
1.1570 + }
1.1571 + }
1.1572 + }
1.1573 +
1.1574 + // We failed to find a dominating test.
1.1575 + // Let's pick a test that might GVN with prior tests.
1.1576 + Node* best_bol = NULL;
1.1577 + BoolTest::mask best_btest = BoolTest::illegal;
1.1578 + for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1.1579 + Node* cmp = cmps[cmpn];
1.1580 + if (cmp == NULL) continue;
1.1581 + for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1.1582 + Node* bol = cmp->fast_out(j);
1.1583 + if (!bol->is_Bool()) continue;
1.1584 + BoolTest::mask btest = bol->as_Bool()->_test._test;
1.1585 + if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1.1586 + if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1.1587 + if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1.1588 + best_bol = bol->as_Bool();
1.1589 + best_btest = btest;
1.1590 + }
1.1591 + }
1.1592 + }
1.1593 +
1.1594 + Node* answer_if_true = NULL;
1.1595 + Node* answer_if_false = NULL;
1.1596 + switch (best_btest) {
1.1597 + default:
1.1598 + if (cmpxy == NULL)
1.1599 + cmpxy = ideal_cmpxy;
1.1600 + best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1.1601 + // and fall through:
1.1602 + case BoolTest::lt: // x < y
1.1603 + case BoolTest::le: // x <= y
1.1604 + answer_if_true = (want_max ? yvalue : xvalue);
1.1605 + answer_if_false = (want_max ? xvalue : yvalue);
1.1606 + break;
1.1607 + case BoolTest::gt: // x > y
1.1608 + case BoolTest::ge: // x >= y
1.1609 + answer_if_true = (want_max ? xvalue : yvalue);
1.1610 + answer_if_false = (want_max ? yvalue : xvalue);
1.1611 + break;
1.1612 + }
1.1613 +
1.1614 + jint hi, lo;
1.1615 + if (want_max) {
1.1616 + // We can sharpen the minimum.
1.1617 + hi = MAX2(txvalue->_hi, tyvalue->_hi);
1.1618 + lo = MAX2(txvalue->_lo, tyvalue->_lo);
1.1619 + } else {
1.1620 + // We can sharpen the maximum.
1.1621 + hi = MIN2(txvalue->_hi, tyvalue->_hi);
1.1622 + lo = MIN2(txvalue->_lo, tyvalue->_lo);
1.1623 + }
1.1624 +
1.1625 + // Use a flow-free graph structure, to avoid creating excess control edges
1.1626 + // which could hinder other optimizations.
1.1627 + // Since Math.min/max is often used with arraycopy, we want
1.1628 + // tightly_coupled_allocation to be able to see beyond min/max expressions.
1.1629 + Node* cmov = CMoveNode::make(C, NULL, best_bol,
1.1630 + answer_if_false, answer_if_true,
1.1631 + TypeInt::make(lo, hi, widen));
1.1632 +
1.1633 + return _gvn.transform(cmov);
1.1634 +
1.1635 + /*
1.1636 + // This is not as desirable as it may seem, since Min and Max
1.1637 + // nodes do not have a full set of optimizations.
1.1638 + // And they would interfere, anyway, with 'if' optimizations
1.1639 + // and with CMoveI canonical forms.
1.1640 + switch (id) {
1.1641 + case vmIntrinsics::_min:
1.1642 + result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1.1643 + case vmIntrinsics::_max:
1.1644 + result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1.1645 + default:
1.1646 + ShouldNotReachHere();
1.1647 + }
1.1648 + */
1.1649 +}
1.1650 +
1.1651 +inline int
1.1652 +LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1.1653 + const TypePtr* base_type = TypePtr::NULL_PTR;
1.1654 + if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1.1655 + if (base_type == NULL) {
1.1656 + // Unknown type.
1.1657 + return Type::AnyPtr;
1.1658 + } else if (base_type == TypePtr::NULL_PTR) {
1.1659 + // Since this is a NULL+long form, we have to switch to a rawptr.
1.1660 + base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1.1661 + offset = MakeConX(0);
1.1662 + return Type::RawPtr;
1.1663 + } else if (base_type->base() == Type::RawPtr) {
1.1664 + return Type::RawPtr;
1.1665 + } else if (base_type->isa_oopptr()) {
1.1666 + // Base is never null => always a heap address.
1.1667 + if (base_type->ptr() == TypePtr::NotNull) {
1.1668 + return Type::OopPtr;
1.1669 + }
1.1670 + // Offset is small => always a heap address.
1.1671 + const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1.1672 + if (offset_type != NULL &&
1.1673 + base_type->offset() == 0 && // (should always be?)
1.1674 + offset_type->_lo >= 0 &&
1.1675 + !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1.1676 + return Type::OopPtr;
1.1677 + }
1.1678 + // Otherwise, it might either be oop+off or NULL+addr.
1.1679 + return Type::AnyPtr;
1.1680 + } else {
1.1681 + // No information:
1.1682 + return Type::AnyPtr;
1.1683 + }
1.1684 +}
1.1685 +
1.1686 +inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1.1687 + int kind = classify_unsafe_addr(base, offset);
1.1688 + if (kind == Type::RawPtr) {
1.1689 + return basic_plus_adr(top(), base, offset);
1.1690 + } else {
1.1691 + return basic_plus_adr(base, offset);
1.1692 + }
1.1693 +}
1.1694 +
1.1695 +//----------------------------inline_reverseBytes_int/long-------------------
1.1696 +// inline Int.reverseBytes(int)
1.1697 +// inline Long.reverseByes(long)
1.1698 +bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
1.1699 + assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
1.1700 + if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
1.1701 + if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
1.1702 + _sp += arg_size(); // restore stack pointer
1.1703 + switch (id) {
1.1704 + case vmIntrinsics::_reverseBytes_i:
1.1705 + push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
1.1706 + break;
1.1707 + case vmIntrinsics::_reverseBytes_l:
1.1708 + push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
1.1709 + break;
1.1710 + default:
1.1711 + ;
1.1712 + }
1.1713 + return true;
1.1714 +}
1.1715 +
1.1716 +//----------------------------inline_unsafe_access----------------------------
1.1717 +
1.1718 +const static BasicType T_ADDRESS_HOLDER = T_LONG;
1.1719 +
1.1720 +// Interpret Unsafe.fieldOffset cookies correctly:
1.1721 +extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
1.1722 +
1.1723 +bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
1.1724 + if (callee()->is_static()) return false; // caller must have the capability!
1.1725 +
1.1726 +#ifndef PRODUCT
1.1727 + {
1.1728 + ResourceMark rm;
1.1729 + // Check the signatures.
1.1730 + ciSignature* sig = signature();
1.1731 +#ifdef ASSERT
1.1732 + if (!is_store) {
1.1733 + // Object getObject(Object base, int/long offset), etc.
1.1734 + BasicType rtype = sig->return_type()->basic_type();
1.1735 + if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
1.1736 + rtype = T_ADDRESS; // it is really a C void*
1.1737 + assert(rtype == type, "getter must return the expected value");
1.1738 + if (!is_native_ptr) {
1.1739 + assert(sig->count() == 2, "oop getter has 2 arguments");
1.1740 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
1.1741 + assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
1.1742 + } else {
1.1743 + assert(sig->count() == 1, "native getter has 1 argument");
1.1744 + assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
1.1745 + }
1.1746 + } else {
1.1747 + // void putObject(Object base, int/long offset, Object x), etc.
1.1748 + assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
1.1749 + if (!is_native_ptr) {
1.1750 + assert(sig->count() == 3, "oop putter has 3 arguments");
1.1751 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
1.1752 + assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
1.1753 + } else {
1.1754 + assert(sig->count() == 2, "native putter has 2 arguments");
1.1755 + assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
1.1756 + }
1.1757 + BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
1.1758 + if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
1.1759 + vtype = T_ADDRESS; // it is really a C void*
1.1760 + assert(vtype == type, "putter must accept the expected value");
1.1761 + }
1.1762 +#endif // ASSERT
1.1763 + }
1.1764 +#endif //PRODUCT
1.1765 +
1.1766 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.1767 +
1.1768 + int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
1.1769 +
1.1770 + // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
1.1771 + int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
1.1772 +
1.1773 + debug_only(int saved_sp = _sp);
1.1774 + _sp += nargs;
1.1775 +
1.1776 + Node* val;
1.1777 + debug_only(val = (Node*)(uintptr_t)-1);
1.1778 +
1.1779 +
1.1780 + if (is_store) {
1.1781 + // Get the value being stored. (Pop it first; it was pushed last.)
1.1782 + switch (type) {
1.1783 + case T_DOUBLE:
1.1784 + case T_LONG:
1.1785 + case T_ADDRESS:
1.1786 + val = pop_pair();
1.1787 + break;
1.1788 + default:
1.1789 + val = pop();
1.1790 + }
1.1791 + }
1.1792 +
1.1793 + // Build address expression. See the code in inline_unsafe_prefetch.
1.1794 + Node *adr;
1.1795 + Node *heap_base_oop = top();
1.1796 + if (!is_native_ptr) {
1.1797 + // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
1.1798 + Node* offset = pop_pair();
1.1799 + // The base is either a Java object or a value produced by Unsafe.staticFieldBase
1.1800 + Node* base = pop();
1.1801 + // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1.1802 + // to be plain byte offsets, which are also the same as those accepted
1.1803 + // by oopDesc::field_base.
1.1804 + assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1.1805 + "fieldOffset must be byte-scaled");
1.1806 + // 32-bit machines ignore the high half!
1.1807 + offset = ConvL2X(offset);
1.1808 + adr = make_unsafe_address(base, offset);
1.1809 + heap_base_oop = base;
1.1810 + } else {
1.1811 + Node* ptr = pop_pair();
1.1812 + // Adjust Java long to machine word:
1.1813 + ptr = ConvL2X(ptr);
1.1814 + adr = make_unsafe_address(NULL, ptr);
1.1815 + }
1.1816 +
1.1817 + // Pop receiver last: it was pushed first.
1.1818 + Node *receiver = pop();
1.1819 +
1.1820 + assert(saved_sp == _sp, "must have correct argument count");
1.1821 +
1.1822 + const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1.1823 +
1.1824 + // First guess at the value type.
1.1825 + const Type *value_type = Type::get_const_basic_type(type);
1.1826 +
1.1827 + // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
1.1828 + // there was not enough information to nail it down.
1.1829 + Compile::AliasType* alias_type = C->alias_type(adr_type);
1.1830 + assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
1.1831 +
1.1832 + // We will need memory barriers unless we can determine a unique
1.1833 + // alias category for this reference. (Note: If for some reason
1.1834 + // the barriers get omitted and the unsafe reference begins to "pollute"
1.1835 + // the alias analysis of the rest of the graph, either Compile::can_alias
1.1836 + // or Compile::must_alias will throw a diagnostic assert.)
1.1837 + bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
1.1838 +
1.1839 + if (!is_store && type == T_OBJECT) {
1.1840 + // Attempt to infer a sharper value type from the offset and base type.
1.1841 + ciKlass* sharpened_klass = NULL;
1.1842 +
1.1843 + // See if it is an instance field, with an object type.
1.1844 + if (alias_type->field() != NULL) {
1.1845 + assert(!is_native_ptr, "native pointer op cannot use a java address");
1.1846 + if (alias_type->field()->type()->is_klass()) {
1.1847 + sharpened_klass = alias_type->field()->type()->as_klass();
1.1848 + }
1.1849 + }
1.1850 +
1.1851 + // See if it is a narrow oop array.
1.1852 + if (adr_type->isa_aryptr()) {
1.1853 + if (adr_type->offset() >= objArrayOopDesc::header_size() * wordSize) {
1.1854 + const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
1.1855 + if (elem_type != NULL) {
1.1856 + sharpened_klass = elem_type->klass();
1.1857 + }
1.1858 + }
1.1859 + }
1.1860 +
1.1861 + if (sharpened_klass != NULL) {
1.1862 + const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
1.1863 +
1.1864 + // Sharpen the value type.
1.1865 + value_type = tjp;
1.1866 +
1.1867 +#ifndef PRODUCT
1.1868 + if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
1.1869 + tty->print(" from base type: "); adr_type->dump();
1.1870 + tty->print(" sharpened value: "); value_type->dump();
1.1871 + }
1.1872 +#endif
1.1873 + }
1.1874 + }
1.1875 +
1.1876 + // Null check on self without removing any arguments. The argument
1.1877 + // null check technically happens in the wrong place, which can lead to
1.1878 + // invalid stack traces when the primitive is inlined into a method
1.1879 + // which handles NullPointerExceptions.
1.1880 + _sp += nargs;
1.1881 + do_null_check(receiver, T_OBJECT);
1.1882 + _sp -= nargs;
1.1883 + if (stopped()) {
1.1884 + return true;
1.1885 + }
1.1886 + // Heap pointers get a null-check from the interpreter,
1.1887 + // as a courtesy. However, this is not guaranteed by Unsafe,
1.1888 + // and it is not possible to fully distinguish unintended nulls
1.1889 + // from intended ones in this API.
1.1890 +
1.1891 + if (is_volatile) {
1.1892 + // We need to emit leading and trailing CPU membars (see below) in
1.1893 + // addition to memory membars when is_volatile. This is a little
1.1894 + // too strong, but avoids the need to insert per-alias-type
1.1895 + // volatile membars (for stores; compare Parse::do_put_xxx), which
1.1896 + // we cannot do effctively here because we probably only have a
1.1897 + // rough approximation of type.
1.1898 + need_mem_bar = true;
1.1899 + // For Stores, place a memory ordering barrier now.
1.1900 + if (is_store)
1.1901 + insert_mem_bar(Op_MemBarRelease);
1.1902 + }
1.1903 +
1.1904 + // Memory barrier to prevent normal and 'unsafe' accesses from
1.1905 + // bypassing each other. Happens after null checks, so the
1.1906 + // exception paths do not take memory state from the memory barrier,
1.1907 + // so there's no problems making a strong assert about mixing users
1.1908 + // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
1.1909 + // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
1.1910 + if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
1.1911 +
1.1912 + if (!is_store) {
1.1913 + Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
1.1914 + // load value and push onto stack
1.1915 + switch (type) {
1.1916 + case T_BOOLEAN:
1.1917 + case T_CHAR:
1.1918 + case T_BYTE:
1.1919 + case T_SHORT:
1.1920 + case T_INT:
1.1921 + case T_FLOAT:
1.1922 + case T_OBJECT:
1.1923 + push( p );
1.1924 + break;
1.1925 + case T_ADDRESS:
1.1926 + // Cast to an int type.
1.1927 + p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
1.1928 + p = ConvX2L(p);
1.1929 + push_pair(p);
1.1930 + break;
1.1931 + case T_DOUBLE:
1.1932 + case T_LONG:
1.1933 + push_pair( p );
1.1934 + break;
1.1935 + default: ShouldNotReachHere();
1.1936 + }
1.1937 + } else {
1.1938 + // place effect of store into memory
1.1939 + switch (type) {
1.1940 + case T_DOUBLE:
1.1941 + val = dstore_rounding(val);
1.1942 + break;
1.1943 + case T_ADDRESS:
1.1944 + // Repackage the long as a pointer.
1.1945 + val = ConvL2X(val);
1.1946 + val = _gvn.transform( new (C, 2) CastX2PNode(val) );
1.1947 + break;
1.1948 + }
1.1949 +
1.1950 + if (type != T_OBJECT ) {
1.1951 + (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
1.1952 + } else {
1.1953 + // Possibly an oop being stored to Java heap or native memory
1.1954 + if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
1.1955 + // oop to Java heap.
1.1956 + (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
1.1957 + } else {
1.1958 +
1.1959 + // We can't tell at compile time if we are storing in the Java heap or outside
1.1960 + // of it. So we need to emit code to conditionally do the proper type of
1.1961 + // store.
1.1962 +
1.1963 + IdealKit kit(gvn(), control(), merged_memory());
1.1964 + kit.declares_done();
1.1965 + // QQQ who knows what probability is here??
1.1966 + kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
1.1967 + (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
1.1968 + } kit.else_(); {
1.1969 + (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
1.1970 + } kit.end_if();
1.1971 + }
1.1972 + }
1.1973 + }
1.1974 +
1.1975 + if (is_volatile) {
1.1976 + if (!is_store)
1.1977 + insert_mem_bar(Op_MemBarAcquire);
1.1978 + else
1.1979 + insert_mem_bar(Op_MemBarVolatile);
1.1980 + }
1.1981 +
1.1982 + if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
1.1983 +
1.1984 + return true;
1.1985 +}
1.1986 +
1.1987 +//----------------------------inline_unsafe_prefetch----------------------------
1.1988 +
1.1989 +bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
1.1990 +#ifndef PRODUCT
1.1991 + {
1.1992 + ResourceMark rm;
1.1993 + // Check the signatures.
1.1994 + ciSignature* sig = signature();
1.1995 +#ifdef ASSERT
1.1996 + // Object getObject(Object base, int/long offset), etc.
1.1997 + BasicType rtype = sig->return_type()->basic_type();
1.1998 + if (!is_native_ptr) {
1.1999 + assert(sig->count() == 2, "oop prefetch has 2 arguments");
1.2000 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
1.2001 + assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
1.2002 + } else {
1.2003 + assert(sig->count() == 1, "native prefetch has 1 argument");
1.2004 + assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
1.2005 + }
1.2006 +#endif // ASSERT
1.2007 + }
1.2008 +#endif // !PRODUCT
1.2009 +
1.2010 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.2011 +
1.2012 + // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
1.2013 + int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
1.2014 +
1.2015 + debug_only(int saved_sp = _sp);
1.2016 + _sp += nargs;
1.2017 +
1.2018 + // Build address expression. See the code in inline_unsafe_access.
1.2019 + Node *adr;
1.2020 + if (!is_native_ptr) {
1.2021 + // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
1.2022 + Node* offset = pop_pair();
1.2023 + // The base is either a Java object or a value produced by Unsafe.staticFieldBase
1.2024 + Node* base = pop();
1.2025 + // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1.2026 + // to be plain byte offsets, which are also the same as those accepted
1.2027 + // by oopDesc::field_base.
1.2028 + assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1.2029 + "fieldOffset must be byte-scaled");
1.2030 + // 32-bit machines ignore the high half!
1.2031 + offset = ConvL2X(offset);
1.2032 + adr = make_unsafe_address(base, offset);
1.2033 + } else {
1.2034 + Node* ptr = pop_pair();
1.2035 + // Adjust Java long to machine word:
1.2036 + ptr = ConvL2X(ptr);
1.2037 + adr = make_unsafe_address(NULL, ptr);
1.2038 + }
1.2039 +
1.2040 + if (is_static) {
1.2041 + assert(saved_sp == _sp, "must have correct argument count");
1.2042 + } else {
1.2043 + // Pop receiver last: it was pushed first.
1.2044 + Node *receiver = pop();
1.2045 + assert(saved_sp == _sp, "must have correct argument count");
1.2046 +
1.2047 + // Null check on self without removing any arguments. The argument
1.2048 + // null check technically happens in the wrong place, which can lead to
1.2049 + // invalid stack traces when the primitive is inlined into a method
1.2050 + // which handles NullPointerExceptions.
1.2051 + _sp += nargs;
1.2052 + do_null_check(receiver, T_OBJECT);
1.2053 + _sp -= nargs;
1.2054 + if (stopped()) {
1.2055 + return true;
1.2056 + }
1.2057 + }
1.2058 +
1.2059 + // Generate the read or write prefetch
1.2060 + Node *prefetch;
1.2061 + if (is_store) {
1.2062 + prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
1.2063 + } else {
1.2064 + prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
1.2065 + }
1.2066 + prefetch->init_req(0, control());
1.2067 + set_i_o(_gvn.transform(prefetch));
1.2068 +
1.2069 + return true;
1.2070 +}
1.2071 +
1.2072 +//----------------------------inline_unsafe_CAS----------------------------
1.2073 +
1.2074 +bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
1.2075 + // This basic scheme here is the same as inline_unsafe_access, but
1.2076 + // differs in enough details that combining them would make the code
1.2077 + // overly confusing. (This is a true fact! I originally combined
1.2078 + // them, but even I was confused by it!) As much code/comments as
1.2079 + // possible are retained from inline_unsafe_access though to make
1.2080 + // the correspondances clearer. - dl
1.2081 +
1.2082 + if (callee()->is_static()) return false; // caller must have the capability!
1.2083 +
1.2084 +#ifndef PRODUCT
1.2085 + {
1.2086 + ResourceMark rm;
1.2087 + // Check the signatures.
1.2088 + ciSignature* sig = signature();
1.2089 +#ifdef ASSERT
1.2090 + BasicType rtype = sig->return_type()->basic_type();
1.2091 + assert(rtype == T_BOOLEAN, "CAS must return boolean");
1.2092 + assert(sig->count() == 4, "CAS has 4 arguments");
1.2093 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
1.2094 + assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
1.2095 +#endif // ASSERT
1.2096 + }
1.2097 +#endif //PRODUCT
1.2098 +
1.2099 + // number of stack slots per value argument (1 or 2)
1.2100 + int type_words = type2size[type];
1.2101 +
1.2102 + // Cannot inline wide CAS on machines that don't support it natively
1.2103 + if (type2aelembytes[type] > BytesPerInt && !VM_Version::supports_cx8())
1.2104 + return false;
1.2105 +
1.2106 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.2107 +
1.2108 + // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
1.2109 + int nargs = 1 + 1 + 2 + type_words + type_words;
1.2110 +
1.2111 + // pop arguments: newval, oldval, offset, base, and receiver
1.2112 + debug_only(int saved_sp = _sp);
1.2113 + _sp += nargs;
1.2114 + Node* newval = (type_words == 1) ? pop() : pop_pair();
1.2115 + Node* oldval = (type_words == 1) ? pop() : pop_pair();
1.2116 + Node *offset = pop_pair();
1.2117 + Node *base = pop();
1.2118 + Node *receiver = pop();
1.2119 + assert(saved_sp == _sp, "must have correct argument count");
1.2120 +
1.2121 + // Null check receiver.
1.2122 + _sp += nargs;
1.2123 + do_null_check(receiver, T_OBJECT);
1.2124 + _sp -= nargs;
1.2125 + if (stopped()) {
1.2126 + return true;
1.2127 + }
1.2128 +
1.2129 + // Build field offset expression.
1.2130 + // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
1.2131 + // to be plain byte offsets, which are also the same as those accepted
1.2132 + // by oopDesc::field_base.
1.2133 + assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
1.2134 + // 32-bit machines ignore the high half of long offsets
1.2135 + offset = ConvL2X(offset);
1.2136 + Node* adr = make_unsafe_address(base, offset);
1.2137 + const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1.2138 +
1.2139 + // (Unlike inline_unsafe_access, there seems no point in trying
1.2140 + // to refine types. Just use the coarse types here.
1.2141 + const Type *value_type = Type::get_const_basic_type(type);
1.2142 + Compile::AliasType* alias_type = C->alias_type(adr_type);
1.2143 + assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
1.2144 + int alias_idx = C->get_alias_index(adr_type);
1.2145 +
1.2146 + // Memory-model-wise, a CAS acts like a little synchronized block,
1.2147 + // so needs barriers on each side. These don't't translate into
1.2148 + // actual barriers on most machines, but we still need rest of
1.2149 + // compiler to respect ordering.
1.2150 +
1.2151 + insert_mem_bar(Op_MemBarRelease);
1.2152 + insert_mem_bar(Op_MemBarCPUOrder);
1.2153 +
1.2154 + // 4984716: MemBars must be inserted before this
1.2155 + // memory node in order to avoid a false
1.2156 + // dependency which will confuse the scheduler.
1.2157 + Node *mem = memory(alias_idx);
1.2158 +
1.2159 + // For now, we handle only those cases that actually exist: ints,
1.2160 + // longs, and Object. Adding others should be straightforward.
1.2161 + Node* cas;
1.2162 + switch(type) {
1.2163 + case T_INT:
1.2164 + cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
1.2165 + break;
1.2166 + case T_LONG:
1.2167 + cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
1.2168 + break;
1.2169 + case T_OBJECT:
1.2170 + // reference stores need a store barrier.
1.2171 + // (They don't if CAS fails, but it isn't worth checking.)
1.2172 + pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT);
1.2173 + cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
1.2174 + post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
1.2175 + break;
1.2176 + default:
1.2177 + ShouldNotReachHere();
1.2178 + break;
1.2179 + }
1.2180 +
1.2181 + // SCMemProjNodes represent the memory state of CAS. Their main
1.2182 + // role is to prevent CAS nodes from being optimized away when their
1.2183 + // results aren't used.
1.2184 + Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
1.2185 + set_memory(proj, alias_idx);
1.2186 +
1.2187 + // Add the trailing membar surrounding the access
1.2188 + insert_mem_bar(Op_MemBarCPUOrder);
1.2189 + insert_mem_bar(Op_MemBarAcquire);
1.2190 +
1.2191 + push(cas);
1.2192 + return true;
1.2193 +}
1.2194 +
1.2195 +bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
1.2196 + // This is another variant of inline_unsafe_access, differing in
1.2197 + // that it always issues store-store ("release") barrier and ensures
1.2198 + // store-atomicity (which only matters for "long").
1.2199 +
1.2200 + if (callee()->is_static()) return false; // caller must have the capability!
1.2201 +
1.2202 +#ifndef PRODUCT
1.2203 + {
1.2204 + ResourceMark rm;
1.2205 + // Check the signatures.
1.2206 + ciSignature* sig = signature();
1.2207 +#ifdef ASSERT
1.2208 + BasicType rtype = sig->return_type()->basic_type();
1.2209 + assert(rtype == T_VOID, "must return void");
1.2210 + assert(sig->count() == 3, "has 3 arguments");
1.2211 + assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
1.2212 + assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
1.2213 +#endif // ASSERT
1.2214 + }
1.2215 +#endif //PRODUCT
1.2216 +
1.2217 + // number of stack slots per value argument (1 or 2)
1.2218 + int type_words = type2size[type];
1.2219 +
1.2220 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.2221 +
1.2222 + // Argument words: "this" plus oop plus offset plus value;
1.2223 + int nargs = 1 + 1 + 2 + type_words;
1.2224 +
1.2225 + // pop arguments: val, offset, base, and receiver
1.2226 + debug_only(int saved_sp = _sp);
1.2227 + _sp += nargs;
1.2228 + Node* val = (type_words == 1) ? pop() : pop_pair();
1.2229 + Node *offset = pop_pair();
1.2230 + Node *base = pop();
1.2231 + Node *receiver = pop();
1.2232 + assert(saved_sp == _sp, "must have correct argument count");
1.2233 +
1.2234 + // Null check receiver.
1.2235 + _sp += nargs;
1.2236 + do_null_check(receiver, T_OBJECT);
1.2237 + _sp -= nargs;
1.2238 + if (stopped()) {
1.2239 + return true;
1.2240 + }
1.2241 +
1.2242 + // Build field offset expression.
1.2243 + assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
1.2244 + // 32-bit machines ignore the high half of long offsets
1.2245 + offset = ConvL2X(offset);
1.2246 + Node* adr = make_unsafe_address(base, offset);
1.2247 + const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
1.2248 + const Type *value_type = Type::get_const_basic_type(type);
1.2249 + Compile::AliasType* alias_type = C->alias_type(adr_type);
1.2250 +
1.2251 + insert_mem_bar(Op_MemBarRelease);
1.2252 + insert_mem_bar(Op_MemBarCPUOrder);
1.2253 + // Ensure that the store is atomic for longs:
1.2254 + bool require_atomic_access = true;
1.2255 + Node* store;
1.2256 + if (type == T_OBJECT) // reference stores need a store barrier.
1.2257 + store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type);
1.2258 + else {
1.2259 + store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
1.2260 + }
1.2261 + insert_mem_bar(Op_MemBarCPUOrder);
1.2262 + return true;
1.2263 +}
1.2264 +
1.2265 +bool LibraryCallKit::inline_unsafe_allocate() {
1.2266 + if (callee()->is_static()) return false; // caller must have the capability!
1.2267 + int nargs = 1 + 1;
1.2268 + assert(signature()->size() == nargs-1, "alloc has 1 argument");
1.2269 + null_check_receiver(callee()); // check then ignore argument(0)
1.2270 + _sp += nargs; // set original stack for use by uncommon_trap
1.2271 + Node* cls = do_null_check(argument(1), T_OBJECT);
1.2272 + _sp -= nargs;
1.2273 + if (stopped()) return true;
1.2274 +
1.2275 + Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
1.2276 + _sp += nargs; // set original stack for use by uncommon_trap
1.2277 + kls = do_null_check(kls, T_OBJECT);
1.2278 + _sp -= nargs;
1.2279 + if (stopped()) return true; // argument was like int.class
1.2280 +
1.2281 + // Note: The argument might still be an illegal value like
1.2282 + // Serializable.class or Object[].class. The runtime will handle it.
1.2283 + // But we must make an explicit check for initialization.
1.2284 + Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
1.2285 + Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
1.2286 + Node* bits = intcon(instanceKlass::fully_initialized);
1.2287 + Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
1.2288 + // The 'test' is non-zero if we need to take a slow path.
1.2289 +
1.2290 + Node* obj = new_instance(kls, test);
1.2291 + push(obj);
1.2292 +
1.2293 + return true;
1.2294 +}
1.2295 +
1.2296 +//------------------------inline_native_time_funcs--------------
1.2297 +// inline code for System.currentTimeMillis() and System.nanoTime()
1.2298 +// these have the same type and signature
1.2299 +bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
1.2300 + address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
1.2301 + CAST_FROM_FN_PTR(address, os::javaTimeMillis);
1.2302 + const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
1.2303 + const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
1.2304 + const TypePtr* no_memory_effects = NULL;
1.2305 + Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
1.2306 + Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
1.2307 +#ifdef ASSERT
1.2308 + Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
1.2309 + assert(value_top == top(), "second value must be top");
1.2310 +#endif
1.2311 + push_pair(value);
1.2312 + return true;
1.2313 +}
1.2314 +
1.2315 +//------------------------inline_native_currentThread------------------
1.2316 +bool LibraryCallKit::inline_native_currentThread() {
1.2317 + Node* junk = NULL;
1.2318 + push(generate_current_thread(junk));
1.2319 + return true;
1.2320 +}
1.2321 +
1.2322 +//------------------------inline_native_isInterrupted------------------
1.2323 +bool LibraryCallKit::inline_native_isInterrupted() {
1.2324 + const int nargs = 1+1; // receiver + boolean
1.2325 + assert(nargs == arg_size(), "sanity");
1.2326 + // Add a fast path to t.isInterrupted(clear_int):
1.2327 + // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
1.2328 + // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
1.2329 + // So, in the common case that the interrupt bit is false,
1.2330 + // we avoid making a call into the VM. Even if the interrupt bit
1.2331 + // is true, if the clear_int argument is false, we avoid the VM call.
1.2332 + // However, if the receiver is not currentThread, we must call the VM,
1.2333 + // because there must be some locking done around the operation.
1.2334 +
1.2335 + // We only go to the fast case code if we pass two guards.
1.2336 + // Paths which do not pass are accumulated in the slow_region.
1.2337 + RegionNode* slow_region = new (C, 1) RegionNode(1);
1.2338 + record_for_igvn(slow_region);
1.2339 + RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
1.2340 + PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
1.2341 + enum { no_int_result_path = 1,
1.2342 + no_clear_result_path = 2,
1.2343 + slow_result_path = 3
1.2344 + };
1.2345 +
1.2346 + // (a) Receiving thread must be the current thread.
1.2347 + Node* rec_thr = argument(0);
1.2348 + Node* tls_ptr = NULL;
1.2349 + Node* cur_thr = generate_current_thread(tls_ptr);
1.2350 + Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
1.2351 + Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
1.2352 +
1.2353 + bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
1.2354 + if (!known_current_thread)
1.2355 + generate_slow_guard(bol_thr, slow_region);
1.2356 +
1.2357 + // (b) Interrupt bit on TLS must be false.
1.2358 + Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
1.2359 + Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
1.2360 + p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
1.2361 + Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT);
1.2362 + Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
1.2363 + Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
1.2364 +
1.2365 + IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1.2366 +
1.2367 + // First fast path: if (!TLS._interrupted) return false;
1.2368 + Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
1.2369 + result_rgn->init_req(no_int_result_path, false_bit);
1.2370 + result_val->init_req(no_int_result_path, intcon(0));
1.2371 +
1.2372 + // drop through to next case
1.2373 + set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
1.2374 +
1.2375 + // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
1.2376 + Node* clr_arg = argument(1);
1.2377 + Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
1.2378 + Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
1.2379 + IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
1.2380 +
1.2381 + // Second fast path: ... else if (!clear_int) return true;
1.2382 + Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
1.2383 + result_rgn->init_req(no_clear_result_path, false_arg);
1.2384 + result_val->init_req(no_clear_result_path, intcon(1));
1.2385 +
1.2386 + // drop through to next case
1.2387 + set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
1.2388 +
1.2389 + // (d) Otherwise, go to the slow path.
1.2390 + slow_region->add_req(control());
1.2391 + set_control( _gvn.transform(slow_region) );
1.2392 +
1.2393 + if (stopped()) {
1.2394 + // There is no slow path.
1.2395 + result_rgn->init_req(slow_result_path, top());
1.2396 + result_val->init_req(slow_result_path, top());
1.2397 + } else {
1.2398 + // non-virtual because it is a private non-static
1.2399 + CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
1.2400 +
1.2401 + Node* slow_val = set_results_for_java_call(slow_call);
1.2402 + // this->control() comes from set_results_for_java_call
1.2403 +
1.2404 + // If we know that the result of the slow call will be true, tell the optimizer!
1.2405 + if (known_current_thread) slow_val = intcon(1);
1.2406 +
1.2407 + Node* fast_io = slow_call->in(TypeFunc::I_O);
1.2408 + Node* fast_mem = slow_call->in(TypeFunc::Memory);
1.2409 + // These two phis are pre-filled with copies of of the fast IO and Memory
1.2410 + Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
1.2411 + Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
1.2412 +
1.2413 + result_rgn->init_req(slow_result_path, control());
1.2414 + io_phi ->init_req(slow_result_path, i_o());
1.2415 + mem_phi ->init_req(slow_result_path, reset_memory());
1.2416 + result_val->init_req(slow_result_path, slow_val);
1.2417 +
1.2418 + set_all_memory( _gvn.transform(mem_phi) );
1.2419 + set_i_o( _gvn.transform(io_phi) );
1.2420 + }
1.2421 +
1.2422 + push_result(result_rgn, result_val);
1.2423 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.2424 +
1.2425 + return true;
1.2426 +}
1.2427 +
1.2428 +//---------------------------load_mirror_from_klass----------------------------
1.2429 +// Given a klass oop, load its java mirror (a java.lang.Class oop).
1.2430 +Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
1.2431 + Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
1.2432 + return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
1.2433 +}
1.2434 +
1.2435 +//-----------------------load_klass_from_mirror_common-------------------------
1.2436 +// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
1.2437 +// Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
1.2438 +// and branch to the given path on the region.
1.2439 +// If never_see_null, take an uncommon trap on null, so we can optimistically
1.2440 +// compile for the non-null case.
1.2441 +// If the region is NULL, force never_see_null = true.
1.2442 +Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
1.2443 + bool never_see_null,
1.2444 + int nargs,
1.2445 + RegionNode* region,
1.2446 + int null_path,
1.2447 + int offset) {
1.2448 + if (region == NULL) never_see_null = true;
1.2449 + Node* p = basic_plus_adr(mirror, offset);
1.2450 + const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
1.2451 + Node* kls = _gvn.transform(new (C, 3) LoadKlassNode(0, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
1.2452 + _sp += nargs; // any deopt will start just before call to enclosing method
1.2453 + Node* null_ctl = top();
1.2454 + kls = null_check_oop(kls, &null_ctl, never_see_null);
1.2455 + if (region != NULL) {
1.2456 + // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
1.2457 + region->init_req(null_path, null_ctl);
1.2458 + } else {
1.2459 + assert(null_ctl == top(), "no loose ends");
1.2460 + }
1.2461 + _sp -= nargs;
1.2462 + return kls;
1.2463 +}
1.2464 +
1.2465 +//--------------------(inline_native_Class_query helpers)---------------------
1.2466 +// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
1.2467 +// Fall through if (mods & mask) == bits, take the guard otherwise.
1.2468 +Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
1.2469 + // Branch around if the given klass has the given modifier bit set.
1.2470 + // Like generate_guard, adds a new path onto the region.
1.2471 + Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
1.2472 + Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
1.2473 + Node* mask = intcon(modifier_mask);
1.2474 + Node* bits = intcon(modifier_bits);
1.2475 + Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
1.2476 + Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
1.2477 + Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
1.2478 + return generate_fair_guard(bol, region);
1.2479 +}
1.2480 +Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
1.2481 + return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
1.2482 +}
1.2483 +
1.2484 +//-------------------------inline_native_Class_query-------------------
1.2485 +bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
1.2486 + int nargs = 1+0; // just the Class mirror, in most cases
1.2487 + const Type* return_type = TypeInt::BOOL;
1.2488 + Node* prim_return_value = top(); // what happens if it's a primitive class?
1.2489 + bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
1.2490 + bool expect_prim = false; // most of these guys expect to work on refs
1.2491 +
1.2492 + enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
1.2493 +
1.2494 + switch (id) {
1.2495 + case vmIntrinsics::_isInstance:
1.2496 + nargs = 1+1; // the Class mirror, plus the object getting queried about
1.2497 + // nothing is an instance of a primitive type
1.2498 + prim_return_value = intcon(0);
1.2499 + break;
1.2500 + case vmIntrinsics::_getModifiers:
1.2501 + prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
1.2502 + assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
1.2503 + return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
1.2504 + break;
1.2505 + case vmIntrinsics::_isInterface:
1.2506 + prim_return_value = intcon(0);
1.2507 + break;
1.2508 + case vmIntrinsics::_isArray:
1.2509 + prim_return_value = intcon(0);
1.2510 + expect_prim = true; // cf. ObjectStreamClass.getClassSignature
1.2511 + break;
1.2512 + case vmIntrinsics::_isPrimitive:
1.2513 + prim_return_value = intcon(1);
1.2514 + expect_prim = true; // obviously
1.2515 + break;
1.2516 + case vmIntrinsics::_getSuperclass:
1.2517 + prim_return_value = null();
1.2518 + return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
1.2519 + break;
1.2520 + case vmIntrinsics::_getComponentType:
1.2521 + prim_return_value = null();
1.2522 + return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
1.2523 + break;
1.2524 + case vmIntrinsics::_getClassAccessFlags:
1.2525 + prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
1.2526 + return_type = TypeInt::INT; // not bool! 6297094
1.2527 + break;
1.2528 + default:
1.2529 + ShouldNotReachHere();
1.2530 + }
1.2531 +
1.2532 + Node* mirror = argument(0);
1.2533 + Node* obj = (nargs <= 1)? top(): argument(1);
1.2534 +
1.2535 + const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
1.2536 + if (mirror_con == NULL) return false; // cannot happen?
1.2537 +
1.2538 +#ifndef PRODUCT
1.2539 + if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
1.2540 + ciType* k = mirror_con->java_mirror_type();
1.2541 + if (k) {
1.2542 + tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
1.2543 + k->print_name();
1.2544 + tty->cr();
1.2545 + }
1.2546 + }
1.2547 +#endif
1.2548 +
1.2549 + // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
1.2550 + RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
1.2551 + record_for_igvn(region);
1.2552 + PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
1.2553 +
1.2554 + // The mirror will never be null of Reflection.getClassAccessFlags, however
1.2555 + // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
1.2556 + // if it is. See bug 4774291.
1.2557 +
1.2558 + // For Reflection.getClassAccessFlags(), the null check occurs in
1.2559 + // the wrong place; see inline_unsafe_access(), above, for a similar
1.2560 + // situation.
1.2561 + _sp += nargs; // set original stack for use by uncommon_trap
1.2562 + mirror = do_null_check(mirror, T_OBJECT);
1.2563 + _sp -= nargs;
1.2564 + // If mirror or obj is dead, only null-path is taken.
1.2565 + if (stopped()) return true;
1.2566 +
1.2567 + if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
1.2568 +
1.2569 + // Now load the mirror's klass metaobject, and null-check it.
1.2570 + // Side-effects region with the control path if the klass is null.
1.2571 + Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
1.2572 + region, _prim_path);
1.2573 + // If kls is null, we have a primitive mirror.
1.2574 + phi->init_req(_prim_path, prim_return_value);
1.2575 + if (stopped()) { push_result(region, phi); return true; }
1.2576 +
1.2577 + Node* p; // handy temp
1.2578 + Node* null_ctl;
1.2579 +
1.2580 + // Now that we have the non-null klass, we can perform the real query.
1.2581 + // For constant classes, the query will constant-fold in LoadNode::Value.
1.2582 + Node* query_value = top();
1.2583 + switch (id) {
1.2584 + case vmIntrinsics::_isInstance:
1.2585 + // nothing is an instance of a primitive type
1.2586 + query_value = gen_instanceof(obj, kls);
1.2587 + break;
1.2588 +
1.2589 + case vmIntrinsics::_getModifiers:
1.2590 + p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
1.2591 + query_value = make_load(NULL, p, TypeInt::INT, T_INT);
1.2592 + break;
1.2593 +
1.2594 + case vmIntrinsics::_isInterface:
1.2595 + // (To verify this code sequence, check the asserts in JVM_IsInterface.)
1.2596 + if (generate_interface_guard(kls, region) != NULL)
1.2597 + // A guard was added. If the guard is taken, it was an interface.
1.2598 + phi->add_req(intcon(1));
1.2599 + // If we fall through, it's a plain class.
1.2600 + query_value = intcon(0);
1.2601 + break;
1.2602 +
1.2603 + case vmIntrinsics::_isArray:
1.2604 + // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
1.2605 + if (generate_array_guard(kls, region) != NULL)
1.2606 + // A guard was added. If the guard is taken, it was an array.
1.2607 + phi->add_req(intcon(1));
1.2608 + // If we fall through, it's a plain class.
1.2609 + query_value = intcon(0);
1.2610 + break;
1.2611 +
1.2612 + case vmIntrinsics::_isPrimitive:
1.2613 + query_value = intcon(0); // "normal" path produces false
1.2614 + break;
1.2615 +
1.2616 + case vmIntrinsics::_getSuperclass:
1.2617 + // The rules here are somewhat unfortunate, but we can still do better
1.2618 + // with random logic than with a JNI call.
1.2619 + // Interfaces store null or Object as _super, but must report null.
1.2620 + // Arrays store an intermediate super as _super, but must report Object.
1.2621 + // Other types can report the actual _super.
1.2622 + // (To verify this code sequence, check the asserts in JVM_IsInterface.)
1.2623 + if (generate_interface_guard(kls, region) != NULL)
1.2624 + // A guard was added. If the guard is taken, it was an interface.
1.2625 + phi->add_req(null());
1.2626 + if (generate_array_guard(kls, region) != NULL)
1.2627 + // A guard was added. If the guard is taken, it was an array.
1.2628 + phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
1.2629 + // If we fall through, it's a plain class. Get its _super.
1.2630 + p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
1.2631 + kls = _gvn.transform(new (C, 3) LoadKlassNode(0, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
1.2632 + null_ctl = top();
1.2633 + kls = null_check_oop(kls, &null_ctl);
1.2634 + if (null_ctl != top()) {
1.2635 + // If the guard is taken, Object.superClass is null (both klass and mirror).
1.2636 + region->add_req(null_ctl);
1.2637 + phi ->add_req(null());
1.2638 + }
1.2639 + if (!stopped()) {
1.2640 + query_value = load_mirror_from_klass(kls);
1.2641 + }
1.2642 + break;
1.2643 +
1.2644 + case vmIntrinsics::_getComponentType:
1.2645 + if (generate_array_guard(kls, region) != NULL) {
1.2646 + // Be sure to pin the oop load to the guard edge just created:
1.2647 + Node* is_array_ctrl = region->in(region->req()-1);
1.2648 + Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
1.2649 + Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
1.2650 + phi->add_req(cmo);
1.2651 + }
1.2652 + query_value = null(); // non-array case is null
1.2653 + break;
1.2654 +
1.2655 + case vmIntrinsics::_getClassAccessFlags:
1.2656 + p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
1.2657 + query_value = make_load(NULL, p, TypeInt::INT, T_INT);
1.2658 + break;
1.2659 +
1.2660 + default:
1.2661 + ShouldNotReachHere();
1.2662 + }
1.2663 +
1.2664 + // Fall-through is the normal case of a query to a real class.
1.2665 + phi->init_req(1, query_value);
1.2666 + region->init_req(1, control());
1.2667 +
1.2668 + push_result(region, phi);
1.2669 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.2670 +
1.2671 + return true;
1.2672 +}
1.2673 +
1.2674 +//--------------------------inline_native_subtype_check------------------------
1.2675 +// This intrinsic takes the JNI calls out of the heart of
1.2676 +// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
1.2677 +bool LibraryCallKit::inline_native_subtype_check() {
1.2678 + int nargs = 1+1; // the Class mirror, plus the other class getting examined
1.2679 +
1.2680 + // Pull both arguments off the stack.
1.2681 + Node* args[2]; // two java.lang.Class mirrors: superc, subc
1.2682 + args[0] = argument(0);
1.2683 + args[1] = argument(1);
1.2684 + Node* klasses[2]; // corresponding Klasses: superk, subk
1.2685 + klasses[0] = klasses[1] = top();
1.2686 +
1.2687 + enum {
1.2688 + // A full decision tree on {superc is prim, subc is prim}:
1.2689 + _prim_0_path = 1, // {P,N} => false
1.2690 + // {P,P} & superc!=subc => false
1.2691 + _prim_same_path, // {P,P} & superc==subc => true
1.2692 + _prim_1_path, // {N,P} => false
1.2693 + _ref_subtype_path, // {N,N} & subtype check wins => true
1.2694 + _both_ref_path, // {N,N} & subtype check loses => false
1.2695 + PATH_LIMIT
1.2696 + };
1.2697 +
1.2698 + RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
1.2699 + Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
1.2700 + record_for_igvn(region);
1.2701 +
1.2702 + const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
1.2703 + const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
1.2704 + int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
1.2705 +
1.2706 + // First null-check both mirrors and load each mirror's klass metaobject.
1.2707 + int which_arg;
1.2708 + for (which_arg = 0; which_arg <= 1; which_arg++) {
1.2709 + Node* arg = args[which_arg];
1.2710 + _sp += nargs; // set original stack for use by uncommon_trap
1.2711 + arg = do_null_check(arg, T_OBJECT);
1.2712 + _sp -= nargs;
1.2713 + if (stopped()) break;
1.2714 + args[which_arg] = _gvn.transform(arg);
1.2715 +
1.2716 + Node* p = basic_plus_adr(arg, class_klass_offset);
1.2717 + Node* kls = new (C, 3) LoadKlassNode(0, immutable_memory(), p, adr_type, kls_type);
1.2718 + klasses[which_arg] = _gvn.transform(kls);
1.2719 + }
1.2720 +
1.2721 + // Having loaded both klasses, test each for null.
1.2722 + bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
1.2723 + for (which_arg = 0; which_arg <= 1; which_arg++) {
1.2724 + Node* kls = klasses[which_arg];
1.2725 + Node* null_ctl = top();
1.2726 + _sp += nargs; // set original stack for use by uncommon_trap
1.2727 + kls = null_check_oop(kls, &null_ctl, never_see_null);
1.2728 + _sp -= nargs;
1.2729 + int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
1.2730 + region->init_req(prim_path, null_ctl);
1.2731 + if (stopped()) break;
1.2732 + klasses[which_arg] = kls;
1.2733 + }
1.2734 +
1.2735 + if (!stopped()) {
1.2736 + // now we have two reference types, in klasses[0..1]
1.2737 + Node* subk = klasses[1]; // the argument to isAssignableFrom
1.2738 + Node* superk = klasses[0]; // the receiver
1.2739 + region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
1.2740 + // now we have a successful reference subtype check
1.2741 + region->set_req(_ref_subtype_path, control());
1.2742 + }
1.2743 +
1.2744 + // If both operands are primitive (both klasses null), then
1.2745 + // we must return true when they are identical primitives.
1.2746 + // It is convenient to test this after the first null klass check.
1.2747 + set_control(region->in(_prim_0_path)); // go back to first null check
1.2748 + if (!stopped()) {
1.2749 + // Since superc is primitive, make a guard for the superc==subc case.
1.2750 + Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
1.2751 + Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
1.2752 + generate_guard(bol_eq, region, PROB_FAIR);
1.2753 + if (region->req() == PATH_LIMIT+1) {
1.2754 + // A guard was added. If the added guard is taken, superc==subc.
1.2755 + region->swap_edges(PATH_LIMIT, _prim_same_path);
1.2756 + region->del_req(PATH_LIMIT);
1.2757 + }
1.2758 + region->set_req(_prim_0_path, control()); // Not equal after all.
1.2759 + }
1.2760 +
1.2761 + // these are the only paths that produce 'true':
1.2762 + phi->set_req(_prim_same_path, intcon(1));
1.2763 + phi->set_req(_ref_subtype_path, intcon(1));
1.2764 +
1.2765 + // pull together the cases:
1.2766 + assert(region->req() == PATH_LIMIT, "sane region");
1.2767 + for (uint i = 1; i < region->req(); i++) {
1.2768 + Node* ctl = region->in(i);
1.2769 + if (ctl == NULL || ctl == top()) {
1.2770 + region->set_req(i, top());
1.2771 + phi ->set_req(i, top());
1.2772 + } else if (phi->in(i) == NULL) {
1.2773 + phi->set_req(i, intcon(0)); // all other paths produce 'false'
1.2774 + }
1.2775 + }
1.2776 +
1.2777 + set_control(_gvn.transform(region));
1.2778 + push(_gvn.transform(phi));
1.2779 +
1.2780 + return true;
1.2781 +}
1.2782 +
1.2783 +//---------------------generate_array_guard_common------------------------
1.2784 +Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
1.2785 + bool obj_array, bool not_array) {
1.2786 + // If obj_array/non_array==false/false:
1.2787 + // Branch around if the given klass is in fact an array (either obj or prim).
1.2788 + // If obj_array/non_array==false/true:
1.2789 + // Branch around if the given klass is not an array klass of any kind.
1.2790 + // If obj_array/non_array==true/true:
1.2791 + // Branch around if the kls is not an oop array (kls is int[], String, etc.)
1.2792 + // If obj_array/non_array==true/false:
1.2793 + // Branch around if the kls is an oop array (Object[] or subtype)
1.2794 + //
1.2795 + // Like generate_guard, adds a new path onto the region.
1.2796 + jint layout_con = 0;
1.2797 + Node* layout_val = get_layout_helper(kls, layout_con);
1.2798 + if (layout_val == NULL) {
1.2799 + bool query = (obj_array
1.2800 + ? Klass::layout_helper_is_objArray(layout_con)
1.2801 + : Klass::layout_helper_is_javaArray(layout_con));
1.2802 + if (query == not_array) {
1.2803 + return NULL; // never a branch
1.2804 + } else { // always a branch
1.2805 + Node* always_branch = control();
1.2806 + if (region != NULL)
1.2807 + region->add_req(always_branch);
1.2808 + set_control(top());
1.2809 + return always_branch;
1.2810 + }
1.2811 + }
1.2812 + // Now test the correct condition.
1.2813 + jint nval = (obj_array
1.2814 + ? ((jint)Klass::_lh_array_tag_type_value
1.2815 + << Klass::_lh_array_tag_shift)
1.2816 + : Klass::_lh_neutral_value);
1.2817 + Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
1.2818 + BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
1.2819 + // invert the test if we are looking for a non-array
1.2820 + if (not_array) btest = BoolTest(btest).negate();
1.2821 + Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
1.2822 + return generate_fair_guard(bol, region);
1.2823 +}
1.2824 +
1.2825 +
1.2826 +//-----------------------inline_native_newArray--------------------------
1.2827 +bool LibraryCallKit::inline_native_newArray() {
1.2828 + int nargs = 2;
1.2829 + Node* mirror = argument(0);
1.2830 + Node* count_val = argument(1);
1.2831 +
1.2832 + _sp += nargs; // set original stack for use by uncommon_trap
1.2833 + mirror = do_null_check(mirror, T_OBJECT);
1.2834 + _sp -= nargs;
1.2835 +
1.2836 + enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
1.2837 + RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
1.2838 + PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
1.2839 + TypeInstPtr::NOTNULL);
1.2840 + PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
1.2841 + PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
1.2842 + TypePtr::BOTTOM);
1.2843 +
1.2844 + bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
1.2845 + Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
1.2846 + nargs,
1.2847 + result_reg, _slow_path);
1.2848 + Node* normal_ctl = control();
1.2849 + Node* no_array_ctl = result_reg->in(_slow_path);
1.2850 +
1.2851 + // Generate code for the slow case. We make a call to newArray().
1.2852 + set_control(no_array_ctl);
1.2853 + if (!stopped()) {
1.2854 + // Either the input type is void.class, or else the
1.2855 + // array klass has not yet been cached. Either the
1.2856 + // ensuing call will throw an exception, or else it
1.2857 + // will cache the array klass for next time.
1.2858 + PreserveJVMState pjvms(this);
1.2859 + CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
1.2860 + Node* slow_result = set_results_for_java_call(slow_call);
1.2861 + // this->control() comes from set_results_for_java_call
1.2862 + result_reg->set_req(_slow_path, control());
1.2863 + result_val->set_req(_slow_path, slow_result);
1.2864 + result_io ->set_req(_slow_path, i_o());
1.2865 + result_mem->set_req(_slow_path, reset_memory());
1.2866 + }
1.2867 +
1.2868 + set_control(normal_ctl);
1.2869 + if (!stopped()) {
1.2870 + // Normal case: The array type has been cached in the java.lang.Class.
1.2871 + // The following call works fine even if the array type is polymorphic.
1.2872 + // It could be a dynamic mix of int[], boolean[], Object[], etc.
1.2873 + _sp += nargs; // set original stack for use by uncommon_trap
1.2874 + Node* obj = new_array(klass_node, count_val);
1.2875 + _sp -= nargs;
1.2876 + result_reg->init_req(_normal_path, control());
1.2877 + result_val->init_req(_normal_path, obj);
1.2878 + result_io ->init_req(_normal_path, i_o());
1.2879 + result_mem->init_req(_normal_path, reset_memory());
1.2880 + }
1.2881 +
1.2882 + // Return the combined state.
1.2883 + set_i_o( _gvn.transform(result_io) );
1.2884 + set_all_memory( _gvn.transform(result_mem) );
1.2885 + push_result(result_reg, result_val);
1.2886 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.2887 +
1.2888 + return true;
1.2889 +}
1.2890 +
1.2891 +//----------------------inline_native_getLength--------------------------
1.2892 +bool LibraryCallKit::inline_native_getLength() {
1.2893 + if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1.2894 +
1.2895 + int nargs = 1;
1.2896 + Node* array = argument(0);
1.2897 +
1.2898 + _sp += nargs; // set original stack for use by uncommon_trap
1.2899 + array = do_null_check(array, T_OBJECT);
1.2900 + _sp -= nargs;
1.2901 +
1.2902 + // If array is dead, only null-path is taken.
1.2903 + if (stopped()) return true;
1.2904 +
1.2905 + // Deoptimize if it is a non-array.
1.2906 + Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
1.2907 +
1.2908 + if (non_array != NULL) {
1.2909 + PreserveJVMState pjvms(this);
1.2910 + set_control(non_array);
1.2911 + _sp += nargs; // push the arguments back on the stack
1.2912 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.2913 + Deoptimization::Action_maybe_recompile);
1.2914 + }
1.2915 +
1.2916 + // If control is dead, only non-array-path is taken.
1.2917 + if (stopped()) return true;
1.2918 +
1.2919 + // The works fine even if the array type is polymorphic.
1.2920 + // It could be a dynamic mix of int[], boolean[], Object[], etc.
1.2921 + push( load_array_length(array) );
1.2922 +
1.2923 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.2924 +
1.2925 + return true;
1.2926 +}
1.2927 +
1.2928 +//------------------------inline_array_copyOf----------------------------
1.2929 +bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
1.2930 + if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1.2931 +
1.2932 + // Restore the stack and pop off the arguments.
1.2933 + int nargs = 3 + (is_copyOfRange? 1: 0);
1.2934 + Node* original = argument(0);
1.2935 + Node* start = is_copyOfRange? argument(1): intcon(0);
1.2936 + Node* end = is_copyOfRange? argument(2): argument(1);
1.2937 + Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
1.2938 +
1.2939 + _sp += nargs; // set original stack for use by uncommon_trap
1.2940 + array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
1.2941 + original = do_null_check(original, T_OBJECT);
1.2942 + _sp -= nargs;
1.2943 +
1.2944 + // Check if a null path was taken unconditionally.
1.2945 + if (stopped()) return true;
1.2946 +
1.2947 + Node* orig_length = load_array_length(original);
1.2948 +
1.2949 + Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs,
1.2950 + NULL, 0);
1.2951 + _sp += nargs; // set original stack for use by uncommon_trap
1.2952 + klass_node = do_null_check(klass_node, T_OBJECT);
1.2953 + _sp -= nargs;
1.2954 +
1.2955 + RegionNode* bailout = new (C, 1) RegionNode(1);
1.2956 + record_for_igvn(bailout);
1.2957 +
1.2958 + // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
1.2959 + // Bail out if that is so.
1.2960 + Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
1.2961 + if (not_objArray != NULL) {
1.2962 + // Improve the klass node's type from the new optimistic assumption:
1.2963 + ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
1.2964 + const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1.2965 + Node* cast = new (C, 2) CastPPNode(klass_node, akls);
1.2966 + cast->init_req(0, control());
1.2967 + klass_node = _gvn.transform(cast);
1.2968 + }
1.2969 +
1.2970 + // Bail out if either start or end is negative.
1.2971 + generate_negative_guard(start, bailout, &start);
1.2972 + generate_negative_guard(end, bailout, &end);
1.2973 +
1.2974 + Node* length = end;
1.2975 + if (_gvn.type(start) != TypeInt::ZERO) {
1.2976 + length = _gvn.transform( new (C, 3) SubINode(end, start) );
1.2977 + }
1.2978 +
1.2979 + // Bail out if length is negative.
1.2980 + // ...Not needed, since the new_array will throw the right exception.
1.2981 + //generate_negative_guard(length, bailout, &length);
1.2982 +
1.2983 + if (bailout->req() > 1) {
1.2984 + PreserveJVMState pjvms(this);
1.2985 + set_control( _gvn.transform(bailout) );
1.2986 + _sp += nargs; // push the arguments back on the stack
1.2987 + uncommon_trap(Deoptimization::Reason_intrinsic,
1.2988 + Deoptimization::Action_maybe_recompile);
1.2989 + }
1.2990 +
1.2991 + if (!stopped()) {
1.2992 + // How many elements will we copy from the original?
1.2993 + // The answer is MinI(orig_length - start, length).
1.2994 + Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
1.2995 + Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
1.2996 +
1.2997 + _sp += nargs; // set original stack for use by uncommon_trap
1.2998 + Node* newcopy = new_array(klass_node, length);
1.2999 + _sp -= nargs;
1.3000 +
1.3001 + // Generate a direct call to the right arraycopy function(s).
1.3002 + // We know the copy is disjoint but we might not know if the
1.3003 + // oop stores need checking.
1.3004 + // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
1.3005 + // This will fail a store-check if x contains any non-nulls.
1.3006 + bool disjoint_bases = true;
1.3007 + bool length_never_negative = true;
1.3008 + generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
1.3009 + original, start, newcopy, intcon(0), moved,
1.3010 + nargs, disjoint_bases, length_never_negative);
1.3011 +
1.3012 + push(newcopy);
1.3013 + }
1.3014 +
1.3015 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3016 +
1.3017 + return true;
1.3018 +}
1.3019 +
1.3020 +
1.3021 +//----------------------generate_virtual_guard---------------------------
1.3022 +// Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
1.3023 +Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
1.3024 + RegionNode* slow_region) {
1.3025 + ciMethod* method = callee();
1.3026 + int vtable_index = method->vtable_index();
1.3027 + // Get the methodOop out of the appropriate vtable entry.
1.3028 + int entry_offset = (instanceKlass::vtable_start_offset() +
1.3029 + vtable_index*vtableEntry::size()) * wordSize +
1.3030 + vtableEntry::method_offset_in_bytes();
1.3031 + Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
1.3032 + Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
1.3033 +
1.3034 + // Compare the target method with the expected method (e.g., Object.hashCode).
1.3035 + const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
1.3036 +
1.3037 + Node* native_call = makecon(native_call_addr);
1.3038 + Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
1.3039 + Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
1.3040 +
1.3041 + return generate_slow_guard(test_native, slow_region);
1.3042 +}
1.3043 +
1.3044 +//-----------------------generate_method_call----------------------------
1.3045 +// Use generate_method_call to make a slow-call to the real
1.3046 +// method if the fast path fails. An alternative would be to
1.3047 +// use a stub like OptoRuntime::slow_arraycopy_Java.
1.3048 +// This only works for expanding the current library call,
1.3049 +// not another intrinsic. (E.g., don't use this for making an
1.3050 +// arraycopy call inside of the copyOf intrinsic.)
1.3051 +CallJavaNode*
1.3052 +LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
1.3053 + // When compiling the intrinsic method itself, do not use this technique.
1.3054 + guarantee(callee() != C->method(), "cannot make slow-call to self");
1.3055 +
1.3056 + ciMethod* method = callee();
1.3057 + // ensure the JVMS we have will be correct for this call
1.3058 + guarantee(method_id == method->intrinsic_id(), "must match");
1.3059 +
1.3060 + const TypeFunc* tf = TypeFunc::make(method);
1.3061 + int tfdc = tf->domain()->cnt();
1.3062 + CallJavaNode* slow_call;
1.3063 + if (is_static) {
1.3064 + assert(!is_virtual, "");
1.3065 + slow_call = new(C, tfdc) CallStaticJavaNode(tf,
1.3066 + SharedRuntime::get_resolve_static_call_stub(),
1.3067 + method, bci());
1.3068 + } else if (is_virtual) {
1.3069 + null_check_receiver(method);
1.3070 + int vtable_index = methodOopDesc::invalid_vtable_index;
1.3071 + if (UseInlineCaches) {
1.3072 + // Suppress the vtable call
1.3073 + } else {
1.3074 + // hashCode and clone are not a miranda methods,
1.3075 + // so the vtable index is fixed.
1.3076 + // No need to use the linkResolver to get it.
1.3077 + vtable_index = method->vtable_index();
1.3078 + }
1.3079 + slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
1.3080 + SharedRuntime::get_resolve_virtual_call_stub(),
1.3081 + method, vtable_index, bci());
1.3082 + } else { // neither virtual nor static: opt_virtual
1.3083 + null_check_receiver(method);
1.3084 + slow_call = new(C, tfdc) CallStaticJavaNode(tf,
1.3085 + SharedRuntime::get_resolve_opt_virtual_call_stub(),
1.3086 + method, bci());
1.3087 + slow_call->set_optimized_virtual(true);
1.3088 + }
1.3089 + set_arguments_for_java_call(slow_call);
1.3090 + set_edges_for_java_call(slow_call);
1.3091 + return slow_call;
1.3092 +}
1.3093 +
1.3094 +
1.3095 +//------------------------------inline_native_hashcode--------------------
1.3096 +// Build special case code for calls to hashCode on an object.
1.3097 +bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
1.3098 + assert(is_static == callee()->is_static(), "correct intrinsic selection");
1.3099 + assert(!(is_virtual && is_static), "either virtual, special, or static");
1.3100 +
1.3101 + enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
1.3102 +
1.3103 + RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
1.3104 + PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
1.3105 + TypeInt::INT);
1.3106 + PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
1.3107 + PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
1.3108 + TypePtr::BOTTOM);
1.3109 + Node* obj = NULL;
1.3110 + if (!is_static) {
1.3111 + // Check for hashing null object
1.3112 + obj = null_check_receiver(callee());
1.3113 + if (stopped()) return true; // unconditionally null
1.3114 + result_reg->init_req(_null_path, top());
1.3115 + result_val->init_req(_null_path, top());
1.3116 + } else {
1.3117 + // Do a null check, and return zero if null.
1.3118 + // System.identityHashCode(null) == 0
1.3119 + obj = argument(0);
1.3120 + Node* null_ctl = top();
1.3121 + obj = null_check_oop(obj, &null_ctl);
1.3122 + result_reg->init_req(_null_path, null_ctl);
1.3123 + result_val->init_req(_null_path, _gvn.intcon(0));
1.3124 + }
1.3125 +
1.3126 + // Unconditionally null? Then return right away.
1.3127 + if (stopped()) {
1.3128 + set_control( result_reg->in(_null_path) );
1.3129 + if (!stopped())
1.3130 + push( result_val ->in(_null_path) );
1.3131 + return true;
1.3132 + }
1.3133 +
1.3134 + // After null check, get the object's klass.
1.3135 + Node* obj_klass = load_object_klass(obj);
1.3136 +
1.3137 + // This call may be virtual (invokevirtual) or bound (invokespecial).
1.3138 + // For each case we generate slightly different code.
1.3139 +
1.3140 + // We only go to the fast case code if we pass a number of guards. The
1.3141 + // paths which do not pass are accumulated in the slow_region.
1.3142 + RegionNode* slow_region = new (C, 1) RegionNode(1);
1.3143 + record_for_igvn(slow_region);
1.3144 +
1.3145 + // If this is a virtual call, we generate a funny guard. We pull out
1.3146 + // the vtable entry corresponding to hashCode() from the target object.
1.3147 + // If the target method which we are calling happens to be the native
1.3148 + // Object hashCode() method, we pass the guard. We do not need this
1.3149 + // guard for non-virtual calls -- the caller is known to be the native
1.3150 + // Object hashCode().
1.3151 + if (is_virtual) {
1.3152 + generate_virtual_guard(obj_klass, slow_region);
1.3153 + }
1.3154 +
1.3155 + // Get the header out of the object, use LoadMarkNode when available
1.3156 + Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
1.3157 + Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
1.3158 + header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
1.3159 +
1.3160 + // Test the header to see if it is unlocked.
1.3161 + Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
1.3162 + Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
1.3163 + Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
1.3164 + Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
1.3165 + Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
1.3166 +
1.3167 + generate_slow_guard(test_unlocked, slow_region);
1.3168 +
1.3169 + // Get the hash value and check to see that it has been properly assigned.
1.3170 + // We depend on hash_mask being at most 32 bits and avoid the use of
1.3171 + // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
1.3172 + // vm: see markOop.hpp.
1.3173 + Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
1.3174 + Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
1.3175 + Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
1.3176 + // This hack lets the hash bits live anywhere in the mark object now, as long
1.3177 + // as the shift drops the relevent bits into the low 32 bits. Note that
1.3178 + // Java spec says that HashCode is an int so there's no point in capturing
1.3179 + // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
1.3180 + hshifted_header = ConvX2I(hshifted_header);
1.3181 + Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
1.3182 +
1.3183 + Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
1.3184 + Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
1.3185 + Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
1.3186 +
1.3187 + generate_slow_guard(test_assigned, slow_region);
1.3188 +
1.3189 + Node* init_mem = reset_memory();
1.3190 + // fill in the rest of the null path:
1.3191 + result_io ->init_req(_null_path, i_o());
1.3192 + result_mem->init_req(_null_path, init_mem);
1.3193 +
1.3194 + result_val->init_req(_fast_path, hash_val);
1.3195 + result_reg->init_req(_fast_path, control());
1.3196 + result_io ->init_req(_fast_path, i_o());
1.3197 + result_mem->init_req(_fast_path, init_mem);
1.3198 +
1.3199 + // Generate code for the slow case. We make a call to hashCode().
1.3200 + set_control(_gvn.transform(slow_region));
1.3201 + if (!stopped()) {
1.3202 + // No need for PreserveJVMState, because we're using up the present state.
1.3203 + set_all_memory(init_mem);
1.3204 + vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
1.3205 + if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
1.3206 + CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
1.3207 + Node* slow_result = set_results_for_java_call(slow_call);
1.3208 + // this->control() comes from set_results_for_java_call
1.3209 + result_reg->init_req(_slow_path, control());
1.3210 + result_val->init_req(_slow_path, slow_result);
1.3211 + result_io ->set_req(_slow_path, i_o());
1.3212 + result_mem ->set_req(_slow_path, reset_memory());
1.3213 + }
1.3214 +
1.3215 + // Return the combined state.
1.3216 + set_i_o( _gvn.transform(result_io) );
1.3217 + set_all_memory( _gvn.transform(result_mem) );
1.3218 + push_result(result_reg, result_val);
1.3219 +
1.3220 + return true;
1.3221 +}
1.3222 +
1.3223 +//---------------------------inline_native_getClass----------------------------
1.3224 +// Build special case code for calls to hashCode on an object.
1.3225 +bool LibraryCallKit::inline_native_getClass() {
1.3226 + Node* obj = null_check_receiver(callee());
1.3227 + if (stopped()) return true;
1.3228 + push( load_mirror_from_klass(load_object_klass(obj)) );
1.3229 + return true;
1.3230 +}
1.3231 +
1.3232 +//-----------------inline_native_Reflection_getCallerClass---------------------
1.3233 +// In the presence of deep enough inlining, getCallerClass() becomes a no-op.
1.3234 +//
1.3235 +// NOTE that this code must perform the same logic as
1.3236 +// vframeStream::security_get_caller_frame in that it must skip
1.3237 +// Method.invoke() and auxiliary frames.
1.3238 +
1.3239 +
1.3240 +
1.3241 +
1.3242 +bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
1.3243 + ciMethod* method = callee();
1.3244 +
1.3245 +#ifndef PRODUCT
1.3246 + if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
1.3247 + tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
1.3248 + }
1.3249 +#endif
1.3250 +
1.3251 + debug_only(int saved_sp = _sp);
1.3252 +
1.3253 + // Argument words: (int depth)
1.3254 + int nargs = 1;
1.3255 +
1.3256 + _sp += nargs;
1.3257 + Node* caller_depth_node = pop();
1.3258 +
1.3259 + assert(saved_sp == _sp, "must have correct argument count");
1.3260 +
1.3261 + // The depth value must be a constant in order for the runtime call
1.3262 + // to be eliminated.
1.3263 + const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
1.3264 + if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
1.3265 +#ifndef PRODUCT
1.3266 + if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
1.3267 + tty->print_cr(" Bailing out because caller depth was not a constant");
1.3268 + }
1.3269 +#endif
1.3270 + return false;
1.3271 + }
1.3272 + // Note that the JVM state at this point does not include the
1.3273 + // getCallerClass() frame which we are trying to inline. The
1.3274 + // semantics of getCallerClass(), however, are that the "first"
1.3275 + // frame is the getCallerClass() frame, so we subtract one from the
1.3276 + // requested depth before continuing. We don't inline requests of
1.3277 + // getCallerClass(0).
1.3278 + int caller_depth = caller_depth_type->get_con() - 1;
1.3279 + if (caller_depth < 0) {
1.3280 +#ifndef PRODUCT
1.3281 + if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
1.3282 + tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
1.3283 + }
1.3284 +#endif
1.3285 + return false;
1.3286 + }
1.3287 +
1.3288 + if (!jvms()->has_method()) {
1.3289 +#ifndef PRODUCT
1.3290 + if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
1.3291 + tty->print_cr(" Bailing out because intrinsic was inlined at top level");
1.3292 + }
1.3293 +#endif
1.3294 + return false;
1.3295 + }
1.3296 + int _depth = jvms()->depth(); // cache call chain depth
1.3297 +
1.3298 + // Walk back up the JVM state to find the caller at the required
1.3299 + // depth. NOTE that this code must perform the same logic as
1.3300 + // vframeStream::security_get_caller_frame in that it must skip
1.3301 + // Method.invoke() and auxiliary frames. Note also that depth is
1.3302 + // 1-based (1 is the bottom of the inlining).
1.3303 + int inlining_depth = _depth;
1.3304 + JVMState* caller_jvms = NULL;
1.3305 +
1.3306 + if (inlining_depth > 0) {
1.3307 + caller_jvms = jvms();
1.3308 + assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
1.3309 + do {
1.3310 + // The following if-tests should be performed in this order
1.3311 + if (is_method_invoke_or_aux_frame(caller_jvms)) {
1.3312 + // Skip a Method.invoke() or auxiliary frame
1.3313 + } else if (caller_depth > 0) {
1.3314 + // Skip real frame
1.3315 + --caller_depth;
1.3316 + } else {
1.3317 + // We're done: reached desired caller after skipping.
1.3318 + break;
1.3319 + }
1.3320 + caller_jvms = caller_jvms->caller();
1.3321 + --inlining_depth;
1.3322 + } while (inlining_depth > 0);
1.3323 + }
1.3324 +
1.3325 + if (inlining_depth == 0) {
1.3326 +#ifndef PRODUCT
1.3327 + if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
1.3328 + tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
1.3329 + tty->print_cr(" JVM state at this point:");
1.3330 + for (int i = _depth; i >= 1; i--) {
1.3331 + tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
1.3332 + }
1.3333 + }
1.3334 +#endif
1.3335 + return false; // Reached end of inlining
1.3336 + }
1.3337 +
1.3338 + // Acquire method holder as java.lang.Class
1.3339 + ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
1.3340 + ciInstance* caller_mirror = caller_klass->java_mirror();
1.3341 + // Push this as a constant
1.3342 + push(makecon(TypeInstPtr::make(caller_mirror)));
1.3343 +#ifndef PRODUCT
1.3344 + if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
1.3345 + 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);
1.3346 + tty->print_cr(" JVM state at this point:");
1.3347 + for (int i = _depth; i >= 1; i--) {
1.3348 + tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
1.3349 + }
1.3350 + }
1.3351 +#endif
1.3352 + return true;
1.3353 +}
1.3354 +
1.3355 +// Helper routine for above
1.3356 +bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
1.3357 + // Is this the Method.invoke method itself?
1.3358 + if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
1.3359 + return true;
1.3360 +
1.3361 + // Is this a helper, defined somewhere underneath MethodAccessorImpl.
1.3362 + ciKlass* k = jvms->method()->holder();
1.3363 + if (k->is_instance_klass()) {
1.3364 + ciInstanceKlass* ik = k->as_instance_klass();
1.3365 + for (; ik != NULL; ik = ik->super()) {
1.3366 + if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
1.3367 + ik == env()->find_system_klass(ik->name())) {
1.3368 + return true;
1.3369 + }
1.3370 + }
1.3371 + }
1.3372 +
1.3373 + return false;
1.3374 +}
1.3375 +
1.3376 +static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
1.3377 + // inline_native_AtomicLong_attemptUpdate() but it has no way of
1.3378 + // computing it since there is no lookup field by name function in the
1.3379 + // CI interface. This is computed and set by inline_native_AtomicLong_get().
1.3380 + // Using a static variable here is safe even if we have multiple compilation
1.3381 + // threads because the offset is constant. At worst the same offset will be
1.3382 + // computed and stored multiple
1.3383 +
1.3384 +bool LibraryCallKit::inline_native_AtomicLong_get() {
1.3385 + // Restore the stack and pop off the argument
1.3386 + _sp+=1;
1.3387 + Node *obj = pop();
1.3388 +
1.3389 + // get the offset of the "value" field. Since the CI interfaces
1.3390 + // does not provide a way to look up a field by name, we scan the bytecodes
1.3391 + // to get the field index. We expect the first 2 instructions of the method
1.3392 + // to be:
1.3393 + // 0 aload_0
1.3394 + // 1 getfield "value"
1.3395 + ciMethod* method = callee();
1.3396 + if (value_field_offset == -1)
1.3397 + {
1.3398 + ciField* value_field;
1.3399 + ciBytecodeStream iter(method);
1.3400 + Bytecodes::Code bc = iter.next();
1.3401 +
1.3402 + if ((bc != Bytecodes::_aload_0) &&
1.3403 + ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
1.3404 + return false;
1.3405 + bc = iter.next();
1.3406 + if (bc != Bytecodes::_getfield)
1.3407 + return false;
1.3408 + bool ignore;
1.3409 + value_field = iter.get_field(ignore);
1.3410 + value_field_offset = value_field->offset_in_bytes();
1.3411 + }
1.3412 +
1.3413 + // Null check without removing any arguments.
1.3414 + _sp++;
1.3415 + obj = do_null_check(obj, T_OBJECT);
1.3416 + _sp--;
1.3417 + // Check for locking null object
1.3418 + if (stopped()) return true;
1.3419 +
1.3420 + Node *adr = basic_plus_adr(obj, obj, value_field_offset);
1.3421 + const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
1.3422 + int alias_idx = C->get_alias_index(adr_type);
1.3423 +
1.3424 + Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
1.3425 +
1.3426 + push_pair(result);
1.3427 +
1.3428 + return true;
1.3429 +}
1.3430 +
1.3431 +bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
1.3432 + // Restore the stack and pop off the arguments
1.3433 + _sp+=5;
1.3434 + Node *newVal = pop_pair();
1.3435 + Node *oldVal = pop_pair();
1.3436 + Node *obj = pop();
1.3437 +
1.3438 + // we need the offset of the "value" field which was computed when
1.3439 + // inlining the get() method. Give up if we don't have it.
1.3440 + if (value_field_offset == -1)
1.3441 + return false;
1.3442 +
1.3443 + // Null check without removing any arguments.
1.3444 + _sp+=5;
1.3445 + obj = do_null_check(obj, T_OBJECT);
1.3446 + _sp-=5;
1.3447 + // Check for locking null object
1.3448 + if (stopped()) return true;
1.3449 +
1.3450 + Node *adr = basic_plus_adr(obj, obj, value_field_offset);
1.3451 + const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
1.3452 + int alias_idx = C->get_alias_index(adr_type);
1.3453 +
1.3454 + Node *result = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
1.3455 + Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(result));
1.3456 + set_memory(store_proj, alias_idx);
1.3457 +
1.3458 + push(result);
1.3459 + return true;
1.3460 +}
1.3461 +
1.3462 +bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
1.3463 + // restore the arguments
1.3464 + _sp += arg_size();
1.3465 +
1.3466 + switch (id) {
1.3467 + case vmIntrinsics::_floatToRawIntBits:
1.3468 + push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
1.3469 + break;
1.3470 +
1.3471 + case vmIntrinsics::_intBitsToFloat:
1.3472 + push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
1.3473 + break;
1.3474 +
1.3475 + case vmIntrinsics::_doubleToRawLongBits:
1.3476 + push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
1.3477 + break;
1.3478 +
1.3479 + case vmIntrinsics::_longBitsToDouble:
1.3480 + push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
1.3481 + break;
1.3482 +
1.3483 + case vmIntrinsics::_doubleToLongBits: {
1.3484 + Node* value = pop_pair();
1.3485 +
1.3486 + // two paths (plus control) merge in a wood
1.3487 + RegionNode *r = new (C, 3) RegionNode(3);
1.3488 + Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
1.3489 +
1.3490 + Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
1.3491 + // Build the boolean node
1.3492 + Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
1.3493 +
1.3494 + // Branch either way.
1.3495 + // NaN case is less traveled, which makes all the difference.
1.3496 + IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1.3497 + Node *opt_isnan = _gvn.transform(ifisnan);
1.3498 + assert( opt_isnan->is_If(), "Expect an IfNode");
1.3499 + IfNode *opt_ifisnan = (IfNode*)opt_isnan;
1.3500 + Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
1.3501 +
1.3502 + set_control(iftrue);
1.3503 +
1.3504 + static const jlong nan_bits = CONST64(0x7ff8000000000000);
1.3505 + Node *slow_result = longcon(nan_bits); // return NaN
1.3506 + phi->init_req(1, _gvn.transform( slow_result ));
1.3507 + r->init_req(1, iftrue);
1.3508 +
1.3509 + // Else fall through
1.3510 + Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
1.3511 + set_control(iffalse);
1.3512 +
1.3513 + phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
1.3514 + r->init_req(2, iffalse);
1.3515 +
1.3516 + // Post merge
1.3517 + set_control(_gvn.transform(r));
1.3518 + record_for_igvn(r);
1.3519 +
1.3520 + Node* result = _gvn.transform(phi);
1.3521 + assert(result->bottom_type()->isa_long(), "must be");
1.3522 + push_pair(result);
1.3523 +
1.3524 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3525 +
1.3526 + break;
1.3527 + }
1.3528 +
1.3529 + case vmIntrinsics::_floatToIntBits: {
1.3530 + Node* value = pop();
1.3531 +
1.3532 + // two paths (plus control) merge in a wood
1.3533 + RegionNode *r = new (C, 3) RegionNode(3);
1.3534 + Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
1.3535 +
1.3536 + Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
1.3537 + // Build the boolean node
1.3538 + Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
1.3539 +
1.3540 + // Branch either way.
1.3541 + // NaN case is less traveled, which makes all the difference.
1.3542 + IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1.3543 + Node *opt_isnan = _gvn.transform(ifisnan);
1.3544 + assert( opt_isnan->is_If(), "Expect an IfNode");
1.3545 + IfNode *opt_ifisnan = (IfNode*)opt_isnan;
1.3546 + Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
1.3547 +
1.3548 + set_control(iftrue);
1.3549 +
1.3550 + static const jint nan_bits = 0x7fc00000;
1.3551 + Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
1.3552 + phi->init_req(1, _gvn.transform( slow_result ));
1.3553 + r->init_req(1, iftrue);
1.3554 +
1.3555 + // Else fall through
1.3556 + Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
1.3557 + set_control(iffalse);
1.3558 +
1.3559 + phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
1.3560 + r->init_req(2, iffalse);
1.3561 +
1.3562 + // Post merge
1.3563 + set_control(_gvn.transform(r));
1.3564 + record_for_igvn(r);
1.3565 +
1.3566 + Node* result = _gvn.transform(phi);
1.3567 + assert(result->bottom_type()->isa_int(), "must be");
1.3568 + push(result);
1.3569 +
1.3570 + C->set_has_split_ifs(true); // Has chance for split-if optimization
1.3571 +
1.3572 + break;
1.3573 + }
1.3574 +
1.3575 + default:
1.3576 + ShouldNotReachHere();
1.3577 + }
1.3578 +
1.3579 + return true;
1.3580 +}
1.3581 +
1.3582 +#ifdef _LP64
1.3583 +#define XTOP ,top() /*additional argument*/
1.3584 +#else //_LP64
1.3585 +#define XTOP /*no additional argument*/
1.3586 +#endif //_LP64
1.3587 +
1.3588 +//----------------------inline_unsafe_copyMemory-------------------------
1.3589 +bool LibraryCallKit::inline_unsafe_copyMemory() {
1.3590 + if (callee()->is_static()) return false; // caller must have the capability!
1.3591 + int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
1.3592 + assert(signature()->size() == nargs-1, "copy has 5 arguments");
1.3593 + null_check_receiver(callee()); // check then ignore argument(0)
1.3594 + if (stopped()) return true;
1.3595 +
1.3596 + C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
1.3597 +
1.3598 + Node* src_ptr = argument(1);
1.3599 + Node* src_off = ConvL2X(argument(2));
1.3600 + assert(argument(3)->is_top(), "2nd half of long");
1.3601 + Node* dst_ptr = argument(4);
1.3602 + Node* dst_off = ConvL2X(argument(5));
1.3603 + assert(argument(6)->is_top(), "2nd half of long");
1.3604 + Node* size = ConvL2X(argument(7));
1.3605 + assert(argument(8)->is_top(), "2nd half of long");
1.3606 +
1.3607 + assert(Unsafe_field_offset_to_byte_offset(11) == 11,
1.3608 + "fieldOffset must be byte-scaled");
1.3609 +
1.3610 + Node* src = make_unsafe_address(src_ptr, src_off);
1.3611 + Node* dst = make_unsafe_address(dst_ptr, dst_off);
1.3612 +
1.3613 + // Conservatively insert a memory barrier on all memory slices.
1.3614 + // Do not let writes of the copy source or destination float below the copy.
1.3615 + insert_mem_bar(Op_MemBarCPUOrder);
1.3616 +
1.3617 + // Call it. Note that the length argument is not scaled.
1.3618 + make_runtime_call(RC_LEAF|RC_NO_FP,
1.3619 + OptoRuntime::fast_arraycopy_Type(),
1.3620 + StubRoutines::unsafe_arraycopy(),
1.3621 + "unsafe_arraycopy",
1.3622 + TypeRawPtr::BOTTOM,
1.3623 + src, dst, size XTOP);
1.3624 +
1.3625 + // Do not let reads of the copy destination float above the copy.
1.3626 + insert_mem_bar(Op_MemBarCPUOrder);
1.3627 +
1.3628 + return true;
1.3629 +}
1.3630 +
1.3631 +
1.3632 +//------------------------inline_native_clone----------------------------
1.3633 +// Here are the simple edge cases:
1.3634 +// null receiver => normal trap
1.3635 +// virtual and clone was overridden => slow path to out-of-line clone
1.3636 +// not cloneable or finalizer => slow path to out-of-line Object.clone
1.3637 +//
1.3638 +// The general case has two steps, allocation and copying.
1.3639 +// Allocation has two cases, and uses GraphKit::new_instance or new_array.
1.3640 +//
1.3641 +// Copying also has two cases, oop arrays and everything else.
1.3642 +// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
1.3643 +// Everything else uses the tight inline loop supplied by CopyArrayNode.
1.3644 +//
1.3645 +// These steps fold up nicely if and when the cloned object's klass
1.3646 +// can be sharply typed as an object array, a type array, or an instance.
1.3647 +//
1.3648 +bool LibraryCallKit::inline_native_clone(bool is_virtual) {
1.3649 + int nargs = 1;
1.3650 + Node* obj = null_check_receiver(callee());
1.3651 + if (stopped()) return true;
1.3652 + Node* obj_klass = load_object_klass(obj);
1.3653 + const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
1.3654 + const TypeOopPtr* toop = ((tklass != NULL)
1.3655 + ? tklass->as_instance_type()
1.3656 + : TypeInstPtr::NOTNULL);
1.3657 +
1.3658 + // Conservatively insert a memory barrier on all memory slices.
1.3659 + // Do not let writes into the original float below the clone.
1.3660 + insert_mem_bar(Op_MemBarCPUOrder);
1.3661 +
1.3662 + // paths into result_reg:
1.3663 + enum {
1.3664 + _slow_path = 1, // out-of-line call to clone method (virtual or not)
1.3665 + _objArray_path, // plain allocation, plus arrayof_oop_arraycopy
1.3666 + _fast_path, // plain allocation, plus a CopyArray operation
1.3667 + PATH_LIMIT
1.3668 + };
1.3669 + RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
1.3670 + PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
1.3671 + TypeInstPtr::NOTNULL);
1.3672 + PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
1.3673 + PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
1.3674 + TypePtr::BOTTOM);
1.3675 + record_for_igvn(result_reg);
1.3676 +
1.3677 + const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1.3678 + int raw_adr_idx = Compile::AliasIdxRaw;
1.3679 + const bool raw_mem_only = true;
1.3680 +
1.3681 + // paths into alloc_reg (on the fast path, just before the CopyArray):
1.3682 + enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT };
1.3683 + RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT);
1.3684 + PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type);
1.3685 + PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X);
1.3686 + PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO);
1.3687 + PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY,
1.3688 + raw_adr_type);
1.3689 + record_for_igvn(alloc_reg);
1.3690 +
1.3691 + bool card_mark = false; // (see below)
1.3692 +
1.3693 + Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
1.3694 + if (array_ctl != NULL) {
1.3695 + // It's an array.
1.3696 + PreserveJVMState pjvms(this);
1.3697 + set_control(array_ctl);
1.3698 + Node* obj_length = load_array_length(obj);
1.3699 + Node* obj_size = NULL;
1.3700 + _sp += nargs; // set original stack for use by uncommon_trap
1.3701 + Node* alloc_obj = new_array(obj_klass, obj_length,
1.3702 + raw_mem_only, &obj_size);
1.3703 + _sp -= nargs;
1.3704 + assert(obj_size != NULL, "");
1.3705 + Node* raw_obj = alloc_obj->in(1);
1.3706 + assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
1.3707 + if (ReduceBulkZeroing) {
1.3708 + AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
1.3709 + if (alloc != NULL) {
1.3710 + // We will be completely responsible for initializing this object.
1.3711 + alloc->maybe_set_complete(&_gvn);
1.3712 + }
1.3713 + }
1.3714 +
1.3715 + if (!use_ReduceInitialCardMarks()) {
1.3716 + // If it is an oop array, it requires very special treatment,
1.3717 + // because card marking is required on each card of the array.
1.3718 + Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
1.3719 + if (is_obja != NULL) {
1.3720 + PreserveJVMState pjvms2(this);
1.3721 + set_control(is_obja);
1.3722 + // Generate a direct call to the right arraycopy function(s).
1.3723 + bool disjoint_bases = true;
1.3724 + bool length_never_negative = true;
1.3725 + generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
1.3726 + obj, intcon(0), alloc_obj, intcon(0),
1.3727 + obj_length, nargs,
1.3728 + disjoint_bases, length_never_negative);
1.3729 + result_reg->init_req(_objArray_path, control());
1.3730 + result_val->init_req(_objArray_path, alloc_obj);
1.3731 + result_i_o ->set_req(_objArray_path, i_o());
1.3732 + result_mem ->set_req(_objArray_path, reset_memory());
1.3733 + }
1.3734 + }
1.3735 + // We can dispense with card marks if we know the allocation
1.3736 + // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
1.3737 + // causes the non-eden paths to simulate a fresh allocation,
1.3738 + // insofar that no further card marks are required to initialize
1.3739 + // the object.
1.3740 +
1.3741 + // Otherwise, there are no card marks to worry about.
1.3742 + alloc_val->init_req(_typeArray_alloc, raw_obj);
1.3743 + alloc_siz->init_req(_typeArray_alloc, obj_size);
1.3744 + alloc_reg->init_req(_typeArray_alloc, control());
1.3745 + alloc_i_o->init_req(_typeArray_alloc, i_o());
1.3746 + alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type));
1.3747 + }
1.3748 +
1.3749 + // We only go to the fast case code if we pass a number of guards.
1.3750 + // The paths which do not pass are accumulated in the slow_region.
1.3751 + RegionNode* slow_region = new (C, 1) RegionNode(1);
1.3752 + record_for_igvn(slow_region);
1.3753 + if (!stopped()) {
1.3754 + // It's an instance. Make the slow-path tests.
1.3755 + // If this is a virtual call, we generate a funny guard. We grab
1.3756 + // the vtable entry corresponding to clone() from the target object.
1.3757 + // If the target method which we are calling happens to be the
1.3758 + // Object clone() method, we pass the guard. We do not need this
1.3759 + // guard for non-virtual calls; the caller is known to be the native
1.3760 + // Object clone().
1.3761 + if (is_virtual) {
1.3762 + generate_virtual_guard(obj_klass, slow_region);
1.3763 + }
1.3764 +
1.3765 + // The object must be cloneable and must not have a finalizer.
1.3766 + // Both of these conditions may be checked in a single test.
1.3767 + // We could optimize the cloneable test further, but we don't care.
1.3768 + generate_access_flags_guard(obj_klass,
1.3769 + // Test both conditions:
1.3770 + JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
1.3771 + // Must be cloneable but not finalizer:
1.3772 + JVM_ACC_IS_CLONEABLE,
1.3773 + slow_region);
1.3774 + }
1.3775 +
1.3776 + if (!stopped()) {
1.3777 + // It's an instance, and it passed the slow-path tests.
1.3778 + PreserveJVMState pjvms(this);
1.3779 + Node* obj_size = NULL;
1.3780 + Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
1.3781 + assert(obj_size != NULL, "");
1.3782 + Node* raw_obj = alloc_obj->in(1);
1.3783 + assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
1.3784 + if (ReduceBulkZeroing) {
1.3785 + AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
1.3786 + if (alloc != NULL && !alloc->maybe_set_complete(&_gvn))
1.3787 + alloc = NULL;
1.3788 + }
1.3789 + if (!use_ReduceInitialCardMarks()) {
1.3790 + // Put in store barrier for any and all oops we are sticking
1.3791 + // into this object. (We could avoid this if we could prove
1.3792 + // that the object type contains no oop fields at all.)
1.3793 + card_mark = true;
1.3794 + }
1.3795 + alloc_val->init_req(_instance_alloc, raw_obj);
1.3796 + alloc_siz->init_req(_instance_alloc, obj_size);
1.3797 + alloc_reg->init_req(_instance_alloc, control());
1.3798 + alloc_i_o->init_req(_instance_alloc, i_o());
1.3799 + alloc_mem->init_req(_instance_alloc, memory(raw_adr_type));
1.3800 + }
1.3801 +
1.3802 + // Generate code for the slow case. We make a call to clone().
1.3803 + set_control(_gvn.transform(slow_region));
1.3804 + if (!stopped()) {
1.3805 + PreserveJVMState pjvms(this);
1.3806 + CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
1.3807 + Node* slow_result = set_results_for_java_call(slow_call);
1.3808 + // this->control() comes from set_results_for_java_call
1.3809 + result_reg->init_req(_slow_path, control());
1.3810 + result_val->init_req(_slow_path, slow_result);
1.3811 + result_i_o ->set_req(_slow_path, i_o());
1.3812 + result_mem ->set_req(_slow_path, reset_memory());
1.3813 + }
1.3814 +
1.3815 + // The object is allocated, as an array and/or an instance. Now copy it.
1.3816 + set_control( _gvn.transform(alloc_reg) );
1.3817 + set_i_o( _gvn.transform(alloc_i_o) );
1.3818 + set_memory( _gvn.transform(alloc_mem), raw_adr_type );
1.3819 + Node* raw_obj = _gvn.transform(alloc_val);
1.3820 +
1.3821 + if (!stopped()) {
1.3822 + // Copy the fastest available way.
1.3823 + // (No need for PreserveJVMState, since we're using it all up now.)
1.3824 + Node* src = obj;
1.3825 + Node* dest = raw_obj;
1.3826 + Node* end = dest;
1.3827 + Node* size = _gvn.transform(alloc_siz);
1.3828 +
1.3829 + // Exclude the header.
1.3830 + int base_off = sizeof(oopDesc);
1.3831 + src = basic_plus_adr(src, base_off);
1.3832 + dest = basic_plus_adr(dest, base_off);
1.3833 + end = basic_plus_adr(end, size);
1.3834 +
1.3835 + // Compute the length also, if needed:
1.3836 + Node* countx = size;
1.3837 + countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
1.3838 + countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
1.3839 +
1.3840 + // Select an appropriate instruction to initialize the range.
1.3841 + // The CopyArray instruction (if supported) can be optimized
1.3842 + // into a discrete set of scalar loads and stores.
1.3843 + bool disjoint_bases = true;
1.3844 + generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
1.3845 + src, NULL, dest, NULL, countx);
1.3846 +
1.3847 + // Now that the object is properly initialized, type it as an oop.
1.3848 + // Use a secondary InitializeNode memory barrier.
1.3849 + InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx,
1.3850 + raw_obj)->as_Initialize();
1.3851 + init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
1.3852 + Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj,
1.3853 + TypeInstPtr::NOTNULL);
1.3854 + new_obj = _gvn.transform(new_obj);
1.3855 +
1.3856 + // If necessary, emit some card marks afterwards. (Non-arrays only.)
1.3857 + if (card_mark) {
1.3858 + Node* no_particular_value = NULL;
1.3859 + Node* no_particular_field = NULL;
1.3860 + post_barrier(control(),
1.3861 + memory(raw_adr_type),
1.3862 + new_obj,
1.3863 + no_particular_field,
1.3864 + raw_adr_idx,
1.3865 + no_particular_value,
1.3866 + T_OBJECT,
1.3867 + false);
1.3868 + }
1.3869 + // Present the results of the slow call.
1.3870 + result_reg->init_req(_fast_path, control());
1.3871 + result_val->init_req(_fast_path, new_obj);
1.3872 + result_i_o ->set_req(_fast_path, i_o());
1.3873 + result_mem ->set_req(_fast_path, reset_memory());
1.3874 + }
1.3875 +
1.3876 + // Return the combined state.
1.3877 + set_control( _gvn.transform(result_reg) );
1.3878 + set_i_o( _gvn.transform(result_i_o) );
1.3879 + set_all_memory( _gvn.transform(result_mem) );
1.3880 +
1.3881 + // Cast the result to a sharper type, since we know what clone does.
1.3882 + Node* new_obj = _gvn.transform(result_val);
1.3883 + Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop);
1.3884 + push(_gvn.transform(cast));
1.3885 +
1.3886 + return true;
1.3887 +}
1.3888 +
1.3889 +
1.3890 +// constants for computing the copy function
1.3891 +enum {
1.3892 + COPYFUNC_UNALIGNED = 0,
1.3893 + COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
1.3894 + COPYFUNC_CONJOINT = 0,
1.3895 + COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
1.3896 +};
1.3897 +
1.3898 +// Note: The condition "disjoint" applies also for overlapping copies
1.3899 +// where an descending copy is permitted (i.e., dest_offset <= src_offset).
1.3900 +static address
1.3901 +select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
1.3902 + int selector =
1.3903 + (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
1.3904 + (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
1.3905 +
1.3906 +#define RETURN_STUB(xxx_arraycopy) { \
1.3907 + name = #xxx_arraycopy; \
1.3908 + return StubRoutines::xxx_arraycopy(); }
1.3909 +
1.3910 + switch (t) {
1.3911 + case T_BYTE:
1.3912 + case T_BOOLEAN:
1.3913 + switch (selector) {
1.3914 + case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
1.3915 + case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
1.3916 + case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
1.3917 + case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
1.3918 + }
1.3919 + case T_CHAR:
1.3920 + case T_SHORT:
1.3921 + switch (selector) {
1.3922 + case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
1.3923 + case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
1.3924 + case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
1.3925 + case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
1.3926 + }
1.3927 + case T_INT:
1.3928 + case T_FLOAT:
1.3929 + switch (selector) {
1.3930 + case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
1.3931 + case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
1.3932 + case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
1.3933 + case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
1.3934 + }
1.3935 + case T_DOUBLE:
1.3936 + case T_LONG:
1.3937 + switch (selector) {
1.3938 + case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
1.3939 + case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
1.3940 + case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
1.3941 + case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
1.3942 + }
1.3943 + case T_ARRAY:
1.3944 + case T_OBJECT:
1.3945 + switch (selector) {
1.3946 + case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
1.3947 + case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
1.3948 + case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
1.3949 + case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
1.3950 + }
1.3951 + default:
1.3952 + ShouldNotReachHere();
1.3953 + return NULL;
1.3954 + }
1.3955 +
1.3956 +#undef RETURN_STUB
1.3957 +}
1.3958 +
1.3959 +//------------------------------basictype2arraycopy----------------------------
1.3960 +address LibraryCallKit::basictype2arraycopy(BasicType t,
1.3961 + Node* src_offset,
1.3962 + Node* dest_offset,
1.3963 + bool disjoint_bases,
1.3964 + const char* &name) {
1.3965 + const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
1.3966 + const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
1.3967 +
1.3968 + bool aligned = false;
1.3969 + bool disjoint = disjoint_bases;
1.3970 +
1.3971 + // if the offsets are the same, we can treat the memory regions as
1.3972 + // disjoint, because either the memory regions are in different arrays,
1.3973 + // or they are identical (which we can treat as disjoint.) We can also
1.3974 + // treat a copy with a destination index less that the source index
1.3975 + // as disjoint since a low->high copy will work correctly in this case.
1.3976 + if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
1.3977 + dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
1.3978 + // both indices are constants
1.3979 + int s_offs = src_offset_inttype->get_con();
1.3980 + int d_offs = dest_offset_inttype->get_con();
1.3981 + int element_size = type2aelembytes[t];
1.3982 + aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
1.3983 + ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
1.3984 + if (s_offs >= d_offs) disjoint = true;
1.3985 + } else if (src_offset == dest_offset && src_offset != NULL) {
1.3986 + // This can occur if the offsets are identical non-constants.
1.3987 + disjoint = true;
1.3988 + }
1.3989 +
1.3990 + return select_arraycopy_function(t, aligned, disjoint, name);
1.3991 +}
1.3992 +
1.3993 +
1.3994 +//------------------------------inline_arraycopy-----------------------
1.3995 +bool LibraryCallKit::inline_arraycopy() {
1.3996 + // Restore the stack and pop off the arguments.
1.3997 + int nargs = 5; // 2 oops, 3 ints, no size_t or long
1.3998 + assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
1.3999 +
1.4000 + Node *src = argument(0);
1.4001 + Node *src_offset = argument(1);
1.4002 + Node *dest = argument(2);
1.4003 + Node *dest_offset = argument(3);
1.4004 + Node *length = argument(4);
1.4005 +
1.4006 + // Compile time checks. If any of these checks cannot be verified at compile time,
1.4007 + // we do not make a fast path for this call. Instead, we let the call remain as it
1.4008 + // is. The checks we choose to mandate at compile time are:
1.4009 + //
1.4010 + // (1) src and dest are arrays.
1.4011 + const Type* src_type = src->Value(&_gvn);
1.4012 + const Type* dest_type = dest->Value(&_gvn);
1.4013 + const TypeAryPtr* top_src = src_type->isa_aryptr();
1.4014 + const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1.4015 + if (top_src == NULL || top_src->klass() == NULL ||
1.4016 + top_dest == NULL || top_dest->klass() == NULL) {
1.4017 + // Conservatively insert a memory barrier on all memory slices.
1.4018 + // Do not let writes into the source float below the arraycopy.
1.4019 + insert_mem_bar(Op_MemBarCPUOrder);
1.4020 +
1.4021 + // Call StubRoutines::generic_arraycopy stub.
1.4022 + generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
1.4023 + src, src_offset, dest, dest_offset, length,
1.4024 + nargs);
1.4025 +
1.4026 + // Do not let reads from the destination float above the arraycopy.
1.4027 + // Since we cannot type the arrays, we don't know which slices
1.4028 + // might be affected. We could restrict this barrier only to those
1.4029 + // memory slices which pertain to array elements--but don't bother.
1.4030 + if (!InsertMemBarAfterArraycopy)
1.4031 + // (If InsertMemBarAfterArraycopy, there is already one in place.)
1.4032 + insert_mem_bar(Op_MemBarCPUOrder);
1.4033 + return true;
1.4034 + }
1.4035 +
1.4036 + // (2) src and dest arrays must have elements of the same BasicType
1.4037 + // Figure out the size and type of the elements we will be copying.
1.4038 + BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
1.4039 + BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1.4040 + if (src_elem == T_ARRAY) src_elem = T_OBJECT;
1.4041 + if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
1.4042 +
1.4043 + if (src_elem != dest_elem || dest_elem == T_VOID) {
1.4044 + // The component types are not the same or are not recognized. Punt.
1.4045 + // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1.4046 + generate_slow_arraycopy(TypePtr::BOTTOM,
1.4047 + src, src_offset, dest, dest_offset, length,
1.4048 + nargs);
1.4049 + return true;
1.4050 + }
1.4051 +
1.4052 + //---------------------------------------------------------------------------
1.4053 + // We will make a fast path for this call to arraycopy.
1.4054 +
1.4055 + // We have the following tests left to perform:
1.4056 + //
1.4057 + // (3) src and dest must not be null.
1.4058 + // (4) src_offset must not be negative.
1.4059 + // (5) dest_offset must not be negative.
1.4060 + // (6) length must not be negative.
1.4061 + // (7) src_offset + length must not exceed length of src.
1.4062 + // (8) dest_offset + length must not exceed length of dest.
1.4063 + // (9) each element of an oop array must be assignable
1.4064 +
1.4065 + RegionNode* slow_region = new (C, 1) RegionNode(1);
1.4066 + record_for_igvn(slow_region);
1.4067 +
1.4068 + // (3) operands must not be null
1.4069 + // We currently perform our null checks with the do_null_check routine.
1.4070 + // This means that the null exceptions will be reported in the caller
1.4071 + // rather than (correctly) reported inside of the native arraycopy call.
1.4072 + // This should be corrected, given time. We do our null check with the
1.4073 + // stack pointer restored.
1.4074 + _sp += nargs;
1.4075 + src = do_null_check(src, T_ARRAY);
1.4076 + dest = do_null_check(dest, T_ARRAY);
1.4077 + _sp -= nargs;
1.4078 +
1.4079 + // (4) src_offset must not be negative.
1.4080 + generate_negative_guard(src_offset, slow_region);
1.4081 +
1.4082 + // (5) dest_offset must not be negative.
1.4083 + generate_negative_guard(dest_offset, slow_region);
1.4084 +
1.4085 + // (6) length must not be negative (moved to generate_arraycopy()).
1.4086 + // generate_negative_guard(length, slow_region);
1.4087 +
1.4088 + // (7) src_offset + length must not exceed length of src.
1.4089 + generate_limit_guard(src_offset, length,
1.4090 + load_array_length(src),
1.4091 + slow_region);
1.4092 +
1.4093 + // (8) dest_offset + length must not exceed length of dest.
1.4094 + generate_limit_guard(dest_offset, length,
1.4095 + load_array_length(dest),
1.4096 + slow_region);
1.4097 +
1.4098 + // (9) each element of an oop array must be assignable
1.4099 + // The generate_arraycopy subroutine checks this.
1.4100 +
1.4101 + // This is where the memory effects are placed:
1.4102 + const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1.4103 + generate_arraycopy(adr_type, dest_elem,
1.4104 + src, src_offset, dest, dest_offset, length,
1.4105 + nargs, false, false, slow_region);
1.4106 +
1.4107 + return true;
1.4108 +}
1.4109 +
1.4110 +//-----------------------------generate_arraycopy----------------------
1.4111 +// Generate an optimized call to arraycopy.
1.4112 +// Caller must guard against non-arrays.
1.4113 +// Caller must determine a common array basic-type for both arrays.
1.4114 +// Caller must validate offsets against array bounds.
1.4115 +// The slow_region has already collected guard failure paths
1.4116 +// (such as out of bounds length or non-conformable array types).
1.4117 +// The generated code has this shape, in general:
1.4118 +//
1.4119 +// if (length == 0) return // via zero_path
1.4120 +// slowval = -1
1.4121 +// if (types unknown) {
1.4122 +// slowval = call generic copy loop
1.4123 +// if (slowval == 0) return // via checked_path
1.4124 +// } else if (indexes in bounds) {
1.4125 +// if ((is object array) && !(array type check)) {
1.4126 +// slowval = call checked copy loop
1.4127 +// if (slowval == 0) return // via checked_path
1.4128 +// } else {
1.4129 +// call bulk copy loop
1.4130 +// return // via fast_path
1.4131 +// }
1.4132 +// }
1.4133 +// // adjust params for remaining work:
1.4134 +// if (slowval != -1) {
1.4135 +// n = -1^slowval; src_offset += n; dest_offset += n; length -= n
1.4136 +// }
1.4137 +// slow_region:
1.4138 +// call slow arraycopy(src, src_offset, dest, dest_offset, length)
1.4139 +// return // via slow_call_path
1.4140 +//
1.4141 +// This routine is used from several intrinsics: System.arraycopy,
1.4142 +// Object.clone (the array subcase), and Arrays.copyOf[Range].
1.4143 +//
1.4144 +void
1.4145 +LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
1.4146 + BasicType basic_elem_type,
1.4147 + Node* src, Node* src_offset,
1.4148 + Node* dest, Node* dest_offset,
1.4149 + Node* copy_length,
1.4150 + int nargs,
1.4151 + bool disjoint_bases,
1.4152 + bool length_never_negative,
1.4153 + RegionNode* slow_region) {
1.4154 +
1.4155 + if (slow_region == NULL) {
1.4156 + slow_region = new(C,1) RegionNode(1);
1.4157 + record_for_igvn(slow_region);
1.4158 + }
1.4159 +
1.4160 + Node* original_dest = dest;
1.4161 + AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
1.4162 + Node* raw_dest = NULL; // used before zeroing, if needed
1.4163 + bool must_clear_dest = false;
1.4164 +
1.4165 + // See if this is the initialization of a newly-allocated array.
1.4166 + // If so, we will take responsibility here for initializing it to zero.
1.4167 + // (Note: Because tightly_coupled_allocation performs checks on the
1.4168 + // out-edges of the dest, we need to avoid making derived pointers
1.4169 + // from it until we have checked its uses.)
1.4170 + if (ReduceBulkZeroing
1.4171 + && !ZeroTLAB // pointless if already zeroed
1.4172 + && basic_elem_type != T_CONFLICT // avoid corner case
1.4173 + && !_gvn.eqv_uncast(src, dest)
1.4174 + && ((alloc = tightly_coupled_allocation(dest, slow_region))
1.4175 + != NULL)
1.4176 + && alloc->maybe_set_complete(&_gvn)) {
1.4177 + // "You break it, you buy it."
1.4178 + InitializeNode* init = alloc->initialization();
1.4179 + assert(init->is_complete(), "we just did this");
1.4180 + assert(dest->Opcode() == Op_CheckCastPP, "sanity");
1.4181 + assert(dest->in(0)->in(0) == init, "dest pinned");
1.4182 + raw_dest = dest->in(1); // grab the raw pointer!
1.4183 + original_dest = dest;
1.4184 + dest = raw_dest;
1.4185 + adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
1.4186 + // Decouple the original InitializeNode, turning it into a simple membar.
1.4187 + // We will build a new one at the end of this routine.
1.4188 + init->set_req(InitializeNode::RawAddress, top());
1.4189 + // From this point on, every exit path is responsible for
1.4190 + // initializing any non-copied parts of the object to zero.
1.4191 + must_clear_dest = true;
1.4192 + } else {
1.4193 + // No zeroing elimination here.
1.4194 + alloc = NULL;
1.4195 + //original_dest = dest;
1.4196 + //must_clear_dest = false;
1.4197 + }
1.4198 +
1.4199 + // Results are placed here:
1.4200 + enum { fast_path = 1, // normal void-returning assembly stub
1.4201 + checked_path = 2, // special assembly stub with cleanup
1.4202 + slow_call_path = 3, // something went wrong; call the VM
1.4203 + zero_path = 4, // bypass when length of copy is zero
1.4204 + bcopy_path = 5, // copy primitive array by 64-bit blocks
1.4205 + PATH_LIMIT = 6
1.4206 + };
1.4207 + RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
1.4208 + PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
1.4209 + PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
1.4210 + record_for_igvn(result_region);
1.4211 + _gvn.set_type_bottom(result_i_o);
1.4212 + _gvn.set_type_bottom(result_memory);
1.4213 + assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
1.4214 +
1.4215 + // The slow_control path:
1.4216 + Node* slow_control;
1.4217 + Node* slow_i_o = i_o();
1.4218 + Node* slow_mem = memory(adr_type);
1.4219 + debug_only(slow_control = (Node*) badAddress);
1.4220 +
1.4221 + // Checked control path:
1.4222 + Node* checked_control = top();
1.4223 + Node* checked_mem = NULL;
1.4224 + Node* checked_i_o = NULL;
1.4225 + Node* checked_value = NULL;
1.4226 +
1.4227 + if (basic_elem_type == T_CONFLICT) {
1.4228 + assert(!must_clear_dest, "");
1.4229 + Node* cv = generate_generic_arraycopy(adr_type,
1.4230 + src, src_offset, dest, dest_offset,
1.4231 + copy_length, nargs);
1.4232 + if (cv == NULL) cv = intcon(-1); // failure (no stub available)
1.4233 + checked_control = control();
1.4234 + checked_i_o = i_o();
1.4235 + checked_mem = memory(adr_type);
1.4236 + checked_value = cv;
1.4237 + set_control(top()); // no fast path
1.4238 + }
1.4239 +
1.4240 + Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
1.4241 + if (not_pos != NULL) {
1.4242 + PreserveJVMState pjvms(this);
1.4243 + set_control(not_pos);
1.4244 +
1.4245 + // (6) length must not be negative.
1.4246 + if (!length_never_negative) {
1.4247 + generate_negative_guard(copy_length, slow_region);
1.4248 + }
1.4249 +
1.4250 + if (!stopped() && must_clear_dest) {
1.4251 + Node* dest_length = alloc->in(AllocateNode::ALength);
1.4252 + if (_gvn.eqv_uncast(copy_length, dest_length)
1.4253 + || _gvn.find_int_con(dest_length, 1) <= 0) {
1.4254 + // There is no zeroing to do.
1.4255 + } else {
1.4256 + // Clear the whole thing since there are no source elements to copy.
1.4257 + generate_clear_array(adr_type, dest, basic_elem_type,
1.4258 + intcon(0), NULL,
1.4259 + alloc->in(AllocateNode::AllocSize));
1.4260 + }
1.4261 + }
1.4262 +
1.4263 + // Present the results of the fast call.
1.4264 + result_region->init_req(zero_path, control());
1.4265 + result_i_o ->init_req(zero_path, i_o());
1.4266 + result_memory->init_req(zero_path, memory(adr_type));
1.4267 + }
1.4268 +
1.4269 + if (!stopped() && must_clear_dest) {
1.4270 + // We have to initialize the *uncopied* part of the array to zero.
1.4271 + // The copy destination is the slice dest[off..off+len]. The other slices
1.4272 + // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
1.4273 + Node* dest_size = alloc->in(AllocateNode::AllocSize);
1.4274 + Node* dest_length = alloc->in(AllocateNode::ALength);
1.4275 + Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
1.4276 + copy_length) );
1.4277 +
1.4278 + // If there is a head section that needs zeroing, do it now.
1.4279 + if (find_int_con(dest_offset, -1) != 0) {
1.4280 + generate_clear_array(adr_type, dest, basic_elem_type,
1.4281 + intcon(0), dest_offset,
1.4282 + NULL);
1.4283 + }
1.4284 +
1.4285 + // Next, perform a dynamic check on the tail length.
1.4286 + // It is often zero, and we can win big if we prove this.
1.4287 + // There are two wins: Avoid generating the ClearArray
1.4288 + // with its attendant messy index arithmetic, and upgrade
1.4289 + // the copy to a more hardware-friendly word size of 64 bits.
1.4290 + Node* tail_ctl = NULL;
1.4291 + if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
1.4292 + Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
1.4293 + Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
1.4294 + tail_ctl = generate_slow_guard(bol_lt, NULL);
1.4295 + assert(tail_ctl != NULL || !stopped(), "must be an outcome");
1.4296 + }
1.4297 +
1.4298 + // At this point, let's assume there is no tail.
1.4299 + if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
1.4300 + // There is no tail. Try an upgrade to a 64-bit copy.
1.4301 + bool didit = false;
1.4302 + { PreserveJVMState pjvms(this);
1.4303 + didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
1.4304 + src, src_offset, dest, dest_offset,
1.4305 + dest_size);
1.4306 + if (didit) {
1.4307 + // Present the results of the block-copying fast call.
1.4308 + result_region->init_req(bcopy_path, control());
1.4309 + result_i_o ->init_req(bcopy_path, i_o());
1.4310 + result_memory->init_req(bcopy_path, memory(adr_type));
1.4311 + }
1.4312 + }
1.4313 + if (didit)
1.4314 + set_control(top()); // no regular fast path
1.4315 + }
1.4316 +
1.4317 + // Clear the tail, if any.
1.4318 + if (tail_ctl != NULL) {
1.4319 + Node* notail_ctl = stopped() ? NULL : control();
1.4320 + set_control(tail_ctl);
1.4321 + if (notail_ctl == NULL) {
1.4322 + generate_clear_array(adr_type, dest, basic_elem_type,
1.4323 + dest_tail, NULL,
1.4324 + dest_size);
1.4325 + } else {
1.4326 + // Make a local merge.
1.4327 + Node* done_ctl = new(C,3) RegionNode(3);
1.4328 + Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
1.4329 + done_ctl->init_req(1, notail_ctl);
1.4330 + done_mem->init_req(1, memory(adr_type));
1.4331 + generate_clear_array(adr_type, dest, basic_elem_type,
1.4332 + dest_tail, NULL,
1.4333 + dest_size);
1.4334 + done_ctl->init_req(2, control());
1.4335 + done_mem->init_req(2, memory(adr_type));
1.4336 + set_control( _gvn.transform(done_ctl) );
1.4337 + set_memory( _gvn.transform(done_mem), adr_type );
1.4338 + }
1.4339 + }
1.4340 + }
1.4341 +
1.4342 + BasicType copy_type = basic_elem_type;
1.4343 + assert(basic_elem_type != T_ARRAY, "caller must fix this");
1.4344 + if (!stopped() && copy_type == T_OBJECT) {
1.4345 + // If src and dest have compatible element types, we can copy bits.
1.4346 + // Types S[] and D[] are compatible if D is a supertype of S.
1.4347 + //
1.4348 + // If they are not, we will use checked_oop_disjoint_arraycopy,
1.4349 + // which performs a fast optimistic per-oop check, and backs off
1.4350 + // further to JVM_ArrayCopy on the first per-oop check that fails.
1.4351 + // (Actually, we don't move raw bits only; the GC requires card marks.)
1.4352 +
1.4353 + // Get the klassOop for both src and dest
1.4354 + Node* src_klass = load_object_klass(src);
1.4355 + Node* dest_klass = load_object_klass(dest);
1.4356 +
1.4357 + // Generate the subtype check.
1.4358 + // This might fold up statically, or then again it might not.
1.4359 + //
1.4360 + // Non-static example: Copying List<String>.elements to a new String[].
1.4361 + // The backing store for a List<String> is always an Object[],
1.4362 + // but its elements are always type String, if the generic types
1.4363 + // are correct at the source level.
1.4364 + //
1.4365 + // Test S[] against D[], not S against D, because (probably)
1.4366 + // the secondary supertype cache is less busy for S[] than S.
1.4367 + // This usually only matters when D is an interface.
1.4368 + Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
1.4369 + // Plug failing path into checked_oop_disjoint_arraycopy
1.4370 + if (not_subtype_ctrl != top()) {
1.4371 + PreserveJVMState pjvms(this);
1.4372 + set_control(not_subtype_ctrl);
1.4373 + // (At this point we can assume disjoint_bases, since types differ.)
1.4374 + int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
1.4375 + Node* p1 = basic_plus_adr(dest_klass, ek_offset);
1.4376 + Node* n1 = new (C, 3) LoadKlassNode(0, immutable_memory(), p1, TypeRawPtr::BOTTOM);
1.4377 + Node* dest_elem_klass = _gvn.transform(n1);
1.4378 + Node* cv = generate_checkcast_arraycopy(adr_type,
1.4379 + dest_elem_klass,
1.4380 + src, src_offset, dest, dest_offset,
1.4381 + copy_length,
1.4382 + nargs);
1.4383 + if (cv == NULL) cv = intcon(-1); // failure (no stub available)
1.4384 + checked_control = control();
1.4385 + checked_i_o = i_o();
1.4386 + checked_mem = memory(adr_type);
1.4387 + checked_value = cv;
1.4388 + }
1.4389 + // At this point we know we do not need type checks on oop stores.
1.4390 +
1.4391 + // Let's see if we need card marks:
1.4392 + if (alloc != NULL && use_ReduceInitialCardMarks()) {
1.4393 + // If we do not need card marks, copy using the jint or jlong stub.
1.4394 + copy_type = LP64_ONLY(T_LONG) NOT_LP64(T_INT);
1.4395 + assert(type2aelembytes[basic_elem_type] == type2aelembytes[copy_type],
1.4396 + "sizes agree");
1.4397 + }
1.4398 + }
1.4399 +
1.4400 + if (!stopped()) {
1.4401 + // Generate the fast path, if possible.
1.4402 + PreserveJVMState pjvms(this);
1.4403 + generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
1.4404 + src, src_offset, dest, dest_offset,
1.4405 + ConvI2X(copy_length));
1.4406 +
1.4407 + // Present the results of the fast call.
1.4408 + result_region->init_req(fast_path, control());
1.4409 + result_i_o ->init_req(fast_path, i_o());
1.4410 + result_memory->init_req(fast_path, memory(adr_type));
1.4411 + }
1.4412 +
1.4413 + // Here are all the slow paths up to this point, in one bundle:
1.4414 + slow_control = top();
1.4415 + if (slow_region != NULL)
1.4416 + slow_control = _gvn.transform(slow_region);
1.4417 + debug_only(slow_region = (RegionNode*)badAddress);
1.4418 +
1.4419 + set_control(checked_control);
1.4420 + if (!stopped()) {
1.4421 + // Clean up after the checked call.
1.4422 + // The returned value is either 0 or -1^K,
1.4423 + // where K = number of partially transferred array elements.
1.4424 + Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
1.4425 + Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1.4426 + IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
1.4427 +
1.4428 + // If it is 0, we are done, so transfer to the end.
1.4429 + Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
1.4430 + result_region->init_req(checked_path, checks_done);
1.4431 + result_i_o ->init_req(checked_path, checked_i_o);
1.4432 + result_memory->init_req(checked_path, checked_mem);
1.4433 +
1.4434 + // If it is not zero, merge into the slow call.
1.4435 + set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
1.4436 + RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
1.4437 + PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
1.4438 + PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
1.4439 + record_for_igvn(slow_reg2);
1.4440 + slow_reg2 ->init_req(1, slow_control);
1.4441 + slow_i_o2 ->init_req(1, slow_i_o);
1.4442 + slow_mem2 ->init_req(1, slow_mem);
1.4443 + slow_reg2 ->init_req(2, control());
1.4444 + slow_i_o2 ->init_req(2, i_o());
1.4445 + slow_mem2 ->init_req(2, memory(adr_type));
1.4446 +
1.4447 + slow_control = _gvn.transform(slow_reg2);
1.4448 + slow_i_o = _gvn.transform(slow_i_o2);
1.4449 + slow_mem = _gvn.transform(slow_mem2);
1.4450 +
1.4451 + if (alloc != NULL) {
1.4452 + // We'll restart from the very beginning, after zeroing the whole thing.
1.4453 + // This can cause double writes, but that's OK since dest is brand new.
1.4454 + // So we ignore the low 31 bits of the value returned from the stub.
1.4455 + } else {
1.4456 + // We must continue the copy exactly where it failed, or else
1.4457 + // another thread might see the wrong number of writes to dest.
1.4458 + Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
1.4459 + Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
1.4460 + slow_offset->init_req(1, intcon(0));
1.4461 + slow_offset->init_req(2, checked_offset);
1.4462 + slow_offset = _gvn.transform(slow_offset);
1.4463 +
1.4464 + // Adjust the arguments by the conditionally incoming offset.
1.4465 + Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
1.4466 + Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
1.4467 + Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
1.4468 +
1.4469 + // Tweak the node variables to adjust the code produced below:
1.4470 + src_offset = src_off_plus;
1.4471 + dest_offset = dest_off_plus;
1.4472 + copy_length = length_minus;
1.4473 + }
1.4474 + }
1.4475 +
1.4476 + set_control(slow_control);
1.4477 + if (!stopped()) {
1.4478 + // Generate the slow path, if needed.
1.4479 + PreserveJVMState pjvms(this); // replace_in_map may trash the map
1.4480 +
1.4481 + set_memory(slow_mem, adr_type);
1.4482 + set_i_o(slow_i_o);
1.4483 +
1.4484 + if (must_clear_dest) {
1.4485 + generate_clear_array(adr_type, dest, basic_elem_type,
1.4486 + intcon(0), NULL,
1.4487 + alloc->in(AllocateNode::AllocSize));
1.4488 + }
1.4489 +
1.4490 + if (dest != original_dest) {
1.4491 + // Promote from rawptr to oop, so it looks right in the call's GC map.
1.4492 + dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest,
1.4493 + TypeInstPtr::NOTNULL) );
1.4494 +
1.4495 + // Edit the call's debug-info to avoid referring to original_dest.
1.4496 + // (The problem with original_dest is that it isn't ready until
1.4497 + // after the InitializeNode completes, but this stuff is before.)
1.4498 + // Substitute in the locally valid dest_oop.
1.4499 + replace_in_map(original_dest, dest);
1.4500 + }
1.4501 +
1.4502 + generate_slow_arraycopy(adr_type,
1.4503 + src, src_offset, dest, dest_offset,
1.4504 + copy_length, nargs);
1.4505 +
1.4506 + result_region->init_req(slow_call_path, control());
1.4507 + result_i_o ->init_req(slow_call_path, i_o());
1.4508 + result_memory->init_req(slow_call_path, memory(adr_type));
1.4509 + }
1.4510 +
1.4511 + // Remove unused edges.
1.4512 + for (uint i = 1; i < result_region->req(); i++) {
1.4513 + if (result_region->in(i) == NULL)
1.4514 + result_region->init_req(i, top());
1.4515 + }
1.4516 +
1.4517 + // Finished; return the combined state.
1.4518 + set_control( _gvn.transform(result_region) );
1.4519 + set_i_o( _gvn.transform(result_i_o) );
1.4520 + set_memory( _gvn.transform(result_memory), adr_type );
1.4521 +
1.4522 + if (dest != original_dest) {
1.4523 + // Pin the "finished" array node after the arraycopy/zeroing operations.
1.4524 + // Use a secondary InitializeNode memory barrier.
1.4525 + InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
1.4526 + Compile::AliasIdxRaw,
1.4527 + raw_dest)->as_Initialize();
1.4528 + init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
1.4529 + _gvn.hash_delete(original_dest);
1.4530 + original_dest->set_req(0, control());
1.4531 + _gvn.hash_find_insert(original_dest); // put back into GVN table
1.4532 + }
1.4533 +
1.4534 + // The memory edges above are precise in order to model effects around
1.4535 + // array copyies accurately to allow value numbering of field loads around
1.4536 + // arraycopy. Such field loads, both before and after, are common in Java
1.4537 + // collections and similar classes involving header/array data structures.
1.4538 + //
1.4539 + // But with low number of register or when some registers are used or killed
1.4540 + // by arraycopy calls it causes registers spilling on stack. See 6544710.
1.4541 + // The next memory barrier is added to avoid it. If the arraycopy can be
1.4542 + // optimized away (which it can, sometimes) then we can manually remove
1.4543 + // the membar also.
1.4544 + if (InsertMemBarAfterArraycopy)
1.4545 + insert_mem_bar(Op_MemBarCPUOrder);
1.4546 +}
1.4547 +
1.4548 +
1.4549 +// Helper function which determines if an arraycopy immediately follows
1.4550 +// an allocation, with no intervening tests or other escapes for the object.
1.4551 +AllocateArrayNode*
1.4552 +LibraryCallKit::tightly_coupled_allocation(Node* ptr,
1.4553 + RegionNode* slow_region) {
1.4554 + if (stopped()) return NULL; // no fast path
1.4555 + if (C->AliasLevel() == 0) return NULL; // no MergeMems around
1.4556 +
1.4557 + AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
1.4558 + if (alloc == NULL) return NULL;
1.4559 +
1.4560 + Node* rawmem = memory(Compile::AliasIdxRaw);
1.4561 + // Is the allocation's memory state untouched?
1.4562 + if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
1.4563 + // Bail out if there have been raw-memory effects since the allocation.
1.4564 + // (Example: There might have been a call or safepoint.)
1.4565 + return NULL;
1.4566 + }
1.4567 + rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
1.4568 + if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
1.4569 + return NULL;
1.4570 + }
1.4571 +
1.4572 + // There must be no unexpected observers of this allocation.
1.4573 + for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
1.4574 + Node* obs = ptr->fast_out(i);
1.4575 + if (obs != this->map()) {
1.4576 + return NULL;
1.4577 + }
1.4578 + }
1.4579 +
1.4580 + // This arraycopy must unconditionally follow the allocation of the ptr.
1.4581 + Node* alloc_ctl = ptr->in(0);
1.4582 + assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
1.4583 +
1.4584 + Node* ctl = control();
1.4585 + while (ctl != alloc_ctl) {
1.4586 + // There may be guards which feed into the slow_region.
1.4587 + // Any other control flow means that we might not get a chance
1.4588 + // to finish initializing the allocated object.
1.4589 + if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
1.4590 + IfNode* iff = ctl->in(0)->as_If();
1.4591 + Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
1.4592 + assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
1.4593 + if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
1.4594 + ctl = iff->in(0); // This test feeds the known slow_region.
1.4595 + continue;
1.4596 + }
1.4597 + // One more try: Various low-level checks bottom out in
1.4598 + // uncommon traps. If the debug-info of the trap omits
1.4599 + // any reference to the allocation, as we've already
1.4600 + // observed, then there can be no objection to the trap.
1.4601 + bool found_trap = false;
1.4602 + for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
1.4603 + Node* obs = not_ctl->fast_out(j);
1.4604 + if (obs->in(0) == not_ctl && obs->is_Call() &&
1.4605 + (obs->as_Call()->entry_point() ==
1.4606 + SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
1.4607 + found_trap = true; break;
1.4608 + }
1.4609 + }
1.4610 + if (found_trap) {
1.4611 + ctl = iff->in(0); // This test feeds a harmless uncommon trap.
1.4612 + continue;
1.4613 + }
1.4614 + }
1.4615 + return NULL;
1.4616 + }
1.4617 +
1.4618 + // If we get this far, we have an allocation which immediately
1.4619 + // precedes the arraycopy, and we can take over zeroing the new object.
1.4620 + // The arraycopy will finish the initialization, and provide
1.4621 + // a new control state to which we will anchor the destination pointer.
1.4622 +
1.4623 + return alloc;
1.4624 +}
1.4625 +
1.4626 +// Helper for initialization of arrays, creating a ClearArray.
1.4627 +// It writes zero bits in [start..end), within the body of an array object.
1.4628 +// The memory effects are all chained onto the 'adr_type' alias category.
1.4629 +//
1.4630 +// Since the object is otherwise uninitialized, we are free
1.4631 +// to put a little "slop" around the edges of the cleared area,
1.4632 +// as long as it does not go back into the array's header,
1.4633 +// or beyond the array end within the heap.
1.4634 +//
1.4635 +// The lower edge can be rounded down to the nearest jint and the
1.4636 +// upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
1.4637 +//
1.4638 +// Arguments:
1.4639 +// adr_type memory slice where writes are generated
1.4640 +// dest oop of the destination array
1.4641 +// basic_elem_type element type of the destination
1.4642 +// slice_idx array index of first element to store
1.4643 +// slice_len number of elements to store (or NULL)
1.4644 +// dest_size total size in bytes of the array object
1.4645 +//
1.4646 +// Exactly one of slice_len or dest_size must be non-NULL.
1.4647 +// If dest_size is non-NULL, zeroing extends to the end of the object.
1.4648 +// If slice_len is non-NULL, the slice_idx value must be a constant.
1.4649 +void
1.4650 +LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
1.4651 + Node* dest,
1.4652 + BasicType basic_elem_type,
1.4653 + Node* slice_idx,
1.4654 + Node* slice_len,
1.4655 + Node* dest_size) {
1.4656 + // one or the other but not both of slice_len and dest_size:
1.4657 + assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
1.4658 + if (slice_len == NULL) slice_len = top();
1.4659 + if (dest_size == NULL) dest_size = top();
1.4660 +
1.4661 + // operate on this memory slice:
1.4662 + Node* mem = memory(adr_type); // memory slice to operate on
1.4663 +
1.4664 + // scaling and rounding of indexes:
1.4665 + int scale = exact_log2(type2aelembytes[basic_elem_type]);
1.4666 + int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1.4667 + int clear_low = (-1 << scale) & (BytesPerInt - 1);
1.4668 + int bump_bit = (-1 << scale) & BytesPerInt;
1.4669 +
1.4670 + // determine constant starts and ends
1.4671 + const intptr_t BIG_NEG = -128;
1.4672 + assert(BIG_NEG + 2*abase < 0, "neg enough");
1.4673 + intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
1.4674 + intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
1.4675 + if (slice_len_con == 0) {
1.4676 + return; // nothing to do here
1.4677 + }
1.4678 + intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
1.4679 + intptr_t end_con = find_intptr_t_con(dest_size, -1);
1.4680 + if (slice_idx_con >= 0 && slice_len_con >= 0) {
1.4681 + assert(end_con < 0, "not two cons");
1.4682 + end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
1.4683 + BytesPerLong);
1.4684 + }
1.4685 +
1.4686 + if (start_con >= 0 && end_con >= 0) {
1.4687 + // Constant start and end. Simple.
1.4688 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.4689 + start_con, end_con, &_gvn);
1.4690 + } else if (start_con >= 0 && dest_size != top()) {
1.4691 + // Constant start, pre-rounded end after the tail of the array.
1.4692 + Node* end = dest_size;
1.4693 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.4694 + start_con, end, &_gvn);
1.4695 + } else if (start_con >= 0 && slice_len != top()) {
1.4696 + // Constant start, non-constant end. End needs rounding up.
1.4697 + // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
1.4698 + intptr_t end_base = abase + (slice_idx_con << scale);
1.4699 + int end_round = (-1 << scale) & (BytesPerLong - 1);
1.4700 + Node* end = ConvI2X(slice_len);
1.4701 + if (scale != 0)
1.4702 + end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
1.4703 + end_base += end_round;
1.4704 + end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
1.4705 + end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
1.4706 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.4707 + start_con, end, &_gvn);
1.4708 + } else if (start_con < 0 && dest_size != top()) {
1.4709 + // Non-constant start, pre-rounded end after the tail of the array.
1.4710 + // This is almost certainly a "round-to-end" operation.
1.4711 + Node* start = slice_idx;
1.4712 + start = ConvI2X(start);
1.4713 + if (scale != 0)
1.4714 + start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
1.4715 + start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
1.4716 + if ((bump_bit | clear_low) != 0) {
1.4717 + int to_clear = (bump_bit | clear_low);
1.4718 + // Align up mod 8, then store a jint zero unconditionally
1.4719 + // just before the mod-8 boundary.
1.4720 + // This would only fail if the first array element were immediately
1.4721 + // after the length field, and were also at an even offset mod 8.
1.4722 + assert(((abase + bump_bit) & ~to_clear) - BytesPerInt
1.4723 + >= arrayOopDesc::length_offset_in_bytes() + BytesPerInt,
1.4724 + "store must not trash length field");
1.4725 +
1.4726 + // Bump 'start' up to (or past) the next jint boundary:
1.4727 + start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
1.4728 + // Round bumped 'start' down to jlong boundary in body of array.
1.4729 + start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
1.4730 + // Store a zero to the immediately preceding jint:
1.4731 + Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-BytesPerInt)) );
1.4732 + Node* p1 = basic_plus_adr(dest, x1);
1.4733 + mem = StoreNode::make(C, control(), mem, p1, adr_type, intcon(0), T_INT);
1.4734 + mem = _gvn.transform(mem);
1.4735 + }
1.4736 +
1.4737 + Node* end = dest_size; // pre-rounded
1.4738 + mem = ClearArrayNode::clear_memory(control(), mem, dest,
1.4739 + start, end, &_gvn);
1.4740 + } else {
1.4741 + // Non-constant start, unrounded non-constant end.
1.4742 + // (Nobody zeroes a random midsection of an array using this routine.)
1.4743 + ShouldNotReachHere(); // fix caller
1.4744 + }
1.4745 +
1.4746 + // Done.
1.4747 + set_memory(mem, adr_type);
1.4748 +}
1.4749 +
1.4750 +
1.4751 +bool
1.4752 +LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
1.4753 + BasicType basic_elem_type,
1.4754 + AllocateNode* alloc,
1.4755 + Node* src, Node* src_offset,
1.4756 + Node* dest, Node* dest_offset,
1.4757 + Node* dest_size) {
1.4758 + // See if there is an advantage from block transfer.
1.4759 + int scale = exact_log2(type2aelembytes[basic_elem_type]);
1.4760 + if (scale >= LogBytesPerLong)
1.4761 + return false; // it is already a block transfer
1.4762 +
1.4763 + // Look at the alignment of the starting offsets.
1.4764 + int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1.4765 + const intptr_t BIG_NEG = -128;
1.4766 + assert(BIG_NEG + 2*abase < 0, "neg enough");
1.4767 +
1.4768 + intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
1.4769 + intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
1.4770 + if (src_off < 0 || dest_off < 0)
1.4771 + // At present, we can only understand constants.
1.4772 + return false;
1.4773 +
1.4774 + if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1.4775 + // Non-aligned; too bad.
1.4776 + // One more chance: Pick off an initial 32-bit word.
1.4777 + // This is a common case, since abase can be odd mod 8.
1.4778 + if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1.4779 + ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1.4780 + Node* sptr = basic_plus_adr(src, src_off);
1.4781 + Node* dptr = basic_plus_adr(dest, dest_off);
1.4782 + Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
1.4783 + store_to_memory(control(), dptr, sval, T_INT, adr_type);
1.4784 + src_off += BytesPerInt;
1.4785 + dest_off += BytesPerInt;
1.4786 + } else {
1.4787 + return false;
1.4788 + }
1.4789 + }
1.4790 + assert(src_off % BytesPerLong == 0, "");
1.4791 + assert(dest_off % BytesPerLong == 0, "");
1.4792 +
1.4793 + // Do this copy by giant steps.
1.4794 + Node* sptr = basic_plus_adr(src, src_off);
1.4795 + Node* dptr = basic_plus_adr(dest, dest_off);
1.4796 + Node* countx = dest_size;
1.4797 + countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
1.4798 + countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
1.4799 +
1.4800 + bool disjoint_bases = true; // since alloc != NULL
1.4801 + generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
1.4802 + sptr, NULL, dptr, NULL, countx);
1.4803 +
1.4804 + return true;
1.4805 +}
1.4806 +
1.4807 +
1.4808 +// Helper function; generates code for the slow case.
1.4809 +// We make a call to a runtime method which emulates the native method,
1.4810 +// but without the native wrapper overhead.
1.4811 +void
1.4812 +LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
1.4813 + Node* src, Node* src_offset,
1.4814 + Node* dest, Node* dest_offset,
1.4815 + Node* copy_length,
1.4816 + int nargs) {
1.4817 + _sp += nargs; // any deopt will start just before call to enclosing method
1.4818 + Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
1.4819 + OptoRuntime::slow_arraycopy_Type(),
1.4820 + OptoRuntime::slow_arraycopy_Java(),
1.4821 + "slow_arraycopy", adr_type,
1.4822 + src, src_offset, dest, dest_offset,
1.4823 + copy_length);
1.4824 + _sp -= nargs;
1.4825 +
1.4826 + // Handle exceptions thrown by this fellow:
1.4827 + make_slow_call_ex(call, env()->Throwable_klass(), false);
1.4828 +}
1.4829 +
1.4830 +// Helper function; generates code for cases requiring runtime checks.
1.4831 +Node*
1.4832 +LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
1.4833 + Node* dest_elem_klass,
1.4834 + Node* src, Node* src_offset,
1.4835 + Node* dest, Node* dest_offset,
1.4836 + Node* copy_length,
1.4837 + int nargs) {
1.4838 + if (stopped()) return NULL;
1.4839 +
1.4840 + address copyfunc_addr = StubRoutines::checkcast_arraycopy();
1.4841 + if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1.4842 + return NULL;
1.4843 + }
1.4844 +
1.4845 + // Pick out the parameters required to perform a store-check
1.4846 + // for the target array. This is an optimistic check. It will
1.4847 + // look in each non-null element's class, at the desired klass's
1.4848 + // super_check_offset, for the desired klass.
1.4849 + int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
1.4850 + Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1.4851 + Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
1.4852 + Node* check_offset = _gvn.transform(n3);
1.4853 + Node* check_value = dest_elem_klass;
1.4854 +
1.4855 + Node* src_start = array_element_address(src, src_offset, T_OBJECT);
1.4856 + Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1.4857 +
1.4858 + // (We know the arrays are never conjoint, because their types differ.)
1.4859 + Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1.4860 + OptoRuntime::checkcast_arraycopy_Type(),
1.4861 + copyfunc_addr, "checkcast_arraycopy", adr_type,
1.4862 + // five arguments, of which two are
1.4863 + // intptr_t (jlong in LP64)
1.4864 + src_start, dest_start,
1.4865 + copy_length XTOP,
1.4866 + check_offset XTOP,
1.4867 + check_value);
1.4868 +
1.4869 + return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1.4870 +}
1.4871 +
1.4872 +
1.4873 +// Helper function; generates code for cases requiring runtime checks.
1.4874 +Node*
1.4875 +LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
1.4876 + Node* src, Node* src_offset,
1.4877 + Node* dest, Node* dest_offset,
1.4878 + Node* copy_length,
1.4879 + int nargs) {
1.4880 + if (stopped()) return NULL;
1.4881 +
1.4882 + address copyfunc_addr = StubRoutines::generic_arraycopy();
1.4883 + if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1.4884 + return NULL;
1.4885 + }
1.4886 +
1.4887 + Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1.4888 + OptoRuntime::generic_arraycopy_Type(),
1.4889 + copyfunc_addr, "generic_arraycopy", adr_type,
1.4890 + src, src_offset, dest, dest_offset, copy_length);
1.4891 +
1.4892 + return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1.4893 +}
1.4894 +
1.4895 +// Helper function; generates the fast out-of-line call to an arraycopy stub.
1.4896 +void
1.4897 +LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
1.4898 + BasicType basic_elem_type,
1.4899 + bool disjoint_bases,
1.4900 + Node* src, Node* src_offset,
1.4901 + Node* dest, Node* dest_offset,
1.4902 + Node* copy_length) {
1.4903 + if (stopped()) return; // nothing to do
1.4904 +
1.4905 + Node* src_start = src;
1.4906 + Node* dest_start = dest;
1.4907 + if (src_offset != NULL || dest_offset != NULL) {
1.4908 + assert(src_offset != NULL && dest_offset != NULL, "");
1.4909 + src_start = array_element_address(src, src_offset, basic_elem_type);
1.4910 + dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1.4911 + }
1.4912 +
1.4913 + // Figure out which arraycopy runtime method to call.
1.4914 + const char* copyfunc_name = "arraycopy";
1.4915 + address copyfunc_addr =
1.4916 + basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1.4917 + disjoint_bases, copyfunc_name);
1.4918 +
1.4919 + // Call it. Note that the count_ix value is not scaled to a byte-size.
1.4920 + make_runtime_call(RC_LEAF|RC_NO_FP,
1.4921 + OptoRuntime::fast_arraycopy_Type(),
1.4922 + copyfunc_addr, copyfunc_name, adr_type,
1.4923 + src_start, dest_start, copy_length XTOP);
1.4924 +}
