Tue, 02 Sep 2014 12:48:45 -0700
8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
Summary: Add new C2 intrinsic for BigInteger::multiplyToLen() on x86 in 64-bit VM.
Reviewed-by: roland
duke@435 | 1 | /* |
mikael@6198 | 2 | * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
stefank@2314 | 25 | #include "precompiled.hpp" |
stefank@2314 | 26 | #include "compiler/compileLog.hpp" |
stefank@2314 | 27 | #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" |
stefank@2314 | 28 | #include "gc_implementation/g1/heapRegion.hpp" |
stefank@2314 | 29 | #include "gc_interface/collectedHeap.hpp" |
stefank@2314 | 30 | #include "memory/barrierSet.hpp" |
stefank@2314 | 31 | #include "memory/cardTableModRefBS.hpp" |
stefank@2314 | 32 | #include "opto/addnode.hpp" |
stefank@2314 | 33 | #include "opto/graphKit.hpp" |
stefank@2314 | 34 | #include "opto/idealKit.hpp" |
stefank@2314 | 35 | #include "opto/locknode.hpp" |
stefank@2314 | 36 | #include "opto/machnode.hpp" |
stefank@2314 | 37 | #include "opto/parse.hpp" |
stefank@2314 | 38 | #include "opto/rootnode.hpp" |
stefank@2314 | 39 | #include "opto/runtime.hpp" |
stefank@2314 | 40 | #include "runtime/deoptimization.hpp" |
stefank@2314 | 41 | #include "runtime/sharedRuntime.hpp" |
duke@435 | 42 | |
duke@435 | 43 | //----------------------------GraphKit----------------------------------------- |
duke@435 | 44 | // Main utility constructor. |
duke@435 | 45 | GraphKit::GraphKit(JVMState* jvms) |
duke@435 | 46 | : Phase(Phase::Parser), |
duke@435 | 47 | _env(C->env()), |
duke@435 | 48 | _gvn(*C->initial_gvn()) |
duke@435 | 49 | { |
duke@435 | 50 | _exceptions = jvms->map()->next_exception(); |
duke@435 | 51 | if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); |
duke@435 | 52 | set_jvms(jvms); |
duke@435 | 53 | } |
duke@435 | 54 | |
duke@435 | 55 | // Private constructor for parser. |
duke@435 | 56 | GraphKit::GraphKit() |
duke@435 | 57 | : Phase(Phase::Parser), |
duke@435 | 58 | _env(C->env()), |
duke@435 | 59 | _gvn(*C->initial_gvn()) |
duke@435 | 60 | { |
duke@435 | 61 | _exceptions = NULL; |
duke@435 | 62 | set_map(NULL); |
duke@435 | 63 | debug_only(_sp = -99); |
duke@435 | 64 | debug_only(set_bci(-99)); |
duke@435 | 65 | } |
duke@435 | 66 | |
duke@435 | 67 | |
duke@435 | 68 | |
duke@435 | 69 | //---------------------------clean_stack--------------------------------------- |
duke@435 | 70 | // Clear away rubbish from the stack area of the JVM state. |
duke@435 | 71 | // This destroys any arguments that may be waiting on the stack. |
duke@435 | 72 | void GraphKit::clean_stack(int from_sp) { |
duke@435 | 73 | SafePointNode* map = this->map(); |
duke@435 | 74 | JVMState* jvms = this->jvms(); |
duke@435 | 75 | int stk_size = jvms->stk_size(); |
duke@435 | 76 | int stkoff = jvms->stkoff(); |
duke@435 | 77 | Node* top = this->top(); |
duke@435 | 78 | for (int i = from_sp; i < stk_size; i++) { |
duke@435 | 79 | if (map->in(stkoff + i) != top) { |
duke@435 | 80 | map->set_req(stkoff + i, top); |
duke@435 | 81 | } |
duke@435 | 82 | } |
duke@435 | 83 | } |
duke@435 | 84 | |
duke@435 | 85 | |
duke@435 | 86 | //--------------------------------sync_jvms----------------------------------- |
duke@435 | 87 | // Make sure our current jvms agrees with our parse state. |
duke@435 | 88 | JVMState* GraphKit::sync_jvms() const { |
duke@435 | 89 | JVMState* jvms = this->jvms(); |
duke@435 | 90 | jvms->set_bci(bci()); // Record the new bci in the JVMState |
duke@435 | 91 | jvms->set_sp(sp()); // Record the new sp in the JVMState |
duke@435 | 92 | assert(jvms_in_sync(), "jvms is now in sync"); |
duke@435 | 93 | return jvms; |
duke@435 | 94 | } |
duke@435 | 95 | |
twisti@4313 | 96 | //--------------------------------sync_jvms_for_reexecute--------------------- |
twisti@4313 | 97 | // Make sure our current jvms agrees with our parse state. This version |
twisti@4313 | 98 | // uses the reexecute_sp for reexecuting bytecodes. |
twisti@4313 | 99 | JVMState* GraphKit::sync_jvms_for_reexecute() { |
twisti@4313 | 100 | JVMState* jvms = this->jvms(); |
twisti@4313 | 101 | jvms->set_bci(bci()); // Record the new bci in the JVMState |
twisti@4313 | 102 | jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState |
twisti@4313 | 103 | return jvms; |
twisti@4313 | 104 | } |
twisti@4313 | 105 | |
duke@435 | 106 | #ifdef ASSERT |
duke@435 | 107 | bool GraphKit::jvms_in_sync() const { |
duke@435 | 108 | Parse* parse = is_Parse(); |
duke@435 | 109 | if (parse == NULL) { |
duke@435 | 110 | if (bci() != jvms()->bci()) return false; |
duke@435 | 111 | if (sp() != (int)jvms()->sp()) return false; |
duke@435 | 112 | return true; |
duke@435 | 113 | } |
duke@435 | 114 | if (jvms()->method() != parse->method()) return false; |
duke@435 | 115 | if (jvms()->bci() != parse->bci()) return false; |
duke@435 | 116 | int jvms_sp = jvms()->sp(); |
duke@435 | 117 | if (jvms_sp != parse->sp()) return false; |
duke@435 | 118 | int jvms_depth = jvms()->depth(); |
duke@435 | 119 | if (jvms_depth != parse->depth()) return false; |
duke@435 | 120 | return true; |
duke@435 | 121 | } |
duke@435 | 122 | |
duke@435 | 123 | // Local helper checks for special internal merge points |
duke@435 | 124 | // used to accumulate and merge exception states. |
duke@435 | 125 | // They are marked by the region's in(0) edge being the map itself. |
duke@435 | 126 | // Such merge points must never "escape" into the parser at large, |
duke@435 | 127 | // until they have been handed to gvn.transform. |
duke@435 | 128 | static bool is_hidden_merge(Node* reg) { |
duke@435 | 129 | if (reg == NULL) return false; |
duke@435 | 130 | if (reg->is_Phi()) { |
duke@435 | 131 | reg = reg->in(0); |
duke@435 | 132 | if (reg == NULL) return false; |
duke@435 | 133 | } |
duke@435 | 134 | return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); |
duke@435 | 135 | } |
duke@435 | 136 | |
duke@435 | 137 | void GraphKit::verify_map() const { |
duke@435 | 138 | if (map() == NULL) return; // null map is OK |
duke@435 | 139 | assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); |
duke@435 | 140 | assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); |
duke@435 | 141 | assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); |
duke@435 | 142 | } |
duke@435 | 143 | |
duke@435 | 144 | void GraphKit::verify_exception_state(SafePointNode* ex_map) { |
duke@435 | 145 | assert(ex_map->next_exception() == NULL, "not already part of a chain"); |
duke@435 | 146 | assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); |
duke@435 | 147 | } |
duke@435 | 148 | #endif |
duke@435 | 149 | |
duke@435 | 150 | //---------------------------stop_and_kill_map--------------------------------- |
duke@435 | 151 | // Set _map to NULL, signalling a stop to further bytecode execution. |
duke@435 | 152 | // First smash the current map's control to a constant, to mark it dead. |
duke@435 | 153 | void GraphKit::stop_and_kill_map() { |
duke@435 | 154 | SafePointNode* dead_map = stop(); |
duke@435 | 155 | if (dead_map != NULL) { |
bharadwaj@4315 | 156 | dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. |
duke@435 | 157 | assert(dead_map->is_killed(), "must be so marked"); |
duke@435 | 158 | } |
duke@435 | 159 | } |
duke@435 | 160 | |
duke@435 | 161 | |
duke@435 | 162 | //--------------------------------stopped-------------------------------------- |
duke@435 | 163 | // Tell if _map is NULL, or control is top. |
duke@435 | 164 | bool GraphKit::stopped() { |
duke@435 | 165 | if (map() == NULL) return true; |
duke@435 | 166 | else if (control() == top()) return true; |
duke@435 | 167 | else return false; |
duke@435 | 168 | } |
duke@435 | 169 | |
duke@435 | 170 | |
duke@435 | 171 | //-----------------------------has_ex_handler---------------------------------- |
duke@435 | 172 | // Tell if this method or any caller method has exception handlers. |
duke@435 | 173 | bool GraphKit::has_ex_handler() { |
duke@435 | 174 | for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { |
duke@435 | 175 | if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { |
duke@435 | 176 | return true; |
duke@435 | 177 | } |
duke@435 | 178 | } |
duke@435 | 179 | return false; |
duke@435 | 180 | } |
duke@435 | 181 | |
duke@435 | 182 | //------------------------------save_ex_oop------------------------------------ |
duke@435 | 183 | // Save an exception without blowing stack contents or other JVM state. |
duke@435 | 184 | void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { |
duke@435 | 185 | assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); |
duke@435 | 186 | ex_map->add_req(ex_oop); |
duke@435 | 187 | debug_only(verify_exception_state(ex_map)); |
duke@435 | 188 | } |
duke@435 | 189 | |
duke@435 | 190 | inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { |
duke@435 | 191 | assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); |
duke@435 | 192 | Node* ex_oop = ex_map->in(ex_map->req()-1); |
duke@435 | 193 | if (clear_it) ex_map->del_req(ex_map->req()-1); |
duke@435 | 194 | return ex_oop; |
duke@435 | 195 | } |
duke@435 | 196 | |
duke@435 | 197 | //-----------------------------saved_ex_oop------------------------------------ |
duke@435 | 198 | // Recover a saved exception from its map. |
duke@435 | 199 | Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { |
duke@435 | 200 | return common_saved_ex_oop(ex_map, false); |
duke@435 | 201 | } |
duke@435 | 202 | |
duke@435 | 203 | //--------------------------clear_saved_ex_oop--------------------------------- |
duke@435 | 204 | // Erase a previously saved exception from its map. |
duke@435 | 205 | Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { |
duke@435 | 206 | return common_saved_ex_oop(ex_map, true); |
duke@435 | 207 | } |
duke@435 | 208 | |
duke@435 | 209 | #ifdef ASSERT |
duke@435 | 210 | //---------------------------has_saved_ex_oop---------------------------------- |
duke@435 | 211 | // Erase a previously saved exception from its map. |
duke@435 | 212 | bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { |
duke@435 | 213 | return ex_map->req() == ex_map->jvms()->endoff()+1; |
duke@435 | 214 | } |
duke@435 | 215 | #endif |
duke@435 | 216 | |
duke@435 | 217 | //-------------------------make_exception_state-------------------------------- |
duke@435 | 218 | // Turn the current JVM state into an exception state, appending the ex_oop. |
duke@435 | 219 | SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { |
duke@435 | 220 | sync_jvms(); |
duke@435 | 221 | SafePointNode* ex_map = stop(); // do not manipulate this map any more |
duke@435 | 222 | set_saved_ex_oop(ex_map, ex_oop); |
duke@435 | 223 | return ex_map; |
duke@435 | 224 | } |
duke@435 | 225 | |
duke@435 | 226 | |
duke@435 | 227 | //--------------------------add_exception_state-------------------------------- |
duke@435 | 228 | // Add an exception to my list of exceptions. |
duke@435 | 229 | void GraphKit::add_exception_state(SafePointNode* ex_map) { |
duke@435 | 230 | if (ex_map == NULL || ex_map->control() == top()) { |
duke@435 | 231 | return; |
duke@435 | 232 | } |
duke@435 | 233 | #ifdef ASSERT |
duke@435 | 234 | verify_exception_state(ex_map); |
duke@435 | 235 | if (has_exceptions()) { |
duke@435 | 236 | assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); |
duke@435 | 237 | } |
duke@435 | 238 | #endif |
duke@435 | 239 | |
duke@435 | 240 | // If there is already an exception of exactly this type, merge with it. |
duke@435 | 241 | // In particular, null-checks and other low-level exceptions common up here. |
duke@435 | 242 | Node* ex_oop = saved_ex_oop(ex_map); |
duke@435 | 243 | const Type* ex_type = _gvn.type(ex_oop); |
duke@435 | 244 | if (ex_oop == top()) { |
duke@435 | 245 | // No action needed. |
duke@435 | 246 | return; |
duke@435 | 247 | } |
duke@435 | 248 | assert(ex_type->isa_instptr(), "exception must be an instance"); |
duke@435 | 249 | for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { |
duke@435 | 250 | const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); |
duke@435 | 251 | // We check sp also because call bytecodes can generate exceptions |
duke@435 | 252 | // both before and after arguments are popped! |
duke@435 | 253 | if (ex_type2 == ex_type |
duke@435 | 254 | && e2->_jvms->sp() == ex_map->_jvms->sp()) { |
duke@435 | 255 | combine_exception_states(ex_map, e2); |
duke@435 | 256 | return; |
duke@435 | 257 | } |
duke@435 | 258 | } |
duke@435 | 259 | |
duke@435 | 260 | // No pre-existing exception of the same type. Chain it on the list. |
duke@435 | 261 | push_exception_state(ex_map); |
duke@435 | 262 | } |
duke@435 | 263 | |
duke@435 | 264 | //-----------------------add_exception_states_from----------------------------- |
duke@435 | 265 | void GraphKit::add_exception_states_from(JVMState* jvms) { |
duke@435 | 266 | SafePointNode* ex_map = jvms->map()->next_exception(); |
duke@435 | 267 | if (ex_map != NULL) { |
duke@435 | 268 | jvms->map()->set_next_exception(NULL); |
duke@435 | 269 | for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { |
duke@435 | 270 | next_map = ex_map->next_exception(); |
duke@435 | 271 | ex_map->set_next_exception(NULL); |
duke@435 | 272 | add_exception_state(ex_map); |
duke@435 | 273 | } |
duke@435 | 274 | } |
duke@435 | 275 | } |
duke@435 | 276 | |
duke@435 | 277 | //-----------------------transfer_exceptions_into_jvms------------------------- |
duke@435 | 278 | JVMState* GraphKit::transfer_exceptions_into_jvms() { |
duke@435 | 279 | if (map() == NULL) { |
duke@435 | 280 | // We need a JVMS to carry the exceptions, but the map has gone away. |
duke@435 | 281 | // Create a scratch JVMS, cloned from any of the exception states... |
duke@435 | 282 | if (has_exceptions()) { |
duke@435 | 283 | _map = _exceptions; |
duke@435 | 284 | _map = clone_map(); |
duke@435 | 285 | _map->set_next_exception(NULL); |
duke@435 | 286 | clear_saved_ex_oop(_map); |
duke@435 | 287 | debug_only(verify_map()); |
duke@435 | 288 | } else { |
duke@435 | 289 | // ...or created from scratch |
duke@435 | 290 | JVMState* jvms = new (C) JVMState(_method, NULL); |
duke@435 | 291 | jvms->set_bci(_bci); |
duke@435 | 292 | jvms->set_sp(_sp); |
kvn@4115 | 293 | jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms)); |
duke@435 | 294 | set_jvms(jvms); |
duke@435 | 295 | for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); |
duke@435 | 296 | set_all_memory(top()); |
duke@435 | 297 | while (map()->req() < jvms->endoff()) map()->add_req(top()); |
duke@435 | 298 | } |
duke@435 | 299 | // (This is a kludge, in case you didn't notice.) |
duke@435 | 300 | set_control(top()); |
duke@435 | 301 | } |
duke@435 | 302 | JVMState* jvms = sync_jvms(); |
duke@435 | 303 | assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); |
duke@435 | 304 | jvms->map()->set_next_exception(_exceptions); |
duke@435 | 305 | _exceptions = NULL; // done with this set of exceptions |
duke@435 | 306 | return jvms; |
duke@435 | 307 | } |
duke@435 | 308 | |
duke@435 | 309 | static inline void add_n_reqs(Node* dstphi, Node* srcphi) { |
duke@435 | 310 | assert(is_hidden_merge(dstphi), "must be a special merge node"); |
duke@435 | 311 | assert(is_hidden_merge(srcphi), "must be a special merge node"); |
duke@435 | 312 | uint limit = srcphi->req(); |
duke@435 | 313 | for (uint i = PhiNode::Input; i < limit; i++) { |
duke@435 | 314 | dstphi->add_req(srcphi->in(i)); |
duke@435 | 315 | } |
duke@435 | 316 | } |
duke@435 | 317 | static inline void add_one_req(Node* dstphi, Node* src) { |
duke@435 | 318 | assert(is_hidden_merge(dstphi), "must be a special merge node"); |
duke@435 | 319 | assert(!is_hidden_merge(src), "must not be a special merge node"); |
duke@435 | 320 | dstphi->add_req(src); |
duke@435 | 321 | } |
duke@435 | 322 | |
duke@435 | 323 | //-----------------------combine_exception_states------------------------------ |
duke@435 | 324 | // This helper function combines exception states by building phis on a |
duke@435 | 325 | // specially marked state-merging region. These regions and phis are |
duke@435 | 326 | // untransformed, and can build up gradually. The region is marked by |
duke@435 | 327 | // having a control input of its exception map, rather than NULL. Such |
duke@435 | 328 | // regions do not appear except in this function, and in use_exception_state. |
duke@435 | 329 | void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { |
duke@435 | 330 | if (failing()) return; // dying anyway... |
duke@435 | 331 | JVMState* ex_jvms = ex_map->_jvms; |
duke@435 | 332 | assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); |
duke@435 | 333 | assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); |
duke@435 | 334 | assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); |
duke@435 | 335 | assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); |
kvn@5110 | 336 | assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); |
duke@435 | 337 | assert(ex_map->req() == phi_map->req(), "matching maps"); |
duke@435 | 338 | uint tos = ex_jvms->stkoff() + ex_jvms->sp(); |
duke@435 | 339 | Node* hidden_merge_mark = root(); |
duke@435 | 340 | Node* region = phi_map->control(); |
duke@435 | 341 | MergeMemNode* phi_mem = phi_map->merged_memory(); |
duke@435 | 342 | MergeMemNode* ex_mem = ex_map->merged_memory(); |
duke@435 | 343 | if (region->in(0) != hidden_merge_mark) { |
duke@435 | 344 | // The control input is not (yet) a specially-marked region in phi_map. |
duke@435 | 345 | // Make it so, and build some phis. |
kvn@4115 | 346 | region = new (C) RegionNode(2); |
duke@435 | 347 | _gvn.set_type(region, Type::CONTROL); |
duke@435 | 348 | region->set_req(0, hidden_merge_mark); // marks an internal ex-state |
duke@435 | 349 | region->init_req(1, phi_map->control()); |
duke@435 | 350 | phi_map->set_control(region); |
duke@435 | 351 | Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); |
duke@435 | 352 | record_for_igvn(io_phi); |
duke@435 | 353 | _gvn.set_type(io_phi, Type::ABIO); |
duke@435 | 354 | phi_map->set_i_o(io_phi); |
duke@435 | 355 | for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { |
duke@435 | 356 | Node* m = mms.memory(); |
duke@435 | 357 | Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); |
duke@435 | 358 | record_for_igvn(m_phi); |
duke@435 | 359 | _gvn.set_type(m_phi, Type::MEMORY); |
duke@435 | 360 | mms.set_memory(m_phi); |
duke@435 | 361 | } |
duke@435 | 362 | } |
duke@435 | 363 | |
duke@435 | 364 | // Either or both of phi_map and ex_map might already be converted into phis. |
duke@435 | 365 | Node* ex_control = ex_map->control(); |
duke@435 | 366 | // if there is special marking on ex_map also, we add multiple edges from src |
duke@435 | 367 | bool add_multiple = (ex_control->in(0) == hidden_merge_mark); |
duke@435 | 368 | // how wide was the destination phi_map, originally? |
duke@435 | 369 | uint orig_width = region->req(); |
duke@435 | 370 | |
duke@435 | 371 | if (add_multiple) { |
duke@435 | 372 | add_n_reqs(region, ex_control); |
duke@435 | 373 | add_n_reqs(phi_map->i_o(), ex_map->i_o()); |
duke@435 | 374 | } else { |
duke@435 | 375 | // ex_map has no merges, so we just add single edges everywhere |
duke@435 | 376 | add_one_req(region, ex_control); |
duke@435 | 377 | add_one_req(phi_map->i_o(), ex_map->i_o()); |
duke@435 | 378 | } |
duke@435 | 379 | for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { |
duke@435 | 380 | if (mms.is_empty()) { |
duke@435 | 381 | // get a copy of the base memory, and patch some inputs into it |
duke@435 | 382 | const TypePtr* adr_type = mms.adr_type(C); |
duke@435 | 383 | Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); |
duke@435 | 384 | assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); |
duke@435 | 385 | mms.set_memory(phi); |
duke@435 | 386 | // Prepare to append interesting stuff onto the newly sliced phi: |
duke@435 | 387 | while (phi->req() > orig_width) phi->del_req(phi->req()-1); |
duke@435 | 388 | } |
duke@435 | 389 | // Append stuff from ex_map: |
duke@435 | 390 | if (add_multiple) { |
duke@435 | 391 | add_n_reqs(mms.memory(), mms.memory2()); |
duke@435 | 392 | } else { |
duke@435 | 393 | add_one_req(mms.memory(), mms.memory2()); |
duke@435 | 394 | } |
duke@435 | 395 | } |
duke@435 | 396 | uint limit = ex_map->req(); |
duke@435 | 397 | for (uint i = TypeFunc::Parms; i < limit; i++) { |
duke@435 | 398 | // Skip everything in the JVMS after tos. (The ex_oop follows.) |
duke@435 | 399 | if (i == tos) i = ex_jvms->monoff(); |
duke@435 | 400 | Node* src = ex_map->in(i); |
duke@435 | 401 | Node* dst = phi_map->in(i); |
duke@435 | 402 | if (src != dst) { |
duke@435 | 403 | PhiNode* phi; |
duke@435 | 404 | if (dst->in(0) != region) { |
duke@435 | 405 | dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); |
duke@435 | 406 | record_for_igvn(phi); |
duke@435 | 407 | _gvn.set_type(phi, phi->type()); |
duke@435 | 408 | phi_map->set_req(i, dst); |
duke@435 | 409 | // Prepare to append interesting stuff onto the new phi: |
duke@435 | 410 | while (dst->req() > orig_width) dst->del_req(dst->req()-1); |
duke@435 | 411 | } else { |
duke@435 | 412 | assert(dst->is_Phi(), "nobody else uses a hidden region"); |
kvn@5110 | 413 | phi = dst->as_Phi(); |
duke@435 | 414 | } |
duke@435 | 415 | if (add_multiple && src->in(0) == ex_control) { |
duke@435 | 416 | // Both are phis. |
duke@435 | 417 | add_n_reqs(dst, src); |
duke@435 | 418 | } else { |
duke@435 | 419 | while (dst->req() < region->req()) add_one_req(dst, src); |
duke@435 | 420 | } |
duke@435 | 421 | const Type* srctype = _gvn.type(src); |
duke@435 | 422 | if (phi->type() != srctype) { |
roland@6313 | 423 | const Type* dsttype = phi->type()->meet_speculative(srctype); |
duke@435 | 424 | if (phi->type() != dsttype) { |
duke@435 | 425 | phi->set_type(dsttype); |
duke@435 | 426 | _gvn.set_type(phi, dsttype); |
duke@435 | 427 | } |
duke@435 | 428 | } |
duke@435 | 429 | } |
duke@435 | 430 | } |
roland@7041 | 431 | phi_map->merge_replaced_nodes_with(ex_map); |
duke@435 | 432 | } |
duke@435 | 433 | |
duke@435 | 434 | //--------------------------use_exception_state-------------------------------- |
duke@435 | 435 | Node* GraphKit::use_exception_state(SafePointNode* phi_map) { |
duke@435 | 436 | if (failing()) { stop(); return top(); } |
duke@435 | 437 | Node* region = phi_map->control(); |
duke@435 | 438 | Node* hidden_merge_mark = root(); |
duke@435 | 439 | assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); |
duke@435 | 440 | Node* ex_oop = clear_saved_ex_oop(phi_map); |
duke@435 | 441 | if (region->in(0) == hidden_merge_mark) { |
duke@435 | 442 | // Special marking for internal ex-states. Process the phis now. |
duke@435 | 443 | region->set_req(0, region); // now it's an ordinary region |
duke@435 | 444 | set_jvms(phi_map->jvms()); // ...so now we can use it as a map |
duke@435 | 445 | // Note: Setting the jvms also sets the bci and sp. |
duke@435 | 446 | set_control(_gvn.transform(region)); |
duke@435 | 447 | uint tos = jvms()->stkoff() + sp(); |
duke@435 | 448 | for (uint i = 1; i < tos; i++) { |
duke@435 | 449 | Node* x = phi_map->in(i); |
duke@435 | 450 | if (x->in(0) == region) { |
duke@435 | 451 | assert(x->is_Phi(), "expected a special phi"); |
duke@435 | 452 | phi_map->set_req(i, _gvn.transform(x)); |
duke@435 | 453 | } |
duke@435 | 454 | } |
duke@435 | 455 | for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { |
duke@435 | 456 | Node* x = mms.memory(); |
duke@435 | 457 | if (x->in(0) == region) { |
duke@435 | 458 | assert(x->is_Phi(), "nobody else uses a hidden region"); |
duke@435 | 459 | mms.set_memory(_gvn.transform(x)); |
duke@435 | 460 | } |
duke@435 | 461 | } |
duke@435 | 462 | if (ex_oop->in(0) == region) { |
duke@435 | 463 | assert(ex_oop->is_Phi(), "expected a special phi"); |
duke@435 | 464 | ex_oop = _gvn.transform(ex_oop); |
duke@435 | 465 | } |
duke@435 | 466 | } else { |
duke@435 | 467 | set_jvms(phi_map->jvms()); |
duke@435 | 468 | } |
duke@435 | 469 | |
duke@435 | 470 | assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); |
duke@435 | 471 | assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); |
duke@435 | 472 | return ex_oop; |
duke@435 | 473 | } |
duke@435 | 474 | |
duke@435 | 475 | //---------------------------------java_bc------------------------------------- |
duke@435 | 476 | Bytecodes::Code GraphKit::java_bc() const { |
duke@435 | 477 | ciMethod* method = this->method(); |
duke@435 | 478 | int bci = this->bci(); |
duke@435 | 479 | if (method != NULL && bci != InvocationEntryBci) |
duke@435 | 480 | return method->java_code_at_bci(bci); |
duke@435 | 481 | else |
duke@435 | 482 | return Bytecodes::_illegal; |
duke@435 | 483 | } |
duke@435 | 484 | |
dcubed@1648 | 485 | void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, |
dcubed@1648 | 486 | bool must_throw) { |
dcubed@1648 | 487 | // if the exception capability is set, then we will generate code |
dcubed@1648 | 488 | // to check the JavaThread.should_post_on_exceptions flag to see |
dcubed@1648 | 489 | // if we actually need to report exception events (for this |
dcubed@1648 | 490 | // thread). If we don't need to report exception events, we will |
dcubed@1648 | 491 | // take the normal fast path provided by add_exception_events. If |
dcubed@1648 | 492 | // exception event reporting is enabled for this thread, we will |
dcubed@1648 | 493 | // take the uncommon_trap in the BuildCutout below. |
dcubed@1648 | 494 | |
dcubed@1648 | 495 | // first must access the should_post_on_exceptions_flag in this thread's JavaThread |
kvn@4115 | 496 | Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); |
dcubed@1648 | 497 | Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); |
goetz@6479 | 498 | Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); |
dcubed@1648 | 499 | |
dcubed@1648 | 500 | // Test the should_post_on_exceptions_flag vs. 0 |
kvn@4115 | 501 | Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); |
kvn@4115 | 502 | Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); |
dcubed@1648 | 503 | |
dcubed@1648 | 504 | // Branch to slow_path if should_post_on_exceptions_flag was true |
dcubed@1648 | 505 | { BuildCutout unless(this, tst, PROB_MAX); |
dcubed@1648 | 506 | // Do not try anything fancy if we're notifying the VM on every throw. |
dcubed@1648 | 507 | // Cf. case Bytecodes::_athrow in parse2.cpp. |
dcubed@1648 | 508 | uncommon_trap(reason, Deoptimization::Action_none, |
dcubed@1648 | 509 | (ciKlass*)NULL, (char*)NULL, must_throw); |
dcubed@1648 | 510 | } |
dcubed@1648 | 511 | |
dcubed@1648 | 512 | } |
dcubed@1648 | 513 | |
duke@435 | 514 | //------------------------------builtin_throw---------------------------------- |
duke@435 | 515 | void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { |
duke@435 | 516 | bool must_throw = true; |
duke@435 | 517 | |
dcubed@1648 | 518 | if (env()->jvmti_can_post_on_exceptions()) { |
dcubed@1648 | 519 | // check if we must post exception events, take uncommon trap if so |
dcubed@1648 | 520 | uncommon_trap_if_should_post_on_exceptions(reason, must_throw); |
dcubed@1648 | 521 | // here if should_post_on_exceptions is false |
dcubed@1648 | 522 | // continue on with the normal codegen |
duke@435 | 523 | } |
duke@435 | 524 | |
duke@435 | 525 | // If this particular condition has not yet happened at this |
duke@435 | 526 | // bytecode, then use the uncommon trap mechanism, and allow for |
duke@435 | 527 | // a future recompilation if several traps occur here. |
duke@435 | 528 | // If the throw is hot, try to use a more complicated inline mechanism |
duke@435 | 529 | // which keeps execution inside the compiled code. |
duke@435 | 530 | bool treat_throw_as_hot = false; |
duke@435 | 531 | ciMethodData* md = method()->method_data(); |
duke@435 | 532 | |
duke@435 | 533 | if (ProfileTraps) { |
duke@435 | 534 | if (too_many_traps(reason)) { |
duke@435 | 535 | treat_throw_as_hot = true; |
duke@435 | 536 | } |
duke@435 | 537 | // (If there is no MDO at all, assume it is early in |
duke@435 | 538 | // execution, and that any deopts are part of the |
duke@435 | 539 | // startup transient, and don't need to be remembered.) |
duke@435 | 540 | |
duke@435 | 541 | // Also, if there is a local exception handler, treat all throws |
duke@435 | 542 | // as hot if there has been at least one in this method. |
duke@435 | 543 | if (C->trap_count(reason) != 0 |
duke@435 | 544 | && method()->method_data()->trap_count(reason) != 0 |
duke@435 | 545 | && has_ex_handler()) { |
duke@435 | 546 | treat_throw_as_hot = true; |
duke@435 | 547 | } |
duke@435 | 548 | } |
duke@435 | 549 | |
duke@435 | 550 | // If this throw happens frequently, an uncommon trap might cause |
duke@435 | 551 | // a performance pothole. If there is a local exception handler, |
duke@435 | 552 | // and if this particular bytecode appears to be deoptimizing often, |
duke@435 | 553 | // let us handle the throw inline, with a preconstructed instance. |
duke@435 | 554 | // Note: If the deopt count has blown up, the uncommon trap |
duke@435 | 555 | // runtime is going to flush this nmethod, not matter what. |
duke@435 | 556 | if (treat_throw_as_hot |
duke@435 | 557 | && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { |
duke@435 | 558 | // If the throw is local, we use a pre-existing instance and |
duke@435 | 559 | // punt on the backtrace. This would lead to a missing backtrace |
duke@435 | 560 | // (a repeat of 4292742) if the backtrace object is ever asked |
duke@435 | 561 | // for its backtrace. |
duke@435 | 562 | // Fixing this remaining case of 4292742 requires some flavor of |
duke@435 | 563 | // escape analysis. Leave that for the future. |
duke@435 | 564 | ciInstance* ex_obj = NULL; |
duke@435 | 565 | switch (reason) { |
duke@435 | 566 | case Deoptimization::Reason_null_check: |
duke@435 | 567 | ex_obj = env()->NullPointerException_instance(); |
duke@435 | 568 | break; |
duke@435 | 569 | case Deoptimization::Reason_div0_check: |
duke@435 | 570 | ex_obj = env()->ArithmeticException_instance(); |
duke@435 | 571 | break; |
duke@435 | 572 | case Deoptimization::Reason_range_check: |
duke@435 | 573 | ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); |
duke@435 | 574 | break; |
duke@435 | 575 | case Deoptimization::Reason_class_check: |
duke@435 | 576 | if (java_bc() == Bytecodes::_aastore) { |
duke@435 | 577 | ex_obj = env()->ArrayStoreException_instance(); |
duke@435 | 578 | } else { |
duke@435 | 579 | ex_obj = env()->ClassCastException_instance(); |
duke@435 | 580 | } |
duke@435 | 581 | break; |
duke@435 | 582 | } |
duke@435 | 583 | if (failing()) { stop(); return; } // exception allocation might fail |
duke@435 | 584 | if (ex_obj != NULL) { |
duke@435 | 585 | // Cheat with a preallocated exception object. |
duke@435 | 586 | if (C->log() != NULL) |
duke@435 | 587 | C->log()->elem("hot_throw preallocated='1' reason='%s'", |
duke@435 | 588 | Deoptimization::trap_reason_name(reason)); |
duke@435 | 589 | const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); |
kvn@599 | 590 | Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); |
duke@435 | 591 | |
duke@435 | 592 | // Clear the detail message of the preallocated exception object. |
duke@435 | 593 | // Weblogic sometimes mutates the detail message of exceptions |
duke@435 | 594 | // using reflection. |
duke@435 | 595 | int offset = java_lang_Throwable::get_detailMessage_offset(); |
duke@435 | 596 | const TypePtr* adr_typ = ex_con->add_offset(offset); |
duke@435 | 597 | |
duke@435 | 598 | Node *adr = basic_plus_adr(ex_node, ex_node, offset); |
kvn@2307 | 599 | const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); |
goetz@6479 | 600 | // Conservatively release stores of object references. |
goetz@6479 | 601 | Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); |
duke@435 | 602 | |
duke@435 | 603 | add_exception_state(make_exception_state(ex_node)); |
duke@435 | 604 | return; |
duke@435 | 605 | } |
duke@435 | 606 | } |
duke@435 | 607 | |
duke@435 | 608 | // %%% Maybe add entry to OptoRuntime which directly throws the exc.? |
duke@435 | 609 | // It won't be much cheaper than bailing to the interp., since we'll |
duke@435 | 610 | // have to pass up all the debug-info, and the runtime will have to |
duke@435 | 611 | // create the stack trace. |
duke@435 | 612 | |
duke@435 | 613 | // Usual case: Bail to interpreter. |
duke@435 | 614 | // Reserve the right to recompile if we haven't seen anything yet. |
duke@435 | 615 | |
roland@6377 | 616 | assert(!Deoptimization::reason_is_speculate(reason), "unsupported"); |
duke@435 | 617 | Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; |
duke@435 | 618 | if (treat_throw_as_hot |
roland@6377 | 619 | && (method()->method_data()->trap_recompiled_at(bci(), NULL) |
duke@435 | 620 | || C->too_many_traps(reason))) { |
duke@435 | 621 | // We cannot afford to take more traps here. Suffer in the interpreter. |
duke@435 | 622 | if (C->log() != NULL) |
duke@435 | 623 | C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", |
duke@435 | 624 | Deoptimization::trap_reason_name(reason), |
duke@435 | 625 | C->trap_count(reason)); |
duke@435 | 626 | action = Deoptimization::Action_none; |
duke@435 | 627 | } |
duke@435 | 628 | |
duke@435 | 629 | // "must_throw" prunes the JVM state to include only the stack, if there |
duke@435 | 630 | // are no local exception handlers. This should cut down on register |
duke@435 | 631 | // allocation time and code size, by drastically reducing the number |
duke@435 | 632 | // of in-edges on the call to the uncommon trap. |
duke@435 | 633 | |
duke@435 | 634 | uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); |
duke@435 | 635 | } |
duke@435 | 636 | |
duke@435 | 637 | |
duke@435 | 638 | //----------------------------PreserveJVMState--------------------------------- |
duke@435 | 639 | PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { |
duke@435 | 640 | debug_only(kit->verify_map()); |
duke@435 | 641 | _kit = kit; |
duke@435 | 642 | _map = kit->map(); // preserve the map |
duke@435 | 643 | _sp = kit->sp(); |
duke@435 | 644 | kit->set_map(clone_map ? kit->clone_map() : NULL); |
duke@435 | 645 | #ifdef ASSERT |
duke@435 | 646 | _bci = kit->bci(); |
duke@435 | 647 | Parse* parser = kit->is_Parse(); |
never@802 | 648 | int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); |
duke@435 | 649 | _block = block; |
duke@435 | 650 | #endif |
duke@435 | 651 | } |
duke@435 | 652 | PreserveJVMState::~PreserveJVMState() { |
duke@435 | 653 | GraphKit* kit = _kit; |
duke@435 | 654 | #ifdef ASSERT |
duke@435 | 655 | assert(kit->bci() == _bci, "bci must not shift"); |
duke@435 | 656 | Parse* parser = kit->is_Parse(); |
never@802 | 657 | int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); |
duke@435 | 658 | assert(block == _block, "block must not shift"); |
duke@435 | 659 | #endif |
duke@435 | 660 | kit->set_map(_map); |
duke@435 | 661 | kit->set_sp(_sp); |
duke@435 | 662 | } |
duke@435 | 663 | |
duke@435 | 664 | |
duke@435 | 665 | //-----------------------------BuildCutout------------------------------------- |
duke@435 | 666 | BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) |
duke@435 | 667 | : PreserveJVMState(kit) |
duke@435 | 668 | { |
duke@435 | 669 | assert(p->is_Con() || p->is_Bool(), "test must be a bool"); |
duke@435 | 670 | SafePointNode* outer_map = _map; // preserved map is caller's |
duke@435 | 671 | SafePointNode* inner_map = kit->map(); |
duke@435 | 672 | IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); |
kvn@4115 | 673 | outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); |
kvn@4115 | 674 | inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); |
duke@435 | 675 | } |
duke@435 | 676 | BuildCutout::~BuildCutout() { |
duke@435 | 677 | GraphKit* kit = _kit; |
duke@435 | 678 | assert(kit->stopped(), "cutout code must stop, throw, return, etc."); |
duke@435 | 679 | } |
duke@435 | 680 | |
cfang@1335 | 681 | //---------------------------PreserveReexecuteState---------------------------- |
cfang@1335 | 682 | PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { |
kvn@1385 | 683 | assert(!kit->stopped(), "must call stopped() before"); |
cfang@1335 | 684 | _kit = kit; |
cfang@1335 | 685 | _sp = kit->sp(); |
cfang@1335 | 686 | _reexecute = kit->jvms()->_reexecute; |
cfang@1335 | 687 | } |
cfang@1335 | 688 | PreserveReexecuteState::~PreserveReexecuteState() { |
kvn@1385 | 689 | if (_kit->stopped()) return; |
cfang@1335 | 690 | _kit->jvms()->_reexecute = _reexecute; |
cfang@1335 | 691 | _kit->set_sp(_sp); |
cfang@1335 | 692 | } |
duke@435 | 693 | |
duke@435 | 694 | //------------------------------clone_map-------------------------------------- |
duke@435 | 695 | // Implementation of PreserveJVMState |
duke@435 | 696 | // |
duke@435 | 697 | // Only clone_map(...) here. If this function is only used in the |
duke@435 | 698 | // PreserveJVMState class we may want to get rid of this extra |
duke@435 | 699 | // function eventually and do it all there. |
duke@435 | 700 | |
duke@435 | 701 | SafePointNode* GraphKit::clone_map() { |
duke@435 | 702 | if (map() == NULL) return NULL; |
duke@435 | 703 | |
duke@435 | 704 | // Clone the memory edge first |
duke@435 | 705 | Node* mem = MergeMemNode::make(C, map()->memory()); |
duke@435 | 706 | gvn().set_type_bottom(mem); |
duke@435 | 707 | |
duke@435 | 708 | SafePointNode *clonemap = (SafePointNode*)map()->clone(); |
duke@435 | 709 | JVMState* jvms = this->jvms(); |
duke@435 | 710 | JVMState* clonejvms = jvms->clone_shallow(C); |
duke@435 | 711 | clonemap->set_memory(mem); |
duke@435 | 712 | clonemap->set_jvms(clonejvms); |
duke@435 | 713 | clonejvms->set_map(clonemap); |
duke@435 | 714 | record_for_igvn(clonemap); |
duke@435 | 715 | gvn().set_type_bottom(clonemap); |
duke@435 | 716 | return clonemap; |
duke@435 | 717 | } |
duke@435 | 718 | |
duke@435 | 719 | |
duke@435 | 720 | //-----------------------------set_map_clone----------------------------------- |
duke@435 | 721 | void GraphKit::set_map_clone(SafePointNode* m) { |
duke@435 | 722 | _map = m; |
duke@435 | 723 | _map = clone_map(); |
duke@435 | 724 | _map->set_next_exception(NULL); |
duke@435 | 725 | debug_only(verify_map()); |
duke@435 | 726 | } |
duke@435 | 727 | |
duke@435 | 728 | |
duke@435 | 729 | //----------------------------kill_dead_locals--------------------------------- |
duke@435 | 730 | // Detect any locals which are known to be dead, and force them to top. |
duke@435 | 731 | void GraphKit::kill_dead_locals() { |
duke@435 | 732 | // Consult the liveness information for the locals. If any |
duke@435 | 733 | // of them are unused, then they can be replaced by top(). This |
duke@435 | 734 | // should help register allocation time and cut down on the size |
duke@435 | 735 | // of the deoptimization information. |
duke@435 | 736 | |
duke@435 | 737 | // This call is made from many of the bytecode handling |
duke@435 | 738 | // subroutines called from the Big Switch in do_one_bytecode. |
duke@435 | 739 | // Every bytecode which might include a slow path is responsible |
duke@435 | 740 | // for killing its dead locals. The more consistent we |
duke@435 | 741 | // are about killing deads, the fewer useless phis will be |
duke@435 | 742 | // constructed for them at various merge points. |
duke@435 | 743 | |
duke@435 | 744 | // bci can be -1 (InvocationEntryBci). We return the entry |
duke@435 | 745 | // liveness for the method. |
duke@435 | 746 | |
duke@435 | 747 | if (method() == NULL || method()->code_size() == 0) { |
duke@435 | 748 | // We are building a graph for a call to a native method. |
duke@435 | 749 | // All locals are live. |
duke@435 | 750 | return; |
duke@435 | 751 | } |
duke@435 | 752 | |
duke@435 | 753 | ResourceMark rm; |
duke@435 | 754 | |
duke@435 | 755 | // Consult the liveness information for the locals. If any |
duke@435 | 756 | // of them are unused, then they can be replaced by top(). This |
duke@435 | 757 | // should help register allocation time and cut down on the size |
duke@435 | 758 | // of the deoptimization information. |
duke@435 | 759 | MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); |
duke@435 | 760 | |
duke@435 | 761 | int len = (int)live_locals.size(); |
duke@435 | 762 | assert(len <= jvms()->loc_size(), "too many live locals"); |
duke@435 | 763 | for (int local = 0; local < len; local++) { |
duke@435 | 764 | if (!live_locals.at(local)) { |
duke@435 | 765 | set_local(local, top()); |
duke@435 | 766 | } |
duke@435 | 767 | } |
duke@435 | 768 | } |
duke@435 | 769 | |
duke@435 | 770 | #ifdef ASSERT |
duke@435 | 771 | //-------------------------dead_locals_are_killed------------------------------ |
duke@435 | 772 | // Return true if all dead locals are set to top in the map. |
duke@435 | 773 | // Used to assert "clean" debug info at various points. |
duke@435 | 774 | bool GraphKit::dead_locals_are_killed() { |
duke@435 | 775 | if (method() == NULL || method()->code_size() == 0) { |
duke@435 | 776 | // No locals need to be dead, so all is as it should be. |
duke@435 | 777 | return true; |
duke@435 | 778 | } |
duke@435 | 779 | |
duke@435 | 780 | // Make sure somebody called kill_dead_locals upstream. |
duke@435 | 781 | ResourceMark rm; |
duke@435 | 782 | for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { |
duke@435 | 783 | if (jvms->loc_size() == 0) continue; // no locals to consult |
duke@435 | 784 | SafePointNode* map = jvms->map(); |
duke@435 | 785 | ciMethod* method = jvms->method(); |
duke@435 | 786 | int bci = jvms->bci(); |
duke@435 | 787 | if (jvms == this->jvms()) { |
duke@435 | 788 | bci = this->bci(); // it might not yet be synched |
duke@435 | 789 | } |
duke@435 | 790 | MethodLivenessResult live_locals = method->liveness_at_bci(bci); |
duke@435 | 791 | int len = (int)live_locals.size(); |
duke@435 | 792 | if (!live_locals.is_valid() || len == 0) |
duke@435 | 793 | // This method is trivial, or is poisoned by a breakpoint. |
duke@435 | 794 | return true; |
duke@435 | 795 | assert(len == jvms->loc_size(), "live map consistent with locals map"); |
duke@435 | 796 | for (int local = 0; local < len; local++) { |
duke@435 | 797 | if (!live_locals.at(local) && map->local(jvms, local) != top()) { |
duke@435 | 798 | if (PrintMiscellaneous && (Verbose || WizardMode)) { |
duke@435 | 799 | tty->print_cr("Zombie local %d: ", local); |
duke@435 | 800 | jvms->dump(); |
duke@435 | 801 | } |
duke@435 | 802 | return false; |
duke@435 | 803 | } |
duke@435 | 804 | } |
duke@435 | 805 | } |
duke@435 | 806 | return true; |
duke@435 | 807 | } |
duke@435 | 808 | |
duke@435 | 809 | #endif //ASSERT |
duke@435 | 810 | |
cfang@1335 | 811 | // Helper function for enforcing certain bytecodes to reexecute if |
cfang@1335 | 812 | // deoptimization happens |
kvn@1687 | 813 | static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { |
cfang@1335 | 814 | ciMethod* cur_method = jvms->method(); |
cfang@1335 | 815 | int cur_bci = jvms->bci(); |
cfang@1335 | 816 | if (cur_method != NULL && cur_bci != InvocationEntryBci) { |
cfang@1335 | 817 | Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); |
kvn@1687 | 818 | return Interpreter::bytecode_should_reexecute(code) || |
kvn@1687 | 819 | is_anewarray && code == Bytecodes::_multianewarray; |
kvn@1687 | 820 | // Reexecute _multianewarray bytecode which was replaced with |
kvn@1687 | 821 | // sequence of [a]newarray. See Parse::do_multianewarray(). |
kvn@1687 | 822 | // |
kvn@1687 | 823 | // Note: interpreter should not have it set since this optimization |
kvn@1687 | 824 | // is limited by dimensions and guarded by flag so in some cases |
kvn@1687 | 825 | // multianewarray() runtime calls will be generated and |
kvn@1687 | 826 | // the bytecode should not be reexecutes (stack will not be reset). |
cfang@1335 | 827 | } else |
cfang@1335 | 828 | return false; |
cfang@1335 | 829 | } |
cfang@1335 | 830 | |
duke@435 | 831 | // Helper function for adding JVMState and debug information to node |
duke@435 | 832 | void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { |
duke@435 | 833 | // Add the safepoint edges to the call (or other safepoint). |
duke@435 | 834 | |
duke@435 | 835 | // Make sure dead locals are set to top. This |
duke@435 | 836 | // should help register allocation time and cut down on the size |
duke@435 | 837 | // of the deoptimization information. |
duke@435 | 838 | assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); |
duke@435 | 839 | |
duke@435 | 840 | // Walk the inline list to fill in the correct set of JVMState's |
duke@435 | 841 | // Also fill in the associated edges for each JVMState. |
duke@435 | 842 | |
twisti@4313 | 843 | // If the bytecode needs to be reexecuted we need to put |
twisti@4313 | 844 | // the arguments back on the stack. |
twisti@4313 | 845 | const bool should_reexecute = jvms()->should_reexecute(); |
twisti@4313 | 846 | JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); |
twisti@4313 | 847 | |
twisti@4313 | 848 | // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to |
twisti@4313 | 849 | // undefined if the bci is different. This is normal for Parse but it |
twisti@4313 | 850 | // should not happen for LibraryCallKit because only one bci is processed. |
twisti@4313 | 851 | assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), |
twisti@4313 | 852 | "in LibraryCallKit the reexecute bit should not change"); |
duke@435 | 853 | |
duke@435 | 854 | // If we are guaranteed to throw, we can prune everything but the |
duke@435 | 855 | // input to the current bytecode. |
duke@435 | 856 | bool can_prune_locals = false; |
duke@435 | 857 | uint stack_slots_not_pruned = 0; |
duke@435 | 858 | int inputs = 0, depth = 0; |
duke@435 | 859 | if (must_throw) { |
duke@435 | 860 | assert(method() == youngest_jvms->method(), "sanity"); |
duke@435 | 861 | if (compute_stack_effects(inputs, depth)) { |
duke@435 | 862 | can_prune_locals = true; |
duke@435 | 863 | stack_slots_not_pruned = inputs; |
duke@435 | 864 | } |
duke@435 | 865 | } |
duke@435 | 866 | |
never@1832 | 867 | if (env()->jvmti_can_access_local_variables()) { |
duke@435 | 868 | // At any safepoint, this method can get breakpointed, which would |
duke@435 | 869 | // then require an immediate deoptimization. |
duke@435 | 870 | can_prune_locals = false; // do not prune locals |
duke@435 | 871 | stack_slots_not_pruned = 0; |
duke@435 | 872 | } |
duke@435 | 873 | |
duke@435 | 874 | // do not scribble on the input jvms |
duke@435 | 875 | JVMState* out_jvms = youngest_jvms->clone_deep(C); |
duke@435 | 876 | call->set_jvms(out_jvms); // Start jvms list for call node |
duke@435 | 877 | |
cfang@1335 | 878 | // For a known set of bytecodes, the interpreter should reexecute them if |
cfang@1335 | 879 | // deoptimization happens. We set the reexecute state for them here |
cfang@1335 | 880 | if (out_jvms->is_reexecute_undefined() && //don't change if already specified |
kvn@1687 | 881 | should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { |
cfang@1335 | 882 | out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed |
cfang@1335 | 883 | } |
cfang@1335 | 884 | |
duke@435 | 885 | // Presize the call: |
twisti@4313 | 886 | DEBUG_ONLY(uint non_debug_edges = call->req()); |
duke@435 | 887 | call->add_req_batch(top(), youngest_jvms->debug_depth()); |
duke@435 | 888 | assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); |
duke@435 | 889 | |
duke@435 | 890 | // Set up edges so that the call looks like this: |
duke@435 | 891 | // Call [state:] ctl io mem fptr retadr |
duke@435 | 892 | // [parms:] parm0 ... parmN |
duke@435 | 893 | // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN |
duke@435 | 894 | // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] |
duke@435 | 895 | // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN |
duke@435 | 896 | // Note that caller debug info precedes callee debug info. |
duke@435 | 897 | |
duke@435 | 898 | // Fill pointer walks backwards from "young:" to "root:" in the diagram above: |
duke@435 | 899 | uint debug_ptr = call->req(); |
duke@435 | 900 | |
duke@435 | 901 | // Loop over the map input edges associated with jvms, add them |
duke@435 | 902 | // to the call node, & reset all offsets to match call node array. |
duke@435 | 903 | for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { |
duke@435 | 904 | uint debug_end = debug_ptr; |
duke@435 | 905 | uint debug_start = debug_ptr - in_jvms->debug_size(); |
duke@435 | 906 | debug_ptr = debug_start; // back up the ptr |
duke@435 | 907 | |
duke@435 | 908 | uint p = debug_start; // walks forward in [debug_start, debug_end) |
duke@435 | 909 | uint j, k, l; |
duke@435 | 910 | SafePointNode* in_map = in_jvms->map(); |
duke@435 | 911 | out_jvms->set_map(call); |
duke@435 | 912 | |
duke@435 | 913 | if (can_prune_locals) { |
duke@435 | 914 | assert(in_jvms->method() == out_jvms->method(), "sanity"); |
duke@435 | 915 | // If the current throw can reach an exception handler in this JVMS, |
duke@435 | 916 | // then we must keep everything live that can reach that handler. |
duke@435 | 917 | // As a quick and dirty approximation, we look for any handlers at all. |
duke@435 | 918 | if (in_jvms->method()->has_exception_handlers()) { |
duke@435 | 919 | can_prune_locals = false; |
duke@435 | 920 | } |
duke@435 | 921 | } |
duke@435 | 922 | |
duke@435 | 923 | // Add the Locals |
duke@435 | 924 | k = in_jvms->locoff(); |
duke@435 | 925 | l = in_jvms->loc_size(); |
duke@435 | 926 | out_jvms->set_locoff(p); |
never@1832 | 927 | if (!can_prune_locals) { |
duke@435 | 928 | for (j = 0; j < l; j++) |
duke@435 | 929 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 930 | } else { |
duke@435 | 931 | p += l; // already set to top above by add_req_batch |
duke@435 | 932 | } |
duke@435 | 933 | |
duke@435 | 934 | // Add the Expression Stack |
duke@435 | 935 | k = in_jvms->stkoff(); |
duke@435 | 936 | l = in_jvms->sp(); |
duke@435 | 937 | out_jvms->set_stkoff(p); |
never@1832 | 938 | if (!can_prune_locals) { |
duke@435 | 939 | for (j = 0; j < l; j++) |
duke@435 | 940 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 941 | } else if (can_prune_locals && stack_slots_not_pruned != 0) { |
duke@435 | 942 | // Divide stack into {S0,...,S1}, where S0 is set to top. |
duke@435 | 943 | uint s1 = stack_slots_not_pruned; |
duke@435 | 944 | stack_slots_not_pruned = 0; // for next iteration |
duke@435 | 945 | if (s1 > l) s1 = l; |
duke@435 | 946 | uint s0 = l - s1; |
duke@435 | 947 | p += s0; // skip the tops preinstalled by add_req_batch |
duke@435 | 948 | for (j = s0; j < l; j++) |
duke@435 | 949 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 950 | } else { |
duke@435 | 951 | p += l; // already set to top above by add_req_batch |
duke@435 | 952 | } |
duke@435 | 953 | |
duke@435 | 954 | // Add the Monitors |
duke@435 | 955 | k = in_jvms->monoff(); |
duke@435 | 956 | l = in_jvms->mon_size(); |
duke@435 | 957 | out_jvms->set_monoff(p); |
duke@435 | 958 | for (j = 0; j < l; j++) |
duke@435 | 959 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 960 | |
kvn@498 | 961 | // Copy any scalar object fields. |
kvn@498 | 962 | k = in_jvms->scloff(); |
kvn@498 | 963 | l = in_jvms->scl_size(); |
kvn@498 | 964 | out_jvms->set_scloff(p); |
kvn@498 | 965 | for (j = 0; j < l; j++) |
kvn@498 | 966 | call->set_req(p++, in_map->in(k+j)); |
kvn@498 | 967 | |
duke@435 | 968 | // Finish the new jvms. |
duke@435 | 969 | out_jvms->set_endoff(p); |
duke@435 | 970 | |
duke@435 | 971 | assert(out_jvms->endoff() == debug_end, "fill ptr must match"); |
duke@435 | 972 | assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); |
duke@435 | 973 | assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); |
duke@435 | 974 | assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); |
kvn@498 | 975 | assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); |
duke@435 | 976 | assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); |
duke@435 | 977 | |
duke@435 | 978 | // Update the two tail pointers in parallel. |
duke@435 | 979 | out_jvms = out_jvms->caller(); |
duke@435 | 980 | in_jvms = in_jvms->caller(); |
duke@435 | 981 | } |
duke@435 | 982 | |
duke@435 | 983 | assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); |
duke@435 | 984 | |
duke@435 | 985 | // Test the correctness of JVMState::debug_xxx accessors: |
duke@435 | 986 | assert(call->jvms()->debug_start() == non_debug_edges, ""); |
duke@435 | 987 | assert(call->jvms()->debug_end() == call->req(), ""); |
duke@435 | 988 | assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); |
duke@435 | 989 | } |
duke@435 | 990 | |
twisti@4313 | 991 | bool GraphKit::compute_stack_effects(int& inputs, int& depth) { |
duke@435 | 992 | Bytecodes::Code code = java_bc(); |
duke@435 | 993 | if (code == Bytecodes::_wide) { |
duke@435 | 994 | code = method()->java_code_at_bci(bci() + 1); |
duke@435 | 995 | } |
duke@435 | 996 | |
duke@435 | 997 | BasicType rtype = T_ILLEGAL; |
duke@435 | 998 | int rsize = 0; |
duke@435 | 999 | |
duke@435 | 1000 | if (code != Bytecodes::_illegal) { |
duke@435 | 1001 | depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 |
duke@435 | 1002 | rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V |
duke@435 | 1003 | if (rtype < T_CONFLICT) |
duke@435 | 1004 | rsize = type2size[rtype]; |
duke@435 | 1005 | } |
duke@435 | 1006 | |
duke@435 | 1007 | switch (code) { |
duke@435 | 1008 | case Bytecodes::_illegal: |
duke@435 | 1009 | return false; |
duke@435 | 1010 | |
duke@435 | 1011 | case Bytecodes::_ldc: |
duke@435 | 1012 | case Bytecodes::_ldc_w: |
duke@435 | 1013 | case Bytecodes::_ldc2_w: |
duke@435 | 1014 | inputs = 0; |
duke@435 | 1015 | break; |
duke@435 | 1016 | |
duke@435 | 1017 | case Bytecodes::_dup: inputs = 1; break; |
duke@435 | 1018 | case Bytecodes::_dup_x1: inputs = 2; break; |
duke@435 | 1019 | case Bytecodes::_dup_x2: inputs = 3; break; |
duke@435 | 1020 | case Bytecodes::_dup2: inputs = 2; break; |
duke@435 | 1021 | case Bytecodes::_dup2_x1: inputs = 3; break; |
duke@435 | 1022 | case Bytecodes::_dup2_x2: inputs = 4; break; |
duke@435 | 1023 | case Bytecodes::_swap: inputs = 2; break; |
duke@435 | 1024 | case Bytecodes::_arraylength: inputs = 1; break; |
duke@435 | 1025 | |
duke@435 | 1026 | case Bytecodes::_getstatic: |
duke@435 | 1027 | case Bytecodes::_putstatic: |
duke@435 | 1028 | case Bytecodes::_getfield: |
duke@435 | 1029 | case Bytecodes::_putfield: |
duke@435 | 1030 | { |
twisti@4313 | 1031 | bool ignored_will_link; |
twisti@4313 | 1032 | ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); |
twisti@4313 | 1033 | int size = field->type()->size(); |
duke@435 | 1034 | bool is_get = (depth >= 0), is_static = (depth & 1); |
twisti@4313 | 1035 | inputs = (is_static ? 0 : 1); |
duke@435 | 1036 | if (is_get) { |
duke@435 | 1037 | depth = size - inputs; |
duke@435 | 1038 | } else { |
duke@435 | 1039 | inputs += size; // putxxx pops the value from the stack |
duke@435 | 1040 | depth = - inputs; |
duke@435 | 1041 | } |
duke@435 | 1042 | } |
duke@435 | 1043 | break; |
duke@435 | 1044 | |
duke@435 | 1045 | case Bytecodes::_invokevirtual: |
duke@435 | 1046 | case Bytecodes::_invokespecial: |
duke@435 | 1047 | case Bytecodes::_invokestatic: |
jrose@1161 | 1048 | case Bytecodes::_invokedynamic: |
duke@435 | 1049 | case Bytecodes::_invokeinterface: |
duke@435 | 1050 | { |
twisti@4021 | 1051 | bool ignored_will_link; |
twisti@4021 | 1052 | ciSignature* declared_signature = NULL; |
twisti@4313 | 1053 | ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); |
twisti@4021 | 1054 | assert(declared_signature != NULL, "cannot be null"); |
twisti@4313 | 1055 | inputs = declared_signature->arg_size_for_bc(code); |
twisti@4021 | 1056 | int size = declared_signature->return_type()->size(); |
duke@435 | 1057 | depth = size - inputs; |
duke@435 | 1058 | } |
duke@435 | 1059 | break; |
duke@435 | 1060 | |
duke@435 | 1061 | case Bytecodes::_multianewarray: |
duke@435 | 1062 | { |
duke@435 | 1063 | ciBytecodeStream iter(method()); |
duke@435 | 1064 | iter.reset_to_bci(bci()); |
duke@435 | 1065 | iter.next(); |
duke@435 | 1066 | inputs = iter.get_dimensions(); |
duke@435 | 1067 | assert(rsize == 1, ""); |
duke@435 | 1068 | depth = rsize - inputs; |
duke@435 | 1069 | } |
duke@435 | 1070 | break; |
duke@435 | 1071 | |
duke@435 | 1072 | case Bytecodes::_ireturn: |
duke@435 | 1073 | case Bytecodes::_lreturn: |
duke@435 | 1074 | case Bytecodes::_freturn: |
duke@435 | 1075 | case Bytecodes::_dreturn: |
duke@435 | 1076 | case Bytecodes::_areturn: |
duke@435 | 1077 | assert(rsize = -depth, ""); |
duke@435 | 1078 | inputs = rsize; |
duke@435 | 1079 | break; |
duke@435 | 1080 | |
duke@435 | 1081 | case Bytecodes::_jsr: |
duke@435 | 1082 | case Bytecodes::_jsr_w: |
duke@435 | 1083 | inputs = 0; |
duke@435 | 1084 | depth = 1; // S.B. depth=1, not zero |
duke@435 | 1085 | break; |
duke@435 | 1086 | |
duke@435 | 1087 | default: |
duke@435 | 1088 | // bytecode produces a typed result |
duke@435 | 1089 | inputs = rsize - depth; |
duke@435 | 1090 | assert(inputs >= 0, ""); |
duke@435 | 1091 | break; |
duke@435 | 1092 | } |
duke@435 | 1093 | |
duke@435 | 1094 | #ifdef ASSERT |
duke@435 | 1095 | // spot check |
duke@435 | 1096 | int outputs = depth + inputs; |
duke@435 | 1097 | assert(outputs >= 0, "sanity"); |
duke@435 | 1098 | switch (code) { |
duke@435 | 1099 | case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; |
duke@435 | 1100 | case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; |
duke@435 | 1101 | case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; |
duke@435 | 1102 | case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; |
duke@435 | 1103 | case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; |
duke@435 | 1104 | } |
duke@435 | 1105 | #endif //ASSERT |
duke@435 | 1106 | |
duke@435 | 1107 | return true; |
duke@435 | 1108 | } |
duke@435 | 1109 | |
duke@435 | 1110 | |
duke@435 | 1111 | |
duke@435 | 1112 | //------------------------------basic_plus_adr--------------------------------- |
duke@435 | 1113 | Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { |
duke@435 | 1114 | // short-circuit a common case |
duke@435 | 1115 | if (offset == intcon(0)) return ptr; |
kvn@4115 | 1116 | return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); |
duke@435 | 1117 | } |
duke@435 | 1118 | |
duke@435 | 1119 | Node* GraphKit::ConvI2L(Node* offset) { |
duke@435 | 1120 | // short-circuit a common case |
duke@435 | 1121 | jint offset_con = find_int_con(offset, Type::OffsetBot); |
duke@435 | 1122 | if (offset_con != Type::OffsetBot) { |
vlivanov@4157 | 1123 | return longcon((jlong) offset_con); |
duke@435 | 1124 | } |
kvn@4115 | 1125 | return _gvn.transform( new (C) ConvI2LNode(offset)); |
duke@435 | 1126 | } |
poonam@6425 | 1127 | |
poonam@6425 | 1128 | Node* GraphKit::ConvI2UL(Node* offset) { |
poonam@6425 | 1129 | juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); |
poonam@6425 | 1130 | if (offset_con != (juint) Type::OffsetBot) { |
poonam@6425 | 1131 | return longcon((julong) offset_con); |
poonam@6425 | 1132 | } |
poonam@6425 | 1133 | Node* conv = _gvn.transform( new (C) ConvI2LNode(offset)); |
poonam@6425 | 1134 | Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) ); |
poonam@6425 | 1135 | return _gvn.transform( new (C) AndLNode(conv, mask) ); |
poonam@6425 | 1136 | } |
poonam@6425 | 1137 | |
duke@435 | 1138 | Node* GraphKit::ConvL2I(Node* offset) { |
duke@435 | 1139 | // short-circuit a common case |
duke@435 | 1140 | jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); |
duke@435 | 1141 | if (offset_con != (jlong)Type::OffsetBot) { |
duke@435 | 1142 | return intcon((int) offset_con); |
duke@435 | 1143 | } |
kvn@4115 | 1144 | return _gvn.transform( new (C) ConvL2INode(offset)); |
duke@435 | 1145 | } |
duke@435 | 1146 | |
duke@435 | 1147 | //-------------------------load_object_klass----------------------------------- |
duke@435 | 1148 | Node* GraphKit::load_object_klass(Node* obj) { |
duke@435 | 1149 | // Special-case a fresh allocation to avoid building nodes: |
duke@435 | 1150 | Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); |
duke@435 | 1151 | if (akls != NULL) return akls; |
duke@435 | 1152 | Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); |
kvn@599 | 1153 | return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); |
duke@435 | 1154 | } |
duke@435 | 1155 | |
duke@435 | 1156 | //-------------------------load_array_length----------------------------------- |
duke@435 | 1157 | Node* GraphKit::load_array_length(Node* array) { |
duke@435 | 1158 | // Special-case a fresh allocation to avoid building nodes: |
rasbold@801 | 1159 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); |
rasbold@801 | 1160 | Node *alen; |
rasbold@801 | 1161 | if (alloc == NULL) { |
rasbold@801 | 1162 | Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); |
kvn@4115 | 1163 | alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); |
rasbold@801 | 1164 | } else { |
rasbold@801 | 1165 | alen = alloc->Ideal_length(); |
kvn@1389 | 1166 | Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); |
rasbold@801 | 1167 | if (ccast != alen) { |
rasbold@801 | 1168 | alen = _gvn.transform(ccast); |
rasbold@801 | 1169 | } |
rasbold@801 | 1170 | } |
rasbold@801 | 1171 | return alen; |
duke@435 | 1172 | } |
duke@435 | 1173 | |
duke@435 | 1174 | //------------------------------do_null_check---------------------------------- |
duke@435 | 1175 | // Helper function to do a NULL pointer check. Returned value is |
duke@435 | 1176 | // the incoming address with NULL casted away. You are allowed to use the |
duke@435 | 1177 | // not-null value only if you are control dependent on the test. |
duke@435 | 1178 | extern int explicit_null_checks_inserted, |
duke@435 | 1179 | explicit_null_checks_elided; |
duke@435 | 1180 | Node* GraphKit::null_check_common(Node* value, BasicType type, |
duke@435 | 1181 | // optional arguments for variations: |
duke@435 | 1182 | bool assert_null, |
duke@435 | 1183 | Node* *null_control) { |
duke@435 | 1184 | assert(!assert_null || null_control == NULL, "not both at once"); |
duke@435 | 1185 | if (stopped()) return top(); |
duke@435 | 1186 | if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { |
duke@435 | 1187 | // For some performance testing, we may wish to suppress null checking. |
duke@435 | 1188 | value = cast_not_null(value); // Make it appear to be non-null (4962416). |
duke@435 | 1189 | return value; |
duke@435 | 1190 | } |
duke@435 | 1191 | explicit_null_checks_inserted++; |
duke@435 | 1192 | |
duke@435 | 1193 | // Construct NULL check |
duke@435 | 1194 | Node *chk = NULL; |
duke@435 | 1195 | switch(type) { |
kvn@4115 | 1196 | case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; |
twisti@4313 | 1197 | case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; |
duke@435 | 1198 | case T_ARRAY : // fall through |
duke@435 | 1199 | type = T_OBJECT; // simplify further tests |
duke@435 | 1200 | case T_OBJECT : { |
duke@435 | 1201 | const Type *t = _gvn.type( value ); |
duke@435 | 1202 | |
kvn@1385 | 1203 | const TypeOopPtr* tp = t->isa_oopptr(); |
kvn@1390 | 1204 | if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() |
duke@435 | 1205 | // Only for do_null_check, not any of its siblings: |
duke@435 | 1206 | && !assert_null && null_control == NULL) { |
duke@435 | 1207 | // Usually, any field access or invocation on an unloaded oop type |
duke@435 | 1208 | // will simply fail to link, since the statically linked class is |
duke@435 | 1209 | // likely also to be unloaded. However, in -Xcomp mode, sometimes |
duke@435 | 1210 | // the static class is loaded but the sharper oop type is not. |
duke@435 | 1211 | // Rather than checking for this obscure case in lots of places, |
duke@435 | 1212 | // we simply observe that a null check on an unloaded class |
duke@435 | 1213 | // will always be followed by a nonsense operation, so we |
duke@435 | 1214 | // can just issue the uncommon trap here. |
duke@435 | 1215 | // Our access to the unloaded class will only be correct |
duke@435 | 1216 | // after it has been loaded and initialized, which requires |
duke@435 | 1217 | // a trip through the interpreter. |
duke@435 | 1218 | #ifndef PRODUCT |
duke@435 | 1219 | if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } |
duke@435 | 1220 | #endif |
duke@435 | 1221 | uncommon_trap(Deoptimization::Reason_unloaded, |
duke@435 | 1222 | Deoptimization::Action_reinterpret, |
duke@435 | 1223 | tp->klass(), "!loaded"); |
duke@435 | 1224 | return top(); |
duke@435 | 1225 | } |
duke@435 | 1226 | |
duke@435 | 1227 | if (assert_null) { |
duke@435 | 1228 | // See if the type is contained in NULL_PTR. |
duke@435 | 1229 | // If so, then the value is already null. |
duke@435 | 1230 | if (t->higher_equal(TypePtr::NULL_PTR)) { |
duke@435 | 1231 | explicit_null_checks_elided++; |
duke@435 | 1232 | return value; // Elided null assert quickly! |
duke@435 | 1233 | } |
duke@435 | 1234 | } else { |
duke@435 | 1235 | // See if mixing in the NULL pointer changes type. |
duke@435 | 1236 | // If so, then the NULL pointer was not allowed in the original |
duke@435 | 1237 | // type. In other words, "value" was not-null. |
roland@6313 | 1238 | if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { |
duke@435 | 1239 | // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... |
duke@435 | 1240 | explicit_null_checks_elided++; |
duke@435 | 1241 | return value; // Elided null check quickly! |
duke@435 | 1242 | } |
duke@435 | 1243 | } |
kvn@4115 | 1244 | chk = new (C) CmpPNode( value, null() ); |
duke@435 | 1245 | break; |
duke@435 | 1246 | } |
duke@435 | 1247 | |
twisti@4313 | 1248 | default: |
twisti@4313 | 1249 | fatal(err_msg_res("unexpected type: %s", type2name(type))); |
duke@435 | 1250 | } |
duke@435 | 1251 | assert(chk != NULL, "sanity check"); |
duke@435 | 1252 | chk = _gvn.transform(chk); |
duke@435 | 1253 | |
duke@435 | 1254 | BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; |
kvn@4115 | 1255 | BoolNode *btst = new (C) BoolNode( chk, btest); |
duke@435 | 1256 | Node *tst = _gvn.transform( btst ); |
duke@435 | 1257 | |
duke@435 | 1258 | //----------- |
twisti@1040 | 1259 | // if peephole optimizations occurred, a prior test existed. |
duke@435 | 1260 | // If a prior test existed, maybe it dominates as we can avoid this test. |
duke@435 | 1261 | if (tst != btst && type == T_OBJECT) { |
duke@435 | 1262 | // At this point we want to scan up the CFG to see if we can |
duke@435 | 1263 | // find an identical test (and so avoid this test altogether). |
duke@435 | 1264 | Node *cfg = control(); |
duke@435 | 1265 | int depth = 0; |
duke@435 | 1266 | while( depth < 16 ) { // Limit search depth for speed |
duke@435 | 1267 | if( cfg->Opcode() == Op_IfTrue && |
duke@435 | 1268 | cfg->in(0)->in(1) == tst ) { |
duke@435 | 1269 | // Found prior test. Use "cast_not_null" to construct an identical |
duke@435 | 1270 | // CastPP (and hence hash to) as already exists for the prior test. |
duke@435 | 1271 | // Return that casted value. |
duke@435 | 1272 | if (assert_null) { |
duke@435 | 1273 | replace_in_map(value, null()); |
duke@435 | 1274 | return null(); // do not issue the redundant test |
duke@435 | 1275 | } |
duke@435 | 1276 | Node *oldcontrol = control(); |
duke@435 | 1277 | set_control(cfg); |
duke@435 | 1278 | Node *res = cast_not_null(value); |
duke@435 | 1279 | set_control(oldcontrol); |
duke@435 | 1280 | explicit_null_checks_elided++; |
duke@435 | 1281 | return res; |
duke@435 | 1282 | } |
duke@435 | 1283 | cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); |
duke@435 | 1284 | if (cfg == NULL) break; // Quit at region nodes |
duke@435 | 1285 | depth++; |
duke@435 | 1286 | } |
duke@435 | 1287 | } |
duke@435 | 1288 | |
duke@435 | 1289 | //----------- |
duke@435 | 1290 | // Branch to failure if null |
duke@435 | 1291 | float ok_prob = PROB_MAX; // a priori estimate: nulls never happen |
duke@435 | 1292 | Deoptimization::DeoptReason reason; |
duke@435 | 1293 | if (assert_null) |
duke@435 | 1294 | reason = Deoptimization::Reason_null_assert; |
duke@435 | 1295 | else if (type == T_OBJECT) |
duke@435 | 1296 | reason = Deoptimization::Reason_null_check; |
duke@435 | 1297 | else |
duke@435 | 1298 | reason = Deoptimization::Reason_div0_check; |
duke@435 | 1299 | |
ysr@777 | 1300 | // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, |
ysr@777 | 1301 | // ciMethodData::has_trap_at will return a conservative -1 if any |
ysr@777 | 1302 | // must-be-null assertion has failed. This could cause performance |
ysr@777 | 1303 | // problems for a method after its first do_null_assert failure. |
ysr@777 | 1304 | // Consider using 'Reason_class_check' instead? |
ysr@777 | 1305 | |
duke@435 | 1306 | // To cause an implicit null check, we set the not-null probability |
twisti@1040 | 1307 | // to the maximum (PROB_MAX). For an explicit check the probability |
duke@435 | 1308 | // is set to a smaller value. |
duke@435 | 1309 | if (null_control != NULL || too_many_traps(reason)) { |
duke@435 | 1310 | // probability is less likely |
duke@435 | 1311 | ok_prob = PROB_LIKELY_MAG(3); |
duke@435 | 1312 | } else if (!assert_null && |
duke@435 | 1313 | (ImplicitNullCheckThreshold > 0) && |
duke@435 | 1314 | method() != NULL && |
duke@435 | 1315 | (method()->method_data()->trap_count(reason) |
duke@435 | 1316 | >= (uint)ImplicitNullCheckThreshold)) { |
duke@435 | 1317 | ok_prob = PROB_LIKELY_MAG(3); |
duke@435 | 1318 | } |
duke@435 | 1319 | |
duke@435 | 1320 | if (null_control != NULL) { |
duke@435 | 1321 | IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); |
kvn@4115 | 1322 | Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); |
kvn@4115 | 1323 | set_control( _gvn.transform( new (C) IfTrueNode(iff))); |
duke@435 | 1324 | if (null_true == top()) |
duke@435 | 1325 | explicit_null_checks_elided++; |
duke@435 | 1326 | (*null_control) = null_true; |
duke@435 | 1327 | } else { |
duke@435 | 1328 | BuildCutout unless(this, tst, ok_prob); |
duke@435 | 1329 | // Check for optimizer eliding test at parse time |
duke@435 | 1330 | if (stopped()) { |
duke@435 | 1331 | // Failure not possible; do not bother making uncommon trap. |
duke@435 | 1332 | explicit_null_checks_elided++; |
duke@435 | 1333 | } else if (assert_null) { |
duke@435 | 1334 | uncommon_trap(reason, |
duke@435 | 1335 | Deoptimization::Action_make_not_entrant, |
duke@435 | 1336 | NULL, "assert_null"); |
duke@435 | 1337 | } else { |
kvn@767 | 1338 | replace_in_map(value, zerocon(type)); |
duke@435 | 1339 | builtin_throw(reason); |
duke@435 | 1340 | } |
duke@435 | 1341 | } |
duke@435 | 1342 | |
duke@435 | 1343 | // Must throw exception, fall-thru not possible? |
duke@435 | 1344 | if (stopped()) { |
duke@435 | 1345 | return top(); // No result |
duke@435 | 1346 | } |
duke@435 | 1347 | |
duke@435 | 1348 | if (assert_null) { |
duke@435 | 1349 | // Cast obj to null on this path. |
duke@435 | 1350 | replace_in_map(value, zerocon(type)); |
duke@435 | 1351 | return zerocon(type); |
duke@435 | 1352 | } |
duke@435 | 1353 | |
duke@435 | 1354 | // Cast obj to not-null on this path, if there is no null_control. |
duke@435 | 1355 | // (If there is a null_control, a non-null value may come back to haunt us.) |
duke@435 | 1356 | if (type == T_OBJECT) { |
duke@435 | 1357 | Node* cast = cast_not_null(value, false); |
duke@435 | 1358 | if (null_control == NULL || (*null_control) == top()) |
duke@435 | 1359 | replace_in_map(value, cast); |
duke@435 | 1360 | value = cast; |
duke@435 | 1361 | } |
duke@435 | 1362 | |
duke@435 | 1363 | return value; |
duke@435 | 1364 | } |
duke@435 | 1365 | |
duke@435 | 1366 | |
duke@435 | 1367 | //------------------------------cast_not_null---------------------------------- |
duke@435 | 1368 | // Cast obj to not-null on this path |
duke@435 | 1369 | Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { |
duke@435 | 1370 | const Type *t = _gvn.type(obj); |
roland@6313 | 1371 | const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); |
duke@435 | 1372 | // Object is already not-null? |
duke@435 | 1373 | if( t == t_not_null ) return obj; |
duke@435 | 1374 | |
kvn@4115 | 1375 | Node *cast = new (C) CastPPNode(obj,t_not_null); |
duke@435 | 1376 | cast->init_req(0, control()); |
duke@435 | 1377 | cast = _gvn.transform( cast ); |
duke@435 | 1378 | |
duke@435 | 1379 | // Scan for instances of 'obj' in the current JVM mapping. |
duke@435 | 1380 | // These instances are known to be not-null after the test. |
duke@435 | 1381 | if (do_replace_in_map) |
duke@435 | 1382 | replace_in_map(obj, cast); |
duke@435 | 1383 | |
duke@435 | 1384 | return cast; // Return casted value |
duke@435 | 1385 | } |
duke@435 | 1386 | |
duke@435 | 1387 | |
duke@435 | 1388 | //--------------------------replace_in_map------------------------------------- |
duke@435 | 1389 | void GraphKit::replace_in_map(Node* old, Node* neww) { |
roland@5981 | 1390 | if (old == neww) { |
roland@5981 | 1391 | return; |
roland@5981 | 1392 | } |
roland@5981 | 1393 | |
roland@5981 | 1394 | map()->replace_edge(old, neww); |
duke@435 | 1395 | |
duke@435 | 1396 | // Note: This operation potentially replaces any edge |
duke@435 | 1397 | // on the map. This includes locals, stack, and monitors |
duke@435 | 1398 | // of the current (innermost) JVM state. |
duke@435 | 1399 | |
roland@7041 | 1400 | // don't let inconsistent types from profiling escape this |
roland@7041 | 1401 | // method |
roland@7041 | 1402 | |
roland@7041 | 1403 | const Type* told = _gvn.type(old); |
roland@7041 | 1404 | const Type* tnew = _gvn.type(neww); |
roland@7041 | 1405 | |
roland@7041 | 1406 | if (!tnew->higher_equal(told)) { |
roland@5981 | 1407 | return; |
roland@5981 | 1408 | } |
roland@5981 | 1409 | |
roland@7041 | 1410 | map()->record_replaced_node(old, neww); |
duke@435 | 1411 | } |
duke@435 | 1412 | |
duke@435 | 1413 | |
duke@435 | 1414 | //============================================================================= |
duke@435 | 1415 | //--------------------------------memory--------------------------------------- |
duke@435 | 1416 | Node* GraphKit::memory(uint alias_idx) { |
duke@435 | 1417 | MergeMemNode* mem = merged_memory(); |
duke@435 | 1418 | Node* p = mem->memory_at(alias_idx); |
duke@435 | 1419 | _gvn.set_type(p, Type::MEMORY); // must be mapped |
duke@435 | 1420 | return p; |
duke@435 | 1421 | } |
duke@435 | 1422 | |
duke@435 | 1423 | //-----------------------------reset_memory------------------------------------ |
duke@435 | 1424 | Node* GraphKit::reset_memory() { |
duke@435 | 1425 | Node* mem = map()->memory(); |
duke@435 | 1426 | // do not use this node for any more parsing! |
duke@435 | 1427 | debug_only( map()->set_memory((Node*)NULL) ); |
duke@435 | 1428 | return _gvn.transform( mem ); |
duke@435 | 1429 | } |
duke@435 | 1430 | |
duke@435 | 1431 | //------------------------------set_all_memory--------------------------------- |
duke@435 | 1432 | void GraphKit::set_all_memory(Node* newmem) { |
duke@435 | 1433 | Node* mergemem = MergeMemNode::make(C, newmem); |
duke@435 | 1434 | gvn().set_type_bottom(mergemem); |
duke@435 | 1435 | map()->set_memory(mergemem); |
duke@435 | 1436 | } |
duke@435 | 1437 | |
duke@435 | 1438 | //------------------------------set_all_memory_call---------------------------- |
never@1515 | 1439 | void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { |
kvn@4115 | 1440 | Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); |
duke@435 | 1441 | set_all_memory(newmem); |
duke@435 | 1442 | } |
duke@435 | 1443 | |
duke@435 | 1444 | //============================================================================= |
duke@435 | 1445 | // |
duke@435 | 1446 | // parser factory methods for MemNodes |
duke@435 | 1447 | // |
duke@435 | 1448 | // These are layered on top of the factory methods in LoadNode and StoreNode, |
duke@435 | 1449 | // and integrate with the parser's memory state and _gvn engine. |
duke@435 | 1450 | // |
duke@435 | 1451 | |
duke@435 | 1452 | // factory methods in "int adr_idx" |
duke@435 | 1453 | Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, |
duke@435 | 1454 | int adr_idx, |
goetz@6479 | 1455 | MemNode::MemOrd mo, bool require_atomic_access) { |
duke@435 | 1456 | assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); |
duke@435 | 1457 | const TypePtr* adr_type = NULL; // debug-mode-only argument |
duke@435 | 1458 | debug_only(adr_type = C->get_adr_type(adr_idx)); |
duke@435 | 1459 | Node* mem = memory(adr_idx); |
duke@435 | 1460 | Node* ld; |
duke@435 | 1461 | if (require_atomic_access && bt == T_LONG) { |
goetz@6479 | 1462 | ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo); |
duke@435 | 1463 | } else { |
goetz@6479 | 1464 | ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo); |
duke@435 | 1465 | } |
kvn@5110 | 1466 | ld = _gvn.transform(ld); |
kvn@5110 | 1467 | if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { |
kvn@5110 | 1468 | // Improve graph before escape analysis and boxing elimination. |
kvn@5110 | 1469 | record_for_igvn(ld); |
kvn@5110 | 1470 | } |
kvn@5110 | 1471 | return ld; |
duke@435 | 1472 | } |
duke@435 | 1473 | |
duke@435 | 1474 | Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, |
duke@435 | 1475 | int adr_idx, |
goetz@6479 | 1476 | MemNode::MemOrd mo, |
duke@435 | 1477 | bool require_atomic_access) { |
duke@435 | 1478 | assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); |
duke@435 | 1479 | const TypePtr* adr_type = NULL; |
duke@435 | 1480 | debug_only(adr_type = C->get_adr_type(adr_idx)); |
duke@435 | 1481 | Node *mem = memory(adr_idx); |
duke@435 | 1482 | Node* st; |
duke@435 | 1483 | if (require_atomic_access && bt == T_LONG) { |
goetz@6479 | 1484 | st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); |
duke@435 | 1485 | } else { |
goetz@6479 | 1486 | st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); |
duke@435 | 1487 | } |
duke@435 | 1488 | st = _gvn.transform(st); |
duke@435 | 1489 | set_memory(st, adr_idx); |
duke@435 | 1490 | // Back-to-back stores can only remove intermediate store with DU info |
duke@435 | 1491 | // so push on worklist for optimizer. |
duke@435 | 1492 | if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) |
duke@435 | 1493 | record_for_igvn(st); |
duke@435 | 1494 | |
duke@435 | 1495 | return st; |
duke@435 | 1496 | } |
duke@435 | 1497 | |
kvn@1286 | 1498 | |
johnc@2781 | 1499 | void GraphKit::pre_barrier(bool do_load, |
johnc@2781 | 1500 | Node* ctl, |
duke@435 | 1501 | Node* obj, |
duke@435 | 1502 | Node* adr, |
kvn@1286 | 1503 | uint adr_idx, |
kvn@1286 | 1504 | Node* val, |
never@1260 | 1505 | const TypeOopPtr* val_type, |
johnc@2781 | 1506 | Node* pre_val, |
duke@435 | 1507 | BasicType bt) { |
johnc@2781 | 1508 | |
duke@435 | 1509 | BarrierSet* bs = Universe::heap()->barrier_set(); |
duke@435 | 1510 | set_control(ctl); |
duke@435 | 1511 | switch (bs->kind()) { |
ysr@777 | 1512 | case BarrierSet::G1SATBCT: |
ysr@777 | 1513 | case BarrierSet::G1SATBCTLogging: |
johnc@2781 | 1514 | g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); |
ysr@777 | 1515 | break; |
duke@435 | 1516 | |
duke@435 | 1517 | case BarrierSet::CardTableModRef: |
duke@435 | 1518 | case BarrierSet::CardTableExtension: |
duke@435 | 1519 | case BarrierSet::ModRef: |
duke@435 | 1520 | break; |
duke@435 | 1521 | |
duke@435 | 1522 | case BarrierSet::Other: |
duke@435 | 1523 | default : |
duke@435 | 1524 | ShouldNotReachHere(); |
duke@435 | 1525 | |
duke@435 | 1526 | } |
duke@435 | 1527 | } |
duke@435 | 1528 | |
kvn@5637 | 1529 | bool GraphKit::can_move_pre_barrier() const { |
kvn@5637 | 1530 | BarrierSet* bs = Universe::heap()->barrier_set(); |
kvn@5637 | 1531 | switch (bs->kind()) { |
kvn@5637 | 1532 | case BarrierSet::G1SATBCT: |
kvn@5637 | 1533 | case BarrierSet::G1SATBCTLogging: |
kvn@5637 | 1534 | return true; // Can move it if no safepoint |
kvn@5637 | 1535 | |
kvn@5637 | 1536 | case BarrierSet::CardTableModRef: |
kvn@5637 | 1537 | case BarrierSet::CardTableExtension: |
kvn@5637 | 1538 | case BarrierSet::ModRef: |
kvn@5637 | 1539 | return true; // There is no pre-barrier |
kvn@5637 | 1540 | |
kvn@5637 | 1541 | case BarrierSet::Other: |
kvn@5637 | 1542 | default : |
kvn@5637 | 1543 | ShouldNotReachHere(); |
kvn@5637 | 1544 | } |
kvn@5637 | 1545 | return false; |
kvn@5637 | 1546 | } |
kvn@5637 | 1547 | |
duke@435 | 1548 | void GraphKit::post_barrier(Node* ctl, |
duke@435 | 1549 | Node* store, |
duke@435 | 1550 | Node* obj, |
duke@435 | 1551 | Node* adr, |
kvn@1286 | 1552 | uint adr_idx, |
kvn@1286 | 1553 | Node* val, |
duke@435 | 1554 | BasicType bt, |
duke@435 | 1555 | bool use_precise) { |
duke@435 | 1556 | BarrierSet* bs = Universe::heap()->barrier_set(); |
duke@435 | 1557 | set_control(ctl); |
duke@435 | 1558 | switch (bs->kind()) { |
ysr@777 | 1559 | case BarrierSet::G1SATBCT: |
ysr@777 | 1560 | case BarrierSet::G1SATBCTLogging: |
kvn@1286 | 1561 | g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); |
ysr@777 | 1562 | break; |
duke@435 | 1563 | |
duke@435 | 1564 | case BarrierSet::CardTableModRef: |
duke@435 | 1565 | case BarrierSet::CardTableExtension: |
cfang@1420 | 1566 | write_barrier_post(store, obj, adr, adr_idx, val, use_precise); |
duke@435 | 1567 | break; |
duke@435 | 1568 | |
duke@435 | 1569 | case BarrierSet::ModRef: |
duke@435 | 1570 | break; |
duke@435 | 1571 | |
duke@435 | 1572 | case BarrierSet::Other: |
duke@435 | 1573 | default : |
duke@435 | 1574 | ShouldNotReachHere(); |
duke@435 | 1575 | |
duke@435 | 1576 | } |
duke@435 | 1577 | } |
duke@435 | 1578 | |
kvn@1286 | 1579 | Node* GraphKit::store_oop(Node* ctl, |
kvn@1286 | 1580 | Node* obj, |
kvn@1286 | 1581 | Node* adr, |
kvn@1286 | 1582 | const TypePtr* adr_type, |
kvn@1286 | 1583 | Node* val, |
kvn@1286 | 1584 | const TypeOopPtr* val_type, |
kvn@1286 | 1585 | BasicType bt, |
goetz@6479 | 1586 | bool use_precise, |
goetz@6479 | 1587 | MemNode::MemOrd mo) { |
kvn@3521 | 1588 | // Transformation of a value which could be NULL pointer (CastPP #NULL) |
kvn@3521 | 1589 | // could be delayed during Parse (for example, in adjust_map_after_if()). |
kvn@3521 | 1590 | // Execute transformation here to avoid barrier generation in such case. |
kvn@3521 | 1591 | if (_gvn.type(val) == TypePtr::NULL_PTR) |
kvn@3521 | 1592 | val = _gvn.makecon(TypePtr::NULL_PTR); |
kvn@1286 | 1593 | |
kvn@1286 | 1594 | set_control(ctl); |
kvn@1286 | 1595 | if (stopped()) return top(); // Dead path ? |
kvn@1286 | 1596 | |
kvn@1286 | 1597 | assert(bt == T_OBJECT, "sanity"); |
kvn@1286 | 1598 | assert(val != NULL, "not dead path"); |
duke@435 | 1599 | uint adr_idx = C->get_alias_index(adr_type); |
kvn@1286 | 1600 | assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); |
kvn@1286 | 1601 | |
johnc@2781 | 1602 | pre_barrier(true /* do_load */, |
johnc@2781 | 1603 | control(), obj, adr, adr_idx, val, val_type, |
johnc@2781 | 1604 | NULL /* pre_val */, |
johnc@2781 | 1605 | bt); |
johnc@2781 | 1606 | |
goetz@6479 | 1607 | Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); |
kvn@1286 | 1608 | post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); |
duke@435 | 1609 | return store; |
duke@435 | 1610 | } |
duke@435 | 1611 | |
kvn@1286 | 1612 | // Could be an array or object we don't know at compile time (unsafe ref.) |
duke@435 | 1613 | Node* GraphKit::store_oop_to_unknown(Node* ctl, |
kvn@1286 | 1614 | Node* obj, // containing obj |
kvn@1286 | 1615 | Node* adr, // actual adress to store val at |
kvn@1286 | 1616 | const TypePtr* adr_type, |
kvn@1286 | 1617 | Node* val, |
goetz@6479 | 1618 | BasicType bt, |
goetz@6479 | 1619 | MemNode::MemOrd mo) { |
never@1260 | 1620 | Compile::AliasType* at = C->alias_type(adr_type); |
never@1260 | 1621 | const TypeOopPtr* val_type = NULL; |
never@1260 | 1622 | if (adr_type->isa_instptr()) { |
never@1260 | 1623 | if (at->field() != NULL) { |
never@1260 | 1624 | // known field. This code is a copy of the do_put_xxx logic. |
never@1260 | 1625 | ciField* field = at->field(); |
never@1260 | 1626 | if (!field->type()->is_loaded()) { |
never@1260 | 1627 | val_type = TypeInstPtr::BOTTOM; |
never@1260 | 1628 | } else { |
never@1260 | 1629 | val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); |
never@1260 | 1630 | } |
never@1260 | 1631 | } |
never@1260 | 1632 | } else if (adr_type->isa_aryptr()) { |
never@1262 | 1633 | val_type = adr_type->is_aryptr()->elem()->make_oopptr(); |
never@1260 | 1634 | } |
never@1260 | 1635 | if (val_type == NULL) { |
never@1260 | 1636 | val_type = TypeInstPtr::BOTTOM; |
never@1260 | 1637 | } |
goetz@6479 | 1638 | return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); |
duke@435 | 1639 | } |
duke@435 | 1640 | |
duke@435 | 1641 | |
duke@435 | 1642 | //-------------------------array_element_address------------------------- |
duke@435 | 1643 | Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, |
duke@435 | 1644 | const TypeInt* sizetype) { |
kvn@464 | 1645 | uint shift = exact_log2(type2aelembytes(elembt)); |
duke@435 | 1646 | uint header = arrayOopDesc::base_offset_in_bytes(elembt); |
duke@435 | 1647 | |
duke@435 | 1648 | // short-circuit a common case (saves lots of confusing waste motion) |
duke@435 | 1649 | jint idx_con = find_int_con(idx, -1); |
duke@435 | 1650 | if (idx_con >= 0) { |
duke@435 | 1651 | intptr_t offset = header + ((intptr_t)idx_con << shift); |
duke@435 | 1652 | return basic_plus_adr(ary, offset); |
duke@435 | 1653 | } |
duke@435 | 1654 | |
duke@435 | 1655 | // must be correct type for alignment purposes |
duke@435 | 1656 | Node* base = basic_plus_adr(ary, header); |
duke@435 | 1657 | #ifdef _LP64 |
duke@435 | 1658 | // The scaled index operand to AddP must be a clean 64-bit value. |
duke@435 | 1659 | // Java allows a 32-bit int to be incremented to a negative |
duke@435 | 1660 | // value, which appears in a 64-bit register as a large |
duke@435 | 1661 | // positive number. Using that large positive number as an |
duke@435 | 1662 | // operand in pointer arithmetic has bad consequences. |
duke@435 | 1663 | // On the other hand, 32-bit overflow is rare, and the possibility |
duke@435 | 1664 | // can often be excluded, if we annotate the ConvI2L node with |
duke@435 | 1665 | // a type assertion that its value is known to be a small positive |
duke@435 | 1666 | // number. (The prior range check has ensured this.) |
duke@435 | 1667 | // This assertion is used by ConvI2LNode::Ideal. |
duke@435 | 1668 | int index_max = max_jint - 1; // array size is max_jint, index is one less |
duke@435 | 1669 | if (sizetype != NULL) index_max = sizetype->_hi - 1; |
duke@435 | 1670 | const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); |
kvn@4115 | 1671 | idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) ); |
duke@435 | 1672 | #endif |
kvn@4115 | 1673 | Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); |
duke@435 | 1674 | return basic_plus_adr(ary, base, scale); |
duke@435 | 1675 | } |
duke@435 | 1676 | |
duke@435 | 1677 | //-------------------------load_array_element------------------------- |
duke@435 | 1678 | Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { |
duke@435 | 1679 | const Type* elemtype = arytype->elem(); |
duke@435 | 1680 | BasicType elembt = elemtype->array_element_basic_type(); |
duke@435 | 1681 | Node* adr = array_element_address(ary, idx, elembt, arytype->size()); |
goetz@6479 | 1682 | Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); |
duke@435 | 1683 | return ld; |
duke@435 | 1684 | } |
duke@435 | 1685 | |
duke@435 | 1686 | //-------------------------set_arguments_for_java_call------------------------- |
duke@435 | 1687 | // Arguments (pre-popped from the stack) are taken from the JVMS. |
duke@435 | 1688 | void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { |
duke@435 | 1689 | // Add the call arguments: |
duke@435 | 1690 | uint nargs = call->method()->arg_size(); |
duke@435 | 1691 | for (uint i = 0; i < nargs; i++) { |
duke@435 | 1692 | Node* arg = argument(i); |
duke@435 | 1693 | call->init_req(i + TypeFunc::Parms, arg); |
duke@435 | 1694 | } |
duke@435 | 1695 | } |
duke@435 | 1696 | |
duke@435 | 1697 | //---------------------------set_edges_for_java_call--------------------------- |
duke@435 | 1698 | // Connect a newly created call into the current JVMS. |
duke@435 | 1699 | // A return value node (if any) is returned from set_edges_for_java_call. |
never@1515 | 1700 | void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { |
duke@435 | 1701 | |
duke@435 | 1702 | // Add the predefined inputs: |
duke@435 | 1703 | call->init_req( TypeFunc::Control, control() ); |
duke@435 | 1704 | call->init_req( TypeFunc::I_O , i_o() ); |
duke@435 | 1705 | call->init_req( TypeFunc::Memory , reset_memory() ); |
duke@435 | 1706 | call->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 1707 | call->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 1708 | |
duke@435 | 1709 | add_safepoint_edges(call, must_throw); |
duke@435 | 1710 | |
duke@435 | 1711 | Node* xcall = _gvn.transform(call); |
duke@435 | 1712 | |
duke@435 | 1713 | if (xcall == top()) { |
duke@435 | 1714 | set_control(top()); |
duke@435 | 1715 | return; |
duke@435 | 1716 | } |
duke@435 | 1717 | assert(xcall == call, "call identity is stable"); |
duke@435 | 1718 | |
duke@435 | 1719 | // Re-use the current map to produce the result. |
duke@435 | 1720 | |
kvn@4115 | 1721 | set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); |
kvn@4115 | 1722 | set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); |
never@1515 | 1723 | set_all_memory_call(xcall, separate_io_proj); |
duke@435 | 1724 | |
duke@435 | 1725 | //return xcall; // no need, caller already has it |
duke@435 | 1726 | } |
duke@435 | 1727 | |
never@1515 | 1728 | Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { |
duke@435 | 1729 | if (stopped()) return top(); // maybe the call folded up? |
duke@435 | 1730 | |
duke@435 | 1731 | // Capture the return value, if any. |
duke@435 | 1732 | Node* ret; |
duke@435 | 1733 | if (call->method() == NULL || |
duke@435 | 1734 | call->method()->return_type()->basic_type() == T_VOID) |
duke@435 | 1735 | ret = top(); |
kvn@4115 | 1736 | else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); |
duke@435 | 1737 | |
duke@435 | 1738 | // Note: Since any out-of-line call can produce an exception, |
duke@435 | 1739 | // we always insert an I_O projection from the call into the result. |
duke@435 | 1740 | |
never@1515 | 1741 | make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); |
never@1515 | 1742 | |
never@1515 | 1743 | if (separate_io_proj) { |
never@1515 | 1744 | // The caller requested separate projections be used by the fall |
never@1515 | 1745 | // through and exceptional paths, so replace the projections for |
never@1515 | 1746 | // the fall through path. |
kvn@4115 | 1747 | set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); |
kvn@4115 | 1748 | set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); |
never@1515 | 1749 | } |
duke@435 | 1750 | return ret; |
duke@435 | 1751 | } |
duke@435 | 1752 | |
duke@435 | 1753 | //--------------------set_predefined_input_for_runtime_call-------------------- |
duke@435 | 1754 | // Reading and setting the memory state is way conservative here. |
duke@435 | 1755 | // The real problem is that I am not doing real Type analysis on memory, |
duke@435 | 1756 | // so I cannot distinguish card mark stores from other stores. Across a GC |
duke@435 | 1757 | // point the Store Barrier and the card mark memory has to agree. I cannot |
duke@435 | 1758 | // have a card mark store and its barrier split across the GC point from |
duke@435 | 1759 | // either above or below. Here I get that to happen by reading ALL of memory. |
duke@435 | 1760 | // A better answer would be to separate out card marks from other memory. |
duke@435 | 1761 | // For now, return the input memory state, so that it can be reused |
duke@435 | 1762 | // after the call, if this call has restricted memory effects. |
duke@435 | 1763 | Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { |
duke@435 | 1764 | // Set fixed predefined input arguments |
duke@435 | 1765 | Node* memory = reset_memory(); |
duke@435 | 1766 | call->init_req( TypeFunc::Control, control() ); |
duke@435 | 1767 | call->init_req( TypeFunc::I_O, top() ); // does no i/o |
duke@435 | 1768 | call->init_req( TypeFunc::Memory, memory ); // may gc ptrs |
duke@435 | 1769 | call->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 1770 | call->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 1771 | return memory; |
duke@435 | 1772 | } |
duke@435 | 1773 | |
duke@435 | 1774 | //-------------------set_predefined_output_for_runtime_call-------------------- |
duke@435 | 1775 | // Set control and memory (not i_o) from the call. |
duke@435 | 1776 | // If keep_mem is not NULL, use it for the output state, |
duke@435 | 1777 | // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. |
duke@435 | 1778 | // If hook_mem is NULL, this call produces no memory effects at all. |
duke@435 | 1779 | // If hook_mem is a Java-visible memory slice (such as arraycopy operands), |
duke@435 | 1780 | // then only that memory slice is taken from the call. |
duke@435 | 1781 | // In the last case, we must put an appropriate memory barrier before |
duke@435 | 1782 | // the call, so as to create the correct anti-dependencies on loads |
duke@435 | 1783 | // preceding the call. |
duke@435 | 1784 | void GraphKit::set_predefined_output_for_runtime_call(Node* call, |
duke@435 | 1785 | Node* keep_mem, |
duke@435 | 1786 | const TypePtr* hook_mem) { |
duke@435 | 1787 | // no i/o |
kvn@4115 | 1788 | set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); |
duke@435 | 1789 | if (keep_mem) { |
duke@435 | 1790 | // First clone the existing memory state |
duke@435 | 1791 | set_all_memory(keep_mem); |
duke@435 | 1792 | if (hook_mem != NULL) { |
duke@435 | 1793 | // Make memory for the call |
kvn@4115 | 1794 | Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); |
duke@435 | 1795 | // Set the RawPtr memory state only. This covers all the heap top/GC stuff |
duke@435 | 1796 | // We also use hook_mem to extract specific effects from arraycopy stubs. |
duke@435 | 1797 | set_memory(mem, hook_mem); |
duke@435 | 1798 | } |
duke@435 | 1799 | // ...else the call has NO memory effects. |
duke@435 | 1800 | |
duke@435 | 1801 | // Make sure the call advertises its memory effects precisely. |
duke@435 | 1802 | // This lets us build accurate anti-dependences in gcm.cpp. |
duke@435 | 1803 | assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), |
duke@435 | 1804 | "call node must be constructed correctly"); |
duke@435 | 1805 | } else { |
duke@435 | 1806 | assert(hook_mem == NULL, ""); |
duke@435 | 1807 | // This is not a "slow path" call; all memory comes from the call. |
duke@435 | 1808 | set_all_memory_call(call); |
duke@435 | 1809 | } |
duke@435 | 1810 | } |
duke@435 | 1811 | |
never@1515 | 1812 | |
never@1515 | 1813 | // Replace the call with the current state of the kit. |
roland@7041 | 1814 | void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { |
never@1515 | 1815 | JVMState* ejvms = NULL; |
never@1515 | 1816 | if (has_exceptions()) { |
never@1515 | 1817 | ejvms = transfer_exceptions_into_jvms(); |
never@1515 | 1818 | } |
never@1515 | 1819 | |
roland@7041 | 1820 | ReplacedNodes replaced_nodes = map()->replaced_nodes(); |
roland@7041 | 1821 | ReplacedNodes replaced_nodes_exception; |
roland@7041 | 1822 | Node* ex_ctl = top(); |
roland@7041 | 1823 | |
never@1515 | 1824 | SafePointNode* final_state = stop(); |
never@1515 | 1825 | |
never@1515 | 1826 | // Find all the needed outputs of this call |
never@1515 | 1827 | CallProjections callprojs; |
never@1515 | 1828 | call->extract_projections(&callprojs, true); |
never@1515 | 1829 | |
roland@4357 | 1830 | Node* init_mem = call->in(TypeFunc::Memory); |
roland@4357 | 1831 | Node* final_mem = final_state->in(TypeFunc::Memory); |
roland@4357 | 1832 | Node* final_ctl = final_state->in(TypeFunc::Control); |
roland@4357 | 1833 | Node* final_io = final_state->in(TypeFunc::I_O); |
roland@4357 | 1834 | |
never@1515 | 1835 | // Replace all the old call edges with the edges from the inlining result |
roland@4357 | 1836 | if (callprojs.fallthrough_catchproj != NULL) { |
roland@4357 | 1837 | C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); |
roland@4357 | 1838 | } |
roland@4357 | 1839 | if (callprojs.fallthrough_memproj != NULL) { |
roland@7041 | 1840 | if (final_mem->is_MergeMem()) { |
roland@7041 | 1841 | // Parser's exits MergeMem was not transformed but may be optimized |
roland@7041 | 1842 | final_mem = _gvn.transform(final_mem); |
roland@7041 | 1843 | } |
roland@4357 | 1844 | C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); |
roland@4357 | 1845 | } |
roland@4357 | 1846 | if (callprojs.fallthrough_ioproj != NULL) { |
roland@4357 | 1847 | C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); |
roland@4357 | 1848 | } |
never@1515 | 1849 | |
never@1515 | 1850 | // Replace the result with the new result if it exists and is used |
never@1515 | 1851 | if (callprojs.resproj != NULL && result != NULL) { |
never@1515 | 1852 | C->gvn_replace_by(callprojs.resproj, result); |
never@1515 | 1853 | } |
never@1515 | 1854 | |
never@1515 | 1855 | if (ejvms == NULL) { |
never@1515 | 1856 | // No exception edges to simply kill off those paths |
roland@4409 | 1857 | if (callprojs.catchall_catchproj != NULL) { |
roland@4409 | 1858 | C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); |
roland@4409 | 1859 | } |
roland@4409 | 1860 | if (callprojs.catchall_memproj != NULL) { |
roland@4409 | 1861 | C->gvn_replace_by(callprojs.catchall_memproj, C->top()); |
roland@4409 | 1862 | } |
roland@4409 | 1863 | if (callprojs.catchall_ioproj != NULL) { |
roland@4409 | 1864 | C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); |
roland@4409 | 1865 | } |
never@1534 | 1866 | // Replace the old exception object with top |
never@1534 | 1867 | if (callprojs.exobj != NULL) { |
never@1534 | 1868 | C->gvn_replace_by(callprojs.exobj, C->top()); |
never@1534 | 1869 | } |
never@1515 | 1870 | } else { |
never@1515 | 1871 | GraphKit ekit(ejvms); |
never@1515 | 1872 | |
never@1515 | 1873 | // Load my combined exception state into the kit, with all phis transformed: |
never@1515 | 1874 | SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); |
roland@7041 | 1875 | replaced_nodes_exception = ex_map->replaced_nodes(); |
never@1515 | 1876 | |
never@1515 | 1877 | Node* ex_oop = ekit.use_exception_state(ex_map); |
roland@7041 | 1878 | |
roland@4409 | 1879 | if (callprojs.catchall_catchproj != NULL) { |
roland@4409 | 1880 | C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); |
roland@7041 | 1881 | ex_ctl = ekit.control(); |
roland@4409 | 1882 | } |
roland@4409 | 1883 | if (callprojs.catchall_memproj != NULL) { |
roland@4409 | 1884 | C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); |
roland@4409 | 1885 | } |
roland@4409 | 1886 | if (callprojs.catchall_ioproj != NULL) { |
roland@4409 | 1887 | C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); |
roland@4409 | 1888 | } |
never@1515 | 1889 | |
never@1515 | 1890 | // Replace the old exception object with the newly created one |
never@1515 | 1891 | if (callprojs.exobj != NULL) { |
never@1515 | 1892 | C->gvn_replace_by(callprojs.exobj, ex_oop); |
never@1515 | 1893 | } |
never@1515 | 1894 | } |
never@1515 | 1895 | |
never@1515 | 1896 | // Disconnect the call from the graph |
bharadwaj@4315 | 1897 | call->disconnect_inputs(NULL, C); |
never@1515 | 1898 | C->gvn_replace_by(call, C->top()); |
never@2141 | 1899 | |
never@2141 | 1900 | // Clean up any MergeMems that feed other MergeMems since the |
never@2141 | 1901 | // optimizer doesn't like that. |
never@2141 | 1902 | if (final_mem->is_MergeMem()) { |
never@2141 | 1903 | Node_List wl; |
never@2141 | 1904 | for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { |
never@2141 | 1905 | Node* m = i.get(); |
never@2141 | 1906 | if (m->is_MergeMem() && !wl.contains(m)) { |
never@2141 | 1907 | wl.push(m); |
never@2141 | 1908 | } |
never@2141 | 1909 | } |
never@2141 | 1910 | while (wl.size() > 0) { |
never@2141 | 1911 | _gvn.transform(wl.pop()); |
never@2141 | 1912 | } |
never@2141 | 1913 | } |
roland@7041 | 1914 | |
roland@7041 | 1915 | if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { |
roland@7041 | 1916 | replaced_nodes.apply(C, final_ctl); |
roland@7041 | 1917 | } |
roland@7041 | 1918 | if (!ex_ctl->is_top() && do_replaced_nodes) { |
roland@7041 | 1919 | replaced_nodes_exception.apply(C, ex_ctl); |
roland@7041 | 1920 | } |
never@1515 | 1921 | } |
never@1515 | 1922 | |
never@1515 | 1923 | |
duke@435 | 1924 | //------------------------------increment_counter------------------------------ |
duke@435 | 1925 | // for statistics: increment a VM counter by 1 |
duke@435 | 1926 | |
duke@435 | 1927 | void GraphKit::increment_counter(address counter_addr) { |
duke@435 | 1928 | Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); |
duke@435 | 1929 | increment_counter(adr1); |
duke@435 | 1930 | } |
duke@435 | 1931 | |
duke@435 | 1932 | void GraphKit::increment_counter(Node* counter_addr) { |
duke@435 | 1933 | int adr_type = Compile::AliasIdxRaw; |
kvn@1964 | 1934 | Node* ctrl = control(); |
goetz@6479 | 1935 | Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); |
kvn@4115 | 1936 | Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); |
goetz@6479 | 1937 | store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); |
duke@435 | 1938 | } |
duke@435 | 1939 | |
duke@435 | 1940 | |
duke@435 | 1941 | //------------------------------uncommon_trap---------------------------------- |
duke@435 | 1942 | // Bail out to the interpreter in mid-method. Implemented by calling the |
duke@435 | 1943 | // uncommon_trap blob. This helper function inserts a runtime call with the |
duke@435 | 1944 | // right debug info. |
duke@435 | 1945 | void GraphKit::uncommon_trap(int trap_request, |
duke@435 | 1946 | ciKlass* klass, const char* comment, |
duke@435 | 1947 | bool must_throw, |
duke@435 | 1948 | bool keep_exact_action) { |
duke@435 | 1949 | if (failing()) stop(); |
duke@435 | 1950 | if (stopped()) return; // trap reachable? |
duke@435 | 1951 | |
duke@435 | 1952 | // Note: If ProfileTraps is true, and if a deopt. actually |
duke@435 | 1953 | // occurs here, the runtime will make sure an MDO exists. There is |
iveresov@2349 | 1954 | // no need to call method()->ensure_method_data() at this point. |
duke@435 | 1955 | |
twisti@4313 | 1956 | // Set the stack pointer to the right value for reexecution: |
twisti@4313 | 1957 | set_sp(reexecute_sp()); |
twisti@4313 | 1958 | |
duke@435 | 1959 | #ifdef ASSERT |
duke@435 | 1960 | if (!must_throw) { |
duke@435 | 1961 | // Make sure the stack has at least enough depth to execute |
duke@435 | 1962 | // the current bytecode. |
twisti@4313 | 1963 | int inputs, ignored_depth; |
twisti@4313 | 1964 | if (compute_stack_effects(inputs, ignored_depth)) { |
twisti@4313 | 1965 | assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", |
twisti@4313 | 1966 | Bytecodes::name(java_bc()), sp(), inputs)); |
duke@435 | 1967 | } |
duke@435 | 1968 | } |
duke@435 | 1969 | #endif |
duke@435 | 1970 | |
duke@435 | 1971 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); |
duke@435 | 1972 | Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); |
duke@435 | 1973 | |
duke@435 | 1974 | switch (action) { |
duke@435 | 1975 | case Deoptimization::Action_maybe_recompile: |
duke@435 | 1976 | case Deoptimization::Action_reinterpret: |
duke@435 | 1977 | // Temporary fix for 6529811 to allow virtual calls to be sure they |
duke@435 | 1978 | // get the chance to go from mono->bi->mega |
duke@435 | 1979 | if (!keep_exact_action && |
duke@435 | 1980 | Deoptimization::trap_request_index(trap_request) < 0 && |
duke@435 | 1981 | too_many_recompiles(reason)) { |
duke@435 | 1982 | // This BCI is causing too many recompilations. |
duke@435 | 1983 | action = Deoptimization::Action_none; |
duke@435 | 1984 | trap_request = Deoptimization::make_trap_request(reason, action); |
duke@435 | 1985 | } else { |
duke@435 | 1986 | C->set_trap_can_recompile(true); |
duke@435 | 1987 | } |
duke@435 | 1988 | break; |
duke@435 | 1989 | case Deoptimization::Action_make_not_entrant: |
duke@435 | 1990 | C->set_trap_can_recompile(true); |
duke@435 | 1991 | break; |
duke@435 | 1992 | #ifdef ASSERT |
duke@435 | 1993 | case Deoptimization::Action_none: |
duke@435 | 1994 | case Deoptimization::Action_make_not_compilable: |
duke@435 | 1995 | break; |
duke@435 | 1996 | default: |
twisti@4313 | 1997 | fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); |
twisti@4313 | 1998 | break; |
duke@435 | 1999 | #endif |
duke@435 | 2000 | } |
duke@435 | 2001 | |
duke@435 | 2002 | if (TraceOptoParse) { |
duke@435 | 2003 | char buf[100]; |
duke@435 | 2004 | tty->print_cr("Uncommon trap %s at bci:%d", |
duke@435 | 2005 | Deoptimization::format_trap_request(buf, sizeof(buf), |
duke@435 | 2006 | trap_request), bci()); |
duke@435 | 2007 | } |
duke@435 | 2008 | |
duke@435 | 2009 | CompileLog* log = C->log(); |
duke@435 | 2010 | if (log != NULL) { |
duke@435 | 2011 | int kid = (klass == NULL)? -1: log->identify(klass); |
duke@435 | 2012 | log->begin_elem("uncommon_trap bci='%d'", bci()); |
duke@435 | 2013 | char buf[100]; |
duke@435 | 2014 | log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), |
duke@435 | 2015 | trap_request)); |
duke@435 | 2016 | if (kid >= 0) log->print(" klass='%d'", kid); |
duke@435 | 2017 | if (comment != NULL) log->print(" comment='%s'", comment); |
duke@435 | 2018 | log->end_elem(); |
duke@435 | 2019 | } |
duke@435 | 2020 | |
duke@435 | 2021 | // Make sure any guarding test views this path as very unlikely |
duke@435 | 2022 | Node *i0 = control()->in(0); |
duke@435 | 2023 | if (i0 != NULL && i0->is_If()) { // Found a guarding if test? |
duke@435 | 2024 | IfNode *iff = i0->as_If(); |
duke@435 | 2025 | float f = iff->_prob; // Get prob |
duke@435 | 2026 | if (control()->Opcode() == Op_IfTrue) { |
duke@435 | 2027 | if (f > PROB_UNLIKELY_MAG(4)) |
duke@435 | 2028 | iff->_prob = PROB_MIN; |
duke@435 | 2029 | } else { |
duke@435 | 2030 | if (f < PROB_LIKELY_MAG(4)) |
duke@435 | 2031 | iff->_prob = PROB_MAX; |
duke@435 | 2032 | } |
duke@435 | 2033 | } |
duke@435 | 2034 | |
duke@435 | 2035 | // Clear out dead values from the debug info. |
duke@435 | 2036 | kill_dead_locals(); |
duke@435 | 2037 | |
duke@435 | 2038 | // Now insert the uncommon trap subroutine call |
twisti@2103 | 2039 | address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); |
duke@435 | 2040 | const TypePtr* no_memory_effects = NULL; |
duke@435 | 2041 | // Pass the index of the class to be loaded |
duke@435 | 2042 | Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | |
duke@435 | 2043 | (must_throw ? RC_MUST_THROW : 0), |
duke@435 | 2044 | OptoRuntime::uncommon_trap_Type(), |
duke@435 | 2045 | call_addr, "uncommon_trap", no_memory_effects, |
duke@435 | 2046 | intcon(trap_request)); |
duke@435 | 2047 | assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, |
duke@435 | 2048 | "must extract request correctly from the graph"); |
duke@435 | 2049 | assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); |
duke@435 | 2050 | |
duke@435 | 2051 | call->set_req(TypeFunc::ReturnAdr, returnadr()); |
duke@435 | 2052 | // The debug info is the only real input to this call. |
duke@435 | 2053 | |
duke@435 | 2054 | // Halt-and-catch fire here. The above call should never return! |
kvn@4115 | 2055 | HaltNode* halt = new(C) HaltNode(control(), frameptr()); |
duke@435 | 2056 | _gvn.set_type_bottom(halt); |
duke@435 | 2057 | root()->add_req(halt); |
duke@435 | 2058 | |
duke@435 | 2059 | stop_and_kill_map(); |
duke@435 | 2060 | } |
duke@435 | 2061 | |
duke@435 | 2062 | |
duke@435 | 2063 | //--------------------------just_allocated_object------------------------------ |
duke@435 | 2064 | // Report the object that was just allocated. |
duke@435 | 2065 | // It must be the case that there are no intervening safepoints. |
duke@435 | 2066 | // We use this to determine if an object is so "fresh" that |
duke@435 | 2067 | // it does not require card marks. |
duke@435 | 2068 | Node* GraphKit::just_allocated_object(Node* current_control) { |
duke@435 | 2069 | if (C->recent_alloc_ctl() == current_control) |
duke@435 | 2070 | return C->recent_alloc_obj(); |
duke@435 | 2071 | return NULL; |
duke@435 | 2072 | } |
duke@435 | 2073 | |
duke@435 | 2074 | |
duke@435 | 2075 | void GraphKit::round_double_arguments(ciMethod* dest_method) { |
duke@435 | 2076 | // (Note: TypeFunc::make has a cache that makes this fast.) |
duke@435 | 2077 | const TypeFunc* tf = TypeFunc::make(dest_method); |
duke@435 | 2078 | int nargs = tf->_domain->_cnt - TypeFunc::Parms; |
duke@435 | 2079 | for (int j = 0; j < nargs; j++) { |
duke@435 | 2080 | const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); |
duke@435 | 2081 | if( targ->basic_type() == T_DOUBLE ) { |
duke@435 | 2082 | // If any parameters are doubles, they must be rounded before |
duke@435 | 2083 | // the call, dstore_rounding does gvn.transform |
duke@435 | 2084 | Node *arg = argument(j); |
duke@435 | 2085 | arg = dstore_rounding(arg); |
duke@435 | 2086 | set_argument(j, arg); |
duke@435 | 2087 | } |
duke@435 | 2088 | } |
duke@435 | 2089 | } |
duke@435 | 2090 | |
roland@5991 | 2091 | /** |
roland@5991 | 2092 | * Record profiling data exact_kls for Node n with the type system so |
roland@5991 | 2093 | * that it can propagate it (speculation) |
roland@5991 | 2094 | * |
roland@5991 | 2095 | * @param n node that the type applies to |
roland@5991 | 2096 | * @param exact_kls type from profiling |
roland@5991 | 2097 | * |
roland@5991 | 2098 | * @return node with improved type |
roland@5991 | 2099 | */ |
roland@5991 | 2100 | Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) { |
roland@6380 | 2101 | const Type* current_type = _gvn.type(n); |
roland@5991 | 2102 | assert(UseTypeSpeculation, "type speculation must be on"); |
roland@6380 | 2103 | |
roland@6380 | 2104 | const TypeOopPtr* speculative = current_type->speculative(); |
roland@6380 | 2105 | |
roland@6380 | 2106 | if (current_type->would_improve_type(exact_kls, jvms()->depth())) { |
roland@5991 | 2107 | const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); |
roland@5991 | 2108 | const TypeOopPtr* xtype = tklass->as_instance_type(); |
roland@5991 | 2109 | assert(xtype->klass_is_exact(), "Should be exact"); |
roland@6380 | 2110 | // record the new speculative type's depth |
roland@6380 | 2111 | speculative = xtype->with_inline_depth(jvms()->depth()); |
roland@6380 | 2112 | } |
roland@6380 | 2113 | |
roland@6380 | 2114 | if (speculative != current_type->speculative()) { |
roland@5991 | 2115 | // Build a type with a speculative type (what we think we know |
roland@5991 | 2116 | // about the type but will need a guard when we use it) |
roland@6380 | 2117 | const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); |
roland@6380 | 2118 | // We're changing the type, we need a new CheckCast node to carry |
roland@6380 | 2119 | // the new type. The new type depends on the control: what |
roland@6380 | 2120 | // profiling tells us is only valid from here as far as we can |
roland@6380 | 2121 | // tell. |
roland@6380 | 2122 | Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); |
roland@5991 | 2123 | cast = _gvn.transform(cast); |
roland@5991 | 2124 | replace_in_map(n, cast); |
roland@5991 | 2125 | n = cast; |
roland@5991 | 2126 | } |
roland@6380 | 2127 | |
roland@5991 | 2128 | return n; |
roland@5991 | 2129 | } |
roland@5991 | 2130 | |
roland@5991 | 2131 | /** |
roland@5991 | 2132 | * Record profiling data from receiver profiling at an invoke with the |
roland@5991 | 2133 | * type system so that it can propagate it (speculation) |
roland@5991 | 2134 | * |
roland@5991 | 2135 | * @param n receiver node |
roland@5991 | 2136 | * |
roland@6377 | 2137 | * @return node with improved type |
roland@5991 | 2138 | */ |
roland@5991 | 2139 | Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { |
roland@5991 | 2140 | if (!UseTypeSpeculation) { |
roland@5991 | 2141 | return n; |
roland@5991 | 2142 | } |
roland@5991 | 2143 | ciKlass* exact_kls = profile_has_unique_klass(); |
roland@5991 | 2144 | return record_profile_for_speculation(n, exact_kls); |
roland@5991 | 2145 | } |
roland@5991 | 2146 | |
roland@5991 | 2147 | /** |
roland@5991 | 2148 | * Record profiling data from argument profiling at an invoke with the |
roland@5991 | 2149 | * type system so that it can propagate it (speculation) |
roland@5991 | 2150 | * |
roland@5991 | 2151 | * @param dest_method target method for the call |
roland@5991 | 2152 | * @param bc what invoke bytecode is this? |
roland@5991 | 2153 | */ |
roland@5991 | 2154 | void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { |
roland@5991 | 2155 | if (!UseTypeSpeculation) { |
roland@5991 | 2156 | return; |
roland@5991 | 2157 | } |
roland@5991 | 2158 | const TypeFunc* tf = TypeFunc::make(dest_method); |
roland@5991 | 2159 | int nargs = tf->_domain->_cnt - TypeFunc::Parms; |
roland@5991 | 2160 | int skip = Bytecodes::has_receiver(bc) ? 1 : 0; |
roland@5991 | 2161 | for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { |
roland@5991 | 2162 | const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); |
roland@5991 | 2163 | if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { |
roland@5991 | 2164 | ciKlass* better_type = method()->argument_profiled_type(bci(), i); |
roland@5991 | 2165 | if (better_type != NULL) { |
roland@5991 | 2166 | record_profile_for_speculation(argument(j), better_type); |
roland@5991 | 2167 | } |
roland@5991 | 2168 | i++; |
roland@5991 | 2169 | } |
roland@5991 | 2170 | } |
roland@5991 | 2171 | } |
roland@5991 | 2172 | |
roland@5991 | 2173 | /** |
roland@5991 | 2174 | * Record profiling data from parameter profiling at an invoke with |
roland@5991 | 2175 | * the type system so that it can propagate it (speculation) |
roland@5991 | 2176 | */ |
roland@5991 | 2177 | void GraphKit::record_profiled_parameters_for_speculation() { |
roland@5991 | 2178 | if (!UseTypeSpeculation) { |
roland@5991 | 2179 | return; |
roland@5991 | 2180 | } |
roland@5991 | 2181 | for (int i = 0, j = 0; i < method()->arg_size() ; i++) { |
roland@5991 | 2182 | if (_gvn.type(local(i))->isa_oopptr()) { |
roland@5991 | 2183 | ciKlass* better_type = method()->parameter_profiled_type(j); |
roland@5991 | 2184 | if (better_type != NULL) { |
roland@5991 | 2185 | record_profile_for_speculation(local(i), better_type); |
roland@5991 | 2186 | } |
roland@5991 | 2187 | j++; |
roland@5991 | 2188 | } |
roland@5991 | 2189 | } |
roland@5991 | 2190 | } |
roland@5991 | 2191 | |
duke@435 | 2192 | void GraphKit::round_double_result(ciMethod* dest_method) { |
duke@435 | 2193 | // A non-strict method may return a double value which has an extended |
duke@435 | 2194 | // exponent, but this must not be visible in a caller which is 'strict' |
duke@435 | 2195 | // If a strict caller invokes a non-strict callee, round a double result |
duke@435 | 2196 | |
duke@435 | 2197 | BasicType result_type = dest_method->return_type()->basic_type(); |
duke@435 | 2198 | assert( method() != NULL, "must have caller context"); |
duke@435 | 2199 | if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { |
duke@435 | 2200 | // Destination method's return value is on top of stack |
duke@435 | 2201 | // dstore_rounding() does gvn.transform |
duke@435 | 2202 | Node *result = pop_pair(); |
duke@435 | 2203 | result = dstore_rounding(result); |
duke@435 | 2204 | push_pair(result); |
duke@435 | 2205 | } |
duke@435 | 2206 | } |
duke@435 | 2207 | |
duke@435 | 2208 | // rounding for strict float precision conformance |
duke@435 | 2209 | Node* GraphKit::precision_rounding(Node* n) { |
duke@435 | 2210 | return UseStrictFP && _method->flags().is_strict() |
duke@435 | 2211 | && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding |
kvn@4115 | 2212 | ? _gvn.transform( new (C) RoundFloatNode(0, n) ) |
duke@435 | 2213 | : n; |
duke@435 | 2214 | } |
duke@435 | 2215 | |
duke@435 | 2216 | // rounding for strict double precision conformance |
duke@435 | 2217 | Node* GraphKit::dprecision_rounding(Node *n) { |
duke@435 | 2218 | return UseStrictFP && _method->flags().is_strict() |
duke@435 | 2219 | && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding |
kvn@4115 | 2220 | ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) |
duke@435 | 2221 | : n; |
duke@435 | 2222 | } |
duke@435 | 2223 | |
duke@435 | 2224 | // rounding for non-strict double stores |
duke@435 | 2225 | Node* GraphKit::dstore_rounding(Node* n) { |
duke@435 | 2226 | return Matcher::strict_fp_requires_explicit_rounding |
duke@435 | 2227 | && UseSSE <= 1 |
kvn@4115 | 2228 | ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) |
duke@435 | 2229 | : n; |
duke@435 | 2230 | } |
duke@435 | 2231 | |
duke@435 | 2232 | //============================================================================= |
duke@435 | 2233 | // Generate a fast path/slow path idiom. Graph looks like: |
duke@435 | 2234 | // [foo] indicates that 'foo' is a parameter |
duke@435 | 2235 | // |
duke@435 | 2236 | // [in] NULL |
duke@435 | 2237 | // \ / |
duke@435 | 2238 | // CmpP |
duke@435 | 2239 | // Bool ne |
duke@435 | 2240 | // If |
duke@435 | 2241 | // / \ |
duke@435 | 2242 | // True False-<2> |
duke@435 | 2243 | // / | |
duke@435 | 2244 | // / cast_not_null |
duke@435 | 2245 | // Load | | ^ |
duke@435 | 2246 | // [fast_test] | | |
duke@435 | 2247 | // gvn to opt_test | | |
duke@435 | 2248 | // / \ | <1> |
duke@435 | 2249 | // True False | |
duke@435 | 2250 | // | \\ | |
duke@435 | 2251 | // [slow_call] \[fast_result] |
duke@435 | 2252 | // Ctl Val \ \ |
duke@435 | 2253 | // | \ \ |
duke@435 | 2254 | // Catch <1> \ \ |
duke@435 | 2255 | // / \ ^ \ \ |
duke@435 | 2256 | // Ex No_Ex | \ \ |
duke@435 | 2257 | // | \ \ | \ <2> \ |
duke@435 | 2258 | // ... \ [slow_res] | | \ [null_result] |
duke@435 | 2259 | // \ \--+--+--- | | |
duke@435 | 2260 | // \ | / \ | / |
duke@435 | 2261 | // --------Region Phi |
duke@435 | 2262 | // |
duke@435 | 2263 | //============================================================================= |
duke@435 | 2264 | // Code is structured as a series of driver functions all called 'do_XXX' that |
duke@435 | 2265 | // call a set of helper functions. Helper functions first, then drivers. |
duke@435 | 2266 | |
duke@435 | 2267 | //------------------------------null_check_oop--------------------------------- |
duke@435 | 2268 | // Null check oop. Set null-path control into Region in slot 3. |
duke@435 | 2269 | // Make a cast-not-nullness use the other not-null control. Return cast. |
duke@435 | 2270 | Node* GraphKit::null_check_oop(Node* value, Node* *null_control, |
roland@5908 | 2271 | bool never_see_null, bool safe_for_replace) { |
duke@435 | 2272 | // Initial NULL check taken path |
duke@435 | 2273 | (*null_control) = top(); |
duke@435 | 2274 | Node* cast = null_check_common(value, T_OBJECT, false, null_control); |
duke@435 | 2275 | |
duke@435 | 2276 | // Generate uncommon_trap: |
duke@435 | 2277 | if (never_see_null && (*null_control) != top()) { |
duke@435 | 2278 | // If we see an unexpected null at a check-cast we record it and force a |
duke@435 | 2279 | // recompile; the offending check-cast will be compiled to handle NULLs. |
duke@435 | 2280 | // If we see more than one offending BCI, then all checkcasts in the |
duke@435 | 2281 | // method will be compiled to handle NULLs. |
duke@435 | 2282 | PreserveJVMState pjvms(this); |
duke@435 | 2283 | set_control(*null_control); |
kvn@767 | 2284 | replace_in_map(value, null()); |
duke@435 | 2285 | uncommon_trap(Deoptimization::Reason_null_check, |
duke@435 | 2286 | Deoptimization::Action_make_not_entrant); |
duke@435 | 2287 | (*null_control) = top(); // NULL path is dead |
duke@435 | 2288 | } |
roland@5908 | 2289 | if ((*null_control) == top() && safe_for_replace) { |
roland@5908 | 2290 | replace_in_map(value, cast); |
roland@5908 | 2291 | } |
duke@435 | 2292 | |
duke@435 | 2293 | // Cast away null-ness on the result |
duke@435 | 2294 | return cast; |
duke@435 | 2295 | } |
duke@435 | 2296 | |
duke@435 | 2297 | //------------------------------opt_iff---------------------------------------- |
duke@435 | 2298 | // Optimize the fast-check IfNode. Set the fast-path region slot 2. |
duke@435 | 2299 | // Return slow-path control. |
duke@435 | 2300 | Node* GraphKit::opt_iff(Node* region, Node* iff) { |
duke@435 | 2301 | IfNode *opt_iff = _gvn.transform(iff)->as_If(); |
duke@435 | 2302 | |
duke@435 | 2303 | // Fast path taken; set region slot 2 |
kvn@4115 | 2304 | Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); |
duke@435 | 2305 | region->init_req(2,fast_taken); // Capture fast-control |
duke@435 | 2306 | |
duke@435 | 2307 | // Fast path not-taken, i.e. slow path |
kvn@4115 | 2308 | Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); |
duke@435 | 2309 | return slow_taken; |
duke@435 | 2310 | } |
duke@435 | 2311 | |
duke@435 | 2312 | //-----------------------------make_runtime_call------------------------------- |
duke@435 | 2313 | Node* GraphKit::make_runtime_call(int flags, |
duke@435 | 2314 | const TypeFunc* call_type, address call_addr, |
duke@435 | 2315 | const char* call_name, |
duke@435 | 2316 | const TypePtr* adr_type, |
duke@435 | 2317 | // The following parms are all optional. |
duke@435 | 2318 | // The first NULL ends the list. |
duke@435 | 2319 | Node* parm0, Node* parm1, |
duke@435 | 2320 | Node* parm2, Node* parm3, |
duke@435 | 2321 | Node* parm4, Node* parm5, |
duke@435 | 2322 | Node* parm6, Node* parm7) { |
duke@435 | 2323 | // Slow-path call |
duke@435 | 2324 | bool is_leaf = !(flags & RC_NO_LEAF); |
duke@435 | 2325 | bool has_io = (!is_leaf && !(flags & RC_NO_IO)); |
duke@435 | 2326 | if (call_name == NULL) { |
duke@435 | 2327 | assert(!is_leaf, "must supply name for leaf"); |
duke@435 | 2328 | call_name = OptoRuntime::stub_name(call_addr); |
duke@435 | 2329 | } |
duke@435 | 2330 | CallNode* call; |
duke@435 | 2331 | if (!is_leaf) { |
kvn@4115 | 2332 | call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, |
duke@435 | 2333 | bci(), adr_type); |
duke@435 | 2334 | } else if (flags & RC_NO_FP) { |
kvn@4115 | 2335 | call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); |
duke@435 | 2336 | } else { |
kvn@4115 | 2337 | call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); |
duke@435 | 2338 | } |
duke@435 | 2339 | |
duke@435 | 2340 | // The following is similar to set_edges_for_java_call, |
duke@435 | 2341 | // except that the memory effects of the call are restricted to AliasIdxRaw. |
duke@435 | 2342 | |
duke@435 | 2343 | // Slow path call has no side-effects, uses few values |
duke@435 | 2344 | bool wide_in = !(flags & RC_NARROW_MEM); |
duke@435 | 2345 | bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); |
duke@435 | 2346 | |
duke@435 | 2347 | Node* prev_mem = NULL; |
duke@435 | 2348 | if (wide_in) { |
duke@435 | 2349 | prev_mem = set_predefined_input_for_runtime_call(call); |
duke@435 | 2350 | } else { |
duke@435 | 2351 | assert(!wide_out, "narrow in => narrow out"); |
duke@435 | 2352 | Node* narrow_mem = memory(adr_type); |
duke@435 | 2353 | prev_mem = reset_memory(); |
duke@435 | 2354 | map()->set_memory(narrow_mem); |
duke@435 | 2355 | set_predefined_input_for_runtime_call(call); |
duke@435 | 2356 | } |
duke@435 | 2357 | |
duke@435 | 2358 | // Hook each parm in order. Stop looking at the first NULL. |
duke@435 | 2359 | if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); |
duke@435 | 2360 | if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); |
duke@435 | 2361 | if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); |
duke@435 | 2362 | if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); |
duke@435 | 2363 | if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); |
duke@435 | 2364 | if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); |
duke@435 | 2365 | if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); |
duke@435 | 2366 | if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); |
duke@435 | 2367 | /* close each nested if ===> */ } } } } } } } } |
duke@435 | 2368 | assert(call->in(call->req()-1) != NULL, "must initialize all parms"); |
duke@435 | 2369 | |
duke@435 | 2370 | if (!is_leaf) { |
duke@435 | 2371 | // Non-leaves can block and take safepoints: |
duke@435 | 2372 | add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); |
duke@435 | 2373 | } |
duke@435 | 2374 | // Non-leaves can throw exceptions: |
duke@435 | 2375 | if (has_io) { |
duke@435 | 2376 | call->set_req(TypeFunc::I_O, i_o()); |
duke@435 | 2377 | } |
duke@435 | 2378 | |
duke@435 | 2379 | if (flags & RC_UNCOMMON) { |
duke@435 | 2380 | // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. |
duke@435 | 2381 | // (An "if" probability corresponds roughly to an unconditional count. |
duke@435 | 2382 | // Sort of.) |
duke@435 | 2383 | call->set_cnt(PROB_UNLIKELY_MAG(4)); |
duke@435 | 2384 | } |
duke@435 | 2385 | |
duke@435 | 2386 | Node* c = _gvn.transform(call); |
duke@435 | 2387 | assert(c == call, "cannot disappear"); |
duke@435 | 2388 | |
duke@435 | 2389 | if (wide_out) { |
duke@435 | 2390 | // Slow path call has full side-effects. |
duke@435 | 2391 | set_predefined_output_for_runtime_call(call); |
duke@435 | 2392 | } else { |
duke@435 | 2393 | // Slow path call has few side-effects, and/or sets few values. |
duke@435 | 2394 | set_predefined_output_for_runtime_call(call, prev_mem, adr_type); |
duke@435 | 2395 | } |
duke@435 | 2396 | |
duke@435 | 2397 | if (has_io) { |
kvn@4115 | 2398 | set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); |
duke@435 | 2399 | } |
duke@435 | 2400 | return call; |
duke@435 | 2401 | |
duke@435 | 2402 | } |
duke@435 | 2403 | |
duke@435 | 2404 | //------------------------------merge_memory----------------------------------- |
duke@435 | 2405 | // Merge memory from one path into the current memory state. |
duke@435 | 2406 | void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { |
duke@435 | 2407 | for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { |
duke@435 | 2408 | Node* old_slice = mms.force_memory(); |
duke@435 | 2409 | Node* new_slice = mms.memory2(); |
duke@435 | 2410 | if (old_slice != new_slice) { |
duke@435 | 2411 | PhiNode* phi; |
kvn@7026 | 2412 | if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { |
kvn@7026 | 2413 | if (mms.is_empty()) { |
kvn@7026 | 2414 | // clone base memory Phi's inputs for this memory slice |
kvn@7026 | 2415 | assert(old_slice == mms.base_memory(), "sanity"); |
kvn@7026 | 2416 | phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); |
kvn@7026 | 2417 | _gvn.set_type(phi, Type::MEMORY); |
kvn@7026 | 2418 | for (uint i = 1; i < phi->req(); i++) { |
kvn@7026 | 2419 | phi->init_req(i, old_slice->in(i)); |
kvn@7026 | 2420 | } |
kvn@7026 | 2421 | } else { |
kvn@7026 | 2422 | phi = old_slice->as_Phi(); // Phi was generated already |
kvn@7026 | 2423 | } |
duke@435 | 2424 | } else { |
duke@435 | 2425 | phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); |
duke@435 | 2426 | _gvn.set_type(phi, Type::MEMORY); |
duke@435 | 2427 | } |
kvn@7026 | 2428 | phi->set_req(new_path, new_slice); |
kvn@7026 | 2429 | mms.set_memory(phi); |
duke@435 | 2430 | } |
duke@435 | 2431 | } |
duke@435 | 2432 | } |
duke@435 | 2433 | |
duke@435 | 2434 | //------------------------------make_slow_call_ex------------------------------ |
duke@435 | 2435 | // Make the exception handler hookups for the slow call |
vlivanov@6733 | 2436 | void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { |
duke@435 | 2437 | if (stopped()) return; |
duke@435 | 2438 | |
duke@435 | 2439 | // Make a catch node with just two handlers: fall-through and catch-all |
kvn@4115 | 2440 | Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); |
kvn@4115 | 2441 | Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); |
kvn@4115 | 2442 | Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); |
kvn@4115 | 2443 | Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); |
duke@435 | 2444 | |
duke@435 | 2445 | { PreserveJVMState pjvms(this); |
duke@435 | 2446 | set_control(excp); |
duke@435 | 2447 | set_i_o(i_o); |
duke@435 | 2448 | |
duke@435 | 2449 | if (excp != top()) { |
vlivanov@6733 | 2450 | if (deoptimize) { |
vlivanov@6733 | 2451 | // Deoptimize if an exception is caught. Don't construct exception state in this case. |
vlivanov@6733 | 2452 | uncommon_trap(Deoptimization::Reason_unhandled, |
vlivanov@6733 | 2453 | Deoptimization::Action_none); |
vlivanov@6733 | 2454 | } else { |
vlivanov@6733 | 2455 | // Create an exception state also. |
vlivanov@6733 | 2456 | // Use an exact type if the caller has specified a specific exception. |
vlivanov@6733 | 2457 | const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); |
vlivanov@6733 | 2458 | Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); |
vlivanov@6733 | 2459 | add_exception_state(make_exception_state(_gvn.transform(ex_oop))); |
vlivanov@6733 | 2460 | } |
duke@435 | 2461 | } |
duke@435 | 2462 | } |
duke@435 | 2463 | |
duke@435 | 2464 | // Get the no-exception control from the CatchNode. |
duke@435 | 2465 | set_control(norm); |
duke@435 | 2466 | } |
duke@435 | 2467 | |
duke@435 | 2468 | |
duke@435 | 2469 | //-------------------------------gen_subtype_check----------------------------- |
duke@435 | 2470 | // Generate a subtyping check. Takes as input the subtype and supertype. |
duke@435 | 2471 | // Returns 2 values: sets the default control() to the true path and returns |
duke@435 | 2472 | // the false path. Only reads invariant memory; sets no (visible) memory. |
duke@435 | 2473 | // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding |
duke@435 | 2474 | // but that's not exposed to the optimizer. This call also doesn't take in an |
duke@435 | 2475 | // Object; if you wish to check an Object you need to load the Object's class |
duke@435 | 2476 | // prior to coming here. |
duke@435 | 2477 | Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { |
duke@435 | 2478 | // Fast check for identical types, perhaps identical constants. |
duke@435 | 2479 | // The types can even be identical non-constants, in cases |
duke@435 | 2480 | // involving Array.newInstance, Object.clone, etc. |
duke@435 | 2481 | if (subklass == superklass) |
duke@435 | 2482 | return top(); // false path is dead; no test needed. |
duke@435 | 2483 | |
duke@435 | 2484 | if (_gvn.type(superklass)->singleton()) { |
duke@435 | 2485 | ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); |
duke@435 | 2486 | ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); |
duke@435 | 2487 | |
duke@435 | 2488 | // In the common case of an exact superklass, try to fold up the |
duke@435 | 2489 | // test before generating code. You may ask, why not just generate |
duke@435 | 2490 | // the code and then let it fold up? The answer is that the generated |
duke@435 | 2491 | // code will necessarily include null checks, which do not always |
duke@435 | 2492 | // completely fold away. If they are also needless, then they turn |
duke@435 | 2493 | // into a performance loss. Example: |
duke@435 | 2494 | // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; |
duke@435 | 2495 | // Here, the type of 'fa' is often exact, so the store check |
duke@435 | 2496 | // of fa[1]=x will fold up, without testing the nullness of x. |
duke@435 | 2497 | switch (static_subtype_check(superk, subk)) { |
duke@435 | 2498 | case SSC_always_false: |
duke@435 | 2499 | { |
duke@435 | 2500 | Node* always_fail = control(); |
duke@435 | 2501 | set_control(top()); |
duke@435 | 2502 | return always_fail; |
duke@435 | 2503 | } |
duke@435 | 2504 | case SSC_always_true: |
duke@435 | 2505 | return top(); |
duke@435 | 2506 | case SSC_easy_test: |
duke@435 | 2507 | { |
duke@435 | 2508 | // Just do a direct pointer compare and be done. |
kvn@4115 | 2509 | Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); |
kvn@4115 | 2510 | Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); |
duke@435 | 2511 | IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
kvn@4115 | 2512 | set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); |
kvn@4115 | 2513 | return _gvn.transform( new(C) IfFalseNode(iff) ); |
duke@435 | 2514 | } |
duke@435 | 2515 | case SSC_full_test: |
duke@435 | 2516 | break; |
duke@435 | 2517 | default: |
duke@435 | 2518 | ShouldNotReachHere(); |
duke@435 | 2519 | } |
duke@435 | 2520 | } |
duke@435 | 2521 | |
duke@435 | 2522 | // %%% Possible further optimization: Even if the superklass is not exact, |
duke@435 | 2523 | // if the subklass is the unique subtype of the superklass, the check |
duke@435 | 2524 | // will always succeed. We could leave a dependency behind to ensure this. |
duke@435 | 2525 | |
duke@435 | 2526 | // First load the super-klass's check-offset |
stefank@3391 | 2527 | Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); |
goetz@6479 | 2528 | Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(), |
goetz@6479 | 2529 | TypeInt::INT, MemNode::unordered)); |
stefank@3391 | 2530 | int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); |
duke@435 | 2531 | bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); |
duke@435 | 2532 | |
duke@435 | 2533 | // Load from the sub-klass's super-class display list, or a 1-word cache of |
duke@435 | 2534 | // the secondary superclass list, or a failing value with a sentinel offset |
duke@435 | 2535 | // if the super-klass is an interface or exceptionally deep in the Java |
duke@435 | 2536 | // hierarchy and we have to scan the secondary superclass list the hard way. |
duke@435 | 2537 | // Worst-case type is a little odd: NULL is allowed as a result (usually |
duke@435 | 2538 | // klass loads can never produce a NULL). |
duke@435 | 2539 | Node *chk_off_X = ConvI2X(chk_off); |
kvn@4115 | 2540 | Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); |
duke@435 | 2541 | // For some types like interfaces the following loadKlass is from a 1-word |
duke@435 | 2542 | // cache which is mutable so can't use immutable memory. Other |
duke@435 | 2543 | // types load from the super-class display table which is immutable. |
duke@435 | 2544 | Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); |
kvn@599 | 2545 | Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); |
duke@435 | 2546 | |
duke@435 | 2547 | // Compile speed common case: ARE a subtype and we canNOT fail |
duke@435 | 2548 | if( superklass == nkls ) |
duke@435 | 2549 | return top(); // false path is dead; no test needed. |
duke@435 | 2550 | |
duke@435 | 2551 | // See if we get an immediate positive hit. Happens roughly 83% of the |
duke@435 | 2552 | // time. Test to see if the value loaded just previously from the subklass |
duke@435 | 2553 | // is exactly the superklass. |
kvn@4115 | 2554 | Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); |
kvn@4115 | 2555 | Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); |
duke@435 | 2556 | IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); |
kvn@4115 | 2557 | Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); |
kvn@4115 | 2558 | set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); |
duke@435 | 2559 | |
duke@435 | 2560 | // Compile speed common case: Check for being deterministic right now. If |
duke@435 | 2561 | // chk_off is a constant and not equal to cacheoff then we are NOT a |
duke@435 | 2562 | // subklass. In this case we need exactly the 1 test above and we can |
duke@435 | 2563 | // return those results immediately. |
duke@435 | 2564 | if (!might_be_cache) { |
duke@435 | 2565 | Node* not_subtype_ctrl = control(); |
duke@435 | 2566 | set_control(iftrue1); // We need exactly the 1 test above |
duke@435 | 2567 | return not_subtype_ctrl; |
duke@435 | 2568 | } |
duke@435 | 2569 | |
duke@435 | 2570 | // Gather the various success & failures here |
kvn@4115 | 2571 | RegionNode *r_ok_subtype = new (C) RegionNode(4); |
duke@435 | 2572 | record_for_igvn(r_ok_subtype); |
kvn@4115 | 2573 | RegionNode *r_not_subtype = new (C) RegionNode(3); |
duke@435 | 2574 | record_for_igvn(r_not_subtype); |
duke@435 | 2575 | |
duke@435 | 2576 | r_ok_subtype->init_req(1, iftrue1); |
duke@435 | 2577 | |
duke@435 | 2578 | // Check for immediate negative hit. Happens roughly 11% of the time (which |
duke@435 | 2579 | // is roughly 63% of the remaining cases). Test to see if the loaded |
duke@435 | 2580 | // check-offset points into the subklass display list or the 1-element |
duke@435 | 2581 | // cache. If it points to the display (and NOT the cache) and the display |
duke@435 | 2582 | // missed then it's not a subtype. |
duke@435 | 2583 | Node *cacheoff = _gvn.intcon(cacheoff_con); |
kvn@4115 | 2584 | Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); |
kvn@4115 | 2585 | Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); |
duke@435 | 2586 | IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); |
kvn@4115 | 2587 | r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); |
kvn@4115 | 2588 | set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); |
duke@435 | 2589 | |
jrose@1079 | 2590 | // Check for self. Very rare to get here, but it is taken 1/3 the time. |
duke@435 | 2591 | // No performance impact (too rare) but allows sharing of secondary arrays |
duke@435 | 2592 | // which has some footprint reduction. |
kvn@4115 | 2593 | Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); |
kvn@4115 | 2594 | Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); |
duke@435 | 2595 | IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); |
kvn@4115 | 2596 | r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); |
kvn@4115 | 2597 | set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); |
duke@435 | 2598 | |
jrose@1079 | 2599 | // -- Roads not taken here: -- |
jrose@1079 | 2600 | // We could also have chosen to perform the self-check at the beginning |
jrose@1079 | 2601 | // of this code sequence, as the assembler does. This would not pay off |
jrose@1079 | 2602 | // the same way, since the optimizer, unlike the assembler, can perform |
jrose@1079 | 2603 | // static type analysis to fold away many successful self-checks. |
jrose@1079 | 2604 | // Non-foldable self checks work better here in second position, because |
jrose@1079 | 2605 | // the initial primary superclass check subsumes a self-check for most |
jrose@1079 | 2606 | // types. An exception would be a secondary type like array-of-interface, |
jrose@1079 | 2607 | // which does not appear in its own primary supertype display. |
jrose@1079 | 2608 | // Finally, we could have chosen to move the self-check into the |
jrose@1079 | 2609 | // PartialSubtypeCheckNode, and from there out-of-line in a platform |
jrose@1079 | 2610 | // dependent manner. But it is worthwhile to have the check here, |
jrose@1079 | 2611 | // where it can be perhaps be optimized. The cost in code space is |
jrose@1079 | 2612 | // small (register compare, branch). |
jrose@1079 | 2613 | |
duke@435 | 2614 | // Now do a linear scan of the secondary super-klass array. Again, no real |
duke@435 | 2615 | // performance impact (too rare) but it's gotta be done. |
duke@435 | 2616 | // Since the code is rarely used, there is no penalty for moving it |
jrose@1079 | 2617 | // out of line, and it can only improve I-cache density. |
jrose@1079 | 2618 | // The decision to inline or out-of-line this final check is platform |
jrose@1079 | 2619 | // dependent, and is found in the AD file definition of PartialSubtypeCheck. |
duke@435 | 2620 | Node* psc = _gvn.transform( |
kvn@4115 | 2621 | new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); |
kvn@4115 | 2622 | |
kvn@4115 | 2623 | Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); |
kvn@4115 | 2624 | Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); |
duke@435 | 2625 | IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); |
kvn@4115 | 2626 | r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); |
kvn@4115 | 2627 | r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); |
duke@435 | 2628 | |
duke@435 | 2629 | // Return false path; set default control to true path. |
duke@435 | 2630 | set_control( _gvn.transform(r_ok_subtype) ); |
duke@435 | 2631 | return _gvn.transform(r_not_subtype); |
duke@435 | 2632 | } |
duke@435 | 2633 | |
duke@435 | 2634 | //----------------------------static_subtype_check----------------------------- |
duke@435 | 2635 | // Shortcut important common cases when superklass is exact: |
duke@435 | 2636 | // (0) superklass is java.lang.Object (can occur in reflective code) |
duke@435 | 2637 | // (1) subklass is already limited to a subtype of superklass => always ok |
duke@435 | 2638 | // (2) subklass does not overlap with superklass => always fail |
duke@435 | 2639 | // (3) superklass has NO subtypes and we can check with a simple compare. |
duke@435 | 2640 | int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { |
duke@435 | 2641 | if (StressReflectiveCode) { |
duke@435 | 2642 | return SSC_full_test; // Let caller generate the general case. |
duke@435 | 2643 | } |
duke@435 | 2644 | |
duke@435 | 2645 | if (superk == env()->Object_klass()) { |
duke@435 | 2646 | return SSC_always_true; // (0) this test cannot fail |
duke@435 | 2647 | } |
duke@435 | 2648 | |
duke@435 | 2649 | ciType* superelem = superk; |
duke@435 | 2650 | if (superelem->is_array_klass()) |
duke@435 | 2651 | superelem = superelem->as_array_klass()->base_element_type(); |
duke@435 | 2652 | |
duke@435 | 2653 | if (!subk->is_interface()) { // cannot trust static interface types yet |
duke@435 | 2654 | if (subk->is_subtype_of(superk)) { |
duke@435 | 2655 | return SSC_always_true; // (1) false path dead; no dynamic test needed |
duke@435 | 2656 | } |
duke@435 | 2657 | if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && |
duke@435 | 2658 | !superk->is_subtype_of(subk)) { |
duke@435 | 2659 | return SSC_always_false; |
duke@435 | 2660 | } |
duke@435 | 2661 | } |
duke@435 | 2662 | |
duke@435 | 2663 | // If casting to an instance klass, it must have no subtypes |
duke@435 | 2664 | if (superk->is_interface()) { |
duke@435 | 2665 | // Cannot trust interfaces yet. |
duke@435 | 2666 | // %%% S.B. superk->nof_implementors() == 1 |
duke@435 | 2667 | } else if (superelem->is_instance_klass()) { |
duke@435 | 2668 | ciInstanceKlass* ik = superelem->as_instance_klass(); |
duke@435 | 2669 | if (!ik->has_subklass() && !ik->is_interface()) { |
duke@435 | 2670 | if (!ik->is_final()) { |
duke@435 | 2671 | // Add a dependency if there is a chance of a later subclass. |
duke@435 | 2672 | C->dependencies()->assert_leaf_type(ik); |
duke@435 | 2673 | } |
duke@435 | 2674 | return SSC_easy_test; // (3) caller can do a simple ptr comparison |
duke@435 | 2675 | } |
duke@435 | 2676 | } else { |
duke@435 | 2677 | // A primitive array type has no subtypes. |
duke@435 | 2678 | return SSC_easy_test; // (3) caller can do a simple ptr comparison |
duke@435 | 2679 | } |
duke@435 | 2680 | |
duke@435 | 2681 | return SSC_full_test; |
duke@435 | 2682 | } |
duke@435 | 2683 | |
duke@435 | 2684 | // Profile-driven exact type check: |
duke@435 | 2685 | Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, |
duke@435 | 2686 | float prob, |
duke@435 | 2687 | Node* *casted_receiver) { |
duke@435 | 2688 | const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); |
duke@435 | 2689 | Node* recv_klass = load_object_klass(receiver); |
duke@435 | 2690 | Node* want_klass = makecon(tklass); |
kvn@4115 | 2691 | Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); |
kvn@4115 | 2692 | Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); |
duke@435 | 2693 | IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); |
kvn@4115 | 2694 | set_control( _gvn.transform( new(C) IfTrueNode (iff) )); |
kvn@4115 | 2695 | Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); |
duke@435 | 2696 | |
duke@435 | 2697 | const TypeOopPtr* recv_xtype = tklass->as_instance_type(); |
duke@435 | 2698 | assert(recv_xtype->klass_is_exact(), ""); |
duke@435 | 2699 | |
duke@435 | 2700 | // Subsume downstream occurrences of receiver with a cast to |
duke@435 | 2701 | // recv_xtype, since now we know what the type will be. |
kvn@4115 | 2702 | Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); |
duke@435 | 2703 | (*casted_receiver) = _gvn.transform(cast); |
duke@435 | 2704 | // (User must make the replace_in_map call.) |
duke@435 | 2705 | |
duke@435 | 2706 | return fail; |
duke@435 | 2707 | } |
duke@435 | 2708 | |
duke@435 | 2709 | |
jrose@2101 | 2710 | //------------------------------seems_never_null------------------------------- |
jrose@2101 | 2711 | // Use null_seen information if it is available from the profile. |
jrose@2101 | 2712 | // If we see an unexpected null at a type check we record it and force a |
jrose@2101 | 2713 | // recompile; the offending check will be recompiled to handle NULLs. |
jrose@2101 | 2714 | // If we see several offending BCIs, then all checks in the |
jrose@2101 | 2715 | // method will be recompiled. |
jrose@2101 | 2716 | bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { |
jrose@2101 | 2717 | if (UncommonNullCast // Cutout for this technique |
jrose@2101 | 2718 | && obj != null() // And not the -Xcomp stupid case? |
jrose@2101 | 2719 | && !too_many_traps(Deoptimization::Reason_null_check) |
jrose@2101 | 2720 | ) { |
jrose@2101 | 2721 | if (data == NULL) |
jrose@2101 | 2722 | // Edge case: no mature data. Be optimistic here. |
jrose@2101 | 2723 | return true; |
jrose@2101 | 2724 | // If the profile has not seen a null, assume it won't happen. |
jrose@2101 | 2725 | assert(java_bc() == Bytecodes::_checkcast || |
jrose@2101 | 2726 | java_bc() == Bytecodes::_instanceof || |
jrose@2101 | 2727 | java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); |
jrose@2101 | 2728 | return !data->as_BitData()->null_seen(); |
jrose@2101 | 2729 | } |
jrose@2101 | 2730 | return false; |
jrose@2101 | 2731 | } |
jrose@2101 | 2732 | |
jrose@2101 | 2733 | //------------------------maybe_cast_profiled_receiver------------------------- |
jrose@2101 | 2734 | // If the profile has seen exactly one type, narrow to exactly that type. |
jrose@2101 | 2735 | // Subsequent type checks will always fold up. |
jrose@2101 | 2736 | Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, |
roland@5991 | 2737 | ciKlass* require_klass, |
roland@6377 | 2738 | ciKlass* spec_klass, |
roland@5991 | 2739 | bool safe_for_replace) { |
jrose@2101 | 2740 | if (!UseTypeProfile || !TypeProfileCasts) return NULL; |
jrose@2101 | 2741 | |
roland@6377 | 2742 | Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check; |
roland@6377 | 2743 | |
jrose@2101 | 2744 | // Make sure we haven't already deoptimized from this tactic. |
roland@6377 | 2745 | if (too_many_traps(reason)) |
jrose@2101 | 2746 | return NULL; |
jrose@2101 | 2747 | |
jrose@2101 | 2748 | // (No, this isn't a call, but it's enough like a virtual call |
jrose@2101 | 2749 | // to use the same ciMethod accessor to get the profile info...) |
roland@5991 | 2750 | // If we have a speculative type use it instead of profiling (which |
roland@5991 | 2751 | // may not help us) |
roland@5991 | 2752 | ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; |
roland@5991 | 2753 | if (exact_kls != NULL) {// no cast failures here |
jrose@2101 | 2754 | if (require_klass == NULL || |
jrose@2101 | 2755 | static_subtype_check(require_klass, exact_kls) == SSC_always_true) { |
roland@5991 | 2756 | // If we narrow the type to match what the type profile sees or |
roland@5991 | 2757 | // the speculative type, we can then remove the rest of the |
roland@5991 | 2758 | // cast. |
jrose@2101 | 2759 | // This is a win, even if the exact_kls is very specific, |
jrose@2101 | 2760 | // because downstream operations, such as method calls, |
jrose@2101 | 2761 | // will often benefit from the sharper type. |
jrose@2101 | 2762 | Node* exact_obj = not_null_obj; // will get updated in place... |
jrose@2101 | 2763 | Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, |
jrose@2101 | 2764 | &exact_obj); |
jrose@2101 | 2765 | { PreserveJVMState pjvms(this); |
jrose@2101 | 2766 | set_control(slow_ctl); |
roland@6377 | 2767 | uncommon_trap(reason, |
jrose@2101 | 2768 | Deoptimization::Action_maybe_recompile); |
jrose@2101 | 2769 | } |
roland@5991 | 2770 | if (safe_for_replace) { |
roland@5991 | 2771 | replace_in_map(not_null_obj, exact_obj); |
roland@5991 | 2772 | } |
jrose@2101 | 2773 | return exact_obj; |
jrose@2101 | 2774 | } |
jrose@2101 | 2775 | // assert(ssc == SSC_always_true)... except maybe the profile lied to us. |
jrose@2101 | 2776 | } |
jrose@2101 | 2777 | |
jrose@2101 | 2778 | return NULL; |
jrose@2101 | 2779 | } |
jrose@2101 | 2780 | |
roland@5991 | 2781 | /** |
roland@5991 | 2782 | * Cast obj to type and emit guard unless we had too many traps here |
roland@5991 | 2783 | * already |
roland@5991 | 2784 | * |
roland@5991 | 2785 | * @param obj node being casted |
roland@5991 | 2786 | * @param type type to cast the node to |
roland@5991 | 2787 | * @param not_null true if we know node cannot be null |
roland@5991 | 2788 | */ |
roland@5991 | 2789 | Node* GraphKit::maybe_cast_profiled_obj(Node* obj, |
roland@5991 | 2790 | ciKlass* type, |
roland@5991 | 2791 | bool not_null) { |
roland@5991 | 2792 | // type == NULL if profiling tells us this object is always null |
roland@5991 | 2793 | if (type != NULL) { |
roland@6377 | 2794 | Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; |
roland@6377 | 2795 | Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check; |
roland@6377 | 2796 | if (!too_many_traps(null_reason) && |
roland@6377 | 2797 | !too_many_traps(class_reason)) { |
roland@5991 | 2798 | Node* not_null_obj = NULL; |
roland@5991 | 2799 | // not_null is true if we know the object is not null and |
roland@5991 | 2800 | // there's no need for a null check |
roland@5991 | 2801 | if (!not_null) { |
roland@5991 | 2802 | Node* null_ctl = top(); |
roland@5991 | 2803 | not_null_obj = null_check_oop(obj, &null_ctl, true, true); |
roland@5991 | 2804 | assert(null_ctl->is_top(), "no null control here"); |
roland@5991 | 2805 | } else { |
roland@5991 | 2806 | not_null_obj = obj; |
roland@5991 | 2807 | } |
roland@5991 | 2808 | |
roland@5991 | 2809 | Node* exact_obj = not_null_obj; |
roland@5991 | 2810 | ciKlass* exact_kls = type; |
roland@5991 | 2811 | Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, |
roland@5991 | 2812 | &exact_obj); |
roland@5991 | 2813 | { |
roland@5991 | 2814 | PreserveJVMState pjvms(this); |
roland@5991 | 2815 | set_control(slow_ctl); |
roland@6377 | 2816 | uncommon_trap(class_reason, |
roland@5991 | 2817 | Deoptimization::Action_maybe_recompile); |
roland@5991 | 2818 | } |
roland@5991 | 2819 | replace_in_map(not_null_obj, exact_obj); |
roland@5991 | 2820 | obj = exact_obj; |
roland@5991 | 2821 | } |
roland@5991 | 2822 | } else { |
roland@5991 | 2823 | if (!too_many_traps(Deoptimization::Reason_null_assert)) { |
roland@5991 | 2824 | Node* exact_obj = null_assert(obj); |
roland@5991 | 2825 | replace_in_map(obj, exact_obj); |
roland@5991 | 2826 | obj = exact_obj; |
roland@5991 | 2827 | } |
roland@5991 | 2828 | } |
roland@5991 | 2829 | return obj; |
roland@5991 | 2830 | } |
jrose@2101 | 2831 | |
duke@435 | 2832 | //-------------------------------gen_instanceof-------------------------------- |
duke@435 | 2833 | // Generate an instance-of idiom. Used by both the instance-of bytecode |
duke@435 | 2834 | // and the reflective instance-of call. |
roland@5991 | 2835 | Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { |
jrose@2101 | 2836 | kill_dead_locals(); // Benefit all the uncommon traps |
duke@435 | 2837 | assert( !stopped(), "dead parse path should be checked in callers" ); |
duke@435 | 2838 | assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), |
duke@435 | 2839 | "must check for not-null not-dead klass in callers"); |
duke@435 | 2840 | |
duke@435 | 2841 | // Make the merge point |
duke@435 | 2842 | enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; |
kvn@4115 | 2843 | RegionNode* region = new(C) RegionNode(PATH_LIMIT); |
kvn@4115 | 2844 | Node* phi = new(C) PhiNode(region, TypeInt::BOOL); |
duke@435 | 2845 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
duke@435 | 2846 | |
jrose@2101 | 2847 | ciProfileData* data = NULL; |
jrose@2101 | 2848 | if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode |
jrose@2101 | 2849 | data = method()->method_data()->bci_to_data(bci()); |
jrose@2101 | 2850 | } |
jrose@2101 | 2851 | bool never_see_null = (ProfileDynamicTypes // aggressive use of profile |
jrose@2101 | 2852 | && seems_never_null(obj, data)); |
jrose@2101 | 2853 | |
duke@435 | 2854 | // Null check; get casted pointer; set region slot 3 |
duke@435 | 2855 | Node* null_ctl = top(); |
roland@5908 | 2856 | Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); |
duke@435 | 2857 | |
duke@435 | 2858 | // If not_null_obj is dead, only null-path is taken |
duke@435 | 2859 | if (stopped()) { // Doing instance-of on a NULL? |
duke@435 | 2860 | set_control(null_ctl); |
duke@435 | 2861 | return intcon(0); |
duke@435 | 2862 | } |
duke@435 | 2863 | region->init_req(_null_path, null_ctl); |
duke@435 | 2864 | phi ->init_req(_null_path, intcon(0)); // Set null path value |
jrose@2101 | 2865 | if (null_ctl == top()) { |
jrose@2101 | 2866 | // Do this eagerly, so that pattern matches like is_diamond_phi |
jrose@2101 | 2867 | // will work even during parsing. |
jrose@2101 | 2868 | assert(_null_path == PATH_LIMIT-1, "delete last"); |
jrose@2101 | 2869 | region->del_req(_null_path); |
jrose@2101 | 2870 | phi ->del_req(_null_path); |
jrose@2101 | 2871 | } |
jrose@2101 | 2872 | |
roland@5991 | 2873 | // Do we know the type check always succeed? |
roland@5991 | 2874 | bool known_statically = false; |
roland@5991 | 2875 | if (_gvn.type(superklass)->singleton()) { |
roland@5991 | 2876 | ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); |
roland@5991 | 2877 | ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); |
roland@5991 | 2878 | if (subk != NULL && subk->is_loaded()) { |
roland@5991 | 2879 | int static_res = static_subtype_check(superk, subk); |
roland@5991 | 2880 | known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); |
jrose@2101 | 2881 | } |
roland@5991 | 2882 | } |
roland@5991 | 2883 | |
roland@5991 | 2884 | if (known_statically && UseTypeSpeculation) { |
roland@6377 | 2885 | // If we know the type check always succeeds then we don't use the |
roland@5991 | 2886 | // profiling data at this bytecode. Don't lose it, feed it to the |
roland@5991 | 2887 | // type system as a speculative type. |
roland@5991 | 2888 | not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); |
roland@5991 | 2889 | } else { |
roland@5991 | 2890 | const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); |
roland@5991 | 2891 | // We may not have profiling here or it may not help us. If we |
roland@5991 | 2892 | // have a speculative type use it to perform an exact cast. |
roland@5991 | 2893 | ciKlass* spec_obj_type = obj_type->speculative_type(); |
roland@5991 | 2894 | if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { |
roland@5991 | 2895 | Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); |
roland@5991 | 2896 | if (stopped()) { // Profile disagrees with this path. |
roland@5991 | 2897 | set_control(null_ctl); // Null is the only remaining possibility. |
roland@5991 | 2898 | return intcon(0); |
roland@5991 | 2899 | } |
roland@5991 | 2900 | if (cast_obj != NULL) { |
roland@5991 | 2901 | not_null_obj = cast_obj; |
roland@5991 | 2902 | } |
roland@5991 | 2903 | } |
jrose@2101 | 2904 | } |
duke@435 | 2905 | |
duke@435 | 2906 | // Load the object's klass |
duke@435 | 2907 | Node* obj_klass = load_object_klass(not_null_obj); |
duke@435 | 2908 | |
duke@435 | 2909 | // Generate the subtype check |
duke@435 | 2910 | Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); |
duke@435 | 2911 | |
duke@435 | 2912 | // Plug in the success path to the general merge in slot 1. |
duke@435 | 2913 | region->init_req(_obj_path, control()); |
duke@435 | 2914 | phi ->init_req(_obj_path, intcon(1)); |
duke@435 | 2915 | |
duke@435 | 2916 | // Plug in the failing path to the general merge in slot 2. |
duke@435 | 2917 | region->init_req(_fail_path, not_subtype_ctrl); |
duke@435 | 2918 | phi ->init_req(_fail_path, intcon(0)); |
duke@435 | 2919 | |
duke@435 | 2920 | // Return final merged results |
duke@435 | 2921 | set_control( _gvn.transform(region) ); |
duke@435 | 2922 | record_for_igvn(region); |
duke@435 | 2923 | return _gvn.transform(phi); |
duke@435 | 2924 | } |
duke@435 | 2925 | |
duke@435 | 2926 | //-------------------------------gen_checkcast--------------------------------- |
duke@435 | 2927 | // Generate a checkcast idiom. Used by both the checkcast bytecode and the |
duke@435 | 2928 | // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the |
duke@435 | 2929 | // uncommon-trap paths work. Adjust stack after this call. |
duke@435 | 2930 | // If failure_control is supplied and not null, it is filled in with |
duke@435 | 2931 | // the control edge for the cast failure. Otherwise, an appropriate |
duke@435 | 2932 | // uncommon trap or exception is thrown. |
duke@435 | 2933 | Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, |
duke@435 | 2934 | Node* *failure_control) { |
duke@435 | 2935 | kill_dead_locals(); // Benefit all the uncommon traps |
duke@435 | 2936 | const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); |
duke@435 | 2937 | const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); |
duke@435 | 2938 | |
duke@435 | 2939 | // Fast cutout: Check the case that the cast is vacuously true. |
duke@435 | 2940 | // This detects the common cases where the test will short-circuit |
duke@435 | 2941 | // away completely. We do this before we perform the null check, |
duke@435 | 2942 | // because if the test is going to turn into zero code, we don't |
duke@435 | 2943 | // want a residual null check left around. (Causes a slowdown, |
duke@435 | 2944 | // for example, in some objArray manipulations, such as a[i]=a[j].) |
duke@435 | 2945 | if (tk->singleton()) { |
duke@435 | 2946 | const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); |
duke@435 | 2947 | if (objtp != NULL && objtp->klass() != NULL) { |
duke@435 | 2948 | switch (static_subtype_check(tk->klass(), objtp->klass())) { |
duke@435 | 2949 | case SSC_always_true: |
roland@5991 | 2950 | // If we know the type check always succeed then we don't use |
roland@5991 | 2951 | // the profiling data at this bytecode. Don't lose it, feed it |
roland@5991 | 2952 | // to the type system as a speculative type. |
roland@5991 | 2953 | return record_profiled_receiver_for_speculation(obj); |
duke@435 | 2954 | case SSC_always_false: |
duke@435 | 2955 | // It needs a null check because a null will *pass* the cast check. |
duke@435 | 2956 | // A non-null value will always produce an exception. |
twisti@4313 | 2957 | return null_assert(obj); |
duke@435 | 2958 | } |
duke@435 | 2959 | } |
duke@435 | 2960 | } |
duke@435 | 2961 | |
duke@435 | 2962 | ciProfileData* data = NULL; |
roland@5908 | 2963 | bool safe_for_replace = false; |
duke@435 | 2964 | if (failure_control == NULL) { // use MDO in regular case only |
duke@435 | 2965 | assert(java_bc() == Bytecodes::_aastore || |
duke@435 | 2966 | java_bc() == Bytecodes::_checkcast, |
duke@435 | 2967 | "interpreter profiles type checks only for these BCs"); |
duke@435 | 2968 | data = method()->method_data()->bci_to_data(bci()); |
roland@5908 | 2969 | safe_for_replace = true; |
duke@435 | 2970 | } |
duke@435 | 2971 | |
duke@435 | 2972 | // Make the merge point |
duke@435 | 2973 | enum { _obj_path = 1, _null_path, PATH_LIMIT }; |
kvn@4115 | 2974 | RegionNode* region = new (C) RegionNode(PATH_LIMIT); |
kvn@4115 | 2975 | Node* phi = new (C) PhiNode(region, toop); |
duke@435 | 2976 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
duke@435 | 2977 | |
duke@435 | 2978 | // Use null-cast information if it is available |
jrose@2101 | 2979 | bool never_see_null = ((failure_control == NULL) // regular case only |
jrose@2101 | 2980 | && seems_never_null(obj, data)); |
duke@435 | 2981 | |
duke@435 | 2982 | // Null check; get casted pointer; set region slot 3 |
duke@435 | 2983 | Node* null_ctl = top(); |
roland@5908 | 2984 | Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); |
duke@435 | 2985 | |
duke@435 | 2986 | // If not_null_obj is dead, only null-path is taken |
duke@435 | 2987 | if (stopped()) { // Doing instance-of on a NULL? |
duke@435 | 2988 | set_control(null_ctl); |
duke@435 | 2989 | return null(); |
duke@435 | 2990 | } |
duke@435 | 2991 | region->init_req(_null_path, null_ctl); |
duke@435 | 2992 | phi ->init_req(_null_path, null()); // Set null path value |
jrose@2101 | 2993 | if (null_ctl == top()) { |
jrose@2101 | 2994 | // Do this eagerly, so that pattern matches like is_diamond_phi |
jrose@2101 | 2995 | // will work even during parsing. |
jrose@2101 | 2996 | assert(_null_path == PATH_LIMIT-1, "delete last"); |
jrose@2101 | 2997 | region->del_req(_null_path); |
jrose@2101 | 2998 | phi ->del_req(_null_path); |
jrose@2101 | 2999 | } |
jrose@2101 | 3000 | |
jrose@2101 | 3001 | Node* cast_obj = NULL; |
iveresov@6350 | 3002 | if (tk->klass_is_exact()) { |
iveresov@6350 | 3003 | // The following optimization tries to statically cast the speculative type of the object |
iveresov@6350 | 3004 | // (for example obtained during profiling) to the type of the superklass and then do a |
iveresov@6350 | 3005 | // dynamic check that the type of the object is what we expect. To work correctly |
iveresov@6350 | 3006 | // for checkcast and aastore the type of superklass should be exact. |
iveresov@6350 | 3007 | const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); |
iveresov@6350 | 3008 | // We may not have profiling here or it may not help us. If we have |
iveresov@6350 | 3009 | // a speculative type use it to perform an exact cast. |
iveresov@6350 | 3010 | ciKlass* spec_obj_type = obj_type->speculative_type(); |
iveresov@6350 | 3011 | if (spec_obj_type != NULL || |
iveresov@6350 | 3012 | (data != NULL && |
iveresov@6350 | 3013 | // Counter has never been decremented (due to cast failure). |
iveresov@6350 | 3014 | // ...This is a reasonable thing to expect. It is true of |
iveresov@6350 | 3015 | // all casts inserted by javac to implement generic types. |
iveresov@6350 | 3016 | data->as_CounterData()->count() >= 0)) { |
iveresov@6350 | 3017 | cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); |
iveresov@6350 | 3018 | if (cast_obj != NULL) { |
iveresov@6350 | 3019 | if (failure_control != NULL) // failure is now impossible |
iveresov@6350 | 3020 | (*failure_control) = top(); |
iveresov@6350 | 3021 | // adjust the type of the phi to the exact klass: |
iveresov@6350 | 3022 | phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); |
iveresov@6350 | 3023 | } |
duke@435 | 3024 | } |
duke@435 | 3025 | } |
duke@435 | 3026 | |
duke@435 | 3027 | if (cast_obj == NULL) { |
duke@435 | 3028 | // Load the object's klass |
duke@435 | 3029 | Node* obj_klass = load_object_klass(not_null_obj); |
duke@435 | 3030 | |
duke@435 | 3031 | // Generate the subtype check |
duke@435 | 3032 | Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); |
duke@435 | 3033 | |
duke@435 | 3034 | // Plug in success path into the merge |
kvn@4115 | 3035 | cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), |
duke@435 | 3036 | not_null_obj, toop)); |
duke@435 | 3037 | // Failure path ends in uncommon trap (or may be dead - failure impossible) |
duke@435 | 3038 | if (failure_control == NULL) { |
duke@435 | 3039 | if (not_subtype_ctrl != top()) { // If failure is possible |
duke@435 | 3040 | PreserveJVMState pjvms(this); |
duke@435 | 3041 | set_control(not_subtype_ctrl); |
duke@435 | 3042 | builtin_throw(Deoptimization::Reason_class_check, obj_klass); |
duke@435 | 3043 | } |
duke@435 | 3044 | } else { |
duke@435 | 3045 | (*failure_control) = not_subtype_ctrl; |
duke@435 | 3046 | } |
duke@435 | 3047 | } |
duke@435 | 3048 | |
duke@435 | 3049 | region->init_req(_obj_path, control()); |
duke@435 | 3050 | phi ->init_req(_obj_path, cast_obj); |
duke@435 | 3051 | |
duke@435 | 3052 | // A merge of NULL or Casted-NotNull obj |
duke@435 | 3053 | Node* res = _gvn.transform(phi); |
duke@435 | 3054 | |
duke@435 | 3055 | // Note I do NOT always 'replace_in_map(obj,result)' here. |
duke@435 | 3056 | // if( tk->klass()->can_be_primary_super() ) |
duke@435 | 3057 | // This means that if I successfully store an Object into an array-of-String |
duke@435 | 3058 | // I 'forget' that the Object is really now known to be a String. I have to |
duke@435 | 3059 | // do this because we don't have true union types for interfaces - if I store |
duke@435 | 3060 | // a Baz into an array-of-Interface and then tell the optimizer it's an |
duke@435 | 3061 | // Interface, I forget that it's also a Baz and cannot do Baz-like field |
duke@435 | 3062 | // references to it. FIX THIS WHEN UNION TYPES APPEAR! |
duke@435 | 3063 | // replace_in_map( obj, res ); |
duke@435 | 3064 | |
duke@435 | 3065 | // Return final merged results |
duke@435 | 3066 | set_control( _gvn.transform(region) ); |
duke@435 | 3067 | record_for_igvn(region); |
duke@435 | 3068 | return res; |
duke@435 | 3069 | } |
duke@435 | 3070 | |
duke@435 | 3071 | //------------------------------next_monitor----------------------------------- |
duke@435 | 3072 | // What number should be given to the next monitor? |
duke@435 | 3073 | int GraphKit::next_monitor() { |
duke@435 | 3074 | int current = jvms()->monitor_depth()* C->sync_stack_slots(); |
duke@435 | 3075 | int next = current + C->sync_stack_slots(); |
duke@435 | 3076 | // Keep the toplevel high water mark current: |
duke@435 | 3077 | if (C->fixed_slots() < next) C->set_fixed_slots(next); |
duke@435 | 3078 | return current; |
duke@435 | 3079 | } |
duke@435 | 3080 | |
duke@435 | 3081 | //------------------------------insert_mem_bar--------------------------------- |
duke@435 | 3082 | // Memory barrier to avoid floating things around |
duke@435 | 3083 | // The membar serves as a pinch point between both control and all memory slices. |
duke@435 | 3084 | Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { |
duke@435 | 3085 | MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); |
duke@435 | 3086 | mb->init_req(TypeFunc::Control, control()); |
duke@435 | 3087 | mb->init_req(TypeFunc::Memory, reset_memory()); |
duke@435 | 3088 | Node* membar = _gvn.transform(mb); |
twisti@4313 | 3089 | set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); |
duke@435 | 3090 | set_all_memory_call(membar); |
duke@435 | 3091 | return membar; |
duke@435 | 3092 | } |
duke@435 | 3093 | |
duke@435 | 3094 | //-------------------------insert_mem_bar_volatile---------------------------- |
duke@435 | 3095 | // Memory barrier to avoid floating things around |
duke@435 | 3096 | // The membar serves as a pinch point between both control and memory(alias_idx). |
duke@435 | 3097 | // If you want to make a pinch point on all memory slices, do not use this |
duke@435 | 3098 | // function (even with AliasIdxBot); use insert_mem_bar() instead. |
duke@435 | 3099 | Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { |
duke@435 | 3100 | // When Parse::do_put_xxx updates a volatile field, it appends a series |
duke@435 | 3101 | // of MemBarVolatile nodes, one for *each* volatile field alias category. |
duke@435 | 3102 | // The first membar is on the same memory slice as the field store opcode. |
duke@435 | 3103 | // This forces the membar to follow the store. (Bug 6500685 broke this.) |
duke@435 | 3104 | // All the other membars (for other volatile slices, including AliasIdxBot, |
duke@435 | 3105 | // which stands for all unknown volatile slices) are control-dependent |
duke@435 | 3106 | // on the first membar. This prevents later volatile loads or stores |
duke@435 | 3107 | // from sliding up past the just-emitted store. |
duke@435 | 3108 | |
duke@435 | 3109 | MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); |
duke@435 | 3110 | mb->set_req(TypeFunc::Control,control()); |
duke@435 | 3111 | if (alias_idx == Compile::AliasIdxBot) { |
duke@435 | 3112 | mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); |
duke@435 | 3113 | } else { |
duke@435 | 3114 | assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); |
duke@435 | 3115 | mb->set_req(TypeFunc::Memory, memory(alias_idx)); |
duke@435 | 3116 | } |
duke@435 | 3117 | Node* membar = _gvn.transform(mb); |
kvn@4115 | 3118 | set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); |
duke@435 | 3119 | if (alias_idx == Compile::AliasIdxBot) { |
kvn@4115 | 3120 | merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); |
duke@435 | 3121 | } else { |
kvn@4115 | 3122 | set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); |
duke@435 | 3123 | } |
duke@435 | 3124 | return membar; |
duke@435 | 3125 | } |
duke@435 | 3126 | |
duke@435 | 3127 | //------------------------------shared_lock------------------------------------ |
duke@435 | 3128 | // Emit locking code. |
duke@435 | 3129 | FastLockNode* GraphKit::shared_lock(Node* obj) { |
duke@435 | 3130 | // bci is either a monitorenter bc or InvocationEntryBci |
duke@435 | 3131 | // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces |
duke@435 | 3132 | assert(SynchronizationEntryBCI == InvocationEntryBci, ""); |
duke@435 | 3133 | |
duke@435 | 3134 | if( !GenerateSynchronizationCode ) |
duke@435 | 3135 | return NULL; // Not locking things? |
duke@435 | 3136 | if (stopped()) // Dead monitor? |
duke@435 | 3137 | return NULL; |
duke@435 | 3138 | |
duke@435 | 3139 | assert(dead_locals_are_killed(), "should kill locals before sync. point"); |
duke@435 | 3140 | |
duke@435 | 3141 | // Box the stack location |
kvn@4115 | 3142 | Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); |
duke@435 | 3143 | Node* mem = reset_memory(); |
duke@435 | 3144 | |
kvn@4115 | 3145 | FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); |
kvn@6429 | 3146 | if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { |
duke@435 | 3147 | // Create the counters for this fast lock. |
duke@435 | 3148 | flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci |
duke@435 | 3149 | } |
kvn@6429 | 3150 | |
kvn@6429 | 3151 | // Create the rtm counters for this fast lock if needed. |
kvn@6429 | 3152 | flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci |
kvn@6429 | 3153 | |
duke@435 | 3154 | // Add monitor to debug info for the slow path. If we block inside the |
duke@435 | 3155 | // slow path and de-opt, we need the monitor hanging around |
duke@435 | 3156 | map()->push_monitor( flock ); |
duke@435 | 3157 | |
duke@435 | 3158 | const TypeFunc *tf = LockNode::lock_type(); |
kvn@4115 | 3159 | LockNode *lock = new (C) LockNode(C, tf); |
duke@435 | 3160 | |
duke@435 | 3161 | lock->init_req( TypeFunc::Control, control() ); |
duke@435 | 3162 | lock->init_req( TypeFunc::Memory , mem ); |
duke@435 | 3163 | lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o |
duke@435 | 3164 | lock->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 3165 | lock->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 3166 | |
duke@435 | 3167 | lock->init_req(TypeFunc::Parms + 0, obj); |
duke@435 | 3168 | lock->init_req(TypeFunc::Parms + 1, box); |
duke@435 | 3169 | lock->init_req(TypeFunc::Parms + 2, flock); |
duke@435 | 3170 | add_safepoint_edges(lock); |
duke@435 | 3171 | |
duke@435 | 3172 | lock = _gvn.transform( lock )->as_Lock(); |
duke@435 | 3173 | |
duke@435 | 3174 | // lock has no side-effects, sets few values |
duke@435 | 3175 | set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); |
duke@435 | 3176 | |
roland@3047 | 3177 | insert_mem_bar(Op_MemBarAcquireLock); |
duke@435 | 3178 | |
duke@435 | 3179 | // Add this to the worklist so that the lock can be eliminated |
duke@435 | 3180 | record_for_igvn(lock); |
duke@435 | 3181 | |
duke@435 | 3182 | #ifndef PRODUCT |
duke@435 | 3183 | if (PrintLockStatistics) { |
duke@435 | 3184 | // Update the counter for this lock. Don't bother using an atomic |
duke@435 | 3185 | // operation since we don't require absolute accuracy. |
duke@435 | 3186 | lock->create_lock_counter(map()->jvms()); |
kvn@1964 | 3187 | increment_counter(lock->counter()->addr()); |
duke@435 | 3188 | } |
duke@435 | 3189 | #endif |
duke@435 | 3190 | |
duke@435 | 3191 | return flock; |
duke@435 | 3192 | } |
duke@435 | 3193 | |
duke@435 | 3194 | |
duke@435 | 3195 | //------------------------------shared_unlock---------------------------------- |
duke@435 | 3196 | // Emit unlocking code. |
duke@435 | 3197 | void GraphKit::shared_unlock(Node* box, Node* obj) { |
duke@435 | 3198 | // bci is either a monitorenter bc or InvocationEntryBci |
duke@435 | 3199 | // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces |
duke@435 | 3200 | assert(SynchronizationEntryBCI == InvocationEntryBci, ""); |
duke@435 | 3201 | |
duke@435 | 3202 | if( !GenerateSynchronizationCode ) |
duke@435 | 3203 | return; |
duke@435 | 3204 | if (stopped()) { // Dead monitor? |
duke@435 | 3205 | map()->pop_monitor(); // Kill monitor from debug info |
duke@435 | 3206 | return; |
duke@435 | 3207 | } |
duke@435 | 3208 | |
duke@435 | 3209 | // Memory barrier to avoid floating things down past the locked region |
roland@3047 | 3210 | insert_mem_bar(Op_MemBarReleaseLock); |
duke@435 | 3211 | |
duke@435 | 3212 | const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); |
kvn@4115 | 3213 | UnlockNode *unlock = new (C) UnlockNode(C, tf); |
duke@435 | 3214 | uint raw_idx = Compile::AliasIdxRaw; |
duke@435 | 3215 | unlock->init_req( TypeFunc::Control, control() ); |
duke@435 | 3216 | unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); |
duke@435 | 3217 | unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o |
duke@435 | 3218 | unlock->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 3219 | unlock->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 3220 | |
duke@435 | 3221 | unlock->init_req(TypeFunc::Parms + 0, obj); |
duke@435 | 3222 | unlock->init_req(TypeFunc::Parms + 1, box); |
duke@435 | 3223 | unlock = _gvn.transform(unlock)->as_Unlock(); |
duke@435 | 3224 | |
duke@435 | 3225 | Node* mem = reset_memory(); |
duke@435 | 3226 | |
duke@435 | 3227 | // unlock has no side-effects, sets few values |
duke@435 | 3228 | set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); |
duke@435 | 3229 | |
duke@435 | 3230 | // Kill monitor from debug info |
duke@435 | 3231 | map()->pop_monitor( ); |
duke@435 | 3232 | } |
duke@435 | 3233 | |
duke@435 | 3234 | //-------------------------------get_layout_helper----------------------------- |
duke@435 | 3235 | // If the given klass is a constant or known to be an array, |
duke@435 | 3236 | // fetch the constant layout helper value into constant_value |
duke@435 | 3237 | // and return (Node*)NULL. Otherwise, load the non-constant |
duke@435 | 3238 | // layout helper value, and return the node which represents it. |
duke@435 | 3239 | // This two-faced routine is useful because allocation sites |
duke@435 | 3240 | // almost always feature constant types. |
duke@435 | 3241 | Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { |
duke@435 | 3242 | const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); |
duke@435 | 3243 | if (!StressReflectiveCode && inst_klass != NULL) { |
duke@435 | 3244 | ciKlass* klass = inst_klass->klass(); |
duke@435 | 3245 | bool xklass = inst_klass->klass_is_exact(); |
duke@435 | 3246 | if (xklass || klass->is_array_klass()) { |
duke@435 | 3247 | jint lhelper = klass->layout_helper(); |
duke@435 | 3248 | if (lhelper != Klass::_lh_neutral_value) { |
duke@435 | 3249 | constant_value = lhelper; |
duke@435 | 3250 | return (Node*) NULL; |
duke@435 | 3251 | } |
duke@435 | 3252 | } |
duke@435 | 3253 | } |
duke@435 | 3254 | constant_value = Klass::_lh_neutral_value; // put in a known value |
stefank@3391 | 3255 | Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); |
goetz@6479 | 3256 | return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); |
duke@435 | 3257 | } |
duke@435 | 3258 | |
duke@435 | 3259 | // We just put in an allocate/initialize with a big raw-memory effect. |
duke@435 | 3260 | // Hook selected additional alias categories on the initialization. |
duke@435 | 3261 | static void hook_memory_on_init(GraphKit& kit, int alias_idx, |
duke@435 | 3262 | MergeMemNode* init_in_merge, |
duke@435 | 3263 | Node* init_out_raw) { |
duke@435 | 3264 | DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); |
duke@435 | 3265 | assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); |
duke@435 | 3266 | |
duke@435 | 3267 | Node* prevmem = kit.memory(alias_idx); |
duke@435 | 3268 | init_in_merge->set_memory_at(alias_idx, prevmem); |
duke@435 | 3269 | kit.set_memory(init_out_raw, alias_idx); |
duke@435 | 3270 | } |
duke@435 | 3271 | |
duke@435 | 3272 | //---------------------------set_output_for_allocation------------------------- |
duke@435 | 3273 | Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, |
vlivanov@6733 | 3274 | const TypeOopPtr* oop_type, |
vlivanov@6733 | 3275 | bool deoptimize_on_exception) { |
duke@435 | 3276 | int rawidx = Compile::AliasIdxRaw; |
duke@435 | 3277 | alloc->set_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 3278 | add_safepoint_edges(alloc); |
duke@435 | 3279 | Node* allocx = _gvn.transform(alloc); |
kvn@4115 | 3280 | set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); |
duke@435 | 3281 | // create memory projection for i_o |
kvn@4115 | 3282 | set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); |
vlivanov@6733 | 3283 | make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); |
duke@435 | 3284 | |
duke@435 | 3285 | // create a memory projection as for the normal control path |
kvn@4115 | 3286 | Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); |
duke@435 | 3287 | set_memory(malloc, rawidx); |
duke@435 | 3288 | |
duke@435 | 3289 | // a normal slow-call doesn't change i_o, but an allocation does |
duke@435 | 3290 | // we create a separate i_o projection for the normal control path |
kvn@4115 | 3291 | set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); |
kvn@4115 | 3292 | Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); |
duke@435 | 3293 | |
duke@435 | 3294 | // put in an initialization barrier |
duke@435 | 3295 | InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, |
duke@435 | 3296 | rawoop)->as_Initialize(); |
duke@435 | 3297 | assert(alloc->initialization() == init, "2-way macro link must work"); |
duke@435 | 3298 | assert(init ->allocation() == alloc, "2-way macro link must work"); |
kvn@2810 | 3299 | { |
duke@435 | 3300 | // Extract memory strands which may participate in the new object's |
duke@435 | 3301 | // initialization, and source them from the new InitializeNode. |
duke@435 | 3302 | // This will allow us to observe initializations when they occur, |
duke@435 | 3303 | // and link them properly (as a group) to the InitializeNode. |
duke@435 | 3304 | assert(init->in(InitializeNode::Memory) == malloc, ""); |
duke@435 | 3305 | MergeMemNode* minit_in = MergeMemNode::make(C, malloc); |
duke@435 | 3306 | init->set_req(InitializeNode::Memory, minit_in); |
duke@435 | 3307 | record_for_igvn(minit_in); // fold it up later, if possible |
duke@435 | 3308 | Node* minit_out = memory(rawidx); |
duke@435 | 3309 | assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); |
duke@435 | 3310 | if (oop_type->isa_aryptr()) { |
duke@435 | 3311 | const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); |
duke@435 | 3312 | int elemidx = C->get_alias_index(telemref); |
duke@435 | 3313 | hook_memory_on_init(*this, elemidx, minit_in, minit_out); |
duke@435 | 3314 | } else if (oop_type->isa_instptr()) { |
duke@435 | 3315 | ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); |
duke@435 | 3316 | for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { |
duke@435 | 3317 | ciField* field = ik->nonstatic_field_at(i); |
kvn@458 | 3318 | if (field->offset() >= TrackedInitializationLimit * HeapWordSize) |
duke@435 | 3319 | continue; // do not bother to track really large numbers of fields |
duke@435 | 3320 | // Find (or create) the alias category for this field: |
duke@435 | 3321 | int fieldidx = C->alias_type(field)->index(); |
duke@435 | 3322 | hook_memory_on_init(*this, fieldidx, minit_in, minit_out); |
duke@435 | 3323 | } |
duke@435 | 3324 | } |
duke@435 | 3325 | } |
duke@435 | 3326 | |
duke@435 | 3327 | // Cast raw oop to the real thing... |
kvn@4115 | 3328 | Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); |
duke@435 | 3329 | javaoop = _gvn.transform(javaoop); |
duke@435 | 3330 | C->set_recent_alloc(control(), javaoop); |
duke@435 | 3331 | assert(just_allocated_object(control()) == javaoop, "just allocated"); |
duke@435 | 3332 | |
duke@435 | 3333 | #ifdef ASSERT |
rasbold@801 | 3334 | { // Verify that the AllocateNode::Ideal_allocation recognizers work: |
rasbold@801 | 3335 | assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, |
rasbold@801 | 3336 | "Ideal_allocation works"); |
rasbold@801 | 3337 | assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, |
rasbold@801 | 3338 | "Ideal_allocation works"); |
duke@435 | 3339 | if (alloc->is_AllocateArray()) { |
rasbold@801 | 3340 | assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), |
rasbold@801 | 3341 | "Ideal_allocation works"); |
rasbold@801 | 3342 | assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), |
rasbold@801 | 3343 | "Ideal_allocation works"); |
duke@435 | 3344 | } else { |
rasbold@801 | 3345 | assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); |
duke@435 | 3346 | } |
duke@435 | 3347 | } |
duke@435 | 3348 | #endif //ASSERT |
duke@435 | 3349 | |
duke@435 | 3350 | return javaoop; |
duke@435 | 3351 | } |
duke@435 | 3352 | |
duke@435 | 3353 | //---------------------------new_instance-------------------------------------- |
duke@435 | 3354 | // This routine takes a klass_node which may be constant (for a static type) |
duke@435 | 3355 | // or may be non-constant (for reflective code). It will work equally well |
duke@435 | 3356 | // for either, and the graph will fold nicely if the optimizer later reduces |
duke@435 | 3357 | // the type to a constant. |
duke@435 | 3358 | // The optional arguments are for specialized use by intrinsics: |
duke@435 | 3359 | // - If 'extra_slow_test' if not null is an extra condition for the slow-path. |
duke@435 | 3360 | // - If 'return_size_val', report the the total object size to the caller. |
vlivanov@6733 | 3361 | // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) |
duke@435 | 3362 | Node* GraphKit::new_instance(Node* klass_node, |
duke@435 | 3363 | Node* extra_slow_test, |
vlivanov@6733 | 3364 | Node* *return_size_val, |
vlivanov@6733 | 3365 | bool deoptimize_on_exception) { |
duke@435 | 3366 | // Compute size in doublewords |
duke@435 | 3367 | // The size is always an integral number of doublewords, represented |
duke@435 | 3368 | // as a positive bytewise size stored in the klass's layout_helper. |
duke@435 | 3369 | // The layout_helper also encodes (in a low bit) the need for a slow path. |
duke@435 | 3370 | jint layout_con = Klass::_lh_neutral_value; |
duke@435 | 3371 | Node* layout_val = get_layout_helper(klass_node, layout_con); |
duke@435 | 3372 | int layout_is_con = (layout_val == NULL); |
duke@435 | 3373 | |
duke@435 | 3374 | if (extra_slow_test == NULL) extra_slow_test = intcon(0); |
duke@435 | 3375 | // Generate the initial go-slow test. It's either ALWAYS (return a |
duke@435 | 3376 | // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective |
duke@435 | 3377 | // case) a computed value derived from the layout_helper. |
duke@435 | 3378 | Node* initial_slow_test = NULL; |
duke@435 | 3379 | if (layout_is_con) { |
duke@435 | 3380 | assert(!StressReflectiveCode, "stress mode does not use these paths"); |
duke@435 | 3381 | bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); |
duke@435 | 3382 | initial_slow_test = must_go_slow? intcon(1): extra_slow_test; |
duke@435 | 3383 | |
duke@435 | 3384 | } else { // reflective case |
duke@435 | 3385 | // This reflective path is used by Unsafe.allocateInstance. |
duke@435 | 3386 | // (It may be stress-tested by specifying StressReflectiveCode.) |
duke@435 | 3387 | // Basically, we want to get into the VM is there's an illegal argument. |
duke@435 | 3388 | Node* bit = intcon(Klass::_lh_instance_slow_path_bit); |
kvn@4115 | 3389 | initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); |
duke@435 | 3390 | if (extra_slow_test != intcon(0)) { |
kvn@4115 | 3391 | initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); |
duke@435 | 3392 | } |
duke@435 | 3393 | // (Macro-expander will further convert this to a Bool, if necessary.) |
duke@435 | 3394 | } |
duke@435 | 3395 | |
duke@435 | 3396 | // Find the size in bytes. This is easy; it's the layout_helper. |
duke@435 | 3397 | // The size value must be valid even if the slow path is taken. |
duke@435 | 3398 | Node* size = NULL; |
duke@435 | 3399 | if (layout_is_con) { |
duke@435 | 3400 | size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); |
duke@435 | 3401 | } else { // reflective case |
duke@435 | 3402 | // This reflective path is used by clone and Unsafe.allocateInstance. |
duke@435 | 3403 | size = ConvI2X(layout_val); |
duke@435 | 3404 | |
duke@435 | 3405 | // Clear the low bits to extract layout_helper_size_in_bytes: |
duke@435 | 3406 | assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); |
duke@435 | 3407 | Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); |
kvn@4115 | 3408 | size = _gvn.transform( new (C) AndXNode(size, mask) ); |
duke@435 | 3409 | } |
duke@435 | 3410 | if (return_size_val != NULL) { |
duke@435 | 3411 | (*return_size_val) = size; |
duke@435 | 3412 | } |
duke@435 | 3413 | |
duke@435 | 3414 | // This is a precise notnull oop of the klass. |
duke@435 | 3415 | // (Actually, it need not be precise if this is a reflective allocation.) |
duke@435 | 3416 | // It's what we cast the result to. |
duke@435 | 3417 | const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); |
duke@435 | 3418 | if (!tklass) tklass = TypeKlassPtr::OBJECT; |
duke@435 | 3419 | const TypeOopPtr* oop_type = tklass->as_instance_type(); |
duke@435 | 3420 | |
duke@435 | 3421 | // Now generate allocation code |
kvn@509 | 3422 | |
kvn@1000 | 3423 | // The entire memory state is needed for slow path of the allocation |
kvn@1000 | 3424 | // since GC and deoptimization can happened. |
kvn@1000 | 3425 | Node *mem = reset_memory(); |
kvn@1000 | 3426 | set_all_memory(mem); // Create new memory state |
kvn@509 | 3427 | |
duke@435 | 3428 | AllocateNode* alloc |
kvn@5110 | 3429 | = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP), |
kvn@4115 | 3430 | control(), mem, i_o(), |
kvn@4115 | 3431 | size, klass_node, |
kvn@4115 | 3432 | initial_slow_test); |
duke@435 | 3433 | |
vlivanov@6733 | 3434 | return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); |
duke@435 | 3435 | } |
duke@435 | 3436 | |
duke@435 | 3437 | //-------------------------------new_array------------------------------------- |
duke@435 | 3438 | // helper for both newarray and anewarray |
duke@435 | 3439 | // The 'length' parameter is (obviously) the length of the array. |
duke@435 | 3440 | // See comments on new_instance for the meaning of the other arguments. |
duke@435 | 3441 | Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) |
duke@435 | 3442 | Node* length, // number of array elements |
cfang@1165 | 3443 | int nargs, // number of arguments to push back for uncommon trap |
vlivanov@6733 | 3444 | Node* *return_size_val, |
vlivanov@6733 | 3445 | bool deoptimize_on_exception) { |
duke@435 | 3446 | jint layout_con = Klass::_lh_neutral_value; |
duke@435 | 3447 | Node* layout_val = get_layout_helper(klass_node, layout_con); |
duke@435 | 3448 | int layout_is_con = (layout_val == NULL); |
duke@435 | 3449 | |
duke@435 | 3450 | if (!layout_is_con && !StressReflectiveCode && |
duke@435 | 3451 | !too_many_traps(Deoptimization::Reason_class_check)) { |
duke@435 | 3452 | // This is a reflective array creation site. |
duke@435 | 3453 | // Optimistically assume that it is a subtype of Object[], |
duke@435 | 3454 | // so that we can fold up all the address arithmetic. |
duke@435 | 3455 | layout_con = Klass::array_layout_helper(T_OBJECT); |
kvn@4115 | 3456 | Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); |
kvn@4115 | 3457 | Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); |
duke@435 | 3458 | { BuildCutout unless(this, bol_lh, PROB_MAX); |
twisti@4313 | 3459 | inc_sp(nargs); |
duke@435 | 3460 | uncommon_trap(Deoptimization::Reason_class_check, |
duke@435 | 3461 | Deoptimization::Action_maybe_recompile); |
duke@435 | 3462 | } |
duke@435 | 3463 | layout_val = NULL; |
duke@435 | 3464 | layout_is_con = true; |
duke@435 | 3465 | } |
duke@435 | 3466 | |
duke@435 | 3467 | // Generate the initial go-slow test. Make sure we do not overflow |
duke@435 | 3468 | // if length is huge (near 2Gig) or negative! We do not need |
duke@435 | 3469 | // exact double-words here, just a close approximation of needed |
duke@435 | 3470 | // double-words. We can't add any offset or rounding bits, lest we |
duke@435 | 3471 | // take a size -1 of bytes and make it positive. Use an unsigned |
duke@435 | 3472 | // compare, so negative sizes look hugely positive. |
duke@435 | 3473 | int fast_size_limit = FastAllocateSizeLimit; |
duke@435 | 3474 | if (layout_is_con) { |
duke@435 | 3475 | assert(!StressReflectiveCode, "stress mode does not use these paths"); |
duke@435 | 3476 | // Increase the size limit if we have exact knowledge of array type. |
duke@435 | 3477 | int log2_esize = Klass::layout_helper_log2_element_size(layout_con); |
duke@435 | 3478 | fast_size_limit <<= (LogBytesPerLong - log2_esize); |
duke@435 | 3479 | } |
duke@435 | 3480 | |
kvn@4115 | 3481 | Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); |
kvn@4115 | 3482 | Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); |
duke@435 | 3483 | if (initial_slow_test->is_Bool()) { |
duke@435 | 3484 | // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. |
duke@435 | 3485 | initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); |
duke@435 | 3486 | } |
duke@435 | 3487 | |
duke@435 | 3488 | // --- Size Computation --- |
duke@435 | 3489 | // array_size = round_to_heap(array_header + (length << elem_shift)); |
duke@435 | 3490 | // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) |
duke@435 | 3491 | // and round_to(x, y) == ((x + y-1) & ~(y-1)) |
duke@435 | 3492 | // The rounding mask is strength-reduced, if possible. |
duke@435 | 3493 | int round_mask = MinObjAlignmentInBytes - 1; |
duke@435 | 3494 | Node* header_size = NULL; |
duke@435 | 3495 | int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
duke@435 | 3496 | // (T_BYTE has the weakest alignment and size restrictions...) |
duke@435 | 3497 | if (layout_is_con) { |
duke@435 | 3498 | int hsize = Klass::layout_helper_header_size(layout_con); |
duke@435 | 3499 | int eshift = Klass::layout_helper_log2_element_size(layout_con); |
duke@435 | 3500 | BasicType etype = Klass::layout_helper_element_type(layout_con); |
duke@435 | 3501 | if ((round_mask & ~right_n_bits(eshift)) == 0) |
duke@435 | 3502 | round_mask = 0; // strength-reduce it if it goes away completely |
duke@435 | 3503 | assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); |
duke@435 | 3504 | assert(header_size_min <= hsize, "generic minimum is smallest"); |
duke@435 | 3505 | header_size_min = hsize; |
duke@435 | 3506 | header_size = intcon(hsize + round_mask); |
duke@435 | 3507 | } else { |
duke@435 | 3508 | Node* hss = intcon(Klass::_lh_header_size_shift); |
duke@435 | 3509 | Node* hsm = intcon(Klass::_lh_header_size_mask); |
kvn@4115 | 3510 | Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); |
kvn@4115 | 3511 | hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); |
duke@435 | 3512 | Node* mask = intcon(round_mask); |
kvn@4115 | 3513 | header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); |
duke@435 | 3514 | } |
duke@435 | 3515 | |
duke@435 | 3516 | Node* elem_shift = NULL; |
duke@435 | 3517 | if (layout_is_con) { |
duke@435 | 3518 | int eshift = Klass::layout_helper_log2_element_size(layout_con); |
duke@435 | 3519 | if (eshift != 0) |
duke@435 | 3520 | elem_shift = intcon(eshift); |
duke@435 | 3521 | } else { |
duke@435 | 3522 | // There is no need to mask or shift this value. |
duke@435 | 3523 | // The semantics of LShiftINode include an implicit mask to 0x1F. |
duke@435 | 3524 | assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); |
duke@435 | 3525 | elem_shift = layout_val; |
duke@435 | 3526 | } |
duke@435 | 3527 | |
duke@435 | 3528 | // Transition to native address size for all offset calculations: |
duke@435 | 3529 | Node* lengthx = ConvI2X(length); |
duke@435 | 3530 | Node* headerx = ConvI2X(header_size); |
duke@435 | 3531 | #ifdef _LP64 |
duke@435 | 3532 | { const TypeLong* tllen = _gvn.find_long_type(lengthx); |
duke@435 | 3533 | if (tllen != NULL && tllen->_lo < 0) { |
duke@435 | 3534 | // Add a manual constraint to a positive range. Cf. array_element_address. |
duke@435 | 3535 | jlong size_max = arrayOopDesc::max_array_length(T_BYTE); |
duke@435 | 3536 | if (size_max > tllen->_hi) size_max = tllen->_hi; |
duke@435 | 3537 | const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); |
kvn@4115 | 3538 | lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon)); |
duke@435 | 3539 | } |
duke@435 | 3540 | } |
duke@435 | 3541 | #endif |
duke@435 | 3542 | |
duke@435 | 3543 | // Combine header size (plus rounding) and body size. Then round down. |
duke@435 | 3544 | // This computation cannot overflow, because it is used only in two |
duke@435 | 3545 | // places, one where the length is sharply limited, and the other |
duke@435 | 3546 | // after a successful allocation. |
duke@435 | 3547 | Node* abody = lengthx; |
duke@435 | 3548 | if (elem_shift != NULL) |
kvn@4115 | 3549 | abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); |
kvn@4115 | 3550 | Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); |
duke@435 | 3551 | if (round_mask != 0) { |
duke@435 | 3552 | Node* mask = MakeConX(~round_mask); |
kvn@4115 | 3553 | size = _gvn.transform( new(C) AndXNode(size, mask) ); |
duke@435 | 3554 | } |
duke@435 | 3555 | // else if round_mask == 0, the size computation is self-rounding |
duke@435 | 3556 | |
duke@435 | 3557 | if (return_size_val != NULL) { |
duke@435 | 3558 | // This is the size |
duke@435 | 3559 | (*return_size_val) = size; |
duke@435 | 3560 | } |
duke@435 | 3561 | |
duke@435 | 3562 | // Now generate allocation code |
kvn@509 | 3563 | |
kvn@1000 | 3564 | // The entire memory state is needed for slow path of the allocation |
kvn@1000 | 3565 | // since GC and deoptimization can happened. |
kvn@1000 | 3566 | Node *mem = reset_memory(); |
kvn@1000 | 3567 | set_all_memory(mem); // Create new memory state |
kvn@509 | 3568 | |
duke@435 | 3569 | // Create the AllocateArrayNode and its result projections |
duke@435 | 3570 | AllocateArrayNode* alloc |
kvn@5110 | 3571 | = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), |
kvn@4115 | 3572 | control(), mem, i_o(), |
kvn@4115 | 3573 | size, klass_node, |
kvn@4115 | 3574 | initial_slow_test, |
kvn@4115 | 3575 | length); |
duke@435 | 3576 | |
duke@435 | 3577 | // Cast to correct type. Note that the klass_node may be constant or not, |
duke@435 | 3578 | // and in the latter case the actual array type will be inexact also. |
duke@435 | 3579 | // (This happens via a non-constant argument to inline_native_newArray.) |
duke@435 | 3580 | // In any case, the value of klass_node provides the desired array type. |
duke@435 | 3581 | const TypeInt* length_type = _gvn.find_int_type(length); |
duke@435 | 3582 | const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); |
duke@435 | 3583 | if (ary_type->isa_aryptr() && length_type != NULL) { |
duke@435 | 3584 | // Try to get a better type than POS for the size |
duke@435 | 3585 | ary_type = ary_type->is_aryptr()->cast_to_size(length_type); |
duke@435 | 3586 | } |
duke@435 | 3587 | |
vlivanov@6733 | 3588 | Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); |
duke@435 | 3589 | |
rasbold@801 | 3590 | // Cast length on remaining path to be as narrow as possible |
rasbold@801 | 3591 | if (map()->find_edge(length) >= 0) { |
rasbold@801 | 3592 | Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); |
rasbold@801 | 3593 | if (ccast != length) { |
rasbold@801 | 3594 | _gvn.set_type_bottom(ccast); |
rasbold@801 | 3595 | record_for_igvn(ccast); |
duke@435 | 3596 | replace_in_map(length, ccast); |
duke@435 | 3597 | } |
duke@435 | 3598 | } |
duke@435 | 3599 | |
duke@435 | 3600 | return javaoop; |
duke@435 | 3601 | } |
duke@435 | 3602 | |
duke@435 | 3603 | // The following "Ideal_foo" functions are placed here because they recognize |
duke@435 | 3604 | // the graph shapes created by the functions immediately above. |
duke@435 | 3605 | |
duke@435 | 3606 | //---------------------------Ideal_allocation---------------------------------- |
duke@435 | 3607 | // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. |
duke@435 | 3608 | AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { |
duke@435 | 3609 | if (ptr == NULL) { // reduce dumb test in callers |
duke@435 | 3610 | return NULL; |
duke@435 | 3611 | } |
kvn@5351 | 3612 | if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast |
kvn@5351 | 3613 | ptr = ptr->in(1); |
kvn@5351 | 3614 | if (ptr == NULL) return NULL; |
kvn@5351 | 3615 | } |
kvn@5351 | 3616 | // Return NULL for allocations with several casts: |
kvn@5351 | 3617 | // j.l.reflect.Array.newInstance(jobject, jint) |
kvn@5351 | 3618 | // Object.clone() |
kvn@5351 | 3619 | // to keep more precise type from last cast. |
duke@435 | 3620 | if (ptr->is_Proj()) { |
duke@435 | 3621 | Node* allo = ptr->in(0); |
duke@435 | 3622 | if (allo != NULL && allo->is_Allocate()) { |
duke@435 | 3623 | return allo->as_Allocate(); |
duke@435 | 3624 | } |
duke@435 | 3625 | } |
duke@435 | 3626 | // Report failure to match. |
duke@435 | 3627 | return NULL; |
duke@435 | 3628 | } |
duke@435 | 3629 | |
duke@435 | 3630 | // Fancy version which also strips off an offset (and reports it to caller). |
duke@435 | 3631 | AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, |
duke@435 | 3632 | intptr_t& offset) { |
duke@435 | 3633 | Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); |
duke@435 | 3634 | if (base == NULL) return NULL; |
duke@435 | 3635 | return Ideal_allocation(base, phase); |
duke@435 | 3636 | } |
duke@435 | 3637 | |
duke@435 | 3638 | // Trace Initialize <- Proj[Parm] <- Allocate |
duke@435 | 3639 | AllocateNode* InitializeNode::allocation() { |
duke@435 | 3640 | Node* rawoop = in(InitializeNode::RawAddress); |
duke@435 | 3641 | if (rawoop->is_Proj()) { |
duke@435 | 3642 | Node* alloc = rawoop->in(0); |
duke@435 | 3643 | if (alloc->is_Allocate()) { |
duke@435 | 3644 | return alloc->as_Allocate(); |
duke@435 | 3645 | } |
duke@435 | 3646 | } |
duke@435 | 3647 | return NULL; |
duke@435 | 3648 | } |
duke@435 | 3649 | |
duke@435 | 3650 | // Trace Allocate -> Proj[Parm] -> Initialize |
duke@435 | 3651 | InitializeNode* AllocateNode::initialization() { |
duke@435 | 3652 | ProjNode* rawoop = proj_out(AllocateNode::RawAddress); |
duke@435 | 3653 | if (rawoop == NULL) return NULL; |
duke@435 | 3654 | for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { |
duke@435 | 3655 | Node* init = rawoop->fast_out(i); |
duke@435 | 3656 | if (init->is_Initialize()) { |
duke@435 | 3657 | assert(init->as_Initialize()->allocation() == this, "2-way link"); |
duke@435 | 3658 | return init->as_Initialize(); |
duke@435 | 3659 | } |
duke@435 | 3660 | } |
duke@435 | 3661 | return NULL; |
duke@435 | 3662 | } |
ysr@777 | 3663 | |
kvn@2665 | 3664 | //----------------------------- loop predicates --------------------------- |
kvn@2665 | 3665 | |
kvn@2665 | 3666 | //------------------------------add_predicate_impl---------------------------- |
kvn@2665 | 3667 | void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { |
kvn@2665 | 3668 | // Too many traps seen? |
kvn@2665 | 3669 | if (too_many_traps(reason)) { |
kvn@2665 | 3670 | #ifdef ASSERT |
kvn@2665 | 3671 | if (TraceLoopPredicate) { |
kvn@2665 | 3672 | int tc = C->trap_count(reason); |
kvn@2665 | 3673 | tty->print("too many traps=%s tcount=%d in ", |
kvn@2665 | 3674 | Deoptimization::trap_reason_name(reason), tc); |
kvn@2665 | 3675 | method()->print(); // which method has too many predicate traps |
kvn@2665 | 3676 | tty->cr(); |
kvn@2665 | 3677 | } |
kvn@2665 | 3678 | #endif |
kvn@2665 | 3679 | // We cannot afford to take more traps here, |
kvn@2665 | 3680 | // do not generate predicate. |
kvn@2665 | 3681 | return; |
kvn@2665 | 3682 | } |
kvn@2665 | 3683 | |
kvn@2665 | 3684 | Node *cont = _gvn.intcon(1); |
kvn@4115 | 3685 | Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); |
kvn@4115 | 3686 | Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); |
kvn@2665 | 3687 | IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); |
kvn@4115 | 3688 | Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); |
kvn@2665 | 3689 | C->add_predicate_opaq(opq); |
kvn@2665 | 3690 | { |
kvn@2665 | 3691 | PreserveJVMState pjvms(this); |
kvn@2665 | 3692 | set_control(iffalse); |
twisti@4313 | 3693 | inc_sp(nargs); |
kvn@2665 | 3694 | uncommon_trap(reason, Deoptimization::Action_maybe_recompile); |
kvn@2665 | 3695 | } |
kvn@4115 | 3696 | Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); |
kvn@2665 | 3697 | set_control(iftrue); |
kvn@2665 | 3698 | } |
kvn@2665 | 3699 | |
kvn@2665 | 3700 | //------------------------------add_predicate--------------------------------- |
kvn@2665 | 3701 | void GraphKit::add_predicate(int nargs) { |
kvn@2665 | 3702 | if (UseLoopPredicate) { |
kvn@2665 | 3703 | add_predicate_impl(Deoptimization::Reason_predicate, nargs); |
kvn@2665 | 3704 | } |
kvn@2877 | 3705 | // loop's limit check predicate should be near the loop. |
kvn@2877 | 3706 | if (LoopLimitCheck) { |
kvn@2877 | 3707 | add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); |
kvn@2877 | 3708 | } |
kvn@2665 | 3709 | } |
kvn@2665 | 3710 | |
kvn@1286 | 3711 | //----------------------------- store barriers ---------------------------- |
kvn@1286 | 3712 | #define __ ideal. |
kvn@1286 | 3713 | |
kvn@1286 | 3714 | void GraphKit::sync_kit(IdealKit& ideal) { |
kvn@2726 | 3715 | set_all_memory(__ merged_memory()); |
kvn@2726 | 3716 | set_i_o(__ i_o()); |
kvn@2726 | 3717 | set_control(__ ctrl()); |
kvn@2726 | 3718 | } |
kvn@2726 | 3719 | |
kvn@2726 | 3720 | void GraphKit::final_sync(IdealKit& ideal) { |
kvn@1286 | 3721 | // Final sync IdealKit and graphKit. |
kvn@2726 | 3722 | sync_kit(ideal); |
kvn@1286 | 3723 | } |
kvn@1286 | 3724 | |
kvn@1286 | 3725 | // vanilla/CMS post barrier |
kvn@1286 | 3726 | // Insert a write-barrier store. This is to let generational GC work; we have |
kvn@1286 | 3727 | // to flag all oop-stores before the next GC point. |
kvn@1286 | 3728 | void GraphKit::write_barrier_post(Node* oop_store, |
kvn@1286 | 3729 | Node* obj, |
kvn@1286 | 3730 | Node* adr, |
cfang@1420 | 3731 | uint adr_idx, |
kvn@1286 | 3732 | Node* val, |
kvn@1286 | 3733 | bool use_precise) { |
kvn@1286 | 3734 | // No store check needed if we're storing a NULL or an old object |
kvn@1286 | 3735 | // (latter case is probably a string constant). The concurrent |
kvn@1286 | 3736 | // mark sweep garbage collector, however, needs to have all nonNull |
kvn@1286 | 3737 | // oop updates flagged via card-marks. |
kvn@1286 | 3738 | if (val != NULL && val->is_Con()) { |
kvn@1286 | 3739 | // must be either an oop or NULL |
kvn@1286 | 3740 | const Type* t = val->bottom_type(); |
kvn@1286 | 3741 | if (t == TypePtr::NULL_PTR || t == Type::TOP) |
kvn@1286 | 3742 | // stores of null never (?) need barriers |
kvn@1286 | 3743 | return; |
kvn@1286 | 3744 | } |
kvn@1286 | 3745 | |
ysr@1462 | 3746 | if (use_ReduceInitialCardMarks() |
ysr@1462 | 3747 | && obj == just_allocated_object(control())) { |
ysr@1462 | 3748 | // We can skip marks on a freshly-allocated object in Eden. |
ysr@1601 | 3749 | // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. |
ysr@1601 | 3750 | // That routine informs GC to take appropriate compensating steps, |
ysr@1601 | 3751 | // upon a slow-path allocation, so as to make this card-mark |
ysr@1601 | 3752 | // elision safe. |
ysr@1462 | 3753 | return; |
ysr@1462 | 3754 | } |
ysr@1462 | 3755 | |
kvn@1286 | 3756 | if (!use_precise) { |
kvn@1286 | 3757 | // All card marks for a (non-array) instance are in one place: |
kvn@1286 | 3758 | adr = obj; |
kvn@1286 | 3759 | } |
kvn@1286 | 3760 | // (Else it's an array (or unknown), and we want more precise card marks.) |
kvn@1286 | 3761 | assert(adr != NULL, ""); |
kvn@1286 | 3762 | |
kvn@2726 | 3763 | IdealKit ideal(this, true); |
kvn@1286 | 3764 | |
kvn@1286 | 3765 | // Convert the pointer to an int prior to doing math on it |
kvn@1286 | 3766 | Node* cast = __ CastPX(__ ctrl(), adr); |
kvn@1286 | 3767 | |
kvn@1286 | 3768 | // Divide by card size |
kvn@1286 | 3769 | assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, |
kvn@1286 | 3770 | "Only one we handle so far."); |
kvn@1286 | 3771 | Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); |
kvn@1286 | 3772 | |
kvn@1286 | 3773 | // Combine card table base and card offset |
kvn@1286 | 3774 | Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); |
kvn@1286 | 3775 | |
kvn@1286 | 3776 | // Get the alias_index for raw card-mark memory |
kvn@1286 | 3777 | int adr_type = Compile::AliasIdxRaw; |
never@2814 | 3778 | Node* zero = __ ConI(0); // Dirty card value |
never@2814 | 3779 | BasicType bt = T_BYTE; |
never@2814 | 3780 | |
never@2814 | 3781 | if (UseCondCardMark) { |
never@2814 | 3782 | // The classic GC reference write barrier is typically implemented |
never@2814 | 3783 | // as a store into the global card mark table. Unfortunately |
never@2814 | 3784 | // unconditional stores can result in false sharing and excessive |
never@2814 | 3785 | // coherence traffic as well as false transactional aborts. |
never@2814 | 3786 | // UseCondCardMark enables MP "polite" conditional card mark |
never@2814 | 3787 | // stores. In theory we could relax the load from ctrl() to |
never@2814 | 3788 | // no_ctrl, but that doesn't buy much latitude. |
never@2814 | 3789 | Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); |
never@2814 | 3790 | __ if_then(card_val, BoolTest::ne, zero); |
never@2814 | 3791 | } |
never@2814 | 3792 | |
kvn@1286 | 3793 | // Smash zero into card |
kvn@1286 | 3794 | if( !UseConcMarkSweepGC ) { |
goetz@6479 | 3795 | __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release); |
kvn@1286 | 3796 | } else { |
kvn@1286 | 3797 | // Specialized path for CM store barrier |
cfang@1420 | 3798 | __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); |
kvn@1286 | 3799 | } |
kvn@1286 | 3800 | |
never@2814 | 3801 | if (UseCondCardMark) { |
never@2814 | 3802 | __ end_if(); |
never@2814 | 3803 | } |
never@2814 | 3804 | |
kvn@1286 | 3805 | // Final sync IdealKit and GraphKit. |
kvn@2726 | 3806 | final_sync(ideal); |
kvn@1286 | 3807 | } |
kvn@1286 | 3808 | |
kvn@1286 | 3809 | // G1 pre/post barriers |
johnc@2781 | 3810 | void GraphKit::g1_write_barrier_pre(bool do_load, |
johnc@2781 | 3811 | Node* obj, |
ysr@777 | 3812 | Node* adr, |
ysr@777 | 3813 | uint alias_idx, |
ysr@777 | 3814 | Node* val, |
never@1260 | 3815 | const TypeOopPtr* val_type, |
johnc@2781 | 3816 | Node* pre_val, |
ysr@777 | 3817 | BasicType bt) { |
johnc@2781 | 3818 | |
johnc@2781 | 3819 | // Some sanity checks |
johnc@2781 | 3820 | // Note: val is unused in this routine. |
johnc@2781 | 3821 | |
johnc@2781 | 3822 | if (do_load) { |
johnc@2781 | 3823 | // We need to generate the load of the previous value |
johnc@2781 | 3824 | assert(obj != NULL, "must have a base"); |
johnc@2781 | 3825 | assert(adr != NULL, "where are loading from?"); |
johnc@2781 | 3826 | assert(pre_val == NULL, "loaded already?"); |
johnc@2781 | 3827 | assert(val_type != NULL, "need a type"); |
johnc@2781 | 3828 | } else { |
johnc@2781 | 3829 | // In this case both val_type and alias_idx are unused. |
johnc@2781 | 3830 | assert(pre_val != NULL, "must be loaded already"); |
kvn@5637 | 3831 | // Nothing to be done if pre_val is null. |
kvn@5637 | 3832 | if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; |
johnc@2781 | 3833 | assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); |
johnc@2781 | 3834 | } |
johnc@2781 | 3835 | assert(bt == T_OBJECT, "or we shouldn't be here"); |
johnc@2781 | 3836 | |
kvn@2726 | 3837 | IdealKit ideal(this, true); |
kvn@1286 | 3838 | |
kvn@1286 | 3839 | Node* tls = __ thread(); // ThreadLocalStorage |
ysr@777 | 3840 | |
ysr@777 | 3841 | Node* no_ctrl = NULL; |
ysr@777 | 3842 | Node* no_base = __ top(); |
johnc@5017 | 3843 | Node* zero = __ ConI(0); |
johnc@5017 | 3844 | Node* zeroX = __ ConX(0); |
ysr@777 | 3845 | |
ysr@777 | 3846 | float likely = PROB_LIKELY(0.999); |
ysr@777 | 3847 | float unlikely = PROB_UNLIKELY(0.999); |
ysr@777 | 3848 | |
ysr@777 | 3849 | BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; |
ysr@777 | 3850 | assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); |
ysr@777 | 3851 | |
ysr@777 | 3852 | // Offsets into the thread |
ysr@777 | 3853 | const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 |
ysr@777 | 3854 | PtrQueue::byte_offset_of_active()); |
ysr@777 | 3855 | const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 |
ysr@777 | 3856 | PtrQueue::byte_offset_of_index()); |
ysr@777 | 3857 | const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 |
ysr@777 | 3858 | PtrQueue::byte_offset_of_buf()); |
johnc@2781 | 3859 | |
ysr@777 | 3860 | // Now the actual pointers into the thread |
kvn@1286 | 3861 | Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); |
kvn@1286 | 3862 | Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); |
kvn@1286 | 3863 | Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); |
ysr@777 | 3864 | |
ysr@777 | 3865 | // Now some of the values |
never@979 | 3866 | Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); |
ysr@777 | 3867 | |
ysr@777 | 3868 | // if (!marking) |
kvn@5925 | 3869 | __ if_then(marking, BoolTest::ne, zero, unlikely); { |
johnc@5017 | 3870 | BasicType index_bt = TypeX_X->basic_type(); |
johnc@5017 | 3871 | assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); |
johnc@5017 | 3872 | Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); |
ysr@777 | 3873 | |
johnc@2781 | 3874 | if (do_load) { |
ysr@777 | 3875 | // load original value |
ysr@777 | 3876 | // alias_idx correct?? |
kvn@5627 | 3877 | pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); |
johnc@2781 | 3878 | } |
johnc@2781 | 3879 | |
johnc@2781 | 3880 | // if (pre_val != NULL) |
johnc@2781 | 3881 | __ if_then(pre_val, BoolTest::ne, null()); { |
johnc@2781 | 3882 | Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); |
ysr@777 | 3883 | |
ysr@777 | 3884 | // is the queue for this thread full? |
johnc@5017 | 3885 | __ if_then(index, BoolTest::ne, zeroX, likely); { |
ysr@777 | 3886 | |
ysr@777 | 3887 | // decrement the index |
johnc@5017 | 3888 | Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); |
ysr@777 | 3889 | |
johnc@2781 | 3890 | // Now get the buffer location we will log the previous value into and store it |
johnc@5017 | 3891 | Node *log_addr = __ AddP(no_base, buffer, next_index); |
goetz@6479 | 3892 | __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); |
ysr@777 | 3893 | // update the index |
goetz@6479 | 3894 | __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); |
ysr@777 | 3895 | |
ysr@777 | 3896 | } __ else_(); { |
ysr@777 | 3897 | |
ysr@777 | 3898 | // logging buffer is full, call the runtime |
ysr@777 | 3899 | const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); |
johnc@2781 | 3900 | __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); |
kvn@1286 | 3901 | } __ end_if(); // (!index) |
johnc@2781 | 3902 | } __ end_if(); // (pre_val != NULL) |
kvn@1286 | 3903 | } __ end_if(); // (!marking) |
kvn@1286 | 3904 | |
kvn@1286 | 3905 | // Final sync IdealKit and GraphKit. |
kvn@2726 | 3906 | final_sync(ideal); |
ysr@777 | 3907 | } |
ysr@777 | 3908 | |
ysr@777 | 3909 | // |
ysr@777 | 3910 | // Update the card table and add card address to the queue |
ysr@777 | 3911 | // |
kvn@1286 | 3912 | void GraphKit::g1_mark_card(IdealKit& ideal, |
kvn@1286 | 3913 | Node* card_adr, |
kvn@1286 | 3914 | Node* oop_store, |
cfang@1420 | 3915 | uint oop_alias_idx, |
kvn@1286 | 3916 | Node* index, |
kvn@1286 | 3917 | Node* index_adr, |
kvn@1286 | 3918 | Node* buffer, |
kvn@1286 | 3919 | const TypeFunc* tf) { |
kvn@1286 | 3920 | |
johnc@5017 | 3921 | Node* zero = __ ConI(0); |
johnc@5017 | 3922 | Node* zeroX = __ ConX(0); |
ysr@777 | 3923 | Node* no_base = __ top(); |
ysr@777 | 3924 | BasicType card_bt = T_BYTE; |
ysr@777 | 3925 | // Smash zero into card. MUST BE ORDERED WRT TO STORE |
cfang@1420 | 3926 | __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); |
ysr@777 | 3927 | |
ysr@777 | 3928 | // Now do the queue work |
johnc@5017 | 3929 | __ if_then(index, BoolTest::ne, zeroX); { |
johnc@5017 | 3930 | |
johnc@5017 | 3931 | Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); |
johnc@5017 | 3932 | Node* log_addr = __ AddP(no_base, buffer, next_index); |
ysr@777 | 3933 | |
goetz@6479 | 3934 | // Order, see storeCM. |
goetz@6479 | 3935 | __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); |
goetz@6479 | 3936 | __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); |
ysr@777 | 3937 | |
ysr@777 | 3938 | } __ else_(); { |
ysr@777 | 3939 | __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); |
ysr@777 | 3940 | } __ end_if(); |
kvn@1286 | 3941 | |
ysr@777 | 3942 | } |
ysr@777 | 3943 | |
kvn@1286 | 3944 | void GraphKit::g1_write_barrier_post(Node* oop_store, |
ysr@777 | 3945 | Node* obj, |
ysr@777 | 3946 | Node* adr, |
ysr@777 | 3947 | uint alias_idx, |
ysr@777 | 3948 | Node* val, |
ysr@777 | 3949 | BasicType bt, |
ysr@777 | 3950 | bool use_precise) { |
ysr@777 | 3951 | // If we are writing a NULL then we need no post barrier |
ysr@777 | 3952 | |
ysr@777 | 3953 | if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { |
ysr@777 | 3954 | // Must be NULL |
ysr@777 | 3955 | const Type* t = val->bottom_type(); |
ysr@777 | 3956 | assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); |
ysr@777 | 3957 | // No post barrier if writing NULLx |
ysr@777 | 3958 | return; |
ysr@777 | 3959 | } |
ysr@777 | 3960 | |
ysr@777 | 3961 | if (!use_precise) { |
ysr@777 | 3962 | // All card marks for a (non-array) instance are in one place: |
ysr@777 | 3963 | adr = obj; |
ysr@777 | 3964 | } |
ysr@777 | 3965 | // (Else it's an array (or unknown), and we want more precise card marks.) |
ysr@777 | 3966 | assert(adr != NULL, ""); |
ysr@777 | 3967 | |
kvn@2726 | 3968 | IdealKit ideal(this, true); |
kvn@1286 | 3969 | |
kvn@1286 | 3970 | Node* tls = __ thread(); // ThreadLocalStorage |
ysr@777 | 3971 | |
ysr@777 | 3972 | Node* no_base = __ top(); |
ysr@777 | 3973 | float likely = PROB_LIKELY(0.999); |
ysr@777 | 3974 | float unlikely = PROB_UNLIKELY(0.999); |
mgerdin@5860 | 3975 | Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); |
mgerdin@5860 | 3976 | Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); |
ysr@777 | 3977 | Node* zeroX = __ ConX(0); |
ysr@777 | 3978 | |
ysr@777 | 3979 | // Get the alias_index for raw card-mark memory |
ysr@777 | 3980 | const TypePtr* card_type = TypeRawPtr::BOTTOM; |
ysr@777 | 3981 | |
ysr@777 | 3982 | const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); |
ysr@777 | 3983 | |
ysr@777 | 3984 | // Offsets into the thread |
ysr@777 | 3985 | const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + |
ysr@777 | 3986 | PtrQueue::byte_offset_of_index()); |
ysr@777 | 3987 | const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + |
ysr@777 | 3988 | PtrQueue::byte_offset_of_buf()); |
ysr@777 | 3989 | |
ysr@777 | 3990 | // Pointers into the thread |
ysr@777 | 3991 | |
kvn@1286 | 3992 | Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); |
kvn@1286 | 3993 | Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); |
ysr@777 | 3994 | |
ysr@777 | 3995 | // Now some values |
johnc@1948 | 3996 | // Use ctrl to avoid hoisting these values past a safepoint, which could |
johnc@1948 | 3997 | // potentially reset these fields in the JavaThread. |
johnc@5017 | 3998 | Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); |
johnc@1948 | 3999 | Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); |
ysr@777 | 4000 | |
ysr@777 | 4001 | // Convert the store obj pointer to an int prior to doing math on it |
ysr@777 | 4002 | // Must use ctrl to prevent "integerized oop" existing across safepoint |
kvn@1286 | 4003 | Node* cast = __ CastPX(__ ctrl(), adr); |
ysr@777 | 4004 | |
ysr@777 | 4005 | // Divide pointer by card size |
ysr@777 | 4006 | Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); |
ysr@777 | 4007 | |
ysr@777 | 4008 | // Combine card table base and card offset |
kvn@1286 | 4009 | Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); |
ysr@777 | 4010 | |
ysr@777 | 4011 | // If we know the value being stored does it cross regions? |
ysr@777 | 4012 | |
ysr@777 | 4013 | if (val != NULL) { |
ysr@777 | 4014 | // Does the store cause us to cross regions? |
ysr@777 | 4015 | |
ysr@777 | 4016 | // Should be able to do an unsigned compare of region_size instead of |
ysr@777 | 4017 | // and extra shift. Do we have an unsigned compare?? |
ysr@777 | 4018 | // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); |
ysr@777 | 4019 | Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); |
ysr@777 | 4020 | |
ysr@777 | 4021 | // if (xor_res == 0) same region so skip |
ysr@777 | 4022 | __ if_then(xor_res, BoolTest::ne, zeroX); { |
ysr@777 | 4023 | |
ysr@777 | 4024 | // No barrier if we are storing a NULL |
ysr@777 | 4025 | __ if_then(val, BoolTest::ne, null(), unlikely); { |
ysr@777 | 4026 | |
ysr@777 | 4027 | // Ok must mark the card if not already dirty |
ysr@777 | 4028 | |
ysr@777 | 4029 | // load the original value of the card |
ysr@777 | 4030 | Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); |
ysr@777 | 4031 | |
mgerdin@5860 | 4032 | __ if_then(card_val, BoolTest::ne, young_card); { |
mgerdin@5860 | 4033 | sync_kit(ideal); |
mgerdin@5860 | 4034 | // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. |
mgerdin@5860 | 4035 | insert_mem_bar(Op_MemBarVolatile, oop_store); |
mgerdin@5860 | 4036 | __ sync_kit(this); |
mgerdin@5860 | 4037 | |
mgerdin@5860 | 4038 | Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); |
mgerdin@5860 | 4039 | __ if_then(card_val_reload, BoolTest::ne, dirty_card); { |
mgerdin@5860 | 4040 | g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); |
mgerdin@5860 | 4041 | } __ end_if(); |
ysr@777 | 4042 | } __ end_if(); |
ysr@777 | 4043 | } __ end_if(); |
ysr@777 | 4044 | } __ end_if(); |
ysr@777 | 4045 | } else { |
kvn@1286 | 4046 | // Object.clone() instrinsic uses this path. |
cfang@1420 | 4047 | g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); |
ysr@777 | 4048 | } |
ysr@777 | 4049 | |
kvn@1286 | 4050 | // Final sync IdealKit and GraphKit. |
kvn@2726 | 4051 | final_sync(ideal); |
kvn@1286 | 4052 | } |
ysr@777 | 4053 | #undef __ |
kvn@3760 | 4054 | |
kvn@3760 | 4055 | |
kvn@3760 | 4056 | |
kvn@3760 | 4057 | Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { |
kvn@3760 | 4058 | if (java_lang_String::has_offset_field()) { |
kvn@3760 | 4059 | int offset_offset = java_lang_String::offset_offset_in_bytes(); |
kvn@3760 | 4060 | const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), |
kvn@3760 | 4061 | false, NULL, 0); |
kvn@3760 | 4062 | const TypePtr* offset_field_type = string_type->add_offset(offset_offset); |
kvn@3760 | 4063 | int offset_field_idx = C->get_alias_index(offset_field_type); |
kvn@3760 | 4064 | return make_load(ctrl, |
kvn@3760 | 4065 | basic_plus_adr(str, str, offset_offset), |
goetz@6479 | 4066 | TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); |
kvn@3760 | 4067 | } else { |
kvn@3760 | 4068 | return intcon(0); |
kvn@3760 | 4069 | } |
kvn@3760 | 4070 | } |
kvn@3760 | 4071 | |
kvn@3760 | 4072 | Node* GraphKit::load_String_length(Node* ctrl, Node* str) { |
kvn@3760 | 4073 | if (java_lang_String::has_count_field()) { |
kvn@3760 | 4074 | int count_offset = java_lang_String::count_offset_in_bytes(); |
kvn@3760 | 4075 | const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), |
kvn@3760 | 4076 | false, NULL, 0); |
kvn@3760 | 4077 | const TypePtr* count_field_type = string_type->add_offset(count_offset); |
kvn@3760 | 4078 | int count_field_idx = C->get_alias_index(count_field_type); |
kvn@3760 | 4079 | return make_load(ctrl, |
kvn@3760 | 4080 | basic_plus_adr(str, str, count_offset), |
goetz@6479 | 4081 | TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); |
kvn@3760 | 4082 | } else { |
kvn@3760 | 4083 | return load_array_length(load_String_value(ctrl, str)); |
kvn@3760 | 4084 | } |
kvn@3760 | 4085 | } |
kvn@3760 | 4086 | |
kvn@3760 | 4087 | Node* GraphKit::load_String_value(Node* ctrl, Node* str) { |
kvn@3760 | 4088 | int value_offset = java_lang_String::value_offset_in_bytes(); |
kvn@3760 | 4089 | const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), |
kvn@3760 | 4090 | false, NULL, 0); |
kvn@3760 | 4091 | const TypePtr* value_field_type = string_type->add_offset(value_offset); |
kvn@3760 | 4092 | const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, |
kvn@3760 | 4093 | TypeAry::make(TypeInt::CHAR,TypeInt::POS), |
kvn@3760 | 4094 | ciTypeArrayKlass::make(T_CHAR), true, 0); |
kvn@3760 | 4095 | int value_field_idx = C->get_alias_index(value_field_type); |
vlivanov@5658 | 4096 | Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), |
goetz@6479 | 4097 | value_type, T_OBJECT, value_field_idx, MemNode::unordered); |
vlivanov@5658 | 4098 | // String.value field is known to be @Stable. |
vlivanov@5658 | 4099 | if (UseImplicitStableValues) { |
vlivanov@5658 | 4100 | load = cast_array_to_stable(load, value_type); |
vlivanov@5658 | 4101 | } |
vlivanov@5658 | 4102 | return load; |
kvn@3760 | 4103 | } |
kvn@3760 | 4104 | |
kvn@3760 | 4105 | void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { |
kvn@3760 | 4106 | int offset_offset = java_lang_String::offset_offset_in_bytes(); |
kvn@3760 | 4107 | const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), |
kvn@3760 | 4108 | false, NULL, 0); |
kvn@3760 | 4109 | const TypePtr* offset_field_type = string_type->add_offset(offset_offset); |
kvn@3760 | 4110 | int offset_field_idx = C->get_alias_index(offset_field_type); |
kvn@3760 | 4111 | store_to_memory(ctrl, basic_plus_adr(str, offset_offset), |
goetz@6479 | 4112 | value, T_INT, offset_field_idx, MemNode::unordered); |
kvn@3760 | 4113 | } |
kvn@3760 | 4114 | |
kvn@3760 | 4115 | void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { |
kvn@3760 | 4116 | int value_offset = java_lang_String::value_offset_in_bytes(); |
kvn@3760 | 4117 | const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), |
kvn@3760 | 4118 | false, NULL, 0); |
kvn@3760 | 4119 | const TypePtr* value_field_type = string_type->add_offset(value_offset); |
twisti@5799 | 4120 | |
twisti@5799 | 4121 | store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, |
goetz@6479 | 4122 | value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); |
kvn@3760 | 4123 | } |
kvn@3760 | 4124 | |
kvn@3760 | 4125 | void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { |
kvn@3760 | 4126 | int count_offset = java_lang_String::count_offset_in_bytes(); |
kvn@3760 | 4127 | const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), |
kvn@3760 | 4128 | false, NULL, 0); |
kvn@3760 | 4129 | const TypePtr* count_field_type = string_type->add_offset(count_offset); |
kvn@3760 | 4130 | int count_field_idx = C->get_alias_index(count_field_type); |
kvn@3760 | 4131 | store_to_memory(ctrl, basic_plus_adr(str, count_offset), |
goetz@6479 | 4132 | value, T_INT, count_field_idx, MemNode::unordered); |
kvn@3760 | 4133 | } |
vlivanov@5658 | 4134 | |
vlivanov@5658 | 4135 | Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { |
vlivanov@5658 | 4136 | // Reify the property as a CastPP node in Ideal graph to comply with monotonicity |
vlivanov@5658 | 4137 | // assumption of CCP analysis. |
vlivanov@5658 | 4138 | return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true))); |
vlivanov@5658 | 4139 | } |