Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
duke@435 | 1 | /* |
kvn@1687 | 2 | * Copyright 2001-2010 Sun Microsystems, Inc. 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 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | #include "incls/_precompiled.incl" |
duke@435 | 26 | #include "incls/_graphKit.cpp.incl" |
duke@435 | 27 | |
duke@435 | 28 | //----------------------------GraphKit----------------------------------------- |
duke@435 | 29 | // Main utility constructor. |
duke@435 | 30 | GraphKit::GraphKit(JVMState* jvms) |
duke@435 | 31 | : Phase(Phase::Parser), |
duke@435 | 32 | _env(C->env()), |
duke@435 | 33 | _gvn(*C->initial_gvn()) |
duke@435 | 34 | { |
duke@435 | 35 | _exceptions = jvms->map()->next_exception(); |
duke@435 | 36 | if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); |
duke@435 | 37 | set_jvms(jvms); |
duke@435 | 38 | } |
duke@435 | 39 | |
duke@435 | 40 | // Private constructor for parser. |
duke@435 | 41 | GraphKit::GraphKit() |
duke@435 | 42 | : Phase(Phase::Parser), |
duke@435 | 43 | _env(C->env()), |
duke@435 | 44 | _gvn(*C->initial_gvn()) |
duke@435 | 45 | { |
duke@435 | 46 | _exceptions = NULL; |
duke@435 | 47 | set_map(NULL); |
duke@435 | 48 | debug_only(_sp = -99); |
duke@435 | 49 | debug_only(set_bci(-99)); |
duke@435 | 50 | } |
duke@435 | 51 | |
duke@435 | 52 | |
duke@435 | 53 | |
duke@435 | 54 | //---------------------------clean_stack--------------------------------------- |
duke@435 | 55 | // Clear away rubbish from the stack area of the JVM state. |
duke@435 | 56 | // This destroys any arguments that may be waiting on the stack. |
duke@435 | 57 | void GraphKit::clean_stack(int from_sp) { |
duke@435 | 58 | SafePointNode* map = this->map(); |
duke@435 | 59 | JVMState* jvms = this->jvms(); |
duke@435 | 60 | int stk_size = jvms->stk_size(); |
duke@435 | 61 | int stkoff = jvms->stkoff(); |
duke@435 | 62 | Node* top = this->top(); |
duke@435 | 63 | for (int i = from_sp; i < stk_size; i++) { |
duke@435 | 64 | if (map->in(stkoff + i) != top) { |
duke@435 | 65 | map->set_req(stkoff + i, top); |
duke@435 | 66 | } |
duke@435 | 67 | } |
duke@435 | 68 | } |
duke@435 | 69 | |
duke@435 | 70 | |
duke@435 | 71 | //--------------------------------sync_jvms----------------------------------- |
duke@435 | 72 | // Make sure our current jvms agrees with our parse state. |
duke@435 | 73 | JVMState* GraphKit::sync_jvms() const { |
duke@435 | 74 | JVMState* jvms = this->jvms(); |
duke@435 | 75 | jvms->set_bci(bci()); // Record the new bci in the JVMState |
duke@435 | 76 | jvms->set_sp(sp()); // Record the new sp in the JVMState |
duke@435 | 77 | assert(jvms_in_sync(), "jvms is now in sync"); |
duke@435 | 78 | return jvms; |
duke@435 | 79 | } |
duke@435 | 80 | |
duke@435 | 81 | #ifdef ASSERT |
duke@435 | 82 | bool GraphKit::jvms_in_sync() const { |
duke@435 | 83 | Parse* parse = is_Parse(); |
duke@435 | 84 | if (parse == NULL) { |
duke@435 | 85 | if (bci() != jvms()->bci()) return false; |
duke@435 | 86 | if (sp() != (int)jvms()->sp()) return false; |
duke@435 | 87 | return true; |
duke@435 | 88 | } |
duke@435 | 89 | if (jvms()->method() != parse->method()) return false; |
duke@435 | 90 | if (jvms()->bci() != parse->bci()) return false; |
duke@435 | 91 | int jvms_sp = jvms()->sp(); |
duke@435 | 92 | if (jvms_sp != parse->sp()) return false; |
duke@435 | 93 | int jvms_depth = jvms()->depth(); |
duke@435 | 94 | if (jvms_depth != parse->depth()) return false; |
duke@435 | 95 | return true; |
duke@435 | 96 | } |
duke@435 | 97 | |
duke@435 | 98 | // Local helper checks for special internal merge points |
duke@435 | 99 | // used to accumulate and merge exception states. |
duke@435 | 100 | // They are marked by the region's in(0) edge being the map itself. |
duke@435 | 101 | // Such merge points must never "escape" into the parser at large, |
duke@435 | 102 | // until they have been handed to gvn.transform. |
duke@435 | 103 | static bool is_hidden_merge(Node* reg) { |
duke@435 | 104 | if (reg == NULL) return false; |
duke@435 | 105 | if (reg->is_Phi()) { |
duke@435 | 106 | reg = reg->in(0); |
duke@435 | 107 | if (reg == NULL) return false; |
duke@435 | 108 | } |
duke@435 | 109 | return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); |
duke@435 | 110 | } |
duke@435 | 111 | |
duke@435 | 112 | void GraphKit::verify_map() const { |
duke@435 | 113 | if (map() == NULL) return; // null map is OK |
duke@435 | 114 | assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); |
duke@435 | 115 | assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); |
duke@435 | 116 | assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); |
duke@435 | 117 | } |
duke@435 | 118 | |
duke@435 | 119 | void GraphKit::verify_exception_state(SafePointNode* ex_map) { |
duke@435 | 120 | assert(ex_map->next_exception() == NULL, "not already part of a chain"); |
duke@435 | 121 | assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); |
duke@435 | 122 | } |
duke@435 | 123 | #endif |
duke@435 | 124 | |
duke@435 | 125 | //---------------------------stop_and_kill_map--------------------------------- |
duke@435 | 126 | // Set _map to NULL, signalling a stop to further bytecode execution. |
duke@435 | 127 | // First smash the current map's control to a constant, to mark it dead. |
duke@435 | 128 | void GraphKit::stop_and_kill_map() { |
duke@435 | 129 | SafePointNode* dead_map = stop(); |
duke@435 | 130 | if (dead_map != NULL) { |
duke@435 | 131 | dead_map->disconnect_inputs(NULL); // Mark the map as killed. |
duke@435 | 132 | assert(dead_map->is_killed(), "must be so marked"); |
duke@435 | 133 | } |
duke@435 | 134 | } |
duke@435 | 135 | |
duke@435 | 136 | |
duke@435 | 137 | //--------------------------------stopped-------------------------------------- |
duke@435 | 138 | // Tell if _map is NULL, or control is top. |
duke@435 | 139 | bool GraphKit::stopped() { |
duke@435 | 140 | if (map() == NULL) return true; |
duke@435 | 141 | else if (control() == top()) return true; |
duke@435 | 142 | else return false; |
duke@435 | 143 | } |
duke@435 | 144 | |
duke@435 | 145 | |
duke@435 | 146 | //-----------------------------has_ex_handler---------------------------------- |
duke@435 | 147 | // Tell if this method or any caller method has exception handlers. |
duke@435 | 148 | bool GraphKit::has_ex_handler() { |
duke@435 | 149 | for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { |
duke@435 | 150 | if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { |
duke@435 | 151 | return true; |
duke@435 | 152 | } |
duke@435 | 153 | } |
duke@435 | 154 | return false; |
duke@435 | 155 | } |
duke@435 | 156 | |
duke@435 | 157 | //------------------------------save_ex_oop------------------------------------ |
duke@435 | 158 | // Save an exception without blowing stack contents or other JVM state. |
duke@435 | 159 | void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { |
duke@435 | 160 | assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); |
duke@435 | 161 | ex_map->add_req(ex_oop); |
duke@435 | 162 | debug_only(verify_exception_state(ex_map)); |
duke@435 | 163 | } |
duke@435 | 164 | |
duke@435 | 165 | inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { |
duke@435 | 166 | assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); |
duke@435 | 167 | Node* ex_oop = ex_map->in(ex_map->req()-1); |
duke@435 | 168 | if (clear_it) ex_map->del_req(ex_map->req()-1); |
duke@435 | 169 | return ex_oop; |
duke@435 | 170 | } |
duke@435 | 171 | |
duke@435 | 172 | //-----------------------------saved_ex_oop------------------------------------ |
duke@435 | 173 | // Recover a saved exception from its map. |
duke@435 | 174 | Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { |
duke@435 | 175 | return common_saved_ex_oop(ex_map, false); |
duke@435 | 176 | } |
duke@435 | 177 | |
duke@435 | 178 | //--------------------------clear_saved_ex_oop--------------------------------- |
duke@435 | 179 | // Erase a previously saved exception from its map. |
duke@435 | 180 | Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { |
duke@435 | 181 | return common_saved_ex_oop(ex_map, true); |
duke@435 | 182 | } |
duke@435 | 183 | |
duke@435 | 184 | #ifdef ASSERT |
duke@435 | 185 | //---------------------------has_saved_ex_oop---------------------------------- |
duke@435 | 186 | // Erase a previously saved exception from its map. |
duke@435 | 187 | bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { |
duke@435 | 188 | return ex_map->req() == ex_map->jvms()->endoff()+1; |
duke@435 | 189 | } |
duke@435 | 190 | #endif |
duke@435 | 191 | |
duke@435 | 192 | //-------------------------make_exception_state-------------------------------- |
duke@435 | 193 | // Turn the current JVM state into an exception state, appending the ex_oop. |
duke@435 | 194 | SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { |
duke@435 | 195 | sync_jvms(); |
duke@435 | 196 | SafePointNode* ex_map = stop(); // do not manipulate this map any more |
duke@435 | 197 | set_saved_ex_oop(ex_map, ex_oop); |
duke@435 | 198 | return ex_map; |
duke@435 | 199 | } |
duke@435 | 200 | |
duke@435 | 201 | |
duke@435 | 202 | //--------------------------add_exception_state-------------------------------- |
duke@435 | 203 | // Add an exception to my list of exceptions. |
duke@435 | 204 | void GraphKit::add_exception_state(SafePointNode* ex_map) { |
duke@435 | 205 | if (ex_map == NULL || ex_map->control() == top()) { |
duke@435 | 206 | return; |
duke@435 | 207 | } |
duke@435 | 208 | #ifdef ASSERT |
duke@435 | 209 | verify_exception_state(ex_map); |
duke@435 | 210 | if (has_exceptions()) { |
duke@435 | 211 | assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); |
duke@435 | 212 | } |
duke@435 | 213 | #endif |
duke@435 | 214 | |
duke@435 | 215 | // If there is already an exception of exactly this type, merge with it. |
duke@435 | 216 | // In particular, null-checks and other low-level exceptions common up here. |
duke@435 | 217 | Node* ex_oop = saved_ex_oop(ex_map); |
duke@435 | 218 | const Type* ex_type = _gvn.type(ex_oop); |
duke@435 | 219 | if (ex_oop == top()) { |
duke@435 | 220 | // No action needed. |
duke@435 | 221 | return; |
duke@435 | 222 | } |
duke@435 | 223 | assert(ex_type->isa_instptr(), "exception must be an instance"); |
duke@435 | 224 | for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { |
duke@435 | 225 | const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); |
duke@435 | 226 | // We check sp also because call bytecodes can generate exceptions |
duke@435 | 227 | // both before and after arguments are popped! |
duke@435 | 228 | if (ex_type2 == ex_type |
duke@435 | 229 | && e2->_jvms->sp() == ex_map->_jvms->sp()) { |
duke@435 | 230 | combine_exception_states(ex_map, e2); |
duke@435 | 231 | return; |
duke@435 | 232 | } |
duke@435 | 233 | } |
duke@435 | 234 | |
duke@435 | 235 | // No pre-existing exception of the same type. Chain it on the list. |
duke@435 | 236 | push_exception_state(ex_map); |
duke@435 | 237 | } |
duke@435 | 238 | |
duke@435 | 239 | //-----------------------add_exception_states_from----------------------------- |
duke@435 | 240 | void GraphKit::add_exception_states_from(JVMState* jvms) { |
duke@435 | 241 | SafePointNode* ex_map = jvms->map()->next_exception(); |
duke@435 | 242 | if (ex_map != NULL) { |
duke@435 | 243 | jvms->map()->set_next_exception(NULL); |
duke@435 | 244 | for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { |
duke@435 | 245 | next_map = ex_map->next_exception(); |
duke@435 | 246 | ex_map->set_next_exception(NULL); |
duke@435 | 247 | add_exception_state(ex_map); |
duke@435 | 248 | } |
duke@435 | 249 | } |
duke@435 | 250 | } |
duke@435 | 251 | |
duke@435 | 252 | //-----------------------transfer_exceptions_into_jvms------------------------- |
duke@435 | 253 | JVMState* GraphKit::transfer_exceptions_into_jvms() { |
duke@435 | 254 | if (map() == NULL) { |
duke@435 | 255 | // We need a JVMS to carry the exceptions, but the map has gone away. |
duke@435 | 256 | // Create a scratch JVMS, cloned from any of the exception states... |
duke@435 | 257 | if (has_exceptions()) { |
duke@435 | 258 | _map = _exceptions; |
duke@435 | 259 | _map = clone_map(); |
duke@435 | 260 | _map->set_next_exception(NULL); |
duke@435 | 261 | clear_saved_ex_oop(_map); |
duke@435 | 262 | debug_only(verify_map()); |
duke@435 | 263 | } else { |
duke@435 | 264 | // ...or created from scratch |
duke@435 | 265 | JVMState* jvms = new (C) JVMState(_method, NULL); |
duke@435 | 266 | jvms->set_bci(_bci); |
duke@435 | 267 | jvms->set_sp(_sp); |
duke@435 | 268 | jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms)); |
duke@435 | 269 | set_jvms(jvms); |
duke@435 | 270 | for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); |
duke@435 | 271 | set_all_memory(top()); |
duke@435 | 272 | while (map()->req() < jvms->endoff()) map()->add_req(top()); |
duke@435 | 273 | } |
duke@435 | 274 | // (This is a kludge, in case you didn't notice.) |
duke@435 | 275 | set_control(top()); |
duke@435 | 276 | } |
duke@435 | 277 | JVMState* jvms = sync_jvms(); |
duke@435 | 278 | assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); |
duke@435 | 279 | jvms->map()->set_next_exception(_exceptions); |
duke@435 | 280 | _exceptions = NULL; // done with this set of exceptions |
duke@435 | 281 | return jvms; |
duke@435 | 282 | } |
duke@435 | 283 | |
duke@435 | 284 | static inline void add_n_reqs(Node* dstphi, Node* srcphi) { |
duke@435 | 285 | assert(is_hidden_merge(dstphi), "must be a special merge node"); |
duke@435 | 286 | assert(is_hidden_merge(srcphi), "must be a special merge node"); |
duke@435 | 287 | uint limit = srcphi->req(); |
duke@435 | 288 | for (uint i = PhiNode::Input; i < limit; i++) { |
duke@435 | 289 | dstphi->add_req(srcphi->in(i)); |
duke@435 | 290 | } |
duke@435 | 291 | } |
duke@435 | 292 | static inline void add_one_req(Node* dstphi, Node* src) { |
duke@435 | 293 | assert(is_hidden_merge(dstphi), "must be a special merge node"); |
duke@435 | 294 | assert(!is_hidden_merge(src), "must not be a special merge node"); |
duke@435 | 295 | dstphi->add_req(src); |
duke@435 | 296 | } |
duke@435 | 297 | |
duke@435 | 298 | //-----------------------combine_exception_states------------------------------ |
duke@435 | 299 | // This helper function combines exception states by building phis on a |
duke@435 | 300 | // specially marked state-merging region. These regions and phis are |
duke@435 | 301 | // untransformed, and can build up gradually. The region is marked by |
duke@435 | 302 | // having a control input of its exception map, rather than NULL. Such |
duke@435 | 303 | // regions do not appear except in this function, and in use_exception_state. |
duke@435 | 304 | void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { |
duke@435 | 305 | if (failing()) return; // dying anyway... |
duke@435 | 306 | JVMState* ex_jvms = ex_map->_jvms; |
duke@435 | 307 | assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); |
duke@435 | 308 | assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); |
duke@435 | 309 | assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); |
duke@435 | 310 | assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); |
duke@435 | 311 | assert(ex_map->req() == phi_map->req(), "matching maps"); |
duke@435 | 312 | uint tos = ex_jvms->stkoff() + ex_jvms->sp(); |
duke@435 | 313 | Node* hidden_merge_mark = root(); |
duke@435 | 314 | Node* region = phi_map->control(); |
duke@435 | 315 | MergeMemNode* phi_mem = phi_map->merged_memory(); |
duke@435 | 316 | MergeMemNode* ex_mem = ex_map->merged_memory(); |
duke@435 | 317 | if (region->in(0) != hidden_merge_mark) { |
duke@435 | 318 | // The control input is not (yet) a specially-marked region in phi_map. |
duke@435 | 319 | // Make it so, and build some phis. |
duke@435 | 320 | region = new (C, 2) RegionNode(2); |
duke@435 | 321 | _gvn.set_type(region, Type::CONTROL); |
duke@435 | 322 | region->set_req(0, hidden_merge_mark); // marks an internal ex-state |
duke@435 | 323 | region->init_req(1, phi_map->control()); |
duke@435 | 324 | phi_map->set_control(region); |
duke@435 | 325 | Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); |
duke@435 | 326 | record_for_igvn(io_phi); |
duke@435 | 327 | _gvn.set_type(io_phi, Type::ABIO); |
duke@435 | 328 | phi_map->set_i_o(io_phi); |
duke@435 | 329 | for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { |
duke@435 | 330 | Node* m = mms.memory(); |
duke@435 | 331 | Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); |
duke@435 | 332 | record_for_igvn(m_phi); |
duke@435 | 333 | _gvn.set_type(m_phi, Type::MEMORY); |
duke@435 | 334 | mms.set_memory(m_phi); |
duke@435 | 335 | } |
duke@435 | 336 | } |
duke@435 | 337 | |
duke@435 | 338 | // Either or both of phi_map and ex_map might already be converted into phis. |
duke@435 | 339 | Node* ex_control = ex_map->control(); |
duke@435 | 340 | // if there is special marking on ex_map also, we add multiple edges from src |
duke@435 | 341 | bool add_multiple = (ex_control->in(0) == hidden_merge_mark); |
duke@435 | 342 | // how wide was the destination phi_map, originally? |
duke@435 | 343 | uint orig_width = region->req(); |
duke@435 | 344 | |
duke@435 | 345 | if (add_multiple) { |
duke@435 | 346 | add_n_reqs(region, ex_control); |
duke@435 | 347 | add_n_reqs(phi_map->i_o(), ex_map->i_o()); |
duke@435 | 348 | } else { |
duke@435 | 349 | // ex_map has no merges, so we just add single edges everywhere |
duke@435 | 350 | add_one_req(region, ex_control); |
duke@435 | 351 | add_one_req(phi_map->i_o(), ex_map->i_o()); |
duke@435 | 352 | } |
duke@435 | 353 | for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { |
duke@435 | 354 | if (mms.is_empty()) { |
duke@435 | 355 | // get a copy of the base memory, and patch some inputs into it |
duke@435 | 356 | const TypePtr* adr_type = mms.adr_type(C); |
duke@435 | 357 | Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); |
duke@435 | 358 | assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); |
duke@435 | 359 | mms.set_memory(phi); |
duke@435 | 360 | // Prepare to append interesting stuff onto the newly sliced phi: |
duke@435 | 361 | while (phi->req() > orig_width) phi->del_req(phi->req()-1); |
duke@435 | 362 | } |
duke@435 | 363 | // Append stuff from ex_map: |
duke@435 | 364 | if (add_multiple) { |
duke@435 | 365 | add_n_reqs(mms.memory(), mms.memory2()); |
duke@435 | 366 | } else { |
duke@435 | 367 | add_one_req(mms.memory(), mms.memory2()); |
duke@435 | 368 | } |
duke@435 | 369 | } |
duke@435 | 370 | uint limit = ex_map->req(); |
duke@435 | 371 | for (uint i = TypeFunc::Parms; i < limit; i++) { |
duke@435 | 372 | // Skip everything in the JVMS after tos. (The ex_oop follows.) |
duke@435 | 373 | if (i == tos) i = ex_jvms->monoff(); |
duke@435 | 374 | Node* src = ex_map->in(i); |
duke@435 | 375 | Node* dst = phi_map->in(i); |
duke@435 | 376 | if (src != dst) { |
duke@435 | 377 | PhiNode* phi; |
duke@435 | 378 | if (dst->in(0) != region) { |
duke@435 | 379 | dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); |
duke@435 | 380 | record_for_igvn(phi); |
duke@435 | 381 | _gvn.set_type(phi, phi->type()); |
duke@435 | 382 | phi_map->set_req(i, dst); |
duke@435 | 383 | // Prepare to append interesting stuff onto the new phi: |
duke@435 | 384 | while (dst->req() > orig_width) dst->del_req(dst->req()-1); |
duke@435 | 385 | } else { |
duke@435 | 386 | assert(dst->is_Phi(), "nobody else uses a hidden region"); |
duke@435 | 387 | phi = (PhiNode*)dst; |
duke@435 | 388 | } |
duke@435 | 389 | if (add_multiple && src->in(0) == ex_control) { |
duke@435 | 390 | // Both are phis. |
duke@435 | 391 | add_n_reqs(dst, src); |
duke@435 | 392 | } else { |
duke@435 | 393 | while (dst->req() < region->req()) add_one_req(dst, src); |
duke@435 | 394 | } |
duke@435 | 395 | const Type* srctype = _gvn.type(src); |
duke@435 | 396 | if (phi->type() != srctype) { |
duke@435 | 397 | const Type* dsttype = phi->type()->meet(srctype); |
duke@435 | 398 | if (phi->type() != dsttype) { |
duke@435 | 399 | phi->set_type(dsttype); |
duke@435 | 400 | _gvn.set_type(phi, dsttype); |
duke@435 | 401 | } |
duke@435 | 402 | } |
duke@435 | 403 | } |
duke@435 | 404 | } |
duke@435 | 405 | } |
duke@435 | 406 | |
duke@435 | 407 | //--------------------------use_exception_state-------------------------------- |
duke@435 | 408 | Node* GraphKit::use_exception_state(SafePointNode* phi_map) { |
duke@435 | 409 | if (failing()) { stop(); return top(); } |
duke@435 | 410 | Node* region = phi_map->control(); |
duke@435 | 411 | Node* hidden_merge_mark = root(); |
duke@435 | 412 | assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); |
duke@435 | 413 | Node* ex_oop = clear_saved_ex_oop(phi_map); |
duke@435 | 414 | if (region->in(0) == hidden_merge_mark) { |
duke@435 | 415 | // Special marking for internal ex-states. Process the phis now. |
duke@435 | 416 | region->set_req(0, region); // now it's an ordinary region |
duke@435 | 417 | set_jvms(phi_map->jvms()); // ...so now we can use it as a map |
duke@435 | 418 | // Note: Setting the jvms also sets the bci and sp. |
duke@435 | 419 | set_control(_gvn.transform(region)); |
duke@435 | 420 | uint tos = jvms()->stkoff() + sp(); |
duke@435 | 421 | for (uint i = 1; i < tos; i++) { |
duke@435 | 422 | Node* x = phi_map->in(i); |
duke@435 | 423 | if (x->in(0) == region) { |
duke@435 | 424 | assert(x->is_Phi(), "expected a special phi"); |
duke@435 | 425 | phi_map->set_req(i, _gvn.transform(x)); |
duke@435 | 426 | } |
duke@435 | 427 | } |
duke@435 | 428 | for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { |
duke@435 | 429 | Node* x = mms.memory(); |
duke@435 | 430 | if (x->in(0) == region) { |
duke@435 | 431 | assert(x->is_Phi(), "nobody else uses a hidden region"); |
duke@435 | 432 | mms.set_memory(_gvn.transform(x)); |
duke@435 | 433 | } |
duke@435 | 434 | } |
duke@435 | 435 | if (ex_oop->in(0) == region) { |
duke@435 | 436 | assert(ex_oop->is_Phi(), "expected a special phi"); |
duke@435 | 437 | ex_oop = _gvn.transform(ex_oop); |
duke@435 | 438 | } |
duke@435 | 439 | } else { |
duke@435 | 440 | set_jvms(phi_map->jvms()); |
duke@435 | 441 | } |
duke@435 | 442 | |
duke@435 | 443 | assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); |
duke@435 | 444 | assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); |
duke@435 | 445 | return ex_oop; |
duke@435 | 446 | } |
duke@435 | 447 | |
duke@435 | 448 | //---------------------------------java_bc------------------------------------- |
duke@435 | 449 | Bytecodes::Code GraphKit::java_bc() const { |
duke@435 | 450 | ciMethod* method = this->method(); |
duke@435 | 451 | int bci = this->bci(); |
duke@435 | 452 | if (method != NULL && bci != InvocationEntryBci) |
duke@435 | 453 | return method->java_code_at_bci(bci); |
duke@435 | 454 | else |
duke@435 | 455 | return Bytecodes::_illegal; |
duke@435 | 456 | } |
duke@435 | 457 | |
dcubed@1648 | 458 | void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, |
dcubed@1648 | 459 | bool must_throw) { |
dcubed@1648 | 460 | // if the exception capability is set, then we will generate code |
dcubed@1648 | 461 | // to check the JavaThread.should_post_on_exceptions flag to see |
dcubed@1648 | 462 | // if we actually need to report exception events (for this |
dcubed@1648 | 463 | // thread). If we don't need to report exception events, we will |
dcubed@1648 | 464 | // take the normal fast path provided by add_exception_events. If |
dcubed@1648 | 465 | // exception event reporting is enabled for this thread, we will |
dcubed@1648 | 466 | // take the uncommon_trap in the BuildCutout below. |
dcubed@1648 | 467 | |
dcubed@1648 | 468 | // first must access the should_post_on_exceptions_flag in this thread's JavaThread |
dcubed@1648 | 469 | Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode()); |
dcubed@1648 | 470 | Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); |
dcubed@1648 | 471 | Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false); |
dcubed@1648 | 472 | |
dcubed@1648 | 473 | // Test the should_post_on_exceptions_flag vs. 0 |
dcubed@1648 | 474 | Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) ); |
dcubed@1648 | 475 | Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) ); |
dcubed@1648 | 476 | |
dcubed@1648 | 477 | // Branch to slow_path if should_post_on_exceptions_flag was true |
dcubed@1648 | 478 | { BuildCutout unless(this, tst, PROB_MAX); |
dcubed@1648 | 479 | // Do not try anything fancy if we're notifying the VM on every throw. |
dcubed@1648 | 480 | // Cf. case Bytecodes::_athrow in parse2.cpp. |
dcubed@1648 | 481 | uncommon_trap(reason, Deoptimization::Action_none, |
dcubed@1648 | 482 | (ciKlass*)NULL, (char*)NULL, must_throw); |
dcubed@1648 | 483 | } |
dcubed@1648 | 484 | |
dcubed@1648 | 485 | } |
dcubed@1648 | 486 | |
duke@435 | 487 | //------------------------------builtin_throw---------------------------------- |
duke@435 | 488 | void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { |
duke@435 | 489 | bool must_throw = true; |
duke@435 | 490 | |
dcubed@1648 | 491 | if (env()->jvmti_can_post_on_exceptions()) { |
dcubed@1648 | 492 | // check if we must post exception events, take uncommon trap if so |
dcubed@1648 | 493 | uncommon_trap_if_should_post_on_exceptions(reason, must_throw); |
dcubed@1648 | 494 | // here if should_post_on_exceptions is false |
dcubed@1648 | 495 | // continue on with the normal codegen |
duke@435 | 496 | } |
duke@435 | 497 | |
duke@435 | 498 | // If this particular condition has not yet happened at this |
duke@435 | 499 | // bytecode, then use the uncommon trap mechanism, and allow for |
duke@435 | 500 | // a future recompilation if several traps occur here. |
duke@435 | 501 | // If the throw is hot, try to use a more complicated inline mechanism |
duke@435 | 502 | // which keeps execution inside the compiled code. |
duke@435 | 503 | bool treat_throw_as_hot = false; |
duke@435 | 504 | ciMethodData* md = method()->method_data(); |
duke@435 | 505 | |
duke@435 | 506 | if (ProfileTraps) { |
duke@435 | 507 | if (too_many_traps(reason)) { |
duke@435 | 508 | treat_throw_as_hot = true; |
duke@435 | 509 | } |
duke@435 | 510 | // (If there is no MDO at all, assume it is early in |
duke@435 | 511 | // execution, and that any deopts are part of the |
duke@435 | 512 | // startup transient, and don't need to be remembered.) |
duke@435 | 513 | |
duke@435 | 514 | // Also, if there is a local exception handler, treat all throws |
duke@435 | 515 | // as hot if there has been at least one in this method. |
duke@435 | 516 | if (C->trap_count(reason) != 0 |
duke@435 | 517 | && method()->method_data()->trap_count(reason) != 0 |
duke@435 | 518 | && has_ex_handler()) { |
duke@435 | 519 | treat_throw_as_hot = true; |
duke@435 | 520 | } |
duke@435 | 521 | } |
duke@435 | 522 | |
duke@435 | 523 | // If this throw happens frequently, an uncommon trap might cause |
duke@435 | 524 | // a performance pothole. If there is a local exception handler, |
duke@435 | 525 | // and if this particular bytecode appears to be deoptimizing often, |
duke@435 | 526 | // let us handle the throw inline, with a preconstructed instance. |
duke@435 | 527 | // Note: If the deopt count has blown up, the uncommon trap |
duke@435 | 528 | // runtime is going to flush this nmethod, not matter what. |
duke@435 | 529 | if (treat_throw_as_hot |
duke@435 | 530 | && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { |
duke@435 | 531 | // If the throw is local, we use a pre-existing instance and |
duke@435 | 532 | // punt on the backtrace. This would lead to a missing backtrace |
duke@435 | 533 | // (a repeat of 4292742) if the backtrace object is ever asked |
duke@435 | 534 | // for its backtrace. |
duke@435 | 535 | // Fixing this remaining case of 4292742 requires some flavor of |
duke@435 | 536 | // escape analysis. Leave that for the future. |
duke@435 | 537 | ciInstance* ex_obj = NULL; |
duke@435 | 538 | switch (reason) { |
duke@435 | 539 | case Deoptimization::Reason_null_check: |
duke@435 | 540 | ex_obj = env()->NullPointerException_instance(); |
duke@435 | 541 | break; |
duke@435 | 542 | case Deoptimization::Reason_div0_check: |
duke@435 | 543 | ex_obj = env()->ArithmeticException_instance(); |
duke@435 | 544 | break; |
duke@435 | 545 | case Deoptimization::Reason_range_check: |
duke@435 | 546 | ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); |
duke@435 | 547 | break; |
duke@435 | 548 | case Deoptimization::Reason_class_check: |
duke@435 | 549 | if (java_bc() == Bytecodes::_aastore) { |
duke@435 | 550 | ex_obj = env()->ArrayStoreException_instance(); |
duke@435 | 551 | } else { |
duke@435 | 552 | ex_obj = env()->ClassCastException_instance(); |
duke@435 | 553 | } |
duke@435 | 554 | break; |
duke@435 | 555 | } |
duke@435 | 556 | if (failing()) { stop(); return; } // exception allocation might fail |
duke@435 | 557 | if (ex_obj != NULL) { |
duke@435 | 558 | // Cheat with a preallocated exception object. |
duke@435 | 559 | if (C->log() != NULL) |
duke@435 | 560 | C->log()->elem("hot_throw preallocated='1' reason='%s'", |
duke@435 | 561 | Deoptimization::trap_reason_name(reason)); |
duke@435 | 562 | const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); |
kvn@599 | 563 | Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); |
duke@435 | 564 | |
duke@435 | 565 | // Clear the detail message of the preallocated exception object. |
duke@435 | 566 | // Weblogic sometimes mutates the detail message of exceptions |
duke@435 | 567 | // using reflection. |
duke@435 | 568 | int offset = java_lang_Throwable::get_detailMessage_offset(); |
duke@435 | 569 | const TypePtr* adr_typ = ex_con->add_offset(offset); |
duke@435 | 570 | |
duke@435 | 571 | Node *adr = basic_plus_adr(ex_node, ex_node, offset); |
duke@435 | 572 | Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT); |
duke@435 | 573 | |
duke@435 | 574 | add_exception_state(make_exception_state(ex_node)); |
duke@435 | 575 | return; |
duke@435 | 576 | } |
duke@435 | 577 | } |
duke@435 | 578 | |
duke@435 | 579 | // %%% Maybe add entry to OptoRuntime which directly throws the exc.? |
duke@435 | 580 | // It won't be much cheaper than bailing to the interp., since we'll |
duke@435 | 581 | // have to pass up all the debug-info, and the runtime will have to |
duke@435 | 582 | // create the stack trace. |
duke@435 | 583 | |
duke@435 | 584 | // Usual case: Bail to interpreter. |
duke@435 | 585 | // Reserve the right to recompile if we haven't seen anything yet. |
duke@435 | 586 | |
duke@435 | 587 | Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; |
duke@435 | 588 | if (treat_throw_as_hot |
duke@435 | 589 | && (method()->method_data()->trap_recompiled_at(bci()) |
duke@435 | 590 | || C->too_many_traps(reason))) { |
duke@435 | 591 | // We cannot afford to take more traps here. Suffer in the interpreter. |
duke@435 | 592 | if (C->log() != NULL) |
duke@435 | 593 | C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", |
duke@435 | 594 | Deoptimization::trap_reason_name(reason), |
duke@435 | 595 | C->trap_count(reason)); |
duke@435 | 596 | action = Deoptimization::Action_none; |
duke@435 | 597 | } |
duke@435 | 598 | |
duke@435 | 599 | // "must_throw" prunes the JVM state to include only the stack, if there |
duke@435 | 600 | // are no local exception handlers. This should cut down on register |
duke@435 | 601 | // allocation time and code size, by drastically reducing the number |
duke@435 | 602 | // of in-edges on the call to the uncommon trap. |
duke@435 | 603 | |
duke@435 | 604 | uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); |
duke@435 | 605 | } |
duke@435 | 606 | |
duke@435 | 607 | |
duke@435 | 608 | //----------------------------PreserveJVMState--------------------------------- |
duke@435 | 609 | PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { |
duke@435 | 610 | debug_only(kit->verify_map()); |
duke@435 | 611 | _kit = kit; |
duke@435 | 612 | _map = kit->map(); // preserve the map |
duke@435 | 613 | _sp = kit->sp(); |
duke@435 | 614 | kit->set_map(clone_map ? kit->clone_map() : NULL); |
duke@435 | 615 | #ifdef ASSERT |
duke@435 | 616 | _bci = kit->bci(); |
duke@435 | 617 | Parse* parser = kit->is_Parse(); |
never@802 | 618 | int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); |
duke@435 | 619 | _block = block; |
duke@435 | 620 | #endif |
duke@435 | 621 | } |
duke@435 | 622 | PreserveJVMState::~PreserveJVMState() { |
duke@435 | 623 | GraphKit* kit = _kit; |
duke@435 | 624 | #ifdef ASSERT |
duke@435 | 625 | assert(kit->bci() == _bci, "bci must not shift"); |
duke@435 | 626 | Parse* parser = kit->is_Parse(); |
never@802 | 627 | int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); |
duke@435 | 628 | assert(block == _block, "block must not shift"); |
duke@435 | 629 | #endif |
duke@435 | 630 | kit->set_map(_map); |
duke@435 | 631 | kit->set_sp(_sp); |
duke@435 | 632 | } |
duke@435 | 633 | |
duke@435 | 634 | |
duke@435 | 635 | //-----------------------------BuildCutout------------------------------------- |
duke@435 | 636 | BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) |
duke@435 | 637 | : PreserveJVMState(kit) |
duke@435 | 638 | { |
duke@435 | 639 | assert(p->is_Con() || p->is_Bool(), "test must be a bool"); |
duke@435 | 640 | SafePointNode* outer_map = _map; // preserved map is caller's |
duke@435 | 641 | SafePointNode* inner_map = kit->map(); |
duke@435 | 642 | IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); |
duke@435 | 643 | outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) )); |
duke@435 | 644 | inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) )); |
duke@435 | 645 | } |
duke@435 | 646 | BuildCutout::~BuildCutout() { |
duke@435 | 647 | GraphKit* kit = _kit; |
duke@435 | 648 | assert(kit->stopped(), "cutout code must stop, throw, return, etc."); |
duke@435 | 649 | } |
duke@435 | 650 | |
cfang@1335 | 651 | //---------------------------PreserveReexecuteState---------------------------- |
cfang@1335 | 652 | PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { |
kvn@1385 | 653 | assert(!kit->stopped(), "must call stopped() before"); |
cfang@1335 | 654 | _kit = kit; |
cfang@1335 | 655 | _sp = kit->sp(); |
cfang@1335 | 656 | _reexecute = kit->jvms()->_reexecute; |
cfang@1335 | 657 | } |
cfang@1335 | 658 | PreserveReexecuteState::~PreserveReexecuteState() { |
kvn@1385 | 659 | if (_kit->stopped()) return; |
cfang@1335 | 660 | _kit->jvms()->_reexecute = _reexecute; |
cfang@1335 | 661 | _kit->set_sp(_sp); |
cfang@1335 | 662 | } |
duke@435 | 663 | |
duke@435 | 664 | //------------------------------clone_map-------------------------------------- |
duke@435 | 665 | // Implementation of PreserveJVMState |
duke@435 | 666 | // |
duke@435 | 667 | // Only clone_map(...) here. If this function is only used in the |
duke@435 | 668 | // PreserveJVMState class we may want to get rid of this extra |
duke@435 | 669 | // function eventually and do it all there. |
duke@435 | 670 | |
duke@435 | 671 | SafePointNode* GraphKit::clone_map() { |
duke@435 | 672 | if (map() == NULL) return NULL; |
duke@435 | 673 | |
duke@435 | 674 | // Clone the memory edge first |
duke@435 | 675 | Node* mem = MergeMemNode::make(C, map()->memory()); |
duke@435 | 676 | gvn().set_type_bottom(mem); |
duke@435 | 677 | |
duke@435 | 678 | SafePointNode *clonemap = (SafePointNode*)map()->clone(); |
duke@435 | 679 | JVMState* jvms = this->jvms(); |
duke@435 | 680 | JVMState* clonejvms = jvms->clone_shallow(C); |
duke@435 | 681 | clonemap->set_memory(mem); |
duke@435 | 682 | clonemap->set_jvms(clonejvms); |
duke@435 | 683 | clonejvms->set_map(clonemap); |
duke@435 | 684 | record_for_igvn(clonemap); |
duke@435 | 685 | gvn().set_type_bottom(clonemap); |
duke@435 | 686 | return clonemap; |
duke@435 | 687 | } |
duke@435 | 688 | |
duke@435 | 689 | |
duke@435 | 690 | //-----------------------------set_map_clone----------------------------------- |
duke@435 | 691 | void GraphKit::set_map_clone(SafePointNode* m) { |
duke@435 | 692 | _map = m; |
duke@435 | 693 | _map = clone_map(); |
duke@435 | 694 | _map->set_next_exception(NULL); |
duke@435 | 695 | debug_only(verify_map()); |
duke@435 | 696 | } |
duke@435 | 697 | |
duke@435 | 698 | |
duke@435 | 699 | //----------------------------kill_dead_locals--------------------------------- |
duke@435 | 700 | // Detect any locals which are known to be dead, and force them to top. |
duke@435 | 701 | void GraphKit::kill_dead_locals() { |
duke@435 | 702 | // Consult the liveness information for the locals. If any |
duke@435 | 703 | // of them are unused, then they can be replaced by top(). This |
duke@435 | 704 | // should help register allocation time and cut down on the size |
duke@435 | 705 | // of the deoptimization information. |
duke@435 | 706 | |
duke@435 | 707 | // This call is made from many of the bytecode handling |
duke@435 | 708 | // subroutines called from the Big Switch in do_one_bytecode. |
duke@435 | 709 | // Every bytecode which might include a slow path is responsible |
duke@435 | 710 | // for killing its dead locals. The more consistent we |
duke@435 | 711 | // are about killing deads, the fewer useless phis will be |
duke@435 | 712 | // constructed for them at various merge points. |
duke@435 | 713 | |
duke@435 | 714 | // bci can be -1 (InvocationEntryBci). We return the entry |
duke@435 | 715 | // liveness for the method. |
duke@435 | 716 | |
duke@435 | 717 | if (method() == NULL || method()->code_size() == 0) { |
duke@435 | 718 | // We are building a graph for a call to a native method. |
duke@435 | 719 | // All locals are live. |
duke@435 | 720 | return; |
duke@435 | 721 | } |
duke@435 | 722 | |
duke@435 | 723 | ResourceMark rm; |
duke@435 | 724 | |
duke@435 | 725 | // Consult the liveness information for the locals. If any |
duke@435 | 726 | // of them are unused, then they can be replaced by top(). This |
duke@435 | 727 | // should help register allocation time and cut down on the size |
duke@435 | 728 | // of the deoptimization information. |
duke@435 | 729 | MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); |
duke@435 | 730 | |
duke@435 | 731 | int len = (int)live_locals.size(); |
duke@435 | 732 | assert(len <= jvms()->loc_size(), "too many live locals"); |
duke@435 | 733 | for (int local = 0; local < len; local++) { |
duke@435 | 734 | if (!live_locals.at(local)) { |
duke@435 | 735 | set_local(local, top()); |
duke@435 | 736 | } |
duke@435 | 737 | } |
duke@435 | 738 | } |
duke@435 | 739 | |
duke@435 | 740 | #ifdef ASSERT |
duke@435 | 741 | //-------------------------dead_locals_are_killed------------------------------ |
duke@435 | 742 | // Return true if all dead locals are set to top in the map. |
duke@435 | 743 | // Used to assert "clean" debug info at various points. |
duke@435 | 744 | bool GraphKit::dead_locals_are_killed() { |
duke@435 | 745 | if (method() == NULL || method()->code_size() == 0) { |
duke@435 | 746 | // No locals need to be dead, so all is as it should be. |
duke@435 | 747 | return true; |
duke@435 | 748 | } |
duke@435 | 749 | |
duke@435 | 750 | // Make sure somebody called kill_dead_locals upstream. |
duke@435 | 751 | ResourceMark rm; |
duke@435 | 752 | for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { |
duke@435 | 753 | if (jvms->loc_size() == 0) continue; // no locals to consult |
duke@435 | 754 | SafePointNode* map = jvms->map(); |
duke@435 | 755 | ciMethod* method = jvms->method(); |
duke@435 | 756 | int bci = jvms->bci(); |
duke@435 | 757 | if (jvms == this->jvms()) { |
duke@435 | 758 | bci = this->bci(); // it might not yet be synched |
duke@435 | 759 | } |
duke@435 | 760 | MethodLivenessResult live_locals = method->liveness_at_bci(bci); |
duke@435 | 761 | int len = (int)live_locals.size(); |
duke@435 | 762 | if (!live_locals.is_valid() || len == 0) |
duke@435 | 763 | // This method is trivial, or is poisoned by a breakpoint. |
duke@435 | 764 | return true; |
duke@435 | 765 | assert(len == jvms->loc_size(), "live map consistent with locals map"); |
duke@435 | 766 | for (int local = 0; local < len; local++) { |
duke@435 | 767 | if (!live_locals.at(local) && map->local(jvms, local) != top()) { |
duke@435 | 768 | if (PrintMiscellaneous && (Verbose || WizardMode)) { |
duke@435 | 769 | tty->print_cr("Zombie local %d: ", local); |
duke@435 | 770 | jvms->dump(); |
duke@435 | 771 | } |
duke@435 | 772 | return false; |
duke@435 | 773 | } |
duke@435 | 774 | } |
duke@435 | 775 | } |
duke@435 | 776 | return true; |
duke@435 | 777 | } |
duke@435 | 778 | |
duke@435 | 779 | #endif //ASSERT |
duke@435 | 780 | |
cfang@1335 | 781 | // Helper function for enforcing certain bytecodes to reexecute if |
cfang@1335 | 782 | // deoptimization happens |
kvn@1687 | 783 | static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { |
cfang@1335 | 784 | ciMethod* cur_method = jvms->method(); |
cfang@1335 | 785 | int cur_bci = jvms->bci(); |
cfang@1335 | 786 | if (cur_method != NULL && cur_bci != InvocationEntryBci) { |
cfang@1335 | 787 | Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); |
kvn@1687 | 788 | return Interpreter::bytecode_should_reexecute(code) || |
kvn@1687 | 789 | is_anewarray && code == Bytecodes::_multianewarray; |
kvn@1687 | 790 | // Reexecute _multianewarray bytecode which was replaced with |
kvn@1687 | 791 | // sequence of [a]newarray. See Parse::do_multianewarray(). |
kvn@1687 | 792 | // |
kvn@1687 | 793 | // Note: interpreter should not have it set since this optimization |
kvn@1687 | 794 | // is limited by dimensions and guarded by flag so in some cases |
kvn@1687 | 795 | // multianewarray() runtime calls will be generated and |
kvn@1687 | 796 | // the bytecode should not be reexecutes (stack will not be reset). |
cfang@1335 | 797 | } else |
cfang@1335 | 798 | return false; |
cfang@1335 | 799 | } |
cfang@1335 | 800 | |
duke@435 | 801 | // Helper function for adding JVMState and debug information to node |
duke@435 | 802 | void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { |
duke@435 | 803 | // Add the safepoint edges to the call (or other safepoint). |
duke@435 | 804 | |
duke@435 | 805 | // Make sure dead locals are set to top. This |
duke@435 | 806 | // should help register allocation time and cut down on the size |
duke@435 | 807 | // of the deoptimization information. |
duke@435 | 808 | assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); |
duke@435 | 809 | |
duke@435 | 810 | // Walk the inline list to fill in the correct set of JVMState's |
duke@435 | 811 | // Also fill in the associated edges for each JVMState. |
duke@435 | 812 | |
duke@435 | 813 | JVMState* youngest_jvms = sync_jvms(); |
duke@435 | 814 | |
duke@435 | 815 | // Do we need debug info here? If it is a SafePoint and this method |
duke@435 | 816 | // cannot de-opt, then we do NOT need any debug info. |
duke@435 | 817 | bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint); |
duke@435 | 818 | |
duke@435 | 819 | // If we are guaranteed to throw, we can prune everything but the |
duke@435 | 820 | // input to the current bytecode. |
duke@435 | 821 | bool can_prune_locals = false; |
duke@435 | 822 | uint stack_slots_not_pruned = 0; |
duke@435 | 823 | int inputs = 0, depth = 0; |
duke@435 | 824 | if (must_throw) { |
duke@435 | 825 | assert(method() == youngest_jvms->method(), "sanity"); |
duke@435 | 826 | if (compute_stack_effects(inputs, depth)) { |
duke@435 | 827 | can_prune_locals = true; |
duke@435 | 828 | stack_slots_not_pruned = inputs; |
duke@435 | 829 | } |
duke@435 | 830 | } |
duke@435 | 831 | |
kvn@1215 | 832 | if (env()->jvmti_can_examine_or_deopt_anywhere()) { |
duke@435 | 833 | // At any safepoint, this method can get breakpointed, which would |
duke@435 | 834 | // then require an immediate deoptimization. |
duke@435 | 835 | full_info = true; |
duke@435 | 836 | can_prune_locals = false; // do not prune locals |
duke@435 | 837 | stack_slots_not_pruned = 0; |
duke@435 | 838 | } |
duke@435 | 839 | |
duke@435 | 840 | // do not scribble on the input jvms |
duke@435 | 841 | JVMState* out_jvms = youngest_jvms->clone_deep(C); |
duke@435 | 842 | call->set_jvms(out_jvms); // Start jvms list for call node |
duke@435 | 843 | |
cfang@1335 | 844 | // For a known set of bytecodes, the interpreter should reexecute them if |
cfang@1335 | 845 | // deoptimization happens. We set the reexecute state for them here |
cfang@1335 | 846 | if (out_jvms->is_reexecute_undefined() && //don't change if already specified |
kvn@1687 | 847 | should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { |
cfang@1335 | 848 | out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed |
cfang@1335 | 849 | } |
cfang@1335 | 850 | |
duke@435 | 851 | // Presize the call: |
duke@435 | 852 | debug_only(uint non_debug_edges = call->req()); |
duke@435 | 853 | call->add_req_batch(top(), youngest_jvms->debug_depth()); |
duke@435 | 854 | assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); |
duke@435 | 855 | |
duke@435 | 856 | // Set up edges so that the call looks like this: |
duke@435 | 857 | // Call [state:] ctl io mem fptr retadr |
duke@435 | 858 | // [parms:] parm0 ... parmN |
duke@435 | 859 | // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN |
duke@435 | 860 | // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] |
duke@435 | 861 | // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN |
duke@435 | 862 | // Note that caller debug info precedes callee debug info. |
duke@435 | 863 | |
duke@435 | 864 | // Fill pointer walks backwards from "young:" to "root:" in the diagram above: |
duke@435 | 865 | uint debug_ptr = call->req(); |
duke@435 | 866 | |
duke@435 | 867 | // Loop over the map input edges associated with jvms, add them |
duke@435 | 868 | // to the call node, & reset all offsets to match call node array. |
duke@435 | 869 | for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { |
duke@435 | 870 | uint debug_end = debug_ptr; |
duke@435 | 871 | uint debug_start = debug_ptr - in_jvms->debug_size(); |
duke@435 | 872 | debug_ptr = debug_start; // back up the ptr |
duke@435 | 873 | |
duke@435 | 874 | uint p = debug_start; // walks forward in [debug_start, debug_end) |
duke@435 | 875 | uint j, k, l; |
duke@435 | 876 | SafePointNode* in_map = in_jvms->map(); |
duke@435 | 877 | out_jvms->set_map(call); |
duke@435 | 878 | |
duke@435 | 879 | if (can_prune_locals) { |
duke@435 | 880 | assert(in_jvms->method() == out_jvms->method(), "sanity"); |
duke@435 | 881 | // If the current throw can reach an exception handler in this JVMS, |
duke@435 | 882 | // then we must keep everything live that can reach that handler. |
duke@435 | 883 | // As a quick and dirty approximation, we look for any handlers at all. |
duke@435 | 884 | if (in_jvms->method()->has_exception_handlers()) { |
duke@435 | 885 | can_prune_locals = false; |
duke@435 | 886 | } |
duke@435 | 887 | } |
duke@435 | 888 | |
duke@435 | 889 | // Add the Locals |
duke@435 | 890 | k = in_jvms->locoff(); |
duke@435 | 891 | l = in_jvms->loc_size(); |
duke@435 | 892 | out_jvms->set_locoff(p); |
duke@435 | 893 | if (full_info && !can_prune_locals) { |
duke@435 | 894 | for (j = 0; j < l; j++) |
duke@435 | 895 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 896 | } else { |
duke@435 | 897 | p += l; // already set to top above by add_req_batch |
duke@435 | 898 | } |
duke@435 | 899 | |
duke@435 | 900 | // Add the Expression Stack |
duke@435 | 901 | k = in_jvms->stkoff(); |
duke@435 | 902 | l = in_jvms->sp(); |
duke@435 | 903 | out_jvms->set_stkoff(p); |
duke@435 | 904 | if (full_info && !can_prune_locals) { |
duke@435 | 905 | for (j = 0; j < l; j++) |
duke@435 | 906 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 907 | } else if (can_prune_locals && stack_slots_not_pruned != 0) { |
duke@435 | 908 | // Divide stack into {S0,...,S1}, where S0 is set to top. |
duke@435 | 909 | uint s1 = stack_slots_not_pruned; |
duke@435 | 910 | stack_slots_not_pruned = 0; // for next iteration |
duke@435 | 911 | if (s1 > l) s1 = l; |
duke@435 | 912 | uint s0 = l - s1; |
duke@435 | 913 | p += s0; // skip the tops preinstalled by add_req_batch |
duke@435 | 914 | for (j = s0; j < l; j++) |
duke@435 | 915 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 916 | } else { |
duke@435 | 917 | p += l; // already set to top above by add_req_batch |
duke@435 | 918 | } |
duke@435 | 919 | |
duke@435 | 920 | // Add the Monitors |
duke@435 | 921 | k = in_jvms->monoff(); |
duke@435 | 922 | l = in_jvms->mon_size(); |
duke@435 | 923 | out_jvms->set_monoff(p); |
duke@435 | 924 | for (j = 0; j < l; j++) |
duke@435 | 925 | call->set_req(p++, in_map->in(k+j)); |
duke@435 | 926 | |
kvn@498 | 927 | // Copy any scalar object fields. |
kvn@498 | 928 | k = in_jvms->scloff(); |
kvn@498 | 929 | l = in_jvms->scl_size(); |
kvn@498 | 930 | out_jvms->set_scloff(p); |
kvn@498 | 931 | for (j = 0; j < l; j++) |
kvn@498 | 932 | call->set_req(p++, in_map->in(k+j)); |
kvn@498 | 933 | |
duke@435 | 934 | // Finish the new jvms. |
duke@435 | 935 | out_jvms->set_endoff(p); |
duke@435 | 936 | |
duke@435 | 937 | assert(out_jvms->endoff() == debug_end, "fill ptr must match"); |
duke@435 | 938 | assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); |
duke@435 | 939 | assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); |
duke@435 | 940 | assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); |
kvn@498 | 941 | assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); |
duke@435 | 942 | assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); |
duke@435 | 943 | |
duke@435 | 944 | // Update the two tail pointers in parallel. |
duke@435 | 945 | out_jvms = out_jvms->caller(); |
duke@435 | 946 | in_jvms = in_jvms->caller(); |
duke@435 | 947 | } |
duke@435 | 948 | |
duke@435 | 949 | assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); |
duke@435 | 950 | |
duke@435 | 951 | // Test the correctness of JVMState::debug_xxx accessors: |
duke@435 | 952 | assert(call->jvms()->debug_start() == non_debug_edges, ""); |
duke@435 | 953 | assert(call->jvms()->debug_end() == call->req(), ""); |
duke@435 | 954 | assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); |
duke@435 | 955 | } |
duke@435 | 956 | |
duke@435 | 957 | bool GraphKit::compute_stack_effects(int& inputs, int& depth) { |
duke@435 | 958 | Bytecodes::Code code = java_bc(); |
duke@435 | 959 | if (code == Bytecodes::_wide) { |
duke@435 | 960 | code = method()->java_code_at_bci(bci() + 1); |
duke@435 | 961 | } |
duke@435 | 962 | |
duke@435 | 963 | BasicType rtype = T_ILLEGAL; |
duke@435 | 964 | int rsize = 0; |
duke@435 | 965 | |
duke@435 | 966 | if (code != Bytecodes::_illegal) { |
duke@435 | 967 | depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 |
duke@435 | 968 | rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V |
duke@435 | 969 | if (rtype < T_CONFLICT) |
duke@435 | 970 | rsize = type2size[rtype]; |
duke@435 | 971 | } |
duke@435 | 972 | |
duke@435 | 973 | switch (code) { |
duke@435 | 974 | case Bytecodes::_illegal: |
duke@435 | 975 | return false; |
duke@435 | 976 | |
duke@435 | 977 | case Bytecodes::_ldc: |
duke@435 | 978 | case Bytecodes::_ldc_w: |
duke@435 | 979 | case Bytecodes::_ldc2_w: |
duke@435 | 980 | inputs = 0; |
duke@435 | 981 | break; |
duke@435 | 982 | |
duke@435 | 983 | case Bytecodes::_dup: inputs = 1; break; |
duke@435 | 984 | case Bytecodes::_dup_x1: inputs = 2; break; |
duke@435 | 985 | case Bytecodes::_dup_x2: inputs = 3; break; |
duke@435 | 986 | case Bytecodes::_dup2: inputs = 2; break; |
duke@435 | 987 | case Bytecodes::_dup2_x1: inputs = 3; break; |
duke@435 | 988 | case Bytecodes::_dup2_x2: inputs = 4; break; |
duke@435 | 989 | case Bytecodes::_swap: inputs = 2; break; |
duke@435 | 990 | case Bytecodes::_arraylength: inputs = 1; break; |
duke@435 | 991 | |
duke@435 | 992 | case Bytecodes::_getstatic: |
duke@435 | 993 | case Bytecodes::_putstatic: |
duke@435 | 994 | case Bytecodes::_getfield: |
duke@435 | 995 | case Bytecodes::_putfield: |
duke@435 | 996 | { |
duke@435 | 997 | bool is_get = (depth >= 0), is_static = (depth & 1); |
duke@435 | 998 | bool ignore; |
duke@435 | 999 | ciBytecodeStream iter(method()); |
duke@435 | 1000 | iter.reset_to_bci(bci()); |
duke@435 | 1001 | iter.next(); |
duke@435 | 1002 | ciField* field = iter.get_field(ignore); |
duke@435 | 1003 | int size = field->type()->size(); |
duke@435 | 1004 | inputs = (is_static ? 0 : 1); |
duke@435 | 1005 | if (is_get) { |
duke@435 | 1006 | depth = size - inputs; |
duke@435 | 1007 | } else { |
duke@435 | 1008 | inputs += size; // putxxx pops the value from the stack |
duke@435 | 1009 | depth = - inputs; |
duke@435 | 1010 | } |
duke@435 | 1011 | } |
duke@435 | 1012 | break; |
duke@435 | 1013 | |
duke@435 | 1014 | case Bytecodes::_invokevirtual: |
duke@435 | 1015 | case Bytecodes::_invokespecial: |
duke@435 | 1016 | case Bytecodes::_invokestatic: |
jrose@1161 | 1017 | case Bytecodes::_invokedynamic: |
duke@435 | 1018 | case Bytecodes::_invokeinterface: |
duke@435 | 1019 | { |
duke@435 | 1020 | bool ignore; |
duke@435 | 1021 | ciBytecodeStream iter(method()); |
duke@435 | 1022 | iter.reset_to_bci(bci()); |
duke@435 | 1023 | iter.next(); |
duke@435 | 1024 | ciMethod* method = iter.get_method(ignore); |
duke@435 | 1025 | inputs = method->arg_size_no_receiver(); |
twisti@1572 | 1026 | // Add a receiver argument, maybe: |
twisti@1572 | 1027 | if (code != Bytecodes::_invokestatic && |
twisti@1572 | 1028 | code != Bytecodes::_invokedynamic) |
twisti@1572 | 1029 | inputs += 1; |
twisti@1572 | 1030 | // (Do not use ciMethod::arg_size(), because |
twisti@1572 | 1031 | // it might be an unloaded method, which doesn't |
twisti@1572 | 1032 | // know whether it is static or not.) |
duke@435 | 1033 | int size = method->return_type()->size(); |
duke@435 | 1034 | depth = size - inputs; |
duke@435 | 1035 | } |
duke@435 | 1036 | break; |
duke@435 | 1037 | |
duke@435 | 1038 | case Bytecodes::_multianewarray: |
duke@435 | 1039 | { |
duke@435 | 1040 | ciBytecodeStream iter(method()); |
duke@435 | 1041 | iter.reset_to_bci(bci()); |
duke@435 | 1042 | iter.next(); |
duke@435 | 1043 | inputs = iter.get_dimensions(); |
duke@435 | 1044 | assert(rsize == 1, ""); |
duke@435 | 1045 | depth = rsize - inputs; |
duke@435 | 1046 | } |
duke@435 | 1047 | break; |
duke@435 | 1048 | |
duke@435 | 1049 | case Bytecodes::_ireturn: |
duke@435 | 1050 | case Bytecodes::_lreturn: |
duke@435 | 1051 | case Bytecodes::_freturn: |
duke@435 | 1052 | case Bytecodes::_dreturn: |
duke@435 | 1053 | case Bytecodes::_areturn: |
duke@435 | 1054 | assert(rsize = -depth, ""); |
duke@435 | 1055 | inputs = rsize; |
duke@435 | 1056 | break; |
duke@435 | 1057 | |
duke@435 | 1058 | case Bytecodes::_jsr: |
duke@435 | 1059 | case Bytecodes::_jsr_w: |
duke@435 | 1060 | inputs = 0; |
duke@435 | 1061 | depth = 1; // S.B. depth=1, not zero |
duke@435 | 1062 | break; |
duke@435 | 1063 | |
duke@435 | 1064 | default: |
duke@435 | 1065 | // bytecode produces a typed result |
duke@435 | 1066 | inputs = rsize - depth; |
duke@435 | 1067 | assert(inputs >= 0, ""); |
duke@435 | 1068 | break; |
duke@435 | 1069 | } |
duke@435 | 1070 | |
duke@435 | 1071 | #ifdef ASSERT |
duke@435 | 1072 | // spot check |
duke@435 | 1073 | int outputs = depth + inputs; |
duke@435 | 1074 | assert(outputs >= 0, "sanity"); |
duke@435 | 1075 | switch (code) { |
duke@435 | 1076 | case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; |
duke@435 | 1077 | case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; |
duke@435 | 1078 | case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; |
duke@435 | 1079 | case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; |
duke@435 | 1080 | case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; |
duke@435 | 1081 | } |
duke@435 | 1082 | #endif //ASSERT |
duke@435 | 1083 | |
duke@435 | 1084 | return true; |
duke@435 | 1085 | } |
duke@435 | 1086 | |
duke@435 | 1087 | |
duke@435 | 1088 | |
duke@435 | 1089 | //------------------------------basic_plus_adr--------------------------------- |
duke@435 | 1090 | Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { |
duke@435 | 1091 | // short-circuit a common case |
duke@435 | 1092 | if (offset == intcon(0)) return ptr; |
duke@435 | 1093 | return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) ); |
duke@435 | 1094 | } |
duke@435 | 1095 | |
duke@435 | 1096 | Node* GraphKit::ConvI2L(Node* offset) { |
duke@435 | 1097 | // short-circuit a common case |
duke@435 | 1098 | jint offset_con = find_int_con(offset, Type::OffsetBot); |
duke@435 | 1099 | if (offset_con != Type::OffsetBot) { |
duke@435 | 1100 | return longcon((long) offset_con); |
duke@435 | 1101 | } |
duke@435 | 1102 | return _gvn.transform( new (C, 2) ConvI2LNode(offset)); |
duke@435 | 1103 | } |
duke@435 | 1104 | Node* GraphKit::ConvL2I(Node* offset) { |
duke@435 | 1105 | // short-circuit a common case |
duke@435 | 1106 | jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); |
duke@435 | 1107 | if (offset_con != (jlong)Type::OffsetBot) { |
duke@435 | 1108 | return intcon((int) offset_con); |
duke@435 | 1109 | } |
duke@435 | 1110 | return _gvn.transform( new (C, 2) ConvL2INode(offset)); |
duke@435 | 1111 | } |
duke@435 | 1112 | |
duke@435 | 1113 | //-------------------------load_object_klass----------------------------------- |
duke@435 | 1114 | Node* GraphKit::load_object_klass(Node* obj) { |
duke@435 | 1115 | // Special-case a fresh allocation to avoid building nodes: |
duke@435 | 1116 | Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); |
duke@435 | 1117 | if (akls != NULL) return akls; |
duke@435 | 1118 | Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); |
kvn@599 | 1119 | return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); |
duke@435 | 1120 | } |
duke@435 | 1121 | |
duke@435 | 1122 | //-------------------------load_array_length----------------------------------- |
duke@435 | 1123 | Node* GraphKit::load_array_length(Node* array) { |
duke@435 | 1124 | // Special-case a fresh allocation to avoid building nodes: |
rasbold@801 | 1125 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); |
rasbold@801 | 1126 | Node *alen; |
rasbold@801 | 1127 | if (alloc == NULL) { |
rasbold@801 | 1128 | Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); |
rasbold@801 | 1129 | alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); |
rasbold@801 | 1130 | } else { |
rasbold@801 | 1131 | alen = alloc->Ideal_length(); |
kvn@1389 | 1132 | Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); |
rasbold@801 | 1133 | if (ccast != alen) { |
rasbold@801 | 1134 | alen = _gvn.transform(ccast); |
rasbold@801 | 1135 | } |
rasbold@801 | 1136 | } |
rasbold@801 | 1137 | return alen; |
duke@435 | 1138 | } |
duke@435 | 1139 | |
duke@435 | 1140 | //------------------------------do_null_check---------------------------------- |
duke@435 | 1141 | // Helper function to do a NULL pointer check. Returned value is |
duke@435 | 1142 | // the incoming address with NULL casted away. You are allowed to use the |
duke@435 | 1143 | // not-null value only if you are control dependent on the test. |
duke@435 | 1144 | extern int explicit_null_checks_inserted, |
duke@435 | 1145 | explicit_null_checks_elided; |
duke@435 | 1146 | Node* GraphKit::null_check_common(Node* value, BasicType type, |
duke@435 | 1147 | // optional arguments for variations: |
duke@435 | 1148 | bool assert_null, |
duke@435 | 1149 | Node* *null_control) { |
duke@435 | 1150 | assert(!assert_null || null_control == NULL, "not both at once"); |
duke@435 | 1151 | if (stopped()) return top(); |
duke@435 | 1152 | if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { |
duke@435 | 1153 | // For some performance testing, we may wish to suppress null checking. |
duke@435 | 1154 | value = cast_not_null(value); // Make it appear to be non-null (4962416). |
duke@435 | 1155 | return value; |
duke@435 | 1156 | } |
duke@435 | 1157 | explicit_null_checks_inserted++; |
duke@435 | 1158 | |
duke@435 | 1159 | // Construct NULL check |
duke@435 | 1160 | Node *chk = NULL; |
duke@435 | 1161 | switch(type) { |
duke@435 | 1162 | case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break; |
duke@435 | 1163 | case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break; |
duke@435 | 1164 | case T_ARRAY : // fall through |
duke@435 | 1165 | type = T_OBJECT; // simplify further tests |
duke@435 | 1166 | case T_OBJECT : { |
duke@435 | 1167 | const Type *t = _gvn.type( value ); |
duke@435 | 1168 | |
kvn@1385 | 1169 | const TypeOopPtr* tp = t->isa_oopptr(); |
kvn@1390 | 1170 | if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() |
duke@435 | 1171 | // Only for do_null_check, not any of its siblings: |
duke@435 | 1172 | && !assert_null && null_control == NULL) { |
duke@435 | 1173 | // Usually, any field access or invocation on an unloaded oop type |
duke@435 | 1174 | // will simply fail to link, since the statically linked class is |
duke@435 | 1175 | // likely also to be unloaded. However, in -Xcomp mode, sometimes |
duke@435 | 1176 | // the static class is loaded but the sharper oop type is not. |
duke@435 | 1177 | // Rather than checking for this obscure case in lots of places, |
duke@435 | 1178 | // we simply observe that a null check on an unloaded class |
duke@435 | 1179 | // will always be followed by a nonsense operation, so we |
duke@435 | 1180 | // can just issue the uncommon trap here. |
duke@435 | 1181 | // Our access to the unloaded class will only be correct |
duke@435 | 1182 | // after it has been loaded and initialized, which requires |
duke@435 | 1183 | // a trip through the interpreter. |
duke@435 | 1184 | #ifndef PRODUCT |
duke@435 | 1185 | if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } |
duke@435 | 1186 | #endif |
duke@435 | 1187 | uncommon_trap(Deoptimization::Reason_unloaded, |
duke@435 | 1188 | Deoptimization::Action_reinterpret, |
duke@435 | 1189 | tp->klass(), "!loaded"); |
duke@435 | 1190 | return top(); |
duke@435 | 1191 | } |
duke@435 | 1192 | |
duke@435 | 1193 | if (assert_null) { |
duke@435 | 1194 | // See if the type is contained in NULL_PTR. |
duke@435 | 1195 | // If so, then the value is already null. |
duke@435 | 1196 | if (t->higher_equal(TypePtr::NULL_PTR)) { |
duke@435 | 1197 | explicit_null_checks_elided++; |
duke@435 | 1198 | return value; // Elided null assert quickly! |
duke@435 | 1199 | } |
duke@435 | 1200 | } else { |
duke@435 | 1201 | // See if mixing in the NULL pointer changes type. |
duke@435 | 1202 | // If so, then the NULL pointer was not allowed in the original |
duke@435 | 1203 | // type. In other words, "value" was not-null. |
duke@435 | 1204 | if (t->meet(TypePtr::NULL_PTR) != t) { |
duke@435 | 1205 | // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... |
duke@435 | 1206 | explicit_null_checks_elided++; |
duke@435 | 1207 | return value; // Elided null check quickly! |
duke@435 | 1208 | } |
duke@435 | 1209 | } |
duke@435 | 1210 | chk = new (C, 3) CmpPNode( value, null() ); |
duke@435 | 1211 | break; |
duke@435 | 1212 | } |
duke@435 | 1213 | |
duke@435 | 1214 | default : ShouldNotReachHere(); |
duke@435 | 1215 | } |
duke@435 | 1216 | assert(chk != NULL, "sanity check"); |
duke@435 | 1217 | chk = _gvn.transform(chk); |
duke@435 | 1218 | |
duke@435 | 1219 | BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; |
duke@435 | 1220 | BoolNode *btst = new (C, 2) BoolNode( chk, btest); |
duke@435 | 1221 | Node *tst = _gvn.transform( btst ); |
duke@435 | 1222 | |
duke@435 | 1223 | //----------- |
twisti@1040 | 1224 | // if peephole optimizations occurred, a prior test existed. |
duke@435 | 1225 | // If a prior test existed, maybe it dominates as we can avoid this test. |
duke@435 | 1226 | if (tst != btst && type == T_OBJECT) { |
duke@435 | 1227 | // At this point we want to scan up the CFG to see if we can |
duke@435 | 1228 | // find an identical test (and so avoid this test altogether). |
duke@435 | 1229 | Node *cfg = control(); |
duke@435 | 1230 | int depth = 0; |
duke@435 | 1231 | while( depth < 16 ) { // Limit search depth for speed |
duke@435 | 1232 | if( cfg->Opcode() == Op_IfTrue && |
duke@435 | 1233 | cfg->in(0)->in(1) == tst ) { |
duke@435 | 1234 | // Found prior test. Use "cast_not_null" to construct an identical |
duke@435 | 1235 | // CastPP (and hence hash to) as already exists for the prior test. |
duke@435 | 1236 | // Return that casted value. |
duke@435 | 1237 | if (assert_null) { |
duke@435 | 1238 | replace_in_map(value, null()); |
duke@435 | 1239 | return null(); // do not issue the redundant test |
duke@435 | 1240 | } |
duke@435 | 1241 | Node *oldcontrol = control(); |
duke@435 | 1242 | set_control(cfg); |
duke@435 | 1243 | Node *res = cast_not_null(value); |
duke@435 | 1244 | set_control(oldcontrol); |
duke@435 | 1245 | explicit_null_checks_elided++; |
duke@435 | 1246 | return res; |
duke@435 | 1247 | } |
duke@435 | 1248 | cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); |
duke@435 | 1249 | if (cfg == NULL) break; // Quit at region nodes |
duke@435 | 1250 | depth++; |
duke@435 | 1251 | } |
duke@435 | 1252 | } |
duke@435 | 1253 | |
duke@435 | 1254 | //----------- |
duke@435 | 1255 | // Branch to failure if null |
duke@435 | 1256 | float ok_prob = PROB_MAX; // a priori estimate: nulls never happen |
duke@435 | 1257 | Deoptimization::DeoptReason reason; |
duke@435 | 1258 | if (assert_null) |
duke@435 | 1259 | reason = Deoptimization::Reason_null_assert; |
duke@435 | 1260 | else if (type == T_OBJECT) |
duke@435 | 1261 | reason = Deoptimization::Reason_null_check; |
duke@435 | 1262 | else |
duke@435 | 1263 | reason = Deoptimization::Reason_div0_check; |
duke@435 | 1264 | |
ysr@777 | 1265 | // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, |
ysr@777 | 1266 | // ciMethodData::has_trap_at will return a conservative -1 if any |
ysr@777 | 1267 | // must-be-null assertion has failed. This could cause performance |
ysr@777 | 1268 | // problems for a method after its first do_null_assert failure. |
ysr@777 | 1269 | // Consider using 'Reason_class_check' instead? |
ysr@777 | 1270 | |
duke@435 | 1271 | // To cause an implicit null check, we set the not-null probability |
twisti@1040 | 1272 | // to the maximum (PROB_MAX). For an explicit check the probability |
duke@435 | 1273 | // is set to a smaller value. |
duke@435 | 1274 | if (null_control != NULL || too_many_traps(reason)) { |
duke@435 | 1275 | // probability is less likely |
duke@435 | 1276 | ok_prob = PROB_LIKELY_MAG(3); |
duke@435 | 1277 | } else if (!assert_null && |
duke@435 | 1278 | (ImplicitNullCheckThreshold > 0) && |
duke@435 | 1279 | method() != NULL && |
duke@435 | 1280 | (method()->method_data()->trap_count(reason) |
duke@435 | 1281 | >= (uint)ImplicitNullCheckThreshold)) { |
duke@435 | 1282 | ok_prob = PROB_LIKELY_MAG(3); |
duke@435 | 1283 | } |
duke@435 | 1284 | |
duke@435 | 1285 | if (null_control != NULL) { |
duke@435 | 1286 | IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); |
duke@435 | 1287 | Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff)); |
duke@435 | 1288 | set_control( _gvn.transform( new (C, 1) IfTrueNode(iff))); |
duke@435 | 1289 | if (null_true == top()) |
duke@435 | 1290 | explicit_null_checks_elided++; |
duke@435 | 1291 | (*null_control) = null_true; |
duke@435 | 1292 | } else { |
duke@435 | 1293 | BuildCutout unless(this, tst, ok_prob); |
duke@435 | 1294 | // Check for optimizer eliding test at parse time |
duke@435 | 1295 | if (stopped()) { |
duke@435 | 1296 | // Failure not possible; do not bother making uncommon trap. |
duke@435 | 1297 | explicit_null_checks_elided++; |
duke@435 | 1298 | } else if (assert_null) { |
duke@435 | 1299 | uncommon_trap(reason, |
duke@435 | 1300 | Deoptimization::Action_make_not_entrant, |
duke@435 | 1301 | NULL, "assert_null"); |
duke@435 | 1302 | } else { |
kvn@767 | 1303 | replace_in_map(value, zerocon(type)); |
duke@435 | 1304 | builtin_throw(reason); |
duke@435 | 1305 | } |
duke@435 | 1306 | } |
duke@435 | 1307 | |
duke@435 | 1308 | // Must throw exception, fall-thru not possible? |
duke@435 | 1309 | if (stopped()) { |
duke@435 | 1310 | return top(); // No result |
duke@435 | 1311 | } |
duke@435 | 1312 | |
duke@435 | 1313 | if (assert_null) { |
duke@435 | 1314 | // Cast obj to null on this path. |
duke@435 | 1315 | replace_in_map(value, zerocon(type)); |
duke@435 | 1316 | return zerocon(type); |
duke@435 | 1317 | } |
duke@435 | 1318 | |
duke@435 | 1319 | // Cast obj to not-null on this path, if there is no null_control. |
duke@435 | 1320 | // (If there is a null_control, a non-null value may come back to haunt us.) |
duke@435 | 1321 | if (type == T_OBJECT) { |
duke@435 | 1322 | Node* cast = cast_not_null(value, false); |
duke@435 | 1323 | if (null_control == NULL || (*null_control) == top()) |
duke@435 | 1324 | replace_in_map(value, cast); |
duke@435 | 1325 | value = cast; |
duke@435 | 1326 | } |
duke@435 | 1327 | |
duke@435 | 1328 | return value; |
duke@435 | 1329 | } |
duke@435 | 1330 | |
duke@435 | 1331 | |
duke@435 | 1332 | //------------------------------cast_not_null---------------------------------- |
duke@435 | 1333 | // Cast obj to not-null on this path |
duke@435 | 1334 | Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { |
duke@435 | 1335 | const Type *t = _gvn.type(obj); |
duke@435 | 1336 | const Type *t_not_null = t->join(TypePtr::NOTNULL); |
duke@435 | 1337 | // Object is already not-null? |
duke@435 | 1338 | if( t == t_not_null ) return obj; |
duke@435 | 1339 | |
duke@435 | 1340 | Node *cast = new (C, 2) CastPPNode(obj,t_not_null); |
duke@435 | 1341 | cast->init_req(0, control()); |
duke@435 | 1342 | cast = _gvn.transform( cast ); |
duke@435 | 1343 | |
duke@435 | 1344 | // Scan for instances of 'obj' in the current JVM mapping. |
duke@435 | 1345 | // These instances are known to be not-null after the test. |
duke@435 | 1346 | if (do_replace_in_map) |
duke@435 | 1347 | replace_in_map(obj, cast); |
duke@435 | 1348 | |
duke@435 | 1349 | return cast; // Return casted value |
duke@435 | 1350 | } |
duke@435 | 1351 | |
duke@435 | 1352 | |
duke@435 | 1353 | //--------------------------replace_in_map------------------------------------- |
duke@435 | 1354 | void GraphKit::replace_in_map(Node* old, Node* neww) { |
duke@435 | 1355 | this->map()->replace_edge(old, neww); |
duke@435 | 1356 | |
duke@435 | 1357 | // Note: This operation potentially replaces any edge |
duke@435 | 1358 | // on the map. This includes locals, stack, and monitors |
duke@435 | 1359 | // of the current (innermost) JVM state. |
duke@435 | 1360 | |
duke@435 | 1361 | // We can consider replacing in caller maps. |
duke@435 | 1362 | // The idea would be that an inlined function's null checks |
duke@435 | 1363 | // can be shared with the entire inlining tree. |
duke@435 | 1364 | // The expense of doing this is that the PreserveJVMState class |
duke@435 | 1365 | // would have to preserve caller states too, with a deep copy. |
duke@435 | 1366 | } |
duke@435 | 1367 | |
duke@435 | 1368 | |
duke@435 | 1369 | |
duke@435 | 1370 | //============================================================================= |
duke@435 | 1371 | //--------------------------------memory--------------------------------------- |
duke@435 | 1372 | Node* GraphKit::memory(uint alias_idx) { |
duke@435 | 1373 | MergeMemNode* mem = merged_memory(); |
duke@435 | 1374 | Node* p = mem->memory_at(alias_idx); |
duke@435 | 1375 | _gvn.set_type(p, Type::MEMORY); // must be mapped |
duke@435 | 1376 | return p; |
duke@435 | 1377 | } |
duke@435 | 1378 | |
duke@435 | 1379 | //-----------------------------reset_memory------------------------------------ |
duke@435 | 1380 | Node* GraphKit::reset_memory() { |
duke@435 | 1381 | Node* mem = map()->memory(); |
duke@435 | 1382 | // do not use this node for any more parsing! |
duke@435 | 1383 | debug_only( map()->set_memory((Node*)NULL) ); |
duke@435 | 1384 | return _gvn.transform( mem ); |
duke@435 | 1385 | } |
duke@435 | 1386 | |
duke@435 | 1387 | //------------------------------set_all_memory--------------------------------- |
duke@435 | 1388 | void GraphKit::set_all_memory(Node* newmem) { |
duke@435 | 1389 | Node* mergemem = MergeMemNode::make(C, newmem); |
duke@435 | 1390 | gvn().set_type_bottom(mergemem); |
duke@435 | 1391 | map()->set_memory(mergemem); |
duke@435 | 1392 | } |
duke@435 | 1393 | |
duke@435 | 1394 | //------------------------------set_all_memory_call---------------------------- |
never@1515 | 1395 | void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { |
never@1515 | 1396 | Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); |
duke@435 | 1397 | set_all_memory(newmem); |
duke@435 | 1398 | } |
duke@435 | 1399 | |
duke@435 | 1400 | //============================================================================= |
duke@435 | 1401 | // |
duke@435 | 1402 | // parser factory methods for MemNodes |
duke@435 | 1403 | // |
duke@435 | 1404 | // These are layered on top of the factory methods in LoadNode and StoreNode, |
duke@435 | 1405 | // and integrate with the parser's memory state and _gvn engine. |
duke@435 | 1406 | // |
duke@435 | 1407 | |
duke@435 | 1408 | // factory methods in "int adr_idx" |
duke@435 | 1409 | Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, |
duke@435 | 1410 | int adr_idx, |
duke@435 | 1411 | bool require_atomic_access) { |
duke@435 | 1412 | assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); |
duke@435 | 1413 | const TypePtr* adr_type = NULL; // debug-mode-only argument |
duke@435 | 1414 | debug_only(adr_type = C->get_adr_type(adr_idx)); |
duke@435 | 1415 | Node* mem = memory(adr_idx); |
duke@435 | 1416 | Node* ld; |
duke@435 | 1417 | if (require_atomic_access && bt == T_LONG) { |
duke@435 | 1418 | ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); |
duke@435 | 1419 | } else { |
coleenp@548 | 1420 | ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); |
duke@435 | 1421 | } |
duke@435 | 1422 | return _gvn.transform(ld); |
duke@435 | 1423 | } |
duke@435 | 1424 | |
duke@435 | 1425 | Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, |
duke@435 | 1426 | int adr_idx, |
duke@435 | 1427 | bool require_atomic_access) { |
duke@435 | 1428 | assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); |
duke@435 | 1429 | const TypePtr* adr_type = NULL; |
duke@435 | 1430 | debug_only(adr_type = C->get_adr_type(adr_idx)); |
duke@435 | 1431 | Node *mem = memory(adr_idx); |
duke@435 | 1432 | Node* st; |
duke@435 | 1433 | if (require_atomic_access && bt == T_LONG) { |
duke@435 | 1434 | st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); |
duke@435 | 1435 | } else { |
coleenp@548 | 1436 | st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); |
duke@435 | 1437 | } |
duke@435 | 1438 | st = _gvn.transform(st); |
duke@435 | 1439 | set_memory(st, adr_idx); |
duke@435 | 1440 | // Back-to-back stores can only remove intermediate store with DU info |
duke@435 | 1441 | // so push on worklist for optimizer. |
duke@435 | 1442 | if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) |
duke@435 | 1443 | record_for_igvn(st); |
duke@435 | 1444 | |
duke@435 | 1445 | return st; |
duke@435 | 1446 | } |
duke@435 | 1447 | |
kvn@1286 | 1448 | |
duke@435 | 1449 | void GraphKit::pre_barrier(Node* ctl, |
duke@435 | 1450 | Node* obj, |
duke@435 | 1451 | Node* adr, |
kvn@1286 | 1452 | uint adr_idx, |
kvn@1286 | 1453 | Node* val, |
never@1260 | 1454 | const TypeOopPtr* val_type, |
duke@435 | 1455 | BasicType bt) { |
duke@435 | 1456 | BarrierSet* bs = Universe::heap()->barrier_set(); |
duke@435 | 1457 | set_control(ctl); |
duke@435 | 1458 | switch (bs->kind()) { |
ysr@777 | 1459 | case BarrierSet::G1SATBCT: |
ysr@777 | 1460 | case BarrierSet::G1SATBCTLogging: |
kvn@1286 | 1461 | g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt); |
ysr@777 | 1462 | break; |
duke@435 | 1463 | |
duke@435 | 1464 | case BarrierSet::CardTableModRef: |
duke@435 | 1465 | case BarrierSet::CardTableExtension: |
duke@435 | 1466 | case BarrierSet::ModRef: |
duke@435 | 1467 | break; |
duke@435 | 1468 | |
duke@435 | 1469 | case BarrierSet::Other: |
duke@435 | 1470 | default : |
duke@435 | 1471 | ShouldNotReachHere(); |
duke@435 | 1472 | |
duke@435 | 1473 | } |
duke@435 | 1474 | } |
duke@435 | 1475 | |
duke@435 | 1476 | void GraphKit::post_barrier(Node* ctl, |
duke@435 | 1477 | Node* store, |
duke@435 | 1478 | Node* obj, |
duke@435 | 1479 | Node* adr, |
kvn@1286 | 1480 | uint adr_idx, |
kvn@1286 | 1481 | Node* val, |
duke@435 | 1482 | BasicType bt, |
duke@435 | 1483 | bool use_precise) { |
duke@435 | 1484 | BarrierSet* bs = Universe::heap()->barrier_set(); |
duke@435 | 1485 | set_control(ctl); |
duke@435 | 1486 | switch (bs->kind()) { |
ysr@777 | 1487 | case BarrierSet::G1SATBCT: |
ysr@777 | 1488 | case BarrierSet::G1SATBCTLogging: |
kvn@1286 | 1489 | g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); |
ysr@777 | 1490 | break; |
duke@435 | 1491 | |
duke@435 | 1492 | case BarrierSet::CardTableModRef: |
duke@435 | 1493 | case BarrierSet::CardTableExtension: |
cfang@1420 | 1494 | write_barrier_post(store, obj, adr, adr_idx, val, use_precise); |
duke@435 | 1495 | break; |
duke@435 | 1496 | |
duke@435 | 1497 | case BarrierSet::ModRef: |
duke@435 | 1498 | break; |
duke@435 | 1499 | |
duke@435 | 1500 | case BarrierSet::Other: |
duke@435 | 1501 | default : |
duke@435 | 1502 | ShouldNotReachHere(); |
duke@435 | 1503 | |
duke@435 | 1504 | } |
duke@435 | 1505 | } |
duke@435 | 1506 | |
kvn@1286 | 1507 | Node* GraphKit::store_oop(Node* ctl, |
kvn@1286 | 1508 | Node* obj, |
kvn@1286 | 1509 | Node* adr, |
kvn@1286 | 1510 | const TypePtr* adr_type, |
kvn@1286 | 1511 | Node* val, |
kvn@1286 | 1512 | const TypeOopPtr* val_type, |
kvn@1286 | 1513 | BasicType bt, |
kvn@1286 | 1514 | bool use_precise) { |
kvn@1286 | 1515 | |
kvn@1286 | 1516 | set_control(ctl); |
kvn@1286 | 1517 | if (stopped()) return top(); // Dead path ? |
kvn@1286 | 1518 | |
kvn@1286 | 1519 | assert(bt == T_OBJECT, "sanity"); |
kvn@1286 | 1520 | assert(val != NULL, "not dead path"); |
duke@435 | 1521 | uint adr_idx = C->get_alias_index(adr_type); |
kvn@1286 | 1522 | assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); |
kvn@1286 | 1523 | |
kvn@1286 | 1524 | pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt); |
kvn@1286 | 1525 | Node* store = store_to_memory(control(), adr, val, bt, adr_idx); |
kvn@1286 | 1526 | post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); |
duke@435 | 1527 | return store; |
duke@435 | 1528 | } |
duke@435 | 1529 | |
kvn@1286 | 1530 | // Could be an array or object we don't know at compile time (unsafe ref.) |
duke@435 | 1531 | Node* GraphKit::store_oop_to_unknown(Node* ctl, |
kvn@1286 | 1532 | Node* obj, // containing obj |
kvn@1286 | 1533 | Node* adr, // actual adress to store val at |
kvn@1286 | 1534 | const TypePtr* adr_type, |
kvn@1286 | 1535 | Node* val, |
kvn@1286 | 1536 | BasicType bt) { |
never@1260 | 1537 | Compile::AliasType* at = C->alias_type(adr_type); |
never@1260 | 1538 | const TypeOopPtr* val_type = NULL; |
never@1260 | 1539 | if (adr_type->isa_instptr()) { |
never@1260 | 1540 | if (at->field() != NULL) { |
never@1260 | 1541 | // known field. This code is a copy of the do_put_xxx logic. |
never@1260 | 1542 | ciField* field = at->field(); |
never@1260 | 1543 | if (!field->type()->is_loaded()) { |
never@1260 | 1544 | val_type = TypeInstPtr::BOTTOM; |
never@1260 | 1545 | } else { |
never@1260 | 1546 | val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); |
never@1260 | 1547 | } |
never@1260 | 1548 | } |
never@1260 | 1549 | } else if (adr_type->isa_aryptr()) { |
never@1262 | 1550 | val_type = adr_type->is_aryptr()->elem()->make_oopptr(); |
never@1260 | 1551 | } |
never@1260 | 1552 | if (val_type == NULL) { |
never@1260 | 1553 | val_type = TypeInstPtr::BOTTOM; |
never@1260 | 1554 | } |
kvn@1286 | 1555 | return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true); |
duke@435 | 1556 | } |
duke@435 | 1557 | |
duke@435 | 1558 | |
duke@435 | 1559 | //-------------------------array_element_address------------------------- |
duke@435 | 1560 | Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, |
duke@435 | 1561 | const TypeInt* sizetype) { |
kvn@464 | 1562 | uint shift = exact_log2(type2aelembytes(elembt)); |
duke@435 | 1563 | uint header = arrayOopDesc::base_offset_in_bytes(elembt); |
duke@435 | 1564 | |
duke@435 | 1565 | // short-circuit a common case (saves lots of confusing waste motion) |
duke@435 | 1566 | jint idx_con = find_int_con(idx, -1); |
duke@435 | 1567 | if (idx_con >= 0) { |
duke@435 | 1568 | intptr_t offset = header + ((intptr_t)idx_con << shift); |
duke@435 | 1569 | return basic_plus_adr(ary, offset); |
duke@435 | 1570 | } |
duke@435 | 1571 | |
duke@435 | 1572 | // must be correct type for alignment purposes |
duke@435 | 1573 | Node* base = basic_plus_adr(ary, header); |
duke@435 | 1574 | #ifdef _LP64 |
duke@435 | 1575 | // The scaled index operand to AddP must be a clean 64-bit value. |
duke@435 | 1576 | // Java allows a 32-bit int to be incremented to a negative |
duke@435 | 1577 | // value, which appears in a 64-bit register as a large |
duke@435 | 1578 | // positive number. Using that large positive number as an |
duke@435 | 1579 | // operand in pointer arithmetic has bad consequences. |
duke@435 | 1580 | // On the other hand, 32-bit overflow is rare, and the possibility |
duke@435 | 1581 | // can often be excluded, if we annotate the ConvI2L node with |
duke@435 | 1582 | // a type assertion that its value is known to be a small positive |
duke@435 | 1583 | // number. (The prior range check has ensured this.) |
duke@435 | 1584 | // This assertion is used by ConvI2LNode::Ideal. |
duke@435 | 1585 | int index_max = max_jint - 1; // array size is max_jint, index is one less |
duke@435 | 1586 | if (sizetype != NULL) index_max = sizetype->_hi - 1; |
duke@435 | 1587 | const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); |
duke@435 | 1588 | idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); |
duke@435 | 1589 | #endif |
duke@435 | 1590 | Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); |
duke@435 | 1591 | return basic_plus_adr(ary, base, scale); |
duke@435 | 1592 | } |
duke@435 | 1593 | |
duke@435 | 1594 | //-------------------------load_array_element------------------------- |
duke@435 | 1595 | Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { |
duke@435 | 1596 | const Type* elemtype = arytype->elem(); |
duke@435 | 1597 | BasicType elembt = elemtype->array_element_basic_type(); |
duke@435 | 1598 | Node* adr = array_element_address(ary, idx, elembt, arytype->size()); |
duke@435 | 1599 | Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); |
duke@435 | 1600 | return ld; |
duke@435 | 1601 | } |
duke@435 | 1602 | |
duke@435 | 1603 | //-------------------------set_arguments_for_java_call------------------------- |
duke@435 | 1604 | // Arguments (pre-popped from the stack) are taken from the JVMS. |
duke@435 | 1605 | void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { |
duke@435 | 1606 | // Add the call arguments: |
duke@435 | 1607 | uint nargs = call->method()->arg_size(); |
duke@435 | 1608 | for (uint i = 0; i < nargs; i++) { |
duke@435 | 1609 | Node* arg = argument(i); |
duke@435 | 1610 | call->init_req(i + TypeFunc::Parms, arg); |
duke@435 | 1611 | } |
duke@435 | 1612 | } |
duke@435 | 1613 | |
duke@435 | 1614 | //---------------------------set_edges_for_java_call--------------------------- |
duke@435 | 1615 | // Connect a newly created call into the current JVMS. |
duke@435 | 1616 | // A return value node (if any) is returned from set_edges_for_java_call. |
never@1515 | 1617 | void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { |
duke@435 | 1618 | |
duke@435 | 1619 | // Add the predefined inputs: |
duke@435 | 1620 | call->init_req( TypeFunc::Control, control() ); |
duke@435 | 1621 | call->init_req( TypeFunc::I_O , i_o() ); |
duke@435 | 1622 | call->init_req( TypeFunc::Memory , reset_memory() ); |
duke@435 | 1623 | call->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 1624 | call->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 1625 | |
duke@435 | 1626 | add_safepoint_edges(call, must_throw); |
duke@435 | 1627 | |
duke@435 | 1628 | Node* xcall = _gvn.transform(call); |
duke@435 | 1629 | |
duke@435 | 1630 | if (xcall == top()) { |
duke@435 | 1631 | set_control(top()); |
duke@435 | 1632 | return; |
duke@435 | 1633 | } |
duke@435 | 1634 | assert(xcall == call, "call identity is stable"); |
duke@435 | 1635 | |
duke@435 | 1636 | // Re-use the current map to produce the result. |
duke@435 | 1637 | |
duke@435 | 1638 | set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); |
never@1515 | 1639 | set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj))); |
never@1515 | 1640 | set_all_memory_call(xcall, separate_io_proj); |
duke@435 | 1641 | |
duke@435 | 1642 | //return xcall; // no need, caller already has it |
duke@435 | 1643 | } |
duke@435 | 1644 | |
never@1515 | 1645 | Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { |
duke@435 | 1646 | if (stopped()) return top(); // maybe the call folded up? |
duke@435 | 1647 | |
duke@435 | 1648 | // Capture the return value, if any. |
duke@435 | 1649 | Node* ret; |
duke@435 | 1650 | if (call->method() == NULL || |
duke@435 | 1651 | call->method()->return_type()->basic_type() == T_VOID) |
duke@435 | 1652 | ret = top(); |
duke@435 | 1653 | else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); |
duke@435 | 1654 | |
duke@435 | 1655 | // Note: Since any out-of-line call can produce an exception, |
duke@435 | 1656 | // we always insert an I_O projection from the call into the result. |
duke@435 | 1657 | |
never@1515 | 1658 | make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); |
never@1515 | 1659 | |
never@1515 | 1660 | if (separate_io_proj) { |
never@1515 | 1661 | // The caller requested separate projections be used by the fall |
never@1515 | 1662 | // through and exceptional paths, so replace the projections for |
never@1515 | 1663 | // the fall through path. |
never@1515 | 1664 | set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) )); |
never@1515 | 1665 | set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) )); |
never@1515 | 1666 | } |
duke@435 | 1667 | return ret; |
duke@435 | 1668 | } |
duke@435 | 1669 | |
duke@435 | 1670 | //--------------------set_predefined_input_for_runtime_call-------------------- |
duke@435 | 1671 | // Reading and setting the memory state is way conservative here. |
duke@435 | 1672 | // The real problem is that I am not doing real Type analysis on memory, |
duke@435 | 1673 | // so I cannot distinguish card mark stores from other stores. Across a GC |
duke@435 | 1674 | // point the Store Barrier and the card mark memory has to agree. I cannot |
duke@435 | 1675 | // have a card mark store and its barrier split across the GC point from |
duke@435 | 1676 | // either above or below. Here I get that to happen by reading ALL of memory. |
duke@435 | 1677 | // A better answer would be to separate out card marks from other memory. |
duke@435 | 1678 | // For now, return the input memory state, so that it can be reused |
duke@435 | 1679 | // after the call, if this call has restricted memory effects. |
duke@435 | 1680 | Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { |
duke@435 | 1681 | // Set fixed predefined input arguments |
duke@435 | 1682 | Node* memory = reset_memory(); |
duke@435 | 1683 | call->init_req( TypeFunc::Control, control() ); |
duke@435 | 1684 | call->init_req( TypeFunc::I_O, top() ); // does no i/o |
duke@435 | 1685 | call->init_req( TypeFunc::Memory, memory ); // may gc ptrs |
duke@435 | 1686 | call->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 1687 | call->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 1688 | return memory; |
duke@435 | 1689 | } |
duke@435 | 1690 | |
duke@435 | 1691 | //-------------------set_predefined_output_for_runtime_call-------------------- |
duke@435 | 1692 | // Set control and memory (not i_o) from the call. |
duke@435 | 1693 | // If keep_mem is not NULL, use it for the output state, |
duke@435 | 1694 | // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. |
duke@435 | 1695 | // If hook_mem is NULL, this call produces no memory effects at all. |
duke@435 | 1696 | // If hook_mem is a Java-visible memory slice (such as arraycopy operands), |
duke@435 | 1697 | // then only that memory slice is taken from the call. |
duke@435 | 1698 | // In the last case, we must put an appropriate memory barrier before |
duke@435 | 1699 | // the call, so as to create the correct anti-dependencies on loads |
duke@435 | 1700 | // preceding the call. |
duke@435 | 1701 | void GraphKit::set_predefined_output_for_runtime_call(Node* call, |
duke@435 | 1702 | Node* keep_mem, |
duke@435 | 1703 | const TypePtr* hook_mem) { |
duke@435 | 1704 | // no i/o |
duke@435 | 1705 | set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); |
duke@435 | 1706 | if (keep_mem) { |
duke@435 | 1707 | // First clone the existing memory state |
duke@435 | 1708 | set_all_memory(keep_mem); |
duke@435 | 1709 | if (hook_mem != NULL) { |
duke@435 | 1710 | // Make memory for the call |
duke@435 | 1711 | Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); |
duke@435 | 1712 | // Set the RawPtr memory state only. This covers all the heap top/GC stuff |
duke@435 | 1713 | // We also use hook_mem to extract specific effects from arraycopy stubs. |
duke@435 | 1714 | set_memory(mem, hook_mem); |
duke@435 | 1715 | } |
duke@435 | 1716 | // ...else the call has NO memory effects. |
duke@435 | 1717 | |
duke@435 | 1718 | // Make sure the call advertises its memory effects precisely. |
duke@435 | 1719 | // This lets us build accurate anti-dependences in gcm.cpp. |
duke@435 | 1720 | assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), |
duke@435 | 1721 | "call node must be constructed correctly"); |
duke@435 | 1722 | } else { |
duke@435 | 1723 | assert(hook_mem == NULL, ""); |
duke@435 | 1724 | // This is not a "slow path" call; all memory comes from the call. |
duke@435 | 1725 | set_all_memory_call(call); |
duke@435 | 1726 | } |
duke@435 | 1727 | } |
duke@435 | 1728 | |
never@1515 | 1729 | |
never@1515 | 1730 | // Replace the call with the current state of the kit. |
never@1515 | 1731 | void GraphKit::replace_call(CallNode* call, Node* result) { |
never@1515 | 1732 | JVMState* ejvms = NULL; |
never@1515 | 1733 | if (has_exceptions()) { |
never@1515 | 1734 | ejvms = transfer_exceptions_into_jvms(); |
never@1515 | 1735 | } |
never@1515 | 1736 | |
never@1515 | 1737 | SafePointNode* final_state = stop(); |
never@1515 | 1738 | |
never@1515 | 1739 | // Find all the needed outputs of this call |
never@1515 | 1740 | CallProjections callprojs; |
never@1515 | 1741 | call->extract_projections(&callprojs, true); |
never@1515 | 1742 | |
never@1515 | 1743 | // Replace all the old call edges with the edges from the inlining result |
never@1515 | 1744 | C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control)); |
never@1515 | 1745 | C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory)); |
never@1515 | 1746 | C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O)); |
never@1515 | 1747 | |
never@1515 | 1748 | // Replace the result with the new result if it exists and is used |
never@1515 | 1749 | if (callprojs.resproj != NULL && result != NULL) { |
never@1515 | 1750 | C->gvn_replace_by(callprojs.resproj, result); |
never@1515 | 1751 | } |
never@1515 | 1752 | |
never@1515 | 1753 | if (ejvms == NULL) { |
never@1515 | 1754 | // No exception edges to simply kill off those paths |
never@1515 | 1755 | C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); |
never@1515 | 1756 | C->gvn_replace_by(callprojs.catchall_memproj, C->top()); |
never@1515 | 1757 | C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); |
never@1534 | 1758 | |
never@1534 | 1759 | // Replace the old exception object with top |
never@1534 | 1760 | if (callprojs.exobj != NULL) { |
never@1534 | 1761 | C->gvn_replace_by(callprojs.exobj, C->top()); |
never@1534 | 1762 | } |
never@1515 | 1763 | } else { |
never@1515 | 1764 | GraphKit ekit(ejvms); |
never@1515 | 1765 | |
never@1515 | 1766 | // Load my combined exception state into the kit, with all phis transformed: |
never@1515 | 1767 | SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); |
never@1515 | 1768 | |
never@1515 | 1769 | Node* ex_oop = ekit.use_exception_state(ex_map); |
never@1515 | 1770 | |
never@1515 | 1771 | C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); |
never@1515 | 1772 | C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); |
never@1515 | 1773 | C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); |
never@1515 | 1774 | |
never@1515 | 1775 | // Replace the old exception object with the newly created one |
never@1515 | 1776 | if (callprojs.exobj != NULL) { |
never@1515 | 1777 | C->gvn_replace_by(callprojs.exobj, ex_oop); |
never@1515 | 1778 | } |
never@1515 | 1779 | } |
never@1515 | 1780 | |
never@1515 | 1781 | // Disconnect the call from the graph |
never@1515 | 1782 | call->disconnect_inputs(NULL); |
never@1515 | 1783 | C->gvn_replace_by(call, C->top()); |
never@1515 | 1784 | } |
never@1515 | 1785 | |
never@1515 | 1786 | |
duke@435 | 1787 | //------------------------------increment_counter------------------------------ |
duke@435 | 1788 | // for statistics: increment a VM counter by 1 |
duke@435 | 1789 | |
duke@435 | 1790 | void GraphKit::increment_counter(address counter_addr) { |
duke@435 | 1791 | Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); |
duke@435 | 1792 | increment_counter(adr1); |
duke@435 | 1793 | } |
duke@435 | 1794 | |
duke@435 | 1795 | void GraphKit::increment_counter(Node* counter_addr) { |
duke@435 | 1796 | int adr_type = Compile::AliasIdxRaw; |
duke@435 | 1797 | Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); |
duke@435 | 1798 | Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); |
duke@435 | 1799 | store_to_memory( NULL, counter_addr, incr, T_INT, adr_type ); |
duke@435 | 1800 | } |
duke@435 | 1801 | |
duke@435 | 1802 | |
duke@435 | 1803 | //------------------------------uncommon_trap---------------------------------- |
duke@435 | 1804 | // Bail out to the interpreter in mid-method. Implemented by calling the |
duke@435 | 1805 | // uncommon_trap blob. This helper function inserts a runtime call with the |
duke@435 | 1806 | // right debug info. |
duke@435 | 1807 | void GraphKit::uncommon_trap(int trap_request, |
duke@435 | 1808 | ciKlass* klass, const char* comment, |
duke@435 | 1809 | bool must_throw, |
duke@435 | 1810 | bool keep_exact_action) { |
duke@435 | 1811 | if (failing()) stop(); |
duke@435 | 1812 | if (stopped()) return; // trap reachable? |
duke@435 | 1813 | |
duke@435 | 1814 | // Note: If ProfileTraps is true, and if a deopt. actually |
duke@435 | 1815 | // occurs here, the runtime will make sure an MDO exists. There is |
duke@435 | 1816 | // no need to call method()->build_method_data() at this point. |
duke@435 | 1817 | |
duke@435 | 1818 | #ifdef ASSERT |
duke@435 | 1819 | if (!must_throw) { |
duke@435 | 1820 | // Make sure the stack has at least enough depth to execute |
duke@435 | 1821 | // the current bytecode. |
duke@435 | 1822 | int inputs, ignore; |
duke@435 | 1823 | if (compute_stack_effects(inputs, ignore)) { |
duke@435 | 1824 | assert(sp() >= inputs, "must have enough JVMS stack to execute"); |
duke@435 | 1825 | // It is a frequent error in library_call.cpp to issue an |
duke@435 | 1826 | // uncommon trap with the _sp value already popped. |
duke@435 | 1827 | } |
duke@435 | 1828 | } |
duke@435 | 1829 | #endif |
duke@435 | 1830 | |
duke@435 | 1831 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); |
duke@435 | 1832 | Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); |
duke@435 | 1833 | |
duke@435 | 1834 | switch (action) { |
duke@435 | 1835 | case Deoptimization::Action_maybe_recompile: |
duke@435 | 1836 | case Deoptimization::Action_reinterpret: |
duke@435 | 1837 | // Temporary fix for 6529811 to allow virtual calls to be sure they |
duke@435 | 1838 | // get the chance to go from mono->bi->mega |
duke@435 | 1839 | if (!keep_exact_action && |
duke@435 | 1840 | Deoptimization::trap_request_index(trap_request) < 0 && |
duke@435 | 1841 | too_many_recompiles(reason)) { |
duke@435 | 1842 | // This BCI is causing too many recompilations. |
duke@435 | 1843 | action = Deoptimization::Action_none; |
duke@435 | 1844 | trap_request = Deoptimization::make_trap_request(reason, action); |
duke@435 | 1845 | } else { |
duke@435 | 1846 | C->set_trap_can_recompile(true); |
duke@435 | 1847 | } |
duke@435 | 1848 | break; |
duke@435 | 1849 | case Deoptimization::Action_make_not_entrant: |
duke@435 | 1850 | C->set_trap_can_recompile(true); |
duke@435 | 1851 | break; |
duke@435 | 1852 | #ifdef ASSERT |
duke@435 | 1853 | case Deoptimization::Action_none: |
duke@435 | 1854 | case Deoptimization::Action_make_not_compilable: |
duke@435 | 1855 | break; |
duke@435 | 1856 | default: |
duke@435 | 1857 | assert(false, "bad action"); |
duke@435 | 1858 | #endif |
duke@435 | 1859 | } |
duke@435 | 1860 | |
duke@435 | 1861 | if (TraceOptoParse) { |
duke@435 | 1862 | char buf[100]; |
duke@435 | 1863 | tty->print_cr("Uncommon trap %s at bci:%d", |
duke@435 | 1864 | Deoptimization::format_trap_request(buf, sizeof(buf), |
duke@435 | 1865 | trap_request), bci()); |
duke@435 | 1866 | } |
duke@435 | 1867 | |
duke@435 | 1868 | CompileLog* log = C->log(); |
duke@435 | 1869 | if (log != NULL) { |
duke@435 | 1870 | int kid = (klass == NULL)? -1: log->identify(klass); |
duke@435 | 1871 | log->begin_elem("uncommon_trap bci='%d'", bci()); |
duke@435 | 1872 | char buf[100]; |
duke@435 | 1873 | log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), |
duke@435 | 1874 | trap_request)); |
duke@435 | 1875 | if (kid >= 0) log->print(" klass='%d'", kid); |
duke@435 | 1876 | if (comment != NULL) log->print(" comment='%s'", comment); |
duke@435 | 1877 | log->end_elem(); |
duke@435 | 1878 | } |
duke@435 | 1879 | |
duke@435 | 1880 | // Make sure any guarding test views this path as very unlikely |
duke@435 | 1881 | Node *i0 = control()->in(0); |
duke@435 | 1882 | if (i0 != NULL && i0->is_If()) { // Found a guarding if test? |
duke@435 | 1883 | IfNode *iff = i0->as_If(); |
duke@435 | 1884 | float f = iff->_prob; // Get prob |
duke@435 | 1885 | if (control()->Opcode() == Op_IfTrue) { |
duke@435 | 1886 | if (f > PROB_UNLIKELY_MAG(4)) |
duke@435 | 1887 | iff->_prob = PROB_MIN; |
duke@435 | 1888 | } else { |
duke@435 | 1889 | if (f < PROB_LIKELY_MAG(4)) |
duke@435 | 1890 | iff->_prob = PROB_MAX; |
duke@435 | 1891 | } |
duke@435 | 1892 | } |
duke@435 | 1893 | |
duke@435 | 1894 | // Clear out dead values from the debug info. |
duke@435 | 1895 | kill_dead_locals(); |
duke@435 | 1896 | |
duke@435 | 1897 | // Now insert the uncommon trap subroutine call |
duke@435 | 1898 | address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin(); |
duke@435 | 1899 | const TypePtr* no_memory_effects = NULL; |
duke@435 | 1900 | // Pass the index of the class to be loaded |
duke@435 | 1901 | Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | |
duke@435 | 1902 | (must_throw ? RC_MUST_THROW : 0), |
duke@435 | 1903 | OptoRuntime::uncommon_trap_Type(), |
duke@435 | 1904 | call_addr, "uncommon_trap", no_memory_effects, |
duke@435 | 1905 | intcon(trap_request)); |
duke@435 | 1906 | assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, |
duke@435 | 1907 | "must extract request correctly from the graph"); |
duke@435 | 1908 | assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); |
duke@435 | 1909 | |
duke@435 | 1910 | call->set_req(TypeFunc::ReturnAdr, returnadr()); |
duke@435 | 1911 | // The debug info is the only real input to this call. |
duke@435 | 1912 | |
duke@435 | 1913 | // Halt-and-catch fire here. The above call should never return! |
duke@435 | 1914 | HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); |
duke@435 | 1915 | _gvn.set_type_bottom(halt); |
duke@435 | 1916 | root()->add_req(halt); |
duke@435 | 1917 | |
duke@435 | 1918 | stop_and_kill_map(); |
duke@435 | 1919 | } |
duke@435 | 1920 | |
duke@435 | 1921 | |
duke@435 | 1922 | //--------------------------just_allocated_object------------------------------ |
duke@435 | 1923 | // Report the object that was just allocated. |
duke@435 | 1924 | // It must be the case that there are no intervening safepoints. |
duke@435 | 1925 | // We use this to determine if an object is so "fresh" that |
duke@435 | 1926 | // it does not require card marks. |
duke@435 | 1927 | Node* GraphKit::just_allocated_object(Node* current_control) { |
duke@435 | 1928 | if (C->recent_alloc_ctl() == current_control) |
duke@435 | 1929 | return C->recent_alloc_obj(); |
duke@435 | 1930 | return NULL; |
duke@435 | 1931 | } |
duke@435 | 1932 | |
duke@435 | 1933 | |
duke@435 | 1934 | void GraphKit::round_double_arguments(ciMethod* dest_method) { |
duke@435 | 1935 | // (Note: TypeFunc::make has a cache that makes this fast.) |
duke@435 | 1936 | const TypeFunc* tf = TypeFunc::make(dest_method); |
duke@435 | 1937 | int nargs = tf->_domain->_cnt - TypeFunc::Parms; |
duke@435 | 1938 | for (int j = 0; j < nargs; j++) { |
duke@435 | 1939 | const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); |
duke@435 | 1940 | if( targ->basic_type() == T_DOUBLE ) { |
duke@435 | 1941 | // If any parameters are doubles, they must be rounded before |
duke@435 | 1942 | // the call, dstore_rounding does gvn.transform |
duke@435 | 1943 | Node *arg = argument(j); |
duke@435 | 1944 | arg = dstore_rounding(arg); |
duke@435 | 1945 | set_argument(j, arg); |
duke@435 | 1946 | } |
duke@435 | 1947 | } |
duke@435 | 1948 | } |
duke@435 | 1949 | |
duke@435 | 1950 | void GraphKit::round_double_result(ciMethod* dest_method) { |
duke@435 | 1951 | // A non-strict method may return a double value which has an extended |
duke@435 | 1952 | // exponent, but this must not be visible in a caller which is 'strict' |
duke@435 | 1953 | // If a strict caller invokes a non-strict callee, round a double result |
duke@435 | 1954 | |
duke@435 | 1955 | BasicType result_type = dest_method->return_type()->basic_type(); |
duke@435 | 1956 | assert( method() != NULL, "must have caller context"); |
duke@435 | 1957 | if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { |
duke@435 | 1958 | // Destination method's return value is on top of stack |
duke@435 | 1959 | // dstore_rounding() does gvn.transform |
duke@435 | 1960 | Node *result = pop_pair(); |
duke@435 | 1961 | result = dstore_rounding(result); |
duke@435 | 1962 | push_pair(result); |
duke@435 | 1963 | } |
duke@435 | 1964 | } |
duke@435 | 1965 | |
duke@435 | 1966 | // rounding for strict float precision conformance |
duke@435 | 1967 | Node* GraphKit::precision_rounding(Node* n) { |
duke@435 | 1968 | return UseStrictFP && _method->flags().is_strict() |
duke@435 | 1969 | && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding |
duke@435 | 1970 | ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) |
duke@435 | 1971 | : n; |
duke@435 | 1972 | } |
duke@435 | 1973 | |
duke@435 | 1974 | // rounding for strict double precision conformance |
duke@435 | 1975 | Node* GraphKit::dprecision_rounding(Node *n) { |
duke@435 | 1976 | return UseStrictFP && _method->flags().is_strict() |
duke@435 | 1977 | && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding |
duke@435 | 1978 | ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) |
duke@435 | 1979 | : n; |
duke@435 | 1980 | } |
duke@435 | 1981 | |
duke@435 | 1982 | // rounding for non-strict double stores |
duke@435 | 1983 | Node* GraphKit::dstore_rounding(Node* n) { |
duke@435 | 1984 | return Matcher::strict_fp_requires_explicit_rounding |
duke@435 | 1985 | && UseSSE <= 1 |
duke@435 | 1986 | ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) |
duke@435 | 1987 | : n; |
duke@435 | 1988 | } |
duke@435 | 1989 | |
duke@435 | 1990 | //============================================================================= |
duke@435 | 1991 | // Generate a fast path/slow path idiom. Graph looks like: |
duke@435 | 1992 | // [foo] indicates that 'foo' is a parameter |
duke@435 | 1993 | // |
duke@435 | 1994 | // [in] NULL |
duke@435 | 1995 | // \ / |
duke@435 | 1996 | // CmpP |
duke@435 | 1997 | // Bool ne |
duke@435 | 1998 | // If |
duke@435 | 1999 | // / \ |
duke@435 | 2000 | // True False-<2> |
duke@435 | 2001 | // / | |
duke@435 | 2002 | // / cast_not_null |
duke@435 | 2003 | // Load | | ^ |
duke@435 | 2004 | // [fast_test] | | |
duke@435 | 2005 | // gvn to opt_test | | |
duke@435 | 2006 | // / \ | <1> |
duke@435 | 2007 | // True False | |
duke@435 | 2008 | // | \\ | |
duke@435 | 2009 | // [slow_call] \[fast_result] |
duke@435 | 2010 | // Ctl Val \ \ |
duke@435 | 2011 | // | \ \ |
duke@435 | 2012 | // Catch <1> \ \ |
duke@435 | 2013 | // / \ ^ \ \ |
duke@435 | 2014 | // Ex No_Ex | \ \ |
duke@435 | 2015 | // | \ \ | \ <2> \ |
duke@435 | 2016 | // ... \ [slow_res] | | \ [null_result] |
duke@435 | 2017 | // \ \--+--+--- | | |
duke@435 | 2018 | // \ | / \ | / |
duke@435 | 2019 | // --------Region Phi |
duke@435 | 2020 | // |
duke@435 | 2021 | //============================================================================= |
duke@435 | 2022 | // Code is structured as a series of driver functions all called 'do_XXX' that |
duke@435 | 2023 | // call a set of helper functions. Helper functions first, then drivers. |
duke@435 | 2024 | |
duke@435 | 2025 | //------------------------------null_check_oop--------------------------------- |
duke@435 | 2026 | // Null check oop. Set null-path control into Region in slot 3. |
duke@435 | 2027 | // Make a cast-not-nullness use the other not-null control. Return cast. |
duke@435 | 2028 | Node* GraphKit::null_check_oop(Node* value, Node* *null_control, |
duke@435 | 2029 | bool never_see_null) { |
duke@435 | 2030 | // Initial NULL check taken path |
duke@435 | 2031 | (*null_control) = top(); |
duke@435 | 2032 | Node* cast = null_check_common(value, T_OBJECT, false, null_control); |
duke@435 | 2033 | |
duke@435 | 2034 | // Generate uncommon_trap: |
duke@435 | 2035 | if (never_see_null && (*null_control) != top()) { |
duke@435 | 2036 | // If we see an unexpected null at a check-cast we record it and force a |
duke@435 | 2037 | // recompile; the offending check-cast will be compiled to handle NULLs. |
duke@435 | 2038 | // If we see more than one offending BCI, then all checkcasts in the |
duke@435 | 2039 | // method will be compiled to handle NULLs. |
duke@435 | 2040 | PreserveJVMState pjvms(this); |
duke@435 | 2041 | set_control(*null_control); |
kvn@767 | 2042 | replace_in_map(value, null()); |
duke@435 | 2043 | uncommon_trap(Deoptimization::Reason_null_check, |
duke@435 | 2044 | Deoptimization::Action_make_not_entrant); |
duke@435 | 2045 | (*null_control) = top(); // NULL path is dead |
duke@435 | 2046 | } |
duke@435 | 2047 | |
duke@435 | 2048 | // Cast away null-ness on the result |
duke@435 | 2049 | return cast; |
duke@435 | 2050 | } |
duke@435 | 2051 | |
duke@435 | 2052 | //------------------------------opt_iff---------------------------------------- |
duke@435 | 2053 | // Optimize the fast-check IfNode. Set the fast-path region slot 2. |
duke@435 | 2054 | // Return slow-path control. |
duke@435 | 2055 | Node* GraphKit::opt_iff(Node* region, Node* iff) { |
duke@435 | 2056 | IfNode *opt_iff = _gvn.transform(iff)->as_If(); |
duke@435 | 2057 | |
duke@435 | 2058 | // Fast path taken; set region slot 2 |
duke@435 | 2059 | Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); |
duke@435 | 2060 | region->init_req(2,fast_taken); // Capture fast-control |
duke@435 | 2061 | |
duke@435 | 2062 | // Fast path not-taken, i.e. slow path |
duke@435 | 2063 | Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); |
duke@435 | 2064 | return slow_taken; |
duke@435 | 2065 | } |
duke@435 | 2066 | |
duke@435 | 2067 | //-----------------------------make_runtime_call------------------------------- |
duke@435 | 2068 | Node* GraphKit::make_runtime_call(int flags, |
duke@435 | 2069 | const TypeFunc* call_type, address call_addr, |
duke@435 | 2070 | const char* call_name, |
duke@435 | 2071 | const TypePtr* adr_type, |
duke@435 | 2072 | // The following parms are all optional. |
duke@435 | 2073 | // The first NULL ends the list. |
duke@435 | 2074 | Node* parm0, Node* parm1, |
duke@435 | 2075 | Node* parm2, Node* parm3, |
duke@435 | 2076 | Node* parm4, Node* parm5, |
duke@435 | 2077 | Node* parm6, Node* parm7) { |
duke@435 | 2078 | // Slow-path call |
duke@435 | 2079 | int size = call_type->domain()->cnt(); |
duke@435 | 2080 | bool is_leaf = !(flags & RC_NO_LEAF); |
duke@435 | 2081 | bool has_io = (!is_leaf && !(flags & RC_NO_IO)); |
duke@435 | 2082 | if (call_name == NULL) { |
duke@435 | 2083 | assert(!is_leaf, "must supply name for leaf"); |
duke@435 | 2084 | call_name = OptoRuntime::stub_name(call_addr); |
duke@435 | 2085 | } |
duke@435 | 2086 | CallNode* call; |
duke@435 | 2087 | if (!is_leaf) { |
duke@435 | 2088 | call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, |
duke@435 | 2089 | bci(), adr_type); |
duke@435 | 2090 | } else if (flags & RC_NO_FP) { |
duke@435 | 2091 | call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); |
duke@435 | 2092 | } else { |
duke@435 | 2093 | call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); |
duke@435 | 2094 | } |
duke@435 | 2095 | |
duke@435 | 2096 | // The following is similar to set_edges_for_java_call, |
duke@435 | 2097 | // except that the memory effects of the call are restricted to AliasIdxRaw. |
duke@435 | 2098 | |
duke@435 | 2099 | // Slow path call has no side-effects, uses few values |
duke@435 | 2100 | bool wide_in = !(flags & RC_NARROW_MEM); |
duke@435 | 2101 | bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); |
duke@435 | 2102 | |
duke@435 | 2103 | Node* prev_mem = NULL; |
duke@435 | 2104 | if (wide_in) { |
duke@435 | 2105 | prev_mem = set_predefined_input_for_runtime_call(call); |
duke@435 | 2106 | } else { |
duke@435 | 2107 | assert(!wide_out, "narrow in => narrow out"); |
duke@435 | 2108 | Node* narrow_mem = memory(adr_type); |
duke@435 | 2109 | prev_mem = reset_memory(); |
duke@435 | 2110 | map()->set_memory(narrow_mem); |
duke@435 | 2111 | set_predefined_input_for_runtime_call(call); |
duke@435 | 2112 | } |
duke@435 | 2113 | |
duke@435 | 2114 | // Hook each parm in order. Stop looking at the first NULL. |
duke@435 | 2115 | if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); |
duke@435 | 2116 | if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); |
duke@435 | 2117 | if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); |
duke@435 | 2118 | if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); |
duke@435 | 2119 | if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); |
duke@435 | 2120 | if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); |
duke@435 | 2121 | if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); |
duke@435 | 2122 | if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); |
duke@435 | 2123 | /* close each nested if ===> */ } } } } } } } } |
duke@435 | 2124 | assert(call->in(call->req()-1) != NULL, "must initialize all parms"); |
duke@435 | 2125 | |
duke@435 | 2126 | if (!is_leaf) { |
duke@435 | 2127 | // Non-leaves can block and take safepoints: |
duke@435 | 2128 | add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); |
duke@435 | 2129 | } |
duke@435 | 2130 | // Non-leaves can throw exceptions: |
duke@435 | 2131 | if (has_io) { |
duke@435 | 2132 | call->set_req(TypeFunc::I_O, i_o()); |
duke@435 | 2133 | } |
duke@435 | 2134 | |
duke@435 | 2135 | if (flags & RC_UNCOMMON) { |
duke@435 | 2136 | // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. |
duke@435 | 2137 | // (An "if" probability corresponds roughly to an unconditional count. |
duke@435 | 2138 | // Sort of.) |
duke@435 | 2139 | call->set_cnt(PROB_UNLIKELY_MAG(4)); |
duke@435 | 2140 | } |
duke@435 | 2141 | |
duke@435 | 2142 | Node* c = _gvn.transform(call); |
duke@435 | 2143 | assert(c == call, "cannot disappear"); |
duke@435 | 2144 | |
duke@435 | 2145 | if (wide_out) { |
duke@435 | 2146 | // Slow path call has full side-effects. |
duke@435 | 2147 | set_predefined_output_for_runtime_call(call); |
duke@435 | 2148 | } else { |
duke@435 | 2149 | // Slow path call has few side-effects, and/or sets few values. |
duke@435 | 2150 | set_predefined_output_for_runtime_call(call, prev_mem, adr_type); |
duke@435 | 2151 | } |
duke@435 | 2152 | |
duke@435 | 2153 | if (has_io) { |
duke@435 | 2154 | set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); |
duke@435 | 2155 | } |
duke@435 | 2156 | return call; |
duke@435 | 2157 | |
duke@435 | 2158 | } |
duke@435 | 2159 | |
duke@435 | 2160 | //------------------------------merge_memory----------------------------------- |
duke@435 | 2161 | // Merge memory from one path into the current memory state. |
duke@435 | 2162 | void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { |
duke@435 | 2163 | for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { |
duke@435 | 2164 | Node* old_slice = mms.force_memory(); |
duke@435 | 2165 | Node* new_slice = mms.memory2(); |
duke@435 | 2166 | if (old_slice != new_slice) { |
duke@435 | 2167 | PhiNode* phi; |
duke@435 | 2168 | if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { |
duke@435 | 2169 | phi = new_slice->as_Phi(); |
duke@435 | 2170 | #ifdef ASSERT |
duke@435 | 2171 | if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) |
duke@435 | 2172 | old_slice = old_slice->in(new_path); |
duke@435 | 2173 | // Caller is responsible for ensuring that any pre-existing |
duke@435 | 2174 | // phis are already aware of old memory. |
duke@435 | 2175 | int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path |
duke@435 | 2176 | assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); |
duke@435 | 2177 | #endif |
duke@435 | 2178 | mms.set_memory(phi); |
duke@435 | 2179 | } else { |
duke@435 | 2180 | phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); |
duke@435 | 2181 | _gvn.set_type(phi, Type::MEMORY); |
duke@435 | 2182 | phi->set_req(new_path, new_slice); |
duke@435 | 2183 | mms.set_memory(_gvn.transform(phi)); // assume it is complete |
duke@435 | 2184 | } |
duke@435 | 2185 | } |
duke@435 | 2186 | } |
duke@435 | 2187 | } |
duke@435 | 2188 | |
duke@435 | 2189 | //------------------------------make_slow_call_ex------------------------------ |
duke@435 | 2190 | // Make the exception handler hookups for the slow call |
duke@435 | 2191 | void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { |
duke@435 | 2192 | if (stopped()) return; |
duke@435 | 2193 | |
duke@435 | 2194 | // Make a catch node with just two handlers: fall-through and catch-all |
duke@435 | 2195 | Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); |
duke@435 | 2196 | Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); |
duke@435 | 2197 | Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); |
duke@435 | 2198 | Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); |
duke@435 | 2199 | |
duke@435 | 2200 | { PreserveJVMState pjvms(this); |
duke@435 | 2201 | set_control(excp); |
duke@435 | 2202 | set_i_o(i_o); |
duke@435 | 2203 | |
duke@435 | 2204 | if (excp != top()) { |
duke@435 | 2205 | // Create an exception state also. |
duke@435 | 2206 | // Use an exact type if the caller has specified a specific exception. |
duke@435 | 2207 | const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); |
duke@435 | 2208 | Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); |
duke@435 | 2209 | add_exception_state(make_exception_state(_gvn.transform(ex_oop))); |
duke@435 | 2210 | } |
duke@435 | 2211 | } |
duke@435 | 2212 | |
duke@435 | 2213 | // Get the no-exception control from the CatchNode. |
duke@435 | 2214 | set_control(norm); |
duke@435 | 2215 | } |
duke@435 | 2216 | |
duke@435 | 2217 | |
duke@435 | 2218 | //-------------------------------gen_subtype_check----------------------------- |
duke@435 | 2219 | // Generate a subtyping check. Takes as input the subtype and supertype. |
duke@435 | 2220 | // Returns 2 values: sets the default control() to the true path and returns |
duke@435 | 2221 | // the false path. Only reads invariant memory; sets no (visible) memory. |
duke@435 | 2222 | // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding |
duke@435 | 2223 | // but that's not exposed to the optimizer. This call also doesn't take in an |
duke@435 | 2224 | // Object; if you wish to check an Object you need to load the Object's class |
duke@435 | 2225 | // prior to coming here. |
duke@435 | 2226 | Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { |
duke@435 | 2227 | // Fast check for identical types, perhaps identical constants. |
duke@435 | 2228 | // The types can even be identical non-constants, in cases |
duke@435 | 2229 | // involving Array.newInstance, Object.clone, etc. |
duke@435 | 2230 | if (subklass == superklass) |
duke@435 | 2231 | return top(); // false path is dead; no test needed. |
duke@435 | 2232 | |
duke@435 | 2233 | if (_gvn.type(superklass)->singleton()) { |
duke@435 | 2234 | ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); |
duke@435 | 2235 | ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); |
duke@435 | 2236 | |
duke@435 | 2237 | // In the common case of an exact superklass, try to fold up the |
duke@435 | 2238 | // test before generating code. You may ask, why not just generate |
duke@435 | 2239 | // the code and then let it fold up? The answer is that the generated |
duke@435 | 2240 | // code will necessarily include null checks, which do not always |
duke@435 | 2241 | // completely fold away. If they are also needless, then they turn |
duke@435 | 2242 | // into a performance loss. Example: |
duke@435 | 2243 | // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; |
duke@435 | 2244 | // Here, the type of 'fa' is often exact, so the store check |
duke@435 | 2245 | // of fa[1]=x will fold up, without testing the nullness of x. |
duke@435 | 2246 | switch (static_subtype_check(superk, subk)) { |
duke@435 | 2247 | case SSC_always_false: |
duke@435 | 2248 | { |
duke@435 | 2249 | Node* always_fail = control(); |
duke@435 | 2250 | set_control(top()); |
duke@435 | 2251 | return always_fail; |
duke@435 | 2252 | } |
duke@435 | 2253 | case SSC_always_true: |
duke@435 | 2254 | return top(); |
duke@435 | 2255 | case SSC_easy_test: |
duke@435 | 2256 | { |
duke@435 | 2257 | // Just do a direct pointer compare and be done. |
duke@435 | 2258 | Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); |
duke@435 | 2259 | Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); |
duke@435 | 2260 | IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); |
duke@435 | 2261 | set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); |
duke@435 | 2262 | return _gvn.transform( new(C, 1) IfFalseNode(iff) ); |
duke@435 | 2263 | } |
duke@435 | 2264 | case SSC_full_test: |
duke@435 | 2265 | break; |
duke@435 | 2266 | default: |
duke@435 | 2267 | ShouldNotReachHere(); |
duke@435 | 2268 | } |
duke@435 | 2269 | } |
duke@435 | 2270 | |
duke@435 | 2271 | // %%% Possible further optimization: Even if the superklass is not exact, |
duke@435 | 2272 | // if the subklass is the unique subtype of the superklass, the check |
duke@435 | 2273 | // will always succeed. We could leave a dependency behind to ensure this. |
duke@435 | 2274 | |
duke@435 | 2275 | // First load the super-klass's check-offset |
duke@435 | 2276 | Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() ); |
duke@435 | 2277 | Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); |
duke@435 | 2278 | int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes(); |
duke@435 | 2279 | bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); |
duke@435 | 2280 | |
duke@435 | 2281 | // Load from the sub-klass's super-class display list, or a 1-word cache of |
duke@435 | 2282 | // the secondary superclass list, or a failing value with a sentinel offset |
duke@435 | 2283 | // if the super-klass is an interface or exceptionally deep in the Java |
duke@435 | 2284 | // hierarchy and we have to scan the secondary superclass list the hard way. |
duke@435 | 2285 | // Worst-case type is a little odd: NULL is allowed as a result (usually |
duke@435 | 2286 | // klass loads can never produce a NULL). |
duke@435 | 2287 | Node *chk_off_X = ConvI2X(chk_off); |
duke@435 | 2288 | Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); |
duke@435 | 2289 | // For some types like interfaces the following loadKlass is from a 1-word |
duke@435 | 2290 | // cache which is mutable so can't use immutable memory. Other |
duke@435 | 2291 | // types load from the super-class display table which is immutable. |
duke@435 | 2292 | Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); |
kvn@599 | 2293 | Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); |
duke@435 | 2294 | |
duke@435 | 2295 | // Compile speed common case: ARE a subtype and we canNOT fail |
duke@435 | 2296 | if( superklass == nkls ) |
duke@435 | 2297 | return top(); // false path is dead; no test needed. |
duke@435 | 2298 | |
duke@435 | 2299 | // See if we get an immediate positive hit. Happens roughly 83% of the |
duke@435 | 2300 | // time. Test to see if the value loaded just previously from the subklass |
duke@435 | 2301 | // is exactly the superklass. |
duke@435 | 2302 | Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); |
duke@435 | 2303 | Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); |
duke@435 | 2304 | IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); |
duke@435 | 2305 | Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); |
duke@435 | 2306 | set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); |
duke@435 | 2307 | |
duke@435 | 2308 | // Compile speed common case: Check for being deterministic right now. If |
duke@435 | 2309 | // chk_off is a constant and not equal to cacheoff then we are NOT a |
duke@435 | 2310 | // subklass. In this case we need exactly the 1 test above and we can |
duke@435 | 2311 | // return those results immediately. |
duke@435 | 2312 | if (!might_be_cache) { |
duke@435 | 2313 | Node* not_subtype_ctrl = control(); |
duke@435 | 2314 | set_control(iftrue1); // We need exactly the 1 test above |
duke@435 | 2315 | return not_subtype_ctrl; |
duke@435 | 2316 | } |
duke@435 | 2317 | |
duke@435 | 2318 | // Gather the various success & failures here |
duke@435 | 2319 | RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); |
duke@435 | 2320 | record_for_igvn(r_ok_subtype); |
duke@435 | 2321 | RegionNode *r_not_subtype = new (C, 3) RegionNode(3); |
duke@435 | 2322 | record_for_igvn(r_not_subtype); |
duke@435 | 2323 | |
duke@435 | 2324 | r_ok_subtype->init_req(1, iftrue1); |
duke@435 | 2325 | |
duke@435 | 2326 | // Check for immediate negative hit. Happens roughly 11% of the time (which |
duke@435 | 2327 | // is roughly 63% of the remaining cases). Test to see if the loaded |
duke@435 | 2328 | // check-offset points into the subklass display list or the 1-element |
duke@435 | 2329 | // cache. If it points to the display (and NOT the cache) and the display |
duke@435 | 2330 | // missed then it's not a subtype. |
duke@435 | 2331 | Node *cacheoff = _gvn.intcon(cacheoff_con); |
duke@435 | 2332 | Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); |
duke@435 | 2333 | Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); |
duke@435 | 2334 | IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); |
duke@435 | 2335 | r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); |
duke@435 | 2336 | set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); |
duke@435 | 2337 | |
jrose@1079 | 2338 | // Check for self. Very rare to get here, but it is taken 1/3 the time. |
duke@435 | 2339 | // No performance impact (too rare) but allows sharing of secondary arrays |
duke@435 | 2340 | // which has some footprint reduction. |
duke@435 | 2341 | Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); |
duke@435 | 2342 | Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); |
duke@435 | 2343 | IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); |
duke@435 | 2344 | r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); |
duke@435 | 2345 | set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); |
duke@435 | 2346 | |
jrose@1079 | 2347 | // -- Roads not taken here: -- |
jrose@1079 | 2348 | // We could also have chosen to perform the self-check at the beginning |
jrose@1079 | 2349 | // of this code sequence, as the assembler does. This would not pay off |
jrose@1079 | 2350 | // the same way, since the optimizer, unlike the assembler, can perform |
jrose@1079 | 2351 | // static type analysis to fold away many successful self-checks. |
jrose@1079 | 2352 | // Non-foldable self checks work better here in second position, because |
jrose@1079 | 2353 | // the initial primary superclass check subsumes a self-check for most |
jrose@1079 | 2354 | // types. An exception would be a secondary type like array-of-interface, |
jrose@1079 | 2355 | // which does not appear in its own primary supertype display. |
jrose@1079 | 2356 | // Finally, we could have chosen to move the self-check into the |
jrose@1079 | 2357 | // PartialSubtypeCheckNode, and from there out-of-line in a platform |
jrose@1079 | 2358 | // dependent manner. But it is worthwhile to have the check here, |
jrose@1079 | 2359 | // where it can be perhaps be optimized. The cost in code space is |
jrose@1079 | 2360 | // small (register compare, branch). |
jrose@1079 | 2361 | |
duke@435 | 2362 | // Now do a linear scan of the secondary super-klass array. Again, no real |
duke@435 | 2363 | // performance impact (too rare) but it's gotta be done. |
duke@435 | 2364 | // Since the code is rarely used, there is no penalty for moving it |
jrose@1079 | 2365 | // out of line, and it can only improve I-cache density. |
jrose@1079 | 2366 | // The decision to inline or out-of-line this final check is platform |
jrose@1079 | 2367 | // dependent, and is found in the AD file definition of PartialSubtypeCheck. |
duke@435 | 2368 | Node* psc = _gvn.transform( |
duke@435 | 2369 | new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); |
duke@435 | 2370 | |
duke@435 | 2371 | Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); |
duke@435 | 2372 | Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); |
duke@435 | 2373 | IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); |
duke@435 | 2374 | r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); |
duke@435 | 2375 | r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); |
duke@435 | 2376 | |
duke@435 | 2377 | // Return false path; set default control to true path. |
duke@435 | 2378 | set_control( _gvn.transform(r_ok_subtype) ); |
duke@435 | 2379 | return _gvn.transform(r_not_subtype); |
duke@435 | 2380 | } |
duke@435 | 2381 | |
duke@435 | 2382 | //----------------------------static_subtype_check----------------------------- |
duke@435 | 2383 | // Shortcut important common cases when superklass is exact: |
duke@435 | 2384 | // (0) superklass is java.lang.Object (can occur in reflective code) |
duke@435 | 2385 | // (1) subklass is already limited to a subtype of superklass => always ok |
duke@435 | 2386 | // (2) subklass does not overlap with superklass => always fail |
duke@435 | 2387 | // (3) superklass has NO subtypes and we can check with a simple compare. |
duke@435 | 2388 | int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { |
duke@435 | 2389 | if (StressReflectiveCode) { |
duke@435 | 2390 | return SSC_full_test; // Let caller generate the general case. |
duke@435 | 2391 | } |
duke@435 | 2392 | |
duke@435 | 2393 | if (superk == env()->Object_klass()) { |
duke@435 | 2394 | return SSC_always_true; // (0) this test cannot fail |
duke@435 | 2395 | } |
duke@435 | 2396 | |
duke@435 | 2397 | ciType* superelem = superk; |
duke@435 | 2398 | if (superelem->is_array_klass()) |
duke@435 | 2399 | superelem = superelem->as_array_klass()->base_element_type(); |
duke@435 | 2400 | |
duke@435 | 2401 | if (!subk->is_interface()) { // cannot trust static interface types yet |
duke@435 | 2402 | if (subk->is_subtype_of(superk)) { |
duke@435 | 2403 | return SSC_always_true; // (1) false path dead; no dynamic test needed |
duke@435 | 2404 | } |
duke@435 | 2405 | if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && |
duke@435 | 2406 | !superk->is_subtype_of(subk)) { |
duke@435 | 2407 | return SSC_always_false; |
duke@435 | 2408 | } |
duke@435 | 2409 | } |
duke@435 | 2410 | |
duke@435 | 2411 | // If casting to an instance klass, it must have no subtypes |
duke@435 | 2412 | if (superk->is_interface()) { |
duke@435 | 2413 | // Cannot trust interfaces yet. |
duke@435 | 2414 | // %%% S.B. superk->nof_implementors() == 1 |
duke@435 | 2415 | } else if (superelem->is_instance_klass()) { |
duke@435 | 2416 | ciInstanceKlass* ik = superelem->as_instance_klass(); |
duke@435 | 2417 | if (!ik->has_subklass() && !ik->is_interface()) { |
duke@435 | 2418 | if (!ik->is_final()) { |
duke@435 | 2419 | // Add a dependency if there is a chance of a later subclass. |
duke@435 | 2420 | C->dependencies()->assert_leaf_type(ik); |
duke@435 | 2421 | } |
duke@435 | 2422 | return SSC_easy_test; // (3) caller can do a simple ptr comparison |
duke@435 | 2423 | } |
duke@435 | 2424 | } else { |
duke@435 | 2425 | // A primitive array type has no subtypes. |
duke@435 | 2426 | return SSC_easy_test; // (3) caller can do a simple ptr comparison |
duke@435 | 2427 | } |
duke@435 | 2428 | |
duke@435 | 2429 | return SSC_full_test; |
duke@435 | 2430 | } |
duke@435 | 2431 | |
duke@435 | 2432 | // Profile-driven exact type check: |
duke@435 | 2433 | Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, |
duke@435 | 2434 | float prob, |
duke@435 | 2435 | Node* *casted_receiver) { |
duke@435 | 2436 | const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); |
duke@435 | 2437 | Node* recv_klass = load_object_klass(receiver); |
duke@435 | 2438 | Node* want_klass = makecon(tklass); |
duke@435 | 2439 | Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); |
duke@435 | 2440 | Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); |
duke@435 | 2441 | IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); |
duke@435 | 2442 | set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); |
duke@435 | 2443 | Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); |
duke@435 | 2444 | |
duke@435 | 2445 | const TypeOopPtr* recv_xtype = tklass->as_instance_type(); |
duke@435 | 2446 | assert(recv_xtype->klass_is_exact(), ""); |
duke@435 | 2447 | |
duke@435 | 2448 | // Subsume downstream occurrences of receiver with a cast to |
duke@435 | 2449 | // recv_xtype, since now we know what the type will be. |
duke@435 | 2450 | Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); |
duke@435 | 2451 | (*casted_receiver) = _gvn.transform(cast); |
duke@435 | 2452 | // (User must make the replace_in_map call.) |
duke@435 | 2453 | |
duke@435 | 2454 | return fail; |
duke@435 | 2455 | } |
duke@435 | 2456 | |
duke@435 | 2457 | |
duke@435 | 2458 | //-------------------------------gen_instanceof-------------------------------- |
duke@435 | 2459 | // Generate an instance-of idiom. Used by both the instance-of bytecode |
duke@435 | 2460 | // and the reflective instance-of call. |
duke@435 | 2461 | Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) { |
duke@435 | 2462 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
duke@435 | 2463 | assert( !stopped(), "dead parse path should be checked in callers" ); |
duke@435 | 2464 | assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), |
duke@435 | 2465 | "must check for not-null not-dead klass in callers"); |
duke@435 | 2466 | |
duke@435 | 2467 | // Make the merge point |
duke@435 | 2468 | enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; |
duke@435 | 2469 | RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
duke@435 | 2470 | Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); |
duke@435 | 2471 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
duke@435 | 2472 | |
duke@435 | 2473 | // Null check; get casted pointer; set region slot 3 |
duke@435 | 2474 | Node* null_ctl = top(); |
duke@435 | 2475 | Node* not_null_obj = null_check_oop(subobj, &null_ctl); |
duke@435 | 2476 | |
duke@435 | 2477 | // If not_null_obj is dead, only null-path is taken |
duke@435 | 2478 | if (stopped()) { // Doing instance-of on a NULL? |
duke@435 | 2479 | set_control(null_ctl); |
duke@435 | 2480 | return intcon(0); |
duke@435 | 2481 | } |
duke@435 | 2482 | region->init_req(_null_path, null_ctl); |
duke@435 | 2483 | phi ->init_req(_null_path, intcon(0)); // Set null path value |
duke@435 | 2484 | |
duke@435 | 2485 | // Load the object's klass |
duke@435 | 2486 | Node* obj_klass = load_object_klass(not_null_obj); |
duke@435 | 2487 | |
duke@435 | 2488 | // Generate the subtype check |
duke@435 | 2489 | Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); |
duke@435 | 2490 | |
duke@435 | 2491 | // Plug in the success path to the general merge in slot 1. |
duke@435 | 2492 | region->init_req(_obj_path, control()); |
duke@435 | 2493 | phi ->init_req(_obj_path, intcon(1)); |
duke@435 | 2494 | |
duke@435 | 2495 | // Plug in the failing path to the general merge in slot 2. |
duke@435 | 2496 | region->init_req(_fail_path, not_subtype_ctrl); |
duke@435 | 2497 | phi ->init_req(_fail_path, intcon(0)); |
duke@435 | 2498 | |
duke@435 | 2499 | // Return final merged results |
duke@435 | 2500 | set_control( _gvn.transform(region) ); |
duke@435 | 2501 | record_for_igvn(region); |
duke@435 | 2502 | return _gvn.transform(phi); |
duke@435 | 2503 | } |
duke@435 | 2504 | |
duke@435 | 2505 | //-------------------------------gen_checkcast--------------------------------- |
duke@435 | 2506 | // Generate a checkcast idiom. Used by both the checkcast bytecode and the |
duke@435 | 2507 | // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the |
duke@435 | 2508 | // uncommon-trap paths work. Adjust stack after this call. |
duke@435 | 2509 | // If failure_control is supplied and not null, it is filled in with |
duke@435 | 2510 | // the control edge for the cast failure. Otherwise, an appropriate |
duke@435 | 2511 | // uncommon trap or exception is thrown. |
duke@435 | 2512 | Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, |
duke@435 | 2513 | Node* *failure_control) { |
duke@435 | 2514 | kill_dead_locals(); // Benefit all the uncommon traps |
duke@435 | 2515 | const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); |
duke@435 | 2516 | const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); |
duke@435 | 2517 | |
duke@435 | 2518 | // Fast cutout: Check the case that the cast is vacuously true. |
duke@435 | 2519 | // This detects the common cases where the test will short-circuit |
duke@435 | 2520 | // away completely. We do this before we perform the null check, |
duke@435 | 2521 | // because if the test is going to turn into zero code, we don't |
duke@435 | 2522 | // want a residual null check left around. (Causes a slowdown, |
duke@435 | 2523 | // for example, in some objArray manipulations, such as a[i]=a[j].) |
duke@435 | 2524 | if (tk->singleton()) { |
duke@435 | 2525 | const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); |
duke@435 | 2526 | if (objtp != NULL && objtp->klass() != NULL) { |
duke@435 | 2527 | switch (static_subtype_check(tk->klass(), objtp->klass())) { |
duke@435 | 2528 | case SSC_always_true: |
duke@435 | 2529 | return obj; |
duke@435 | 2530 | case SSC_always_false: |
duke@435 | 2531 | // It needs a null check because a null will *pass* the cast check. |
duke@435 | 2532 | // A non-null value will always produce an exception. |
duke@435 | 2533 | return do_null_assert(obj, T_OBJECT); |
duke@435 | 2534 | } |
duke@435 | 2535 | } |
duke@435 | 2536 | } |
duke@435 | 2537 | |
duke@435 | 2538 | ciProfileData* data = NULL; |
duke@435 | 2539 | if (failure_control == NULL) { // use MDO in regular case only |
duke@435 | 2540 | assert(java_bc() == Bytecodes::_aastore || |
duke@435 | 2541 | java_bc() == Bytecodes::_checkcast, |
duke@435 | 2542 | "interpreter profiles type checks only for these BCs"); |
duke@435 | 2543 | data = method()->method_data()->bci_to_data(bci()); |
duke@435 | 2544 | } |
duke@435 | 2545 | |
duke@435 | 2546 | // Make the merge point |
duke@435 | 2547 | enum { _obj_path = 1, _null_path, PATH_LIMIT }; |
duke@435 | 2548 | RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); |
duke@435 | 2549 | Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); |
duke@435 | 2550 | C->set_has_split_ifs(true); // Has chance for split-if optimization |
duke@435 | 2551 | |
duke@435 | 2552 | // Use null-cast information if it is available |
duke@435 | 2553 | bool never_see_null = false; |
duke@435 | 2554 | // If we see an unexpected null at a check-cast we record it and force a |
duke@435 | 2555 | // recompile; the offending check-cast will be compiled to handle NULLs. |
duke@435 | 2556 | // If we see several offending BCIs, then all checkcasts in the |
duke@435 | 2557 | // method will be compiled to handle NULLs. |
duke@435 | 2558 | if (UncommonNullCast // Cutout for this technique |
duke@435 | 2559 | && failure_control == NULL // regular case |
duke@435 | 2560 | && obj != null() // And not the -Xcomp stupid case? |
duke@435 | 2561 | && !too_many_traps(Deoptimization::Reason_null_check)) { |
duke@435 | 2562 | // Finally, check the "null_seen" bit from the interpreter. |
duke@435 | 2563 | if (data == NULL || !data->as_BitData()->null_seen()) { |
duke@435 | 2564 | never_see_null = true; |
duke@435 | 2565 | } |
duke@435 | 2566 | } |
duke@435 | 2567 | |
duke@435 | 2568 | // Null check; get casted pointer; set region slot 3 |
duke@435 | 2569 | Node* null_ctl = top(); |
duke@435 | 2570 | Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); |
duke@435 | 2571 | |
duke@435 | 2572 | // If not_null_obj is dead, only null-path is taken |
duke@435 | 2573 | if (stopped()) { // Doing instance-of on a NULL? |
duke@435 | 2574 | set_control(null_ctl); |
duke@435 | 2575 | return null(); |
duke@435 | 2576 | } |
duke@435 | 2577 | region->init_req(_null_path, null_ctl); |
duke@435 | 2578 | phi ->init_req(_null_path, null()); // Set null path value |
duke@435 | 2579 | |
duke@435 | 2580 | Node* cast_obj = NULL; // the casted version of the object |
duke@435 | 2581 | |
duke@435 | 2582 | // If the profile has seen exactly one type, narrow to that type. |
duke@435 | 2583 | // (The subsequent subtype check will always fold up.) |
duke@435 | 2584 | if (UseTypeProfile && TypeProfileCasts && data != NULL && |
duke@435 | 2585 | // Counter has never been decremented (due to cast failure). |
duke@435 | 2586 | // ...This is a reasonable thing to expect. It is true of |
duke@435 | 2587 | // all casts inserted by javac to implement generic types. |
duke@435 | 2588 | data->as_CounterData()->count() >= 0 && |
duke@435 | 2589 | !too_many_traps(Deoptimization::Reason_class_check)) { |
duke@435 | 2590 | // (No, this isn't a call, but it's enough like a virtual call |
duke@435 | 2591 | // to use the same ciMethod accessor to get the profile info...) |
duke@435 | 2592 | ciCallProfile profile = method()->call_profile_at_bci(bci()); |
duke@435 | 2593 | if (profile.count() >= 0 && // no cast failures here |
duke@435 | 2594 | profile.has_receiver(0) && |
duke@435 | 2595 | profile.morphism() == 1) { |
duke@435 | 2596 | ciKlass* exact_kls = profile.receiver(0); |
duke@435 | 2597 | int ssc = static_subtype_check(tk->klass(), exact_kls); |
duke@435 | 2598 | if (ssc == SSC_always_true) { |
duke@435 | 2599 | // If we narrow the type to match what the type profile sees, |
duke@435 | 2600 | // we can then remove the rest of the cast. |
duke@435 | 2601 | // This is a win, even if the exact_kls is very specific, |
duke@435 | 2602 | // because downstream operations, such as method calls, |
duke@435 | 2603 | // will often benefit from the sharper type. |
duke@435 | 2604 | Node* exact_obj = not_null_obj; // will get updated in place... |
duke@435 | 2605 | Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, |
duke@435 | 2606 | &exact_obj); |
duke@435 | 2607 | { PreserveJVMState pjvms(this); |
duke@435 | 2608 | set_control(slow_ctl); |
duke@435 | 2609 | uncommon_trap(Deoptimization::Reason_class_check, |
duke@435 | 2610 | Deoptimization::Action_maybe_recompile); |
duke@435 | 2611 | } |
duke@435 | 2612 | if (failure_control != NULL) // failure is now impossible |
duke@435 | 2613 | (*failure_control) = top(); |
duke@435 | 2614 | replace_in_map(not_null_obj, exact_obj); |
duke@435 | 2615 | // adjust the type of the phi to the exact klass: |
duke@435 | 2616 | phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR)); |
duke@435 | 2617 | cast_obj = exact_obj; |
duke@435 | 2618 | } |
duke@435 | 2619 | // assert(cast_obj != NULL)... except maybe the profile lied to us. |
duke@435 | 2620 | } |
duke@435 | 2621 | } |
duke@435 | 2622 | |
duke@435 | 2623 | if (cast_obj == NULL) { |
duke@435 | 2624 | // Load the object's klass |
duke@435 | 2625 | Node* obj_klass = load_object_klass(not_null_obj); |
duke@435 | 2626 | |
duke@435 | 2627 | // Generate the subtype check |
duke@435 | 2628 | Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); |
duke@435 | 2629 | |
duke@435 | 2630 | // Plug in success path into the merge |
duke@435 | 2631 | cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), |
duke@435 | 2632 | not_null_obj, toop)); |
duke@435 | 2633 | // Failure path ends in uncommon trap (or may be dead - failure impossible) |
duke@435 | 2634 | if (failure_control == NULL) { |
duke@435 | 2635 | if (not_subtype_ctrl != top()) { // If failure is possible |
duke@435 | 2636 | PreserveJVMState pjvms(this); |
duke@435 | 2637 | set_control(not_subtype_ctrl); |
duke@435 | 2638 | builtin_throw(Deoptimization::Reason_class_check, obj_klass); |
duke@435 | 2639 | } |
duke@435 | 2640 | } else { |
duke@435 | 2641 | (*failure_control) = not_subtype_ctrl; |
duke@435 | 2642 | } |
duke@435 | 2643 | } |
duke@435 | 2644 | |
duke@435 | 2645 | region->init_req(_obj_path, control()); |
duke@435 | 2646 | phi ->init_req(_obj_path, cast_obj); |
duke@435 | 2647 | |
duke@435 | 2648 | // A merge of NULL or Casted-NotNull obj |
duke@435 | 2649 | Node* res = _gvn.transform(phi); |
duke@435 | 2650 | |
duke@435 | 2651 | // Note I do NOT always 'replace_in_map(obj,result)' here. |
duke@435 | 2652 | // if( tk->klass()->can_be_primary_super() ) |
duke@435 | 2653 | // This means that if I successfully store an Object into an array-of-String |
duke@435 | 2654 | // I 'forget' that the Object is really now known to be a String. I have to |
duke@435 | 2655 | // do this because we don't have true union types for interfaces - if I store |
duke@435 | 2656 | // a Baz into an array-of-Interface and then tell the optimizer it's an |
duke@435 | 2657 | // Interface, I forget that it's also a Baz and cannot do Baz-like field |
duke@435 | 2658 | // references to it. FIX THIS WHEN UNION TYPES APPEAR! |
duke@435 | 2659 | // replace_in_map( obj, res ); |
duke@435 | 2660 | |
duke@435 | 2661 | // Return final merged results |
duke@435 | 2662 | set_control( _gvn.transform(region) ); |
duke@435 | 2663 | record_for_igvn(region); |
duke@435 | 2664 | return res; |
duke@435 | 2665 | } |
duke@435 | 2666 | |
duke@435 | 2667 | //------------------------------next_monitor----------------------------------- |
duke@435 | 2668 | // What number should be given to the next monitor? |
duke@435 | 2669 | int GraphKit::next_monitor() { |
duke@435 | 2670 | int current = jvms()->monitor_depth()* C->sync_stack_slots(); |
duke@435 | 2671 | int next = current + C->sync_stack_slots(); |
duke@435 | 2672 | // Keep the toplevel high water mark current: |
duke@435 | 2673 | if (C->fixed_slots() < next) C->set_fixed_slots(next); |
duke@435 | 2674 | return current; |
duke@435 | 2675 | } |
duke@435 | 2676 | |
duke@435 | 2677 | //------------------------------insert_mem_bar--------------------------------- |
duke@435 | 2678 | // Memory barrier to avoid floating things around |
duke@435 | 2679 | // The membar serves as a pinch point between both control and all memory slices. |
duke@435 | 2680 | Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { |
duke@435 | 2681 | MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); |
duke@435 | 2682 | mb->init_req(TypeFunc::Control, control()); |
duke@435 | 2683 | mb->init_req(TypeFunc::Memory, reset_memory()); |
duke@435 | 2684 | Node* membar = _gvn.transform(mb); |
duke@435 | 2685 | set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); |
duke@435 | 2686 | set_all_memory_call(membar); |
duke@435 | 2687 | return membar; |
duke@435 | 2688 | } |
duke@435 | 2689 | |
duke@435 | 2690 | //-------------------------insert_mem_bar_volatile---------------------------- |
duke@435 | 2691 | // Memory barrier to avoid floating things around |
duke@435 | 2692 | // The membar serves as a pinch point between both control and memory(alias_idx). |
duke@435 | 2693 | // If you want to make a pinch point on all memory slices, do not use this |
duke@435 | 2694 | // function (even with AliasIdxBot); use insert_mem_bar() instead. |
duke@435 | 2695 | Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { |
duke@435 | 2696 | // When Parse::do_put_xxx updates a volatile field, it appends a series |
duke@435 | 2697 | // of MemBarVolatile nodes, one for *each* volatile field alias category. |
duke@435 | 2698 | // The first membar is on the same memory slice as the field store opcode. |
duke@435 | 2699 | // This forces the membar to follow the store. (Bug 6500685 broke this.) |
duke@435 | 2700 | // All the other membars (for other volatile slices, including AliasIdxBot, |
duke@435 | 2701 | // which stands for all unknown volatile slices) are control-dependent |
duke@435 | 2702 | // on the first membar. This prevents later volatile loads or stores |
duke@435 | 2703 | // from sliding up past the just-emitted store. |
duke@435 | 2704 | |
duke@435 | 2705 | MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); |
duke@435 | 2706 | mb->set_req(TypeFunc::Control,control()); |
duke@435 | 2707 | if (alias_idx == Compile::AliasIdxBot) { |
duke@435 | 2708 | mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); |
duke@435 | 2709 | } else { |
duke@435 | 2710 | assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); |
duke@435 | 2711 | mb->set_req(TypeFunc::Memory, memory(alias_idx)); |
duke@435 | 2712 | } |
duke@435 | 2713 | Node* membar = _gvn.transform(mb); |
duke@435 | 2714 | set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); |
duke@435 | 2715 | if (alias_idx == Compile::AliasIdxBot) { |
duke@435 | 2716 | merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); |
duke@435 | 2717 | } else { |
duke@435 | 2718 | set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); |
duke@435 | 2719 | } |
duke@435 | 2720 | return membar; |
duke@435 | 2721 | } |
duke@435 | 2722 | |
duke@435 | 2723 | //------------------------------shared_lock------------------------------------ |
duke@435 | 2724 | // Emit locking code. |
duke@435 | 2725 | FastLockNode* GraphKit::shared_lock(Node* obj) { |
duke@435 | 2726 | // bci is either a monitorenter bc or InvocationEntryBci |
duke@435 | 2727 | // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces |
duke@435 | 2728 | assert(SynchronizationEntryBCI == InvocationEntryBci, ""); |
duke@435 | 2729 | |
duke@435 | 2730 | if( !GenerateSynchronizationCode ) |
duke@435 | 2731 | return NULL; // Not locking things? |
duke@435 | 2732 | if (stopped()) // Dead monitor? |
duke@435 | 2733 | return NULL; |
duke@435 | 2734 | |
duke@435 | 2735 | assert(dead_locals_are_killed(), "should kill locals before sync. point"); |
duke@435 | 2736 | |
duke@435 | 2737 | // Box the stack location |
duke@435 | 2738 | Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); |
duke@435 | 2739 | Node* mem = reset_memory(); |
duke@435 | 2740 | |
duke@435 | 2741 | FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); |
duke@435 | 2742 | if (PrintPreciseBiasedLockingStatistics) { |
duke@435 | 2743 | // Create the counters for this fast lock. |
duke@435 | 2744 | flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci |
duke@435 | 2745 | } |
duke@435 | 2746 | // Add monitor to debug info for the slow path. If we block inside the |
duke@435 | 2747 | // slow path and de-opt, we need the monitor hanging around |
duke@435 | 2748 | map()->push_monitor( flock ); |
duke@435 | 2749 | |
duke@435 | 2750 | const TypeFunc *tf = LockNode::lock_type(); |
duke@435 | 2751 | LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); |
duke@435 | 2752 | |
duke@435 | 2753 | lock->init_req( TypeFunc::Control, control() ); |
duke@435 | 2754 | lock->init_req( TypeFunc::Memory , mem ); |
duke@435 | 2755 | lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o |
duke@435 | 2756 | lock->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 2757 | lock->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 2758 | |
duke@435 | 2759 | lock->init_req(TypeFunc::Parms + 0, obj); |
duke@435 | 2760 | lock->init_req(TypeFunc::Parms + 1, box); |
duke@435 | 2761 | lock->init_req(TypeFunc::Parms + 2, flock); |
duke@435 | 2762 | add_safepoint_edges(lock); |
duke@435 | 2763 | |
duke@435 | 2764 | lock = _gvn.transform( lock )->as_Lock(); |
duke@435 | 2765 | |
duke@435 | 2766 | // lock has no side-effects, sets few values |
duke@435 | 2767 | set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); |
duke@435 | 2768 | |
duke@435 | 2769 | insert_mem_bar(Op_MemBarAcquire); |
duke@435 | 2770 | |
duke@435 | 2771 | // Add this to the worklist so that the lock can be eliminated |
duke@435 | 2772 | record_for_igvn(lock); |
duke@435 | 2773 | |
duke@435 | 2774 | #ifndef PRODUCT |
duke@435 | 2775 | if (PrintLockStatistics) { |
duke@435 | 2776 | // Update the counter for this lock. Don't bother using an atomic |
duke@435 | 2777 | // operation since we don't require absolute accuracy. |
duke@435 | 2778 | lock->create_lock_counter(map()->jvms()); |
duke@435 | 2779 | int adr_type = Compile::AliasIdxRaw; |
duke@435 | 2780 | Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr())); |
duke@435 | 2781 | Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); |
duke@435 | 2782 | Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); |
duke@435 | 2783 | store_to_memory(control(), counter_addr, incr, T_INT, adr_type); |
duke@435 | 2784 | } |
duke@435 | 2785 | #endif |
duke@435 | 2786 | |
duke@435 | 2787 | return flock; |
duke@435 | 2788 | } |
duke@435 | 2789 | |
duke@435 | 2790 | |
duke@435 | 2791 | //------------------------------shared_unlock---------------------------------- |
duke@435 | 2792 | // Emit unlocking code. |
duke@435 | 2793 | void GraphKit::shared_unlock(Node* box, Node* obj) { |
duke@435 | 2794 | // bci is either a monitorenter bc or InvocationEntryBci |
duke@435 | 2795 | // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces |
duke@435 | 2796 | assert(SynchronizationEntryBCI == InvocationEntryBci, ""); |
duke@435 | 2797 | |
duke@435 | 2798 | if( !GenerateSynchronizationCode ) |
duke@435 | 2799 | return; |
duke@435 | 2800 | if (stopped()) { // Dead monitor? |
duke@435 | 2801 | map()->pop_monitor(); // Kill monitor from debug info |
duke@435 | 2802 | return; |
duke@435 | 2803 | } |
duke@435 | 2804 | |
duke@435 | 2805 | // Memory barrier to avoid floating things down past the locked region |
duke@435 | 2806 | insert_mem_bar(Op_MemBarRelease); |
duke@435 | 2807 | |
duke@435 | 2808 | const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); |
duke@435 | 2809 | UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); |
duke@435 | 2810 | uint raw_idx = Compile::AliasIdxRaw; |
duke@435 | 2811 | unlock->init_req( TypeFunc::Control, control() ); |
duke@435 | 2812 | unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); |
duke@435 | 2813 | unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o |
duke@435 | 2814 | unlock->init_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 2815 | unlock->init_req( TypeFunc::ReturnAdr, top() ); |
duke@435 | 2816 | |
duke@435 | 2817 | unlock->init_req(TypeFunc::Parms + 0, obj); |
duke@435 | 2818 | unlock->init_req(TypeFunc::Parms + 1, box); |
duke@435 | 2819 | unlock = _gvn.transform(unlock)->as_Unlock(); |
duke@435 | 2820 | |
duke@435 | 2821 | Node* mem = reset_memory(); |
duke@435 | 2822 | |
duke@435 | 2823 | // unlock has no side-effects, sets few values |
duke@435 | 2824 | set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); |
duke@435 | 2825 | |
duke@435 | 2826 | // Kill monitor from debug info |
duke@435 | 2827 | map()->pop_monitor( ); |
duke@435 | 2828 | } |
duke@435 | 2829 | |
duke@435 | 2830 | //-------------------------------get_layout_helper----------------------------- |
duke@435 | 2831 | // If the given klass is a constant or known to be an array, |
duke@435 | 2832 | // fetch the constant layout helper value into constant_value |
duke@435 | 2833 | // and return (Node*)NULL. Otherwise, load the non-constant |
duke@435 | 2834 | // layout helper value, and return the node which represents it. |
duke@435 | 2835 | // This two-faced routine is useful because allocation sites |
duke@435 | 2836 | // almost always feature constant types. |
duke@435 | 2837 | Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { |
duke@435 | 2838 | const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); |
duke@435 | 2839 | if (!StressReflectiveCode && inst_klass != NULL) { |
duke@435 | 2840 | ciKlass* klass = inst_klass->klass(); |
duke@435 | 2841 | bool xklass = inst_klass->klass_is_exact(); |
duke@435 | 2842 | if (xklass || klass->is_array_klass()) { |
duke@435 | 2843 | jint lhelper = klass->layout_helper(); |
duke@435 | 2844 | if (lhelper != Klass::_lh_neutral_value) { |
duke@435 | 2845 | constant_value = lhelper; |
duke@435 | 2846 | return (Node*) NULL; |
duke@435 | 2847 | } |
duke@435 | 2848 | } |
duke@435 | 2849 | } |
duke@435 | 2850 | constant_value = Klass::_lh_neutral_value; // put in a known value |
duke@435 | 2851 | Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc)); |
duke@435 | 2852 | return make_load(NULL, lhp, TypeInt::INT, T_INT); |
duke@435 | 2853 | } |
duke@435 | 2854 | |
duke@435 | 2855 | // We just put in an allocate/initialize with a big raw-memory effect. |
duke@435 | 2856 | // Hook selected additional alias categories on the initialization. |
duke@435 | 2857 | static void hook_memory_on_init(GraphKit& kit, int alias_idx, |
duke@435 | 2858 | MergeMemNode* init_in_merge, |
duke@435 | 2859 | Node* init_out_raw) { |
duke@435 | 2860 | DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); |
duke@435 | 2861 | assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); |
duke@435 | 2862 | |
duke@435 | 2863 | Node* prevmem = kit.memory(alias_idx); |
duke@435 | 2864 | init_in_merge->set_memory_at(alias_idx, prevmem); |
duke@435 | 2865 | kit.set_memory(init_out_raw, alias_idx); |
duke@435 | 2866 | } |
duke@435 | 2867 | |
duke@435 | 2868 | //---------------------------set_output_for_allocation------------------------- |
duke@435 | 2869 | Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, |
duke@435 | 2870 | const TypeOopPtr* oop_type, |
duke@435 | 2871 | bool raw_mem_only) { |
duke@435 | 2872 | int rawidx = Compile::AliasIdxRaw; |
duke@435 | 2873 | alloc->set_req( TypeFunc::FramePtr, frameptr() ); |
duke@435 | 2874 | add_safepoint_edges(alloc); |
duke@435 | 2875 | Node* allocx = _gvn.transform(alloc); |
duke@435 | 2876 | set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); |
duke@435 | 2877 | // create memory projection for i_o |
duke@435 | 2878 | set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); |
duke@435 | 2879 | make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); |
duke@435 | 2880 | |
duke@435 | 2881 | // create a memory projection as for the normal control path |
duke@435 | 2882 | Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); |
duke@435 | 2883 | set_memory(malloc, rawidx); |
duke@435 | 2884 | |
duke@435 | 2885 | // a normal slow-call doesn't change i_o, but an allocation does |
duke@435 | 2886 | // we create a separate i_o projection for the normal control path |
duke@435 | 2887 | set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); |
duke@435 | 2888 | Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); |
duke@435 | 2889 | |
duke@435 | 2890 | // put in an initialization barrier |
duke@435 | 2891 | InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, |
duke@435 | 2892 | rawoop)->as_Initialize(); |
duke@435 | 2893 | assert(alloc->initialization() == init, "2-way macro link must work"); |
duke@435 | 2894 | assert(init ->allocation() == alloc, "2-way macro link must work"); |
duke@435 | 2895 | if (ReduceFieldZeroing && !raw_mem_only) { |
duke@435 | 2896 | // Extract memory strands which may participate in the new object's |
duke@435 | 2897 | // initialization, and source them from the new InitializeNode. |
duke@435 | 2898 | // This will allow us to observe initializations when they occur, |
duke@435 | 2899 | // and link them properly (as a group) to the InitializeNode. |
duke@435 | 2900 | assert(init->in(InitializeNode::Memory) == malloc, ""); |
duke@435 | 2901 | MergeMemNode* minit_in = MergeMemNode::make(C, malloc); |
duke@435 | 2902 | init->set_req(InitializeNode::Memory, minit_in); |
duke@435 | 2903 | record_for_igvn(minit_in); // fold it up later, if possible |
duke@435 | 2904 | Node* minit_out = memory(rawidx); |
duke@435 | 2905 | assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); |
duke@435 | 2906 | if (oop_type->isa_aryptr()) { |
duke@435 | 2907 | const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); |
duke@435 | 2908 | int elemidx = C->get_alias_index(telemref); |
duke@435 | 2909 | hook_memory_on_init(*this, elemidx, minit_in, minit_out); |
duke@435 | 2910 | } else if (oop_type->isa_instptr()) { |
duke@435 | 2911 | ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); |
duke@435 | 2912 | for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { |
duke@435 | 2913 | ciField* field = ik->nonstatic_field_at(i); |
kvn@458 | 2914 | if (field->offset() >= TrackedInitializationLimit * HeapWordSize) |
duke@435 | 2915 | continue; // do not bother to track really large numbers of fields |
duke@435 | 2916 | // Find (or create) the alias category for this field: |
duke@435 | 2917 | int fieldidx = C->alias_type(field)->index(); |
duke@435 | 2918 | hook_memory_on_init(*this, fieldidx, minit_in, minit_out); |
duke@435 | 2919 | } |
duke@435 | 2920 | } |
duke@435 | 2921 | } |
duke@435 | 2922 | |
duke@435 | 2923 | // Cast raw oop to the real thing... |
duke@435 | 2924 | Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); |
duke@435 | 2925 | javaoop = _gvn.transform(javaoop); |
duke@435 | 2926 | C->set_recent_alloc(control(), javaoop); |
duke@435 | 2927 | assert(just_allocated_object(control()) == javaoop, "just allocated"); |
duke@435 | 2928 | |
duke@435 | 2929 | #ifdef ASSERT |
rasbold@801 | 2930 | { // Verify that the AllocateNode::Ideal_allocation recognizers work: |
rasbold@801 | 2931 | assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, |
rasbold@801 | 2932 | "Ideal_allocation works"); |
rasbold@801 | 2933 | assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, |
rasbold@801 | 2934 | "Ideal_allocation works"); |
duke@435 | 2935 | if (alloc->is_AllocateArray()) { |
rasbold@801 | 2936 | assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), |
rasbold@801 | 2937 | "Ideal_allocation works"); |
rasbold@801 | 2938 | assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), |
rasbold@801 | 2939 | "Ideal_allocation works"); |
duke@435 | 2940 | } else { |
rasbold@801 | 2941 | assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); |
duke@435 | 2942 | } |
duke@435 | 2943 | } |
duke@435 | 2944 | #endif //ASSERT |
duke@435 | 2945 | |
duke@435 | 2946 | return javaoop; |
duke@435 | 2947 | } |
duke@435 | 2948 | |
duke@435 | 2949 | //---------------------------new_instance-------------------------------------- |
duke@435 | 2950 | // This routine takes a klass_node which may be constant (for a static type) |
duke@435 | 2951 | // or may be non-constant (for reflective code). It will work equally well |
duke@435 | 2952 | // for either, and the graph will fold nicely if the optimizer later reduces |
duke@435 | 2953 | // the type to a constant. |
duke@435 | 2954 | // The optional arguments are for specialized use by intrinsics: |
duke@435 | 2955 | // - If 'extra_slow_test' if not null is an extra condition for the slow-path. |
duke@435 | 2956 | // - If 'raw_mem_only', do not cast the result to an oop. |
duke@435 | 2957 | // - If 'return_size_val', report the the total object size to the caller. |
duke@435 | 2958 | Node* GraphKit::new_instance(Node* klass_node, |
duke@435 | 2959 | Node* extra_slow_test, |
duke@435 | 2960 | bool raw_mem_only, // affect only raw memory |
duke@435 | 2961 | Node* *return_size_val) { |
duke@435 | 2962 | // Compute size in doublewords |
duke@435 | 2963 | // The size is always an integral number of doublewords, represented |
duke@435 | 2964 | // as a positive bytewise size stored in the klass's layout_helper. |
duke@435 | 2965 | // The layout_helper also encodes (in a low bit) the need for a slow path. |
duke@435 | 2966 | jint layout_con = Klass::_lh_neutral_value; |
duke@435 | 2967 | Node* layout_val = get_layout_helper(klass_node, layout_con); |
duke@435 | 2968 | int layout_is_con = (layout_val == NULL); |
duke@435 | 2969 | |
duke@435 | 2970 | if (extra_slow_test == NULL) extra_slow_test = intcon(0); |
duke@435 | 2971 | // Generate the initial go-slow test. It's either ALWAYS (return a |
duke@435 | 2972 | // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective |
duke@435 | 2973 | // case) a computed value derived from the layout_helper. |
duke@435 | 2974 | Node* initial_slow_test = NULL; |
duke@435 | 2975 | if (layout_is_con) { |
duke@435 | 2976 | assert(!StressReflectiveCode, "stress mode does not use these paths"); |
duke@435 | 2977 | bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); |
duke@435 | 2978 | initial_slow_test = must_go_slow? intcon(1): extra_slow_test; |
duke@435 | 2979 | |
duke@435 | 2980 | } else { // reflective case |
duke@435 | 2981 | // This reflective path is used by Unsafe.allocateInstance. |
duke@435 | 2982 | // (It may be stress-tested by specifying StressReflectiveCode.) |
duke@435 | 2983 | // Basically, we want to get into the VM is there's an illegal argument. |
duke@435 | 2984 | Node* bit = intcon(Klass::_lh_instance_slow_path_bit); |
duke@435 | 2985 | initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); |
duke@435 | 2986 | if (extra_slow_test != intcon(0)) { |
duke@435 | 2987 | initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); |
duke@435 | 2988 | } |
duke@435 | 2989 | // (Macro-expander will further convert this to a Bool, if necessary.) |
duke@435 | 2990 | } |
duke@435 | 2991 | |
duke@435 | 2992 | // Find the size in bytes. This is easy; it's the layout_helper. |
duke@435 | 2993 | // The size value must be valid even if the slow path is taken. |
duke@435 | 2994 | Node* size = NULL; |
duke@435 | 2995 | if (layout_is_con) { |
duke@435 | 2996 | size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); |
duke@435 | 2997 | } else { // reflective case |
duke@435 | 2998 | // This reflective path is used by clone and Unsafe.allocateInstance. |
duke@435 | 2999 | size = ConvI2X(layout_val); |
duke@435 | 3000 | |
duke@435 | 3001 | // Clear the low bits to extract layout_helper_size_in_bytes: |
duke@435 | 3002 | assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); |
duke@435 | 3003 | Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); |
duke@435 | 3004 | size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); |
duke@435 | 3005 | } |
duke@435 | 3006 | if (return_size_val != NULL) { |
duke@435 | 3007 | (*return_size_val) = size; |
duke@435 | 3008 | } |
duke@435 | 3009 | |
duke@435 | 3010 | // This is a precise notnull oop of the klass. |
duke@435 | 3011 | // (Actually, it need not be precise if this is a reflective allocation.) |
duke@435 | 3012 | // It's what we cast the result to. |
duke@435 | 3013 | const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); |
duke@435 | 3014 | if (!tklass) tklass = TypeKlassPtr::OBJECT; |
duke@435 | 3015 | const TypeOopPtr* oop_type = tklass->as_instance_type(); |
duke@435 | 3016 | |
duke@435 | 3017 | // Now generate allocation code |
kvn@509 | 3018 | |
kvn@1000 | 3019 | // The entire memory state is needed for slow path of the allocation |
kvn@1000 | 3020 | // since GC and deoptimization can happened. |
kvn@1000 | 3021 | Node *mem = reset_memory(); |
kvn@1000 | 3022 | set_all_memory(mem); // Create new memory state |
kvn@509 | 3023 | |
duke@435 | 3024 | AllocateNode* alloc |
duke@435 | 3025 | = new (C, AllocateNode::ParmLimit) |
duke@435 | 3026 | AllocateNode(C, AllocateNode::alloc_type(), |
kvn@509 | 3027 | control(), mem, i_o(), |
duke@435 | 3028 | size, klass_node, |
duke@435 | 3029 | initial_slow_test); |
duke@435 | 3030 | |
duke@435 | 3031 | return set_output_for_allocation(alloc, oop_type, raw_mem_only); |
duke@435 | 3032 | } |
duke@435 | 3033 | |
duke@435 | 3034 | //-------------------------------new_array------------------------------------- |
duke@435 | 3035 | // helper for both newarray and anewarray |
duke@435 | 3036 | // The 'length' parameter is (obviously) the length of the array. |
duke@435 | 3037 | // See comments on new_instance for the meaning of the other arguments. |
duke@435 | 3038 | Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) |
duke@435 | 3039 | Node* length, // number of array elements |
cfang@1165 | 3040 | int nargs, // number of arguments to push back for uncommon trap |
duke@435 | 3041 | bool raw_mem_only, // affect only raw memory |
duke@435 | 3042 | Node* *return_size_val) { |
duke@435 | 3043 | jint layout_con = Klass::_lh_neutral_value; |
duke@435 | 3044 | Node* layout_val = get_layout_helper(klass_node, layout_con); |
duke@435 | 3045 | int layout_is_con = (layout_val == NULL); |
duke@435 | 3046 | |
duke@435 | 3047 | if (!layout_is_con && !StressReflectiveCode && |
duke@435 | 3048 | !too_many_traps(Deoptimization::Reason_class_check)) { |
duke@435 | 3049 | // This is a reflective array creation site. |
duke@435 | 3050 | // Optimistically assume that it is a subtype of Object[], |
duke@435 | 3051 | // so that we can fold up all the address arithmetic. |
duke@435 | 3052 | layout_con = Klass::array_layout_helper(T_OBJECT); |
duke@435 | 3053 | Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); |
duke@435 | 3054 | Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); |
duke@435 | 3055 | { BuildCutout unless(this, bol_lh, PROB_MAX); |
cfang@1165 | 3056 | _sp += nargs; |
duke@435 | 3057 | uncommon_trap(Deoptimization::Reason_class_check, |
duke@435 | 3058 | Deoptimization::Action_maybe_recompile); |
duke@435 | 3059 | } |
duke@435 | 3060 | layout_val = NULL; |
duke@435 | 3061 | layout_is_con = true; |
duke@435 | 3062 | } |
duke@435 | 3063 | |
duke@435 | 3064 | // Generate the initial go-slow test. Make sure we do not overflow |
duke@435 | 3065 | // if length is huge (near 2Gig) or negative! We do not need |
duke@435 | 3066 | // exact double-words here, just a close approximation of needed |
duke@435 | 3067 | // double-words. We can't add any offset or rounding bits, lest we |
duke@435 | 3068 | // take a size -1 of bytes and make it positive. Use an unsigned |
duke@435 | 3069 | // compare, so negative sizes look hugely positive. |
duke@435 | 3070 | int fast_size_limit = FastAllocateSizeLimit; |
duke@435 | 3071 | if (layout_is_con) { |
duke@435 | 3072 | assert(!StressReflectiveCode, "stress mode does not use these paths"); |
duke@435 | 3073 | // Increase the size limit if we have exact knowledge of array type. |
duke@435 | 3074 | int log2_esize = Klass::layout_helper_log2_element_size(layout_con); |
duke@435 | 3075 | fast_size_limit <<= (LogBytesPerLong - log2_esize); |
duke@435 | 3076 | } |
duke@435 | 3077 | |
duke@435 | 3078 | Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); |
duke@435 | 3079 | Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); |
duke@435 | 3080 | if (initial_slow_test->is_Bool()) { |
duke@435 | 3081 | // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. |
duke@435 | 3082 | initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); |
duke@435 | 3083 | } |
duke@435 | 3084 | |
duke@435 | 3085 | // --- Size Computation --- |
duke@435 | 3086 | // array_size = round_to_heap(array_header + (length << elem_shift)); |
duke@435 | 3087 | // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) |
duke@435 | 3088 | // and round_to(x, y) == ((x + y-1) & ~(y-1)) |
duke@435 | 3089 | // The rounding mask is strength-reduced, if possible. |
duke@435 | 3090 | int round_mask = MinObjAlignmentInBytes - 1; |
duke@435 | 3091 | Node* header_size = NULL; |
duke@435 | 3092 | int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
duke@435 | 3093 | // (T_BYTE has the weakest alignment and size restrictions...) |
duke@435 | 3094 | if (layout_is_con) { |
duke@435 | 3095 | int hsize = Klass::layout_helper_header_size(layout_con); |
duke@435 | 3096 | int eshift = Klass::layout_helper_log2_element_size(layout_con); |
duke@435 | 3097 | BasicType etype = Klass::layout_helper_element_type(layout_con); |
duke@435 | 3098 | if ((round_mask & ~right_n_bits(eshift)) == 0) |
duke@435 | 3099 | round_mask = 0; // strength-reduce it if it goes away completely |
duke@435 | 3100 | assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); |
duke@435 | 3101 | assert(header_size_min <= hsize, "generic minimum is smallest"); |
duke@435 | 3102 | header_size_min = hsize; |
duke@435 | 3103 | header_size = intcon(hsize + round_mask); |
duke@435 | 3104 | } else { |
duke@435 | 3105 | Node* hss = intcon(Klass::_lh_header_size_shift); |
duke@435 | 3106 | Node* hsm = intcon(Klass::_lh_header_size_mask); |
duke@435 | 3107 | Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); |
duke@435 | 3108 | hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); |
duke@435 | 3109 | Node* mask = intcon(round_mask); |
duke@435 | 3110 | header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); |
duke@435 | 3111 | } |
duke@435 | 3112 | |
duke@435 | 3113 | Node* elem_shift = NULL; |
duke@435 | 3114 | if (layout_is_con) { |
duke@435 | 3115 | int eshift = Klass::layout_helper_log2_element_size(layout_con); |
duke@435 | 3116 | if (eshift != 0) |
duke@435 | 3117 | elem_shift = intcon(eshift); |
duke@435 | 3118 | } else { |
duke@435 | 3119 | // There is no need to mask or shift this value. |
duke@435 | 3120 | // The semantics of LShiftINode include an implicit mask to 0x1F. |
duke@435 | 3121 | assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); |
duke@435 | 3122 | elem_shift = layout_val; |
duke@435 | 3123 | } |
duke@435 | 3124 | |
duke@435 | 3125 | // Transition to native address size for all offset calculations: |
duke@435 | 3126 | Node* lengthx = ConvI2X(length); |
duke@435 | 3127 | Node* headerx = ConvI2X(header_size); |
duke@435 | 3128 | #ifdef _LP64 |
duke@435 | 3129 | { const TypeLong* tllen = _gvn.find_long_type(lengthx); |
duke@435 | 3130 | if (tllen != NULL && tllen->_lo < 0) { |
duke@435 | 3131 | // Add a manual constraint to a positive range. Cf. array_element_address. |
duke@435 | 3132 | jlong size_max = arrayOopDesc::max_array_length(T_BYTE); |
duke@435 | 3133 | if (size_max > tllen->_hi) size_max = tllen->_hi; |
duke@435 | 3134 | const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); |
duke@435 | 3135 | lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); |
duke@435 | 3136 | } |
duke@435 | 3137 | } |
duke@435 | 3138 | #endif |
duke@435 | 3139 | |
duke@435 | 3140 | // Combine header size (plus rounding) and body size. Then round down. |
duke@435 | 3141 | // This computation cannot overflow, because it is used only in two |
duke@435 | 3142 | // places, one where the length is sharply limited, and the other |
duke@435 | 3143 | // after a successful allocation. |
duke@435 | 3144 | Node* abody = lengthx; |
duke@435 | 3145 | if (elem_shift != NULL) |
duke@435 | 3146 | abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); |
duke@435 | 3147 | Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); |
duke@435 | 3148 | if (round_mask != 0) { |
duke@435 | 3149 | Node* mask = MakeConX(~round_mask); |
duke@435 | 3150 | size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); |
duke@435 | 3151 | } |
duke@435 | 3152 | // else if round_mask == 0, the size computation is self-rounding |
duke@435 | 3153 | |
duke@435 | 3154 | if (return_size_val != NULL) { |
duke@435 | 3155 | // This is the size |
duke@435 | 3156 | (*return_size_val) = size; |
duke@435 | 3157 | } |
duke@435 | 3158 | |
duke@435 | 3159 | // Now generate allocation code |
kvn@509 | 3160 | |
kvn@1000 | 3161 | // The entire memory state is needed for slow path of the allocation |
kvn@1000 | 3162 | // since GC and deoptimization can happened. |
kvn@1000 | 3163 | Node *mem = reset_memory(); |
kvn@1000 | 3164 | set_all_memory(mem); // Create new memory state |
kvn@509 | 3165 | |
duke@435 | 3166 | // Create the AllocateArrayNode and its result projections |
duke@435 | 3167 | AllocateArrayNode* alloc |
duke@435 | 3168 | = new (C, AllocateArrayNode::ParmLimit) |
duke@435 | 3169 | AllocateArrayNode(C, AllocateArrayNode::alloc_type(), |
kvn@509 | 3170 | control(), mem, i_o(), |
duke@435 | 3171 | size, klass_node, |
duke@435 | 3172 | initial_slow_test, |
duke@435 | 3173 | length); |
duke@435 | 3174 | |
duke@435 | 3175 | // Cast to correct type. Note that the klass_node may be constant or not, |
duke@435 | 3176 | // and in the latter case the actual array type will be inexact also. |
duke@435 | 3177 | // (This happens via a non-constant argument to inline_native_newArray.) |
duke@435 | 3178 | // In any case, the value of klass_node provides the desired array type. |
duke@435 | 3179 | const TypeInt* length_type = _gvn.find_int_type(length); |
duke@435 | 3180 | const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); |
duke@435 | 3181 | if (ary_type->isa_aryptr() && length_type != NULL) { |
duke@435 | 3182 | // Try to get a better type than POS for the size |
duke@435 | 3183 | ary_type = ary_type->is_aryptr()->cast_to_size(length_type); |
duke@435 | 3184 | } |
duke@435 | 3185 | |
duke@435 | 3186 | Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only); |
duke@435 | 3187 | |
rasbold@801 | 3188 | // Cast length on remaining path to be as narrow as possible |
rasbold@801 | 3189 | if (map()->find_edge(length) >= 0) { |
rasbold@801 | 3190 | Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); |
rasbold@801 | 3191 | if (ccast != length) { |
rasbold@801 | 3192 | _gvn.set_type_bottom(ccast); |
rasbold@801 | 3193 | record_for_igvn(ccast); |
duke@435 | 3194 | replace_in_map(length, ccast); |
duke@435 | 3195 | } |
duke@435 | 3196 | } |
duke@435 | 3197 | |
duke@435 | 3198 | return javaoop; |
duke@435 | 3199 | } |
duke@435 | 3200 | |
duke@435 | 3201 | // The following "Ideal_foo" functions are placed here because they recognize |
duke@435 | 3202 | // the graph shapes created by the functions immediately above. |
duke@435 | 3203 | |
duke@435 | 3204 | //---------------------------Ideal_allocation---------------------------------- |
duke@435 | 3205 | // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. |
duke@435 | 3206 | AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { |
duke@435 | 3207 | if (ptr == NULL) { // reduce dumb test in callers |
duke@435 | 3208 | return NULL; |
duke@435 | 3209 | } |
duke@435 | 3210 | if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast |
duke@435 | 3211 | ptr = ptr->in(1); |
duke@435 | 3212 | if (ptr == NULL) return NULL; |
duke@435 | 3213 | } |
duke@435 | 3214 | if (ptr->is_Proj()) { |
duke@435 | 3215 | Node* allo = ptr->in(0); |
duke@435 | 3216 | if (allo != NULL && allo->is_Allocate()) { |
duke@435 | 3217 | return allo->as_Allocate(); |
duke@435 | 3218 | } |
duke@435 | 3219 | } |
duke@435 | 3220 | // Report failure to match. |
duke@435 | 3221 | return NULL; |
duke@435 | 3222 | } |
duke@435 | 3223 | |
duke@435 | 3224 | // Fancy version which also strips off an offset (and reports it to caller). |
duke@435 | 3225 | AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, |
duke@435 | 3226 | intptr_t& offset) { |
duke@435 | 3227 | Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); |
duke@435 | 3228 | if (base == NULL) return NULL; |
duke@435 | 3229 | return Ideal_allocation(base, phase); |
duke@435 | 3230 | } |
duke@435 | 3231 | |
duke@435 | 3232 | // Trace Initialize <- Proj[Parm] <- Allocate |
duke@435 | 3233 | AllocateNode* InitializeNode::allocation() { |
duke@435 | 3234 | Node* rawoop = in(InitializeNode::RawAddress); |
duke@435 | 3235 | if (rawoop->is_Proj()) { |
duke@435 | 3236 | Node* alloc = rawoop->in(0); |
duke@435 | 3237 | if (alloc->is_Allocate()) { |
duke@435 | 3238 | return alloc->as_Allocate(); |
duke@435 | 3239 | } |
duke@435 | 3240 | } |
duke@435 | 3241 | return NULL; |
duke@435 | 3242 | } |
duke@435 | 3243 | |
duke@435 | 3244 | // Trace Allocate -> Proj[Parm] -> Initialize |
duke@435 | 3245 | InitializeNode* AllocateNode::initialization() { |
duke@435 | 3246 | ProjNode* rawoop = proj_out(AllocateNode::RawAddress); |
duke@435 | 3247 | if (rawoop == NULL) return NULL; |
duke@435 | 3248 | for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { |
duke@435 | 3249 | Node* init = rawoop->fast_out(i); |
duke@435 | 3250 | if (init->is_Initialize()) { |
duke@435 | 3251 | assert(init->as_Initialize()->allocation() == this, "2-way link"); |
duke@435 | 3252 | return init->as_Initialize(); |
duke@435 | 3253 | } |
duke@435 | 3254 | } |
duke@435 | 3255 | return NULL; |
duke@435 | 3256 | } |
ysr@777 | 3257 | |
kvn@1286 | 3258 | //----------------------------- store barriers ---------------------------- |
kvn@1286 | 3259 | #define __ ideal. |
kvn@1286 | 3260 | |
kvn@1286 | 3261 | void GraphKit::sync_kit(IdealKit& ideal) { |
kvn@1286 | 3262 | // Final sync IdealKit and graphKit. |
kvn@1286 | 3263 | __ drain_delay_transform(); |
kvn@1286 | 3264 | set_all_memory(__ merged_memory()); |
kvn@1286 | 3265 | set_control(__ ctrl()); |
kvn@1286 | 3266 | } |
kvn@1286 | 3267 | |
kvn@1286 | 3268 | // vanilla/CMS post barrier |
kvn@1286 | 3269 | // Insert a write-barrier store. This is to let generational GC work; we have |
kvn@1286 | 3270 | // to flag all oop-stores before the next GC point. |
kvn@1286 | 3271 | void GraphKit::write_barrier_post(Node* oop_store, |
kvn@1286 | 3272 | Node* obj, |
kvn@1286 | 3273 | Node* adr, |
cfang@1420 | 3274 | uint adr_idx, |
kvn@1286 | 3275 | Node* val, |
kvn@1286 | 3276 | bool use_precise) { |
kvn@1286 | 3277 | // No store check needed if we're storing a NULL or an old object |
kvn@1286 | 3278 | // (latter case is probably a string constant). The concurrent |
kvn@1286 | 3279 | // mark sweep garbage collector, however, needs to have all nonNull |
kvn@1286 | 3280 | // oop updates flagged via card-marks. |
kvn@1286 | 3281 | if (val != NULL && val->is_Con()) { |
kvn@1286 | 3282 | // must be either an oop or NULL |
kvn@1286 | 3283 | const Type* t = val->bottom_type(); |
kvn@1286 | 3284 | if (t == TypePtr::NULL_PTR || t == Type::TOP) |
kvn@1286 | 3285 | // stores of null never (?) need barriers |
kvn@1286 | 3286 | return; |
kvn@1286 | 3287 | ciObject* con = t->is_oopptr()->const_oop(); |
kvn@1286 | 3288 | if (con != NULL |
kvn@1286 | 3289 | && con->is_perm() |
kvn@1286 | 3290 | && Universe::heap()->can_elide_permanent_oop_store_barriers()) |
kvn@1286 | 3291 | // no store barrier needed, because no old-to-new ref created |
kvn@1286 | 3292 | return; |
kvn@1286 | 3293 | } |
kvn@1286 | 3294 | |
ysr@1462 | 3295 | if (use_ReduceInitialCardMarks() |
ysr@1462 | 3296 | && obj == just_allocated_object(control())) { |
ysr@1462 | 3297 | // We can skip marks on a freshly-allocated object in Eden. |
ysr@1601 | 3298 | // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. |
ysr@1601 | 3299 | // That routine informs GC to take appropriate compensating steps, |
ysr@1601 | 3300 | // upon a slow-path allocation, so as to make this card-mark |
ysr@1601 | 3301 | // elision safe. |
ysr@1462 | 3302 | return; |
ysr@1462 | 3303 | } |
ysr@1462 | 3304 | |
kvn@1286 | 3305 | if (!use_precise) { |
kvn@1286 | 3306 | // All card marks for a (non-array) instance are in one place: |
kvn@1286 | 3307 | adr = obj; |
kvn@1286 | 3308 | } |
kvn@1286 | 3309 | // (Else it's an array (or unknown), and we want more precise card marks.) |
kvn@1286 | 3310 | assert(adr != NULL, ""); |
kvn@1286 | 3311 | |
kvn@1286 | 3312 | IdealKit ideal(gvn(), control(), merged_memory(), true); |
kvn@1286 | 3313 | |
kvn@1286 | 3314 | // Convert the pointer to an int prior to doing math on it |
kvn@1286 | 3315 | Node* cast = __ CastPX(__ ctrl(), adr); |
kvn@1286 | 3316 | |
kvn@1286 | 3317 | // Divide by card size |
kvn@1286 | 3318 | assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, |
kvn@1286 | 3319 | "Only one we handle so far."); |
kvn@1286 | 3320 | Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); |
kvn@1286 | 3321 | |
kvn@1286 | 3322 | // Combine card table base and card offset |
kvn@1286 | 3323 | Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); |
kvn@1286 | 3324 | |
kvn@1286 | 3325 | // Get the alias_index for raw card-mark memory |
kvn@1286 | 3326 | int adr_type = Compile::AliasIdxRaw; |
kvn@1286 | 3327 | // Smash zero into card |
kvn@1286 | 3328 | Node* zero = __ ConI(0); |
kvn@1286 | 3329 | BasicType bt = T_BYTE; |
kvn@1286 | 3330 | if( !UseConcMarkSweepGC ) { |
kvn@1286 | 3331 | __ store(__ ctrl(), card_adr, zero, bt, adr_type); |
kvn@1286 | 3332 | } else { |
kvn@1286 | 3333 | // Specialized path for CM store barrier |
cfang@1420 | 3334 | __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); |
kvn@1286 | 3335 | } |
kvn@1286 | 3336 | |
kvn@1286 | 3337 | // Final sync IdealKit and GraphKit. |
kvn@1286 | 3338 | sync_kit(ideal); |
kvn@1286 | 3339 | } |
kvn@1286 | 3340 | |
kvn@1286 | 3341 | // G1 pre/post barriers |
ysr@777 | 3342 | void GraphKit::g1_write_barrier_pre(Node* obj, |
ysr@777 | 3343 | Node* adr, |
ysr@777 | 3344 | uint alias_idx, |
ysr@777 | 3345 | Node* val, |
never@1260 | 3346 | const TypeOopPtr* val_type, |
ysr@777 | 3347 | BasicType bt) { |
ysr@777 | 3348 | IdealKit ideal(gvn(), control(), merged_memory(), true); |
kvn@1286 | 3349 | |
kvn@1286 | 3350 | Node* tls = __ thread(); // ThreadLocalStorage |
ysr@777 | 3351 | |
ysr@777 | 3352 | Node* no_ctrl = NULL; |
ysr@777 | 3353 | Node* no_base = __ top(); |
ysr@777 | 3354 | Node* zero = __ ConI(0); |
ysr@777 | 3355 | |
ysr@777 | 3356 | float likely = PROB_LIKELY(0.999); |
ysr@777 | 3357 | float unlikely = PROB_UNLIKELY(0.999); |
ysr@777 | 3358 | |
ysr@777 | 3359 | BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; |
ysr@777 | 3360 | assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); |
ysr@777 | 3361 | |
ysr@777 | 3362 | // Offsets into the thread |
ysr@777 | 3363 | const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 |
ysr@777 | 3364 | PtrQueue::byte_offset_of_active()); |
ysr@777 | 3365 | const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 |
ysr@777 | 3366 | PtrQueue::byte_offset_of_index()); |
ysr@777 | 3367 | const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 |
ysr@777 | 3368 | PtrQueue::byte_offset_of_buf()); |
ysr@777 | 3369 | // Now the actual pointers into the thread |
ysr@777 | 3370 | |
ysr@777 | 3371 | // set_control( ctl); |
ysr@777 | 3372 | |
kvn@1286 | 3373 | Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); |
kvn@1286 | 3374 | Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); |
kvn@1286 | 3375 | Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); |
ysr@777 | 3376 | |
ysr@777 | 3377 | // Now some of the values |
ysr@777 | 3378 | |
never@979 | 3379 | Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); |
ysr@777 | 3380 | |
ysr@777 | 3381 | // if (!marking) |
ysr@777 | 3382 | __ if_then(marking, BoolTest::ne, zero); { |
never@979 | 3383 | Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); |
ysr@777 | 3384 | |
ysr@777 | 3385 | const Type* t1 = adr->bottom_type(); |
ysr@777 | 3386 | const Type* t2 = val->bottom_type(); |
ysr@777 | 3387 | |
ysr@777 | 3388 | Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx); |
ysr@777 | 3389 | // if (orig != NULL) |
ysr@777 | 3390 | __ if_then(orig, BoolTest::ne, null()); { |
never@979 | 3391 | Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); |
ysr@777 | 3392 | |
ysr@777 | 3393 | // load original value |
ysr@777 | 3394 | // alias_idx correct?? |
ysr@777 | 3395 | |
ysr@777 | 3396 | // is the queue for this thread full? |
ysr@777 | 3397 | __ if_then(index, BoolTest::ne, zero, likely); { |
ysr@777 | 3398 | |
ysr@777 | 3399 | // decrement the index |
ysr@777 | 3400 | Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); |
ysr@777 | 3401 | Node* next_indexX = next_index; |
ysr@777 | 3402 | #ifdef _LP64 |
kvn@1286 | 3403 | // We could refine the type for what it's worth |
kvn@1286 | 3404 | // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); |
kvn@1286 | 3405 | next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); |
kvn@1286 | 3406 | #endif |
ysr@777 | 3407 | |
ysr@777 | 3408 | // Now get the buffer location we will log the original value into and store it |
ysr@777 | 3409 | Node *log_addr = __ AddP(no_base, buffer, next_indexX); |
ysr@777 | 3410 | __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw); |
ysr@777 | 3411 | |
ysr@777 | 3412 | // update the index |
kvn@1286 | 3413 | __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); |
ysr@777 | 3414 | |
ysr@777 | 3415 | } __ else_(); { |
ysr@777 | 3416 | |
ysr@777 | 3417 | // logging buffer is full, call the runtime |
ysr@777 | 3418 | const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); |
kvn@1286 | 3419 | __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls); |
kvn@1286 | 3420 | } __ end_if(); // (!index) |
kvn@1286 | 3421 | } __ end_if(); // (orig != NULL) |
kvn@1286 | 3422 | } __ end_if(); // (!marking) |
kvn@1286 | 3423 | |
kvn@1286 | 3424 | // Final sync IdealKit and GraphKit. |
kvn@1286 | 3425 | sync_kit(ideal); |
ysr@777 | 3426 | } |
ysr@777 | 3427 | |
ysr@777 | 3428 | // |
ysr@777 | 3429 | // Update the card table and add card address to the queue |
ysr@777 | 3430 | // |
kvn@1286 | 3431 | void GraphKit::g1_mark_card(IdealKit& ideal, |
kvn@1286 | 3432 | Node* card_adr, |
kvn@1286 | 3433 | Node* oop_store, |
cfang@1420 | 3434 | uint oop_alias_idx, |
kvn@1286 | 3435 | Node* index, |
kvn@1286 | 3436 | Node* index_adr, |
kvn@1286 | 3437 | Node* buffer, |
kvn@1286 | 3438 | const TypeFunc* tf) { |
kvn@1286 | 3439 | |
ysr@777 | 3440 | Node* zero = __ ConI(0); |
ysr@777 | 3441 | Node* no_base = __ top(); |
ysr@777 | 3442 | BasicType card_bt = T_BYTE; |
ysr@777 | 3443 | // Smash zero into card. MUST BE ORDERED WRT TO STORE |
cfang@1420 | 3444 | __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); |
ysr@777 | 3445 | |
ysr@777 | 3446 | // Now do the queue work |
ysr@777 | 3447 | __ if_then(index, BoolTest::ne, zero); { |
ysr@777 | 3448 | |
kvn@1286 | 3449 | Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); |
ysr@777 | 3450 | Node* next_indexX = next_index; |
ysr@777 | 3451 | #ifdef _LP64 |
ysr@777 | 3452 | // We could refine the type for what it's worth |
ysr@777 | 3453 | // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); |
ysr@777 | 3454 | next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); |
ysr@777 | 3455 | #endif // _LP64 |
ysr@777 | 3456 | Node* log_addr = __ AddP(no_base, buffer, next_indexX); |
ysr@777 | 3457 | |
ysr@777 | 3458 | __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); |
ysr@777 | 3459 | __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); |
ysr@777 | 3460 | |
ysr@777 | 3461 | } __ else_(); { |
ysr@777 | 3462 | __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); |
ysr@777 | 3463 | } __ end_if(); |
kvn@1286 | 3464 | |
ysr@777 | 3465 | } |
ysr@777 | 3466 | |
kvn@1286 | 3467 | void GraphKit::g1_write_barrier_post(Node* oop_store, |
ysr@777 | 3468 | Node* obj, |
ysr@777 | 3469 | Node* adr, |
ysr@777 | 3470 | uint alias_idx, |
ysr@777 | 3471 | Node* val, |
ysr@777 | 3472 | BasicType bt, |
ysr@777 | 3473 | bool use_precise) { |
ysr@777 | 3474 | // If we are writing a NULL then we need no post barrier |
ysr@777 | 3475 | |
ysr@777 | 3476 | if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { |
ysr@777 | 3477 | // Must be NULL |
ysr@777 | 3478 | const Type* t = val->bottom_type(); |
ysr@777 | 3479 | assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); |
ysr@777 | 3480 | // No post barrier if writing NULLx |
ysr@777 | 3481 | return; |
ysr@777 | 3482 | } |
ysr@777 | 3483 | |
ysr@777 | 3484 | if (!use_precise) { |
ysr@777 | 3485 | // All card marks for a (non-array) instance are in one place: |
ysr@777 | 3486 | adr = obj; |
ysr@777 | 3487 | } |
ysr@777 | 3488 | // (Else it's an array (or unknown), and we want more precise card marks.) |
ysr@777 | 3489 | assert(adr != NULL, ""); |
ysr@777 | 3490 | |
ysr@777 | 3491 | IdealKit ideal(gvn(), control(), merged_memory(), true); |
kvn@1286 | 3492 | |
kvn@1286 | 3493 | Node* tls = __ thread(); // ThreadLocalStorage |
ysr@777 | 3494 | |
ysr@777 | 3495 | Node* no_ctrl = NULL; |
ysr@777 | 3496 | Node* no_base = __ top(); |
ysr@777 | 3497 | float likely = PROB_LIKELY(0.999); |
ysr@777 | 3498 | float unlikely = PROB_UNLIKELY(0.999); |
ysr@777 | 3499 | Node* zero = __ ConI(0); |
ysr@777 | 3500 | Node* zeroX = __ ConX(0); |
ysr@777 | 3501 | |
ysr@777 | 3502 | // Get the alias_index for raw card-mark memory |
ysr@777 | 3503 | const TypePtr* card_type = TypeRawPtr::BOTTOM; |
ysr@777 | 3504 | |
ysr@777 | 3505 | const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); |
ysr@777 | 3506 | |
ysr@777 | 3507 | // Offsets into the thread |
ysr@777 | 3508 | const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + |
ysr@777 | 3509 | PtrQueue::byte_offset_of_index()); |
ysr@777 | 3510 | const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + |
ysr@777 | 3511 | PtrQueue::byte_offset_of_buf()); |
ysr@777 | 3512 | |
ysr@777 | 3513 | // Pointers into the thread |
ysr@777 | 3514 | |
kvn@1286 | 3515 | Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); |
kvn@1286 | 3516 | Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); |
ysr@777 | 3517 | |
ysr@777 | 3518 | // Now some values |
ysr@777 | 3519 | |
ysr@777 | 3520 | Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); |
ysr@777 | 3521 | Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); |
ysr@777 | 3522 | |
ysr@777 | 3523 | |
ysr@777 | 3524 | // Convert the store obj pointer to an int prior to doing math on it |
ysr@777 | 3525 | // Must use ctrl to prevent "integerized oop" existing across safepoint |
kvn@1286 | 3526 | Node* cast = __ CastPX(__ ctrl(), adr); |
ysr@777 | 3527 | |
ysr@777 | 3528 | // Divide pointer by card size |
ysr@777 | 3529 | Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); |
ysr@777 | 3530 | |
ysr@777 | 3531 | // Combine card table base and card offset |
kvn@1286 | 3532 | Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); |
ysr@777 | 3533 | |
ysr@777 | 3534 | // If we know the value being stored does it cross regions? |
ysr@777 | 3535 | |
ysr@777 | 3536 | if (val != NULL) { |
ysr@777 | 3537 | // Does the store cause us to cross regions? |
ysr@777 | 3538 | |
ysr@777 | 3539 | // Should be able to do an unsigned compare of region_size instead of |
ysr@777 | 3540 | // and extra shift. Do we have an unsigned compare?? |
ysr@777 | 3541 | // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); |
ysr@777 | 3542 | Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); |
ysr@777 | 3543 | |
ysr@777 | 3544 | // if (xor_res == 0) same region so skip |
ysr@777 | 3545 | __ if_then(xor_res, BoolTest::ne, zeroX); { |
ysr@777 | 3546 | |
ysr@777 | 3547 | // No barrier if we are storing a NULL |
ysr@777 | 3548 | __ if_then(val, BoolTest::ne, null(), unlikely); { |
ysr@777 | 3549 | |
ysr@777 | 3550 | // Ok must mark the card if not already dirty |
ysr@777 | 3551 | |
ysr@777 | 3552 | // load the original value of the card |
ysr@777 | 3553 | Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); |
ysr@777 | 3554 | |
ysr@777 | 3555 | __ if_then(card_val, BoolTest::ne, zero); { |
cfang@1420 | 3556 | g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); |
ysr@777 | 3557 | } __ end_if(); |
ysr@777 | 3558 | } __ end_if(); |
ysr@777 | 3559 | } __ end_if(); |
ysr@777 | 3560 | } else { |
kvn@1286 | 3561 | // Object.clone() instrinsic uses this path. |
cfang@1420 | 3562 | g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); |
ysr@777 | 3563 | } |
ysr@777 | 3564 | |
kvn@1286 | 3565 | // Final sync IdealKit and GraphKit. |
kvn@1286 | 3566 | sync_kit(ideal); |
kvn@1286 | 3567 | } |
ysr@777 | 3568 | #undef __ |