Mercurial > hg > truffle
annotate src/share/vm/opto/graphKit.cpp @ 196:d1605aabd0a1 jdk7-b30
6719955: Update copyright year
Summary: Update copyright year for files that have been modified in 2008
Reviewed-by: ohair, tbell
| author | xdono |
|---|---|
| date | Wed, 02 Jul 2008 12:55:16 -0700 |
| parents | c436414a719e |
| children | c792b641b8bd 1ee8caae33af |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 196 | 2 * Copyright 2001-2008 Sun Microsystems, Inc. All Rights Reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
| 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 21 * have any questions. | |
| 22 * | |
| 23 */ | |
| 24 | |
| 25 #include "incls/_precompiled.incl" | |
| 26 #include "incls/_graphKit.cpp.incl" | |
| 27 | |
| 28 //----------------------------GraphKit----------------------------------------- | |
| 29 // Main utility constructor. | |
| 30 GraphKit::GraphKit(JVMState* jvms) | |
| 31 : Phase(Phase::Parser), | |
| 32 _env(C->env()), | |
| 33 _gvn(*C->initial_gvn()) | |
| 34 { | |
| 35 _exceptions = jvms->map()->next_exception(); | |
| 36 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); | |
| 37 set_jvms(jvms); | |
| 38 } | |
| 39 | |
| 40 // Private constructor for parser. | |
| 41 GraphKit::GraphKit() | |
| 42 : Phase(Phase::Parser), | |
| 43 _env(C->env()), | |
| 44 _gvn(*C->initial_gvn()) | |
| 45 { | |
| 46 _exceptions = NULL; | |
| 47 set_map(NULL); | |
| 48 debug_only(_sp = -99); | |
| 49 debug_only(set_bci(-99)); | |
| 50 } | |
| 51 | |
| 52 | |
| 53 | |
| 54 //---------------------------clean_stack--------------------------------------- | |
| 55 // Clear away rubbish from the stack area of the JVM state. | |
| 56 // This destroys any arguments that may be waiting on the stack. | |
| 57 void GraphKit::clean_stack(int from_sp) { | |
| 58 SafePointNode* map = this->map(); | |
| 59 JVMState* jvms = this->jvms(); | |
| 60 int stk_size = jvms->stk_size(); | |
| 61 int stkoff = jvms->stkoff(); | |
| 62 Node* top = this->top(); | |
| 63 for (int i = from_sp; i < stk_size; i++) { | |
| 64 if (map->in(stkoff + i) != top) { | |
| 65 map->set_req(stkoff + i, top); | |
| 66 } | |
| 67 } | |
| 68 } | |
| 69 | |
| 70 | |
| 71 //--------------------------------sync_jvms----------------------------------- | |
| 72 // Make sure our current jvms agrees with our parse state. | |
| 73 JVMState* GraphKit::sync_jvms() const { | |
| 74 JVMState* jvms = this->jvms(); | |
| 75 jvms->set_bci(bci()); // Record the new bci in the JVMState | |
| 76 jvms->set_sp(sp()); // Record the new sp in the JVMState | |
| 77 assert(jvms_in_sync(), "jvms is now in sync"); | |
| 78 return jvms; | |
| 79 } | |
| 80 | |
| 81 #ifdef ASSERT | |
| 82 bool GraphKit::jvms_in_sync() const { | |
| 83 Parse* parse = is_Parse(); | |
| 84 if (parse == NULL) { | |
| 85 if (bci() != jvms()->bci()) return false; | |
| 86 if (sp() != (int)jvms()->sp()) return false; | |
| 87 return true; | |
| 88 } | |
| 89 if (jvms()->method() != parse->method()) return false; | |
| 90 if (jvms()->bci() != parse->bci()) return false; | |
| 91 int jvms_sp = jvms()->sp(); | |
| 92 if (jvms_sp != parse->sp()) return false; | |
| 93 int jvms_depth = jvms()->depth(); | |
| 94 if (jvms_depth != parse->depth()) return false; | |
| 95 return true; | |
| 96 } | |
| 97 | |
| 98 // Local helper checks for special internal merge points | |
| 99 // used to accumulate and merge exception states. | |
| 100 // They are marked by the region's in(0) edge being the map itself. | |
| 101 // Such merge points must never "escape" into the parser at large, | |
| 102 // until they have been handed to gvn.transform. | |
| 103 static bool is_hidden_merge(Node* reg) { | |
| 104 if (reg == NULL) return false; | |
| 105 if (reg->is_Phi()) { | |
| 106 reg = reg->in(0); | |
| 107 if (reg == NULL) return false; | |
| 108 } | |
| 109 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); | |
| 110 } | |
| 111 | |
| 112 void GraphKit::verify_map() const { | |
| 113 if (map() == NULL) return; // null map is OK | |
| 114 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); | |
| 115 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); | |
| 116 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); | |
| 117 } | |
| 118 | |
| 119 void GraphKit::verify_exception_state(SafePointNode* ex_map) { | |
| 120 assert(ex_map->next_exception() == NULL, "not already part of a chain"); | |
| 121 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); | |
| 122 } | |
| 123 #endif | |
| 124 | |
| 125 //---------------------------stop_and_kill_map--------------------------------- | |
| 126 // Set _map to NULL, signalling a stop to further bytecode execution. | |
| 127 // First smash the current map's control to a constant, to mark it dead. | |
| 128 void GraphKit::stop_and_kill_map() { | |
| 129 SafePointNode* dead_map = stop(); | |
| 130 if (dead_map != NULL) { | |
| 131 dead_map->disconnect_inputs(NULL); // Mark the map as killed. | |
| 132 assert(dead_map->is_killed(), "must be so marked"); | |
| 133 } | |
| 134 } | |
| 135 | |
| 136 | |
| 137 //--------------------------------stopped-------------------------------------- | |
| 138 // Tell if _map is NULL, or control is top. | |
| 139 bool GraphKit::stopped() { | |
| 140 if (map() == NULL) return true; | |
| 141 else if (control() == top()) return true; | |
| 142 else return false; | |
| 143 } | |
| 144 | |
| 145 | |
| 146 //-----------------------------has_ex_handler---------------------------------- | |
| 147 // Tell if this method or any caller method has exception handlers. | |
| 148 bool GraphKit::has_ex_handler() { | |
| 149 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { | |
| 150 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { | |
| 151 return true; | |
| 152 } | |
| 153 } | |
| 154 return false; | |
| 155 } | |
| 156 | |
| 157 //------------------------------save_ex_oop------------------------------------ | |
| 158 // Save an exception without blowing stack contents or other JVM state. | |
| 159 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { | |
| 160 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); | |
| 161 ex_map->add_req(ex_oop); | |
| 162 debug_only(verify_exception_state(ex_map)); | |
| 163 } | |
| 164 | |
| 165 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { | |
| 166 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); | |
| 167 Node* ex_oop = ex_map->in(ex_map->req()-1); | |
| 168 if (clear_it) ex_map->del_req(ex_map->req()-1); | |
| 169 return ex_oop; | |
| 170 } | |
| 171 | |
| 172 //-----------------------------saved_ex_oop------------------------------------ | |
| 173 // Recover a saved exception from its map. | |
| 174 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { | |
| 175 return common_saved_ex_oop(ex_map, false); | |
| 176 } | |
| 177 | |
| 178 //--------------------------clear_saved_ex_oop--------------------------------- | |
| 179 // Erase a previously saved exception from its map. | |
| 180 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { | |
| 181 return common_saved_ex_oop(ex_map, true); | |
| 182 } | |
| 183 | |
| 184 #ifdef ASSERT | |
| 185 //---------------------------has_saved_ex_oop---------------------------------- | |
| 186 // Erase a previously saved exception from its map. | |
| 187 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { | |
| 188 return ex_map->req() == ex_map->jvms()->endoff()+1; | |
| 189 } | |
| 190 #endif | |
| 191 | |
| 192 //-------------------------make_exception_state-------------------------------- | |
| 193 // Turn the current JVM state into an exception state, appending the ex_oop. | |
| 194 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { | |
| 195 sync_jvms(); | |
| 196 SafePointNode* ex_map = stop(); // do not manipulate this map any more | |
| 197 set_saved_ex_oop(ex_map, ex_oop); | |
| 198 return ex_map; | |
| 199 } | |
| 200 | |
| 201 | |
| 202 //--------------------------add_exception_state-------------------------------- | |
| 203 // Add an exception to my list of exceptions. | |
| 204 void GraphKit::add_exception_state(SafePointNode* ex_map) { | |
| 205 if (ex_map == NULL || ex_map->control() == top()) { | |
| 206 return; | |
| 207 } | |
| 208 #ifdef ASSERT | |
| 209 verify_exception_state(ex_map); | |
| 210 if (has_exceptions()) { | |
| 211 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); | |
| 212 } | |
| 213 #endif | |
| 214 | |
| 215 // If there is already an exception of exactly this type, merge with it. | |
| 216 // In particular, null-checks and other low-level exceptions common up here. | |
| 217 Node* ex_oop = saved_ex_oop(ex_map); | |
| 218 const Type* ex_type = _gvn.type(ex_oop); | |
| 219 if (ex_oop == top()) { | |
| 220 // No action needed. | |
| 221 return; | |
| 222 } | |
| 223 assert(ex_type->isa_instptr(), "exception must be an instance"); | |
| 224 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { | |
| 225 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); | |
| 226 // We check sp also because call bytecodes can generate exceptions | |
| 227 // both before and after arguments are popped! | |
| 228 if (ex_type2 == ex_type | |
| 229 && e2->_jvms->sp() == ex_map->_jvms->sp()) { | |
| 230 combine_exception_states(ex_map, e2); | |
| 231 return; | |
| 232 } | |
| 233 } | |
| 234 | |
| 235 // No pre-existing exception of the same type. Chain it on the list. | |
| 236 push_exception_state(ex_map); | |
| 237 } | |
| 238 | |
| 239 //-----------------------add_exception_states_from----------------------------- | |
| 240 void GraphKit::add_exception_states_from(JVMState* jvms) { | |
| 241 SafePointNode* ex_map = jvms->map()->next_exception(); | |
| 242 if (ex_map != NULL) { | |
| 243 jvms->map()->set_next_exception(NULL); | |
| 244 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { | |
| 245 next_map = ex_map->next_exception(); | |
| 246 ex_map->set_next_exception(NULL); | |
| 247 add_exception_state(ex_map); | |
| 248 } | |
| 249 } | |
| 250 } | |
| 251 | |
| 252 //-----------------------transfer_exceptions_into_jvms------------------------- | |
| 253 JVMState* GraphKit::transfer_exceptions_into_jvms() { | |
| 254 if (map() == NULL) { | |
| 255 // We need a JVMS to carry the exceptions, but the map has gone away. | |
| 256 // Create a scratch JVMS, cloned from any of the exception states... | |
| 257 if (has_exceptions()) { | |
| 258 _map = _exceptions; | |
| 259 _map = clone_map(); | |
| 260 _map->set_next_exception(NULL); | |
| 261 clear_saved_ex_oop(_map); | |
| 262 debug_only(verify_map()); | |
| 263 } else { | |
| 264 // ...or created from scratch | |
| 265 JVMState* jvms = new (C) JVMState(_method, NULL); | |
| 266 jvms->set_bci(_bci); | |
| 267 jvms->set_sp(_sp); | |
| 268 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms)); | |
| 269 set_jvms(jvms); | |
| 270 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); | |
| 271 set_all_memory(top()); | |
| 272 while (map()->req() < jvms->endoff()) map()->add_req(top()); | |
| 273 } | |
| 274 // (This is a kludge, in case you didn't notice.) | |
| 275 set_control(top()); | |
| 276 } | |
| 277 JVMState* jvms = sync_jvms(); | |
| 278 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); | |
| 279 jvms->map()->set_next_exception(_exceptions); | |
| 280 _exceptions = NULL; // done with this set of exceptions | |
| 281 return jvms; | |
| 282 } | |
| 283 | |
| 284 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { | |
| 285 assert(is_hidden_merge(dstphi), "must be a special merge node"); | |
| 286 assert(is_hidden_merge(srcphi), "must be a special merge node"); | |
| 287 uint limit = srcphi->req(); | |
| 288 for (uint i = PhiNode::Input; i < limit; i++) { | |
| 289 dstphi->add_req(srcphi->in(i)); | |
| 290 } | |
| 291 } | |
| 292 static inline void add_one_req(Node* dstphi, Node* src) { | |
| 293 assert(is_hidden_merge(dstphi), "must be a special merge node"); | |
| 294 assert(!is_hidden_merge(src), "must not be a special merge node"); | |
| 295 dstphi->add_req(src); | |
| 296 } | |
| 297 | |
| 298 //-----------------------combine_exception_states------------------------------ | |
| 299 // This helper function combines exception states by building phis on a | |
| 300 // specially marked state-merging region. These regions and phis are | |
| 301 // untransformed, and can build up gradually. The region is marked by | |
| 302 // having a control input of its exception map, rather than NULL. Such | |
| 303 // regions do not appear except in this function, and in use_exception_state. | |
| 304 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { | |
| 305 if (failing()) return; // dying anyway... | |
| 306 JVMState* ex_jvms = ex_map->_jvms; | |
| 307 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); | |
| 308 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); | |
| 309 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); | |
| 310 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); | |
| 311 assert(ex_map->req() == phi_map->req(), "matching maps"); | |
| 312 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); | |
| 313 Node* hidden_merge_mark = root(); | |
| 314 Node* region = phi_map->control(); | |
| 315 MergeMemNode* phi_mem = phi_map->merged_memory(); | |
| 316 MergeMemNode* ex_mem = ex_map->merged_memory(); | |
| 317 if (region->in(0) != hidden_merge_mark) { | |
| 318 // The control input is not (yet) a specially-marked region in phi_map. | |
| 319 // Make it so, and build some phis. | |
| 320 region = new (C, 2) RegionNode(2); | |
| 321 _gvn.set_type(region, Type::CONTROL); | |
| 322 region->set_req(0, hidden_merge_mark); // marks an internal ex-state | |
| 323 region->init_req(1, phi_map->control()); | |
| 324 phi_map->set_control(region); | |
| 325 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); | |
| 326 record_for_igvn(io_phi); | |
| 327 _gvn.set_type(io_phi, Type::ABIO); | |
| 328 phi_map->set_i_o(io_phi); | |
| 329 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { | |
| 330 Node* m = mms.memory(); | |
| 331 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); | |
| 332 record_for_igvn(m_phi); | |
| 333 _gvn.set_type(m_phi, Type::MEMORY); | |
| 334 mms.set_memory(m_phi); | |
| 335 } | |
| 336 } | |
| 337 | |
| 338 // Either or both of phi_map and ex_map might already be converted into phis. | |
| 339 Node* ex_control = ex_map->control(); | |
| 340 // if there is special marking on ex_map also, we add multiple edges from src | |
| 341 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); | |
| 342 // how wide was the destination phi_map, originally? | |
| 343 uint orig_width = region->req(); | |
| 344 | |
| 345 if (add_multiple) { | |
| 346 add_n_reqs(region, ex_control); | |
| 347 add_n_reqs(phi_map->i_o(), ex_map->i_o()); | |
| 348 } else { | |
| 349 // ex_map has no merges, so we just add single edges everywhere | |
| 350 add_one_req(region, ex_control); | |
| 351 add_one_req(phi_map->i_o(), ex_map->i_o()); | |
| 352 } | |
| 353 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { | |
| 354 if (mms.is_empty()) { | |
| 355 // get a copy of the base memory, and patch some inputs into it | |
| 356 const TypePtr* adr_type = mms.adr_type(C); | |
| 357 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); | |
| 358 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); | |
| 359 mms.set_memory(phi); | |
| 360 // Prepare to append interesting stuff onto the newly sliced phi: | |
| 361 while (phi->req() > orig_width) phi->del_req(phi->req()-1); | |
| 362 } | |
| 363 // Append stuff from ex_map: | |
| 364 if (add_multiple) { | |
| 365 add_n_reqs(mms.memory(), mms.memory2()); | |
| 366 } else { | |
| 367 add_one_req(mms.memory(), mms.memory2()); | |
| 368 } | |
| 369 } | |
| 370 uint limit = ex_map->req(); | |
| 371 for (uint i = TypeFunc::Parms; i < limit; i++) { | |
| 372 // Skip everything in the JVMS after tos. (The ex_oop follows.) | |
| 373 if (i == tos) i = ex_jvms->monoff(); | |
| 374 Node* src = ex_map->in(i); | |
| 375 Node* dst = phi_map->in(i); | |
| 376 if (src != dst) { | |
| 377 PhiNode* phi; | |
| 378 if (dst->in(0) != region) { | |
| 379 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); | |
| 380 record_for_igvn(phi); | |
| 381 _gvn.set_type(phi, phi->type()); | |
| 382 phi_map->set_req(i, dst); | |
| 383 // Prepare to append interesting stuff onto the new phi: | |
| 384 while (dst->req() > orig_width) dst->del_req(dst->req()-1); | |
| 385 } else { | |
| 386 assert(dst->is_Phi(), "nobody else uses a hidden region"); | |
| 387 phi = (PhiNode*)dst; | |
| 388 } | |
| 389 if (add_multiple && src->in(0) == ex_control) { | |
| 390 // Both are phis. | |
| 391 add_n_reqs(dst, src); | |
| 392 } else { | |
| 393 while (dst->req() < region->req()) add_one_req(dst, src); | |
| 394 } | |
| 395 const Type* srctype = _gvn.type(src); | |
| 396 if (phi->type() != srctype) { | |
| 397 const Type* dsttype = phi->type()->meet(srctype); | |
| 398 if (phi->type() != dsttype) { | |
| 399 phi->set_type(dsttype); | |
| 400 _gvn.set_type(phi, dsttype); | |
| 401 } | |
| 402 } | |
| 403 } | |
| 404 } | |
| 405 } | |
| 406 | |
| 407 //--------------------------use_exception_state-------------------------------- | |
| 408 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { | |
| 409 if (failing()) { stop(); return top(); } | |
| 410 Node* region = phi_map->control(); | |
| 411 Node* hidden_merge_mark = root(); | |
| 412 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); | |
| 413 Node* ex_oop = clear_saved_ex_oop(phi_map); | |
| 414 if (region->in(0) == hidden_merge_mark) { | |
| 415 // Special marking for internal ex-states. Process the phis now. | |
| 416 region->set_req(0, region); // now it's an ordinary region | |
| 417 set_jvms(phi_map->jvms()); // ...so now we can use it as a map | |
| 418 // Note: Setting the jvms also sets the bci and sp. | |
| 419 set_control(_gvn.transform(region)); | |
| 420 uint tos = jvms()->stkoff() + sp(); | |
| 421 for (uint i = 1; i < tos; i++) { | |
| 422 Node* x = phi_map->in(i); | |
| 423 if (x->in(0) == region) { | |
| 424 assert(x->is_Phi(), "expected a special phi"); | |
| 425 phi_map->set_req(i, _gvn.transform(x)); | |
| 426 } | |
| 427 } | |
| 428 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { | |
| 429 Node* x = mms.memory(); | |
| 430 if (x->in(0) == region) { | |
| 431 assert(x->is_Phi(), "nobody else uses a hidden region"); | |
| 432 mms.set_memory(_gvn.transform(x)); | |
| 433 } | |
| 434 } | |
| 435 if (ex_oop->in(0) == region) { | |
| 436 assert(ex_oop->is_Phi(), "expected a special phi"); | |
| 437 ex_oop = _gvn.transform(ex_oop); | |
| 438 } | |
| 439 } else { | |
| 440 set_jvms(phi_map->jvms()); | |
| 441 } | |
| 442 | |
| 443 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); | |
| 444 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); | |
| 445 return ex_oop; | |
| 446 } | |
| 447 | |
| 448 //---------------------------------java_bc------------------------------------- | |
| 449 Bytecodes::Code GraphKit::java_bc() const { | |
| 450 ciMethod* method = this->method(); | |
| 451 int bci = this->bci(); | |
| 452 if (method != NULL && bci != InvocationEntryBci) | |
| 453 return method->java_code_at_bci(bci); | |
| 454 else | |
| 455 return Bytecodes::_illegal; | |
| 456 } | |
| 457 | |
| 458 //------------------------------builtin_throw---------------------------------- | |
| 459 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { | |
| 460 bool must_throw = true; | |
| 461 | |
| 462 if (JvmtiExport::can_post_exceptions()) { | |
| 463 // Do not try anything fancy if we're notifying the VM on every throw. | |
| 464 // Cf. case Bytecodes::_athrow in parse2.cpp. | |
| 465 uncommon_trap(reason, Deoptimization::Action_none, | |
| 466 (ciKlass*)NULL, (char*)NULL, must_throw); | |
| 467 return; | |
| 468 } | |
| 469 | |
| 470 // If this particular condition has not yet happened at this | |
| 471 // bytecode, then use the uncommon trap mechanism, and allow for | |
| 472 // a future recompilation if several traps occur here. | |
| 473 // If the throw is hot, try to use a more complicated inline mechanism | |
| 474 // which keeps execution inside the compiled code. | |
| 475 bool treat_throw_as_hot = false; | |
| 476 ciMethodData* md = method()->method_data(); | |
| 477 | |
| 478 if (ProfileTraps) { | |
| 479 if (too_many_traps(reason)) { | |
| 480 treat_throw_as_hot = true; | |
| 481 } | |
| 482 // (If there is no MDO at all, assume it is early in | |
| 483 // execution, and that any deopts are part of the | |
| 484 // startup transient, and don't need to be remembered.) | |
| 485 | |
| 486 // Also, if there is a local exception handler, treat all throws | |
| 487 // as hot if there has been at least one in this method. | |
| 488 if (C->trap_count(reason) != 0 | |
| 489 && method()->method_data()->trap_count(reason) != 0 | |
| 490 && has_ex_handler()) { | |
| 491 treat_throw_as_hot = true; | |
| 492 } | |
| 493 } | |
| 494 | |
| 495 // If this throw happens frequently, an uncommon trap might cause | |
| 496 // a performance pothole. If there is a local exception handler, | |
| 497 // and if this particular bytecode appears to be deoptimizing often, | |
| 498 // let us handle the throw inline, with a preconstructed instance. | |
| 499 // Note: If the deopt count has blown up, the uncommon trap | |
| 500 // runtime is going to flush this nmethod, not matter what. | |
| 501 if (treat_throw_as_hot | |
| 502 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { | |
| 503 // If the throw is local, we use a pre-existing instance and | |
| 504 // punt on the backtrace. This would lead to a missing backtrace | |
| 505 // (a repeat of 4292742) if the backtrace object is ever asked | |
| 506 // for its backtrace. | |
| 507 // Fixing this remaining case of 4292742 requires some flavor of | |
| 508 // escape analysis. Leave that for the future. | |
| 509 ciInstance* ex_obj = NULL; | |
| 510 switch (reason) { | |
| 511 case Deoptimization::Reason_null_check: | |
| 512 ex_obj = env()->NullPointerException_instance(); | |
| 513 break; | |
| 514 case Deoptimization::Reason_div0_check: | |
| 515 ex_obj = env()->ArithmeticException_instance(); | |
| 516 break; | |
| 517 case Deoptimization::Reason_range_check: | |
| 518 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); | |
| 519 break; | |
| 520 case Deoptimization::Reason_class_check: | |
| 521 if (java_bc() == Bytecodes::_aastore) { | |
| 522 ex_obj = env()->ArrayStoreException_instance(); | |
| 523 } else { | |
| 524 ex_obj = env()->ClassCastException_instance(); | |
| 525 } | |
| 526 break; | |
| 527 } | |
| 528 if (failing()) { stop(); return; } // exception allocation might fail | |
| 529 if (ex_obj != NULL) { | |
| 530 // Cheat with a preallocated exception object. | |
| 531 if (C->log() != NULL) | |
| 532 C->log()->elem("hot_throw preallocated='1' reason='%s'", | |
| 533 Deoptimization::trap_reason_name(reason)); | |
| 534 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
113
diff
changeset
|
535 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); |
| 0 | 536 |
| 537 // Clear the detail message of the preallocated exception object. | |
| 538 // Weblogic sometimes mutates the detail message of exceptions | |
| 539 // using reflection. | |
| 540 int offset = java_lang_Throwable::get_detailMessage_offset(); | |
| 541 const TypePtr* adr_typ = ex_con->add_offset(offset); | |
| 542 | |
| 543 Node *adr = basic_plus_adr(ex_node, ex_node, offset); | |
| 544 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT); | |
| 545 | |
| 546 add_exception_state(make_exception_state(ex_node)); | |
| 547 return; | |
| 548 } | |
| 549 } | |
| 550 | |
| 551 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? | |
| 552 // It won't be much cheaper than bailing to the interp., since we'll | |
| 553 // have to pass up all the debug-info, and the runtime will have to | |
| 554 // create the stack trace. | |
| 555 | |
| 556 // Usual case: Bail to interpreter. | |
| 557 // Reserve the right to recompile if we haven't seen anything yet. | |
| 558 | |
| 559 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; | |
| 560 if (treat_throw_as_hot | |
| 561 && (method()->method_data()->trap_recompiled_at(bci()) | |
| 562 || C->too_many_traps(reason))) { | |
| 563 // We cannot afford to take more traps here. Suffer in the interpreter. | |
| 564 if (C->log() != NULL) | |
| 565 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", | |
| 566 Deoptimization::trap_reason_name(reason), | |
| 567 C->trap_count(reason)); | |
| 568 action = Deoptimization::Action_none; | |
| 569 } | |
| 570 | |
| 571 // "must_throw" prunes the JVM state to include only the stack, if there | |
| 572 // are no local exception handlers. This should cut down on register | |
| 573 // allocation time and code size, by drastically reducing the number | |
| 574 // of in-edges on the call to the uncommon trap. | |
| 575 | |
| 576 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); | |
| 577 } | |
| 578 | |
| 579 | |
| 580 //----------------------------PreserveJVMState--------------------------------- | |
| 581 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { | |
| 582 debug_only(kit->verify_map()); | |
| 583 _kit = kit; | |
| 584 _map = kit->map(); // preserve the map | |
| 585 _sp = kit->sp(); | |
| 586 kit->set_map(clone_map ? kit->clone_map() : NULL); | |
| 587 #ifdef ASSERT | |
| 588 _bci = kit->bci(); | |
| 589 Parse* parser = kit->is_Parse(); | |
| 590 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->pre_order(); | |
| 591 _block = block; | |
| 592 #endif | |
| 593 } | |
| 594 PreserveJVMState::~PreserveJVMState() { | |
| 595 GraphKit* kit = _kit; | |
| 596 #ifdef ASSERT | |
| 597 assert(kit->bci() == _bci, "bci must not shift"); | |
| 598 Parse* parser = kit->is_Parse(); | |
| 599 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->pre_order(); | |
| 600 assert(block == _block, "block must not shift"); | |
| 601 #endif | |
| 602 kit->set_map(_map); | |
| 603 kit->set_sp(_sp); | |
| 604 } | |
| 605 | |
| 606 | |
| 607 //-----------------------------BuildCutout------------------------------------- | |
| 608 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) | |
| 609 : PreserveJVMState(kit) | |
| 610 { | |
| 611 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); | |
| 612 SafePointNode* outer_map = _map; // preserved map is caller's | |
| 613 SafePointNode* inner_map = kit->map(); | |
| 614 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); | |
| 615 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) )); | |
| 616 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) )); | |
| 617 } | |
| 618 BuildCutout::~BuildCutout() { | |
| 619 GraphKit* kit = _kit; | |
| 620 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); | |
| 621 } | |
| 622 | |
| 623 | |
| 624 //------------------------------clone_map-------------------------------------- | |
| 625 // Implementation of PreserveJVMState | |
| 626 // | |
| 627 // Only clone_map(...) here. If this function is only used in the | |
| 628 // PreserveJVMState class we may want to get rid of this extra | |
| 629 // function eventually and do it all there. | |
| 630 | |
| 631 SafePointNode* GraphKit::clone_map() { | |
| 632 if (map() == NULL) return NULL; | |
| 633 | |
| 634 // Clone the memory edge first | |
| 635 Node* mem = MergeMemNode::make(C, map()->memory()); | |
| 636 gvn().set_type_bottom(mem); | |
| 637 | |
| 638 SafePointNode *clonemap = (SafePointNode*)map()->clone(); | |
| 639 JVMState* jvms = this->jvms(); | |
| 640 JVMState* clonejvms = jvms->clone_shallow(C); | |
| 641 clonemap->set_memory(mem); | |
| 642 clonemap->set_jvms(clonejvms); | |
| 643 clonejvms->set_map(clonemap); | |
| 644 record_for_igvn(clonemap); | |
| 645 gvn().set_type_bottom(clonemap); | |
| 646 return clonemap; | |
| 647 } | |
| 648 | |
| 649 | |
| 650 //-----------------------------set_map_clone----------------------------------- | |
| 651 void GraphKit::set_map_clone(SafePointNode* m) { | |
| 652 _map = m; | |
| 653 _map = clone_map(); | |
| 654 _map->set_next_exception(NULL); | |
| 655 debug_only(verify_map()); | |
| 656 } | |
| 657 | |
| 658 | |
| 659 //----------------------------kill_dead_locals--------------------------------- | |
| 660 // Detect any locals which are known to be dead, and force them to top. | |
| 661 void GraphKit::kill_dead_locals() { | |
| 662 // Consult the liveness information for the locals. If any | |
| 663 // of them are unused, then they can be replaced by top(). This | |
| 664 // should help register allocation time and cut down on the size | |
| 665 // of the deoptimization information. | |
| 666 | |
| 667 // This call is made from many of the bytecode handling | |
| 668 // subroutines called from the Big Switch in do_one_bytecode. | |
| 669 // Every bytecode which might include a slow path is responsible | |
| 670 // for killing its dead locals. The more consistent we | |
| 671 // are about killing deads, the fewer useless phis will be | |
| 672 // constructed for them at various merge points. | |
| 673 | |
| 674 // bci can be -1 (InvocationEntryBci). We return the entry | |
| 675 // liveness for the method. | |
| 676 | |
| 677 if (method() == NULL || method()->code_size() == 0) { | |
| 678 // We are building a graph for a call to a native method. | |
| 679 // All locals are live. | |
| 680 return; | |
| 681 } | |
| 682 | |
| 683 ResourceMark rm; | |
| 684 | |
| 685 // Consult the liveness information for the locals. If any | |
| 686 // of them are unused, then they can be replaced by top(). This | |
| 687 // should help register allocation time and cut down on the size | |
| 688 // of the deoptimization information. | |
| 689 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); | |
| 690 | |
| 691 int len = (int)live_locals.size(); | |
| 692 assert(len <= jvms()->loc_size(), "too many live locals"); | |
| 693 for (int local = 0; local < len; local++) { | |
| 694 if (!live_locals.at(local)) { | |
| 695 set_local(local, top()); | |
| 696 } | |
| 697 } | |
| 698 } | |
| 699 | |
| 700 #ifdef ASSERT | |
| 701 //-------------------------dead_locals_are_killed------------------------------ | |
| 702 // Return true if all dead locals are set to top in the map. | |
| 703 // Used to assert "clean" debug info at various points. | |
| 704 bool GraphKit::dead_locals_are_killed() { | |
| 705 if (method() == NULL || method()->code_size() == 0) { | |
| 706 // No locals need to be dead, so all is as it should be. | |
| 707 return true; | |
| 708 } | |
| 709 | |
| 710 // Make sure somebody called kill_dead_locals upstream. | |
| 711 ResourceMark rm; | |
| 712 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { | |
| 713 if (jvms->loc_size() == 0) continue; // no locals to consult | |
| 714 SafePointNode* map = jvms->map(); | |
| 715 ciMethod* method = jvms->method(); | |
| 716 int bci = jvms->bci(); | |
| 717 if (jvms == this->jvms()) { | |
| 718 bci = this->bci(); // it might not yet be synched | |
| 719 } | |
| 720 MethodLivenessResult live_locals = method->liveness_at_bci(bci); | |
| 721 int len = (int)live_locals.size(); | |
| 722 if (!live_locals.is_valid() || len == 0) | |
| 723 // This method is trivial, or is poisoned by a breakpoint. | |
| 724 return true; | |
| 725 assert(len == jvms->loc_size(), "live map consistent with locals map"); | |
| 726 for (int local = 0; local < len; local++) { | |
| 727 if (!live_locals.at(local) && map->local(jvms, local) != top()) { | |
| 728 if (PrintMiscellaneous && (Verbose || WizardMode)) { | |
| 729 tty->print_cr("Zombie local %d: ", local); | |
| 730 jvms->dump(); | |
| 731 } | |
| 732 return false; | |
| 733 } | |
| 734 } | |
| 735 } | |
| 736 return true; | |
| 737 } | |
| 738 | |
| 739 #endif //ASSERT | |
| 740 | |
| 741 // Helper function for adding JVMState and debug information to node | |
| 742 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { | |
| 743 // Add the safepoint edges to the call (or other safepoint). | |
| 744 | |
| 745 // Make sure dead locals are set to top. This | |
| 746 // should help register allocation time and cut down on the size | |
| 747 // of the deoptimization information. | |
| 748 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); | |
| 749 | |
| 750 // Walk the inline list to fill in the correct set of JVMState's | |
| 751 // Also fill in the associated edges for each JVMState. | |
| 752 | |
| 753 JVMState* youngest_jvms = sync_jvms(); | |
| 754 | |
| 755 // Do we need debug info here? If it is a SafePoint and this method | |
| 756 // cannot de-opt, then we do NOT need any debug info. | |
| 757 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint); | |
| 758 | |
| 759 // If we are guaranteed to throw, we can prune everything but the | |
| 760 // input to the current bytecode. | |
| 761 bool can_prune_locals = false; | |
| 762 uint stack_slots_not_pruned = 0; | |
| 763 int inputs = 0, depth = 0; | |
| 764 if (must_throw) { | |
| 765 assert(method() == youngest_jvms->method(), "sanity"); | |
| 766 if (compute_stack_effects(inputs, depth)) { | |
| 767 can_prune_locals = true; | |
| 768 stack_slots_not_pruned = inputs; | |
| 769 } | |
| 770 } | |
| 771 | |
| 772 if (JvmtiExport::can_examine_or_deopt_anywhere()) { | |
| 773 // At any safepoint, this method can get breakpointed, which would | |
| 774 // then require an immediate deoptimization. | |
| 775 full_info = true; | |
| 776 can_prune_locals = false; // do not prune locals | |
| 777 stack_slots_not_pruned = 0; | |
| 778 } | |
| 779 | |
| 780 // do not scribble on the input jvms | |
| 781 JVMState* out_jvms = youngest_jvms->clone_deep(C); | |
| 782 call->set_jvms(out_jvms); // Start jvms list for call node | |
| 783 | |
| 784 // Presize the call: | |
| 785 debug_only(uint non_debug_edges = call->req()); | |
| 786 call->add_req_batch(top(), youngest_jvms->debug_depth()); | |
| 787 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); | |
| 788 | |
| 789 // Set up edges so that the call looks like this: | |
| 790 // Call [state:] ctl io mem fptr retadr | |
| 791 // [parms:] parm0 ... parmN | |
| 792 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN | |
| 793 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] | |
| 794 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN | |
| 795 // Note that caller debug info precedes callee debug info. | |
| 796 | |
| 797 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: | |
| 798 uint debug_ptr = call->req(); | |
| 799 | |
| 800 // Loop over the map input edges associated with jvms, add them | |
| 801 // to the call node, & reset all offsets to match call node array. | |
| 802 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { | |
| 803 uint debug_end = debug_ptr; | |
| 804 uint debug_start = debug_ptr - in_jvms->debug_size(); | |
| 805 debug_ptr = debug_start; // back up the ptr | |
| 806 | |
| 807 uint p = debug_start; // walks forward in [debug_start, debug_end) | |
| 808 uint j, k, l; | |
| 809 SafePointNode* in_map = in_jvms->map(); | |
| 810 out_jvms->set_map(call); | |
| 811 | |
| 812 if (can_prune_locals) { | |
| 813 assert(in_jvms->method() == out_jvms->method(), "sanity"); | |
| 814 // If the current throw can reach an exception handler in this JVMS, | |
| 815 // then we must keep everything live that can reach that handler. | |
| 816 // As a quick and dirty approximation, we look for any handlers at all. | |
| 817 if (in_jvms->method()->has_exception_handlers()) { | |
| 818 can_prune_locals = false; | |
| 819 } | |
| 820 } | |
| 821 | |
| 822 // Add the Locals | |
| 823 k = in_jvms->locoff(); | |
| 824 l = in_jvms->loc_size(); | |
| 825 out_jvms->set_locoff(p); | |
| 826 if (full_info && !can_prune_locals) { | |
| 827 for (j = 0; j < l; j++) | |
| 828 call->set_req(p++, in_map->in(k+j)); | |
| 829 } else { | |
| 830 p += l; // already set to top above by add_req_batch | |
| 831 } | |
| 832 | |
| 833 // Add the Expression Stack | |
| 834 k = in_jvms->stkoff(); | |
| 835 l = in_jvms->sp(); | |
| 836 out_jvms->set_stkoff(p); | |
| 837 if (full_info && !can_prune_locals) { | |
| 838 for (j = 0; j < l; j++) | |
| 839 call->set_req(p++, in_map->in(k+j)); | |
| 840 } else if (can_prune_locals && stack_slots_not_pruned != 0) { | |
| 841 // Divide stack into {S0,...,S1}, where S0 is set to top. | |
| 842 uint s1 = stack_slots_not_pruned; | |
| 843 stack_slots_not_pruned = 0; // for next iteration | |
| 844 if (s1 > l) s1 = l; | |
| 845 uint s0 = l - s1; | |
| 846 p += s0; // skip the tops preinstalled by add_req_batch | |
| 847 for (j = s0; j < l; j++) | |
| 848 call->set_req(p++, in_map->in(k+j)); | |
| 849 } else { | |
| 850 p += l; // already set to top above by add_req_batch | |
| 851 } | |
| 852 | |
| 853 // Add the Monitors | |
| 854 k = in_jvms->monoff(); | |
| 855 l = in_jvms->mon_size(); | |
| 856 out_jvms->set_monoff(p); | |
| 857 for (j = 0; j < l; j++) | |
| 858 call->set_req(p++, in_map->in(k+j)); | |
| 859 | |
|
63
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
860 // Copy any scalar object fields. |
|
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
861 k = in_jvms->scloff(); |
|
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
862 l = in_jvms->scl_size(); |
|
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
863 out_jvms->set_scloff(p); |
|
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
864 for (j = 0; j < l; j++) |
|
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
865 call->set_req(p++, in_map->in(k+j)); |
|
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
866 |
| 0 | 867 // Finish the new jvms. |
| 868 out_jvms->set_endoff(p); | |
| 869 | |
| 870 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); | |
| 871 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); | |
| 872 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); | |
| 873 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); | |
|
63
eac007780a58
6671807: (Escape Analysis) Add new ideal node to represent the state of a scalarized object at a safepoint
kvn
parents:
29
diff
changeset
|
874 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); |
| 0 | 875 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); |
| 876 | |
| 877 // Update the two tail pointers in parallel. | |
| 878 out_jvms = out_jvms->caller(); | |
| 879 in_jvms = in_jvms->caller(); | |
| 880 } | |
| 881 | |
| 882 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); | |
| 883 | |
| 884 // Test the correctness of JVMState::debug_xxx accessors: | |
| 885 assert(call->jvms()->debug_start() == non_debug_edges, ""); | |
| 886 assert(call->jvms()->debug_end() == call->req(), ""); | |
| 887 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); | |
| 888 } | |
| 889 | |
| 890 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { | |
| 891 Bytecodes::Code code = java_bc(); | |
| 892 if (code == Bytecodes::_wide) { | |
| 893 code = method()->java_code_at_bci(bci() + 1); | |
| 894 } | |
| 895 | |
| 896 BasicType rtype = T_ILLEGAL; | |
| 897 int rsize = 0; | |
| 898 | |
| 899 if (code != Bytecodes::_illegal) { | |
| 900 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 | |
| 901 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V | |
| 902 if (rtype < T_CONFLICT) | |
| 903 rsize = type2size[rtype]; | |
| 904 } | |
| 905 | |
| 906 switch (code) { | |
| 907 case Bytecodes::_illegal: | |
| 908 return false; | |
| 909 | |
| 910 case Bytecodes::_ldc: | |
| 911 case Bytecodes::_ldc_w: | |
| 912 case Bytecodes::_ldc2_w: | |
| 913 inputs = 0; | |
| 914 break; | |
| 915 | |
| 916 case Bytecodes::_dup: inputs = 1; break; | |
| 917 case Bytecodes::_dup_x1: inputs = 2; break; | |
| 918 case Bytecodes::_dup_x2: inputs = 3; break; | |
| 919 case Bytecodes::_dup2: inputs = 2; break; | |
| 920 case Bytecodes::_dup2_x1: inputs = 3; break; | |
| 921 case Bytecodes::_dup2_x2: inputs = 4; break; | |
| 922 case Bytecodes::_swap: inputs = 2; break; | |
| 923 case Bytecodes::_arraylength: inputs = 1; break; | |
| 924 | |
| 925 case Bytecodes::_getstatic: | |
| 926 case Bytecodes::_putstatic: | |
| 927 case Bytecodes::_getfield: | |
| 928 case Bytecodes::_putfield: | |
| 929 { | |
| 930 bool is_get = (depth >= 0), is_static = (depth & 1); | |
| 931 bool ignore; | |
| 932 ciBytecodeStream iter(method()); | |
| 933 iter.reset_to_bci(bci()); | |
| 934 iter.next(); | |
| 935 ciField* field = iter.get_field(ignore); | |
| 936 int size = field->type()->size(); | |
| 937 inputs = (is_static ? 0 : 1); | |
| 938 if (is_get) { | |
| 939 depth = size - inputs; | |
| 940 } else { | |
| 941 inputs += size; // putxxx pops the value from the stack | |
| 942 depth = - inputs; | |
| 943 } | |
| 944 } | |
| 945 break; | |
| 946 | |
| 947 case Bytecodes::_invokevirtual: | |
| 948 case Bytecodes::_invokespecial: | |
| 949 case Bytecodes::_invokestatic: | |
| 950 case Bytecodes::_invokeinterface: | |
| 951 { | |
| 952 bool is_static = (depth == 0); | |
| 953 bool ignore; | |
| 954 ciBytecodeStream iter(method()); | |
| 955 iter.reset_to_bci(bci()); | |
| 956 iter.next(); | |
| 957 ciMethod* method = iter.get_method(ignore); | |
| 958 inputs = method->arg_size_no_receiver(); | |
| 959 if (!is_static) inputs += 1; | |
| 960 int size = method->return_type()->size(); | |
| 961 depth = size - inputs; | |
| 962 } | |
| 963 break; | |
| 964 | |
| 965 case Bytecodes::_multianewarray: | |
| 966 { | |
| 967 ciBytecodeStream iter(method()); | |
| 968 iter.reset_to_bci(bci()); | |
| 969 iter.next(); | |
| 970 inputs = iter.get_dimensions(); | |
| 971 assert(rsize == 1, ""); | |
| 972 depth = rsize - inputs; | |
| 973 } | |
| 974 break; | |
| 975 | |
| 976 case Bytecodes::_ireturn: | |
| 977 case Bytecodes::_lreturn: | |
| 978 case Bytecodes::_freturn: | |
| 979 case Bytecodes::_dreturn: | |
| 980 case Bytecodes::_areturn: | |
| 981 assert(rsize = -depth, ""); | |
| 982 inputs = rsize; | |
| 983 break; | |
| 984 | |
| 985 case Bytecodes::_jsr: | |
| 986 case Bytecodes::_jsr_w: | |
| 987 inputs = 0; | |
| 988 depth = 1; // S.B. depth=1, not zero | |
| 989 break; | |
| 990 | |
| 991 default: | |
| 992 // bytecode produces a typed result | |
| 993 inputs = rsize - depth; | |
| 994 assert(inputs >= 0, ""); | |
| 995 break; | |
| 996 } | |
| 997 | |
| 998 #ifdef ASSERT | |
| 999 // spot check | |
| 1000 int outputs = depth + inputs; | |
| 1001 assert(outputs >= 0, "sanity"); | |
| 1002 switch (code) { | |
| 1003 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; | |
| 1004 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; | |
| 1005 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; | |
| 1006 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; | |
| 1007 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; | |
| 1008 } | |
| 1009 #endif //ASSERT | |
| 1010 | |
| 1011 return true; | |
| 1012 } | |
| 1013 | |
| 1014 | |
| 1015 | |
| 1016 //------------------------------basic_plus_adr--------------------------------- | |
| 1017 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { | |
| 1018 // short-circuit a common case | |
| 1019 if (offset == intcon(0)) return ptr; | |
| 1020 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) ); | |
| 1021 } | |
| 1022 | |
| 1023 Node* GraphKit::ConvI2L(Node* offset) { | |
| 1024 // short-circuit a common case | |
| 1025 jint offset_con = find_int_con(offset, Type::OffsetBot); | |
| 1026 if (offset_con != Type::OffsetBot) { | |
| 1027 return longcon((long) offset_con); | |
| 1028 } | |
| 1029 return _gvn.transform( new (C, 2) ConvI2LNode(offset)); | |
| 1030 } | |
| 1031 Node* GraphKit::ConvL2I(Node* offset) { | |
| 1032 // short-circuit a common case | |
| 1033 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); | |
| 1034 if (offset_con != (jlong)Type::OffsetBot) { | |
| 1035 return intcon((int) offset_con); | |
| 1036 } | |
| 1037 return _gvn.transform( new (C, 2) ConvL2INode(offset)); | |
| 1038 } | |
| 1039 | |
| 1040 //-------------------------load_object_klass----------------------------------- | |
| 1041 Node* GraphKit::load_object_klass(Node* obj) { | |
| 1042 // Special-case a fresh allocation to avoid building nodes: | |
| 1043 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); | |
| 1044 if (akls != NULL) return akls; | |
| 1045 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
113
diff
changeset
|
1046 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); |
| 0 | 1047 } |
| 1048 | |
| 1049 //-------------------------load_array_length----------------------------------- | |
| 1050 Node* GraphKit::load_array_length(Node* array) { | |
| 1051 // Special-case a fresh allocation to avoid building nodes: | |
| 1052 Node* alen = AllocateArrayNode::Ideal_length(array, &_gvn); | |
| 1053 if (alen != NULL) return alen; | |
| 1054 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); | |
| 1055 return _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); | |
| 1056 } | |
| 1057 | |
| 1058 //------------------------------do_null_check---------------------------------- | |
| 1059 // Helper function to do a NULL pointer check. Returned value is | |
| 1060 // the incoming address with NULL casted away. You are allowed to use the | |
| 1061 // not-null value only if you are control dependent on the test. | |
| 1062 extern int explicit_null_checks_inserted, | |
| 1063 explicit_null_checks_elided; | |
| 1064 Node* GraphKit::null_check_common(Node* value, BasicType type, | |
| 1065 // optional arguments for variations: | |
| 1066 bool assert_null, | |
| 1067 Node* *null_control) { | |
| 1068 assert(!assert_null || null_control == NULL, "not both at once"); | |
| 1069 if (stopped()) return top(); | |
| 1070 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { | |
| 1071 // For some performance testing, we may wish to suppress null checking. | |
| 1072 value = cast_not_null(value); // Make it appear to be non-null (4962416). | |
| 1073 return value; | |
| 1074 } | |
| 1075 explicit_null_checks_inserted++; | |
| 1076 | |
| 1077 // Construct NULL check | |
| 1078 Node *chk = NULL; | |
| 1079 switch(type) { | |
| 1080 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break; | |
| 1081 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break; | |
| 1082 case T_ARRAY : // fall through | |
| 1083 type = T_OBJECT; // simplify further tests | |
| 1084 case T_OBJECT : { | |
| 1085 const Type *t = _gvn.type( value ); | |
| 1086 | |
| 1087 const TypeInstPtr* tp = t->isa_instptr(); | |
| 1088 if (tp != NULL && !tp->klass()->is_loaded() | |
| 1089 // Only for do_null_check, not any of its siblings: | |
| 1090 && !assert_null && null_control == NULL) { | |
| 1091 // Usually, any field access or invocation on an unloaded oop type | |
| 1092 // will simply fail to link, since the statically linked class is | |
| 1093 // likely also to be unloaded. However, in -Xcomp mode, sometimes | |
| 1094 // the static class is loaded but the sharper oop type is not. | |
| 1095 // Rather than checking for this obscure case in lots of places, | |
| 1096 // we simply observe that a null check on an unloaded class | |
| 1097 // will always be followed by a nonsense operation, so we | |
| 1098 // can just issue the uncommon trap here. | |
| 1099 // Our access to the unloaded class will only be correct | |
| 1100 // after it has been loaded and initialized, which requires | |
| 1101 // a trip through the interpreter. | |
| 1102 #ifndef PRODUCT | |
| 1103 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } | |
| 1104 #endif | |
| 1105 uncommon_trap(Deoptimization::Reason_unloaded, | |
| 1106 Deoptimization::Action_reinterpret, | |
| 1107 tp->klass(), "!loaded"); | |
| 1108 return top(); | |
| 1109 } | |
| 1110 | |
| 1111 if (assert_null) { | |
| 1112 // See if the type is contained in NULL_PTR. | |
| 1113 // If so, then the value is already null. | |
| 1114 if (t->higher_equal(TypePtr::NULL_PTR)) { | |
| 1115 explicit_null_checks_elided++; | |
| 1116 return value; // Elided null assert quickly! | |
| 1117 } | |
| 1118 } else { | |
| 1119 // See if mixing in the NULL pointer changes type. | |
| 1120 // If so, then the NULL pointer was not allowed in the original | |
| 1121 // type. In other words, "value" was not-null. | |
| 1122 if (t->meet(TypePtr::NULL_PTR) != t) { | |
| 1123 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... | |
| 1124 explicit_null_checks_elided++; | |
| 1125 return value; // Elided null check quickly! | |
| 1126 } | |
| 1127 } | |
| 1128 chk = new (C, 3) CmpPNode( value, null() ); | |
| 1129 break; | |
| 1130 } | |
| 1131 | |
| 1132 default : ShouldNotReachHere(); | |
| 1133 } | |
| 1134 assert(chk != NULL, "sanity check"); | |
| 1135 chk = _gvn.transform(chk); | |
| 1136 | |
| 1137 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; | |
| 1138 BoolNode *btst = new (C, 2) BoolNode( chk, btest); | |
| 1139 Node *tst = _gvn.transform( btst ); | |
| 1140 | |
| 1141 //----------- | |
| 1142 // if peephole optimizations occured, a prior test existed. | |
| 1143 // If a prior test existed, maybe it dominates as we can avoid this test. | |
| 1144 if (tst != btst && type == T_OBJECT) { | |
| 1145 // At this point we want to scan up the CFG to see if we can | |
| 1146 // find an identical test (and so avoid this test altogether). | |
| 1147 Node *cfg = control(); | |
| 1148 int depth = 0; | |
| 1149 while( depth < 16 ) { // Limit search depth for speed | |
| 1150 if( cfg->Opcode() == Op_IfTrue && | |
| 1151 cfg->in(0)->in(1) == tst ) { | |
| 1152 // Found prior test. Use "cast_not_null" to construct an identical | |
| 1153 // CastPP (and hence hash to) as already exists for the prior test. | |
| 1154 // Return that casted value. | |
| 1155 if (assert_null) { | |
| 1156 replace_in_map(value, null()); | |
| 1157 return null(); // do not issue the redundant test | |
| 1158 } | |
| 1159 Node *oldcontrol = control(); | |
| 1160 set_control(cfg); | |
| 1161 Node *res = cast_not_null(value); | |
| 1162 set_control(oldcontrol); | |
| 1163 explicit_null_checks_elided++; | |
| 1164 return res; | |
| 1165 } | |
| 1166 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); | |
| 1167 if (cfg == NULL) break; // Quit at region nodes | |
| 1168 depth++; | |
| 1169 } | |
| 1170 } | |
| 1171 | |
| 1172 //----------- | |
| 1173 // Branch to failure if null | |
| 1174 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen | |
| 1175 Deoptimization::DeoptReason reason; | |
| 1176 if (assert_null) | |
| 1177 reason = Deoptimization::Reason_null_assert; | |
| 1178 else if (type == T_OBJECT) | |
| 1179 reason = Deoptimization::Reason_null_check; | |
| 1180 else | |
| 1181 reason = Deoptimization::Reason_div0_check; | |
| 1182 | |
| 1183 // To cause an implicit null check, we set the not-null probability | |
| 1184 // to the maximum (PROB_MAX). For an explicit check the probablity | |
| 1185 // is set to a smaller value. | |
| 1186 if (null_control != NULL || too_many_traps(reason)) { | |
| 1187 // probability is less likely | |
| 1188 ok_prob = PROB_LIKELY_MAG(3); | |
| 1189 } else if (!assert_null && | |
| 1190 (ImplicitNullCheckThreshold > 0) && | |
| 1191 method() != NULL && | |
| 1192 (method()->method_data()->trap_count(reason) | |
| 1193 >= (uint)ImplicitNullCheckThreshold)) { | |
| 1194 ok_prob = PROB_LIKELY_MAG(3); | |
| 1195 } | |
| 1196 | |
| 1197 if (null_control != NULL) { | |
| 1198 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); | |
| 1199 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff)); | |
| 1200 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff))); | |
| 1201 if (null_true == top()) | |
| 1202 explicit_null_checks_elided++; | |
| 1203 (*null_control) = null_true; | |
| 1204 } else { | |
| 1205 BuildCutout unless(this, tst, ok_prob); | |
| 1206 // Check for optimizer eliding test at parse time | |
| 1207 if (stopped()) { | |
| 1208 // Failure not possible; do not bother making uncommon trap. | |
| 1209 explicit_null_checks_elided++; | |
| 1210 } else if (assert_null) { | |
| 1211 uncommon_trap(reason, | |
| 1212 Deoptimization::Action_make_not_entrant, | |
| 1213 NULL, "assert_null"); | |
| 1214 } else { | |
| 1215 builtin_throw(reason); | |
| 1216 } | |
| 1217 } | |
| 1218 | |
| 1219 // Must throw exception, fall-thru not possible? | |
| 1220 if (stopped()) { | |
| 1221 return top(); // No result | |
| 1222 } | |
| 1223 | |
| 1224 if (assert_null) { | |
| 1225 // Cast obj to null on this path. | |
| 1226 replace_in_map(value, zerocon(type)); | |
| 1227 return zerocon(type); | |
| 1228 } | |
| 1229 | |
| 1230 // Cast obj to not-null on this path, if there is no null_control. | |
| 1231 // (If there is a null_control, a non-null value may come back to haunt us.) | |
| 1232 if (type == T_OBJECT) { | |
| 1233 Node* cast = cast_not_null(value, false); | |
| 1234 if (null_control == NULL || (*null_control) == top()) | |
| 1235 replace_in_map(value, cast); | |
| 1236 value = cast; | |
| 1237 } | |
| 1238 | |
| 1239 return value; | |
| 1240 } | |
| 1241 | |
| 1242 | |
| 1243 //------------------------------cast_not_null---------------------------------- | |
| 1244 // Cast obj to not-null on this path | |
| 1245 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { | |
| 1246 const Type *t = _gvn.type(obj); | |
| 1247 const Type *t_not_null = t->join(TypePtr::NOTNULL); | |
| 1248 // Object is already not-null? | |
| 1249 if( t == t_not_null ) return obj; | |
| 1250 | |
| 1251 Node *cast = new (C, 2) CastPPNode(obj,t_not_null); | |
| 1252 cast->init_req(0, control()); | |
| 1253 cast = _gvn.transform( cast ); | |
| 1254 | |
| 1255 // Scan for instances of 'obj' in the current JVM mapping. | |
| 1256 // These instances are known to be not-null after the test. | |
| 1257 if (do_replace_in_map) | |
| 1258 replace_in_map(obj, cast); | |
| 1259 | |
| 1260 return cast; // Return casted value | |
| 1261 } | |
| 1262 | |
| 1263 | |
| 1264 //--------------------------replace_in_map------------------------------------- | |
| 1265 void GraphKit::replace_in_map(Node* old, Node* neww) { | |
| 1266 this->map()->replace_edge(old, neww); | |
| 1267 | |
| 1268 // Note: This operation potentially replaces any edge | |
| 1269 // on the map. This includes locals, stack, and monitors | |
| 1270 // of the current (innermost) JVM state. | |
| 1271 | |
| 1272 // We can consider replacing in caller maps. | |
| 1273 // The idea would be that an inlined function's null checks | |
| 1274 // can be shared with the entire inlining tree. | |
| 1275 // The expense of doing this is that the PreserveJVMState class | |
| 1276 // would have to preserve caller states too, with a deep copy. | |
| 1277 } | |
| 1278 | |
| 1279 | |
| 1280 | |
| 1281 //============================================================================= | |
| 1282 //--------------------------------memory--------------------------------------- | |
| 1283 Node* GraphKit::memory(uint alias_idx) { | |
| 1284 MergeMemNode* mem = merged_memory(); | |
| 1285 Node* p = mem->memory_at(alias_idx); | |
| 1286 _gvn.set_type(p, Type::MEMORY); // must be mapped | |
| 1287 return p; | |
| 1288 } | |
| 1289 | |
| 1290 //-----------------------------reset_memory------------------------------------ | |
| 1291 Node* GraphKit::reset_memory() { | |
| 1292 Node* mem = map()->memory(); | |
| 1293 // do not use this node for any more parsing! | |
| 1294 debug_only( map()->set_memory((Node*)NULL) ); | |
| 1295 return _gvn.transform( mem ); | |
| 1296 } | |
| 1297 | |
| 1298 //------------------------------set_all_memory--------------------------------- | |
| 1299 void GraphKit::set_all_memory(Node* newmem) { | |
| 1300 Node* mergemem = MergeMemNode::make(C, newmem); | |
| 1301 gvn().set_type_bottom(mergemem); | |
| 1302 map()->set_memory(mergemem); | |
| 1303 } | |
| 1304 | |
| 1305 //------------------------------set_all_memory_call---------------------------- | |
| 1306 void GraphKit::set_all_memory_call(Node* call) { | |
| 1307 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); | |
| 1308 set_all_memory(newmem); | |
| 1309 } | |
| 1310 | |
| 1311 //============================================================================= | |
| 1312 // | |
| 1313 // parser factory methods for MemNodes | |
| 1314 // | |
| 1315 // These are layered on top of the factory methods in LoadNode and StoreNode, | |
| 1316 // and integrate with the parser's memory state and _gvn engine. | |
| 1317 // | |
| 1318 | |
| 1319 // factory methods in "int adr_idx" | |
| 1320 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, | |
| 1321 int adr_idx, | |
| 1322 bool require_atomic_access) { | |
| 1323 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); | |
| 1324 const TypePtr* adr_type = NULL; // debug-mode-only argument | |
| 1325 debug_only(adr_type = C->get_adr_type(adr_idx)); | |
| 1326 Node* mem = memory(adr_idx); | |
| 1327 Node* ld; | |
| 1328 if (require_atomic_access && bt == T_LONG) { | |
| 1329 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); | |
| 1330 } else { | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
1331 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); |
| 0 | 1332 } |
| 1333 return _gvn.transform(ld); | |
| 1334 } | |
| 1335 | |
| 1336 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, | |
| 1337 int adr_idx, | |
| 1338 bool require_atomic_access) { | |
| 1339 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); | |
| 1340 const TypePtr* adr_type = NULL; | |
| 1341 debug_only(adr_type = C->get_adr_type(adr_idx)); | |
| 1342 Node *mem = memory(adr_idx); | |
| 1343 Node* st; | |
| 1344 if (require_atomic_access && bt == T_LONG) { | |
| 1345 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); | |
| 1346 } else { | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
1347 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); |
| 0 | 1348 } |
| 1349 st = _gvn.transform(st); | |
| 1350 set_memory(st, adr_idx); | |
| 1351 // Back-to-back stores can only remove intermediate store with DU info | |
| 1352 // so push on worklist for optimizer. | |
| 1353 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) | |
| 1354 record_for_igvn(st); | |
| 1355 | |
| 1356 return st; | |
| 1357 } | |
| 1358 | |
| 1359 void GraphKit::pre_barrier(Node* ctl, | |
| 1360 Node* obj, | |
| 1361 Node* adr, | |
| 1362 uint adr_idx, | |
| 1363 Node *val, | |
| 1364 const Type* val_type, | |
| 1365 BasicType bt) { | |
| 1366 BarrierSet* bs = Universe::heap()->barrier_set(); | |
| 1367 set_control(ctl); | |
| 1368 switch (bs->kind()) { | |
| 1369 | |
| 1370 case BarrierSet::CardTableModRef: | |
| 1371 case BarrierSet::CardTableExtension: | |
| 1372 case BarrierSet::ModRef: | |
| 1373 break; | |
| 1374 | |
| 1375 case BarrierSet::Other: | |
| 1376 default : | |
| 1377 ShouldNotReachHere(); | |
| 1378 | |
| 1379 } | |
| 1380 } | |
| 1381 | |
| 1382 void GraphKit::post_barrier(Node* ctl, | |
| 1383 Node* store, | |
| 1384 Node* obj, | |
| 1385 Node* adr, | |
| 1386 uint adr_idx, | |
| 1387 Node *val, | |
| 1388 BasicType bt, | |
| 1389 bool use_precise) { | |
| 1390 BarrierSet* bs = Universe::heap()->barrier_set(); | |
| 1391 set_control(ctl); | |
| 1392 switch (bs->kind()) { | |
| 1393 | |
| 1394 case BarrierSet::CardTableModRef: | |
| 1395 case BarrierSet::CardTableExtension: | |
| 1396 write_barrier_post(store, obj, adr, val, use_precise); | |
| 1397 break; | |
| 1398 | |
| 1399 case BarrierSet::ModRef: | |
| 1400 break; | |
| 1401 | |
| 1402 case BarrierSet::Other: | |
| 1403 default : | |
| 1404 ShouldNotReachHere(); | |
| 1405 | |
| 1406 } | |
| 1407 } | |
| 1408 | |
| 1409 Node* GraphKit::store_oop_to_object(Node* ctl, | |
| 1410 Node* obj, | |
| 1411 Node* adr, | |
| 1412 const TypePtr* adr_type, | |
| 1413 Node *val, | |
| 1414 const Type* val_type, | |
| 1415 BasicType bt) { | |
| 1416 uint adr_idx = C->get_alias_index(adr_type); | |
| 1417 Node* store; | |
| 1418 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt); | |
| 1419 store = store_to_memory(control(), adr, val, bt, adr_idx); | |
| 1420 post_barrier(control(), store, obj, adr, adr_idx, val, bt, false); | |
| 1421 return store; | |
| 1422 } | |
| 1423 | |
| 1424 Node* GraphKit::store_oop_to_array(Node* ctl, | |
| 1425 Node* obj, | |
| 1426 Node* adr, | |
| 1427 const TypePtr* adr_type, | |
| 1428 Node *val, | |
| 1429 const Type* val_type, | |
| 1430 BasicType bt) { | |
| 1431 uint adr_idx = C->get_alias_index(adr_type); | |
| 1432 Node* store; | |
| 1433 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt); | |
| 1434 store = store_to_memory(control(), adr, val, bt, adr_idx); | |
| 1435 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true); | |
| 1436 return store; | |
| 1437 } | |
| 1438 | |
| 1439 Node* GraphKit::store_oop_to_unknown(Node* ctl, | |
| 1440 Node* obj, | |
| 1441 Node* adr, | |
| 1442 const TypePtr* adr_type, | |
| 1443 Node *val, | |
| 1444 const Type* val_type, | |
| 1445 BasicType bt) { | |
| 1446 uint adr_idx = C->get_alias_index(adr_type); | |
| 1447 Node* store; | |
| 1448 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt); | |
| 1449 store = store_to_memory(control(), adr, val, bt, adr_idx); | |
| 1450 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true); | |
| 1451 return store; | |
| 1452 } | |
| 1453 | |
| 1454 | |
| 1455 //-------------------------array_element_address------------------------- | |
| 1456 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, | |
| 1457 const TypeInt* sizetype) { | |
|
29
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
23
diff
changeset
|
1458 uint shift = exact_log2(type2aelembytes(elembt)); |
| 0 | 1459 uint header = arrayOopDesc::base_offset_in_bytes(elembt); |
| 1460 | |
| 1461 // short-circuit a common case (saves lots of confusing waste motion) | |
| 1462 jint idx_con = find_int_con(idx, -1); | |
| 1463 if (idx_con >= 0) { | |
| 1464 intptr_t offset = header + ((intptr_t)idx_con << shift); | |
| 1465 return basic_plus_adr(ary, offset); | |
| 1466 } | |
| 1467 | |
| 1468 // must be correct type for alignment purposes | |
| 1469 Node* base = basic_plus_adr(ary, header); | |
| 1470 #ifdef _LP64 | |
| 1471 // The scaled index operand to AddP must be a clean 64-bit value. | |
| 1472 // Java allows a 32-bit int to be incremented to a negative | |
| 1473 // value, which appears in a 64-bit register as a large | |
| 1474 // positive number. Using that large positive number as an | |
| 1475 // operand in pointer arithmetic has bad consequences. | |
| 1476 // On the other hand, 32-bit overflow is rare, and the possibility | |
| 1477 // can often be excluded, if we annotate the ConvI2L node with | |
| 1478 // a type assertion that its value is known to be a small positive | |
| 1479 // number. (The prior range check has ensured this.) | |
| 1480 // This assertion is used by ConvI2LNode::Ideal. | |
| 1481 int index_max = max_jint - 1; // array size is max_jint, index is one less | |
| 1482 if (sizetype != NULL) index_max = sizetype->_hi - 1; | |
| 1483 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); | |
| 1484 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); | |
| 1485 #endif | |
| 1486 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); | |
| 1487 return basic_plus_adr(ary, base, scale); | |
| 1488 } | |
| 1489 | |
| 1490 //-------------------------load_array_element------------------------- | |
| 1491 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { | |
| 1492 const Type* elemtype = arytype->elem(); | |
| 1493 BasicType elembt = elemtype->array_element_basic_type(); | |
| 1494 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); | |
| 1495 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); | |
| 1496 return ld; | |
| 1497 } | |
| 1498 | |
| 1499 //-------------------------set_arguments_for_java_call------------------------- | |
| 1500 // Arguments (pre-popped from the stack) are taken from the JVMS. | |
| 1501 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { | |
| 1502 // Add the call arguments: | |
| 1503 uint nargs = call->method()->arg_size(); | |
| 1504 for (uint i = 0; i < nargs; i++) { | |
| 1505 Node* arg = argument(i); | |
| 1506 call->init_req(i + TypeFunc::Parms, arg); | |
| 1507 } | |
| 1508 } | |
| 1509 | |
| 1510 //---------------------------set_edges_for_java_call--------------------------- | |
| 1511 // Connect a newly created call into the current JVMS. | |
| 1512 // A return value node (if any) is returned from set_edges_for_java_call. | |
| 1513 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) { | |
| 1514 | |
| 1515 // Add the predefined inputs: | |
| 1516 call->init_req( TypeFunc::Control, control() ); | |
| 1517 call->init_req( TypeFunc::I_O , i_o() ); | |
| 1518 call->init_req( TypeFunc::Memory , reset_memory() ); | |
| 1519 call->init_req( TypeFunc::FramePtr, frameptr() ); | |
| 1520 call->init_req( TypeFunc::ReturnAdr, top() ); | |
| 1521 | |
| 1522 add_safepoint_edges(call, must_throw); | |
| 1523 | |
| 1524 Node* xcall = _gvn.transform(call); | |
| 1525 | |
| 1526 if (xcall == top()) { | |
| 1527 set_control(top()); | |
| 1528 return; | |
| 1529 } | |
| 1530 assert(xcall == call, "call identity is stable"); | |
| 1531 | |
| 1532 // Re-use the current map to produce the result. | |
| 1533 | |
| 1534 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); | |
| 1535 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O ))); | |
| 1536 set_all_memory_call(xcall); | |
| 1537 | |
| 1538 //return xcall; // no need, caller already has it | |
| 1539 } | |
| 1540 | |
| 1541 Node* GraphKit::set_results_for_java_call(CallJavaNode* call) { | |
| 1542 if (stopped()) return top(); // maybe the call folded up? | |
| 1543 | |
| 1544 // Capture the return value, if any. | |
| 1545 Node* ret; | |
| 1546 if (call->method() == NULL || | |
| 1547 call->method()->return_type()->basic_type() == T_VOID) | |
| 1548 ret = top(); | |
| 1549 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); | |
| 1550 | |
| 1551 // Note: Since any out-of-line call can produce an exception, | |
| 1552 // we always insert an I_O projection from the call into the result. | |
| 1553 | |
| 1554 make_slow_call_ex(call, env()->Throwable_klass(), false); | |
| 1555 | |
| 1556 return ret; | |
| 1557 } | |
| 1558 | |
| 1559 //--------------------set_predefined_input_for_runtime_call-------------------- | |
| 1560 // Reading and setting the memory state is way conservative here. | |
| 1561 // The real problem is that I am not doing real Type analysis on memory, | |
| 1562 // so I cannot distinguish card mark stores from other stores. Across a GC | |
| 1563 // point the Store Barrier and the card mark memory has to agree. I cannot | |
| 1564 // have a card mark store and its barrier split across the GC point from | |
| 1565 // either above or below. Here I get that to happen by reading ALL of memory. | |
| 1566 // A better answer would be to separate out card marks from other memory. | |
| 1567 // For now, return the input memory state, so that it can be reused | |
| 1568 // after the call, if this call has restricted memory effects. | |
| 1569 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { | |
| 1570 // Set fixed predefined input arguments | |
| 1571 Node* memory = reset_memory(); | |
| 1572 call->init_req( TypeFunc::Control, control() ); | |
| 1573 call->init_req( TypeFunc::I_O, top() ); // does no i/o | |
| 1574 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs | |
| 1575 call->init_req( TypeFunc::FramePtr, frameptr() ); | |
| 1576 call->init_req( TypeFunc::ReturnAdr, top() ); | |
| 1577 return memory; | |
| 1578 } | |
| 1579 | |
| 1580 //-------------------set_predefined_output_for_runtime_call-------------------- | |
| 1581 // Set control and memory (not i_o) from the call. | |
| 1582 // If keep_mem is not NULL, use it for the output state, | |
| 1583 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. | |
| 1584 // If hook_mem is NULL, this call produces no memory effects at all. | |
| 1585 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), | |
| 1586 // then only that memory slice is taken from the call. | |
| 1587 // In the last case, we must put an appropriate memory barrier before | |
| 1588 // the call, so as to create the correct anti-dependencies on loads | |
| 1589 // preceding the call. | |
| 1590 void GraphKit::set_predefined_output_for_runtime_call(Node* call, | |
| 1591 Node* keep_mem, | |
| 1592 const TypePtr* hook_mem) { | |
| 1593 // no i/o | |
| 1594 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); | |
| 1595 if (keep_mem) { | |
| 1596 // First clone the existing memory state | |
| 1597 set_all_memory(keep_mem); | |
| 1598 if (hook_mem != NULL) { | |
| 1599 // Make memory for the call | |
| 1600 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); | |
| 1601 // Set the RawPtr memory state only. This covers all the heap top/GC stuff | |
| 1602 // We also use hook_mem to extract specific effects from arraycopy stubs. | |
| 1603 set_memory(mem, hook_mem); | |
| 1604 } | |
| 1605 // ...else the call has NO memory effects. | |
| 1606 | |
| 1607 // Make sure the call advertises its memory effects precisely. | |
| 1608 // This lets us build accurate anti-dependences in gcm.cpp. | |
| 1609 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), | |
| 1610 "call node must be constructed correctly"); | |
| 1611 } else { | |
| 1612 assert(hook_mem == NULL, ""); | |
| 1613 // This is not a "slow path" call; all memory comes from the call. | |
| 1614 set_all_memory_call(call); | |
| 1615 } | |
| 1616 } | |
| 1617 | |
| 1618 //------------------------------increment_counter------------------------------ | |
| 1619 // for statistics: increment a VM counter by 1 | |
| 1620 | |
| 1621 void GraphKit::increment_counter(address counter_addr) { | |
| 1622 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); | |
| 1623 increment_counter(adr1); | |
| 1624 } | |
| 1625 | |
| 1626 void GraphKit::increment_counter(Node* counter_addr) { | |
| 1627 int adr_type = Compile::AliasIdxRaw; | |
| 1628 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); | |
| 1629 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); | |
| 1630 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type ); | |
| 1631 } | |
| 1632 | |
| 1633 | |
| 1634 //------------------------------uncommon_trap---------------------------------- | |
| 1635 // Bail out to the interpreter in mid-method. Implemented by calling the | |
| 1636 // uncommon_trap blob. This helper function inserts a runtime call with the | |
| 1637 // right debug info. | |
| 1638 void GraphKit::uncommon_trap(int trap_request, | |
| 1639 ciKlass* klass, const char* comment, | |
| 1640 bool must_throw, | |
| 1641 bool keep_exact_action) { | |
| 1642 if (failing()) stop(); | |
| 1643 if (stopped()) return; // trap reachable? | |
| 1644 | |
| 1645 // Note: If ProfileTraps is true, and if a deopt. actually | |
| 1646 // occurs here, the runtime will make sure an MDO exists. There is | |
| 1647 // no need to call method()->build_method_data() at this point. | |
| 1648 | |
| 1649 #ifdef ASSERT | |
| 1650 if (!must_throw) { | |
| 1651 // Make sure the stack has at least enough depth to execute | |
| 1652 // the current bytecode. | |
| 1653 int inputs, ignore; | |
| 1654 if (compute_stack_effects(inputs, ignore)) { | |
| 1655 assert(sp() >= inputs, "must have enough JVMS stack to execute"); | |
| 1656 // It is a frequent error in library_call.cpp to issue an | |
| 1657 // uncommon trap with the _sp value already popped. | |
| 1658 } | |
| 1659 } | |
| 1660 #endif | |
| 1661 | |
| 1662 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); | |
| 1663 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); | |
| 1664 | |
| 1665 switch (action) { | |
| 1666 case Deoptimization::Action_maybe_recompile: | |
| 1667 case Deoptimization::Action_reinterpret: | |
| 1668 // Temporary fix for 6529811 to allow virtual calls to be sure they | |
| 1669 // get the chance to go from mono->bi->mega | |
| 1670 if (!keep_exact_action && | |
| 1671 Deoptimization::trap_request_index(trap_request) < 0 && | |
| 1672 too_many_recompiles(reason)) { | |
| 1673 // This BCI is causing too many recompilations. | |
| 1674 action = Deoptimization::Action_none; | |
| 1675 trap_request = Deoptimization::make_trap_request(reason, action); | |
| 1676 } else { | |
| 1677 C->set_trap_can_recompile(true); | |
| 1678 } | |
| 1679 break; | |
| 1680 case Deoptimization::Action_make_not_entrant: | |
| 1681 C->set_trap_can_recompile(true); | |
| 1682 break; | |
| 1683 #ifdef ASSERT | |
| 1684 case Deoptimization::Action_none: | |
| 1685 case Deoptimization::Action_make_not_compilable: | |
| 1686 break; | |
| 1687 default: | |
| 1688 assert(false, "bad action"); | |
| 1689 #endif | |
| 1690 } | |
| 1691 | |
| 1692 if (TraceOptoParse) { | |
| 1693 char buf[100]; | |
| 1694 tty->print_cr("Uncommon trap %s at bci:%d", | |
| 1695 Deoptimization::format_trap_request(buf, sizeof(buf), | |
| 1696 trap_request), bci()); | |
| 1697 } | |
| 1698 | |
| 1699 CompileLog* log = C->log(); | |
| 1700 if (log != NULL) { | |
| 1701 int kid = (klass == NULL)? -1: log->identify(klass); | |
| 1702 log->begin_elem("uncommon_trap bci='%d'", bci()); | |
| 1703 char buf[100]; | |
| 1704 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), | |
| 1705 trap_request)); | |
| 1706 if (kid >= 0) log->print(" klass='%d'", kid); | |
| 1707 if (comment != NULL) log->print(" comment='%s'", comment); | |
| 1708 log->end_elem(); | |
| 1709 } | |
| 1710 | |
| 1711 // Make sure any guarding test views this path as very unlikely | |
| 1712 Node *i0 = control()->in(0); | |
| 1713 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? | |
| 1714 IfNode *iff = i0->as_If(); | |
| 1715 float f = iff->_prob; // Get prob | |
| 1716 if (control()->Opcode() == Op_IfTrue) { | |
| 1717 if (f > PROB_UNLIKELY_MAG(4)) | |
| 1718 iff->_prob = PROB_MIN; | |
| 1719 } else { | |
| 1720 if (f < PROB_LIKELY_MAG(4)) | |
| 1721 iff->_prob = PROB_MAX; | |
| 1722 } | |
| 1723 } | |
| 1724 | |
| 1725 // Clear out dead values from the debug info. | |
| 1726 kill_dead_locals(); | |
| 1727 | |
| 1728 // Now insert the uncommon trap subroutine call | |
| 1729 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin(); | |
| 1730 const TypePtr* no_memory_effects = NULL; | |
| 1731 // Pass the index of the class to be loaded | |
| 1732 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | | |
| 1733 (must_throw ? RC_MUST_THROW : 0), | |
| 1734 OptoRuntime::uncommon_trap_Type(), | |
| 1735 call_addr, "uncommon_trap", no_memory_effects, | |
| 1736 intcon(trap_request)); | |
| 1737 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, | |
| 1738 "must extract request correctly from the graph"); | |
| 1739 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); | |
| 1740 | |
| 1741 call->set_req(TypeFunc::ReturnAdr, returnadr()); | |
| 1742 // The debug info is the only real input to this call. | |
| 1743 | |
| 1744 // Halt-and-catch fire here. The above call should never return! | |
| 1745 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); | |
| 1746 _gvn.set_type_bottom(halt); | |
| 1747 root()->add_req(halt); | |
| 1748 | |
| 1749 stop_and_kill_map(); | |
| 1750 } | |
| 1751 | |
| 1752 | |
| 1753 //--------------------------just_allocated_object------------------------------ | |
| 1754 // Report the object that was just allocated. | |
| 1755 // It must be the case that there are no intervening safepoints. | |
| 1756 // We use this to determine if an object is so "fresh" that | |
| 1757 // it does not require card marks. | |
| 1758 Node* GraphKit::just_allocated_object(Node* current_control) { | |
| 1759 if (C->recent_alloc_ctl() == current_control) | |
| 1760 return C->recent_alloc_obj(); | |
| 1761 return NULL; | |
| 1762 } | |
| 1763 | |
| 1764 | |
| 1765 //------------------------------store_barrier---------------------------------- | |
| 1766 // Insert a write-barrier store. This is to let generational GC work; we have | |
| 1767 // to flag all oop-stores before the next GC point. | |
| 1768 void GraphKit::write_barrier_post(Node* oop_store, Node* obj, Node* adr, | |
| 1769 Node* val, bool use_precise) { | |
| 1770 // No store check needed if we're storing a NULL or an old object | |
| 1771 // (latter case is probably a string constant). The concurrent | |
| 1772 // mark sweep garbage collector, however, needs to have all nonNull | |
| 1773 // oop updates flagged via card-marks. | |
| 1774 if (val != NULL && val->is_Con()) { | |
| 1775 // must be either an oop or NULL | |
| 1776 const Type* t = val->bottom_type(); | |
| 1777 if (t == TypePtr::NULL_PTR || t == Type::TOP) | |
| 1778 // stores of null never (?) need barriers | |
| 1779 return; | |
| 1780 ciObject* con = t->is_oopptr()->const_oop(); | |
| 1781 if (con != NULL | |
| 1782 && con->is_perm() | |
| 1783 && Universe::heap()->can_elide_permanent_oop_store_barriers()) | |
| 1784 // no store barrier needed, because no old-to-new ref created | |
| 1785 return; | |
| 1786 } | |
| 1787 | |
| 1788 if (use_ReduceInitialCardMarks() | |
| 1789 && obj == just_allocated_object(control())) { | |
| 1790 // We can skip marks on a freshly-allocated object. | |
| 1791 // Keep this code in sync with do_eager_card_mark in runtime.cpp. | |
| 1792 // That routine eagerly marks the occasional object which is produced | |
| 1793 // by the slow path, so that we don't have to do it here. | |
| 1794 return; | |
| 1795 } | |
| 1796 | |
| 1797 if (!use_precise) { | |
| 1798 // All card marks for a (non-array) instance are in one place: | |
| 1799 adr = obj; | |
| 1800 } | |
| 1801 // (Else it's an array (or unknown), and we want more precise card marks.) | |
| 1802 assert(adr != NULL, ""); | |
| 1803 | |
| 1804 // Get the alias_index for raw card-mark memory | |
| 1805 int adr_type = Compile::AliasIdxRaw; | |
| 1806 // Convert the pointer to an int prior to doing math on it | |
| 1807 Node* cast = _gvn.transform(new (C, 2) CastP2XNode(control(), adr)); | |
| 1808 // Divide by card size | |
| 1809 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, | |
| 1810 "Only one we handle so far."); | |
| 1811 CardTableModRefBS* ct = | |
| 1812 (CardTableModRefBS*)(Universe::heap()->barrier_set()); | |
| 1813 Node *b = _gvn.transform(new (C, 3) URShiftXNode( cast, _gvn.intcon(CardTableModRefBS::card_shift) )); | |
| 1814 // We store into a byte array, so do not bother to left-shift by zero | |
| 1815 // Get base of card map | |
| 1816 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), | |
| 1817 "adjust this code"); | |
| 1818 Node *c = makecon(TypeRawPtr::make((address)ct->byte_map_base)); | |
| 1819 // Combine | |
| 1820 Node *sb_ctl = control(); | |
| 1821 Node *sb_adr = _gvn.transform(new (C, 4) AddPNode( top()/*no base ptr*/, c, b )); | |
| 1822 Node *sb_val = _gvn.intcon(0); | |
| 1823 // Smash zero into card | |
| 1824 if( !UseConcMarkSweepGC ) { | |
| 1825 BasicType bt = T_BYTE; | |
| 1826 store_to_memory(sb_ctl, sb_adr, sb_val, bt, adr_type); | |
| 1827 } else { | |
| 1828 // Specialized path for CM store barrier | |
| 1829 cms_card_mark( sb_ctl, sb_adr, sb_val, oop_store); | |
| 1830 } | |
| 1831 } | |
| 1832 | |
| 1833 // Specialized path for CMS store barrier | |
| 1834 void GraphKit::cms_card_mark(Node* ctl, Node* adr, Node* val, Node *oop_store) { | |
| 1835 BasicType bt = T_BYTE; | |
| 1836 int adr_idx = Compile::AliasIdxRaw; | |
| 1837 Node* mem = memory(adr_idx); | |
| 1838 | |
| 1839 // The type input is NULL in PRODUCT builds | |
| 1840 const TypePtr* type = NULL; | |
| 1841 debug_only(type = C->get_adr_type(adr_idx)); | |
| 1842 | |
| 1843 // Add required edge to oop_store, optimizer does not support precedence edges. | |
| 1844 // Convert required edge to precedence edge before allocation. | |
| 1845 Node *store = _gvn.transform( new (C, 5) StoreCMNode(ctl, mem, adr, type, val, oop_store) ); | |
| 1846 set_memory(store, adr_idx); | |
| 1847 | |
| 1848 // For CMS, back-to-back card-marks can only remove the first one | |
| 1849 // and this requires DU info. Push on worklist for optimizer. | |
| 1850 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) | |
| 1851 record_for_igvn(store); | |
| 1852 } | |
| 1853 | |
| 1854 | |
| 1855 void GraphKit::round_double_arguments(ciMethod* dest_method) { | |
| 1856 // (Note: TypeFunc::make has a cache that makes this fast.) | |
| 1857 const TypeFunc* tf = TypeFunc::make(dest_method); | |
| 1858 int nargs = tf->_domain->_cnt - TypeFunc::Parms; | |
| 1859 for (int j = 0; j < nargs; j++) { | |
| 1860 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); | |
| 1861 if( targ->basic_type() == T_DOUBLE ) { | |
| 1862 // If any parameters are doubles, they must be rounded before | |
| 1863 // the call, dstore_rounding does gvn.transform | |
| 1864 Node *arg = argument(j); | |
| 1865 arg = dstore_rounding(arg); | |
| 1866 set_argument(j, arg); | |
| 1867 } | |
| 1868 } | |
| 1869 } | |
| 1870 | |
| 1871 void GraphKit::round_double_result(ciMethod* dest_method) { | |
| 1872 // A non-strict method may return a double value which has an extended | |
| 1873 // exponent, but this must not be visible in a caller which is 'strict' | |
| 1874 // If a strict caller invokes a non-strict callee, round a double result | |
| 1875 | |
| 1876 BasicType result_type = dest_method->return_type()->basic_type(); | |
| 1877 assert( method() != NULL, "must have caller context"); | |
| 1878 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { | |
| 1879 // Destination method's return value is on top of stack | |
| 1880 // dstore_rounding() does gvn.transform | |
| 1881 Node *result = pop_pair(); | |
| 1882 result = dstore_rounding(result); | |
| 1883 push_pair(result); | |
| 1884 } | |
| 1885 } | |
| 1886 | |
| 1887 // rounding for strict float precision conformance | |
| 1888 Node* GraphKit::precision_rounding(Node* n) { | |
| 1889 return UseStrictFP && _method->flags().is_strict() | |
| 1890 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding | |
| 1891 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) | |
| 1892 : n; | |
| 1893 } | |
| 1894 | |
| 1895 // rounding for strict double precision conformance | |
| 1896 Node* GraphKit::dprecision_rounding(Node *n) { | |
| 1897 return UseStrictFP && _method->flags().is_strict() | |
| 1898 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding | |
| 1899 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) | |
| 1900 : n; | |
| 1901 } | |
| 1902 | |
| 1903 // rounding for non-strict double stores | |
| 1904 Node* GraphKit::dstore_rounding(Node* n) { | |
| 1905 return Matcher::strict_fp_requires_explicit_rounding | |
| 1906 && UseSSE <= 1 | |
| 1907 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) | |
| 1908 : n; | |
| 1909 } | |
| 1910 | |
| 1911 //============================================================================= | |
| 1912 // Generate a fast path/slow path idiom. Graph looks like: | |
| 1913 // [foo] indicates that 'foo' is a parameter | |
| 1914 // | |
| 1915 // [in] NULL | |
| 1916 // \ / | |
| 1917 // CmpP | |
| 1918 // Bool ne | |
| 1919 // If | |
| 1920 // / \ | |
| 1921 // True False-<2> | |
| 1922 // / | | |
| 1923 // / cast_not_null | |
| 1924 // Load | | ^ | |
| 1925 // [fast_test] | | | |
| 1926 // gvn to opt_test | | | |
| 1927 // / \ | <1> | |
| 1928 // True False | | |
| 1929 // | \\ | | |
| 1930 // [slow_call] \[fast_result] | |
| 1931 // Ctl Val \ \ | |
| 1932 // | \ \ | |
| 1933 // Catch <1> \ \ | |
| 1934 // / \ ^ \ \ | |
| 1935 // Ex No_Ex | \ \ | |
| 1936 // | \ \ | \ <2> \ | |
| 1937 // ... \ [slow_res] | | \ [null_result] | |
| 1938 // \ \--+--+--- | | | |
| 1939 // \ | / \ | / | |
| 1940 // --------Region Phi | |
| 1941 // | |
| 1942 //============================================================================= | |
| 1943 // Code is structured as a series of driver functions all called 'do_XXX' that | |
| 1944 // call a set of helper functions. Helper functions first, then drivers. | |
| 1945 | |
| 1946 //------------------------------null_check_oop--------------------------------- | |
| 1947 // Null check oop. Set null-path control into Region in slot 3. | |
| 1948 // Make a cast-not-nullness use the other not-null control. Return cast. | |
| 1949 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, | |
| 1950 bool never_see_null) { | |
| 1951 // Initial NULL check taken path | |
| 1952 (*null_control) = top(); | |
| 1953 Node* cast = null_check_common(value, T_OBJECT, false, null_control); | |
| 1954 | |
| 1955 // Generate uncommon_trap: | |
| 1956 if (never_see_null && (*null_control) != top()) { | |
| 1957 // If we see an unexpected null at a check-cast we record it and force a | |
| 1958 // recompile; the offending check-cast will be compiled to handle NULLs. | |
| 1959 // If we see more than one offending BCI, then all checkcasts in the | |
| 1960 // method will be compiled to handle NULLs. | |
| 1961 PreserveJVMState pjvms(this); | |
| 1962 set_control(*null_control); | |
| 1963 uncommon_trap(Deoptimization::Reason_null_check, | |
| 1964 Deoptimization::Action_make_not_entrant); | |
| 1965 (*null_control) = top(); // NULL path is dead | |
| 1966 } | |
| 1967 | |
| 1968 // Cast away null-ness on the result | |
| 1969 return cast; | |
| 1970 } | |
| 1971 | |
| 1972 //------------------------------opt_iff---------------------------------------- | |
| 1973 // Optimize the fast-check IfNode. Set the fast-path region slot 2. | |
| 1974 // Return slow-path control. | |
| 1975 Node* GraphKit::opt_iff(Node* region, Node* iff) { | |
| 1976 IfNode *opt_iff = _gvn.transform(iff)->as_If(); | |
| 1977 | |
| 1978 // Fast path taken; set region slot 2 | |
| 1979 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); | |
| 1980 region->init_req(2,fast_taken); // Capture fast-control | |
| 1981 | |
| 1982 // Fast path not-taken, i.e. slow path | |
| 1983 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); | |
| 1984 return slow_taken; | |
| 1985 } | |
| 1986 | |
| 1987 //-----------------------------make_runtime_call------------------------------- | |
| 1988 Node* GraphKit::make_runtime_call(int flags, | |
| 1989 const TypeFunc* call_type, address call_addr, | |
| 1990 const char* call_name, | |
| 1991 const TypePtr* adr_type, | |
| 1992 // The following parms are all optional. | |
| 1993 // The first NULL ends the list. | |
| 1994 Node* parm0, Node* parm1, | |
| 1995 Node* parm2, Node* parm3, | |
| 1996 Node* parm4, Node* parm5, | |
| 1997 Node* parm6, Node* parm7) { | |
| 1998 // Slow-path call | |
| 1999 int size = call_type->domain()->cnt(); | |
| 2000 bool is_leaf = !(flags & RC_NO_LEAF); | |
| 2001 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); | |
| 2002 if (call_name == NULL) { | |
| 2003 assert(!is_leaf, "must supply name for leaf"); | |
| 2004 call_name = OptoRuntime::stub_name(call_addr); | |
| 2005 } | |
| 2006 CallNode* call; | |
| 2007 if (!is_leaf) { | |
| 2008 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, | |
| 2009 bci(), adr_type); | |
| 2010 } else if (flags & RC_NO_FP) { | |
| 2011 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); | |
| 2012 } else { | |
| 2013 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); | |
| 2014 } | |
| 2015 | |
| 2016 // The following is similar to set_edges_for_java_call, | |
| 2017 // except that the memory effects of the call are restricted to AliasIdxRaw. | |
| 2018 | |
| 2019 // Slow path call has no side-effects, uses few values | |
| 2020 bool wide_in = !(flags & RC_NARROW_MEM); | |
| 2021 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); | |
| 2022 | |
| 2023 Node* prev_mem = NULL; | |
| 2024 if (wide_in) { | |
| 2025 prev_mem = set_predefined_input_for_runtime_call(call); | |
| 2026 } else { | |
| 2027 assert(!wide_out, "narrow in => narrow out"); | |
| 2028 Node* narrow_mem = memory(adr_type); | |
| 2029 prev_mem = reset_memory(); | |
| 2030 map()->set_memory(narrow_mem); | |
| 2031 set_predefined_input_for_runtime_call(call); | |
| 2032 } | |
| 2033 | |
| 2034 // Hook each parm in order. Stop looking at the first NULL. | |
| 2035 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); | |
| 2036 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); | |
| 2037 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); | |
| 2038 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); | |
| 2039 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); | |
| 2040 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); | |
| 2041 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); | |
| 2042 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); | |
| 2043 /* close each nested if ===> */ } } } } } } } } | |
| 2044 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); | |
| 2045 | |
| 2046 if (!is_leaf) { | |
| 2047 // Non-leaves can block and take safepoints: | |
| 2048 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); | |
| 2049 } | |
| 2050 // Non-leaves can throw exceptions: | |
| 2051 if (has_io) { | |
| 2052 call->set_req(TypeFunc::I_O, i_o()); | |
| 2053 } | |
| 2054 | |
| 2055 if (flags & RC_UNCOMMON) { | |
| 2056 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. | |
| 2057 // (An "if" probability corresponds roughly to an unconditional count. | |
| 2058 // Sort of.) | |
| 2059 call->set_cnt(PROB_UNLIKELY_MAG(4)); | |
| 2060 } | |
| 2061 | |
| 2062 Node* c = _gvn.transform(call); | |
| 2063 assert(c == call, "cannot disappear"); | |
| 2064 | |
| 2065 if (wide_out) { | |
| 2066 // Slow path call has full side-effects. | |
| 2067 set_predefined_output_for_runtime_call(call); | |
| 2068 } else { | |
| 2069 // Slow path call has few side-effects, and/or sets few values. | |
| 2070 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); | |
| 2071 } | |
| 2072 | |
| 2073 if (has_io) { | |
| 2074 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); | |
| 2075 } | |
| 2076 return call; | |
| 2077 | |
| 2078 } | |
| 2079 | |
| 2080 //------------------------------merge_memory----------------------------------- | |
| 2081 // Merge memory from one path into the current memory state. | |
| 2082 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { | |
| 2083 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { | |
| 2084 Node* old_slice = mms.force_memory(); | |
| 2085 Node* new_slice = mms.memory2(); | |
| 2086 if (old_slice != new_slice) { | |
| 2087 PhiNode* phi; | |
| 2088 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { | |
| 2089 phi = new_slice->as_Phi(); | |
| 2090 #ifdef ASSERT | |
| 2091 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) | |
| 2092 old_slice = old_slice->in(new_path); | |
| 2093 // Caller is responsible for ensuring that any pre-existing | |
| 2094 // phis are already aware of old memory. | |
| 2095 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path | |
| 2096 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); | |
| 2097 #endif | |
| 2098 mms.set_memory(phi); | |
| 2099 } else { | |
| 2100 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); | |
| 2101 _gvn.set_type(phi, Type::MEMORY); | |
| 2102 phi->set_req(new_path, new_slice); | |
| 2103 mms.set_memory(_gvn.transform(phi)); // assume it is complete | |
| 2104 } | |
| 2105 } | |
| 2106 } | |
| 2107 } | |
| 2108 | |
| 2109 //------------------------------make_slow_call_ex------------------------------ | |
| 2110 // Make the exception handler hookups for the slow call | |
| 2111 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { | |
| 2112 if (stopped()) return; | |
| 2113 | |
| 2114 // Make a catch node with just two handlers: fall-through and catch-all | |
| 2115 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); | |
| 2116 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); | |
| 2117 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); | |
| 2118 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); | |
| 2119 | |
| 2120 { PreserveJVMState pjvms(this); | |
| 2121 set_control(excp); | |
| 2122 set_i_o(i_o); | |
| 2123 | |
| 2124 if (excp != top()) { | |
| 2125 // Create an exception state also. | |
| 2126 // Use an exact type if the caller has specified a specific exception. | |
| 2127 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); | |
| 2128 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); | |
| 2129 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); | |
| 2130 } | |
| 2131 } | |
| 2132 | |
| 2133 // Get the no-exception control from the CatchNode. | |
| 2134 set_control(norm); | |
| 2135 } | |
| 2136 | |
| 2137 | |
| 2138 //-------------------------------gen_subtype_check----------------------------- | |
| 2139 // Generate a subtyping check. Takes as input the subtype and supertype. | |
| 2140 // Returns 2 values: sets the default control() to the true path and returns | |
| 2141 // the false path. Only reads invariant memory; sets no (visible) memory. | |
| 2142 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding | |
| 2143 // but that's not exposed to the optimizer. This call also doesn't take in an | |
| 2144 // Object; if you wish to check an Object you need to load the Object's class | |
| 2145 // prior to coming here. | |
| 2146 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { | |
| 2147 // Fast check for identical types, perhaps identical constants. | |
| 2148 // The types can even be identical non-constants, in cases | |
| 2149 // involving Array.newInstance, Object.clone, etc. | |
| 2150 if (subklass == superklass) | |
| 2151 return top(); // false path is dead; no test needed. | |
| 2152 | |
| 2153 if (_gvn.type(superklass)->singleton()) { | |
| 2154 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); | |
| 2155 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); | |
| 2156 | |
| 2157 // In the common case of an exact superklass, try to fold up the | |
| 2158 // test before generating code. You may ask, why not just generate | |
| 2159 // the code and then let it fold up? The answer is that the generated | |
| 2160 // code will necessarily include null checks, which do not always | |
| 2161 // completely fold away. If they are also needless, then they turn | |
| 2162 // into a performance loss. Example: | |
| 2163 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; | |
| 2164 // Here, the type of 'fa' is often exact, so the store check | |
| 2165 // of fa[1]=x will fold up, without testing the nullness of x. | |
| 2166 switch (static_subtype_check(superk, subk)) { | |
| 2167 case SSC_always_false: | |
| 2168 { | |
| 2169 Node* always_fail = control(); | |
| 2170 set_control(top()); | |
| 2171 return always_fail; | |
| 2172 } | |
| 2173 case SSC_always_true: | |
| 2174 return top(); | |
| 2175 case SSC_easy_test: | |
| 2176 { | |
| 2177 // Just do a direct pointer compare and be done. | |
| 2178 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); | |
| 2179 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); | |
| 2180 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); | |
| 2181 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); | |
| 2182 return _gvn.transform( new(C, 1) IfFalseNode(iff) ); | |
| 2183 } | |
| 2184 case SSC_full_test: | |
| 2185 break; | |
| 2186 default: | |
| 2187 ShouldNotReachHere(); | |
| 2188 } | |
| 2189 } | |
| 2190 | |
| 2191 // %%% Possible further optimization: Even if the superklass is not exact, | |
| 2192 // if the subklass is the unique subtype of the superklass, the check | |
| 2193 // will always succeed. We could leave a dependency behind to ensure this. | |
| 2194 | |
| 2195 // First load the super-klass's check-offset | |
| 2196 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() ); | |
| 2197 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); | |
| 2198 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes(); | |
| 2199 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); | |
| 2200 | |
| 2201 // Load from the sub-klass's super-class display list, or a 1-word cache of | |
| 2202 // the secondary superclass list, or a failing value with a sentinel offset | |
| 2203 // if the super-klass is an interface or exceptionally deep in the Java | |
| 2204 // hierarchy and we have to scan the secondary superclass list the hard way. | |
| 2205 // Worst-case type is a little odd: NULL is allowed as a result (usually | |
| 2206 // klass loads can never produce a NULL). | |
| 2207 Node *chk_off_X = ConvI2X(chk_off); | |
| 2208 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); | |
| 2209 // For some types like interfaces the following loadKlass is from a 1-word | |
| 2210 // cache which is mutable so can't use immutable memory. Other | |
| 2211 // types load from the super-class display table which is immutable. | |
| 2212 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); | |
|
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
113
diff
changeset
|
2213 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); |
| 0 | 2214 |
| 2215 // Compile speed common case: ARE a subtype and we canNOT fail | |
| 2216 if( superklass == nkls ) | |
| 2217 return top(); // false path is dead; no test needed. | |
| 2218 | |
| 2219 // See if we get an immediate positive hit. Happens roughly 83% of the | |
| 2220 // time. Test to see if the value loaded just previously from the subklass | |
| 2221 // is exactly the superklass. | |
| 2222 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); | |
| 2223 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); | |
| 2224 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); | |
| 2225 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); | |
| 2226 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); | |
| 2227 | |
| 2228 // Compile speed common case: Check for being deterministic right now. If | |
| 2229 // chk_off is a constant and not equal to cacheoff then we are NOT a | |
| 2230 // subklass. In this case we need exactly the 1 test above and we can | |
| 2231 // return those results immediately. | |
| 2232 if (!might_be_cache) { | |
| 2233 Node* not_subtype_ctrl = control(); | |
| 2234 set_control(iftrue1); // We need exactly the 1 test above | |
| 2235 return not_subtype_ctrl; | |
| 2236 } | |
| 2237 | |
| 2238 // Gather the various success & failures here | |
| 2239 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); | |
| 2240 record_for_igvn(r_ok_subtype); | |
| 2241 RegionNode *r_not_subtype = new (C, 3) RegionNode(3); | |
| 2242 record_for_igvn(r_not_subtype); | |
| 2243 | |
| 2244 r_ok_subtype->init_req(1, iftrue1); | |
| 2245 | |
| 2246 // Check for immediate negative hit. Happens roughly 11% of the time (which | |
| 2247 // is roughly 63% of the remaining cases). Test to see if the loaded | |
| 2248 // check-offset points into the subklass display list or the 1-element | |
| 2249 // cache. If it points to the display (and NOT the cache) and the display | |
| 2250 // missed then it's not a subtype. | |
| 2251 Node *cacheoff = _gvn.intcon(cacheoff_con); | |
| 2252 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); | |
| 2253 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); | |
| 2254 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); | |
| 2255 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); | |
| 2256 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); | |
| 2257 | |
| 2258 // Check for self. Very rare to get here, but its taken 1/3 the time. | |
| 2259 // No performance impact (too rare) but allows sharing of secondary arrays | |
| 2260 // which has some footprint reduction. | |
| 2261 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); | |
| 2262 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); | |
| 2263 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); | |
| 2264 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); | |
| 2265 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); | |
| 2266 | |
| 2267 // Now do a linear scan of the secondary super-klass array. Again, no real | |
| 2268 // performance impact (too rare) but it's gotta be done. | |
| 2269 // (The stub also contains the self-check of subklass == superklass. | |
| 2270 // Since the code is rarely used, there is no penalty for moving it | |
| 2271 // out of line, and it can only improve I-cache density.) | |
| 2272 Node* psc = _gvn.transform( | |
| 2273 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); | |
| 2274 | |
| 2275 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); | |
| 2276 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); | |
| 2277 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); | |
| 2278 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); | |
| 2279 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); | |
| 2280 | |
| 2281 // Return false path; set default control to true path. | |
| 2282 set_control( _gvn.transform(r_ok_subtype) ); | |
| 2283 return _gvn.transform(r_not_subtype); | |
| 2284 } | |
| 2285 | |
| 2286 //----------------------------static_subtype_check----------------------------- | |
| 2287 // Shortcut important common cases when superklass is exact: | |
| 2288 // (0) superklass is java.lang.Object (can occur in reflective code) | |
| 2289 // (1) subklass is already limited to a subtype of superklass => always ok | |
| 2290 // (2) subklass does not overlap with superklass => always fail | |
| 2291 // (3) superklass has NO subtypes and we can check with a simple compare. | |
| 2292 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { | |
| 2293 if (StressReflectiveCode) { | |
| 2294 return SSC_full_test; // Let caller generate the general case. | |
| 2295 } | |
| 2296 | |
| 2297 if (superk == env()->Object_klass()) { | |
| 2298 return SSC_always_true; // (0) this test cannot fail | |
| 2299 } | |
| 2300 | |
| 2301 ciType* superelem = superk; | |
| 2302 if (superelem->is_array_klass()) | |
| 2303 superelem = superelem->as_array_klass()->base_element_type(); | |
| 2304 | |
| 2305 if (!subk->is_interface()) { // cannot trust static interface types yet | |
| 2306 if (subk->is_subtype_of(superk)) { | |
| 2307 return SSC_always_true; // (1) false path dead; no dynamic test needed | |
| 2308 } | |
| 2309 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && | |
| 2310 !superk->is_subtype_of(subk)) { | |
| 2311 return SSC_always_false; | |
| 2312 } | |
| 2313 } | |
| 2314 | |
| 2315 // If casting to an instance klass, it must have no subtypes | |
| 2316 if (superk->is_interface()) { | |
| 2317 // Cannot trust interfaces yet. | |
| 2318 // %%% S.B. superk->nof_implementors() == 1 | |
| 2319 } else if (superelem->is_instance_klass()) { | |
| 2320 ciInstanceKlass* ik = superelem->as_instance_klass(); | |
| 2321 if (!ik->has_subklass() && !ik->is_interface()) { | |
| 2322 if (!ik->is_final()) { | |
| 2323 // Add a dependency if there is a chance of a later subclass. | |
| 2324 C->dependencies()->assert_leaf_type(ik); | |
| 2325 } | |
| 2326 return SSC_easy_test; // (3) caller can do a simple ptr comparison | |
| 2327 } | |
| 2328 } else { | |
| 2329 // A primitive array type has no subtypes. | |
| 2330 return SSC_easy_test; // (3) caller can do a simple ptr comparison | |
| 2331 } | |
| 2332 | |
| 2333 return SSC_full_test; | |
| 2334 } | |
| 2335 | |
| 2336 // Profile-driven exact type check: | |
| 2337 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, | |
| 2338 float prob, | |
| 2339 Node* *casted_receiver) { | |
| 2340 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); | |
| 2341 Node* recv_klass = load_object_klass(receiver); | |
| 2342 Node* want_klass = makecon(tklass); | |
| 2343 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); | |
| 2344 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); | |
| 2345 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); | |
| 2346 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); | |
| 2347 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); | |
| 2348 | |
| 2349 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); | |
| 2350 assert(recv_xtype->klass_is_exact(), ""); | |
| 2351 | |
| 2352 // Subsume downstream occurrences of receiver with a cast to | |
| 2353 // recv_xtype, since now we know what the type will be. | |
| 2354 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); | |
| 2355 (*casted_receiver) = _gvn.transform(cast); | |
| 2356 // (User must make the replace_in_map call.) | |
| 2357 | |
| 2358 return fail; | |
| 2359 } | |
| 2360 | |
| 2361 | |
| 2362 //-------------------------------gen_instanceof-------------------------------- | |
| 2363 // Generate an instance-of idiom. Used by both the instance-of bytecode | |
| 2364 // and the reflective instance-of call. | |
| 2365 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) { | |
| 2366 C->set_has_split_ifs(true); // Has chance for split-if optimization | |
| 2367 assert( !stopped(), "dead parse path should be checked in callers" ); | |
| 2368 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), | |
| 2369 "must check for not-null not-dead klass in callers"); | |
| 2370 | |
| 2371 // Make the merge point | |
| 2372 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; | |
| 2373 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); | |
| 2374 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); | |
| 2375 C->set_has_split_ifs(true); // Has chance for split-if optimization | |
| 2376 | |
| 2377 // Null check; get casted pointer; set region slot 3 | |
| 2378 Node* null_ctl = top(); | |
| 2379 Node* not_null_obj = null_check_oop(subobj, &null_ctl); | |
| 2380 | |
| 2381 // If not_null_obj is dead, only null-path is taken | |
| 2382 if (stopped()) { // Doing instance-of on a NULL? | |
| 2383 set_control(null_ctl); | |
| 2384 return intcon(0); | |
| 2385 } | |
| 2386 region->init_req(_null_path, null_ctl); | |
| 2387 phi ->init_req(_null_path, intcon(0)); // Set null path value | |
| 2388 | |
| 2389 // Load the object's klass | |
| 2390 Node* obj_klass = load_object_klass(not_null_obj); | |
| 2391 | |
| 2392 // Generate the subtype check | |
| 2393 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); | |
| 2394 | |
| 2395 // Plug in the success path to the general merge in slot 1. | |
| 2396 region->init_req(_obj_path, control()); | |
| 2397 phi ->init_req(_obj_path, intcon(1)); | |
| 2398 | |
| 2399 // Plug in the failing path to the general merge in slot 2. | |
| 2400 region->init_req(_fail_path, not_subtype_ctrl); | |
| 2401 phi ->init_req(_fail_path, intcon(0)); | |
| 2402 | |
| 2403 // Return final merged results | |
| 2404 set_control( _gvn.transform(region) ); | |
| 2405 record_for_igvn(region); | |
| 2406 return _gvn.transform(phi); | |
| 2407 } | |
| 2408 | |
| 2409 //-------------------------------gen_checkcast--------------------------------- | |
| 2410 // Generate a checkcast idiom. Used by both the checkcast bytecode and the | |
| 2411 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the | |
| 2412 // uncommon-trap paths work. Adjust stack after this call. | |
| 2413 // If failure_control is supplied and not null, it is filled in with | |
| 2414 // the control edge for the cast failure. Otherwise, an appropriate | |
| 2415 // uncommon trap or exception is thrown. | |
| 2416 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, | |
| 2417 Node* *failure_control) { | |
| 2418 kill_dead_locals(); // Benefit all the uncommon traps | |
| 2419 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); | |
| 2420 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); | |
| 2421 | |
| 2422 // Fast cutout: Check the case that the cast is vacuously true. | |
| 2423 // This detects the common cases where the test will short-circuit | |
| 2424 // away completely. We do this before we perform the null check, | |
| 2425 // because if the test is going to turn into zero code, we don't | |
| 2426 // want a residual null check left around. (Causes a slowdown, | |
| 2427 // for example, in some objArray manipulations, such as a[i]=a[j].) | |
| 2428 if (tk->singleton()) { | |
| 2429 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); | |
| 2430 if (objtp != NULL && objtp->klass() != NULL) { | |
| 2431 switch (static_subtype_check(tk->klass(), objtp->klass())) { | |
| 2432 case SSC_always_true: | |
| 2433 return obj; | |
| 2434 case SSC_always_false: | |
| 2435 // It needs a null check because a null will *pass* the cast check. | |
| 2436 // A non-null value will always produce an exception. | |
| 2437 return do_null_assert(obj, T_OBJECT); | |
| 2438 } | |
| 2439 } | |
| 2440 } | |
| 2441 | |
| 2442 ciProfileData* data = NULL; | |
| 2443 if (failure_control == NULL) { // use MDO in regular case only | |
| 2444 assert(java_bc() == Bytecodes::_aastore || | |
| 2445 java_bc() == Bytecodes::_checkcast, | |
| 2446 "interpreter profiles type checks only for these BCs"); | |
| 2447 data = method()->method_data()->bci_to_data(bci()); | |
| 2448 } | |
| 2449 | |
| 2450 // Make the merge point | |
| 2451 enum { _obj_path = 1, _null_path, PATH_LIMIT }; | |
| 2452 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); | |
| 2453 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); | |
| 2454 C->set_has_split_ifs(true); // Has chance for split-if optimization | |
| 2455 | |
| 2456 // Use null-cast information if it is available | |
| 2457 bool never_see_null = false; | |
| 2458 // If we see an unexpected null at a check-cast we record it and force a | |
| 2459 // recompile; the offending check-cast will be compiled to handle NULLs. | |
| 2460 // If we see several offending BCIs, then all checkcasts in the | |
| 2461 // method will be compiled to handle NULLs. | |
| 2462 if (UncommonNullCast // Cutout for this technique | |
| 2463 && failure_control == NULL // regular case | |
| 2464 && obj != null() // And not the -Xcomp stupid case? | |
| 2465 && !too_many_traps(Deoptimization::Reason_null_check)) { | |
| 2466 // Finally, check the "null_seen" bit from the interpreter. | |
| 2467 if (data == NULL || !data->as_BitData()->null_seen()) { | |
| 2468 never_see_null = true; | |
| 2469 } | |
| 2470 } | |
| 2471 | |
| 2472 // Null check; get casted pointer; set region slot 3 | |
| 2473 Node* null_ctl = top(); | |
| 2474 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); | |
| 2475 | |
| 2476 // If not_null_obj is dead, only null-path is taken | |
| 2477 if (stopped()) { // Doing instance-of on a NULL? | |
| 2478 set_control(null_ctl); | |
| 2479 return null(); | |
| 2480 } | |
| 2481 region->init_req(_null_path, null_ctl); | |
| 2482 phi ->init_req(_null_path, null()); // Set null path value | |
| 2483 | |
| 2484 Node* cast_obj = NULL; // the casted version of the object | |
| 2485 | |
| 2486 // If the profile has seen exactly one type, narrow to that type. | |
| 2487 // (The subsequent subtype check will always fold up.) | |
| 2488 if (UseTypeProfile && TypeProfileCasts && data != NULL && | |
| 2489 // Counter has never been decremented (due to cast failure). | |
| 2490 // ...This is a reasonable thing to expect. It is true of | |
| 2491 // all casts inserted by javac to implement generic types. | |
| 2492 data->as_CounterData()->count() >= 0 && | |
| 2493 !too_many_traps(Deoptimization::Reason_class_check)) { | |
| 2494 // (No, this isn't a call, but it's enough like a virtual call | |
| 2495 // to use the same ciMethod accessor to get the profile info...) | |
| 2496 ciCallProfile profile = method()->call_profile_at_bci(bci()); | |
| 2497 if (profile.count() >= 0 && // no cast failures here | |
| 2498 profile.has_receiver(0) && | |
| 2499 profile.morphism() == 1) { | |
| 2500 ciKlass* exact_kls = profile.receiver(0); | |
| 2501 int ssc = static_subtype_check(tk->klass(), exact_kls); | |
| 2502 if (ssc == SSC_always_true) { | |
| 2503 // If we narrow the type to match what the type profile sees, | |
| 2504 // we can then remove the rest of the cast. | |
| 2505 // This is a win, even if the exact_kls is very specific, | |
| 2506 // because downstream operations, such as method calls, | |
| 2507 // will often benefit from the sharper type. | |
| 2508 Node* exact_obj = not_null_obj; // will get updated in place... | |
| 2509 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, | |
| 2510 &exact_obj); | |
| 2511 { PreserveJVMState pjvms(this); | |
| 2512 set_control(slow_ctl); | |
| 2513 uncommon_trap(Deoptimization::Reason_class_check, | |
| 2514 Deoptimization::Action_maybe_recompile); | |
| 2515 } | |
| 2516 if (failure_control != NULL) // failure is now impossible | |
| 2517 (*failure_control) = top(); | |
| 2518 replace_in_map(not_null_obj, exact_obj); | |
| 2519 // adjust the type of the phi to the exact klass: | |
| 2520 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR)); | |
| 2521 cast_obj = exact_obj; | |
| 2522 } | |
| 2523 // assert(cast_obj != NULL)... except maybe the profile lied to us. | |
| 2524 } | |
| 2525 } | |
| 2526 | |
| 2527 if (cast_obj == NULL) { | |
| 2528 // Load the object's klass | |
| 2529 Node* obj_klass = load_object_klass(not_null_obj); | |
| 2530 | |
| 2531 // Generate the subtype check | |
| 2532 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); | |
| 2533 | |
| 2534 // Plug in success path into the merge | |
| 2535 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), | |
| 2536 not_null_obj, toop)); | |
| 2537 // Failure path ends in uncommon trap (or may be dead - failure impossible) | |
| 2538 if (failure_control == NULL) { | |
| 2539 if (not_subtype_ctrl != top()) { // If failure is possible | |
| 2540 PreserveJVMState pjvms(this); | |
| 2541 set_control(not_subtype_ctrl); | |
| 2542 builtin_throw(Deoptimization::Reason_class_check, obj_klass); | |
| 2543 } | |
| 2544 } else { | |
| 2545 (*failure_control) = not_subtype_ctrl; | |
| 2546 } | |
| 2547 } | |
| 2548 | |
| 2549 region->init_req(_obj_path, control()); | |
| 2550 phi ->init_req(_obj_path, cast_obj); | |
| 2551 | |
| 2552 // A merge of NULL or Casted-NotNull obj | |
| 2553 Node* res = _gvn.transform(phi); | |
| 2554 | |
| 2555 // Note I do NOT always 'replace_in_map(obj,result)' here. | |
| 2556 // if( tk->klass()->can_be_primary_super() ) | |
| 2557 // This means that if I successfully store an Object into an array-of-String | |
| 2558 // I 'forget' that the Object is really now known to be a String. I have to | |
| 2559 // do this because we don't have true union types for interfaces - if I store | |
| 2560 // a Baz into an array-of-Interface and then tell the optimizer it's an | |
| 2561 // Interface, I forget that it's also a Baz and cannot do Baz-like field | |
| 2562 // references to it. FIX THIS WHEN UNION TYPES APPEAR! | |
| 2563 // replace_in_map( obj, res ); | |
| 2564 | |
| 2565 // Return final merged results | |
| 2566 set_control( _gvn.transform(region) ); | |
| 2567 record_for_igvn(region); | |
| 2568 return res; | |
| 2569 } | |
| 2570 | |
| 2571 //------------------------------next_monitor----------------------------------- | |
| 2572 // What number should be given to the next monitor? | |
| 2573 int GraphKit::next_monitor() { | |
| 2574 int current = jvms()->monitor_depth()* C->sync_stack_slots(); | |
| 2575 int next = current + C->sync_stack_slots(); | |
| 2576 // Keep the toplevel high water mark current: | |
| 2577 if (C->fixed_slots() < next) C->set_fixed_slots(next); | |
| 2578 return current; | |
| 2579 } | |
| 2580 | |
| 2581 //------------------------------insert_mem_bar--------------------------------- | |
| 2582 // Memory barrier to avoid floating things around | |
| 2583 // The membar serves as a pinch point between both control and all memory slices. | |
| 2584 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { | |
| 2585 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); | |
| 2586 mb->init_req(TypeFunc::Control, control()); | |
| 2587 mb->init_req(TypeFunc::Memory, reset_memory()); | |
| 2588 Node* membar = _gvn.transform(mb); | |
| 2589 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); | |
| 2590 set_all_memory_call(membar); | |
| 2591 return membar; | |
| 2592 } | |
| 2593 | |
| 2594 //-------------------------insert_mem_bar_volatile---------------------------- | |
| 2595 // Memory barrier to avoid floating things around | |
| 2596 // The membar serves as a pinch point between both control and memory(alias_idx). | |
| 2597 // If you want to make a pinch point on all memory slices, do not use this | |
| 2598 // function (even with AliasIdxBot); use insert_mem_bar() instead. | |
| 2599 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { | |
| 2600 // When Parse::do_put_xxx updates a volatile field, it appends a series | |
| 2601 // of MemBarVolatile nodes, one for *each* volatile field alias category. | |
| 2602 // The first membar is on the same memory slice as the field store opcode. | |
| 2603 // This forces the membar to follow the store. (Bug 6500685 broke this.) | |
| 2604 // All the other membars (for other volatile slices, including AliasIdxBot, | |
| 2605 // which stands for all unknown volatile slices) are control-dependent | |
| 2606 // on the first membar. This prevents later volatile loads or stores | |
| 2607 // from sliding up past the just-emitted store. | |
| 2608 | |
| 2609 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); | |
| 2610 mb->set_req(TypeFunc::Control,control()); | |
| 2611 if (alias_idx == Compile::AliasIdxBot) { | |
| 2612 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); | |
| 2613 } else { | |
| 2614 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); | |
| 2615 mb->set_req(TypeFunc::Memory, memory(alias_idx)); | |
| 2616 } | |
| 2617 Node* membar = _gvn.transform(mb); | |
| 2618 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); | |
| 2619 if (alias_idx == Compile::AliasIdxBot) { | |
| 2620 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); | |
| 2621 } else { | |
| 2622 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); | |
| 2623 } | |
| 2624 return membar; | |
| 2625 } | |
| 2626 | |
| 2627 //------------------------------shared_lock------------------------------------ | |
| 2628 // Emit locking code. | |
| 2629 FastLockNode* GraphKit::shared_lock(Node* obj) { | |
| 2630 // bci is either a monitorenter bc or InvocationEntryBci | |
| 2631 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces | |
| 2632 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); | |
| 2633 | |
| 2634 if( !GenerateSynchronizationCode ) | |
| 2635 return NULL; // Not locking things? | |
| 2636 if (stopped()) // Dead monitor? | |
| 2637 return NULL; | |
| 2638 | |
| 2639 assert(dead_locals_are_killed(), "should kill locals before sync. point"); | |
| 2640 | |
| 2641 // Box the stack location | |
| 2642 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); | |
| 2643 Node* mem = reset_memory(); | |
| 2644 | |
| 2645 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); | |
| 2646 if (PrintPreciseBiasedLockingStatistics) { | |
| 2647 // Create the counters for this fast lock. | |
| 2648 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci | |
| 2649 } | |
| 2650 // Add monitor to debug info for the slow path. If we block inside the | |
| 2651 // slow path and de-opt, we need the monitor hanging around | |
| 2652 map()->push_monitor( flock ); | |
| 2653 | |
| 2654 const TypeFunc *tf = LockNode::lock_type(); | |
| 2655 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); | |
| 2656 | |
| 2657 lock->init_req( TypeFunc::Control, control() ); | |
| 2658 lock->init_req( TypeFunc::Memory , mem ); | |
| 2659 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o | |
| 2660 lock->init_req( TypeFunc::FramePtr, frameptr() ); | |
| 2661 lock->init_req( TypeFunc::ReturnAdr, top() ); | |
| 2662 | |
| 2663 lock->init_req(TypeFunc::Parms + 0, obj); | |
| 2664 lock->init_req(TypeFunc::Parms + 1, box); | |
| 2665 lock->init_req(TypeFunc::Parms + 2, flock); | |
| 2666 add_safepoint_edges(lock); | |
| 2667 | |
| 2668 lock = _gvn.transform( lock )->as_Lock(); | |
| 2669 | |
| 2670 // lock has no side-effects, sets few values | |
| 2671 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); | |
| 2672 | |
| 2673 insert_mem_bar(Op_MemBarAcquire); | |
| 2674 | |
| 2675 // Add this to the worklist so that the lock can be eliminated | |
| 2676 record_for_igvn(lock); | |
| 2677 | |
| 2678 #ifndef PRODUCT | |
| 2679 if (PrintLockStatistics) { | |
| 2680 // Update the counter for this lock. Don't bother using an atomic | |
| 2681 // operation since we don't require absolute accuracy. | |
| 2682 lock->create_lock_counter(map()->jvms()); | |
| 2683 int adr_type = Compile::AliasIdxRaw; | |
| 2684 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr())); | |
| 2685 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); | |
| 2686 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); | |
| 2687 store_to_memory(control(), counter_addr, incr, T_INT, adr_type); | |
| 2688 } | |
| 2689 #endif | |
| 2690 | |
| 2691 return flock; | |
| 2692 } | |
| 2693 | |
| 2694 | |
| 2695 //------------------------------shared_unlock---------------------------------- | |
| 2696 // Emit unlocking code. | |
| 2697 void GraphKit::shared_unlock(Node* box, Node* obj) { | |
| 2698 // bci is either a monitorenter bc or InvocationEntryBci | |
| 2699 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces | |
| 2700 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); | |
| 2701 | |
| 2702 if( !GenerateSynchronizationCode ) | |
| 2703 return; | |
| 2704 if (stopped()) { // Dead monitor? | |
| 2705 map()->pop_monitor(); // Kill monitor from debug info | |
| 2706 return; | |
| 2707 } | |
| 2708 | |
| 2709 // Memory barrier to avoid floating things down past the locked region | |
| 2710 insert_mem_bar(Op_MemBarRelease); | |
| 2711 | |
| 2712 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); | |
| 2713 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); | |
| 2714 uint raw_idx = Compile::AliasIdxRaw; | |
| 2715 unlock->init_req( TypeFunc::Control, control() ); | |
| 2716 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); | |
| 2717 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o | |
| 2718 unlock->init_req( TypeFunc::FramePtr, frameptr() ); | |
| 2719 unlock->init_req( TypeFunc::ReturnAdr, top() ); | |
| 2720 | |
| 2721 unlock->init_req(TypeFunc::Parms + 0, obj); | |
| 2722 unlock->init_req(TypeFunc::Parms + 1, box); | |
| 2723 unlock = _gvn.transform(unlock)->as_Unlock(); | |
| 2724 | |
| 2725 Node* mem = reset_memory(); | |
| 2726 | |
| 2727 // unlock has no side-effects, sets few values | |
| 2728 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); | |
| 2729 | |
| 2730 // Kill monitor from debug info | |
| 2731 map()->pop_monitor( ); | |
| 2732 } | |
| 2733 | |
| 2734 //-------------------------------get_layout_helper----------------------------- | |
| 2735 // If the given klass is a constant or known to be an array, | |
| 2736 // fetch the constant layout helper value into constant_value | |
| 2737 // and return (Node*)NULL. Otherwise, load the non-constant | |
| 2738 // layout helper value, and return the node which represents it. | |
| 2739 // This two-faced routine is useful because allocation sites | |
| 2740 // almost always feature constant types. | |
| 2741 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { | |
| 2742 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); | |
| 2743 if (!StressReflectiveCode && inst_klass != NULL) { | |
| 2744 ciKlass* klass = inst_klass->klass(); | |
| 2745 bool xklass = inst_klass->klass_is_exact(); | |
| 2746 if (xklass || klass->is_array_klass()) { | |
| 2747 jint lhelper = klass->layout_helper(); | |
| 2748 if (lhelper != Klass::_lh_neutral_value) { | |
| 2749 constant_value = lhelper; | |
| 2750 return (Node*) NULL; | |
| 2751 } | |
| 2752 } | |
| 2753 } | |
| 2754 constant_value = Klass::_lh_neutral_value; // put in a known value | |
| 2755 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc)); | |
| 2756 return make_load(NULL, lhp, TypeInt::INT, T_INT); | |
| 2757 } | |
| 2758 | |
| 2759 // We just put in an allocate/initialize with a big raw-memory effect. | |
| 2760 // Hook selected additional alias categories on the initialization. | |
| 2761 static void hook_memory_on_init(GraphKit& kit, int alias_idx, | |
| 2762 MergeMemNode* init_in_merge, | |
| 2763 Node* init_out_raw) { | |
| 2764 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); | |
| 2765 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); | |
| 2766 | |
| 2767 Node* prevmem = kit.memory(alias_idx); | |
| 2768 init_in_merge->set_memory_at(alias_idx, prevmem); | |
| 2769 kit.set_memory(init_out_raw, alias_idx); | |
| 2770 } | |
| 2771 | |
| 2772 //---------------------------set_output_for_allocation------------------------- | |
| 2773 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, | |
| 2774 const TypeOopPtr* oop_type, | |
| 2775 bool raw_mem_only) { | |
| 2776 int rawidx = Compile::AliasIdxRaw; | |
| 2777 alloc->set_req( TypeFunc::FramePtr, frameptr() ); | |
| 2778 add_safepoint_edges(alloc); | |
| 2779 Node* allocx = _gvn.transform(alloc); | |
| 2780 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); | |
| 2781 // create memory projection for i_o | |
| 2782 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); | |
| 2783 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); | |
| 2784 | |
| 2785 // create a memory projection as for the normal control path | |
| 2786 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); | |
| 2787 set_memory(malloc, rawidx); | |
| 2788 | |
| 2789 // a normal slow-call doesn't change i_o, but an allocation does | |
| 2790 // we create a separate i_o projection for the normal control path | |
| 2791 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); | |
| 2792 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); | |
| 2793 | |
| 2794 // put in an initialization barrier | |
| 2795 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, | |
| 2796 rawoop)->as_Initialize(); | |
| 2797 assert(alloc->initialization() == init, "2-way macro link must work"); | |
| 2798 assert(init ->allocation() == alloc, "2-way macro link must work"); | |
| 2799 if (ReduceFieldZeroing && !raw_mem_only) { | |
| 2800 // Extract memory strands which may participate in the new object's | |
| 2801 // initialization, and source them from the new InitializeNode. | |
| 2802 // This will allow us to observe initializations when they occur, | |
| 2803 // and link them properly (as a group) to the InitializeNode. | |
| 2804 assert(init->in(InitializeNode::Memory) == malloc, ""); | |
| 2805 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); | |
| 2806 init->set_req(InitializeNode::Memory, minit_in); | |
| 2807 record_for_igvn(minit_in); // fold it up later, if possible | |
| 2808 Node* minit_out = memory(rawidx); | |
| 2809 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); | |
| 2810 if (oop_type->isa_aryptr()) { | |
| 2811 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); | |
| 2812 int elemidx = C->get_alias_index(telemref); | |
| 2813 hook_memory_on_init(*this, elemidx, minit_in, minit_out); | |
| 2814 } else if (oop_type->isa_instptr()) { | |
| 2815 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); | |
| 2816 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { | |
| 2817 ciField* field = ik->nonstatic_field_at(i); | |
|
23
9bdad1bb1c31
6621098: "* HeapWordSize" for TrackedInitializationLimit is missing
kvn
parents:
0
diff
changeset
|
2818 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) |
| 0 | 2819 continue; // do not bother to track really large numbers of fields |
| 2820 // Find (or create) the alias category for this field: | |
| 2821 int fieldidx = C->alias_type(field)->index(); | |
| 2822 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); | |
| 2823 } | |
| 2824 } | |
| 2825 } | |
| 2826 | |
| 2827 // Cast raw oop to the real thing... | |
| 2828 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); | |
| 2829 javaoop = _gvn.transform(javaoop); | |
| 2830 C->set_recent_alloc(control(), javaoop); | |
| 2831 assert(just_allocated_object(control()) == javaoop, "just allocated"); | |
| 2832 | |
| 2833 #ifdef ASSERT | |
| 2834 { // Verify that the AllocateNode::Ideal_foo recognizers work: | |
| 2835 Node* kn = alloc->in(AllocateNode::KlassNode); | |
| 2836 Node* ln = alloc->in(AllocateNode::ALength); | |
| 2837 assert(AllocateNode::Ideal_klass(rawoop, &_gvn) == kn, | |
| 2838 "Ideal_klass works"); | |
| 2839 assert(AllocateNode::Ideal_klass(javaoop, &_gvn) == kn, | |
| 2840 "Ideal_klass works"); | |
| 2841 if (alloc->is_AllocateArray()) { | |
| 2842 assert(AllocateArrayNode::Ideal_length(rawoop, &_gvn) == ln, | |
| 2843 "Ideal_length works"); | |
| 2844 assert(AllocateArrayNode::Ideal_length(javaoop, &_gvn) == ln, | |
| 2845 "Ideal_length works"); | |
| 2846 } else { | |
| 2847 assert(ln->is_top(), "no length, please"); | |
| 2848 } | |
| 2849 } | |
| 2850 #endif //ASSERT | |
| 2851 | |
| 2852 return javaoop; | |
| 2853 } | |
| 2854 | |
| 2855 //---------------------------new_instance-------------------------------------- | |
| 2856 // This routine takes a klass_node which may be constant (for a static type) | |
| 2857 // or may be non-constant (for reflective code). It will work equally well | |
| 2858 // for either, and the graph will fold nicely if the optimizer later reduces | |
| 2859 // the type to a constant. | |
| 2860 // The optional arguments are for specialized use by intrinsics: | |
| 2861 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. | |
| 2862 // - If 'raw_mem_only', do not cast the result to an oop. | |
| 2863 // - If 'return_size_val', report the the total object size to the caller. | |
| 2864 Node* GraphKit::new_instance(Node* klass_node, | |
| 2865 Node* extra_slow_test, | |
| 2866 bool raw_mem_only, // affect only raw memory | |
| 2867 Node* *return_size_val) { | |
| 2868 // Compute size in doublewords | |
| 2869 // The size is always an integral number of doublewords, represented | |
| 2870 // as a positive bytewise size stored in the klass's layout_helper. | |
| 2871 // The layout_helper also encodes (in a low bit) the need for a slow path. | |
| 2872 jint layout_con = Klass::_lh_neutral_value; | |
| 2873 Node* layout_val = get_layout_helper(klass_node, layout_con); | |
| 2874 int layout_is_con = (layout_val == NULL); | |
| 2875 | |
| 2876 if (extra_slow_test == NULL) extra_slow_test = intcon(0); | |
| 2877 // Generate the initial go-slow test. It's either ALWAYS (return a | |
| 2878 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective | |
| 2879 // case) a computed value derived from the layout_helper. | |
| 2880 Node* initial_slow_test = NULL; | |
| 2881 if (layout_is_con) { | |
| 2882 assert(!StressReflectiveCode, "stress mode does not use these paths"); | |
| 2883 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); | |
| 2884 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; | |
| 2885 | |
| 2886 } else { // reflective case | |
| 2887 // This reflective path is used by Unsafe.allocateInstance. | |
| 2888 // (It may be stress-tested by specifying StressReflectiveCode.) | |
| 2889 // Basically, we want to get into the VM is there's an illegal argument. | |
| 2890 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); | |
| 2891 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); | |
| 2892 if (extra_slow_test != intcon(0)) { | |
| 2893 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); | |
| 2894 } | |
| 2895 // (Macro-expander will further convert this to a Bool, if necessary.) | |
| 2896 } | |
| 2897 | |
| 2898 // Find the size in bytes. This is easy; it's the layout_helper. | |
| 2899 // The size value must be valid even if the slow path is taken. | |
| 2900 Node* size = NULL; | |
| 2901 if (layout_is_con) { | |
| 2902 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); | |
| 2903 } else { // reflective case | |
| 2904 // This reflective path is used by clone and Unsafe.allocateInstance. | |
| 2905 size = ConvI2X(layout_val); | |
| 2906 | |
| 2907 // Clear the low bits to extract layout_helper_size_in_bytes: | |
| 2908 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); | |
| 2909 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); | |
| 2910 size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); | |
| 2911 } | |
| 2912 if (return_size_val != NULL) { | |
| 2913 (*return_size_val) = size; | |
| 2914 } | |
| 2915 | |
| 2916 // This is a precise notnull oop of the klass. | |
| 2917 // (Actually, it need not be precise if this is a reflective allocation.) | |
| 2918 // It's what we cast the result to. | |
| 2919 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); | |
| 2920 if (!tklass) tklass = TypeKlassPtr::OBJECT; | |
| 2921 const TypeOopPtr* oop_type = tklass->as_instance_type(); | |
| 2922 | |
| 2923 // Now generate allocation code | |
|
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2924 |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2925 // With escape analysis, the entire memory state is needed to be able to |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2926 // eliminate the allocation. If the allocations cannot be eliminated, this |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2927 // will be optimized to the raw slice when the allocation is expanded. |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2928 Node *mem; |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2929 if (C->do_escape_analysis()) { |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2930 mem = reset_memory(); |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2931 set_all_memory(mem); |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2932 } else { |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2933 mem = memory(Compile::AliasIdxRaw); |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2934 } |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2935 |
| 0 | 2936 AllocateNode* alloc |
| 2937 = new (C, AllocateNode::ParmLimit) | |
| 2938 AllocateNode(C, AllocateNode::alloc_type(), | |
|
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
2939 control(), mem, i_o(), |
| 0 | 2940 size, klass_node, |
| 2941 initial_slow_test); | |
| 2942 | |
| 2943 return set_output_for_allocation(alloc, oop_type, raw_mem_only); | |
| 2944 } | |
| 2945 | |
| 2946 //-------------------------------new_array------------------------------------- | |
| 2947 // helper for both newarray and anewarray | |
| 2948 // The 'length' parameter is (obviously) the length of the array. | |
| 2949 // See comments on new_instance for the meaning of the other arguments. | |
| 2950 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) | |
| 2951 Node* length, // number of array elements | |
| 2952 bool raw_mem_only, // affect only raw memory | |
| 2953 Node* *return_size_val) { | |
| 2954 jint layout_con = Klass::_lh_neutral_value; | |
| 2955 Node* layout_val = get_layout_helper(klass_node, layout_con); | |
| 2956 int layout_is_con = (layout_val == NULL); | |
| 2957 | |
| 2958 if (!layout_is_con && !StressReflectiveCode && | |
| 2959 !too_many_traps(Deoptimization::Reason_class_check)) { | |
| 2960 // This is a reflective array creation site. | |
| 2961 // Optimistically assume that it is a subtype of Object[], | |
| 2962 // so that we can fold up all the address arithmetic. | |
| 2963 layout_con = Klass::array_layout_helper(T_OBJECT); | |
| 2964 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); | |
| 2965 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); | |
| 2966 { BuildCutout unless(this, bol_lh, PROB_MAX); | |
| 2967 uncommon_trap(Deoptimization::Reason_class_check, | |
| 2968 Deoptimization::Action_maybe_recompile); | |
| 2969 } | |
| 2970 layout_val = NULL; | |
| 2971 layout_is_con = true; | |
| 2972 } | |
| 2973 | |
| 2974 // Generate the initial go-slow test. Make sure we do not overflow | |
| 2975 // if length is huge (near 2Gig) or negative! We do not need | |
| 2976 // exact double-words here, just a close approximation of needed | |
| 2977 // double-words. We can't add any offset or rounding bits, lest we | |
| 2978 // take a size -1 of bytes and make it positive. Use an unsigned | |
| 2979 // compare, so negative sizes look hugely positive. | |
| 2980 int fast_size_limit = FastAllocateSizeLimit; | |
| 2981 if (layout_is_con) { | |
| 2982 assert(!StressReflectiveCode, "stress mode does not use these paths"); | |
| 2983 // Increase the size limit if we have exact knowledge of array type. | |
| 2984 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); | |
| 2985 fast_size_limit <<= (LogBytesPerLong - log2_esize); | |
| 2986 } | |
| 2987 | |
| 2988 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); | |
| 2989 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); | |
| 2990 if (initial_slow_test->is_Bool()) { | |
| 2991 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. | |
| 2992 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); | |
| 2993 } | |
| 2994 | |
| 2995 // --- Size Computation --- | |
| 2996 // array_size = round_to_heap(array_header + (length << elem_shift)); | |
| 2997 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) | |
| 2998 // and round_to(x, y) == ((x + y-1) & ~(y-1)) | |
| 2999 // The rounding mask is strength-reduced, if possible. | |
| 3000 int round_mask = MinObjAlignmentInBytes - 1; | |
| 3001 Node* header_size = NULL; | |
| 3002 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); | |
| 3003 // (T_BYTE has the weakest alignment and size restrictions...) | |
| 3004 if (layout_is_con) { | |
| 3005 int hsize = Klass::layout_helper_header_size(layout_con); | |
| 3006 int eshift = Klass::layout_helper_log2_element_size(layout_con); | |
| 3007 BasicType etype = Klass::layout_helper_element_type(layout_con); | |
| 3008 if ((round_mask & ~right_n_bits(eshift)) == 0) | |
| 3009 round_mask = 0; // strength-reduce it if it goes away completely | |
| 3010 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); | |
| 3011 assert(header_size_min <= hsize, "generic minimum is smallest"); | |
| 3012 header_size_min = hsize; | |
| 3013 header_size = intcon(hsize + round_mask); | |
| 3014 } else { | |
| 3015 Node* hss = intcon(Klass::_lh_header_size_shift); | |
| 3016 Node* hsm = intcon(Klass::_lh_header_size_mask); | |
| 3017 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); | |
| 3018 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); | |
| 3019 Node* mask = intcon(round_mask); | |
| 3020 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); | |
| 3021 } | |
| 3022 | |
| 3023 Node* elem_shift = NULL; | |
| 3024 if (layout_is_con) { | |
| 3025 int eshift = Klass::layout_helper_log2_element_size(layout_con); | |
| 3026 if (eshift != 0) | |
| 3027 elem_shift = intcon(eshift); | |
| 3028 } else { | |
| 3029 // There is no need to mask or shift this value. | |
| 3030 // The semantics of LShiftINode include an implicit mask to 0x1F. | |
| 3031 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); | |
| 3032 elem_shift = layout_val; | |
| 3033 } | |
| 3034 | |
| 3035 // Transition to native address size for all offset calculations: | |
| 3036 Node* lengthx = ConvI2X(length); | |
| 3037 Node* headerx = ConvI2X(header_size); | |
| 3038 #ifdef _LP64 | |
| 3039 { const TypeLong* tllen = _gvn.find_long_type(lengthx); | |
| 3040 if (tllen != NULL && tllen->_lo < 0) { | |
| 3041 // Add a manual constraint to a positive range. Cf. array_element_address. | |
| 3042 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); | |
| 3043 if (size_max > tllen->_hi) size_max = tllen->_hi; | |
| 3044 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); | |
| 3045 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); | |
| 3046 } | |
| 3047 } | |
| 3048 #endif | |
| 3049 | |
| 3050 // Combine header size (plus rounding) and body size. Then round down. | |
| 3051 // This computation cannot overflow, because it is used only in two | |
| 3052 // places, one where the length is sharply limited, and the other | |
| 3053 // after a successful allocation. | |
| 3054 Node* abody = lengthx; | |
| 3055 if (elem_shift != NULL) | |
| 3056 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); | |
| 3057 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); | |
| 3058 if (round_mask != 0) { | |
| 3059 Node* mask = MakeConX(~round_mask); | |
| 3060 size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); | |
| 3061 } | |
| 3062 // else if round_mask == 0, the size computation is self-rounding | |
| 3063 | |
| 3064 if (return_size_val != NULL) { | |
| 3065 // This is the size | |
| 3066 (*return_size_val) = size; | |
| 3067 } | |
| 3068 | |
| 3069 // Now generate allocation code | |
|
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3070 |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3071 // With escape analysis, the entire memory state is needed to be able to |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3072 // eliminate the allocation. If the allocations cannot be eliminated, this |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3073 // will be optimized to the raw slice when the allocation is expanded. |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3074 Node *mem; |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3075 if (C->do_escape_analysis()) { |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3076 mem = reset_memory(); |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3077 set_all_memory(mem); |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3078 } else { |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3079 mem = memory(Compile::AliasIdxRaw); |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3080 } |
|
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3081 |
| 0 | 3082 // Create the AllocateArrayNode and its result projections |
| 3083 AllocateArrayNode* alloc | |
| 3084 = new (C, AllocateArrayNode::ParmLimit) | |
| 3085 AllocateArrayNode(C, AllocateArrayNode::alloc_type(), | |
|
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
63
diff
changeset
|
3086 control(), mem, i_o(), |
| 0 | 3087 size, klass_node, |
| 3088 initial_slow_test, | |
| 3089 length); | |
| 3090 | |
| 3091 // Cast to correct type. Note that the klass_node may be constant or not, | |
| 3092 // and in the latter case the actual array type will be inexact also. | |
| 3093 // (This happens via a non-constant argument to inline_native_newArray.) | |
| 3094 // In any case, the value of klass_node provides the desired array type. | |
| 3095 const TypeInt* length_type = _gvn.find_int_type(length); | |
| 3096 const TypeInt* narrow_length_type = NULL; | |
| 3097 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); | |
| 3098 if (ary_type->isa_aryptr() && length_type != NULL) { | |
| 3099 // Try to get a better type than POS for the size | |
| 3100 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); | |
| 3101 narrow_length_type = ary_type->is_aryptr()->size(); | |
| 3102 if (narrow_length_type == length_type) | |
| 3103 narrow_length_type = NULL; | |
| 3104 } | |
| 3105 | |
| 3106 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only); | |
| 3107 | |
| 3108 // Cast length on remaining path to be positive: | |
| 3109 if (narrow_length_type != NULL) { | |
| 3110 Node* ccast = new (C, 2) CastIINode(length, narrow_length_type); | |
| 3111 ccast->set_req(0, control()); | |
| 3112 _gvn.set_type_bottom(ccast); | |
| 3113 record_for_igvn(ccast); | |
| 3114 if (map()->find_edge(length) >= 0) { | |
| 3115 replace_in_map(length, ccast); | |
| 3116 } | |
| 3117 } | |
| 3118 | |
| 3119 return javaoop; | |
| 3120 } | |
| 3121 | |
| 3122 // The following "Ideal_foo" functions are placed here because they recognize | |
| 3123 // the graph shapes created by the functions immediately above. | |
| 3124 | |
| 3125 //---------------------------Ideal_allocation---------------------------------- | |
| 3126 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. | |
| 3127 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { | |
| 3128 if (ptr == NULL) { // reduce dumb test in callers | |
| 3129 return NULL; | |
| 3130 } | |
| 3131 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast | |
| 3132 ptr = ptr->in(1); | |
| 3133 if (ptr == NULL) return NULL; | |
| 3134 } | |
| 3135 if (ptr->is_Proj()) { | |
| 3136 Node* allo = ptr->in(0); | |
| 3137 if (allo != NULL && allo->is_Allocate()) { | |
| 3138 return allo->as_Allocate(); | |
| 3139 } | |
| 3140 } | |
| 3141 // Report failure to match. | |
| 3142 return NULL; | |
| 3143 } | |
| 3144 | |
| 3145 // Fancy version which also strips off an offset (and reports it to caller). | |
| 3146 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, | |
| 3147 intptr_t& offset) { | |
| 3148 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); | |
| 3149 if (base == NULL) return NULL; | |
| 3150 return Ideal_allocation(base, phase); | |
| 3151 } | |
| 3152 | |
| 3153 // Trace Initialize <- Proj[Parm] <- Allocate | |
| 3154 AllocateNode* InitializeNode::allocation() { | |
| 3155 Node* rawoop = in(InitializeNode::RawAddress); | |
| 3156 if (rawoop->is_Proj()) { | |
| 3157 Node* alloc = rawoop->in(0); | |
| 3158 if (alloc->is_Allocate()) { | |
| 3159 return alloc->as_Allocate(); | |
| 3160 } | |
| 3161 } | |
| 3162 return NULL; | |
| 3163 } | |
| 3164 | |
| 3165 // Trace Allocate -> Proj[Parm] -> Initialize | |
| 3166 InitializeNode* AllocateNode::initialization() { | |
| 3167 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); | |
| 3168 if (rawoop == NULL) return NULL; | |
| 3169 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { | |
| 3170 Node* init = rawoop->fast_out(i); | |
| 3171 if (init->is_Initialize()) { | |
| 3172 assert(init->as_Initialize()->allocation() == this, "2-way link"); | |
| 3173 return init->as_Initialize(); | |
| 3174 } | |
| 3175 } | |
| 3176 return NULL; | |
| 3177 } |