1 /* 2 * Copyright (c) 2015, 2019, Red Hat, Inc. All rights reserved. 3 * 4 * This code is free software; you can redistribute it and/or modify it 5 * under the terms of the GNU General Public License version 2 only, as 6 * published by the Free Software Foundation. 7 * 8 * This code is distributed in the hope that it will be useful, but WITHOUT 9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 11 * version 2 for more details (a copy is included in the LICENSE file that 12 * accompanied this code). 13 * 14 * You should have received a copy of the GNU General Public License version 15 * 2 along with this work; if not, write to the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 17 * 18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 19 * or visit www.oracle.com if you need additional information or have any 20 * questions. 21 * 22 */ 23 24 #include "precompiled.hpp" 25 26 #include "gc/shenandoah/c2/shenandoahSupport.hpp" 27 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 28 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp" 29 #include "gc/shenandoah/shenandoahForwarding.hpp" 30 #include "gc/shenandoah/shenandoahHeap.hpp" 31 #include "gc/shenandoah/shenandoahHeapRegion.hpp" 32 #include "gc/shenandoah/shenandoahRuntime.hpp" 33 #include "gc/shenandoah/shenandoahThreadLocalData.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/block.hpp" 36 #include "opto/callnode.hpp" 37 #include "opto/castnode.hpp" 38 #include "opto/movenode.hpp" 39 #include "opto/phaseX.hpp" 40 #include "opto/rootnode.hpp" 41 #include "opto/runtime.hpp" 42 #include "opto/subnode.hpp" 43 44 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) { 45 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state(); 46 if ((state->enqueue_barriers_count() + 47 state->load_reference_barriers_count()) > 0) { 48 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion; 49 C->clear_major_progress(); 50 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand); 51 if (C->failing()) return false; 52 PhaseIdealLoop::verify(igvn); 53 DEBUG_ONLY(verify_raw_mem(C->root());) 54 if (attempt_more_loopopts) { 55 C->set_major_progress(); 56 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) { 57 return false; 58 } 59 C->clear_major_progress(); 60 } 61 } 62 return true; 63 } 64 65 bool ShenandoahBarrierC2Support::is_heap_state_test(Node* iff, int mask) { 66 if (!UseShenandoahGC) { 67 return false; 68 } 69 assert(iff->is_If(), "bad input"); 70 if (iff->Opcode() != Op_If) { 71 return false; 72 } 73 Node* bol = iff->in(1); 74 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) { 75 return false; 76 } 77 Node* cmp = bol->in(1); 78 if (cmp->Opcode() != Op_CmpI) { 79 return false; 80 } 81 Node* in1 = cmp->in(1); 82 Node* in2 = cmp->in(2); 83 if (in2->find_int_con(-1) != 0) { 84 return false; 85 } 86 if (in1->Opcode() != Op_AndI) { 87 return false; 88 } 89 in2 = in1->in(2); 90 if (in2->find_int_con(-1) != mask) { 91 return false; 92 } 93 in1 = in1->in(1); 94 95 return is_gc_state_load(in1); 96 } 97 98 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) { 99 return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED); 100 } 101 102 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) { 103 if (!UseShenandoahGC) { 104 return false; 105 } 106 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) { 107 return false; 108 } 109 Node* addp = n->in(MemNode::Address); 110 if (!addp->is_AddP()) { 111 return false; 112 } 113 Node* base = addp->in(AddPNode::Address); 114 Node* off = addp->in(AddPNode::Offset); 115 if (base->Opcode() != Op_ThreadLocal) { 116 return false; 117 } 118 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) { 119 return false; 120 } 121 return true; 122 } 123 124 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) { 125 assert(phase->is_dominator(stop, start), "bad inputs"); 126 ResourceMark rm; 127 Unique_Node_List wq; 128 wq.push(start); 129 for (uint next = 0; next < wq.size(); next++) { 130 Node *m = wq.at(next); 131 if (m == stop) { 132 continue; 133 } 134 if (m->is_SafePoint() && !m->is_CallLeaf()) { 135 return true; 136 } 137 if (m->is_Region()) { 138 for (uint i = 1; i < m->req(); i++) { 139 wq.push(m->in(i)); 140 } 141 } else { 142 wq.push(m->in(0)); 143 } 144 } 145 return false; 146 } 147 148 bool ShenandoahBarrierC2Support::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) { 149 assert(is_gc_state_load(n), "inconsistent"); 150 Node* addp = n->in(MemNode::Address); 151 Node* dominator = NULL; 152 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) { 153 Node* u = addp->fast_out(i); 154 assert(is_gc_state_load(u), "inconsistent"); 155 if (u != n && phase->is_dominator(u->in(0), n->in(0))) { 156 if (dominator == NULL) { 157 dominator = u; 158 } else { 159 if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) { 160 dominator = u; 161 } 162 } 163 } 164 } 165 if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) { 166 return false; 167 } 168 phase->igvn().replace_node(n, dominator); 169 170 return true; 171 } 172 173 #ifdef ASSERT 174 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) { 175 assert(phis.size() == 0, ""); 176 177 while (true) { 178 if (in->bottom_type() == TypePtr::NULL_PTR) { 179 if (trace) {tty->print_cr("NULL");} 180 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) { 181 if (trace) {tty->print_cr("Non oop");} 182 } else if (t == ShenandoahLoad && ShenandoahOptimizeStableFinals && 183 in->bottom_type()->make_ptr()->isa_aryptr() && 184 in->bottom_type()->make_ptr()->is_aryptr()->is_stable()) { 185 if (trace) {tty->print_cr("Stable array load");} 186 } else { 187 if (in->is_ConstraintCast()) { 188 in = in->in(1); 189 continue; 190 } else if (in->is_AddP()) { 191 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access"); 192 in = in->in(AddPNode::Address); 193 continue; 194 } else if (in->is_Con()) { 195 if (trace) { 196 tty->print("Found constant"); 197 in->dump(); 198 } 199 } else if (in->Opcode() == Op_Parm) { 200 if (trace) { 201 tty->print("Found argument"); 202 } 203 } else if (in->Opcode() == Op_CreateEx) { 204 if (trace) { 205 tty->print("Found create-exception"); 206 } 207 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) { 208 if (trace) { 209 tty->print("Found raw LoadP (OSR argument?)"); 210 } 211 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 212 if (t == ShenandoahOopStore) { 213 uint i = 0; 214 for (; i < phis.size(); i++) { 215 Node* n = phis.node_at(i); 216 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) { 217 break; 218 } 219 } 220 if (i == phis.size()) { 221 return false; 222 } 223 } 224 barriers_used.push(in); 225 if (trace) {tty->print("Found barrier"); in->dump();} 226 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) { 227 if (t != ShenandoahOopStore) { 228 in = in->in(1); 229 continue; 230 } 231 if (trace) {tty->print("Found enqueue barrier"); in->dump();} 232 phis.push(in, in->req()); 233 in = in->in(1); 234 continue; 235 } else if (in->is_Proj() && in->in(0)->is_Allocate()) { 236 if (trace) { 237 tty->print("Found alloc"); 238 in->in(0)->dump(); 239 } 240 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) { 241 if (trace) { 242 tty->print("Found Java call"); 243 } 244 } else if (in->is_Phi()) { 245 if (!visited.test_set(in->_idx)) { 246 if (trace) {tty->print("Pushed phi:"); in->dump();} 247 phis.push(in, 2); 248 in = in->in(1); 249 continue; 250 } 251 if (trace) {tty->print("Already seen phi:"); in->dump();} 252 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) { 253 if (!visited.test_set(in->_idx)) { 254 if (trace) {tty->print("Pushed cmovep:"); in->dump();} 255 phis.push(in, CMoveNode::IfTrue); 256 in = in->in(CMoveNode::IfFalse); 257 continue; 258 } 259 if (trace) {tty->print("Already seen cmovep:"); in->dump();} 260 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) { 261 in = in->in(1); 262 continue; 263 } else { 264 return false; 265 } 266 } 267 bool cont = false; 268 while (phis.is_nonempty()) { 269 uint idx = phis.index(); 270 Node* phi = phis.node(); 271 if (idx >= phi->req()) { 272 if (trace) {tty->print("Popped phi:"); phi->dump();} 273 phis.pop(); 274 continue; 275 } 276 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();} 277 in = phi->in(idx); 278 phis.set_index(idx+1); 279 cont = true; 280 break; 281 } 282 if (!cont) { 283 break; 284 } 285 } 286 return true; 287 } 288 289 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) { 290 if (n1 != NULL) { 291 n1->dump(+10); 292 } 293 if (n2 != NULL) { 294 n2->dump(+10); 295 } 296 fatal("%s", msg); 297 } 298 299 void ShenandoahBarrierC2Support::verify(RootNode* root) { 300 ResourceMark rm; 301 Unique_Node_List wq; 302 GrowableArray<Node*> barriers; 303 Unique_Node_List barriers_used; 304 Node_Stack phis(0); 305 VectorSet visited(Thread::current()->resource_area()); 306 const bool trace = false; 307 const bool verify_no_useless_barrier = false; 308 309 wq.push(root); 310 for (uint next = 0; next < wq.size(); next++) { 311 Node *n = wq.at(next); 312 if (n->is_Load()) { 313 const bool trace = false; 314 if (trace) {tty->print("Verifying"); n->dump();} 315 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) { 316 if (trace) {tty->print_cr("Load range/klass");} 317 } else { 318 const TypePtr* adr_type = n->as_Load()->adr_type(); 319 320 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) { 321 if (trace) {tty->print_cr("Mark load");} 322 } else if (adr_type->isa_instptr() && 323 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) && 324 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) { 325 if (trace) {tty->print_cr("Reference.get()");} 326 } else { 327 bool verify = true; 328 if (adr_type->isa_instptr()) { 329 const TypeInstPtr* tinst = adr_type->is_instptr(); 330 ciKlass* k = tinst->klass(); 331 assert(k->is_instance_klass(), ""); 332 ciInstanceKlass* ik = (ciInstanceKlass*)k; 333 int offset = adr_type->offset(); 334 335 if ((ik->debug_final_field_at(offset) && ShenandoahOptimizeInstanceFinals) || 336 (ik->debug_stable_field_at(offset) && ShenandoahOptimizeStableFinals)) { 337 if (trace) {tty->print_cr("Final/stable");} 338 verify = false; 339 } else if (k == ciEnv::current()->Class_klass() && 340 tinst->const_oop() != NULL && 341 tinst->offset() >= (ik->size_helper() * wordSize)) { 342 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 343 ciField* field = k->get_field_by_offset(tinst->offset(), true); 344 if ((ShenandoahOptimizeStaticFinals && field->is_final()) || 345 (ShenandoahOptimizeStableFinals && field->is_stable())) { 346 verify = false; 347 } 348 } 349 } 350 351 if (verify && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) { 352 report_verify_failure("Shenandoah verification: Load should have barriers", n); 353 } 354 } 355 } 356 } else if (n->is_Store()) { 357 const bool trace = false; 358 359 if (trace) {tty->print("Verifying"); n->dump();} 360 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) { 361 Node* adr = n->in(MemNode::Address); 362 bool verify = true; 363 364 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) { 365 adr = adr->in(AddPNode::Address); 366 if (adr->is_AddP()) { 367 assert(adr->in(AddPNode::Base)->is_top(), ""); 368 adr = adr->in(AddPNode::Address); 369 if (adr->Opcode() == Op_LoadP && 370 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() && 371 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 372 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) { 373 if (trace) {tty->print_cr("SATB prebarrier");} 374 verify = false; 375 } 376 } 377 } 378 379 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) { 380 report_verify_failure("Shenandoah verification: Store should have barriers", n); 381 } 382 } 383 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) { 384 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n); 385 } 386 } else if (n->Opcode() == Op_CmpP) { 387 const bool trace = false; 388 389 Node* in1 = n->in(1); 390 Node* in2 = n->in(2); 391 if (in1->bottom_type()->isa_oopptr()) { 392 if (trace) {tty->print("Verifying"); n->dump();} 393 394 bool mark_inputs = false; 395 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR || 396 (in1->is_Con() || in2->is_Con())) { 397 if (trace) {tty->print_cr("Comparison against a constant");} 398 mark_inputs = true; 399 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) || 400 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) { 401 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");} 402 mark_inputs = true; 403 } else { 404 assert(in2->bottom_type()->isa_oopptr(), ""); 405 406 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) || 407 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) { 408 report_verify_failure("Shenandoah verification: Cmp should have barriers", n); 409 } 410 } 411 if (verify_no_useless_barrier && 412 mark_inputs && 413 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) || 414 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) { 415 phis.clear(); 416 visited.Reset(); 417 } 418 } 419 } else if (n->is_LoadStore()) { 420 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() && 421 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) { 422 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n); 423 } 424 425 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) { 426 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n); 427 } 428 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) { 429 CallNode* call = n->as_Call(); 430 431 static struct { 432 const char* name; 433 struct { 434 int pos; 435 verify_type t; 436 } args[6]; 437 } calls[] = { 438 "aescrypt_encryptBlock", 439 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 440 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 441 "aescrypt_decryptBlock", 442 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 443 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 444 "multiplyToLen", 445 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore }, 446 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 447 "squareToLen", 448 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone}, 449 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 450 "montgomery_multiply", 451 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, 452 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 453 "montgomery_square", 454 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore }, 455 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 456 "mulAdd", 457 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, 458 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 459 "vectorizedMismatch", 460 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, 461 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 462 "updateBytesCRC32", 463 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 464 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 465 "updateBytesAdler32", 466 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 467 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 468 "updateBytesCRC32C", 469 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone}, 470 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 471 "counterMode_AESCrypt", 472 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 473 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } }, 474 "cipherBlockChaining_encryptAESCrypt", 475 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 476 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 477 "cipherBlockChaining_decryptAESCrypt", 478 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 479 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 480 "shenandoah_clone_barrier", 481 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 482 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 483 "ghash_processBlocks", 484 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, 485 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 486 "sha1_implCompress", 487 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 488 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 489 "sha256_implCompress", 490 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 491 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 492 "sha512_implCompress", 493 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 494 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 495 "sha1_implCompressMB", 496 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 497 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 498 "sha256_implCompressMB", 499 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 500 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 501 "sha512_implCompressMB", 502 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 503 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 504 "encodeBlock", 505 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone }, 506 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 507 }; 508 509 if (call->is_call_to_arraycopystub()) { 510 Node* dest = NULL; 511 const TypeTuple* args = n->as_Call()->_tf->domain(); 512 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) { 513 if (args->field_at(i)->isa_ptr()) { 514 j++; 515 if (j == 2) { 516 dest = n->in(i); 517 break; 518 } 519 } 520 } 521 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) || 522 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) { 523 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n); 524 } 525 } else if (strlen(call->_name) > 5 && 526 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) { 527 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) { 528 report_verify_failure("Shenandoah verification: _fill should have barriers", n); 529 } 530 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) { 531 // skip 532 } else { 533 const int calls_len = sizeof(calls) / sizeof(calls[0]); 534 int i = 0; 535 for (; i < calls_len; i++) { 536 if (!strcmp(calls[i].name, call->_name)) { 537 break; 538 } 539 } 540 if (i != calls_len) { 541 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]); 542 for (uint j = 0; j < args_len; j++) { 543 int pos = calls[i].args[j].pos; 544 if (pos == -1) { 545 break; 546 } 547 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) { 548 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n); 549 } 550 } 551 for (uint j = TypeFunc::Parms; j < call->req(); j++) { 552 if (call->in(j)->bottom_type()->make_ptr() && 553 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) { 554 uint k = 0; 555 for (; k < args_len && calls[i].args[k].pos != (int)j; k++); 556 if (k == args_len) { 557 fatal("arg %d for call %s not covered", j, call->_name); 558 } 559 } 560 } 561 } else { 562 for (uint j = TypeFunc::Parms; j < call->req(); j++) { 563 if (call->in(j)->bottom_type()->make_ptr() && 564 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) { 565 fatal("%s not covered", call->_name); 566 } 567 } 568 } 569 } 570 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 571 // skip 572 } else if (n->is_AddP() 573 || n->is_Phi() 574 || n->is_ConstraintCast() 575 || n->Opcode() == Op_Return 576 || n->Opcode() == Op_CMoveP 577 || n->Opcode() == Op_CMoveN 578 || n->Opcode() == Op_Rethrow 579 || n->is_MemBar() 580 || n->Opcode() == Op_Conv2B 581 || n->Opcode() == Op_SafePoint 582 || n->is_CallJava() 583 || n->Opcode() == Op_Unlock 584 || n->Opcode() == Op_EncodeP 585 || n->Opcode() == Op_DecodeN) { 586 // nothing to do 587 } else { 588 static struct { 589 int opcode; 590 struct { 591 int pos; 592 verify_type t; 593 } inputs[2]; 594 } others[] = { 595 Op_FastLock, 596 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} }, 597 Op_Lock, 598 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} }, 599 Op_ArrayCopy, 600 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } }, 601 Op_StrCompressedCopy, 602 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 603 Op_StrInflatedCopy, 604 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 605 Op_AryEq, 606 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } }, 607 Op_StrIndexOf, 608 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } }, 609 Op_StrComp, 610 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } }, 611 Op_StrEquals, 612 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } }, 613 Op_EncodeISOArray, 614 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 615 Op_HasNegatives, 616 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} }, 617 Op_CastP2X, 618 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} }, 619 Op_StrIndexOfChar, 620 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } }, 621 }; 622 623 const int others_len = sizeof(others) / sizeof(others[0]); 624 int i = 0; 625 for (; i < others_len; i++) { 626 if (others[i].opcode == n->Opcode()) { 627 break; 628 } 629 } 630 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req(); 631 if (i != others_len) { 632 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]); 633 for (uint j = 0; j < inputs_len; j++) { 634 int pos = others[i].inputs[j].pos; 635 if (pos == -1) { 636 break; 637 } 638 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) { 639 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n); 640 } 641 } 642 for (uint j = 1; j < stop; j++) { 643 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() && 644 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) { 645 uint k = 0; 646 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++); 647 if (k == inputs_len) { 648 fatal("arg %d for node %s not covered", j, n->Name()); 649 } 650 } 651 } 652 } else { 653 for (uint j = 1; j < stop; j++) { 654 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() && 655 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) { 656 fatal("%s not covered", n->Name()); 657 } 658 } 659 } 660 } 661 662 if (n->is_SafePoint()) { 663 SafePointNode* sfpt = n->as_SafePoint(); 664 if (verify_no_useless_barrier && sfpt->jvms() != NULL) { 665 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) { 666 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) { 667 phis.clear(); 668 visited.Reset(); 669 } 670 } 671 } 672 } 673 } 674 675 if (verify_no_useless_barrier) { 676 for (int i = 0; i < barriers.length(); i++) { 677 Node* n = barriers.at(i); 678 if (!barriers_used.member(n)) { 679 tty->print("XXX useless barrier"); n->dump(-2); 680 ShouldNotReachHere(); 681 } 682 } 683 } 684 } 685 #endif 686 687 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) { 688 // That both nodes have the same control is not sufficient to prove 689 // domination, verify that there's no path from d to n 690 ResourceMark rm; 691 Unique_Node_List wq; 692 wq.push(d); 693 for (uint next = 0; next < wq.size(); next++) { 694 Node *m = wq.at(next); 695 if (m == n) { 696 return false; 697 } 698 if (m->is_Phi() && m->in(0)->is_Loop()) { 699 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control"); 700 } else { 701 for (uint i = 0; i < m->req(); i++) { 702 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) { 703 wq.push(m->in(i)); 704 } 705 } 706 } 707 } 708 return true; 709 } 710 711 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) { 712 if (d_c != n_c) { 713 return phase->is_dominator(d_c, n_c); 714 } 715 return is_dominator_same_ctrl(d_c, d, n, phase); 716 } 717 718 Node* next_mem(Node* mem, int alias) { 719 Node* res = NULL; 720 if (mem->is_Proj()) { 721 res = mem->in(0); 722 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 723 res = mem->in(TypeFunc::Memory); 724 } else if (mem->is_Phi()) { 725 res = mem->in(1); 726 } else if (mem->is_MergeMem()) { 727 res = mem->as_MergeMem()->memory_at(alias); 728 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 729 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier"); 730 res = mem->in(MemNode::Memory); 731 } else { 732 #ifdef ASSERT 733 mem->dump(); 734 #endif 735 ShouldNotReachHere(); 736 } 737 return res; 738 } 739 740 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) { 741 Node* iffproj = NULL; 742 while (c != dom) { 743 Node* next = phase->idom(c); 744 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?"); 745 if (c->is_Region()) { 746 ResourceMark rm; 747 Unique_Node_List wq; 748 wq.push(c); 749 for (uint i = 0; i < wq.size(); i++) { 750 Node *n = wq.at(i); 751 if (n == next) { 752 continue; 753 } 754 if (n->is_Region()) { 755 for (uint j = 1; j < n->req(); j++) { 756 wq.push(n->in(j)); 757 } 758 } else { 759 wq.push(n->in(0)); 760 } 761 } 762 for (uint i = 0; i < wq.size(); i++) { 763 Node *n = wq.at(i); 764 assert(n->is_CFG(), ""); 765 if (n->is_Multi()) { 766 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 767 Node* u = n->fast_out(j); 768 if (u->is_CFG()) { 769 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 770 return NodeSentinel; 771 } 772 } 773 } 774 } 775 } 776 } else if (c->is_Proj()) { 777 if (c->is_IfProj()) { 778 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) { 779 // continue; 780 } else { 781 if (!allow_one_proj) { 782 return NodeSentinel; 783 } 784 if (iffproj == NULL) { 785 iffproj = c; 786 } else { 787 return NodeSentinel; 788 } 789 } 790 } else if (c->Opcode() == Op_JumpProj) { 791 return NodeSentinel; // unsupported 792 } else if (c->Opcode() == Op_CatchProj) { 793 return NodeSentinel; // unsupported 794 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) { 795 return NodeSentinel; // unsupported 796 } else { 797 assert(next->unique_ctrl_out() == c, "unsupported branch pattern"); 798 } 799 } 800 c = next; 801 } 802 return iffproj; 803 } 804 805 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) { 806 ResourceMark rm; 807 VectorSet wq(Thread::current()->resource_area()); 808 wq.set(mem->_idx); 809 mem_ctrl = phase->ctrl_or_self(mem); 810 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) { 811 mem = next_mem(mem, alias); 812 if (wq.test_set(mem->_idx)) { 813 return NULL; 814 } 815 mem_ctrl = phase->ctrl_or_self(mem); 816 } 817 if (mem->is_MergeMem()) { 818 mem = mem->as_MergeMem()->memory_at(alias); 819 mem_ctrl = phase->ctrl_or_self(mem); 820 } 821 return mem; 822 } 823 824 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) { 825 Node* mem = NULL; 826 Node* c = ctrl; 827 do { 828 if (c->is_Region()) { 829 Node* phi_bottom = NULL; 830 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) { 831 Node* u = c->fast_out(i); 832 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) { 833 if (u->adr_type() == TypePtr::BOTTOM) { 834 mem = u; 835 } 836 } 837 } 838 } else { 839 if (c->is_Call() && c->as_Call()->adr_type() != NULL) { 840 CallProjections projs; 841 c->as_Call()->extract_projections(&projs, true, false); 842 if (projs.fallthrough_memproj != NULL) { 843 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) { 844 if (projs.catchall_memproj == NULL) { 845 mem = projs.fallthrough_memproj; 846 } else { 847 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) { 848 mem = projs.fallthrough_memproj; 849 } else { 850 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier"); 851 mem = projs.catchall_memproj; 852 } 853 } 854 } 855 } else { 856 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory); 857 if (proj != NULL && 858 proj->adr_type() == TypePtr::BOTTOM) { 859 mem = proj; 860 } 861 } 862 } else { 863 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 864 Node* u = c->fast_out(i); 865 if (u->is_Proj() && 866 u->bottom_type() == Type::MEMORY && 867 u->adr_type() == TypePtr::BOTTOM) { 868 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), ""); 869 assert(mem == NULL, "only one proj"); 870 mem = u; 871 } 872 } 873 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected"); 874 } 875 } 876 c = phase->idom(c); 877 } while (mem == NULL); 878 return mem; 879 } 880 881 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) { 882 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 883 Node* u = n->fast_out(i); 884 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) { 885 uses.push(u); 886 } 887 } 888 } 889 890 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) { 891 OuterStripMinedLoopEndNode* le = inner->outer_loop_end(); 892 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl)); 893 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl)); 894 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt); 895 phase->register_control(new_le, phase->get_loop(le), le->in(0)); 896 phase->lazy_replace(outer, new_outer); 897 phase->lazy_replace(le, new_le); 898 inner->clear_strip_mined(); 899 } 900 901 void ShenandoahBarrierC2Support::test_heap_stable(Node*& ctrl, Node* raw_mem, Node*& heap_stable_ctrl, 902 PhaseIdealLoop* phase) { 903 IdealLoopTree* loop = phase->get_loop(ctrl); 904 Node* thread = new ThreadLocalNode(); 905 phase->register_new_node(thread, ctrl); 906 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 907 phase->set_ctrl(offset, phase->C->root()); 908 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset); 909 phase->register_new_node(gc_state_addr, ctrl); 910 uint gc_state_idx = Compile::AliasIdxRaw; 911 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument 912 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx)); 913 914 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered); 915 phase->register_new_node(gc_state, ctrl); 916 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)); 917 phase->register_new_node(heap_stable_and, ctrl); 918 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT)); 919 phase->register_new_node(heap_stable_cmp, ctrl); 920 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne); 921 phase->register_new_node(heap_stable_test, ctrl); 922 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 923 phase->register_control(heap_stable_iff, loop, ctrl); 924 925 heap_stable_ctrl = new IfFalseNode(heap_stable_iff); 926 phase->register_control(heap_stable_ctrl, loop, heap_stable_iff); 927 ctrl = new IfTrueNode(heap_stable_iff); 928 phase->register_control(ctrl, loop, heap_stable_iff); 929 930 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape"); 931 } 932 933 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) { 934 const Type* val_t = phase->igvn().type(val); 935 if (val_t->meet(TypePtr::NULL_PTR) == val_t) { 936 IdealLoopTree* loop = phase->get_loop(ctrl); 937 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT)); 938 phase->register_new_node(null_cmp, ctrl); 939 Node* null_test = new BoolNode(null_cmp, BoolTest::ne); 940 phase->register_new_node(null_test, ctrl); 941 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 942 phase->register_control(null_iff, loop, ctrl); 943 ctrl = new IfTrueNode(null_iff); 944 phase->register_control(ctrl, loop, null_iff); 945 null_ctrl = new IfFalseNode(null_iff); 946 phase->register_control(null_ctrl, loop, null_iff); 947 } 948 } 949 950 Node* ShenandoahBarrierC2Support::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) { 951 IdealLoopTree *loop = phase->get_loop(c); 952 Node* iff = unc_ctrl->in(0); 953 assert(iff->is_If(), "broken"); 954 Node* new_iff = iff->clone(); 955 new_iff->set_req(0, c); 956 phase->register_control(new_iff, loop, c); 957 Node* iffalse = new IfFalseNode(new_iff->as_If()); 958 phase->register_control(iffalse, loop, new_iff); 959 Node* iftrue = new IfTrueNode(new_iff->as_If()); 960 phase->register_control(iftrue, loop, new_iff); 961 c = iftrue; 962 const Type *t = phase->igvn().type(val); 963 assert(val->Opcode() == Op_CastPP, "expect cast to non null here"); 964 Node* uncasted_val = val->in(1); 965 val = new CastPPNode(uncasted_val, t); 966 val->init_req(0, c); 967 phase->register_new_node(val, c); 968 return val; 969 } 970 971 void ShenandoahBarrierC2Support::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl, 972 Unique_Node_List& uses, PhaseIdealLoop* phase) { 973 IfNode* iff = unc_ctrl->in(0)->as_If(); 974 Node* proj = iff->proj_out(0); 975 assert(proj != unc_ctrl, "bad projection"); 976 Node* use = proj->unique_ctrl_out(); 977 978 assert(use == unc || use->is_Region(), "what else?"); 979 980 uses.clear(); 981 if (use == unc) { 982 phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use)); 983 for (uint i = 1; i < unc->req(); i++) { 984 Node* n = unc->in(i); 985 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) { 986 uses.push(n); 987 } 988 } 989 } else { 990 assert(use->is_Region(), "what else?"); 991 uint idx = 1; 992 for (; use->in(idx) != proj; idx++); 993 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 994 Node* u = use->fast_out(i); 995 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) { 996 uses.push(u->in(idx)); 997 } 998 } 999 } 1000 for(uint next = 0; next < uses.size(); next++ ) { 1001 Node *n = uses.at(next); 1002 assert(phase->get_ctrl(n) == proj, "bad control"); 1003 phase->set_ctrl_and_loop(n, new_unc_ctrl); 1004 if (n->in(0) == proj) { 1005 phase->igvn().replace_input_of(n, 0, new_unc_ctrl); 1006 } 1007 for (uint i = 0; i < n->req(); i++) { 1008 Node* m = n->in(i); 1009 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) { 1010 uses.push(m); 1011 } 1012 } 1013 } 1014 1015 phase->igvn().rehash_node_delayed(use); 1016 int nb = use->replace_edge(proj, new_unc_ctrl); 1017 assert(nb == 1, "only use expected"); 1018 } 1019 1020 void ShenandoahBarrierC2Support::in_cset_fast_test(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) { 1021 IdealLoopTree *loop = phase->get_loop(ctrl); 1022 Node* raw_rbtrue = new CastP2XNode(ctrl, val); 1023 phase->register_new_node(raw_rbtrue, ctrl); 1024 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint())); 1025 phase->register_new_node(cset_offset, ctrl); 1026 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr())); 1027 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root()); 1028 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset); 1029 phase->register_new_node(in_cset_fast_test_adr, ctrl); 1030 uint in_cset_fast_test_idx = Compile::AliasIdxRaw; 1031 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument 1032 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx)); 1033 Node* in_cset_fast_test_load = new LoadBNode(ctrl, raw_mem, in_cset_fast_test_adr, in_cset_fast_test_adr_type, TypeInt::BYTE, MemNode::unordered); 1034 phase->register_new_node(in_cset_fast_test_load, ctrl); 1035 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT)); 1036 phase->register_new_node(in_cset_fast_test_cmp, ctrl); 1037 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::eq); 1038 phase->register_new_node(in_cset_fast_test_test, ctrl); 1039 IfNode* in_cset_fast_test_iff = new IfNode(ctrl, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 1040 phase->register_control(in_cset_fast_test_iff, loop, ctrl); 1041 1042 not_cset_ctrl = new IfTrueNode(in_cset_fast_test_iff); 1043 phase->register_control(not_cset_ctrl, loop, in_cset_fast_test_iff); 1044 1045 ctrl = new IfFalseNode(in_cset_fast_test_iff); 1046 phase->register_control(ctrl, loop, in_cset_fast_test_iff); 1047 } 1048 1049 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node*& result_mem, Node* raw_mem, bool is_native, PhaseIdealLoop* phase) { 1050 IdealLoopTree*loop = phase->get_loop(ctrl); 1051 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr()->cast_to_nonconst(); 1052 1053 // The slow path stub consumes and produces raw memory in addition 1054 // to the existing memory edges 1055 Node* base = find_bottom_mem(ctrl, phase); 1056 MergeMemNode* mm = MergeMemNode::make(base); 1057 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 1058 phase->register_new_node(mm, ctrl); 1059 1060 address calladdr = is_native ? CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_native) 1061 : CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier); 1062 const char* name = is_native ? "oop_load_from_native_barrier" : "load_reference_barrier"; 1063 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM); 1064 call->init_req(TypeFunc::Control, ctrl); 1065 call->init_req(TypeFunc::I_O, phase->C->top()); 1066 call->init_req(TypeFunc::Memory, mm); 1067 call->init_req(TypeFunc::FramePtr, phase->C->top()); 1068 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 1069 call->init_req(TypeFunc::Parms, val); 1070 phase->register_control(call, loop, ctrl); 1071 ctrl = new ProjNode(call, TypeFunc::Control); 1072 phase->register_control(ctrl, loop, call); 1073 result_mem = new ProjNode(call, TypeFunc::Memory); 1074 phase->register_new_node(result_mem, call); 1075 val = new ProjNode(call, TypeFunc::Parms); 1076 phase->register_new_node(val, call); 1077 val = new CheckCastPPNode(ctrl, val, obj_type); 1078 phase->register_new_node(val, ctrl); 1079 } 1080 1081 void ShenandoahBarrierC2Support::fix_ctrl(Node* barrier, Node* region, const MemoryGraphFixer& fixer, Unique_Node_List& uses, Unique_Node_List& uses_to_ignore, uint last, PhaseIdealLoop* phase) { 1082 Node* ctrl = phase->get_ctrl(barrier); 1083 Node* init_raw_mem = fixer.find_mem(ctrl, barrier); 1084 1085 // Update the control of all nodes that should be after the 1086 // barrier control flow 1087 uses.clear(); 1088 // Every node that is control dependent on the barrier's input 1089 // control will be after the expanded barrier. The raw memory (if 1090 // its memory is control dependent on the barrier's input control) 1091 // must stay above the barrier. 1092 uses_to_ignore.clear(); 1093 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) { 1094 uses_to_ignore.push(init_raw_mem); 1095 } 1096 for (uint next = 0; next < uses_to_ignore.size(); next++) { 1097 Node *n = uses_to_ignore.at(next); 1098 for (uint i = 0; i < n->req(); i++) { 1099 Node* in = n->in(i); 1100 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) { 1101 uses_to_ignore.push(in); 1102 } 1103 } 1104 } 1105 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) { 1106 Node* u = ctrl->fast_out(i); 1107 if (u->_idx < last && 1108 u != barrier && 1109 !uses_to_ignore.member(u) && 1110 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) && 1111 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) { 1112 Node* old_c = phase->ctrl_or_self(u); 1113 Node* c = old_c; 1114 if (c != ctrl || 1115 is_dominator_same_ctrl(old_c, barrier, u, phase) || 1116 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) { 1117 phase->igvn().rehash_node_delayed(u); 1118 int nb = u->replace_edge(ctrl, region); 1119 if (u->is_CFG()) { 1120 if (phase->idom(u) == ctrl) { 1121 phase->set_idom(u, region, phase->dom_depth(region)); 1122 } 1123 } else if (phase->get_ctrl(u) == ctrl) { 1124 assert(u != init_raw_mem, "should leave input raw mem above the barrier"); 1125 uses.push(u); 1126 } 1127 assert(nb == 1, "more than 1 ctrl input?"); 1128 --i, imax -= nb; 1129 } 1130 } 1131 } 1132 } 1133 1134 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) { 1135 Node* region = NULL; 1136 while (c != ctrl) { 1137 if (c->is_Region()) { 1138 region = c; 1139 } 1140 c = phase->idom(c); 1141 } 1142 assert(region != NULL, ""); 1143 Node* phi = new PhiNode(region, n->bottom_type()); 1144 for (uint j = 1; j < region->req(); j++) { 1145 Node* in = region->in(j); 1146 if (phase->is_dominator(projs.fallthrough_catchproj, in)) { 1147 phi->init_req(j, n); 1148 } else if (phase->is_dominator(projs.catchall_catchproj, in)) { 1149 phi->init_req(j, n_clone); 1150 } else { 1151 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase)); 1152 } 1153 } 1154 phase->register_new_node(phi, region); 1155 return phi; 1156 } 1157 1158 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) { 1159 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state(); 1160 1161 Unique_Node_List uses; 1162 for (int i = 0; i < state->enqueue_barriers_count(); i++) { 1163 Node* barrier = state->enqueue_barrier(i); 1164 Node* ctrl = phase->get_ctrl(barrier); 1165 IdealLoopTree* loop = phase->get_loop(ctrl); 1166 if (loop->_head->is_OuterStripMinedLoop()) { 1167 // Expanding a barrier here will break loop strip mining 1168 // verification. Transform the loop so the loop nest doesn't 1169 // appear as strip mined. 1170 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop(); 1171 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase); 1172 } 1173 } 1174 1175 Node_Stack stack(0); 1176 Node_List clones; 1177 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) { 1178 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1179 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1180 continue; 1181 } 1182 1183 Node* ctrl = phase->get_ctrl(lrb); 1184 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 1185 1186 CallStaticJavaNode* unc = NULL; 1187 Node* unc_ctrl = NULL; 1188 Node* uncasted_val = val; 1189 1190 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) { 1191 Node* u = lrb->fast_out(i); 1192 if (u->Opcode() == Op_CastPP && 1193 u->in(0) != NULL && 1194 phase->is_dominator(u->in(0), ctrl)) { 1195 const Type* u_t = phase->igvn().type(u); 1196 1197 if (u_t->meet(TypePtr::NULL_PTR) != u_t && 1198 u->in(0)->Opcode() == Op_IfTrue && 1199 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 1200 u->in(0)->in(0)->is_If() && 1201 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool && 1202 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne && 1203 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 1204 u->in(0)->in(0)->in(1)->in(1)->in(1) == val && 1205 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 1206 IdealLoopTree* loop = phase->get_loop(ctrl); 1207 IdealLoopTree* unc_loop = phase->get_loop(u->in(0)); 1208 1209 if (!unc_loop->is_member(loop)) { 1210 continue; 1211 } 1212 1213 Node* branch = no_branches(ctrl, u->in(0), false, phase); 1214 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch"); 1215 if (branch == NodeSentinel) { 1216 continue; 1217 } 1218 1219 phase->igvn().replace_input_of(u, 1, val); 1220 phase->igvn().replace_input_of(lrb, ShenandoahLoadReferenceBarrierNode::ValueIn, u); 1221 phase->set_ctrl(u, u->in(0)); 1222 phase->set_ctrl(lrb, u->in(0)); 1223 unc = u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1224 unc_ctrl = u->in(0); 1225 val = u; 1226 1227 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 1228 Node* u = val->fast_out(j); 1229 if (u == lrb) continue; 1230 phase->igvn().rehash_node_delayed(u); 1231 int nb = u->replace_edge(val, lrb); 1232 --j; jmax -= nb; 1233 } 1234 1235 RegionNode* r = new RegionNode(3); 1236 IfNode* iff = unc_ctrl->in(0)->as_If(); 1237 1238 Node* ctrl_use = unc_ctrl->unique_ctrl_out(); 1239 Node* unc_ctrl_clone = unc_ctrl->clone(); 1240 phase->register_control(unc_ctrl_clone, loop, iff); 1241 Node* c = unc_ctrl_clone; 1242 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase); 1243 r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0)); 1244 1245 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0)); 1246 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl)); 1247 phase->lazy_replace(c, unc_ctrl); 1248 c = NULL;; 1249 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone); 1250 phase->set_ctrl(val, unc_ctrl_clone); 1251 1252 IfNode* new_iff = new_cast->in(0)->in(0)->as_If(); 1253 fix_null_check(unc, unc_ctrl_clone, r, uses, phase); 1254 Node* iff_proj = iff->proj_out(0); 1255 r->init_req(2, iff_proj); 1256 phase->register_control(r, phase->ltree_root(), iff); 1257 1258 Node* new_bol = new_iff->in(1)->clone(); 1259 Node* new_cmp = new_bol->in(1)->clone(); 1260 assert(new_cmp->Opcode() == Op_CmpP, "broken"); 1261 assert(new_cmp->in(1) == val->in(1), "broken"); 1262 new_bol->set_req(1, new_cmp); 1263 new_cmp->set_req(1, lrb); 1264 phase->register_new_node(new_bol, new_iff->in(0)); 1265 phase->register_new_node(new_cmp, new_iff->in(0)); 1266 phase->igvn().replace_input_of(new_iff, 1, new_bol); 1267 phase->igvn().replace_input_of(new_cast, 1, lrb); 1268 1269 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) { 1270 Node* u = lrb->fast_out(i); 1271 if (u == new_cast || u == new_cmp) { 1272 continue; 1273 } 1274 phase->igvn().rehash_node_delayed(u); 1275 int nb = u->replace_edge(lrb, new_cast); 1276 assert(nb > 0, "no update?"); 1277 --i; imax -= nb; 1278 } 1279 1280 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 1281 Node* u = val->fast_out(i); 1282 if (u == lrb) { 1283 continue; 1284 } 1285 phase->igvn().rehash_node_delayed(u); 1286 int nb = u->replace_edge(val, new_cast); 1287 assert(nb > 0, "no update?"); 1288 --i; imax -= nb; 1289 } 1290 1291 ctrl = unc_ctrl_clone; 1292 phase->set_ctrl_and_loop(lrb, ctrl); 1293 break; 1294 } 1295 } 1296 } 1297 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) { 1298 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava(); 1299 CallProjections projs; 1300 call->extract_projections(&projs, false, false); 1301 1302 Node* lrb_clone = lrb->clone(); 1303 phase->register_new_node(lrb_clone, projs.catchall_catchproj); 1304 phase->set_ctrl(lrb, projs.fallthrough_catchproj); 1305 1306 stack.push(lrb, 0); 1307 clones.push(lrb_clone); 1308 1309 do { 1310 assert(stack.size() == clones.size(), ""); 1311 Node* n = stack.node(); 1312 #ifdef ASSERT 1313 if (n->is_Load()) { 1314 Node* mem = n->in(MemNode::Memory); 1315 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) { 1316 Node* u = mem->fast_out(j); 1317 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?"); 1318 } 1319 } 1320 #endif 1321 uint idx = stack.index(); 1322 Node* n_clone = clones.at(clones.size()-1); 1323 if (idx < n->outcnt()) { 1324 Node* u = n->raw_out(idx); 1325 Node* c = phase->ctrl_or_self(u); 1326 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) { 1327 stack.set_index(idx+1); 1328 assert(!u->is_CFG(), ""); 1329 stack.push(u, 0); 1330 Node* u_clone = u->clone(); 1331 int nb = u_clone->replace_edge(n, n_clone); 1332 assert(nb > 0, "should have replaced some uses"); 1333 phase->register_new_node(u_clone, projs.catchall_catchproj); 1334 clones.push(u_clone); 1335 phase->set_ctrl(u, projs.fallthrough_catchproj); 1336 } else { 1337 bool replaced = false; 1338 if (u->is_Phi()) { 1339 for (uint k = 1; k < u->req(); k++) { 1340 if (u->in(k) == n) { 1341 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) { 1342 phase->igvn().replace_input_of(u, k, n_clone); 1343 replaced = true; 1344 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) { 1345 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase)); 1346 replaced = true; 1347 } 1348 } 1349 } 1350 } else { 1351 if (phase->is_dominator(projs.catchall_catchproj, c)) { 1352 phase->igvn().rehash_node_delayed(u); 1353 int nb = u->replace_edge(n, n_clone); 1354 assert(nb > 0, "should have replaced some uses"); 1355 replaced = true; 1356 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) { 1357 phase->igvn().rehash_node_delayed(u); 1358 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase)); 1359 assert(nb > 0, "should have replaced some uses"); 1360 replaced = true; 1361 } 1362 } 1363 if (!replaced) { 1364 stack.set_index(idx+1); 1365 } 1366 } 1367 } else { 1368 stack.pop(); 1369 clones.pop(); 1370 } 1371 } while (stack.size() > 0); 1372 assert(stack.size() == 0 && clones.size() == 0, ""); 1373 } 1374 } 1375 1376 for (int i = 0; i < state->load_reference_barriers_count(); i++) { 1377 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1378 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1379 continue; 1380 } 1381 Node* ctrl = phase->get_ctrl(lrb); 1382 IdealLoopTree* loop = phase->get_loop(ctrl); 1383 if (loop->_head->is_OuterStripMinedLoop()) { 1384 // Expanding a barrier here will break loop strip mining 1385 // verification. Transform the loop so the loop nest doesn't 1386 // appear as strip mined. 1387 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop(); 1388 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase); 1389 } 1390 } 1391 1392 // Expand load-reference-barriers 1393 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase); 1394 Unique_Node_List uses_to_ignore; 1395 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) { 1396 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1397 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1398 phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn)); 1399 continue; 1400 } 1401 uint last = phase->C->unique(); 1402 Node* ctrl = phase->get_ctrl(lrb); 1403 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 1404 1405 1406 Node* orig_ctrl = ctrl; 1407 1408 Node* raw_mem = fixer.find_mem(ctrl, lrb); 1409 Node* init_raw_mem = raw_mem; 1410 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 1411 1412 IdealLoopTree *loop = phase->get_loop(ctrl); 1413 CallStaticJavaNode* unc = lrb->pin_and_expand_null_check(phase->igvn()); 1414 Node* unc_ctrl = NULL; 1415 if (unc != NULL) { 1416 if (val->in(ShenandoahLoadReferenceBarrierNode::Control) != ctrl) { 1417 unc = NULL; 1418 } else { 1419 unc_ctrl = val->in(ShenandoahLoadReferenceBarrierNode::Control); 1420 } 1421 } 1422 1423 Node* uncasted_val = val; 1424 if (unc != NULL) { 1425 uncasted_val = val->in(1); 1426 } 1427 1428 Node* heap_stable_ctrl = NULL; 1429 Node* null_ctrl = NULL; 1430 1431 assert(val->bottom_type()->make_oopptr(), "need oop"); 1432 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant"); 1433 1434 enum { _heap_stable = 1, _not_cset, _fwded, _evac_path, _null_path, PATH_LIMIT }; 1435 Node* region = new RegionNode(PATH_LIMIT); 1436 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr()); 1437 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1438 1439 // Stable path. 1440 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase); 1441 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If(); 1442 1443 // Heap stable case 1444 region->init_req(_heap_stable, heap_stable_ctrl); 1445 val_phi->init_req(_heap_stable, uncasted_val); 1446 raw_mem_phi->init_req(_heap_stable, raw_mem); 1447 1448 Node* reg2_ctrl = NULL; 1449 // Null case 1450 test_null(ctrl, val, null_ctrl, phase); 1451 if (null_ctrl != NULL) { 1452 reg2_ctrl = null_ctrl->in(0); 1453 region->init_req(_null_path, null_ctrl); 1454 val_phi->init_req(_null_path, uncasted_val); 1455 raw_mem_phi->init_req(_null_path, raw_mem); 1456 } else { 1457 region->del_req(_null_path); 1458 val_phi->del_req(_null_path); 1459 raw_mem_phi->del_req(_null_path); 1460 } 1461 1462 // Test for in-cset. 1463 // Wires !in_cset(obj) to slot 2 of region and phis 1464 Node* not_cset_ctrl = NULL; 1465 in_cset_fast_test(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase); 1466 if (not_cset_ctrl != NULL) { 1467 if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0); 1468 region->init_req(_not_cset, not_cset_ctrl); 1469 val_phi->init_req(_not_cset, uncasted_val); 1470 raw_mem_phi->init_req(_not_cset, raw_mem); 1471 } 1472 1473 // Resolve object when orig-value is in cset. 1474 // Make the unconditional resolve for fwdptr. 1475 Node* new_val = uncasted_val; 1476 if (unc_ctrl != NULL) { 1477 // Clone the null check in this branch to allow implicit null check 1478 new_val = clone_null_check(ctrl, val, unc_ctrl, phase); 1479 fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase); 1480 1481 IfNode* iff = unc_ctrl->in(0)->as_If(); 1482 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1)); 1483 } 1484 Node* addr = new AddPNode(new_val, uncasted_val, phase->igvn().MakeConX(oopDesc::mark_offset_in_bytes())); 1485 phase->register_new_node(addr, ctrl); 1486 assert(new_val->bottom_type()->isa_oopptr(), "what else?"); 1487 Node* markword = new LoadXNode(ctrl, raw_mem, addr, TypeRawPtr::BOTTOM, TypeX_X, MemNode::unordered); 1488 phase->register_new_node(markword, ctrl); 1489 1490 // Test if object is forwarded. This is the case if lowest two bits are set. 1491 Node* masked = new AndXNode(markword, phase->igvn().MakeConX(markOopDesc::lock_mask_in_place)); 1492 phase->register_new_node(masked, ctrl); 1493 Node* cmp = new CmpXNode(masked, phase->igvn().MakeConX(markOopDesc::marked_value)); 1494 phase->register_new_node(cmp, ctrl); 1495 1496 // Only branch to LRB stub if object is not forwarded; otherwise reply with fwd ptr 1497 Node* bol = new BoolNode(cmp, BoolTest::eq); // Equals 3 means it's forwarded 1498 phase->register_new_node(bol, ctrl); 1499 1500 IfNode* iff = new IfNode(ctrl, bol, PROB_LIKELY(0.999), COUNT_UNKNOWN); 1501 phase->register_control(iff, loop, ctrl); 1502 Node* if_fwd = new IfTrueNode(iff); 1503 phase->register_control(if_fwd, loop, iff); 1504 Node* if_not_fwd = new IfFalseNode(iff); 1505 phase->register_control(if_not_fwd, loop, iff); 1506 1507 // Decode forward pointer: since we already have the lowest bits, we can just subtract them 1508 // from the mark word without the need for large immediate mask. 1509 Node* masked2 = new SubXNode(markword, masked); 1510 phase->register_new_node(masked2, if_fwd); 1511 Node* fwdraw = new CastX2PNode(masked2); 1512 fwdraw->init_req(0, if_fwd); 1513 phase->register_new_node(fwdraw, if_fwd); 1514 Node* fwd = new CheckCastPPNode(NULL, fwdraw, val->bottom_type()); 1515 phase->register_new_node(fwd, if_fwd); 1516 1517 // Wire up not-equal-path in slots 3. 1518 region->init_req(_fwded, if_fwd); 1519 val_phi->init_req(_fwded, fwd); 1520 raw_mem_phi->init_req(_fwded, raw_mem); 1521 1522 // Call lrb-stub and wire up that path in slots 4 1523 Node* result_mem = NULL; 1524 ctrl = if_not_fwd; 1525 fwd = new_val; 1526 call_lrb_stub(ctrl, fwd, result_mem, raw_mem, lrb->is_native(), phase); 1527 region->init_req(_evac_path, ctrl); 1528 val_phi->init_req(_evac_path, fwd); 1529 raw_mem_phi->init_req(_evac_path, result_mem); 1530 1531 phase->register_control(region, loop, heap_stable_iff); 1532 Node* out_val = val_phi; 1533 phase->register_new_node(val_phi, region); 1534 phase->register_new_node(raw_mem_phi, region); 1535 1536 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase); 1537 1538 ctrl = orig_ctrl; 1539 1540 if (unc != NULL) { 1541 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 1542 Node* u = val->fast_out(i); 1543 Node* c = phase->ctrl_or_self(u); 1544 if (u != lrb && (c != ctrl || is_dominator_same_ctrl(c, lrb, u, phase))) { 1545 phase->igvn().rehash_node_delayed(u); 1546 int nb = u->replace_edge(val, out_val); 1547 --i, imax -= nb; 1548 } 1549 } 1550 if (val->outcnt() == 0) { 1551 phase->igvn()._worklist.push(val); 1552 } 1553 } 1554 phase->igvn().replace_node(lrb, out_val); 1555 1556 follow_barrier_uses(out_val, ctrl, uses, phase); 1557 1558 for(uint next = 0; next < uses.size(); next++ ) { 1559 Node *n = uses.at(next); 1560 assert(phase->get_ctrl(n) == ctrl, "bad control"); 1561 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 1562 phase->set_ctrl(n, region); 1563 follow_barrier_uses(n, ctrl, uses, phase); 1564 } 1565 1566 // The slow path call produces memory: hook the raw memory phi 1567 // from the expanded load reference barrier with the rest of the graph 1568 // which may require adding memory phis at every post dominated 1569 // region and at enclosing loop heads. Use the memory state 1570 // collected in memory_nodes to fix the memory graph. Update that 1571 // memory state as we go. 1572 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses); 1573 } 1574 // Done expanding load-reference-barriers. 1575 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced"); 1576 1577 for (int i = state->enqueue_barriers_count() - 1; i >= 0; i--) { 1578 Node* barrier = state->enqueue_barrier(i); 1579 Node* pre_val = barrier->in(1); 1580 1581 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) { 1582 ShouldNotReachHere(); 1583 continue; 1584 } 1585 1586 Node* ctrl = phase->get_ctrl(barrier); 1587 1588 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) { 1589 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move"); 1590 ctrl = ctrl->in(0)->in(0); 1591 phase->set_ctrl(barrier, ctrl); 1592 } else if (ctrl->is_CallRuntime()) { 1593 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move"); 1594 ctrl = ctrl->in(0); 1595 phase->set_ctrl(barrier, ctrl); 1596 } 1597 1598 Node* init_ctrl = ctrl; 1599 IdealLoopTree* loop = phase->get_loop(ctrl); 1600 Node* raw_mem = fixer.find_mem(ctrl, barrier); 1601 Node* init_raw_mem = raw_mem; 1602 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 1603 Node* heap_stable_ctrl = NULL; 1604 Node* null_ctrl = NULL; 1605 uint last = phase->C->unique(); 1606 1607 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT }; 1608 Node* region = new RegionNode(PATH_LIMIT); 1609 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1610 1611 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 }; 1612 Node* region2 = new RegionNode(PATH_LIMIT2); 1613 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1614 1615 // Stable path. 1616 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase); 1617 region->init_req(_heap_stable, heap_stable_ctrl); 1618 phi->init_req(_heap_stable, raw_mem); 1619 1620 // Null path 1621 Node* reg2_ctrl = NULL; 1622 test_null(ctrl, pre_val, null_ctrl, phase); 1623 if (null_ctrl != NULL) { 1624 reg2_ctrl = null_ctrl->in(0); 1625 region2->init_req(_null_path, null_ctrl); 1626 phi2->init_req(_null_path, raw_mem); 1627 } else { 1628 region2->del_req(_null_path); 1629 phi2->del_req(_null_path); 1630 } 1631 1632 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 1633 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 1634 Node* thread = new ThreadLocalNode(); 1635 phase->register_new_node(thread, ctrl); 1636 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset)); 1637 phase->register_new_node(buffer_adr, ctrl); 1638 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset)); 1639 phase->register_new_node(index_adr, ctrl); 1640 1641 BasicType index_bt = TypeX_X->basic_type(); 1642 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 1643 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 1644 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered); 1645 phase->register_new_node(index, ctrl); 1646 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0)); 1647 phase->register_new_node(index_cmp, ctrl); 1648 Node* index_test = new BoolNode(index_cmp, BoolTest::ne); 1649 phase->register_new_node(index_test, ctrl); 1650 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 1651 if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff; 1652 phase->register_control(queue_full_iff, loop, ctrl); 1653 Node* not_full = new IfTrueNode(queue_full_iff); 1654 phase->register_control(not_full, loop, queue_full_iff); 1655 Node* full = new IfFalseNode(queue_full_iff); 1656 phase->register_control(full, loop, queue_full_iff); 1657 1658 ctrl = not_full; 1659 1660 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t))); 1661 phase->register_new_node(next_index, ctrl); 1662 1663 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered); 1664 phase->register_new_node(buffer, ctrl); 1665 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index); 1666 phase->register_new_node(log_addr, ctrl); 1667 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered); 1668 phase->register_new_node(log_store, ctrl); 1669 // update the index 1670 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered); 1671 phase->register_new_node(index_update, ctrl); 1672 1673 // Fast-path case 1674 region2->init_req(_fast_path, ctrl); 1675 phi2->init_req(_fast_path, index_update); 1676 1677 ctrl = full; 1678 1679 Node* base = find_bottom_mem(ctrl, phase); 1680 1681 MergeMemNode* mm = MergeMemNode::make(base); 1682 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 1683 phase->register_new_node(mm, ctrl); 1684 1685 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", TypeRawPtr::BOTTOM); 1686 call->init_req(TypeFunc::Control, ctrl); 1687 call->init_req(TypeFunc::I_O, phase->C->top()); 1688 call->init_req(TypeFunc::Memory, mm); 1689 call->init_req(TypeFunc::FramePtr, phase->C->top()); 1690 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 1691 call->init_req(TypeFunc::Parms, pre_val); 1692 call->init_req(TypeFunc::Parms+1, thread); 1693 phase->register_control(call, loop, ctrl); 1694 1695 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control); 1696 phase->register_control(ctrl_proj, loop, call); 1697 Node* mem_proj = new ProjNode(call, TypeFunc::Memory); 1698 phase->register_new_node(mem_proj, call); 1699 1700 // Slow-path case 1701 region2->init_req(_slow_path, ctrl_proj); 1702 phi2->init_req(_slow_path, mem_proj); 1703 1704 phase->register_control(region2, loop, reg2_ctrl); 1705 phase->register_new_node(phi2, region2); 1706 1707 region->init_req(_heap_unstable, region2); 1708 phi->init_req(_heap_unstable, phi2); 1709 1710 phase->register_control(region, loop, heap_stable_ctrl->in(0)); 1711 phase->register_new_node(phi, region); 1712 1713 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase); 1714 for(uint next = 0; next < uses.size(); next++ ) { 1715 Node *n = uses.at(next); 1716 assert(phase->get_ctrl(n) == init_ctrl, "bad control"); 1717 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 1718 phase->set_ctrl(n, region); 1719 follow_barrier_uses(n, init_ctrl, uses, phase); 1720 } 1721 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses); 1722 1723 phase->igvn().replace_node(barrier, pre_val); 1724 } 1725 assert(state->enqueue_barriers_count() == 0, "all enqueue barrier nodes should have been replaced"); 1726 1727 } 1728 1729 void ShenandoahBarrierC2Support::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) { 1730 IdealLoopTree *loop = phase->get_loop(iff); 1731 Node* loop_head = loop->_head; 1732 Node* entry_c = loop_head->in(LoopNode::EntryControl); 1733 1734 Node* bol = iff->in(1); 1735 Node* cmp = bol->in(1); 1736 Node* andi = cmp->in(1); 1737 Node* load = andi->in(1); 1738 1739 assert(is_gc_state_load(load), "broken"); 1740 if (!phase->is_dominator(load->in(0), entry_c)) { 1741 Node* mem_ctrl = NULL; 1742 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase); 1743 load = load->clone(); 1744 load->set_req(MemNode::Memory, mem); 1745 load->set_req(0, entry_c); 1746 phase->register_new_node(load, entry_c); 1747 andi = andi->clone(); 1748 andi->set_req(1, load); 1749 phase->register_new_node(andi, entry_c); 1750 cmp = cmp->clone(); 1751 cmp->set_req(1, andi); 1752 phase->register_new_node(cmp, entry_c); 1753 bol = bol->clone(); 1754 bol->set_req(1, cmp); 1755 phase->register_new_node(bol, entry_c); 1756 1757 Node* old_bol =iff->in(1); 1758 phase->igvn().replace_input_of(iff, 1, bol); 1759 } 1760 } 1761 1762 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) { 1763 if (!n->is_If() || n->is_CountedLoopEnd()) { 1764 return false; 1765 } 1766 Node* region = n->in(0); 1767 1768 if (!region->is_Region()) { 1769 return false; 1770 } 1771 Node* dom = phase->idom(region); 1772 if (!dom->is_If()) { 1773 return false; 1774 } 1775 1776 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) { 1777 return false; 1778 } 1779 1780 IfNode* dom_if = dom->as_If(); 1781 Node* proj_true = dom_if->proj_out(1); 1782 Node* proj_false = dom_if->proj_out(0); 1783 1784 for (uint i = 1; i < region->req(); i++) { 1785 if (phase->is_dominator(proj_true, region->in(i))) { 1786 continue; 1787 } 1788 if (phase->is_dominator(proj_false, region->in(i))) { 1789 continue; 1790 } 1791 return false; 1792 } 1793 1794 return true; 1795 } 1796 1797 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) { 1798 assert(is_heap_stable_test(n), "no other tests"); 1799 if (identical_backtoback_ifs(n, phase)) { 1800 Node* n_ctrl = n->in(0); 1801 if (phase->can_split_if(n_ctrl)) { 1802 IfNode* dom_if = phase->idom(n_ctrl)->as_If(); 1803 if (is_heap_stable_test(n)) { 1804 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1); 1805 assert(is_gc_state_load(gc_state_load), "broken"); 1806 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1); 1807 assert(is_gc_state_load(dom_gc_state_load), "broken"); 1808 if (gc_state_load != dom_gc_state_load) { 1809 phase->igvn().replace_node(gc_state_load, dom_gc_state_load); 1810 } 1811 } 1812 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1813 Node* proj_true = dom_if->proj_out(1); 1814 Node* proj_false = dom_if->proj_out(0); 1815 Node* con_true = phase->igvn().makecon(TypeInt::ONE); 1816 Node* con_false = phase->igvn().makecon(TypeInt::ZERO); 1817 1818 for (uint i = 1; i < n_ctrl->req(); i++) { 1819 if (phase->is_dominator(proj_true, n_ctrl->in(i))) { 1820 bolphi->init_req(i, con_true); 1821 } else { 1822 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1823 bolphi->init_req(i, con_false); 1824 } 1825 } 1826 phase->register_new_node(bolphi, n_ctrl); 1827 phase->igvn().replace_input_of(n, 1, bolphi); 1828 phase->do_split_if(n); 1829 } 1830 } 1831 } 1832 1833 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) { 1834 // Find first invariant test that doesn't exit the loop 1835 LoopNode *head = loop->_head->as_Loop(); 1836 IfNode* unswitch_iff = NULL; 1837 Node* n = head->in(LoopNode::LoopBackControl); 1838 int loop_has_sfpts = -1; 1839 while (n != head) { 1840 Node* n_dom = phase->idom(n); 1841 if (n->is_Region()) { 1842 if (n_dom->is_If()) { 1843 IfNode* iff = n_dom->as_If(); 1844 if (iff->in(1)->is_Bool()) { 1845 BoolNode* bol = iff->in(1)->as_Bool(); 1846 if (bol->in(1)->is_Cmp()) { 1847 // If condition is invariant and not a loop exit, 1848 // then found reason to unswitch. 1849 if (is_heap_stable_test(iff) && 1850 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) { 1851 assert(!loop->is_loop_exit(iff), "both branches should be in the loop"); 1852 if (loop_has_sfpts == -1) { 1853 for(uint i = 0; i < loop->_body.size(); i++) { 1854 Node *m = loop->_body[i]; 1855 if (m->is_SafePoint() && !m->is_CallLeaf()) { 1856 loop_has_sfpts = 1; 1857 break; 1858 } 1859 } 1860 if (loop_has_sfpts == -1) { 1861 loop_has_sfpts = 0; 1862 } 1863 } 1864 if (!loop_has_sfpts) { 1865 unswitch_iff = iff; 1866 } 1867 } 1868 } 1869 } 1870 } 1871 } 1872 n = n_dom; 1873 } 1874 return unswitch_iff; 1875 } 1876 1877 1878 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) { 1879 Node_List heap_stable_tests; 1880 Node_List gc_state_loads; 1881 stack.push(phase->C->start(), 0); 1882 do { 1883 Node* n = stack.node(); 1884 uint i = stack.index(); 1885 1886 if (i < n->outcnt()) { 1887 Node* u = n->raw_out(i); 1888 stack.set_index(i+1); 1889 if (!visited.test_set(u->_idx)) { 1890 stack.push(u, 0); 1891 } 1892 } else { 1893 stack.pop(); 1894 if (ShenandoahCommonGCStateLoads && is_gc_state_load(n)) { 1895 gc_state_loads.push(n); 1896 } 1897 if (n->is_If() && is_heap_stable_test(n)) { 1898 heap_stable_tests.push(n); 1899 } 1900 } 1901 } while (stack.size() > 0); 1902 1903 bool progress; 1904 do { 1905 progress = false; 1906 for (uint i = 0; i < gc_state_loads.size(); i++) { 1907 Node* n = gc_state_loads.at(i); 1908 if (n->outcnt() != 0) { 1909 progress |= try_common_gc_state_load(n, phase); 1910 } 1911 } 1912 } while (progress); 1913 1914 for (uint i = 0; i < heap_stable_tests.size(); i++) { 1915 Node* n = heap_stable_tests.at(i); 1916 assert(is_heap_stable_test(n), "only evacuation test"); 1917 merge_back_to_back_tests(n, phase); 1918 } 1919 1920 if (!phase->C->major_progress()) { 1921 VectorSet seen(Thread::current()->resource_area()); 1922 for (uint i = 0; i < heap_stable_tests.size(); i++) { 1923 Node* n = heap_stable_tests.at(i); 1924 IdealLoopTree* loop = phase->get_loop(n); 1925 if (loop != phase->ltree_root() && 1926 loop->_child == NULL && 1927 !loop->_irreducible) { 1928 LoopNode* head = loop->_head->as_Loop(); 1929 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) && 1930 !seen.test_set(head->_idx)) { 1931 IfNode* iff = find_unswitching_candidate(loop, phase); 1932 if (iff != NULL) { 1933 Node* bol = iff->in(1); 1934 if (head->is_strip_mined()) { 1935 head->verify_strip_mined(0); 1936 } 1937 move_heap_stable_test_out_of_loop(iff, phase); 1938 1939 AutoNodeBudget node_budget(phase); 1940 1941 if (loop->policy_unswitching(phase)) { 1942 if (head->is_strip_mined()) { 1943 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop(); 1944 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase); 1945 } 1946 phase->do_unswitching(loop, old_new); 1947 } else { 1948 // Not proceeding with unswitching. Move load back in 1949 // the loop. 1950 phase->igvn().replace_input_of(iff, 1, bol); 1951 } 1952 } 1953 } 1954 } 1955 } 1956 } 1957 } 1958 1959 #ifdef ASSERT 1960 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) { 1961 const bool trace = false; 1962 ResourceMark rm; 1963 Unique_Node_List nodes; 1964 Unique_Node_List controls; 1965 Unique_Node_List memories; 1966 1967 nodes.push(root); 1968 for (uint next = 0; next < nodes.size(); next++) { 1969 Node *n = nodes.at(next); 1970 if (ShenandoahBarrierSetC2::is_shenandoah_lrb_call(n)) { 1971 controls.push(n); 1972 if (trace) { tty->print("XXXXXX verifying"); n->dump(); } 1973 for (uint next2 = 0; next2 < controls.size(); next2++) { 1974 Node *m = controls.at(next2); 1975 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 1976 Node* u = m->fast_out(i); 1977 if (u->is_CFG() && !u->is_Root() && 1978 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) && 1979 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) { 1980 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); } 1981 controls.push(u); 1982 } 1983 } 1984 } 1985 memories.push(n->as_Call()->proj_out(TypeFunc::Memory)); 1986 for (uint next2 = 0; next2 < memories.size(); next2++) { 1987 Node *m = memories.at(next2); 1988 assert(m->bottom_type() == Type::MEMORY, ""); 1989 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 1990 Node* u = m->fast_out(i); 1991 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) { 1992 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 1993 memories.push(u); 1994 } else if (u->is_LoadStore()) { 1995 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); } 1996 memories.push(u->find_out_with(Op_SCMemProj)); 1997 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) { 1998 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 1999 memories.push(u); 2000 } else if (u->is_Phi()) { 2001 assert(u->bottom_type() == Type::MEMORY, ""); 2002 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) { 2003 assert(controls.member(u->in(0)), ""); 2004 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 2005 memories.push(u); 2006 } 2007 } else if (u->is_SafePoint() || u->is_MemBar()) { 2008 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2009 Node* uu = u->fast_out(j); 2010 if (uu->bottom_type() == Type::MEMORY) { 2011 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); } 2012 memories.push(uu); 2013 } 2014 } 2015 } 2016 } 2017 } 2018 for (uint next2 = 0; next2 < controls.size(); next2++) { 2019 Node *m = controls.at(next2); 2020 if (m->is_Region()) { 2021 bool all_in = true; 2022 for (uint i = 1; i < m->req(); i++) { 2023 if (!controls.member(m->in(i))) { 2024 all_in = false; 2025 break; 2026 } 2027 } 2028 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); } 2029 bool found_phi = false; 2030 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) { 2031 Node* u = m->fast_out(j); 2032 if (u->is_Phi() && memories.member(u)) { 2033 found_phi = true; 2034 for (uint i = 1; i < u->req() && found_phi; i++) { 2035 Node* k = u->in(i); 2036 if (memories.member(k) != controls.member(m->in(i))) { 2037 found_phi = false; 2038 } 2039 } 2040 } 2041 } 2042 assert(found_phi || all_in, ""); 2043 } 2044 } 2045 controls.clear(); 2046 memories.clear(); 2047 } 2048 for( uint i = 0; i < n->len(); ++i ) { 2049 Node *m = n->in(i); 2050 if (m != NULL) { 2051 nodes.push(m); 2052 } 2053 } 2054 } 2055 } 2056 #endif 2057 2058 ShenandoahEnqueueBarrierNode::ShenandoahEnqueueBarrierNode(Node* val) : Node(NULL, val) { 2059 ShenandoahBarrierSetC2::bsc2()->state()->add_enqueue_barrier(this); 2060 } 2061 2062 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const { 2063 if (in(1) == NULL || in(1)->is_top()) { 2064 return Type::TOP; 2065 } 2066 const Type* t = in(1)->bottom_type(); 2067 if (t == TypePtr::NULL_PTR) { 2068 return t; 2069 } 2070 return t->is_oopptr()->cast_to_nonconst(); 2071 } 2072 2073 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const { 2074 if (in(1) == NULL) { 2075 return Type::TOP; 2076 } 2077 const Type* t = phase->type(in(1)); 2078 if (t == Type::TOP) { 2079 return Type::TOP; 2080 } 2081 if (t == TypePtr::NULL_PTR) { 2082 return t; 2083 } 2084 return t->is_oopptr()->cast_to_nonconst(); 2085 } 2086 2087 int ShenandoahEnqueueBarrierNode::needed(Node* n) { 2088 if (n == NULL || 2089 n->is_Allocate() || 2090 n->Opcode() == Op_ShenandoahEnqueueBarrier || 2091 n->bottom_type() == TypePtr::NULL_PTR || 2092 (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) { 2093 return NotNeeded; 2094 } 2095 if (n->is_Phi() || 2096 n->is_CMove()) { 2097 return MaybeNeeded; 2098 } 2099 return Needed; 2100 } 2101 2102 Node* ShenandoahEnqueueBarrierNode::next(Node* n) { 2103 for (;;) { 2104 if (n == NULL) { 2105 return n; 2106 } else if (n->bottom_type() == TypePtr::NULL_PTR) { 2107 return n; 2108 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) { 2109 return n; 2110 } else if (n->is_ConstraintCast() || 2111 n->Opcode() == Op_DecodeN || 2112 n->Opcode() == Op_EncodeP) { 2113 n = n->in(1); 2114 } else if (n->is_Proj()) { 2115 n = n->in(0); 2116 } else { 2117 return n; 2118 } 2119 } 2120 ShouldNotReachHere(); 2121 return NULL; 2122 } 2123 2124 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) { 2125 PhaseIterGVN* igvn = phase->is_IterGVN(); 2126 2127 Node* n = next(in(1)); 2128 2129 int cont = needed(n); 2130 2131 if (cont == NotNeeded) { 2132 return in(1); 2133 } else if (cont == MaybeNeeded) { 2134 if (igvn == NULL) { 2135 phase->record_for_igvn(this); 2136 return this; 2137 } else { 2138 ResourceMark rm; 2139 Unique_Node_List wq; 2140 uint wq_i = 0; 2141 2142 for (;;) { 2143 if (n->is_Phi()) { 2144 for (uint i = 1; i < n->req(); i++) { 2145 Node* m = n->in(i); 2146 if (m != NULL) { 2147 wq.push(m); 2148 } 2149 } 2150 } else { 2151 assert(n->is_CMove(), "nothing else here"); 2152 Node* m = n->in(CMoveNode::IfFalse); 2153 wq.push(m); 2154 m = n->in(CMoveNode::IfTrue); 2155 wq.push(m); 2156 } 2157 Node* orig_n = NULL; 2158 do { 2159 if (wq_i >= wq.size()) { 2160 return in(1); 2161 } 2162 n = wq.at(wq_i); 2163 wq_i++; 2164 orig_n = n; 2165 n = next(n); 2166 cont = needed(n); 2167 if (cont == Needed) { 2168 return this; 2169 } 2170 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n))); 2171 } 2172 } 2173 } 2174 2175 return this; 2176 } 2177 2178 #ifdef ASSERT 2179 static bool has_never_branch(Node* root) { 2180 for (uint i = 1; i < root->req(); i++) { 2181 Node* in = root->in(i); 2182 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) { 2183 return true; 2184 } 2185 } 2186 return false; 2187 } 2188 #endif 2189 2190 void MemoryGraphFixer::collect_memory_nodes() { 2191 Node_Stack stack(0); 2192 VectorSet visited(Thread::current()->resource_area()); 2193 Node_List regions; 2194 2195 // Walk the raw memory graph and create a mapping from CFG node to 2196 // memory node. Exclude phis for now. 2197 stack.push(_phase->C->root(), 1); 2198 do { 2199 Node* n = stack.node(); 2200 int opc = n->Opcode(); 2201 uint i = stack.index(); 2202 if (i < n->req()) { 2203 Node* mem = NULL; 2204 if (opc == Op_Root) { 2205 Node* in = n->in(i); 2206 int in_opc = in->Opcode(); 2207 if (in_opc == Op_Return || in_opc == Op_Rethrow) { 2208 mem = in->in(TypeFunc::Memory); 2209 } else if (in_opc == Op_Halt) { 2210 if (!in->in(0)->is_Region()) { 2211 Node* proj = in->in(0); 2212 assert(proj->is_Proj(), ""); 2213 Node* in = proj->in(0); 2214 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), ""); 2215 if (in->is_CallStaticJava()) { 2216 mem = in->in(TypeFunc::Memory); 2217 } else if (in->Opcode() == Op_Catch) { 2218 Node* call = in->in(0)->in(0); 2219 assert(call->is_Call(), ""); 2220 mem = call->in(TypeFunc::Memory); 2221 } else if (in->Opcode() == Op_NeverBranch) { 2222 ResourceMark rm; 2223 Unique_Node_List wq; 2224 wq.push(in); 2225 wq.push(in->as_Multi()->proj_out(0)); 2226 for (uint j = 1; j < wq.size(); j++) { 2227 Node* c = wq.at(j); 2228 assert(!c->is_Root(), "shouldn't leave loop"); 2229 if (c->is_SafePoint()) { 2230 assert(mem == NULL, "only one safepoint"); 2231 mem = c->in(TypeFunc::Memory); 2232 } 2233 for (DUIterator_Fast kmax, k = c->fast_outs(kmax); k < kmax; k++) { 2234 Node* u = c->fast_out(k); 2235 if (u->is_CFG()) { 2236 wq.push(u); 2237 } 2238 } 2239 } 2240 assert(mem != NULL, "should have found safepoint"); 2241 } 2242 } 2243 } else { 2244 #ifdef ASSERT 2245 n->dump(); 2246 in->dump(); 2247 #endif 2248 ShouldNotReachHere(); 2249 } 2250 } else { 2251 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, ""); 2252 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, ""); 2253 mem = n->in(i); 2254 } 2255 i++; 2256 stack.set_index(i); 2257 if (mem == NULL) { 2258 continue; 2259 } 2260 for (;;) { 2261 if (visited.test_set(mem->_idx) || mem->is_Start()) { 2262 break; 2263 } 2264 if (mem->is_Phi()) { 2265 stack.push(mem, 2); 2266 mem = mem->in(1); 2267 } else if (mem->is_Proj()) { 2268 stack.push(mem, mem->req()); 2269 mem = mem->in(0); 2270 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 2271 mem = mem->in(TypeFunc::Memory); 2272 } else if (mem->is_MergeMem()) { 2273 MergeMemNode* mm = mem->as_MergeMem(); 2274 mem = mm->memory_at(_alias); 2275 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 2276 assert(_alias == Compile::AliasIdxRaw, ""); 2277 stack.push(mem, mem->req()); 2278 mem = mem->in(MemNode::Memory); 2279 } else { 2280 #ifdef ASSERT 2281 mem->dump(); 2282 #endif 2283 ShouldNotReachHere(); 2284 } 2285 } 2286 } else { 2287 if (n->is_Phi()) { 2288 // Nothing 2289 } else if (!n->is_Root()) { 2290 Node* c = get_ctrl(n); 2291 _memory_nodes.map(c->_idx, n); 2292 } 2293 stack.pop(); 2294 } 2295 } while(stack.is_nonempty()); 2296 2297 // Iterate over CFG nodes in rpo and propagate memory state to 2298 // compute memory state at regions, creating new phis if needed. 2299 Node_List rpo_list; 2300 visited.Clear(); 2301 _phase->rpo(_phase->C->root(), stack, visited, rpo_list); 2302 Node* root = rpo_list.pop(); 2303 assert(root == _phase->C->root(), ""); 2304 2305 const bool trace = false; 2306 #ifdef ASSERT 2307 if (trace) { 2308 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2309 Node* c = rpo_list.at(i); 2310 if (_memory_nodes[c->_idx] != NULL) { 2311 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); 2312 } 2313 } 2314 } 2315 #endif 2316 uint last = _phase->C->unique(); 2317 2318 #ifdef ASSERT 2319 uint8_t max_depth = 0; 2320 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) { 2321 IdealLoopTree* lpt = iter.current(); 2322 max_depth = MAX2(max_depth, lpt->_nest); 2323 } 2324 #endif 2325 2326 bool progress = true; 2327 int iteration = 0; 2328 Node_List dead_phis; 2329 while (progress) { 2330 progress = false; 2331 iteration++; 2332 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), ""); 2333 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); } 2334 IdealLoopTree* last_updated_ilt = NULL; 2335 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2336 Node* c = rpo_list.at(i); 2337 2338 Node* prev_mem = _memory_nodes[c->_idx]; 2339 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2340 Node* prev_region = regions[c->_idx]; 2341 Node* unique = NULL; 2342 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) { 2343 Node* m = _memory_nodes[c->in(j)->_idx]; 2344 assert(m != NULL || (c->is_Loop() && j == LoopNode::LoopBackControl && iteration == 1) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "expect memory state"); 2345 if (m != NULL) { 2346 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) { 2347 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), ""); 2348 // continue 2349 } else if (unique == NULL) { 2350 unique = m; 2351 } else if (m == unique) { 2352 // continue 2353 } else { 2354 unique = NodeSentinel; 2355 } 2356 } 2357 } 2358 assert(unique != NULL, "empty phi???"); 2359 if (unique != NodeSentinel) { 2360 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) { 2361 dead_phis.push(prev_region); 2362 } 2363 regions.map(c->_idx, unique); 2364 } else { 2365 Node* phi = NULL; 2366 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) { 2367 phi = prev_region; 2368 for (uint k = 1; k < c->req(); k++) { 2369 Node* m = _memory_nodes[c->in(k)->_idx]; 2370 assert(m != NULL, "expect memory state"); 2371 phi->set_req(k, m); 2372 } 2373 } else { 2374 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) { 2375 Node* u = c->fast_out(j); 2376 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 2377 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) { 2378 phi = u; 2379 for (uint k = 1; k < c->req() && phi != NULL; k++) { 2380 Node* m = _memory_nodes[c->in(k)->_idx]; 2381 assert(m != NULL, "expect memory state"); 2382 if (u->in(k) != m) { 2383 phi = NULL; 2384 } 2385 } 2386 } 2387 } 2388 if (phi == NULL) { 2389 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias)); 2390 for (uint k = 1; k < c->req(); k++) { 2391 Node* m = _memory_nodes[c->in(k)->_idx]; 2392 assert(m != NULL, "expect memory state"); 2393 phi->init_req(k, m); 2394 } 2395 } 2396 } 2397 assert(phi != NULL, ""); 2398 regions.map(c->_idx, phi); 2399 } 2400 Node* current_region = regions[c->_idx]; 2401 if (current_region != prev_region) { 2402 progress = true; 2403 if (prev_region == prev_mem) { 2404 _memory_nodes.map(c->_idx, current_region); 2405 } 2406 } 2407 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) { 2408 Node* m = _memory_nodes[_phase->idom(c)->_idx]; 2409 assert(m != NULL, "expect memory state"); 2410 if (m != prev_mem) { 2411 _memory_nodes.map(c->_idx, m); 2412 progress = true; 2413 } 2414 } 2415 #ifdef ASSERT 2416 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); } 2417 #endif 2418 } 2419 } 2420 2421 // Replace existing phi with computed memory state for that region 2422 // if different (could be a new phi or a dominating memory node if 2423 // that phi was found to be useless). 2424 while (dead_phis.size() > 0) { 2425 Node* n = dead_phis.pop(); 2426 n->replace_by(_phase->C->top()); 2427 n->destruct(); 2428 } 2429 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2430 Node* c = rpo_list.at(i); 2431 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2432 Node* n = regions[c->_idx]; 2433 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) { 2434 _phase->register_new_node(n, c); 2435 } 2436 } 2437 } 2438 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2439 Node* c = rpo_list.at(i); 2440 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2441 Node* n = regions[c->_idx]; 2442 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 2443 Node* u = c->fast_out(i); 2444 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 2445 u != n) { 2446 if (u->adr_type() == TypePtr::BOTTOM) { 2447 fix_memory_uses(u, n, n, c); 2448 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 2449 _phase->lazy_replace(u, n); 2450 --i; --imax; 2451 } 2452 } 2453 } 2454 } 2455 } 2456 } 2457 2458 Node* MemoryGraphFixer::get_ctrl(Node* n) const { 2459 Node* c = _phase->get_ctrl(n); 2460 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) { 2461 assert(c == n->in(0), ""); 2462 CallNode* call = c->as_Call(); 2463 CallProjections projs; 2464 call->extract_projections(&projs, true, false); 2465 if (projs.catchall_memproj != NULL) { 2466 if (projs.fallthrough_memproj == n) { 2467 c = projs.fallthrough_catchproj; 2468 } else { 2469 assert(projs.catchall_memproj == n, ""); 2470 c = projs.catchall_catchproj; 2471 } 2472 } 2473 } 2474 return c; 2475 } 2476 2477 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const { 2478 if (_phase->has_ctrl(n)) 2479 return get_ctrl(n); 2480 else { 2481 assert (n->is_CFG(), "must be a CFG node"); 2482 return n; 2483 } 2484 } 2485 2486 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const { 2487 return m != NULL && get_ctrl(m) == c; 2488 } 2489 2490 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const { 2491 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, ""); 2492 Node* mem = _memory_nodes[ctrl->_idx]; 2493 Node* c = ctrl; 2494 while (!mem_is_valid(mem, c) && 2495 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) { 2496 c = _phase->idom(c); 2497 mem = _memory_nodes[c->_idx]; 2498 } 2499 if (n != NULL && mem_is_valid(mem, c)) { 2500 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) { 2501 mem = next_mem(mem, _alias); 2502 } 2503 if (mem->is_MergeMem()) { 2504 mem = mem->as_MergeMem()->memory_at(_alias); 2505 } 2506 if (!mem_is_valid(mem, c)) { 2507 do { 2508 c = _phase->idom(c); 2509 mem = _memory_nodes[c->_idx]; 2510 } while (!mem_is_valid(mem, c) && 2511 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))); 2512 } 2513 } 2514 assert(mem->bottom_type() == Type::MEMORY, ""); 2515 return mem; 2516 } 2517 2518 bool MemoryGraphFixer::has_mem_phi(Node* region) const { 2519 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2520 Node* use = region->fast_out(i); 2521 if (use->is_Phi() && use->bottom_type() == Type::MEMORY && 2522 (_phase->C->get_alias_index(use->adr_type()) == _alias)) { 2523 return true; 2524 } 2525 } 2526 return false; 2527 } 2528 2529 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) { 2530 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, ""); 2531 const bool trace = false; 2532 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); }); 2533 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); }); 2534 GrowableArray<Node*> phis; 2535 if (mem_for_ctrl != mem) { 2536 Node* old = mem_for_ctrl; 2537 Node* prev = NULL; 2538 while (old != mem) { 2539 prev = old; 2540 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) { 2541 assert(_alias == Compile::AliasIdxRaw, ""); 2542 old = old->in(MemNode::Memory); 2543 } else if (old->Opcode() == Op_SCMemProj) { 2544 assert(_alias == Compile::AliasIdxRaw, ""); 2545 old = old->in(0); 2546 } else { 2547 ShouldNotReachHere(); 2548 } 2549 } 2550 assert(prev != NULL, ""); 2551 if (new_ctrl != ctrl) { 2552 _memory_nodes.map(ctrl->_idx, mem); 2553 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl); 2554 } 2555 uint input = (uint)MemNode::Memory; 2556 _phase->igvn().replace_input_of(prev, input, new_mem); 2557 } else { 2558 uses.clear(); 2559 _memory_nodes.map(new_ctrl->_idx, new_mem); 2560 uses.push(new_ctrl); 2561 for(uint next = 0; next < uses.size(); next++ ) { 2562 Node *n = uses.at(next); 2563 assert(n->is_CFG(), ""); 2564 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); }); 2565 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2566 Node* u = n->fast_out(i); 2567 if (!u->is_Root() && u->is_CFG() && u != n) { 2568 Node* m = _memory_nodes[u->_idx]; 2569 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) && 2570 !has_mem_phi(u) && 2571 u->unique_ctrl_out()->Opcode() != Op_Halt) { 2572 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); }); 2573 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); }); 2574 2575 if (!mem_is_valid(m, u) || !m->is_Phi()) { 2576 bool push = true; 2577 bool create_phi = true; 2578 if (_phase->is_dominator(new_ctrl, u)) { 2579 create_phi = false; 2580 } else if (!_phase->C->has_irreducible_loop()) { 2581 IdealLoopTree* loop = _phase->get_loop(ctrl); 2582 bool do_check = true; 2583 IdealLoopTree* l = loop; 2584 create_phi = false; 2585 while (l != _phase->ltree_root()) { 2586 Node* head = l->_head; 2587 if (head->in(0) == NULL) { 2588 head = _phase->get_ctrl(head); 2589 } 2590 if (_phase->is_dominator(head, u) && _phase->is_dominator(_phase->idom(u), head)) { 2591 create_phi = true; 2592 do_check = false; 2593 break; 2594 } 2595 l = l->_parent; 2596 } 2597 2598 if (do_check) { 2599 assert(!create_phi, ""); 2600 IdealLoopTree* u_loop = _phase->get_loop(u); 2601 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) { 2602 Node* c = ctrl; 2603 while (!_phase->is_dominator(c, u_loop->tail())) { 2604 c = _phase->idom(c); 2605 } 2606 if (!_phase->is_dominator(c, u)) { 2607 do_check = false; 2608 } 2609 } 2610 } 2611 2612 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) { 2613 create_phi = true; 2614 } 2615 } 2616 if (create_phi) { 2617 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias)); 2618 _phase->register_new_node(phi, u); 2619 phis.push(phi); 2620 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); }); 2621 if (!mem_is_valid(m, u)) { 2622 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); }); 2623 _memory_nodes.map(u->_idx, phi); 2624 } else { 2625 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); }); 2626 for (;;) { 2627 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), ""); 2628 Node* next = NULL; 2629 if (m->is_Proj()) { 2630 next = m->in(0); 2631 } else { 2632 assert(m->is_Mem() || m->is_LoadStore(), ""); 2633 assert(_alias == Compile::AliasIdxRaw, ""); 2634 next = m->in(MemNode::Memory); 2635 } 2636 if (_phase->get_ctrl(next) != u) { 2637 break; 2638 } 2639 if (next->is_MergeMem()) { 2640 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, ""); 2641 break; 2642 } 2643 if (next->is_Phi()) { 2644 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, ""); 2645 break; 2646 } 2647 m = next; 2648 } 2649 2650 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); }); 2651 assert(m->is_Mem() || m->is_LoadStore(), ""); 2652 uint input = (uint)MemNode::Memory; 2653 _phase->igvn().replace_input_of(m, input, phi); 2654 push = false; 2655 } 2656 } else { 2657 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); }); 2658 } 2659 if (push) { 2660 uses.push(u); 2661 } 2662 } 2663 } else if (!mem_is_valid(m, u) && 2664 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) { 2665 uses.push(u); 2666 } 2667 } 2668 } 2669 } 2670 for (int i = 0; i < phis.length(); i++) { 2671 Node* n = phis.at(i); 2672 Node* r = n->in(0); 2673 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); }); 2674 for (uint j = 1; j < n->req(); j++) { 2675 Node* m = find_mem(r->in(j), NULL); 2676 _phase->igvn().replace_input_of(n, j, m); 2677 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); }); 2678 } 2679 } 2680 } 2681 uint last = _phase->C->unique(); 2682 MergeMemNode* mm = NULL; 2683 int alias = _alias; 2684 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); }); 2685 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 2686 Node* u = mem->out(i); 2687 if (u->_idx < last) { 2688 if (u->is_Mem()) { 2689 if (_phase->C->get_alias_index(u->adr_type()) == alias) { 2690 Node* m = find_mem(_phase->get_ctrl(u), u); 2691 if (m != mem) { 2692 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2693 _phase->igvn().replace_input_of(u, MemNode::Memory, m); 2694 --i; 2695 } 2696 } 2697 } else if (u->is_MergeMem()) { 2698 MergeMemNode* u_mm = u->as_MergeMem(); 2699 if (u_mm->memory_at(alias) == mem) { 2700 MergeMemNode* newmm = NULL; 2701 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2702 Node* uu = u->fast_out(j); 2703 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 2704 if (uu->is_Phi()) { 2705 assert(uu->adr_type() == TypePtr::BOTTOM, ""); 2706 Node* region = uu->in(0); 2707 int nb = 0; 2708 for (uint k = 1; k < uu->req(); k++) { 2709 if (uu->in(k) == u) { 2710 Node* m = find_mem(region->in(k), NULL); 2711 if (m != mem) { 2712 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); }); 2713 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 2714 if (newmm != u) { 2715 _phase->igvn().replace_input_of(uu, k, newmm); 2716 nb++; 2717 --jmax; 2718 } 2719 } 2720 } 2721 } 2722 if (nb > 0) { 2723 --j; 2724 } 2725 } else { 2726 Node* m = find_mem(_phase->ctrl_or_self(uu), uu); 2727 if (m != mem) { 2728 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); }); 2729 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 2730 if (newmm != u) { 2731 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 2732 --j, --jmax; 2733 } 2734 } 2735 } 2736 } 2737 } 2738 } else if (u->is_Phi()) { 2739 assert(u->bottom_type() == Type::MEMORY, "what else?"); 2740 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) { 2741 Node* region = u->in(0); 2742 bool replaced = false; 2743 for (uint j = 1; j < u->req(); j++) { 2744 if (u->in(j) == mem) { 2745 Node* m = find_mem(region->in(j), NULL); 2746 Node* nnew = m; 2747 if (m != mem) { 2748 if (u->adr_type() == TypePtr::BOTTOM) { 2749 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m)); 2750 nnew = mm; 2751 } 2752 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); }); 2753 _phase->igvn().replace_input_of(u, j, nnew); 2754 replaced = true; 2755 } 2756 } 2757 } 2758 if (replaced) { 2759 --i; 2760 } 2761 } 2762 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 2763 u->adr_type() == NULL) { 2764 assert(u->adr_type() != NULL || 2765 u->Opcode() == Op_Rethrow || 2766 u->Opcode() == Op_Return || 2767 u->Opcode() == Op_SafePoint || 2768 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 2769 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 2770 u->Opcode() == Op_CallLeaf, ""); 2771 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2772 if (m != mem) { 2773 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m)); 2774 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 2775 --i; 2776 } 2777 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) { 2778 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2779 if (m != mem) { 2780 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2781 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 2782 --i; 2783 } 2784 } else if (u->adr_type() != TypePtr::BOTTOM && 2785 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) { 2786 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2787 assert(m != mem, ""); 2788 // u is on the wrong slice... 2789 assert(u->is_ClearArray(), ""); 2790 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2791 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 2792 --i; 2793 } 2794 } 2795 } 2796 #ifdef ASSERT 2797 assert(new_mem->outcnt() > 0, ""); 2798 for (int i = 0; i < phis.length(); i++) { 2799 Node* n = phis.at(i); 2800 assert(n->outcnt() > 0, "new phi must have uses now"); 2801 } 2802 #endif 2803 } 2804 2805 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const { 2806 MergeMemNode* mm = MergeMemNode::make(mem); 2807 mm->set_memory_at(_alias, rep_proj); 2808 _phase->register_new_node(mm, rep_ctrl); 2809 return mm; 2810 } 2811 2812 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const { 2813 MergeMemNode* newmm = NULL; 2814 MergeMemNode* u_mm = u->as_MergeMem(); 2815 Node* c = _phase->get_ctrl(u); 2816 if (_phase->is_dominator(c, rep_ctrl)) { 2817 c = rep_ctrl; 2818 } else { 2819 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other"); 2820 } 2821 if (u->outcnt() == 1) { 2822 if (u->req() > (uint)_alias && u->in(_alias) == mem) { 2823 _phase->igvn().replace_input_of(u, _alias, rep_proj); 2824 --i; 2825 } else { 2826 _phase->igvn().rehash_node_delayed(u); 2827 u_mm->set_memory_at(_alias, rep_proj); 2828 } 2829 newmm = u_mm; 2830 _phase->set_ctrl_and_loop(u, c); 2831 } else { 2832 // can't simply clone u and then change one of its input because 2833 // it adds and then removes an edge which messes with the 2834 // DUIterator 2835 newmm = MergeMemNode::make(u_mm->base_memory()); 2836 for (uint j = 0; j < u->req(); j++) { 2837 if (j < newmm->req()) { 2838 if (j == (uint)_alias) { 2839 newmm->set_req(j, rep_proj); 2840 } else if (newmm->in(j) != u->in(j)) { 2841 newmm->set_req(j, u->in(j)); 2842 } 2843 } else if (j == (uint)_alias) { 2844 newmm->add_req(rep_proj); 2845 } else { 2846 newmm->add_req(u->in(j)); 2847 } 2848 } 2849 if ((uint)_alias >= u->req()) { 2850 newmm->set_memory_at(_alias, rep_proj); 2851 } 2852 _phase->register_new_node(newmm, c); 2853 } 2854 return newmm; 2855 } 2856 2857 bool MemoryGraphFixer::should_process_phi(Node* phi) const { 2858 if (phi->adr_type() == TypePtr::BOTTOM) { 2859 Node* region = phi->in(0); 2860 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) { 2861 Node* uu = region->fast_out(j); 2862 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) { 2863 return false; 2864 } 2865 } 2866 return true; 2867 } 2868 return _phase->C->get_alias_index(phi->adr_type()) == _alias; 2869 } 2870 2871 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const { 2872 uint last = _phase-> C->unique(); 2873 MergeMemNode* mm = NULL; 2874 assert(mem->bottom_type() == Type::MEMORY, ""); 2875 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 2876 Node* u = mem->out(i); 2877 if (u != replacement && u->_idx < last) { 2878 if (u->is_MergeMem()) { 2879 MergeMemNode* u_mm = u->as_MergeMem(); 2880 if (u_mm->memory_at(_alias) == mem) { 2881 MergeMemNode* newmm = NULL; 2882 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2883 Node* uu = u->fast_out(j); 2884 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 2885 if (uu->is_Phi()) { 2886 if (should_process_phi(uu)) { 2887 Node* region = uu->in(0); 2888 int nb = 0; 2889 for (uint k = 1; k < uu->req(); k++) { 2890 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) { 2891 if (newmm == NULL) { 2892 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 2893 } 2894 if (newmm != u) { 2895 _phase->igvn().replace_input_of(uu, k, newmm); 2896 nb++; 2897 --jmax; 2898 } 2899 } 2900 } 2901 if (nb > 0) { 2902 --j; 2903 } 2904 } 2905 } else { 2906 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) { 2907 if (newmm == NULL) { 2908 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 2909 } 2910 if (newmm != u) { 2911 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 2912 --j, --jmax; 2913 } 2914 } 2915 } 2916 } 2917 } 2918 } else if (u->is_Phi()) { 2919 assert(u->bottom_type() == Type::MEMORY, "what else?"); 2920 Node* region = u->in(0); 2921 if (should_process_phi(u)) { 2922 bool replaced = false; 2923 for (uint j = 1; j < u->req(); j++) { 2924 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) { 2925 Node* nnew = rep_proj; 2926 if (u->adr_type() == TypePtr::BOTTOM) { 2927 if (mm == NULL) { 2928 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 2929 } 2930 nnew = mm; 2931 } 2932 _phase->igvn().replace_input_of(u, j, nnew); 2933 replaced = true; 2934 } 2935 } 2936 if (replaced) { 2937 --i; 2938 } 2939 2940 } 2941 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 2942 u->adr_type() == NULL) { 2943 assert(u->adr_type() != NULL || 2944 u->Opcode() == Op_Rethrow || 2945 u->Opcode() == Op_Return || 2946 u->Opcode() == Op_SafePoint || 2947 u->Opcode() == Op_StoreIConditional || 2948 u->Opcode() == Op_StoreLConditional || 2949 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 2950 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 2951 u->Opcode() == Op_CallLeaf, "%s", u->Name()); 2952 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 2953 if (mm == NULL) { 2954 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 2955 } 2956 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 2957 --i; 2958 } 2959 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 2960 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 2961 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj); 2962 --i; 2963 } 2964 } 2965 } 2966 } 2967 } 2968 2969 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, bool native) 2970 : Node(ctrl, obj), _native(native) { 2971 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this); 2972 } 2973 2974 bool ShenandoahLoadReferenceBarrierNode::is_native() const { 2975 return _native; 2976 } 2977 2978 uint ShenandoahLoadReferenceBarrierNode::size_of() const { 2979 return sizeof(*this); 2980 } 2981 2982 uint ShenandoahLoadReferenceBarrierNode::hash() const { 2983 return Node::hash() + (_native ? 1 : 0); 2984 } 2985 2986 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const { 2987 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier && 2988 _native == ((const ShenandoahLoadReferenceBarrierNode&)n)._native; 2989 } 2990 2991 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const { 2992 if (in(ValueIn) == NULL || in(ValueIn)->is_top()) { 2993 return Type::TOP; 2994 } 2995 const Type* t = in(ValueIn)->bottom_type(); 2996 if (t == TypePtr::NULL_PTR) { 2997 return t; 2998 } 2999 return t->is_oopptr(); 3000 } 3001 3002 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const { 3003 // Either input is TOP ==> the result is TOP 3004 const Type *t2 = phase->type(in(ValueIn)); 3005 if( t2 == Type::TOP ) return Type::TOP; 3006 3007 if (t2 == TypePtr::NULL_PTR) { 3008 return t2; 3009 } 3010 3011 const Type* type = t2->is_oopptr()/*->cast_to_nonconst()*/; 3012 return type; 3013 } 3014 3015 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) { 3016 Node* value = in(ValueIn); 3017 if (!needs_barrier(phase, value)) { 3018 return value; 3019 } 3020 return this; 3021 } 3022 3023 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) { 3024 Unique_Node_List visited; 3025 return needs_barrier_impl(phase, n, visited); 3026 } 3027 3028 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) { 3029 if (n == NULL) return false; 3030 if (visited.member(n)) { 3031 return false; // Been there. 3032 } 3033 visited.push(n); 3034 3035 if (n->is_Allocate()) { 3036 // tty->print_cr("optimize barrier on alloc"); 3037 return false; 3038 } 3039 if (n->is_Call()) { 3040 // tty->print_cr("optimize barrier on call"); 3041 return false; 3042 } 3043 3044 const Type* type = phase->type(n); 3045 if (type == Type::TOP) { 3046 return false; 3047 } 3048 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) { 3049 // tty->print_cr("optimize barrier on null"); 3050 return false; 3051 } 3052 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) { 3053 // tty->print_cr("optimize barrier on constant"); 3054 return false; 3055 } 3056 3057 switch (n->Opcode()) { 3058 case Op_AddP: 3059 return true; // TODO: Can refine? 3060 case Op_LoadP: 3061 case Op_ShenandoahCompareAndExchangeN: 3062 case Op_ShenandoahCompareAndExchangeP: 3063 case Op_CompareAndExchangeN: 3064 case Op_CompareAndExchangeP: 3065 case Op_GetAndSetN: 3066 case Op_GetAndSetP: 3067 return true; 3068 case Op_Phi: { 3069 for (uint i = 1; i < n->req(); i++) { 3070 if (needs_barrier_impl(phase, n->in(i), visited)) return true; 3071 } 3072 return false; 3073 } 3074 case Op_CheckCastPP: 3075 case Op_CastPP: 3076 return needs_barrier_impl(phase, n->in(1), visited); 3077 case Op_Proj: 3078 return needs_barrier_impl(phase, n->in(0), visited); 3079 case Op_ShenandoahLoadReferenceBarrier: 3080 // tty->print_cr("optimize barrier on barrier"); 3081 return false; 3082 case Op_Parm: 3083 // tty->print_cr("optimize barrier on input arg"); 3084 return false; 3085 case Op_DecodeN: 3086 case Op_EncodeP: 3087 return needs_barrier_impl(phase, n->in(1), visited); 3088 case Op_LoadN: 3089 return true; 3090 case Op_CMoveN: 3091 case Op_CMoveP: 3092 return needs_barrier_impl(phase, n->in(2), visited) || 3093 needs_barrier_impl(phase, n->in(3), visited); 3094 case Op_ShenandoahEnqueueBarrier: 3095 return needs_barrier_impl(phase, n->in(1), visited); 3096 default: 3097 break; 3098 } 3099 #ifdef ASSERT 3100 tty->print("need barrier on?: "); 3101 tty->print_cr("ins:"); 3102 n->dump(2); 3103 tty->print_cr("outs:"); 3104 n->dump(-2); 3105 ShouldNotReachHere(); 3106 #endif 3107 return true; 3108 } 3109 3110 ShenandoahLoadReferenceBarrierNode::Strength ShenandoahLoadReferenceBarrierNode::get_barrier_strength() { 3111 Unique_Node_List visited; 3112 Node_Stack stack(0); 3113 stack.push(this, 0); 3114 Strength strength = NONE; 3115 while (strength != STRONG && stack.size() > 0) { 3116 Node* n = stack.node(); 3117 if (visited.member(n)) { 3118 stack.pop(); 3119 continue; 3120 } 3121 visited.push(n); 3122 bool visit_users = false; 3123 switch (n->Opcode()) { 3124 case Op_StoreN: 3125 case Op_StoreP: { 3126 strength = STRONG; 3127 break; 3128 } 3129 case Op_CmpP: { 3130 if (!n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) && 3131 !n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) { 3132 strength = STRONG; 3133 } 3134 break; 3135 } 3136 case Op_CallStaticJava: { 3137 strength = STRONG; 3138 break; 3139 } 3140 case Op_CallDynamicJava: 3141 case Op_CallLeaf: 3142 case Op_CallLeafNoFP: 3143 case Op_CompareAndSwapL: 3144 case Op_CompareAndSwapI: 3145 case Op_CompareAndSwapB: 3146 case Op_CompareAndSwapS: 3147 case Op_CompareAndSwapN: 3148 case Op_CompareAndSwapP: 3149 case Op_CompareAndExchangeL: 3150 case Op_CompareAndExchangeI: 3151 case Op_CompareAndExchangeB: 3152 case Op_CompareAndExchangeS: 3153 case Op_CompareAndExchangeN: 3154 case Op_CompareAndExchangeP: 3155 case Op_WeakCompareAndSwapL: 3156 case Op_WeakCompareAndSwapI: 3157 case Op_WeakCompareAndSwapB: 3158 case Op_WeakCompareAndSwapS: 3159 case Op_WeakCompareAndSwapN: 3160 case Op_WeakCompareAndSwapP: 3161 case Op_ShenandoahCompareAndSwapN: 3162 case Op_ShenandoahCompareAndSwapP: 3163 case Op_ShenandoahWeakCompareAndSwapN: 3164 case Op_ShenandoahWeakCompareAndSwapP: 3165 case Op_ShenandoahCompareAndExchangeN: 3166 case Op_ShenandoahCompareAndExchangeP: 3167 case Op_GetAndSetL: 3168 case Op_GetAndSetI: 3169 case Op_GetAndSetB: 3170 case Op_GetAndSetS: 3171 case Op_GetAndSetP: 3172 case Op_GetAndSetN: 3173 case Op_GetAndAddL: 3174 case Op_GetAndAddI: 3175 case Op_GetAndAddB: 3176 case Op_GetAndAddS: 3177 case Op_ShenandoahEnqueueBarrier: 3178 case Op_FastLock: 3179 case Op_FastUnlock: 3180 case Op_Rethrow: 3181 case Op_Return: 3182 case Op_StoreB: 3183 case Op_StoreC: 3184 case Op_StoreD: 3185 case Op_StoreF: 3186 case Op_StoreL: 3187 case Op_StoreLConditional: 3188 case Op_StoreI: 3189 case Op_StoreIConditional: 3190 case Op_StoreVector: 3191 case Op_StrInflatedCopy: 3192 case Op_StrCompressedCopy: 3193 case Op_EncodeP: 3194 case Op_CastP2X: 3195 case Op_SafePoint: 3196 case Op_EncodeISOArray: 3197 strength = STRONG; 3198 break; 3199 case Op_LoadB: 3200 case Op_LoadUB: 3201 case Op_LoadUS: 3202 case Op_LoadD: 3203 case Op_LoadF: 3204 case Op_LoadL: 3205 case Op_LoadI: 3206 case Op_LoadS: 3207 case Op_LoadN: 3208 case Op_LoadP: 3209 case Op_LoadVector: { 3210 const TypePtr* adr_type = n->adr_type(); 3211 int alias_idx = Compile::current()->get_alias_index(adr_type); 3212 Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx); 3213 ciField* field = alias_type->field(); 3214 bool is_static = field != NULL && field->is_static(); 3215 bool is_final = field != NULL && field->is_final(); 3216 bool is_stable = field != NULL && field->is_stable(); 3217 if (ShenandoahOptimizeStaticFinals && is_static && is_final) { 3218 // Leave strength as is. 3219 } else if (ShenandoahOptimizeInstanceFinals && !is_static && is_final) { 3220 // Leave strength as is. 3221 } else if (ShenandoahOptimizeStableFinals && (is_stable || (adr_type->isa_aryptr() && adr_type->isa_aryptr()->is_stable()))) { 3222 // Leave strength as is. 3223 } else { 3224 strength = WEAK; 3225 } 3226 break; 3227 } 3228 case Op_AryEq: { 3229 Node* n1 = n->in(2); 3230 Node* n2 = n->in(3); 3231 if (!ShenandoahOptimizeStableFinals || 3232 !n1->bottom_type()->isa_aryptr() || !n1->bottom_type()->isa_aryptr()->is_stable() || 3233 !n2->bottom_type()->isa_aryptr() || !n2->bottom_type()->isa_aryptr()->is_stable()) { 3234 strength = WEAK; 3235 } 3236 break; 3237 } 3238 case Op_StrEquals: 3239 case Op_StrComp: 3240 case Op_StrIndexOf: 3241 case Op_StrIndexOfChar: 3242 case Op_HasNegatives: 3243 if (!ShenandoahOptimizeStableFinals) { 3244 strength = WEAK; 3245 } 3246 break; 3247 case Op_Conv2B: 3248 case Op_LoadRange: 3249 case Op_LoadKlass: 3250 case Op_LoadNKlass: 3251 // NONE, i.e. leave current strength as is 3252 break; 3253 case Op_AddP: 3254 case Op_CheckCastPP: 3255 case Op_CastPP: 3256 case Op_CMoveP: 3257 case Op_Phi: 3258 case Op_ShenandoahLoadReferenceBarrier: 3259 visit_users = true; 3260 break; 3261 default: { 3262 #ifdef ASSERT 3263 tty->print_cr("Unknown node in get_barrier_strength:"); 3264 n->dump(1); 3265 ShouldNotReachHere(); 3266 #else 3267 strength = STRONG; 3268 #endif 3269 } 3270 } 3271 #ifdef ASSERT 3272 /* 3273 if (strength == STRONG) { 3274 tty->print("strengthening node: "); 3275 n->dump(); 3276 } 3277 */ 3278 #endif 3279 stack.pop(); 3280 if (visit_users) { 3281 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3282 Node* user = n->fast_out(i); 3283 if (user != NULL) { 3284 stack.push(user, 0); 3285 } 3286 } 3287 } 3288 } 3289 return strength; 3290 } 3291 3292 CallStaticJavaNode* ShenandoahLoadReferenceBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) { 3293 Node* val = in(ValueIn); 3294 3295 const Type* val_t = igvn.type(val); 3296 3297 if (val_t->meet(TypePtr::NULL_PTR) != val_t && 3298 val->Opcode() == Op_CastPP && 3299 val->in(0) != NULL && 3300 val->in(0)->Opcode() == Op_IfTrue && 3301 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 3302 val->in(0)->in(0)->is_If() && 3303 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool && 3304 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne && 3305 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 3306 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) && 3307 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 3308 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), ""); 3309 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 3310 return unc; 3311 } 3312 return NULL; 3313 }