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 for( uint i = 0; i < n->len(); ++i ) { 674 Node *m = n->in(i); 675 if (m == NULL) continue; 676 677 // In most cases, inputs should be known to be non null. If it's 678 // not the case, it could be a missing cast_not_null() in an 679 // intrinsic or support might be needed in AddPNode::Ideal() to 680 // avoid a NULL+offset input. 681 if (!(n->is_Phi() || 682 (n->is_SafePoint() && (!n->is_CallRuntime() || !strcmp(n->as_Call()->_name, "shenandoah_wb_pre") || !strcmp(n->as_Call()->_name, "unsafe_arraycopy"))) || 683 n->Opcode() == Op_CmpP || 684 n->Opcode() == Op_CmpN || 685 (n->Opcode() == Op_StoreP && i == StoreNode::ValueIn) || 686 (n->Opcode() == Op_StoreN && i == StoreNode::ValueIn) || 687 n->is_ConstraintCast() || 688 n->Opcode() == Op_Return || 689 n->Opcode() == Op_Conv2B || 690 n->is_AddP() || 691 n->Opcode() == Op_CMoveP || 692 n->Opcode() == Op_CMoveN || 693 n->Opcode() == Op_Rethrow || 694 n->is_MemBar() || 695 n->is_Mem() || 696 n->Opcode() == Op_AryEq || 697 n->Opcode() == Op_SCMemProj || 698 n->Opcode() == Op_EncodeP || 699 n->Opcode() == Op_DecodeN || 700 n->Opcode() == Op_ShenandoahEnqueueBarrier || 701 n->Opcode() == Op_ShenandoahLoadReferenceBarrier)) { 702 if (m->bottom_type()->make_oopptr() && m->bottom_type()->make_oopptr()->meet(TypePtr::NULL_PTR) == m->bottom_type()) { 703 report_verify_failure("Shenandoah verification: null input", n, m); 704 } 705 } 706 707 wq.push(m); 708 } 709 } 710 711 if (verify_no_useless_barrier) { 712 for (int i = 0; i < barriers.length(); i++) { 713 Node* n = barriers.at(i); 714 if (!barriers_used.member(n)) { 715 tty->print("XXX useless barrier"); n->dump(-2); 716 ShouldNotReachHere(); 717 } 718 } 719 } 720 } 721 #endif 722 723 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) { 724 // That both nodes have the same control is not sufficient to prove 725 // domination, verify that there's no path from d to n 726 ResourceMark rm; 727 Unique_Node_List wq; 728 wq.push(d); 729 for (uint next = 0; next < wq.size(); next++) { 730 Node *m = wq.at(next); 731 if (m == n) { 732 return false; 733 } 734 if (m->is_Phi() && m->in(0)->is_Loop()) { 735 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control"); 736 } else { 737 for (uint i = 0; i < m->req(); i++) { 738 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) { 739 wq.push(m->in(i)); 740 } 741 } 742 } 743 } 744 return true; 745 } 746 747 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) { 748 if (d_c != n_c) { 749 return phase->is_dominator(d_c, n_c); 750 } 751 return is_dominator_same_ctrl(d_c, d, n, phase); 752 } 753 754 Node* next_mem(Node* mem, int alias) { 755 Node* res = NULL; 756 if (mem->is_Proj()) { 757 res = mem->in(0); 758 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 759 res = mem->in(TypeFunc::Memory); 760 } else if (mem->is_Phi()) { 761 res = mem->in(1); 762 } else if (mem->is_MergeMem()) { 763 res = mem->as_MergeMem()->memory_at(alias); 764 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 765 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier"); 766 res = mem->in(MemNode::Memory); 767 } else { 768 #ifdef ASSERT 769 mem->dump(); 770 #endif 771 ShouldNotReachHere(); 772 } 773 return res; 774 } 775 776 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) { 777 Node* iffproj = NULL; 778 while (c != dom) { 779 Node* next = phase->idom(c); 780 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?"); 781 if (c->is_Region()) { 782 ResourceMark rm; 783 Unique_Node_List wq; 784 wq.push(c); 785 for (uint i = 0; i < wq.size(); i++) { 786 Node *n = wq.at(i); 787 if (n == next) { 788 continue; 789 } 790 if (n->is_Region()) { 791 for (uint j = 1; j < n->req(); j++) { 792 wq.push(n->in(j)); 793 } 794 } else { 795 wq.push(n->in(0)); 796 } 797 } 798 for (uint i = 0; i < wq.size(); i++) { 799 Node *n = wq.at(i); 800 assert(n->is_CFG(), ""); 801 if (n->is_Multi()) { 802 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 803 Node* u = n->fast_out(j); 804 if (u->is_CFG()) { 805 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 806 return NodeSentinel; 807 } 808 } 809 } 810 } 811 } 812 } else if (c->is_Proj()) { 813 if (c->is_IfProj()) { 814 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) { 815 // continue; 816 } else { 817 if (!allow_one_proj) { 818 return NodeSentinel; 819 } 820 if (iffproj == NULL) { 821 iffproj = c; 822 } else { 823 return NodeSentinel; 824 } 825 } 826 } else if (c->Opcode() == Op_JumpProj) { 827 return NodeSentinel; // unsupported 828 } else if (c->Opcode() == Op_CatchProj) { 829 return NodeSentinel; // unsupported 830 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) { 831 return NodeSentinel; // unsupported 832 } else { 833 assert(next->unique_ctrl_out() == c, "unsupported branch pattern"); 834 } 835 } 836 c = next; 837 } 838 return iffproj; 839 } 840 841 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) { 842 ResourceMark rm; 843 VectorSet wq(Thread::current()->resource_area()); 844 wq.set(mem->_idx); 845 mem_ctrl = phase->ctrl_or_self(mem); 846 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) { 847 mem = next_mem(mem, alias); 848 if (wq.test_set(mem->_idx)) { 849 return NULL; 850 } 851 mem_ctrl = phase->ctrl_or_self(mem); 852 } 853 if (mem->is_MergeMem()) { 854 mem = mem->as_MergeMem()->memory_at(alias); 855 mem_ctrl = phase->ctrl_or_self(mem); 856 } 857 return mem; 858 } 859 860 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) { 861 Node* mem = NULL; 862 Node* c = ctrl; 863 do { 864 if (c->is_Region()) { 865 Node* phi_bottom = NULL; 866 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) { 867 Node* u = c->fast_out(i); 868 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) { 869 if (u->adr_type() == TypePtr::BOTTOM) { 870 mem = u; 871 } 872 } 873 } 874 } else { 875 if (c->is_Call() && c->as_Call()->adr_type() != NULL) { 876 CallProjections projs; 877 c->as_Call()->extract_projections(&projs, true, false); 878 if (projs.fallthrough_memproj != NULL) { 879 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) { 880 if (projs.catchall_memproj == NULL) { 881 mem = projs.fallthrough_memproj; 882 } else { 883 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) { 884 mem = projs.fallthrough_memproj; 885 } else { 886 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier"); 887 mem = projs.catchall_memproj; 888 } 889 } 890 } 891 } else { 892 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory); 893 if (proj != NULL && 894 proj->adr_type() == TypePtr::BOTTOM) { 895 mem = proj; 896 } 897 } 898 } else { 899 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 900 Node* u = c->fast_out(i); 901 if (u->is_Proj() && 902 u->bottom_type() == Type::MEMORY && 903 u->adr_type() == TypePtr::BOTTOM) { 904 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), ""); 905 assert(mem == NULL, "only one proj"); 906 mem = u; 907 } 908 } 909 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected"); 910 } 911 } 912 c = phase->idom(c); 913 } while (mem == NULL); 914 return mem; 915 } 916 917 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) { 918 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 919 Node* u = n->fast_out(i); 920 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) { 921 uses.push(u); 922 } 923 } 924 } 925 926 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) { 927 OuterStripMinedLoopEndNode* le = inner->outer_loop_end(); 928 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl)); 929 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl)); 930 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt); 931 phase->register_control(new_le, phase->get_loop(le), le->in(0)); 932 phase->lazy_replace(outer, new_outer); 933 phase->lazy_replace(le, new_le); 934 inner->clear_strip_mined(); 935 } 936 937 void ShenandoahBarrierC2Support::test_heap_stable(Node*& ctrl, Node* raw_mem, Node*& heap_stable_ctrl, 938 PhaseIdealLoop* phase) { 939 IdealLoopTree* loop = phase->get_loop(ctrl); 940 Node* thread = new ThreadLocalNode(); 941 phase->register_new_node(thread, ctrl); 942 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 943 phase->set_ctrl(offset, phase->C->root()); 944 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset); 945 phase->register_new_node(gc_state_addr, ctrl); 946 uint gc_state_idx = Compile::AliasIdxRaw; 947 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument 948 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx)); 949 950 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered); 951 phase->register_new_node(gc_state, ctrl); 952 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)); 953 phase->register_new_node(heap_stable_and, ctrl); 954 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT)); 955 phase->register_new_node(heap_stable_cmp, ctrl); 956 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne); 957 phase->register_new_node(heap_stable_test, ctrl); 958 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 959 phase->register_control(heap_stable_iff, loop, ctrl); 960 961 heap_stable_ctrl = new IfFalseNode(heap_stable_iff); 962 phase->register_control(heap_stable_ctrl, loop, heap_stable_iff); 963 ctrl = new IfTrueNode(heap_stable_iff); 964 phase->register_control(ctrl, loop, heap_stable_iff); 965 966 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape"); 967 } 968 969 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) { 970 const Type* val_t = phase->igvn().type(val); 971 if (val_t->meet(TypePtr::NULL_PTR) == val_t) { 972 IdealLoopTree* loop = phase->get_loop(ctrl); 973 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT)); 974 phase->register_new_node(null_cmp, ctrl); 975 Node* null_test = new BoolNode(null_cmp, BoolTest::ne); 976 phase->register_new_node(null_test, ctrl); 977 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 978 phase->register_control(null_iff, loop, ctrl); 979 ctrl = new IfTrueNode(null_iff); 980 phase->register_control(ctrl, loop, null_iff); 981 null_ctrl = new IfFalseNode(null_iff); 982 phase->register_control(null_ctrl, loop, null_iff); 983 } 984 } 985 986 Node* ShenandoahBarrierC2Support::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) { 987 IdealLoopTree *loop = phase->get_loop(c); 988 Node* iff = unc_ctrl->in(0); 989 assert(iff->is_If(), "broken"); 990 Node* new_iff = iff->clone(); 991 new_iff->set_req(0, c); 992 phase->register_control(new_iff, loop, c); 993 Node* iffalse = new IfFalseNode(new_iff->as_If()); 994 phase->register_control(iffalse, loop, new_iff); 995 Node* iftrue = new IfTrueNode(new_iff->as_If()); 996 phase->register_control(iftrue, loop, new_iff); 997 c = iftrue; 998 const Type *t = phase->igvn().type(val); 999 assert(val->Opcode() == Op_CastPP, "expect cast to non null here"); 1000 Node* uncasted_val = val->in(1); 1001 val = new CastPPNode(uncasted_val, t); 1002 val->init_req(0, c); 1003 phase->register_new_node(val, c); 1004 return val; 1005 } 1006 1007 void ShenandoahBarrierC2Support::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl, 1008 Unique_Node_List& uses, PhaseIdealLoop* phase) { 1009 IfNode* iff = unc_ctrl->in(0)->as_If(); 1010 Node* proj = iff->proj_out(0); 1011 assert(proj != unc_ctrl, "bad projection"); 1012 Node* use = proj->unique_ctrl_out(); 1013 1014 assert(use == unc || use->is_Region(), "what else?"); 1015 1016 uses.clear(); 1017 if (use == unc) { 1018 phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use)); 1019 for (uint i = 1; i < unc->req(); i++) { 1020 Node* n = unc->in(i); 1021 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) { 1022 uses.push(n); 1023 } 1024 } 1025 } else { 1026 assert(use->is_Region(), "what else?"); 1027 uint idx = 1; 1028 for (; use->in(idx) != proj; idx++); 1029 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 1030 Node* u = use->fast_out(i); 1031 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) { 1032 uses.push(u->in(idx)); 1033 } 1034 } 1035 } 1036 for(uint next = 0; next < uses.size(); next++ ) { 1037 Node *n = uses.at(next); 1038 assert(phase->get_ctrl(n) == proj, "bad control"); 1039 phase->set_ctrl_and_loop(n, new_unc_ctrl); 1040 if (n->in(0) == proj) { 1041 phase->igvn().replace_input_of(n, 0, new_unc_ctrl); 1042 } 1043 for (uint i = 0; i < n->req(); i++) { 1044 Node* m = n->in(i); 1045 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) { 1046 uses.push(m); 1047 } 1048 } 1049 } 1050 1051 phase->igvn().rehash_node_delayed(use); 1052 int nb = use->replace_edge(proj, new_unc_ctrl); 1053 assert(nb == 1, "only use expected"); 1054 } 1055 1056 void ShenandoahBarrierC2Support::in_cset_fast_test(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) { 1057 IdealLoopTree *loop = phase->get_loop(ctrl); 1058 Node* raw_rbtrue = new CastP2XNode(ctrl, val); 1059 phase->register_new_node(raw_rbtrue, ctrl); 1060 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint())); 1061 phase->register_new_node(cset_offset, ctrl); 1062 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr())); 1063 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root()); 1064 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset); 1065 phase->register_new_node(in_cset_fast_test_adr, ctrl); 1066 uint in_cset_fast_test_idx = Compile::AliasIdxRaw; 1067 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument 1068 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx)); 1069 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); 1070 phase->register_new_node(in_cset_fast_test_load, ctrl); 1071 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT)); 1072 phase->register_new_node(in_cset_fast_test_cmp, ctrl); 1073 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::eq); 1074 phase->register_new_node(in_cset_fast_test_test, ctrl); 1075 IfNode* in_cset_fast_test_iff = new IfNode(ctrl, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 1076 phase->register_control(in_cset_fast_test_iff, loop, ctrl); 1077 1078 not_cset_ctrl = new IfTrueNode(in_cset_fast_test_iff); 1079 phase->register_control(not_cset_ctrl, loop, in_cset_fast_test_iff); 1080 1081 ctrl = new IfFalseNode(in_cset_fast_test_iff); 1082 phase->register_control(ctrl, loop, in_cset_fast_test_iff); 1083 } 1084 1085 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node*& result_mem, Node* raw_mem, PhaseIdealLoop* phase) { 1086 IdealLoopTree*loop = phase->get_loop(ctrl); 1087 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr()->cast_to_nonconst(); 1088 1089 // The slow path stub consumes and produces raw memory in addition 1090 // to the existing memory edges 1091 Node* base = find_bottom_mem(ctrl, phase); 1092 MergeMemNode* mm = MergeMemNode::make(base); 1093 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 1094 phase->register_new_node(mm, ctrl); 1095 1096 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_write_barrier_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT), "shenandoah_write_barrier", TypeRawPtr::BOTTOM); 1097 call->init_req(TypeFunc::Control, ctrl); 1098 call->init_req(TypeFunc::I_O, phase->C->top()); 1099 call->init_req(TypeFunc::Memory, mm); 1100 call->init_req(TypeFunc::FramePtr, phase->C->top()); 1101 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 1102 call->init_req(TypeFunc::Parms, val); 1103 phase->register_control(call, loop, ctrl); 1104 ctrl = new ProjNode(call, TypeFunc::Control); 1105 phase->register_control(ctrl, loop, call); 1106 result_mem = new ProjNode(call, TypeFunc::Memory); 1107 phase->register_new_node(result_mem, call); 1108 val = new ProjNode(call, TypeFunc::Parms); 1109 phase->register_new_node(val, call); 1110 val = new CheckCastPPNode(ctrl, val, obj_type); 1111 phase->register_new_node(val, ctrl); 1112 } 1113 1114 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) { 1115 Node* ctrl = phase->get_ctrl(barrier); 1116 Node* init_raw_mem = fixer.find_mem(ctrl, barrier); 1117 1118 // Update the control of all nodes that should be after the 1119 // barrier control flow 1120 uses.clear(); 1121 // Every node that is control dependent on the barrier's input 1122 // control will be after the expanded barrier. The raw memory (if 1123 // its memory is control dependent on the barrier's input control) 1124 // must stay above the barrier. 1125 uses_to_ignore.clear(); 1126 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) { 1127 uses_to_ignore.push(init_raw_mem); 1128 } 1129 for (uint next = 0; next < uses_to_ignore.size(); next++) { 1130 Node *n = uses_to_ignore.at(next); 1131 for (uint i = 0; i < n->req(); i++) { 1132 Node* in = n->in(i); 1133 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) { 1134 uses_to_ignore.push(in); 1135 } 1136 } 1137 } 1138 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) { 1139 Node* u = ctrl->fast_out(i); 1140 if (u->_idx < last && 1141 u != barrier && 1142 !uses_to_ignore.member(u) && 1143 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) && 1144 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) { 1145 Node* old_c = phase->ctrl_or_self(u); 1146 Node* c = old_c; 1147 if (c != ctrl || 1148 is_dominator_same_ctrl(old_c, barrier, u, phase) || 1149 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) { 1150 phase->igvn().rehash_node_delayed(u); 1151 int nb = u->replace_edge(ctrl, region); 1152 if (u->is_CFG()) { 1153 if (phase->idom(u) == ctrl) { 1154 phase->set_idom(u, region, phase->dom_depth(region)); 1155 } 1156 } else if (phase->get_ctrl(u) == ctrl) { 1157 assert(u != init_raw_mem, "should leave input raw mem above the barrier"); 1158 uses.push(u); 1159 } 1160 assert(nb == 1, "more than 1 ctrl input?"); 1161 --i, imax -= nb; 1162 } 1163 } 1164 } 1165 } 1166 1167 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) { 1168 Node* region = NULL; 1169 while (c != ctrl) { 1170 if (c->is_Region()) { 1171 region = c; 1172 } 1173 c = phase->idom(c); 1174 } 1175 assert(region != NULL, ""); 1176 Node* phi = new PhiNode(region, n->bottom_type()); 1177 for (uint j = 1; j < region->req(); j++) { 1178 Node* in = region->in(j); 1179 if (phase->is_dominator(projs.fallthrough_catchproj, in)) { 1180 phi->init_req(j, n); 1181 } else if (phase->is_dominator(projs.catchall_catchproj, in)) { 1182 phi->init_req(j, n_clone); 1183 } else { 1184 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase)); 1185 } 1186 } 1187 phase->register_new_node(phi, region); 1188 return phi; 1189 } 1190 1191 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) { 1192 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state(); 1193 1194 // Collect raw memory state at CFG points in the entire graph and 1195 // record it in memory_nodes. Optimize the raw memory graph in the 1196 // process. Optimizing the memory graph also makes the memory graph 1197 // simpler. 1198 GrowableArray<MemoryGraphFixer*> memory_graph_fixers; 1199 1200 Unique_Node_List uses; 1201 for (int i = 0; i < state->enqueue_barriers_count(); i++) { 1202 Node* barrier = state->enqueue_barrier(i); 1203 Node* ctrl = phase->get_ctrl(barrier); 1204 IdealLoopTree* loop = phase->get_loop(ctrl); 1205 if (loop->_head->is_OuterStripMinedLoop()) { 1206 // Expanding a barrier here will break loop strip mining 1207 // verification. Transform the loop so the loop nest doesn't 1208 // appear as strip mined. 1209 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop(); 1210 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase); 1211 } 1212 } 1213 1214 Node_Stack stack(0); 1215 Node_List clones; 1216 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) { 1217 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1218 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1219 continue; 1220 } 1221 1222 Node* ctrl = phase->get_ctrl(lrb); 1223 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 1224 1225 CallStaticJavaNode* unc = NULL; 1226 Node* unc_ctrl = NULL; 1227 Node* uncasted_val = val; 1228 1229 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) { 1230 Node* u = lrb->fast_out(i); 1231 if (u->Opcode() == Op_CastPP && 1232 u->in(0) != NULL && 1233 phase->is_dominator(u->in(0), ctrl)) { 1234 const Type* u_t = phase->igvn().type(u); 1235 1236 if (u_t->meet(TypePtr::NULL_PTR) != u_t && 1237 u->in(0)->Opcode() == Op_IfTrue && 1238 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 1239 u->in(0)->in(0)->is_If() && 1240 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool && 1241 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne && 1242 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 1243 u->in(0)->in(0)->in(1)->in(1)->in(1) == val && 1244 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 1245 IdealLoopTree* loop = phase->get_loop(ctrl); 1246 IdealLoopTree* unc_loop = phase->get_loop(u->in(0)); 1247 1248 if (!unc_loop->is_member(loop)) { 1249 continue; 1250 } 1251 1252 Node* branch = no_branches(ctrl, u->in(0), false, phase); 1253 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch"); 1254 if (branch == NodeSentinel) { 1255 continue; 1256 } 1257 1258 phase->igvn().replace_input_of(u, 1, val); 1259 phase->igvn().replace_input_of(lrb, ShenandoahLoadReferenceBarrierNode::ValueIn, u); 1260 phase->set_ctrl(u, u->in(0)); 1261 phase->set_ctrl(lrb, u->in(0)); 1262 unc = u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1263 unc_ctrl = u->in(0); 1264 val = u; 1265 1266 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 1267 Node* u = val->fast_out(j); 1268 if (u == lrb) continue; 1269 phase->igvn().rehash_node_delayed(u); 1270 int nb = u->replace_edge(val, lrb); 1271 --j; jmax -= nb; 1272 } 1273 1274 RegionNode* r = new RegionNode(3); 1275 IfNode* iff = unc_ctrl->in(0)->as_If(); 1276 1277 Node* ctrl_use = unc_ctrl->unique_ctrl_out(); 1278 Node* unc_ctrl_clone = unc_ctrl->clone(); 1279 phase->register_control(unc_ctrl_clone, loop, iff); 1280 Node* c = unc_ctrl_clone; 1281 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase); 1282 r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0)); 1283 1284 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0)); 1285 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl)); 1286 phase->lazy_replace(c, unc_ctrl); 1287 c = NULL;; 1288 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone); 1289 phase->set_ctrl(val, unc_ctrl_clone); 1290 1291 IfNode* new_iff = new_cast->in(0)->in(0)->as_If(); 1292 fix_null_check(unc, unc_ctrl_clone, r, uses, phase); 1293 Node* iff_proj = iff->proj_out(0); 1294 r->init_req(2, iff_proj); 1295 phase->register_control(r, phase->ltree_root(), iff); 1296 1297 Node* new_bol = new_iff->in(1)->clone(); 1298 Node* new_cmp = new_bol->in(1)->clone(); 1299 assert(new_cmp->Opcode() == Op_CmpP, "broken"); 1300 assert(new_cmp->in(1) == val->in(1), "broken"); 1301 new_bol->set_req(1, new_cmp); 1302 new_cmp->set_req(1, lrb); 1303 phase->register_new_node(new_bol, new_iff->in(0)); 1304 phase->register_new_node(new_cmp, new_iff->in(0)); 1305 phase->igvn().replace_input_of(new_iff, 1, new_bol); 1306 phase->igvn().replace_input_of(new_cast, 1, lrb); 1307 1308 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) { 1309 Node* u = lrb->fast_out(i); 1310 if (u == new_cast || u == new_cmp) { 1311 continue; 1312 } 1313 phase->igvn().rehash_node_delayed(u); 1314 int nb = u->replace_edge(lrb, new_cast); 1315 assert(nb > 0, "no update?"); 1316 --i; imax -= nb; 1317 } 1318 1319 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 1320 Node* u = val->fast_out(i); 1321 if (u == lrb) { 1322 continue; 1323 } 1324 phase->igvn().rehash_node_delayed(u); 1325 int nb = u->replace_edge(val, new_cast); 1326 assert(nb > 0, "no update?"); 1327 --i; imax -= nb; 1328 } 1329 1330 ctrl = unc_ctrl_clone; 1331 phase->set_ctrl_and_loop(lrb, ctrl); 1332 break; 1333 } 1334 } 1335 } 1336 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) { 1337 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava(); 1338 CallProjections projs; 1339 call->extract_projections(&projs, false, false); 1340 1341 Node* lrb_clone = lrb->clone(); 1342 phase->register_new_node(lrb_clone, projs.catchall_catchproj); 1343 phase->set_ctrl(lrb, projs.fallthrough_catchproj); 1344 1345 stack.push(lrb, 0); 1346 clones.push(lrb_clone); 1347 1348 do { 1349 assert(stack.size() == clones.size(), ""); 1350 Node* n = stack.node(); 1351 #ifdef ASSERT 1352 if (n->is_Load()) { 1353 Node* mem = n->in(MemNode::Memory); 1354 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) { 1355 Node* u = mem->fast_out(j); 1356 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?"); 1357 } 1358 } 1359 #endif 1360 uint idx = stack.index(); 1361 Node* n_clone = clones.at(clones.size()-1); 1362 if (idx < n->outcnt()) { 1363 Node* u = n->raw_out(idx); 1364 Node* c = phase->ctrl_or_self(u); 1365 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) { 1366 stack.set_index(idx+1); 1367 assert(!u->is_CFG(), ""); 1368 stack.push(u, 0); 1369 Node* u_clone = u->clone(); 1370 int nb = u_clone->replace_edge(n, n_clone); 1371 assert(nb > 0, "should have replaced some uses"); 1372 phase->register_new_node(u_clone, projs.catchall_catchproj); 1373 clones.push(u_clone); 1374 phase->set_ctrl(u, projs.fallthrough_catchproj); 1375 } else { 1376 bool replaced = false; 1377 if (u->is_Phi()) { 1378 for (uint k = 1; k < u->req(); k++) { 1379 if (u->in(k) == n) { 1380 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) { 1381 phase->igvn().replace_input_of(u, k, n_clone); 1382 replaced = true; 1383 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) { 1384 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase)); 1385 replaced = true; 1386 } 1387 } 1388 } 1389 } else { 1390 if (phase->is_dominator(projs.catchall_catchproj, c)) { 1391 phase->igvn().rehash_node_delayed(u); 1392 int nb = u->replace_edge(n, n_clone); 1393 assert(nb > 0, "should have replaced some uses"); 1394 replaced = true; 1395 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) { 1396 phase->igvn().rehash_node_delayed(u); 1397 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase)); 1398 assert(nb > 0, "should have replaced some uses"); 1399 replaced = true; 1400 } 1401 } 1402 if (!replaced) { 1403 stack.set_index(idx+1); 1404 } 1405 } 1406 } else { 1407 stack.pop(); 1408 clones.pop(); 1409 } 1410 } while (stack.size() > 0); 1411 assert(stack.size() == 0 && clones.size() == 0, ""); 1412 } 1413 } 1414 1415 // Expand load-reference-barriers 1416 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase); 1417 Unique_Node_List uses_to_ignore; 1418 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) { 1419 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1420 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1421 phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn)); 1422 continue; 1423 } 1424 uint last = phase->C->unique(); 1425 Node* ctrl = phase->get_ctrl(lrb); 1426 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 1427 1428 1429 Node* orig_ctrl = ctrl; 1430 1431 Node* raw_mem = fixer.find_mem(ctrl, lrb); 1432 Node* init_raw_mem = raw_mem; 1433 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 1434 // int alias = phase->C->get_alias_index(lrb->adr_type()); 1435 1436 IdealLoopTree *loop = phase->get_loop(ctrl); 1437 CallStaticJavaNode* unc = lrb->pin_and_expand_null_check(phase->igvn()); 1438 Node* unc_ctrl = NULL; 1439 if (unc != NULL) { 1440 if (val->in(ShenandoahLoadReferenceBarrierNode::Control) != ctrl) { 1441 unc = NULL; 1442 } else { 1443 unc_ctrl = val->in(ShenandoahLoadReferenceBarrierNode::Control); 1444 } 1445 } 1446 1447 Node* uncasted_val = val; 1448 if (unc != NULL) { 1449 uncasted_val = val->in(1); 1450 } 1451 1452 Node* heap_stable_ctrl = NULL; 1453 Node* null_ctrl = NULL; 1454 1455 assert(val->bottom_type()->make_oopptr(), "need oop"); 1456 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant"); 1457 1458 enum { _heap_stable = 1, _not_cset, _not_equal, _evac_path, _null_path, PATH_LIMIT }; 1459 Node* region = new RegionNode(PATH_LIMIT); 1460 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr()); 1461 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1462 1463 // Stable path. 1464 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase); 1465 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If(); 1466 1467 // Heap stable case 1468 region->init_req(_heap_stable, heap_stable_ctrl); 1469 val_phi->init_req(_heap_stable, uncasted_val); 1470 raw_mem_phi->init_req(_heap_stable, raw_mem); 1471 1472 Node* reg2_ctrl = NULL; 1473 // Null case 1474 test_null(ctrl, val, null_ctrl, phase); 1475 if (null_ctrl != NULL) { 1476 reg2_ctrl = null_ctrl->in(0); 1477 region->init_req(_null_path, null_ctrl); 1478 val_phi->init_req(_null_path, uncasted_val); 1479 raw_mem_phi->init_req(_null_path, raw_mem); 1480 } else { 1481 region->del_req(_null_path); 1482 val_phi->del_req(_null_path); 1483 raw_mem_phi->del_req(_null_path); 1484 } 1485 1486 // Test for in-cset. 1487 // Wires !in_cset(obj) to slot 2 of region and phis 1488 Node* not_cset_ctrl = NULL; 1489 in_cset_fast_test(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase); 1490 if (not_cset_ctrl != NULL) { 1491 if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0); 1492 region->init_req(_not_cset, not_cset_ctrl); 1493 val_phi->init_req(_not_cset, uncasted_val); 1494 raw_mem_phi->init_req(_not_cset, raw_mem); 1495 } 1496 1497 // Resolve object when orig-value is in cset. 1498 // Make the unconditional resolve for fwdptr. 1499 Node* new_val = uncasted_val; 1500 if (unc_ctrl != NULL) { 1501 // Clone the null check in this branch to allow implicit null check 1502 new_val = clone_null_check(ctrl, val, unc_ctrl, phase); 1503 fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase); 1504 1505 IfNode* iff = unc_ctrl->in(0)->as_If(); 1506 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1)); 1507 } 1508 Node* addr = new AddPNode(new_val, uncasted_val, phase->igvn().MakeConX(ShenandoahForwarding::byte_offset())); 1509 phase->register_new_node(addr, ctrl); 1510 assert(val->bottom_type()->isa_oopptr(), "what else?"); 1511 const TypePtr* obj_type = val->bottom_type()->is_oopptr(); 1512 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 1513 Node* fwd = new LoadPNode(ctrl, raw_mem, addr, adr_type, obj_type, MemNode::unordered); 1514 phase->register_new_node(fwd, ctrl); 1515 1516 // Only branch to LRB stub if object is not forwarded; otherwise reply with fwd ptr 1517 Node* cmp = new CmpPNode(fwd, new_val); 1518 phase->register_new_node(cmp, ctrl); 1519 Node* bol = new BoolNode(cmp, BoolTest::eq); 1520 phase->register_new_node(bol, ctrl); 1521 1522 IfNode* iff = new IfNode(ctrl, bol, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 1523 if (reg2_ctrl == NULL) reg2_ctrl = iff; 1524 phase->register_control(iff, loop, ctrl); 1525 Node* if_not_eq = new IfFalseNode(iff); 1526 phase->register_control(if_not_eq, loop, iff); 1527 Node* if_eq = new IfTrueNode(iff); 1528 phase->register_control(if_eq, loop, iff); 1529 1530 // Wire up not-equal-path in slots 3. 1531 region->init_req(_not_equal, if_not_eq); 1532 val_phi->init_req(_not_equal, fwd); 1533 raw_mem_phi->init_req(_not_equal, raw_mem); 1534 1535 // Call wb-stub and wire up that path in slots 4 1536 Node* result_mem = NULL; 1537 ctrl = if_eq; 1538 call_lrb_stub(ctrl, fwd, result_mem, raw_mem, phase); 1539 region->init_req(_evac_path, ctrl); 1540 val_phi->init_req(_evac_path, fwd); 1541 raw_mem_phi->init_req(_evac_path, result_mem); 1542 1543 phase->register_control(region, loop, heap_stable_iff); 1544 Node* out_val = val_phi; 1545 phase->register_new_node(val_phi, region); 1546 phase->register_new_node(raw_mem_phi, region); 1547 1548 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase); 1549 1550 ctrl = orig_ctrl; 1551 1552 if (unc != NULL) { 1553 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 1554 Node* u = val->fast_out(i); 1555 Node* c = phase->ctrl_or_self(u); 1556 if (u != lrb && (c != ctrl || is_dominator_same_ctrl(c, lrb, u, phase))) { 1557 phase->igvn().rehash_node_delayed(u); 1558 int nb = u->replace_edge(val, out_val); 1559 --i, imax -= nb; 1560 } 1561 } 1562 if (val->outcnt() == 0) { 1563 phase->igvn()._worklist.push(val); 1564 } 1565 } 1566 phase->igvn().replace_node(lrb, out_val); 1567 1568 follow_barrier_uses(out_val, ctrl, uses, phase); 1569 1570 for(uint next = 0; next < uses.size(); next++ ) { 1571 Node *n = uses.at(next); 1572 assert(phase->get_ctrl(n) == ctrl, "bad control"); 1573 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 1574 phase->set_ctrl(n, region); 1575 follow_barrier_uses(n, ctrl, uses, phase); 1576 } 1577 1578 // The slow path call produces memory: hook the raw memory phi 1579 // from the expanded load reference barrier with the rest of the graph 1580 // which may require adding memory phis at every post dominated 1581 // region and at enclosing loop heads. Use the memory state 1582 // collected in memory_nodes to fix the memory graph. Update that 1583 // memory state as we go. 1584 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses); 1585 } 1586 // Done expanding load-reference-barriers. 1587 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced"); 1588 1589 for (int i = state->enqueue_barriers_count() - 1; i >= 0; i--) { 1590 Node* barrier = state->enqueue_barrier(i); 1591 Node* pre_val = barrier->in(1); 1592 1593 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) { 1594 ShouldNotReachHere(); 1595 continue; 1596 } 1597 1598 Node* ctrl = phase->get_ctrl(barrier); 1599 1600 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) { 1601 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move"); 1602 ctrl = ctrl->in(0)->in(0); 1603 phase->set_ctrl(barrier, ctrl); 1604 } else if (ctrl->is_CallRuntime()) { 1605 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move"); 1606 ctrl = ctrl->in(0); 1607 phase->set_ctrl(barrier, ctrl); 1608 } 1609 1610 Node* init_ctrl = ctrl; 1611 IdealLoopTree* loop = phase->get_loop(ctrl); 1612 Node* raw_mem = fixer.find_mem(ctrl, barrier); 1613 Node* init_raw_mem = raw_mem; 1614 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 1615 Node* heap_stable_ctrl = NULL; 1616 Node* null_ctrl = NULL; 1617 uint last = phase->C->unique(); 1618 1619 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT }; 1620 Node* region = new RegionNode(PATH_LIMIT); 1621 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1622 1623 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 }; 1624 Node* region2 = new RegionNode(PATH_LIMIT2); 1625 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1626 1627 // Stable path. 1628 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase); 1629 region->init_req(_heap_stable, heap_stable_ctrl); 1630 phi->init_req(_heap_stable, raw_mem); 1631 1632 // Null path 1633 Node* reg2_ctrl = NULL; 1634 test_null(ctrl, pre_val, null_ctrl, phase); 1635 if (null_ctrl != NULL) { 1636 reg2_ctrl = null_ctrl->in(0); 1637 region2->init_req(_null_path, null_ctrl); 1638 phi2->init_req(_null_path, raw_mem); 1639 } else { 1640 region2->del_req(_null_path); 1641 phi2->del_req(_null_path); 1642 } 1643 1644 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 1645 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 1646 Node* thread = new ThreadLocalNode(); 1647 phase->register_new_node(thread, ctrl); 1648 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset)); 1649 phase->register_new_node(buffer_adr, ctrl); 1650 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset)); 1651 phase->register_new_node(index_adr, ctrl); 1652 1653 BasicType index_bt = TypeX_X->basic_type(); 1654 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 1655 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 1656 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered); 1657 phase->register_new_node(index, ctrl); 1658 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0)); 1659 phase->register_new_node(index_cmp, ctrl); 1660 Node* index_test = new BoolNode(index_cmp, BoolTest::ne); 1661 phase->register_new_node(index_test, ctrl); 1662 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 1663 if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff; 1664 phase->register_control(queue_full_iff, loop, ctrl); 1665 Node* not_full = new IfTrueNode(queue_full_iff); 1666 phase->register_control(not_full, loop, queue_full_iff); 1667 Node* full = new IfFalseNode(queue_full_iff); 1668 phase->register_control(full, loop, queue_full_iff); 1669 1670 ctrl = not_full; 1671 1672 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t))); 1673 phase->register_new_node(next_index, ctrl); 1674 1675 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered); 1676 phase->register_new_node(buffer, ctrl); 1677 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index); 1678 phase->register_new_node(log_addr, ctrl); 1679 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered); 1680 phase->register_new_node(log_store, ctrl); 1681 // update the index 1682 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered); 1683 phase->register_new_node(index_update, ctrl); 1684 1685 // Fast-path case 1686 region2->init_req(_fast_path, ctrl); 1687 phi2->init_req(_fast_path, index_update); 1688 1689 ctrl = full; 1690 1691 Node* base = find_bottom_mem(ctrl, phase); 1692 1693 MergeMemNode* mm = MergeMemNode::make(base); 1694 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 1695 phase->register_new_node(mm, ctrl); 1696 1697 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); 1698 call->init_req(TypeFunc::Control, ctrl); 1699 call->init_req(TypeFunc::I_O, phase->C->top()); 1700 call->init_req(TypeFunc::Memory, mm); 1701 call->init_req(TypeFunc::FramePtr, phase->C->top()); 1702 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 1703 call->init_req(TypeFunc::Parms, pre_val); 1704 call->init_req(TypeFunc::Parms+1, thread); 1705 phase->register_control(call, loop, ctrl); 1706 1707 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control); 1708 phase->register_control(ctrl_proj, loop, call); 1709 Node* mem_proj = new ProjNode(call, TypeFunc::Memory); 1710 phase->register_new_node(mem_proj, call); 1711 1712 // Slow-path case 1713 region2->init_req(_slow_path, ctrl_proj); 1714 phi2->init_req(_slow_path, mem_proj); 1715 1716 phase->register_control(region2, loop, reg2_ctrl); 1717 phase->register_new_node(phi2, region2); 1718 1719 region->init_req(_heap_unstable, region2); 1720 phi->init_req(_heap_unstable, phi2); 1721 1722 phase->register_control(region, loop, heap_stable_ctrl->in(0)); 1723 phase->register_new_node(phi, region); 1724 1725 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase); 1726 for(uint next = 0; next < uses.size(); next++ ) { 1727 Node *n = uses.at(next); 1728 assert(phase->get_ctrl(n) == init_ctrl, "bad control"); 1729 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 1730 phase->set_ctrl(n, region); 1731 follow_barrier_uses(n, init_ctrl, uses, phase); 1732 } 1733 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses); 1734 1735 phase->igvn().replace_node(barrier, pre_val); 1736 } 1737 assert(state->enqueue_barriers_count() == 0, "all enqueue barrier nodes should have been replaced"); 1738 1739 } 1740 1741 void ShenandoahBarrierC2Support::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) { 1742 IdealLoopTree *loop = phase->get_loop(iff); 1743 Node* loop_head = loop->_head; 1744 Node* entry_c = loop_head->in(LoopNode::EntryControl); 1745 1746 Node* bol = iff->in(1); 1747 Node* cmp = bol->in(1); 1748 Node* andi = cmp->in(1); 1749 Node* load = andi->in(1); 1750 1751 assert(is_gc_state_load(load), "broken"); 1752 if (!phase->is_dominator(load->in(0), entry_c)) { 1753 Node* mem_ctrl = NULL; 1754 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase); 1755 load = load->clone(); 1756 load->set_req(MemNode::Memory, mem); 1757 load->set_req(0, entry_c); 1758 phase->register_new_node(load, entry_c); 1759 andi = andi->clone(); 1760 andi->set_req(1, load); 1761 phase->register_new_node(andi, entry_c); 1762 cmp = cmp->clone(); 1763 cmp->set_req(1, andi); 1764 phase->register_new_node(cmp, entry_c); 1765 bol = bol->clone(); 1766 bol->set_req(1, cmp); 1767 phase->register_new_node(bol, entry_c); 1768 1769 Node* old_bol =iff->in(1); 1770 phase->igvn().replace_input_of(iff, 1, bol); 1771 } 1772 } 1773 1774 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) { 1775 if (!n->is_If() || n->is_CountedLoopEnd()) { 1776 return false; 1777 } 1778 Node* region = n->in(0); 1779 1780 if (!region->is_Region()) { 1781 return false; 1782 } 1783 Node* dom = phase->idom(region); 1784 if (!dom->is_If()) { 1785 return false; 1786 } 1787 1788 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) { 1789 return false; 1790 } 1791 1792 IfNode* dom_if = dom->as_If(); 1793 Node* proj_true = dom_if->proj_out(1); 1794 Node* proj_false = dom_if->proj_out(0); 1795 1796 for (uint i = 1; i < region->req(); i++) { 1797 if (phase->is_dominator(proj_true, region->in(i))) { 1798 continue; 1799 } 1800 if (phase->is_dominator(proj_false, region->in(i))) { 1801 continue; 1802 } 1803 return false; 1804 } 1805 1806 return true; 1807 } 1808 1809 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) { 1810 assert(is_heap_stable_test(n), "no other tests"); 1811 if (identical_backtoback_ifs(n, phase)) { 1812 Node* n_ctrl = n->in(0); 1813 if (phase->can_split_if(n_ctrl)) { 1814 IfNode* dom_if = phase->idom(n_ctrl)->as_If(); 1815 if (is_heap_stable_test(n)) { 1816 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1); 1817 assert(is_gc_state_load(gc_state_load), "broken"); 1818 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1); 1819 assert(is_gc_state_load(dom_gc_state_load), "broken"); 1820 if (gc_state_load != dom_gc_state_load) { 1821 phase->igvn().replace_node(gc_state_load, dom_gc_state_load); 1822 } 1823 } 1824 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1825 Node* proj_true = dom_if->proj_out(1); 1826 Node* proj_false = dom_if->proj_out(0); 1827 Node* con_true = phase->igvn().makecon(TypeInt::ONE); 1828 Node* con_false = phase->igvn().makecon(TypeInt::ZERO); 1829 1830 for (uint i = 1; i < n_ctrl->req(); i++) { 1831 if (phase->is_dominator(proj_true, n_ctrl->in(i))) { 1832 bolphi->init_req(i, con_true); 1833 } else { 1834 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1835 bolphi->init_req(i, con_false); 1836 } 1837 } 1838 phase->register_new_node(bolphi, n_ctrl); 1839 phase->igvn().replace_input_of(n, 1, bolphi); 1840 phase->do_split_if(n); 1841 } 1842 } 1843 } 1844 1845 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) { 1846 // Find first invariant test that doesn't exit the loop 1847 LoopNode *head = loop->_head->as_Loop(); 1848 IfNode* unswitch_iff = NULL; 1849 Node* n = head->in(LoopNode::LoopBackControl); 1850 int loop_has_sfpts = -1; 1851 while (n != head) { 1852 Node* n_dom = phase->idom(n); 1853 if (n->is_Region()) { 1854 if (n_dom->is_If()) { 1855 IfNode* iff = n_dom->as_If(); 1856 if (iff->in(1)->is_Bool()) { 1857 BoolNode* bol = iff->in(1)->as_Bool(); 1858 if (bol->in(1)->is_Cmp()) { 1859 // If condition is invariant and not a loop exit, 1860 // then found reason to unswitch. 1861 if (is_heap_stable_test(iff) && 1862 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) { 1863 assert(!loop->is_loop_exit(iff), "both branches should be in the loop"); 1864 if (loop_has_sfpts == -1) { 1865 for(uint i = 0; i < loop->_body.size(); i++) { 1866 Node *m = loop->_body[i]; 1867 if (m->is_SafePoint() && !m->is_CallLeaf()) { 1868 loop_has_sfpts = 1; 1869 break; 1870 } 1871 } 1872 if (loop_has_sfpts == -1) { 1873 loop_has_sfpts = 0; 1874 } 1875 } 1876 if (!loop_has_sfpts) { 1877 unswitch_iff = iff; 1878 } 1879 } 1880 } 1881 } 1882 } 1883 } 1884 n = n_dom; 1885 } 1886 return unswitch_iff; 1887 } 1888 1889 1890 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) { 1891 Node_List heap_stable_tests; 1892 Node_List gc_state_loads; 1893 stack.push(phase->C->start(), 0); 1894 do { 1895 Node* n = stack.node(); 1896 uint i = stack.index(); 1897 1898 if (i < n->outcnt()) { 1899 Node* u = n->raw_out(i); 1900 stack.set_index(i+1); 1901 if (!visited.test_set(u->_idx)) { 1902 stack.push(u, 0); 1903 } 1904 } else { 1905 stack.pop(); 1906 if (ShenandoahCommonGCStateLoads && is_gc_state_load(n)) { 1907 gc_state_loads.push(n); 1908 } 1909 if (n->is_If() && is_heap_stable_test(n)) { 1910 heap_stable_tests.push(n); 1911 } 1912 } 1913 } while (stack.size() > 0); 1914 1915 bool progress; 1916 do { 1917 progress = false; 1918 for (uint i = 0; i < gc_state_loads.size(); i++) { 1919 Node* n = gc_state_loads.at(i); 1920 if (n->outcnt() != 0) { 1921 progress |= try_common_gc_state_load(n, phase); 1922 } 1923 } 1924 } while (progress); 1925 1926 for (uint i = 0; i < heap_stable_tests.size(); i++) { 1927 Node* n = heap_stable_tests.at(i); 1928 assert(is_heap_stable_test(n), "only evacuation test"); 1929 merge_back_to_back_tests(n, phase); 1930 } 1931 1932 if (!phase->C->major_progress()) { 1933 VectorSet seen(Thread::current()->resource_area()); 1934 for (uint i = 0; i < heap_stable_tests.size(); i++) { 1935 Node* n = heap_stable_tests.at(i); 1936 IdealLoopTree* loop = phase->get_loop(n); 1937 if (loop != phase->ltree_root() && 1938 loop->_child == NULL && 1939 !loop->_irreducible) { 1940 LoopNode* head = loop->_head->as_Loop(); 1941 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) && 1942 !seen.test_set(head->_idx)) { 1943 IfNode* iff = find_unswitching_candidate(loop, phase); 1944 if (iff != NULL) { 1945 Node* bol = iff->in(1); 1946 if (head->is_strip_mined()) { 1947 head->verify_strip_mined(0); 1948 } 1949 move_heap_stable_test_out_of_loop(iff, phase); 1950 1951 AutoNodeBudget node_budget(phase); 1952 1953 if (loop->policy_unswitching(phase)) { 1954 if (head->is_strip_mined()) { 1955 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop(); 1956 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase); 1957 } 1958 phase->do_unswitching(loop, old_new); 1959 } else { 1960 // Not proceeding with unswitching. Move load back in 1961 // the loop. 1962 phase->igvn().replace_input_of(iff, 1, bol); 1963 } 1964 } 1965 } 1966 } 1967 } 1968 } 1969 } 1970 1971 #ifdef ASSERT 1972 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) { 1973 const bool trace = false; 1974 ResourceMark rm; 1975 Unique_Node_List nodes; 1976 Unique_Node_List controls; 1977 Unique_Node_List memories; 1978 1979 nodes.push(root); 1980 for (uint next = 0; next < nodes.size(); next++) { 1981 Node *n = nodes.at(next); 1982 if (ShenandoahBarrierSetC2::is_shenandoah_wb_call(n)) { 1983 controls.push(n); 1984 if (trace) { tty->print("XXXXXX verifying"); n->dump(); } 1985 for (uint next2 = 0; next2 < controls.size(); next2++) { 1986 Node *m = controls.at(next2); 1987 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 1988 Node* u = m->fast_out(i); 1989 if (u->is_CFG() && !u->is_Root() && 1990 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) && 1991 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) { 1992 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); } 1993 controls.push(u); 1994 } 1995 } 1996 } 1997 memories.push(n->as_Call()->proj_out(TypeFunc::Memory)); 1998 for (uint next2 = 0; next2 < memories.size(); next2++) { 1999 Node *m = memories.at(next2); 2000 assert(m->bottom_type() == Type::MEMORY, ""); 2001 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2002 Node* u = m->fast_out(i); 2003 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) { 2004 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 2005 memories.push(u); 2006 } else if (u->is_LoadStore()) { 2007 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); } 2008 memories.push(u->find_out_with(Op_SCMemProj)); 2009 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) { 2010 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 2011 memories.push(u); 2012 } else if (u->is_Phi()) { 2013 assert(u->bottom_type() == Type::MEMORY, ""); 2014 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) { 2015 assert(controls.member(u->in(0)), ""); 2016 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 2017 memories.push(u); 2018 } 2019 } else if (u->is_SafePoint() || u->is_MemBar()) { 2020 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2021 Node* uu = u->fast_out(j); 2022 if (uu->bottom_type() == Type::MEMORY) { 2023 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); } 2024 memories.push(uu); 2025 } 2026 } 2027 } 2028 } 2029 } 2030 for (uint next2 = 0; next2 < controls.size(); next2++) { 2031 Node *m = controls.at(next2); 2032 if (m->is_Region()) { 2033 bool all_in = true; 2034 for (uint i = 1; i < m->req(); i++) { 2035 if (!controls.member(m->in(i))) { 2036 all_in = false; 2037 break; 2038 } 2039 } 2040 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); } 2041 bool found_phi = false; 2042 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) { 2043 Node* u = m->fast_out(j); 2044 if (u->is_Phi() && memories.member(u)) { 2045 found_phi = true; 2046 for (uint i = 1; i < u->req() && found_phi; i++) { 2047 Node* k = u->in(i); 2048 if (memories.member(k) != controls.member(m->in(i))) { 2049 found_phi = false; 2050 } 2051 } 2052 } 2053 } 2054 assert(found_phi || all_in, ""); 2055 } 2056 } 2057 controls.clear(); 2058 memories.clear(); 2059 } 2060 for( uint i = 0; i < n->len(); ++i ) { 2061 Node *m = n->in(i); 2062 if (m != NULL) { 2063 nodes.push(m); 2064 } 2065 } 2066 } 2067 } 2068 #endif 2069 2070 ShenandoahEnqueueBarrierNode::ShenandoahEnqueueBarrierNode(Node* val) : Node(NULL, val) { 2071 ShenandoahBarrierSetC2::bsc2()->state()->add_enqueue_barrier(this); 2072 } 2073 2074 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const { 2075 if (in(1) == NULL || in(1)->is_top()) { 2076 return Type::TOP; 2077 } 2078 const Type* t = in(1)->bottom_type(); 2079 if (t == TypePtr::NULL_PTR) { 2080 return t; 2081 } 2082 return t->is_oopptr()->cast_to_nonconst(); 2083 } 2084 2085 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const { 2086 if (in(1) == NULL) { 2087 return Type::TOP; 2088 } 2089 const Type* t = phase->type(in(1)); 2090 if (t == Type::TOP) { 2091 return Type::TOP; 2092 } 2093 if (t == TypePtr::NULL_PTR) { 2094 return t; 2095 } 2096 return t->is_oopptr()->cast_to_nonconst(); 2097 } 2098 2099 int ShenandoahEnqueueBarrierNode::needed(Node* n) { 2100 if (n == NULL || 2101 n->is_Allocate() || 2102 n->Opcode() == Op_ShenandoahEnqueueBarrier || 2103 n->bottom_type() == TypePtr::NULL_PTR || 2104 (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) { 2105 return NotNeeded; 2106 } 2107 if (n->is_Phi() || 2108 n->is_CMove()) { 2109 return MaybeNeeded; 2110 } 2111 return Needed; 2112 } 2113 2114 Node* ShenandoahEnqueueBarrierNode::next(Node* n) { 2115 for (;;) { 2116 if (n == NULL) { 2117 return n; 2118 } else if (n->bottom_type() == TypePtr::NULL_PTR) { 2119 return n; 2120 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) { 2121 return n; 2122 } else if (n->is_ConstraintCast() || 2123 n->Opcode() == Op_DecodeN || 2124 n->Opcode() == Op_EncodeP) { 2125 n = n->in(1); 2126 } else if (n->is_Proj()) { 2127 n = n->in(0); 2128 } else { 2129 return n; 2130 } 2131 } 2132 ShouldNotReachHere(); 2133 return NULL; 2134 } 2135 2136 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) { 2137 PhaseIterGVN* igvn = phase->is_IterGVN(); 2138 2139 Node* n = next(in(1)); 2140 2141 int cont = needed(n); 2142 2143 if (cont == NotNeeded) { 2144 return in(1); 2145 } else if (cont == MaybeNeeded) { 2146 if (igvn == NULL) { 2147 phase->record_for_igvn(this); 2148 return this; 2149 } else { 2150 ResourceMark rm; 2151 Unique_Node_List wq; 2152 uint wq_i = 0; 2153 2154 for (;;) { 2155 if (n->is_Phi()) { 2156 for (uint i = 1; i < n->req(); i++) { 2157 Node* m = n->in(i); 2158 if (m != NULL) { 2159 wq.push(m); 2160 } 2161 } 2162 } else { 2163 assert(n->is_CMove(), "nothing else here"); 2164 Node* m = n->in(CMoveNode::IfFalse); 2165 wq.push(m); 2166 m = n->in(CMoveNode::IfTrue); 2167 wq.push(m); 2168 } 2169 Node* orig_n = NULL; 2170 do { 2171 if (wq_i >= wq.size()) { 2172 return in(1); 2173 } 2174 n = wq.at(wq_i); 2175 wq_i++; 2176 orig_n = n; 2177 n = next(n); 2178 cont = needed(n); 2179 if (cont == Needed) { 2180 return this; 2181 } 2182 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n))); 2183 } 2184 } 2185 } 2186 2187 return this; 2188 } 2189 2190 #ifdef ASSERT 2191 static bool has_never_branch(Node* root) { 2192 for (uint i = 1; i < root->req(); i++) { 2193 Node* in = root->in(i); 2194 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) { 2195 return true; 2196 } 2197 } 2198 return false; 2199 } 2200 #endif 2201 2202 void MemoryGraphFixer::collect_memory_nodes() { 2203 Node_Stack stack(0); 2204 VectorSet visited(Thread::current()->resource_area()); 2205 Node_List regions; 2206 2207 // Walk the raw memory graph and create a mapping from CFG node to 2208 // memory node. Exclude phis for now. 2209 stack.push(_phase->C->root(), 1); 2210 do { 2211 Node* n = stack.node(); 2212 int opc = n->Opcode(); 2213 uint i = stack.index(); 2214 if (i < n->req()) { 2215 Node* mem = NULL; 2216 if (opc == Op_Root) { 2217 Node* in = n->in(i); 2218 int in_opc = in->Opcode(); 2219 if (in_opc == Op_Return || in_opc == Op_Rethrow) { 2220 mem = in->in(TypeFunc::Memory); 2221 } else if (in_opc == Op_Halt) { 2222 if (!in->in(0)->is_Region()) { 2223 Node* proj = in->in(0); 2224 assert(proj->is_Proj(), ""); 2225 Node* in = proj->in(0); 2226 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), ""); 2227 if (in->is_CallStaticJava()) { 2228 mem = in->in(TypeFunc::Memory); 2229 } else if (in->Opcode() == Op_Catch) { 2230 Node* call = in->in(0)->in(0); 2231 assert(call->is_Call(), ""); 2232 mem = call->in(TypeFunc::Memory); 2233 } else if (in->Opcode() == Op_NeverBranch) { 2234 ResourceMark rm; 2235 Unique_Node_List wq; 2236 wq.push(in); 2237 wq.push(in->as_Multi()->proj_out(0)); 2238 for (uint j = 1; j < wq.size(); j++) { 2239 Node* c = wq.at(j); 2240 assert(!c->is_Root(), "shouldn't leave loop"); 2241 if (c->is_SafePoint()) { 2242 assert(mem == NULL, "only one safepoint"); 2243 mem = c->in(TypeFunc::Memory); 2244 } 2245 for (DUIterator_Fast kmax, k = c->fast_outs(kmax); k < kmax; k++) { 2246 Node* u = c->fast_out(k); 2247 if (u->is_CFG()) { 2248 wq.push(u); 2249 } 2250 } 2251 } 2252 assert(mem != NULL, "should have found safepoint"); 2253 } 2254 } 2255 } else { 2256 #ifdef ASSERT 2257 n->dump(); 2258 in->dump(); 2259 #endif 2260 ShouldNotReachHere(); 2261 } 2262 } else { 2263 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, ""); 2264 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, ""); 2265 mem = n->in(i); 2266 } 2267 i++; 2268 stack.set_index(i); 2269 if (mem == NULL) { 2270 continue; 2271 } 2272 for (;;) { 2273 if (visited.test_set(mem->_idx) || mem->is_Start()) { 2274 break; 2275 } 2276 if (mem->is_Phi()) { 2277 stack.push(mem, 2); 2278 mem = mem->in(1); 2279 } else if (mem->is_Proj()) { 2280 stack.push(mem, mem->req()); 2281 mem = mem->in(0); 2282 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 2283 mem = mem->in(TypeFunc::Memory); 2284 } else if (mem->is_MergeMem()) { 2285 MergeMemNode* mm = mem->as_MergeMem(); 2286 mem = mm->memory_at(_alias); 2287 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 2288 assert(_alias == Compile::AliasIdxRaw, ""); 2289 stack.push(mem, mem->req()); 2290 mem = mem->in(MemNode::Memory); 2291 } else { 2292 #ifdef ASSERT 2293 mem->dump(); 2294 #endif 2295 ShouldNotReachHere(); 2296 } 2297 } 2298 } else { 2299 if (n->is_Phi()) { 2300 // Nothing 2301 } else if (!n->is_Root()) { 2302 Node* c = get_ctrl(n); 2303 _memory_nodes.map(c->_idx, n); 2304 } 2305 stack.pop(); 2306 } 2307 } while(stack.is_nonempty()); 2308 2309 // Iterate over CFG nodes in rpo and propagate memory state to 2310 // compute memory state at regions, creating new phis if needed. 2311 Node_List rpo_list; 2312 visited.Clear(); 2313 _phase->rpo(_phase->C->root(), stack, visited, rpo_list); 2314 Node* root = rpo_list.pop(); 2315 assert(root == _phase->C->root(), ""); 2316 2317 const bool trace = false; 2318 #ifdef ASSERT 2319 if (trace) { 2320 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2321 Node* c = rpo_list.at(i); 2322 if (_memory_nodes[c->_idx] != NULL) { 2323 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); 2324 } 2325 } 2326 } 2327 #endif 2328 uint last = _phase->C->unique(); 2329 2330 #ifdef ASSERT 2331 uint8_t max_depth = 0; 2332 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) { 2333 IdealLoopTree* lpt = iter.current(); 2334 max_depth = MAX2(max_depth, lpt->_nest); 2335 } 2336 #endif 2337 2338 bool progress = true; 2339 int iteration = 0; 2340 Node_List dead_phis; 2341 while (progress) { 2342 progress = false; 2343 iteration++; 2344 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), ""); 2345 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); } 2346 IdealLoopTree* last_updated_ilt = NULL; 2347 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2348 Node* c = rpo_list.at(i); 2349 2350 Node* prev_mem = _memory_nodes[c->_idx]; 2351 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2352 Node* prev_region = regions[c->_idx]; 2353 Node* unique = NULL; 2354 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) { 2355 Node* m = _memory_nodes[c->in(j)->_idx]; 2356 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"); 2357 if (m != NULL) { 2358 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) { 2359 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), ""); 2360 // continue 2361 } else if (unique == NULL) { 2362 unique = m; 2363 } else if (m == unique) { 2364 // continue 2365 } else { 2366 unique = NodeSentinel; 2367 } 2368 } 2369 } 2370 assert(unique != NULL, "empty phi???"); 2371 if (unique != NodeSentinel) { 2372 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) { 2373 dead_phis.push(prev_region); 2374 } 2375 regions.map(c->_idx, unique); 2376 } else { 2377 Node* phi = NULL; 2378 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) { 2379 phi = prev_region; 2380 for (uint k = 1; k < c->req(); k++) { 2381 Node* m = _memory_nodes[c->in(k)->_idx]; 2382 assert(m != NULL, "expect memory state"); 2383 phi->set_req(k, m); 2384 } 2385 } else { 2386 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) { 2387 Node* u = c->fast_out(j); 2388 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 2389 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) { 2390 phi = u; 2391 for (uint k = 1; k < c->req() && phi != NULL; k++) { 2392 Node* m = _memory_nodes[c->in(k)->_idx]; 2393 assert(m != NULL, "expect memory state"); 2394 if (u->in(k) != m) { 2395 phi = NULL; 2396 } 2397 } 2398 } 2399 } 2400 if (phi == NULL) { 2401 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias)); 2402 for (uint k = 1; k < c->req(); k++) { 2403 Node* m = _memory_nodes[c->in(k)->_idx]; 2404 assert(m != NULL, "expect memory state"); 2405 phi->init_req(k, m); 2406 } 2407 } 2408 } 2409 assert(phi != NULL, ""); 2410 regions.map(c->_idx, phi); 2411 } 2412 Node* current_region = regions[c->_idx]; 2413 if (current_region != prev_region) { 2414 progress = true; 2415 if (prev_region == prev_mem) { 2416 _memory_nodes.map(c->_idx, current_region); 2417 } 2418 } 2419 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) { 2420 Node* m = _memory_nodes[_phase->idom(c)->_idx]; 2421 assert(m != NULL, "expect memory state"); 2422 if (m != prev_mem) { 2423 _memory_nodes.map(c->_idx, m); 2424 progress = true; 2425 } 2426 } 2427 #ifdef ASSERT 2428 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); } 2429 #endif 2430 } 2431 } 2432 2433 // Replace existing phi with computed memory state for that region 2434 // if different (could be a new phi or a dominating memory node if 2435 // that phi was found to be useless). 2436 while (dead_phis.size() > 0) { 2437 Node* n = dead_phis.pop(); 2438 n->replace_by(_phase->C->top()); 2439 n->destruct(); 2440 } 2441 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2442 Node* c = rpo_list.at(i); 2443 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2444 Node* n = regions[c->_idx]; 2445 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) { 2446 _phase->register_new_node(n, c); 2447 } 2448 } 2449 } 2450 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2451 Node* c = rpo_list.at(i); 2452 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2453 Node* n = regions[c->_idx]; 2454 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 2455 Node* u = c->fast_out(i); 2456 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 2457 u != n) { 2458 if (u->adr_type() == TypePtr::BOTTOM) { 2459 fix_memory_uses(u, n, n, c); 2460 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 2461 _phase->lazy_replace(u, n); 2462 --i; --imax; 2463 } 2464 } 2465 } 2466 } 2467 } 2468 } 2469 2470 Node* MemoryGraphFixer::get_ctrl(Node* n) const { 2471 Node* c = _phase->get_ctrl(n); 2472 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) { 2473 assert(c == n->in(0), ""); 2474 CallNode* call = c->as_Call(); 2475 CallProjections projs; 2476 call->extract_projections(&projs, true, false); 2477 if (projs.catchall_memproj != NULL) { 2478 if (projs.fallthrough_memproj == n) { 2479 c = projs.fallthrough_catchproj; 2480 } else { 2481 assert(projs.catchall_memproj == n, ""); 2482 c = projs.catchall_catchproj; 2483 } 2484 } 2485 } 2486 return c; 2487 } 2488 2489 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const { 2490 if (_phase->has_ctrl(n)) 2491 return get_ctrl(n); 2492 else { 2493 assert (n->is_CFG(), "must be a CFG node"); 2494 return n; 2495 } 2496 } 2497 2498 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const { 2499 return m != NULL && get_ctrl(m) == c; 2500 } 2501 2502 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const { 2503 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, ""); 2504 Node* mem = _memory_nodes[ctrl->_idx]; 2505 Node* c = ctrl; 2506 while (!mem_is_valid(mem, c) && 2507 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) { 2508 c = _phase->idom(c); 2509 mem = _memory_nodes[c->_idx]; 2510 } 2511 if (n != NULL && mem_is_valid(mem, c)) { 2512 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) { 2513 mem = next_mem(mem, _alias); 2514 } 2515 if (mem->is_MergeMem()) { 2516 mem = mem->as_MergeMem()->memory_at(_alias); 2517 } 2518 if (!mem_is_valid(mem, c)) { 2519 do { 2520 c = _phase->idom(c); 2521 mem = _memory_nodes[c->_idx]; 2522 } while (!mem_is_valid(mem, c) && 2523 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))); 2524 } 2525 } 2526 assert(mem->bottom_type() == Type::MEMORY, ""); 2527 return mem; 2528 } 2529 2530 bool MemoryGraphFixer::has_mem_phi(Node* region) const { 2531 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2532 Node* use = region->fast_out(i); 2533 if (use->is_Phi() && use->bottom_type() == Type::MEMORY && 2534 (_phase->C->get_alias_index(use->adr_type()) == _alias)) { 2535 return true; 2536 } 2537 } 2538 return false; 2539 } 2540 2541 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) { 2542 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, ""); 2543 const bool trace = false; 2544 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); }); 2545 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); }); 2546 GrowableArray<Node*> phis; 2547 if (mem_for_ctrl != mem) { 2548 Node* old = mem_for_ctrl; 2549 Node* prev = NULL; 2550 while (old != mem) { 2551 prev = old; 2552 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) { 2553 assert(_alias == Compile::AliasIdxRaw, ""); 2554 old = old->in(MemNode::Memory); 2555 } else if (old->Opcode() == Op_SCMemProj) { 2556 assert(_alias == Compile::AliasIdxRaw, ""); 2557 old = old->in(0); 2558 } else { 2559 ShouldNotReachHere(); 2560 } 2561 } 2562 assert(prev != NULL, ""); 2563 if (new_ctrl != ctrl) { 2564 _memory_nodes.map(ctrl->_idx, mem); 2565 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl); 2566 } 2567 uint input = (uint)MemNode::Memory; 2568 _phase->igvn().replace_input_of(prev, input, new_mem); 2569 } else { 2570 uses.clear(); 2571 _memory_nodes.map(new_ctrl->_idx, new_mem); 2572 uses.push(new_ctrl); 2573 for(uint next = 0; next < uses.size(); next++ ) { 2574 Node *n = uses.at(next); 2575 assert(n->is_CFG(), ""); 2576 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); }); 2577 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2578 Node* u = n->fast_out(i); 2579 if (!u->is_Root() && u->is_CFG() && u != n) { 2580 Node* m = _memory_nodes[u->_idx]; 2581 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) && 2582 !has_mem_phi(u) && 2583 u->unique_ctrl_out()->Opcode() != Op_Halt) { 2584 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); }); 2585 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); }); 2586 2587 if (!mem_is_valid(m, u) || !m->is_Phi()) { 2588 bool push = true; 2589 bool create_phi = true; 2590 if (_phase->is_dominator(new_ctrl, u)) { 2591 create_phi = false; 2592 } else if (!_phase->C->has_irreducible_loop()) { 2593 IdealLoopTree* loop = _phase->get_loop(ctrl); 2594 bool do_check = true; 2595 IdealLoopTree* l = loop; 2596 create_phi = false; 2597 while (l != _phase->ltree_root()) { 2598 if (_phase->is_dominator(l->_head, u) && _phase->is_dominator(_phase->idom(u), l->_head)) { 2599 create_phi = true; 2600 do_check = false; 2601 break; 2602 } 2603 l = l->_parent; 2604 } 2605 2606 if (do_check) { 2607 assert(!create_phi, ""); 2608 IdealLoopTree* u_loop = _phase->get_loop(u); 2609 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) { 2610 Node* c = ctrl; 2611 while (!_phase->is_dominator(c, u_loop->tail())) { 2612 c = _phase->idom(c); 2613 } 2614 if (!_phase->is_dominator(c, u)) { 2615 do_check = false; 2616 } 2617 } 2618 } 2619 2620 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) { 2621 create_phi = true; 2622 } 2623 } 2624 if (create_phi) { 2625 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias)); 2626 _phase->register_new_node(phi, u); 2627 phis.push(phi); 2628 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); }); 2629 if (!mem_is_valid(m, u)) { 2630 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); }); 2631 _memory_nodes.map(u->_idx, phi); 2632 } else { 2633 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); }); 2634 for (;;) { 2635 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), ""); 2636 Node* next = NULL; 2637 if (m->is_Proj()) { 2638 next = m->in(0); 2639 } else { 2640 assert(m->is_Mem() || m->is_LoadStore(), ""); 2641 assert(_alias == Compile::AliasIdxRaw, ""); 2642 next = m->in(MemNode::Memory); 2643 } 2644 if (_phase->get_ctrl(next) != u) { 2645 break; 2646 } 2647 if (next->is_MergeMem()) { 2648 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, ""); 2649 break; 2650 } 2651 if (next->is_Phi()) { 2652 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, ""); 2653 break; 2654 } 2655 m = next; 2656 } 2657 2658 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); }); 2659 assert(m->is_Mem() || m->is_LoadStore(), ""); 2660 uint input = (uint)MemNode::Memory; 2661 _phase->igvn().replace_input_of(m, input, phi); 2662 push = false; 2663 } 2664 } else { 2665 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); }); 2666 } 2667 if (push) { 2668 uses.push(u); 2669 } 2670 } 2671 } else if (!mem_is_valid(m, u) && 2672 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) { 2673 uses.push(u); 2674 } 2675 } 2676 } 2677 } 2678 for (int i = 0; i < phis.length(); i++) { 2679 Node* n = phis.at(i); 2680 Node* r = n->in(0); 2681 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); }); 2682 for (uint j = 1; j < n->req(); j++) { 2683 Node* m = find_mem(r->in(j), NULL); 2684 _phase->igvn().replace_input_of(n, j, m); 2685 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); }); 2686 } 2687 } 2688 } 2689 uint last = _phase->C->unique(); 2690 MergeMemNode* mm = NULL; 2691 int alias = _alias; 2692 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); }); 2693 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 2694 Node* u = mem->out(i); 2695 if (u->_idx < last) { 2696 if (u->is_Mem()) { 2697 if (_phase->C->get_alias_index(u->adr_type()) == alias) { 2698 Node* m = find_mem(_phase->get_ctrl(u), u); 2699 if (m != mem) { 2700 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2701 _phase->igvn().replace_input_of(u, MemNode::Memory, m); 2702 --i; 2703 } 2704 } 2705 } else if (u->is_MergeMem()) { 2706 MergeMemNode* u_mm = u->as_MergeMem(); 2707 if (u_mm->memory_at(alias) == mem) { 2708 MergeMemNode* newmm = NULL; 2709 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2710 Node* uu = u->fast_out(j); 2711 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 2712 if (uu->is_Phi()) { 2713 assert(uu->adr_type() == TypePtr::BOTTOM, ""); 2714 Node* region = uu->in(0); 2715 int nb = 0; 2716 for (uint k = 1; k < uu->req(); k++) { 2717 if (uu->in(k) == u) { 2718 Node* m = find_mem(region->in(k), NULL); 2719 if (m != mem) { 2720 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); }); 2721 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 2722 if (newmm != u) { 2723 _phase->igvn().replace_input_of(uu, k, newmm); 2724 nb++; 2725 --jmax; 2726 } 2727 } 2728 } 2729 } 2730 if (nb > 0) { 2731 --j; 2732 } 2733 } else { 2734 Node* m = find_mem(_phase->ctrl_or_self(uu), uu); 2735 if (m != mem) { 2736 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); }); 2737 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 2738 if (newmm != u) { 2739 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 2740 --j, --jmax; 2741 } 2742 } 2743 } 2744 } 2745 } 2746 } else if (u->is_Phi()) { 2747 assert(u->bottom_type() == Type::MEMORY, "what else?"); 2748 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) { 2749 Node* region = u->in(0); 2750 bool replaced = false; 2751 for (uint j = 1; j < u->req(); j++) { 2752 if (u->in(j) == mem) { 2753 Node* m = find_mem(region->in(j), NULL); 2754 Node* nnew = m; 2755 if (m != mem) { 2756 if (u->adr_type() == TypePtr::BOTTOM) { 2757 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m)); 2758 nnew = mm; 2759 } 2760 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); }); 2761 _phase->igvn().replace_input_of(u, j, nnew); 2762 replaced = true; 2763 } 2764 } 2765 } 2766 if (replaced) { 2767 --i; 2768 } 2769 } 2770 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 2771 u->adr_type() == NULL) { 2772 assert(u->adr_type() != NULL || 2773 u->Opcode() == Op_Rethrow || 2774 u->Opcode() == Op_Return || 2775 u->Opcode() == Op_SafePoint || 2776 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 2777 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 2778 u->Opcode() == Op_CallLeaf, ""); 2779 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2780 if (m != mem) { 2781 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m)); 2782 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 2783 --i; 2784 } 2785 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) { 2786 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2787 if (m != mem) { 2788 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2789 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 2790 --i; 2791 } 2792 } else if (u->adr_type() != TypePtr::BOTTOM && 2793 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) { 2794 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2795 assert(m != mem, ""); 2796 // u is on the wrong slice... 2797 assert(u->is_ClearArray(), ""); 2798 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2799 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 2800 --i; 2801 } 2802 } 2803 } 2804 #ifdef ASSERT 2805 assert(new_mem->outcnt() > 0, ""); 2806 for (int i = 0; i < phis.length(); i++) { 2807 Node* n = phis.at(i); 2808 assert(n->outcnt() > 0, "new phi must have uses now"); 2809 } 2810 #endif 2811 } 2812 2813 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const { 2814 MergeMemNode* mm = MergeMemNode::make(mem); 2815 mm->set_memory_at(_alias, rep_proj); 2816 _phase->register_new_node(mm, rep_ctrl); 2817 return mm; 2818 } 2819 2820 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const { 2821 MergeMemNode* newmm = NULL; 2822 MergeMemNode* u_mm = u->as_MergeMem(); 2823 Node* c = _phase->get_ctrl(u); 2824 if (_phase->is_dominator(c, rep_ctrl)) { 2825 c = rep_ctrl; 2826 } else { 2827 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other"); 2828 } 2829 if (u->outcnt() == 1) { 2830 if (u->req() > (uint)_alias && u->in(_alias) == mem) { 2831 _phase->igvn().replace_input_of(u, _alias, rep_proj); 2832 --i; 2833 } else { 2834 _phase->igvn().rehash_node_delayed(u); 2835 u_mm->set_memory_at(_alias, rep_proj); 2836 } 2837 newmm = u_mm; 2838 _phase->set_ctrl_and_loop(u, c); 2839 } else { 2840 // can't simply clone u and then change one of its input because 2841 // it adds and then removes an edge which messes with the 2842 // DUIterator 2843 newmm = MergeMemNode::make(u_mm->base_memory()); 2844 for (uint j = 0; j < u->req(); j++) { 2845 if (j < newmm->req()) { 2846 if (j == (uint)_alias) { 2847 newmm->set_req(j, rep_proj); 2848 } else if (newmm->in(j) != u->in(j)) { 2849 newmm->set_req(j, u->in(j)); 2850 } 2851 } else if (j == (uint)_alias) { 2852 newmm->add_req(rep_proj); 2853 } else { 2854 newmm->add_req(u->in(j)); 2855 } 2856 } 2857 if ((uint)_alias >= u->req()) { 2858 newmm->set_memory_at(_alias, rep_proj); 2859 } 2860 _phase->register_new_node(newmm, c); 2861 } 2862 return newmm; 2863 } 2864 2865 bool MemoryGraphFixer::should_process_phi(Node* phi) const { 2866 if (phi->adr_type() == TypePtr::BOTTOM) { 2867 Node* region = phi->in(0); 2868 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) { 2869 Node* uu = region->fast_out(j); 2870 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) { 2871 return false; 2872 } 2873 } 2874 return true; 2875 } 2876 return _phase->C->get_alias_index(phi->adr_type()) == _alias; 2877 } 2878 2879 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const { 2880 uint last = _phase-> C->unique(); 2881 MergeMemNode* mm = NULL; 2882 assert(mem->bottom_type() == Type::MEMORY, ""); 2883 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 2884 Node* u = mem->out(i); 2885 if (u != replacement && u->_idx < last) { 2886 if (u->is_MergeMem()) { 2887 MergeMemNode* u_mm = u->as_MergeMem(); 2888 if (u_mm->memory_at(_alias) == mem) { 2889 MergeMemNode* newmm = NULL; 2890 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2891 Node* uu = u->fast_out(j); 2892 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 2893 if (uu->is_Phi()) { 2894 if (should_process_phi(uu)) { 2895 Node* region = uu->in(0); 2896 int nb = 0; 2897 for (uint k = 1; k < uu->req(); k++) { 2898 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) { 2899 if (newmm == NULL) { 2900 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 2901 } 2902 if (newmm != u) { 2903 _phase->igvn().replace_input_of(uu, k, newmm); 2904 nb++; 2905 --jmax; 2906 } 2907 } 2908 } 2909 if (nb > 0) { 2910 --j; 2911 } 2912 } 2913 } else { 2914 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) { 2915 if (newmm == NULL) { 2916 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 2917 } 2918 if (newmm != u) { 2919 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 2920 --j, --jmax; 2921 } 2922 } 2923 } 2924 } 2925 } 2926 } else if (u->is_Phi()) { 2927 assert(u->bottom_type() == Type::MEMORY, "what else?"); 2928 Node* region = u->in(0); 2929 if (should_process_phi(u)) { 2930 bool replaced = false; 2931 for (uint j = 1; j < u->req(); j++) { 2932 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) { 2933 Node* nnew = rep_proj; 2934 if (u->adr_type() == TypePtr::BOTTOM) { 2935 if (mm == NULL) { 2936 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 2937 } 2938 nnew = mm; 2939 } 2940 _phase->igvn().replace_input_of(u, j, nnew); 2941 replaced = true; 2942 } 2943 } 2944 if (replaced) { 2945 --i; 2946 } 2947 2948 } 2949 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 2950 u->adr_type() == NULL) { 2951 assert(u->adr_type() != NULL || 2952 u->Opcode() == Op_Rethrow || 2953 u->Opcode() == Op_Return || 2954 u->Opcode() == Op_SafePoint || 2955 u->Opcode() == Op_StoreLConditional || 2956 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 2957 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 2958 u->Opcode() == Op_CallLeaf, ""); 2959 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 2960 if (mm == NULL) { 2961 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 2962 } 2963 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 2964 --i; 2965 } 2966 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 2967 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 2968 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj); 2969 --i; 2970 } 2971 } 2972 } 2973 } 2974 } 2975 2976 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj) 2977 : Node(ctrl, obj) { 2978 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this); 2979 } 2980 2981 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const { 2982 if (in(ValueIn) == NULL || in(ValueIn)->is_top()) { 2983 return Type::TOP; 2984 } 2985 const Type* t = in(ValueIn)->bottom_type(); 2986 if (t == TypePtr::NULL_PTR) { 2987 return t; 2988 } 2989 return t->is_oopptr(); 2990 } 2991 2992 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const { 2993 // Either input is TOP ==> the result is TOP 2994 const Type *t2 = phase->type(in(ValueIn)); 2995 if( t2 == Type::TOP ) return Type::TOP; 2996 2997 if (t2 == TypePtr::NULL_PTR) { 2998 return t2; 2999 } 3000 3001 const Type* type = t2->is_oopptr()/*->cast_to_nonconst()*/; 3002 return type; 3003 } 3004 3005 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) { 3006 Node* value = in(ValueIn); 3007 if (!needs_barrier(phase, value)) { 3008 return value; 3009 } 3010 return this; 3011 } 3012 3013 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) { 3014 Unique_Node_List visited; 3015 return needs_barrier_impl(phase, n, visited); 3016 } 3017 3018 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) { 3019 if (n == NULL) return false; 3020 if (visited.member(n)) { 3021 return false; // Been there. 3022 } 3023 visited.push(n); 3024 3025 if (n->is_Allocate()) { 3026 // tty->print_cr("optimize barrier on alloc"); 3027 return false; 3028 } 3029 if (n->is_Call()) { 3030 // tty->print_cr("optimize barrier on call"); 3031 return false; 3032 } 3033 3034 const Type* type = phase->type(n); 3035 if (type == Type::TOP) { 3036 return false; 3037 } 3038 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) { 3039 // tty->print_cr("optimize barrier on null"); 3040 return false; 3041 } 3042 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) { 3043 // tty->print_cr("optimize barrier on constant"); 3044 return false; 3045 } 3046 3047 switch (n->Opcode()) { 3048 case Op_AddP: 3049 return true; // TODO: Can refine? 3050 case Op_LoadP: 3051 case Op_ShenandoahCompareAndExchangeN: 3052 case Op_ShenandoahCompareAndExchangeP: 3053 case Op_CompareAndExchangeN: 3054 case Op_CompareAndExchangeP: 3055 case Op_GetAndSetN: 3056 case Op_GetAndSetP: 3057 return true; 3058 case Op_Phi: { 3059 for (uint i = 1; i < n->req(); i++) { 3060 if (needs_barrier_impl(phase, n->in(i), visited)) return true; 3061 } 3062 return false; 3063 } 3064 case Op_CheckCastPP: 3065 case Op_CastPP: 3066 return needs_barrier_impl(phase, n->in(1), visited); 3067 case Op_Proj: 3068 return needs_barrier_impl(phase, n->in(0), visited); 3069 case Op_ShenandoahLoadReferenceBarrier: 3070 // tty->print_cr("optimize barrier on barrier"); 3071 return false; 3072 case Op_Parm: 3073 // tty->print_cr("optimize barrier on input arg"); 3074 return false; 3075 case Op_DecodeN: 3076 case Op_EncodeP: 3077 return needs_barrier_impl(phase, n->in(1), visited); 3078 case Op_LoadN: 3079 return true; 3080 case Op_CMoveP: 3081 return needs_barrier_impl(phase, n->in(2), visited) || 3082 needs_barrier_impl(phase, n->in(3), visited); 3083 case Op_ShenandoahEnqueueBarrier: 3084 return needs_barrier_impl(phase, n->in(1), visited); 3085 default: 3086 break; 3087 } 3088 #ifdef ASSERT 3089 tty->print("need barrier on?: "); 3090 tty->print_cr("ins:"); 3091 n->dump(2); 3092 tty->print_cr("outs:"); 3093 n->dump(-2); 3094 ShouldNotReachHere(); 3095 #endif 3096 return true; 3097 } 3098 3099 ShenandoahLoadReferenceBarrierNode::Strength ShenandoahLoadReferenceBarrierNode::get_barrier_strength() { 3100 Unique_Node_List visited; 3101 Node_Stack stack(0); 3102 stack.push(this, 0); 3103 Strength strength = NONE; 3104 while (strength != STRONG && stack.size() > 0) { 3105 Node* n = stack.node(); 3106 if (visited.member(n)) { 3107 stack.pop(); 3108 continue; 3109 } 3110 visited.push(n); 3111 bool visit_users = false; 3112 switch (n->Opcode()) { 3113 case Op_StoreN: 3114 case Op_StoreP: { 3115 strength = STRONG; 3116 break; 3117 } 3118 case Op_CmpP: { 3119 if (!n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) && 3120 !n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) { 3121 strength = STRONG; 3122 } 3123 break; 3124 } 3125 case Op_CallStaticJava: { 3126 // If it's an deopt-call we don't need barriers because 3127 // the LRB will be applied when unpacking the deopt frame. 3128 if (n->as_CallStaticJava()->uncommon_trap_request() == 0) { 3129 strength = STRONG; 3130 } 3131 break; 3132 } 3133 case Op_CallDynamicJava: 3134 case Op_CallLeaf: 3135 case Op_CallLeafNoFP: 3136 case Op_CompareAndSwapL: 3137 case Op_CompareAndSwapI: 3138 case Op_CompareAndSwapB: 3139 case Op_CompareAndSwapS: 3140 case Op_CompareAndSwapN: 3141 case Op_CompareAndSwapP: 3142 case Op_CompareAndExchangeL: 3143 case Op_CompareAndExchangeI: 3144 case Op_CompareAndExchangeB: 3145 case Op_CompareAndExchangeS: 3146 case Op_CompareAndExchangeN: 3147 case Op_CompareAndExchangeP: 3148 case Op_WeakCompareAndSwapL: 3149 case Op_WeakCompareAndSwapI: 3150 case Op_WeakCompareAndSwapB: 3151 case Op_WeakCompareAndSwapS: 3152 case Op_WeakCompareAndSwapN: 3153 case Op_WeakCompareAndSwapP: 3154 case Op_ShenandoahCompareAndSwapN: 3155 case Op_ShenandoahCompareAndSwapP: 3156 case Op_ShenandoahWeakCompareAndSwapN: 3157 case Op_ShenandoahWeakCompareAndSwapP: 3158 case Op_ShenandoahCompareAndExchangeN: 3159 case Op_ShenandoahCompareAndExchangeP: 3160 case Op_GetAndSetL: 3161 case Op_GetAndSetI: 3162 case Op_GetAndSetB: 3163 case Op_GetAndSetS: 3164 case Op_GetAndSetP: 3165 case Op_GetAndSetN: 3166 case Op_GetAndAddL: 3167 case Op_GetAndAddI: 3168 case Op_GetAndAddB: 3169 case Op_GetAndAddS: 3170 case Op_ShenandoahEnqueueBarrier: 3171 case Op_FastLock: 3172 case Op_FastUnlock: 3173 case Op_Rethrow: 3174 case Op_Return: 3175 case Op_StoreB: 3176 case Op_StoreC: 3177 case Op_StoreD: 3178 case Op_StoreF: 3179 case Op_StoreL: 3180 case Op_StoreLConditional: 3181 case Op_StoreI: 3182 case Op_StoreVector: 3183 case Op_StrInflatedCopy: 3184 case Op_StrCompressedCopy: 3185 case Op_EncodeP: 3186 case Op_CastP2X: 3187 case Op_SafePoint: 3188 case Op_EncodeISOArray: 3189 strength = STRONG; 3190 break; 3191 case Op_LoadB: 3192 case Op_LoadUB: 3193 case Op_LoadUS: 3194 case Op_LoadD: 3195 case Op_LoadF: 3196 case Op_LoadL: 3197 case Op_LoadI: 3198 case Op_LoadS: 3199 case Op_LoadN: 3200 case Op_LoadP: 3201 case Op_LoadVector: { 3202 const TypePtr* adr_type = n->adr_type(); 3203 int alias_idx = Compile::current()->get_alias_index(adr_type); 3204 Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx); 3205 ciField* field = alias_type->field(); 3206 bool is_static = field != NULL && field->is_static(); 3207 bool is_final = field != NULL && field->is_final(); 3208 bool is_stable = field != NULL && field->is_stable(); 3209 if (ShenandoahOptimizeStaticFinals && is_static && is_final) { 3210 // Leave strength as is. 3211 } else if (ShenandoahOptimizeInstanceFinals && !is_static && is_final) { 3212 // Leave strength as is. 3213 } else if (ShenandoahOptimizeStableFinals && (is_stable || (adr_type->isa_aryptr() && adr_type->isa_aryptr()->is_stable()))) { 3214 // Leave strength as is. 3215 } else { 3216 strength = WEAK; 3217 } 3218 break; 3219 } 3220 case Op_AryEq: { 3221 Node* n1 = n->in(2); 3222 Node* n2 = n->in(3); 3223 if (!ShenandoahOptimizeStableFinals || 3224 !n1->bottom_type()->isa_aryptr() || !n1->bottom_type()->isa_aryptr()->is_stable() || 3225 !n2->bottom_type()->isa_aryptr() || !n2->bottom_type()->isa_aryptr()->is_stable()) { 3226 strength = WEAK; 3227 } 3228 break; 3229 } 3230 case Op_StrEquals: 3231 case Op_StrComp: 3232 case Op_StrIndexOf: 3233 case Op_StrIndexOfChar: 3234 if (!ShenandoahOptimizeStableFinals) { 3235 strength = WEAK; 3236 } 3237 break; 3238 case Op_Conv2B: 3239 case Op_LoadRange: 3240 case Op_LoadKlass: 3241 case Op_LoadNKlass: 3242 // NONE, i.e. leave current strength as is 3243 break; 3244 case Op_AddP: 3245 case Op_CheckCastPP: 3246 case Op_CastPP: 3247 case Op_CMoveP: 3248 case Op_Phi: 3249 case Op_ShenandoahLoadReferenceBarrier: 3250 visit_users = true; 3251 break; 3252 default: { 3253 #ifdef ASSERT 3254 tty->print_cr("Unknown node in get_barrier_strength:"); 3255 n->dump(1); 3256 ShouldNotReachHere(); 3257 #else 3258 strength = STRONG; 3259 #endif 3260 } 3261 } 3262 #ifdef ASSERT 3263 /* 3264 if (strength == STRONG) { 3265 tty->print("strengthening node: "); 3266 n->dump(); 3267 } 3268 */ 3269 #endif 3270 stack.pop(); 3271 if (visit_users) { 3272 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3273 Node* user = n->fast_out(i); 3274 if (user != NULL) { 3275 stack.push(user, 0); 3276 } 3277 } 3278 } 3279 } 3280 return strength; 3281 } 3282 3283 CallStaticJavaNode* ShenandoahLoadReferenceBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) { 3284 Node* val = in(ValueIn); 3285 3286 const Type* val_t = igvn.type(val); 3287 3288 if (val_t->meet(TypePtr::NULL_PTR) != val_t && 3289 val->Opcode() == Op_CastPP && 3290 val->in(0) != NULL && 3291 val->in(0)->Opcode() == Op_IfTrue && 3292 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 3293 val->in(0)->in(0)->is_If() && 3294 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool && 3295 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne && 3296 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 3297 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) && 3298 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 3299 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), ""); 3300 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 3301 return unc; 3302 } 3303 return NULL; 3304 }