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 int cnt = 0; 57 if (!C->optimize_loops(cnt, igvn, LoopOptsShenandoahPostExpand)) { 58 return false; 59 } 60 C->clear_major_progress(); 61 } 62 } 63 return true; 64 } 65 66 bool ShenandoahBarrierC2Support::is_heap_state_test(Node* iff, int mask) { 67 if (!UseShenandoahGC) { 68 return false; 69 } 70 assert(iff->is_If(), "bad input"); 71 if (iff->Opcode() != Op_If) { 72 return false; 73 } 74 Node* bol = iff->in(1); 75 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) { 76 return false; 77 } 78 Node* cmp = bol->in(1); 79 if (cmp->Opcode() != Op_CmpI) { 80 return false; 81 } 82 Node* in1 = cmp->in(1); 83 Node* in2 = cmp->in(2); 84 if (in2->find_int_con(-1) != 0) { 85 return false; 86 } 87 if (in1->Opcode() != Op_AndI) { 88 return false; 89 } 90 in2 = in1->in(2); 91 if (in2->find_int_con(-1) != mask) { 92 return false; 93 } 94 in1 = in1->in(1); 95 96 return is_gc_state_load(in1); 97 } 98 99 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) { 100 return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED); 101 } 102 103 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) { 104 if (!UseShenandoahGC) { 105 return false; 106 } 107 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) { 108 return false; 109 } 110 Node* addp = n->in(MemNode::Address); 111 if (!addp->is_AddP()) { 112 return false; 113 } 114 Node* base = addp->in(AddPNode::Address); 115 Node* off = addp->in(AddPNode::Offset); 116 if (base->Opcode() != Op_ThreadLocal) { 117 return false; 118 } 119 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) { 120 return false; 121 } 122 return true; 123 } 124 125 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) { 126 assert(phase->is_dominator(stop, start), "bad inputs"); 127 ResourceMark rm; 128 Unique_Node_List wq; 129 wq.push(start); 130 for (uint next = 0; next < wq.size(); next++) { 131 Node *m = wq.at(next); 132 if (m == stop) { 133 continue; 134 } 135 if (m->is_SafePoint() && !m->is_CallLeaf()) { 136 return true; 137 } 138 if (m->is_Region()) { 139 for (uint i = 1; i < m->req(); i++) { 140 wq.push(m->in(i)); 141 } 142 } else { 143 wq.push(m->in(0)); 144 } 145 } 146 return false; 147 } 148 149 bool ShenandoahBarrierC2Support::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) { 150 assert(is_gc_state_load(n), "inconsistent"); 151 Node* addp = n->in(MemNode::Address); 152 Node* dominator = NULL; 153 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) { 154 Node* u = addp->fast_out(i); 155 assert(is_gc_state_load(u), "inconsistent"); 156 if (u != n && phase->is_dominator(u->in(0), n->in(0))) { 157 if (dominator == NULL) { 158 dominator = u; 159 } else { 160 if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) { 161 dominator = u; 162 } 163 } 164 } 165 } 166 if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) { 167 return false; 168 } 169 phase->igvn().replace_node(n, dominator); 170 171 return true; 172 } 173 174 #ifdef ASSERT 175 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) { 176 assert(phis.size() == 0, ""); 177 178 while (true) { 179 if (in->bottom_type() == TypePtr::NULL_PTR) { 180 if (trace) {tty->print_cr("NULL");} 181 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) { 182 if (trace) {tty->print_cr("Non oop");} 183 } else if (in->bottom_type()->make_ptr()->make_oopptr() == TypeInstPtr::MIRROR) { 184 if (trace) {tty->print_cr("Java mirror");} 185 } else { 186 if (in->is_ConstraintCast()) { 187 in = in->in(1); 188 continue; 189 } else if (in->is_AddP()) { 190 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access"); 191 in = in->in(AddPNode::Address); 192 continue; 193 } else if (in->is_Con()) { 194 if (trace) { 195 tty->print("Found constant"); 196 in->dump(); 197 } 198 } else if (in->Opcode() == Op_Parm) { 199 if (trace) { 200 tty->print("Found argument"); 201 } 202 } else if (in->Opcode() == Op_CreateEx) { 203 if (trace) { 204 tty->print("Found create-exception"); 205 } 206 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) { 207 if (trace) { 208 tty->print("Found raw LoadP (OSR argument?)"); 209 } 210 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 211 if (t == ShenandoahOopStore) { 212 uint i = 0; 213 for (; i < phis.size(); i++) { 214 Node* n = phis.node_at(i); 215 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) { 216 break; 217 } 218 } 219 if (i == phis.size()) { 220 return false; 221 } 222 } 223 barriers_used.push(in); 224 if (trace) {tty->print("Found barrier"); in->dump();} 225 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) { 226 if (t != ShenandoahOopStore) { 227 in = in->in(1); 228 continue; 229 } 230 if (trace) {tty->print("Found enqueue barrier"); in->dump();} 231 phis.push(in, in->req()); 232 in = in->in(1); 233 continue; 234 } else if (in->is_Proj() && in->in(0)->is_Allocate()) { 235 if (trace) { 236 tty->print("Found alloc"); 237 in->in(0)->dump(); 238 } 239 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) { 240 if (trace) { 241 tty->print("Found Java call"); 242 } 243 } else if (in->is_Phi()) { 244 if (!visited.test_set(in->_idx)) { 245 if (trace) {tty->print("Pushed phi:"); in->dump();} 246 phis.push(in, 2); 247 in = in->in(1); 248 continue; 249 } 250 if (trace) {tty->print("Already seen phi:"); in->dump();} 251 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) { 252 if (!visited.test_set(in->_idx)) { 253 if (trace) {tty->print("Pushed cmovep:"); in->dump();} 254 phis.push(in, CMoveNode::IfTrue); 255 in = in->in(CMoveNode::IfFalse); 256 continue; 257 } 258 if (trace) {tty->print("Already seen cmovep:"); in->dump();} 259 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) { 260 in = in->in(1); 261 continue; 262 } else { 263 return false; 264 } 265 } 266 bool cont = false; 267 while (phis.is_nonempty()) { 268 uint idx = phis.index(); 269 Node* phi = phis.node(); 270 if (idx >= phi->req()) { 271 if (trace) {tty->print("Popped phi:"); phi->dump();} 272 phis.pop(); 273 continue; 274 } 275 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();} 276 in = phi->in(idx); 277 phis.set_index(idx+1); 278 cont = true; 279 break; 280 } 281 if (!cont) { 282 break; 283 } 284 } 285 return true; 286 } 287 288 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) { 289 if (n1 != NULL) { 290 n1->dump(+10); 291 } 292 if (n2 != NULL) { 293 n2->dump(+10); 294 } 295 fatal("%s", msg); 296 } 297 298 void ShenandoahBarrierC2Support::verify(RootNode* root) { 299 ResourceMark rm; 300 Unique_Node_List wq; 301 GrowableArray<Node*> barriers; 302 Unique_Node_List barriers_used; 303 Node_Stack phis(0); 304 VectorSet visited(Thread::current()->resource_area()); 305 const bool trace = false; 306 const bool verify_no_useless_barrier = false; 307 308 wq.push(root); 309 for (uint next = 0; next < wq.size(); next++) { 310 Node *n = wq.at(next); 311 if (n->is_Load()) { 312 const bool trace = false; 313 if (trace) {tty->print("Verifying"); n->dump();} 314 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) { 315 if (trace) {tty->print_cr("Load range/klass");} 316 } else { 317 const TypePtr* adr_type = n->as_Load()->adr_type(); 318 319 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) { 320 if (trace) {tty->print_cr("Mark load");} 321 } else if (adr_type->isa_instptr() && 322 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) && 323 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) { 324 if (trace) {tty->print_cr("Reference.get()");} 325 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) { 326 report_verify_failure("Shenandoah verification: Load should have barriers", n); 327 } 328 } 329 } else if (n->is_Store()) { 330 const bool trace = false; 331 332 if (trace) {tty->print("Verifying"); n->dump();} 333 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) { 334 Node* adr = n->in(MemNode::Address); 335 bool verify = true; 336 337 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) { 338 adr = adr->in(AddPNode::Address); 339 if (adr->is_AddP()) { 340 assert(adr->in(AddPNode::Base)->is_top(), ""); 341 adr = adr->in(AddPNode::Address); 342 if (adr->Opcode() == Op_LoadP && 343 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() && 344 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 345 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) { 346 if (trace) {tty->print_cr("SATB prebarrier");} 347 verify = false; 348 } 349 } 350 } 351 352 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) { 353 report_verify_failure("Shenandoah verification: Store should have barriers", n); 354 } 355 } 356 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) { 357 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n); 358 } 359 } else if (n->Opcode() == Op_CmpP) { 360 const bool trace = false; 361 362 Node* in1 = n->in(1); 363 Node* in2 = n->in(2); 364 if (in1->bottom_type()->isa_oopptr()) { 365 if (trace) {tty->print("Verifying"); n->dump();} 366 367 bool mark_inputs = false; 368 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR || 369 (in1->is_Con() || in2->is_Con())) { 370 if (trace) {tty->print_cr("Comparison against a constant");} 371 mark_inputs = true; 372 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) || 373 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) { 374 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");} 375 mark_inputs = true; 376 } else { 377 assert(in2->bottom_type()->isa_oopptr(), ""); 378 379 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) || 380 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) { 381 report_verify_failure("Shenandoah verification: Cmp should have barriers", n); 382 } 383 } 384 if (verify_no_useless_barrier && 385 mark_inputs && 386 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) || 387 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) { 388 phis.clear(); 389 visited.Reset(); 390 } 391 } 392 } else if (n->is_LoadStore()) { 393 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() && 394 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) { 395 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n); 396 } 397 398 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) { 399 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n); 400 } 401 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) { 402 CallNode* call = n->as_Call(); 403 404 static struct { 405 const char* name; 406 struct { 407 int pos; 408 verify_type t; 409 } args[6]; 410 } calls[] = { 411 "aescrypt_encryptBlock", 412 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 413 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 414 "aescrypt_decryptBlock", 415 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 416 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 417 "multiplyToLen", 418 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore }, 419 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 420 "squareToLen", 421 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone}, 422 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 423 "montgomery_multiply", 424 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, 425 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 426 "montgomery_square", 427 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore }, 428 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 429 "mulAdd", 430 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, 431 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 432 "vectorizedMismatch", 433 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, 434 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 435 "updateBytesCRC32", 436 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 437 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 438 "updateBytesAdler32", 439 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 440 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 441 "updateBytesCRC32C", 442 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone}, 443 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 444 "counterMode_AESCrypt", 445 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 446 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } }, 447 "cipherBlockChaining_encryptAESCrypt", 448 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 449 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 450 "cipherBlockChaining_decryptAESCrypt", 451 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 452 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 453 "shenandoah_clone_barrier", 454 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 455 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 456 "ghash_processBlocks", 457 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, 458 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 459 "sha1_implCompress", 460 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 461 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 462 "sha256_implCompress", 463 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 464 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 465 "sha512_implCompress", 466 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 467 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 468 "sha1_implCompressMB", 469 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 470 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 471 "sha256_implCompressMB", 472 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 473 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 474 "sha512_implCompressMB", 475 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 476 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 477 "encodeBlock", 478 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone }, 479 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 480 }; 481 482 if (call->is_call_to_arraycopystub()) { 483 Node* dest = NULL; 484 const TypeTuple* args = n->as_Call()->_tf->domain(); 485 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) { 486 if (args->field_at(i)->isa_ptr()) { 487 j++; 488 if (j == 2) { 489 dest = n->in(i); 490 break; 491 } 492 } 493 } 494 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) || 495 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) { 496 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n); 497 } 498 } else if (strlen(call->_name) > 5 && 499 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) { 500 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) { 501 report_verify_failure("Shenandoah verification: _fill should have barriers", n); 502 } 503 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) { 504 // skip 505 } else { 506 const int calls_len = sizeof(calls) / sizeof(calls[0]); 507 int i = 0; 508 for (; i < calls_len; i++) { 509 if (!strcmp(calls[i].name, call->_name)) { 510 break; 511 } 512 } 513 if (i != calls_len) { 514 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]); 515 for (uint j = 0; j < args_len; j++) { 516 int pos = calls[i].args[j].pos; 517 if (pos == -1) { 518 break; 519 } 520 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) { 521 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n); 522 } 523 } 524 for (uint j = TypeFunc::Parms; j < call->req(); j++) { 525 if (call->in(j)->bottom_type()->make_ptr() && 526 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) { 527 uint k = 0; 528 for (; k < args_len && calls[i].args[k].pos != (int)j; k++); 529 if (k == args_len) { 530 fatal("arg %d for call %s not covered", j, call->_name); 531 } 532 } 533 } 534 } else { 535 for (uint j = TypeFunc::Parms; j < call->req(); j++) { 536 if (call->in(j)->bottom_type()->make_ptr() && 537 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) { 538 fatal("%s not covered", call->_name); 539 } 540 } 541 } 542 } 543 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) { 544 // skip 545 } else if (n->is_AddP() 546 || n->is_Phi() 547 || n->is_ConstraintCast() 548 || n->Opcode() == Op_Return 549 || n->Opcode() == Op_CMoveP 550 || n->Opcode() == Op_CMoveN 551 || n->Opcode() == Op_Rethrow 552 || n->is_MemBar() 553 || n->Opcode() == Op_Conv2B 554 || n->Opcode() == Op_SafePoint 555 || n->is_CallJava() 556 || n->Opcode() == Op_Unlock 557 || n->Opcode() == Op_EncodeP 558 || n->Opcode() == Op_DecodeN) { 559 // nothing to do 560 } else { 561 static struct { 562 int opcode; 563 struct { 564 int pos; 565 verify_type t; 566 } inputs[2]; 567 } others[] = { 568 Op_FastLock, 569 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} }, 570 Op_Lock, 571 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} }, 572 Op_ArrayCopy, 573 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } }, 574 Op_StrCompressedCopy, 575 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 576 Op_StrInflatedCopy, 577 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 578 Op_AryEq, 579 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } }, 580 Op_StrIndexOf, 581 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } }, 582 Op_StrComp, 583 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } }, 584 Op_StrEquals, 585 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } }, 586 Op_EncodeISOArray, 587 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 588 Op_HasNegatives, 589 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} }, 590 Op_CastP2X, 591 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} }, 592 Op_StrIndexOfChar, 593 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } }, 594 }; 595 596 const int others_len = sizeof(others) / sizeof(others[0]); 597 int i = 0; 598 for (; i < others_len; i++) { 599 if (others[i].opcode == n->Opcode()) { 600 break; 601 } 602 } 603 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req(); 604 if (i != others_len) { 605 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]); 606 for (uint j = 0; j < inputs_len; j++) { 607 int pos = others[i].inputs[j].pos; 608 if (pos == -1) { 609 break; 610 } 611 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) { 612 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n); 613 } 614 } 615 for (uint j = 1; j < stop; j++) { 616 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() && 617 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) { 618 uint k = 0; 619 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++); 620 if (k == inputs_len) { 621 fatal("arg %d for node %s not covered", j, n->Name()); 622 } 623 } 624 } 625 } else { 626 for (uint j = 1; j < stop; j++) { 627 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() && 628 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) { 629 fatal("%s not covered", n->Name()); 630 } 631 } 632 } 633 } 634 635 if (n->is_SafePoint()) { 636 SafePointNode* sfpt = n->as_SafePoint(); 637 if (verify_no_useless_barrier && sfpt->jvms() != NULL) { 638 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) { 639 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) { 640 phis.clear(); 641 visited.Reset(); 642 } 643 } 644 } 645 } 646 } 647 648 if (verify_no_useless_barrier) { 649 for (int i = 0; i < barriers.length(); i++) { 650 Node* n = barriers.at(i); 651 if (!barriers_used.member(n)) { 652 tty->print("XXX useless barrier"); n->dump(-2); 653 ShouldNotReachHere(); 654 } 655 } 656 } 657 } 658 #endif 659 660 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) { 661 // That both nodes have the same control is not sufficient to prove 662 // domination, verify that there's no path from d to n 663 ResourceMark rm; 664 Unique_Node_List wq; 665 wq.push(d); 666 for (uint next = 0; next < wq.size(); next++) { 667 Node *m = wq.at(next); 668 if (m == n) { 669 return false; 670 } 671 if (m->is_Phi() && m->in(0)->is_Loop()) { 672 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control"); 673 } else { 674 for (uint i = 0; i < m->req(); i++) { 675 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) { 676 wq.push(m->in(i)); 677 } 678 } 679 } 680 } 681 return true; 682 } 683 684 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) { 685 if (d_c != n_c) { 686 return phase->is_dominator(d_c, n_c); 687 } 688 return is_dominator_same_ctrl(d_c, d, n, phase); 689 } 690 691 Node* next_mem(Node* mem, int alias) { 692 Node* res = NULL; 693 if (mem->is_Proj()) { 694 res = mem->in(0); 695 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 696 res = mem->in(TypeFunc::Memory); 697 } else if (mem->is_Phi()) { 698 res = mem->in(1); 699 } else if (mem->is_MergeMem()) { 700 res = mem->as_MergeMem()->memory_at(alias); 701 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 702 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier"); 703 res = mem->in(MemNode::Memory); 704 } else { 705 #ifdef ASSERT 706 mem->dump(); 707 #endif 708 ShouldNotReachHere(); 709 } 710 return res; 711 } 712 713 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) { 714 Node* iffproj = NULL; 715 while (c != dom) { 716 Node* next = phase->idom(c); 717 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?"); 718 if (c->is_Region()) { 719 ResourceMark rm; 720 Unique_Node_List wq; 721 wq.push(c); 722 for (uint i = 0; i < wq.size(); i++) { 723 Node *n = wq.at(i); 724 if (n == next) { 725 continue; 726 } 727 if (n->is_Region()) { 728 for (uint j = 1; j < n->req(); j++) { 729 wq.push(n->in(j)); 730 } 731 } else { 732 wq.push(n->in(0)); 733 } 734 } 735 for (uint i = 0; i < wq.size(); i++) { 736 Node *n = wq.at(i); 737 assert(n->is_CFG(), ""); 738 if (n->is_Multi()) { 739 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 740 Node* u = n->fast_out(j); 741 if (u->is_CFG()) { 742 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 743 return NodeSentinel; 744 } 745 } 746 } 747 } 748 } 749 } else if (c->is_Proj()) { 750 if (c->is_IfProj()) { 751 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) { 752 // continue; 753 } else { 754 if (!allow_one_proj) { 755 return NodeSentinel; 756 } 757 if (iffproj == NULL) { 758 iffproj = c; 759 } else { 760 return NodeSentinel; 761 } 762 } 763 } else if (c->Opcode() == Op_JumpProj) { 764 return NodeSentinel; // unsupported 765 } else if (c->Opcode() == Op_CatchProj) { 766 return NodeSentinel; // unsupported 767 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) { 768 return NodeSentinel; // unsupported 769 } else { 770 assert(next->unique_ctrl_out() == c, "unsupported branch pattern"); 771 } 772 } 773 c = next; 774 } 775 return iffproj; 776 } 777 778 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) { 779 ResourceMark rm; 780 VectorSet wq(Thread::current()->resource_area()); 781 wq.set(mem->_idx); 782 mem_ctrl = phase->ctrl_or_self(mem); 783 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) { 784 mem = next_mem(mem, alias); 785 if (wq.test_set(mem->_idx)) { 786 return NULL; 787 } 788 mem_ctrl = phase->ctrl_or_self(mem); 789 } 790 if (mem->is_MergeMem()) { 791 mem = mem->as_MergeMem()->memory_at(alias); 792 mem_ctrl = phase->ctrl_or_self(mem); 793 } 794 return mem; 795 } 796 797 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) { 798 Node* mem = NULL; 799 Node* c = ctrl; 800 do { 801 if (c->is_Region()) { 802 Node* phi_bottom = NULL; 803 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) { 804 Node* u = c->fast_out(i); 805 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) { 806 if (u->adr_type() == TypePtr::BOTTOM) { 807 mem = u; 808 } 809 } 810 } 811 } else { 812 if (c->is_Call() && c->as_Call()->adr_type() != NULL) { 813 CallProjections projs; 814 c->as_Call()->extract_projections(&projs, true, false); 815 if (projs.fallthrough_memproj != NULL) { 816 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) { 817 if (projs.catchall_memproj == NULL) { 818 mem = projs.fallthrough_memproj; 819 } else { 820 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) { 821 mem = projs.fallthrough_memproj; 822 } else { 823 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier"); 824 mem = projs.catchall_memproj; 825 } 826 } 827 } 828 } else { 829 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory); 830 if (proj != NULL && 831 proj->adr_type() == TypePtr::BOTTOM) { 832 mem = proj; 833 } 834 } 835 } else { 836 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 837 Node* u = c->fast_out(i); 838 if (u->is_Proj() && 839 u->bottom_type() == Type::MEMORY && 840 u->adr_type() == TypePtr::BOTTOM) { 841 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), ""); 842 assert(mem == NULL, "only one proj"); 843 mem = u; 844 } 845 } 846 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected"); 847 } 848 } 849 c = phase->idom(c); 850 } while (mem == NULL); 851 return mem; 852 } 853 854 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) { 855 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 856 Node* u = n->fast_out(i); 857 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) { 858 uses.push(u); 859 } 860 } 861 } 862 863 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) { 864 OuterStripMinedLoopEndNode* le = inner->outer_loop_end(); 865 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl)); 866 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl)); 867 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt); 868 phase->register_control(new_le, phase->get_loop(le), le->in(0)); 869 phase->lazy_replace(outer, new_outer); 870 phase->lazy_replace(le, new_le); 871 inner->clear_strip_mined(); 872 } 873 874 void ShenandoahBarrierC2Support::test_heap_stable(Node*& ctrl, Node* raw_mem, Node*& heap_stable_ctrl, 875 PhaseIdealLoop* phase) { 876 IdealLoopTree* loop = phase->get_loop(ctrl); 877 Node* thread = new ThreadLocalNode(); 878 phase->register_new_node(thread, ctrl); 879 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 880 phase->set_ctrl(offset, phase->C->root()); 881 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset); 882 phase->register_new_node(gc_state_addr, ctrl); 883 uint gc_state_idx = Compile::AliasIdxRaw; 884 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument 885 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx)); 886 887 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered); 888 phase->register_new_node(gc_state, ctrl); 889 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)); 890 phase->register_new_node(heap_stable_and, ctrl); 891 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT)); 892 phase->register_new_node(heap_stable_cmp, ctrl); 893 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne); 894 phase->register_new_node(heap_stable_test, ctrl); 895 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 896 phase->register_control(heap_stable_iff, loop, ctrl); 897 898 heap_stable_ctrl = new IfFalseNode(heap_stable_iff); 899 phase->register_control(heap_stable_ctrl, loop, heap_stable_iff); 900 ctrl = new IfTrueNode(heap_stable_iff); 901 phase->register_control(ctrl, loop, heap_stable_iff); 902 903 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape"); 904 } 905 906 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) { 907 const Type* val_t = phase->igvn().type(val); 908 if (val_t->meet(TypePtr::NULL_PTR) == val_t) { 909 IdealLoopTree* loop = phase->get_loop(ctrl); 910 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT)); 911 phase->register_new_node(null_cmp, ctrl); 912 Node* null_test = new BoolNode(null_cmp, BoolTest::ne); 913 phase->register_new_node(null_test, ctrl); 914 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 915 phase->register_control(null_iff, loop, ctrl); 916 ctrl = new IfTrueNode(null_iff); 917 phase->register_control(ctrl, loop, null_iff); 918 null_ctrl = new IfFalseNode(null_iff); 919 phase->register_control(null_ctrl, loop, null_iff); 920 } 921 } 922 923 Node* ShenandoahBarrierC2Support::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) { 924 IdealLoopTree *loop = phase->get_loop(c); 925 Node* iff = unc_ctrl->in(0); 926 assert(iff->is_If(), "broken"); 927 Node* new_iff = iff->clone(); 928 new_iff->set_req(0, c); 929 phase->register_control(new_iff, loop, c); 930 Node* iffalse = new IfFalseNode(new_iff->as_If()); 931 phase->register_control(iffalse, loop, new_iff); 932 Node* iftrue = new IfTrueNode(new_iff->as_If()); 933 phase->register_control(iftrue, loop, new_iff); 934 c = iftrue; 935 const Type *t = phase->igvn().type(val); 936 assert(val->Opcode() == Op_CastPP, "expect cast to non null here"); 937 Node* uncasted_val = val->in(1); 938 val = new CastPPNode(uncasted_val, t); 939 val->init_req(0, c); 940 phase->register_new_node(val, c); 941 return val; 942 } 943 944 void ShenandoahBarrierC2Support::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl, 945 Unique_Node_List& uses, PhaseIdealLoop* phase) { 946 IfNode* iff = unc_ctrl->in(0)->as_If(); 947 Node* proj = iff->proj_out(0); 948 assert(proj != unc_ctrl, "bad projection"); 949 Node* use = proj->unique_ctrl_out(); 950 951 assert(use == unc || use->is_Region(), "what else?"); 952 953 uses.clear(); 954 if (use == unc) { 955 phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use)); 956 for (uint i = 1; i < unc->req(); i++) { 957 Node* n = unc->in(i); 958 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) { 959 uses.push(n); 960 } 961 } 962 } else { 963 assert(use->is_Region(), "what else?"); 964 uint idx = 1; 965 for (; use->in(idx) != proj; idx++); 966 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 967 Node* u = use->fast_out(i); 968 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) { 969 uses.push(u->in(idx)); 970 } 971 } 972 } 973 for(uint next = 0; next < uses.size(); next++ ) { 974 Node *n = uses.at(next); 975 assert(phase->get_ctrl(n) == proj, "bad control"); 976 phase->set_ctrl_and_loop(n, new_unc_ctrl); 977 if (n->in(0) == proj) { 978 phase->igvn().replace_input_of(n, 0, new_unc_ctrl); 979 } 980 for (uint i = 0; i < n->req(); i++) { 981 Node* m = n->in(i); 982 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) { 983 uses.push(m); 984 } 985 } 986 } 987 988 phase->igvn().rehash_node_delayed(use); 989 int nb = use->replace_edge(proj, new_unc_ctrl); 990 assert(nb == 1, "only use expected"); 991 } 992 993 void ShenandoahBarrierC2Support::in_cset_fast_test(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) { 994 IdealLoopTree *loop = phase->get_loop(ctrl); 995 Node* raw_rbtrue = new CastP2XNode(ctrl, val); 996 phase->register_new_node(raw_rbtrue, ctrl); 997 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint())); 998 phase->register_new_node(cset_offset, ctrl); 999 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr())); 1000 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root()); 1001 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset); 1002 phase->register_new_node(in_cset_fast_test_adr, ctrl); 1003 uint in_cset_fast_test_idx = Compile::AliasIdxRaw; 1004 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument 1005 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx)); 1006 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); 1007 phase->register_new_node(in_cset_fast_test_load, ctrl); 1008 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT)); 1009 phase->register_new_node(in_cset_fast_test_cmp, ctrl); 1010 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::eq); 1011 phase->register_new_node(in_cset_fast_test_test, ctrl); 1012 IfNode* in_cset_fast_test_iff = new IfNode(ctrl, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 1013 phase->register_control(in_cset_fast_test_iff, loop, ctrl); 1014 1015 not_cset_ctrl = new IfTrueNode(in_cset_fast_test_iff); 1016 phase->register_control(not_cset_ctrl, loop, in_cset_fast_test_iff); 1017 1018 ctrl = new IfFalseNode(in_cset_fast_test_iff); 1019 phase->register_control(ctrl, loop, in_cset_fast_test_iff); 1020 } 1021 1022 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr, Node*& result_mem, Node* raw_mem, bool is_native, PhaseIdealLoop* phase) { 1023 IdealLoopTree*loop = phase->get_loop(ctrl); 1024 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr(); 1025 1026 // The slow path stub consumes and produces raw memory in addition 1027 // to the existing memory edges 1028 Node* base = find_bottom_mem(ctrl, phase); 1029 MergeMemNode* mm = MergeMemNode::make(base); 1030 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 1031 phase->register_new_node(mm, ctrl); 1032 1033 address target = LP64_ONLY(UseCompressedOops) NOT_LP64(false) ? 1034 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow) : 1035 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier); 1036 1037 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), 1038 target, 1039 "shenandoah_load_reference_barrier", TypeRawPtr::BOTTOM); 1040 call->init_req(TypeFunc::Control, ctrl); 1041 call->init_req(TypeFunc::I_O, phase->C->top()); 1042 call->init_req(TypeFunc::Memory, mm); 1043 call->init_req(TypeFunc::FramePtr, phase->C->top()); 1044 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 1045 call->init_req(TypeFunc::Parms, val); 1046 call->init_req(TypeFunc::Parms+1, load_addr); 1047 phase->register_control(call, loop, ctrl); 1048 ctrl = new ProjNode(call, TypeFunc::Control); 1049 phase->register_control(ctrl, loop, call); 1050 result_mem = new ProjNode(call, TypeFunc::Memory); 1051 phase->register_new_node(result_mem, call); 1052 val = new ProjNode(call, TypeFunc::Parms); 1053 phase->register_new_node(val, call); 1054 val = new CheckCastPPNode(ctrl, val, obj_type); 1055 phase->register_new_node(val, ctrl); 1056 } 1057 1058 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) { 1059 Node* ctrl = phase->get_ctrl(barrier); 1060 Node* init_raw_mem = fixer.find_mem(ctrl, barrier); 1061 1062 // Update the control of all nodes that should be after the 1063 // barrier control flow 1064 uses.clear(); 1065 // Every node that is control dependent on the barrier's input 1066 // control will be after the expanded barrier. The raw memory (if 1067 // its memory is control dependent on the barrier's input control) 1068 // must stay above the barrier. 1069 uses_to_ignore.clear(); 1070 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) { 1071 uses_to_ignore.push(init_raw_mem); 1072 } 1073 for (uint next = 0; next < uses_to_ignore.size(); next++) { 1074 Node *n = uses_to_ignore.at(next); 1075 for (uint i = 0; i < n->req(); i++) { 1076 Node* in = n->in(i); 1077 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) { 1078 uses_to_ignore.push(in); 1079 } 1080 } 1081 } 1082 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) { 1083 Node* u = ctrl->fast_out(i); 1084 if (u->_idx < last && 1085 u != barrier && 1086 !uses_to_ignore.member(u) && 1087 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) && 1088 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) { 1089 Node* old_c = phase->ctrl_or_self(u); 1090 Node* c = old_c; 1091 if (c != ctrl || 1092 is_dominator_same_ctrl(old_c, barrier, u, phase) || 1093 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) { 1094 phase->igvn().rehash_node_delayed(u); 1095 int nb = u->replace_edge(ctrl, region); 1096 if (u->is_CFG()) { 1097 if (phase->idom(u) == ctrl) { 1098 phase->set_idom(u, region, phase->dom_depth(region)); 1099 } 1100 } else if (phase->get_ctrl(u) == ctrl) { 1101 assert(u != init_raw_mem, "should leave input raw mem above the barrier"); 1102 uses.push(u); 1103 } 1104 assert(nb == 1, "more than 1 ctrl input?"); 1105 --i, imax -= nb; 1106 } 1107 } 1108 } 1109 } 1110 1111 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) { 1112 Node* region = NULL; 1113 while (c != ctrl) { 1114 if (c->is_Region()) { 1115 region = c; 1116 } 1117 c = phase->idom(c); 1118 } 1119 assert(region != NULL, ""); 1120 Node* phi = new PhiNode(region, n->bottom_type()); 1121 for (uint j = 1; j < region->req(); j++) { 1122 Node* in = region->in(j); 1123 if (phase->is_dominator(projs.fallthrough_catchproj, in)) { 1124 phi->init_req(j, n); 1125 } else if (phase->is_dominator(projs.catchall_catchproj, in)) { 1126 phi->init_req(j, n_clone); 1127 } else { 1128 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase)); 1129 } 1130 } 1131 phase->register_new_node(phi, region); 1132 return phi; 1133 } 1134 1135 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) { 1136 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state(); 1137 1138 Unique_Node_List uses; 1139 for (int i = 0; i < state->enqueue_barriers_count(); i++) { 1140 Node* barrier = state->enqueue_barrier(i); 1141 Node* ctrl = phase->get_ctrl(barrier); 1142 IdealLoopTree* loop = phase->get_loop(ctrl); 1143 if (loop->_head->is_OuterStripMinedLoop()) { 1144 // Expanding a barrier here will break loop strip mining 1145 // verification. Transform the loop so the loop nest doesn't 1146 // appear as strip mined. 1147 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop(); 1148 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase); 1149 } 1150 } 1151 1152 Node_Stack stack(0); 1153 Node_List clones; 1154 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) { 1155 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1156 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1157 continue; 1158 } 1159 1160 Node* ctrl = phase->get_ctrl(lrb); 1161 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 1162 1163 CallStaticJavaNode* unc = NULL; 1164 Node* unc_ctrl = NULL; 1165 Node* uncasted_val = val; 1166 1167 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) { 1168 Node* u = lrb->fast_out(i); 1169 if (u->Opcode() == Op_CastPP && 1170 u->in(0) != NULL && 1171 phase->is_dominator(u->in(0), ctrl)) { 1172 const Type* u_t = phase->igvn().type(u); 1173 1174 if (u_t->meet(TypePtr::NULL_PTR) != u_t && 1175 u->in(0)->Opcode() == Op_IfTrue && 1176 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 1177 u->in(0)->in(0)->is_If() && 1178 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool && 1179 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne && 1180 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 1181 u->in(0)->in(0)->in(1)->in(1)->in(1) == val && 1182 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 1183 IdealLoopTree* loop = phase->get_loop(ctrl); 1184 IdealLoopTree* unc_loop = phase->get_loop(u->in(0)); 1185 1186 if (!unc_loop->is_member(loop)) { 1187 continue; 1188 } 1189 1190 Node* branch = no_branches(ctrl, u->in(0), false, phase); 1191 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch"); 1192 if (branch == NodeSentinel) { 1193 continue; 1194 } 1195 1196 phase->igvn().replace_input_of(u, 1, val); 1197 phase->igvn().replace_input_of(lrb, ShenandoahLoadReferenceBarrierNode::ValueIn, u); 1198 phase->set_ctrl(u, u->in(0)); 1199 phase->set_ctrl(lrb, u->in(0)); 1200 unc = u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1201 unc_ctrl = u->in(0); 1202 val = u; 1203 1204 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 1205 Node* u = val->fast_out(j); 1206 if (u == lrb) continue; 1207 phase->igvn().rehash_node_delayed(u); 1208 int nb = u->replace_edge(val, lrb); 1209 --j; jmax -= nb; 1210 } 1211 1212 RegionNode* r = new RegionNode(3); 1213 IfNode* iff = unc_ctrl->in(0)->as_If(); 1214 1215 Node* ctrl_use = unc_ctrl->unique_ctrl_out(); 1216 Node* unc_ctrl_clone = unc_ctrl->clone(); 1217 phase->register_control(unc_ctrl_clone, loop, iff); 1218 Node* c = unc_ctrl_clone; 1219 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase); 1220 r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0)); 1221 1222 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0)); 1223 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl)); 1224 phase->lazy_replace(c, unc_ctrl); 1225 c = NULL;; 1226 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone); 1227 phase->set_ctrl(val, unc_ctrl_clone); 1228 1229 IfNode* new_iff = new_cast->in(0)->in(0)->as_If(); 1230 fix_null_check(unc, unc_ctrl_clone, r, uses, phase); 1231 Node* iff_proj = iff->proj_out(0); 1232 r->init_req(2, iff_proj); 1233 phase->register_control(r, phase->ltree_root(), iff); 1234 1235 Node* new_bol = new_iff->in(1)->clone(); 1236 Node* new_cmp = new_bol->in(1)->clone(); 1237 assert(new_cmp->Opcode() == Op_CmpP, "broken"); 1238 assert(new_cmp->in(1) == val->in(1), "broken"); 1239 new_bol->set_req(1, new_cmp); 1240 new_cmp->set_req(1, lrb); 1241 phase->register_new_node(new_bol, new_iff->in(0)); 1242 phase->register_new_node(new_cmp, new_iff->in(0)); 1243 phase->igvn().replace_input_of(new_iff, 1, new_bol); 1244 phase->igvn().replace_input_of(new_cast, 1, lrb); 1245 1246 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) { 1247 Node* u = lrb->fast_out(i); 1248 if (u == new_cast || u == new_cmp) { 1249 continue; 1250 } 1251 phase->igvn().rehash_node_delayed(u); 1252 int nb = u->replace_edge(lrb, new_cast); 1253 assert(nb > 0, "no update?"); 1254 --i; imax -= nb; 1255 } 1256 1257 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 1258 Node* u = val->fast_out(i); 1259 if (u == lrb) { 1260 continue; 1261 } 1262 phase->igvn().rehash_node_delayed(u); 1263 int nb = u->replace_edge(val, new_cast); 1264 assert(nb > 0, "no update?"); 1265 --i; imax -= nb; 1266 } 1267 1268 ctrl = unc_ctrl_clone; 1269 phase->set_ctrl_and_loop(lrb, ctrl); 1270 break; 1271 } 1272 } 1273 } 1274 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) { 1275 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava(); 1276 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 1277 // The rethrow call may have too many projections to be 1278 // properly handled here. Given there's no reason for a 1279 // barrier to depend on the call, move it above the call 1280 if (phase->get_ctrl(val) == ctrl) { 1281 assert(val->Opcode() == Op_DecodeN, "unexpected node"); 1282 assert(phase->is_dominator(phase->get_ctrl(val->in(1)), call->in(0)), "Load is too low"); 1283 phase->set_ctrl(val, call->in(0)); 1284 } 1285 phase->set_ctrl(lrb, call->in(0)); 1286 continue; 1287 } 1288 CallProjections projs; 1289 call->extract_projections(&projs, false, false); 1290 1291 Node* lrb_clone = lrb->clone(); 1292 phase->register_new_node(lrb_clone, projs.catchall_catchproj); 1293 phase->set_ctrl(lrb, projs.fallthrough_catchproj); 1294 1295 stack.push(lrb, 0); 1296 clones.push(lrb_clone); 1297 1298 do { 1299 assert(stack.size() == clones.size(), ""); 1300 Node* n = stack.node(); 1301 #ifdef ASSERT 1302 if (n->is_Load()) { 1303 Node* mem = n->in(MemNode::Memory); 1304 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) { 1305 Node* u = mem->fast_out(j); 1306 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?"); 1307 } 1308 } 1309 #endif 1310 uint idx = stack.index(); 1311 Node* n_clone = clones.at(clones.size()-1); 1312 if (idx < n->outcnt()) { 1313 Node* u = n->raw_out(idx); 1314 Node* c = phase->ctrl_or_self(u); 1315 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) { 1316 stack.set_index(idx+1); 1317 assert(!u->is_CFG(), ""); 1318 stack.push(u, 0); 1319 Node* u_clone = u->clone(); 1320 int nb = u_clone->replace_edge(n, n_clone); 1321 assert(nb > 0, "should have replaced some uses"); 1322 phase->register_new_node(u_clone, projs.catchall_catchproj); 1323 clones.push(u_clone); 1324 phase->set_ctrl(u, projs.fallthrough_catchproj); 1325 } else { 1326 bool replaced = false; 1327 if (u->is_Phi()) { 1328 for (uint k = 1; k < u->req(); k++) { 1329 if (u->in(k) == n) { 1330 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) { 1331 phase->igvn().replace_input_of(u, k, n_clone); 1332 replaced = true; 1333 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) { 1334 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase)); 1335 replaced = true; 1336 } 1337 } 1338 } 1339 } else { 1340 if (phase->is_dominator(projs.catchall_catchproj, c)) { 1341 phase->igvn().rehash_node_delayed(u); 1342 int nb = u->replace_edge(n, n_clone); 1343 assert(nb > 0, "should have replaced some uses"); 1344 replaced = true; 1345 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) { 1346 phase->igvn().rehash_node_delayed(u); 1347 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase)); 1348 assert(nb > 0, "should have replaced some uses"); 1349 replaced = true; 1350 } 1351 } 1352 if (!replaced) { 1353 stack.set_index(idx+1); 1354 } 1355 } 1356 } else { 1357 stack.pop(); 1358 clones.pop(); 1359 } 1360 } while (stack.size() > 0); 1361 assert(stack.size() == 0 && clones.size() == 0, ""); 1362 } 1363 } 1364 1365 for (int i = 0; i < state->load_reference_barriers_count(); i++) { 1366 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1367 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1368 continue; 1369 } 1370 Node* ctrl = phase->get_ctrl(lrb); 1371 IdealLoopTree* loop = phase->get_loop(ctrl); 1372 if (loop->_head->is_OuterStripMinedLoop()) { 1373 // Expanding a barrier here will break loop strip mining 1374 // verification. Transform the loop so the loop nest doesn't 1375 // appear as strip mined. 1376 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop(); 1377 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase); 1378 } 1379 } 1380 1381 // Expand load-reference-barriers 1382 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase); 1383 Unique_Node_List uses_to_ignore; 1384 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) { 1385 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i); 1386 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) { 1387 phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn)); 1388 continue; 1389 } 1390 uint last = phase->C->unique(); 1391 Node* ctrl = phase->get_ctrl(lrb); 1392 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn); 1393 1394 1395 Node* orig_ctrl = ctrl; 1396 1397 Node* raw_mem = fixer.find_mem(ctrl, lrb); 1398 Node* init_raw_mem = raw_mem; 1399 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 1400 1401 IdealLoopTree *loop = phase->get_loop(ctrl); 1402 CallStaticJavaNode* unc = lrb->pin_and_expand_null_check(phase->igvn()); 1403 Node* unc_ctrl = NULL; 1404 if (unc != NULL) { 1405 if (val->in(ShenandoahLoadReferenceBarrierNode::Control) != ctrl) { 1406 unc = NULL; 1407 } else { 1408 unc_ctrl = val->in(ShenandoahLoadReferenceBarrierNode::Control); 1409 } 1410 } 1411 1412 Node* uncasted_val = val; 1413 if (unc != NULL) { 1414 uncasted_val = val->in(1); 1415 } 1416 1417 Node* heap_stable_ctrl = NULL; 1418 Node* null_ctrl = NULL; 1419 1420 assert(val->bottom_type()->make_oopptr(), "need oop"); 1421 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant"); 1422 1423 enum { _heap_stable = 1, _not_cset, _evac_path, _null_path, PATH_LIMIT }; 1424 Node* region = new RegionNode(PATH_LIMIT); 1425 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr()); 1426 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1427 1428 // Stable path. 1429 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase); 1430 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If(); 1431 1432 // Heap stable case 1433 region->init_req(_heap_stable, heap_stable_ctrl); 1434 val_phi->init_req(_heap_stable, uncasted_val); 1435 raw_mem_phi->init_req(_heap_stable, raw_mem); 1436 1437 Node* reg2_ctrl = NULL; 1438 // Null case 1439 test_null(ctrl, val, null_ctrl, phase); 1440 if (null_ctrl != NULL) { 1441 reg2_ctrl = null_ctrl->in(0); 1442 region->init_req(_null_path, null_ctrl); 1443 val_phi->init_req(_null_path, uncasted_val); 1444 raw_mem_phi->init_req(_null_path, raw_mem); 1445 } else { 1446 region->del_req(_null_path); 1447 val_phi->del_req(_null_path); 1448 raw_mem_phi->del_req(_null_path); 1449 } 1450 1451 // Test for in-cset. 1452 // Wires !in_cset(obj) to slot 2 of region and phis 1453 Node* not_cset_ctrl = NULL; 1454 in_cset_fast_test(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase); 1455 if (not_cset_ctrl != NULL) { 1456 if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0); 1457 region->init_req(_not_cset, not_cset_ctrl); 1458 val_phi->init_req(_not_cset, uncasted_val); 1459 raw_mem_phi->init_req(_not_cset, raw_mem); 1460 } 1461 1462 // Resolve object when orig-value is in cset. 1463 // Make the unconditional resolve for fwdptr. 1464 Node* new_val = uncasted_val; 1465 if (unc_ctrl != NULL) { 1466 // Clone the null check in this branch to allow implicit null check 1467 new_val = clone_null_check(ctrl, val, unc_ctrl, phase); 1468 fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase); 1469 1470 IfNode* iff = unc_ctrl->in(0)->as_If(); 1471 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1)); 1472 } 1473 1474 // Call lrb-stub and wire up that path in slots 4 1475 Node* result_mem = NULL; 1476 1477 Node* fwd = new_val; 1478 Node* addr; 1479 if (ShenandoahSelfFixing) { 1480 VectorSet visited(Thread::current()->resource_area()); 1481 addr = get_load_addr(phase, visited, lrb); 1482 } else { 1483 addr = phase->igvn().zerocon(T_OBJECT); 1484 } 1485 if (addr->Opcode() == Op_AddP) { 1486 Node* orig_base = addr->in(AddPNode::Base); 1487 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), true); 1488 phase->register_new_node(base, ctrl); 1489 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) { 1490 // Field access 1491 addr = addr->clone(); 1492 addr->set_req(AddPNode::Base, base); 1493 addr->set_req(AddPNode::Address, base); 1494 phase->register_new_node(addr, ctrl); 1495 } else { 1496 Node* addr2 = addr->in(AddPNode::Address); 1497 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) && 1498 addr2->in(AddPNode::Base) == orig_base) { 1499 addr2 = addr2->clone(); 1500 addr2->set_req(AddPNode::Base, base); 1501 addr2->set_req(AddPNode::Address, base); 1502 phase->register_new_node(addr2, ctrl); 1503 addr = addr->clone(); 1504 addr->set_req(AddPNode::Base, base); 1505 addr->set_req(AddPNode::Address, addr2); 1506 phase->register_new_node(addr, ctrl); 1507 } 1508 } 1509 } 1510 call_lrb_stub(ctrl, fwd, addr, result_mem, raw_mem, false, phase); 1511 region->init_req(_evac_path, ctrl); 1512 val_phi->init_req(_evac_path, fwd); 1513 raw_mem_phi->init_req(_evac_path, result_mem); 1514 1515 phase->register_control(region, loop, heap_stable_iff); 1516 Node* out_val = val_phi; 1517 phase->register_new_node(val_phi, region); 1518 phase->register_new_node(raw_mem_phi, region); 1519 1520 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase); 1521 1522 ctrl = orig_ctrl; 1523 1524 if (unc != NULL) { 1525 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 1526 Node* u = val->fast_out(i); 1527 Node* c = phase->ctrl_or_self(u); 1528 if (u != lrb && (c != ctrl || is_dominator_same_ctrl(c, lrb, u, phase))) { 1529 phase->igvn().rehash_node_delayed(u); 1530 int nb = u->replace_edge(val, out_val); 1531 --i, imax -= nb; 1532 } 1533 } 1534 if (val->outcnt() == 0) { 1535 phase->igvn()._worklist.push(val); 1536 } 1537 } 1538 phase->igvn().replace_node(lrb, out_val); 1539 1540 follow_barrier_uses(out_val, ctrl, uses, phase); 1541 1542 for(uint next = 0; next < uses.size(); next++ ) { 1543 Node *n = uses.at(next); 1544 assert(phase->get_ctrl(n) == ctrl, "bad control"); 1545 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 1546 phase->set_ctrl(n, region); 1547 follow_barrier_uses(n, ctrl, uses, phase); 1548 } 1549 1550 // The slow path call produces memory: hook the raw memory phi 1551 // from the expanded load reference barrier with the rest of the graph 1552 // which may require adding memory phis at every post dominated 1553 // region and at enclosing loop heads. Use the memory state 1554 // collected in memory_nodes to fix the memory graph. Update that 1555 // memory state as we go. 1556 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses); 1557 } 1558 // Done expanding load-reference-barriers. 1559 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced"); 1560 1561 for (int i = state->enqueue_barriers_count() - 1; i >= 0; i--) { 1562 Node* barrier = state->enqueue_barrier(i); 1563 Node* pre_val = barrier->in(1); 1564 1565 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) { 1566 ShouldNotReachHere(); 1567 continue; 1568 } 1569 1570 Node* ctrl = phase->get_ctrl(barrier); 1571 1572 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) { 1573 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move"); 1574 ctrl = ctrl->in(0)->in(0); 1575 phase->set_ctrl(barrier, ctrl); 1576 } else if (ctrl->is_CallRuntime()) { 1577 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move"); 1578 ctrl = ctrl->in(0); 1579 phase->set_ctrl(barrier, ctrl); 1580 } 1581 1582 Node* init_ctrl = ctrl; 1583 IdealLoopTree* loop = phase->get_loop(ctrl); 1584 Node* raw_mem = fixer.find_mem(ctrl, barrier); 1585 Node* init_raw_mem = raw_mem; 1586 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 1587 Node* heap_stable_ctrl = NULL; 1588 Node* null_ctrl = NULL; 1589 uint last = phase->C->unique(); 1590 1591 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT }; 1592 Node* region = new RegionNode(PATH_LIMIT); 1593 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1594 1595 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 }; 1596 Node* region2 = new RegionNode(PATH_LIMIT2); 1597 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 1598 1599 // Stable path. 1600 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase); 1601 region->init_req(_heap_stable, heap_stable_ctrl); 1602 phi->init_req(_heap_stable, raw_mem); 1603 1604 // Null path 1605 Node* reg2_ctrl = NULL; 1606 test_null(ctrl, pre_val, null_ctrl, phase); 1607 if (null_ctrl != NULL) { 1608 reg2_ctrl = null_ctrl->in(0); 1609 region2->init_req(_null_path, null_ctrl); 1610 phi2->init_req(_null_path, raw_mem); 1611 } else { 1612 region2->del_req(_null_path); 1613 phi2->del_req(_null_path); 1614 } 1615 1616 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 1617 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 1618 Node* thread = new ThreadLocalNode(); 1619 phase->register_new_node(thread, ctrl); 1620 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset)); 1621 phase->register_new_node(buffer_adr, ctrl); 1622 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset)); 1623 phase->register_new_node(index_adr, ctrl); 1624 1625 BasicType index_bt = TypeX_X->basic_type(); 1626 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size."); 1627 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 1628 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered); 1629 phase->register_new_node(index, ctrl); 1630 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0)); 1631 phase->register_new_node(index_cmp, ctrl); 1632 Node* index_test = new BoolNode(index_cmp, BoolTest::ne); 1633 phase->register_new_node(index_test, ctrl); 1634 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 1635 if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff; 1636 phase->register_control(queue_full_iff, loop, ctrl); 1637 Node* not_full = new IfTrueNode(queue_full_iff); 1638 phase->register_control(not_full, loop, queue_full_iff); 1639 Node* full = new IfFalseNode(queue_full_iff); 1640 phase->register_control(full, loop, queue_full_iff); 1641 1642 ctrl = not_full; 1643 1644 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t))); 1645 phase->register_new_node(next_index, ctrl); 1646 1647 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered); 1648 phase->register_new_node(buffer, ctrl); 1649 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index); 1650 phase->register_new_node(log_addr, ctrl); 1651 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered); 1652 phase->register_new_node(log_store, ctrl); 1653 // update the index 1654 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered); 1655 phase->register_new_node(index_update, ctrl); 1656 1657 // Fast-path case 1658 region2->init_req(_fast_path, ctrl); 1659 phi2->init_req(_fast_path, index_update); 1660 1661 ctrl = full; 1662 1663 Node* base = find_bottom_mem(ctrl, phase); 1664 1665 MergeMemNode* mm = MergeMemNode::make(base); 1666 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 1667 phase->register_new_node(mm, ctrl); 1668 1669 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); 1670 call->init_req(TypeFunc::Control, ctrl); 1671 call->init_req(TypeFunc::I_O, phase->C->top()); 1672 call->init_req(TypeFunc::Memory, mm); 1673 call->init_req(TypeFunc::FramePtr, phase->C->top()); 1674 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 1675 call->init_req(TypeFunc::Parms, pre_val); 1676 call->init_req(TypeFunc::Parms+1, thread); 1677 phase->register_control(call, loop, ctrl); 1678 1679 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control); 1680 phase->register_control(ctrl_proj, loop, call); 1681 Node* mem_proj = new ProjNode(call, TypeFunc::Memory); 1682 phase->register_new_node(mem_proj, call); 1683 1684 // Slow-path case 1685 region2->init_req(_slow_path, ctrl_proj); 1686 phi2->init_req(_slow_path, mem_proj); 1687 1688 phase->register_control(region2, loop, reg2_ctrl); 1689 phase->register_new_node(phi2, region2); 1690 1691 region->init_req(_heap_unstable, region2); 1692 phi->init_req(_heap_unstable, phi2); 1693 1694 phase->register_control(region, loop, heap_stable_ctrl->in(0)); 1695 phase->register_new_node(phi, region); 1696 1697 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase); 1698 for(uint next = 0; next < uses.size(); next++ ) { 1699 Node *n = uses.at(next); 1700 assert(phase->get_ctrl(n) == init_ctrl, "bad control"); 1701 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 1702 phase->set_ctrl(n, region); 1703 follow_barrier_uses(n, init_ctrl, uses, phase); 1704 } 1705 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses); 1706 1707 phase->igvn().replace_node(barrier, pre_val); 1708 } 1709 assert(state->enqueue_barriers_count() == 0, "all enqueue barrier nodes should have been replaced"); 1710 1711 } 1712 1713 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) { 1714 if (visited.test_set(in->_idx)) { 1715 return NULL; 1716 } 1717 switch (in->Opcode()) { 1718 case Op_Proj: 1719 return get_load_addr(phase, visited, in->in(0)); 1720 case Op_CastPP: 1721 case Op_CheckCastPP: 1722 case Op_DecodeN: 1723 case Op_EncodeP: 1724 return get_load_addr(phase, visited, in->in(1)); 1725 case Op_LoadN: 1726 case Op_LoadP: 1727 return in->in(MemNode::Address); 1728 case Op_CompareAndExchangeN: 1729 case Op_CompareAndExchangeP: 1730 case Op_GetAndSetN: 1731 case Op_GetAndSetP: 1732 case Op_ShenandoahCompareAndExchangeP: 1733 case Op_ShenandoahCompareAndExchangeN: 1734 // Those instructions would just have stored a different 1735 // value into the field. No use to attempt to fix it at this point. 1736 return phase->igvn().zerocon(T_OBJECT); 1737 case Op_CMoveP: 1738 case Op_CMoveN: { 1739 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue)); 1740 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse)); 1741 // Handle unambiguous cases: single address reported on both branches. 1742 if (t != NULL && f == NULL) return t; 1743 if (t == NULL && f != NULL) return f; 1744 if (t != NULL && t == f) return t; 1745 // Ambiguity. 1746 return phase->igvn().zerocon(T_OBJECT); 1747 } 1748 case Op_Phi: { 1749 Node* addr = NULL; 1750 for (uint i = 1; i < in->req(); i++) { 1751 Node* addr1 = get_load_addr(phase, visited, in->in(i)); 1752 if (addr == NULL) { 1753 addr = addr1; 1754 } 1755 if (addr != addr1) { 1756 return phase->igvn().zerocon(T_OBJECT); 1757 } 1758 } 1759 return addr; 1760 } 1761 case Op_ShenandoahLoadReferenceBarrier: 1762 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn)); 1763 case Op_ShenandoahEnqueueBarrier: 1764 return get_load_addr(phase, visited, in->in(1)); 1765 case Op_CallDynamicJava: 1766 case Op_CallLeaf: 1767 case Op_CallStaticJava: 1768 case Op_ConN: 1769 case Op_ConP: 1770 case Op_Parm: 1771 case Op_CreateEx: 1772 return phase->igvn().zerocon(T_OBJECT); 1773 default: 1774 #ifdef ASSERT 1775 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]); 1776 #endif 1777 return phase->igvn().zerocon(T_OBJECT); 1778 } 1779 1780 } 1781 1782 void ShenandoahBarrierC2Support::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) { 1783 IdealLoopTree *loop = phase->get_loop(iff); 1784 Node* loop_head = loop->_head; 1785 Node* entry_c = loop_head->in(LoopNode::EntryControl); 1786 1787 Node* bol = iff->in(1); 1788 Node* cmp = bol->in(1); 1789 Node* andi = cmp->in(1); 1790 Node* load = andi->in(1); 1791 1792 assert(is_gc_state_load(load), "broken"); 1793 if (!phase->is_dominator(load->in(0), entry_c)) { 1794 Node* mem_ctrl = NULL; 1795 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase); 1796 load = load->clone(); 1797 load->set_req(MemNode::Memory, mem); 1798 load->set_req(0, entry_c); 1799 phase->register_new_node(load, entry_c); 1800 andi = andi->clone(); 1801 andi->set_req(1, load); 1802 phase->register_new_node(andi, entry_c); 1803 cmp = cmp->clone(); 1804 cmp->set_req(1, andi); 1805 phase->register_new_node(cmp, entry_c); 1806 bol = bol->clone(); 1807 bol->set_req(1, cmp); 1808 phase->register_new_node(bol, entry_c); 1809 1810 Node* old_bol =iff->in(1); 1811 phase->igvn().replace_input_of(iff, 1, bol); 1812 } 1813 } 1814 1815 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) { 1816 if (!n->is_If() || n->is_CountedLoopEnd()) { 1817 return false; 1818 } 1819 Node* region = n->in(0); 1820 1821 if (!region->is_Region()) { 1822 return false; 1823 } 1824 Node* dom = phase->idom(region); 1825 if (!dom->is_If()) { 1826 return false; 1827 } 1828 1829 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) { 1830 return false; 1831 } 1832 1833 IfNode* dom_if = dom->as_If(); 1834 Node* proj_true = dom_if->proj_out(1); 1835 Node* proj_false = dom_if->proj_out(0); 1836 1837 for (uint i = 1; i < region->req(); i++) { 1838 if (phase->is_dominator(proj_true, region->in(i))) { 1839 continue; 1840 } 1841 if (phase->is_dominator(proj_false, region->in(i))) { 1842 continue; 1843 } 1844 return false; 1845 } 1846 1847 return true; 1848 } 1849 1850 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) { 1851 assert(is_heap_stable_test(n), "no other tests"); 1852 if (identical_backtoback_ifs(n, phase)) { 1853 Node* n_ctrl = n->in(0); 1854 if (phase->can_split_if(n_ctrl)) { 1855 IfNode* dom_if = phase->idom(n_ctrl)->as_If(); 1856 if (is_heap_stable_test(n)) { 1857 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1); 1858 assert(is_gc_state_load(gc_state_load), "broken"); 1859 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1); 1860 assert(is_gc_state_load(dom_gc_state_load), "broken"); 1861 if (gc_state_load != dom_gc_state_load) { 1862 phase->igvn().replace_node(gc_state_load, dom_gc_state_load); 1863 } 1864 } 1865 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1866 Node* proj_true = dom_if->proj_out(1); 1867 Node* proj_false = dom_if->proj_out(0); 1868 Node* con_true = phase->igvn().makecon(TypeInt::ONE); 1869 Node* con_false = phase->igvn().makecon(TypeInt::ZERO); 1870 1871 for (uint i = 1; i < n_ctrl->req(); i++) { 1872 if (phase->is_dominator(proj_true, n_ctrl->in(i))) { 1873 bolphi->init_req(i, con_true); 1874 } else { 1875 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1876 bolphi->init_req(i, con_false); 1877 } 1878 } 1879 phase->register_new_node(bolphi, n_ctrl); 1880 phase->igvn().replace_input_of(n, 1, bolphi); 1881 phase->do_split_if(n); 1882 } 1883 } 1884 } 1885 1886 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) { 1887 // Find first invariant test that doesn't exit the loop 1888 LoopNode *head = loop->_head->as_Loop(); 1889 IfNode* unswitch_iff = NULL; 1890 Node* n = head->in(LoopNode::LoopBackControl); 1891 int loop_has_sfpts = -1; 1892 while (n != head) { 1893 Node* n_dom = phase->idom(n); 1894 if (n->is_Region()) { 1895 if (n_dom->is_If()) { 1896 IfNode* iff = n_dom->as_If(); 1897 if (iff->in(1)->is_Bool()) { 1898 BoolNode* bol = iff->in(1)->as_Bool(); 1899 if (bol->in(1)->is_Cmp()) { 1900 // If condition is invariant and not a loop exit, 1901 // then found reason to unswitch. 1902 if (is_heap_stable_test(iff) && 1903 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) { 1904 assert(!loop->is_loop_exit(iff), "both branches should be in the loop"); 1905 if (loop_has_sfpts == -1) { 1906 for(uint i = 0; i < loop->_body.size(); i++) { 1907 Node *m = loop->_body[i]; 1908 if (m->is_SafePoint() && !m->is_CallLeaf()) { 1909 loop_has_sfpts = 1; 1910 break; 1911 } 1912 } 1913 if (loop_has_sfpts == -1) { 1914 loop_has_sfpts = 0; 1915 } 1916 } 1917 if (!loop_has_sfpts) { 1918 unswitch_iff = iff; 1919 } 1920 } 1921 } 1922 } 1923 } 1924 } 1925 n = n_dom; 1926 } 1927 return unswitch_iff; 1928 } 1929 1930 1931 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) { 1932 Node_List heap_stable_tests; 1933 Node_List gc_state_loads; 1934 stack.push(phase->C->start(), 0); 1935 do { 1936 Node* n = stack.node(); 1937 uint i = stack.index(); 1938 1939 if (i < n->outcnt()) { 1940 Node* u = n->raw_out(i); 1941 stack.set_index(i+1); 1942 if (!visited.test_set(u->_idx)) { 1943 stack.push(u, 0); 1944 } 1945 } else { 1946 stack.pop(); 1947 if (ShenandoahCommonGCStateLoads && is_gc_state_load(n)) { 1948 gc_state_loads.push(n); 1949 } 1950 if (n->is_If() && is_heap_stable_test(n)) { 1951 heap_stable_tests.push(n); 1952 } 1953 } 1954 } while (stack.size() > 0); 1955 1956 bool progress; 1957 do { 1958 progress = false; 1959 for (uint i = 0; i < gc_state_loads.size(); i++) { 1960 Node* n = gc_state_loads.at(i); 1961 if (n->outcnt() != 0) { 1962 progress |= try_common_gc_state_load(n, phase); 1963 } 1964 } 1965 } while (progress); 1966 1967 for (uint i = 0; i < heap_stable_tests.size(); i++) { 1968 Node* n = heap_stable_tests.at(i); 1969 assert(is_heap_stable_test(n), "only evacuation test"); 1970 merge_back_to_back_tests(n, phase); 1971 } 1972 1973 if (!phase->C->major_progress()) { 1974 VectorSet seen(Thread::current()->resource_area()); 1975 for (uint i = 0; i < heap_stable_tests.size(); i++) { 1976 Node* n = heap_stable_tests.at(i); 1977 IdealLoopTree* loop = phase->get_loop(n); 1978 if (loop != phase->ltree_root() && 1979 loop->_child == NULL && 1980 !loop->_irreducible) { 1981 LoopNode* head = loop->_head->as_Loop(); 1982 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) && 1983 !seen.test_set(head->_idx)) { 1984 IfNode* iff = find_unswitching_candidate(loop, phase); 1985 if (iff != NULL) { 1986 Node* bol = iff->in(1); 1987 if (head->is_strip_mined()) { 1988 head->verify_strip_mined(0); 1989 } 1990 move_heap_stable_test_out_of_loop(iff, phase); 1991 if (loop->policy_unswitching(phase)) { 1992 if (head->is_strip_mined()) { 1993 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop(); 1994 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase); 1995 } 1996 phase->do_unswitching(loop, old_new); 1997 } else { 1998 // Not proceeding with unswitching. Move load back in 1999 // the loop. 2000 phase->igvn().replace_input_of(iff, 1, bol); 2001 } 2002 } 2003 } 2004 } 2005 } 2006 } 2007 } 2008 2009 #ifdef ASSERT 2010 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) { 2011 const bool trace = false; 2012 ResourceMark rm; 2013 Unique_Node_List nodes; 2014 Unique_Node_List controls; 2015 Unique_Node_List memories; 2016 2017 nodes.push(root); 2018 for (uint next = 0; next < nodes.size(); next++) { 2019 Node *n = nodes.at(next); 2020 if (ShenandoahBarrierSetC2::is_shenandoah_lrb_call(n)) { 2021 controls.push(n); 2022 if (trace) { tty->print("XXXXXX verifying"); n->dump(); } 2023 for (uint next2 = 0; next2 < controls.size(); next2++) { 2024 Node *m = controls.at(next2); 2025 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2026 Node* u = m->fast_out(i); 2027 if (u->is_CFG() && !u->is_Root() && 2028 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) && 2029 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) { 2030 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); } 2031 controls.push(u); 2032 } 2033 } 2034 } 2035 memories.push(n->as_Call()->proj_out(TypeFunc::Memory)); 2036 for (uint next2 = 0; next2 < memories.size(); next2++) { 2037 Node *m = memories.at(next2); 2038 assert(m->bottom_type() == Type::MEMORY, ""); 2039 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2040 Node* u = m->fast_out(i); 2041 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) { 2042 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 2043 memories.push(u); 2044 } else if (u->is_LoadStore()) { 2045 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); } 2046 memories.push(u->find_out_with(Op_SCMemProj)); 2047 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) { 2048 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 2049 memories.push(u); 2050 } else if (u->is_Phi()) { 2051 assert(u->bottom_type() == Type::MEMORY, ""); 2052 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) { 2053 assert(controls.member(u->in(0)), ""); 2054 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 2055 memories.push(u); 2056 } 2057 } else if (u->is_SafePoint() || u->is_MemBar()) { 2058 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2059 Node* uu = u->fast_out(j); 2060 if (uu->bottom_type() == Type::MEMORY) { 2061 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); } 2062 memories.push(uu); 2063 } 2064 } 2065 } 2066 } 2067 } 2068 for (uint next2 = 0; next2 < controls.size(); next2++) { 2069 Node *m = controls.at(next2); 2070 if (m->is_Region()) { 2071 bool all_in = true; 2072 for (uint i = 1; i < m->req(); i++) { 2073 if (!controls.member(m->in(i))) { 2074 all_in = false; 2075 break; 2076 } 2077 } 2078 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); } 2079 bool found_phi = false; 2080 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) { 2081 Node* u = m->fast_out(j); 2082 if (u->is_Phi() && memories.member(u)) { 2083 found_phi = true; 2084 for (uint i = 1; i < u->req() && found_phi; i++) { 2085 Node* k = u->in(i); 2086 if (memories.member(k) != controls.member(m->in(i))) { 2087 found_phi = false; 2088 } 2089 } 2090 } 2091 } 2092 assert(found_phi || all_in, ""); 2093 } 2094 } 2095 controls.clear(); 2096 memories.clear(); 2097 } 2098 for( uint i = 0; i < n->len(); ++i ) { 2099 Node *m = n->in(i); 2100 if (m != NULL) { 2101 nodes.push(m); 2102 } 2103 } 2104 } 2105 } 2106 #endif 2107 2108 ShenandoahEnqueueBarrierNode::ShenandoahEnqueueBarrierNode(Node* val) : Node(NULL, val) { 2109 ShenandoahBarrierSetC2::bsc2()->state()->add_enqueue_barrier(this); 2110 } 2111 2112 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const { 2113 if (in(1) == NULL || in(1)->is_top()) { 2114 return Type::TOP; 2115 } 2116 const Type* t = in(1)->bottom_type(); 2117 if (t == TypePtr::NULL_PTR) { 2118 return t; 2119 } 2120 return t->is_oopptr(); 2121 } 2122 2123 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const { 2124 if (in(1) == NULL) { 2125 return Type::TOP; 2126 } 2127 const Type* t = phase->type(in(1)); 2128 if (t == Type::TOP) { 2129 return Type::TOP; 2130 } 2131 if (t == TypePtr::NULL_PTR) { 2132 return t; 2133 } 2134 return t->is_oopptr(); 2135 } 2136 2137 int ShenandoahEnqueueBarrierNode::needed(Node* n) { 2138 if (n == NULL || 2139 n->is_Allocate() || 2140 n->Opcode() == Op_ShenandoahEnqueueBarrier || 2141 n->bottom_type() == TypePtr::NULL_PTR || 2142 (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) { 2143 return NotNeeded; 2144 } 2145 if (n->is_Phi() || 2146 n->is_CMove()) { 2147 return MaybeNeeded; 2148 } 2149 return Needed; 2150 } 2151 2152 Node* ShenandoahEnqueueBarrierNode::next(Node* n) { 2153 for (;;) { 2154 if (n == NULL) { 2155 return n; 2156 } else if (n->bottom_type() == TypePtr::NULL_PTR) { 2157 return n; 2158 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) { 2159 return n; 2160 } else if (n->is_ConstraintCast() || 2161 n->Opcode() == Op_DecodeN || 2162 n->Opcode() == Op_EncodeP) { 2163 n = n->in(1); 2164 } else if (n->is_Proj()) { 2165 n = n->in(0); 2166 } else { 2167 return n; 2168 } 2169 } 2170 ShouldNotReachHere(); 2171 return NULL; 2172 } 2173 2174 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) { 2175 PhaseIterGVN* igvn = phase->is_IterGVN(); 2176 2177 Node* n = next(in(1)); 2178 2179 int cont = needed(n); 2180 2181 if (cont == NotNeeded) { 2182 return in(1); 2183 } else if (cont == MaybeNeeded) { 2184 if (igvn == NULL) { 2185 phase->record_for_igvn(this); 2186 return this; 2187 } else { 2188 ResourceMark rm; 2189 Unique_Node_List wq; 2190 uint wq_i = 0; 2191 2192 for (;;) { 2193 if (n->is_Phi()) { 2194 for (uint i = 1; i < n->req(); i++) { 2195 Node* m = n->in(i); 2196 if (m != NULL) { 2197 wq.push(m); 2198 } 2199 } 2200 } else { 2201 assert(n->is_CMove(), "nothing else here"); 2202 Node* m = n->in(CMoveNode::IfFalse); 2203 wq.push(m); 2204 m = n->in(CMoveNode::IfTrue); 2205 wq.push(m); 2206 } 2207 Node* orig_n = NULL; 2208 do { 2209 if (wq_i >= wq.size()) { 2210 return in(1); 2211 } 2212 n = wq.at(wq_i); 2213 wq_i++; 2214 orig_n = n; 2215 n = next(n); 2216 cont = needed(n); 2217 if (cont == Needed) { 2218 return this; 2219 } 2220 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n))); 2221 } 2222 } 2223 } 2224 2225 return this; 2226 } 2227 2228 #ifdef ASSERT 2229 static bool has_never_branch(Node* root) { 2230 for (uint i = 1; i < root->req(); i++) { 2231 Node* in = root->in(i); 2232 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) { 2233 return true; 2234 } 2235 } 2236 return false; 2237 } 2238 #endif 2239 2240 void MemoryGraphFixer::collect_memory_nodes() { 2241 Node_Stack stack(0); 2242 VectorSet visited(Thread::current()->resource_area()); 2243 Node_List regions; 2244 2245 // Walk the raw memory graph and create a mapping from CFG node to 2246 // memory node. Exclude phis for now. 2247 stack.push(_phase->C->root(), 1); 2248 do { 2249 Node* n = stack.node(); 2250 int opc = n->Opcode(); 2251 uint i = stack.index(); 2252 if (i < n->req()) { 2253 Node* mem = NULL; 2254 if (opc == Op_Root) { 2255 Node* in = n->in(i); 2256 int in_opc = in->Opcode(); 2257 if (in_opc == Op_Return || in_opc == Op_Rethrow) { 2258 mem = in->in(TypeFunc::Memory); 2259 } else if (in_opc == Op_Halt) { 2260 if (!in->in(0)->is_Region()) { 2261 Node* proj = in->in(0); 2262 assert(proj->is_Proj(), ""); 2263 Node* in = proj->in(0); 2264 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), ""); 2265 if (in->is_CallStaticJava()) { 2266 mem = in->in(TypeFunc::Memory); 2267 } else if (in->Opcode() == Op_Catch) { 2268 Node* call = in->in(0)->in(0); 2269 assert(call->is_Call(), ""); 2270 mem = call->in(TypeFunc::Memory); 2271 } else if (in->Opcode() == Op_NeverBranch) { 2272 ResourceMark rm; 2273 Unique_Node_List wq; 2274 wq.push(in); 2275 wq.push(in->as_Multi()->proj_out(0)); 2276 for (uint j = 1; j < wq.size(); j++) { 2277 Node* c = wq.at(j); 2278 assert(!c->is_Root(), "shouldn't leave loop"); 2279 if (c->is_SafePoint()) { 2280 assert(mem == NULL, "only one safepoint"); 2281 mem = c->in(TypeFunc::Memory); 2282 } 2283 for (DUIterator_Fast kmax, k = c->fast_outs(kmax); k < kmax; k++) { 2284 Node* u = c->fast_out(k); 2285 if (u->is_CFG()) { 2286 wq.push(u); 2287 } 2288 } 2289 } 2290 assert(mem != NULL, "should have found safepoint"); 2291 } 2292 } 2293 } else { 2294 #ifdef ASSERT 2295 n->dump(); 2296 in->dump(); 2297 #endif 2298 ShouldNotReachHere(); 2299 } 2300 } else { 2301 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, ""); 2302 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, ""); 2303 mem = n->in(i); 2304 } 2305 i++; 2306 stack.set_index(i); 2307 if (mem == NULL) { 2308 continue; 2309 } 2310 for (;;) { 2311 if (visited.test_set(mem->_idx) || mem->is_Start()) { 2312 break; 2313 } 2314 if (mem->is_Phi()) { 2315 stack.push(mem, 2); 2316 mem = mem->in(1); 2317 } else if (mem->is_Proj()) { 2318 stack.push(mem, mem->req()); 2319 mem = mem->in(0); 2320 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 2321 mem = mem->in(TypeFunc::Memory); 2322 } else if (mem->is_MergeMem()) { 2323 MergeMemNode* mm = mem->as_MergeMem(); 2324 mem = mm->memory_at(_alias); 2325 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 2326 assert(_alias == Compile::AliasIdxRaw, ""); 2327 stack.push(mem, mem->req()); 2328 mem = mem->in(MemNode::Memory); 2329 } else { 2330 #ifdef ASSERT 2331 mem->dump(); 2332 #endif 2333 ShouldNotReachHere(); 2334 } 2335 } 2336 } else { 2337 if (n->is_Phi()) { 2338 // Nothing 2339 } else if (!n->is_Root()) { 2340 Node* c = get_ctrl(n); 2341 _memory_nodes.map(c->_idx, n); 2342 } 2343 stack.pop(); 2344 } 2345 } while(stack.is_nonempty()); 2346 2347 // Iterate over CFG nodes in rpo and propagate memory state to 2348 // compute memory state at regions, creating new phis if needed. 2349 Node_List rpo_list; 2350 visited.Clear(); 2351 _phase->rpo(_phase->C->root(), stack, visited, rpo_list); 2352 Node* root = rpo_list.pop(); 2353 assert(root == _phase->C->root(), ""); 2354 2355 const bool trace = false; 2356 #ifdef ASSERT 2357 if (trace) { 2358 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2359 Node* c = rpo_list.at(i); 2360 if (_memory_nodes[c->_idx] != NULL) { 2361 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); 2362 } 2363 } 2364 } 2365 #endif 2366 uint last = _phase->C->unique(); 2367 2368 #ifdef ASSERT 2369 uint8_t max_depth = 0; 2370 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) { 2371 IdealLoopTree* lpt = iter.current(); 2372 max_depth = MAX2(max_depth, lpt->_nest); 2373 } 2374 #endif 2375 2376 bool progress = true; 2377 int iteration = 0; 2378 Node_List dead_phis; 2379 while (progress) { 2380 progress = false; 2381 iteration++; 2382 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), ""); 2383 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); } 2384 IdealLoopTree* last_updated_ilt = NULL; 2385 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2386 Node* c = rpo_list.at(i); 2387 2388 Node* prev_mem = _memory_nodes[c->_idx]; 2389 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2390 Node* prev_region = regions[c->_idx]; 2391 Node* unique = NULL; 2392 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) { 2393 Node* m = _memory_nodes[c->in(j)->_idx]; 2394 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"); 2395 if (m != NULL) { 2396 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) { 2397 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), ""); 2398 // continue 2399 } else if (unique == NULL) { 2400 unique = m; 2401 } else if (m == unique) { 2402 // continue 2403 } else { 2404 unique = NodeSentinel; 2405 } 2406 } 2407 } 2408 assert(unique != NULL, "empty phi???"); 2409 if (unique != NodeSentinel) { 2410 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) { 2411 dead_phis.push(prev_region); 2412 } 2413 regions.map(c->_idx, unique); 2414 } else { 2415 Node* phi = NULL; 2416 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) { 2417 phi = prev_region; 2418 for (uint k = 1; k < c->req(); k++) { 2419 Node* m = _memory_nodes[c->in(k)->_idx]; 2420 assert(m != NULL, "expect memory state"); 2421 phi->set_req(k, m); 2422 } 2423 } else { 2424 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) { 2425 Node* u = c->fast_out(j); 2426 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 2427 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) { 2428 phi = u; 2429 for (uint k = 1; k < c->req() && phi != NULL; k++) { 2430 Node* m = _memory_nodes[c->in(k)->_idx]; 2431 assert(m != NULL, "expect memory state"); 2432 if (u->in(k) != m) { 2433 phi = NULL; 2434 } 2435 } 2436 } 2437 } 2438 if (phi == NULL) { 2439 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias)); 2440 for (uint k = 1; k < c->req(); k++) { 2441 Node* m = _memory_nodes[c->in(k)->_idx]; 2442 assert(m != NULL, "expect memory state"); 2443 phi->init_req(k, m); 2444 } 2445 } 2446 } 2447 assert(phi != NULL, ""); 2448 regions.map(c->_idx, phi); 2449 } 2450 Node* current_region = regions[c->_idx]; 2451 if (current_region != prev_region) { 2452 progress = true; 2453 if (prev_region == prev_mem) { 2454 _memory_nodes.map(c->_idx, current_region); 2455 } 2456 } 2457 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) { 2458 Node* m = _memory_nodes[_phase->idom(c)->_idx]; 2459 assert(m != NULL, "expect memory state"); 2460 if (m != prev_mem) { 2461 _memory_nodes.map(c->_idx, m); 2462 progress = true; 2463 } 2464 } 2465 #ifdef ASSERT 2466 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); } 2467 #endif 2468 } 2469 } 2470 2471 // Replace existing phi with computed memory state for that region 2472 // if different (could be a new phi or a dominating memory node if 2473 // that phi was found to be useless). 2474 while (dead_phis.size() > 0) { 2475 Node* n = dead_phis.pop(); 2476 n->replace_by(_phase->C->top()); 2477 n->destruct(); 2478 } 2479 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2480 Node* c = rpo_list.at(i); 2481 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2482 Node* n = regions[c->_idx]; 2483 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) { 2484 _phase->register_new_node(n, c); 2485 } 2486 } 2487 } 2488 for (int i = rpo_list.size() - 1; i >= 0; i--) { 2489 Node* c = rpo_list.at(i); 2490 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 2491 Node* n = regions[c->_idx]; 2492 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 2493 Node* u = c->fast_out(i); 2494 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 2495 u != n) { 2496 if (u->adr_type() == TypePtr::BOTTOM) { 2497 fix_memory_uses(u, n, n, c); 2498 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 2499 _phase->lazy_replace(u, n); 2500 --i; --imax; 2501 } 2502 } 2503 } 2504 } 2505 } 2506 } 2507 2508 Node* MemoryGraphFixer::get_ctrl(Node* n) const { 2509 Node* c = _phase->get_ctrl(n); 2510 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) { 2511 assert(c == n->in(0), ""); 2512 CallNode* call = c->as_Call(); 2513 CallProjections projs; 2514 call->extract_projections(&projs, true, false); 2515 if (projs.catchall_memproj != NULL) { 2516 if (projs.fallthrough_memproj == n) { 2517 c = projs.fallthrough_catchproj; 2518 } else { 2519 assert(projs.catchall_memproj == n, ""); 2520 c = projs.catchall_catchproj; 2521 } 2522 } 2523 } 2524 return c; 2525 } 2526 2527 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const { 2528 if (_phase->has_ctrl(n)) 2529 return get_ctrl(n); 2530 else { 2531 assert (n->is_CFG(), "must be a CFG node"); 2532 return n; 2533 } 2534 } 2535 2536 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const { 2537 return m != NULL && get_ctrl(m) == c; 2538 } 2539 2540 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const { 2541 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, ""); 2542 Node* mem = _memory_nodes[ctrl->_idx]; 2543 Node* c = ctrl; 2544 while (!mem_is_valid(mem, c) && 2545 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) { 2546 c = _phase->idom(c); 2547 mem = _memory_nodes[c->_idx]; 2548 } 2549 if (n != NULL && mem_is_valid(mem, c)) { 2550 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) { 2551 mem = next_mem(mem, _alias); 2552 } 2553 if (mem->is_MergeMem()) { 2554 mem = mem->as_MergeMem()->memory_at(_alias); 2555 } 2556 if (!mem_is_valid(mem, c)) { 2557 do { 2558 c = _phase->idom(c); 2559 mem = _memory_nodes[c->_idx]; 2560 } while (!mem_is_valid(mem, c) && 2561 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))); 2562 } 2563 } 2564 assert(mem->bottom_type() == Type::MEMORY, ""); 2565 return mem; 2566 } 2567 2568 bool MemoryGraphFixer::has_mem_phi(Node* region) const { 2569 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 2570 Node* use = region->fast_out(i); 2571 if (use->is_Phi() && use->bottom_type() == Type::MEMORY && 2572 (_phase->C->get_alias_index(use->adr_type()) == _alias)) { 2573 return true; 2574 } 2575 } 2576 return false; 2577 } 2578 2579 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) { 2580 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, ""); 2581 const bool trace = false; 2582 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); }); 2583 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); }); 2584 GrowableArray<Node*> phis; 2585 if (mem_for_ctrl != mem) { 2586 Node* old = mem_for_ctrl; 2587 Node* prev = NULL; 2588 while (old != mem) { 2589 prev = old; 2590 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) { 2591 assert(_alias == Compile::AliasIdxRaw, ""); 2592 old = old->in(MemNode::Memory); 2593 } else if (old->Opcode() == Op_SCMemProj) { 2594 assert(_alias == Compile::AliasIdxRaw, ""); 2595 old = old->in(0); 2596 } else { 2597 ShouldNotReachHere(); 2598 } 2599 } 2600 assert(prev != NULL, ""); 2601 if (new_ctrl != ctrl) { 2602 _memory_nodes.map(ctrl->_idx, mem); 2603 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl); 2604 } 2605 uint input = (uint)MemNode::Memory; 2606 _phase->igvn().replace_input_of(prev, input, new_mem); 2607 } else { 2608 uses.clear(); 2609 _memory_nodes.map(new_ctrl->_idx, new_mem); 2610 uses.push(new_ctrl); 2611 for(uint next = 0; next < uses.size(); next++ ) { 2612 Node *n = uses.at(next); 2613 assert(n->is_CFG(), ""); 2614 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); }); 2615 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2616 Node* u = n->fast_out(i); 2617 if (!u->is_Root() && u->is_CFG() && u != n) { 2618 Node* m = _memory_nodes[u->_idx]; 2619 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) && 2620 !has_mem_phi(u) && 2621 u->unique_ctrl_out()->Opcode() != Op_Halt) { 2622 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); }); 2623 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); }); 2624 2625 if (!mem_is_valid(m, u) || !m->is_Phi()) { 2626 bool push = true; 2627 bool create_phi = true; 2628 if (_phase->is_dominator(new_ctrl, u)) { 2629 create_phi = false; 2630 } else if (!_phase->C->has_irreducible_loop()) { 2631 IdealLoopTree* loop = _phase->get_loop(ctrl); 2632 bool do_check = true; 2633 IdealLoopTree* l = loop; 2634 create_phi = false; 2635 while (l != _phase->ltree_root()) { 2636 Node* head = l->_head; 2637 if (head->in(0) == NULL) { 2638 head = _phase->get_ctrl(head); 2639 } 2640 if (_phase->is_dominator(head, u) && _phase->is_dominator(_phase->idom(u), head)) { 2641 create_phi = true; 2642 do_check = false; 2643 break; 2644 } 2645 l = l->_parent; 2646 } 2647 2648 if (do_check) { 2649 assert(!create_phi, ""); 2650 IdealLoopTree* u_loop = _phase->get_loop(u); 2651 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) { 2652 Node* c = ctrl; 2653 while (!_phase->is_dominator(c, u_loop->tail())) { 2654 c = _phase->idom(c); 2655 } 2656 if (!_phase->is_dominator(c, u)) { 2657 do_check = false; 2658 } 2659 } 2660 } 2661 2662 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) { 2663 create_phi = true; 2664 } 2665 } 2666 if (create_phi) { 2667 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias)); 2668 _phase->register_new_node(phi, u); 2669 phis.push(phi); 2670 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); }); 2671 if (!mem_is_valid(m, u)) { 2672 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); }); 2673 _memory_nodes.map(u->_idx, phi); 2674 } else { 2675 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); }); 2676 for (;;) { 2677 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), ""); 2678 Node* next = NULL; 2679 if (m->is_Proj()) { 2680 next = m->in(0); 2681 } else { 2682 assert(m->is_Mem() || m->is_LoadStore(), ""); 2683 assert(_alias == Compile::AliasIdxRaw, ""); 2684 next = m->in(MemNode::Memory); 2685 } 2686 if (_phase->get_ctrl(next) != u) { 2687 break; 2688 } 2689 if (next->is_MergeMem()) { 2690 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, ""); 2691 break; 2692 } 2693 if (next->is_Phi()) { 2694 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, ""); 2695 break; 2696 } 2697 m = next; 2698 } 2699 2700 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); }); 2701 assert(m->is_Mem() || m->is_LoadStore(), ""); 2702 uint input = (uint)MemNode::Memory; 2703 _phase->igvn().replace_input_of(m, input, phi); 2704 push = false; 2705 } 2706 } else { 2707 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); }); 2708 } 2709 if (push) { 2710 uses.push(u); 2711 } 2712 } 2713 } else if (!mem_is_valid(m, u) && 2714 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) { 2715 uses.push(u); 2716 } 2717 } 2718 } 2719 } 2720 for (int i = 0; i < phis.length(); i++) { 2721 Node* n = phis.at(i); 2722 Node* r = n->in(0); 2723 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); }); 2724 for (uint j = 1; j < n->req(); j++) { 2725 Node* m = find_mem(r->in(j), NULL); 2726 _phase->igvn().replace_input_of(n, j, m); 2727 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); }); 2728 } 2729 } 2730 } 2731 uint last = _phase->C->unique(); 2732 MergeMemNode* mm = NULL; 2733 int alias = _alias; 2734 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); }); 2735 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 2736 Node* u = mem->out(i); 2737 if (u->_idx < last) { 2738 if (u->is_Mem()) { 2739 if (_phase->C->get_alias_index(u->adr_type()) == alias) { 2740 Node* m = find_mem(_phase->get_ctrl(u), u); 2741 if (m != mem) { 2742 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2743 _phase->igvn().replace_input_of(u, MemNode::Memory, m); 2744 --i; 2745 } 2746 } 2747 } else if (u->is_MergeMem()) { 2748 MergeMemNode* u_mm = u->as_MergeMem(); 2749 if (u_mm->memory_at(alias) == mem) { 2750 MergeMemNode* newmm = NULL; 2751 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2752 Node* uu = u->fast_out(j); 2753 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 2754 if (uu->is_Phi()) { 2755 assert(uu->adr_type() == TypePtr::BOTTOM, ""); 2756 Node* region = uu->in(0); 2757 int nb = 0; 2758 for (uint k = 1; k < uu->req(); k++) { 2759 if (uu->in(k) == u) { 2760 Node* m = find_mem(region->in(k), NULL); 2761 if (m != mem) { 2762 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); }); 2763 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 2764 if (newmm != u) { 2765 _phase->igvn().replace_input_of(uu, k, newmm); 2766 nb++; 2767 --jmax; 2768 } 2769 } 2770 } 2771 } 2772 if (nb > 0) { 2773 --j; 2774 } 2775 } else { 2776 Node* m = find_mem(_phase->ctrl_or_self(uu), uu); 2777 if (m != mem) { 2778 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); }); 2779 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 2780 if (newmm != u) { 2781 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 2782 --j, --jmax; 2783 } 2784 } 2785 } 2786 } 2787 } 2788 } else if (u->is_Phi()) { 2789 assert(u->bottom_type() == Type::MEMORY, "what else?"); 2790 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) { 2791 Node* region = u->in(0); 2792 bool replaced = false; 2793 for (uint j = 1; j < u->req(); j++) { 2794 if (u->in(j) == mem) { 2795 Node* m = find_mem(region->in(j), NULL); 2796 Node* nnew = m; 2797 if (m != mem) { 2798 if (u->adr_type() == TypePtr::BOTTOM) { 2799 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m)); 2800 nnew = mm; 2801 } 2802 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); }); 2803 _phase->igvn().replace_input_of(u, j, nnew); 2804 replaced = true; 2805 } 2806 } 2807 } 2808 if (replaced) { 2809 --i; 2810 } 2811 } 2812 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 2813 u->adr_type() == NULL) { 2814 assert(u->adr_type() != NULL || 2815 u->Opcode() == Op_Rethrow || 2816 u->Opcode() == Op_Return || 2817 u->Opcode() == Op_SafePoint || 2818 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 2819 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 2820 u->Opcode() == Op_CallLeaf, ""); 2821 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2822 if (m != mem) { 2823 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m)); 2824 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 2825 --i; 2826 } 2827 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) { 2828 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2829 if (m != mem) { 2830 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2831 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 2832 --i; 2833 } 2834 } else if (u->adr_type() != TypePtr::BOTTOM && 2835 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) { 2836 Node* m = find_mem(_phase->ctrl_or_self(u), u); 2837 assert(m != mem, ""); 2838 // u is on the wrong slice... 2839 assert(u->is_ClearArray(), ""); 2840 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 2841 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 2842 --i; 2843 } 2844 } 2845 } 2846 #ifdef ASSERT 2847 assert(new_mem->outcnt() > 0, ""); 2848 for (int i = 0; i < phis.length(); i++) { 2849 Node* n = phis.at(i); 2850 assert(n->outcnt() > 0, "new phi must have uses now"); 2851 } 2852 #endif 2853 } 2854 2855 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const { 2856 MergeMemNode* mm = MergeMemNode::make(mem); 2857 mm->set_memory_at(_alias, rep_proj); 2858 _phase->register_new_node(mm, rep_ctrl); 2859 return mm; 2860 } 2861 2862 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const { 2863 MergeMemNode* newmm = NULL; 2864 MergeMemNode* u_mm = u->as_MergeMem(); 2865 Node* c = _phase->get_ctrl(u); 2866 if (_phase->is_dominator(c, rep_ctrl)) { 2867 c = rep_ctrl; 2868 } else { 2869 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other"); 2870 } 2871 if (u->outcnt() == 1) { 2872 if (u->req() > (uint)_alias && u->in(_alias) == mem) { 2873 _phase->igvn().replace_input_of(u, _alias, rep_proj); 2874 --i; 2875 } else { 2876 _phase->igvn().rehash_node_delayed(u); 2877 u_mm->set_memory_at(_alias, rep_proj); 2878 } 2879 newmm = u_mm; 2880 _phase->set_ctrl_and_loop(u, c); 2881 } else { 2882 // can't simply clone u and then change one of its input because 2883 // it adds and then removes an edge which messes with the 2884 // DUIterator 2885 newmm = MergeMemNode::make(u_mm->base_memory()); 2886 for (uint j = 0; j < u->req(); j++) { 2887 if (j < newmm->req()) { 2888 if (j == (uint)_alias) { 2889 newmm->set_req(j, rep_proj); 2890 } else if (newmm->in(j) != u->in(j)) { 2891 newmm->set_req(j, u->in(j)); 2892 } 2893 } else if (j == (uint)_alias) { 2894 newmm->add_req(rep_proj); 2895 } else { 2896 newmm->add_req(u->in(j)); 2897 } 2898 } 2899 if ((uint)_alias >= u->req()) { 2900 newmm->set_memory_at(_alias, rep_proj); 2901 } 2902 _phase->register_new_node(newmm, c); 2903 } 2904 return newmm; 2905 } 2906 2907 bool MemoryGraphFixer::should_process_phi(Node* phi) const { 2908 if (phi->adr_type() == TypePtr::BOTTOM) { 2909 Node* region = phi->in(0); 2910 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) { 2911 Node* uu = region->fast_out(j); 2912 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) { 2913 return false; 2914 } 2915 } 2916 return true; 2917 } 2918 return _phase->C->get_alias_index(phi->adr_type()) == _alias; 2919 } 2920 2921 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const { 2922 uint last = _phase-> C->unique(); 2923 MergeMemNode* mm = NULL; 2924 assert(mem->bottom_type() == Type::MEMORY, ""); 2925 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 2926 Node* u = mem->out(i); 2927 if (u != replacement && u->_idx < last) { 2928 if (u->is_MergeMem()) { 2929 MergeMemNode* u_mm = u->as_MergeMem(); 2930 if (u_mm->memory_at(_alias) == mem) { 2931 MergeMemNode* newmm = NULL; 2932 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 2933 Node* uu = u->fast_out(j); 2934 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 2935 if (uu->is_Phi()) { 2936 if (should_process_phi(uu)) { 2937 Node* region = uu->in(0); 2938 int nb = 0; 2939 for (uint k = 1; k < uu->req(); k++) { 2940 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) { 2941 if (newmm == NULL) { 2942 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 2943 } 2944 if (newmm != u) { 2945 _phase->igvn().replace_input_of(uu, k, newmm); 2946 nb++; 2947 --jmax; 2948 } 2949 } 2950 } 2951 if (nb > 0) { 2952 --j; 2953 } 2954 } 2955 } else { 2956 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) { 2957 if (newmm == NULL) { 2958 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 2959 } 2960 if (newmm != u) { 2961 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 2962 --j, --jmax; 2963 } 2964 } 2965 } 2966 } 2967 } 2968 } else if (u->is_Phi()) { 2969 assert(u->bottom_type() == Type::MEMORY, "what else?"); 2970 Node* region = u->in(0); 2971 if (should_process_phi(u)) { 2972 bool replaced = false; 2973 for (uint j = 1; j < u->req(); j++) { 2974 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) { 2975 Node* nnew = rep_proj; 2976 if (u->adr_type() == TypePtr::BOTTOM) { 2977 if (mm == NULL) { 2978 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 2979 } 2980 nnew = mm; 2981 } 2982 _phase->igvn().replace_input_of(u, j, nnew); 2983 replaced = true; 2984 } 2985 } 2986 if (replaced) { 2987 --i; 2988 } 2989 2990 } 2991 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 2992 u->adr_type() == NULL) { 2993 assert(u->adr_type() != NULL || 2994 u->Opcode() == Op_Rethrow || 2995 u->Opcode() == Op_Return || 2996 u->Opcode() == Op_SafePoint || 2997 u->Opcode() == Op_StoreIConditional || 2998 u->Opcode() == Op_StoreLConditional || 2999 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 3000 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 3001 u->Opcode() == Op_CallLeaf, "%s", u->Name()); 3002 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 3003 if (mm == NULL) { 3004 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 3005 } 3006 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 3007 --i; 3008 } 3009 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 3010 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 3011 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj); 3012 --i; 3013 } 3014 } 3015 } 3016 } 3017 } 3018 3019 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj) 3020 : Node(ctrl, obj) { 3021 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this); 3022 } 3023 3024 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const { 3025 if (in(ValueIn) == NULL || in(ValueIn)->is_top()) { 3026 return Type::TOP; 3027 } 3028 const Type* t = in(ValueIn)->bottom_type(); 3029 if (t == TypePtr::NULL_PTR) { 3030 return t; 3031 } 3032 return t->is_oopptr(); 3033 } 3034 3035 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const { 3036 // Either input is TOP ==> the result is TOP 3037 const Type *t2 = phase->type(in(ValueIn)); 3038 if( t2 == Type::TOP ) return Type::TOP; 3039 3040 if (t2 == TypePtr::NULL_PTR) { 3041 return t2; 3042 } 3043 3044 const Type* type = t2->is_oopptr(); 3045 return type; 3046 } 3047 3048 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) { 3049 Node* value = in(ValueIn); 3050 if (!needs_barrier(phase, value)) { 3051 return value; 3052 } 3053 return this; 3054 } 3055 3056 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) { 3057 Unique_Node_List visited; 3058 return needs_barrier_impl(phase, n, visited); 3059 } 3060 3061 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) { 3062 if (n == NULL) return false; 3063 if (visited.member(n)) { 3064 return false; // Been there. 3065 } 3066 visited.push(n); 3067 3068 if (n->is_Allocate()) { 3069 // tty->print_cr("optimize barrier on alloc"); 3070 return false; 3071 } 3072 if (n->is_Call()) { 3073 // tty->print_cr("optimize barrier on call"); 3074 return false; 3075 } 3076 3077 const Type* type = phase->type(n); 3078 if (type == Type::TOP) { 3079 return false; 3080 } 3081 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) { 3082 // tty->print_cr("optimize barrier on null"); 3083 return false; 3084 } 3085 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) { 3086 // tty->print_cr("optimize barrier on constant"); 3087 return false; 3088 } 3089 3090 switch (n->Opcode()) { 3091 case Op_AddP: 3092 return true; // TODO: Can refine? 3093 case Op_LoadP: 3094 case Op_ShenandoahCompareAndExchangeN: 3095 case Op_ShenandoahCompareAndExchangeP: 3096 case Op_CompareAndExchangeN: 3097 case Op_CompareAndExchangeP: 3098 case Op_GetAndSetN: 3099 case Op_GetAndSetP: 3100 return true; 3101 case Op_Phi: { 3102 for (uint i = 1; i < n->req(); i++) { 3103 if (needs_barrier_impl(phase, n->in(i), visited)) return true; 3104 } 3105 return false; 3106 } 3107 case Op_CheckCastPP: 3108 case Op_CastPP: 3109 return needs_barrier_impl(phase, n->in(1), visited); 3110 case Op_Proj: 3111 return needs_barrier_impl(phase, n->in(0), visited); 3112 case Op_ShenandoahLoadReferenceBarrier: 3113 // tty->print_cr("optimize barrier on barrier"); 3114 return false; 3115 case Op_Parm: 3116 // tty->print_cr("optimize barrier on input arg"); 3117 return false; 3118 case Op_DecodeN: 3119 case Op_EncodeP: 3120 return needs_barrier_impl(phase, n->in(1), visited); 3121 case Op_LoadN: 3122 return true; 3123 case Op_CMoveN: 3124 case Op_CMoveP: 3125 return needs_barrier_impl(phase, n->in(2), visited) || 3126 needs_barrier_impl(phase, n->in(3), visited); 3127 case Op_ShenandoahEnqueueBarrier: 3128 return needs_barrier_impl(phase, n->in(1), visited); 3129 case Op_CreateEx: 3130 return false; 3131 default: 3132 break; 3133 } 3134 #ifdef ASSERT 3135 tty->print("need barrier on?: "); 3136 tty->print_cr("ins:"); 3137 n->dump(2); 3138 tty->print_cr("outs:"); 3139 n->dump(-2); 3140 ShouldNotReachHere(); 3141 #endif 3142 return true; 3143 } 3144 3145 ShenandoahLoadReferenceBarrierNode::Strength ShenandoahLoadReferenceBarrierNode::get_barrier_strength() { 3146 Unique_Node_List visited; 3147 Node_Stack stack(0); 3148 stack.push(this, 0); 3149 3150 // Look for strongest strength: go over nodes looking for STRONG ones. 3151 // Stop once we encountered STRONG. Otherwise, walk until we ran out of nodes, 3152 // and then the overall strength is NONE. 3153 Strength strength = NONE; 3154 while (strength != STRONG && stack.size() > 0) { 3155 Node* n = stack.node(); 3156 if (visited.member(n)) { 3157 stack.pop(); 3158 continue; 3159 } 3160 visited.push(n); 3161 bool visit_users = false; 3162 switch (n->Opcode()) { 3163 case Op_CallStaticJava: 3164 case Op_CallDynamicJava: 3165 case Op_CallLeaf: 3166 case Op_CallLeafNoFP: 3167 case Op_CompareAndSwapL: 3168 case Op_CompareAndSwapI: 3169 case Op_CompareAndSwapB: 3170 case Op_CompareAndSwapS: 3171 case Op_CompareAndSwapN: 3172 case Op_CompareAndSwapP: 3173 case Op_CompareAndExchangeL: 3174 case Op_CompareAndExchangeI: 3175 case Op_CompareAndExchangeB: 3176 case Op_CompareAndExchangeS: 3177 case Op_CompareAndExchangeN: 3178 case Op_CompareAndExchangeP: 3179 case Op_WeakCompareAndSwapL: 3180 case Op_WeakCompareAndSwapI: 3181 case Op_WeakCompareAndSwapB: 3182 case Op_WeakCompareAndSwapS: 3183 case Op_WeakCompareAndSwapN: 3184 case Op_WeakCompareAndSwapP: 3185 case Op_ShenandoahCompareAndSwapN: 3186 case Op_ShenandoahCompareAndSwapP: 3187 case Op_ShenandoahWeakCompareAndSwapN: 3188 case Op_ShenandoahWeakCompareAndSwapP: 3189 case Op_ShenandoahCompareAndExchangeN: 3190 case Op_ShenandoahCompareAndExchangeP: 3191 case Op_GetAndSetL: 3192 case Op_GetAndSetI: 3193 case Op_GetAndSetB: 3194 case Op_GetAndSetS: 3195 case Op_GetAndSetP: 3196 case Op_GetAndSetN: 3197 case Op_GetAndAddL: 3198 case Op_GetAndAddI: 3199 case Op_GetAndAddB: 3200 case Op_GetAndAddS: 3201 case Op_ShenandoahEnqueueBarrier: 3202 case Op_FastLock: 3203 case Op_FastUnlock: 3204 case Op_Rethrow: 3205 case Op_Return: 3206 case Op_StoreB: 3207 case Op_StoreC: 3208 case Op_StoreD: 3209 case Op_StoreF: 3210 case Op_StoreL: 3211 case Op_StoreLConditional: 3212 case Op_StoreI: 3213 case Op_StoreIConditional: 3214 case Op_StoreN: 3215 case Op_StoreP: 3216 case Op_StoreVector: 3217 case Op_StrInflatedCopy: 3218 case Op_StrCompressedCopy: 3219 case Op_EncodeP: 3220 case Op_CastP2X: 3221 case Op_SafePoint: 3222 case Op_EncodeISOArray: 3223 case Op_AryEq: 3224 case Op_StrEquals: 3225 case Op_StrComp: 3226 case Op_StrIndexOf: 3227 case Op_StrIndexOfChar: 3228 case Op_HasNegatives: 3229 // Known to require barriers 3230 strength = STRONG; 3231 break; 3232 case Op_CmpP: { 3233 if (n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) || 3234 n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) { 3235 // One of the sides is known null, no need for barrier. 3236 } else { 3237 strength = STRONG; 3238 } 3239 break; 3240 } 3241 case Op_LoadB: 3242 case Op_LoadUB: 3243 case Op_LoadUS: 3244 case Op_LoadD: 3245 case Op_LoadF: 3246 case Op_LoadL: 3247 case Op_LoadI: 3248 case Op_LoadS: 3249 case Op_LoadN: 3250 case Op_LoadP: 3251 case Op_LoadVector: { 3252 const TypePtr* adr_type = n->adr_type(); 3253 int alias_idx = Compile::current()->get_alias_index(adr_type); 3254 Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx); 3255 ciField* field = alias_type->field(); 3256 bool is_static = field != NULL && field->is_static(); 3257 bool is_final = field != NULL && field->is_final(); 3258 3259 if (ShenandoahOptimizeStaticFinals && is_static && is_final) { 3260 // Loading the constant does not require barriers: it should be handled 3261 // as part of GC roots already. 3262 } else { 3263 strength = STRONG; 3264 } 3265 break; 3266 } 3267 case Op_Conv2B: 3268 case Op_LoadRange: 3269 case Op_LoadKlass: 3270 case Op_LoadNKlass: 3271 // Do not require barriers 3272 break; 3273 case Op_AddP: 3274 case Op_CheckCastPP: 3275 case Op_CastPP: 3276 case Op_CMoveP: 3277 case Op_Phi: 3278 case Op_ShenandoahLoadReferenceBarrier: 3279 // Whether or not these need the barriers depends on their users 3280 visit_users = true; 3281 break; 3282 default: { 3283 #ifdef ASSERT 3284 fatal("Unknown node in get_barrier_strength: %s", NodeClassNames[n->Opcode()]); 3285 #else 3286 // Default to strong: better to have excess barriers, rather than miss some. 3287 strength = STRONG; 3288 #endif 3289 } 3290 } 3291 3292 stack.pop(); 3293 if (visit_users) { 3294 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3295 Node* user = n->fast_out(i); 3296 if (user != NULL) { 3297 stack.push(user, 0); 3298 } 3299 } 3300 } 3301 } 3302 return strength; 3303 } 3304 3305 CallStaticJavaNode* ShenandoahLoadReferenceBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) { 3306 Node* val = in(ValueIn); 3307 3308 const Type* val_t = igvn.type(val); 3309 3310 if (val_t->meet(TypePtr::NULL_PTR) != val_t && 3311 val->Opcode() == Op_CastPP && 3312 val->in(0) != NULL && 3313 val->in(0)->Opcode() == Op_IfTrue && 3314 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 3315 val->in(0)->in(0)->is_If() && 3316 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool && 3317 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne && 3318 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 3319 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) && 3320 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 3321 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), ""); 3322 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 3323 return unc; 3324 } 3325 return NULL; 3326 }