1 /* 2 * Copyright (c) 1998, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.inline.hpp" 27 #include "gc/shared/gc_globals.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "oops/compressedOops.hpp" 30 #include "opto/ad.hpp" 31 #include "opto/block.hpp" 32 #include "opto/c2compiler.hpp" 33 #include "opto/callnode.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/machnode.hpp" 36 #include "opto/runtime.hpp" 37 #include "opto/chaitin.hpp" 38 #include "runtime/os.inline.hpp" 39 #include "runtime/sharedRuntime.hpp" 40 41 // Optimization - Graph Style 42 43 // Check whether val is not-null-decoded compressed oop, 44 // i.e. will grab into the base of the heap if it represents null. 45 static bool accesses_heap_base_zone(Node *val) { 46 if (CompressedOops::base() != nullptr) { // Implies UseCompressedOops. 47 if (val && val->is_Mach()) { 48 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) { 49 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that 50 // decode null to point to the heap base (Decode_NN). 51 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) { 52 return true; 53 } 54 } 55 // Must recognize load operation with Decode matched in memory operand. 56 // We should not reach here except for PPC/AIX, as os::zero_page_read_protected() 57 // returns true everywhere else. On PPC, no such memory operands 58 // exist, therefore we did not yet implement a check for such operands. 59 NOT_AIX(Unimplemented()); 60 } 61 } 62 return false; 63 } 64 65 static bool needs_explicit_null_check_for_read(Node *val) { 66 // On some OSes (AIX) the page at address 0 is only write protected. 67 // If so, only Store operations will trap. 68 if (os::zero_page_read_protected()) { 69 return false; // Implicit null check will work. 70 } 71 // Also a read accessing the base of a heap-based compressed heap will trap. 72 if (accesses_heap_base_zone(val) && // Hits the base zone page. 73 CompressedOops::use_implicit_null_checks()) { // Base zone page is protected. 74 return false; 75 } 76 77 return true; 78 } 79 80 //------------------------------implicit_null_check---------------------------- 81 // Detect implicit-null-check opportunities. Basically, find null checks 82 // with suitable memory ops nearby. Use the memory op to do the null check. 83 // I can generate a memory op if there is not one nearby. 84 // The proj is the control projection for the not-null case. 85 // The val is the pointer being checked for nullness or 86 // decodeHeapOop_not_null node if it did not fold into address. 87 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { 88 // Assume if null check need for 0 offset then always needed 89 // Intel solaris doesn't support any null checks yet and no 90 // mechanism exists (yet) to set the switches at an os_cpu level 91 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 92 93 // Make sure the ptr-is-null path appears to be uncommon! 94 float f = block->end()->as_MachIf()->_prob; 95 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 96 if( f > PROB_UNLIKELY_MAG(4) ) return; 97 98 uint bidx = 0; // Capture index of value into memop 99 bool was_store; // Memory op is a store op 100 101 // Get the successor block for if the test ptr is non-null 102 Block* not_null_block; // this one goes with the proj 103 Block* null_block; 104 if (block->get_node(block->number_of_nodes()-1) == proj) { 105 null_block = block->_succs[0]; 106 not_null_block = block->_succs[1]; 107 } else { 108 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other"); 109 not_null_block = block->_succs[0]; 110 null_block = block->_succs[1]; 111 } 112 while (null_block->is_Empty() == Block::empty_with_goto) { 113 null_block = null_block->_succs[0]; 114 } 115 116 // Search the exception block for an uncommon trap. 117 // (See Parse::do_if and Parse::do_ifnull for the reason 118 // we need an uncommon trap. Briefly, we need a way to 119 // detect failure of this optimization, as in 6366351.) 120 { 121 bool found_trap = false; 122 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { 123 Node* nn = null_block->get_node(i1); 124 if (nn->is_MachCall() && 125 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 126 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 127 if (trtype->isa_int() && trtype->is_int()->is_con()) { 128 jint tr_con = trtype->is_int()->get_con(); 129 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 130 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 131 assert((int)reason < (int)BitsPerInt, "recode bit map"); 132 if (is_set_nth_bit(allowed_reasons, (int) reason) 133 && action != Deoptimization::Action_none) { 134 // This uncommon trap is sure to recompile, eventually. 135 // When that happens, C->too_many_traps will prevent 136 // this transformation from happening again. 137 found_trap = true; 138 } 139 } 140 break; 141 } 142 } 143 if (!found_trap) { 144 // We did not find an uncommon trap. 145 return; 146 } 147 } 148 149 // Check for decodeHeapOop_not_null node which did not fold into address 150 bool is_decoden = ((intptr_t)val) & 1; 151 val = (Node*)(((intptr_t)val) & ~1); 152 153 assert(!is_decoden || (val->in(0) == nullptr) && val->is_Mach() && 154 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 155 156 // Search the successor block for a load or store who's base value is also 157 // the tested value. There may be several. 158 MachNode *best = nullptr; // Best found so far 159 for (DUIterator i = val->outs(); val->has_out(i); i++) { 160 Node *m = val->out(i); 161 if( !m->is_Mach() ) continue; 162 MachNode *mach = m->as_Mach(); 163 was_store = false; 164 int iop = mach->ideal_Opcode(); 165 switch( iop ) { 166 case Op_LoadB: 167 case Op_LoadUB: 168 case Op_LoadUS: 169 case Op_LoadD: 170 case Op_LoadF: 171 case Op_LoadI: 172 case Op_LoadL: 173 case Op_LoadP: 174 case Op_LoadN: 175 case Op_LoadS: 176 case Op_LoadKlass: 177 case Op_LoadNKlass: 178 case Op_LoadRange: 179 case Op_LoadD_unaligned: 180 case Op_LoadL_unaligned: 181 assert(mach->in(2) == val, "should be address"); 182 break; 183 case Op_StoreB: 184 case Op_StoreC: 185 case Op_StoreCM: 186 case Op_StoreD: 187 case Op_StoreF: 188 case Op_StoreI: 189 case Op_StoreL: 190 case Op_StoreP: 191 case Op_StoreN: 192 case Op_StoreNKlass: 193 was_store = true; // Memory op is a store op 194 // Stores will have their address in slot 2 (memory in slot 1). 195 // If the value being nul-checked is in another slot, it means we 196 // are storing the checked value, which does NOT check the value! 197 if( mach->in(2) != val ) continue; 198 break; // Found a memory op? 199 case Op_StrComp: 200 case Op_StrEquals: 201 case Op_StrIndexOf: 202 case Op_StrIndexOfChar: 203 case Op_AryEq: 204 case Op_VectorizedHashCode: 205 case Op_StrInflatedCopy: 206 case Op_StrCompressedCopy: 207 case Op_EncodeISOArray: 208 case Op_CountPositives: 209 // Not a legit memory op for implicit null check regardless of 210 // embedded loads 211 continue; 212 default: // Also check for embedded loads 213 if( !mach->needs_anti_dependence_check() ) 214 continue; // Not an memory op; skip it 215 if( must_clone[iop] ) { 216 // Do not move nodes which produce flags because 217 // RA will try to clone it to place near branch and 218 // it will cause recompilation, see clone_node(). 219 continue; 220 } 221 { 222 // Check that value is used in memory address in 223 // instructions with embedded load (CmpP val1,(val2+off)). 224 Node* base; 225 Node* index; 226 const MachOper* oper = mach->memory_inputs(base, index); 227 if (oper == nullptr || oper == (MachOper*)-1) { 228 continue; // Not an memory op; skip it 229 } 230 if (val == base || 231 (val == index && val->bottom_type()->isa_narrowoop())) { 232 break; // Found it 233 } else { 234 continue; // Skip it 235 } 236 } 237 break; 238 } 239 240 // On some OSes (AIX) the page at address 0 is only write protected. 241 // If so, only Store operations will trap. 242 // But a read accessing the base of a heap-based compressed heap will trap. 243 if (!was_store && needs_explicit_null_check_for_read(val)) { 244 continue; 245 } 246 247 // Check that node's control edge is not-null block's head or dominates it, 248 // otherwise we can't hoist it because there are other control dependencies. 249 Node* ctrl = mach->in(0); 250 if (ctrl != nullptr && !(ctrl == not_null_block->head() || 251 get_block_for_node(ctrl)->dominates(not_null_block))) { 252 continue; 253 } 254 255 // check if the offset is not too high for implicit exception 256 { 257 intptr_t offset = 0; 258 const TypePtr *adr_type = nullptr; // Do not need this return value here 259 const Node* base = mach->get_base_and_disp(offset, adr_type); 260 if (base == nullptr || base == NodeSentinel) { 261 // Narrow oop address doesn't have base, only index. 262 // Give up if offset is beyond page size or if heap base is not protected. 263 if (val->bottom_type()->isa_narrowoop() && 264 (MacroAssembler::needs_explicit_null_check(offset) || 265 !CompressedOops::use_implicit_null_checks())) 266 continue; 267 // cannot reason about it; is probably not implicit null exception 268 } else { 269 const TypePtr* tptr; 270 if ((UseCompressedOops || UseCompressedClassPointers) && 271 (CompressedOops::shift() == 0 || CompressedKlassPointers::shift() == 0)) { 272 // 32-bits narrow oop can be the base of address expressions 273 tptr = base->get_ptr_type(); 274 } else { 275 // only regular oops are expected here 276 tptr = base->bottom_type()->is_ptr(); 277 } 278 // Give up if offset is not a compile-time constant. 279 if (offset == Type::OffsetBot || tptr->offset() == Type::OffsetBot) 280 continue; 281 offset += tptr->offset(); // correct if base is offsetted 282 // Give up if reference is beyond page size. 283 if (MacroAssembler::needs_explicit_null_check(offset)) 284 continue; 285 // Give up if base is a decode node and the heap base is not protected. 286 if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN && 287 !CompressedOops::use_implicit_null_checks()) 288 continue; 289 } 290 } 291 292 // Check ctrl input to see if the null-check dominates the memory op 293 Block *cb = get_block_for_node(mach); 294 cb = cb->_idom; // Always hoist at least 1 block 295 if( !was_store ) { // Stores can be hoisted only one block 296 while( cb->_dom_depth > (block->_dom_depth + 1)) 297 cb = cb->_idom; // Hoist loads as far as we want 298 // The non-null-block should dominate the memory op, too. Live 299 // range spilling will insert a spill in the non-null-block if it is 300 // needs to spill the memory op for an implicit null check. 301 if (cb->_dom_depth == (block->_dom_depth + 1)) { 302 if (cb != not_null_block) continue; 303 cb = cb->_idom; 304 } 305 } 306 if( cb != block ) continue; 307 308 // Found a memory user; see if it can be hoisted to check-block 309 uint vidx = 0; // Capture index of value into memop 310 uint j; 311 for( j = mach->req()-1; j > 0; j-- ) { 312 if( mach->in(j) == val ) { 313 vidx = j; 314 // Ignore DecodeN val which could be hoisted to where needed. 315 if( is_decoden ) continue; 316 } 317 // Block of memory-op input 318 Block* inb = get_block_for_node(mach->in(j)); 319 if (mach->in(j)->is_Con() && mach->in(j)->req() == 1 && inb == get_block_for_node(mach)) { 320 // Ignore constant loads scheduled in the same block (we can simply hoist them as well) 321 continue; 322 } 323 Block *b = block; // Start from nul check 324 while( b != inb && b->_dom_depth > inb->_dom_depth ) 325 b = b->_idom; // search upwards for input 326 // See if input dominates null check 327 if( b != inb ) 328 break; 329 } 330 if( j > 0 ) 331 continue; 332 Block *mb = get_block_for_node(mach); 333 // Hoisting stores requires more checks for the anti-dependence case. 334 // Give up hoisting if we have to move the store past any load. 335 if (was_store) { 336 // Make sure control does not do a merge (would have to check allpaths) 337 if (mb->num_preds() != 2) { 338 continue; 339 } 340 // mach is a store, hence block is the immediate dominator of mb. 341 // Due to the null-check shape of block (where its successors cannot re-join), 342 // block must be the direct predecessor of mb. 343 assert(get_block_for_node(mb->pred(1)) == block, "Unexpected predecessor block"); 344 uint k; 345 uint num_nodes = mb->number_of_nodes(); 346 for (k = 1; k < num_nodes; k++) { 347 Node *n = mb->get_node(k); 348 if (n->needs_anti_dependence_check() && 349 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory)) { 350 break; // Found anti-dependent load 351 } 352 } 353 if (k < num_nodes) { 354 continue; // Found anti-dependent load 355 } 356 } 357 358 // Make sure this memory op is not already being used for a NullCheck 359 Node *e = mb->end(); 360 if( e->is_MachNullCheck() && e->in(1) == mach ) 361 continue; // Already being used as a null check 362 363 // Found a candidate! Pick one with least dom depth - the highest 364 // in the dom tree should be closest to the null check. 365 if (best == nullptr || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 366 best = mach; 367 bidx = vidx; 368 } 369 } 370 // No candidate! 371 if (best == nullptr) { 372 return; 373 } 374 375 // ---- Found an implicit null check 376 #ifndef PRODUCT 377 extern uint implicit_null_checks; 378 implicit_null_checks++; 379 #endif 380 381 if( is_decoden ) { 382 // Check if we need to hoist decodeHeapOop_not_null first. 383 Block *valb = get_block_for_node(val); 384 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 385 // Hoist it up to the end of the test block together with its inputs if they exist. 386 for (uint i = 2; i < val->req(); i++) { 387 // DecodeN has 2 regular inputs + optional MachTemp or load Base inputs. 388 Node *temp = val->in(i); 389 Block *tempb = get_block_for_node(temp); 390 if (!tempb->dominates(block)) { 391 assert(block->dominates(tempb), "sanity check: temp node placement"); 392 // We only expect nodes without further inputs, like MachTemp or load Base. 393 assert(temp->req() == 0 || (temp->req() == 1 && temp->in(0) == (Node*)C->root()), 394 "need for recursive hoisting not expected"); 395 tempb->find_remove(temp); 396 block->add_inst(temp); 397 map_node_to_block(temp, block); 398 } 399 } 400 valb->find_remove(val); 401 block->add_inst(val); 402 map_node_to_block(val, block); 403 // DecodeN on x86 may kill flags. Check for flag-killing projections 404 // that also need to be hoisted. 405 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 406 Node* n = val->fast_out(j); 407 if( n->is_MachProj() ) { 408 get_block_for_node(n)->find_remove(n); 409 block->add_inst(n); 410 map_node_to_block(n, block); 411 } 412 } 413 } 414 } 415 416 // Hoist constant load inputs as well. 417 for (uint i = 1; i < best->req(); ++i) { 418 Node* n = best->in(i); 419 if (n->is_Con() && get_block_for_node(n) == get_block_for_node(best)) { 420 get_block_for_node(n)->find_remove(n); 421 block->add_inst(n); 422 map_node_to_block(n, block); 423 // Constant loads may kill flags (for example, when XORing a register). 424 // Check for flag-killing projections that also need to be hoisted. 425 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 426 Node* proj = n->fast_out(j); 427 if (proj->is_MachProj()) { 428 get_block_for_node(proj)->find_remove(proj); 429 block->add_inst(proj); 430 map_node_to_block(proj, block); 431 } 432 } 433 } 434 } 435 436 // Hoist the memory candidate up to the end of the test block. 437 Block *old_block = get_block_for_node(best); 438 old_block->find_remove(best); 439 block->add_inst(best); 440 map_node_to_block(best, block); 441 442 // Move the control dependence if it is pinned to not-null block. 443 // Don't change it in other cases: null or dominating control. 444 Node* ctrl = best->in(0); 445 if (ctrl != nullptr && get_block_for_node(ctrl) == not_null_block) { 446 // Set it to control edge of null check. 447 best->set_req(0, proj->in(0)->in(0)); 448 } 449 450 // Check for flag-killing projections that also need to be hoisted 451 // Should be DU safe because no edge updates. 452 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 453 Node* n = best->fast_out(j); 454 if( n->is_MachProj() ) { 455 get_block_for_node(n)->find_remove(n); 456 block->add_inst(n); 457 map_node_to_block(n, block); 458 } 459 } 460 461 // proj==Op_True --> ne test; proj==Op_False --> eq test. 462 // One of two graph shapes got matched: 463 // (IfTrue (If (Bool NE (CmpP ptr null)))) 464 // (IfFalse (If (Bool EQ (CmpP ptr null)))) 465 // null checks are always branch-if-eq. If we see a IfTrue projection 466 // then we are replacing a 'ne' test with a 'eq' null check test. 467 // We need to flip the projections to keep the same semantics. 468 if( proj->Opcode() == Op_IfTrue ) { 469 // Swap order of projections in basic block to swap branch targets 470 Node *tmp1 = block->get_node(block->end_idx()+1); 471 Node *tmp2 = block->get_node(block->end_idx()+2); 472 block->map_node(tmp2, block->end_idx()+1); 473 block->map_node(tmp1, block->end_idx()+2); 474 Node *tmp = new Node(C->top()); // Use not null input 475 tmp1->replace_by(tmp); 476 tmp2->replace_by(tmp1); 477 tmp->replace_by(tmp2); 478 tmp->destruct(nullptr); 479 } 480 481 // Remove the existing null check; use a new implicit null check instead. 482 // Since schedule-local needs precise def-use info, we need to correct 483 // it as well. 484 Node *old_tst = proj->in(0); 485 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx); 486 block->map_node(nul_chk, block->end_idx()); 487 map_node_to_block(nul_chk, block); 488 // Redirect users of old_test to nul_chk 489 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 490 old_tst->last_out(i2)->set_req(0, nul_chk); 491 // Clean-up any dead code 492 for (uint i3 = 0; i3 < old_tst->req(); i3++) { 493 Node* in = old_tst->in(i3); 494 old_tst->set_req(i3, nullptr); 495 if (in->outcnt() == 0) { 496 // Remove dead input node 497 in->disconnect_inputs(C); 498 block->find_remove(in); 499 } 500 } 501 502 latency_from_uses(nul_chk); 503 latency_from_uses(best); 504 505 // insert anti-dependences to defs in this block 506 if (! best->needs_anti_dependence_check()) { 507 for (uint k = 1; k < block->number_of_nodes(); k++) { 508 Node *n = block->get_node(k); 509 if (n->needs_anti_dependence_check() && 510 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) { 511 // Found anti-dependent load 512 insert_anti_dependences(block, n); 513 } 514 } 515 } 516 } 517 518 519 //------------------------------select----------------------------------------- 520 // Select a nice fellow from the worklist to schedule next. If there is only one 521 // choice, then use it. CreateEx nodes that are initially ready must start their 522 // blocks and are given the highest priority, by being placed at the beginning 523 // of the worklist. Next after initially-ready CreateEx nodes are projections, 524 // which must follow their parents, and CreateEx nodes with local input 525 // dependencies. Next are constants and CheckCastPP nodes. There are a number of 526 // other special cases, for instructions that consume condition codes, et al. 527 // These are chosen immediately. Some instructions are required to immediately 528 // precede the last instruction in the block, and these are taken last. Of the 529 // remaining cases (most), choose the instruction with the greatest latency 530 // (that is, the most number of pseudo-cycles required to the end of the 531 // routine). If there is a tie, choose the instruction with the most inputs. 532 Node* PhaseCFG::select( 533 Block* block, 534 Node_List &worklist, 535 GrowableArray<int> &ready_cnt, 536 VectorSet &next_call, 537 uint sched_slot, 538 intptr_t* recalc_pressure_nodes) { 539 540 // If only a single entry on the stack, use it 541 uint cnt = worklist.size(); 542 if (cnt == 1) { 543 Node *n = worklist[0]; 544 worklist.map(0,worklist.pop()); 545 return n; 546 } 547 548 uint choice = 0; // Bigger is most important 549 uint latency = 0; // Bigger is scheduled first 550 uint score = 0; // Bigger is better 551 int idx = -1; // Index in worklist 552 int cand_cnt = 0; // Candidate count 553 bool block_size_threshold_ok = (recalc_pressure_nodes != nullptr) && (block->number_of_nodes() > 10); 554 555 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 556 // Order in worklist is used to break ties. 557 // See caller for how this is used to delay scheduling 558 // of induction variable increments to after the other 559 // uses of the phi are scheduled. 560 Node *n = worklist[i]; // Get Node on worklist 561 562 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 563 if (iop == Op_CreateEx || n->is_Proj()) { 564 // CreateEx nodes that are initially ready must start the block (after Phi 565 // and Parm nodes which are pre-scheduled) and get top priority. This is 566 // currently enforced by placing them at the beginning of the initial 567 // worklist and selecting them eagerly here. After these, projections and 568 // other CreateEx nodes are selected with equal priority. 569 worklist.map(i,worklist.pop()); 570 return n; 571 } 572 573 if (n->Opcode() == Op_Con || iop == Op_CheckCastPP) { 574 // Constants and CheckCastPP nodes have higher priority than the rest of 575 // the nodes tested below. Record as current winner, but keep looking for 576 // higher-priority nodes in the worklist. 577 choice = 4; 578 // Latency and score are only used to break ties among low-priority nodes. 579 latency = 0; 580 score = 0; 581 idx = i; 582 continue; 583 } 584 585 // Final call in a block must be adjacent to 'catch' 586 Node *e = block->end(); 587 if( e->is_Catch() && e->in(0)->in(0) == n ) 588 continue; 589 590 // Memory op for an implicit null check has to be at the end of the block 591 if( e->is_MachNullCheck() && e->in(1) == n ) 592 continue; 593 594 // Schedule IV increment last. 595 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd) { 596 // Cmp might be matched into CountedLoopEnd node. 597 Node *cmp = (e->in(1)->ideal_reg() == Op_RegFlags) ? e->in(1) : e; 598 if (cmp->req() > 1 && cmp->in(1) == n && n->is_iteratively_computed()) { 599 continue; 600 } 601 } 602 603 uint n_choice = 2; 604 605 // See if this instruction is consumed by a branch. If so, then (as the 606 // branch is the last instruction in the basic block) force it to the 607 // end of the basic block 608 if ( must_clone[iop] ) { 609 // See if any use is a branch 610 bool found_machif = false; 611 612 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 613 Node* use = n->fast_out(j); 614 615 // The use is a conditional branch, make them adjacent 616 if (use->is_MachIf() && get_block_for_node(use) == block) { 617 found_machif = true; 618 break; 619 } 620 621 // More than this instruction pending for successor to be ready, 622 // don't choose this if other opportunities are ready 623 if (ready_cnt.at(use->_idx) > 1) 624 n_choice = 1; 625 } 626 627 // loop terminated, prefer not to use this instruction 628 if (found_machif) 629 continue; 630 } 631 632 // See if this has a predecessor that is "must_clone", i.e. sets the 633 // condition code. If so, choose this first 634 for (uint j = 0; j < n->req() ; j++) { 635 Node *inn = n->in(j); 636 if (inn) { 637 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 638 n_choice = 3; 639 break; 640 } 641 } 642 } 643 644 // MachTemps should be scheduled last so they are near their uses 645 if (n->is_MachTemp()) { 646 n_choice = 1; 647 } 648 649 uint n_latency = get_latency_for_node(n); 650 uint n_score = n->req(); // Many inputs get high score to break ties 651 652 if (OptoRegScheduling && block_size_threshold_ok) { 653 if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) { 654 _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit()); 655 _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit()); 656 // simulate the notion that we just picked this node to schedule 657 n->add_flag(Node::Flag_is_scheduled); 658 // now calculate its effect upon the graph if we did 659 adjust_register_pressure(n, block, recalc_pressure_nodes, false); 660 // return its state for finalize in case somebody else wins 661 n->remove_flag(Node::Flag_is_scheduled); 662 // now save the two final pressure components of register pressure, limiting pressure calcs to short size 663 short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure(); 664 short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure(); 665 recalc_pressure_nodes[n->_idx] = int_pressure; 666 recalc_pressure_nodes[n->_idx] |= (float_pressure << 16); 667 } 668 669 if (_scheduling_for_pressure) { 670 latency = n_latency; 671 if (n_choice != 3) { 672 // Now evaluate each register pressure component based on threshold in the score. 673 // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks 674 // on a single instruction, but we might see it shrink on both banks. 675 // For each use of register that has a register class that is over the high pressure limit, we build n_score up for 676 // live ranges that terminate on this instruction. 677 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 678 short int_pressure = (short)recalc_pressure_nodes[n->_idx]; 679 n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score; 680 } 681 if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 682 short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16); 683 n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score; 684 } 685 } else { 686 // make sure we choose these candidates 687 score = 0; 688 } 689 } 690 } 691 692 // Keep best latency found 693 cand_cnt++; 694 if (choice < n_choice || 695 (choice == n_choice && 696 ((StressLCM && C->randomized_select(cand_cnt)) || 697 (!StressLCM && 698 (latency < n_latency || 699 (latency == n_latency && 700 (score < n_score))))))) { 701 choice = n_choice; 702 latency = n_latency; 703 score = n_score; 704 idx = i; // Also keep index in worklist 705 } 706 } // End of for all ready nodes in worklist 707 708 guarantee(idx >= 0, "index should be set"); 709 Node *n = worklist[(uint)idx]; // Get the winner 710 711 worklist.map((uint)idx, worklist.pop()); // Compress worklist 712 return n; 713 } 714 715 //-------------------------adjust_register_pressure---------------------------- 716 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) { 717 PhaseLive* liveinfo = _regalloc->get_live(); 718 IndexSet* liveout = liveinfo->live(block); 719 // first adjust the register pressure for the sources 720 for (uint i = 1; i < n->req(); i++) { 721 bool lrg_ends = false; 722 Node *src_n = n->in(i); 723 if (src_n == nullptr) continue; 724 if (!src_n->is_Mach()) continue; 725 uint src = _regalloc->_lrg_map.find(src_n); 726 if (src == 0) continue; 727 LRG& lrg_src = _regalloc->lrgs(src); 728 // detect if the live range ends or not 729 if (liveout->member(src) == false) { 730 lrg_ends = true; 731 for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) { 732 Node* m = src_n->fast_out(j); // Get user 733 if (m == n) continue; 734 if (!m->is_Mach()) continue; 735 MachNode *mach = m->as_Mach(); 736 bool src_matches = false; 737 int iop = mach->ideal_Opcode(); 738 739 switch (iop) { 740 case Op_StoreB: 741 case Op_StoreC: 742 case Op_StoreCM: 743 case Op_StoreD: 744 case Op_StoreF: 745 case Op_StoreI: 746 case Op_StoreL: 747 case Op_StoreP: 748 case Op_StoreN: 749 case Op_StoreVector: 750 case Op_StoreVectorMasked: 751 case Op_StoreVectorScatter: 752 case Op_StoreVectorScatterMasked: 753 case Op_StoreNKlass: 754 for (uint k = 1; k < m->req(); k++) { 755 Node *in = m->in(k); 756 if (in == src_n) { 757 src_matches = true; 758 break; 759 } 760 } 761 break; 762 763 default: 764 src_matches = true; 765 break; 766 } 767 768 // If we have a store as our use, ignore the non source operands 769 if (src_matches == false) continue; 770 771 // Mark every unscheduled use which is not n with a recalculation 772 if ((get_block_for_node(m) == block) && (!m->is_scheduled())) { 773 if (finalize_mode && !m->is_Phi()) { 774 recalc_pressure_nodes[m->_idx] = 0x7fff7fff; 775 } 776 lrg_ends = false; 777 } 778 } 779 } 780 // if none, this live range ends and we can adjust register pressure 781 if (lrg_ends) { 782 if (finalize_mode) { 783 _regalloc->lower_pressure(block, 0, lrg_src, nullptr, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 784 } else { 785 _regalloc->lower_pressure(block, 0, lrg_src, nullptr, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 786 } 787 } 788 } 789 790 // now add the register pressure from the dest and evaluate which heuristic we should use: 791 // 1.) The default, latency scheduling 792 // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks 793 uint dst = _regalloc->_lrg_map.find(n); 794 if (dst != 0) { 795 LRG& lrg_dst = _regalloc->lrgs(dst); 796 if (finalize_mode) { 797 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 798 // check to see if we fall over the register pressure cliff here 799 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 800 _scheduling_for_pressure = true; 801 } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 802 _scheduling_for_pressure = true; 803 } else { 804 // restore latency scheduling mode 805 _scheduling_for_pressure = false; 806 } 807 } else { 808 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 809 } 810 } 811 } 812 813 //------------------------------set_next_call---------------------------------- 814 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 815 if( next_call.test_set(n->_idx) ) return; 816 for( uint i=0; i<n->len(); i++ ) { 817 Node *m = n->in(i); 818 if( !m ) continue; // must see all nodes in block that precede call 819 if (get_block_for_node(m) == block) { 820 set_next_call(block, m, next_call); 821 } 822 } 823 } 824 825 //------------------------------needed_for_next_call--------------------------- 826 // Set the flag 'next_call' for each Node that is needed for the next call to 827 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 828 // next subroutine call get priority - basically it moves things NOT needed 829 // for the next call till after the call. This prevents me from trying to 830 // carry lots of stuff live across a call. 831 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 832 // Find the next control-defining Node in this block 833 Node* call = nullptr; 834 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 835 Node* m = this_call->fast_out(i); 836 if (get_block_for_node(m) == block && // Local-block user 837 m != this_call && // Not self-start node 838 m->is_MachCall()) { 839 call = m; 840 break; 841 } 842 } 843 if (call == nullptr) return; // No next call (e.g., block end is near) 844 // Set next-call for all inputs to this call 845 set_next_call(block, call, next_call); 846 } 847 848 //------------------------------add_call_kills------------------------------------- 849 // helper function that adds caller save registers to MachProjNode 850 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 851 // Fill in the kill mask for the call 852 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 853 if( !regs.Member(r) ) { // Not already defined by the call 854 // Save-on-call register? 855 if ((save_policy[r] == 'C') || 856 (save_policy[r] == 'A') || 857 ((save_policy[r] == 'E') && exclude_soe)) { 858 proj->_rout.Insert(r); 859 } 860 } 861 } 862 } 863 864 865 //------------------------------sched_call------------------------------------- 866 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 867 RegMask regs; 868 869 // Schedule all the users of the call right now. All the users are 870 // projection Nodes, so they must be scheduled next to the call. 871 // Collect all the defined registers. 872 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 873 Node* n = mcall->fast_out(i); 874 assert( n->is_MachProj(), "" ); 875 int n_cnt = ready_cnt.at(n->_idx)-1; 876 ready_cnt.at_put(n->_idx, n_cnt); 877 assert( n_cnt == 0, "" ); 878 // Schedule next to call 879 block->map_node(n, node_cnt++); 880 // Collect defined registers 881 regs.OR(n->out_RegMask()); 882 // Check for scheduling the next control-definer 883 if( n->bottom_type() == Type::CONTROL ) 884 // Warm up next pile of heuristic bits 885 needed_for_next_call(block, n, next_call); 886 887 // Children of projections are now all ready 888 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 889 Node* m = n->fast_out(j); // Get user 890 if(get_block_for_node(m) != block) { 891 continue; 892 } 893 if( m->is_Phi() ) continue; 894 int m_cnt = ready_cnt.at(m->_idx) - 1; 895 ready_cnt.at_put(m->_idx, m_cnt); 896 if( m_cnt == 0 ) 897 worklist.push(m); 898 } 899 900 } 901 902 // Act as if the call defines the Frame Pointer. 903 // Certainly the FP is alive and well after the call. 904 regs.Insert(_matcher.c_frame_pointer()); 905 906 // Set all registers killed and not already defined by the call. 907 uint r_cnt = mcall->tf()->range_cc()->cnt(); 908 int op = mcall->ideal_Opcode(); 909 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 910 map_node_to_block(proj, block); 911 block->insert_node(proj, node_cnt++); 912 913 // Select the right register save policy. 914 const char *save_policy = nullptr; 915 switch (op) { 916 case Op_CallRuntime: 917 case Op_CallLeaf: 918 case Op_CallLeafNoFP: 919 case Op_CallLeafVector: 920 // Calling C code so use C calling convention 921 save_policy = _matcher._c_reg_save_policy; 922 break; 923 924 case Op_CallStaticJava: 925 case Op_CallDynamicJava: 926 // Calling Java code so use Java calling convention 927 save_policy = _matcher._register_save_policy; 928 break; 929 930 default: 931 ShouldNotReachHere(); 932 } 933 934 // When using CallRuntime mark SOE registers as killed by the call 935 // so values that could show up in the RegisterMap aren't live in a 936 // callee saved register since the register wouldn't know where to 937 // find them. CallLeaf and CallLeafNoFP are ok because they can't 938 // have debug info on them. Strictly speaking this only needs to be 939 // done for oops since idealreg2debugmask takes care of debug info 940 // references but there no way to handle oops differently than other 941 // pointers as far as the kill mask goes. 942 bool exclude_soe = op == Op_CallRuntime; 943 944 // If the call is a MethodHandle invoke, we need to exclude the 945 // register which is used to save the SP value over MH invokes from 946 // the mask. Otherwise this register could be used for 947 // deoptimization information. 948 if (op == Op_CallStaticJava) { 949 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 950 if (mcallstaticjava->_method_handle_invoke) 951 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 952 } 953 954 add_call_kills(proj, regs, save_policy, exclude_soe); 955 956 return node_cnt; 957 } 958 959 960 //------------------------------schedule_local--------------------------------- 961 // Topological sort within a block. Someday become a real scheduler. 962 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) { 963 // Already "sorted" are the block start Node (as the first entry), and 964 // the block-ending Node and any trailing control projections. We leave 965 // these alone. PhiNodes and ParmNodes are made to follow the block start 966 // Node. Everything else gets topo-sorted. 967 968 #ifndef PRODUCT 969 if (trace_opto_pipelining()) { 970 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 971 for (uint i = 0;i < block->number_of_nodes(); i++) { 972 tty->print("# "); 973 block->get_node(i)->dump(); 974 } 975 tty->print_cr("#"); 976 } 977 #endif 978 979 // RootNode is already sorted 980 if (block->number_of_nodes() == 1) { 981 return true; 982 } 983 984 bool block_size_threshold_ok = (recalc_pressure_nodes != nullptr) && (block->number_of_nodes() > 10); 985 986 // We track the uses of local definitions as input dependences so that 987 // we know when a given instruction is available to be scheduled. 988 uint i; 989 if (OptoRegScheduling && block_size_threshold_ok) { 990 for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc 991 Node *n = block->get_node(i); 992 n->remove_flag(Node::Flag_is_scheduled); 993 if (!n->is_Phi()) { 994 recalc_pressure_nodes[n->_idx] = 0x7fff7fff; 995 } 996 } 997 } 998 999 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 1000 uint node_cnt = block->end_idx(); 1001 uint phi_cnt = 1; 1002 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 1003 Node *n = block->get_node(i); 1004 if( n->is_Phi() || // Found a PhiNode or ParmNode 1005 (n->is_Proj() && n->in(0) == block->head()) ) { 1006 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 1007 block->map_node(block->get_node(phi_cnt), i); 1008 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 1009 if (OptoRegScheduling && block_size_threshold_ok) { 1010 // mark n as scheduled 1011 n->add_flag(Node::Flag_is_scheduled); 1012 } 1013 } else { // All others 1014 // Count block-local inputs to 'n' 1015 uint cnt = n->len(); // Input count 1016 uint local = 0; 1017 for( uint j=0; j<cnt; j++ ) { 1018 Node *m = n->in(j); 1019 if( m && get_block_for_node(m) == block && !m->is_top() ) 1020 local++; // One more block-local input 1021 } 1022 ready_cnt.at_put(n->_idx, local); // Count em up 1023 1024 #ifdef ASSERT 1025 if (UseG1GC) { 1026 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 1027 // Check the precedence edges 1028 for (uint prec = n->req(); prec < n->len(); prec++) { 1029 Node* oop_store = n->in(prec); 1030 if (oop_store != nullptr) { 1031 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 1032 } 1033 } 1034 } 1035 } 1036 #endif 1037 1038 // A few node types require changing a required edge to a precedence edge 1039 // before allocation. 1040 if( n->is_Mach() && n->req() > TypeFunc::Parms && 1041 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 1042 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 1043 // MemBarAcquire could be created without Precedent edge. 1044 // del_req() replaces the specified edge with the last input edge 1045 // and then removes the last edge. If the specified edge > number of 1046 // edges the last edge will be moved outside of the input edges array 1047 // and the edge will be lost. This is why this code should be 1048 // executed only when Precedent (== TypeFunc::Parms) edge is present. 1049 Node *x = n->in(TypeFunc::Parms); 1050 if (x != nullptr && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) { 1051 // Old edge to node within same block will get removed, but no precedence 1052 // edge will get added because it already exists. Update ready count. 1053 int cnt = ready_cnt.at(n->_idx); 1054 assert(cnt > 1, "MemBar node %d must not get ready here", n->_idx); 1055 ready_cnt.at_put(n->_idx, cnt-1); 1056 } 1057 n->del_req(TypeFunc::Parms); 1058 n->add_prec(x); 1059 } 1060 } 1061 } 1062 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 1063 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 1064 1065 // All the prescheduled guys do not hold back internal nodes 1066 uint i3; 1067 for (i3 = 0; i3 < phi_cnt; i3++) { // For all pre-scheduled 1068 Node *n = block->get_node(i3); // Get pre-scheduled 1069 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1070 Node* m = n->fast_out(j); 1071 if (get_block_for_node(m) == block) { // Local-block user 1072 int m_cnt = ready_cnt.at(m->_idx)-1; 1073 if (OptoRegScheduling && block_size_threshold_ok) { 1074 // mark m as scheduled 1075 if (m_cnt < 0) { 1076 m->add_flag(Node::Flag_is_scheduled); 1077 } 1078 } 1079 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 1080 } 1081 } 1082 } 1083 1084 Node_List delay; 1085 // Make a worklist 1086 Node_List worklist; 1087 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 1088 Node *m = block->get_node(i4); 1089 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 1090 if (m->is_iteratively_computed()) { 1091 // Push induction variable increments last to allow other uses 1092 // of the phi to be scheduled first. The select() method breaks 1093 // ties in scheduling by worklist order. 1094 delay.push(m); 1095 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 1096 // Place CreateEx nodes that are initially ready at the beginning of the 1097 // worklist so they are selected first and scheduled at the block start. 1098 worklist.insert(0, m); 1099 } else { 1100 worklist.push(m); // Then on to worklist! 1101 } 1102 } 1103 } 1104 while (delay.size()) { 1105 Node* d = delay.pop(); 1106 worklist.push(d); 1107 } 1108 1109 if (OptoRegScheduling && block_size_threshold_ok) { 1110 // To stage register pressure calculations we need to examine the live set variables 1111 // breaking them up by register class to compartmentalize the calculations. 1112 _regalloc->_sched_int_pressure.init(Matcher::int_pressure_limit()); 1113 _regalloc->_sched_float_pressure.init(Matcher::float_pressure_limit()); 1114 _regalloc->_scratch_int_pressure.init(Matcher::int_pressure_limit()); 1115 _regalloc->_scratch_float_pressure.init(Matcher::float_pressure_limit()); 1116 1117 _regalloc->compute_entry_block_pressure(block); 1118 } 1119 1120 // Warm up the 'next_call' heuristic bits 1121 needed_for_next_call(block, block->head(), next_call); 1122 1123 #ifndef PRODUCT 1124 if (trace_opto_pipelining()) { 1125 for (uint j=0; j< block->number_of_nodes(); j++) { 1126 Node *n = block->get_node(j); 1127 int idx = n->_idx; 1128 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 1129 tty->print("latency:%3d ", get_latency_for_node(n)); 1130 tty->print("%4d: %s\n", idx, n->Name()); 1131 } 1132 } 1133 #endif 1134 1135 uint max_idx = (uint)ready_cnt.length(); 1136 // Pull from worklist and schedule 1137 while( worklist.size() ) { // Worklist is not ready 1138 1139 #ifndef PRODUCT 1140 if (trace_opto_pipelining()) { 1141 tty->print("# ready list:"); 1142 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1143 Node *n = worklist[i]; // Get Node on worklist 1144 tty->print(" %d", n->_idx); 1145 } 1146 tty->cr(); 1147 } 1148 #endif 1149 1150 // Select and pop a ready guy from worklist 1151 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes); 1152 block->map_node(n, phi_cnt++); // Schedule him next 1153 1154 if (OptoRegScheduling && block_size_threshold_ok) { 1155 n->add_flag(Node::Flag_is_scheduled); 1156 1157 // Now adjust the resister pressure with the node we selected 1158 if (!n->is_Phi()) { 1159 adjust_register_pressure(n, block, recalc_pressure_nodes, true); 1160 } 1161 } 1162 1163 #ifndef PRODUCT 1164 if (trace_opto_pipelining()) { 1165 tty->print("# select %d: %s", n->_idx, n->Name()); 1166 tty->print(", latency:%d", get_latency_for_node(n)); 1167 n->dump(); 1168 if (Verbose) { 1169 tty->print("# ready list:"); 1170 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1171 Node *n = worklist[i]; // Get Node on worklist 1172 tty->print(" %d", n->_idx); 1173 } 1174 tty->cr(); 1175 } 1176 } 1177 1178 #endif 1179 if( n->is_MachCall() ) { 1180 MachCallNode *mcall = n->as_MachCall(); 1181 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 1182 continue; 1183 } 1184 1185 if (n->is_Mach() && n->as_Mach()->has_call()) { 1186 RegMask regs; 1187 regs.Insert(_matcher.c_frame_pointer()); 1188 regs.OR(n->out_RegMask()); 1189 1190 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 1191 map_node_to_block(proj, block); 1192 block->insert_node(proj, phi_cnt++); 1193 1194 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 1195 } 1196 1197 // Children are now all ready 1198 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 1199 Node* m = n->fast_out(i5); // Get user 1200 if (get_block_for_node(m) != block) { 1201 continue; 1202 } 1203 if( m->is_Phi() ) continue; 1204 if (m->_idx >= max_idx) { // new node, skip it 1205 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 1206 continue; 1207 } 1208 int m_cnt = ready_cnt.at(m->_idx) - 1; 1209 ready_cnt.at_put(m->_idx, m_cnt); 1210 if( m_cnt == 0 ) 1211 worklist.push(m); 1212 } 1213 } 1214 1215 if( phi_cnt != block->end_idx() ) { 1216 // did not schedule all. Retry, Bailout, or Die 1217 if (C->subsume_loads() == true && !C->failing()) { 1218 // Retry with subsume_loads == false 1219 // If this is the first failure, the sentinel string will "stick" 1220 // to the Compile object, and the C2Compiler will see it and retry. 1221 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 1222 } else { 1223 assert(false, "graph should be schedulable"); 1224 } 1225 // assert( phi_cnt == end_idx(), "did not schedule all" ); 1226 return false; 1227 } 1228 1229 if (OptoRegScheduling && block_size_threshold_ok) { 1230 _regalloc->compute_exit_block_pressure(block); 1231 block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure(); 1232 block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure(); 1233 } 1234 1235 #ifndef PRODUCT 1236 if (trace_opto_pipelining()) { 1237 tty->print_cr("#"); 1238 tty->print_cr("# after schedule_local"); 1239 for (uint i = 0;i < block->number_of_nodes();i++) { 1240 tty->print("# "); 1241 block->get_node(i)->dump(); 1242 } 1243 tty->print_cr("# "); 1244 1245 if (OptoRegScheduling && block_size_threshold_ok) { 1246 tty->print_cr("# pressure info : %d", block->_pre_order); 1247 _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info"); 1248 _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info"); 1249 } 1250 tty->cr(); 1251 } 1252 #endif 1253 1254 return true; 1255 } 1256 1257 //--------------------------catch_cleanup_fix_all_inputs----------------------- 1258 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 1259 for (uint l = 0; l < use->len(); l++) { 1260 if (use->in(l) == old_def) { 1261 if (l < use->req()) { 1262 use->set_req(l, new_def); 1263 } else { 1264 use->rm_prec(l); 1265 use->add_prec(new_def); 1266 l--; 1267 } 1268 } 1269 } 1270 } 1271 1272 //------------------------------catch_cleanup_find_cloned_def------------------ 1273 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1274 assert( use_blk != def_blk, "Inter-block cleanup only"); 1275 1276 // The use is some block below the Catch. Find and return the clone of the def 1277 // that dominates the use. If there is no clone in a dominating block, then 1278 // create a phi for the def in a dominating block. 1279 1280 // Find which successor block dominates this use. The successor 1281 // blocks must all be single-entry (from the Catch only; I will have 1282 // split blocks to make this so), hence they all dominate. 1283 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 1284 use_blk = use_blk->_idom; 1285 1286 // Find the successor 1287 Node *fixup = nullptr; 1288 1289 uint j; 1290 for( j = 0; j < def_blk->_num_succs; j++ ) 1291 if( use_blk == def_blk->_succs[j] ) 1292 break; 1293 1294 if( j == def_blk->_num_succs ) { 1295 // Block at same level in dom-tree is not a successor. It needs a 1296 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 1297 Node_Array inputs; 1298 for(uint k = 1; k < use_blk->num_preds(); k++) { 1299 Block* block = get_block_for_node(use_blk->pred(k)); 1300 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 1301 } 1302 1303 // Check to see if the use_blk already has an identical phi inserted. 1304 // If it exists, it will be at the first position since all uses of a 1305 // def are processed together. 1306 Node *phi = use_blk->get_node(1); 1307 if( phi->is_Phi() ) { 1308 fixup = phi; 1309 for (uint k = 1; k < use_blk->num_preds(); k++) { 1310 if (phi->in(k) != inputs[k]) { 1311 // Not a match 1312 fixup = nullptr; 1313 break; 1314 } 1315 } 1316 } 1317 1318 // If an existing PhiNode was not found, make a new one. 1319 if (fixup == nullptr) { 1320 Node *new_phi = PhiNode::make(use_blk->head(), def); 1321 use_blk->insert_node(new_phi, 1); 1322 map_node_to_block(new_phi, use_blk); 1323 for (uint k = 1; k < use_blk->num_preds(); k++) { 1324 new_phi->set_req(k, inputs[k]); 1325 } 1326 fixup = new_phi; 1327 } 1328 1329 } else { 1330 // Found the use just below the Catch. Make it use the clone. 1331 fixup = use_blk->get_node(n_clone_idx); 1332 } 1333 1334 return fixup; 1335 } 1336 1337 //--------------------------catch_cleanup_intra_block-------------------------- 1338 // Fix all input edges in use that reference "def". The use is in the same 1339 // block as the def and both have been cloned in each successor block. 1340 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1341 1342 // Both the use and def have been cloned. For each successor block, 1343 // get the clone of the use, and make its input the clone of the def 1344 // found in that block. 1345 1346 uint use_idx = blk->find_node(use); 1347 uint offset_idx = use_idx - beg; 1348 for( uint k = 0; k < blk->_num_succs; k++ ) { 1349 // Get clone in each successor block 1350 Block *sb = blk->_succs[k]; 1351 Node *clone = sb->get_node(offset_idx+1); 1352 assert( clone->Opcode() == use->Opcode(), "" ); 1353 1354 // Make use-clone reference the def-clone 1355 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1356 } 1357 } 1358 1359 //------------------------------catch_cleanup_inter_block--------------------- 1360 // Fix all input edges in use that reference "def". The use is in a different 1361 // block than the def. 1362 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1363 if( !use_blk ) return; // Can happen if the use is a precedence edge 1364 1365 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1366 catch_cleanup_fix_all_inputs(use, def, new_def); 1367 } 1368 1369 //------------------------------call_catch_cleanup----------------------------- 1370 // If we inserted any instructions between a Call and his CatchNode, 1371 // clone the instructions on all paths below the Catch. 1372 void PhaseCFG::call_catch_cleanup(Block* block) { 1373 1374 // End of region to clone 1375 uint end = block->end_idx(); 1376 if( !block->get_node(end)->is_Catch() ) return; 1377 // Start of region to clone 1378 uint beg = end; 1379 while(!block->get_node(beg-1)->is_MachProj() || 1380 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1381 beg--; 1382 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1383 } 1384 // Range of inserted instructions is [beg, end) 1385 if( beg == end ) return; 1386 1387 // Clone along all Catch output paths. Clone area between the 'beg' and 1388 // 'end' indices. 1389 for( uint i = 0; i < block->_num_succs; i++ ) { 1390 Block *sb = block->_succs[i]; 1391 // Clone the entire area; ignoring the edge fixup for now. 1392 for( uint j = end; j > beg; j-- ) { 1393 Node *clone = block->get_node(j-1)->clone(); 1394 sb->insert_node(clone, 1); 1395 map_node_to_block(clone, sb); 1396 if (clone->needs_anti_dependence_check()) { 1397 insert_anti_dependences(sb, clone); 1398 } 1399 } 1400 } 1401 1402 1403 // Fixup edges. Check the def-use info per cloned Node 1404 for(uint i2 = beg; i2 < end; i2++ ) { 1405 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1406 Node *n = block->get_node(i2); // Node that got cloned 1407 // Need DU safe iterator because of edge manipulation in calls. 1408 Unique_Node_List* out = new Unique_Node_List(); 1409 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1410 out->push(n->fast_out(j1)); 1411 } 1412 uint max = out->size(); 1413 for (uint j = 0; j < max; j++) {// For all users 1414 Node *use = out->pop(); 1415 Block *buse = get_block_for_node(use); 1416 if( use->is_Phi() ) { 1417 for( uint k = 1; k < use->req(); k++ ) 1418 if( use->in(k) == n ) { 1419 Block* b = get_block_for_node(buse->pred(k)); 1420 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1421 use->set_req(k, fixup); 1422 } 1423 } else { 1424 if (block == buse) { 1425 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1426 } else { 1427 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1428 } 1429 } 1430 } // End for all users 1431 1432 } // End of for all Nodes in cloned area 1433 1434 // Remove the now-dead cloned ops 1435 for(uint i3 = beg; i3 < end; i3++ ) { 1436 block->get_node(beg)->disconnect_inputs(C); 1437 block->remove_node(beg); 1438 } 1439 1440 // If the successor blocks have a CreateEx node, move it back to the top 1441 for (uint i4 = 0; i4 < block->_num_succs; i4++) { 1442 Block *sb = block->_succs[i4]; 1443 uint new_cnt = end - beg; 1444 // Remove any newly created, but dead, nodes by traversing their schedule 1445 // backwards. Here, a dead node is a node whose only outputs (if any) are 1446 // unused projections. 1447 for (uint j = new_cnt; j > 0; j--) { 1448 Node *n = sb->get_node(j); 1449 // Individual projections are examined together with all siblings when 1450 // their parent is visited. 1451 if (n->is_Proj()) { 1452 continue; 1453 } 1454 bool dead = true; 1455 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1456 Node* out = n->fast_out(i); 1457 // n is live if it has a non-projection output or a used projection. 1458 if (!out->is_Proj() || out->outcnt() > 0) { 1459 dead = false; 1460 break; 1461 } 1462 } 1463 if (dead) { 1464 // n's only outputs (if any) are unused projections scheduled next to n 1465 // (see PhaseCFG::select()). Remove these projections backwards. 1466 for (uint k = j + n->outcnt(); k > j; k--) { 1467 Node* proj = sb->get_node(k); 1468 assert(proj->is_Proj() && proj->in(0) == n, 1469 "projection should correspond to dead node"); 1470 proj->disconnect_inputs(C); 1471 sb->remove_node(k); 1472 new_cnt--; 1473 } 1474 // Now remove the node itself. 1475 n->disconnect_inputs(C); 1476 sb->remove_node(j); 1477 new_cnt--; 1478 } 1479 } 1480 // If any newly created nodes remain, move the CreateEx node to the top 1481 if (new_cnt > 0) { 1482 Node *cex = sb->get_node(1+new_cnt); 1483 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1484 sb->remove_node(1+new_cnt); 1485 sb->insert_node(cex, 1); 1486 } 1487 } 1488 } 1489 }