lcm.cpp revision 642:660978a2a31a
1/* 2 * Copyright 1998-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25// Optimization - Graph Style 26 27#include "incls/_precompiled.incl" 28#include "incls/_lcm.cpp.incl" 29 30//------------------------------implicit_null_check---------------------------- 31// Detect implicit-null-check opportunities. Basically, find NULL checks 32// with suitable memory ops nearby. Use the memory op to do the NULL check. 33// I can generate a memory op if there is not one nearby. 34// The proj is the control projection for the not-null case. 35// The val is the pointer being checked for nullness. 36void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) { 37 // Assume if null check need for 0 offset then always needed 38 // Intel solaris doesn't support any null checks yet and no 39 // mechanism exists (yet) to set the switches at an os_cpu level 40 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 41 42 // Make sure the ptr-is-null path appears to be uncommon! 43 float f = end()->as_MachIf()->_prob; 44 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 45 if( f > PROB_UNLIKELY_MAG(4) ) return; 46 47 uint bidx = 0; // Capture index of value into memop 48 bool was_store; // Memory op is a store op 49 50 // Get the successor block for if the test ptr is non-null 51 Block* not_null_block; // this one goes with the proj 52 Block* null_block; 53 if (_nodes[_nodes.size()-1] == proj) { 54 null_block = _succs[0]; 55 not_null_block = _succs[1]; 56 } else { 57 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other"); 58 not_null_block = _succs[0]; 59 null_block = _succs[1]; 60 } 61 while (null_block->is_Empty() == Block::empty_with_goto) { 62 null_block = null_block->_succs[0]; 63 } 64 65 // Search the exception block for an uncommon trap. 66 // (See Parse::do_if and Parse::do_ifnull for the reason 67 // we need an uncommon trap. Briefly, we need a way to 68 // detect failure of this optimization, as in 6366351.) 69 { 70 bool found_trap = false; 71 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) { 72 Node* nn = null_block->_nodes[i1]; 73 if (nn->is_MachCall() && 74 nn->as_MachCall()->entry_point() == 75 SharedRuntime::uncommon_trap_blob()->instructions_begin()) { 76 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 77 if (trtype->isa_int() && trtype->is_int()->is_con()) { 78 jint tr_con = trtype->is_int()->get_con(); 79 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 80 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 81 assert((int)reason < (int)BitsPerInt, "recode bit map"); 82 if (is_set_nth_bit(allowed_reasons, (int) reason) 83 && action != Deoptimization::Action_none) { 84 // This uncommon trap is sure to recompile, eventually. 85 // When that happens, C->too_many_traps will prevent 86 // this transformation from happening again. 87 found_trap = true; 88 } 89 } 90 break; 91 } 92 } 93 if (!found_trap) { 94 // We did not find an uncommon trap. 95 return; 96 } 97 } 98 99 // Search the successor block for a load or store who's base value is also 100 // the tested value. There may be several. 101 Node_List *out = new Node_List(Thread::current()->resource_area()); 102 MachNode *best = NULL; // Best found so far 103 for (DUIterator i = val->outs(); val->has_out(i); i++) { 104 Node *m = val->out(i); 105 if( !m->is_Mach() ) continue; 106 MachNode *mach = m->as_Mach(); 107 was_store = false; 108 switch( mach->ideal_Opcode() ) { 109 case Op_LoadB: 110 case Op_LoadUS: 111 case Op_LoadD: 112 case Op_LoadF: 113 case Op_LoadI: 114 case Op_LoadL: 115 case Op_LoadP: 116 case Op_LoadN: 117 case Op_LoadS: 118 case Op_LoadKlass: 119 case Op_LoadNKlass: 120 case Op_LoadRange: 121 case Op_LoadD_unaligned: 122 case Op_LoadL_unaligned: 123 break; 124 case Op_StoreB: 125 case Op_StoreC: 126 case Op_StoreCM: 127 case Op_StoreD: 128 case Op_StoreF: 129 case Op_StoreI: 130 case Op_StoreL: 131 case Op_StoreP: 132 case Op_StoreN: 133 was_store = true; // Memory op is a store op 134 // Stores will have their address in slot 2 (memory in slot 1). 135 // If the value being nul-checked is in another slot, it means we 136 // are storing the checked value, which does NOT check the value! 137 if( mach->in(2) != val ) continue; 138 break; // Found a memory op? 139 case Op_StrComp: 140 case Op_AryEq: 141 // Not a legit memory op for implicit null check regardless of 142 // embedded loads 143 continue; 144 default: // Also check for embedded loads 145 if( !mach->needs_anti_dependence_check() ) 146 continue; // Not an memory op; skip it 147 break; 148 } 149 // check if the offset is not too high for implicit exception 150 { 151 intptr_t offset = 0; 152 const TypePtr *adr_type = NULL; // Do not need this return value here 153 const Node* base = mach->get_base_and_disp(offset, adr_type); 154 if (base == NULL || base == NodeSentinel) { 155 // Narrow oop address doesn't have base, only index 156 if( val->bottom_type()->isa_narrowoop() && 157 MacroAssembler::needs_explicit_null_check(offset) ) 158 continue; // Give up if offset is beyond page size 159 // cannot reason about it; is probably not implicit null exception 160 } else { 161 const TypePtr* tptr; 162 if (UseCompressedOops && Universe::narrow_oop_shift() == 0) { 163 // 32-bits narrow oop can be the base of address expressions 164 tptr = base->bottom_type()->make_ptr(); 165 } else { 166 // only regular oops are expected here 167 tptr = base->bottom_type()->is_ptr(); 168 } 169 // Give up if offset is not a compile-time constant 170 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 171 continue; 172 offset += tptr->_offset; // correct if base is offseted 173 if( MacroAssembler::needs_explicit_null_check(offset) ) 174 continue; // Give up is reference is beyond 4K page size 175 } 176 } 177 178 // Check ctrl input to see if the null-check dominates the memory op 179 Block *cb = cfg->_bbs[mach->_idx]; 180 cb = cb->_idom; // Always hoist at least 1 block 181 if( !was_store ) { // Stores can be hoisted only one block 182 while( cb->_dom_depth > (_dom_depth + 1)) 183 cb = cb->_idom; // Hoist loads as far as we want 184 // The non-null-block should dominate the memory op, too. Live 185 // range spilling will insert a spill in the non-null-block if it is 186 // needs to spill the memory op for an implicit null check. 187 if (cb->_dom_depth == (_dom_depth + 1)) { 188 if (cb != not_null_block) continue; 189 cb = cb->_idom; 190 } 191 } 192 if( cb != this ) continue; 193 194 // Found a memory user; see if it can be hoisted to check-block 195 uint vidx = 0; // Capture index of value into memop 196 uint j; 197 for( j = mach->req()-1; j > 0; j-- ) { 198 if( mach->in(j) == val ) vidx = j; 199 // Block of memory-op input 200 Block *inb = cfg->_bbs[mach->in(j)->_idx]; 201 Block *b = this; // Start from nul check 202 while( b != inb && b->_dom_depth > inb->_dom_depth ) 203 b = b->_idom; // search upwards for input 204 // See if input dominates null check 205 if( b != inb ) 206 break; 207 } 208 if( j > 0 ) 209 continue; 210 Block *mb = cfg->_bbs[mach->_idx]; 211 // Hoisting stores requires more checks for the anti-dependence case. 212 // Give up hoisting if we have to move the store past any load. 213 if( was_store ) { 214 Block *b = mb; // Start searching here for a local load 215 // mach use (faulting) trying to hoist 216 // n might be blocker to hoisting 217 while( b != this ) { 218 uint k; 219 for( k = 1; k < b->_nodes.size(); k++ ) { 220 Node *n = b->_nodes[k]; 221 if( n->needs_anti_dependence_check() && 222 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 223 break; // Found anti-dependent load 224 } 225 if( k < b->_nodes.size() ) 226 break; // Found anti-dependent load 227 // Make sure control does not do a merge (would have to check allpaths) 228 if( b->num_preds() != 2 ) break; 229 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block 230 } 231 if( b != this ) continue; 232 } 233 234 // Make sure this memory op is not already being used for a NullCheck 235 Node *e = mb->end(); 236 if( e->is_MachNullCheck() && e->in(1) == mach ) 237 continue; // Already being used as a NULL check 238 239 // Found a candidate! Pick one with least dom depth - the highest 240 // in the dom tree should be closest to the null check. 241 if( !best || 242 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) { 243 best = mach; 244 bidx = vidx; 245 246 } 247 } 248 // No candidate! 249 if( !best ) return; 250 251 // ---- Found an implicit null check 252 extern int implicit_null_checks; 253 implicit_null_checks++; 254 255 // Hoist the memory candidate up to the end of the test block. 256 Block *old_block = cfg->_bbs[best->_idx]; 257 old_block->find_remove(best); 258 add_inst(best); 259 cfg->_bbs.map(best->_idx,this); 260 261 // Move the control dependence 262 if (best->in(0) && best->in(0) == old_block->_nodes[0]) 263 best->set_req(0, _nodes[0]); 264 265 // Check for flag-killing projections that also need to be hoisted 266 // Should be DU safe because no edge updates. 267 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 268 Node* n = best->fast_out(j); 269 if( n->Opcode() == Op_MachProj ) { 270 cfg->_bbs[n->_idx]->find_remove(n); 271 add_inst(n); 272 cfg->_bbs.map(n->_idx,this); 273 } 274 } 275 276 Compile *C = cfg->C; 277 // proj==Op_True --> ne test; proj==Op_False --> eq test. 278 // One of two graph shapes got matched: 279 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 280 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 281 // NULL checks are always branch-if-eq. If we see a IfTrue projection 282 // then we are replacing a 'ne' test with a 'eq' NULL check test. 283 // We need to flip the projections to keep the same semantics. 284 if( proj->Opcode() == Op_IfTrue ) { 285 // Swap order of projections in basic block to swap branch targets 286 Node *tmp1 = _nodes[end_idx()+1]; 287 Node *tmp2 = _nodes[end_idx()+2]; 288 _nodes.map(end_idx()+1, tmp2); 289 _nodes.map(end_idx()+2, tmp1); 290 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input 291 tmp1->replace_by(tmp); 292 tmp2->replace_by(tmp1); 293 tmp->replace_by(tmp2); 294 tmp->destruct(); 295 } 296 297 // Remove the existing null check; use a new implicit null check instead. 298 // Since schedule-local needs precise def-use info, we need to correct 299 // it as well. 300 Node *old_tst = proj->in(0); 301 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); 302 _nodes.map(end_idx(),nul_chk); 303 cfg->_bbs.map(nul_chk->_idx,this); 304 // Redirect users of old_test to nul_chk 305 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 306 old_tst->last_out(i2)->set_req(0, nul_chk); 307 // Clean-up any dead code 308 for (uint i3 = 0; i3 < old_tst->req(); i3++) 309 old_tst->set_req(i3, NULL); 310 311 cfg->latency_from_uses(nul_chk); 312 cfg->latency_from_uses(best); 313} 314 315 316//------------------------------select----------------------------------------- 317// Select a nice fellow from the worklist to schedule next. If there is only 318// one choice, then use it. Projections take top priority for correctness 319// reasons - if I see a projection, then it is next. There are a number of 320// other special cases, for instructions that consume condition codes, et al. 321// These are chosen immediately. Some instructions are required to immediately 322// precede the last instruction in the block, and these are taken last. Of the 323// remaining cases (most), choose the instruction with the greatest latency 324// (that is, the most number of pseudo-cycles required to the end of the 325// routine). If there is a tie, choose the instruction with the most inputs. 326Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) { 327 328 // If only a single entry on the stack, use it 329 uint cnt = worklist.size(); 330 if (cnt == 1) { 331 Node *n = worklist[0]; 332 worklist.map(0,worklist.pop()); 333 return n; 334 } 335 336 uint choice = 0; // Bigger is most important 337 uint latency = 0; // Bigger is scheduled first 338 uint score = 0; // Bigger is better 339 int idx = -1; // Index in worklist 340 341 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 342 // Order in worklist is used to break ties. 343 // See caller for how this is used to delay scheduling 344 // of induction variable increments to after the other 345 // uses of the phi are scheduled. 346 Node *n = worklist[i]; // Get Node on worklist 347 348 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 349 if( n->is_Proj() || // Projections always win 350 n->Opcode()== Op_Con || // So does constant 'Top' 351 iop == Op_CreateEx || // Create-exception must start block 352 iop == Op_CheckCastPP 353 ) { 354 worklist.map(i,worklist.pop()); 355 return n; 356 } 357 358 // Final call in a block must be adjacent to 'catch' 359 Node *e = end(); 360 if( e->is_Catch() && e->in(0)->in(0) == n ) 361 continue; 362 363 // Memory op for an implicit null check has to be at the end of the block 364 if( e->is_MachNullCheck() && e->in(1) == n ) 365 continue; 366 367 uint n_choice = 2; 368 369 // See if this instruction is consumed by a branch. If so, then (as the 370 // branch is the last instruction in the basic block) force it to the 371 // end of the basic block 372 if ( must_clone[iop] ) { 373 // See if any use is a branch 374 bool found_machif = false; 375 376 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 377 Node* use = n->fast_out(j); 378 379 // The use is a conditional branch, make them adjacent 380 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) { 381 found_machif = true; 382 break; 383 } 384 385 // More than this instruction pending for successor to be ready, 386 // don't choose this if other opportunities are ready 387 if (ready_cnt[use->_idx] > 1) 388 n_choice = 1; 389 } 390 391 // loop terminated, prefer not to use this instruction 392 if (found_machif) 393 continue; 394 } 395 396 // See if this has a predecessor that is "must_clone", i.e. sets the 397 // condition code. If so, choose this first 398 for (uint j = 0; j < n->req() ; j++) { 399 Node *inn = n->in(j); 400 if (inn) { 401 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 402 n_choice = 3; 403 break; 404 } 405 } 406 } 407 408 // MachTemps should be scheduled last so they are near their uses 409 if (n->is_MachTemp()) { 410 n_choice = 1; 411 } 412 413 uint n_latency = cfg->_node_latency.at_grow(n->_idx); 414 uint n_score = n->req(); // Many inputs get high score to break ties 415 416 // Keep best latency found 417 if( choice < n_choice || 418 ( choice == n_choice && 419 ( latency < n_latency || 420 ( latency == n_latency && 421 ( score < n_score ))))) { 422 choice = n_choice; 423 latency = n_latency; 424 score = n_score; 425 idx = i; // Also keep index in worklist 426 } 427 } // End of for all ready nodes in worklist 428 429 assert(idx >= 0, "index should be set"); 430 Node *n = worklist[(uint)idx]; // Get the winner 431 432 worklist.map((uint)idx, worklist.pop()); // Compress worklist 433 return n; 434} 435 436 437//------------------------------set_next_call---------------------------------- 438void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) { 439 if( next_call.test_set(n->_idx) ) return; 440 for( uint i=0; i<n->len(); i++ ) { 441 Node *m = n->in(i); 442 if( !m ) continue; // must see all nodes in block that precede call 443 if( bbs[m->_idx] == this ) 444 set_next_call( m, next_call, bbs ); 445 } 446} 447 448//------------------------------needed_for_next_call--------------------------- 449// Set the flag 'next_call' for each Node that is needed for the next call to 450// be scheduled. This flag lets me bias scheduling so Nodes needed for the 451// next subroutine call get priority - basically it moves things NOT needed 452// for the next call till after the call. This prevents me from trying to 453// carry lots of stuff live across a call. 454void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) { 455 // Find the next control-defining Node in this block 456 Node* call = NULL; 457 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 458 Node* m = this_call->fast_out(i); 459 if( bbs[m->_idx] == this && // Local-block user 460 m != this_call && // Not self-start node 461 m->is_Call() ) 462 call = m; 463 break; 464 } 465 if (call == NULL) return; // No next call (e.g., block end is near) 466 // Set next-call for all inputs to this call 467 set_next_call(call, next_call, bbs); 468} 469 470//------------------------------sched_call------------------------------------- 471uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) { 472 RegMask regs; 473 474 // Schedule all the users of the call right now. All the users are 475 // projection Nodes, so they must be scheduled next to the call. 476 // Collect all the defined registers. 477 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 478 Node* n = mcall->fast_out(i); 479 assert( n->Opcode()==Op_MachProj, "" ); 480 --ready_cnt[n->_idx]; 481 assert( !ready_cnt[n->_idx], "" ); 482 // Schedule next to call 483 _nodes.map(node_cnt++, n); 484 // Collect defined registers 485 regs.OR(n->out_RegMask()); 486 // Check for scheduling the next control-definer 487 if( n->bottom_type() == Type::CONTROL ) 488 // Warm up next pile of heuristic bits 489 needed_for_next_call(n, next_call, bbs); 490 491 // Children of projections are now all ready 492 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 493 Node* m = n->fast_out(j); // Get user 494 if( bbs[m->_idx] != this ) continue; 495 if( m->is_Phi() ) continue; 496 if( !--ready_cnt[m->_idx] ) 497 worklist.push(m); 498 } 499 500 } 501 502 // Act as if the call defines the Frame Pointer. 503 // Certainly the FP is alive and well after the call. 504 regs.Insert(matcher.c_frame_pointer()); 505 506 // Set all registers killed and not already defined by the call. 507 uint r_cnt = mcall->tf()->range()->cnt(); 508 int op = mcall->ideal_Opcode(); 509 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 510 bbs.map(proj->_idx,this); 511 _nodes.insert(node_cnt++, proj); 512 513 // Select the right register save policy. 514 const char * save_policy; 515 switch (op) { 516 case Op_CallRuntime: 517 case Op_CallLeaf: 518 case Op_CallLeafNoFP: 519 // Calling C code so use C calling convention 520 save_policy = matcher._c_reg_save_policy; 521 break; 522 523 case Op_CallStaticJava: 524 case Op_CallDynamicJava: 525 // Calling Java code so use Java calling convention 526 save_policy = matcher._register_save_policy; 527 break; 528 529 default: 530 ShouldNotReachHere(); 531 } 532 533 // When using CallRuntime mark SOE registers as killed by the call 534 // so values that could show up in the RegisterMap aren't live in a 535 // callee saved register since the register wouldn't know where to 536 // find them. CallLeaf and CallLeafNoFP are ok because they can't 537 // have debug info on them. Strictly speaking this only needs to be 538 // done for oops since idealreg2debugmask takes care of debug info 539 // references but there no way to handle oops differently than other 540 // pointers as far as the kill mask goes. 541 bool exclude_soe = op == Op_CallRuntime; 542 543 // Fill in the kill mask for the call 544 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 545 if( !regs.Member(r) ) { // Not already defined by the call 546 // Save-on-call register? 547 if ((save_policy[r] == 'C') || 548 (save_policy[r] == 'A') || 549 ((save_policy[r] == 'E') && exclude_soe)) { 550 proj->_rout.Insert(r); 551 } 552 } 553 } 554 555 return node_cnt; 556} 557 558 559//------------------------------schedule_local--------------------------------- 560// Topological sort within a block. Someday become a real scheduler. 561bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) { 562 // Already "sorted" are the block start Node (as the first entry), and 563 // the block-ending Node and any trailing control projections. We leave 564 // these alone. PhiNodes and ParmNodes are made to follow the block start 565 // Node. Everything else gets topo-sorted. 566 567#ifndef PRODUCT 568 if (cfg->trace_opto_pipelining()) { 569 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order); 570 for (uint i = 0;i < _nodes.size();i++) { 571 tty->print("# "); 572 _nodes[i]->fast_dump(); 573 } 574 tty->print_cr("#"); 575 } 576#endif 577 578 // RootNode is already sorted 579 if( _nodes.size() == 1 ) return true; 580 581 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 582 uint node_cnt = end_idx(); 583 uint phi_cnt = 1; 584 uint i; 585 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 586 Node *n = _nodes[i]; 587 if( n->is_Phi() || // Found a PhiNode or ParmNode 588 (n->is_Proj() && n->in(0) == head()) ) { 589 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 590 _nodes.map(i,_nodes[phi_cnt]); 591 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front 592 } else { // All others 593 // Count block-local inputs to 'n' 594 uint cnt = n->len(); // Input count 595 uint local = 0; 596 for( uint j=0; j<cnt; j++ ) { 597 Node *m = n->in(j); 598 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() ) 599 local++; // One more block-local input 600 } 601 ready_cnt[n->_idx] = local; // Count em up 602 603 // A few node types require changing a required edge to a precedence edge 604 // before allocation. 605 if( UseConcMarkSweepGC || UseG1GC ) { 606 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 607 // Note: Required edges with an index greater than oper_input_base 608 // are not supported by the allocator. 609 // Note2: Can only depend on unmatched edge being last, 610 // can not depend on its absolute position. 611 Node *oop_store = n->in(n->req() - 1); 612 n->del_req(n->req() - 1); 613 n->add_prec(oop_store); 614 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark"); 615 } 616 } 617 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire && 618 n->req() > TypeFunc::Parms ) { 619 // MemBarAcquire could be created without Precedent edge. 620 // del_req() replaces the specified edge with the last input edge 621 // and then removes the last edge. If the specified edge > number of 622 // edges the last edge will be moved outside of the input edges array 623 // and the edge will be lost. This is why this code should be 624 // executed only when Precedent (== TypeFunc::Parms) edge is present. 625 Node *x = n->in(TypeFunc::Parms); 626 n->del_req(TypeFunc::Parms); 627 n->add_prec(x); 628 } 629 } 630 } 631 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count 632 ready_cnt[_nodes[i2]->_idx] = 0; 633 634 // All the prescheduled guys do not hold back internal nodes 635 uint i3; 636 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 637 Node *n = _nodes[i3]; // Get pre-scheduled 638 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 639 Node* m = n->fast_out(j); 640 if( cfg->_bbs[m->_idx] ==this ) // Local-block user 641 ready_cnt[m->_idx]--; // Fix ready count 642 } 643 } 644 645 Node_List delay; 646 // Make a worklist 647 Node_List worklist; 648 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 649 Node *m = _nodes[i4]; 650 if( !ready_cnt[m->_idx] ) { // Zero ready count? 651 if (m->is_iteratively_computed()) { 652 // Push induction variable increments last to allow other uses 653 // of the phi to be scheduled first. The select() method breaks 654 // ties in scheduling by worklist order. 655 delay.push(m); 656 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 657 // Force the CreateEx to the top of the list so it's processed 658 // first and ends up at the start of the block. 659 worklist.insert(0, m); 660 } else { 661 worklist.push(m); // Then on to worklist! 662 } 663 } 664 } 665 while (delay.size()) { 666 Node* d = delay.pop(); 667 worklist.push(d); 668 } 669 670 // Warm up the 'next_call' heuristic bits 671 needed_for_next_call(_nodes[0], next_call, cfg->_bbs); 672 673#ifndef PRODUCT 674 if (cfg->trace_opto_pipelining()) { 675 for (uint j=0; j<_nodes.size(); j++) { 676 Node *n = _nodes[j]; 677 int idx = n->_idx; 678 tty->print("# ready cnt:%3d ", ready_cnt[idx]); 679 tty->print("latency:%3d ", cfg->_node_latency.at_grow(idx)); 680 tty->print("%4d: %s\n", idx, n->Name()); 681 } 682 } 683#endif 684 685 // Pull from worklist and schedule 686 while( worklist.size() ) { // Worklist is not ready 687 688#ifndef PRODUCT 689 if (cfg->trace_opto_pipelining()) { 690 tty->print("# ready list:"); 691 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 692 Node *n = worklist[i]; // Get Node on worklist 693 tty->print(" %d", n->_idx); 694 } 695 tty->cr(); 696 } 697#endif 698 699 // Select and pop a ready guy from worklist 700 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt); 701 _nodes.map(phi_cnt++,n); // Schedule him next 702 703#ifndef PRODUCT 704 if (cfg->trace_opto_pipelining()) { 705 tty->print("# select %d: %s", n->_idx, n->Name()); 706 tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx)); 707 n->dump(); 708 if (Verbose) { 709 tty->print("# ready list:"); 710 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 711 Node *n = worklist[i]; // Get Node on worklist 712 tty->print(" %d", n->_idx); 713 } 714 tty->cr(); 715 } 716 } 717 718#endif 719 if( n->is_MachCall() ) { 720 MachCallNode *mcall = n->as_MachCall(); 721 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call); 722 continue; 723 } 724 // Children are now all ready 725 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 726 Node* m = n->fast_out(i5); // Get user 727 if( cfg->_bbs[m->_idx] != this ) continue; 728 if( m->is_Phi() ) continue; 729 if( !--ready_cnt[m->_idx] ) 730 worklist.push(m); 731 } 732 } 733 734 if( phi_cnt != end_idx() ) { 735 // did not schedule all. Retry, Bailout, or Die 736 Compile* C = matcher.C; 737 if (C->subsume_loads() == true && !C->failing()) { 738 // Retry with subsume_loads == false 739 // If this is the first failure, the sentinel string will "stick" 740 // to the Compile object, and the C2Compiler will see it and retry. 741 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 742 } 743 // assert( phi_cnt == end_idx(), "did not schedule all" ); 744 return false; 745 } 746 747#ifndef PRODUCT 748 if (cfg->trace_opto_pipelining()) { 749 tty->print_cr("#"); 750 tty->print_cr("# after schedule_local"); 751 for (uint i = 0;i < _nodes.size();i++) { 752 tty->print("# "); 753 _nodes[i]->fast_dump(); 754 } 755 tty->cr(); 756 } 757#endif 758 759 760 return true; 761} 762 763//--------------------------catch_cleanup_fix_all_inputs----------------------- 764static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 765 for (uint l = 0; l < use->len(); l++) { 766 if (use->in(l) == old_def) { 767 if (l < use->req()) { 768 use->set_req(l, new_def); 769 } else { 770 use->rm_prec(l); 771 use->add_prec(new_def); 772 l--; 773 } 774 } 775 } 776} 777 778//------------------------------catch_cleanup_find_cloned_def------------------ 779static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { 780 assert( use_blk != def_blk, "Inter-block cleanup only"); 781 782 // The use is some block below the Catch. Find and return the clone of the def 783 // that dominates the use. If there is no clone in a dominating block, then 784 // create a phi for the def in a dominating block. 785 786 // Find which successor block dominates this use. The successor 787 // blocks must all be single-entry (from the Catch only; I will have 788 // split blocks to make this so), hence they all dominate. 789 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 790 use_blk = use_blk->_idom; 791 792 // Find the successor 793 Node *fixup = NULL; 794 795 uint j; 796 for( j = 0; j < def_blk->_num_succs; j++ ) 797 if( use_blk == def_blk->_succs[j] ) 798 break; 799 800 if( j == def_blk->_num_succs ) { 801 // Block at same level in dom-tree is not a successor. It needs a 802 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 803 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 804 for(uint k = 1; k < use_blk->num_preds(); k++) { 805 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx)); 806 } 807 808 // Check to see if the use_blk already has an identical phi inserted. 809 // If it exists, it will be at the first position since all uses of a 810 // def are processed together. 811 Node *phi = use_blk->_nodes[1]; 812 if( phi->is_Phi() ) { 813 fixup = phi; 814 for (uint k = 1; k < use_blk->num_preds(); k++) { 815 if (phi->in(k) != inputs[k]) { 816 // Not a match 817 fixup = NULL; 818 break; 819 } 820 } 821 } 822 823 // If an existing PhiNode was not found, make a new one. 824 if (fixup == NULL) { 825 Node *new_phi = PhiNode::make(use_blk->head(), def); 826 use_blk->_nodes.insert(1, new_phi); 827 bbs.map(new_phi->_idx, use_blk); 828 for (uint k = 1; k < use_blk->num_preds(); k++) { 829 new_phi->set_req(k, inputs[k]); 830 } 831 fixup = new_phi; 832 } 833 834 } else { 835 // Found the use just below the Catch. Make it use the clone. 836 fixup = use_blk->_nodes[n_clone_idx]; 837 } 838 839 return fixup; 840} 841 842//--------------------------catch_cleanup_intra_block-------------------------- 843// Fix all input edges in use that reference "def". The use is in the same 844// block as the def and both have been cloned in each successor block. 845static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 846 847 // Both the use and def have been cloned. For each successor block, 848 // get the clone of the use, and make its input the clone of the def 849 // found in that block. 850 851 uint use_idx = blk->find_node(use); 852 uint offset_idx = use_idx - beg; 853 for( uint k = 0; k < blk->_num_succs; k++ ) { 854 // Get clone in each successor block 855 Block *sb = blk->_succs[k]; 856 Node *clone = sb->_nodes[offset_idx+1]; 857 assert( clone->Opcode() == use->Opcode(), "" ); 858 859 // Make use-clone reference the def-clone 860 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]); 861 } 862} 863 864//------------------------------catch_cleanup_inter_block--------------------- 865// Fix all input edges in use that reference "def". The use is in a different 866// block than the def. 867static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { 868 if( !use_blk ) return; // Can happen if the use is a precedence edge 869 870 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx); 871 catch_cleanup_fix_all_inputs(use, def, new_def); 872} 873 874//------------------------------call_catch_cleanup----------------------------- 875// If we inserted any instructions between a Call and his CatchNode, 876// clone the instructions on all paths below the Catch. 877void Block::call_catch_cleanup(Block_Array &bbs) { 878 879 // End of region to clone 880 uint end = end_idx(); 881 if( !_nodes[end]->is_Catch() ) return; 882 // Start of region to clone 883 uint beg = end; 884 while( _nodes[beg-1]->Opcode() != Op_MachProj || 885 !_nodes[beg-1]->in(0)->is_Call() ) { 886 beg--; 887 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 888 } 889 // Range of inserted instructions is [beg, end) 890 if( beg == end ) return; 891 892 // Clone along all Catch output paths. Clone area between the 'beg' and 893 // 'end' indices. 894 for( uint i = 0; i < _num_succs; i++ ) { 895 Block *sb = _succs[i]; 896 // Clone the entire area; ignoring the edge fixup for now. 897 for( uint j = end; j > beg; j-- ) { 898 Node *clone = _nodes[j-1]->clone(); 899 sb->_nodes.insert( 1, clone ); 900 bbs.map(clone->_idx,sb); 901 } 902 } 903 904 905 // Fixup edges. Check the def-use info per cloned Node 906 for(uint i2 = beg; i2 < end; i2++ ) { 907 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 908 Node *n = _nodes[i2]; // Node that got cloned 909 // Need DU safe iterator because of edge manipulation in calls. 910 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 911 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 912 out->push(n->fast_out(j1)); 913 } 914 uint max = out->size(); 915 for (uint j = 0; j < max; j++) {// For all users 916 Node *use = out->pop(); 917 Block *buse = bbs[use->_idx]; 918 if( use->is_Phi() ) { 919 for( uint k = 1; k < use->req(); k++ ) 920 if( use->in(k) == n ) { 921 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx); 922 use->set_req(k, fixup); 923 } 924 } else { 925 if (this == buse) { 926 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx); 927 } else { 928 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx); 929 } 930 } 931 } // End for all users 932 933 } // End of for all Nodes in cloned area 934 935 // Remove the now-dead cloned ops 936 for(uint i3 = beg; i3 < end; i3++ ) { 937 _nodes[beg]->disconnect_inputs(NULL); 938 _nodes.remove(beg); 939 } 940 941 // If the successor blocks have a CreateEx node, move it back to the top 942 for(uint i4 = 0; i4 < _num_succs; i4++ ) { 943 Block *sb = _succs[i4]; 944 uint new_cnt = end - beg; 945 // Remove any newly created, but dead, nodes. 946 for( uint j = new_cnt; j > 0; j-- ) { 947 Node *n = sb->_nodes[j]; 948 if (n->outcnt() == 0 && 949 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 950 n->disconnect_inputs(NULL); 951 sb->_nodes.remove(j); 952 new_cnt--; 953 } 954 } 955 // If any newly created nodes remain, move the CreateEx node to the top 956 if (new_cnt > 0) { 957 Node *cex = sb->_nodes[1+new_cnt]; 958 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 959 sb->_nodes.remove(1+new_cnt); 960 sb->_nodes.insert(1,cex); 961 } 962 } 963 } 964} 965