lcm.cpp revision 579:0fbdb4381b99
1/* 2 * Copyright 1998-2009 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 = base->bottom_type()->is_ptr(); 162 // Give up if offset is not a compile-time constant 163 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 164 continue; 165 offset += tptr->_offset; // correct if base is offseted 166 if( MacroAssembler::needs_explicit_null_check(offset) ) 167 continue; // Give up is reference is beyond 4K page size 168 } 169 } 170 171 // Check ctrl input to see if the null-check dominates the memory op 172 Block *cb = cfg->_bbs[mach->_idx]; 173 cb = cb->_idom; // Always hoist at least 1 block 174 if( !was_store ) { // Stores can be hoisted only one block 175 while( cb->_dom_depth > (_dom_depth + 1)) 176 cb = cb->_idom; // Hoist loads as far as we want 177 // The non-null-block should dominate the memory op, too. Live 178 // range spilling will insert a spill in the non-null-block if it is 179 // needs to spill the memory op for an implicit null check. 180 if (cb->_dom_depth == (_dom_depth + 1)) { 181 if (cb != not_null_block) continue; 182 cb = cb->_idom; 183 } 184 } 185 if( cb != this ) continue; 186 187 // Found a memory user; see if it can be hoisted to check-block 188 uint vidx = 0; // Capture index of value into memop 189 uint j; 190 for( j = mach->req()-1; j > 0; j-- ) { 191 if( mach->in(j) == val ) vidx = j; 192 // Block of memory-op input 193 Block *inb = cfg->_bbs[mach->in(j)->_idx]; 194 Block *b = this; // Start from nul check 195 while( b != inb && b->_dom_depth > inb->_dom_depth ) 196 b = b->_idom; // search upwards for input 197 // See if input dominates null check 198 if( b != inb ) 199 break; 200 } 201 if( j > 0 ) 202 continue; 203 Block *mb = cfg->_bbs[mach->_idx]; 204 // Hoisting stores requires more checks for the anti-dependence case. 205 // Give up hoisting if we have to move the store past any load. 206 if( was_store ) { 207 Block *b = mb; // Start searching here for a local load 208 // mach use (faulting) trying to hoist 209 // n might be blocker to hoisting 210 while( b != this ) { 211 uint k; 212 for( k = 1; k < b->_nodes.size(); k++ ) { 213 Node *n = b->_nodes[k]; 214 if( n->needs_anti_dependence_check() && 215 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 216 break; // Found anti-dependent load 217 } 218 if( k < b->_nodes.size() ) 219 break; // Found anti-dependent load 220 // Make sure control does not do a merge (would have to check allpaths) 221 if( b->num_preds() != 2 ) break; 222 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block 223 } 224 if( b != this ) continue; 225 } 226 227 // Make sure this memory op is not already being used for a NullCheck 228 Node *e = mb->end(); 229 if( e->is_MachNullCheck() && e->in(1) == mach ) 230 continue; // Already being used as a NULL check 231 232 // Found a candidate! Pick one with least dom depth - the highest 233 // in the dom tree should be closest to the null check. 234 if( !best || 235 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) { 236 best = mach; 237 bidx = vidx; 238 239 } 240 } 241 // No candidate! 242 if( !best ) return; 243 244 // ---- Found an implicit null check 245 extern int implicit_null_checks; 246 implicit_null_checks++; 247 248 // Hoist the memory candidate up to the end of the test block. 249 Block *old_block = cfg->_bbs[best->_idx]; 250 old_block->find_remove(best); 251 add_inst(best); 252 cfg->_bbs.map(best->_idx,this); 253 254 // Move the control dependence 255 if (best->in(0) && best->in(0) == old_block->_nodes[0]) 256 best->set_req(0, _nodes[0]); 257 258 // Check for flag-killing projections that also need to be hoisted 259 // Should be DU safe because no edge updates. 260 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 261 Node* n = best->fast_out(j); 262 if( n->Opcode() == Op_MachProj ) { 263 cfg->_bbs[n->_idx]->find_remove(n); 264 add_inst(n); 265 cfg->_bbs.map(n->_idx,this); 266 } 267 } 268 269 Compile *C = cfg->C; 270 // proj==Op_True --> ne test; proj==Op_False --> eq test. 271 // One of two graph shapes got matched: 272 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 273 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 274 // NULL checks are always branch-if-eq. If we see a IfTrue projection 275 // then we are replacing a 'ne' test with a 'eq' NULL check test. 276 // We need to flip the projections to keep the same semantics. 277 if( proj->Opcode() == Op_IfTrue ) { 278 // Swap order of projections in basic block to swap branch targets 279 Node *tmp1 = _nodes[end_idx()+1]; 280 Node *tmp2 = _nodes[end_idx()+2]; 281 _nodes.map(end_idx()+1, tmp2); 282 _nodes.map(end_idx()+2, tmp1); 283 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input 284 tmp1->replace_by(tmp); 285 tmp2->replace_by(tmp1); 286 tmp->replace_by(tmp2); 287 tmp->destruct(); 288 } 289 290 // Remove the existing null check; use a new implicit null check instead. 291 // Since schedule-local needs precise def-use info, we need to correct 292 // it as well. 293 Node *old_tst = proj->in(0); 294 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); 295 _nodes.map(end_idx(),nul_chk); 296 cfg->_bbs.map(nul_chk->_idx,this); 297 // Redirect users of old_test to nul_chk 298 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 299 old_tst->last_out(i2)->set_req(0, nul_chk); 300 // Clean-up any dead code 301 for (uint i3 = 0; i3 < old_tst->req(); i3++) 302 old_tst->set_req(i3, NULL); 303 304 cfg->latency_from_uses(nul_chk); 305 cfg->latency_from_uses(best); 306} 307 308 309//------------------------------select----------------------------------------- 310// Select a nice fellow from the worklist to schedule next. If there is only 311// one choice, then use it. Projections take top priority for correctness 312// reasons - if I see a projection, then it is next. There are a number of 313// other special cases, for instructions that consume condition codes, et al. 314// These are chosen immediately. Some instructions are required to immediately 315// precede the last instruction in the block, and these are taken last. Of the 316// remaining cases (most), choose the instruction with the greatest latency 317// (that is, the most number of pseudo-cycles required to the end of the 318// routine). If there is a tie, choose the instruction with the most inputs. 319Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) { 320 321 // If only a single entry on the stack, use it 322 uint cnt = worklist.size(); 323 if (cnt == 1) { 324 Node *n = worklist[0]; 325 worklist.map(0,worklist.pop()); 326 return n; 327 } 328 329 uint choice = 0; // Bigger is most important 330 uint latency = 0; // Bigger is scheduled first 331 uint score = 0; // Bigger is better 332 int idx = -1; // Index in worklist 333 334 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 335 // Order in worklist is used to break ties. 336 // See caller for how this is used to delay scheduling 337 // of induction variable increments to after the other 338 // uses of the phi are scheduled. 339 Node *n = worklist[i]; // Get Node on worklist 340 341 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 342 if( n->is_Proj() || // Projections always win 343 n->Opcode()== Op_Con || // So does constant 'Top' 344 iop == Op_CreateEx || // Create-exception must start block 345 iop == Op_CheckCastPP 346 ) { 347 worklist.map(i,worklist.pop()); 348 return n; 349 } 350 351 // Final call in a block must be adjacent to 'catch' 352 Node *e = end(); 353 if( e->is_Catch() && e->in(0)->in(0) == n ) 354 continue; 355 356 // Memory op for an implicit null check has to be at the end of the block 357 if( e->is_MachNullCheck() && e->in(1) == n ) 358 continue; 359 360 uint n_choice = 2; 361 362 // See if this instruction is consumed by a branch. If so, then (as the 363 // branch is the last instruction in the basic block) force it to the 364 // end of the basic block 365 if ( must_clone[iop] ) { 366 // See if any use is a branch 367 bool found_machif = false; 368 369 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 370 Node* use = n->fast_out(j); 371 372 // The use is a conditional branch, make them adjacent 373 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) { 374 found_machif = true; 375 break; 376 } 377 378 // More than this instruction pending for successor to be ready, 379 // don't choose this if other opportunities are ready 380 if (ready_cnt[use->_idx] > 1) 381 n_choice = 1; 382 } 383 384 // loop terminated, prefer not to use this instruction 385 if (found_machif) 386 continue; 387 } 388 389 // See if this has a predecessor that is "must_clone", i.e. sets the 390 // condition code. If so, choose this first 391 for (uint j = 0; j < n->req() ; j++) { 392 Node *inn = n->in(j); 393 if (inn) { 394 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 395 n_choice = 3; 396 break; 397 } 398 } 399 } 400 401 // MachTemps should be scheduled last so they are near their uses 402 if (n->is_MachTemp()) { 403 n_choice = 1; 404 } 405 406 uint n_latency = cfg->_node_latency.at_grow(n->_idx); 407 uint n_score = n->req(); // Many inputs get high score to break ties 408 409 // Keep best latency found 410 if( choice < n_choice || 411 ( choice == n_choice && 412 ( latency < n_latency || 413 ( latency == n_latency && 414 ( score < n_score ))))) { 415 choice = n_choice; 416 latency = n_latency; 417 score = n_score; 418 idx = i; // Also keep index in worklist 419 } 420 } // End of for all ready nodes in worklist 421 422 assert(idx >= 0, "index should be set"); 423 Node *n = worklist[(uint)idx]; // Get the winner 424 425 worklist.map((uint)idx, worklist.pop()); // Compress worklist 426 return n; 427} 428 429 430//------------------------------set_next_call---------------------------------- 431void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) { 432 if( next_call.test_set(n->_idx) ) return; 433 for( uint i=0; i<n->len(); i++ ) { 434 Node *m = n->in(i); 435 if( !m ) continue; // must see all nodes in block that precede call 436 if( bbs[m->_idx] == this ) 437 set_next_call( m, next_call, bbs ); 438 } 439} 440 441//------------------------------needed_for_next_call--------------------------- 442// Set the flag 'next_call' for each Node that is needed for the next call to 443// be scheduled. This flag lets me bias scheduling so Nodes needed for the 444// next subroutine call get priority - basically it moves things NOT needed 445// for the next call till after the call. This prevents me from trying to 446// carry lots of stuff live across a call. 447void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) { 448 // Find the next control-defining Node in this block 449 Node* call = NULL; 450 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 451 Node* m = this_call->fast_out(i); 452 if( bbs[m->_idx] == this && // Local-block user 453 m != this_call && // Not self-start node 454 m->is_Call() ) 455 call = m; 456 break; 457 } 458 if (call == NULL) return; // No next call (e.g., block end is near) 459 // Set next-call for all inputs to this call 460 set_next_call(call, next_call, bbs); 461} 462 463//------------------------------sched_call------------------------------------- 464uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) { 465 RegMask regs; 466 467 // Schedule all the users of the call right now. All the users are 468 // projection Nodes, so they must be scheduled next to the call. 469 // Collect all the defined registers. 470 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 471 Node* n = mcall->fast_out(i); 472 assert( n->Opcode()==Op_MachProj, "" ); 473 --ready_cnt[n->_idx]; 474 assert( !ready_cnt[n->_idx], "" ); 475 // Schedule next to call 476 _nodes.map(node_cnt++, n); 477 // Collect defined registers 478 regs.OR(n->out_RegMask()); 479 // Check for scheduling the next control-definer 480 if( n->bottom_type() == Type::CONTROL ) 481 // Warm up next pile of heuristic bits 482 needed_for_next_call(n, next_call, bbs); 483 484 // Children of projections are now all ready 485 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 486 Node* m = n->fast_out(j); // Get user 487 if( bbs[m->_idx] != this ) continue; 488 if( m->is_Phi() ) continue; 489 if( !--ready_cnt[m->_idx] ) 490 worklist.push(m); 491 } 492 493 } 494 495 // Act as if the call defines the Frame Pointer. 496 // Certainly the FP is alive and well after the call. 497 regs.Insert(matcher.c_frame_pointer()); 498 499 // Set all registers killed and not already defined by the call. 500 uint r_cnt = mcall->tf()->range()->cnt(); 501 int op = mcall->ideal_Opcode(); 502 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 503 bbs.map(proj->_idx,this); 504 _nodes.insert(node_cnt++, proj); 505 506 // Select the right register save policy. 507 const char * save_policy; 508 switch (op) { 509 case Op_CallRuntime: 510 case Op_CallLeaf: 511 case Op_CallLeafNoFP: 512 // Calling C code so use C calling convention 513 save_policy = matcher._c_reg_save_policy; 514 break; 515 516 case Op_CallStaticJava: 517 case Op_CallDynamicJava: 518 // Calling Java code so use Java calling convention 519 save_policy = matcher._register_save_policy; 520 break; 521 522 default: 523 ShouldNotReachHere(); 524 } 525 526 // When using CallRuntime mark SOE registers as killed by the call 527 // so values that could show up in the RegisterMap aren't live in a 528 // callee saved register since the register wouldn't know where to 529 // find them. CallLeaf and CallLeafNoFP are ok because they can't 530 // have debug info on them. Strictly speaking this only needs to be 531 // done for oops since idealreg2debugmask takes care of debug info 532 // references but there no way to handle oops differently than other 533 // pointers as far as the kill mask goes. 534 bool exclude_soe = op == Op_CallRuntime; 535 536 // Fill in the kill mask for the call 537 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 538 if( !regs.Member(r) ) { // Not already defined by the call 539 // Save-on-call register? 540 if ((save_policy[r] == 'C') || 541 (save_policy[r] == 'A') || 542 ((save_policy[r] == 'E') && exclude_soe)) { 543 proj->_rout.Insert(r); 544 } 545 } 546 } 547 548 return node_cnt; 549} 550 551 552//------------------------------schedule_local--------------------------------- 553// Topological sort within a block. Someday become a real scheduler. 554bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) { 555 // Already "sorted" are the block start Node (as the first entry), and 556 // the block-ending Node and any trailing control projections. We leave 557 // these alone. PhiNodes and ParmNodes are made to follow the block start 558 // Node. Everything else gets topo-sorted. 559 560#ifndef PRODUCT 561 if (cfg->trace_opto_pipelining()) { 562 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order); 563 for (uint i = 0;i < _nodes.size();i++) { 564 tty->print("# "); 565 _nodes[i]->fast_dump(); 566 } 567 tty->print_cr("#"); 568 } 569#endif 570 571 // RootNode is already sorted 572 if( _nodes.size() == 1 ) return true; 573 574 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 575 uint node_cnt = end_idx(); 576 uint phi_cnt = 1; 577 uint i; 578 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 579 Node *n = _nodes[i]; 580 if( n->is_Phi() || // Found a PhiNode or ParmNode 581 (n->is_Proj() && n->in(0) == head()) ) { 582 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 583 _nodes.map(i,_nodes[phi_cnt]); 584 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front 585 } else { // All others 586 // Count block-local inputs to 'n' 587 uint cnt = n->len(); // Input count 588 uint local = 0; 589 for( uint j=0; j<cnt; j++ ) { 590 Node *m = n->in(j); 591 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() ) 592 local++; // One more block-local input 593 } 594 ready_cnt[n->_idx] = local; // Count em up 595 596 // A few node types require changing a required edge to a precedence edge 597 // before allocation. 598 if( UseConcMarkSweepGC || UseG1GC ) { 599 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 600 // Note: Required edges with an index greater than oper_input_base 601 // are not supported by the allocator. 602 // Note2: Can only depend on unmatched edge being last, 603 // can not depend on its absolute position. 604 Node *oop_store = n->in(n->req() - 1); 605 n->del_req(n->req() - 1); 606 n->add_prec(oop_store); 607 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark"); 608 } 609 } 610 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire && 611 n->req() > TypeFunc::Parms ) { 612 // MemBarAcquire could be created without Precedent edge. 613 // del_req() replaces the specified edge with the last input edge 614 // and then removes the last edge. If the specified edge > number of 615 // edges the last edge will be moved outside of the input edges array 616 // and the edge will be lost. This is why this code should be 617 // executed only when Precedent (== TypeFunc::Parms) edge is present. 618 Node *x = n->in(TypeFunc::Parms); 619 n->del_req(TypeFunc::Parms); 620 n->add_prec(x); 621 } 622 } 623 } 624 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count 625 ready_cnt[_nodes[i2]->_idx] = 0; 626 627 // All the prescheduled guys do not hold back internal nodes 628 uint i3; 629 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 630 Node *n = _nodes[i3]; // Get pre-scheduled 631 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 632 Node* m = n->fast_out(j); 633 if( cfg->_bbs[m->_idx] ==this ) // Local-block user 634 ready_cnt[m->_idx]--; // Fix ready count 635 } 636 } 637 638 Node_List delay; 639 // Make a worklist 640 Node_List worklist; 641 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 642 Node *m = _nodes[i4]; 643 if( !ready_cnt[m->_idx] ) { // Zero ready count? 644 if (m->is_iteratively_computed()) { 645 // Push induction variable increments last to allow other uses 646 // of the phi to be scheduled first. The select() method breaks 647 // ties in scheduling by worklist order. 648 delay.push(m); 649 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 650 // Force the CreateEx to the top of the list so it's processed 651 // first and ends up at the start of the block. 652 worklist.insert(0, m); 653 } else { 654 worklist.push(m); // Then on to worklist! 655 } 656 } 657 } 658 while (delay.size()) { 659 Node* d = delay.pop(); 660 worklist.push(d); 661 } 662 663 // Warm up the 'next_call' heuristic bits 664 needed_for_next_call(_nodes[0], next_call, cfg->_bbs); 665 666#ifndef PRODUCT 667 if (cfg->trace_opto_pipelining()) { 668 for (uint j=0; j<_nodes.size(); j++) { 669 Node *n = _nodes[j]; 670 int idx = n->_idx; 671 tty->print("# ready cnt:%3d ", ready_cnt[idx]); 672 tty->print("latency:%3d ", cfg->_node_latency.at_grow(idx)); 673 tty->print("%4d: %s\n", idx, n->Name()); 674 } 675 } 676#endif 677 678 // Pull from worklist and schedule 679 while( worklist.size() ) { // Worklist is not ready 680 681#ifndef PRODUCT 682 if (cfg->trace_opto_pipelining()) { 683 tty->print("# ready list:"); 684 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 685 Node *n = worklist[i]; // Get Node on worklist 686 tty->print(" %d", n->_idx); 687 } 688 tty->cr(); 689 } 690#endif 691 692 // Select and pop a ready guy from worklist 693 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt); 694 _nodes.map(phi_cnt++,n); // Schedule him next 695 696#ifndef PRODUCT 697 if (cfg->trace_opto_pipelining()) { 698 tty->print("# select %d: %s", n->_idx, n->Name()); 699 tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx)); 700 n->dump(); 701 if (Verbose) { 702 tty->print("# ready list:"); 703 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 704 Node *n = worklist[i]; // Get Node on worklist 705 tty->print(" %d", n->_idx); 706 } 707 tty->cr(); 708 } 709 } 710 711#endif 712 if( n->is_MachCall() ) { 713 MachCallNode *mcall = n->as_MachCall(); 714 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call); 715 continue; 716 } 717 // Children are now all ready 718 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 719 Node* m = n->fast_out(i5); // Get user 720 if( cfg->_bbs[m->_idx] != this ) continue; 721 if( m->is_Phi() ) continue; 722 if( !--ready_cnt[m->_idx] ) 723 worklist.push(m); 724 } 725 } 726 727 if( phi_cnt != end_idx() ) { 728 // did not schedule all. Retry, Bailout, or Die 729 Compile* C = matcher.C; 730 if (C->subsume_loads() == true && !C->failing()) { 731 // Retry with subsume_loads == false 732 // If this is the first failure, the sentinel string will "stick" 733 // to the Compile object, and the C2Compiler will see it and retry. 734 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 735 } 736 // assert( phi_cnt == end_idx(), "did not schedule all" ); 737 return false; 738 } 739 740#ifndef PRODUCT 741 if (cfg->trace_opto_pipelining()) { 742 tty->print_cr("#"); 743 tty->print_cr("# after schedule_local"); 744 for (uint i = 0;i < _nodes.size();i++) { 745 tty->print("# "); 746 _nodes[i]->fast_dump(); 747 } 748 tty->cr(); 749 } 750#endif 751 752 753 return true; 754} 755 756//--------------------------catch_cleanup_fix_all_inputs----------------------- 757static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 758 for (uint l = 0; l < use->len(); l++) { 759 if (use->in(l) == old_def) { 760 if (l < use->req()) { 761 use->set_req(l, new_def); 762 } else { 763 use->rm_prec(l); 764 use->add_prec(new_def); 765 l--; 766 } 767 } 768 } 769} 770 771//------------------------------catch_cleanup_find_cloned_def------------------ 772static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { 773 assert( use_blk != def_blk, "Inter-block cleanup only"); 774 775 // The use is some block below the Catch. Find and return the clone of the def 776 // that dominates the use. If there is no clone in a dominating block, then 777 // create a phi for the def in a dominating block. 778 779 // Find which successor block dominates this use. The successor 780 // blocks must all be single-entry (from the Catch only; I will have 781 // split blocks to make this so), hence they all dominate. 782 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 783 use_blk = use_blk->_idom; 784 785 // Find the successor 786 Node *fixup = NULL; 787 788 uint j; 789 for( j = 0; j < def_blk->_num_succs; j++ ) 790 if( use_blk == def_blk->_succs[j] ) 791 break; 792 793 if( j == def_blk->_num_succs ) { 794 // Block at same level in dom-tree is not a successor. It needs a 795 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 796 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 797 for(uint k = 1; k < use_blk->num_preds(); k++) { 798 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx)); 799 } 800 801 // Check to see if the use_blk already has an identical phi inserted. 802 // If it exists, it will be at the first position since all uses of a 803 // def are processed together. 804 Node *phi = use_blk->_nodes[1]; 805 if( phi->is_Phi() ) { 806 fixup = phi; 807 for (uint k = 1; k < use_blk->num_preds(); k++) { 808 if (phi->in(k) != inputs[k]) { 809 // Not a match 810 fixup = NULL; 811 break; 812 } 813 } 814 } 815 816 // If an existing PhiNode was not found, make a new one. 817 if (fixup == NULL) { 818 Node *new_phi = PhiNode::make(use_blk->head(), def); 819 use_blk->_nodes.insert(1, new_phi); 820 bbs.map(new_phi->_idx, use_blk); 821 for (uint k = 1; k < use_blk->num_preds(); k++) { 822 new_phi->set_req(k, inputs[k]); 823 } 824 fixup = new_phi; 825 } 826 827 } else { 828 // Found the use just below the Catch. Make it use the clone. 829 fixup = use_blk->_nodes[n_clone_idx]; 830 } 831 832 return fixup; 833} 834 835//--------------------------catch_cleanup_intra_block-------------------------- 836// Fix all input edges in use that reference "def". The use is in the same 837// block as the def and both have been cloned in each successor block. 838static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 839 840 // Both the use and def have been cloned. For each successor block, 841 // get the clone of the use, and make its input the clone of the def 842 // found in that block. 843 844 uint use_idx = blk->find_node(use); 845 uint offset_idx = use_idx - beg; 846 for( uint k = 0; k < blk->_num_succs; k++ ) { 847 // Get clone in each successor block 848 Block *sb = blk->_succs[k]; 849 Node *clone = sb->_nodes[offset_idx+1]; 850 assert( clone->Opcode() == use->Opcode(), "" ); 851 852 // Make use-clone reference the def-clone 853 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]); 854 } 855} 856 857//------------------------------catch_cleanup_inter_block--------------------- 858// Fix all input edges in use that reference "def". The use is in a different 859// block than the def. 860static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { 861 if( !use_blk ) return; // Can happen if the use is a precedence edge 862 863 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx); 864 catch_cleanup_fix_all_inputs(use, def, new_def); 865} 866 867//------------------------------call_catch_cleanup----------------------------- 868// If we inserted any instructions between a Call and his CatchNode, 869// clone the instructions on all paths below the Catch. 870void Block::call_catch_cleanup(Block_Array &bbs) { 871 872 // End of region to clone 873 uint end = end_idx(); 874 if( !_nodes[end]->is_Catch() ) return; 875 // Start of region to clone 876 uint beg = end; 877 while( _nodes[beg-1]->Opcode() != Op_MachProj || 878 !_nodes[beg-1]->in(0)->is_Call() ) { 879 beg--; 880 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 881 } 882 // Range of inserted instructions is [beg, end) 883 if( beg == end ) return; 884 885 // Clone along all Catch output paths. Clone area between the 'beg' and 886 // 'end' indices. 887 for( uint i = 0; i < _num_succs; i++ ) { 888 Block *sb = _succs[i]; 889 // Clone the entire area; ignoring the edge fixup for now. 890 for( uint j = end; j > beg; j-- ) { 891 Node *clone = _nodes[j-1]->clone(); 892 sb->_nodes.insert( 1, clone ); 893 bbs.map(clone->_idx,sb); 894 } 895 } 896 897 898 // Fixup edges. Check the def-use info per cloned Node 899 for(uint i2 = beg; i2 < end; i2++ ) { 900 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 901 Node *n = _nodes[i2]; // Node that got cloned 902 // Need DU safe iterator because of edge manipulation in calls. 903 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 904 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 905 out->push(n->fast_out(j1)); 906 } 907 uint max = out->size(); 908 for (uint j = 0; j < max; j++) {// For all users 909 Node *use = out->pop(); 910 Block *buse = bbs[use->_idx]; 911 if( use->is_Phi() ) { 912 for( uint k = 1; k < use->req(); k++ ) 913 if( use->in(k) == n ) { 914 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx); 915 use->set_req(k, fixup); 916 } 917 } else { 918 if (this == buse) { 919 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx); 920 } else { 921 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx); 922 } 923 } 924 } // End for all users 925 926 } // End of for all Nodes in cloned area 927 928 // Remove the now-dead cloned ops 929 for(uint i3 = beg; i3 < end; i3++ ) { 930 _nodes[beg]->disconnect_inputs(NULL); 931 _nodes.remove(beg); 932 } 933 934 // If the successor blocks have a CreateEx node, move it back to the top 935 for(uint i4 = 0; i4 < _num_succs; i4++ ) { 936 Block *sb = _succs[i4]; 937 uint new_cnt = end - beg; 938 // Remove any newly created, but dead, nodes. 939 for( uint j = new_cnt; j > 0; j-- ) { 940 Node *n = sb->_nodes[j]; 941 if (n->outcnt() == 0 && 942 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 943 n->disconnect_inputs(NULL); 944 sb->_nodes.remove(j); 945 new_cnt--; 946 } 947 } 948 // If any newly created nodes remain, move the CreateEx node to the top 949 if (new_cnt > 0) { 950 Node *cex = sb->_nodes[1+new_cnt]; 951 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 952 sb->_nodes.remove(1+new_cnt); 953 sb->_nodes.insert(1,cex); 954 } 955 } 956 } 957} 958