lcm.cpp revision 1499:e9ff18c4ace7
1/* 2 * Copyright (c) 1998, 2009, 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// 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 or 36// decodeHeapOop_not_null node if it did not fold into address. 37void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) { 38 // Assume if null check need for 0 offset then always needed 39 // Intel solaris doesn't support any null checks yet and no 40 // mechanism exists (yet) to set the switches at an os_cpu level 41 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 42 43 // Make sure the ptr-is-null path appears to be uncommon! 44 float f = end()->as_MachIf()->_prob; 45 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 46 if( f > PROB_UNLIKELY_MAG(4) ) return; 47 48 uint bidx = 0; // Capture index of value into memop 49 bool was_store; // Memory op is a store op 50 51 // Get the successor block for if the test ptr is non-null 52 Block* not_null_block; // this one goes with the proj 53 Block* null_block; 54 if (_nodes[_nodes.size()-1] == proj) { 55 null_block = _succs[0]; 56 not_null_block = _succs[1]; 57 } else { 58 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other"); 59 not_null_block = _succs[0]; 60 null_block = _succs[1]; 61 } 62 while (null_block->is_Empty() == Block::empty_with_goto) { 63 null_block = null_block->_succs[0]; 64 } 65 66 // Search the exception block for an uncommon trap. 67 // (See Parse::do_if and Parse::do_ifnull for the reason 68 // we need an uncommon trap. Briefly, we need a way to 69 // detect failure of this optimization, as in 6366351.) 70 { 71 bool found_trap = false; 72 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) { 73 Node* nn = null_block->_nodes[i1]; 74 if (nn->is_MachCall() && 75 nn->as_MachCall()->entry_point() == 76 SharedRuntime::uncommon_trap_blob()->instructions_begin()) { 77 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 78 if (trtype->isa_int() && trtype->is_int()->is_con()) { 79 jint tr_con = trtype->is_int()->get_con(); 80 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 81 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 82 assert((int)reason < (int)BitsPerInt, "recode bit map"); 83 if (is_set_nth_bit(allowed_reasons, (int) reason) 84 && action != Deoptimization::Action_none) { 85 // This uncommon trap is sure to recompile, eventually. 86 // When that happens, C->too_many_traps will prevent 87 // this transformation from happening again. 88 found_trap = true; 89 } 90 } 91 break; 92 } 93 } 94 if (!found_trap) { 95 // We did not find an uncommon trap. 96 return; 97 } 98 } 99 100 // Check for decodeHeapOop_not_null node which did not fold into address 101 bool is_decoden = ((intptr_t)val) & 1; 102 val = (Node*)(((intptr_t)val) & ~1); 103 104 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 105 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 106 107 // Search the successor block for a load or store who's base value is also 108 // the tested value. There may be several. 109 Node_List *out = new Node_List(Thread::current()->resource_area()); 110 MachNode *best = NULL; // Best found so far 111 for (DUIterator i = val->outs(); val->has_out(i); i++) { 112 Node *m = val->out(i); 113 if( !m->is_Mach() ) continue; 114 MachNode *mach = m->as_Mach(); 115 was_store = false; 116 switch( mach->ideal_Opcode() ) { 117 case Op_LoadB: 118 case Op_LoadUS: 119 case Op_LoadD: 120 case Op_LoadF: 121 case Op_LoadI: 122 case Op_LoadL: 123 case Op_LoadP: 124 case Op_LoadN: 125 case Op_LoadS: 126 case Op_LoadKlass: 127 case Op_LoadNKlass: 128 case Op_LoadRange: 129 case Op_LoadD_unaligned: 130 case Op_LoadL_unaligned: 131 assert(mach->in(2) == val, "should be address"); 132 break; 133 case Op_StoreB: 134 case Op_StoreC: 135 case Op_StoreCM: 136 case Op_StoreD: 137 case Op_StoreF: 138 case Op_StoreI: 139 case Op_StoreL: 140 case Op_StoreP: 141 case Op_StoreN: 142 was_store = true; // Memory op is a store op 143 // Stores will have their address in slot 2 (memory in slot 1). 144 // If the value being nul-checked is in another slot, it means we 145 // are storing the checked value, which does NOT check the value! 146 if( mach->in(2) != val ) continue; 147 break; // Found a memory op? 148 case Op_StrComp: 149 case Op_StrEquals: 150 case Op_StrIndexOf: 151 case Op_AryEq: 152 // Not a legit memory op for implicit null check regardless of 153 // embedded loads 154 continue; 155 default: // Also check for embedded loads 156 if( !mach->needs_anti_dependence_check() ) 157 continue; // Not an memory op; skip it 158 { 159 // Check that value is used in memory address in 160 // instructions with embedded load (CmpP val1,(val2+off)). 161 Node* base; 162 Node* index; 163 const MachOper* oper = mach->memory_inputs(base, index); 164 if (oper == NULL || oper == (MachOper*)-1) { 165 continue; // Not an memory op; skip it 166 } 167 if (val == base || 168 val == index && val->bottom_type()->isa_narrowoop()) { 169 break; // Found it 170 } else { 171 continue; // Skip it 172 } 173 } 174 break; 175 } 176 // check if the offset is not too high for implicit exception 177 { 178 intptr_t offset = 0; 179 const TypePtr *adr_type = NULL; // Do not need this return value here 180 const Node* base = mach->get_base_and_disp(offset, adr_type); 181 if (base == NULL || base == NodeSentinel) { 182 // Narrow oop address doesn't have base, only index 183 if( val->bottom_type()->isa_narrowoop() && 184 MacroAssembler::needs_explicit_null_check(offset) ) 185 continue; // Give up if offset is beyond page size 186 // cannot reason about it; is probably not implicit null exception 187 } else { 188 const TypePtr* tptr; 189 if (UseCompressedOops && Universe::narrow_oop_shift() == 0) { 190 // 32-bits narrow oop can be the base of address expressions 191 tptr = base->bottom_type()->make_ptr(); 192 } else { 193 // only regular oops are expected here 194 tptr = base->bottom_type()->is_ptr(); 195 } 196 // Give up if offset is not a compile-time constant 197 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 198 continue; 199 offset += tptr->_offset; // correct if base is offseted 200 if( MacroAssembler::needs_explicit_null_check(offset) ) 201 continue; // Give up is reference is beyond 4K page size 202 } 203 } 204 205 // Check ctrl input to see if the null-check dominates the memory op 206 Block *cb = cfg->_bbs[mach->_idx]; 207 cb = cb->_idom; // Always hoist at least 1 block 208 if( !was_store ) { // Stores can be hoisted only one block 209 while( cb->_dom_depth > (_dom_depth + 1)) 210 cb = cb->_idom; // Hoist loads as far as we want 211 // The non-null-block should dominate the memory op, too. Live 212 // range spilling will insert a spill in the non-null-block if it is 213 // needs to spill the memory op for an implicit null check. 214 if (cb->_dom_depth == (_dom_depth + 1)) { 215 if (cb != not_null_block) continue; 216 cb = cb->_idom; 217 } 218 } 219 if( cb != this ) continue; 220 221 // Found a memory user; see if it can be hoisted to check-block 222 uint vidx = 0; // Capture index of value into memop 223 uint j; 224 for( j = mach->req()-1; j > 0; j-- ) { 225 if( mach->in(j) == val ) { 226 vidx = j; 227 // Ignore DecodeN val which could be hoisted to where needed. 228 if( is_decoden ) continue; 229 } 230 // Block of memory-op input 231 Block *inb = cfg->_bbs[mach->in(j)->_idx]; 232 Block *b = this; // Start from nul check 233 while( b != inb && b->_dom_depth > inb->_dom_depth ) 234 b = b->_idom; // search upwards for input 235 // See if input dominates null check 236 if( b != inb ) 237 break; 238 } 239 if( j > 0 ) 240 continue; 241 Block *mb = cfg->_bbs[mach->_idx]; 242 // Hoisting stores requires more checks for the anti-dependence case. 243 // Give up hoisting if we have to move the store past any load. 244 if( was_store ) { 245 Block *b = mb; // Start searching here for a local load 246 // mach use (faulting) trying to hoist 247 // n might be blocker to hoisting 248 while( b != this ) { 249 uint k; 250 for( k = 1; k < b->_nodes.size(); k++ ) { 251 Node *n = b->_nodes[k]; 252 if( n->needs_anti_dependence_check() && 253 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 254 break; // Found anti-dependent load 255 } 256 if( k < b->_nodes.size() ) 257 break; // Found anti-dependent load 258 // Make sure control does not do a merge (would have to check allpaths) 259 if( b->num_preds() != 2 ) break; 260 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block 261 } 262 if( b != this ) continue; 263 } 264 265 // Make sure this memory op is not already being used for a NullCheck 266 Node *e = mb->end(); 267 if( e->is_MachNullCheck() && e->in(1) == mach ) 268 continue; // Already being used as a NULL check 269 270 // Found a candidate! Pick one with least dom depth - the highest 271 // in the dom tree should be closest to the null check. 272 if( !best || 273 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) { 274 best = mach; 275 bidx = vidx; 276 277 } 278 } 279 // No candidate! 280 if( !best ) return; 281 282 // ---- Found an implicit null check 283 extern int implicit_null_checks; 284 implicit_null_checks++; 285 286 if( is_decoden ) { 287 // Check if we need to hoist decodeHeapOop_not_null first. 288 Block *valb = cfg->_bbs[val->_idx]; 289 if( this != valb && this->_dom_depth < valb->_dom_depth ) { 290 // Hoist it up to the end of the test block. 291 valb->find_remove(val); 292 this->add_inst(val); 293 cfg->_bbs.map(val->_idx,this); 294 // DecodeN on x86 may kill flags. Check for flag-killing projections 295 // that also need to be hoisted. 296 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 297 Node* n = val->fast_out(j); 298 if( n->Opcode() == Op_MachProj ) { 299 cfg->_bbs[n->_idx]->find_remove(n); 300 this->add_inst(n); 301 cfg->_bbs.map(n->_idx,this); 302 } 303 } 304 } 305 } 306 // Hoist the memory candidate up to the end of the test block. 307 Block *old_block = cfg->_bbs[best->_idx]; 308 old_block->find_remove(best); 309 add_inst(best); 310 cfg->_bbs.map(best->_idx,this); 311 312 // Move the control dependence 313 if (best->in(0) && best->in(0) == old_block->_nodes[0]) 314 best->set_req(0, _nodes[0]); 315 316 // Check for flag-killing projections that also need to be hoisted 317 // Should be DU safe because no edge updates. 318 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 319 Node* n = best->fast_out(j); 320 if( n->Opcode() == Op_MachProj ) { 321 cfg->_bbs[n->_idx]->find_remove(n); 322 add_inst(n); 323 cfg->_bbs.map(n->_idx,this); 324 } 325 } 326 327 Compile *C = cfg->C; 328 // proj==Op_True --> ne test; proj==Op_False --> eq test. 329 // One of two graph shapes got matched: 330 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 331 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 332 // NULL checks are always branch-if-eq. If we see a IfTrue projection 333 // then we are replacing a 'ne' test with a 'eq' NULL check test. 334 // We need to flip the projections to keep the same semantics. 335 if( proj->Opcode() == Op_IfTrue ) { 336 // Swap order of projections in basic block to swap branch targets 337 Node *tmp1 = _nodes[end_idx()+1]; 338 Node *tmp2 = _nodes[end_idx()+2]; 339 _nodes.map(end_idx()+1, tmp2); 340 _nodes.map(end_idx()+2, tmp1); 341 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input 342 tmp1->replace_by(tmp); 343 tmp2->replace_by(tmp1); 344 tmp->replace_by(tmp2); 345 tmp->destruct(); 346 } 347 348 // Remove the existing null check; use a new implicit null check instead. 349 // Since schedule-local needs precise def-use info, we need to correct 350 // it as well. 351 Node *old_tst = proj->in(0); 352 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); 353 _nodes.map(end_idx(),nul_chk); 354 cfg->_bbs.map(nul_chk->_idx,this); 355 // Redirect users of old_test to nul_chk 356 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 357 old_tst->last_out(i2)->set_req(0, nul_chk); 358 // Clean-up any dead code 359 for (uint i3 = 0; i3 < old_tst->req(); i3++) 360 old_tst->set_req(i3, NULL); 361 362 cfg->latency_from_uses(nul_chk); 363 cfg->latency_from_uses(best); 364} 365 366 367//------------------------------select----------------------------------------- 368// Select a nice fellow from the worklist to schedule next. If there is only 369// one choice, then use it. Projections take top priority for correctness 370// reasons - if I see a projection, then it is next. There are a number of 371// other special cases, for instructions that consume condition codes, et al. 372// These are chosen immediately. Some instructions are required to immediately 373// precede the last instruction in the block, and these are taken last. Of the 374// remaining cases (most), choose the instruction with the greatest latency 375// (that is, the most number of pseudo-cycles required to the end of the 376// routine). If there is a tie, choose the instruction with the most inputs. 377Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) { 378 379 // If only a single entry on the stack, use it 380 uint cnt = worklist.size(); 381 if (cnt == 1) { 382 Node *n = worklist[0]; 383 worklist.map(0,worklist.pop()); 384 return n; 385 } 386 387 uint choice = 0; // Bigger is most important 388 uint latency = 0; // Bigger is scheduled first 389 uint score = 0; // Bigger is better 390 int idx = -1; // Index in worklist 391 392 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 393 // Order in worklist is used to break ties. 394 // See caller for how this is used to delay scheduling 395 // of induction variable increments to after the other 396 // uses of the phi are scheduled. 397 Node *n = worklist[i]; // Get Node on worklist 398 399 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 400 if( n->is_Proj() || // Projections always win 401 n->Opcode()== Op_Con || // So does constant 'Top' 402 iop == Op_CreateEx || // Create-exception must start block 403 iop == Op_CheckCastPP 404 ) { 405 worklist.map(i,worklist.pop()); 406 return n; 407 } 408 409 // Final call in a block must be adjacent to 'catch' 410 Node *e = end(); 411 if( e->is_Catch() && e->in(0)->in(0) == n ) 412 continue; 413 414 // Memory op for an implicit null check has to be at the end of the block 415 if( e->is_MachNullCheck() && e->in(1) == n ) 416 continue; 417 418 uint n_choice = 2; 419 420 // See if this instruction is consumed by a branch. If so, then (as the 421 // branch is the last instruction in the basic block) force it to the 422 // end of the basic block 423 if ( must_clone[iop] ) { 424 // See if any use is a branch 425 bool found_machif = false; 426 427 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 428 Node* use = n->fast_out(j); 429 430 // The use is a conditional branch, make them adjacent 431 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) { 432 found_machif = true; 433 break; 434 } 435 436 // More than this instruction pending for successor to be ready, 437 // don't choose this if other opportunities are ready 438 if (ready_cnt[use->_idx] > 1) 439 n_choice = 1; 440 } 441 442 // loop terminated, prefer not to use this instruction 443 if (found_machif) 444 continue; 445 } 446 447 // See if this has a predecessor that is "must_clone", i.e. sets the 448 // condition code. If so, choose this first 449 for (uint j = 0; j < n->req() ; j++) { 450 Node *inn = n->in(j); 451 if (inn) { 452 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 453 n_choice = 3; 454 break; 455 } 456 } 457 } 458 459 // MachTemps should be scheduled last so they are near their uses 460 if (n->is_MachTemp()) { 461 n_choice = 1; 462 } 463 464 uint n_latency = cfg->_node_latency.at_grow(n->_idx); 465 uint n_score = n->req(); // Many inputs get high score to break ties 466 467 // Keep best latency found 468 if( choice < n_choice || 469 ( choice == n_choice && 470 ( latency < n_latency || 471 ( latency == n_latency && 472 ( score < n_score ))))) { 473 choice = n_choice; 474 latency = n_latency; 475 score = n_score; 476 idx = i; // Also keep index in worklist 477 } 478 } // End of for all ready nodes in worklist 479 480 assert(idx >= 0, "index should be set"); 481 Node *n = worklist[(uint)idx]; // Get the winner 482 483 worklist.map((uint)idx, worklist.pop()); // Compress worklist 484 return n; 485} 486 487 488//------------------------------set_next_call---------------------------------- 489void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) { 490 if( next_call.test_set(n->_idx) ) return; 491 for( uint i=0; i<n->len(); i++ ) { 492 Node *m = n->in(i); 493 if( !m ) continue; // must see all nodes in block that precede call 494 if( bbs[m->_idx] == this ) 495 set_next_call( m, next_call, bbs ); 496 } 497} 498 499//------------------------------needed_for_next_call--------------------------- 500// Set the flag 'next_call' for each Node that is needed for the next call to 501// be scheduled. This flag lets me bias scheduling so Nodes needed for the 502// next subroutine call get priority - basically it moves things NOT needed 503// for the next call till after the call. This prevents me from trying to 504// carry lots of stuff live across a call. 505void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) { 506 // Find the next control-defining Node in this block 507 Node* call = NULL; 508 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 509 Node* m = this_call->fast_out(i); 510 if( bbs[m->_idx] == this && // Local-block user 511 m != this_call && // Not self-start node 512 m->is_Call() ) 513 call = m; 514 break; 515 } 516 if (call == NULL) return; // No next call (e.g., block end is near) 517 // Set next-call for all inputs to this call 518 set_next_call(call, next_call, bbs); 519} 520 521//------------------------------sched_call------------------------------------- 522uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) { 523 RegMask regs; 524 525 // Schedule all the users of the call right now. All the users are 526 // projection Nodes, so they must be scheduled next to the call. 527 // Collect all the defined registers. 528 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 529 Node* n = mcall->fast_out(i); 530 assert( n->Opcode()==Op_MachProj, "" ); 531 --ready_cnt[n->_idx]; 532 assert( !ready_cnt[n->_idx], "" ); 533 // Schedule next to call 534 _nodes.map(node_cnt++, n); 535 // Collect defined registers 536 regs.OR(n->out_RegMask()); 537 // Check for scheduling the next control-definer 538 if( n->bottom_type() == Type::CONTROL ) 539 // Warm up next pile of heuristic bits 540 needed_for_next_call(n, next_call, bbs); 541 542 // Children of projections are now all ready 543 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 544 Node* m = n->fast_out(j); // Get user 545 if( bbs[m->_idx] != this ) continue; 546 if( m->is_Phi() ) continue; 547 if( !--ready_cnt[m->_idx] ) 548 worklist.push(m); 549 } 550 551 } 552 553 // Act as if the call defines the Frame Pointer. 554 // Certainly the FP is alive and well after the call. 555 regs.Insert(matcher.c_frame_pointer()); 556 557 // Set all registers killed and not already defined by the call. 558 uint r_cnt = mcall->tf()->range()->cnt(); 559 int op = mcall->ideal_Opcode(); 560 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 561 bbs.map(proj->_idx,this); 562 _nodes.insert(node_cnt++, proj); 563 564 // Select the right register save policy. 565 const char * save_policy; 566 switch (op) { 567 case Op_CallRuntime: 568 case Op_CallLeaf: 569 case Op_CallLeafNoFP: 570 // Calling C code so use C calling convention 571 save_policy = matcher._c_reg_save_policy; 572 break; 573 574 case Op_CallStaticJava: 575 case Op_CallDynamicJava: 576 // Calling Java code so use Java calling convention 577 save_policy = matcher._register_save_policy; 578 break; 579 580 default: 581 ShouldNotReachHere(); 582 } 583 584 // When using CallRuntime mark SOE registers as killed by the call 585 // so values that could show up in the RegisterMap aren't live in a 586 // callee saved register since the register wouldn't know where to 587 // find them. CallLeaf and CallLeafNoFP are ok because they can't 588 // have debug info on them. Strictly speaking this only needs to be 589 // done for oops since idealreg2debugmask takes care of debug info 590 // references but there no way to handle oops differently than other 591 // pointers as far as the kill mask goes. 592 bool exclude_soe = op == Op_CallRuntime; 593 594 // If the call is a MethodHandle invoke, we need to exclude the 595 // register which is used to save the SP value over MH invokes from 596 // the mask. Otherwise this register could be used for 597 // deoptimization information. 598 if (op == Op_CallStaticJava) { 599 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 600 if (mcallstaticjava->_method_handle_invoke) 601 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 602 } 603 604 // Fill in the kill mask for the call 605 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 606 if( !regs.Member(r) ) { // Not already defined by the call 607 // Save-on-call register? 608 if ((save_policy[r] == 'C') || 609 (save_policy[r] == 'A') || 610 ((save_policy[r] == 'E') && exclude_soe)) { 611 proj->_rout.Insert(r); 612 } 613 } 614 } 615 616 return node_cnt; 617} 618 619 620//------------------------------schedule_local--------------------------------- 621// Topological sort within a block. Someday become a real scheduler. 622bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) { 623 // Already "sorted" are the block start Node (as the first entry), and 624 // the block-ending Node and any trailing control projections. We leave 625 // these alone. PhiNodes and ParmNodes are made to follow the block start 626 // Node. Everything else gets topo-sorted. 627 628#ifndef PRODUCT 629 if (cfg->trace_opto_pipelining()) { 630 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order); 631 for (uint i = 0;i < _nodes.size();i++) { 632 tty->print("# "); 633 _nodes[i]->fast_dump(); 634 } 635 tty->print_cr("#"); 636 } 637#endif 638 639 // RootNode is already sorted 640 if( _nodes.size() == 1 ) return true; 641 642 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 643 uint node_cnt = end_idx(); 644 uint phi_cnt = 1; 645 uint i; 646 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 647 Node *n = _nodes[i]; 648 if( n->is_Phi() || // Found a PhiNode or ParmNode 649 (n->is_Proj() && n->in(0) == head()) ) { 650 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 651 _nodes.map(i,_nodes[phi_cnt]); 652 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front 653 } else { // All others 654 // Count block-local inputs to 'n' 655 uint cnt = n->len(); // Input count 656 uint local = 0; 657 for( uint j=0; j<cnt; j++ ) { 658 Node *m = n->in(j); 659 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() ) 660 local++; // One more block-local input 661 } 662 ready_cnt[n->_idx] = local; // Count em up 663 664 // A few node types require changing a required edge to a precedence edge 665 // before allocation. 666 if( UseConcMarkSweepGC || UseG1GC ) { 667 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 668 // Note: Required edges with an index greater than oper_input_base 669 // are not supported by the allocator. 670 // Note2: Can only depend on unmatched edge being last, 671 // can not depend on its absolute position. 672 Node *oop_store = n->in(n->req() - 1); 673 n->del_req(n->req() - 1); 674 n->add_prec(oop_store); 675 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark"); 676 } 677 } 678 if( n->is_Mach() && n->req() > TypeFunc::Parms && 679 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 680 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 681 // MemBarAcquire could be created without Precedent edge. 682 // del_req() replaces the specified edge with the last input edge 683 // and then removes the last edge. If the specified edge > number of 684 // edges the last edge will be moved outside of the input edges array 685 // and the edge will be lost. This is why this code should be 686 // executed only when Precedent (== TypeFunc::Parms) edge is present. 687 Node *x = n->in(TypeFunc::Parms); 688 n->del_req(TypeFunc::Parms); 689 n->add_prec(x); 690 } 691 } 692 } 693 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count 694 ready_cnt[_nodes[i2]->_idx] = 0; 695 696 // All the prescheduled guys do not hold back internal nodes 697 uint i3; 698 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 699 Node *n = _nodes[i3]; // Get pre-scheduled 700 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 701 Node* m = n->fast_out(j); 702 if( cfg->_bbs[m->_idx] ==this ) // Local-block user 703 ready_cnt[m->_idx]--; // Fix ready count 704 } 705 } 706 707 Node_List delay; 708 // Make a worklist 709 Node_List worklist; 710 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 711 Node *m = _nodes[i4]; 712 if( !ready_cnt[m->_idx] ) { // Zero ready count? 713 if (m->is_iteratively_computed()) { 714 // Push induction variable increments last to allow other uses 715 // of the phi to be scheduled first. The select() method breaks 716 // ties in scheduling by worklist order. 717 delay.push(m); 718 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 719 // Force the CreateEx to the top of the list so it's processed 720 // first and ends up at the start of the block. 721 worklist.insert(0, m); 722 } else { 723 worklist.push(m); // Then on to worklist! 724 } 725 } 726 } 727 while (delay.size()) { 728 Node* d = delay.pop(); 729 worklist.push(d); 730 } 731 732 // Warm up the 'next_call' heuristic bits 733 needed_for_next_call(_nodes[0], next_call, cfg->_bbs); 734 735#ifndef PRODUCT 736 if (cfg->trace_opto_pipelining()) { 737 for (uint j=0; j<_nodes.size(); j++) { 738 Node *n = _nodes[j]; 739 int idx = n->_idx; 740 tty->print("# ready cnt:%3d ", ready_cnt[idx]); 741 tty->print("latency:%3d ", cfg->_node_latency.at_grow(idx)); 742 tty->print("%4d: %s\n", idx, n->Name()); 743 } 744 } 745#endif 746 747 // Pull from worklist and schedule 748 while( worklist.size() ) { // Worklist is not ready 749 750#ifndef PRODUCT 751 if (cfg->trace_opto_pipelining()) { 752 tty->print("# ready list:"); 753 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 754 Node *n = worklist[i]; // Get Node on worklist 755 tty->print(" %d", n->_idx); 756 } 757 tty->cr(); 758 } 759#endif 760 761 // Select and pop a ready guy from worklist 762 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt); 763 _nodes.map(phi_cnt++,n); // Schedule him next 764 765#ifndef PRODUCT 766 if (cfg->trace_opto_pipelining()) { 767 tty->print("# select %d: %s", n->_idx, n->Name()); 768 tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx)); 769 n->dump(); 770 if (Verbose) { 771 tty->print("# ready list:"); 772 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 773 Node *n = worklist[i]; // Get Node on worklist 774 tty->print(" %d", n->_idx); 775 } 776 tty->cr(); 777 } 778 } 779 780#endif 781 if( n->is_MachCall() ) { 782 MachCallNode *mcall = n->as_MachCall(); 783 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call); 784 continue; 785 } 786 // Children are now all ready 787 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 788 Node* m = n->fast_out(i5); // Get user 789 if( cfg->_bbs[m->_idx] != this ) continue; 790 if( m->is_Phi() ) continue; 791 if( !--ready_cnt[m->_idx] ) 792 worklist.push(m); 793 } 794 } 795 796 if( phi_cnt != end_idx() ) { 797 // did not schedule all. Retry, Bailout, or Die 798 Compile* C = matcher.C; 799 if (C->subsume_loads() == true && !C->failing()) { 800 // Retry with subsume_loads == false 801 // If this is the first failure, the sentinel string will "stick" 802 // to the Compile object, and the C2Compiler will see it and retry. 803 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 804 } 805 // assert( phi_cnt == end_idx(), "did not schedule all" ); 806 return false; 807 } 808 809#ifndef PRODUCT 810 if (cfg->trace_opto_pipelining()) { 811 tty->print_cr("#"); 812 tty->print_cr("# after schedule_local"); 813 for (uint i = 0;i < _nodes.size();i++) { 814 tty->print("# "); 815 _nodes[i]->fast_dump(); 816 } 817 tty->cr(); 818 } 819#endif 820 821 822 return true; 823} 824 825//--------------------------catch_cleanup_fix_all_inputs----------------------- 826static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 827 for (uint l = 0; l < use->len(); l++) { 828 if (use->in(l) == old_def) { 829 if (l < use->req()) { 830 use->set_req(l, new_def); 831 } else { 832 use->rm_prec(l); 833 use->add_prec(new_def); 834 l--; 835 } 836 } 837 } 838} 839 840//------------------------------catch_cleanup_find_cloned_def------------------ 841static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { 842 assert( use_blk != def_blk, "Inter-block cleanup only"); 843 844 // The use is some block below the Catch. Find and return the clone of the def 845 // that dominates the use. If there is no clone in a dominating block, then 846 // create a phi for the def in a dominating block. 847 848 // Find which successor block dominates this use. The successor 849 // blocks must all be single-entry (from the Catch only; I will have 850 // split blocks to make this so), hence they all dominate. 851 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 852 use_blk = use_blk->_idom; 853 854 // Find the successor 855 Node *fixup = NULL; 856 857 uint j; 858 for( j = 0; j < def_blk->_num_succs; j++ ) 859 if( use_blk == def_blk->_succs[j] ) 860 break; 861 862 if( j == def_blk->_num_succs ) { 863 // Block at same level in dom-tree is not a successor. It needs a 864 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 865 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 866 for(uint k = 1; k < use_blk->num_preds(); k++) { 867 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx)); 868 } 869 870 // Check to see if the use_blk already has an identical phi inserted. 871 // If it exists, it will be at the first position since all uses of a 872 // def are processed together. 873 Node *phi = use_blk->_nodes[1]; 874 if( phi->is_Phi() ) { 875 fixup = phi; 876 for (uint k = 1; k < use_blk->num_preds(); k++) { 877 if (phi->in(k) != inputs[k]) { 878 // Not a match 879 fixup = NULL; 880 break; 881 } 882 } 883 } 884 885 // If an existing PhiNode was not found, make a new one. 886 if (fixup == NULL) { 887 Node *new_phi = PhiNode::make(use_blk->head(), def); 888 use_blk->_nodes.insert(1, new_phi); 889 bbs.map(new_phi->_idx, use_blk); 890 for (uint k = 1; k < use_blk->num_preds(); k++) { 891 new_phi->set_req(k, inputs[k]); 892 } 893 fixup = new_phi; 894 } 895 896 } else { 897 // Found the use just below the Catch. Make it use the clone. 898 fixup = use_blk->_nodes[n_clone_idx]; 899 } 900 901 return fixup; 902} 903 904//--------------------------catch_cleanup_intra_block-------------------------- 905// Fix all input edges in use that reference "def". The use is in the same 906// block as the def and both have been cloned in each successor block. 907static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 908 909 // Both the use and def have been cloned. For each successor block, 910 // get the clone of the use, and make its input the clone of the def 911 // found in that block. 912 913 uint use_idx = blk->find_node(use); 914 uint offset_idx = use_idx - beg; 915 for( uint k = 0; k < blk->_num_succs; k++ ) { 916 // Get clone in each successor block 917 Block *sb = blk->_succs[k]; 918 Node *clone = sb->_nodes[offset_idx+1]; 919 assert( clone->Opcode() == use->Opcode(), "" ); 920 921 // Make use-clone reference the def-clone 922 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]); 923 } 924} 925 926//------------------------------catch_cleanup_inter_block--------------------- 927// Fix all input edges in use that reference "def". The use is in a different 928// block than the def. 929static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { 930 if( !use_blk ) return; // Can happen if the use is a precedence edge 931 932 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx); 933 catch_cleanup_fix_all_inputs(use, def, new_def); 934} 935 936//------------------------------call_catch_cleanup----------------------------- 937// If we inserted any instructions between a Call and his CatchNode, 938// clone the instructions on all paths below the Catch. 939void Block::call_catch_cleanup(Block_Array &bbs) { 940 941 // End of region to clone 942 uint end = end_idx(); 943 if( !_nodes[end]->is_Catch() ) return; 944 // Start of region to clone 945 uint beg = end; 946 while( _nodes[beg-1]->Opcode() != Op_MachProj || 947 !_nodes[beg-1]->in(0)->is_Call() ) { 948 beg--; 949 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 950 } 951 // Range of inserted instructions is [beg, end) 952 if( beg == end ) return; 953 954 // Clone along all Catch output paths. Clone area between the 'beg' and 955 // 'end' indices. 956 for( uint i = 0; i < _num_succs; i++ ) { 957 Block *sb = _succs[i]; 958 // Clone the entire area; ignoring the edge fixup for now. 959 for( uint j = end; j > beg; j-- ) { 960 Node *clone = _nodes[j-1]->clone(); 961 sb->_nodes.insert( 1, clone ); 962 bbs.map(clone->_idx,sb); 963 } 964 } 965 966 967 // Fixup edges. Check the def-use info per cloned Node 968 for(uint i2 = beg; i2 < end; i2++ ) { 969 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 970 Node *n = _nodes[i2]; // Node that got cloned 971 // Need DU safe iterator because of edge manipulation in calls. 972 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 973 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 974 out->push(n->fast_out(j1)); 975 } 976 uint max = out->size(); 977 for (uint j = 0; j < max; j++) {// For all users 978 Node *use = out->pop(); 979 Block *buse = bbs[use->_idx]; 980 if( use->is_Phi() ) { 981 for( uint k = 1; k < use->req(); k++ ) 982 if( use->in(k) == n ) { 983 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx); 984 use->set_req(k, fixup); 985 } 986 } else { 987 if (this == buse) { 988 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx); 989 } else { 990 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx); 991 } 992 } 993 } // End for all users 994 995 } // End of for all Nodes in cloned area 996 997 // Remove the now-dead cloned ops 998 for(uint i3 = beg; i3 < end; i3++ ) { 999 _nodes[beg]->disconnect_inputs(NULL); 1000 _nodes.remove(beg); 1001 } 1002 1003 // If the successor blocks have a CreateEx node, move it back to the top 1004 for(uint i4 = 0; i4 < _num_succs; i4++ ) { 1005 Block *sb = _succs[i4]; 1006 uint new_cnt = end - beg; 1007 // Remove any newly created, but dead, nodes. 1008 for( uint j = new_cnt; j > 0; j-- ) { 1009 Node *n = sb->_nodes[j]; 1010 if (n->outcnt() == 0 && 1011 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1012 n->disconnect_inputs(NULL); 1013 sb->_nodes.remove(j); 1014 new_cnt--; 1015 } 1016 } 1017 // If any newly created nodes remain, move the CreateEx node to the top 1018 if (new_cnt > 0) { 1019 Node *cex = sb->_nodes[1+new_cnt]; 1020 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1021 sb->_nodes.remove(1+new_cnt); 1022 sb->_nodes.insert(1,cex); 1023 } 1024 } 1025 } 1026} 1027