1 /* 2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/block.hpp" 28 #include "opto/c2compiler.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/cfgnode.hpp" 31 #include "opto/machnode.hpp" 32 #include "opto/runtime.hpp" 33 #ifdef TARGET_ARCH_MODEL_x86_32 34 # include "adfiles/ad_x86_32.hpp" 35 #endif 36 #ifdef TARGET_ARCH_MODEL_x86_64 37 # include "adfiles/ad_x86_64.hpp" 38 #endif 39 #ifdef TARGET_ARCH_MODEL_sparc 40 # include "adfiles/ad_sparc.hpp" 41 #endif 42 #ifdef TARGET_ARCH_MODEL_zero 43 # include "adfiles/ad_zero.hpp" 44 #endif 45 #ifdef TARGET_ARCH_MODEL_arm 46 # include "adfiles/ad_arm.hpp" 47 #endif 48 #ifdef TARGET_ARCH_MODEL_ppc 49 # include "adfiles/ad_ppc.hpp" 50 #endif 51 52 // Optimization - Graph Style 53 54 //------------------------------implicit_null_check---------------------------- 55 // Detect implicit-null-check opportunities. Basically, find NULL checks 56 // with suitable memory ops nearby. Use the memory op to do the NULL check. 57 // I can generate a memory op if there is not one nearby. 58 // The proj is the control projection for the not-null case. 59 // The val is the pointer being checked for nullness or 60 // decodeHeapOop_not_null node if it did not fold into address. 61 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) { 62 // Assume if null check need for 0 offset then always needed 63 // Intel solaris doesn't support any null checks yet and no 64 // mechanism exists (yet) to set the switches at an os_cpu level 65 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 66 67 // Make sure the ptr-is-null path appears to be uncommon! 68 float f = end()->as_MachIf()->_prob; 69 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 70 if( f > PROB_UNLIKELY_MAG(4) ) return; 71 72 uint bidx = 0; // Capture index of value into memop 73 bool was_store; // Memory op is a store op 74 75 // Get the successor block for if the test ptr is non-null 76 Block* not_null_block; // this one goes with the proj 77 Block* null_block; 78 if (_nodes[_nodes.size()-1] == proj) { 79 null_block = _succs[0]; 80 not_null_block = _succs[1]; 81 } else { 82 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other"); 83 not_null_block = _succs[0]; 84 null_block = _succs[1]; 85 } 86 while (null_block->is_Empty() == Block::empty_with_goto) { 87 null_block = null_block->_succs[0]; 88 } 89 90 // Search the exception block for an uncommon trap. 91 // (See Parse::do_if and Parse::do_ifnull for the reason 92 // we need an uncommon trap. Briefly, we need a way to 93 // detect failure of this optimization, as in 6366351.) 94 { 95 bool found_trap = false; 96 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) { 97 Node* nn = null_block->_nodes[i1]; 98 if (nn->is_MachCall() && 99 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 100 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 101 if (trtype->isa_int() && trtype->is_int()->is_con()) { 102 jint tr_con = trtype->is_int()->get_con(); 103 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 104 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 105 assert((int)reason < (int)BitsPerInt, "recode bit map"); 106 if (is_set_nth_bit(allowed_reasons, (int) reason) 107 && action != Deoptimization::Action_none) { 108 // This uncommon trap is sure to recompile, eventually. 109 // When that happens, C->too_many_traps will prevent 110 // this transformation from happening again. 111 found_trap = true; 112 } 113 } 114 break; 115 } 116 } 117 if (!found_trap) { 118 // We did not find an uncommon trap. 119 return; 120 } 121 } 122 123 // Check for decodeHeapOop_not_null node which did not fold into address 124 bool is_decoden = ((intptr_t)val) & 1; 125 val = (Node*)(((intptr_t)val) & ~1); 126 127 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 128 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 129 130 // Search the successor block for a load or store who's base value is also 131 // the tested value. There may be several. 132 Node_List *out = new Node_List(Thread::current()->resource_area()); 133 MachNode *best = NULL; // Best found so far 134 for (DUIterator i = val->outs(); val->has_out(i); i++) { 135 Node *m = val->out(i); 136 if( !m->is_Mach() ) continue; 137 MachNode *mach = m->as_Mach(); 138 was_store = false; 139 int iop = mach->ideal_Opcode(); 140 switch( iop ) { 141 case Op_LoadB: 142 case Op_LoadUB: 143 case Op_LoadUS: 144 case Op_LoadD: 145 case Op_LoadF: 146 case Op_LoadI: 147 case Op_LoadL: 148 case Op_LoadP: 149 case Op_LoadN: 150 case Op_LoadS: 151 case Op_LoadKlass: 152 case Op_LoadNKlass: 153 case Op_LoadRange: 154 case Op_LoadD_unaligned: 155 case Op_LoadL_unaligned: 156 assert(mach->in(2) == val, "should be address"); 157 break; 158 case Op_StoreB: 159 case Op_StoreC: 160 case Op_StoreCM: 161 case Op_StoreD: 162 case Op_StoreF: 163 case Op_StoreI: 164 case Op_StoreL: 165 case Op_StoreP: 166 case Op_StoreN: 167 case Op_StoreNKlass: 168 was_store = true; // Memory op is a store op 169 // Stores will have their address in slot 2 (memory in slot 1). 170 // If the value being nul-checked is in another slot, it means we 171 // are storing the checked value, which does NOT check the value! 172 if( mach->in(2) != val ) continue; 173 break; // Found a memory op? 174 case Op_StrComp: 175 case Op_StrEquals: 176 case Op_StrIndexOf: 177 case Op_AryEq: 178 case Op_EncodeISOArray: 179 // Not a legit memory op for implicit null check regardless of 180 // embedded loads 181 continue; 182 default: // Also check for embedded loads 183 if( !mach->needs_anti_dependence_check() ) 184 continue; // Not an memory op; skip it 185 if( must_clone[iop] ) { 186 // Do not move nodes which produce flags because 187 // RA will try to clone it to place near branch and 188 // it will cause recompilation, see clone_node(). 189 continue; 190 } 191 { 192 // Check that value is used in memory address in 193 // instructions with embedded load (CmpP val1,(val2+off)). 194 Node* base; 195 Node* index; 196 const MachOper* oper = mach->memory_inputs(base, index); 197 if (oper == NULL || oper == (MachOper*)-1) { 198 continue; // Not an memory op; skip it 199 } 200 if (val == base || 201 val == index && val->bottom_type()->isa_narrowoop()) { 202 break; // Found it 203 } else { 204 continue; // Skip it 205 } 206 } 207 break; 208 } 209 // check if the offset is not too high for implicit exception 210 { 211 intptr_t offset = 0; 212 const TypePtr *adr_type = NULL; // Do not need this return value here 213 const Node* base = mach->get_base_and_disp(offset, adr_type); 214 if (base == NULL || base == NodeSentinel) { 215 // Narrow oop address doesn't have base, only index 216 if( val->bottom_type()->isa_narrowoop() && 217 MacroAssembler::needs_explicit_null_check(offset) ) 218 continue; // Give up if offset is beyond page size 219 // cannot reason about it; is probably not implicit null exception 220 } else { 221 const TypePtr* tptr; 222 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 223 Universe::narrow_klass_shift() == 0)) { 224 // 32-bits narrow oop can be the base of address expressions 225 tptr = base->get_ptr_type(); 226 } else { 227 // only regular oops are expected here 228 tptr = base->bottom_type()->is_ptr(); 229 } 230 // Give up if offset is not a compile-time constant 231 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 232 continue; 233 offset += tptr->_offset; // correct if base is offseted 234 if( MacroAssembler::needs_explicit_null_check(offset) ) 235 continue; // Give up is reference is beyond 4K page size 236 } 237 } 238 239 // Check ctrl input to see if the null-check dominates the memory op 240 Block *cb = cfg->get_block_for_node(mach); 241 cb = cb->_idom; // Always hoist at least 1 block 242 if( !was_store ) { // Stores can be hoisted only one block 243 while( cb->_dom_depth > (_dom_depth + 1)) 244 cb = cb->_idom; // Hoist loads as far as we want 245 // The non-null-block should dominate the memory op, too. Live 246 // range spilling will insert a spill in the non-null-block if it is 247 // needs to spill the memory op for an implicit null check. 248 if (cb->_dom_depth == (_dom_depth + 1)) { 249 if (cb != not_null_block) continue; 250 cb = cb->_idom; 251 } 252 } 253 if( cb != this ) continue; 254 255 // Found a memory user; see if it can be hoisted to check-block 256 uint vidx = 0; // Capture index of value into memop 257 uint j; 258 for( j = mach->req()-1; j > 0; j-- ) { 259 if( mach->in(j) == val ) { 260 vidx = j; 261 // Ignore DecodeN val which could be hoisted to where needed. 262 if( is_decoden ) continue; 263 } 264 // Block of memory-op input 265 Block *inb = cfg->get_block_for_node(mach->in(j)); 266 Block *b = this; // Start from nul check 267 while( b != inb && b->_dom_depth > inb->_dom_depth ) 268 b = b->_idom; // search upwards for input 269 // See if input dominates null check 270 if( b != inb ) 271 break; 272 } 273 if( j > 0 ) 274 continue; 275 Block *mb = cfg->get_block_for_node(mach); 276 // Hoisting stores requires more checks for the anti-dependence case. 277 // Give up hoisting if we have to move the store past any load. 278 if( was_store ) { 279 Block *b = mb; // Start searching here for a local load 280 // mach use (faulting) trying to hoist 281 // n might be blocker to hoisting 282 while( b != this ) { 283 uint k; 284 for( k = 1; k < b->_nodes.size(); k++ ) { 285 Node *n = b->_nodes[k]; 286 if( n->needs_anti_dependence_check() && 287 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 288 break; // Found anti-dependent load 289 } 290 if( k < b->_nodes.size() ) 291 break; // Found anti-dependent load 292 // Make sure control does not do a merge (would have to check allpaths) 293 if( b->num_preds() != 2 ) break; 294 b = cfg->get_block_for_node(b->pred(1)); // Move up to predecessor block 295 } 296 if( b != this ) continue; 297 } 298 299 // Make sure this memory op is not already being used for a NullCheck 300 Node *e = mb->end(); 301 if( e->is_MachNullCheck() && e->in(1) == mach ) 302 continue; // Already being used as a NULL check 303 304 // Found a candidate! Pick one with least dom depth - the highest 305 // in the dom tree should be closest to the null check. 306 if (best == NULL || cfg->get_block_for_node(mach)->_dom_depth < cfg->get_block_for_node(best)->_dom_depth) { 307 best = mach; 308 bidx = vidx; 309 } 310 } 311 // No candidate! 312 if (best == NULL) { 313 return; 314 } 315 316 // ---- Found an implicit null check 317 extern int implicit_null_checks; 318 implicit_null_checks++; 319 320 if( is_decoden ) { 321 // Check if we need to hoist decodeHeapOop_not_null first. 322 Block *valb = cfg->get_block_for_node(val); 323 if( this != valb && this->_dom_depth < valb->_dom_depth ) { 324 // Hoist it up to the end of the test block. 325 valb->find_remove(val); 326 this->add_inst(val); 327 cfg->map_node_to_block(val, this); 328 // DecodeN on x86 may kill flags. Check for flag-killing projections 329 // that also need to be hoisted. 330 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 331 Node* n = val->fast_out(j); 332 if( n->is_MachProj() ) { 333 cfg->get_block_for_node(n)->find_remove(n); 334 this->add_inst(n); 335 cfg->map_node_to_block(n, this); 336 } 337 } 338 } 339 } 340 // Hoist the memory candidate up to the end of the test block. 341 Block *old_block = cfg->get_block_for_node(best); 342 old_block->find_remove(best); 343 add_inst(best); 344 cfg->map_node_to_block(best, this); 345 346 // Move the control dependence 347 if (best->in(0) && best->in(0) == old_block->_nodes[0]) 348 best->set_req(0, _nodes[0]); 349 350 // Check for flag-killing projections that also need to be hoisted 351 // Should be DU safe because no edge updates. 352 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 353 Node* n = best->fast_out(j); 354 if( n->is_MachProj() ) { 355 cfg->get_block_for_node(n)->find_remove(n); 356 add_inst(n); 357 cfg->map_node_to_block(n, this); 358 } 359 } 360 361 Compile *C = cfg->C; 362 // proj==Op_True --> ne test; proj==Op_False --> eq test. 363 // One of two graph shapes got matched: 364 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 365 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 366 // NULL checks are always branch-if-eq. If we see a IfTrue projection 367 // then we are replacing a 'ne' test with a 'eq' NULL check test. 368 // We need to flip the projections to keep the same semantics. 369 if( proj->Opcode() == Op_IfTrue ) { 370 // Swap order of projections in basic block to swap branch targets 371 Node *tmp1 = _nodes[end_idx()+1]; 372 Node *tmp2 = _nodes[end_idx()+2]; 373 _nodes.map(end_idx()+1, tmp2); 374 _nodes.map(end_idx()+2, tmp1); 375 Node *tmp = new (C) Node(C->top()); // Use not NULL input 376 tmp1->replace_by(tmp); 377 tmp2->replace_by(tmp1); 378 tmp->replace_by(tmp2); 379 tmp->destruct(); 380 } 381 382 // Remove the existing null check; use a new implicit null check instead. 383 // Since schedule-local needs precise def-use info, we need to correct 384 // it as well. 385 Node *old_tst = proj->in(0); 386 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); 387 _nodes.map(end_idx(),nul_chk); 388 cfg->map_node_to_block(nul_chk, this); 389 // Redirect users of old_test to nul_chk 390 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 391 old_tst->last_out(i2)->set_req(0, nul_chk); 392 // Clean-up any dead code 393 for (uint i3 = 0; i3 < old_tst->req(); i3++) 394 old_tst->set_req(i3, NULL); 395 396 cfg->latency_from_uses(nul_chk); 397 cfg->latency_from_uses(best); 398 } 399 400 401 //------------------------------select----------------------------------------- 402 // Select a nice fellow from the worklist to schedule next. If there is only 403 // one choice, then use it. Projections take top priority for correctness 404 // reasons - if I see a projection, then it is next. There are a number of 405 // other special cases, for instructions that consume condition codes, et al. 406 // These are chosen immediately. Some instructions are required to immediately 407 // precede the last instruction in the block, and these are taken last. Of the 408 // remaining cases (most), choose the instruction with the greatest latency 409 // (that is, the most number of pseudo-cycles required to the end of the 410 // routine). If there is a tie, choose the instruction with the most inputs. 411 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 412 413 // If only a single entry on the stack, use it 414 uint cnt = worklist.size(); 415 if (cnt == 1) { 416 Node *n = worklist[0]; 417 worklist.map(0,worklist.pop()); 418 return n; 419 } 420 421 uint choice = 0; // Bigger is most important 422 uint latency = 0; // Bigger is scheduled first 423 uint score = 0; // Bigger is better 424 int idx = -1; // Index in worklist 425 int cand_cnt = 0; // Candidate count 426 427 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 428 // Order in worklist is used to break ties. 429 // See caller for how this is used to delay scheduling 430 // of induction variable increments to after the other 431 // uses of the phi are scheduled. 432 Node *n = worklist[i]; // Get Node on worklist 433 434 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 435 if( n->is_Proj() || // Projections always win 436 n->Opcode()== Op_Con || // So does constant 'Top' 437 iop == Op_CreateEx || // Create-exception must start block 438 iop == Op_CheckCastPP 439 ) { 440 worklist.map(i,worklist.pop()); 441 return n; 442 } 443 444 // Final call in a block must be adjacent to 'catch' 445 Node *e = end(); 446 if( e->is_Catch() && e->in(0)->in(0) == n ) 447 continue; 448 449 // Memory op for an implicit null check has to be at the end of the block 450 if( e->is_MachNullCheck() && e->in(1) == n ) 451 continue; 452 453 // Schedule IV increment last. 454 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 455 e->in(1)->in(1) == n && n->is_iteratively_computed()) 456 continue; 457 458 uint n_choice = 2; 459 460 // See if this instruction is consumed by a branch. If so, then (as the 461 // branch is the last instruction in the basic block) force it to the 462 // end of the basic block 463 if ( must_clone[iop] ) { 464 // See if any use is a branch 465 bool found_machif = false; 466 467 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 468 Node* use = n->fast_out(j); 469 470 // The use is a conditional branch, make them adjacent 471 if (use->is_MachIf() && cfg->get_block_for_node(use) == this) { 472 found_machif = true; 473 break; 474 } 475 476 // More than this instruction pending for successor to be ready, 477 // don't choose this if other opportunities are ready 478 if (ready_cnt.at(use->_idx) > 1) 479 n_choice = 1; 480 } 481 482 // loop terminated, prefer not to use this instruction 483 if (found_machif) 484 continue; 485 } 486 487 // See if this has a predecessor that is "must_clone", i.e. sets the 488 // condition code. If so, choose this first 489 for (uint j = 0; j < n->req() ; j++) { 490 Node *inn = n->in(j); 491 if (inn) { 492 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 493 n_choice = 3; 494 break; 495 } 496 } 497 } 498 499 // MachTemps should be scheduled last so they are near their uses 500 if (n->is_MachTemp()) { 501 n_choice = 1; 502 } 503 504 uint n_latency = cfg->_node_latency->at_grow(n->_idx); 505 uint n_score = n->req(); // Many inputs get high score to break ties 506 507 // Keep best latency found 508 cand_cnt++; 509 if (choice < n_choice || 510 (choice == n_choice && 511 ((StressLCM && Compile::randomized_select(cand_cnt)) || 512 (!StressLCM && 513 (latency < n_latency || 514 (latency == n_latency && 515 (score < n_score))))))) { 516 choice = n_choice; 517 latency = n_latency; 518 score = n_score; 519 idx = i; // Also keep index in worklist 520 } 521 } // End of for all ready nodes in worklist 522 523 assert(idx >= 0, "index should be set"); 524 Node *n = worklist[(uint)idx]; // Get the winner 525 526 worklist.map((uint)idx, worklist.pop()); // Compress worklist 527 return n; 528 } 529 530 531 //------------------------------set_next_call---------------------------------- 532 void Block::set_next_call( Node *n, VectorSet &next_call, PhaseCFG* cfg) { 533 if( next_call.test_set(n->_idx) ) return; 534 for( uint i=0; i<n->len(); i++ ) { 535 Node *m = n->in(i); 536 if( !m ) continue; // must see all nodes in block that precede call 537 if (cfg->get_block_for_node(m) == this) { 538 set_next_call(m, next_call, cfg); 539 } 540 } 541 } 542 543 //------------------------------needed_for_next_call--------------------------- 544 // Set the flag 'next_call' for each Node that is needed for the next call to 545 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 546 // next subroutine call get priority - basically it moves things NOT needed 547 // for the next call till after the call. This prevents me from trying to 548 // carry lots of stuff live across a call. 549 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, PhaseCFG* cfg) { 550 // Find the next control-defining Node in this block 551 Node* call = NULL; 552 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 553 Node* m = this_call->fast_out(i); 554 if(cfg->get_block_for_node(m) == this && // Local-block user 555 m != this_call && // Not self-start node 556 m->is_MachCall() ) 557 call = m; 558 break; 559 } 560 if (call == NULL) return; // No next call (e.g., block end is near) 561 // Set next-call for all inputs to this call 562 set_next_call(call, next_call, cfg); 563 } 564 565 //------------------------------add_call_kills------------------------------------- 566 void Block::add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 567 // Fill in the kill mask for the call 568 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 569 if( !regs.Member(r) ) { // Not already defined by the call 570 // Save-on-call register? 571 if ((save_policy[r] == 'C') || 572 (save_policy[r] == 'A') || 573 ((save_policy[r] == 'E') && exclude_soe)) { 574 proj->_rout.Insert(r); 575 } 576 } 577 } 578 } 579 580 581 //------------------------------sched_call------------------------------------- 582 uint Block::sched_call( Matcher &matcher, PhaseCFG* cfg, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call ) { 583 RegMask regs; 584 585 // Schedule all the users of the call right now. All the users are 586 // projection Nodes, so they must be scheduled next to the call. 587 // Collect all the defined registers. 588 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 589 Node* n = mcall->fast_out(i); 590 assert( n->is_MachProj(), "" ); 591 int n_cnt = ready_cnt.at(n->_idx)-1; 592 ready_cnt.at_put(n->_idx, n_cnt); 593 assert( n_cnt == 0, "" ); 594 // Schedule next to call 595 _nodes.map(node_cnt++, n); 596 // Collect defined registers 597 regs.OR(n->out_RegMask()); 598 // Check for scheduling the next control-definer 599 if( n->bottom_type() == Type::CONTROL ) 600 // Warm up next pile of heuristic bits 601 needed_for_next_call(n, next_call, cfg); 602 603 // Children of projections are now all ready 604 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 605 Node* m = n->fast_out(j); // Get user 606 if(cfg->get_block_for_node(m) != this) { 607 continue; 608 } 609 if( m->is_Phi() ) continue; 610 int m_cnt = ready_cnt.at(m->_idx)-1; 611 ready_cnt.at_put(m->_idx, m_cnt); 612 if( m_cnt == 0 ) 613 worklist.push(m); 614 } 615 616 } 617 618 // Act as if the call defines the Frame Pointer. 619 // Certainly the FP is alive and well after the call. 620 regs.Insert(matcher.c_frame_pointer()); 621 622 // Set all registers killed and not already defined by the call. 623 uint r_cnt = mcall->tf()->range()->cnt(); 624 int op = mcall->ideal_Opcode(); 625 MachProjNode *proj = new (matcher.C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 626 cfg->map_node_to_block(proj, this); 627 _nodes.insert(node_cnt++, proj); 628 629 // Select the right register save policy. 630 const char * save_policy; 631 switch (op) { 632 case Op_CallRuntime: 633 case Op_CallLeaf: 634 case Op_CallLeafNoFP: 635 // Calling C code so use C calling convention 636 save_policy = matcher._c_reg_save_policy; 637 break; 638 639 case Op_CallStaticJava: 640 case Op_CallDynamicJava: 641 // Calling Java code so use Java calling convention 642 save_policy = matcher._register_save_policy; 643 break; 644 645 default: 646 ShouldNotReachHere(); 647 } 648 649 // When using CallRuntime mark SOE registers as killed by the call 650 // so values that could show up in the RegisterMap aren't live in a 651 // callee saved register since the register wouldn't know where to 652 // find them. CallLeaf and CallLeafNoFP are ok because they can't 653 // have debug info on them. Strictly speaking this only needs to be 654 // done for oops since idealreg2debugmask takes care of debug info 655 // references but there no way to handle oops differently than other 656 // pointers as far as the kill mask goes. 657 bool exclude_soe = op == Op_CallRuntime; 658 659 // If the call is a MethodHandle invoke, we need to exclude the 660 // register which is used to save the SP value over MH invokes from 661 // the mask. Otherwise this register could be used for 662 // deoptimization information. 663 if (op == Op_CallStaticJava) { 664 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 665 if (mcallstaticjava->_method_handle_invoke) 666 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 667 } 668 669 add_call_kills(proj, regs, save_policy, exclude_soe); 670 671 return node_cnt; 672 } 673 674 675 //------------------------------schedule_local--------------------------------- 676 // Topological sort within a block. Someday become a real scheduler. 677 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, GrowableArray<int> &ready_cnt, VectorSet &next_call) { 678 // Already "sorted" are the block start Node (as the first entry), and 679 // the block-ending Node and any trailing control projections. We leave 680 // these alone. PhiNodes and ParmNodes are made to follow the block start 681 // Node. Everything else gets topo-sorted. 682 683 #ifndef PRODUCT 684 if (cfg->trace_opto_pipelining()) { 685 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order); 686 for (uint i = 0;i < _nodes.size();i++) { 687 tty->print("# "); 688 _nodes[i]->fast_dump(); 689 } 690 tty->print_cr("#"); 691 } 692 #endif 693 694 // RootNode is already sorted 695 if( _nodes.size() == 1 ) return true; 696 697 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 698 uint node_cnt = end_idx(); 699 uint phi_cnt = 1; 700 uint i; 701 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 702 Node *n = _nodes[i]; 703 if( n->is_Phi() || // Found a PhiNode or ParmNode 704 (n->is_Proj() && n->in(0) == head()) ) { 705 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 706 _nodes.map(i,_nodes[phi_cnt]); 707 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front 708 } else { // All others 709 // Count block-local inputs to 'n' 710 uint cnt = n->len(); // Input count 711 uint local = 0; 712 for( uint j=0; j<cnt; j++ ) { 713 Node *m = n->in(j); 714 if( m && cfg->get_block_for_node(m) == this && !m->is_top() ) 715 local++; // One more block-local input 716 } 717 ready_cnt.at_put(n->_idx, local); // Count em up 718 719 #ifdef ASSERT 720 if( UseConcMarkSweepGC || UseG1GC ) { 721 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 722 // Check the precedence edges 723 for (uint prec = n->req(); prec < n->len(); prec++) { 724 Node* oop_store = n->in(prec); 725 if (oop_store != NULL) { 726 assert(cfg->get_block_for_node(oop_store)->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark"); 727 } 728 } 729 } 730 } 731 #endif 732 733 // A few node types require changing a required edge to a precedence edge 734 // before allocation. 735 if( n->is_Mach() && n->req() > TypeFunc::Parms && 736 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 737 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 738 // MemBarAcquire could be created without Precedent edge. 739 // del_req() replaces the specified edge with the last input edge 740 // and then removes the last edge. If the specified edge > number of 741 // edges the last edge will be moved outside of the input edges array 742 // and the edge will be lost. This is why this code should be 743 // executed only when Precedent (== TypeFunc::Parms) edge is present. 744 Node *x = n->in(TypeFunc::Parms); 745 n->del_req(TypeFunc::Parms); 746 n->add_prec(x); 747 } 748 } 749 } 750 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count 751 ready_cnt.at_put(_nodes[i2]->_idx, 0); 752 753 // All the prescheduled guys do not hold back internal nodes 754 uint i3; 755 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 756 Node *n = _nodes[i3]; // Get pre-scheduled 757 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 758 Node* m = n->fast_out(j); 759 if (cfg->get_block_for_node(m) == this) { // Local-block user 760 int m_cnt = ready_cnt.at(m->_idx)-1; 761 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 762 } 763 } 764 } 765 766 Node_List delay; 767 // Make a worklist 768 Node_List worklist; 769 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 770 Node *m = _nodes[i4]; 771 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 772 if (m->is_iteratively_computed()) { 773 // Push induction variable increments last to allow other uses 774 // of the phi to be scheduled first. The select() method breaks 775 // ties in scheduling by worklist order. 776 delay.push(m); 777 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 778 // Force the CreateEx to the top of the list so it's processed 779 // first and ends up at the start of the block. 780 worklist.insert(0, m); 781 } else { 782 worklist.push(m); // Then on to worklist! 783 } 784 } 785 } 786 while (delay.size()) { 787 Node* d = delay.pop(); 788 worklist.push(d); 789 } 790 791 // Warm up the 'next_call' heuristic bits 792 needed_for_next_call(_nodes[0], next_call, cfg); 793 794 #ifndef PRODUCT 795 if (cfg->trace_opto_pipelining()) { 796 for (uint j=0; j<_nodes.size(); j++) { 797 Node *n = _nodes[j]; 798 int idx = n->_idx; 799 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 800 tty->print("latency:%3d ", cfg->_node_latency->at_grow(idx)); 801 tty->print("%4d: %s\n", idx, n->Name()); 802 } 803 } 804 #endif 805 806 uint max_idx = (uint)ready_cnt.length(); 807 // Pull from worklist and schedule 808 while( worklist.size() ) { // Worklist is not ready 809 810 #ifndef PRODUCT 811 if (cfg->trace_opto_pipelining()) { 812 tty->print("# ready list:"); 813 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 814 Node *n = worklist[i]; // Get Node on worklist 815 tty->print(" %d", n->_idx); 816 } 817 tty->cr(); 818 } 819 #endif 820 821 // Select and pop a ready guy from worklist 822 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt); 823 _nodes.map(phi_cnt++,n); // Schedule him next 824 825 #ifndef PRODUCT 826 if (cfg->trace_opto_pipelining()) { 827 tty->print("# select %d: %s", n->_idx, n->Name()); 828 tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx)); 829 n->dump(); 830 if (Verbose) { 831 tty->print("# ready list:"); 832 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 833 Node *n = worklist[i]; // Get Node on worklist 834 tty->print(" %d", n->_idx); 835 } 836 tty->cr(); 837 } 838 } 839 840 #endif 841 if( n->is_MachCall() ) { 842 MachCallNode *mcall = n->as_MachCall(); 843 phi_cnt = sched_call(matcher, cfg, phi_cnt, worklist, ready_cnt, mcall, next_call); 844 continue; 845 } 846 847 if (n->is_Mach() && n->as_Mach()->has_call()) { 848 RegMask regs; 849 regs.Insert(matcher.c_frame_pointer()); 850 regs.OR(n->out_RegMask()); 851 852 MachProjNode *proj = new (matcher.C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 853 cfg->map_node_to_block(proj, this); 854 _nodes.insert(phi_cnt++, proj); 855 856 add_call_kills(proj, regs, matcher._c_reg_save_policy, false); 857 } 858 859 // Children are now all ready 860 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 861 Node* m = n->fast_out(i5); // Get user 862 if (cfg->get_block_for_node(m) != this) { 863 continue; 864 } 865 if( m->is_Phi() ) continue; 866 if (m->_idx >= max_idx) { // new node, skip it 867 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 868 continue; 869 } 870 int m_cnt = ready_cnt.at(m->_idx)-1; 871 ready_cnt.at_put(m->_idx, m_cnt); 872 if( m_cnt == 0 ) 873 worklist.push(m); 874 } 875 } 876 877 if( phi_cnt != end_idx() ) { 878 // did not schedule all. Retry, Bailout, or Die 879 Compile* C = matcher.C; 880 if (C->subsume_loads() == true && !C->failing()) { 881 // Retry with subsume_loads == false 882 // If this is the first failure, the sentinel string will "stick" 883 // to the Compile object, and the C2Compiler will see it and retry. 884 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 885 } 886 // assert( phi_cnt == end_idx(), "did not schedule all" ); 887 return false; 888 } 889 890 #ifndef PRODUCT 891 if (cfg->trace_opto_pipelining()) { 892 tty->print_cr("#"); 893 tty->print_cr("# after schedule_local"); 894 for (uint i = 0;i < _nodes.size();i++) { 895 tty->print("# "); 896 _nodes[i]->fast_dump(); 897 } 898 tty->cr(); 899 } 900 #endif 901 902 903 return true; 904 } 905 906 //--------------------------catch_cleanup_fix_all_inputs----------------------- 907 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 908 for (uint l = 0; l < use->len(); l++) { 909 if (use->in(l) == old_def) { 910 if (l < use->req()) { 911 use->set_req(l, new_def); 912 } else { 913 use->rm_prec(l); 914 use->add_prec(new_def); 915 l--; 916 } 917 } 918 } 919 } 920 921 //------------------------------catch_cleanup_find_cloned_def------------------ 922 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, PhaseCFG* cfg, int n_clone_idx) { 923 assert( use_blk != def_blk, "Inter-block cleanup only"); 924 925 // The use is some block below the Catch. Find and return the clone of the def 926 // that dominates the use. If there is no clone in a dominating block, then 927 // create a phi for the def in a dominating block. 928 929 // Find which successor block dominates this use. The successor 930 // blocks must all be single-entry (from the Catch only; I will have 931 // split blocks to make this so), hence they all dominate. 932 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 933 use_blk = use_blk->_idom; 934 935 // Find the successor 936 Node *fixup = NULL; 937 938 uint j; 939 for( j = 0; j < def_blk->_num_succs; j++ ) 940 if( use_blk == def_blk->_succs[j] ) 941 break; 942 943 if( j == def_blk->_num_succs ) { 944 // Block at same level in dom-tree is not a successor. It needs a 945 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 946 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 947 for(uint k = 1; k < use_blk->num_preds(); k++) { 948 Block* block = cfg->get_block_for_node(use_blk->pred(k)); 949 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, cfg, n_clone_idx)); 950 } 951 952 // Check to see if the use_blk already has an identical phi inserted. 953 // If it exists, it will be at the first position since all uses of a 954 // def are processed together. 955 Node *phi = use_blk->_nodes[1]; 956 if( phi->is_Phi() ) { 957 fixup = phi; 958 for (uint k = 1; k < use_blk->num_preds(); k++) { 959 if (phi->in(k) != inputs[k]) { 960 // Not a match 961 fixup = NULL; 962 break; 963 } 964 } 965 } 966 967 // If an existing PhiNode was not found, make a new one. 968 if (fixup == NULL) { 969 Node *new_phi = PhiNode::make(use_blk->head(), def); 970 use_blk->_nodes.insert(1, new_phi); 971 cfg->map_node_to_block(new_phi, use_blk); 972 for (uint k = 1; k < use_blk->num_preds(); k++) { 973 new_phi->set_req(k, inputs[k]); 974 } 975 fixup = new_phi; 976 } 977 978 } else { 979 // Found the use just below the Catch. Make it use the clone. 980 fixup = use_blk->_nodes[n_clone_idx]; 981 } 982 983 return fixup; 984 } 985 986 //--------------------------catch_cleanup_intra_block-------------------------- 987 // Fix all input edges in use that reference "def". The use is in the same 988 // block as the def and both have been cloned in each successor block. 989 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 990 991 // Both the use and def have been cloned. For each successor block, 992 // get the clone of the use, and make its input the clone of the def 993 // found in that block. 994 995 uint use_idx = blk->find_node(use); 996 uint offset_idx = use_idx - beg; 997 for( uint k = 0; k < blk->_num_succs; k++ ) { 998 // Get clone in each successor block 999 Block *sb = blk->_succs[k]; 1000 Node *clone = sb->_nodes[offset_idx+1]; 1001 assert( clone->Opcode() == use->Opcode(), "" ); 1002 1003 // Make use-clone reference the def-clone 1004 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]); 1005 } 1006 } 1007 1008 //------------------------------catch_cleanup_inter_block--------------------- 1009 // Fix all input edges in use that reference "def". The use is in a different 1010 // block than the def. 1011 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, PhaseCFG* cfg, int n_clone_idx) { 1012 if( !use_blk ) return; // Can happen if the use is a precedence edge 1013 1014 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, cfg, n_clone_idx); 1015 catch_cleanup_fix_all_inputs(use, def, new_def); 1016 } 1017 1018 //------------------------------call_catch_cleanup----------------------------- 1019 // If we inserted any instructions between a Call and his CatchNode, 1020 // clone the instructions on all paths below the Catch. 1021 void Block::call_catch_cleanup(PhaseCFG* cfg, Compile* C) { 1022 1023 // End of region to clone 1024 uint end = end_idx(); 1025 if( !_nodes[end]->is_Catch() ) return; 1026 // Start of region to clone 1027 uint beg = end; 1028 while(!_nodes[beg-1]->is_MachProj() || 1029 !_nodes[beg-1]->in(0)->is_MachCall() ) { 1030 beg--; 1031 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1032 } 1033 // Range of inserted instructions is [beg, end) 1034 if( beg == end ) return; 1035 1036 // Clone along all Catch output paths. Clone area between the 'beg' and 1037 // 'end' indices. 1038 for( uint i = 0; i < _num_succs; i++ ) { 1039 Block *sb = _succs[i]; 1040 // Clone the entire area; ignoring the edge fixup for now. 1041 for( uint j = end; j > beg; j-- ) { 1042 // It is safe here to clone a node with anti_dependence 1043 // since clones dominate on each path. 1044 Node *clone = _nodes[j-1]->clone(); 1045 sb->_nodes.insert( 1, clone ); 1046 cfg->map_node_to_block(clone, sb); 1047 } 1048 } 1049 1050 1051 // Fixup edges. Check the def-use info per cloned Node 1052 for(uint i2 = beg; i2 < end; i2++ ) { 1053 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1054 Node *n = _nodes[i2]; // Node that got cloned 1055 // Need DU safe iterator because of edge manipulation in calls. 1056 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1057 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1058 out->push(n->fast_out(j1)); 1059 } 1060 uint max = out->size(); 1061 for (uint j = 0; j < max; j++) {// For all users 1062 Node *use = out->pop(); 1063 Block *buse = cfg->get_block_for_node(use); 1064 if( use->is_Phi() ) { 1065 for( uint k = 1; k < use->req(); k++ ) 1066 if( use->in(k) == n ) { 1067 Block* block = cfg->get_block_for_node(buse->pred(k)); 1068 Node *fixup = catch_cleanup_find_cloned_def(block, n, this, cfg, n_clone_idx); 1069 use->set_req(k, fixup); 1070 } 1071 } else { 1072 if (this == buse) { 1073 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx); 1074 } else { 1075 catch_cleanup_inter_block(use, buse, n, this, cfg, n_clone_idx); 1076 } 1077 } 1078 } // End for all users 1079 1080 } // End of for all Nodes in cloned area 1081 1082 // Remove the now-dead cloned ops 1083 for(uint i3 = beg; i3 < end; i3++ ) { 1084 _nodes[beg]->disconnect_inputs(NULL, C); 1085 _nodes.remove(beg); 1086 } 1087 1088 // If the successor blocks have a CreateEx node, move it back to the top 1089 for(uint i4 = 0; i4 < _num_succs; i4++ ) { 1090 Block *sb = _succs[i4]; 1091 uint new_cnt = end - beg; 1092 // Remove any newly created, but dead, nodes. 1093 for( uint j = new_cnt; j > 0; j-- ) { 1094 Node *n = sb->_nodes[j]; 1095 if (n->outcnt() == 0 && 1096 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1097 n->disconnect_inputs(NULL, C); 1098 sb->_nodes.remove(j); 1099 new_cnt--; 1100 } 1101 } 1102 // If any newly created nodes remain, move the CreateEx node to the top 1103 if (new_cnt > 0) { 1104 Node *cex = sb->_nodes[1+new_cnt]; 1105 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1106 sb->_nodes.remove(1+new_cnt); 1107 sb->_nodes.insert(1,cex); 1108 } 1109 } 1110 } 1111 }