1 /* 2 * Copyright (c) 1998, 2014, 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/ad.hpp" 28 #include "opto/block.hpp" 29 #include "opto/c2compiler.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/machnode.hpp" 33 #include "opto/runtime.hpp" 34 35 // Optimization - Graph Style 36 37 // Check whether val is not-null-decoded compressed oop, 38 // i.e. will grab into the base of the heap if it represents NULL. 39 static bool accesses_heap_base_zone(Node *val) { 40 if (Universe::narrow_oop_base() > 0) { // Implies UseCompressedOops. 41 if (val && val->is_Mach()) { 42 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) { 43 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that 44 // decode NULL to point to the heap base (Decode_NN). 45 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) { 46 return true; 47 } 48 } 49 // Must recognize load operation with Decode matched in memory operand. 50 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected() 51 // returns true everywhere else. On PPC, no such memory operands 52 // exist, therefore we did not yet implement a check for such operands. 53 NOT_AIX(Unimplemented()); 54 } 55 } 56 return false; 57 } 58 59 static bool needs_explicit_null_check_for_read(Node *val) { 60 // On some OSes (AIX) the page at address 0 is only write protected. 61 // If so, only Store operations will trap. 62 if (os::zero_page_read_protected()) { 63 return false; // Implicit null check will work. 64 } 65 // Also a read accessing the base of a heap-based compressed heap will trap. 66 if (accesses_heap_base_zone(val) && // Hits the base zone page. 67 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected. 68 return false; 69 } 70 71 return true; 72 } 73 74 //------------------------------implicit_null_check---------------------------- 75 // Detect implicit-null-check opportunities. Basically, find NULL checks 76 // with suitable memory ops nearby. Use the memory op to do the NULL check. 77 // I can generate a memory op if there is not one nearby. 78 // The proj is the control projection for the not-null case. 79 // The val is the pointer being checked for nullness or 80 // decodeHeapOop_not_null node if it did not fold into address. 81 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { 82 // Assume if null check need for 0 offset then always needed 83 // Intel solaris doesn't support any null checks yet and no 84 // mechanism exists (yet) to set the switches at an os_cpu level 85 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 86 87 // Make sure the ptr-is-null path appears to be uncommon! 88 float f = block->end()->as_MachIf()->_prob; 89 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 90 if( f > PROB_UNLIKELY_MAG(4) ) return; 91 92 uint bidx = 0; // Capture index of value into memop 93 bool was_store; // Memory op is a store op 94 95 // Get the successor block for if the test ptr is non-null 96 Block* not_null_block; // this one goes with the proj 97 Block* null_block; 98 if (block->get_node(block->number_of_nodes()-1) == proj) { 99 null_block = block->_succs[0]; 100 not_null_block = block->_succs[1]; 101 } else { 102 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other"); 103 not_null_block = block->_succs[0]; 104 null_block = block->_succs[1]; 105 } 106 while (null_block->is_Empty() == Block::empty_with_goto) { 107 null_block = null_block->_succs[0]; 108 } 109 110 // Search the exception block for an uncommon trap. 111 // (See Parse::do_if and Parse::do_ifnull for the reason 112 // we need an uncommon trap. Briefly, we need a way to 113 // detect failure of this optimization, as in 6366351.) 114 { 115 bool found_trap = false; 116 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { 117 Node* nn = null_block->get_node(i1); 118 if (nn->is_MachCall() && 119 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 120 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 121 if (trtype->isa_int() && trtype->is_int()->is_con()) { 122 jint tr_con = trtype->is_int()->get_con(); 123 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 124 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 125 assert((int)reason < (int)BitsPerInt, "recode bit map"); 126 if (is_set_nth_bit(allowed_reasons, (int) reason) 127 && action != Deoptimization::Action_none) { 128 // This uncommon trap is sure to recompile, eventually. 129 // When that happens, C->too_many_traps will prevent 130 // this transformation from happening again. 131 found_trap = true; 132 } 133 } 134 break; 135 } 136 } 137 if (!found_trap) { 138 // We did not find an uncommon trap. 139 return; 140 } 141 } 142 143 // Check for decodeHeapOop_not_null node which did not fold into address 144 bool is_decoden = ((intptr_t)val) & 1; 145 val = (Node*)(((intptr_t)val) & ~1); 146 147 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 148 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 149 150 // Search the successor block for a load or store who's base value is also 151 // the tested value. There may be several. 152 Node_List *out = new Node_List(Thread::current()->resource_area()); 153 MachNode *best = NULL; // Best found so far 154 for (DUIterator i = val->outs(); val->has_out(i); i++) { 155 Node *m = val->out(i); 156 if( !m->is_Mach() ) continue; 157 MachNode *mach = m->as_Mach(); 158 was_store = false; 159 int iop = mach->ideal_Opcode(); 160 switch( iop ) { 161 case Op_LoadB: 162 case Op_LoadUB: 163 case Op_LoadUS: 164 case Op_LoadD: 165 case Op_LoadF: 166 case Op_LoadI: 167 case Op_LoadL: 168 case Op_LoadP: 169 case Op_LoadN: 170 case Op_LoadS: 171 case Op_LoadKlass: 172 case Op_LoadNKlass: 173 case Op_LoadRange: 174 case Op_LoadD_unaligned: 175 case Op_LoadL_unaligned: 176 assert(mach->in(2) == val, "should be address"); 177 break; 178 case Op_StoreB: 179 case Op_StoreC: 180 case Op_StoreCM: 181 case Op_StoreD: 182 case Op_StoreF: 183 case Op_StoreI: 184 case Op_StoreL: 185 case Op_StoreP: 186 case Op_StoreN: 187 case Op_StoreNKlass: 188 was_store = true; // Memory op is a store op 189 // Stores will have their address in slot 2 (memory in slot 1). 190 // If the value being nul-checked is in another slot, it means we 191 // are storing the checked value, which does NOT check the value! 192 if( mach->in(2) != val ) continue; 193 break; // Found a memory op? 194 case Op_StrComp: 195 case Op_StrEquals: 196 case Op_StrIndexOf: 197 case Op_AryEq: 198 case Op_EncodeISOArray: 199 // Not a legit memory op for implicit null check regardless of 200 // embedded loads 201 continue; 202 default: // Also check for embedded loads 203 if( !mach->needs_anti_dependence_check() ) 204 continue; // Not an memory op; skip it 205 if( must_clone[iop] ) { 206 // Do not move nodes which produce flags because 207 // RA will try to clone it to place near branch and 208 // it will cause recompilation, see clone_node(). 209 continue; 210 } 211 { 212 // Check that value is used in memory address in 213 // instructions with embedded load (CmpP val1,(val2+off)). 214 Node* base; 215 Node* index; 216 const MachOper* oper = mach->memory_inputs(base, index); 217 if (oper == NULL || oper == (MachOper*)-1) { 218 continue; // Not an memory op; skip it 219 } 220 if (val == base || 221 val == index && val->bottom_type()->isa_narrowoop()) { 222 break; // Found it 223 } else { 224 continue; // Skip it 225 } 226 } 227 break; 228 } 229 230 // On some OSes (AIX) the page at address 0 is only write protected. 231 // If so, only Store operations will trap. 232 // But a read accessing the base of a heap-based compressed heap will trap. 233 if (!was_store && needs_explicit_null_check_for_read(val)) { 234 continue; 235 } 236 237 // check if the offset is not too high for implicit exception 238 { 239 intptr_t offset = 0; 240 const TypePtr *adr_type = NULL; // Do not need this return value here 241 const Node* base = mach->get_base_and_disp(offset, adr_type); 242 if (base == NULL || base == NodeSentinel) { 243 // Narrow oop address doesn't have base, only index 244 if( val->bottom_type()->isa_narrowoop() && 245 MacroAssembler::needs_explicit_null_check(offset) ) 246 continue; // Give up if offset is beyond page size 247 // cannot reason about it; is probably not implicit null exception 248 } else { 249 const TypePtr* tptr; 250 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 251 Universe::narrow_klass_shift() == 0)) { 252 // 32-bits narrow oop can be the base of address expressions 253 tptr = base->get_ptr_type(); 254 } else { 255 // only regular oops are expected here 256 tptr = base->bottom_type()->is_ptr(); 257 } 258 // Give up if offset is not a compile-time constant 259 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 260 continue; 261 offset += tptr->_offset; // correct if base is offseted 262 if( MacroAssembler::needs_explicit_null_check(offset) ) 263 continue; // Give up is reference is beyond 4K page size 264 } 265 } 266 267 // Check ctrl input to see if the null-check dominates the memory op 268 Block *cb = get_block_for_node(mach); 269 cb = cb->_idom; // Always hoist at least 1 block 270 if( !was_store ) { // Stores can be hoisted only one block 271 while( cb->_dom_depth > (block->_dom_depth + 1)) 272 cb = cb->_idom; // Hoist loads as far as we want 273 // The non-null-block should dominate the memory op, too. Live 274 // range spilling will insert a spill in the non-null-block if it is 275 // needs to spill the memory op for an implicit null check. 276 if (cb->_dom_depth == (block->_dom_depth + 1)) { 277 if (cb != not_null_block) continue; 278 cb = cb->_idom; 279 } 280 } 281 if( cb != block ) continue; 282 283 // Found a memory user; see if it can be hoisted to check-block 284 uint vidx = 0; // Capture index of value into memop 285 uint j; 286 for( j = mach->req()-1; j > 0; j-- ) { 287 if( mach->in(j) == val ) { 288 vidx = j; 289 // Ignore DecodeN val which could be hoisted to where needed. 290 if( is_decoden ) continue; 291 } 292 // Block of memory-op input 293 Block *inb = get_block_for_node(mach->in(j)); 294 Block *b = block; // Start from nul check 295 while( b != inb && b->_dom_depth > inb->_dom_depth ) 296 b = b->_idom; // search upwards for input 297 // See if input dominates null check 298 if( b != inb ) 299 break; 300 } 301 if( j > 0 ) 302 continue; 303 Block *mb = get_block_for_node(mach); 304 // Hoisting stores requires more checks for the anti-dependence case. 305 // Give up hoisting if we have to move the store past any load. 306 if( was_store ) { 307 Block *b = mb; // Start searching here for a local load 308 // mach use (faulting) trying to hoist 309 // n might be blocker to hoisting 310 while( b != block ) { 311 uint k; 312 for( k = 1; k < b->number_of_nodes(); k++ ) { 313 Node *n = b->get_node(k); 314 if( n->needs_anti_dependence_check() && 315 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 316 break; // Found anti-dependent load 317 } 318 if( k < b->number_of_nodes() ) 319 break; // Found anti-dependent load 320 // Make sure control does not do a merge (would have to check allpaths) 321 if( b->num_preds() != 2 ) break; 322 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 323 } 324 if( b != block ) continue; 325 } 326 327 // Make sure this memory op is not already being used for a NullCheck 328 Node *e = mb->end(); 329 if( e->is_MachNullCheck() && e->in(1) == mach ) 330 continue; // Already being used as a NULL check 331 332 // Found a candidate! Pick one with least dom depth - the highest 333 // in the dom tree should be closest to the null check. 334 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 335 best = mach; 336 bidx = vidx; 337 } 338 } 339 // No candidate! 340 if (best == NULL) { 341 return; 342 } 343 344 // ---- Found an implicit null check 345 extern int implicit_null_checks; 346 implicit_null_checks++; 347 348 if( is_decoden ) { 349 // Check if we need to hoist decodeHeapOop_not_null first. 350 Block *valb = get_block_for_node(val); 351 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 352 // Hoist it up to the end of the test block. 353 valb->find_remove(val); 354 block->add_inst(val); 355 map_node_to_block(val, block); 356 // DecodeN on x86 may kill flags. Check for flag-killing projections 357 // that also need to be hoisted. 358 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 359 Node* n = val->fast_out(j); 360 if( n->is_MachProj() ) { 361 get_block_for_node(n)->find_remove(n); 362 block->add_inst(n); 363 map_node_to_block(n, block); 364 } 365 } 366 } 367 } 368 // Hoist the memory candidate up to the end of the test block. 369 Block *old_block = get_block_for_node(best); 370 old_block->find_remove(best); 371 block->add_inst(best); 372 map_node_to_block(best, block); 373 374 // Move the control dependence 375 if (best->in(0) && best->in(0) == old_block->head()) 376 best->set_req(0, block->head()); 377 378 // Check for flag-killing projections that also need to be hoisted 379 // Should be DU safe because no edge updates. 380 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 381 Node* n = best->fast_out(j); 382 if( n->is_MachProj() ) { 383 get_block_for_node(n)->find_remove(n); 384 block->add_inst(n); 385 map_node_to_block(n, block); 386 } 387 } 388 389 // proj==Op_True --> ne test; proj==Op_False --> eq test. 390 // One of two graph shapes got matched: 391 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 392 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 393 // NULL checks are always branch-if-eq. If we see a IfTrue projection 394 // then we are replacing a 'ne' test with a 'eq' NULL check test. 395 // We need to flip the projections to keep the same semantics. 396 if( proj->Opcode() == Op_IfTrue ) { 397 // Swap order of projections in basic block to swap branch targets 398 Node *tmp1 = block->get_node(block->end_idx()+1); 399 Node *tmp2 = block->get_node(block->end_idx()+2); 400 block->map_node(tmp2, block->end_idx()+1); 401 block->map_node(tmp1, block->end_idx()+2); 402 Node *tmp = new Node(C->top()); // Use not NULL input 403 tmp1->replace_by(tmp); 404 tmp2->replace_by(tmp1); 405 tmp->replace_by(tmp2); 406 tmp->destruct(); 407 } 408 409 // Remove the existing null check; use a new implicit null check instead. 410 // Since schedule-local needs precise def-use info, we need to correct 411 // it as well. 412 Node *old_tst = proj->in(0); 413 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx); 414 block->map_node(nul_chk, block->end_idx()); 415 map_node_to_block(nul_chk, block); 416 // Redirect users of old_test to nul_chk 417 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 418 old_tst->last_out(i2)->set_req(0, nul_chk); 419 // Clean-up any dead code 420 for (uint i3 = 0; i3 < old_tst->req(); i3++) 421 old_tst->set_req(i3, NULL); 422 423 latency_from_uses(nul_chk); 424 latency_from_uses(best); 425 } 426 427 428 //------------------------------select----------------------------------------- 429 // Select a nice fellow from the worklist to schedule next. If there is only 430 // one choice, then use it. Projections take top priority for correctness 431 // reasons - if I see a projection, then it is next. There are a number of 432 // other special cases, for instructions that consume condition codes, et al. 433 // These are chosen immediately. Some instructions are required to immediately 434 // precede the last instruction in the block, and these are taken last. Of the 435 // remaining cases (most), choose the instruction with the greatest latency 436 // (that is, the most number of pseudo-cycles required to the end of the 437 // routine). If there is a tie, choose the instruction with the most inputs. 438 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 439 440 // If only a single entry on the stack, use it 441 uint cnt = worklist.size(); 442 if (cnt == 1) { 443 Node *n = worklist[0]; 444 worklist.map(0,worklist.pop()); 445 return n; 446 } 447 448 uint choice = 0; // Bigger is most important 449 uint latency = 0; // Bigger is scheduled first 450 uint score = 0; // Bigger is better 451 int idx = -1; // Index in worklist 452 int cand_cnt = 0; // Candidate count 453 454 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 455 // Order in worklist is used to break ties. 456 // See caller for how this is used to delay scheduling 457 // of induction variable increments to after the other 458 // uses of the phi are scheduled. 459 Node *n = worklist[i]; // Get Node on worklist 460 461 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 462 if( n->is_Proj() || // Projections always win 463 n->Opcode()== Op_Con || // So does constant 'Top' 464 iop == Op_CreateEx || // Create-exception must start block 465 iop == Op_CheckCastPP 466 ) { 467 // select the node n 468 // remove n from worklist and retain the order of remaining nodes 469 worklist.remove((uint)i); 470 return n; 471 } 472 473 // Final call in a block must be adjacent to 'catch' 474 Node *e = block->end(); 475 if( e->is_Catch() && e->in(0)->in(0) == n ) 476 continue; 477 478 // Memory op for an implicit null check has to be at the end of the block 479 if( e->is_MachNullCheck() && e->in(1) == n ) 480 continue; 481 482 // Schedule IV increment last. 483 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 484 e->in(1)->in(1) == n && n->is_iteratively_computed()) 485 continue; 486 487 uint n_choice = 2; 488 489 // See if this instruction is consumed by a branch. If so, then (as the 490 // branch is the last instruction in the basic block) force it to the 491 // end of the basic block 492 if ( must_clone[iop] ) { 493 // See if any use is a branch 494 bool found_machif = false; 495 496 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 497 Node* use = n->fast_out(j); 498 499 // The use is a conditional branch, make them adjacent 500 if (use->is_MachIf() && get_block_for_node(use) == block) { 501 found_machif = true; 502 break; 503 } 504 505 // More than this instruction pending for successor to be ready, 506 // don't choose this if other opportunities are ready 507 if (ready_cnt.at(use->_idx) > 1) 508 n_choice = 1; 509 } 510 511 // loop terminated, prefer not to use this instruction 512 if (found_machif) 513 continue; 514 } 515 516 // See if this has a predecessor that is "must_clone", i.e. sets the 517 // condition code. If so, choose this first 518 for (uint j = 0; j < n->req() ; j++) { 519 Node *inn = n->in(j); 520 if (inn) { 521 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 522 n_choice = 3; 523 break; 524 } 525 } 526 } 527 528 // MachTemps should be scheduled last so they are near their uses 529 if (n->is_MachTemp()) { 530 n_choice = 1; 531 } 532 533 uint n_latency = get_latency_for_node(n); 534 uint n_score = n->req(); // Many inputs get high score to break ties 535 536 // Keep best latency found 537 cand_cnt++; 538 if (choice < n_choice || 539 (choice == n_choice && 540 ((StressLCM && Compile::randomized_select(cand_cnt)) || 541 (!StressLCM && 542 (latency < n_latency || 543 (latency == n_latency && 544 (score < n_score))))))) { 545 choice = n_choice; 546 latency = n_latency; 547 score = n_score; 548 idx = i; // Also keep index in worklist 549 } 550 } // End of for all ready nodes in worklist 551 552 assert(idx >= 0, "index should be set"); 553 Node *n = worklist[(uint)idx]; // Get the winner 554 555 // select the node n 556 // remove n from worklist and retain the order of remaining nodes 557 worklist.remove((uint)idx); 558 return n; 559 } 560 561 562 //------------------------------set_next_call---------------------------------- 563 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 564 if( next_call.test_set(n->_idx) ) return; 565 for( uint i=0; i<n->len(); i++ ) { 566 Node *m = n->in(i); 567 if( !m ) continue; // must see all nodes in block that precede call 568 if (get_block_for_node(m) == block) { 569 set_next_call(block, m, next_call); 570 } 571 } 572 } 573 574 //------------------------------needed_for_next_call--------------------------- 575 // Set the flag 'next_call' for each Node that is needed for the next call to 576 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 577 // next subroutine call get priority - basically it moves things NOT needed 578 // for the next call till after the call. This prevents me from trying to 579 // carry lots of stuff live across a call. 580 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 581 // Find the next control-defining Node in this block 582 Node* call = NULL; 583 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 584 Node* m = this_call->fast_out(i); 585 if (get_block_for_node(m) == block && // Local-block user 586 m != this_call && // Not self-start node 587 m->is_MachCall()) { 588 call = m; 589 break; 590 } 591 } 592 if (call == NULL) return; // No next call (e.g., block end is near) 593 // Set next-call for all inputs to this call 594 set_next_call(block, call, next_call); 595 } 596 597 //------------------------------add_call_kills------------------------------------- 598 // helper function that adds caller save registers to MachProjNode 599 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 600 // Fill in the kill mask for the call 601 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 602 if( !regs.Member(r) ) { // Not already defined by the call 603 // Save-on-call register? 604 if ((save_policy[r] == 'C') || 605 (save_policy[r] == 'A') || 606 ((save_policy[r] == 'E') && exclude_soe)) { 607 proj->_rout.Insert(r); 608 } 609 } 610 } 611 } 612 613 614 //------------------------------sched_call------------------------------------- 615 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 616 RegMask regs; 617 618 // Schedule all the users of the call right now. All the users are 619 // projection Nodes, so they must be scheduled next to the call. 620 // Collect all the defined registers. 621 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 622 Node* n = mcall->fast_out(i); 623 assert( n->is_MachProj(), "" ); 624 int n_cnt = ready_cnt.at(n->_idx)-1; 625 ready_cnt.at_put(n->_idx, n_cnt); 626 assert( n_cnt == 0, "" ); 627 // Schedule next to call 628 block->map_node(n, node_cnt++); 629 // Collect defined registers 630 regs.OR(n->out_RegMask()); 631 // Check for scheduling the next control-definer 632 if( n->bottom_type() == Type::CONTROL ) 633 // Warm up next pile of heuristic bits 634 needed_for_next_call(block, n, next_call); 635 636 // Children of projections are now all ready 637 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 638 Node* m = n->fast_out(j); // Get user 639 if(get_block_for_node(m) != block) { 640 continue; 641 } 642 if( m->is_Phi() ) continue; 643 int m_cnt = ready_cnt.at(m->_idx)-1; 644 ready_cnt.at_put(m->_idx, m_cnt); 645 if( m_cnt == 0 ) 646 worklist.push(m); 647 } 648 649 } 650 651 // Act as if the call defines the Frame Pointer. 652 // Certainly the FP is alive and well after the call. 653 regs.Insert(_matcher.c_frame_pointer()); 654 655 // Set all registers killed and not already defined by the call. 656 uint r_cnt = mcall->tf()->range()->cnt(); 657 int op = mcall->ideal_Opcode(); 658 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 659 map_node_to_block(proj, block); 660 block->insert_node(proj, node_cnt++); 661 662 // Select the right register save policy. 663 const char * save_policy; 664 switch (op) { 665 case Op_CallRuntime: 666 case Op_CallLeaf: 667 case Op_CallLeafNoFP: 668 // Calling C code so use C calling convention 669 save_policy = _matcher._c_reg_save_policy; 670 break; 671 672 case Op_CallStaticJava: 673 case Op_CallDynamicJava: 674 // Calling Java code so use Java calling convention 675 save_policy = _matcher._register_save_policy; 676 break; 677 678 default: 679 ShouldNotReachHere(); 680 } 681 682 // When using CallRuntime mark SOE registers as killed by the call 683 // so values that could show up in the RegisterMap aren't live in a 684 // callee saved register since the register wouldn't know where to 685 // find them. CallLeaf and CallLeafNoFP are ok because they can't 686 // have debug info on them. Strictly speaking this only needs to be 687 // done for oops since idealreg2debugmask takes care of debug info 688 // references but there no way to handle oops differently than other 689 // pointers as far as the kill mask goes. 690 bool exclude_soe = op == Op_CallRuntime; 691 692 // If the call is a MethodHandle invoke, we need to exclude the 693 // register which is used to save the SP value over MH invokes from 694 // the mask. Otherwise this register could be used for 695 // deoptimization information. 696 if (op == Op_CallStaticJava) { 697 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 698 if (mcallstaticjava->_method_handle_invoke) 699 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 700 } 701 702 add_call_kills(proj, regs, save_policy, exclude_soe); 703 704 return node_cnt; 705 } 706 707 708 //------------------------------schedule_local--------------------------------- 709 // Topological sort within a block. Someday become a real scheduler. 710 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) { 711 // Already "sorted" are the block start Node (as the first entry), and 712 // the block-ending Node and any trailing control projections. We leave 713 // these alone. PhiNodes and ParmNodes are made to follow the block start 714 // Node. Everything else gets topo-sorted. 715 716 #ifndef PRODUCT 717 if (trace_opto_pipelining()) { 718 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 719 for (uint i = 0;i < block->number_of_nodes(); i++) { 720 tty->print("# "); 721 block->get_node(i)->fast_dump(); 722 } 723 tty->print_cr("#"); 724 } 725 #endif 726 727 // RootNode is already sorted 728 if (block->number_of_nodes() == 1) { 729 return true; 730 } 731 732 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 733 uint node_cnt = block->end_idx(); 734 uint phi_cnt = 1; 735 uint i; 736 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 737 Node *n = block->get_node(i); 738 if( n->is_Phi() || // Found a PhiNode or ParmNode 739 (n->is_Proj() && n->in(0) == block->head()) ) { 740 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 741 block->map_node(block->get_node(phi_cnt), i); 742 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 743 } else { // All others 744 // Count block-local inputs to 'n' 745 uint cnt = n->len(); // Input count 746 uint local = 0; 747 for( uint j=0; j<cnt; j++ ) { 748 Node *m = n->in(j); 749 if( m && get_block_for_node(m) == block && !m->is_top() ) 750 local++; // One more block-local input 751 } 752 ready_cnt.at_put(n->_idx, local); // Count em up 753 754 #ifdef ASSERT 755 if( UseConcMarkSweepGC || UseG1GC ) { 756 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 757 // Check the precedence edges 758 for (uint prec = n->req(); prec < n->len(); prec++) { 759 Node* oop_store = n->in(prec); 760 if (oop_store != NULL) { 761 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 762 } 763 } 764 } 765 } 766 #endif 767 768 // A few node types require changing a required edge to a precedence edge 769 // before allocation. 770 if( n->is_Mach() && n->req() > TypeFunc::Parms && 771 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 772 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 773 // MemBarAcquire could be created without Precedent edge. 774 // del_req() replaces the specified edge with the last input edge 775 // and then removes the last edge. If the specified edge > number of 776 // edges the last edge will be moved outside of the input edges array 777 // and the edge will be lost. This is why this code should be 778 // executed only when Precedent (== TypeFunc::Parms) edge is present. 779 Node *x = n->in(TypeFunc::Parms); 780 n->del_req(TypeFunc::Parms); 781 n->add_prec(x); 782 } 783 } 784 } 785 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 786 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 787 788 // All the prescheduled guys do not hold back internal nodes 789 uint i3; 790 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 791 Node *n = block->get_node(i3); // Get pre-scheduled 792 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 793 Node* m = n->fast_out(j); 794 if (get_block_for_node(m) == block) { // Local-block user 795 int m_cnt = ready_cnt.at(m->_idx)-1; 796 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 797 } 798 } 799 } 800 801 Node_List delay; 802 // Make a worklist 803 Node_List worklist; 804 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 805 Node *m = block->get_node(i4); 806 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 807 if (m->is_iteratively_computed()) { 808 // Push induction variable increments last to allow other uses 809 // of the phi to be scheduled first. The select() method breaks 810 // ties in scheduling by worklist order. 811 delay.push(m); 812 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 813 // Force the CreateEx to the top of the list so it's processed 814 // first and ends up at the start of the block. 815 worklist.insert(0, m); 816 } else { 817 worklist.push(m); // Then on to worklist! 818 } 819 } 820 } 821 while (delay.size()) { 822 Node* d = delay.pop(); 823 worklist.push(d); 824 } 825 826 // Warm up the 'next_call' heuristic bits 827 needed_for_next_call(block, block->head(), next_call); 828 829 #ifndef PRODUCT 830 if (trace_opto_pipelining()) { 831 for (uint j=0; j< block->number_of_nodes(); j++) { 832 Node *n = block->get_node(j); 833 int idx = n->_idx; 834 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 835 tty->print("latency:%3d ", get_latency_for_node(n)); 836 tty->print("%4d: %s\n", idx, n->Name()); 837 } 838 } 839 #endif 840 841 uint max_idx = (uint)ready_cnt.length(); 842 // Pull from worklist and schedule 843 while( worklist.size() ) { // Worklist is not ready 844 845 #ifndef PRODUCT 846 if (trace_opto_pipelining()) { 847 tty->print("# ready list:"); 848 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 849 Node *n = worklist[i]; // Get Node on worklist 850 tty->print(" %d", n->_idx); 851 } 852 tty->cr(); 853 } 854 #endif 855 856 // Select and pop a ready guy from worklist 857 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt); 858 block->map_node(n, phi_cnt++); // Schedule him next 859 860 #ifndef PRODUCT 861 if (trace_opto_pipelining()) { 862 tty->print("# select %d: %s", n->_idx, n->Name()); 863 tty->print(", latency:%d", get_latency_for_node(n)); 864 n->dump(); 865 if (Verbose) { 866 tty->print("# ready list:"); 867 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 868 Node *n = worklist[i]; // Get Node on worklist 869 tty->print(" %d", n->_idx); 870 } 871 tty->cr(); 872 } 873 } 874 875 #endif 876 if( n->is_MachCall() ) { 877 MachCallNode *mcall = n->as_MachCall(); 878 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 879 continue; 880 } 881 882 if (n->is_Mach() && n->as_Mach()->has_call()) { 883 RegMask regs; 884 regs.Insert(_matcher.c_frame_pointer()); 885 regs.OR(n->out_RegMask()); 886 887 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 888 map_node_to_block(proj, block); 889 block->insert_node(proj, phi_cnt++); 890 891 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 892 } 893 894 // Children are now all ready 895 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 896 Node* m = n->fast_out(i5); // Get user 897 if (get_block_for_node(m) != block) { 898 continue; 899 } 900 if( m->is_Phi() ) continue; 901 if (m->_idx >= max_idx) { // new node, skip it 902 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 903 continue; 904 } 905 int m_cnt = ready_cnt.at(m->_idx)-1; 906 ready_cnt.at_put(m->_idx, m_cnt); 907 if( m_cnt == 0 ) 908 worklist.push(m); 909 } 910 } 911 912 if( phi_cnt != block->end_idx() ) { 913 // did not schedule all. Retry, Bailout, or Die 914 if (C->subsume_loads() == true && !C->failing()) { 915 // Retry with subsume_loads == false 916 // If this is the first failure, the sentinel string will "stick" 917 // to the Compile object, and the C2Compiler will see it and retry. 918 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 919 } 920 // assert( phi_cnt == end_idx(), "did not schedule all" ); 921 return false; 922 } 923 924 #ifndef PRODUCT 925 if (trace_opto_pipelining()) { 926 tty->print_cr("#"); 927 tty->print_cr("# after schedule_local"); 928 for (uint i = 0;i < block->number_of_nodes();i++) { 929 tty->print("# "); 930 block->get_node(i)->fast_dump(); 931 } 932 tty->cr(); 933 } 934 #endif 935 936 937 return true; 938 } 939 940 //--------------------------catch_cleanup_fix_all_inputs----------------------- 941 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 942 for (uint l = 0; l < use->len(); l++) { 943 if (use->in(l) == old_def) { 944 if (l < use->req()) { 945 use->set_req(l, new_def); 946 } else { 947 use->rm_prec(l); 948 use->add_prec(new_def); 949 l--; 950 } 951 } 952 } 953 } 954 955 //------------------------------catch_cleanup_find_cloned_def------------------ 956 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 957 assert( use_blk != def_blk, "Inter-block cleanup only"); 958 959 // The use is some block below the Catch. Find and return the clone of the def 960 // that dominates the use. If there is no clone in a dominating block, then 961 // create a phi for the def in a dominating block. 962 963 // Find which successor block dominates this use. The successor 964 // blocks must all be single-entry (from the Catch only; I will have 965 // split blocks to make this so), hence they all dominate. 966 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 967 use_blk = use_blk->_idom; 968 969 // Find the successor 970 Node *fixup = NULL; 971 972 uint j; 973 for( j = 0; j < def_blk->_num_succs; j++ ) 974 if( use_blk == def_blk->_succs[j] ) 975 break; 976 977 if( j == def_blk->_num_succs ) { 978 // Block at same level in dom-tree is not a successor. It needs a 979 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 980 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 981 for(uint k = 1; k < use_blk->num_preds(); k++) { 982 Block* block = get_block_for_node(use_blk->pred(k)); 983 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 984 } 985 986 // Check to see if the use_blk already has an identical phi inserted. 987 // If it exists, it will be at the first position since all uses of a 988 // def are processed together. 989 Node *phi = use_blk->get_node(1); 990 if( phi->is_Phi() ) { 991 fixup = phi; 992 for (uint k = 1; k < use_blk->num_preds(); k++) { 993 if (phi->in(k) != inputs[k]) { 994 // Not a match 995 fixup = NULL; 996 break; 997 } 998 } 999 } 1000 1001 // If an existing PhiNode was not found, make a new one. 1002 if (fixup == NULL) { 1003 Node *new_phi = PhiNode::make(use_blk->head(), def); 1004 use_blk->insert_node(new_phi, 1); 1005 map_node_to_block(new_phi, use_blk); 1006 for (uint k = 1; k < use_blk->num_preds(); k++) { 1007 new_phi->set_req(k, inputs[k]); 1008 } 1009 fixup = new_phi; 1010 } 1011 1012 } else { 1013 // Found the use just below the Catch. Make it use the clone. 1014 fixup = use_blk->get_node(n_clone_idx); 1015 } 1016 1017 return fixup; 1018 } 1019 1020 //--------------------------catch_cleanup_intra_block-------------------------- 1021 // Fix all input edges in use that reference "def". The use is in the same 1022 // block as the def and both have been cloned in each successor block. 1023 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1024 1025 // Both the use and def have been cloned. For each successor block, 1026 // get the clone of the use, and make its input the clone of the def 1027 // found in that block. 1028 1029 uint use_idx = blk->find_node(use); 1030 uint offset_idx = use_idx - beg; 1031 for( uint k = 0; k < blk->_num_succs; k++ ) { 1032 // Get clone in each successor block 1033 Block *sb = blk->_succs[k]; 1034 Node *clone = sb->get_node(offset_idx+1); 1035 assert( clone->Opcode() == use->Opcode(), "" ); 1036 1037 // Make use-clone reference the def-clone 1038 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1039 } 1040 } 1041 1042 //------------------------------catch_cleanup_inter_block--------------------- 1043 // Fix all input edges in use that reference "def". The use is in a different 1044 // block than the def. 1045 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1046 if( !use_blk ) return; // Can happen if the use is a precedence edge 1047 1048 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1049 catch_cleanup_fix_all_inputs(use, def, new_def); 1050 } 1051 1052 //------------------------------call_catch_cleanup----------------------------- 1053 // If we inserted any instructions between a Call and his CatchNode, 1054 // clone the instructions on all paths below the Catch. 1055 void PhaseCFG::call_catch_cleanup(Block* block) { 1056 1057 // End of region to clone 1058 uint end = block->end_idx(); 1059 if( !block->get_node(end)->is_Catch() ) return; 1060 // Start of region to clone 1061 uint beg = end; 1062 while(!block->get_node(beg-1)->is_MachProj() || 1063 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1064 beg--; 1065 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1066 } 1067 // Range of inserted instructions is [beg, end) 1068 if( beg == end ) return; 1069 1070 // Clone along all Catch output paths. Clone area between the 'beg' and 1071 // 'end' indices. 1072 for( uint i = 0; i < block->_num_succs; i++ ) { 1073 Block *sb = block->_succs[i]; 1074 // Clone the entire area; ignoring the edge fixup for now. 1075 for( uint j = end; j > beg; j-- ) { 1076 // It is safe here to clone a node with anti_dependence 1077 // since clones dominate on each path. 1078 Node *clone = block->get_node(j-1)->clone(); 1079 sb->insert_node(clone, 1); 1080 map_node_to_block(clone, sb); 1081 } 1082 } 1083 1084 1085 // Fixup edges. Check the def-use info per cloned Node 1086 for(uint i2 = beg; i2 < end; i2++ ) { 1087 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1088 Node *n = block->get_node(i2); // Node that got cloned 1089 // Need DU safe iterator because of edge manipulation in calls. 1090 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1091 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1092 out->push(n->fast_out(j1)); 1093 } 1094 uint max = out->size(); 1095 for (uint j = 0; j < max; j++) {// For all users 1096 Node *use = out->pop(); 1097 Block *buse = get_block_for_node(use); 1098 if( use->is_Phi() ) { 1099 for( uint k = 1; k < use->req(); k++ ) 1100 if( use->in(k) == n ) { 1101 Block* b = get_block_for_node(buse->pred(k)); 1102 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1103 use->set_req(k, fixup); 1104 } 1105 } else { 1106 if (block == buse) { 1107 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1108 } else { 1109 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1110 } 1111 } 1112 } // End for all users 1113 1114 } // End of for all Nodes in cloned area 1115 1116 // Remove the now-dead cloned ops 1117 for(uint i3 = beg; i3 < end; i3++ ) { 1118 block->get_node(beg)->disconnect_inputs(NULL, C); 1119 block->remove_node(beg); 1120 } 1121 1122 // If the successor blocks have a CreateEx node, move it back to the top 1123 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1124 Block *sb = block->_succs[i4]; 1125 uint new_cnt = end - beg; 1126 // Remove any newly created, but dead, nodes. 1127 for( uint j = new_cnt; j > 0; j-- ) { 1128 Node *n = sb->get_node(j); 1129 if (n->outcnt() == 0 && 1130 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1131 n->disconnect_inputs(NULL, C); 1132 sb->remove_node(j); 1133 new_cnt--; 1134 } 1135 } 1136 // If any newly created nodes remain, move the CreateEx node to the top 1137 if (new_cnt > 0) { 1138 Node *cex = sb->get_node(1+new_cnt); 1139 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1140 sb->remove_node(1+new_cnt); 1141 sb->insert_node(cex, 1); 1142 } 1143 } 1144 } 1145 }