1 /* 2 * Copyright (c) 1998, 2015, 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 #include "runtime/sharedRuntime.hpp" 35 36 // Optimization - Graph Style 37 38 // Check whether val is not-null-decoded compressed oop, 39 // i.e. will grab into the base of the heap if it represents NULL. 40 static bool accesses_heap_base_zone(Node *val) { 41 if (Universe::narrow_oop_base() > 0) { // Implies UseCompressedOops. 42 if (val && val->is_Mach()) { 43 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) { 44 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that 45 // decode NULL to point to the heap base (Decode_NN). 46 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) { 47 return true; 48 } 49 } 50 // Must recognize load operation with Decode matched in memory operand. 51 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected() 52 // returns true everywhere else. On PPC, no such memory operands 53 // exist, therefore we did not yet implement a check for such operands. 54 NOT_AIX(Unimplemented()); 55 } 56 } 57 return false; 58 } 59 60 static bool needs_explicit_null_check_for_read(Node *val) { 61 // On some OSes (AIX) the page at address 0 is only write protected. 62 // If so, only Store operations will trap. 63 if (os::zero_page_read_protected()) { 64 return false; // Implicit null check will work. 65 } 66 // Also a read accessing the base of a heap-based compressed heap will trap. 67 if (accesses_heap_base_zone(val) && // Hits the base zone page. 68 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected. 69 return false; 70 } 71 72 return true; 73 } 74 75 //------------------------------implicit_null_check---------------------------- 76 // Detect implicit-null-check opportunities. Basically, find NULL checks 77 // with suitable memory ops nearby. Use the memory op to do the NULL check. 78 // I can generate a memory op if there is not one nearby. 79 // The proj is the control projection for the not-null case. 80 // The val is the pointer being checked for nullness or 81 // decodeHeapOop_not_null node if it did not fold into address. 82 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { 83 // Assume if null check need for 0 offset then always needed 84 // Intel solaris doesn't support any null checks yet and no 85 // mechanism exists (yet) to set the switches at an os_cpu level 86 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 87 88 // Make sure the ptr-is-null path appears to be uncommon! 89 float f = block->end()->as_MachIf()->_prob; 90 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 91 if( f > PROB_UNLIKELY_MAG(4) ) return; 92 93 uint bidx = 0; // Capture index of value into memop 94 bool was_store; // Memory op is a store op 95 96 // Get the successor block for if the test ptr is non-null 97 Block* not_null_block; // this one goes with the proj 98 Block* null_block; 99 if (block->get_node(block->number_of_nodes()-1) == proj) { 100 null_block = block->_succs[0]; 101 not_null_block = block->_succs[1]; 102 } else { 103 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other"); 104 not_null_block = block->_succs[0]; 105 null_block = block->_succs[1]; 106 } 107 while (null_block->is_Empty() == Block::empty_with_goto) { 108 null_block = null_block->_succs[0]; 109 } 110 111 // Search the exception block for an uncommon trap. 112 // (See Parse::do_if and Parse::do_ifnull for the reason 113 // we need an uncommon trap. Briefly, we need a way to 114 // detect failure of this optimization, as in 6366351.) 115 { 116 bool found_trap = false; 117 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { 118 Node* nn = null_block->get_node(i1); 119 if (nn->is_MachCall() && 120 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 121 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 122 if (trtype->isa_int() && trtype->is_int()->is_con()) { 123 jint tr_con = trtype->is_int()->get_con(); 124 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 125 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 126 assert((int)reason < (int)BitsPerInt, "recode bit map"); 127 if (is_set_nth_bit(allowed_reasons, (int) reason) 128 && action != Deoptimization::Action_none) { 129 // This uncommon trap is sure to recompile, eventually. 130 // When that happens, C->too_many_traps will prevent 131 // this transformation from happening again. 132 found_trap = true; 133 } 134 } 135 break; 136 } 137 } 138 if (!found_trap) { 139 // We did not find an uncommon trap. 140 return; 141 } 142 } 143 144 // Check for decodeHeapOop_not_null node which did not fold into address 145 bool is_decoden = ((intptr_t)val) & 1; 146 val = (Node*)(((intptr_t)val) & ~1); 147 148 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 149 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 150 151 // Search the successor block for a load or store who's base value is also 152 // the tested value. There may be several. 153 Node_List *out = new Node_List(Thread::current()->resource_area()); 154 MachNode *best = NULL; // Best found so far 155 for (DUIterator i = val->outs(); val->has_out(i); i++) { 156 Node *m = val->out(i); 157 if( !m->is_Mach() ) continue; 158 MachNode *mach = m->as_Mach(); 159 was_store = false; 160 int iop = mach->ideal_Opcode(); 161 switch( iop ) { 162 case Op_LoadB: 163 case Op_LoadUB: 164 case Op_LoadUS: 165 case Op_LoadD: 166 case Op_LoadF: 167 case Op_LoadI: 168 case Op_LoadL: 169 case Op_LoadP: 170 case Op_LoadN: 171 case Op_LoadS: 172 case Op_LoadKlass: 173 case Op_LoadNKlass: 174 case Op_LoadRange: 175 case Op_LoadD_unaligned: 176 case Op_LoadL_unaligned: 177 case Op_ShenandoahReadBarrier: 178 case Op_ShenandoahWriteBarrier: 179 assert(mach->in(2) == val, "should be address"); 180 break; 181 case Op_StoreB: 182 case Op_StoreC: 183 case Op_StoreCM: 184 case Op_StoreD: 185 case Op_StoreF: 186 case Op_StoreI: 187 case Op_StoreL: 188 case Op_StoreP: 189 case Op_StoreN: 190 case Op_StoreNKlass: 191 was_store = true; // Memory op is a store op 192 // Stores will have their address in slot 2 (memory in slot 1). 193 // If the value being nul-checked is in another slot, it means we 194 // are storing the checked value, which does NOT check the value! 195 if( mach->in(2) != val ) continue; 196 break; // Found a memory op? 197 case Op_StrComp: 198 case Op_StrEquals: 199 case Op_StrIndexOf: 200 case Op_AryEq: 201 case Op_EncodeISOArray: 202 // Not a legit memory op for implicit null check regardless of 203 // embedded loads 204 continue; 205 default: // Also check for embedded loads 206 if( !mach->needs_anti_dependence_check() ) 207 continue; // Not an memory op; skip it 208 if( must_clone[iop] ) { 209 // Do not move nodes which produce flags because 210 // RA will try to clone it to place near branch and 211 // it will cause recompilation, see clone_node(). 212 continue; 213 } 214 { 215 // Check that value is used in memory address in 216 // instructions with embedded load (CmpP val1,(val2+off)). 217 Node* base; 218 Node* index; 219 const MachOper* oper = mach->memory_inputs(base, index); 220 if (oper == NULL || oper == (MachOper*)-1) { 221 continue; // Not an memory op; skip it 222 } 223 if (val == base || 224 val == index && val->bottom_type()->isa_narrowoop()) { 225 break; // Found it 226 } else { 227 continue; // Skip it 228 } 229 } 230 break; 231 } 232 233 // On some OSes (AIX) the page at address 0 is only write protected. 234 // If so, only Store operations will trap. 235 // But a read accessing the base of a heap-based compressed heap will trap. 236 if (!was_store && needs_explicit_null_check_for_read(val)) { 237 continue; 238 } 239 240 // check if the offset is not too high for implicit exception 241 { 242 intptr_t offset = 0; 243 const TypePtr *adr_type = NULL; // Do not need this return value here 244 const Node* base = mach->get_base_and_disp(offset, adr_type); 245 if (base == NULL || base == NodeSentinel) { 246 // Narrow oop address doesn't have base, only index 247 if( val->bottom_type()->isa_narrowoop() && 248 MacroAssembler::needs_explicit_null_check(offset) ) 249 continue; // Give up if offset is beyond page size 250 // cannot reason about it; is probably not implicit null exception 251 } else { 252 const TypePtr* tptr; 253 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 254 Universe::narrow_klass_shift() == 0)) { 255 // 32-bits narrow oop can be the base of address expressions 256 tptr = base->get_ptr_type(); 257 } else { 258 // only regular oops are expected here 259 tptr = base->bottom_type()->is_ptr(); 260 } 261 // Give up if offset is not a compile-time constant 262 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 263 continue; 264 offset += tptr->_offset; // correct if base is offseted 265 if( MacroAssembler::needs_explicit_null_check(offset) ) 266 continue; // Give up is reference is beyond 4K page size 267 } 268 } 269 270 // Check ctrl input to see if the null-check dominates the memory op 271 Block *cb = get_block_for_node(mach); 272 cb = cb->_idom; // Always hoist at least 1 block 273 if( !was_store ) { // Stores can be hoisted only one block 274 while( cb->_dom_depth > (block->_dom_depth + 1)) 275 cb = cb->_idom; // Hoist loads as far as we want 276 // The non-null-block should dominate the memory op, too. Live 277 // range spilling will insert a spill in the non-null-block if it is 278 // needs to spill the memory op for an implicit null check. 279 if (cb->_dom_depth == (block->_dom_depth + 1)) { 280 if (cb != not_null_block) continue; 281 cb = cb->_idom; 282 } 283 } 284 if( cb != block ) continue; 285 286 // Found a memory user; see if it can be hoisted to check-block 287 uint vidx = 0; // Capture index of value into memop 288 uint j; 289 for( j = mach->req()-1; j > 0; j-- ) { 290 if( mach->in(j) == val ) { 291 vidx = j; 292 // Ignore DecodeN val which could be hoisted to where needed. 293 if( is_decoden ) continue; 294 } 295 // Block of memory-op input 296 Block *inb = get_block_for_node(mach->in(j)); 297 Block *b = block; // Start from nul check 298 while( b != inb && b->_dom_depth > inb->_dom_depth ) 299 b = b->_idom; // search upwards for input 300 // See if input dominates null check 301 if( b != inb ) 302 break; 303 } 304 if( j > 0 ) 305 continue; 306 Block *mb = get_block_for_node(mach); 307 // Hoisting stores requires more checks for the anti-dependence case. 308 // Give up hoisting if we have to move the store past any load. 309 if( was_store ) { 310 Block *b = mb; // Start searching here for a local load 311 // mach use (faulting) trying to hoist 312 // n might be blocker to hoisting 313 while( b != block ) { 314 uint k; 315 for( k = 1; k < b->number_of_nodes(); k++ ) { 316 Node *n = b->get_node(k); 317 if( n->needs_anti_dependence_check() && 318 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 319 break; // Found anti-dependent load 320 } 321 if( k < b->number_of_nodes() ) 322 break; // Found anti-dependent load 323 // Make sure control does not do a merge (would have to check allpaths) 324 if( b->num_preds() != 2 ) break; 325 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 326 } 327 if( b != block ) continue; 328 } 329 330 // Make sure this memory op is not already being used for a NullCheck 331 Node *e = mb->end(); 332 if( e->is_MachNullCheck() && e->in(1) == mach ) 333 continue; // Already being used as a NULL check 334 335 // Found a candidate! Pick one with least dom depth - the highest 336 // in the dom tree should be closest to the null check. 337 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 338 best = mach; 339 bidx = vidx; 340 } 341 } 342 // No candidate! 343 if (best == NULL) { 344 return; 345 } 346 347 // ---- Found an implicit null check 348 extern int implicit_null_checks; 349 implicit_null_checks++; 350 351 if( is_decoden ) { 352 // Check if we need to hoist decodeHeapOop_not_null first. 353 Block *valb = get_block_for_node(val); 354 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 355 // Hoist it up to the end of the test block. 356 valb->find_remove(val); 357 block->add_inst(val); 358 map_node_to_block(val, block); 359 // DecodeN on x86 may kill flags. Check for flag-killing projections 360 // that also need to be hoisted. 361 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 362 Node* n = val->fast_out(j); 363 if( n->is_MachProj() ) { 364 get_block_for_node(n)->find_remove(n); 365 block->add_inst(n); 366 map_node_to_block(n, block); 367 } 368 } 369 } 370 } 371 // Hoist the memory candidate up to the end of the test block. 372 Block *old_block = get_block_for_node(best); 373 old_block->find_remove(best); 374 block->add_inst(best); 375 map_node_to_block(best, block); 376 377 // Move the control dependence 378 if (best->in(0) && best->in(0) == old_block->head()) 379 best->set_req(0, block->head()); 380 381 // Check for flag-killing projections that also need to be hoisted 382 // Should be DU safe because no edge updates. 383 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 384 Node* n = best->fast_out(j); 385 if( n->is_MachProj() || n->Opcode() == Op_ShenandoahWBMemProj) { 386 get_block_for_node(n)->find_remove(n); 387 block->add_inst(n); 388 map_node_to_block(n, block); 389 } 390 } 391 392 // proj==Op_True --> ne test; proj==Op_False --> eq test. 393 // One of two graph shapes got matched: 394 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 395 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 396 // NULL checks are always branch-if-eq. If we see a IfTrue projection 397 // then we are replacing a 'ne' test with a 'eq' NULL check test. 398 // We need to flip the projections to keep the same semantics. 399 if( proj->Opcode() == Op_IfTrue ) { 400 // Swap order of projections in basic block to swap branch targets 401 Node *tmp1 = block->get_node(block->end_idx()+1); 402 Node *tmp2 = block->get_node(block->end_idx()+2); 403 block->map_node(tmp2, block->end_idx()+1); 404 block->map_node(tmp1, block->end_idx()+2); 405 Node *tmp = new Node(C->top()); // Use not NULL input 406 tmp1->replace_by(tmp); 407 tmp2->replace_by(tmp1); 408 tmp->replace_by(tmp2); 409 tmp->destruct(); 410 } 411 412 // Remove the existing null check; use a new implicit null check instead. 413 // Since schedule-local needs precise def-use info, we need to correct 414 // it as well. 415 Node *old_tst = proj->in(0); 416 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx); 417 block->map_node(nul_chk, block->end_idx()); 418 map_node_to_block(nul_chk, block); 419 // Redirect users of old_test to nul_chk 420 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 421 old_tst->last_out(i2)->set_req(0, nul_chk); 422 // Clean-up any dead code 423 for (uint i3 = 0; i3 < old_tst->req(); i3++) { 424 Node* in = old_tst->in(i3); 425 old_tst->set_req(i3, NULL); 426 if (in->outcnt() == 0) { 427 // Remove dead input node 428 in->disconnect_inputs(NULL, C); 429 block->find_remove(in); 430 } 431 } 432 433 latency_from_uses(nul_chk); 434 latency_from_uses(best); 435 } 436 437 438 //------------------------------select----------------------------------------- 439 // Select a nice fellow from the worklist to schedule next. If there is only 440 // one choice, then use it. Projections take top priority for correctness 441 // reasons - if I see a projection, then it is next. There are a number of 442 // other special cases, for instructions that consume condition codes, et al. 443 // These are chosen immediately. Some instructions are required to immediately 444 // precede the last instruction in the block, and these are taken last. Of the 445 // remaining cases (most), choose the instruction with the greatest latency 446 // (that is, the most number of pseudo-cycles required to the end of the 447 // routine). If there is a tie, choose the instruction with the most inputs. 448 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 449 450 // If only a single entry on the stack, use it 451 uint cnt = worklist.size(); 452 if (cnt == 1) { 453 Node *n = worklist[0]; 454 worklist.map(0,worklist.pop()); 455 return n; 456 } 457 458 uint choice = 0; // Bigger is most important 459 uint latency = 0; // Bigger is scheduled first 460 uint score = 0; // Bigger is better 461 int idx = -1; // Index in worklist 462 int cand_cnt = 0; // Candidate count 463 464 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 465 // Order in worklist is used to break ties. 466 // See caller for how this is used to delay scheduling 467 // of induction variable increments to after the other 468 // uses of the phi are scheduled. 469 Node *n = worklist[i]; // Get Node on worklist 470 471 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 472 if( n->is_Proj() || // Projections always win 473 n->Opcode()== Op_Con || // So does constant 'Top' 474 iop == Op_CreateEx || // Create-exception must start block 475 iop == Op_CheckCastPP 476 ) { 477 worklist.map(i,worklist.pop()); 478 return n; 479 } 480 481 // Final call in a block must be adjacent to 'catch' 482 Node *e = block->end(); 483 if( e->is_Catch() && e->in(0)->in(0) == n ) 484 continue; 485 486 // Memory op for an implicit null check has to be at the end of the block 487 if( e->is_MachNullCheck() && e->in(1) == n ) 488 continue; 489 490 // Schedule IV increment last. 491 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 492 e->in(1)->in(1) == n && n->is_iteratively_computed()) 493 continue; 494 495 uint n_choice = 2; 496 497 // See if this instruction is consumed by a branch. If so, then (as the 498 // branch is the last instruction in the basic block) force it to the 499 // end of the basic block 500 if ( must_clone[iop] ) { 501 // See if any use is a branch 502 bool found_machif = false; 503 504 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 505 Node* use = n->fast_out(j); 506 507 // The use is a conditional branch, make them adjacent 508 if (use->is_MachIf() && get_block_for_node(use) == block) { 509 found_machif = true; 510 break; 511 } 512 513 // More than this instruction pending for successor to be ready, 514 // don't choose this if other opportunities are ready 515 if (ready_cnt.at(use->_idx) > 1) 516 n_choice = 1; 517 } 518 519 // loop terminated, prefer not to use this instruction 520 if (found_machif) 521 continue; 522 } 523 524 // See if this has a predecessor that is "must_clone", i.e. sets the 525 // condition code. If so, choose this first 526 for (uint j = 0; j < n->req() ; j++) { 527 Node *inn = n->in(j); 528 if (inn) { 529 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 530 n_choice = 3; 531 break; 532 } 533 } 534 } 535 536 // MachTemps should be scheduled last so they are near their uses 537 if (n->is_MachTemp()) { 538 n_choice = 1; 539 } 540 541 uint n_latency = get_latency_for_node(n); 542 uint n_score = n->req(); // Many inputs get high score to break ties 543 544 // Keep best latency found 545 cand_cnt++; 546 if (choice < n_choice || 547 (choice == n_choice && 548 ((StressLCM && Compile::randomized_select(cand_cnt)) || 549 (!StressLCM && 550 (latency < n_latency || 551 (latency == n_latency && 552 (score < n_score))))))) { 553 choice = n_choice; 554 latency = n_latency; 555 score = n_score; 556 idx = i; // Also keep index in worklist 557 } 558 } // End of for all ready nodes in worklist 559 560 assert(idx >= 0, "index should be set"); 561 Node *n = worklist[(uint)idx]; // Get the winner 562 563 worklist.map((uint)idx, worklist.pop()); // Compress worklist 564 return n; 565 } 566 567 568 //------------------------------set_next_call---------------------------------- 569 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 570 if( next_call.test_set(n->_idx) ) return; 571 for( uint i=0; i<n->len(); i++ ) { 572 Node *m = n->in(i); 573 if( !m ) continue; // must see all nodes in block that precede call 574 if (get_block_for_node(m) == block) { 575 set_next_call(block, m, next_call); 576 } 577 } 578 } 579 580 //------------------------------needed_for_next_call--------------------------- 581 // Set the flag 'next_call' for each Node that is needed for the next call to 582 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 583 // next subroutine call get priority - basically it moves things NOT needed 584 // for the next call till after the call. This prevents me from trying to 585 // carry lots of stuff live across a call. 586 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 587 // Find the next control-defining Node in this block 588 Node* call = NULL; 589 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 590 Node* m = this_call->fast_out(i); 591 if (get_block_for_node(m) == block && // Local-block user 592 m != this_call && // Not self-start node 593 m->is_MachCall()) { 594 call = m; 595 break; 596 } 597 } 598 if (call == NULL) return; // No next call (e.g., block end is near) 599 // Set next-call for all inputs to this call 600 set_next_call(block, call, next_call); 601 } 602 603 //------------------------------add_call_kills------------------------------------- 604 // helper function that adds caller save registers to MachProjNode 605 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 606 // Fill in the kill mask for the call 607 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 608 if( !regs.Member(r) ) { // Not already defined by the call 609 // Save-on-call register? 610 if ((save_policy[r] == 'C') || 611 (save_policy[r] == 'A') || 612 ((save_policy[r] == 'E') && exclude_soe)) { 613 proj->_rout.Insert(r); 614 } 615 } 616 } 617 } 618 619 620 //------------------------------sched_call------------------------------------- 621 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 622 RegMask regs; 623 624 // Schedule all the users of the call right now. All the users are 625 // projection Nodes, so they must be scheduled next to the call. 626 // Collect all the defined registers. 627 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 628 Node* n = mcall->fast_out(i); 629 assert( n->is_MachProj(), "" ); 630 int n_cnt = ready_cnt.at(n->_idx)-1; 631 ready_cnt.at_put(n->_idx, n_cnt); 632 assert( n_cnt == 0, "" ); 633 // Schedule next to call 634 block->map_node(n, node_cnt++); 635 // Collect defined registers 636 regs.OR(n->out_RegMask()); 637 // Check for scheduling the next control-definer 638 if( n->bottom_type() == Type::CONTROL ) 639 // Warm up next pile of heuristic bits 640 needed_for_next_call(block, n, next_call); 641 642 // Children of projections are now all ready 643 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 644 Node* m = n->fast_out(j); // Get user 645 if(get_block_for_node(m) != block) { 646 continue; 647 } 648 if( m->is_Phi() ) continue; 649 int m_cnt = ready_cnt.at(m->_idx)-1; 650 ready_cnt.at_put(m->_idx, m_cnt); 651 if( m_cnt == 0 ) 652 worklist.push(m); 653 } 654 655 } 656 657 // Act as if the call defines the Frame Pointer. 658 // Certainly the FP is alive and well after the call. 659 regs.Insert(_matcher.c_frame_pointer()); 660 661 // Set all registers killed and not already defined by the call. 662 uint r_cnt = mcall->tf()->range()->cnt(); 663 int op = mcall->ideal_Opcode(); 664 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 665 map_node_to_block(proj, block); 666 block->insert_node(proj, node_cnt++); 667 668 // Select the right register save policy. 669 const char * save_policy; 670 switch (op) { 671 case Op_CallRuntime: 672 case Op_CallLeaf: 673 case Op_CallLeafNoFP: 674 // Calling C code so use C calling convention 675 save_policy = _matcher._c_reg_save_policy; 676 break; 677 678 case Op_CallStaticJava: 679 case Op_CallDynamicJava: 680 // Calling Java code so use Java calling convention 681 save_policy = _matcher._register_save_policy; 682 break; 683 684 default: 685 ShouldNotReachHere(); 686 } 687 688 // When using CallRuntime mark SOE registers as killed by the call 689 // so values that could show up in the RegisterMap aren't live in a 690 // callee saved register since the register wouldn't know where to 691 // find them. CallLeaf and CallLeafNoFP are ok because they can't 692 // have debug info on them. Strictly speaking this only needs to be 693 // done for oops since idealreg2debugmask takes care of debug info 694 // references but there no way to handle oops differently than other 695 // pointers as far as the kill mask goes. 696 bool exclude_soe = op == Op_CallRuntime; 697 698 // If the call is a MethodHandle invoke, we need to exclude the 699 // register which is used to save the SP value over MH invokes from 700 // the mask. Otherwise this register could be used for 701 // deoptimization information. 702 if (op == Op_CallStaticJava) { 703 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 704 if (mcallstaticjava->_method_handle_invoke) 705 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 706 } 707 708 add_call_kills(proj, regs, save_policy, exclude_soe); 709 710 return node_cnt; 711 } 712 713 714 //------------------------------schedule_local--------------------------------- 715 // Topological sort within a block. Someday become a real scheduler. 716 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) { 717 // Already "sorted" are the block start Node (as the first entry), and 718 // the block-ending Node and any trailing control projections. We leave 719 // these alone. PhiNodes and ParmNodes are made to follow the block start 720 // Node. Everything else gets topo-sorted. 721 722 #ifndef PRODUCT 723 if (trace_opto_pipelining()) { 724 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 725 for (uint i = 0;i < block->number_of_nodes(); i++) { 726 tty->print("# "); 727 block->get_node(i)->fast_dump(); 728 } 729 tty->print_cr("#"); 730 } 731 #endif 732 733 // RootNode is already sorted 734 if (block->number_of_nodes() == 1) { 735 return true; 736 } 737 738 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 739 uint node_cnt = block->end_idx(); 740 uint phi_cnt = 1; 741 uint i; 742 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 743 Node *n = block->get_node(i); 744 if( n->is_Phi() || // Found a PhiNode or ParmNode 745 (n->is_Proj() && n->in(0) == block->head()) ) { 746 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 747 block->map_node(block->get_node(phi_cnt), i); 748 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 749 } else { // All others 750 // Count block-local inputs to 'n' 751 uint cnt = n->len(); // Input count 752 uint local = 0; 753 for( uint j=0; j<cnt; j++ ) { 754 Node *m = n->in(j); 755 if( m && get_block_for_node(m) == block && !m->is_top() ) 756 local++; // One more block-local input 757 } 758 ready_cnt.at_put(n->_idx, local); // Count em up 759 760 #ifdef ASSERT 761 if( UseConcMarkSweepGC || UseG1GC ) { 762 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 763 // Check the precedence edges 764 for (uint prec = n->req(); prec < n->len(); prec++) { 765 Node* oop_store = n->in(prec); 766 if (oop_store != NULL) { 767 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 768 } 769 } 770 } 771 } 772 #endif 773 774 // A few node types require changing a required edge to a precedence edge 775 // before allocation. 776 if( n->is_Mach() && n->req() > TypeFunc::Parms && 777 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 778 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 779 // MemBarAcquire could be created without Precedent edge. 780 // del_req() replaces the specified edge with the last input edge 781 // and then removes the last edge. If the specified edge > number of 782 // edges the last edge will be moved outside of the input edges array 783 // and the edge will be lost. This is why this code should be 784 // executed only when Precedent (== TypeFunc::Parms) edge is present. 785 Node *x = n->in(TypeFunc::Parms); 786 n->del_req(TypeFunc::Parms); 787 n->add_prec(x); 788 } 789 } 790 } 791 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 792 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 793 794 // All the prescheduled guys do not hold back internal nodes 795 uint i3; 796 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 797 Node *n = block->get_node(i3); // Get pre-scheduled 798 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 799 Node* m = n->fast_out(j); 800 if (get_block_for_node(m) == block) { // Local-block user 801 int m_cnt = ready_cnt.at(m->_idx)-1; 802 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 803 } 804 } 805 } 806 807 Node_List delay; 808 // Make a worklist 809 Node_List worklist; 810 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 811 Node *m = block->get_node(i4); 812 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 813 if (m->is_iteratively_computed()) { 814 // Push induction variable increments last to allow other uses 815 // of the phi to be scheduled first. The select() method breaks 816 // ties in scheduling by worklist order. 817 delay.push(m); 818 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 819 // Force the CreateEx to the top of the list so it's processed 820 // first and ends up at the start of the block. 821 worklist.insert(0, m); 822 } else { 823 worklist.push(m); // Then on to worklist! 824 } 825 } 826 } 827 while (delay.size()) { 828 Node* d = delay.pop(); 829 worklist.push(d); 830 } 831 832 // Warm up the 'next_call' heuristic bits 833 needed_for_next_call(block, block->head(), next_call); 834 835 #ifndef PRODUCT 836 if (trace_opto_pipelining()) { 837 for (uint j=0; j< block->number_of_nodes(); j++) { 838 Node *n = block->get_node(j); 839 int idx = n->_idx; 840 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 841 tty->print("latency:%3d ", get_latency_for_node(n)); 842 tty->print("%4d: %s\n", idx, n->Name()); 843 } 844 } 845 #endif 846 847 uint max_idx = (uint)ready_cnt.length(); 848 // Pull from worklist and schedule 849 while( worklist.size() ) { // Worklist is not ready 850 851 #ifndef PRODUCT 852 if (trace_opto_pipelining()) { 853 tty->print("# ready list:"); 854 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 855 Node *n = worklist[i]; // Get Node on worklist 856 tty->print(" %d", n->_idx); 857 } 858 tty->cr(); 859 } 860 #endif 861 862 // Select and pop a ready guy from worklist 863 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt); 864 block->map_node(n, phi_cnt++); // Schedule him next 865 866 #ifndef PRODUCT 867 if (trace_opto_pipelining()) { 868 tty->print("# select %d: %s", n->_idx, n->Name()); 869 tty->print(", latency:%d", get_latency_for_node(n)); 870 n->dump(); 871 if (Verbose) { 872 tty->print("# ready list:"); 873 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 874 Node *n = worklist[i]; // Get Node on worklist 875 tty->print(" %d", n->_idx); 876 } 877 tty->cr(); 878 } 879 } 880 881 #endif 882 if( n->is_MachCall() ) { 883 MachCallNode *mcall = n->as_MachCall(); 884 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 885 continue; 886 } 887 888 if (n->is_Mach() && n->as_Mach()->has_call()) { 889 RegMask regs; 890 regs.Insert(_matcher.c_frame_pointer()); 891 regs.OR(n->out_RegMask()); 892 893 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 894 map_node_to_block(proj, block); 895 block->insert_node(proj, phi_cnt++); 896 897 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 898 } 899 900 // Children are now all ready 901 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 902 Node* m = n->fast_out(i5); // Get user 903 if (get_block_for_node(m) != block) { 904 continue; 905 } 906 if( m->is_Phi() ) continue; 907 if (m->_idx >= max_idx) { // new node, skip it 908 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 909 continue; 910 } 911 int m_cnt = ready_cnt.at(m->_idx)-1; 912 ready_cnt.at_put(m->_idx, m_cnt); 913 if( m_cnt == 0 ) 914 worklist.push(m); 915 } 916 } 917 918 if( phi_cnt != block->end_idx() ) { 919 // did not schedule all. Retry, Bailout, or Die 920 if (C->subsume_loads() == true && !C->failing()) { 921 // Retry with subsume_loads == false 922 // If this is the first failure, the sentinel string will "stick" 923 // to the Compile object, and the C2Compiler will see it and retry. 924 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 925 } 926 // assert( phi_cnt == end_idx(), "did not schedule all" ); 927 return false; 928 } 929 930 #ifndef PRODUCT 931 if (trace_opto_pipelining()) { 932 tty->print_cr("#"); 933 tty->print_cr("# after schedule_local"); 934 for (uint i = 0;i < block->number_of_nodes();i++) { 935 tty->print("# "); 936 block->get_node(i)->fast_dump(); 937 } 938 tty->cr(); 939 } 940 #endif 941 942 943 return true; 944 } 945 946 //--------------------------catch_cleanup_fix_all_inputs----------------------- 947 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 948 for (uint l = 0; l < use->len(); l++) { 949 if (use->in(l) == old_def) { 950 if (l < use->req()) { 951 use->set_req(l, new_def); 952 } else { 953 use->rm_prec(l); 954 use->add_prec(new_def); 955 l--; 956 } 957 } 958 } 959 } 960 961 //------------------------------catch_cleanup_find_cloned_def------------------ 962 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 963 assert( use_blk != def_blk, "Inter-block cleanup only"); 964 965 // The use is some block below the Catch. Find and return the clone of the def 966 // that dominates the use. If there is no clone in a dominating block, then 967 // create a phi for the def in a dominating block. 968 969 // Find which successor block dominates this use. The successor 970 // blocks must all be single-entry (from the Catch only; I will have 971 // split blocks to make this so), hence they all dominate. 972 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 973 use_blk = use_blk->_idom; 974 975 // Find the successor 976 Node *fixup = NULL; 977 978 uint j; 979 for( j = 0; j < def_blk->_num_succs; j++ ) 980 if( use_blk == def_blk->_succs[j] ) 981 break; 982 983 if( j == def_blk->_num_succs ) { 984 // Block at same level in dom-tree is not a successor. It needs a 985 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 986 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 987 for(uint k = 1; k < use_blk->num_preds(); k++) { 988 Block* block = get_block_for_node(use_blk->pred(k)); 989 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 990 } 991 992 // Check to see if the use_blk already has an identical phi inserted. 993 // If it exists, it will be at the first position since all uses of a 994 // def are processed together. 995 Node *phi = use_blk->get_node(1); 996 if( phi->is_Phi() ) { 997 fixup = phi; 998 for (uint k = 1; k < use_blk->num_preds(); k++) { 999 if (phi->in(k) != inputs[k]) { 1000 // Not a match 1001 fixup = NULL; 1002 break; 1003 } 1004 } 1005 } 1006 1007 // If an existing PhiNode was not found, make a new one. 1008 if (fixup == NULL) { 1009 Node *new_phi = PhiNode::make(use_blk->head(), def); 1010 use_blk->insert_node(new_phi, 1); 1011 map_node_to_block(new_phi, use_blk); 1012 for (uint k = 1; k < use_blk->num_preds(); k++) { 1013 new_phi->set_req(k, inputs[k]); 1014 } 1015 fixup = new_phi; 1016 } 1017 1018 } else { 1019 // Found the use just below the Catch. Make it use the clone. 1020 fixup = use_blk->get_node(n_clone_idx); 1021 } 1022 1023 return fixup; 1024 } 1025 1026 //--------------------------catch_cleanup_intra_block-------------------------- 1027 // Fix all input edges in use that reference "def". The use is in the same 1028 // block as the def and both have been cloned in each successor block. 1029 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1030 1031 // Both the use and def have been cloned. For each successor block, 1032 // get the clone of the use, and make its input the clone of the def 1033 // found in that block. 1034 1035 uint use_idx = blk->find_node(use); 1036 uint offset_idx = use_idx - beg; 1037 for( uint k = 0; k < blk->_num_succs; k++ ) { 1038 // Get clone in each successor block 1039 Block *sb = blk->_succs[k]; 1040 Node *clone = sb->get_node(offset_idx+1); 1041 assert( clone->Opcode() == use->Opcode(), "" ); 1042 1043 // Make use-clone reference the def-clone 1044 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1045 } 1046 } 1047 1048 //------------------------------catch_cleanup_inter_block--------------------- 1049 // Fix all input edges in use that reference "def". The use is in a different 1050 // block than the def. 1051 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1052 if( !use_blk ) return; // Can happen if the use is a precedence edge 1053 1054 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1055 catch_cleanup_fix_all_inputs(use, def, new_def); 1056 } 1057 1058 //------------------------------call_catch_cleanup----------------------------- 1059 // If we inserted any instructions between a Call and his CatchNode, 1060 // clone the instructions on all paths below the Catch. 1061 void PhaseCFG::call_catch_cleanup(Block* block) { 1062 1063 // End of region to clone 1064 uint end = block->end_idx(); 1065 if( !block->get_node(end)->is_Catch() ) return; 1066 // Start of region to clone 1067 uint beg = end; 1068 while(!block->get_node(beg-1)->is_MachProj() || 1069 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1070 beg--; 1071 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1072 } 1073 // Range of inserted instructions is [beg, end) 1074 if( beg == end ) return; 1075 1076 // Clone along all Catch output paths. Clone area between the 'beg' and 1077 // 'end' indices. 1078 for( uint i = 0; i < block->_num_succs; i++ ) { 1079 Block *sb = block->_succs[i]; 1080 // Clone the entire area; ignoring the edge fixup for now. 1081 for( uint j = end; j > beg; j-- ) { 1082 // It is safe here to clone a node with anti_dependence 1083 // since clones dominate on each path. 1084 Node *clone = block->get_node(j-1)->clone(); 1085 sb->insert_node(clone, 1); 1086 map_node_to_block(clone, sb); 1087 } 1088 } 1089 1090 1091 // Fixup edges. Check the def-use info per cloned Node 1092 for(uint i2 = beg; i2 < end; i2++ ) { 1093 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1094 Node *n = block->get_node(i2); // Node that got cloned 1095 // Need DU safe iterator because of edge manipulation in calls. 1096 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1097 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1098 out->push(n->fast_out(j1)); 1099 } 1100 uint max = out->size(); 1101 for (uint j = 0; j < max; j++) {// For all users 1102 Node *use = out->pop(); 1103 Block *buse = get_block_for_node(use); 1104 if( use->is_Phi() ) { 1105 for( uint k = 1; k < use->req(); k++ ) 1106 if( use->in(k) == n ) { 1107 Block* b = get_block_for_node(buse->pred(k)); 1108 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1109 use->set_req(k, fixup); 1110 } 1111 } else { 1112 if (block == buse) { 1113 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1114 } else { 1115 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1116 } 1117 } 1118 } // End for all users 1119 1120 } // End of for all Nodes in cloned area 1121 1122 // Remove the now-dead cloned ops 1123 for(uint i3 = beg; i3 < end; i3++ ) { 1124 block->get_node(beg)->disconnect_inputs(NULL, C); 1125 block->remove_node(beg); 1126 } 1127 1128 // If the successor blocks have a CreateEx node, move it back to the top 1129 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1130 Block *sb = block->_succs[i4]; 1131 uint new_cnt = end - beg; 1132 // Remove any newly created, but dead, nodes. 1133 for( uint j = new_cnt; j > 0; j-- ) { 1134 Node *n = sb->get_node(j); 1135 if (n->outcnt() == 0 && 1136 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1137 n->disconnect_inputs(NULL, C); 1138 sb->remove_node(j); 1139 new_cnt--; 1140 } 1141 } 1142 // If any newly created nodes remain, move the CreateEx node to the top 1143 if (new_cnt > 0) { 1144 Node *cex = sb->get_node(1+new_cnt); 1145 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1146 sb->remove_node(1+new_cnt); 1147 sb->insert_node(cex, 1); 1148 } 1149 } 1150 } 1151 }