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 Node* in = old_tst->in(i3); 422 old_tst->set_req(i3, NULL); 423 if (in->outcnt() == 0) { 424 // Remove dead input node 425 in->disconnect_inputs(NULL, C); 426 block->find_remove(in); 427 } 428 } 429 430 latency_from_uses(nul_chk); 431 latency_from_uses(best); 432 } 433 434 435 //------------------------------select----------------------------------------- 436 // Select a nice fellow from the worklist to schedule next. If there is only 437 // one choice, then use it. Projections take top priority for correctness 438 // reasons - if I see a projection, then it is next. There are a number of 439 // other special cases, for instructions that consume condition codes, et al. 440 // These are chosen immediately. Some instructions are required to immediately 441 // precede the last instruction in the block, and these are taken last. Of the 442 // remaining cases (most), choose the instruction with the greatest latency 443 // (that is, the most number of pseudo-cycles required to the end of the 444 // routine). If there is a tie, choose the instruction with the most inputs. 445 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 446 447 // If only a single entry on the stack, use it 448 uint cnt = worklist.size(); 449 if (cnt == 1) { 450 Node *n = worklist[0]; 451 worklist.map(0,worklist.pop()); 452 return n; 453 } 454 455 uint choice = 0; // Bigger is most important 456 uint latency = 0; // Bigger is scheduled first 457 uint score = 0; // Bigger is better 458 int idx = -1; // Index in worklist 459 int cand_cnt = 0; // Candidate count 460 461 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 462 // Order in worklist is used to break ties. 463 // See caller for how this is used to delay scheduling 464 // of induction variable increments to after the other 465 // uses of the phi are scheduled. 466 Node *n = worklist[i]; // Get Node on worklist 467 468 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 469 if( n->is_Proj() || // Projections always win 470 n->Opcode()== Op_Con || // So does constant 'Top' 471 iop == Op_CreateEx || // Create-exception must start block 472 iop == Op_CheckCastPP 473 ) { 474 worklist.map(i,worklist.pop()); 475 return n; 476 } 477 478 // Final call in a block must be adjacent to 'catch' 479 Node *e = block->end(); 480 if( e->is_Catch() && e->in(0)->in(0) == n ) 481 continue; 482 483 // Memory op for an implicit null check has to be at the end of the block 484 if( e->is_MachNullCheck() && e->in(1) == n ) 485 continue; 486 487 // Schedule IV increment last. 488 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 489 e->in(1)->in(1) == n && n->is_iteratively_computed()) 490 continue; 491 492 uint n_choice = 2; 493 494 // See if this instruction is consumed by a branch. If so, then (as the 495 // branch is the last instruction in the basic block) force it to the 496 // end of the basic block 497 if ( must_clone[iop] ) { 498 // See if any use is a branch 499 bool found_machif = false; 500 501 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 502 Node* use = n->fast_out(j); 503 504 // The use is a conditional branch, make them adjacent 505 if (use->is_MachIf() && get_block_for_node(use) == block) { 506 found_machif = true; 507 break; 508 } 509 510 // More than this instruction pending for successor to be ready, 511 // don't choose this if other opportunities are ready 512 if (ready_cnt.at(use->_idx) > 1) 513 n_choice = 1; 514 } 515 516 // loop terminated, prefer not to use this instruction 517 if (found_machif) 518 continue; 519 } 520 521 // See if this has a predecessor that is "must_clone", i.e. sets the 522 // condition code. If so, choose this first 523 for (uint j = 0; j < n->req() ; j++) { 524 Node *inn = n->in(j); 525 if (inn) { 526 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 527 n_choice = 3; 528 break; 529 } 530 } 531 } 532 533 // MachTemps should be scheduled last so they are near their uses 534 if (n->is_MachTemp()) { 535 n_choice = 1; 536 } 537 538 uint n_latency = get_latency_for_node(n); 539 uint n_score = n->req(); // Many inputs get high score to break ties 540 541 // Keep best latency found 542 cand_cnt++; 543 if (choice < n_choice || 544 (choice == n_choice && 545 ((StressLCM && Compile::randomized_select(cand_cnt)) || 546 (!StressLCM && 547 (latency < n_latency || 548 (latency == n_latency && 549 (score < n_score))))))) { 550 choice = n_choice; 551 latency = n_latency; 552 score = n_score; 553 idx = i; // Also keep index in worklist 554 } 555 } // End of for all ready nodes in worklist 556 557 assert(idx >= 0, "index should be set"); 558 Node *n = worklist[(uint)idx]; // Get the winner 559 560 worklist.map((uint)idx, worklist.pop()); // Compress worklist 561 return n; 562 } 563 564 565 //------------------------------set_next_call---------------------------------- 566 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 567 if( next_call.test_set(n->_idx) ) return; 568 for( uint i=0; i<n->len(); i++ ) { 569 Node *m = n->in(i); 570 if( !m ) continue; // must see all nodes in block that precede call 571 if (get_block_for_node(m) == block) { 572 set_next_call(block, m, next_call); 573 } 574 } 575 } 576 577 //------------------------------needed_for_next_call--------------------------- 578 // Set the flag 'next_call' for each Node that is needed for the next call to 579 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 580 // next subroutine call get priority - basically it moves things NOT needed 581 // for the next call till after the call. This prevents me from trying to 582 // carry lots of stuff live across a call. 583 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 584 // Find the next control-defining Node in this block 585 Node* call = NULL; 586 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 587 Node* m = this_call->fast_out(i); 588 if (get_block_for_node(m) == block && // Local-block user 589 m != this_call && // Not self-start node 590 m->is_MachCall()) { 591 call = m; 592 break; 593 } 594 } 595 if (call == NULL) return; // No next call (e.g., block end is near) 596 // Set next-call for all inputs to this call 597 set_next_call(block, call, next_call); 598 } 599 600 //------------------------------add_call_kills------------------------------------- 601 // helper function that adds caller save registers to MachProjNode 602 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 603 // Fill in the kill mask for the call 604 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 605 if( !regs.Member(r) ) { // Not already defined by the call 606 // Save-on-call register? 607 if ((save_policy[r] == 'C') || 608 (save_policy[r] == 'A') || 609 ((save_policy[r] == 'E') && exclude_soe)) { 610 proj->_rout.Insert(r); 611 } 612 } 613 } 614 } 615 616 617 //------------------------------sched_call------------------------------------- 618 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 619 RegMask regs; 620 621 // Schedule all the users of the call right now. All the users are 622 // projection Nodes, so they must be scheduled next to the call. 623 // Collect all the defined registers. 624 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 625 Node* n = mcall->fast_out(i); 626 assert( n->is_MachProj(), "" ); 627 int n_cnt = ready_cnt.at(n->_idx)-1; 628 ready_cnt.at_put(n->_idx, n_cnt); 629 assert( n_cnt == 0, "" ); 630 // Schedule next to call 631 block->map_node(n, node_cnt++); 632 // Collect defined registers 633 regs.OR(n->out_RegMask()); 634 // Check for scheduling the next control-definer 635 if( n->bottom_type() == Type::CONTROL ) 636 // Warm up next pile of heuristic bits 637 needed_for_next_call(block, n, next_call); 638 639 // Children of projections are now all ready 640 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 641 Node* m = n->fast_out(j); // Get user 642 if(get_block_for_node(m) != block) { 643 continue; 644 } 645 if( m->is_Phi() ) continue; 646 int m_cnt = ready_cnt.at(m->_idx)-1; 647 ready_cnt.at_put(m->_idx, m_cnt); 648 if( m_cnt == 0 ) 649 worklist.push(m); 650 } 651 652 } 653 654 // Act as if the call defines the Frame Pointer. 655 // Certainly the FP is alive and well after the call. 656 regs.Insert(_matcher.c_frame_pointer()); 657 658 // Set all registers killed and not already defined by the call. 659 uint r_cnt = mcall->tf()->range()->cnt(); 660 int op = mcall->ideal_Opcode(); 661 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 662 map_node_to_block(proj, block); 663 block->insert_node(proj, node_cnt++); 664 665 // Select the right register save policy. 666 const char * save_policy; 667 switch (op) { 668 case Op_CallRuntime: 669 case Op_CallLeaf: 670 case Op_CallLeafNoFP: 671 // Calling C code so use C calling convention 672 save_policy = _matcher._c_reg_save_policy; 673 break; 674 675 case Op_CallStaticJava: 676 case Op_CallDynamicJava: 677 // Calling Java code so use Java calling convention 678 save_policy = _matcher._register_save_policy; 679 break; 680 681 default: 682 ShouldNotReachHere(); 683 } 684 685 // When using CallRuntime mark SOE registers as killed by the call 686 // so values that could show up in the RegisterMap aren't live in a 687 // callee saved register since the register wouldn't know where to 688 // find them. CallLeaf and CallLeafNoFP are ok because they can't 689 // have debug info on them. Strictly speaking this only needs to be 690 // done for oops since idealreg2debugmask takes care of debug info 691 // references but there no way to handle oops differently than other 692 // pointers as far as the kill mask goes. 693 bool exclude_soe = op == Op_CallRuntime; 694 695 // If the call is a MethodHandle invoke, we need to exclude the 696 // register which is used to save the SP value over MH invokes from 697 // the mask. Otherwise this register could be used for 698 // deoptimization information. 699 if (op == Op_CallStaticJava) { 700 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 701 if (mcallstaticjava->_method_handle_invoke) 702 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 703 } 704 705 add_call_kills(proj, regs, save_policy, exclude_soe); 706 707 return node_cnt; 708 } 709 710 711 //------------------------------schedule_local--------------------------------- 712 // Topological sort within a block. Someday become a real scheduler. 713 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) { 714 // Already "sorted" are the block start Node (as the first entry), and 715 // the block-ending Node and any trailing control projections. We leave 716 // these alone. PhiNodes and ParmNodes are made to follow the block start 717 // Node. Everything else gets topo-sorted. 718 719 #ifndef PRODUCT 720 if (trace_opto_pipelining()) { 721 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 722 for (uint i = 0;i < block->number_of_nodes(); i++) { 723 tty->print("# "); 724 block->get_node(i)->fast_dump(); 725 } 726 tty->print_cr("#"); 727 } 728 #endif 729 730 // RootNode is already sorted 731 if (block->number_of_nodes() == 1) { 732 return true; 733 } 734 735 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 736 uint node_cnt = block->end_idx(); 737 uint phi_cnt = 1; 738 uint i; 739 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 740 Node *n = block->get_node(i); 741 if( n->is_Phi() || // Found a PhiNode or ParmNode 742 (n->is_Proj() && n->in(0) == block->head()) ) { 743 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 744 block->map_node(block->get_node(phi_cnt), i); 745 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 746 } else { // All others 747 // Count block-local inputs to 'n' 748 uint cnt = n->len(); // Input count 749 uint local = 0; 750 for( uint j=0; j<cnt; j++ ) { 751 Node *m = n->in(j); 752 if( m && get_block_for_node(m) == block && !m->is_top() ) 753 local++; // One more block-local input 754 } 755 ready_cnt.at_put(n->_idx, local); // Count em up 756 757 #ifdef ASSERT 758 if( UseConcMarkSweepGC || UseG1GC ) { 759 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 760 // Check the precedence edges 761 for (uint prec = n->req(); prec < n->len(); prec++) { 762 Node* oop_store = n->in(prec); 763 if (oop_store != NULL) { 764 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 765 } 766 } 767 } 768 } 769 #endif 770 771 // A few node types require changing a required edge to a precedence edge 772 // before allocation. 773 if( n->is_Mach() && n->req() > TypeFunc::Parms && 774 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 775 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 776 // MemBarAcquire could be created without Precedent edge. 777 // del_req() replaces the specified edge with the last input edge 778 // and then removes the last edge. If the specified edge > number of 779 // edges the last edge will be moved outside of the input edges array 780 // and the edge will be lost. This is why this code should be 781 // executed only when Precedent (== TypeFunc::Parms) edge is present. 782 Node *x = n->in(TypeFunc::Parms); 783 n->del_req(TypeFunc::Parms); 784 n->add_prec(x); 785 } 786 } 787 } 788 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 789 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 790 791 // All the prescheduled guys do not hold back internal nodes 792 uint i3; 793 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 794 Node *n = block->get_node(i3); // Get pre-scheduled 795 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 796 Node* m = n->fast_out(j); 797 if (get_block_for_node(m) == block) { // Local-block user 798 int m_cnt = ready_cnt.at(m->_idx)-1; 799 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 800 } 801 } 802 } 803 804 Node_List delay; 805 // Make a worklist 806 Node_List worklist; 807 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 808 Node *m = block->get_node(i4); 809 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 810 if (m->is_iteratively_computed()) { 811 // Push induction variable increments last to allow other uses 812 // of the phi to be scheduled first. The select() method breaks 813 // ties in scheduling by worklist order. 814 delay.push(m); 815 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 816 // Force the CreateEx to the top of the list so it's processed 817 // first and ends up at the start of the block. 818 worklist.insert(0, m); 819 } else { 820 worklist.push(m); // Then on to worklist! 821 } 822 } 823 } 824 while (delay.size()) { 825 Node* d = delay.pop(); 826 worklist.push(d); 827 } 828 829 // Warm up the 'next_call' heuristic bits 830 needed_for_next_call(block, block->head(), next_call); 831 832 #ifndef PRODUCT 833 if (trace_opto_pipelining()) { 834 for (uint j=0; j< block->number_of_nodes(); j++) { 835 Node *n = block->get_node(j); 836 int idx = n->_idx; 837 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 838 tty->print("latency:%3d ", get_latency_for_node(n)); 839 tty->print("%4d: %s\n", idx, n->Name()); 840 } 841 } 842 #endif 843 844 uint max_idx = (uint)ready_cnt.length(); 845 // Pull from worklist and schedule 846 while( worklist.size() ) { // Worklist is not ready 847 848 #ifndef PRODUCT 849 if (trace_opto_pipelining()) { 850 tty->print("# ready list:"); 851 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 852 Node *n = worklist[i]; // Get Node on worklist 853 tty->print(" %d", n->_idx); 854 } 855 tty->cr(); 856 } 857 #endif 858 859 // Select and pop a ready guy from worklist 860 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt); 861 block->map_node(n, phi_cnt++); // Schedule him next 862 863 #ifndef PRODUCT 864 if (trace_opto_pipelining()) { 865 tty->print("# select %d: %s", n->_idx, n->Name()); 866 tty->print(", latency:%d", get_latency_for_node(n)); 867 n->dump(); 868 if (Verbose) { 869 tty->print("# ready list:"); 870 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 871 Node *n = worklist[i]; // Get Node on worklist 872 tty->print(" %d", n->_idx); 873 } 874 tty->cr(); 875 } 876 } 877 878 #endif 879 if( n->is_MachCall() ) { 880 MachCallNode *mcall = n->as_MachCall(); 881 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 882 continue; 883 } 884 885 if (n->is_Mach() && n->as_Mach()->has_call()) { 886 RegMask regs; 887 regs.Insert(_matcher.c_frame_pointer()); 888 regs.OR(n->out_RegMask()); 889 890 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 891 map_node_to_block(proj, block); 892 block->insert_node(proj, phi_cnt++); 893 894 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 895 } 896 897 // Children are now all ready 898 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 899 Node* m = n->fast_out(i5); // Get user 900 if (get_block_for_node(m) != block) { 901 continue; 902 } 903 if( m->is_Phi() ) continue; 904 if (m->_idx >= max_idx) { // new node, skip it 905 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 906 continue; 907 } 908 int m_cnt = ready_cnt.at(m->_idx)-1; 909 ready_cnt.at_put(m->_idx, m_cnt); 910 if( m_cnt == 0 ) 911 worklist.push(m); 912 } 913 } 914 915 if( phi_cnt != block->end_idx() ) { 916 // did not schedule all. Retry, Bailout, or Die 917 if (C->subsume_loads() == true && !C->failing()) { 918 // Retry with subsume_loads == false 919 // If this is the first failure, the sentinel string will "stick" 920 // to the Compile object, and the C2Compiler will see it and retry. 921 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 922 } 923 // assert( phi_cnt == end_idx(), "did not schedule all" ); 924 return false; 925 } 926 927 #ifndef PRODUCT 928 if (trace_opto_pipelining()) { 929 tty->print_cr("#"); 930 tty->print_cr("# after schedule_local"); 931 for (uint i = 0;i < block->number_of_nodes();i++) { 932 tty->print("# "); 933 block->get_node(i)->fast_dump(); 934 } 935 tty->cr(); 936 } 937 #endif 938 939 940 return true; 941 } 942 943 //--------------------------catch_cleanup_fix_all_inputs----------------------- 944 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 945 for (uint l = 0; l < use->len(); l++) { 946 if (use->in(l) == old_def) { 947 if (l < use->req()) { 948 use->set_req(l, new_def); 949 } else { 950 use->rm_prec(l); 951 use->add_prec(new_def); 952 l--; 953 } 954 } 955 } 956 } 957 958 //------------------------------catch_cleanup_find_cloned_def------------------ 959 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 960 assert( use_blk != def_blk, "Inter-block cleanup only"); 961 962 // The use is some block below the Catch. Find and return the clone of the def 963 // that dominates the use. If there is no clone in a dominating block, then 964 // create a phi for the def in a dominating block. 965 966 // Find which successor block dominates this use. The successor 967 // blocks must all be single-entry (from the Catch only; I will have 968 // split blocks to make this so), hence they all dominate. 969 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 970 use_blk = use_blk->_idom; 971 972 // Find the successor 973 Node *fixup = NULL; 974 975 uint j; 976 for( j = 0; j < def_blk->_num_succs; j++ ) 977 if( use_blk == def_blk->_succs[j] ) 978 break; 979 980 if( j == def_blk->_num_succs ) { 981 // Block at same level in dom-tree is not a successor. It needs a 982 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 983 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 984 for(uint k = 1; k < use_blk->num_preds(); k++) { 985 Block* block = get_block_for_node(use_blk->pred(k)); 986 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 987 } 988 989 // Check to see if the use_blk already has an identical phi inserted. 990 // If it exists, it will be at the first position since all uses of a 991 // def are processed together. 992 Node *phi = use_blk->get_node(1); 993 if( phi->is_Phi() ) { 994 fixup = phi; 995 for (uint k = 1; k < use_blk->num_preds(); k++) { 996 if (phi->in(k) != inputs[k]) { 997 // Not a match 998 fixup = NULL; 999 break; 1000 } 1001 } 1002 } 1003 1004 // If an existing PhiNode was not found, make a new one. 1005 if (fixup == NULL) { 1006 Node *new_phi = PhiNode::make(use_blk->head(), def); 1007 use_blk->insert_node(new_phi, 1); 1008 map_node_to_block(new_phi, use_blk); 1009 for (uint k = 1; k < use_blk->num_preds(); k++) { 1010 new_phi->set_req(k, inputs[k]); 1011 } 1012 fixup = new_phi; 1013 } 1014 1015 } else { 1016 // Found the use just below the Catch. Make it use the clone. 1017 fixup = use_blk->get_node(n_clone_idx); 1018 } 1019 1020 return fixup; 1021 } 1022 1023 //--------------------------catch_cleanup_intra_block-------------------------- 1024 // Fix all input edges in use that reference "def". The use is in the same 1025 // block as the def and both have been cloned in each successor block. 1026 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1027 1028 // Both the use and def have been cloned. For each successor block, 1029 // get the clone of the use, and make its input the clone of the def 1030 // found in that block. 1031 1032 uint use_idx = blk->find_node(use); 1033 uint offset_idx = use_idx - beg; 1034 for( uint k = 0; k < blk->_num_succs; k++ ) { 1035 // Get clone in each successor block 1036 Block *sb = blk->_succs[k]; 1037 Node *clone = sb->get_node(offset_idx+1); 1038 assert( clone->Opcode() == use->Opcode(), "" ); 1039 1040 // Make use-clone reference the def-clone 1041 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1042 } 1043 } 1044 1045 //------------------------------catch_cleanup_inter_block--------------------- 1046 // Fix all input edges in use that reference "def". The use is in a different 1047 // block than the def. 1048 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1049 if( !use_blk ) return; // Can happen if the use is a precedence edge 1050 1051 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1052 catch_cleanup_fix_all_inputs(use, def, new_def); 1053 } 1054 1055 //------------------------------call_catch_cleanup----------------------------- 1056 // If we inserted any instructions between a Call and his CatchNode, 1057 // clone the instructions on all paths below the Catch. 1058 void PhaseCFG::call_catch_cleanup(Block* block) { 1059 1060 // End of region to clone 1061 uint end = block->end_idx(); 1062 if( !block->get_node(end)->is_Catch() ) return; 1063 // Start of region to clone 1064 uint beg = end; 1065 while(!block->get_node(beg-1)->is_MachProj() || 1066 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1067 beg--; 1068 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1069 } 1070 // Range of inserted instructions is [beg, end) 1071 if( beg == end ) return; 1072 1073 // Clone along all Catch output paths. Clone area between the 'beg' and 1074 // 'end' indices. 1075 for( uint i = 0; i < block->_num_succs; i++ ) { 1076 Block *sb = block->_succs[i]; 1077 // Clone the entire area; ignoring the edge fixup for now. 1078 for( uint j = end; j > beg; j-- ) { 1079 // It is safe here to clone a node with anti_dependence 1080 // since clones dominate on each path. 1081 Node *clone = block->get_node(j-1)->clone(); 1082 sb->insert_node(clone, 1); 1083 map_node_to_block(clone, sb); 1084 } 1085 } 1086 1087 1088 // Fixup edges. Check the def-use info per cloned Node 1089 for(uint i2 = beg; i2 < end; i2++ ) { 1090 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1091 Node *n = block->get_node(i2); // Node that got cloned 1092 // Need DU safe iterator because of edge manipulation in calls. 1093 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1094 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1095 out->push(n->fast_out(j1)); 1096 } 1097 uint max = out->size(); 1098 for (uint j = 0; j < max; j++) {// For all users 1099 Node *use = out->pop(); 1100 Block *buse = get_block_for_node(use); 1101 if( use->is_Phi() ) { 1102 for( uint k = 1; k < use->req(); k++ ) 1103 if( use->in(k) == n ) { 1104 Block* b = get_block_for_node(buse->pred(k)); 1105 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1106 use->set_req(k, fixup); 1107 } 1108 } else { 1109 if (block == buse) { 1110 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1111 } else { 1112 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1113 } 1114 } 1115 } // End for all users 1116 1117 } // End of for all Nodes in cloned area 1118 1119 // Remove the now-dead cloned ops 1120 for(uint i3 = beg; i3 < end; i3++ ) { 1121 block->get_node(beg)->disconnect_inputs(NULL, C); 1122 block->remove_node(beg); 1123 } 1124 1125 // If the successor blocks have a CreateEx node, move it back to the top 1126 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1127 Block *sb = block->_succs[i4]; 1128 uint new_cnt = end - beg; 1129 // Remove any newly created, but dead, nodes. 1130 for( uint j = new_cnt; j > 0; j-- ) { 1131 Node *n = sb->get_node(j); 1132 if (n->outcnt() == 0 && 1133 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1134 n->disconnect_inputs(NULL, C); 1135 sb->remove_node(j); 1136 new_cnt--; 1137 } 1138 } 1139 // If any newly created nodes remain, move the CreateEx node to the top 1140 if (new_cnt > 0) { 1141 Node *cex = sb->get_node(1+new_cnt); 1142 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1143 sb->remove_node(1+new_cnt); 1144 sb->insert_node(cex, 1); 1145 } 1146 } 1147 } 1148 }