1 /* 2 * Copyright (c) 1998, 2016, 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 "opto/chaitin.hpp" 35 #include "runtime/sharedRuntime.hpp" 36 37 // Optimization - Graph Style 38 39 // Check whether val is not-null-decoded compressed oop, 40 // i.e. will grab into the base of the heap if it represents NULL. 41 static bool accesses_heap_base_zone(Node *val) { 42 if (Universe::narrow_oop_base() > 0) { // Implies UseCompressedOops. 43 if (val && val->is_Mach()) { 44 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) { 45 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that 46 // decode NULL to point to the heap base (Decode_NN). 47 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) { 48 return true; 49 } 50 } 51 // Must recognize load operation with Decode matched in memory operand. 52 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected() 53 // returns true everywhere else. On PPC, no such memory operands 54 // exist, therefore we did not yet implement a check for such operands. 55 NOT_AIX(Unimplemented()); 56 } 57 } 58 return false; 59 } 60 61 static bool needs_explicit_null_check_for_read(Node *val) { 62 // On some OSes (AIX) the page at address 0 is only write protected. 63 // If so, only Store operations will trap. 64 if (os::zero_page_read_protected()) { 65 return false; // Implicit null check will work. 66 } 67 // Also a read accessing the base of a heap-based compressed heap will trap. 68 if (accesses_heap_base_zone(val) && // Hits the base zone page. 69 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected. 70 return false; 71 } 72 73 return true; 74 } 75 76 //------------------------------implicit_null_check---------------------------- 77 // Detect implicit-null-check opportunities. Basically, find NULL checks 78 // with suitable memory ops nearby. Use the memory op to do the NULL check. 79 // I can generate a memory op if there is not one nearby. 80 // The proj is the control projection for the not-null case. 81 // The val is the pointer being checked for nullness or 82 // decodeHeapOop_not_null node if it did not fold into address. 83 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { 84 // Assume if null check need for 0 offset then always needed 85 // Intel solaris doesn't support any null checks yet and no 86 // mechanism exists (yet) to set the switches at an os_cpu level 87 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 88 89 // Make sure the ptr-is-null path appears to be uncommon! 90 float f = block->end()->as_MachIf()->_prob; 91 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 92 if( f > PROB_UNLIKELY_MAG(4) ) return; 93 94 uint bidx = 0; // Capture index of value into memop 95 bool was_store; // Memory op is a store op 96 97 // Get the successor block for if the test ptr is non-null 98 Block* not_null_block; // this one goes with the proj 99 Block* null_block; 100 if (block->get_node(block->number_of_nodes()-1) == proj) { 101 null_block = block->_succs[0]; 102 not_null_block = block->_succs[1]; 103 } else { 104 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other"); 105 not_null_block = block->_succs[0]; 106 null_block = block->_succs[1]; 107 } 108 while (null_block->is_Empty() == Block::empty_with_goto) { 109 null_block = null_block->_succs[0]; 110 } 111 112 // Search the exception block for an uncommon trap. 113 // (See Parse::do_if and Parse::do_ifnull for the reason 114 // we need an uncommon trap. Briefly, we need a way to 115 // detect failure of this optimization, as in 6366351.) 116 { 117 bool found_trap = false; 118 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { 119 Node* nn = null_block->get_node(i1); 120 if (nn->is_MachCall() && 121 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 122 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 123 if (trtype->isa_int() && trtype->is_int()->is_con()) { 124 jint tr_con = trtype->is_int()->get_con(); 125 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 126 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 127 assert((int)reason < (int)BitsPerInt, "recode bit map"); 128 if (is_set_nth_bit(allowed_reasons, (int) reason) 129 && action != Deoptimization::Action_none) { 130 // This uncommon trap is sure to recompile, eventually. 131 // When that happens, C->too_many_traps will prevent 132 // this transformation from happening again. 133 found_trap = true; 134 } 135 } 136 break; 137 } 138 } 139 if (!found_trap) { 140 // We did not find an uncommon trap. 141 return; 142 } 143 } 144 145 // Check for decodeHeapOop_not_null node which did not fold into address 146 bool is_decoden = ((intptr_t)val) & 1; 147 val = (Node*)(((intptr_t)val) & ~1); 148 149 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 150 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 151 152 // Search the successor block for a load or store who's base value is also 153 // the tested value. There may be several. 154 Node_List *out = new Node_List(Thread::current()->resource_area()); 155 MachNode *best = NULL; // Best found so far 156 for (DUIterator i = val->outs(); val->has_out(i); i++) { 157 Node *m = val->out(i); 158 if( !m->is_Mach() ) continue; 159 MachNode *mach = m->as_Mach(); 160 was_store = false; 161 int iop = mach->ideal_Opcode(); 162 switch( iop ) { 163 case Op_LoadB: 164 case Op_LoadUB: 165 case Op_LoadUS: 166 case Op_LoadD: 167 case Op_LoadF: 168 case Op_LoadI: 169 case Op_LoadL: 170 case Op_LoadP: 171 case Op_LoadN: 172 case Op_LoadS: 173 case Op_LoadKlass: 174 case Op_LoadNKlass: 175 case Op_LoadRange: 176 case Op_LoadD_unaligned: 177 case Op_LoadL_unaligned: 178 assert(mach->in(2) == val, "should be address"); 179 break; 180 case Op_StoreB: 181 case Op_StoreC: 182 case Op_StoreCM: 183 case Op_StoreD: 184 case Op_StoreF: 185 case Op_StoreI: 186 case Op_StoreL: 187 case Op_StoreP: 188 case Op_StoreN: 189 case Op_StoreNKlass: 190 was_store = true; // Memory op is a store op 191 // Stores will have their address in slot 2 (memory in slot 1). 192 // If the value being nul-checked is in another slot, it means we 193 // are storing the checked value, which does NOT check the value! 194 if( mach->in(2) != val ) continue; 195 break; // Found a memory op? 196 case Op_StrComp: 197 case Op_StrEquals: 198 case Op_StrIndexOf: 199 case Op_StrIndexOfChar: 200 case Op_AryEq: 201 case Op_StrInflatedCopy: 202 case Op_StrCompressedCopy: 203 case Op_EncodeISOArray: 204 case Op_HasNegatives: 205 // Not a legit memory op for implicit null check regardless of 206 // embedded loads 207 continue; 208 default: // Also check for embedded loads 209 if( !mach->needs_anti_dependence_check() ) 210 continue; // Not an memory op; skip it 211 if( must_clone[iop] ) { 212 // Do not move nodes which produce flags because 213 // RA will try to clone it to place near branch and 214 // it will cause recompilation, see clone_node(). 215 continue; 216 } 217 { 218 // Check that value is used in memory address in 219 // instructions with embedded load (CmpP val1,(val2+off)). 220 Node* base; 221 Node* index; 222 const MachOper* oper = mach->memory_inputs(base, index); 223 if (oper == NULL || oper == (MachOper*)-1) { 224 continue; // Not an memory op; skip it 225 } 226 if (val == base || 227 val == index && val->bottom_type()->isa_narrowoop()) { 228 break; // Found it 229 } else { 230 continue; // Skip it 231 } 232 } 233 break; 234 } 235 236 // On some OSes (AIX) the page at address 0 is only write protected. 237 // If so, only Store operations will trap. 238 // But a read accessing the base of a heap-based compressed heap will trap. 239 if (!was_store && needs_explicit_null_check_for_read(val)) { 240 continue; 241 } 242 243 // Check that node's control edge is not-null block's head or dominates it, 244 // otherwise we can't hoist it because there are other control dependencies. 245 Node* ctrl = mach->in(0); 246 if (ctrl != NULL && !(ctrl == not_null_block->head() || 247 get_block_for_node(ctrl)->dominates(not_null_block))) { 248 continue; 249 } 250 251 // check if the offset is not too high for implicit exception 252 { 253 intptr_t offset = 0; 254 const TypePtr *adr_type = NULL; // Do not need this return value here 255 const Node* base = mach->get_base_and_disp(offset, adr_type); 256 if (base == NULL || base == NodeSentinel) { 257 // Narrow oop address doesn't have base, only index 258 if( val->bottom_type()->isa_narrowoop() && 259 MacroAssembler::needs_explicit_null_check(offset) ) 260 continue; // Give up if offset is beyond page size 261 // cannot reason about it; is probably not implicit null exception 262 } else { 263 const TypePtr* tptr; 264 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 265 Universe::narrow_klass_shift() == 0)) { 266 // 32-bits narrow oop can be the base of address expressions 267 tptr = base->get_ptr_type(); 268 } else { 269 // only regular oops are expected here 270 tptr = base->bottom_type()->is_ptr(); 271 } 272 // Give up if offset is not a compile-time constant 273 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 274 continue; 275 offset += tptr->_offset; // correct if base is offseted 276 if( MacroAssembler::needs_explicit_null_check(offset) ) 277 continue; // Give up is reference is beyond 4K page size 278 } 279 } 280 281 // Check ctrl input to see if the null-check dominates the memory op 282 Block *cb = get_block_for_node(mach); 283 cb = cb->_idom; // Always hoist at least 1 block 284 if( !was_store ) { // Stores can be hoisted only one block 285 while( cb->_dom_depth > (block->_dom_depth + 1)) 286 cb = cb->_idom; // Hoist loads as far as we want 287 // The non-null-block should dominate the memory op, too. Live 288 // range spilling will insert a spill in the non-null-block if it is 289 // needs to spill the memory op for an implicit null check. 290 if (cb->_dom_depth == (block->_dom_depth + 1)) { 291 if (cb != not_null_block) continue; 292 cb = cb->_idom; 293 } 294 } 295 if( cb != block ) continue; 296 297 // Found a memory user; see if it can be hoisted to check-block 298 uint vidx = 0; // Capture index of value into memop 299 uint j; 300 for( j = mach->req()-1; j > 0; j-- ) { 301 if( mach->in(j) == val ) { 302 vidx = j; 303 // Ignore DecodeN val which could be hoisted to where needed. 304 if( is_decoden ) continue; 305 } 306 // Block of memory-op input 307 Block *inb = get_block_for_node(mach->in(j)); 308 Block *b = block; // Start from nul check 309 while( b != inb && b->_dom_depth > inb->_dom_depth ) 310 b = b->_idom; // search upwards for input 311 // See if input dominates null check 312 if( b != inb ) 313 break; 314 } 315 if( j > 0 ) 316 continue; 317 Block *mb = get_block_for_node(mach); 318 // Hoisting stores requires more checks for the anti-dependence case. 319 // Give up hoisting if we have to move the store past any load. 320 if( was_store ) { 321 Block *b = mb; // Start searching here for a local load 322 // mach use (faulting) trying to hoist 323 // n might be blocker to hoisting 324 while( b != block ) { 325 uint k; 326 for( k = 1; k < b->number_of_nodes(); k++ ) { 327 Node *n = b->get_node(k); 328 if( n->needs_anti_dependence_check() && 329 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 330 break; // Found anti-dependent load 331 } 332 if( k < b->number_of_nodes() ) 333 break; // Found anti-dependent load 334 // Make sure control does not do a merge (would have to check allpaths) 335 if( b->num_preds() != 2 ) break; 336 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 337 } 338 if( b != block ) continue; 339 } 340 341 // Make sure this memory op is not already being used for a NullCheck 342 Node *e = mb->end(); 343 if( e->is_MachNullCheck() && e->in(1) == mach ) 344 continue; // Already being used as a NULL check 345 346 // Found a candidate! Pick one with least dom depth - the highest 347 // in the dom tree should be closest to the null check. 348 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 349 best = mach; 350 bidx = vidx; 351 } 352 } 353 // No candidate! 354 if (best == NULL) { 355 return; 356 } 357 358 // ---- Found an implicit null check 359 #ifndef PRODUCT 360 extern int implicit_null_checks; 361 implicit_null_checks++; 362 #endif 363 364 if( is_decoden ) { 365 // Check if we need to hoist decodeHeapOop_not_null first. 366 Block *valb = get_block_for_node(val); 367 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 368 // Hoist it up to the end of the test block. 369 valb->find_remove(val); 370 block->add_inst(val); 371 map_node_to_block(val, block); 372 // DecodeN on x86 may kill flags. Check for flag-killing projections 373 // that also need to be hoisted. 374 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 375 Node* n = val->fast_out(j); 376 if( n->is_MachProj() ) { 377 get_block_for_node(n)->find_remove(n); 378 block->add_inst(n); 379 map_node_to_block(n, block); 380 } 381 } 382 } 383 } 384 // Hoist the memory candidate up to the end of the test block. 385 Block *old_block = get_block_for_node(best); 386 old_block->find_remove(best); 387 block->add_inst(best); 388 map_node_to_block(best, block); 389 390 // Move the control dependence if it is pinned to not-null block. 391 // Don't change it in other cases: NULL or dominating control. 392 if (best->in(0) == not_null_block->head()) { 393 // Set it to control edge of null check. 394 best->set_req(0, proj->in(0)->in(0)); 395 } 396 397 // Check for flag-killing projections that also need to be hoisted 398 // Should be DU safe because no edge updates. 399 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 400 Node* n = best->fast_out(j); 401 if( n->is_MachProj() ) { 402 get_block_for_node(n)->find_remove(n); 403 block->add_inst(n); 404 map_node_to_block(n, block); 405 } 406 } 407 408 // proj==Op_True --> ne test; proj==Op_False --> eq test. 409 // One of two graph shapes got matched: 410 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 411 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 412 // NULL checks are always branch-if-eq. If we see a IfTrue projection 413 // then we are replacing a 'ne' test with a 'eq' NULL check test. 414 // We need to flip the projections to keep the same semantics. 415 if( proj->Opcode() == Op_IfTrue ) { 416 // Swap order of projections in basic block to swap branch targets 417 Node *tmp1 = block->get_node(block->end_idx()+1); 418 Node *tmp2 = block->get_node(block->end_idx()+2); 419 block->map_node(tmp2, block->end_idx()+1); 420 block->map_node(tmp1, block->end_idx()+2); 421 Node *tmp = new Node(C->top()); // Use not NULL input 422 tmp1->replace_by(tmp); 423 tmp2->replace_by(tmp1); 424 tmp->replace_by(tmp2); 425 tmp->destruct(); 426 } 427 428 // Remove the existing null check; use a new implicit null check instead. 429 // Since schedule-local needs precise def-use info, we need to correct 430 // it as well. 431 Node *old_tst = proj->in(0); 432 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx); 433 block->map_node(nul_chk, block->end_idx()); 434 map_node_to_block(nul_chk, block); 435 // Redirect users of old_test to nul_chk 436 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 437 old_tst->last_out(i2)->set_req(0, nul_chk); 438 // Clean-up any dead code 439 for (uint i3 = 0; i3 < old_tst->req(); i3++) { 440 Node* in = old_tst->in(i3); 441 old_tst->set_req(i3, NULL); 442 if (in->outcnt() == 0) { 443 // Remove dead input node 444 in->disconnect_inputs(NULL, C); 445 block->find_remove(in); 446 } 447 } 448 449 latency_from_uses(nul_chk); 450 latency_from_uses(best); 451 452 // insert anti-dependences to defs in this block 453 if (! best->needs_anti_dependence_check()) { 454 for (uint k = 1; k < block->number_of_nodes(); k++) { 455 Node *n = block->get_node(k); 456 if (n->needs_anti_dependence_check() && 457 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) { 458 // Found anti-dependent load 459 insert_anti_dependences(block, n); 460 } 461 } 462 } 463 } 464 465 466 //------------------------------select----------------------------------------- 467 // Select a nice fellow from the worklist to schedule next. If there is only 468 // one choice, then use it. Projections take top priority for correctness 469 // reasons - if I see a projection, then it is next. There are a number of 470 // other special cases, for instructions that consume condition codes, et al. 471 // These are chosen immediately. Some instructions are required to immediately 472 // precede the last instruction in the block, and these are taken last. Of the 473 // remaining cases (most), choose the instruction with the greatest latency 474 // (that is, the most number of pseudo-cycles required to the end of the 475 // routine). If there is a tie, choose the instruction with the most inputs. 476 Node* PhaseCFG::select( 477 Block* block, 478 Node_List &worklist, 479 GrowableArray<int> &ready_cnt, 480 VectorSet &next_call, 481 uint sched_slot, 482 intptr_t* recalc_pressure_nodes) { 483 484 // If only a single entry on the stack, use it 485 uint cnt = worklist.size(); 486 if (cnt == 1) { 487 Node *n = worklist[0]; 488 worklist.map(0,worklist.pop()); 489 return n; 490 } 491 492 uint choice = 0; // Bigger is most important 493 uint latency = 0; // Bigger is scheduled first 494 uint score = 0; // Bigger is better 495 int idx = -1; // Index in worklist 496 int cand_cnt = 0; // Candidate count 497 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; 498 499 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 500 // Order in worklist is used to break ties. 501 // See caller for how this is used to delay scheduling 502 // of induction variable increments to after the other 503 // uses of the phi are scheduled. 504 Node *n = worklist[i]; // Get Node on worklist 505 506 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 507 if( n->is_Proj() || // Projections always win 508 n->Opcode()== Op_Con || // So does constant 'Top' 509 iop == Op_CreateEx || // Create-exception must start block 510 iop == Op_CheckCastPP 511 ) { 512 worklist.map(i,worklist.pop()); 513 return n; 514 } 515 516 // Final call in a block must be adjacent to 'catch' 517 Node *e = block->end(); 518 if( e->is_Catch() && e->in(0)->in(0) == n ) 519 continue; 520 521 // Memory op for an implicit null check has to be at the end of the block 522 if( e->is_MachNullCheck() && e->in(1) == n ) 523 continue; 524 525 // Schedule IV increment last. 526 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd) { 527 // Cmp might be matched into CountedLoopEnd node. 528 Node *cmp = (e->in(1)->ideal_reg() == Op_RegFlags) ? e->in(1) : e; 529 if (cmp->req() > 1 && cmp->in(1) == n && n->is_iteratively_computed()) { 530 continue; 531 } 532 } 533 534 uint n_choice = 2; 535 536 // See if this instruction is consumed by a branch. If so, then (as the 537 // branch is the last instruction in the basic block) force it to the 538 // end of the basic block 539 if ( must_clone[iop] ) { 540 // See if any use is a branch 541 bool found_machif = false; 542 543 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 544 Node* use = n->fast_out(j); 545 546 // The use is a conditional branch, make them adjacent 547 if (use->is_MachIf() && get_block_for_node(use) == block) { 548 found_machif = true; 549 break; 550 } 551 552 // More than this instruction pending for successor to be ready, 553 // don't choose this if other opportunities are ready 554 if (ready_cnt.at(use->_idx) > 1) 555 n_choice = 1; 556 } 557 558 // loop terminated, prefer not to use this instruction 559 if (found_machif) 560 continue; 561 } 562 563 // See if this has a predecessor that is "must_clone", i.e. sets the 564 // condition code. If so, choose this first 565 for (uint j = 0; j < n->req() ; j++) { 566 Node *inn = n->in(j); 567 if (inn) { 568 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 569 n_choice = 3; 570 break; 571 } 572 } 573 } 574 575 // MachTemps should be scheduled last so they are near their uses 576 if (n->is_MachTemp()) { 577 n_choice = 1; 578 } 579 580 uint n_latency = get_latency_for_node(n); 581 uint n_score = n->req(); // Many inputs get high score to break ties 582 583 if (OptoRegScheduling && block_size_threshold_ok) { 584 if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) { 585 _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit()); 586 _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit()); 587 // simulate the notion that we just picked this node to schedule 588 n->add_flag(Node::Flag_is_scheduled); 589 // now caculate its effect upon the graph if we did 590 adjust_register_pressure(n, block, recalc_pressure_nodes, false); 591 // return its state for finalize in case somebody else wins 592 n->remove_flag(Node::Flag_is_scheduled); 593 // now save the two final pressure components of register pressure, limiting pressure calcs to short size 594 short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure(); 595 short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure(); 596 recalc_pressure_nodes[n->_idx] = int_pressure; 597 recalc_pressure_nodes[n->_idx] |= (float_pressure << 16); 598 } 599 600 if (_scheduling_for_pressure) { 601 latency = n_latency; 602 if (n_choice != 3) { 603 // Now evaluate each register pressure component based on threshold in the score. 604 // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks 605 // on a single instruction, but we might see it shrink on both banks. 606 // For each use of register that has a register class that is over the high pressure limit, we build n_score up for 607 // live ranges that terminate on this instruction. 608 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 609 short int_pressure = (short)recalc_pressure_nodes[n->_idx]; 610 n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score; 611 } 612 if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 613 short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16); 614 n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score; 615 } 616 } else { 617 // make sure we choose these candidates 618 score = 0; 619 } 620 } 621 } 622 623 // Keep best latency found 624 cand_cnt++; 625 if (choice < n_choice || 626 (choice == n_choice && 627 ((StressLCM && Compile::randomized_select(cand_cnt)) || 628 (!StressLCM && 629 (latency < n_latency || 630 (latency == n_latency && 631 (score < n_score))))))) { 632 choice = n_choice; 633 latency = n_latency; 634 score = n_score; 635 idx = i; // Also keep index in worklist 636 } 637 } // End of for all ready nodes in worklist 638 639 assert(idx >= 0, "index should be set"); 640 Node *n = worklist[(uint)idx]; // Get the winner 641 642 worklist.map((uint)idx, worklist.pop()); // Compress worklist 643 return n; 644 } 645 646 //-------------------------adjust_register_pressure---------------------------- 647 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) { 648 PhaseLive* liveinfo = _regalloc->get_live(); 649 IndexSet* liveout = liveinfo->live(block); 650 // first adjust the register pressure for the sources 651 for (uint i = 1; i < n->req(); i++) { 652 bool lrg_ends = false; 653 Node *src_n = n->in(i); 654 if (src_n == NULL) continue; 655 if (!src_n->is_Mach()) continue; 656 uint src = _regalloc->_lrg_map.find(src_n); 657 if (src == 0) continue; 658 LRG& lrg_src = _regalloc->lrgs(src); 659 // detect if the live range ends or not 660 if (liveout->member(src) == false) { 661 lrg_ends = true; 662 for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) { 663 Node* m = src_n->fast_out(j); // Get user 664 if (m == n) continue; 665 if (!m->is_Mach()) continue; 666 MachNode *mach = m->as_Mach(); 667 bool src_matches = false; 668 int iop = mach->ideal_Opcode(); 669 670 switch (iop) { 671 case Op_StoreB: 672 case Op_StoreC: 673 case Op_StoreCM: 674 case Op_StoreD: 675 case Op_StoreF: 676 case Op_StoreI: 677 case Op_StoreL: 678 case Op_StoreP: 679 case Op_StoreN: 680 case Op_StoreVector: 681 case Op_StoreNKlass: 682 for (uint k = 1; k < m->req(); k++) { 683 Node *in = m->in(k); 684 if (in == src_n) { 685 src_matches = true; 686 break; 687 } 688 } 689 break; 690 691 default: 692 src_matches = true; 693 break; 694 } 695 696 // If we have a store as our use, ignore the non source operands 697 if (src_matches == false) continue; 698 699 // Mark every unscheduled use which is not n with a recalculation 700 if ((get_block_for_node(m) == block) && (!m->is_scheduled())) { 701 if (finalize_mode && !m->is_Phi()) { 702 recalc_pressure_nodes[m->_idx] = 0x7fff7fff; 703 } 704 lrg_ends = false; 705 } 706 } 707 } 708 // if none, this live range ends and we can adjust register pressure 709 if (lrg_ends) { 710 if (finalize_mode) { 711 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 712 } else { 713 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 714 } 715 } 716 } 717 718 // now add the register pressure from the dest and evaluate which heuristic we should use: 719 // 1.) The default, latency scheduling 720 // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks 721 uint dst = _regalloc->_lrg_map.find(n); 722 if (dst != 0) { 723 LRG& lrg_dst = _regalloc->lrgs(dst); 724 if (finalize_mode) { 725 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 726 // check to see if we fall over the register pressure cliff here 727 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 728 _scheduling_for_pressure = true; 729 } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 730 _scheduling_for_pressure = true; 731 } else { 732 // restore latency scheduling mode 733 _scheduling_for_pressure = false; 734 } 735 } else { 736 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 737 } 738 } 739 } 740 741 //------------------------------set_next_call---------------------------------- 742 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 743 if( next_call.test_set(n->_idx) ) return; 744 for( uint i=0; i<n->len(); i++ ) { 745 Node *m = n->in(i); 746 if( !m ) continue; // must see all nodes in block that precede call 747 if (get_block_for_node(m) == block) { 748 set_next_call(block, m, next_call); 749 } 750 } 751 } 752 753 //------------------------------needed_for_next_call--------------------------- 754 // Set the flag 'next_call' for each Node that is needed for the next call to 755 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 756 // next subroutine call get priority - basically it moves things NOT needed 757 // for the next call till after the call. This prevents me from trying to 758 // carry lots of stuff live across a call. 759 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 760 // Find the next control-defining Node in this block 761 Node* call = NULL; 762 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 763 Node* m = this_call->fast_out(i); 764 if (get_block_for_node(m) == block && // Local-block user 765 m != this_call && // Not self-start node 766 m->is_MachCall()) { 767 call = m; 768 break; 769 } 770 } 771 if (call == NULL) return; // No next call (e.g., block end is near) 772 // Set next-call for all inputs to this call 773 set_next_call(block, call, next_call); 774 } 775 776 //------------------------------add_call_kills------------------------------------- 777 // helper function that adds caller save registers to MachProjNode 778 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 779 // Fill in the kill mask for the call 780 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 781 if( !regs.Member(r) ) { // Not already defined by the call 782 // Save-on-call register? 783 if ((save_policy[r] == 'C') || 784 (save_policy[r] == 'A') || 785 ((save_policy[r] == 'E') && exclude_soe)) { 786 proj->_rout.Insert(r); 787 } 788 } 789 } 790 } 791 792 793 //------------------------------sched_call------------------------------------- 794 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 795 RegMask regs; 796 797 // Schedule all the users of the call right now. All the users are 798 // projection Nodes, so they must be scheduled next to the call. 799 // Collect all the defined registers. 800 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 801 Node* n = mcall->fast_out(i); 802 assert( n->is_MachProj(), "" ); 803 int n_cnt = ready_cnt.at(n->_idx)-1; 804 ready_cnt.at_put(n->_idx, n_cnt); 805 assert( n_cnt == 0, "" ); 806 // Schedule next to call 807 block->map_node(n, node_cnt++); 808 // Collect defined registers 809 regs.OR(n->out_RegMask()); 810 // Check for scheduling the next control-definer 811 if( n->bottom_type() == Type::CONTROL ) 812 // Warm up next pile of heuristic bits 813 needed_for_next_call(block, n, next_call); 814 815 // Children of projections are now all ready 816 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 817 Node* m = n->fast_out(j); // Get user 818 if(get_block_for_node(m) != block) { 819 continue; 820 } 821 if( m->is_Phi() ) continue; 822 int m_cnt = ready_cnt.at(m->_idx) - 1; 823 ready_cnt.at_put(m->_idx, m_cnt); 824 if( m_cnt == 0 ) 825 worklist.push(m); 826 } 827 828 } 829 830 // Act as if the call defines the Frame Pointer. 831 // Certainly the FP is alive and well after the call. 832 regs.Insert(_matcher.c_frame_pointer()); 833 834 // Set all registers killed and not already defined by the call. 835 uint r_cnt = mcall->tf()->range()->cnt(); 836 int op = mcall->ideal_Opcode(); 837 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 838 map_node_to_block(proj, block); 839 block->insert_node(proj, node_cnt++); 840 841 // Select the right register save policy. 842 const char *save_policy = NULL; 843 switch (op) { 844 case Op_CallRuntime: 845 case Op_CallLeaf: 846 case Op_CallLeafNoFP: 847 // Calling C code so use C calling convention 848 save_policy = _matcher._c_reg_save_policy; 849 break; 850 851 case Op_CallStaticJava: 852 case Op_CallDynamicJava: 853 // Calling Java code so use Java calling convention 854 save_policy = _matcher._register_save_policy; 855 break; 856 857 default: 858 ShouldNotReachHere(); 859 } 860 861 // When using CallRuntime mark SOE registers as killed by the call 862 // so values that could show up in the RegisterMap aren't live in a 863 // callee saved register since the register wouldn't know where to 864 // find them. CallLeaf and CallLeafNoFP are ok because they can't 865 // have debug info on them. Strictly speaking this only needs to be 866 // done for oops since idealreg2debugmask takes care of debug info 867 // references but there no way to handle oops differently than other 868 // pointers as far as the kill mask goes. 869 bool exclude_soe = op == Op_CallRuntime; 870 871 // If the call is a MethodHandle invoke, we need to exclude the 872 // register which is used to save the SP value over MH invokes from 873 // the mask. Otherwise this register could be used for 874 // deoptimization information. 875 if (op == Op_CallStaticJava) { 876 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 877 if (mcallstaticjava->_method_handle_invoke) 878 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 879 } 880 881 add_call_kills(proj, regs, save_policy, exclude_soe); 882 883 return node_cnt; 884 } 885 886 887 //------------------------------schedule_local--------------------------------- 888 // Topological sort within a block. Someday become a real scheduler. 889 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) { 890 // Already "sorted" are the block start Node (as the first entry), and 891 // the block-ending Node and any trailing control projections. We leave 892 // these alone. PhiNodes and ParmNodes are made to follow the block start 893 // Node. Everything else gets topo-sorted. 894 895 #ifndef PRODUCT 896 if (trace_opto_pipelining()) { 897 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 898 for (uint i = 0;i < block->number_of_nodes(); i++) { 899 tty->print("# "); 900 block->get_node(i)->fast_dump(); 901 } 902 tty->print_cr("#"); 903 } 904 #endif 905 906 // RootNode is already sorted 907 if (block->number_of_nodes() == 1) { 908 return true; 909 } 910 911 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; 912 913 // We track the uses of local definitions as input dependences so that 914 // we know when a given instruction is avialable to be scheduled. 915 uint i; 916 if (OptoRegScheduling && block_size_threshold_ok) { 917 for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc 918 Node *n = block->get_node(i); 919 n->remove_flag(Node::Flag_is_scheduled); 920 if (!n->is_Phi()) { 921 recalc_pressure_nodes[n->_idx] = 0x7fff7fff; 922 } 923 } 924 } 925 926 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 927 uint node_cnt = block->end_idx(); 928 uint phi_cnt = 1; 929 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 930 Node *n = block->get_node(i); 931 if( n->is_Phi() || // Found a PhiNode or ParmNode 932 (n->is_Proj() && n->in(0) == block->head()) ) { 933 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 934 block->map_node(block->get_node(phi_cnt), i); 935 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 936 if (OptoRegScheduling && block_size_threshold_ok) { 937 // mark n as scheduled 938 n->add_flag(Node::Flag_is_scheduled); 939 } 940 } else { // All others 941 // Count block-local inputs to 'n' 942 uint cnt = n->len(); // Input count 943 uint local = 0; 944 for( uint j=0; j<cnt; j++ ) { 945 Node *m = n->in(j); 946 if( m && get_block_for_node(m) == block && !m->is_top() ) 947 local++; // One more block-local input 948 } 949 ready_cnt.at_put(n->_idx, local); // Count em up 950 951 #ifdef ASSERT 952 if( UseConcMarkSweepGC || UseG1GC ) { 953 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 954 // Check the precedence edges 955 for (uint prec = n->req(); prec < n->len(); prec++) { 956 Node* oop_store = n->in(prec); 957 if (oop_store != NULL) { 958 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 959 } 960 } 961 } 962 } 963 #endif 964 965 // A few node types require changing a required edge to a precedence edge 966 // before allocation. 967 if( n->is_Mach() && n->req() > TypeFunc::Parms && 968 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 969 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 970 // MemBarAcquire could be created without Precedent edge. 971 // del_req() replaces the specified edge with the last input edge 972 // and then removes the last edge. If the specified edge > number of 973 // edges the last edge will be moved outside of the input edges array 974 // and the edge will be lost. This is why this code should be 975 // executed only when Precedent (== TypeFunc::Parms) edge is present. 976 Node *x = n->in(TypeFunc::Parms); 977 if (x != NULL && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) { 978 // Old edge to node within same block will get removed, but no precedence 979 // edge will get added because it already exists. Update ready count. 980 int cnt = ready_cnt.at(n->_idx); 981 assert(cnt > 1, "MemBar node %d must not get ready here", n->_idx); 982 ready_cnt.at_put(n->_idx, cnt-1); 983 } 984 n->del_req(TypeFunc::Parms); 985 n->add_prec(x); 986 } 987 } 988 } 989 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 990 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 991 992 // All the prescheduled guys do not hold back internal nodes 993 uint i3; 994 for (i3 = 0; i3 < phi_cnt; i3++) { // For all pre-scheduled 995 Node *n = block->get_node(i3); // Get pre-scheduled 996 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 997 Node* m = n->fast_out(j); 998 if (get_block_for_node(m) == block) { // Local-block user 999 int m_cnt = ready_cnt.at(m->_idx)-1; 1000 if (OptoRegScheduling && block_size_threshold_ok) { 1001 // mark m as scheduled 1002 if (m_cnt < 0) { 1003 m->add_flag(Node::Flag_is_scheduled); 1004 } 1005 } 1006 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 1007 } 1008 } 1009 } 1010 1011 Node_List delay; 1012 // Make a worklist 1013 Node_List worklist; 1014 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 1015 Node *m = block->get_node(i4); 1016 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 1017 if (m->is_iteratively_computed()) { 1018 // Push induction variable increments last to allow other uses 1019 // of the phi to be scheduled first. The select() method breaks 1020 // ties in scheduling by worklist order. 1021 delay.push(m); 1022 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 1023 // Force the CreateEx to the top of the list so it's processed 1024 // first and ends up at the start of the block. 1025 worklist.insert(0, m); 1026 } else { 1027 worklist.push(m); // Then on to worklist! 1028 } 1029 } 1030 } 1031 while (delay.size()) { 1032 Node* d = delay.pop(); 1033 worklist.push(d); 1034 } 1035 1036 if (OptoRegScheduling && block_size_threshold_ok) { 1037 // To stage register pressure calculations we need to examine the live set variables 1038 // breaking them up by register class to compartmentalize the calculations. 1039 uint float_pressure = Matcher::float_pressure(FLOATPRESSURE); 1040 _regalloc->_sched_int_pressure.init(INTPRESSURE); 1041 _regalloc->_sched_float_pressure.init(float_pressure); 1042 _regalloc->_scratch_int_pressure.init(INTPRESSURE); 1043 _regalloc->_scratch_float_pressure.init(float_pressure); 1044 1045 _regalloc->compute_entry_block_pressure(block); 1046 } 1047 1048 // Warm up the 'next_call' heuristic bits 1049 needed_for_next_call(block, block->head(), next_call); 1050 1051 #ifndef PRODUCT 1052 if (trace_opto_pipelining()) { 1053 for (uint j=0; j< block->number_of_nodes(); j++) { 1054 Node *n = block->get_node(j); 1055 int idx = n->_idx; 1056 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 1057 tty->print("latency:%3d ", get_latency_for_node(n)); 1058 tty->print("%4d: %s\n", idx, n->Name()); 1059 } 1060 } 1061 #endif 1062 1063 uint max_idx = (uint)ready_cnt.length(); 1064 // Pull from worklist and schedule 1065 while( worklist.size() ) { // Worklist is not ready 1066 1067 #ifndef PRODUCT 1068 if (trace_opto_pipelining()) { 1069 tty->print("# ready list:"); 1070 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1071 Node *n = worklist[i]; // Get Node on worklist 1072 tty->print(" %d", n->_idx); 1073 } 1074 tty->cr(); 1075 } 1076 #endif 1077 1078 // Select and pop a ready guy from worklist 1079 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes); 1080 block->map_node(n, phi_cnt++); // Schedule him next 1081 1082 if (OptoRegScheduling && block_size_threshold_ok) { 1083 n->add_flag(Node::Flag_is_scheduled); 1084 1085 // Now adjust the resister pressure with the node we selected 1086 if (!n->is_Phi()) { 1087 adjust_register_pressure(n, block, recalc_pressure_nodes, true); 1088 } 1089 } 1090 1091 #ifndef PRODUCT 1092 if (trace_opto_pipelining()) { 1093 tty->print("# select %d: %s", n->_idx, n->Name()); 1094 tty->print(", latency:%d", get_latency_for_node(n)); 1095 n->dump(); 1096 if (Verbose) { 1097 tty->print("# ready list:"); 1098 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1099 Node *n = worklist[i]; // Get Node on worklist 1100 tty->print(" %d", n->_idx); 1101 } 1102 tty->cr(); 1103 } 1104 } 1105 1106 #endif 1107 if( n->is_MachCall() ) { 1108 MachCallNode *mcall = n->as_MachCall(); 1109 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 1110 continue; 1111 } 1112 1113 if (n->is_Mach() && n->as_Mach()->has_call()) { 1114 RegMask regs; 1115 regs.Insert(_matcher.c_frame_pointer()); 1116 regs.OR(n->out_RegMask()); 1117 1118 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 1119 map_node_to_block(proj, block); 1120 block->insert_node(proj, phi_cnt++); 1121 1122 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 1123 } 1124 1125 // Children are now all ready 1126 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 1127 Node* m = n->fast_out(i5); // Get user 1128 if (get_block_for_node(m) != block) { 1129 continue; 1130 } 1131 if( m->is_Phi() ) continue; 1132 if (m->_idx >= max_idx) { // new node, skip it 1133 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 1134 continue; 1135 } 1136 int m_cnt = ready_cnt.at(m->_idx) - 1; 1137 ready_cnt.at_put(m->_idx, m_cnt); 1138 if( m_cnt == 0 ) 1139 worklist.push(m); 1140 } 1141 } 1142 1143 if( phi_cnt != block->end_idx() ) { 1144 // did not schedule all. Retry, Bailout, or Die 1145 if (C->subsume_loads() == true && !C->failing()) { 1146 // Retry with subsume_loads == false 1147 // If this is the first failure, the sentinel string will "stick" 1148 // to the Compile object, and the C2Compiler will see it and retry. 1149 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 1150 } else { 1151 assert(false, "graph should be schedulable"); 1152 } 1153 // assert( phi_cnt == end_idx(), "did not schedule all" ); 1154 return false; 1155 } 1156 1157 if (OptoRegScheduling && block_size_threshold_ok) { 1158 _regalloc->compute_exit_block_pressure(block); 1159 block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure(); 1160 block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure(); 1161 } 1162 1163 #ifndef PRODUCT 1164 if (trace_opto_pipelining()) { 1165 tty->print_cr("#"); 1166 tty->print_cr("# after schedule_local"); 1167 for (uint i = 0;i < block->number_of_nodes();i++) { 1168 tty->print("# "); 1169 block->get_node(i)->fast_dump(); 1170 } 1171 tty->print_cr("# "); 1172 1173 if (OptoRegScheduling && block_size_threshold_ok) { 1174 tty->print_cr("# pressure info : %d", block->_pre_order); 1175 _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info"); 1176 _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info"); 1177 } 1178 tty->cr(); 1179 } 1180 #endif 1181 1182 return true; 1183 } 1184 1185 //--------------------------catch_cleanup_fix_all_inputs----------------------- 1186 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 1187 for (uint l = 0; l < use->len(); l++) { 1188 if (use->in(l) == old_def) { 1189 if (l < use->req()) { 1190 use->set_req(l, new_def); 1191 } else { 1192 use->rm_prec(l); 1193 use->add_prec(new_def); 1194 l--; 1195 } 1196 } 1197 } 1198 } 1199 1200 //------------------------------catch_cleanup_find_cloned_def------------------ 1201 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1202 assert( use_blk != def_blk, "Inter-block cleanup only"); 1203 1204 // The use is some block below the Catch. Find and return the clone of the def 1205 // that dominates the use. If there is no clone in a dominating block, then 1206 // create a phi for the def in a dominating block. 1207 1208 // Find which successor block dominates this use. The successor 1209 // blocks must all be single-entry (from the Catch only; I will have 1210 // split blocks to make this so), hence they all dominate. 1211 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 1212 use_blk = use_blk->_idom; 1213 1214 // Find the successor 1215 Node *fixup = NULL; 1216 1217 uint j; 1218 for( j = 0; j < def_blk->_num_succs; j++ ) 1219 if( use_blk == def_blk->_succs[j] ) 1220 break; 1221 1222 if( j == def_blk->_num_succs ) { 1223 // Block at same level in dom-tree is not a successor. It needs a 1224 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 1225 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 1226 for(uint k = 1; k < use_blk->num_preds(); k++) { 1227 Block* block = get_block_for_node(use_blk->pred(k)); 1228 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 1229 } 1230 1231 // Check to see if the use_blk already has an identical phi inserted. 1232 // If it exists, it will be at the first position since all uses of a 1233 // def are processed together. 1234 Node *phi = use_blk->get_node(1); 1235 if( phi->is_Phi() ) { 1236 fixup = phi; 1237 for (uint k = 1; k < use_blk->num_preds(); k++) { 1238 if (phi->in(k) != inputs[k]) { 1239 // Not a match 1240 fixup = NULL; 1241 break; 1242 } 1243 } 1244 } 1245 1246 // If an existing PhiNode was not found, make a new one. 1247 if (fixup == NULL) { 1248 Node *new_phi = PhiNode::make(use_blk->head(), def); 1249 use_blk->insert_node(new_phi, 1); 1250 map_node_to_block(new_phi, use_blk); 1251 for (uint k = 1; k < use_blk->num_preds(); k++) { 1252 new_phi->set_req(k, inputs[k]); 1253 } 1254 fixup = new_phi; 1255 } 1256 1257 } else { 1258 // Found the use just below the Catch. Make it use the clone. 1259 fixup = use_blk->get_node(n_clone_idx); 1260 } 1261 1262 return fixup; 1263 } 1264 1265 //--------------------------catch_cleanup_intra_block-------------------------- 1266 // Fix all input edges in use that reference "def". The use is in the same 1267 // block as the def and both have been cloned in each successor block. 1268 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1269 1270 // Both the use and def have been cloned. For each successor block, 1271 // get the clone of the use, and make its input the clone of the def 1272 // found in that block. 1273 1274 uint use_idx = blk->find_node(use); 1275 uint offset_idx = use_idx - beg; 1276 for( uint k = 0; k < blk->_num_succs; k++ ) { 1277 // Get clone in each successor block 1278 Block *sb = blk->_succs[k]; 1279 Node *clone = sb->get_node(offset_idx+1); 1280 assert( clone->Opcode() == use->Opcode(), "" ); 1281 1282 // Make use-clone reference the def-clone 1283 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1284 } 1285 } 1286 1287 //------------------------------catch_cleanup_inter_block--------------------- 1288 // Fix all input edges in use that reference "def". The use is in a different 1289 // block than the def. 1290 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1291 if( !use_blk ) return; // Can happen if the use is a precedence edge 1292 1293 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1294 catch_cleanup_fix_all_inputs(use, def, new_def); 1295 } 1296 1297 //------------------------------call_catch_cleanup----------------------------- 1298 // If we inserted any instructions between a Call and his CatchNode, 1299 // clone the instructions on all paths below the Catch. 1300 void PhaseCFG::call_catch_cleanup(Block* block) { 1301 1302 // End of region to clone 1303 uint end = block->end_idx(); 1304 if( !block->get_node(end)->is_Catch() ) return; 1305 // Start of region to clone 1306 uint beg = end; 1307 while(!block->get_node(beg-1)->is_MachProj() || 1308 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1309 beg--; 1310 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1311 } 1312 // Range of inserted instructions is [beg, end) 1313 if( beg == end ) return; 1314 1315 // Clone along all Catch output paths. Clone area between the 'beg' and 1316 // 'end' indices. 1317 for( uint i = 0; i < block->_num_succs; i++ ) { 1318 Block *sb = block->_succs[i]; 1319 // Clone the entire area; ignoring the edge fixup for now. 1320 for( uint j = end; j > beg; j-- ) { 1321 Node *clone = block->get_node(j-1)->clone(); 1322 sb->insert_node(clone, 1); 1323 map_node_to_block(clone, sb); 1324 if (clone->needs_anti_dependence_check()) { 1325 insert_anti_dependences(sb, clone); 1326 } 1327 } 1328 } 1329 1330 1331 // Fixup edges. Check the def-use info per cloned Node 1332 for(uint i2 = beg; i2 < end; i2++ ) { 1333 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1334 Node *n = block->get_node(i2); // Node that got cloned 1335 // Need DU safe iterator because of edge manipulation in calls. 1336 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1337 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1338 out->push(n->fast_out(j1)); 1339 } 1340 uint max = out->size(); 1341 for (uint j = 0; j < max; j++) {// For all users 1342 Node *use = out->pop(); 1343 Block *buse = get_block_for_node(use); 1344 if( use->is_Phi() ) { 1345 for( uint k = 1; k < use->req(); k++ ) 1346 if( use->in(k) == n ) { 1347 Block* b = get_block_for_node(buse->pred(k)); 1348 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1349 use->set_req(k, fixup); 1350 } 1351 } else { 1352 if (block == buse) { 1353 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1354 } else { 1355 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1356 } 1357 } 1358 } // End for all users 1359 1360 } // End of for all Nodes in cloned area 1361 1362 // Remove the now-dead cloned ops 1363 for(uint i3 = beg; i3 < end; i3++ ) { 1364 block->get_node(beg)->disconnect_inputs(NULL, C); 1365 block->remove_node(beg); 1366 } 1367 1368 // If the successor blocks have a CreateEx node, move it back to the top 1369 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1370 Block *sb = block->_succs[i4]; 1371 uint new_cnt = end - beg; 1372 // Remove any newly created, but dead, nodes. 1373 for( uint j = new_cnt; j > 0; j-- ) { 1374 Node *n = sb->get_node(j); 1375 if (n->outcnt() == 0 && 1376 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1377 n->disconnect_inputs(NULL, C); 1378 sb->remove_node(j); 1379 new_cnt--; 1380 } 1381 } 1382 // If any newly created nodes remain, move the CreateEx node to the top 1383 if (new_cnt > 0) { 1384 Node *cex = sb->get_node(1+new_cnt); 1385 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1386 sb->remove_node(1+new_cnt); 1387 sb->insert_node(cex, 1); 1388 } 1389 } 1390 } 1391 }