1 /* 2 * Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/ad.hpp" 28 #include "opto/block.hpp" 29 #include "opto/c2compiler.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/machnode.hpp" 33 #include "opto/runtime.hpp" 34 #include "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_AryEq: 200 case Op_EncodeISOArray: 201 // Not a legit memory op for implicit null check regardless of 202 // embedded loads 203 continue; 204 default: // Also check for embedded loads 205 if( !mach->needs_anti_dependence_check() ) 206 continue; // Not an memory op; skip it 207 if( must_clone[iop] ) { 208 // Do not move nodes which produce flags because 209 // RA will try to clone it to place near branch and 210 // it will cause recompilation, see clone_node(). 211 continue; 212 } 213 { 214 // Check that value is used in memory address in 215 // instructions with embedded load (CmpP val1,(val2+off)). 216 Node* base; 217 Node* index; 218 const MachOper* oper = mach->memory_inputs(base, index); 219 if (oper == NULL || oper == (MachOper*)-1) { 220 continue; // Not an memory op; skip it 221 } 222 if (val == base || 223 val == index && val->bottom_type()->isa_narrowoop()) { 224 break; // Found it 225 } else { 226 continue; // Skip it 227 } 228 } 229 break; 230 } 231 232 // On some OSes (AIX) the page at address 0 is only write protected. 233 // If so, only Store operations will trap. 234 // But a read accessing the base of a heap-based compressed heap will trap. 235 if (!was_store && needs_explicit_null_check_for_read(val)) { 236 continue; 237 } 238 239 // check if the offset is not too high for implicit exception 240 { 241 intptr_t offset = 0; 242 const TypePtr *adr_type = NULL; // Do not need this return value here 243 const Node* base = mach->get_base_and_disp(offset, adr_type); 244 if (base == NULL || base == NodeSentinel) { 245 // Narrow oop address doesn't have base, only index 246 if( val->bottom_type()->isa_narrowoop() && 247 MacroAssembler::needs_explicit_null_check(offset) ) 248 continue; // Give up if offset is beyond page size 249 // cannot reason about it; is probably not implicit null exception 250 } else { 251 const TypePtr* tptr; 252 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 253 Universe::narrow_klass_shift() == 0)) { 254 // 32-bits narrow oop can be the base of address expressions 255 tptr = base->get_ptr_type(); 256 } else { 257 // only regular oops are expected here 258 tptr = base->bottom_type()->is_ptr(); 259 } 260 // Give up if offset is not a compile-time constant 261 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 262 continue; 263 offset += tptr->_offset; // correct if base is offseted 264 if( MacroAssembler::needs_explicit_null_check(offset) ) 265 continue; // Give up is reference is beyond 4K page size 266 } 267 } 268 269 // Check ctrl input to see if the null-check dominates the memory op 270 Block *cb = get_block_for_node(mach); 271 cb = cb->_idom; // Always hoist at least 1 block 272 if( !was_store ) { // Stores can be hoisted only one block 273 while( cb->_dom_depth > (block->_dom_depth + 1)) 274 cb = cb->_idom; // Hoist loads as far as we want 275 // The non-null-block should dominate the memory op, too. Live 276 // range spilling will insert a spill in the non-null-block if it is 277 // needs to spill the memory op for an implicit null check. 278 if (cb->_dom_depth == (block->_dom_depth + 1)) { 279 if (cb != not_null_block) continue; 280 cb = cb->_idom; 281 } 282 } 283 if( cb != block ) continue; 284 285 // Found a memory user; see if it can be hoisted to check-block 286 uint vidx = 0; // Capture index of value into memop 287 uint j; 288 for( j = mach->req()-1; j > 0; j-- ) { 289 if( mach->in(j) == val ) { 290 vidx = j; 291 // Ignore DecodeN val which could be hoisted to where needed. 292 if( is_decoden ) continue; 293 } 294 // Block of memory-op input 295 Block *inb = get_block_for_node(mach->in(j)); 296 Block *b = block; // Start from nul check 297 while( b != inb && b->_dom_depth > inb->_dom_depth ) 298 b = b->_idom; // search upwards for input 299 // See if input dominates null check 300 if( b != inb ) 301 break; 302 } 303 if( j > 0 ) 304 continue; 305 Block *mb = get_block_for_node(mach); 306 // Hoisting stores requires more checks for the anti-dependence case. 307 // Give up hoisting if we have to move the store past any load. 308 if( was_store ) { 309 Block *b = mb; // Start searching here for a local load 310 // mach use (faulting) trying to hoist 311 // n might be blocker to hoisting 312 while( b != block ) { 313 uint k; 314 for( k = 1; k < b->number_of_nodes(); k++ ) { 315 Node *n = b->get_node(k); 316 if( n->needs_anti_dependence_check() && 317 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 318 break; // Found anti-dependent load 319 } 320 if( k < b->number_of_nodes() ) 321 break; // Found anti-dependent load 322 // Make sure control does not do a merge (would have to check allpaths) 323 if( b->num_preds() != 2 ) break; 324 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 325 } 326 if( b != block ) continue; 327 } 328 329 // Make sure this memory op is not already being used for a NullCheck 330 Node *e = mb->end(); 331 if( e->is_MachNullCheck() && e->in(1) == mach ) 332 continue; // Already being used as a NULL check 333 334 // Found a candidate! Pick one with least dom depth - the highest 335 // in the dom tree should be closest to the null check. 336 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 337 best = mach; 338 bidx = vidx; 339 } 340 } 341 // No candidate! 342 if (best == NULL) { 343 return; 344 } 345 346 // ---- Found an implicit null check 347 extern int implicit_null_checks; 348 implicit_null_checks++; 349 350 if( is_decoden ) { 351 // Check if we need to hoist decodeHeapOop_not_null first. 352 Block *valb = get_block_for_node(val); 353 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 354 // Hoist it up to the end of the test block. 355 valb->find_remove(val); 356 block->add_inst(val); 357 map_node_to_block(val, block); 358 // DecodeN on x86 may kill flags. Check for flag-killing projections 359 // that also need to be hoisted. 360 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 361 Node* n = val->fast_out(j); 362 if( n->is_MachProj() ) { 363 get_block_for_node(n)->find_remove(n); 364 block->add_inst(n); 365 map_node_to_block(n, block); 366 } 367 } 368 } 369 } 370 // Hoist the memory candidate up to the end of the test block. 371 Block *old_block = get_block_for_node(best); 372 old_block->find_remove(best); 373 block->add_inst(best); 374 map_node_to_block(best, block); 375 376 // Move the control dependence 377 if (best->in(0) && best->in(0) == old_block->head()) 378 best->set_req(0, block->head()); 379 380 // Check for flag-killing projections that also need to be hoisted 381 // Should be DU safe because no edge updates. 382 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 383 Node* n = best->fast_out(j); 384 if( n->is_MachProj() ) { 385 get_block_for_node(n)->find_remove(n); 386 block->add_inst(n); 387 map_node_to_block(n, block); 388 } 389 } 390 391 // proj==Op_True --> ne test; proj==Op_False --> eq test. 392 // One of two graph shapes got matched: 393 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 394 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 395 // NULL checks are always branch-if-eq. If we see a IfTrue projection 396 // then we are replacing a 'ne' test with a 'eq' NULL check test. 397 // We need to flip the projections to keep the same semantics. 398 if( proj->Opcode() == Op_IfTrue ) { 399 // Swap order of projections in basic block to swap branch targets 400 Node *tmp1 = block->get_node(block->end_idx()+1); 401 Node *tmp2 = block->get_node(block->end_idx()+2); 402 block->map_node(tmp2, block->end_idx()+1); 403 block->map_node(tmp1, block->end_idx()+2); 404 Node *tmp = new Node(C->top()); // Use not NULL input 405 tmp1->replace_by(tmp); 406 tmp2->replace_by(tmp1); 407 tmp->replace_by(tmp2); 408 tmp->destruct(); 409 } 410 411 // Remove the existing null check; use a new implicit null check instead. 412 // Since schedule-local needs precise def-use info, we need to correct 413 // it as well. 414 Node *old_tst = proj->in(0); 415 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx); 416 block->map_node(nul_chk, block->end_idx()); 417 map_node_to_block(nul_chk, block); 418 // Redirect users of old_test to nul_chk 419 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 420 old_tst->last_out(i2)->set_req(0, nul_chk); 421 // Clean-up any dead code 422 for (uint i3 = 0; i3 < old_tst->req(); i3++) { 423 Node* in = old_tst->in(i3); 424 old_tst->set_req(i3, NULL); 425 if (in->outcnt() == 0) { 426 // Remove dead input node 427 in->disconnect_inputs(NULL, C); 428 block->find_remove(in); 429 } 430 } 431 432 latency_from_uses(nul_chk); 433 latency_from_uses(best); 434 } 435 436 437 //------------------------------select----------------------------------------- 438 // Select a nice fellow from the worklist to schedule next. If there is only 439 // one choice, then use it. Projections take top priority for correctness 440 // reasons - if I see a projection, then it is next. There are a number of 441 // other special cases, for instructions that consume condition codes, et al. 442 // These are chosen immediately. Some instructions are required to immediately 443 // precede the last instruction in the block, and these are taken last. Of the 444 // remaining cases (most), choose the instruction with the greatest latency 445 // (that is, the most number of pseudo-cycles required to the end of the 446 // routine). If there is a tie, choose the instruction with the most inputs. 447 Node* PhaseCFG::select( 448 Block* block, 449 Node_List &worklist, 450 GrowableArray<int> &ready_cnt, 451 VectorSet &next_call, 452 uint sched_slot, 453 intptr_t* recalc_pressure_nodes) { 454 455 // If only a single entry on the stack, use it 456 uint cnt = worklist.size(); 457 if (cnt == 1) { 458 Node *n = worklist[0]; 459 worklist.map(0,worklist.pop()); 460 return n; 461 } 462 463 uint choice = 0; // Bigger is most important 464 uint latency = 0; // Bigger is scheduled first 465 uint score = 0; // Bigger is better 466 int idx = -1; // Index in worklist 467 int cand_cnt = 0; // Candidate count 468 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; 469 470 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 471 // Order in worklist is used to break ties. 472 // See caller for how this is used to delay scheduling 473 // of induction variable increments to after the other 474 // uses of the phi are scheduled. 475 Node *n = worklist[i]; // Get Node on worklist 476 477 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 478 if( n->is_Proj() || // Projections always win 479 n->Opcode()== Op_Con || // So does constant 'Top' 480 iop == Op_CreateEx || // Create-exception must start block 481 iop == Op_CheckCastPP 482 ) { 483 worklist.map(i,worklist.pop()); 484 return n; 485 } 486 487 // Final call in a block must be adjacent to 'catch' 488 Node *e = block->end(); 489 if( e->is_Catch() && e->in(0)->in(0) == n ) 490 continue; 491 492 // Memory op for an implicit null check has to be at the end of the block 493 if( e->is_MachNullCheck() && e->in(1) == n ) 494 continue; 495 496 // Schedule IV increment last. 497 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 498 e->in(1)->in(1) == n && n->is_iteratively_computed()) 499 continue; 500 501 uint n_choice = 2; 502 503 // See if this instruction is consumed by a branch. If so, then (as the 504 // branch is the last instruction in the basic block) force it to the 505 // end of the basic block 506 if ( must_clone[iop] ) { 507 // See if any use is a branch 508 bool found_machif = false; 509 510 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 511 Node* use = n->fast_out(j); 512 513 // The use is a conditional branch, make them adjacent 514 if (use->is_MachIf() && get_block_for_node(use) == block) { 515 found_machif = true; 516 break; 517 } 518 519 // More than this instruction pending for successor to be ready, 520 // don't choose this if other opportunities are ready 521 if (ready_cnt.at(use->_idx) > 1) 522 n_choice = 1; 523 } 524 525 // loop terminated, prefer not to use this instruction 526 if (found_machif) 527 continue; 528 } 529 530 // See if this has a predecessor that is "must_clone", i.e. sets the 531 // condition code. If so, choose this first 532 for (uint j = 0; j < n->req() ; j++) { 533 Node *inn = n->in(j); 534 if (inn) { 535 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 536 n_choice = 3; 537 break; 538 } 539 } 540 } 541 542 // MachTemps should be scheduled last so they are near their uses 543 if (n->is_MachTemp()) { 544 n_choice = 1; 545 } 546 547 uint n_latency = get_latency_for_node(n); 548 uint n_score = n->req(); // Many inputs get high score to break ties 549 550 if (OptoRegScheduling && block_size_threshold_ok) { 551 if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) { 552 _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit()); 553 _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit()); 554 // simulate the notion that we just picked this node to schedule 555 n->add_flag(Node::Flag_is_scheduled); 556 // now caculate its effect upon the graph if we did 557 adjust_register_pressure(n, block, recalc_pressure_nodes, false); 558 // return its state for finalize in case somebody else wins 559 n->remove_flag(Node::Flag_is_scheduled); 560 // now save the two final pressure components of register pressure, limiting pressure calcs to short size 561 short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure(); 562 short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure(); 563 recalc_pressure_nodes[n->_idx] = int_pressure; 564 recalc_pressure_nodes[n->_idx] |= (float_pressure << 16); 565 } 566 567 if (_scheduling_for_pressure) { 568 latency = n_latency; 569 if (n_choice != 3) { 570 // Now evaluate each register pressure component based on threshold in the score. 571 // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks 572 // on a single instruction, but we might see it shrink on both banks. 573 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 574 short int_pressure = (short)recalc_pressure_nodes[n->_idx]; 575 n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score; 576 } 577 if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 578 short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16); 579 n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score; 580 } 581 } else { 582 // make sure we choose these candidates 583 score = 0; 584 } 585 } 586 } 587 588 // Keep best latency found 589 cand_cnt++; 590 if (choice < n_choice || 591 (choice == n_choice && 592 ((StressLCM && Compile::randomized_select(cand_cnt)) || 593 (!StressLCM && 594 (latency < n_latency || 595 (latency == n_latency && 596 (score < n_score))))))) { 597 choice = n_choice; 598 latency = n_latency; 599 score = n_score; 600 idx = i; // Also keep index in worklist 601 } 602 } // End of for all ready nodes in worklist 603 604 assert(idx >= 0, "index should be set"); 605 Node *n = worklist[(uint)idx]; // Get the winner 606 607 worklist.map((uint)idx, worklist.pop()); // Compress worklist 608 return n; 609 } 610 611 //-------------------------adjust_register_pressure---------------------------- 612 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) { 613 PhaseLive* liveinfo = _regalloc->get_live(); 614 IndexSet* liveout = liveinfo->live(block); 615 // first adjust the register pressure for the sources 616 for (uint i = 1; i < n->req(); i++) { 617 bool lrg_ends = false; 618 Node *src_n = n->in(i); 619 if (src_n == NULL) continue; 620 if (!src_n->is_Mach()) continue; 621 uint src = _regalloc->_lrg_map.find(src_n); 622 if (src == 0) continue; 623 LRG& lrg_src = _regalloc->lrgs(src); 624 // detect if the live range ends or not 625 if (liveout->member(src) == false) { 626 lrg_ends = true; 627 for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) { 628 Node* m = src_n->fast_out(j); // Get user 629 if (m == n) continue; 630 if (!m->is_Mach()) continue; 631 MachNode *mach = m->as_Mach(); 632 bool src_matches = false; 633 int iop = mach->ideal_Opcode(); 634 635 switch (iop) { 636 case Op_StoreB: 637 case Op_StoreC: 638 case Op_StoreCM: 639 case Op_StoreD: 640 case Op_StoreF: 641 case Op_StoreI: 642 case Op_StoreL: 643 case Op_StoreP: 644 case Op_StoreN: 645 case Op_StoreVector: 646 case Op_StoreNKlass: 647 for (uint k = 1; k < m->req(); k++) { 648 Node *in = m->in(k); 649 if (in == src_n) { 650 src_matches = true; 651 break; 652 } 653 } 654 break; 655 656 default: 657 src_matches = true; 658 break; 659 } 660 661 // If we have a store as our use, ignore the non source operands 662 if (src_matches == false) continue; 663 664 // Mark every unscheduled use which is not n with a recalculation 665 if ((get_block_for_node(m) == block) && (!m->is_scheduled())) { 666 if (finalize_mode && !m->is_Phi()) { 667 recalc_pressure_nodes[m->_idx] = 0x7fff7fff; 668 } 669 lrg_ends = false; 670 } 671 } 672 } 673 // if none, this live range ends and we can adjust register pressure 674 if (lrg_ends) { 675 if (finalize_mode) { 676 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 677 } else { 678 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 679 } 680 } 681 } 682 683 // now add the register pressure from the dest and evaluate which heuristic we should use: 684 // 1.) The default, latency scheduling 685 // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks 686 uint dst = _regalloc->_lrg_map.find(n); 687 if (dst != 0) { 688 LRG& lrg_dst = _regalloc->lrgs(dst); 689 if (finalize_mode) { 690 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 691 // check to see if we fall over the register pressure cliff here 692 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 693 _scheduling_for_pressure = true; 694 } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 695 _scheduling_for_pressure = true; 696 } else { 697 // restore latency scheduling mode 698 _scheduling_for_pressure = false; 699 } 700 } else { 701 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 702 } 703 } 704 } 705 706 //------------------------------set_next_call---------------------------------- 707 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 708 if( next_call.test_set(n->_idx) ) return; 709 for( uint i=0; i<n->len(); i++ ) { 710 Node *m = n->in(i); 711 if( !m ) continue; // must see all nodes in block that precede call 712 if (get_block_for_node(m) == block) { 713 set_next_call(block, m, next_call); 714 } 715 } 716 } 717 718 //------------------------------needed_for_next_call--------------------------- 719 // Set the flag 'next_call' for each Node that is needed for the next call to 720 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 721 // next subroutine call get priority - basically it moves things NOT needed 722 // for the next call till after the call. This prevents me from trying to 723 // carry lots of stuff live across a call. 724 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 725 // Find the next control-defining Node in this block 726 Node* call = NULL; 727 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 728 Node* m = this_call->fast_out(i); 729 if (get_block_for_node(m) == block && // Local-block user 730 m != this_call && // Not self-start node 731 m->is_MachCall()) { 732 call = m; 733 break; 734 } 735 } 736 if (call == NULL) return; // No next call (e.g., block end is near) 737 // Set next-call for all inputs to this call 738 set_next_call(block, call, next_call); 739 } 740 741 //------------------------------add_call_kills------------------------------------- 742 // helper function that adds caller save registers to MachProjNode 743 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 744 // Fill in the kill mask for the call 745 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 746 if( !regs.Member(r) ) { // Not already defined by the call 747 // Save-on-call register? 748 if ((save_policy[r] == 'C') || 749 (save_policy[r] == 'A') || 750 ((save_policy[r] == 'E') && exclude_soe)) { 751 proj->_rout.Insert(r); 752 } 753 } 754 } 755 } 756 757 758 //------------------------------sched_call------------------------------------- 759 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 760 RegMask regs; 761 762 // Schedule all the users of the call right now. All the users are 763 // projection Nodes, so they must be scheduled next to the call. 764 // Collect all the defined registers. 765 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 766 Node* n = mcall->fast_out(i); 767 assert( n->is_MachProj(), "" ); 768 int n_cnt = ready_cnt.at(n->_idx)-1; 769 ready_cnt.at_put(n->_idx, n_cnt); 770 assert( n_cnt == 0, "" ); 771 // Schedule next to call 772 block->map_node(n, node_cnt++); 773 // Collect defined registers 774 regs.OR(n->out_RegMask()); 775 // Check for scheduling the next control-definer 776 if( n->bottom_type() == Type::CONTROL ) 777 // Warm up next pile of heuristic bits 778 needed_for_next_call(block, n, next_call); 779 780 // Children of projections are now all ready 781 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 782 Node* m = n->fast_out(j); // Get user 783 if(get_block_for_node(m) != block) { 784 continue; 785 } 786 if( m->is_Phi() ) continue; 787 int m_cnt = ready_cnt.at(m->_idx) - 1; 788 ready_cnt.at_put(m->_idx, m_cnt); 789 if( m_cnt == 0 ) 790 worklist.push(m); 791 } 792 793 } 794 795 // Act as if the call defines the Frame Pointer. 796 // Certainly the FP is alive and well after the call. 797 regs.Insert(_matcher.c_frame_pointer()); 798 799 // Set all registers killed and not already defined by the call. 800 uint r_cnt = mcall->tf()->range()->cnt(); 801 int op = mcall->ideal_Opcode(); 802 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 803 map_node_to_block(proj, block); 804 block->insert_node(proj, node_cnt++); 805 806 // Select the right register save policy. 807 const char * save_policy; 808 switch (op) { 809 case Op_CallRuntime: 810 case Op_CallLeaf: 811 case Op_CallLeafNoFP: 812 // Calling C code so use C calling convention 813 save_policy = _matcher._c_reg_save_policy; 814 break; 815 816 case Op_CallStaticJava: 817 case Op_CallDynamicJava: 818 // Calling Java code so use Java calling convention 819 save_policy = _matcher._register_save_policy; 820 break; 821 822 default: 823 ShouldNotReachHere(); 824 } 825 826 // When using CallRuntime mark SOE registers as killed by the call 827 // so values that could show up in the RegisterMap aren't live in a 828 // callee saved register since the register wouldn't know where to 829 // find them. CallLeaf and CallLeafNoFP are ok because they can't 830 // have debug info on them. Strictly speaking this only needs to be 831 // done for oops since idealreg2debugmask takes care of debug info 832 // references but there no way to handle oops differently than other 833 // pointers as far as the kill mask goes. 834 bool exclude_soe = op == Op_CallRuntime; 835 836 // If the call is a MethodHandle invoke, we need to exclude the 837 // register which is used to save the SP value over MH invokes from 838 // the mask. Otherwise this register could be used for 839 // deoptimization information. 840 if (op == Op_CallStaticJava) { 841 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 842 if (mcallstaticjava->_method_handle_invoke) 843 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 844 } 845 846 add_call_kills(proj, regs, save_policy, exclude_soe); 847 848 return node_cnt; 849 } 850 851 852 //------------------------------schedule_local--------------------------------- 853 // Topological sort within a block. Someday become a real scheduler. 854 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) { 855 // Already "sorted" are the block start Node (as the first entry), and 856 // the block-ending Node and any trailing control projections. We leave 857 // these alone. PhiNodes and ParmNodes are made to follow the block start 858 // Node. Everything else gets topo-sorted. 859 860 #ifndef PRODUCT 861 if (trace_opto_pipelining()) { 862 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 863 for (uint i = 0;i < block->number_of_nodes(); i++) { 864 tty->print("# "); 865 block->get_node(i)->fast_dump(); 866 } 867 tty->print_cr("#"); 868 } 869 #endif 870 871 // RootNode is already sorted 872 if (block->number_of_nodes() == 1) { 873 return true; 874 } 875 876 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; 877 878 // We track the uses of local definitions as input dependences so that 879 // we know when a given instruction is avialable to be scheduled. 880 uint i; 881 if (OptoRegScheduling && block_size_threshold_ok) { 882 for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc 883 Node *n = block->get_node(i); 884 n->remove_flag(Node::Flag_is_scheduled); 885 if (!n->is_Phi()) { 886 recalc_pressure_nodes[n->_idx] = 0x7fff7fff; 887 } 888 } 889 } 890 891 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 892 uint node_cnt = block->end_idx(); 893 uint phi_cnt = 1; 894 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 895 Node *n = block->get_node(i); 896 if( n->is_Phi() || // Found a PhiNode or ParmNode 897 (n->is_Proj() && n->in(0) == block->head()) ) { 898 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 899 block->map_node(block->get_node(phi_cnt), i); 900 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 901 if (OptoRegScheduling && block_size_threshold_ok) { 902 // mark n as scheduled 903 n->add_flag(Node::Flag_is_scheduled); 904 } 905 } else { // All others 906 // Count block-local inputs to 'n' 907 uint cnt = n->len(); // Input count 908 uint local = 0; 909 for( uint j=0; j<cnt; j++ ) { 910 Node *m = n->in(j); 911 if( m && get_block_for_node(m) == block && !m->is_top() ) 912 local++; // One more block-local input 913 } 914 ready_cnt.at_put(n->_idx, local); // Count em up 915 916 #ifdef ASSERT 917 if( UseConcMarkSweepGC || UseG1GC ) { 918 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 919 // Check the precedence edges 920 for (uint prec = n->req(); prec < n->len(); prec++) { 921 Node* oop_store = n->in(prec); 922 if (oop_store != NULL) { 923 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 924 } 925 } 926 } 927 } 928 #endif 929 930 // A few node types require changing a required edge to a precedence edge 931 // before allocation. 932 if( n->is_Mach() && n->req() > TypeFunc::Parms && 933 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 934 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 935 // MemBarAcquire could be created without Precedent edge. 936 // del_req() replaces the specified edge with the last input edge 937 // and then removes the last edge. If the specified edge > number of 938 // edges the last edge will be moved outside of the input edges array 939 // and the edge will be lost. This is why this code should be 940 // executed only when Precedent (== TypeFunc::Parms) edge is present. 941 Node *x = n->in(TypeFunc::Parms); 942 n->del_req(TypeFunc::Parms); 943 n->add_prec(x); 944 } 945 } 946 } 947 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 948 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 949 950 // All the prescheduled guys do not hold back internal nodes 951 uint i3; 952 for (i3 = 0; i3 < phi_cnt; i3++) { // For all pre-scheduled 953 Node *n = block->get_node(i3); // Get pre-scheduled 954 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 955 Node* m = n->fast_out(j); 956 if (get_block_for_node(m) == block) { // Local-block user 957 int m_cnt = ready_cnt.at(m->_idx)-1; 958 if (OptoRegScheduling && block_size_threshold_ok) { 959 // mark m as scheduled 960 if (m_cnt < 0) { 961 m->add_flag(Node::Flag_is_scheduled); 962 } 963 } 964 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 965 } 966 } 967 } 968 969 Node_List delay; 970 // Make a worklist 971 Node_List worklist; 972 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 973 Node *m = block->get_node(i4); 974 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 975 if (m->is_iteratively_computed()) { 976 // Push induction variable increments last to allow other uses 977 // of the phi to be scheduled first. The select() method breaks 978 // ties in scheduling by worklist order. 979 delay.push(m); 980 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 981 // Force the CreateEx to the top of the list so it's processed 982 // first and ends up at the start of the block. 983 worklist.insert(0, m); 984 } else { 985 worklist.push(m); // Then on to worklist! 986 } 987 } 988 } 989 while (delay.size()) { 990 Node* d = delay.pop(); 991 worklist.push(d); 992 } 993 994 if (OptoRegScheduling && block_size_threshold_ok) { 995 // To stage register pressure calculations we need to examine the live set variables 996 // breaking them up by register class to compartmentalize the calculations. 997 uint float_pressure = FLOATPRESSURE; 998 #ifdef _LP64 999 if (UseAVX > 2) { 1000 float_pressure *= 2; 1001 } 1002 #endif 1003 _regalloc->_sched_int_pressure.init(INTPRESSURE); 1004 _regalloc->_sched_float_pressure.init(float_pressure); 1005 _regalloc->_scratch_int_pressure.init(INTPRESSURE); 1006 _regalloc->_scratch_float_pressure.init(float_pressure); 1007 1008 _regalloc->compute_entry_block_pressure(block); 1009 } 1010 1011 // Warm up the 'next_call' heuristic bits 1012 needed_for_next_call(block, block->head(), next_call); 1013 1014 #ifndef PRODUCT 1015 if (trace_opto_pipelining()) { 1016 for (uint j=0; j< block->number_of_nodes(); j++) { 1017 Node *n = block->get_node(j); 1018 int idx = n->_idx; 1019 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 1020 tty->print("latency:%3d ", get_latency_for_node(n)); 1021 tty->print("%4d: %s\n", idx, n->Name()); 1022 } 1023 } 1024 #endif 1025 1026 uint max_idx = (uint)ready_cnt.length(); 1027 // Pull from worklist and schedule 1028 while( worklist.size() ) { // Worklist is not ready 1029 1030 #ifndef PRODUCT 1031 if (trace_opto_pipelining()) { 1032 tty->print("# ready list:"); 1033 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1034 Node *n = worklist[i]; // Get Node on worklist 1035 tty->print(" %d", n->_idx); 1036 } 1037 tty->cr(); 1038 } 1039 #endif 1040 1041 // Select and pop a ready guy from worklist 1042 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes); 1043 block->map_node(n, phi_cnt++); // Schedule him next 1044 1045 if (OptoRegScheduling && block_size_threshold_ok) { 1046 n->add_flag(Node::Flag_is_scheduled); 1047 1048 // Now adjust the resister pressure with the node we selected 1049 if (!n->is_Phi()) { 1050 adjust_register_pressure(n, block, recalc_pressure_nodes, true); 1051 } 1052 } 1053 1054 #ifndef PRODUCT 1055 if (trace_opto_pipelining()) { 1056 tty->print("# select %d: %s", n->_idx, n->Name()); 1057 tty->print(", latency:%d", get_latency_for_node(n)); 1058 n->dump(); 1059 if (Verbose) { 1060 tty->print("# ready list:"); 1061 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1062 Node *n = worklist[i]; // Get Node on worklist 1063 tty->print(" %d", n->_idx); 1064 } 1065 tty->cr(); 1066 } 1067 } 1068 1069 #endif 1070 if( n->is_MachCall() ) { 1071 MachCallNode *mcall = n->as_MachCall(); 1072 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 1073 continue; 1074 } 1075 1076 if (n->is_Mach() && n->as_Mach()->has_call()) { 1077 RegMask regs; 1078 regs.Insert(_matcher.c_frame_pointer()); 1079 regs.OR(n->out_RegMask()); 1080 1081 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 1082 map_node_to_block(proj, block); 1083 block->insert_node(proj, phi_cnt++); 1084 1085 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 1086 } 1087 1088 // Children are now all ready 1089 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 1090 Node* m = n->fast_out(i5); // Get user 1091 if (get_block_for_node(m) != block) { 1092 continue; 1093 } 1094 if( m->is_Phi() ) continue; 1095 if (m->_idx >= max_idx) { // new node, skip it 1096 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 1097 continue; 1098 } 1099 int m_cnt = ready_cnt.at(m->_idx) - 1; 1100 ready_cnt.at_put(m->_idx, m_cnt); 1101 if( m_cnt == 0 ) 1102 worklist.push(m); 1103 } 1104 } 1105 1106 if( phi_cnt != block->end_idx() ) { 1107 // did not schedule all. Retry, Bailout, or Die 1108 if (C->subsume_loads() == true && !C->failing()) { 1109 // Retry with subsume_loads == false 1110 // If this is the first failure, the sentinel string will "stick" 1111 // to the Compile object, and the C2Compiler will see it and retry. 1112 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 1113 } 1114 // assert( phi_cnt == end_idx(), "did not schedule all" ); 1115 return false; 1116 } 1117 1118 if (OptoRegScheduling && block_size_threshold_ok) { 1119 _regalloc->compute_exit_block_pressure(block); 1120 block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure(); 1121 block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure(); 1122 } 1123 1124 #ifndef PRODUCT 1125 if (trace_opto_pipelining()) { 1126 tty->print_cr("#"); 1127 tty->print_cr("# after schedule_local"); 1128 for (uint i = 0;i < block->number_of_nodes();i++) { 1129 tty->print("# "); 1130 block->get_node(i)->fast_dump(); 1131 } 1132 tty->print_cr("# "); 1133 1134 if (OptoRegScheduling && block_size_threshold_ok) { 1135 tty->print_cr("# pressure info : %d", block->_pre_order); 1136 _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info"); 1137 _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info"); 1138 } 1139 tty->cr(); 1140 } 1141 #endif 1142 1143 return true; 1144 } 1145 1146 //--------------------------catch_cleanup_fix_all_inputs----------------------- 1147 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 1148 for (uint l = 0; l < use->len(); l++) { 1149 if (use->in(l) == old_def) { 1150 if (l < use->req()) { 1151 use->set_req(l, new_def); 1152 } else { 1153 use->rm_prec(l); 1154 use->add_prec(new_def); 1155 l--; 1156 } 1157 } 1158 } 1159 } 1160 1161 //------------------------------catch_cleanup_find_cloned_def------------------ 1162 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1163 assert( use_blk != def_blk, "Inter-block cleanup only"); 1164 1165 // The use is some block below the Catch. Find and return the clone of the def 1166 // that dominates the use. If there is no clone in a dominating block, then 1167 // create a phi for the def in a dominating block. 1168 1169 // Find which successor block dominates this use. The successor 1170 // blocks must all be single-entry (from the Catch only; I will have 1171 // split blocks to make this so), hence they all dominate. 1172 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 1173 use_blk = use_blk->_idom; 1174 1175 // Find the successor 1176 Node *fixup = NULL; 1177 1178 uint j; 1179 for( j = 0; j < def_blk->_num_succs; j++ ) 1180 if( use_blk == def_blk->_succs[j] ) 1181 break; 1182 1183 if( j == def_blk->_num_succs ) { 1184 // Block at same level in dom-tree is not a successor. It needs a 1185 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 1186 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 1187 for(uint k = 1; k < use_blk->num_preds(); k++) { 1188 Block* block = get_block_for_node(use_blk->pred(k)); 1189 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 1190 } 1191 1192 // Check to see if the use_blk already has an identical phi inserted. 1193 // If it exists, it will be at the first position since all uses of a 1194 // def are processed together. 1195 Node *phi = use_blk->get_node(1); 1196 if( phi->is_Phi() ) { 1197 fixup = phi; 1198 for (uint k = 1; k < use_blk->num_preds(); k++) { 1199 if (phi->in(k) != inputs[k]) { 1200 // Not a match 1201 fixup = NULL; 1202 break; 1203 } 1204 } 1205 } 1206 1207 // If an existing PhiNode was not found, make a new one. 1208 if (fixup == NULL) { 1209 Node *new_phi = PhiNode::make(use_blk->head(), def); 1210 use_blk->insert_node(new_phi, 1); 1211 map_node_to_block(new_phi, use_blk); 1212 for (uint k = 1; k < use_blk->num_preds(); k++) { 1213 new_phi->set_req(k, inputs[k]); 1214 } 1215 fixup = new_phi; 1216 } 1217 1218 } else { 1219 // Found the use just below the Catch. Make it use the clone. 1220 fixup = use_blk->get_node(n_clone_idx); 1221 } 1222 1223 return fixup; 1224 } 1225 1226 //--------------------------catch_cleanup_intra_block-------------------------- 1227 // Fix all input edges in use that reference "def". The use is in the same 1228 // block as the def and both have been cloned in each successor block. 1229 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1230 1231 // Both the use and def have been cloned. For each successor block, 1232 // get the clone of the use, and make its input the clone of the def 1233 // found in that block. 1234 1235 uint use_idx = blk->find_node(use); 1236 uint offset_idx = use_idx - beg; 1237 for( uint k = 0; k < blk->_num_succs; k++ ) { 1238 // Get clone in each successor block 1239 Block *sb = blk->_succs[k]; 1240 Node *clone = sb->get_node(offset_idx+1); 1241 assert( clone->Opcode() == use->Opcode(), "" ); 1242 1243 // Make use-clone reference the def-clone 1244 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1245 } 1246 } 1247 1248 //------------------------------catch_cleanup_inter_block--------------------- 1249 // Fix all input edges in use that reference "def". The use is in a different 1250 // block than the def. 1251 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1252 if( !use_blk ) return; // Can happen if the use is a precedence edge 1253 1254 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1255 catch_cleanup_fix_all_inputs(use, def, new_def); 1256 } 1257 1258 //------------------------------call_catch_cleanup----------------------------- 1259 // If we inserted any instructions between a Call and his CatchNode, 1260 // clone the instructions on all paths below the Catch. 1261 void PhaseCFG::call_catch_cleanup(Block* block) { 1262 1263 // End of region to clone 1264 uint end = block->end_idx(); 1265 if( !block->get_node(end)->is_Catch() ) return; 1266 // Start of region to clone 1267 uint beg = end; 1268 while(!block->get_node(beg-1)->is_MachProj() || 1269 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1270 beg--; 1271 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1272 } 1273 // Range of inserted instructions is [beg, end) 1274 if( beg == end ) return; 1275 1276 // Clone along all Catch output paths. Clone area between the 'beg' and 1277 // 'end' indices. 1278 for( uint i = 0; i < block->_num_succs; i++ ) { 1279 Block *sb = block->_succs[i]; 1280 // Clone the entire area; ignoring the edge fixup for now. 1281 for( uint j = end; j > beg; j-- ) { 1282 // It is safe here to clone a node with anti_dependence 1283 // since clones dominate on each path. 1284 Node *clone = block->get_node(j-1)->clone(); 1285 sb->insert_node(clone, 1); 1286 map_node_to_block(clone, sb); 1287 } 1288 } 1289 1290 1291 // Fixup edges. Check the def-use info per cloned Node 1292 for(uint i2 = beg; i2 < end; i2++ ) { 1293 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1294 Node *n = block->get_node(i2); // Node that got cloned 1295 // Need DU safe iterator because of edge manipulation in calls. 1296 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1297 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1298 out->push(n->fast_out(j1)); 1299 } 1300 uint max = out->size(); 1301 for (uint j = 0; j < max; j++) {// For all users 1302 Node *use = out->pop(); 1303 Block *buse = get_block_for_node(use); 1304 if( use->is_Phi() ) { 1305 for( uint k = 1; k < use->req(); k++ ) 1306 if( use->in(k) == n ) { 1307 Block* b = get_block_for_node(buse->pred(k)); 1308 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1309 use->set_req(k, fixup); 1310 } 1311 } else { 1312 if (block == buse) { 1313 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1314 } else { 1315 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1316 } 1317 } 1318 } // End for all users 1319 1320 } // End of for all Nodes in cloned area 1321 1322 // Remove the now-dead cloned ops 1323 for(uint i3 = beg; i3 < end; i3++ ) { 1324 block->get_node(beg)->disconnect_inputs(NULL, C); 1325 block->remove_node(beg); 1326 } 1327 1328 // If the successor blocks have a CreateEx node, move it back to the top 1329 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1330 Block *sb = block->_succs[i4]; 1331 uint new_cnt = end - beg; 1332 // Remove any newly created, but dead, nodes. 1333 for( uint j = new_cnt; j > 0; j-- ) { 1334 Node *n = sb->get_node(j); 1335 if (n->outcnt() == 0 && 1336 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1337 n->disconnect_inputs(NULL, C); 1338 sb->remove_node(j); 1339 new_cnt--; 1340 } 1341 } 1342 // If any newly created nodes remain, move the CreateEx node to the top 1343 if (new_cnt > 0) { 1344 Node *cex = sb->get_node(1+new_cnt); 1345 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1346 sb->remove_node(1+new_cnt); 1347 sb->insert_node(cex, 1); 1348 } 1349 } 1350 } 1351 }