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 !Universe::narrow_oop_use_implicit_null_checks())) 261 continue; // Give up if offset is beyond page size 262 // cannot reason about it; is probably not implicit null exception 263 } else { 264 const TypePtr* tptr; 265 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 266 Universe::narrow_klass_shift() == 0)) { 267 // 32-bits narrow oop can be the base of address expressions 268 tptr = base->get_ptr_type(); 269 } else { 270 // only regular oops are expected here 271 tptr = base->bottom_type()->is_ptr(); 272 } 273 // Give up if offset is not a compile-time constant 274 if (offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot) 275 continue; 276 offset += tptr->_offset; // correct if base is offseted 277 if (MacroAssembler::needs_explicit_null_check(offset)) 278 continue; // Give up is reference is beyond 4K page size 279 // Access to non-protected heap base 280 if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN && 281 !Universe::narrow_oop_use_implicit_null_checks()) 282 continue; 283 } 284 } 285 286 // Check ctrl input to see if the null-check dominates the memory op 287 Block *cb = get_block_for_node(mach); 288 cb = cb->_idom; // Always hoist at least 1 block 289 if( !was_store ) { // Stores can be hoisted only one block 290 while( cb->_dom_depth > (block->_dom_depth + 1)) 291 cb = cb->_idom; // Hoist loads as far as we want 292 // The non-null-block should dominate the memory op, too. Live 293 // range spilling will insert a spill in the non-null-block if it is 294 // needs to spill the memory op for an implicit null check. 295 if (cb->_dom_depth == (block->_dom_depth + 1)) { 296 if (cb != not_null_block) continue; 297 cb = cb->_idom; 298 } 299 } 300 if( cb != block ) continue; 301 302 // Found a memory user; see if it can be hoisted to check-block 303 uint vidx = 0; // Capture index of value into memop 304 uint j; 305 for( j = mach->req()-1; j > 0; j-- ) { 306 if( mach->in(j) == val ) { 307 vidx = j; 308 // Ignore DecodeN val which could be hoisted to where needed. 309 if( is_decoden ) continue; 310 } 311 // Block of memory-op input 312 Block *inb = get_block_for_node(mach->in(j)); 313 Block *b = block; // Start from nul check 314 while( b != inb && b->_dom_depth > inb->_dom_depth ) 315 b = b->_idom; // search upwards for input 316 // See if input dominates null check 317 if( b != inb ) 318 break; 319 } 320 if( j > 0 ) 321 continue; 322 Block *mb = get_block_for_node(mach); 323 // Hoisting stores requires more checks for the anti-dependence case. 324 // Give up hoisting if we have to move the store past any load. 325 if( was_store ) { 326 Block *b = mb; // Start searching here for a local load 327 // mach use (faulting) trying to hoist 328 // n might be blocker to hoisting 329 while( b != block ) { 330 uint k; 331 for( k = 1; k < b->number_of_nodes(); k++ ) { 332 Node *n = b->get_node(k); 333 if( n->needs_anti_dependence_check() && 334 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 335 break; // Found anti-dependent load 336 } 337 if( k < b->number_of_nodes() ) 338 break; // Found anti-dependent load 339 // Make sure control does not do a merge (would have to check allpaths) 340 if( b->num_preds() != 2 ) break; 341 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 342 } 343 if( b != block ) continue; 344 } 345 346 // Make sure this memory op is not already being used for a NullCheck 347 Node *e = mb->end(); 348 if( e->is_MachNullCheck() && e->in(1) == mach ) 349 continue; // Already being used as a NULL check 350 351 // Found a candidate! Pick one with least dom depth - the highest 352 // in the dom tree should be closest to the null check. 353 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 354 best = mach; 355 bidx = vidx; 356 } 357 } 358 // No candidate! 359 if (best == NULL) { 360 return; 361 } 362 363 // ---- Found an implicit null check 364 #ifndef PRODUCT 365 extern int implicit_null_checks; 366 implicit_null_checks++; 367 #endif 368 369 if( is_decoden ) { 370 // Check if we need to hoist decodeHeapOop_not_null first. 371 Block *valb = get_block_for_node(val); 372 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 373 // Hoist it up to the end of the test block. 374 valb->find_remove(val); 375 block->add_inst(val); 376 map_node_to_block(val, block); 377 // DecodeN on x86 may kill flags. Check for flag-killing projections 378 // that also need to be hoisted. 379 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 380 Node* n = val->fast_out(j); 381 if( n->is_MachProj() ) { 382 get_block_for_node(n)->find_remove(n); 383 block->add_inst(n); 384 map_node_to_block(n, block); 385 } 386 } 387 } 388 } 389 // Hoist the memory candidate up to the end of the test block. 390 Block *old_block = get_block_for_node(best); 391 old_block->find_remove(best); 392 block->add_inst(best); 393 map_node_to_block(best, block); 394 395 // Move the control dependence if it is pinned to not-null block. 396 // Don't change it in other cases: NULL or dominating control. 397 if (best->in(0) == not_null_block->head()) { 398 // Set it to control edge of null check. 399 best->set_req(0, proj->in(0)->in(0)); 400 } 401 402 // Check for flag-killing projections that also need to be hoisted 403 // Should be DU safe because no edge updates. 404 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 405 Node* n = best->fast_out(j); 406 if( n->is_MachProj() ) { 407 get_block_for_node(n)->find_remove(n); 408 block->add_inst(n); 409 map_node_to_block(n, block); 410 } 411 } 412 413 // proj==Op_True --> ne test; proj==Op_False --> eq test. 414 // One of two graph shapes got matched: 415 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 416 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 417 // NULL checks are always branch-if-eq. If we see a IfTrue projection 418 // then we are replacing a 'ne' test with a 'eq' NULL check test. 419 // We need to flip the projections to keep the same semantics. 420 if( proj->Opcode() == Op_IfTrue ) { 421 // Swap order of projections in basic block to swap branch targets 422 Node *tmp1 = block->get_node(block->end_idx()+1); 423 Node *tmp2 = block->get_node(block->end_idx()+2); 424 block->map_node(tmp2, block->end_idx()+1); 425 block->map_node(tmp1, block->end_idx()+2); 426 Node *tmp = new Node(C->top()); // Use not NULL input 427 tmp1->replace_by(tmp); 428 tmp2->replace_by(tmp1); 429 tmp->replace_by(tmp2); 430 tmp->destruct(); 431 } 432 433 // Remove the existing null check; use a new implicit null check instead. 434 // Since schedule-local needs precise def-use info, we need to correct 435 // it as well. 436 Node *old_tst = proj->in(0); 437 MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx); 438 block->map_node(nul_chk, block->end_idx()); 439 map_node_to_block(nul_chk, block); 440 // Redirect users of old_test to nul_chk 441 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 442 old_tst->last_out(i2)->set_req(0, nul_chk); 443 // Clean-up any dead code 444 for (uint i3 = 0; i3 < old_tst->req(); i3++) { 445 Node* in = old_tst->in(i3); 446 old_tst->set_req(i3, NULL); 447 if (in->outcnt() == 0) { 448 // Remove dead input node 449 in->disconnect_inputs(NULL, C); 450 block->find_remove(in); 451 } 452 } 453 454 latency_from_uses(nul_chk); 455 latency_from_uses(best); 456 457 // insert anti-dependences to defs in this block 458 if (! best->needs_anti_dependence_check()) { 459 for (uint k = 1; k < block->number_of_nodes(); k++) { 460 Node *n = block->get_node(k); 461 if (n->needs_anti_dependence_check() && 462 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) { 463 // Found anti-dependent load 464 insert_anti_dependences(block, n); 465 } 466 } 467 } 468 } 469 470 471 //------------------------------select----------------------------------------- 472 // Select a nice fellow from the worklist to schedule next. If there is only 473 // one choice, then use it. Projections take top priority for correctness 474 // reasons - if I see a projection, then it is next. There are a number of 475 // other special cases, for instructions that consume condition codes, et al. 476 // These are chosen immediately. Some instructions are required to immediately 477 // precede the last instruction in the block, and these are taken last. Of the 478 // remaining cases (most), choose the instruction with the greatest latency 479 // (that is, the most number of pseudo-cycles required to the end of the 480 // routine). If there is a tie, choose the instruction with the most inputs. 481 Node* PhaseCFG::select( 482 Block* block, 483 Node_List &worklist, 484 GrowableArray<int> &ready_cnt, 485 VectorSet &next_call, 486 uint sched_slot, 487 intptr_t* recalc_pressure_nodes) { 488 489 // If only a single entry on the stack, use it 490 uint cnt = worklist.size(); 491 if (cnt == 1) { 492 Node *n = worklist[0]; 493 worklist.map(0,worklist.pop()); 494 return n; 495 } 496 497 uint choice = 0; // Bigger is most important 498 uint latency = 0; // Bigger is scheduled first 499 uint score = 0; // Bigger is better 500 int idx = -1; // Index in worklist 501 int cand_cnt = 0; // Candidate count 502 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; 503 504 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 505 // Order in worklist is used to break ties. 506 // See caller for how this is used to delay scheduling 507 // of induction variable increments to after the other 508 // uses of the phi are scheduled. 509 Node *n = worklist[i]; // Get Node on worklist 510 511 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 512 if( n->is_Proj() || // Projections always win 513 n->Opcode()== Op_Con || // So does constant 'Top' 514 iop == Op_CreateEx || // Create-exception must start block 515 iop == Op_CheckCastPP 516 ) { 517 worklist.map(i,worklist.pop()); 518 return n; 519 } 520 521 // Final call in a block must be adjacent to 'catch' 522 Node *e = block->end(); 523 if( e->is_Catch() && e->in(0)->in(0) == n ) 524 continue; 525 526 // Memory op for an implicit null check has to be at the end of the block 527 if( e->is_MachNullCheck() && e->in(1) == n ) 528 continue; 529 530 // Schedule IV increment last. 531 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd) { 532 // Cmp might be matched into CountedLoopEnd node. 533 Node *cmp = (e->in(1)->ideal_reg() == Op_RegFlags) ? e->in(1) : e; 534 if (cmp->req() > 1 && cmp->in(1) == n && n->is_iteratively_computed()) { 535 continue; 536 } 537 } 538 539 uint n_choice = 2; 540 541 // See if this instruction is consumed by a branch. If so, then (as the 542 // branch is the last instruction in the basic block) force it to the 543 // end of the basic block 544 if ( must_clone[iop] ) { 545 // See if any use is a branch 546 bool found_machif = false; 547 548 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 549 Node* use = n->fast_out(j); 550 551 // The use is a conditional branch, make them adjacent 552 if (use->is_MachIf() && get_block_for_node(use) == block) { 553 found_machif = true; 554 break; 555 } 556 557 // More than this instruction pending for successor to be ready, 558 // don't choose this if other opportunities are ready 559 if (ready_cnt.at(use->_idx) > 1) 560 n_choice = 1; 561 } 562 563 // loop terminated, prefer not to use this instruction 564 if (found_machif) 565 continue; 566 } 567 568 // See if this has a predecessor that is "must_clone", i.e. sets the 569 // condition code. If so, choose this first 570 for (uint j = 0; j < n->req() ; j++) { 571 Node *inn = n->in(j); 572 if (inn) { 573 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 574 n_choice = 3; 575 break; 576 } 577 } 578 } 579 580 // MachTemps should be scheduled last so they are near their uses 581 if (n->is_MachTemp()) { 582 n_choice = 1; 583 } 584 585 uint n_latency = get_latency_for_node(n); 586 uint n_score = n->req(); // Many inputs get high score to break ties 587 588 if (OptoRegScheduling && block_size_threshold_ok) { 589 if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) { 590 _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit()); 591 _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit()); 592 // simulate the notion that we just picked this node to schedule 593 n->add_flag(Node::Flag_is_scheduled); 594 // now caculate its effect upon the graph if we did 595 adjust_register_pressure(n, block, recalc_pressure_nodes, false); 596 // return its state for finalize in case somebody else wins 597 n->remove_flag(Node::Flag_is_scheduled); 598 // now save the two final pressure components of register pressure, limiting pressure calcs to short size 599 short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure(); 600 short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure(); 601 recalc_pressure_nodes[n->_idx] = int_pressure; 602 recalc_pressure_nodes[n->_idx] |= (float_pressure << 16); 603 } 604 605 if (_scheduling_for_pressure) { 606 latency = n_latency; 607 if (n_choice != 3) { 608 // Now evaluate each register pressure component based on threshold in the score. 609 // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks 610 // on a single instruction, but we might see it shrink on both banks. 611 // For each use of register that has a register class that is over the high pressure limit, we build n_score up for 612 // live ranges that terminate on this instruction. 613 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 614 short int_pressure = (short)recalc_pressure_nodes[n->_idx]; 615 n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score; 616 } 617 if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 618 short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16); 619 n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score; 620 } 621 } else { 622 // make sure we choose these candidates 623 score = 0; 624 } 625 } 626 } 627 628 // Keep best latency found 629 cand_cnt++; 630 if (choice < n_choice || 631 (choice == n_choice && 632 ((StressLCM && Compile::randomized_select(cand_cnt)) || 633 (!StressLCM && 634 (latency < n_latency || 635 (latency == n_latency && 636 (score < n_score))))))) { 637 choice = n_choice; 638 latency = n_latency; 639 score = n_score; 640 idx = i; // Also keep index in worklist 641 } 642 } // End of for all ready nodes in worklist 643 644 assert(idx >= 0, "index should be set"); 645 Node *n = worklist[(uint)idx]; // Get the winner 646 647 worklist.map((uint)idx, worklist.pop()); // Compress worklist 648 return n; 649 } 650 651 //-------------------------adjust_register_pressure---------------------------- 652 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) { 653 PhaseLive* liveinfo = _regalloc->get_live(); 654 IndexSet* liveout = liveinfo->live(block); 655 // first adjust the register pressure for the sources 656 for (uint i = 1; i < n->req(); i++) { 657 bool lrg_ends = false; 658 Node *src_n = n->in(i); 659 if (src_n == NULL) continue; 660 if (!src_n->is_Mach()) continue; 661 uint src = _regalloc->_lrg_map.find(src_n); 662 if (src == 0) continue; 663 LRG& lrg_src = _regalloc->lrgs(src); 664 // detect if the live range ends or not 665 if (liveout->member(src) == false) { 666 lrg_ends = true; 667 for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) { 668 Node* m = src_n->fast_out(j); // Get user 669 if (m == n) continue; 670 if (!m->is_Mach()) continue; 671 MachNode *mach = m->as_Mach(); 672 bool src_matches = false; 673 int iop = mach->ideal_Opcode(); 674 675 switch (iop) { 676 case Op_StoreB: 677 case Op_StoreC: 678 case Op_StoreCM: 679 case Op_StoreD: 680 case Op_StoreF: 681 case Op_StoreI: 682 case Op_StoreL: 683 case Op_StoreP: 684 case Op_StoreN: 685 case Op_StoreVector: 686 case Op_StoreNKlass: 687 for (uint k = 1; k < m->req(); k++) { 688 Node *in = m->in(k); 689 if (in == src_n) { 690 src_matches = true; 691 break; 692 } 693 } 694 break; 695 696 default: 697 src_matches = true; 698 break; 699 } 700 701 // If we have a store as our use, ignore the non source operands 702 if (src_matches == false) continue; 703 704 // Mark every unscheduled use which is not n with a recalculation 705 if ((get_block_for_node(m) == block) && (!m->is_scheduled())) { 706 if (finalize_mode && !m->is_Phi()) { 707 recalc_pressure_nodes[m->_idx] = 0x7fff7fff; 708 } 709 lrg_ends = false; 710 } 711 } 712 } 713 // if none, this live range ends and we can adjust register pressure 714 if (lrg_ends) { 715 if (finalize_mode) { 716 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 717 } else { 718 _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 719 } 720 } 721 } 722 723 // now add the register pressure from the dest and evaluate which heuristic we should use: 724 // 1.) The default, latency scheduling 725 // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks 726 uint dst = _regalloc->_lrg_map.find(n); 727 if (dst != 0) { 728 LRG& lrg_dst = _regalloc->lrgs(dst); 729 if (finalize_mode) { 730 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); 731 // check to see if we fall over the register pressure cliff here 732 if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { 733 _scheduling_for_pressure = true; 734 } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { 735 _scheduling_for_pressure = true; 736 } else { 737 // restore latency scheduling mode 738 _scheduling_for_pressure = false; 739 } 740 } else { 741 _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); 742 } 743 } 744 } 745 746 //------------------------------set_next_call---------------------------------- 747 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 748 if( next_call.test_set(n->_idx) ) return; 749 for( uint i=0; i<n->len(); i++ ) { 750 Node *m = n->in(i); 751 if( !m ) continue; // must see all nodes in block that precede call 752 if (get_block_for_node(m) == block) { 753 set_next_call(block, m, next_call); 754 } 755 } 756 } 757 758 //------------------------------needed_for_next_call--------------------------- 759 // Set the flag 'next_call' for each Node that is needed for the next call to 760 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 761 // next subroutine call get priority - basically it moves things NOT needed 762 // for the next call till after the call. This prevents me from trying to 763 // carry lots of stuff live across a call. 764 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 765 // Find the next control-defining Node in this block 766 Node* call = NULL; 767 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 768 Node* m = this_call->fast_out(i); 769 if (get_block_for_node(m) == block && // Local-block user 770 m != this_call && // Not self-start node 771 m->is_MachCall()) { 772 call = m; 773 break; 774 } 775 } 776 if (call == NULL) return; // No next call (e.g., block end is near) 777 // Set next-call for all inputs to this call 778 set_next_call(block, call, next_call); 779 } 780 781 //------------------------------add_call_kills------------------------------------- 782 // helper function that adds caller save registers to MachProjNode 783 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 784 // Fill in the kill mask for the call 785 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 786 if( !regs.Member(r) ) { // Not already defined by the call 787 // Save-on-call register? 788 if ((save_policy[r] == 'C') || 789 (save_policy[r] == 'A') || 790 ((save_policy[r] == 'E') && exclude_soe)) { 791 proj->_rout.Insert(r); 792 } 793 } 794 } 795 } 796 797 798 //------------------------------sched_call------------------------------------- 799 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 800 RegMask regs; 801 802 // Schedule all the users of the call right now. All the users are 803 // projection Nodes, so they must be scheduled next to the call. 804 // Collect all the defined registers. 805 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 806 Node* n = mcall->fast_out(i); 807 assert( n->is_MachProj(), "" ); 808 int n_cnt = ready_cnt.at(n->_idx)-1; 809 ready_cnt.at_put(n->_idx, n_cnt); 810 assert( n_cnt == 0, "" ); 811 // Schedule next to call 812 block->map_node(n, node_cnt++); 813 // Collect defined registers 814 regs.OR(n->out_RegMask()); 815 // Check for scheduling the next control-definer 816 if( n->bottom_type() == Type::CONTROL ) 817 // Warm up next pile of heuristic bits 818 needed_for_next_call(block, n, next_call); 819 820 // Children of projections are now all ready 821 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 822 Node* m = n->fast_out(j); // Get user 823 if(get_block_for_node(m) != block) { 824 continue; 825 } 826 if( m->is_Phi() ) continue; 827 int m_cnt = ready_cnt.at(m->_idx) - 1; 828 ready_cnt.at_put(m->_idx, m_cnt); 829 if( m_cnt == 0 ) 830 worklist.push(m); 831 } 832 833 } 834 835 // Act as if the call defines the Frame Pointer. 836 // Certainly the FP is alive and well after the call. 837 regs.Insert(_matcher.c_frame_pointer()); 838 839 // Set all registers killed and not already defined by the call. 840 uint r_cnt = mcall->tf()->range()->cnt(); 841 int op = mcall->ideal_Opcode(); 842 MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 843 map_node_to_block(proj, block); 844 block->insert_node(proj, node_cnt++); 845 846 // Select the right register save policy. 847 const char *save_policy = NULL; 848 switch (op) { 849 case Op_CallRuntime: 850 case Op_CallLeaf: 851 case Op_CallLeafNoFP: 852 // Calling C code so use C calling convention 853 save_policy = _matcher._c_reg_save_policy; 854 break; 855 856 case Op_CallStaticJava: 857 case Op_CallDynamicJava: 858 // Calling Java code so use Java calling convention 859 save_policy = _matcher._register_save_policy; 860 break; 861 862 default: 863 ShouldNotReachHere(); 864 } 865 866 // When using CallRuntime mark SOE registers as killed by the call 867 // so values that could show up in the RegisterMap aren't live in a 868 // callee saved register since the register wouldn't know where to 869 // find them. CallLeaf and CallLeafNoFP are ok because they can't 870 // have debug info on them. Strictly speaking this only needs to be 871 // done for oops since idealreg2debugmask takes care of debug info 872 // references but there no way to handle oops differently than other 873 // pointers as far as the kill mask goes. 874 bool exclude_soe = op == Op_CallRuntime; 875 876 // If the call is a MethodHandle invoke, we need to exclude the 877 // register which is used to save the SP value over MH invokes from 878 // the mask. Otherwise this register could be used for 879 // deoptimization information. 880 if (op == Op_CallStaticJava) { 881 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 882 if (mcallstaticjava->_method_handle_invoke) 883 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 884 } 885 886 add_call_kills(proj, regs, save_policy, exclude_soe); 887 888 return node_cnt; 889 } 890 891 892 //------------------------------schedule_local--------------------------------- 893 // Topological sort within a block. Someday become a real scheduler. 894 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) { 895 // Already "sorted" are the block start Node (as the first entry), and 896 // the block-ending Node and any trailing control projections. We leave 897 // these alone. PhiNodes and ParmNodes are made to follow the block start 898 // Node. Everything else gets topo-sorted. 899 900 #ifndef PRODUCT 901 if (trace_opto_pipelining()) { 902 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 903 for (uint i = 0;i < block->number_of_nodes(); i++) { 904 tty->print("# "); 905 block->get_node(i)->fast_dump(); 906 } 907 tty->print_cr("#"); 908 } 909 #endif 910 911 // RootNode is already sorted 912 if (block->number_of_nodes() == 1) { 913 return true; 914 } 915 916 bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; 917 918 // We track the uses of local definitions as input dependences so that 919 // we know when a given instruction is avialable to be scheduled. 920 uint i; 921 if (OptoRegScheduling && block_size_threshold_ok) { 922 for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc 923 Node *n = block->get_node(i); 924 n->remove_flag(Node::Flag_is_scheduled); 925 if (!n->is_Phi()) { 926 recalc_pressure_nodes[n->_idx] = 0x7fff7fff; 927 } 928 } 929 } 930 931 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 932 uint node_cnt = block->end_idx(); 933 uint phi_cnt = 1; 934 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 935 Node *n = block->get_node(i); 936 if( n->is_Phi() || // Found a PhiNode or ParmNode 937 (n->is_Proj() && n->in(0) == block->head()) ) { 938 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 939 block->map_node(block->get_node(phi_cnt), i); 940 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 941 if (OptoRegScheduling && block_size_threshold_ok) { 942 // mark n as scheduled 943 n->add_flag(Node::Flag_is_scheduled); 944 } 945 } else { // All others 946 // Count block-local inputs to 'n' 947 uint cnt = n->len(); // Input count 948 uint local = 0; 949 for( uint j=0; j<cnt; j++ ) { 950 Node *m = n->in(j); 951 if( m && get_block_for_node(m) == block && !m->is_top() ) 952 local++; // One more block-local input 953 } 954 ready_cnt.at_put(n->_idx, local); // Count em up 955 956 #ifdef ASSERT 957 if( UseConcMarkSweepGC || UseG1GC ) { 958 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 959 // Check the precedence edges 960 for (uint prec = n->req(); prec < n->len(); prec++) { 961 Node* oop_store = n->in(prec); 962 if (oop_store != NULL) { 963 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 964 } 965 } 966 } 967 } 968 #endif 969 970 // A few node types require changing a required edge to a precedence edge 971 // before allocation. 972 if( n->is_Mach() && n->req() > TypeFunc::Parms && 973 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 974 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 975 // MemBarAcquire could be created without Precedent edge. 976 // del_req() replaces the specified edge with the last input edge 977 // and then removes the last edge. If the specified edge > number of 978 // edges the last edge will be moved outside of the input edges array 979 // and the edge will be lost. This is why this code should be 980 // executed only when Precedent (== TypeFunc::Parms) edge is present. 981 Node *x = n->in(TypeFunc::Parms); 982 if (x != NULL && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) { 983 // Old edge to node within same block will get removed, but no precedence 984 // edge will get added because it already exists. Update ready count. 985 int cnt = ready_cnt.at(n->_idx); 986 assert(cnt > 1, "MemBar node %d must not get ready here", n->_idx); 987 ready_cnt.at_put(n->_idx, cnt-1); 988 } 989 n->del_req(TypeFunc::Parms); 990 n->add_prec(x); 991 } 992 } 993 } 994 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 995 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 996 997 // All the prescheduled guys do not hold back internal nodes 998 uint i3; 999 for (i3 = 0; i3 < phi_cnt; i3++) { // For all pre-scheduled 1000 Node *n = block->get_node(i3); // Get pre-scheduled 1001 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1002 Node* m = n->fast_out(j); 1003 if (get_block_for_node(m) == block) { // Local-block user 1004 int m_cnt = ready_cnt.at(m->_idx)-1; 1005 if (OptoRegScheduling && block_size_threshold_ok) { 1006 // mark m as scheduled 1007 if (m_cnt < 0) { 1008 m->add_flag(Node::Flag_is_scheduled); 1009 } 1010 } 1011 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 1012 } 1013 } 1014 } 1015 1016 Node_List delay; 1017 // Make a worklist 1018 Node_List worklist; 1019 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 1020 Node *m = block->get_node(i4); 1021 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 1022 if (m->is_iteratively_computed()) { 1023 // Push induction variable increments last to allow other uses 1024 // of the phi to be scheduled first. The select() method breaks 1025 // ties in scheduling by worklist order. 1026 delay.push(m); 1027 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 1028 // Force the CreateEx to the top of the list so it's processed 1029 // first and ends up at the start of the block. 1030 worklist.insert(0, m); 1031 } else { 1032 worklist.push(m); // Then on to worklist! 1033 } 1034 } 1035 } 1036 while (delay.size()) { 1037 Node* d = delay.pop(); 1038 worklist.push(d); 1039 } 1040 1041 if (OptoRegScheduling && block_size_threshold_ok) { 1042 // To stage register pressure calculations we need to examine the live set variables 1043 // breaking them up by register class to compartmentalize the calculations. 1044 uint float_pressure = Matcher::float_pressure(FLOATPRESSURE); 1045 _regalloc->_sched_int_pressure.init(INTPRESSURE); 1046 _regalloc->_sched_float_pressure.init(float_pressure); 1047 _regalloc->_scratch_int_pressure.init(INTPRESSURE); 1048 _regalloc->_scratch_float_pressure.init(float_pressure); 1049 1050 _regalloc->compute_entry_block_pressure(block); 1051 } 1052 1053 // Warm up the 'next_call' heuristic bits 1054 needed_for_next_call(block, block->head(), next_call); 1055 1056 #ifndef PRODUCT 1057 if (trace_opto_pipelining()) { 1058 for (uint j=0; j< block->number_of_nodes(); j++) { 1059 Node *n = block->get_node(j); 1060 int idx = n->_idx; 1061 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 1062 tty->print("latency:%3d ", get_latency_for_node(n)); 1063 tty->print("%4d: %s\n", idx, n->Name()); 1064 } 1065 } 1066 #endif 1067 1068 uint max_idx = (uint)ready_cnt.length(); 1069 // Pull from worklist and schedule 1070 while( worklist.size() ) { // Worklist is not ready 1071 1072 #ifndef PRODUCT 1073 if (trace_opto_pipelining()) { 1074 tty->print("# ready list:"); 1075 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1076 Node *n = worklist[i]; // Get Node on worklist 1077 tty->print(" %d", n->_idx); 1078 } 1079 tty->cr(); 1080 } 1081 #endif 1082 1083 // Select and pop a ready guy from worklist 1084 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes); 1085 block->map_node(n, phi_cnt++); // Schedule him next 1086 1087 if (OptoRegScheduling && block_size_threshold_ok) { 1088 n->add_flag(Node::Flag_is_scheduled); 1089 1090 // Now adjust the resister pressure with the node we selected 1091 if (!n->is_Phi()) { 1092 adjust_register_pressure(n, block, recalc_pressure_nodes, true); 1093 } 1094 } 1095 1096 #ifndef PRODUCT 1097 if (trace_opto_pipelining()) { 1098 tty->print("# select %d: %s", n->_idx, n->Name()); 1099 tty->print(", latency:%d", get_latency_for_node(n)); 1100 n->dump(); 1101 if (Verbose) { 1102 tty->print("# ready list:"); 1103 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 1104 Node *n = worklist[i]; // Get Node on worklist 1105 tty->print(" %d", n->_idx); 1106 } 1107 tty->cr(); 1108 } 1109 } 1110 1111 #endif 1112 if( n->is_MachCall() ) { 1113 MachCallNode *mcall = n->as_MachCall(); 1114 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 1115 continue; 1116 } 1117 1118 if (n->is_Mach() && n->as_Mach()->has_call()) { 1119 RegMask regs; 1120 regs.Insert(_matcher.c_frame_pointer()); 1121 regs.OR(n->out_RegMask()); 1122 1123 MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 1124 map_node_to_block(proj, block); 1125 block->insert_node(proj, phi_cnt++); 1126 1127 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 1128 } 1129 1130 // Children are now all ready 1131 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 1132 Node* m = n->fast_out(i5); // Get user 1133 if (get_block_for_node(m) != block) { 1134 continue; 1135 } 1136 if( m->is_Phi() ) continue; 1137 if (m->_idx >= max_idx) { // new node, skip it 1138 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 1139 continue; 1140 } 1141 int m_cnt = ready_cnt.at(m->_idx) - 1; 1142 ready_cnt.at_put(m->_idx, m_cnt); 1143 if( m_cnt == 0 ) 1144 worklist.push(m); 1145 } 1146 } 1147 1148 if( phi_cnt != block->end_idx() ) { 1149 // did not schedule all. Retry, Bailout, or Die 1150 if (C->subsume_loads() == true && !C->failing()) { 1151 // Retry with subsume_loads == false 1152 // If this is the first failure, the sentinel string will "stick" 1153 // to the Compile object, and the C2Compiler will see it and retry. 1154 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 1155 } else { 1156 assert(false, "graph should be schedulable"); 1157 } 1158 // assert( phi_cnt == end_idx(), "did not schedule all" ); 1159 return false; 1160 } 1161 1162 if (OptoRegScheduling && block_size_threshold_ok) { 1163 _regalloc->compute_exit_block_pressure(block); 1164 block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure(); 1165 block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure(); 1166 } 1167 1168 #ifndef PRODUCT 1169 if (trace_opto_pipelining()) { 1170 tty->print_cr("#"); 1171 tty->print_cr("# after schedule_local"); 1172 for (uint i = 0;i < block->number_of_nodes();i++) { 1173 tty->print("# "); 1174 block->get_node(i)->fast_dump(); 1175 } 1176 tty->print_cr("# "); 1177 1178 if (OptoRegScheduling && block_size_threshold_ok) { 1179 tty->print_cr("# pressure info : %d", block->_pre_order); 1180 _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info"); 1181 _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info"); 1182 } 1183 tty->cr(); 1184 } 1185 #endif 1186 1187 return true; 1188 } 1189 1190 //--------------------------catch_cleanup_fix_all_inputs----------------------- 1191 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 1192 for (uint l = 0; l < use->len(); l++) { 1193 if (use->in(l) == old_def) { 1194 if (l < use->req()) { 1195 use->set_req(l, new_def); 1196 } else { 1197 use->rm_prec(l); 1198 use->add_prec(new_def); 1199 l--; 1200 } 1201 } 1202 } 1203 } 1204 1205 //------------------------------catch_cleanup_find_cloned_def------------------ 1206 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1207 assert( use_blk != def_blk, "Inter-block cleanup only"); 1208 1209 // The use is some block below the Catch. Find and return the clone of the def 1210 // that dominates the use. If there is no clone in a dominating block, then 1211 // create a phi for the def in a dominating block. 1212 1213 // Find which successor block dominates this use. The successor 1214 // blocks must all be single-entry (from the Catch only; I will have 1215 // split blocks to make this so), hence they all dominate. 1216 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 1217 use_blk = use_blk->_idom; 1218 1219 // Find the successor 1220 Node *fixup = NULL; 1221 1222 uint j; 1223 for( j = 0; j < def_blk->_num_succs; j++ ) 1224 if( use_blk == def_blk->_succs[j] ) 1225 break; 1226 1227 if( j == def_blk->_num_succs ) { 1228 // Block at same level in dom-tree is not a successor. It needs a 1229 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 1230 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 1231 for(uint k = 1; k < use_blk->num_preds(); k++) { 1232 Block* block = get_block_for_node(use_blk->pred(k)); 1233 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 1234 } 1235 1236 // Check to see if the use_blk already has an identical phi inserted. 1237 // If it exists, it will be at the first position since all uses of a 1238 // def are processed together. 1239 Node *phi = use_blk->get_node(1); 1240 if( phi->is_Phi() ) { 1241 fixup = phi; 1242 for (uint k = 1; k < use_blk->num_preds(); k++) { 1243 if (phi->in(k) != inputs[k]) { 1244 // Not a match 1245 fixup = NULL; 1246 break; 1247 } 1248 } 1249 } 1250 1251 // If an existing PhiNode was not found, make a new one. 1252 if (fixup == NULL) { 1253 Node *new_phi = PhiNode::make(use_blk->head(), def); 1254 use_blk->insert_node(new_phi, 1); 1255 map_node_to_block(new_phi, use_blk); 1256 for (uint k = 1; k < use_blk->num_preds(); k++) { 1257 new_phi->set_req(k, inputs[k]); 1258 } 1259 fixup = new_phi; 1260 } 1261 1262 } else { 1263 // Found the use just below the Catch. Make it use the clone. 1264 fixup = use_blk->get_node(n_clone_idx); 1265 } 1266 1267 return fixup; 1268 } 1269 1270 //--------------------------catch_cleanup_intra_block-------------------------- 1271 // Fix all input edges in use that reference "def". The use is in the same 1272 // block as the def and both have been cloned in each successor block. 1273 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1274 1275 // Both the use and def have been cloned. For each successor block, 1276 // get the clone of the use, and make its input the clone of the def 1277 // found in that block. 1278 1279 uint use_idx = blk->find_node(use); 1280 uint offset_idx = use_idx - beg; 1281 for( uint k = 0; k < blk->_num_succs; k++ ) { 1282 // Get clone in each successor block 1283 Block *sb = blk->_succs[k]; 1284 Node *clone = sb->get_node(offset_idx+1); 1285 assert( clone->Opcode() == use->Opcode(), "" ); 1286 1287 // Make use-clone reference the def-clone 1288 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1289 } 1290 } 1291 1292 //------------------------------catch_cleanup_inter_block--------------------- 1293 // Fix all input edges in use that reference "def". The use is in a different 1294 // block than the def. 1295 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1296 if( !use_blk ) return; // Can happen if the use is a precedence edge 1297 1298 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1299 catch_cleanup_fix_all_inputs(use, def, new_def); 1300 } 1301 1302 //------------------------------call_catch_cleanup----------------------------- 1303 // If we inserted any instructions between a Call and his CatchNode, 1304 // clone the instructions on all paths below the Catch. 1305 void PhaseCFG::call_catch_cleanup(Block* block) { 1306 1307 // End of region to clone 1308 uint end = block->end_idx(); 1309 if( !block->get_node(end)->is_Catch() ) return; 1310 // Start of region to clone 1311 uint beg = end; 1312 while(!block->get_node(beg-1)->is_MachProj() || 1313 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1314 beg--; 1315 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1316 } 1317 // Range of inserted instructions is [beg, end) 1318 if( beg == end ) return; 1319 1320 // Clone along all Catch output paths. Clone area between the 'beg' and 1321 // 'end' indices. 1322 for( uint i = 0; i < block->_num_succs; i++ ) { 1323 Block *sb = block->_succs[i]; 1324 // Clone the entire area; ignoring the edge fixup for now. 1325 for( uint j = end; j > beg; j-- ) { 1326 Node *clone = block->get_node(j-1)->clone(); 1327 sb->insert_node(clone, 1); 1328 map_node_to_block(clone, sb); 1329 if (clone->needs_anti_dependence_check()) { 1330 insert_anti_dependences(sb, clone); 1331 } 1332 } 1333 } 1334 1335 1336 // Fixup edges. Check the def-use info per cloned Node 1337 for(uint i2 = beg; i2 < end; i2++ ) { 1338 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1339 Node *n = block->get_node(i2); // Node that got cloned 1340 // Need DU safe iterator because of edge manipulation in calls. 1341 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1342 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1343 out->push(n->fast_out(j1)); 1344 } 1345 uint max = out->size(); 1346 for (uint j = 0; j < max; j++) {// For all users 1347 Node *use = out->pop(); 1348 Block *buse = get_block_for_node(use); 1349 if( use->is_Phi() ) { 1350 for( uint k = 1; k < use->req(); k++ ) 1351 if( use->in(k) == n ) { 1352 Block* b = get_block_for_node(buse->pred(k)); 1353 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1354 use->set_req(k, fixup); 1355 } 1356 } else { 1357 if (block == buse) { 1358 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1359 } else { 1360 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1361 } 1362 } 1363 } // End for all users 1364 1365 } // End of for all Nodes in cloned area 1366 1367 // Remove the now-dead cloned ops 1368 for(uint i3 = beg; i3 < end; i3++ ) { 1369 block->get_node(beg)->disconnect_inputs(NULL, C); 1370 block->remove_node(beg); 1371 } 1372 1373 // If the successor blocks have a CreateEx node, move it back to the top 1374 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1375 Block *sb = block->_succs[i4]; 1376 uint new_cnt = end - beg; 1377 // Remove any newly created, but dead, nodes. 1378 for( uint j = new_cnt; j > 0; j-- ) { 1379 Node *n = sb->get_node(j); 1380 if (n->outcnt() == 0 && 1381 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1382 n->disconnect_inputs(NULL, C); 1383 sb->remove_node(j); 1384 new_cnt--; 1385 } 1386 } 1387 // If any newly created nodes remain, move the CreateEx node to the top 1388 if (new_cnt > 0) { 1389 Node *cex = sb->get_node(1+new_cnt); 1390 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1391 sb->remove_node(1+new_cnt); 1392 sb->insert_node(cex, 1); 1393 } 1394 } 1395 } 1396 }