1 /* 2 * Copyright (c) 1998, 2013, 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/block.hpp" 28 #include "opto/c2compiler.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/cfgnode.hpp" 31 #include "opto/machnode.hpp" 32 #include "opto/runtime.hpp" 33 #ifdef TARGET_ARCH_MODEL_x86_32 34 # include "adfiles/ad_x86_32.hpp" 35 #endif 36 #ifdef TARGET_ARCH_MODEL_x86_64 37 # include "adfiles/ad_x86_64.hpp" 38 #endif 39 #ifdef TARGET_ARCH_MODEL_sparc 40 # include "adfiles/ad_sparc.hpp" 41 #endif 42 #ifdef TARGET_ARCH_MODEL_zero 43 # include "adfiles/ad_zero.hpp" 44 #endif 45 #ifdef TARGET_ARCH_MODEL_arm 46 # include "adfiles/ad_arm.hpp" 47 #endif 48 #ifdef TARGET_ARCH_MODEL_ppc 49 # include "adfiles/ad_ppc.hpp" 50 #endif 51 52 // Optimization - Graph Style 53 54 //------------------------------implicit_null_check---------------------------- 55 // Detect implicit-null-check opportunities. Basically, find NULL checks 56 // with suitable memory ops nearby. Use the memory op to do the NULL check. 57 // I can generate a memory op if there is not one nearby. 58 // The proj is the control projection for the not-null case. 59 // The val is the pointer being checked for nullness or 60 // decodeHeapOop_not_null node if it did not fold into address. 61 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { 62 // Assume if null check need for 0 offset then always needed 63 // Intel solaris doesn't support any null checks yet and no 64 // mechanism exists (yet) to set the switches at an os_cpu level 65 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 66 67 // Make sure the ptr-is-null path appears to be uncommon! 68 float f = block->end()->as_MachIf()->_prob; 69 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 70 if( f > PROB_UNLIKELY_MAG(4) ) return; 71 72 uint bidx = 0; // Capture index of value into memop 73 bool was_store; // Memory op is a store op 74 75 // Get the successor block for if the test ptr is non-null 76 Block* not_null_block; // this one goes with the proj 77 Block* null_block; 78 if (block->get_node(block->number_of_nodes()-1) == proj) { 79 null_block = block->_succs[0]; 80 not_null_block = block->_succs[1]; 81 } else { 82 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other"); 83 not_null_block = block->_succs[0]; 84 null_block = block->_succs[1]; 85 } 86 while (null_block->is_Empty() == Block::empty_with_goto) { 87 null_block = null_block->_succs[0]; 88 } 89 90 // Search the exception block for an uncommon trap. 91 // (See Parse::do_if and Parse::do_ifnull for the reason 92 // we need an uncommon trap. Briefly, we need a way to 93 // detect failure of this optimization, as in 6366351.) 94 { 95 bool found_trap = false; 96 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { 97 Node* nn = null_block->get_node(i1); 98 if (nn->is_MachCall() && 99 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 100 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 101 if (trtype->isa_int() && trtype->is_int()->is_con()) { 102 jint tr_con = trtype->is_int()->get_con(); 103 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 104 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 105 assert((int)reason < (int)BitsPerInt, "recode bit map"); 106 if (is_set_nth_bit(allowed_reasons, (int) reason) 107 && action != Deoptimization::Action_none) { 108 // This uncommon trap is sure to recompile, eventually. 109 // When that happens, C->too_many_traps will prevent 110 // this transformation from happening again. 111 found_trap = true; 112 } 113 } 114 break; 115 } 116 } 117 if (!found_trap) { 118 // We did not find an uncommon trap. 119 return; 120 } 121 } 122 123 // Check for decodeHeapOop_not_null node which did not fold into address 124 bool is_decoden = ((intptr_t)val) & 1; 125 val = (Node*)(((intptr_t)val) & ~1); 126 127 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 128 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 129 130 // Search the successor block for a load or store who's base value is also 131 // the tested value. There may be several. 132 Node_List *out = new Node_List(Thread::current()->resource_area()); 133 MachNode *best = NULL; // Best found so far 134 for (DUIterator i = val->outs(); val->has_out(i); i++) { 135 Node *m = val->out(i); 136 if( !m->is_Mach() ) continue; 137 MachNode *mach = m->as_Mach(); 138 was_store = false; 139 int iop = mach->ideal_Opcode(); 140 switch( iop ) { 141 case Op_LoadB: 142 case Op_LoadUB: 143 case Op_LoadUS: 144 case Op_LoadD: 145 case Op_LoadF: 146 case Op_LoadI: 147 case Op_LoadL: 148 case Op_LoadP: 149 case Op_LoadN: 150 case Op_LoadS: 151 case Op_LoadKlass: 152 case Op_LoadNKlass: 153 case Op_LoadRange: 154 case Op_LoadD_unaligned: 155 case Op_LoadL_unaligned: 156 assert(mach->in(2) == val, "should be address"); 157 break; 158 case Op_StoreB: 159 case Op_StoreC: 160 case Op_StoreCM: 161 case Op_StoreD: 162 case Op_StoreF: 163 case Op_StoreI: 164 case Op_StoreL: 165 case Op_StoreP: 166 case Op_StoreN: 167 case Op_StoreNKlass: 168 was_store = true; // Memory op is a store op 169 // Stores will have their address in slot 2 (memory in slot 1). 170 // If the value being nul-checked is in another slot, it means we 171 // are storing the checked value, which does NOT check the value! 172 if( mach->in(2) != val ) continue; 173 break; // Found a memory op? 174 case Op_StrComp: 175 case Op_StrEquals: 176 case Op_StrIndexOf: 177 case Op_AryEq: 178 case Op_EncodeISOArray: 179 // Not a legit memory op for implicit null check regardless of 180 // embedded loads 181 continue; 182 default: // Also check for embedded loads 183 if( !mach->needs_anti_dependence_check() ) 184 continue; // Not an memory op; skip it 185 if( must_clone[iop] ) { 186 // Do not move nodes which produce flags because 187 // RA will try to clone it to place near branch and 188 // it will cause recompilation, see clone_node(). 189 continue; 190 } 191 { 192 // Check that value is used in memory address in 193 // instructions with embedded load (CmpP val1,(val2+off)). 194 Node* base; 195 Node* index; 196 const MachOper* oper = mach->memory_inputs(base, index); 197 if (oper == NULL || oper == (MachOper*)-1) { 198 continue; // Not an memory op; skip it 199 } 200 if (val == base || 201 val == index && val->bottom_type()->isa_narrowoop()) { 202 break; // Found it 203 } else { 204 continue; // Skip it 205 } 206 } 207 break; 208 } 209 // check if the offset is not too high for implicit exception 210 { 211 intptr_t offset = 0; 212 const TypePtr *adr_type = NULL; // Do not need this return value here 213 const Node* base = mach->get_base_and_disp(offset, adr_type); 214 if (base == NULL || base == NodeSentinel) { 215 // Narrow oop address doesn't have base, only index 216 if( val->bottom_type()->isa_narrowoop() && 217 MacroAssembler::needs_explicit_null_check(offset) ) 218 continue; // Give up if offset is beyond page size 219 // cannot reason about it; is probably not implicit null exception 220 } else { 221 const TypePtr* tptr; 222 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 223 Universe::narrow_klass_shift() == 0)) { 224 // 32-bits narrow oop can be the base of address expressions 225 tptr = base->get_ptr_type(); 226 } else { 227 // only regular oops are expected here 228 tptr = base->bottom_type()->is_ptr(); 229 } 230 // Give up if offset is not a compile-time constant 231 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 232 continue; 233 offset += tptr->_offset; // correct if base is offseted 234 if( MacroAssembler::needs_explicit_null_check(offset) ) 235 continue; // Give up is reference is beyond 4K page size 236 } 237 } 238 239 // Check ctrl input to see if the null-check dominates the memory op 240 Block *cb = get_block_for_node(mach); 241 cb = cb->_idom; // Always hoist at least 1 block 242 if( !was_store ) { // Stores can be hoisted only one block 243 while( cb->_dom_depth > (block->_dom_depth + 1)) 244 cb = cb->_idom; // Hoist loads as far as we want 245 // The non-null-block should dominate the memory op, too. Live 246 // range spilling will insert a spill in the non-null-block if it is 247 // needs to spill the memory op for an implicit null check. 248 if (cb->_dom_depth == (block->_dom_depth + 1)) { 249 if (cb != not_null_block) continue; 250 cb = cb->_idom; 251 } 252 } 253 if( cb != block ) continue; 254 255 // Found a memory user; see if it can be hoisted to check-block 256 uint vidx = 0; // Capture index of value into memop 257 uint j; 258 for( j = mach->req()-1; j > 0; j-- ) { 259 if( mach->in(j) == val ) { 260 vidx = j; 261 // Ignore DecodeN val which could be hoisted to where needed. 262 if( is_decoden ) continue; 263 } 264 // Block of memory-op input 265 Block *inb = get_block_for_node(mach->in(j)); 266 Block *b = block; // Start from nul check 267 while( b != inb && b->_dom_depth > inb->_dom_depth ) 268 b = b->_idom; // search upwards for input 269 // See if input dominates null check 270 if( b != inb ) 271 break; 272 } 273 if( j > 0 ) 274 continue; 275 Block *mb = get_block_for_node(mach); 276 // Hoisting stores requires more checks for the anti-dependence case. 277 // Give up hoisting if we have to move the store past any load. 278 if( was_store ) { 279 Block *b = mb; // Start searching here for a local load 280 // mach use (faulting) trying to hoist 281 // n might be blocker to hoisting 282 while( b != block ) { 283 uint k; 284 for( k = 1; k < b->number_of_nodes(); k++ ) { 285 Node *n = b->get_node(k); 286 if( n->needs_anti_dependence_check() && 287 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 288 break; // Found anti-dependent load 289 } 290 if( k < b->number_of_nodes() ) 291 break; // Found anti-dependent load 292 // Make sure control does not do a merge (would have to check allpaths) 293 if( b->num_preds() != 2 ) break; 294 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 295 } 296 if( b != block ) continue; 297 } 298 299 // Make sure this memory op is not already being used for a NullCheck 300 Node *e = mb->end(); 301 if( e->is_MachNullCheck() && e->in(1) == mach ) 302 continue; // Already being used as a NULL check 303 304 // Found a candidate! Pick one with least dom depth - the highest 305 // in the dom tree should be closest to the null check. 306 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 307 best = mach; 308 bidx = vidx; 309 } 310 } 311 // No candidate! 312 if (best == NULL) { 313 return; 314 } 315 316 // ---- Found an implicit null check 317 extern int implicit_null_checks; 318 implicit_null_checks++; 319 320 if( is_decoden ) { 321 // Check if we need to hoist decodeHeapOop_not_null first. 322 Block *valb = get_block_for_node(val); 323 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 324 // Hoist it up to the end of the test block. 325 valb->find_remove(val); 326 block->add_inst(val); 327 map_node_to_block(val, block); 328 // DecodeN on x86 may kill flags. Check for flag-killing projections 329 // that also need to be hoisted. 330 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 331 Node* n = val->fast_out(j); 332 if( n->is_MachProj() ) { 333 get_block_for_node(n)->find_remove(n); 334 block->add_inst(n); 335 map_node_to_block(n, block); 336 } 337 } 338 } 339 } 340 // Hoist the memory candidate up to the end of the test block. 341 Block *old_block = get_block_for_node(best); 342 old_block->find_remove(best); 343 block->add_inst(best); 344 map_node_to_block(best, block); 345 346 // Move the control dependence 347 if (best->in(0) && best->in(0) == old_block->head()) 348 best->set_req(0, block->head()); 349 350 // Check for flag-killing projections that also need to be hoisted 351 // Should be DU safe because no edge updates. 352 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 353 Node* n = best->fast_out(j); 354 if( n->is_MachProj() ) { 355 get_block_for_node(n)->find_remove(n); 356 block->add_inst(n); 357 map_node_to_block(n, block); 358 } 359 } 360 361 // proj==Op_True --> ne test; proj==Op_False --> eq test. 362 // One of two graph shapes got matched: 363 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 364 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 365 // NULL checks are always branch-if-eq. If we see a IfTrue projection 366 // then we are replacing a 'ne' test with a 'eq' NULL check test. 367 // We need to flip the projections to keep the same semantics. 368 if( proj->Opcode() == Op_IfTrue ) { 369 // Swap order of projections in basic block to swap branch targets 370 Node *tmp1 = block->get_node(block->end_idx()+1); 371 Node *tmp2 = block->get_node(block->end_idx()+2); 372 block->map_node(tmp2, block->end_idx()+1); 373 block->map_node(tmp1, block->end_idx()+2); 374 Node *tmp = new (C) Node(C->top()); // Use not NULL input 375 tmp1->replace_by(tmp); 376 tmp2->replace_by(tmp1); 377 tmp->replace_by(tmp2); 378 tmp->destruct(); 379 } 380 381 // Remove the existing null check; use a new implicit null check instead. 382 // Since schedule-local needs precise def-use info, we need to correct 383 // it as well. 384 Node *old_tst = proj->in(0); 385 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); 386 block->map_node(nul_chk, block->end_idx()); 387 map_node_to_block(nul_chk, block); 388 // Redirect users of old_test to nul_chk 389 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 390 old_tst->last_out(i2)->set_req(0, nul_chk); 391 // Clean-up any dead code 392 for (uint i3 = 0; i3 < old_tst->req(); i3++) 393 old_tst->set_req(i3, NULL); 394 395 latency_from_uses(nul_chk); 396 latency_from_uses(best); 397 } 398 399 400 //------------------------------select----------------------------------------- 401 // Select a nice fellow from the worklist to schedule next. If there is only 402 // one choice, then use it. Projections take top priority for correctness 403 // reasons - if I see a projection, then it is next. There are a number of 404 // other special cases, for instructions that consume condition codes, et al. 405 // These are chosen immediately. Some instructions are required to immediately 406 // precede the last instruction in the block, and these are taken last. Of the 407 // remaining cases (most), choose the instruction with the greatest latency 408 // (that is, the most number of pseudo-cycles required to the end of the 409 // routine). If there is a tie, choose the instruction with the most inputs. 410 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 411 412 // If only a single entry on the stack, use it 413 uint cnt = worklist.size(); 414 if (cnt == 1) { 415 Node *n = worklist[0]; 416 worklist.map(0,worklist.pop()); 417 return n; 418 } 419 420 uint choice = 0; // Bigger is most important 421 uint latency = 0; // Bigger is scheduled first 422 uint score = 0; // Bigger is better 423 int idx = -1; // Index in worklist 424 int cand_cnt = 0; // Candidate count 425 426 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 427 // Order in worklist is used to break ties. 428 // See caller for how this is used to delay scheduling 429 // of induction variable increments to after the other 430 // uses of the phi are scheduled. 431 Node *n = worklist[i]; // Get Node on worklist 432 433 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 434 if( n->is_Proj() || // Projections always win 435 n->Opcode()== Op_Con || // So does constant 'Top' 436 iop == Op_CreateEx || // Create-exception must start block 437 iop == Op_CheckCastPP 438 ) { 439 worklist.map(i,worklist.pop()); 440 return n; 441 } 442 443 // Final call in a block must be adjacent to 'catch' 444 Node *e = block->end(); 445 if( e->is_Catch() && e->in(0)->in(0) == n ) 446 continue; 447 448 // Memory op for an implicit null check has to be at the end of the block 449 if( e->is_MachNullCheck() && e->in(1) == n ) 450 continue; 451 452 // Schedule IV increment last. 453 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 454 e->in(1)->in(1) == n && n->is_iteratively_computed()) 455 continue; 456 457 uint n_choice = 2; 458 459 // See if this instruction is consumed by a branch. If so, then (as the 460 // branch is the last instruction in the basic block) force it to the 461 // end of the basic block 462 if ( must_clone[iop] ) { 463 // See if any use is a branch 464 bool found_machif = false; 465 466 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 467 Node* use = n->fast_out(j); 468 469 // The use is a conditional branch, make them adjacent 470 if (use->is_MachIf() && get_block_for_node(use) == block) { 471 found_machif = true; 472 break; 473 } 474 475 // For nodes that produce a FlagsProj, make the node adjacent to the 476 // use of the FlagsProj 477 if (use->is_FlagsProj() && get_block_for_node(use) == block) { 478 found_machif = true; 479 break; 480 } 481 482 // More than this instruction pending for successor to be ready, 483 // don't choose this if other opportunities are ready 484 if (ready_cnt.at(use->_idx) > 1) 485 n_choice = 1; 486 } 487 488 // loop terminated, prefer not to use this instruction 489 if (found_machif) 490 continue; 491 } 492 493 // See if this has a predecessor that is "must_clone", i.e. sets the 494 // condition code. If so, choose this first 495 for (uint j = 0; j < n->req() ; j++) { 496 Node *inn = n->in(j); 497 if (inn) { 498 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 499 n_choice = 3; 500 break; 501 } 502 } 503 } 504 505 // MachTemps should be scheduled last so they are near their uses 506 if (n->is_MachTemp()) { 507 n_choice = 1; 508 } 509 510 uint n_latency = get_latency_for_node(n); 511 uint n_score = n->req(); // Many inputs get high score to break ties 512 513 // Keep best latency found 514 cand_cnt++; 515 if (choice < n_choice || 516 (choice == n_choice && 517 ((StressLCM && Compile::randomized_select(cand_cnt)) || 518 (!StressLCM && 519 (latency < n_latency || 520 (latency == n_latency && 521 (score < n_score))))))) { 522 choice = n_choice; 523 latency = n_latency; 524 score = n_score; 525 idx = i; // Also keep index in worklist 526 } 527 } // End of for all ready nodes in worklist 528 529 assert(idx >= 0, "index should be set"); 530 Node *n = worklist[(uint)idx]; // Get the winner 531 532 worklist.map((uint)idx, worklist.pop()); // Compress worklist 533 return n; 534 } 535 536 537 //------------------------------set_next_call---------------------------------- 538 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 539 if( next_call.test_set(n->_idx) ) return; 540 for( uint i=0; i<n->len(); i++ ) { 541 Node *m = n->in(i); 542 if( !m ) continue; // must see all nodes in block that precede call 543 if (get_block_for_node(m) == block) { 544 set_next_call(block, m, next_call); 545 } 546 } 547 } 548 549 //------------------------------needed_for_next_call--------------------------- 550 // Set the flag 'next_call' for each Node that is needed for the next call to 551 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 552 // next subroutine call get priority - basically it moves things NOT needed 553 // for the next call till after the call. This prevents me from trying to 554 // carry lots of stuff live across a call. 555 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 556 // Find the next control-defining Node in this block 557 Node* call = NULL; 558 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 559 Node* m = this_call->fast_out(i); 560 if (get_block_for_node(m) == block && // Local-block user 561 m != this_call && // Not self-start node 562 m->is_MachCall()) { 563 call = m; 564 break; 565 } 566 } 567 if (call == NULL) return; // No next call (e.g., block end is near) 568 // Set next-call for all inputs to this call 569 set_next_call(block, call, next_call); 570 } 571 572 //------------------------------add_call_kills------------------------------------- 573 // helper function that adds caller save registers to MachProjNode 574 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 575 // Fill in the kill mask for the call 576 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 577 if( !regs.Member(r) ) { // Not already defined by the call 578 // Save-on-call register? 579 if ((save_policy[r] == 'C') || 580 (save_policy[r] == 'A') || 581 ((save_policy[r] == 'E') && exclude_soe)) { 582 proj->_rout.Insert(r); 583 } 584 } 585 } 586 } 587 588 589 //------------------------------sched_call------------------------------------- 590 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 591 RegMask regs; 592 593 // Schedule all the users of the call right now. All the users are 594 // projection Nodes, so they must be scheduled next to the call. 595 // Collect all the defined registers. 596 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 597 Node* n = mcall->fast_out(i); 598 assert( n->is_MachProj(), "" ); 599 int n_cnt = ready_cnt.at(n->_idx)-1; 600 ready_cnt.at_put(n->_idx, n_cnt); 601 assert( n_cnt == 0, "" ); 602 // Schedule next to call 603 block->map_node(n, node_cnt++); 604 // Collect defined registers 605 regs.OR(n->out_RegMask()); 606 // Check for scheduling the next control-definer 607 if( n->bottom_type() == Type::CONTROL ) 608 // Warm up next pile of heuristic bits 609 needed_for_next_call(block, n, next_call); 610 611 // Children of projections are now all ready 612 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 613 Node* m = n->fast_out(j); // Get user 614 if(get_block_for_node(m) != block) { 615 continue; 616 } 617 if( m->is_Phi() ) continue; 618 int m_cnt = ready_cnt.at(m->_idx)-1; 619 ready_cnt.at_put(m->_idx, m_cnt); 620 if( m_cnt == 0 ) 621 worklist.push(m); 622 } 623 624 } 625 626 // Act as if the call defines the Frame Pointer. 627 // Certainly the FP is alive and well after the call. 628 regs.Insert(_matcher.c_frame_pointer()); 629 630 // Set all registers killed and not already defined by the call. 631 uint r_cnt = mcall->tf()->range()->cnt(); 632 int op = mcall->ideal_Opcode(); 633 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 634 map_node_to_block(proj, block); 635 block->insert_node(proj, node_cnt++); 636 637 // Select the right register save policy. 638 const char * save_policy; 639 switch (op) { 640 case Op_CallRuntime: 641 case Op_CallLeaf: 642 case Op_CallLeafNoFP: 643 // Calling C code so use C calling convention 644 save_policy = _matcher._c_reg_save_policy; 645 break; 646 647 case Op_CallStaticJava: 648 case Op_CallDynamicJava: 649 // Calling Java code so use Java calling convention 650 save_policy = _matcher._register_save_policy; 651 break; 652 653 default: 654 ShouldNotReachHere(); 655 } 656 657 // When using CallRuntime mark SOE registers as killed by the call 658 // so values that could show up in the RegisterMap aren't live in a 659 // callee saved register since the register wouldn't know where to 660 // find them. CallLeaf and CallLeafNoFP are ok because they can't 661 // have debug info on them. Strictly speaking this only needs to be 662 // done for oops since idealreg2debugmask takes care of debug info 663 // references but there no way to handle oops differently than other 664 // pointers as far as the kill mask goes. 665 bool exclude_soe = op == Op_CallRuntime; 666 667 // If the call is a MethodHandle invoke, we need to exclude the 668 // register which is used to save the SP value over MH invokes from 669 // the mask. Otherwise this register could be used for 670 // deoptimization information. 671 if (op == Op_CallStaticJava) { 672 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 673 if (mcallstaticjava->_method_handle_invoke) 674 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 675 } 676 677 add_call_kills(proj, regs, save_policy, exclude_soe); 678 679 return node_cnt; 680 } 681 682 683 //------------------------------schedule_local--------------------------------- 684 // Topological sort within a block. Someday become a real scheduler. 685 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) { 686 // Already "sorted" are the block start Node (as the first entry), and 687 // the block-ending Node and any trailing control projections. We leave 688 // these alone. PhiNodes and ParmNodes are made to follow the block start 689 // Node. Everything else gets topo-sorted. 690 691 #ifndef PRODUCT 692 if (trace_opto_pipelining()) { 693 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 694 for (uint i = 0;i < block->number_of_nodes(); i++) { 695 tty->print("# "); 696 block->get_node(i)->fast_dump(); 697 } 698 tty->print_cr("#"); 699 } 700 #endif 701 702 // RootNode is already sorted 703 if (block->number_of_nodes() == 1) { 704 return true; 705 } 706 707 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 708 uint node_cnt = block->end_idx(); 709 uint phi_cnt = 1; 710 uint i; 711 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 712 Node *n = block->get_node(i); 713 if( n->is_Phi() || // Found a PhiNode or ParmNode 714 (n->is_Proj() && n->in(0) == block->head()) ) { 715 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 716 block->map_node(block->get_node(phi_cnt), i); 717 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 718 } else { // All others 719 // Count block-local inputs to 'n' 720 uint cnt = n->len(); // Input count 721 uint local = 0; 722 for( uint j=0; j<cnt; j++ ) { 723 Node *m = n->in(j); 724 if( m && get_block_for_node(m) == block && !m->is_top() ) 725 local++; // One more block-local input 726 } 727 ready_cnt.at_put(n->_idx, local); // Count em up 728 729 #ifdef ASSERT 730 if( UseConcMarkSweepGC || UseG1GC ) { 731 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 732 // Check the precedence edges 733 for (uint prec = n->req(); prec < n->len(); prec++) { 734 Node* oop_store = n->in(prec); 735 if (oop_store != NULL) { 736 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 737 } 738 } 739 } 740 } 741 #endif 742 743 // A few node types require changing a required edge to a precedence edge 744 // before allocation. 745 if( n->is_Mach() && n->req() > TypeFunc::Parms && 746 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 747 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 748 // MemBarAcquire could be created without Precedent edge. 749 // del_req() replaces the specified edge with the last input edge 750 // and then removes the last edge. If the specified edge > number of 751 // edges the last edge will be moved outside of the input edges array 752 // and the edge will be lost. This is why this code should be 753 // executed only when Precedent (== TypeFunc::Parms) edge is present. 754 Node *x = n->in(TypeFunc::Parms); 755 n->del_req(TypeFunc::Parms); 756 n->add_prec(x); 757 } 758 } 759 } 760 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 761 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 762 763 // All the prescheduled guys do not hold back internal nodes 764 uint i3; 765 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 766 Node *n = block->get_node(i3); // Get pre-scheduled 767 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 768 Node* m = n->fast_out(j); 769 if (get_block_for_node(m) == block) { // Local-block user 770 int m_cnt = ready_cnt.at(m->_idx)-1; 771 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 772 } 773 } 774 } 775 776 Node_List delay; 777 // Make a worklist 778 Node_List worklist; 779 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 780 Node *m = block->get_node(i4); 781 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 782 if (m->is_iteratively_computed()) { 783 // Push induction variable increments last to allow other uses 784 // of the phi to be scheduled first. The select() method breaks 785 // ties in scheduling by worklist order. 786 delay.push(m); 787 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 788 // Force the CreateEx to the top of the list so it's processed 789 // first and ends up at the start of the block. 790 worklist.insert(0, m); 791 } else { 792 worklist.push(m); // Then on to worklist! 793 } 794 } 795 } 796 while (delay.size()) { 797 Node* d = delay.pop(); 798 worklist.push(d); 799 } 800 801 // Warm up the 'next_call' heuristic bits 802 needed_for_next_call(block, block->head(), next_call); 803 804 #ifndef PRODUCT 805 if (trace_opto_pipelining()) { 806 for (uint j=0; j< block->number_of_nodes(); j++) { 807 Node *n = block->get_node(j); 808 int idx = n->_idx; 809 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 810 tty->print("latency:%3d ", get_latency_for_node(n)); 811 tty->print("%4d: %s\n", idx, n->Name()); 812 } 813 } 814 #endif 815 816 uint max_idx = (uint)ready_cnt.length(); 817 // Pull from worklist and schedule 818 while( worklist.size() ) { // Worklist is not ready 819 820 #ifndef PRODUCT 821 if (trace_opto_pipelining()) { 822 tty->print("# ready list:"); 823 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 824 Node *n = worklist[i]; // Get Node on worklist 825 tty->print(" %d", n->_idx); 826 } 827 tty->cr(); 828 } 829 #endif 830 831 // Select and pop a ready guy from worklist 832 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt); 833 block->map_node(n, phi_cnt++); // Schedule him next 834 835 #ifndef PRODUCT 836 if (trace_opto_pipelining()) { 837 tty->print("# select %d: %s", n->_idx, n->Name()); 838 tty->print(", latency:%d", get_latency_for_node(n)); 839 n->dump(); 840 if (Verbose) { 841 tty->print("# ready list:"); 842 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 843 Node *n = worklist[i]; // Get Node on worklist 844 tty->print(" %d", n->_idx); 845 } 846 tty->cr(); 847 } 848 } 849 850 #endif 851 if( n->is_MachCall() ) { 852 MachCallNode *mcall = n->as_MachCall(); 853 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 854 continue; 855 } 856 857 if (n->is_Mach() && n->as_Mach()->has_call()) { 858 RegMask regs; 859 regs.Insert(_matcher.c_frame_pointer()); 860 regs.OR(n->out_RegMask()); 861 862 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 863 map_node_to_block(proj, block); 864 block->insert_node(proj, phi_cnt++); 865 866 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 867 } 868 869 // Children are now all ready 870 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 871 Node* m = n->fast_out(i5); // Get user 872 if (get_block_for_node(m) != block) { 873 continue; 874 } 875 if( m->is_Phi() ) continue; 876 if (m->_idx >= max_idx) { // new node, skip it 877 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 878 continue; 879 } 880 int m_cnt = ready_cnt.at(m->_idx)-1; 881 ready_cnt.at_put(m->_idx, m_cnt); 882 if( m_cnt == 0 ) 883 worklist.push(m); 884 } 885 } 886 887 if( phi_cnt != block->end_idx() ) { 888 // did not schedule all. Retry, Bailout, or Die 889 if (C->subsume_loads() == true && !C->failing()) { 890 // Retry with subsume_loads == false 891 // If this is the first failure, the sentinel string will "stick" 892 // to the Compile object, and the C2Compiler will see it and retry. 893 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 894 } 895 // assert( phi_cnt == end_idx(), "did not schedule all" ); 896 return false; 897 } 898 899 #ifndef PRODUCT 900 if (trace_opto_pipelining()) { 901 tty->print_cr("#"); 902 tty->print_cr("# after schedule_local"); 903 for (uint i = 0;i < block->number_of_nodes();i++) { 904 tty->print("# "); 905 block->get_node(i)->fast_dump(); 906 } 907 tty->cr(); 908 } 909 #endif 910 911 912 return true; 913 } 914 915 //--------------------------catch_cleanup_fix_all_inputs----------------------- 916 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 917 for (uint l = 0; l < use->len(); l++) { 918 if (use->in(l) == old_def) { 919 if (l < use->req()) { 920 use->set_req(l, new_def); 921 } else { 922 use->rm_prec(l); 923 use->add_prec(new_def); 924 l--; 925 } 926 } 927 } 928 } 929 930 //------------------------------catch_cleanup_find_cloned_def------------------ 931 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 932 assert( use_blk != def_blk, "Inter-block cleanup only"); 933 934 // The use is some block below the Catch. Find and return the clone of the def 935 // that dominates the use. If there is no clone in a dominating block, then 936 // create a phi for the def in a dominating block. 937 938 // Find which successor block dominates this use. The successor 939 // blocks must all be single-entry (from the Catch only; I will have 940 // split blocks to make this so), hence they all dominate. 941 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 942 use_blk = use_blk->_idom; 943 944 // Find the successor 945 Node *fixup = NULL; 946 947 uint j; 948 for( j = 0; j < def_blk->_num_succs; j++ ) 949 if( use_blk == def_blk->_succs[j] ) 950 break; 951 952 if( j == def_blk->_num_succs ) { 953 // Block at same level in dom-tree is not a successor. It needs a 954 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 955 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 956 for(uint k = 1; k < use_blk->num_preds(); k++) { 957 Block* block = get_block_for_node(use_blk->pred(k)); 958 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 959 } 960 961 // Check to see if the use_blk already has an identical phi inserted. 962 // If it exists, it will be at the first position since all uses of a 963 // def are processed together. 964 Node *phi = use_blk->get_node(1); 965 if( phi->is_Phi() ) { 966 fixup = phi; 967 for (uint k = 1; k < use_blk->num_preds(); k++) { 968 if (phi->in(k) != inputs[k]) { 969 // Not a match 970 fixup = NULL; 971 break; 972 } 973 } 974 } 975 976 // If an existing PhiNode was not found, make a new one. 977 if (fixup == NULL) { 978 Node *new_phi = PhiNode::make(use_blk->head(), def); 979 use_blk->insert_node(new_phi, 1); 980 map_node_to_block(new_phi, use_blk); 981 for (uint k = 1; k < use_blk->num_preds(); k++) { 982 new_phi->set_req(k, inputs[k]); 983 } 984 fixup = new_phi; 985 } 986 987 } else { 988 // Found the use just below the Catch. Make it use the clone. 989 fixup = use_blk->get_node(n_clone_idx); 990 } 991 992 return fixup; 993 } 994 995 //--------------------------catch_cleanup_intra_block-------------------------- 996 // Fix all input edges in use that reference "def". The use is in the same 997 // block as the def and both have been cloned in each successor block. 998 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 999 1000 // Both the use and def have been cloned. For each successor block, 1001 // get the clone of the use, and make its input the clone of the def 1002 // found in that block. 1003 1004 uint use_idx = blk->find_node(use); 1005 uint offset_idx = use_idx - beg; 1006 for( uint k = 0; k < blk->_num_succs; k++ ) { 1007 // Get clone in each successor block 1008 Block *sb = blk->_succs[k]; 1009 Node *clone = sb->get_node(offset_idx+1); 1010 assert( clone->Opcode() == use->Opcode(), "" ); 1011 1012 // Make use-clone reference the def-clone 1013 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1014 } 1015 } 1016 1017 //------------------------------catch_cleanup_inter_block--------------------- 1018 // Fix all input edges in use that reference "def". The use is in a different 1019 // block than the def. 1020 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1021 if( !use_blk ) return; // Can happen if the use is a precedence edge 1022 1023 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1024 catch_cleanup_fix_all_inputs(use, def, new_def); 1025 } 1026 1027 //------------------------------call_catch_cleanup----------------------------- 1028 // If we inserted any instructions between a Call and his CatchNode, 1029 // clone the instructions on all paths below the Catch. 1030 void PhaseCFG::call_catch_cleanup(Block* block) { 1031 1032 // End of region to clone 1033 uint end = block->end_idx(); 1034 if( !block->get_node(end)->is_Catch() ) return; 1035 // Start of region to clone 1036 uint beg = end; 1037 while(!block->get_node(beg-1)->is_MachProj() || 1038 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1039 beg--; 1040 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1041 } 1042 // Range of inserted instructions is [beg, end) 1043 if( beg == end ) return; 1044 1045 // Clone along all Catch output paths. Clone area between the 'beg' and 1046 // 'end' indices. 1047 for( uint i = 0; i < block->_num_succs; i++ ) { 1048 Block *sb = block->_succs[i]; 1049 // Clone the entire area; ignoring the edge fixup for now. 1050 for( uint j = end; j > beg; j-- ) { 1051 // It is safe here to clone a node with anti_dependence 1052 // since clones dominate on each path. 1053 Node *clone = block->get_node(j-1)->clone(); 1054 sb->insert_node(clone, 1); 1055 map_node_to_block(clone, sb); 1056 } 1057 } 1058 1059 1060 // Fixup edges. Check the def-use info per cloned Node 1061 for(uint i2 = beg; i2 < end; i2++ ) { 1062 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1063 Node *n = block->get_node(i2); // Node that got cloned 1064 // Need DU safe iterator because of edge manipulation in calls. 1065 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1066 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1067 out->push(n->fast_out(j1)); 1068 } 1069 uint max = out->size(); 1070 for (uint j = 0; j < max; j++) {// For all users 1071 Node *use = out->pop(); 1072 Block *buse = get_block_for_node(use); 1073 if( use->is_Phi() ) { 1074 for( uint k = 1; k < use->req(); k++ ) 1075 if( use->in(k) == n ) { 1076 Block* b = get_block_for_node(buse->pred(k)); 1077 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1078 use->set_req(k, fixup); 1079 } 1080 } else { 1081 if (block == buse) { 1082 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1083 } else { 1084 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1085 } 1086 } 1087 } // End for all users 1088 1089 } // End of for all Nodes in cloned area 1090 1091 // Remove the now-dead cloned ops 1092 for(uint i3 = beg; i3 < end; i3++ ) { 1093 block->get_node(beg)->disconnect_inputs(NULL, C); 1094 block->remove_node(beg); 1095 } 1096 1097 // If the successor blocks have a CreateEx node, move it back to the top 1098 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1099 Block *sb = block->_succs[i4]; 1100 uint new_cnt = end - beg; 1101 // Remove any newly created, but dead, nodes. 1102 for( uint j = new_cnt; j > 0; j-- ) { 1103 Node *n = sb->get_node(j); 1104 if (n->outcnt() == 0 && 1105 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1106 n->disconnect_inputs(NULL, C); 1107 sb->remove_node(j); 1108 new_cnt--; 1109 } 1110 } 1111 // If any newly created nodes remain, move the CreateEx node to the top 1112 if (new_cnt > 0) { 1113 Node *cex = sb->get_node(1+new_cnt); 1114 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1115 sb->remove_node(1+new_cnt); 1116 sb->insert_node(cex, 1); 1117 } 1118 } 1119 } 1120 }