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/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 #if defined AD_MD_HPP 34 # include AD_MD_HPP 35 #elif defined TARGET_ARCH_MODEL_x86_32 36 # include "adfiles/ad_x86_32.hpp" 37 #elif defined TARGET_ARCH_MODEL_x86_64 38 # include "adfiles/ad_x86_64.hpp" 39 #elif defined TARGET_ARCH_MODEL_aarch64 40 # include "adfiles/ad_aarch64.hpp" 41 #elif defined TARGET_ARCH_MODEL_sparc 42 # include "adfiles/ad_sparc.hpp" 43 #elif defined TARGET_ARCH_MODEL_zero 44 # include "adfiles/ad_zero.hpp" 45 #elif defined TARGET_ARCH_MODEL_ppc_64 46 # include "adfiles/ad_ppc_64.hpp" 47 #endif 48 49 // Optimization - Graph Style 50 51 // Check whether val is not-null-decoded compressed oop, 52 // i.e. will grab into the base of the heap if it represents NULL. 53 static bool accesses_heap_base_zone(Node *val) { 54 if (Universe::narrow_oop_base() != NULL) { // Implies UseCompressedOops. 55 if (val && val->is_Mach()) { 56 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) { 57 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that 58 // decode NULL to point to the heap base (Decode_NN). 59 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) { 60 return true; 61 } 62 } 63 // Must recognize load operation with Decode matched in memory operand. 64 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected() 65 // returns true everywhere else. On PPC, no such memory operands 66 // exist, therefore we did not yet implement a check for such operands. 67 NOT_AIX(Unimplemented()); 68 } 69 } 70 return false; 71 } 72 73 static bool needs_explicit_null_check_for_read(Node *val) { 74 // On some OSes (AIX) the page at address 0 is only write protected. 75 // If so, only Store operations will trap. 76 if (os::zero_page_read_protected()) { 77 return false; // Implicit null check will work. 78 } 79 // Also a read accessing the base of a heap-based compressed heap will trap. 80 if (accesses_heap_base_zone(val) && // Hits the base zone page. 81 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected. 82 return false; 83 } 84 85 return true; 86 } 87 88 //------------------------------implicit_null_check---------------------------- 89 // Detect implicit-null-check opportunities. Basically, find NULL checks 90 // with suitable memory ops nearby. Use the memory op to do the NULL check. 91 // I can generate a memory op if there is not one nearby. 92 // The proj is the control projection for the not-null case. 93 // The val is the pointer being checked for nullness or 94 // decodeHeapOop_not_null node if it did not fold into address. 95 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { 96 // Assume if null check need for 0 offset then always needed 97 // Intel solaris doesn't support any null checks yet and no 98 // mechanism exists (yet) to set the switches at an os_cpu level 99 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 100 101 // Make sure the ptr-is-null path appears to be uncommon! 102 float f = block->end()->as_MachIf()->_prob; 103 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 104 if( f > PROB_UNLIKELY_MAG(4) ) return; 105 106 uint bidx = 0; // Capture index of value into memop 107 bool was_store; // Memory op is a store op 108 109 // Get the successor block for if the test ptr is non-null 110 Block* not_null_block; // this one goes with the proj 111 Block* null_block; 112 if (block->get_node(block->number_of_nodes()-1) == proj) { 113 null_block = block->_succs[0]; 114 not_null_block = block->_succs[1]; 115 } else { 116 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other"); 117 not_null_block = block->_succs[0]; 118 null_block = block->_succs[1]; 119 } 120 while (null_block->is_Empty() == Block::empty_with_goto) { 121 null_block = null_block->_succs[0]; 122 } 123 124 // Search the exception block for an uncommon trap. 125 // (See Parse::do_if and Parse::do_ifnull for the reason 126 // we need an uncommon trap. Briefly, we need a way to 127 // detect failure of this optimization, as in 6366351.) 128 { 129 bool found_trap = false; 130 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { 131 Node* nn = null_block->get_node(i1); 132 if (nn->is_MachCall() && 133 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 134 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 135 if (trtype->isa_int() && trtype->is_int()->is_con()) { 136 jint tr_con = trtype->is_int()->get_con(); 137 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 138 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 139 assert((int)reason < (int)BitsPerInt, "recode bit map"); 140 if (is_set_nth_bit(allowed_reasons, (int) reason) 141 && action != Deoptimization::Action_none) { 142 // This uncommon trap is sure to recompile, eventually. 143 // When that happens, C->too_many_traps will prevent 144 // this transformation from happening again. 145 found_trap = true; 146 } 147 } 148 break; 149 } 150 } 151 if (!found_trap) { 152 // We did not find an uncommon trap. 153 return; 154 } 155 } 156 157 // Check for decodeHeapOop_not_null node which did not fold into address 158 bool is_decoden = ((intptr_t)val) & 1; 159 val = (Node*)(((intptr_t)val) & ~1); 160 161 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 162 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 163 164 // Search the successor block for a load or store who's base value is also 165 // the tested value. There may be several. 166 Node_List *out = new Node_List(Thread::current()->resource_area()); 167 MachNode *best = NULL; // Best found so far 168 for (DUIterator i = val->outs(); val->has_out(i); i++) { 169 Node *m = val->out(i); 170 if( !m->is_Mach() ) continue; 171 MachNode *mach = m->as_Mach(); 172 was_store = false; 173 int iop = mach->ideal_Opcode(); 174 switch( iop ) { 175 case Op_LoadB: 176 case Op_LoadUB: 177 case Op_LoadUS: 178 case Op_LoadD: 179 case Op_LoadF: 180 case Op_LoadI: 181 case Op_LoadL: 182 case Op_LoadP: 183 case Op_LoadN: 184 case Op_LoadS: 185 case Op_LoadKlass: 186 case Op_LoadNKlass: 187 case Op_LoadRange: 188 case Op_LoadD_unaligned: 189 case Op_LoadL_unaligned: 190 assert(mach->in(2) == val, "should be address"); 191 break; 192 case Op_StoreB: 193 case Op_StoreC: 194 case Op_StoreCM: 195 case Op_StoreD: 196 case Op_StoreF: 197 case Op_StoreI: 198 case Op_StoreL: 199 case Op_StoreP: 200 case Op_StoreN: 201 case Op_StoreNKlass: 202 was_store = true; // Memory op is a store op 203 // Stores will have their address in slot 2 (memory in slot 1). 204 // If the value being nul-checked is in another slot, it means we 205 // are storing the checked value, which does NOT check the value! 206 if( mach->in(2) != val ) continue; 207 break; // Found a memory op? 208 case Op_StrComp: 209 case Op_StrEquals: 210 case Op_StrIndexOf: 211 case Op_AryEq: 212 case Op_EncodeISOArray: 213 // Not a legit memory op for implicit null check regardless of 214 // embedded loads 215 continue; 216 default: // Also check for embedded loads 217 if( !mach->needs_anti_dependence_check() ) 218 continue; // Not an memory op; skip it 219 if( must_clone[iop] ) { 220 // Do not move nodes which produce flags because 221 // RA will try to clone it to place near branch and 222 // it will cause recompilation, see clone_node(). 223 continue; 224 } 225 { 226 // Check that value is used in memory address in 227 // instructions with embedded load (CmpP val1,(val2+off)). 228 Node* base; 229 Node* index; 230 const MachOper* oper = mach->memory_inputs(base, index); 231 if (oper == NULL || oper == (MachOper*)-1) { 232 continue; // Not an memory op; skip it 233 } 234 if (val == base || 235 val == index && val->bottom_type()->isa_narrowoop()) { 236 break; // Found it 237 } else { 238 continue; // Skip it 239 } 240 } 241 break; 242 } 243 244 // On some OSes (AIX) the page at address 0 is only write protected. 245 // If so, only Store operations will trap. 246 // But a read accessing the base of a heap-based compressed heap will trap. 247 if (!was_store && needs_explicit_null_check_for_read(val)) { 248 continue; 249 } 250 251 // Check that node's control edge is not-null block's head or dominates it, 252 // otherwise we can't hoist it because there are other control dependencies. 253 Node* ctrl = mach->in(0); 254 if (ctrl != NULL && !(ctrl == not_null_block->head() || 255 get_block_for_node(ctrl)->dominates(not_null_block))) { 256 continue; 257 } 258 259 // check if the offset is not too high for implicit exception 260 { 261 intptr_t offset = 0; 262 const TypePtr *adr_type = NULL; // Do not need this return value here 263 const Node* base = mach->get_base_and_disp(offset, adr_type); 264 if (base == NULL || base == NodeSentinel) { 265 // Narrow oop address doesn't have base, only index 266 if( val->bottom_type()->isa_narrowoop() && 267 MacroAssembler::needs_explicit_null_check(offset) ) 268 continue; // Give up if offset is beyond page size 269 // cannot reason about it; is probably not implicit null exception 270 } else { 271 const TypePtr* tptr; 272 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 273 Universe::narrow_klass_shift() == 0)) { 274 // 32-bits narrow oop can be the base of address expressions 275 tptr = base->get_ptr_type(); 276 } else { 277 // only regular oops are expected here 278 tptr = base->bottom_type()->is_ptr(); 279 } 280 // Give up if offset is not a compile-time constant 281 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 282 continue; 283 offset += tptr->_offset; // correct if base is offseted 284 if( MacroAssembler::needs_explicit_null_check(offset) ) 285 continue; // Give up is reference is beyond 4K page size 286 } 287 } 288 289 // Check ctrl input to see if the null-check dominates the memory op 290 Block *cb = get_block_for_node(mach); 291 cb = cb->_idom; // Always hoist at least 1 block 292 if( !was_store ) { // Stores can be hoisted only one block 293 while( cb->_dom_depth > (block->_dom_depth + 1)) 294 cb = cb->_idom; // Hoist loads as far as we want 295 // The non-null-block should dominate the memory op, too. Live 296 // range spilling will insert a spill in the non-null-block if it is 297 // needs to spill the memory op for an implicit null check. 298 if (cb->_dom_depth == (block->_dom_depth + 1)) { 299 if (cb != not_null_block) continue; 300 cb = cb->_idom; 301 } 302 } 303 if( cb != block ) continue; 304 305 // Found a memory user; see if it can be hoisted to check-block 306 uint vidx = 0; // Capture index of value into memop 307 uint j; 308 for( j = mach->req()-1; j > 0; j-- ) { 309 if( mach->in(j) == val ) { 310 vidx = j; 311 // Ignore DecodeN val which could be hoisted to where needed. 312 if( is_decoden ) continue; 313 } 314 // Block of memory-op input 315 Block *inb = get_block_for_node(mach->in(j)); 316 Block *b = block; // Start from nul check 317 while( b != inb && b->_dom_depth > inb->_dom_depth ) 318 b = b->_idom; // search upwards for input 319 // See if input dominates null check 320 if( b != inb ) 321 break; 322 } 323 if( j > 0 ) 324 continue; 325 Block *mb = get_block_for_node(mach); 326 // Hoisting stores requires more checks for the anti-dependence case. 327 // Give up hoisting if we have to move the store past any load. 328 if( was_store ) { 329 Block *b = mb; // Start searching here for a local load 330 // mach use (faulting) trying to hoist 331 // n might be blocker to hoisting 332 while( b != block ) { 333 uint k; 334 for( k = 1; k < b->number_of_nodes(); k++ ) { 335 Node *n = b->get_node(k); 336 if( n->needs_anti_dependence_check() && 337 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 338 break; // Found anti-dependent load 339 } 340 if( k < b->number_of_nodes() ) 341 break; // Found anti-dependent load 342 // Make sure control does not do a merge (would have to check allpaths) 343 if( b->num_preds() != 2 ) break; 344 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 345 } 346 if( b != block ) continue; 347 } 348 349 // Make sure this memory op is not already being used for a NullCheck 350 Node *e = mb->end(); 351 if( e->is_MachNullCheck() && e->in(1) == mach ) 352 continue; // Already being used as a NULL check 353 354 // Found a candidate! Pick one with least dom depth - the highest 355 // in the dom tree should be closest to the null check. 356 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 357 best = mach; 358 bidx = vidx; 359 } 360 } 361 // No candidate! 362 if (best == NULL) { 363 return; 364 } 365 366 // ---- Found an implicit null check 367 extern int implicit_null_checks; 368 implicit_null_checks++; 369 370 if( is_decoden ) { 371 // Check if we need to hoist decodeHeapOop_not_null first. 372 Block *valb = get_block_for_node(val); 373 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 374 // Hoist it up to the end of the test block. 375 valb->find_remove(val); 376 block->add_inst(val); 377 map_node_to_block(val, block); 378 // DecodeN on x86 may kill flags. Check for flag-killing projections 379 // that also need to be hoisted. 380 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 381 Node* n = val->fast_out(j); 382 if( n->is_MachProj() ) { 383 get_block_for_node(n)->find_remove(n); 384 block->add_inst(n); 385 map_node_to_block(n, block); 386 } 387 } 388 } 389 } 390 // Hoist the memory candidate up to the end of the test block. 391 Block *old_block = get_block_for_node(best); 392 old_block->find_remove(best); 393 block->add_inst(best); 394 map_node_to_block(best, block); 395 396 // Move the control dependence if it is pinned to not-null block. 397 // Don't change it in other cases: NULL or dominating control. 398 if (best->in(0) == not_null_block->head()) { 399 // Set it to control edge of null check. 400 best->set_req(0, proj->in(0)->in(0)); 401 } 402 403 // Check for flag-killing projections that also need to be hoisted 404 // Should be DU safe because no edge updates. 405 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 406 Node* n = best->fast_out(j); 407 if( n->is_MachProj() ) { 408 get_block_for_node(n)->find_remove(n); 409 block->add_inst(n); 410 map_node_to_block(n, block); 411 } 412 } 413 414 // proj==Op_True --> ne test; proj==Op_False --> eq test. 415 // One of two graph shapes got matched: 416 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 417 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 418 // NULL checks are always branch-if-eq. If we see a IfTrue projection 419 // then we are replacing a 'ne' test with a 'eq' NULL check test. 420 // We need to flip the projections to keep the same semantics. 421 if( proj->Opcode() == Op_IfTrue ) { 422 // Swap order of projections in basic block to swap branch targets 423 Node *tmp1 = block->get_node(block->end_idx()+1); 424 Node *tmp2 = block->get_node(block->end_idx()+2); 425 block->map_node(tmp2, block->end_idx()+1); 426 block->map_node(tmp1, block->end_idx()+2); 427 Node *tmp = new (C) Node(C->top()); // Use not NULL input 428 tmp1->replace_by(tmp); 429 tmp2->replace_by(tmp1); 430 tmp->replace_by(tmp2); 431 tmp->destruct(); 432 } 433 434 // Remove the existing null check; use a new implicit null check instead. 435 // Since schedule-local needs precise def-use info, we need to correct 436 // it as well. 437 Node *old_tst = proj->in(0); 438 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); 439 block->map_node(nul_chk, block->end_idx()); 440 map_node_to_block(nul_chk, block); 441 // Redirect users of old_test to nul_chk 442 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 443 old_tst->last_out(i2)->set_req(0, nul_chk); 444 // Clean-up any dead code 445 for (uint i3 = 0; i3 < old_tst->req(); i3++) { 446 Node* in = old_tst->in(i3); 447 old_tst->set_req(i3, NULL); 448 if (in->outcnt() == 0) { 449 // Remove dead input node 450 in->disconnect_inputs(NULL, C); 451 block->find_remove(in); 452 } 453 } 454 455 latency_from_uses(nul_chk); 456 latency_from_uses(best); 457 458 // insert anti-dependences to defs in this block 459 if (! best->needs_anti_dependence_check()) { 460 for (uint k = 1; k < block->number_of_nodes(); k++) { 461 Node *n = block->get_node(k); 462 if (n->needs_anti_dependence_check() && 463 n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) { 464 // Found anti-dependent load 465 insert_anti_dependences(block, n); 466 } 467 } 468 } 469 } 470 471 472 //------------------------------select----------------------------------------- 473 // Select a nice fellow from the worklist to schedule next. If there is only 474 // one choice, then use it. Projections take top priority for correctness 475 // reasons - if I see a projection, then it is next. There are a number of 476 // other special cases, for instructions that consume condition codes, et al. 477 // These are chosen immediately. Some instructions are required to immediately 478 // precede the last instruction in the block, and these are taken last. Of the 479 // remaining cases (most), choose the instruction with the greatest latency 480 // (that is, the most number of pseudo-cycles required to the end of the 481 // routine). If there is a tie, choose the instruction with the most inputs. 482 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 483 484 // If only a single entry on the stack, use it 485 uint cnt = worklist.size(); 486 if (cnt == 1) { 487 Node *n = worklist[0]; 488 worklist.map(0,worklist.pop()); 489 return n; 490 } 491 492 uint choice = 0; // Bigger is most important 493 uint latency = 0; // Bigger is scheduled first 494 uint score = 0; // Bigger is better 495 int idx = -1; // Index in worklist 496 int cand_cnt = 0; // Candidate count 497 498 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 499 // Order in worklist is used to break ties. 500 // See caller for how this is used to delay scheduling 501 // of induction variable increments to after the other 502 // uses of the phi are scheduled. 503 Node *n = worklist[i]; // Get Node on worklist 504 505 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 506 if( n->is_Proj() || // Projections always win 507 n->Opcode()== Op_Con || // So does constant 'Top' 508 iop == Op_CreateEx || // Create-exception must start block 509 iop == Op_CheckCastPP 510 ) { 511 worklist.map(i,worklist.pop()); 512 return n; 513 } 514 515 // Final call in a block must be adjacent to 'catch' 516 Node *e = block->end(); 517 if( e->is_Catch() && e->in(0)->in(0) == n ) 518 continue; 519 520 // Memory op for an implicit null check has to be at the end of the block 521 if( e->is_MachNullCheck() && e->in(1) == n ) 522 continue; 523 524 // Schedule IV increment last. 525 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 526 e->in(1)->in(1) == n && n->is_iteratively_computed()) 527 continue; 528 529 uint n_choice = 2; 530 531 // See if this instruction is consumed by a branch. If so, then (as the 532 // branch is the last instruction in the basic block) force it to the 533 // end of the basic block 534 if ( must_clone[iop] ) { 535 // See if any use is a branch 536 bool found_machif = false; 537 538 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 539 Node* use = n->fast_out(j); 540 541 // The use is a conditional branch, make them adjacent 542 if (use->is_MachIf() && get_block_for_node(use) == block) { 543 found_machif = true; 544 break; 545 } 546 547 // More than this instruction pending for successor to be ready, 548 // don't choose this if other opportunities are ready 549 if (ready_cnt.at(use->_idx) > 1) 550 n_choice = 1; 551 } 552 553 // loop terminated, prefer not to use this instruction 554 if (found_machif) 555 continue; 556 } 557 558 // See if this has a predecessor that is "must_clone", i.e. sets the 559 // condition code. If so, choose this first 560 for (uint j = 0; j < n->req() ; j++) { 561 Node *inn = n->in(j); 562 if (inn) { 563 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 564 n_choice = 3; 565 break; 566 } 567 } 568 } 569 570 // MachTemps should be scheduled last so they are near their uses 571 if (n->is_MachTemp()) { 572 n_choice = 1; 573 } 574 575 uint n_latency = get_latency_for_node(n); 576 uint n_score = n->req(); // Many inputs get high score to break ties 577 578 // Keep best latency found 579 cand_cnt++; 580 if (choice < n_choice || 581 (choice == n_choice && 582 ((StressLCM && Compile::randomized_select(cand_cnt)) || 583 (!StressLCM && 584 (latency < n_latency || 585 (latency == n_latency && 586 (score < n_score))))))) { 587 choice = n_choice; 588 latency = n_latency; 589 score = n_score; 590 idx = i; // Also keep index in worklist 591 } 592 } // End of for all ready nodes in worklist 593 594 assert(idx >= 0, "index should be set"); 595 Node *n = worklist[(uint)idx]; // Get the winner 596 597 worklist.map((uint)idx, worklist.pop()); // Compress worklist 598 return n; 599 } 600 601 602 //------------------------------set_next_call---------------------------------- 603 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 604 if( next_call.test_set(n->_idx) ) return; 605 for( uint i=0; i<n->len(); i++ ) { 606 Node *m = n->in(i); 607 if( !m ) continue; // must see all nodes in block that precede call 608 if (get_block_for_node(m) == block) { 609 set_next_call(block, m, next_call); 610 } 611 } 612 } 613 614 //------------------------------needed_for_next_call--------------------------- 615 // Set the flag 'next_call' for each Node that is needed for the next call to 616 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 617 // next subroutine call get priority - basically it moves things NOT needed 618 // for the next call till after the call. This prevents me from trying to 619 // carry lots of stuff live across a call. 620 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 621 // Find the next control-defining Node in this block 622 Node* call = NULL; 623 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 624 Node* m = this_call->fast_out(i); 625 if (get_block_for_node(m) == block && // Local-block user 626 m != this_call && // Not self-start node 627 m->is_MachCall()) { 628 call = m; 629 break; 630 } 631 } 632 if (call == NULL) return; // No next call (e.g., block end is near) 633 // Set next-call for all inputs to this call 634 set_next_call(block, call, next_call); 635 } 636 637 //------------------------------add_call_kills------------------------------------- 638 // helper function that adds caller save registers to MachProjNode 639 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 640 // Fill in the kill mask for the call 641 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 642 if( !regs.Member(r) ) { // Not already defined by the call 643 // Save-on-call register? 644 if ((save_policy[r] == 'C') || 645 (save_policy[r] == 'A') || 646 ((save_policy[r] == 'E') && exclude_soe)) { 647 proj->_rout.Insert(r); 648 } 649 } 650 } 651 } 652 653 654 //------------------------------sched_call------------------------------------- 655 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 656 RegMask regs; 657 658 // Schedule all the users of the call right now. All the users are 659 // projection Nodes, so they must be scheduled next to the call. 660 // Collect all the defined registers. 661 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 662 Node* n = mcall->fast_out(i); 663 assert( n->is_MachProj(), "" ); 664 int n_cnt = ready_cnt.at(n->_idx)-1; 665 ready_cnt.at_put(n->_idx, n_cnt); 666 assert( n_cnt == 0, "" ); 667 // Schedule next to call 668 block->map_node(n, node_cnt++); 669 // Collect defined registers 670 regs.OR(n->out_RegMask()); 671 // Check for scheduling the next control-definer 672 if( n->bottom_type() == Type::CONTROL ) 673 // Warm up next pile of heuristic bits 674 needed_for_next_call(block, n, next_call); 675 676 // Children of projections are now all ready 677 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 678 Node* m = n->fast_out(j); // Get user 679 if(get_block_for_node(m) != block) { 680 continue; 681 } 682 if( m->is_Phi() ) continue; 683 int m_cnt = ready_cnt.at(m->_idx)-1; 684 ready_cnt.at_put(m->_idx, m_cnt); 685 if( m_cnt == 0 ) 686 worklist.push(m); 687 } 688 689 } 690 691 // Act as if the call defines the Frame Pointer. 692 // Certainly the FP is alive and well after the call. 693 regs.Insert(_matcher.c_frame_pointer()); 694 695 // Set all registers killed and not already defined by the call. 696 uint r_cnt = mcall->tf()->range()->cnt(); 697 int op = mcall->ideal_Opcode(); 698 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 699 map_node_to_block(proj, block); 700 block->insert_node(proj, node_cnt++); 701 702 // Select the right register save policy. 703 const char *save_policy = NULL; 704 switch (op) { 705 case Op_CallRuntime: 706 case Op_CallLeaf: 707 case Op_CallLeafNoFP: 708 // Calling C code so use C calling convention 709 save_policy = _matcher._c_reg_save_policy; 710 break; 711 712 case Op_CallStaticJava: 713 case Op_CallDynamicJava: 714 // Calling Java code so use Java calling convention 715 save_policy = _matcher._register_save_policy; 716 break; 717 718 default: 719 ShouldNotReachHere(); 720 } 721 722 // When using CallRuntime mark SOE registers as killed by the call 723 // so values that could show up in the RegisterMap aren't live in a 724 // callee saved register since the register wouldn't know where to 725 // find them. CallLeaf and CallLeafNoFP are ok because they can't 726 // have debug info on them. Strictly speaking this only needs to be 727 // done for oops since idealreg2debugmask takes care of debug info 728 // references but there no way to handle oops differently than other 729 // pointers as far as the kill mask goes. 730 bool exclude_soe = op == Op_CallRuntime; 731 732 // If the call is a MethodHandle invoke, we need to exclude the 733 // register which is used to save the SP value over MH invokes from 734 // the mask. Otherwise this register could be used for 735 // deoptimization information. 736 if (op == Op_CallStaticJava) { 737 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 738 if (mcallstaticjava->_method_handle_invoke) 739 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 740 } 741 742 add_call_kills(proj, regs, save_policy, exclude_soe); 743 744 return node_cnt; 745 } 746 747 748 //------------------------------schedule_local--------------------------------- 749 // Topological sort within a block. Someday become a real scheduler. 750 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) { 751 // Already "sorted" are the block start Node (as the first entry), and 752 // the block-ending Node and any trailing control projections. We leave 753 // these alone. PhiNodes and ParmNodes are made to follow the block start 754 // Node. Everything else gets topo-sorted. 755 756 #ifndef PRODUCT 757 if (trace_opto_pipelining()) { 758 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 759 for (uint i = 0;i < block->number_of_nodes(); i++) { 760 tty->print("# "); 761 block->get_node(i)->fast_dump(); 762 } 763 tty->print_cr("#"); 764 } 765 #endif 766 767 // RootNode is already sorted 768 if (block->number_of_nodes() == 1) { 769 return true; 770 } 771 772 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 773 uint node_cnt = block->end_idx(); 774 uint phi_cnt = 1; 775 uint i; 776 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 777 Node *n = block->get_node(i); 778 if( n->is_Phi() || // Found a PhiNode or ParmNode 779 (n->is_Proj() && n->in(0) == block->head()) ) { 780 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 781 block->map_node(block->get_node(phi_cnt), i); 782 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 783 } else { // All others 784 // Count block-local inputs to 'n' 785 uint cnt = n->len(); // Input count 786 uint local = 0; 787 for( uint j=0; j<cnt; j++ ) { 788 Node *m = n->in(j); 789 if( m && get_block_for_node(m) == block && !m->is_top() ) 790 local++; // One more block-local input 791 } 792 ready_cnt.at_put(n->_idx, local); // Count em up 793 794 #ifdef ASSERT 795 if( UseConcMarkSweepGC || UseG1GC ) { 796 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 797 // Check the precedence edges 798 for (uint prec = n->req(); prec < n->len(); prec++) { 799 Node* oop_store = n->in(prec); 800 if (oop_store != NULL) { 801 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 802 } 803 } 804 } 805 } 806 #endif 807 808 // A few node types require changing a required edge to a precedence edge 809 // before allocation. 810 if( n->is_Mach() && n->req() > TypeFunc::Parms && 811 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 812 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 813 // MemBarAcquire could be created without Precedent edge. 814 // del_req() replaces the specified edge with the last input edge 815 // and then removes the last edge. If the specified edge > number of 816 // edges the last edge will be moved outside of the input edges array 817 // and the edge will be lost. This is why this code should be 818 // executed only when Precedent (== TypeFunc::Parms) edge is present. 819 Node *x = n->in(TypeFunc::Parms); 820 if (x != NULL && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) { 821 // Old edge to node within same block will get removed, but no precedence 822 // edge will get added because it already exists. Update ready count. 823 int cnt = ready_cnt.at(n->_idx); 824 assert(cnt > 1, err_msg("MemBar node %d must not get ready here", n->_idx)); 825 ready_cnt.at_put(n->_idx, cnt-1); 826 } 827 n->del_req(TypeFunc::Parms); 828 n->add_prec(x); 829 } 830 } 831 } 832 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 833 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 834 835 // All the prescheduled guys do not hold back internal nodes 836 uint i3; 837 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 838 Node *n = block->get_node(i3); // Get pre-scheduled 839 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 840 Node* m = n->fast_out(j); 841 if (get_block_for_node(m) == block) { // Local-block user 842 int m_cnt = ready_cnt.at(m->_idx)-1; 843 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 844 } 845 } 846 } 847 848 Node_List delay; 849 // Make a worklist 850 Node_List worklist; 851 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 852 Node *m = block->get_node(i4); 853 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 854 if (m->is_iteratively_computed()) { 855 // Push induction variable increments last to allow other uses 856 // of the phi to be scheduled first. The select() method breaks 857 // ties in scheduling by worklist order. 858 delay.push(m); 859 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 860 // Force the CreateEx to the top of the list so it's processed 861 // first and ends up at the start of the block. 862 worklist.insert(0, m); 863 } else { 864 worklist.push(m); // Then on to worklist! 865 } 866 } 867 } 868 while (delay.size()) { 869 Node* d = delay.pop(); 870 worklist.push(d); 871 } 872 873 // Warm up the 'next_call' heuristic bits 874 needed_for_next_call(block, block->head(), next_call); 875 876 #ifndef PRODUCT 877 if (trace_opto_pipelining()) { 878 for (uint j=0; j< block->number_of_nodes(); j++) { 879 Node *n = block->get_node(j); 880 int idx = n->_idx; 881 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 882 tty->print("latency:%3d ", get_latency_for_node(n)); 883 tty->print("%4d: %s\n", idx, n->Name()); 884 } 885 } 886 #endif 887 888 uint max_idx = (uint)ready_cnt.length(); 889 // Pull from worklist and schedule 890 while( worklist.size() ) { // Worklist is not ready 891 892 #ifndef PRODUCT 893 if (trace_opto_pipelining()) { 894 tty->print("# ready list:"); 895 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 896 Node *n = worklist[i]; // Get Node on worklist 897 tty->print(" %d", n->_idx); 898 } 899 tty->cr(); 900 } 901 #endif 902 903 // Select and pop a ready guy from worklist 904 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt); 905 block->map_node(n, phi_cnt++); // Schedule him next 906 907 #ifndef PRODUCT 908 if (trace_opto_pipelining()) { 909 tty->print("# select %d: %s", n->_idx, n->Name()); 910 tty->print(", latency:%d", get_latency_for_node(n)); 911 n->dump(); 912 if (Verbose) { 913 tty->print("# ready list:"); 914 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 915 Node *n = worklist[i]; // Get Node on worklist 916 tty->print(" %d", n->_idx); 917 } 918 tty->cr(); 919 } 920 } 921 922 #endif 923 if( n->is_MachCall() ) { 924 MachCallNode *mcall = n->as_MachCall(); 925 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 926 continue; 927 } 928 929 if (n->is_Mach() && n->as_Mach()->has_call()) { 930 RegMask regs; 931 regs.Insert(_matcher.c_frame_pointer()); 932 regs.OR(n->out_RegMask()); 933 934 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 935 map_node_to_block(proj, block); 936 block->insert_node(proj, phi_cnt++); 937 938 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 939 } 940 941 // Children are now all ready 942 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 943 Node* m = n->fast_out(i5); // Get user 944 if (get_block_for_node(m) != block) { 945 continue; 946 } 947 if( m->is_Phi() ) continue; 948 if (m->_idx >= max_idx) { // new node, skip it 949 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 950 continue; 951 } 952 int m_cnt = ready_cnt.at(m->_idx)-1; 953 ready_cnt.at_put(m->_idx, m_cnt); 954 if( m_cnt == 0 ) 955 worklist.push(m); 956 } 957 } 958 959 if( phi_cnt != block->end_idx() ) { 960 // did not schedule all. Retry, Bailout, or Die 961 if (C->subsume_loads() == true && !C->failing()) { 962 // Retry with subsume_loads == false 963 // If this is the first failure, the sentinel string will "stick" 964 // to the Compile object, and the C2Compiler will see it and retry. 965 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 966 } else { 967 assert(false, "graph should be schedulable"); 968 } 969 // assert( phi_cnt == end_idx(), "did not schedule all" ); 970 return false; 971 } 972 973 #ifndef PRODUCT 974 if (trace_opto_pipelining()) { 975 tty->print_cr("#"); 976 tty->print_cr("# after schedule_local"); 977 for (uint i = 0;i < block->number_of_nodes();i++) { 978 tty->print("# "); 979 block->get_node(i)->fast_dump(); 980 } 981 tty->cr(); 982 } 983 #endif 984 985 986 return true; 987 } 988 989 //--------------------------catch_cleanup_fix_all_inputs----------------------- 990 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 991 for (uint l = 0; l < use->len(); l++) { 992 if (use->in(l) == old_def) { 993 if (l < use->req()) { 994 use->set_req(l, new_def); 995 } else { 996 use->rm_prec(l); 997 use->add_prec(new_def); 998 l--; 999 } 1000 } 1001 } 1002 } 1003 1004 //------------------------------catch_cleanup_find_cloned_def------------------ 1005 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1006 assert( use_blk != def_blk, "Inter-block cleanup only"); 1007 1008 // The use is some block below the Catch. Find and return the clone of the def 1009 // that dominates the use. If there is no clone in a dominating block, then 1010 // create a phi for the def in a dominating block. 1011 1012 // Find which successor block dominates this use. The successor 1013 // blocks must all be single-entry (from the Catch only; I will have 1014 // split blocks to make this so), hence they all dominate. 1015 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 1016 use_blk = use_blk->_idom; 1017 1018 // Find the successor 1019 Node *fixup = NULL; 1020 1021 uint j; 1022 for( j = 0; j < def_blk->_num_succs; j++ ) 1023 if( use_blk == def_blk->_succs[j] ) 1024 break; 1025 1026 if( j == def_blk->_num_succs ) { 1027 // Block at same level in dom-tree is not a successor. It needs a 1028 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 1029 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 1030 for(uint k = 1; k < use_blk->num_preds(); k++) { 1031 Block* block = get_block_for_node(use_blk->pred(k)); 1032 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 1033 } 1034 1035 // Check to see if the use_blk already has an identical phi inserted. 1036 // If it exists, it will be at the first position since all uses of a 1037 // def are processed together. 1038 Node *phi = use_blk->get_node(1); 1039 if( phi->is_Phi() ) { 1040 fixup = phi; 1041 for (uint k = 1; k < use_blk->num_preds(); k++) { 1042 if (phi->in(k) != inputs[k]) { 1043 // Not a match 1044 fixup = NULL; 1045 break; 1046 } 1047 } 1048 } 1049 1050 // If an existing PhiNode was not found, make a new one. 1051 if (fixup == NULL) { 1052 Node *new_phi = PhiNode::make(use_blk->head(), def); 1053 use_blk->insert_node(new_phi, 1); 1054 map_node_to_block(new_phi, use_blk); 1055 for (uint k = 1; k < use_blk->num_preds(); k++) { 1056 new_phi->set_req(k, inputs[k]); 1057 } 1058 fixup = new_phi; 1059 } 1060 1061 } else { 1062 // Found the use just below the Catch. Make it use the clone. 1063 fixup = use_blk->get_node(n_clone_idx); 1064 } 1065 1066 return fixup; 1067 } 1068 1069 //--------------------------catch_cleanup_intra_block-------------------------- 1070 // Fix all input edges in use that reference "def". The use is in the same 1071 // block as the def and both have been cloned in each successor block. 1072 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1073 1074 // Both the use and def have been cloned. For each successor block, 1075 // get the clone of the use, and make its input the clone of the def 1076 // found in that block. 1077 1078 uint use_idx = blk->find_node(use); 1079 uint offset_idx = use_idx - beg; 1080 for( uint k = 0; k < blk->_num_succs; k++ ) { 1081 // Get clone in each successor block 1082 Block *sb = blk->_succs[k]; 1083 Node *clone = sb->get_node(offset_idx+1); 1084 assert( clone->Opcode() == use->Opcode(), "" ); 1085 1086 // Make use-clone reference the def-clone 1087 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1088 } 1089 } 1090 1091 //------------------------------catch_cleanup_inter_block--------------------- 1092 // Fix all input edges in use that reference "def". The use is in a different 1093 // block than the def. 1094 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1095 if( !use_blk ) return; // Can happen if the use is a precedence edge 1096 1097 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1098 catch_cleanup_fix_all_inputs(use, def, new_def); 1099 } 1100 1101 //------------------------------call_catch_cleanup----------------------------- 1102 // If we inserted any instructions between a Call and his CatchNode, 1103 // clone the instructions on all paths below the Catch. 1104 void PhaseCFG::call_catch_cleanup(Block* block) { 1105 1106 // End of region to clone 1107 uint end = block->end_idx(); 1108 if( !block->get_node(end)->is_Catch() ) return; 1109 // Start of region to clone 1110 uint beg = end; 1111 while(!block->get_node(beg-1)->is_MachProj() || 1112 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1113 beg--; 1114 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1115 } 1116 // Range of inserted instructions is [beg, end) 1117 if( beg == end ) return; 1118 1119 // Clone along all Catch output paths. Clone area between the 'beg' and 1120 // 'end' indices. 1121 for( uint i = 0; i < block->_num_succs; i++ ) { 1122 Block *sb = block->_succs[i]; 1123 // Clone the entire area; ignoring the edge fixup for now. 1124 for( uint j = end; j > beg; j-- ) { 1125 Node *clone = block->get_node(j-1)->clone(); 1126 sb->insert_node(clone, 1); 1127 map_node_to_block(clone, sb); 1128 if (clone->needs_anti_dependence_check()) { 1129 insert_anti_dependences(sb, clone); 1130 } 1131 } 1132 } 1133 1134 1135 // Fixup edges. Check the def-use info per cloned Node 1136 for(uint i2 = beg; i2 < end; i2++ ) { 1137 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1138 Node *n = block->get_node(i2); // Node that got cloned 1139 // Need DU safe iterator because of edge manipulation in calls. 1140 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1141 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1142 out->push(n->fast_out(j1)); 1143 } 1144 uint max = out->size(); 1145 for (uint j = 0; j < max; j++) {// For all users 1146 Node *use = out->pop(); 1147 Block *buse = get_block_for_node(use); 1148 if( use->is_Phi() ) { 1149 for( uint k = 1; k < use->req(); k++ ) 1150 if( use->in(k) == n ) { 1151 Block* b = get_block_for_node(buse->pred(k)); 1152 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1153 use->set_req(k, fixup); 1154 } 1155 } else { 1156 if (block == buse) { 1157 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1158 } else { 1159 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1160 } 1161 } 1162 } // End for all users 1163 1164 } // End of for all Nodes in cloned area 1165 1166 // Remove the now-dead cloned ops 1167 for(uint i3 = beg; i3 < end; i3++ ) { 1168 block->get_node(beg)->disconnect_inputs(NULL, C); 1169 block->remove_node(beg); 1170 } 1171 1172 // If the successor blocks have a CreateEx node, move it back to the top 1173 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1174 Block *sb = block->_succs[i4]; 1175 uint new_cnt = end - beg; 1176 // Remove any newly created, but dead, nodes. 1177 for( uint j = new_cnt; j > 0; j-- ) { 1178 Node *n = sb->get_node(j); 1179 if (n->outcnt() == 0 && 1180 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1181 n->disconnect_inputs(NULL, C); 1182 sb->remove_node(j); 1183 new_cnt--; 1184 } 1185 } 1186 // If any newly created nodes remain, move the CreateEx node to the top 1187 if (new_cnt > 0) { 1188 Node *cex = sb->get_node(1+new_cnt); 1189 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1190 sb->remove_node(1+new_cnt); 1191 sb->insert_node(cex, 1); 1192 } 1193 } 1194 } 1195 }