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_32 49 # include "adfiles/ad_ppc_32.hpp" 50 #endif 51 #ifdef TARGET_ARCH_MODEL_ppc_64 52 # include "adfiles/ad_ppc_64.hpp" 53 #endif 54 55 // Optimization - Graph Style 56 57 // Check whether val is not-null-decoded compressed oop, 58 // i.e. will grab into the base of the heap if it represents NULL. 59 static bool accesses_heap_base_zone(Node *val) { 60 if (Universe::narrow_oop_base() > 0) { // Implies UseCompressedOops. 61 if (val && val->is_Mach()) { 62 if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) { 63 // This assumes all Decodes with TypePtr::NotNull are matched to nodes that 64 // decode NULL to point to the heap base (Decode_NN). 65 if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) { 66 return true; 67 } 68 } 69 // Must recognize load operation with Decode matched in memory operand. 70 // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected() 71 // returns true everywhere else. On PPC, no such memory operands 72 // exist, therefore we did not yet implement a check for such operands. 73 NOT_AIX(Unimplemented()); 74 } 75 } 76 return false; 77 } 78 79 static bool needs_explicit_null_check_for_read(Node *val) { 80 // On some OSes (AIX) the page at address 0 is only write protected. 81 // If so, only Store operations will trap. 82 if (os::zero_page_read_protected()) { 83 return false; // Implicit null check will work. 84 } 85 // Also a read accessing the base of a heap-based compressed heap will trap. 86 if (accesses_heap_base_zone(val) && // Hits the base zone page. 87 Universe::narrow_oop_use_implicit_null_checks()) { // Base zone page is protected. 88 return false; 89 } 90 91 return true; 92 } 93 94 //------------------------------implicit_null_check---------------------------- 95 // Detect implicit-null-check opportunities. Basically, find NULL checks 96 // with suitable memory ops nearby. Use the memory op to do the NULL check. 97 // I can generate a memory op if there is not one nearby. 98 // The proj is the control projection for the not-null case. 99 // The val is the pointer being checked for nullness or 100 // decodeHeapOop_not_null node if it did not fold into address. 101 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { 102 // Assume if null check need for 0 offset then always needed 103 // Intel solaris doesn't support any null checks yet and no 104 // mechanism exists (yet) to set the switches at an os_cpu level 105 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; 106 107 // Make sure the ptr-is-null path appears to be uncommon! 108 float f = block->end()->as_MachIf()->_prob; 109 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; 110 if( f > PROB_UNLIKELY_MAG(4) ) return; 111 112 uint bidx = 0; // Capture index of value into memop 113 bool was_store; // Memory op is a store op 114 115 // Get the successor block for if the test ptr is non-null 116 Block* not_null_block; // this one goes with the proj 117 Block* null_block; 118 if (block->get_node(block->number_of_nodes()-1) == proj) { 119 null_block = block->_succs[0]; 120 not_null_block = block->_succs[1]; 121 } else { 122 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other"); 123 not_null_block = block->_succs[0]; 124 null_block = block->_succs[1]; 125 } 126 while (null_block->is_Empty() == Block::empty_with_goto) { 127 null_block = null_block->_succs[0]; 128 } 129 130 // Search the exception block for an uncommon trap. 131 // (See Parse::do_if and Parse::do_ifnull for the reason 132 // we need an uncommon trap. Briefly, we need a way to 133 // detect failure of this optimization, as in 6366351.) 134 { 135 bool found_trap = false; 136 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { 137 Node* nn = null_block->get_node(i1); 138 if (nn->is_MachCall() && 139 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 140 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); 141 if (trtype->isa_int() && trtype->is_int()->is_con()) { 142 jint tr_con = trtype->is_int()->get_con(); 143 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 144 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 145 assert((int)reason < (int)BitsPerInt, "recode bit map"); 146 if (is_set_nth_bit(allowed_reasons, (int) reason) 147 && action != Deoptimization::Action_none) { 148 // This uncommon trap is sure to recompile, eventually. 149 // When that happens, C->too_many_traps will prevent 150 // this transformation from happening again. 151 found_trap = true; 152 } 153 } 154 break; 155 } 156 } 157 if (!found_trap) { 158 // We did not find an uncommon trap. 159 return; 160 } 161 } 162 163 // Check for decodeHeapOop_not_null node which did not fold into address 164 bool is_decoden = ((intptr_t)val) & 1; 165 val = (Node*)(((intptr_t)val) & ~1); 166 167 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && 168 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); 169 170 // Search the successor block for a load or store who's base value is also 171 // the tested value. There may be several. 172 Node_List *out = new Node_List(Thread::current()->resource_area()); 173 MachNode *best = NULL; // Best found so far 174 for (DUIterator i = val->outs(); val->has_out(i); i++) { 175 Node *m = val->out(i); 176 if( !m->is_Mach() ) continue; 177 MachNode *mach = m->as_Mach(); 178 was_store = false; 179 int iop = mach->ideal_Opcode(); 180 switch( iop ) { 181 case Op_LoadB: 182 case Op_LoadUB: 183 case Op_LoadUS: 184 case Op_LoadD: 185 case Op_LoadF: 186 case Op_LoadI: 187 case Op_LoadL: 188 case Op_LoadP: 189 case Op_LoadN: 190 case Op_LoadS: 191 case Op_LoadKlass: 192 case Op_LoadNKlass: 193 case Op_LoadRange: 194 case Op_LoadD_unaligned: 195 case Op_LoadL_unaligned: 196 assert(mach->in(2) == val, "should be address"); 197 break; 198 case Op_StoreB: 199 case Op_StoreC: 200 case Op_StoreCM: 201 case Op_StoreD: 202 case Op_StoreF: 203 case Op_StoreI: 204 case Op_StoreL: 205 case Op_StoreP: 206 case Op_StoreN: 207 case Op_StoreNKlass: 208 was_store = true; // Memory op is a store op 209 // Stores will have their address in slot 2 (memory in slot 1). 210 // If the value being nul-checked is in another slot, it means we 211 // are storing the checked value, which does NOT check the value! 212 if( mach->in(2) != val ) continue; 213 break; // Found a memory op? 214 case Op_StrComp: 215 case Op_StrEquals: 216 case Op_StrIndexOf: 217 case Op_AryEq: 218 case Op_EncodeISOArray: 219 // Not a legit memory op for implicit null check regardless of 220 // embedded loads 221 continue; 222 default: // Also check for embedded loads 223 if( !mach->needs_anti_dependence_check() ) 224 continue; // Not an memory op; skip it 225 if( must_clone[iop] ) { 226 // Do not move nodes which produce flags because 227 // RA will try to clone it to place near branch and 228 // it will cause recompilation, see clone_node(). 229 continue; 230 } 231 { 232 // Check that value is used in memory address in 233 // instructions with embedded load (CmpP val1,(val2+off)). 234 Node* base; 235 Node* index; 236 const MachOper* oper = mach->memory_inputs(base, index); 237 if (oper == NULL || oper == (MachOper*)-1) { 238 continue; // Not an memory op; skip it 239 } 240 if (val == base || 241 val == index && val->bottom_type()->isa_narrowoop()) { 242 break; // Found it 243 } else { 244 continue; // Skip it 245 } 246 } 247 break; 248 } 249 250 // On some OSes (AIX) the page at address 0 is only write protected. 251 // If so, only Store operations will trap. 252 // But a read accessing the base of a heap-based compressed heap will trap. 253 if (!was_store && needs_explicit_null_check_for_read(val)) { 254 continue; 255 } 256 257 // check if the offset is not too high for implicit exception 258 { 259 intptr_t offset = 0; 260 const TypePtr *adr_type = NULL; // Do not need this return value here 261 const Node* base = mach->get_base_and_disp(offset, adr_type); 262 if (base == NULL || base == NodeSentinel) { 263 // Narrow oop address doesn't have base, only index 264 if( val->bottom_type()->isa_narrowoop() && 265 MacroAssembler::needs_explicit_null_check(offset) ) 266 continue; // Give up if offset is beyond page size 267 // cannot reason about it; is probably not implicit null exception 268 } else { 269 const TypePtr* tptr; 270 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 || 271 Universe::narrow_klass_shift() == 0)) { 272 // 32-bits narrow oop can be the base of address expressions 273 tptr = base->get_ptr_type(); 274 } else { 275 // only regular oops are expected here 276 tptr = base->bottom_type()->is_ptr(); 277 } 278 // Give up if offset is not a compile-time constant 279 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) 280 continue; 281 offset += tptr->_offset; // correct if base is offseted 282 if( MacroAssembler::needs_explicit_null_check(offset) ) 283 continue; // Give up is reference is beyond 4K page size 284 } 285 } 286 287 // Check ctrl input to see if the null-check dominates the memory op 288 Block *cb = get_block_for_node(mach); 289 cb = cb->_idom; // Always hoist at least 1 block 290 if( !was_store ) { // Stores can be hoisted only one block 291 while( cb->_dom_depth > (block->_dom_depth + 1)) 292 cb = cb->_idom; // Hoist loads as far as we want 293 // The non-null-block should dominate the memory op, too. Live 294 // range spilling will insert a spill in the non-null-block if it is 295 // needs to spill the memory op for an implicit null check. 296 if (cb->_dom_depth == (block->_dom_depth + 1)) { 297 if (cb != not_null_block) continue; 298 cb = cb->_idom; 299 } 300 } 301 if( cb != block ) continue; 302 303 // Found a memory user; see if it can be hoisted to check-block 304 uint vidx = 0; // Capture index of value into memop 305 uint j; 306 for( j = mach->req()-1; j > 0; j-- ) { 307 if( mach->in(j) == val ) { 308 vidx = j; 309 // Ignore DecodeN val which could be hoisted to where needed. 310 if( is_decoden ) continue; 311 } 312 // Block of memory-op input 313 Block *inb = get_block_for_node(mach->in(j)); 314 Block *b = block; // Start from nul check 315 while( b != inb && b->_dom_depth > inb->_dom_depth ) 316 b = b->_idom; // search upwards for input 317 // See if input dominates null check 318 if( b != inb ) 319 break; 320 } 321 if( j > 0 ) 322 continue; 323 Block *mb = get_block_for_node(mach); 324 // Hoisting stores requires more checks for the anti-dependence case. 325 // Give up hoisting if we have to move the store past any load. 326 if( was_store ) { 327 Block *b = mb; // Start searching here for a local load 328 // mach use (faulting) trying to hoist 329 // n might be blocker to hoisting 330 while( b != block ) { 331 uint k; 332 for( k = 1; k < b->number_of_nodes(); k++ ) { 333 Node *n = b->get_node(k); 334 if( n->needs_anti_dependence_check() && 335 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) 336 break; // Found anti-dependent load 337 } 338 if( k < b->number_of_nodes() ) 339 break; // Found anti-dependent load 340 // Make sure control does not do a merge (would have to check allpaths) 341 if( b->num_preds() != 2 ) break; 342 b = get_block_for_node(b->pred(1)); // Move up to predecessor block 343 } 344 if( b != block ) continue; 345 } 346 347 // Make sure this memory op is not already being used for a NullCheck 348 Node *e = mb->end(); 349 if( e->is_MachNullCheck() && e->in(1) == mach ) 350 continue; // Already being used as a NULL check 351 352 // Found a candidate! Pick one with least dom depth - the highest 353 // in the dom tree should be closest to the null check. 354 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { 355 best = mach; 356 bidx = vidx; 357 } 358 } 359 // No candidate! 360 if (best == NULL) { 361 return; 362 } 363 364 // ---- Found an implicit null check 365 extern int implicit_null_checks; 366 implicit_null_checks++; 367 368 if( is_decoden ) { 369 // Check if we need to hoist decodeHeapOop_not_null first. 370 Block *valb = get_block_for_node(val); 371 if( block != valb && block->_dom_depth < valb->_dom_depth ) { 372 // Hoist it up to the end of the test block. 373 valb->find_remove(val); 374 block->add_inst(val); 375 map_node_to_block(val, block); 376 // DecodeN on x86 may kill flags. Check for flag-killing projections 377 // that also need to be hoisted. 378 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 379 Node* n = val->fast_out(j); 380 if( n->is_MachProj() ) { 381 get_block_for_node(n)->find_remove(n); 382 block->add_inst(n); 383 map_node_to_block(n, block); 384 } 385 } 386 } 387 } 388 // Hoist the memory candidate up to the end of the test block. 389 Block *old_block = get_block_for_node(best); 390 old_block->find_remove(best); 391 block->add_inst(best); 392 map_node_to_block(best, block); 393 394 // Move the control dependence 395 if (best->in(0) && best->in(0) == old_block->head()) 396 best->set_req(0, block->head()); 397 398 // Check for flag-killing projections that also need to be hoisted 399 // Should be DU safe because no edge updates. 400 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { 401 Node* n = best->fast_out(j); 402 if( n->is_MachProj() ) { 403 get_block_for_node(n)->find_remove(n); 404 block->add_inst(n); 405 map_node_to_block(n, block); 406 } 407 } 408 409 // proj==Op_True --> ne test; proj==Op_False --> eq test. 410 // One of two graph shapes got matched: 411 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) 412 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) 413 // NULL checks are always branch-if-eq. If we see a IfTrue projection 414 // then we are replacing a 'ne' test with a 'eq' NULL check test. 415 // We need to flip the projections to keep the same semantics. 416 if( proj->Opcode() == Op_IfTrue ) { 417 // Swap order of projections in basic block to swap branch targets 418 Node *tmp1 = block->get_node(block->end_idx()+1); 419 Node *tmp2 = block->get_node(block->end_idx()+2); 420 block->map_node(tmp2, block->end_idx()+1); 421 block->map_node(tmp1, block->end_idx()+2); 422 Node *tmp = new (C) Node(C->top()); // Use not NULL input 423 tmp1->replace_by(tmp); 424 tmp2->replace_by(tmp1); 425 tmp->replace_by(tmp2); 426 tmp->destruct(); 427 } 428 429 // Remove the existing null check; use a new implicit null check instead. 430 // Since schedule-local needs precise def-use info, we need to correct 431 // it as well. 432 Node *old_tst = proj->in(0); 433 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); 434 block->map_node(nul_chk, block->end_idx()); 435 map_node_to_block(nul_chk, block); 436 // Redirect users of old_test to nul_chk 437 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) 438 old_tst->last_out(i2)->set_req(0, nul_chk); 439 // Clean-up any dead code 440 for (uint i3 = 0; i3 < old_tst->req(); i3++) 441 old_tst->set_req(i3, NULL); 442 443 latency_from_uses(nul_chk); 444 latency_from_uses(best); 445 } 446 447 448 //------------------------------select----------------------------------------- 449 // Select a nice fellow from the worklist to schedule next. If there is only 450 // one choice, then use it. Projections take top priority for correctness 451 // reasons - if I see a projection, then it is next. There are a number of 452 // other special cases, for instructions that consume condition codes, et al. 453 // These are chosen immediately. Some instructions are required to immediately 454 // precede the last instruction in the block, and these are taken last. Of the 455 // remaining cases (most), choose the instruction with the greatest latency 456 // (that is, the most number of pseudo-cycles required to the end of the 457 // routine). If there is a tie, choose the instruction with the most inputs. 458 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 459 460 // If only a single entry on the stack, use it 461 uint cnt = worklist.size(); 462 if (cnt == 1) { 463 Node *n = worklist[0]; 464 worklist.map(0,worklist.pop()); 465 return n; 466 } 467 468 uint choice = 0; // Bigger is most important 469 uint latency = 0; // Bigger is scheduled first 470 uint score = 0; // Bigger is better 471 int idx = -1; // Index in worklist 472 int cand_cnt = 0; // Candidate count 473 474 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 475 // Order in worklist is used to break ties. 476 // See caller for how this is used to delay scheduling 477 // of induction variable increments to after the other 478 // uses of the phi are scheduled. 479 Node *n = worklist[i]; // Get Node on worklist 480 481 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 482 if( n->is_Proj() || // Projections always win 483 n->Opcode()== Op_Con || // So does constant 'Top' 484 iop == Op_CreateEx || // Create-exception must start block 485 iop == Op_CheckCastPP 486 ) { 487 // select the node n 488 // remove n from worklist and retain the order of remaining nodes 489 worklist.remove((uint)i); 490 return n; 491 } 492 493 // Final call in a block must be adjacent to 'catch' 494 Node *e = block->end(); 495 if( e->is_Catch() && e->in(0)->in(0) == n ) 496 continue; 497 498 // Memory op for an implicit null check has to be at the end of the block 499 if( e->is_MachNullCheck() && e->in(1) == n ) 500 continue; 501 502 // Schedule IV increment last. 503 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 504 e->in(1)->in(1) == n && n->is_iteratively_computed()) 505 continue; 506 507 uint n_choice = 2; 508 509 // See if this instruction is consumed by a branch. If so, then (as the 510 // branch is the last instruction in the basic block) force it to the 511 // end of the basic block 512 if ( must_clone[iop] ) { 513 // See if any use is a branch 514 bool found_machif = false; 515 516 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 517 Node* use = n->fast_out(j); 518 519 // The use is a conditional branch, make them adjacent 520 if (use->is_MachIf() && get_block_for_node(use) == block) { 521 found_machif = true; 522 break; 523 } 524 525 // More than this instruction pending for successor to be ready, 526 // don't choose this if other opportunities are ready 527 if (ready_cnt.at(use->_idx) > 1) 528 n_choice = 1; 529 } 530 531 // loop terminated, prefer not to use this instruction 532 if (found_machif) 533 continue; 534 } 535 536 // See if this has a predecessor that is "must_clone", i.e. sets the 537 // condition code. If so, choose this first 538 for (uint j = 0; j < n->req() ; j++) { 539 Node *inn = n->in(j); 540 if (inn) { 541 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 542 n_choice = 3; 543 break; 544 } 545 } 546 } 547 548 // MachTemps should be scheduled last so they are near their uses 549 if (n->is_MachTemp()) { 550 n_choice = 1; 551 } 552 553 uint n_latency = get_latency_for_node(n); 554 uint n_score = n->req(); // Many inputs get high score to break ties 555 556 // Keep best latency found 557 cand_cnt++; 558 if (choice < n_choice || 559 (choice == n_choice && 560 ((StressLCM && Compile::randomized_select(cand_cnt)) || 561 (!StressLCM && 562 (latency < n_latency || 563 (latency == n_latency && 564 (score < n_score))))))) { 565 choice = n_choice; 566 latency = n_latency; 567 score = n_score; 568 idx = i; // Also keep index in worklist 569 } 570 } // End of for all ready nodes in worklist 571 572 assert(idx >= 0, "index should be set"); 573 Node *n = worklist[(uint)idx]; // Get the winner 574 575 // select the node n 576 // remove n from worklist and retain the order of remaining nodes 577 worklist.remove((uint)idx); 578 return n; 579 } 580 581 582 //------------------------------set_next_call---------------------------------- 583 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 584 if( next_call.test_set(n->_idx) ) return; 585 for( uint i=0; i<n->len(); i++ ) { 586 Node *m = n->in(i); 587 if( !m ) continue; // must see all nodes in block that precede call 588 if (get_block_for_node(m) == block) { 589 set_next_call(block, m, next_call); 590 } 591 } 592 } 593 594 //------------------------------needed_for_next_call--------------------------- 595 // Set the flag 'next_call' for each Node that is needed for the next call to 596 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 597 // next subroutine call get priority - basically it moves things NOT needed 598 // for the next call till after the call. This prevents me from trying to 599 // carry lots of stuff live across a call. 600 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 601 // Find the next control-defining Node in this block 602 Node* call = NULL; 603 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 604 Node* m = this_call->fast_out(i); 605 if (get_block_for_node(m) == block && // Local-block user 606 m != this_call && // Not self-start node 607 m->is_MachCall()) { 608 call = m; 609 break; 610 } 611 } 612 if (call == NULL) return; // No next call (e.g., block end is near) 613 // Set next-call for all inputs to this call 614 set_next_call(block, call, next_call); 615 } 616 617 //------------------------------add_call_kills------------------------------------- 618 // helper function that adds caller save registers to MachProjNode 619 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 620 // Fill in the kill mask for the call 621 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 622 if( !regs.Member(r) ) { // Not already defined by the call 623 // Save-on-call register? 624 if ((save_policy[r] == 'C') || 625 (save_policy[r] == 'A') || 626 ((save_policy[r] == 'E') && exclude_soe)) { 627 proj->_rout.Insert(r); 628 } 629 } 630 } 631 } 632 633 634 //------------------------------sched_call------------------------------------- 635 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 636 RegMask regs; 637 638 // Schedule all the users of the call right now. All the users are 639 // projection Nodes, so they must be scheduled next to the call. 640 // Collect all the defined registers. 641 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 642 Node* n = mcall->fast_out(i); 643 assert( n->is_MachProj(), "" ); 644 int n_cnt = ready_cnt.at(n->_idx)-1; 645 ready_cnt.at_put(n->_idx, n_cnt); 646 assert( n_cnt == 0, "" ); 647 // Schedule next to call 648 block->map_node(n, node_cnt++); 649 // Collect defined registers 650 regs.OR(n->out_RegMask()); 651 // Check for scheduling the next control-definer 652 if( n->bottom_type() == Type::CONTROL ) 653 // Warm up next pile of heuristic bits 654 needed_for_next_call(block, n, next_call); 655 656 // Children of projections are now all ready 657 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 658 Node* m = n->fast_out(j); // Get user 659 if(get_block_for_node(m) != block) { 660 continue; 661 } 662 if( m->is_Phi() ) continue; 663 int m_cnt = ready_cnt.at(m->_idx)-1; 664 ready_cnt.at_put(m->_idx, m_cnt); 665 if( m_cnt == 0 ) 666 worklist.push(m); 667 } 668 669 } 670 671 // Act as if the call defines the Frame Pointer. 672 // Certainly the FP is alive and well after the call. 673 regs.Insert(_matcher.c_frame_pointer()); 674 675 // Set all registers killed and not already defined by the call. 676 uint r_cnt = mcall->tf()->range()->cnt(); 677 int op = mcall->ideal_Opcode(); 678 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 679 map_node_to_block(proj, block); 680 block->insert_node(proj, node_cnt++); 681 682 // Select the right register save policy. 683 const char * save_policy; 684 switch (op) { 685 case Op_CallRuntime: 686 case Op_CallLeaf: 687 case Op_CallLeafNoFP: 688 // Calling C code so use C calling convention 689 save_policy = _matcher._c_reg_save_policy; 690 break; 691 692 case Op_CallStaticJava: 693 case Op_CallDynamicJava: 694 // Calling Java code so use Java calling convention 695 save_policy = _matcher._register_save_policy; 696 break; 697 698 default: 699 ShouldNotReachHere(); 700 } 701 702 // When using CallRuntime mark SOE registers as killed by the call 703 // so values that could show up in the RegisterMap aren't live in a 704 // callee saved register since the register wouldn't know where to 705 // find them. CallLeaf and CallLeafNoFP are ok because they can't 706 // have debug info on them. Strictly speaking this only needs to be 707 // done for oops since idealreg2debugmask takes care of debug info 708 // references but there no way to handle oops differently than other 709 // pointers as far as the kill mask goes. 710 bool exclude_soe = op == Op_CallRuntime; 711 712 // If the call is a MethodHandle invoke, we need to exclude the 713 // register which is used to save the SP value over MH invokes from 714 // the mask. Otherwise this register could be used for 715 // deoptimization information. 716 if (op == Op_CallStaticJava) { 717 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 718 if (mcallstaticjava->_method_handle_invoke) 719 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 720 } 721 722 add_call_kills(proj, regs, save_policy, exclude_soe); 723 724 return node_cnt; 725 } 726 727 728 //------------------------------schedule_local--------------------------------- 729 // Topological sort within a block. Someday become a real scheduler. 730 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) { 731 // Already "sorted" are the block start Node (as the first entry), and 732 // the block-ending Node and any trailing control projections. We leave 733 // these alone. PhiNodes and ParmNodes are made to follow the block start 734 // Node. Everything else gets topo-sorted. 735 736 #ifndef PRODUCT 737 if (trace_opto_pipelining()) { 738 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 739 for (uint i = 0;i < block->number_of_nodes(); i++) { 740 tty->print("# "); 741 block->get_node(i)->fast_dump(); 742 } 743 tty->print_cr("#"); 744 } 745 #endif 746 747 // RootNode is already sorted 748 if (block->number_of_nodes() == 1) { 749 return true; 750 } 751 752 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 753 uint node_cnt = block->end_idx(); 754 uint phi_cnt = 1; 755 uint i; 756 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 757 Node *n = block->get_node(i); 758 if( n->is_Phi() || // Found a PhiNode or ParmNode 759 (n->is_Proj() && n->in(0) == block->head()) ) { 760 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 761 block->map_node(block->get_node(phi_cnt), i); 762 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 763 } else { // All others 764 // Count block-local inputs to 'n' 765 uint cnt = n->len(); // Input count 766 uint local = 0; 767 for( uint j=0; j<cnt; j++ ) { 768 Node *m = n->in(j); 769 if( m && get_block_for_node(m) == block && !m->is_top() ) 770 local++; // One more block-local input 771 } 772 ready_cnt.at_put(n->_idx, local); // Count em up 773 774 #ifdef ASSERT 775 if( UseConcMarkSweepGC || UseG1GC ) { 776 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 777 // Check the precedence edges 778 for (uint prec = n->req(); prec < n->len(); prec++) { 779 Node* oop_store = n->in(prec); 780 if (oop_store != NULL) { 781 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 782 } 783 } 784 } 785 } 786 #endif 787 788 // A few node types require changing a required edge to a precedence edge 789 // before allocation. 790 if( n->is_Mach() && n->req() > TypeFunc::Parms && 791 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 792 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 793 // MemBarAcquire could be created without Precedent edge. 794 // del_req() replaces the specified edge with the last input edge 795 // and then removes the last edge. If the specified edge > number of 796 // edges the last edge will be moved outside of the input edges array 797 // and the edge will be lost. This is why this code should be 798 // executed only when Precedent (== TypeFunc::Parms) edge is present. 799 Node *x = n->in(TypeFunc::Parms); 800 n->del_req(TypeFunc::Parms); 801 n->add_prec(x); 802 } 803 } 804 } 805 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 806 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 807 808 // All the prescheduled guys do not hold back internal nodes 809 uint i3; 810 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 811 Node *n = block->get_node(i3); // Get pre-scheduled 812 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 813 Node* m = n->fast_out(j); 814 if (get_block_for_node(m) == block) { // Local-block user 815 int m_cnt = ready_cnt.at(m->_idx)-1; 816 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 817 } 818 } 819 } 820 821 Node_List delay; 822 // Make a worklist 823 Node_List worklist; 824 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 825 Node *m = block->get_node(i4); 826 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 827 if (m->is_iteratively_computed()) { 828 // Push induction variable increments last to allow other uses 829 // of the phi to be scheduled first. The select() method breaks 830 // ties in scheduling by worklist order. 831 delay.push(m); 832 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 833 // Force the CreateEx to the top of the list so it's processed 834 // first and ends up at the start of the block. 835 worklist.insert(0, m); 836 } else { 837 worklist.push(m); // Then on to worklist! 838 } 839 } 840 } 841 while (delay.size()) { 842 Node* d = delay.pop(); 843 worklist.push(d); 844 } 845 846 // Warm up the 'next_call' heuristic bits 847 needed_for_next_call(block, block->head(), next_call); 848 849 #ifndef PRODUCT 850 if (trace_opto_pipelining()) { 851 for (uint j=0; j< block->number_of_nodes(); j++) { 852 Node *n = block->get_node(j); 853 int idx = n->_idx; 854 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 855 tty->print("latency:%3d ", get_latency_for_node(n)); 856 tty->print("%4d: %s\n", idx, n->Name()); 857 } 858 } 859 #endif 860 861 uint max_idx = (uint)ready_cnt.length(); 862 // Pull from worklist and schedule 863 while( worklist.size() ) { // Worklist is not ready 864 865 #ifndef PRODUCT 866 if (trace_opto_pipelining()) { 867 tty->print("# ready list:"); 868 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 869 Node *n = worklist[i]; // Get Node on worklist 870 tty->print(" %d", n->_idx); 871 } 872 tty->cr(); 873 } 874 #endif 875 876 // Select and pop a ready guy from worklist 877 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt); 878 block->map_node(n, phi_cnt++); // Schedule him next 879 880 #ifndef PRODUCT 881 if (trace_opto_pipelining()) { 882 tty->print("# select %d: %s", n->_idx, n->Name()); 883 tty->print(", latency:%d", get_latency_for_node(n)); 884 n->dump(); 885 if (Verbose) { 886 tty->print("# ready list:"); 887 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 888 Node *n = worklist[i]; // Get Node on worklist 889 tty->print(" %d", n->_idx); 890 } 891 tty->cr(); 892 } 893 } 894 895 #endif 896 if( n->is_MachCall() ) { 897 MachCallNode *mcall = n->as_MachCall(); 898 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 899 continue; 900 } 901 902 if (n->is_Mach() && n->as_Mach()->has_call()) { 903 RegMask regs; 904 regs.Insert(_matcher.c_frame_pointer()); 905 regs.OR(n->out_RegMask()); 906 907 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 908 map_node_to_block(proj, block); 909 block->insert_node(proj, phi_cnt++); 910 911 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 912 } 913 914 // Children are now all ready 915 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 916 Node* m = n->fast_out(i5); // Get user 917 if (get_block_for_node(m) != block) { 918 continue; 919 } 920 if( m->is_Phi() ) continue; 921 if (m->_idx >= max_idx) { // new node, skip it 922 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 923 continue; 924 } 925 int m_cnt = ready_cnt.at(m->_idx)-1; 926 ready_cnt.at_put(m->_idx, m_cnt); 927 if( m_cnt == 0 ) 928 worklist.push(m); 929 } 930 } 931 932 if( phi_cnt != block->end_idx() ) { 933 // did not schedule all. Retry, Bailout, or Die 934 if (C->subsume_loads() == true && !C->failing()) { 935 // Retry with subsume_loads == false 936 // If this is the first failure, the sentinel string will "stick" 937 // to the Compile object, and the C2Compiler will see it and retry. 938 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 939 } 940 // assert( phi_cnt == end_idx(), "did not schedule all" ); 941 return false; 942 } 943 944 #ifndef PRODUCT 945 if (trace_opto_pipelining()) { 946 tty->print_cr("#"); 947 tty->print_cr("# after schedule_local"); 948 for (uint i = 0;i < block->number_of_nodes();i++) { 949 tty->print("# "); 950 block->get_node(i)->fast_dump(); 951 } 952 tty->cr(); 953 } 954 #endif 955 956 957 return true; 958 } 959 960 //--------------------------catch_cleanup_fix_all_inputs----------------------- 961 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 962 for (uint l = 0; l < use->len(); l++) { 963 if (use->in(l) == old_def) { 964 if (l < use->req()) { 965 use->set_req(l, new_def); 966 } else { 967 use->rm_prec(l); 968 use->add_prec(new_def); 969 l--; 970 } 971 } 972 } 973 } 974 975 //------------------------------catch_cleanup_find_cloned_def------------------ 976 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 977 assert( use_blk != def_blk, "Inter-block cleanup only"); 978 979 // The use is some block below the Catch. Find and return the clone of the def 980 // that dominates the use. If there is no clone in a dominating block, then 981 // create a phi for the def in a dominating block. 982 983 // Find which successor block dominates this use. The successor 984 // blocks must all be single-entry (from the Catch only; I will have 985 // split blocks to make this so), hence they all dominate. 986 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 987 use_blk = use_blk->_idom; 988 989 // Find the successor 990 Node *fixup = NULL; 991 992 uint j; 993 for( j = 0; j < def_blk->_num_succs; j++ ) 994 if( use_blk == def_blk->_succs[j] ) 995 break; 996 997 if( j == def_blk->_num_succs ) { 998 // Block at same level in dom-tree is not a successor. It needs a 999 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 1000 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 1001 for(uint k = 1; k < use_blk->num_preds(); k++) { 1002 Block* block = get_block_for_node(use_blk->pred(k)); 1003 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 1004 } 1005 1006 // Check to see if the use_blk already has an identical phi inserted. 1007 // If it exists, it will be at the first position since all uses of a 1008 // def are processed together. 1009 Node *phi = use_blk->get_node(1); 1010 if( phi->is_Phi() ) { 1011 fixup = phi; 1012 for (uint k = 1; k < use_blk->num_preds(); k++) { 1013 if (phi->in(k) != inputs[k]) { 1014 // Not a match 1015 fixup = NULL; 1016 break; 1017 } 1018 } 1019 } 1020 1021 // If an existing PhiNode was not found, make a new one. 1022 if (fixup == NULL) { 1023 Node *new_phi = PhiNode::make(use_blk->head(), def); 1024 use_blk->insert_node(new_phi, 1); 1025 map_node_to_block(new_phi, use_blk); 1026 for (uint k = 1; k < use_blk->num_preds(); k++) { 1027 new_phi->set_req(k, inputs[k]); 1028 } 1029 fixup = new_phi; 1030 } 1031 1032 } else { 1033 // Found the use just below the Catch. Make it use the clone. 1034 fixup = use_blk->get_node(n_clone_idx); 1035 } 1036 1037 return fixup; 1038 } 1039 1040 //--------------------------catch_cleanup_intra_block-------------------------- 1041 // Fix all input edges in use that reference "def". The use is in the same 1042 // block as the def and both have been cloned in each successor block. 1043 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1044 1045 // Both the use and def have been cloned. For each successor block, 1046 // get the clone of the use, and make its input the clone of the def 1047 // found in that block. 1048 1049 uint use_idx = blk->find_node(use); 1050 uint offset_idx = use_idx - beg; 1051 for( uint k = 0; k < blk->_num_succs; k++ ) { 1052 // Get clone in each successor block 1053 Block *sb = blk->_succs[k]; 1054 Node *clone = sb->get_node(offset_idx+1); 1055 assert( clone->Opcode() == use->Opcode(), "" ); 1056 1057 // Make use-clone reference the def-clone 1058 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1059 } 1060 } 1061 1062 //------------------------------catch_cleanup_inter_block--------------------- 1063 // Fix all input edges in use that reference "def". The use is in a different 1064 // block than the def. 1065 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1066 if( !use_blk ) return; // Can happen if the use is a precedence edge 1067 1068 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1069 catch_cleanup_fix_all_inputs(use, def, new_def); 1070 } 1071 1072 //------------------------------call_catch_cleanup----------------------------- 1073 // If we inserted any instructions between a Call and his CatchNode, 1074 // clone the instructions on all paths below the Catch. 1075 void PhaseCFG::call_catch_cleanup(Block* block) { 1076 1077 // End of region to clone 1078 uint end = block->end_idx(); 1079 if( !block->get_node(end)->is_Catch() ) return; 1080 // Start of region to clone 1081 uint beg = end; 1082 while(!block->get_node(beg-1)->is_MachProj() || 1083 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1084 beg--; 1085 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1086 } 1087 // Range of inserted instructions is [beg, end) 1088 if( beg == end ) return; 1089 1090 // Clone along all Catch output paths. Clone area between the 'beg' and 1091 // 'end' indices. 1092 for( uint i = 0; i < block->_num_succs; i++ ) { 1093 Block *sb = block->_succs[i]; 1094 // Clone the entire area; ignoring the edge fixup for now. 1095 for( uint j = end; j > beg; j-- ) { 1096 // It is safe here to clone a node with anti_dependence 1097 // since clones dominate on each path. 1098 Node *clone = block->get_node(j-1)->clone(); 1099 sb->insert_node(clone, 1); 1100 map_node_to_block(clone, sb); 1101 } 1102 } 1103 1104 1105 // Fixup edges. Check the def-use info per cloned Node 1106 for(uint i2 = beg; i2 < end; i2++ ) { 1107 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1108 Node *n = block->get_node(i2); // Node that got cloned 1109 // Need DU safe iterator because of edge manipulation in calls. 1110 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1111 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1112 out->push(n->fast_out(j1)); 1113 } 1114 uint max = out->size(); 1115 for (uint j = 0; j < max; j++) {// For all users 1116 Node *use = out->pop(); 1117 Block *buse = get_block_for_node(use); 1118 if( use->is_Phi() ) { 1119 for( uint k = 1; k < use->req(); k++ ) 1120 if( use->in(k) == n ) { 1121 Block* b = get_block_for_node(buse->pred(k)); 1122 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1123 use->set_req(k, fixup); 1124 } 1125 } else { 1126 if (block == buse) { 1127 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1128 } else { 1129 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1130 } 1131 } 1132 } // End for all users 1133 1134 } // End of for all Nodes in cloned area 1135 1136 // Remove the now-dead cloned ops 1137 for(uint i3 = beg; i3 < end; i3++ ) { 1138 block->get_node(beg)->disconnect_inputs(NULL, C); 1139 block->remove_node(beg); 1140 } 1141 1142 // If the successor blocks have a CreateEx node, move it back to the top 1143 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1144 Block *sb = block->_succs[i4]; 1145 uint new_cnt = end - beg; 1146 // Remove any newly created, but dead, nodes. 1147 for( uint j = new_cnt; j > 0; j-- ) { 1148 Node *n = sb->get_node(j); 1149 if (n->outcnt() == 0 && 1150 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1151 n->disconnect_inputs(NULL, C); 1152 sb->remove_node(j); 1153 new_cnt--; 1154 } 1155 } 1156 // If any newly created nodes remain, move the CreateEx node to the top 1157 if (new_cnt > 0) { 1158 Node *cex = sb->get_node(1+new_cnt); 1159 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1160 sb->remove_node(1+new_cnt); 1161 sb->insert_node(cex, 1); 1162 } 1163 } 1164 } 1165 }