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 Node* in = old_tst->in(i3); 442 old_tst->set_req(i3, NULL); 443 if (in->outcnt() == 0) { 444 // Remove dead input node 445 in->disconnect_inputs(NULL, C); 446 block->find_remove(in); 447 } 448 } 449 450 latency_from_uses(nul_chk); 451 latency_from_uses(best); 452 } 453 454 455 //------------------------------select----------------------------------------- 456 // Select a nice fellow from the worklist to schedule next. If there is only 457 // one choice, then use it. Projections take top priority for correctness 458 // reasons - if I see a projection, then it is next. There are a number of 459 // other special cases, for instructions that consume condition codes, et al. 460 // These are chosen immediately. Some instructions are required to immediately 461 // precede the last instruction in the block, and these are taken last. Of the 462 // remaining cases (most), choose the instruction with the greatest latency 463 // (that is, the most number of pseudo-cycles required to the end of the 464 // routine). If there is a tie, choose the instruction with the most inputs. 465 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) { 466 467 // If only a single entry on the stack, use it 468 uint cnt = worklist.size(); 469 if (cnt == 1) { 470 Node *n = worklist[0]; 471 worklist.map(0,worklist.pop()); 472 return n; 473 } 474 475 uint choice = 0; // Bigger is most important 476 uint latency = 0; // Bigger is scheduled first 477 uint score = 0; // Bigger is better 478 int idx = -1; // Index in worklist 479 int cand_cnt = 0; // Candidate count 480 481 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist 482 // Order in worklist is used to break ties. 483 // See caller for how this is used to delay scheduling 484 // of induction variable increments to after the other 485 // uses of the phi are scheduled. 486 Node *n = worklist[i]; // Get Node on worklist 487 488 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; 489 if( n->is_Proj() || // Projections always win 490 n->Opcode()== Op_Con || // So does constant 'Top' 491 iop == Op_CreateEx || // Create-exception must start block 492 iop == Op_CheckCastPP 493 ) { 494 worklist.map(i,worklist.pop()); 495 return n; 496 } 497 498 // Final call in a block must be adjacent to 'catch' 499 Node *e = block->end(); 500 if( e->is_Catch() && e->in(0)->in(0) == n ) 501 continue; 502 503 // Memory op for an implicit null check has to be at the end of the block 504 if( e->is_MachNullCheck() && e->in(1) == n ) 505 continue; 506 507 // Schedule IV increment last. 508 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd && 509 e->in(1)->in(1) == n && n->is_iteratively_computed()) 510 continue; 511 512 uint n_choice = 2; 513 514 // See if this instruction is consumed by a branch. If so, then (as the 515 // branch is the last instruction in the basic block) force it to the 516 // end of the basic block 517 if ( must_clone[iop] ) { 518 // See if any use is a branch 519 bool found_machif = false; 520 521 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 522 Node* use = n->fast_out(j); 523 524 // The use is a conditional branch, make them adjacent 525 if (use->is_MachIf() && get_block_for_node(use) == block) { 526 found_machif = true; 527 break; 528 } 529 530 // More than this instruction pending for successor to be ready, 531 // don't choose this if other opportunities are ready 532 if (ready_cnt.at(use->_idx) > 1) 533 n_choice = 1; 534 } 535 536 // loop terminated, prefer not to use this instruction 537 if (found_machif) 538 continue; 539 } 540 541 // See if this has a predecessor that is "must_clone", i.e. sets the 542 // condition code. If so, choose this first 543 for (uint j = 0; j < n->req() ; j++) { 544 Node *inn = n->in(j); 545 if (inn) { 546 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { 547 n_choice = 3; 548 break; 549 } 550 } 551 } 552 553 // MachTemps should be scheduled last so they are near their uses 554 if (n->is_MachTemp()) { 555 n_choice = 1; 556 } 557 558 uint n_latency = get_latency_for_node(n); 559 uint n_score = n->req(); // Many inputs get high score to break ties 560 561 // Keep best latency found 562 cand_cnt++; 563 if (choice < n_choice || 564 (choice == n_choice && 565 ((StressLCM && Compile::randomized_select(cand_cnt)) || 566 (!StressLCM && 567 (latency < n_latency || 568 (latency == n_latency && 569 (score < n_score))))))) { 570 choice = n_choice; 571 latency = n_latency; 572 score = n_score; 573 idx = i; // Also keep index in worklist 574 } 575 } // End of for all ready nodes in worklist 576 577 assert(idx >= 0, "index should be set"); 578 Node *n = worklist[(uint)idx]; // Get the winner 579 580 worklist.map((uint)idx, worklist.pop()); // Compress worklist 581 return n; 582 } 583 584 585 //------------------------------set_next_call---------------------------------- 586 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { 587 if( next_call.test_set(n->_idx) ) return; 588 for( uint i=0; i<n->len(); i++ ) { 589 Node *m = n->in(i); 590 if( !m ) continue; // must see all nodes in block that precede call 591 if (get_block_for_node(m) == block) { 592 set_next_call(block, m, next_call); 593 } 594 } 595 } 596 597 //------------------------------needed_for_next_call--------------------------- 598 // Set the flag 'next_call' for each Node that is needed for the next call to 599 // be scheduled. This flag lets me bias scheduling so Nodes needed for the 600 // next subroutine call get priority - basically it moves things NOT needed 601 // for the next call till after the call. This prevents me from trying to 602 // carry lots of stuff live across a call. 603 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { 604 // Find the next control-defining Node in this block 605 Node* call = NULL; 606 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { 607 Node* m = this_call->fast_out(i); 608 if (get_block_for_node(m) == block && // Local-block user 609 m != this_call && // Not self-start node 610 m->is_MachCall()) { 611 call = m; 612 break; 613 } 614 } 615 if (call == NULL) return; // No next call (e.g., block end is near) 616 // Set next-call for all inputs to this call 617 set_next_call(block, call, next_call); 618 } 619 620 //------------------------------add_call_kills------------------------------------- 621 // helper function that adds caller save registers to MachProjNode 622 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { 623 // Fill in the kill mask for the call 624 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { 625 if( !regs.Member(r) ) { // Not already defined by the call 626 // Save-on-call register? 627 if ((save_policy[r] == 'C') || 628 (save_policy[r] == 'A') || 629 ((save_policy[r] == 'E') && exclude_soe)) { 630 proj->_rout.Insert(r); 631 } 632 } 633 } 634 } 635 636 637 //------------------------------sched_call------------------------------------- 638 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { 639 RegMask regs; 640 641 // Schedule all the users of the call right now. All the users are 642 // projection Nodes, so they must be scheduled next to the call. 643 // Collect all the defined registers. 644 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { 645 Node* n = mcall->fast_out(i); 646 assert( n->is_MachProj(), "" ); 647 int n_cnt = ready_cnt.at(n->_idx)-1; 648 ready_cnt.at_put(n->_idx, n_cnt); 649 assert( n_cnt == 0, "" ); 650 // Schedule next to call 651 block->map_node(n, node_cnt++); 652 // Collect defined registers 653 regs.OR(n->out_RegMask()); 654 // Check for scheduling the next control-definer 655 if( n->bottom_type() == Type::CONTROL ) 656 // Warm up next pile of heuristic bits 657 needed_for_next_call(block, n, next_call); 658 659 // Children of projections are now all ready 660 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 661 Node* m = n->fast_out(j); // Get user 662 if(get_block_for_node(m) != block) { 663 continue; 664 } 665 if( m->is_Phi() ) continue; 666 int m_cnt = ready_cnt.at(m->_idx)-1; 667 ready_cnt.at_put(m->_idx, m_cnt); 668 if( m_cnt == 0 ) 669 worklist.push(m); 670 } 671 672 } 673 674 // Act as if the call defines the Frame Pointer. 675 // Certainly the FP is alive and well after the call. 676 regs.Insert(_matcher.c_frame_pointer()); 677 678 // Set all registers killed and not already defined by the call. 679 uint r_cnt = mcall->tf()->range()->cnt(); 680 int op = mcall->ideal_Opcode(); 681 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); 682 map_node_to_block(proj, block); 683 block->insert_node(proj, node_cnt++); 684 685 // Select the right register save policy. 686 const char * save_policy; 687 switch (op) { 688 case Op_CallRuntime: 689 case Op_CallLeaf: 690 case Op_CallLeafNoFP: 691 // Calling C code so use C calling convention 692 save_policy = _matcher._c_reg_save_policy; 693 break; 694 695 case Op_CallStaticJava: 696 case Op_CallDynamicJava: 697 // Calling Java code so use Java calling convention 698 save_policy = _matcher._register_save_policy; 699 break; 700 701 default: 702 ShouldNotReachHere(); 703 } 704 705 // When using CallRuntime mark SOE registers as killed by the call 706 // so values that could show up in the RegisterMap aren't live in a 707 // callee saved register since the register wouldn't know where to 708 // find them. CallLeaf and CallLeafNoFP are ok because they can't 709 // have debug info on them. Strictly speaking this only needs to be 710 // done for oops since idealreg2debugmask takes care of debug info 711 // references but there no way to handle oops differently than other 712 // pointers as far as the kill mask goes. 713 bool exclude_soe = op == Op_CallRuntime; 714 715 // If the call is a MethodHandle invoke, we need to exclude the 716 // register which is used to save the SP value over MH invokes from 717 // the mask. Otherwise this register could be used for 718 // deoptimization information. 719 if (op == Op_CallStaticJava) { 720 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; 721 if (mcallstaticjava->_method_handle_invoke) 722 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); 723 } 724 725 add_call_kills(proj, regs, save_policy, exclude_soe); 726 727 return node_cnt; 728 } 729 730 731 //------------------------------schedule_local--------------------------------- 732 // Topological sort within a block. Someday become a real scheduler. 733 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) { 734 // Already "sorted" are the block start Node (as the first entry), and 735 // the block-ending Node and any trailing control projections. We leave 736 // these alone. PhiNodes and ParmNodes are made to follow the block start 737 // Node. Everything else gets topo-sorted. 738 739 #ifndef PRODUCT 740 if (trace_opto_pipelining()) { 741 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order); 742 for (uint i = 0;i < block->number_of_nodes(); i++) { 743 tty->print("# "); 744 block->get_node(i)->fast_dump(); 745 } 746 tty->print_cr("#"); 747 } 748 #endif 749 750 // RootNode is already sorted 751 if (block->number_of_nodes() == 1) { 752 return true; 753 } 754 755 // Move PhiNodes and ParmNodes from 1 to cnt up to the start 756 uint node_cnt = block->end_idx(); 757 uint phi_cnt = 1; 758 uint i; 759 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi 760 Node *n = block->get_node(i); 761 if( n->is_Phi() || // Found a PhiNode or ParmNode 762 (n->is_Proj() && n->in(0) == block->head()) ) { 763 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt 764 block->map_node(block->get_node(phi_cnt), i); 765 block->map_node(n, phi_cnt++); // swap Phi/Parm up front 766 } else { // All others 767 // Count block-local inputs to 'n' 768 uint cnt = n->len(); // Input count 769 uint local = 0; 770 for( uint j=0; j<cnt; j++ ) { 771 Node *m = n->in(j); 772 if( m && get_block_for_node(m) == block && !m->is_top() ) 773 local++; // One more block-local input 774 } 775 ready_cnt.at_put(n->_idx, local); // Count em up 776 777 #ifdef ASSERT 778 if( UseConcMarkSweepGC || UseG1GC ) { 779 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { 780 // Check the precedence edges 781 for (uint prec = n->req(); prec < n->len(); prec++) { 782 Node* oop_store = n->in(prec); 783 if (oop_store != NULL) { 784 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark"); 785 } 786 } 787 } 788 } 789 #endif 790 791 // A few node types require changing a required edge to a precedence edge 792 // before allocation. 793 if( n->is_Mach() && n->req() > TypeFunc::Parms && 794 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || 795 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { 796 // MemBarAcquire could be created without Precedent edge. 797 // del_req() replaces the specified edge with the last input edge 798 // and then removes the last edge. If the specified edge > number of 799 // edges the last edge will be moved outside of the input edges array 800 // and the edge will be lost. This is why this code should be 801 // executed only when Precedent (== TypeFunc::Parms) edge is present. 802 Node *x = n->in(TypeFunc::Parms); 803 n->del_req(TypeFunc::Parms); 804 n->add_prec(x); 805 } 806 } 807 } 808 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count 809 ready_cnt.at_put(block->get_node(i2)->_idx, 0); 810 811 // All the prescheduled guys do not hold back internal nodes 812 uint i3; 813 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled 814 Node *n = block->get_node(i3); // Get pre-scheduled 815 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 816 Node* m = n->fast_out(j); 817 if (get_block_for_node(m) == block) { // Local-block user 818 int m_cnt = ready_cnt.at(m->_idx)-1; 819 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count 820 } 821 } 822 } 823 824 Node_List delay; 825 // Make a worklist 826 Node_List worklist; 827 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist 828 Node *m = block->get_node(i4); 829 if( !ready_cnt.at(m->_idx) ) { // Zero ready count? 830 if (m->is_iteratively_computed()) { 831 // Push induction variable increments last to allow other uses 832 // of the phi to be scheduled first. The select() method breaks 833 // ties in scheduling by worklist order. 834 delay.push(m); 835 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { 836 // Force the CreateEx to the top of the list so it's processed 837 // first and ends up at the start of the block. 838 worklist.insert(0, m); 839 } else { 840 worklist.push(m); // Then on to worklist! 841 } 842 } 843 } 844 while (delay.size()) { 845 Node* d = delay.pop(); 846 worklist.push(d); 847 } 848 849 // Warm up the 'next_call' heuristic bits 850 needed_for_next_call(block, block->head(), next_call); 851 852 #ifndef PRODUCT 853 if (trace_opto_pipelining()) { 854 for (uint j=0; j< block->number_of_nodes(); j++) { 855 Node *n = block->get_node(j); 856 int idx = n->_idx; 857 tty->print("# ready cnt:%3d ", ready_cnt.at(idx)); 858 tty->print("latency:%3d ", get_latency_for_node(n)); 859 tty->print("%4d: %s\n", idx, n->Name()); 860 } 861 } 862 #endif 863 864 uint max_idx = (uint)ready_cnt.length(); 865 // Pull from worklist and schedule 866 while( worklist.size() ) { // Worklist is not ready 867 868 #ifndef PRODUCT 869 if (trace_opto_pipelining()) { 870 tty->print("# ready list:"); 871 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 872 Node *n = worklist[i]; // Get Node on worklist 873 tty->print(" %d", n->_idx); 874 } 875 tty->cr(); 876 } 877 #endif 878 879 // Select and pop a ready guy from worklist 880 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt); 881 block->map_node(n, phi_cnt++); // Schedule him next 882 883 #ifndef PRODUCT 884 if (trace_opto_pipelining()) { 885 tty->print("# select %d: %s", n->_idx, n->Name()); 886 tty->print(", latency:%d", get_latency_for_node(n)); 887 n->dump(); 888 if (Verbose) { 889 tty->print("# ready list:"); 890 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist 891 Node *n = worklist[i]; // Get Node on worklist 892 tty->print(" %d", n->_idx); 893 } 894 tty->cr(); 895 } 896 } 897 898 #endif 899 if( n->is_MachCall() ) { 900 MachCallNode *mcall = n->as_MachCall(); 901 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); 902 continue; 903 } 904 905 if (n->is_Mach() && n->as_Mach()->has_call()) { 906 RegMask regs; 907 regs.Insert(_matcher.c_frame_pointer()); 908 regs.OR(n->out_RegMask()); 909 910 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); 911 map_node_to_block(proj, block); 912 block->insert_node(proj, phi_cnt++); 913 914 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); 915 } 916 917 // Children are now all ready 918 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { 919 Node* m = n->fast_out(i5); // Get user 920 if (get_block_for_node(m) != block) { 921 continue; 922 } 923 if( m->is_Phi() ) continue; 924 if (m->_idx >= max_idx) { // new node, skip it 925 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types"); 926 continue; 927 } 928 int m_cnt = ready_cnt.at(m->_idx)-1; 929 ready_cnt.at_put(m->_idx, m_cnt); 930 if( m_cnt == 0 ) 931 worklist.push(m); 932 } 933 } 934 935 if( phi_cnt != block->end_idx() ) { 936 // did not schedule all. Retry, Bailout, or Die 937 if (C->subsume_loads() == true && !C->failing()) { 938 // Retry with subsume_loads == false 939 // If this is the first failure, the sentinel string will "stick" 940 // to the Compile object, and the C2Compiler will see it and retry. 941 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 942 } 943 // assert( phi_cnt == end_idx(), "did not schedule all" ); 944 return false; 945 } 946 947 #ifndef PRODUCT 948 if (trace_opto_pipelining()) { 949 tty->print_cr("#"); 950 tty->print_cr("# after schedule_local"); 951 for (uint i = 0;i < block->number_of_nodes();i++) { 952 tty->print("# "); 953 block->get_node(i)->fast_dump(); 954 } 955 tty->cr(); 956 } 957 #endif 958 959 960 return true; 961 } 962 963 //--------------------------catch_cleanup_fix_all_inputs----------------------- 964 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { 965 for (uint l = 0; l < use->len(); l++) { 966 if (use->in(l) == old_def) { 967 if (l < use->req()) { 968 use->set_req(l, new_def); 969 } else { 970 use->rm_prec(l); 971 use->add_prec(new_def); 972 l--; 973 } 974 } 975 } 976 } 977 978 //------------------------------catch_cleanup_find_cloned_def------------------ 979 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 980 assert( use_blk != def_blk, "Inter-block cleanup only"); 981 982 // The use is some block below the Catch. Find and return the clone of the def 983 // that dominates the use. If there is no clone in a dominating block, then 984 // create a phi for the def in a dominating block. 985 986 // Find which successor block dominates this use. The successor 987 // blocks must all be single-entry (from the Catch only; I will have 988 // split blocks to make this so), hence they all dominate. 989 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) 990 use_blk = use_blk->_idom; 991 992 // Find the successor 993 Node *fixup = NULL; 994 995 uint j; 996 for( j = 0; j < def_blk->_num_succs; j++ ) 997 if( use_blk == def_blk->_succs[j] ) 998 break; 999 1000 if( j == def_blk->_num_succs ) { 1001 // Block at same level in dom-tree is not a successor. It needs a 1002 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. 1003 Node_Array inputs = new Node_List(Thread::current()->resource_area()); 1004 for(uint k = 1; k < use_blk->num_preds(); k++) { 1005 Block* block = get_block_for_node(use_blk->pred(k)); 1006 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); 1007 } 1008 1009 // Check to see if the use_blk already has an identical phi inserted. 1010 // If it exists, it will be at the first position since all uses of a 1011 // def are processed together. 1012 Node *phi = use_blk->get_node(1); 1013 if( phi->is_Phi() ) { 1014 fixup = phi; 1015 for (uint k = 1; k < use_blk->num_preds(); k++) { 1016 if (phi->in(k) != inputs[k]) { 1017 // Not a match 1018 fixup = NULL; 1019 break; 1020 } 1021 } 1022 } 1023 1024 // If an existing PhiNode was not found, make a new one. 1025 if (fixup == NULL) { 1026 Node *new_phi = PhiNode::make(use_blk->head(), def); 1027 use_blk->insert_node(new_phi, 1); 1028 map_node_to_block(new_phi, use_blk); 1029 for (uint k = 1; k < use_blk->num_preds(); k++) { 1030 new_phi->set_req(k, inputs[k]); 1031 } 1032 fixup = new_phi; 1033 } 1034 1035 } else { 1036 // Found the use just below the Catch. Make it use the clone. 1037 fixup = use_blk->get_node(n_clone_idx); 1038 } 1039 1040 return fixup; 1041 } 1042 1043 //--------------------------catch_cleanup_intra_block-------------------------- 1044 // Fix all input edges in use that reference "def". The use is in the same 1045 // block as the def and both have been cloned in each successor block. 1046 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { 1047 1048 // Both the use and def have been cloned. For each successor block, 1049 // get the clone of the use, and make its input the clone of the def 1050 // found in that block. 1051 1052 uint use_idx = blk->find_node(use); 1053 uint offset_idx = use_idx - beg; 1054 for( uint k = 0; k < blk->_num_succs; k++ ) { 1055 // Get clone in each successor block 1056 Block *sb = blk->_succs[k]; 1057 Node *clone = sb->get_node(offset_idx+1); 1058 assert( clone->Opcode() == use->Opcode(), "" ); 1059 1060 // Make use-clone reference the def-clone 1061 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); 1062 } 1063 } 1064 1065 //------------------------------catch_cleanup_inter_block--------------------- 1066 // Fix all input edges in use that reference "def". The use is in a different 1067 // block than the def. 1068 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { 1069 if( !use_blk ) return; // Can happen if the use is a precedence edge 1070 1071 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); 1072 catch_cleanup_fix_all_inputs(use, def, new_def); 1073 } 1074 1075 //------------------------------call_catch_cleanup----------------------------- 1076 // If we inserted any instructions between a Call and his CatchNode, 1077 // clone the instructions on all paths below the Catch. 1078 void PhaseCFG::call_catch_cleanup(Block* block) { 1079 1080 // End of region to clone 1081 uint end = block->end_idx(); 1082 if( !block->get_node(end)->is_Catch() ) return; 1083 // Start of region to clone 1084 uint beg = end; 1085 while(!block->get_node(beg-1)->is_MachProj() || 1086 !block->get_node(beg-1)->in(0)->is_MachCall() ) { 1087 beg--; 1088 assert(beg > 0,"Catch cleanup walking beyond block boundary"); 1089 } 1090 // Range of inserted instructions is [beg, end) 1091 if( beg == end ) return; 1092 1093 // Clone along all Catch output paths. Clone area between the 'beg' and 1094 // 'end' indices. 1095 for( uint i = 0; i < block->_num_succs; i++ ) { 1096 Block *sb = block->_succs[i]; 1097 // Clone the entire area; ignoring the edge fixup for now. 1098 for( uint j = end; j > beg; j-- ) { 1099 // It is safe here to clone a node with anti_dependence 1100 // since clones dominate on each path. 1101 Node *clone = block->get_node(j-1)->clone(); 1102 sb->insert_node(clone, 1); 1103 map_node_to_block(clone, sb); 1104 } 1105 } 1106 1107 1108 // Fixup edges. Check the def-use info per cloned Node 1109 for(uint i2 = beg; i2 < end; i2++ ) { 1110 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block 1111 Node *n = block->get_node(i2); // Node that got cloned 1112 // Need DU safe iterator because of edge manipulation in calls. 1113 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); 1114 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { 1115 out->push(n->fast_out(j1)); 1116 } 1117 uint max = out->size(); 1118 for (uint j = 0; j < max; j++) {// For all users 1119 Node *use = out->pop(); 1120 Block *buse = get_block_for_node(use); 1121 if( use->is_Phi() ) { 1122 for( uint k = 1; k < use->req(); k++ ) 1123 if( use->in(k) == n ) { 1124 Block* b = get_block_for_node(buse->pred(k)); 1125 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); 1126 use->set_req(k, fixup); 1127 } 1128 } else { 1129 if (block == buse) { 1130 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); 1131 } else { 1132 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); 1133 } 1134 } 1135 } // End for all users 1136 1137 } // End of for all Nodes in cloned area 1138 1139 // Remove the now-dead cloned ops 1140 for(uint i3 = beg; i3 < end; i3++ ) { 1141 block->get_node(beg)->disconnect_inputs(NULL, C); 1142 block->remove_node(beg); 1143 } 1144 1145 // If the successor blocks have a CreateEx node, move it back to the top 1146 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { 1147 Block *sb = block->_succs[i4]; 1148 uint new_cnt = end - beg; 1149 // Remove any newly created, but dead, nodes. 1150 for( uint j = new_cnt; j > 0; j-- ) { 1151 Node *n = sb->get_node(j); 1152 if (n->outcnt() == 0 && 1153 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ 1154 n->disconnect_inputs(NULL, C); 1155 sb->remove_node(j); 1156 new_cnt--; 1157 } 1158 } 1159 // If any newly created nodes remain, move the CreateEx node to the top 1160 if (new_cnt > 0) { 1161 Node *cex = sb->get_node(1+new_cnt); 1162 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { 1163 sb->remove_node(1+new_cnt); 1164 sb->insert_node(cex, 1); 1165 } 1166 } 1167 } 1168 }